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
1801.06703	1	1801	2018-01-20T17:37:25	Decision Rules for Robotic Mobile Fulfillment Systems	[ "cs.MA" ]	The Robotic Mobile Fulfillment Systems (RMFS) is a new type of robotized, parts-to-picker material handling system, designed especially for e-commerce warehouses. Robots bring movable shelves, called pods, to workstations where inventory is put on or removed from the pods. This paper simulates both the pick and replenishment process and studies the order assignment, pod selection and pod storage assignment problems by evaluating multiple decision rules per problem. The discrete event simulation uses realistic robot movements and keeps track of every unit of inventory on every pod. We analyze seven performance measures, e.g. throughput capacity and order due time, and find that the unit throughput is strongly correlated with the other performance measures. We vary the number of robots, the number of pick stations, the number of SKUs (stock keeping units), the order size and whether returns need processing or not. The decision rules for pick order assignment have a strong impact on the unit throughput rate. This is not the case for replenishment order assignment, pod selection and pod storage. Furthermore, for warehouses with a large number of SKUs, more robots are needed for a high unit throughput rate, even if the number of pods and the dimensions of the storage area remain the same. Lastly, processing return orders only affects the unit throughput rate for warehouse with a large number of SKUs and large pick orders.	cs.MA	cs	Decision Rules for Robotic Mobile Fulfillment Systems M. Merschformann T. Lamballais M.B.M. de Koster L. Suhl January 23, 2018 Abstract The Robotic Mobile Fulfillment Systems (RMFS) is a new type of robotized, parts-to-picker material handling system, designed especially for e-commerce warehouses. Robots bring movable shelves, called pods, to workstations where inventory is put on or removed from the pods. This paper simulates both the pick and replenishment process and stud- ies the order assignment, pod selection and pod storage assignment problems by evaluating multiple decision rules per problem. The dis- crete event simulation uses realistic robot movements and keeps track of every unit of inventory on every pod. We analyze seven performance measures, e.g. throughput capacity and order due time, and find that the unit throughput is strongly correlated with the other performance measures. We vary the number of robots, the number of pick stations, the number of SKUs (stock keeping units), the order size and whether returns need processing or not. The decision rules for pick order assign- ment have a strong impact on the unit throughput rate. This is not the case for replenishment order assignment, pod selection and pod storage. Furthermore, for warehouses with a large number of SKUs, more robots are needed for a high unit throughput rate, even if the number of pods and the dimensions of the storage area remain the same. Lastly, pro- cessing return orders only affects the unit throughput rate for warehouse with a large number of SKUs and large pick orders. 1 Introduction The rise of e-commerce has created the need for new warehousing systems. Traditional, manual picker-to-parts systems work best when orders are large, i.e. consist of many SKUs so that consolidation has to be organized well. How- ever, e-commerce orders are typically small and e-commerce warehouses are often large as they need to contain large assortments of products, which results in long walking distances for the pickers. In contrast to manual picker-to-part 1 systems, automated parts-to-picker systems eliminate the time pickers spend traveling. Thus, they can achieve higher pick rates. The Robotic Mobile Fulfillment System (RMFS) is an automated parts-to- picker system. Robots transport movable shelves, called "pods", that contain the inventory, back and forth between the storage area and the workstations. As RMFSs eliminate picker walking time, high pick rates can be expected. Implementations suggest that pick rates can improve substantially compared to manual picker-to-parts operations, see also [17]. The systems are mainly used by Amazon, which bought the company that invented the RMFS, Kiva Systems, and has since deployed it in its warehouses (see [3]). Recently, com- petitors such as Swisslog, Interlink, GreyOrange, Mobile Industrial Robots and Scallog have been rolling out their versions of an RMFS. The RMFS is described in more detail in [5] and [18]. They mention that numerous operational decisions problems are yet to be examined in depth, for example the assignment of customer orders to workstations or of pods to a storage locations. Each of these decision problems comes with a trade-off. An order may be assigned to a workstation if it is nearing its due time, but as- signing another order that has lines in common with other orders assigned to that workstation may result in more picks per pod and hence a reduction in the number of pod trips. Furthermore, assigning a pod to a storage location that is close to the workstation reduces travel time, but keeping the inventory sorted by assigning pods to favorable storage location if they are likely to be needed in the near future may reduce travel times more. These trade-offs are linked to the number of robots in the system. As an example, with more robots, more trips can be done and hence the order due times can become a more important criterion than the number of picks per pod when selecting a pod to be transported to a workstation. The trade-offs are also linked to the resources and conditions in the warehouse. For example, the more SKUs a warehouse contains, the more difficult it becomes to assign orders to pick stations in such a way that multiple products can be picked from a single pod. As these examples indicate, a need exists for finding methods to address the de- cision problems in an RMFS, for research on the performance of RMFSs across performance measures, and for examining performance while varying aspects like the number of robots. This paper addresses this need. We study the pick order assignment, replenishment order assignment, pick pod selection, re- plenishment pod selection, and pod storage assignment decision problems and propose several decision rules for each. To see which trade-offs in performance may exist, we use different performance measures. Furthermore, we vary three aspects of the RMFS, namely whether or not return orders need to be pro- cessed, the size of the orders, and the number of SKUs in the warehouse. This study focuses on both the pick process and the replenishment process, because a more efficient replenishment process frees up robots for pick tasks. Lastly, the number of pick stations and the number of robots per pick station is varied. 2 Varying these numbers shows how many pick stations and robots are needed to provide pickers with a near continuous supply of pods. Section 2 describes the RMFS in more detail, Section 3 points out related work, Section 4 the decision problems, Section 5 the decision rules, and Sec- tion 6 describes the realistic simulation built for evaluating the decision rules, while Section 7 explains the evaluation framework, Section 8 shows the results of the analysis, and Section 9 provides conclusions and directions for future research. 2 The Robotic Mobile Fulfillment System An RMFS consists of shelves on which products are stored (called pods), robots that can move underneath and also carry them (see Figure 1a), and work stations. After handling a pod at a station it can be returned to a different storage location than where it was retrieved from, hence, inventory can be sorted continuously throughout the day. (a) Robot carrying a pod (see [5]) (b) The internal storage / retrieval process in RMFSs (red: robot & pod movement Figure 1: The essential elements of an RMFS Figure 1b shows the storage and retrieval processes, where the robots transport pods between the workstations and the storage area. Starting at the replen- ishment station, in the example, two replenishment orders with 4 and 8 units of two SKUs (green & orange) are stored on a pod that was retrieved from the inventory by a robot. The blue SKU also relevant to the process is already available on the pod in focus. After the pod was handled at the station it is stored in inventory again. Next, if the pod is selected for picking at a pick station, it is brought to that station. The operator at the station then picks the units matching the open order lines at the station from the pod and puts them into the bins for the respective pick orders. As soon as a pick order is completed it leaves the pick station and is handled by further warehouse sys- tems. If zoning is in place at the warehouse, the pick order may only be a part of a larger customer order and must be consolidated further with the other 3 ReplenishmentStorage tripInventoryRetrieval tripOrder pickingRetrieval tripStorage trip2 replenishmentorders received1 pick ordercompletedpartial pick orders in a following sortation process. If the customer order is already completely fulfilled at the pick station, it may be packed into a carton and prepared for shipping immediately with no further handling. The latter may only be possible in e-commerce operations where lines per order are small. Each pair of storage and retrieval trip is one robot cycle in an RMFS. During one cycle the robot does not set-down or leave the rack until it is returned to a storage location. Note that, the pod may be brought to further replenishment or pick stations between the retrieval and the storage trip, if further replen- ishment or immediate picking can be done with it. For the sake of clarity we limited the visits per cycle to one station in the example above. While the operation of the robot is cyclic the flow of the inventory units through the system starts at a replenishment station (by storing a replenishment order) and exits at a pick station (by fulfilling a pick order). However, in contrast to other systems there is quite some overhead inventory movement, because all contained units, not only needed ones, are moved when a pod is brought to a station. The same happens during replenishment operations, if non-empty pods are moved to a replenishment station. Robots navigate their paths through the warehouse using a waypoint system, which is laid out as a grid. A path is a sequence of connected waypoints and all robots have to be guided concurrently along their paths while avoiding collisions and deadlocks. Robots that are not carrying a pod can move under- neath stationary pods and hence take other paths than robots that do carry pods, because the latter cannot use occupied storage locations. The system layout is depicted in Figure 2 and consists of a storage area where the pods are stored, pick and replenishment stations grouped around the storage area, maneuvering areas between the storage area and the workstations, and per workstation a buffer area. A robot carries a pod from the storage area, via the maneuvering area, to the buffer area of the destination workstation. Only one pod is picked or replenished simultaneously. Workers at the replenishment stations replenish the pods with new inventory. In contrast, workers at the pick stations pick product units to fulfill orders. A picker picks for multiple unfinished/incomplete pick orders at the same time. For both operations the robots need to stop with a pod at a waypoint representing the access point of the respective station. In the buffer area next to each workstation, robots carrying pods can wait for their turn. In the middle of the layout a number of waypoints is used as possible storage locations where pods can be put when they are not used. Every storage location is directly reachable from an aisle and access to a storage location cannot be blocked by stored pods. Travel in the aisles is single-directional to avoid gridlock and reduce congestion. The system has the ability to adapt to changing demand conditions. E.g., if order arrival rates of some SKUs drop, pods containing those SKUs can be relocated further away from the pick stations. This relocation frees up storage locations near the pick stations for pods containing SKUs with high order ar- 4 rival rates. Pods can be relocated when returning from a workstation, hence the inventory can be continually sorted in response to changing demand. Figure 2: A top view of an RMFS layout 3 Related Work To this date no detailed discrete event simulation based research has been done for RMFS. Moreover, most research on RMFSs to date uses queueing networks to study design questions on the strategic level. This work aims to close the gap by delivering insights about RMFS using a very detailed simula- tion framework that integrates most dynamic effects an operator faces. Next, we first outline the queuing network based research and close this section with simulation based work. [14] create queueing networks similar to earlier queueing networks used for autonomous vehicle storage and retrieval systems (AVS/RS) and automated storage and retrieval systems (AS/RS) (see [8] and [16]). Their queueing net- works capture both pick and replenishment operations but cannot model robot movement realistically. They estimate the order throughput time for single-line orders. [10] create a different queueing network for both single- and multi-line orders, with and without zoning in the storage area, that captures only the pick operations, but that does include realistic robot movement. Their model can accurately estimate the expected order cycle time, workstation utilization and robot utilization. [10] determine how the storage area dimensions and the workstation placement around the storage area affect the maximum order throughput, by evaluating a large number of possible designs. [11] develop a queueing network that addresses problems on a tactical level. They show the 5 bu(cid:31)er area pathsmaneuvering area pathsstorage area pathsrobotpodreplenishment stationpick stationwaypointeffect of the number of pods per SKU and of the replenishment level of a pod on order throughput, and they show what the optimal ratio of the number of pick stations to the number of replenishment stations is. They find that it is better to replenish pods before they are entirely empty, even with multiple pods per SKU. [19] use semi-open queueing networks to analyze the policy for assigning robots to pick stations. The authors find that the random policy is significantly outperformed by the proposed handling-speeds-based assignment rule when facing varying service rates of the pickers. [20] build a semi-open queueing network for evaluating the effects of battery management in RMFS. The strategies of battery swapping, automated plug-in charging and inductive charging at the pick station are compared. The authors come to the conclusion that battery swapping is generally more expensive than plug-in charging while inductive charging outperforms both in throughput and costs, if robot prices and retrieval times are low. [5] and [18] mention several decision problems on the operational level that they encountered in practice. One of the few studies that address decision problems on the operational level is by [2]. They provide methods for opti- mally batching the pick orders and sequencing both the pick orders and the pods transported to the stations. They show that an optimized pick order processing requires only half the number of robots that a pick order process based on simple decision rules would need. [15] utilize a simulation based ap- proach in order to optimize the warehouse layout of a manual order picking system for an industrial partner. The authors devise an integrated approach taking on to certain design decisions as well as selecting control policies. The simulation is thereby used "as a solution tool and an evaluation system" (see [15]). [4] use a simulation based approach for evaluating the performance of policy sets for manual order picking systems. The authors make use of DEA as a tool for obtaining a comparable performance indicator among the policy sets. [1] use a discrete event simulation approach similar to this work for assessing storage policies for Automated Grid-based Storage systems. The authors find that even simple strategies improve the system efficiency, which encourages research on more complex strategies. [9] develop a Markow decision process (MDP) model for addressing the resource reallocation problem, i.e., the prob- lem of deciding how many workers and robots to allocate to the pick process and replenishment process continually throughout time. The assumptions re- lated to replenishment differ strongly across the papers mentioned above, and the number of approaches to replenishment in practical applications is diverse as well. 4 Decision Problems This section introduces the decision problems considered in this paper and places them within the context of other decision problems in an RMFS. Re- quests to the system occur via pick orders or replenishment orders. Upon 6 Figure 3: Hierarchical overview of the decision problems and their relations receipt, pallets are broken up into smaller parts consisting of multiple units of one SKU. A replenishment order is a request to place one such part, i.e. a number of units of one specific product, on a pod. We structure the decisions at the operational level in four steps: (1) Order Assignment (OA), the assignment of pick or replenishment orders to worksta- tions, (2) Task Creation (TC), the creation of tasks for the robots, (3) Task Allocation (TA), the allocation of tasks to robots, and (4) Path Planning (PP), the creation of paths along which the robots will move. There are two kinds of Order Assignment decisions: the assignment of pick orders to pick stations, called the Pick Order Assignment (POA) problem, and the assignment of re- plenishment orders to replenishment stations, called the Replenishment Order Assignment (ROA) problem. In the second step, a task is defined as trans- porting a specific pod to a specific workstation and back to a specific storage 7 StrategicLevel-DecisionsattimeofwarehouseconstructionTacticalLevel-Decisionsatstartworkweek,workdayorshiftOperationalLevel-Decisionsinreal-time(1)OrderAssignment(OA)(2)TaskCreation(TC)(3)TaskAllocation(TA)(4)PathPlanning(PP)WorkstationPlacementStorageAreaDimensioningReplenishmentPodSelection(RPS)(2.2)PodStorageAssignment(PSA)PickOrderAssignment(POA)ReplenishmentOrderAssignment(ROA)RatiobetweenNo.ofpickandNo.ofreplenishmentstationsResourceReallocationNumberofPodsperSKUReplenishmentLevel(2.1)PodSelection(PS)PickPodSelection(PPS)location. Therefore, for each workstation, the Task Creation decision prob- lem includes the two subproblems of (2.1) deciding which pod to select for transportation, the Pod Selection (PS) decision problem, and (2.2) deciding at which storage location to return the pod, the Pod Storage Assignment (PSA) decision problem. The Pod Selection (PS) decision problem differs for the pick and replenishment process, because for the pick process the due times of the pick orders is important in selecting a pod. Pod selection in the pick pro- cess is called Pick Pod Selection (PPS) and pod selection in the replenishment process is called Replenishment Pod Selection (RPS). Task Creation uses the pick order and replenishment order assignments to select suitable pods and subsequently converts the requests for the selected pods into tasks for pod transportation between the workstations and the storage area. Task Alloca- tion creates a trip by building a sequence of tasks for the robots to execute. These sequenced tasks implicitly define trips and serve as input for the Path Planning algorithms, where a path is generated for a robot to follow. Figure 3 shows an overview of the decision problems at the strategic, tactical and operational level in an RMFS, with the problems addressed in this paper in bold. As can be seen in Figure 3, this paper focuses on decision prob- lems at the operational level. We use the term "decision rule" to refer to a fairly simple method to solve a decision problem. The aim of this paper is to evaluate several decision rules per decision problem. Some decision rules may closely resemble common best practices, whereas others may be more specific to RMFS. The Task Allocation decision problem is intertwined with the Path Planning decision problem, which has been addressed by [12]. Therefore we do not consider the Task Allocation and Path Planning decision problems. We do address Pick Order Assignment (POA), Replenishment Order Assignment (ROA), Pick Pod Selection (PPS), Replenishment Pod Selection (RPS), and Pod Storage Assignment (PSA). For Pick Order Assignment, we assume there is a constant backlog, and the pick stations are always filled to full capacity with pick orders. Whenever a pick order is fulfilled and leaves its pick sta- tion, a pick order has to be selected from the backlog and assigned to the pick station. For replenishment orders, we assume that the sequence of re- plenishment orders inbound to the system cannot be altered anymore. This assumption resembles the situation in conventional conveyor-based material handling components that do not allow sequence modification but only load routing. Moreover, we aim to avoid taking decision problems outside of the system's boundaries into account, e.g., different dispatching rules of preceding systems. The replenishment stations have a finite capacity. If a replenishment order arrives and multiple replenishment station have capacity left, the ROA decision rule determines to which replenishment station the replenishment or- der is assigned. If no place is available, replenishment orders are put in a replenishment order backlog. When a replenishment order is fulfilled at one of the replenishment stations, a new replenishment order is chosen from the replenishment order backlog according to the FCFS rule. Table 1 summarizes 8 the decision problems addressed in this paper. At this point we also introduce the concept of "pile-on" (sometimes also called "hit-rate"). Pile-on as a concept refers to the average number of units that are picked from a pod every time a pod is presented to a picker at a pick station. Pile-on as a metric measures the number of units (across all SKUs) picked from a pod when presented to a picker at a pick station, averaged across every visit of a pod to a pick station during the entire time horizon. In other words, pile-on is measured in "units picked per pod visit to a pick station". The higher the pile-on is, the fewer pods need to be transported between the pick stations and the storage area, which may reduce the number of robots needed. Table 1: Decision Problems Abb. Name POA Pick Order Assignment Description Choosing a pick order from the backlog ROA Replenishment Order Assignment PPS Pick Pod Selection RPS Replenishment Pod Selection PSA Pod Storage Assignment Selecting the replenishment station for the next replen- ishment order Selecting a pod to transport to a pick station Select a pod for the next re- plenishment order Choosing a storage location for a pod Trigger When another pick order is fulfilled and leaves the pick station, creating room for the next pick order to be assigned When a replenishment order arrives at the system and one or more replenish- ment stations have capacity left When a robot working for a pick station needs a new task Depends on the ROA decision rule When a pod leaves a workstation 5 Decision Rules To solve the operational problems, we define several decision rules per decision problem that are evaluated in a realistic simulation. Several Path Planning algorithms for the RMFS are compared in [12], therefore this decision prob- lem will not be addressed in this paper. Thus, we selected WHCA∗ v, one of the best performing algorithms from the paper, as the path planning engine for this work. Additionally, we fix the Task Allocation algorithm to a simple method that first assigns two-thirds of the robots to pick operations and the rest to replenishment operations. Then, it aims to equally distribute the robots across the respective stations. This means a robot will only do tasks related to the station it is assigned to. This section will therefore only describe decision rules for the Pick Order Assignment, Replenishment Order Assignment, Pick Pod Selection, Replenishment Pod Selection and Pod Storage Assignment de- cision problems. While replenishment and pick operations are similar in the sense that high 9 throughput should be achieved with few resources, the main asymmetry be- tween both is that for the former the goal is to fill the inventory as quickly as possible and for the latter to empty it as quickly as possible. This means that for replenishment operations we aim to replenish pods fast to have them available for pick operations early while preparing pod content such that it allows for a high pile-on during pick operations. For pick operations we aim to achieve a high pile-on and keeping trips short to fulfill as many orders as pos- sible while also considering due times of the pick orders. Furthermore, we do not allow the sequence of replenishment orders to be modified. In contrast, for pick orders we allow to arbitrarily choose one order from the backlog. Lastly, pick orders have due times. All of this leads to different strategies we focus on per decision problem, instead of fully symmetric rules between pick and replenishment decision problems. For a more precise description of some of the rules we introduce the notation shown in Table 2. Table 2: Overview of the symbols used in the rule descriptions s s Required units necessary to fulfill line i of order o Set of all pods Set of pods heading to station s Set of all SKUs Set of pick orders in backlog Set of pick orders assigned to station s Symbol Explanation P P I I OB OS C(p, i) Number of units of SKU i contained in pod p L(o, i) D(o, i) Remaining units necessary to fulfill line i of order o tD o tS o t Due time of order o Time of assignment to the station of order o Time of deciding 5.1 Pick Order Assignment Rules A pick station has to be chosen for every pick order submitted to the system and the pick order itself has to be chosen from the order backlog. In this work, we consider a pick order backlog of constant size, i.e., as soon as an order is removed from the backlog a new one is generated to replace it. This and the immediate replacement of orders completed at a station lead to only one option available to assign any pick order to: the slot of the just completed order. Hence, the choice of station is not a degree of freedom in this work. The rare occasions of multiple orders to be completed at the same time are handled by assigning the orders to the pick stations randomly. Hence, we only investigate rules for selecting the next pick order from the backlog to fill the only open slot 10 at a station. We devise six rules to solve this problem: "Random", "FCFS", "Due-Time", "Fast-Lane", "Common-Lines" and "Pod-Match": Random The Random rule randomly selects a next pick order from the back- log and is used as a benchmark. FCFS The FCFS rule assigns the pick order that was first received. The rationale behind this is to keep pick order throughput times short. Due-Time The Due-Time rule selects the pick order with the earliest due time from the backlog and assigns it to a station. This is a greedy approach aiming to finish the pick orders before their deadline. Fast-Lane The Fast-Lane rule randomly selects a pick order from the back- log like the Random rule, but keeps one slot at each pick station open for immediately completeable pick orders. I.e., only pick orders (o), for whom all lines and all units of inventory are available on the next pod (pn) will be assigned to this station's "fast-lane" order slot (see Equa- tion (1)). Thus, orders assigned to the "fast-lane" slot are processed shortly after assignment. The next pod of the station is either a not completely processed pod the picker is currently working on or the next pod in the station's queue, if no such pod is available. In cases where no pod reached the station's queue yet, we consider the pod with the shortest remaining path to estimate the next pod. When facing multiple options we use a random tie-breaker. Note that this rule can be com- bined with any other proposed POA rule. The reason we combine it with random selection is to better assess the impact of the idea itself. ∀i ∈ I : L(o, i) ≤ C(pn, i) (1) Common-Lines The Common-Lines rule compares the station's (s) currently assigned pick orders with all orders from the backlog and selects the one with most lines in common for assignment (see Equation (2)). The rationale behind this is to increase pile-on by exploiting synergies among the pick orders. When facing multiple options we use a random tie- breaker. (cid:33) 1 L(o, i) > 0 ∧ L(o(cid:48), i) > 0 0 otherwise (2) (cid:32)(cid:40) (cid:88) (cid:88) o(cid:48)∈OS s i∈I argmax o∈OB Pod-Match The Pod-Match rule selects the pick order from the backlog that matches best the pods heading to the station (s) at the moment of assign- ment best. I.e., the more units of the pick order are already available in the pods the better the match (see Equation (3)). When facing multiple options we use a random tie-breaker. (min (C(p, i), D(o, i))) (3) argmax o∈OB (cid:88) (cid:88) p∈P I s i∈I 11 5.2 Replenishment Order Assignment Rules As a result of the assumptions that replenishment orders arrive in a fixed sequence, we investigate only two different approaches for assigning replenish- ment orders to the stations, i.e., immediate Random assignment and batching of customer orders that go on the same pod. Hence, we construct two rules for replenishment assignment: "Random" and "Pod-Batch": Random The Random rule randomly selects a next station with sufficient remaining capacity to allocate incoming replenishment orders to. If no such station is available, the order will wait until one becomes available again. Pod-Batch The Pod-Batch rule tries to use a pod already selected to go to a replenishment station for assigning the next replenishment order. In other words, the Pod-Batch rule first waits for the Replenishment Pod Selection (Section 5.4) rule to decide which orders are assigned to which pod, and then uses the same replenishment station for the orders of one pod. If the replenishment orders do not fit one station, they wait until a station with sufficient capacity becomes available. During this time all consecutive orders are also blocked, because the sequence cannot be altered. 5.3 Pick Pod Selection Rules Every time a robot working for a pick station s requests a next task, a pod suitable for picking at pick station s must be selected. We require for all rules that at least one unit can be picked from the pod. This means that no pod is brought to a station completely in vain and additionally it implies a pile-on of at least 1. The six PPS rules used in this paper are the "Random", "Nearest", "Pile-on", "Demand", "Lateness", and "Age" rules: Random The Random rule randomly selects a pod that offers at least one useful unit for picking. Nearest The Nearest rule selects the pod which has the least estimated path time towards the station according to the path planning algorithm and that offers at least one useful unit for picking. Pile-on The Pile-on rule selects the pod that offers most units necessary to fulfill the orders at the station (see Equation (4)). Ties are broken by favoring pods with which more orders can be completed. If ties still persist, they are broken randomly. argmax p∈P i∈I o∈OS s (min (C(p, i), D(o, i))) (4) (cid:88) (cid:88) 12 Demand The Demand rule selects the pod whose content is most demanded considering the current pick order backlog situation, i.e. the pod with most units demanded in the backlog is chosen (see Equation (5)). Ties are broken randomly. min (C(p, i), D(o, i)) (5) (cid:88) (cid:88) i∈I o∈OB argmax p∈P (6) (7) Lateness The Lateness rule aims to finish late pick orders by selecting a pod that offers units needed to fulfill open order lines with most lateness at the station, i.e., for one order the time the order is late is summed as fractions of the open picks (see Equation (6)). If no order is late, the resulting ties are broken by using the same metric but replacing o , thus, selecting pods for orders whose due times max(cid:0)t − tD (cid:18)min (C(p, i), D(o, i)) (cid:80) i(cid:48)∈I D(o, i(cid:48)) o , 0(cid:1)(cid:19) max(cid:0)t − tD are most imminent. o , 0(cid:1) with tD (cid:88) (cid:88) argmax p∈P i∈I o∈OS s Age The Age rule aims to finish the oldest pick orders of a station by selecting a pod that offers units needed to fulfill the oldest open order lines, i.e. for one order the time the order spent assigned to the station is summed as fractions of the open picks (see Equation (7)) (cid:88) (cid:88) i∈I o∈OS s (cid:18)min (C(p, i), D(o, i)) (cid:80) i(cid:48)∈I D(o, i(cid:48)) (cid:1)(cid:19) (cid:0)t − tS o argmax p∈P 5.4 Replenishment Pod Selection Rules For every replenishment order, a suitable pod with sufficient remaining storage capacity needs to be chosen. The decision is taken right before the replenish- ment order is assigned to a replenishment station. Depending on the selected ROA and RPS rules both are either invoked simultaneously or, if there is a de- pendency between the two, one after the other. An example for the latter case is the combination of the PodBatch ROA rule with the Emptiest RPS rule, because the PodBatch rule relies on an already selected pod for the replenish- ment order. Since Replenishment Pod Selection determines the composition of the pods, it offers many possibilities to create pods with different features, e.g. high frequency pods that combine frequently ordered products, or family-based pods combining products that are often ordered together. If all replenishment orders assigned to the same pod are assigned to the same replenishment sta- tion, only one trip is necessary to place all replenishment orders on the pod, which reduces the number of robot movements. The five RPS rules used in this paper are the "Random", "Emptiest", "Near- est", "Least-Demand" and "Class" rules: 13 Random The Random rule selects a random pod with sufficient remaining capacity. Emptiest The Emptiest rule assigns replenishment orders to the emptiest pod and reuses the same pod for subsequent replenishment orders until it is full or used at a station. Nearest The Nearest rule assigns an incoming replenishment order to the nearest pod with sufficient remaining capacity. Least-Demand With the Least-Demand rule an incoming replenishment or- der is assigned to the pod currently offering the least demanded inventory, i.e. the pod with the least units offered when compared to the aggre- gated demand by assigned and backlogged pick orders is selected. Thus, this pod is not useful for pick-operations at the time of selection and by this it is not disadvantageous to block it for replenishment operations. Class The Class rule assigns incoming replenishment orders to a pod of the same class as the replenishment order, i.e. fast moving SKUs to pods with other fast moving SKUs. The classes are built by a background mechanism for which the cumulative relative amount of pods per class are given. For this work we use "0.1, 0.3, 1.0", i.e., three classes where the first class holds 10 % of the pods for the highest frequency SKUs, the second class holds 20 % and the last class holds the remaining ones, which are the ones with the lowest frequency SKUs. To assign a replenishment order of a certain class, the emptiest pod is selected from the pods of that particular class. Similar to the Emptiest rule, a selected pod is used for the subsequent incoming replenishment orders of the same class until no more replenishment orders fit the pod or until the respective pod completes its visit to a replenishment station. 5.5 Pod Storage Assignment Rules For each pod an unoccupied storage location has to be selected, every time after visiting a pick or replenishment station. PSA is an important aspect of the RMFS, because being able to change the storage location of pods after every visit to a workstation is what makes continuous automatic sorting possible. For PSA, five decision rules are examined, namely the "Random", "Fixed", "Nearest", "Station-Based" and "Class" rules. Random The Random rule chooses a random free storage location. Fixed The Fixed rule maintains the initially assigned storage location for all pods. 14 Nearest The Nearest rule stores pods at the nearest unoccupied storage lo- cation in terms of shortest estimated path time. This path time is de- termined using an A∗ algorithm that takes the time needed for turning the robot (with or without pod) into account. Station-Based The Station-based rule is a variant on the Nearest rule, i.e. instead of bringing the pod to a storage location that is nearest to the robot's position the storage location with shortest path time to a pick station is selected. The greatest difference with the Nearest rule is in the storage locations chosen for pods returning from a visit to a replenish- ment station. Class The Class rule brings pods back to storage locations of the same class, where classes are constructed in a similar fashion as in the RPS decision problem, but based on the shortest path time to a pick station. Within a class, a storage location for a pod is selected analogously to the Nearest rule. Table 3 provides an overview of the decision rules per decision problem and shows how the decision rules are labeled across decision problems. Note that choosing a rule for one decision problem may jeopardize strategies chosen for others. For example, a random Pick Order Assignment may have a negative impact on a Turnover-based approach for assigning replenishment orders to storage locations, because it does not respect the units currently positioned near the pick station while assigning orders to it. Hence, a selection respecting mutual influences has to be done to provide an efficient compilation of rules that is able to adequately overcome the planning problems in such a system. Table 3: Overview of the Decision Rules per Decision Problem Decision Problem Decision Rules POA Random, FCFS, Due-Time, Fast-Lane, Common-Lines, Pod-Match ROA Random, Pod-Batch PPS Random, Nearest, Pile-on, Demand, Lateness, Age RPS Random, Emptiest, Nearest, Least-Demand, Class PSA Random, Fixed, Nearest, Station-Based, Class 6 Simulation Framework In this work we use our simulation framework, called "RAWSim-O", which is inspired by the work of [7]. A more detailed description of the framework can be found in [13] while the source code is available at https://github.com/ merschformann/RAWSim-O. Similar to [7], we use an agent-based and event- driven simulation focusing at a detailed view of the system. The basic simula- tion process is managed by the core simulator instance (see Figure 4a), which 15 (a) Overview of the simulation process. (b) Visualization screenshot Figure 4: RAWSim-O simulation framework is responsible for obtaining the next event and updating the agents. Agents can either represent real entities like robots and stations or virtual entities like process managers, e.g. for emulating order processes. Every decision that has to be made is passed to the corresponding controller. The controller can either immediately decide or can buffer multiple requests in order to optimize and release the decision later on. However, in this work we only consider ad-hoc decision rules with the former approach. To allow visual feedback, the ongoing simulation can optionally be rendered in 2D and 3D. The implementation was done in C#. The level of detail of the simulation is especially high for the simulated move- ment behavior of the robots. We consider the robot's momentum by emulating acceleration and deceleration behavior, collision avoidance and turning speed (see Table 5). The emulation employs a continuous time-horizon. The times for activities other than robot movement, e.g. lifting or storing a pod, or pick- ing one unit at a pick station, are constant (see Table 5). The waypoints allow the emulated robot behavior to match real robot behaviour. Robots that do not carry a pod can traverse underneath stored pods by using the waypoints at which the pods are stored. Furthermore, in the buffers of the workstations, robots can take short-cuts if the buffer is (partially) empty. Information about the system's state is tracked in a high level of detail, be- cause some decision rules differ with regard to the information they require. For example, all pods and all units on all pods are tracked exactly. Incoming information is divided into a static and a dynamic category. Static informa- tion includes everything describing a system instance and is completely given at start. Static information therefore includes the number and composition of pick stations and replenishment stations, the pods, the robots, and the way- point system used for robot navigation. All of the decision rules proposed in this work differ in their computational complexity and therefore also in the computational time they require to reach decision. They are, however, simple enough to be considered as ad-hoc decisions even for large system sizes. 16 StartEndInitializeSimulatorControllerVisualizerFrameworkDecision imminentDecideRenderOutput resultsFetch informationAgentUpdateOptimizeIn contrast to static information, the dynamic information is not completely known beforehand, but becomes available over time. This is the case for in- coming pick orders and replenishment orders submitted to the system over time by external processes. While each replenishment order consists of a num- ber of physical units of one SKU, each pick order consists of a set of order lines, each for one SKU, with corresponding units necessary to fulfill the line. We assume for both pick and replenishment orders, that there is a constant order backlog. A constant order backlog means that when an order from the backlog is assigned to a workstation, it is immediately replaced by a newly generated order. By keeping the order backlogs constant, we aim to analyze the system's behavior under constant pressure. However, it also leads to the phenomenon that the system's storage space utilization (utilized space divided by total space available) in the storage area is affected by the performance of the decision rules controlling it, because no further virtual manager steers the process. E.g., if a combination of rules is replenishing quickly, the storage space utilization will increase. In contrast, it will decrease, if the rules are re- plenishing slowly. Situations in which the storage space utilization is nearing 100%, and only few storage places for new replenishment orders are available, lead to an inefficient replenishment process. To avoid such situations, we pause replenishment order generation, if storage space utilization exceeds 85 % and it is continued after it drops below 65 % again. Analogously, we pause the pick order generation, if storage space utilization drops below 10 % and re- sume after it exceeds 60 % again. The latter is done to avoid draining the inventory completely. Since in both cases either the replenishment stations or the pick stations will become inactive due to no further orders to process, the robots will be reassigned to the remaining active stations. This redistribution of robots across the active stations is done at any time a station becomes active or inactive, i.e. at the beginning and end of order generation pauses. If a new replenishment order is received, first the rules for ROA and RPS are responsible for choosing a replenishment station and a pod (see Figure 5). The time the decision is taken depends on the active rules. The execution of the assignment can earliest be done as soon as there is sufficient capacity on a pod and a station available. The commit of the assignment technically results in an insertion request (shown as red cylinders), i.e. a request that requires a robot to bring the pod to the workstation. Multiple of these requests are then combined to an insertion task and assigned to a robot by a TA rule. Similarly, after the POA rule selects a pick order from the backlog and the assignment is committed to a pick station, an extraction request (shown as blue cylinders) is generated, i.e. a request that requires bringing a suitable pod to the chosen station. Up to this point, the physical units of SKUs for fulfilling the pick order are not yet chosen. Instead, the decision is postponed and taken right before combining different requests to extraction tasks by PPS and assigning them to robots by TA. This allows the implemented rules to exploit more information when choosing a pod for picking. Hence, in this work we consider PPS as a 17 decision closely interlinked with TA. Furthermore, the system generates store requests (shown as orange cylinders) each time a pod has to be transported to a storage location. The PSA rule only decides the storage location for a pod that is not needed anymore and has to be returned to the storage area. If all requests are already being handled by other robots, the robot will be assigned an idle task, thus, the robot dwells at a dwelling point until needed. Dwelling points can be used to reduce congestion effects if there are only a few active stations compared to the number of robots, e.g. robots waiting at a storage location block others that try to pass by. For this, the robot will park at a free storage location to avoid causing conflicts with other robots. The dwell point policy uses locations in the middle of the storage area to avoid blocking prominent storage locations next to the stations. Another type of task would be charging, which is necessary when robots run low on battery, however, in this work we assume the battery capacity to be infinite, so this type of task is ignored. All of the tasks result in trips (shown as green cylinders), which are planned by a path planning algorithm and executed by the robots. The only exception is when a pod can be used for another task at the same station. The trips are planned by a PP algorithm and the resulting paths are executed by the robots. Figure 5 shows an abstract overview of these dependencies. The exact times at which the decisions are taken depend on the respective rules, e.g. the Pod-Batch ROA rule assigns a batch of replenishment orders to the first pick station offering sufficient space while the Random ROA rule imme- diately assigns single replenishment orders to the first station with sufficient capacity available. However, all of the rules have in common that they make assignments greedily while adhering to certain capacity constraints (station capacity, pod capacity, etc.). Figure 5: Order of decisions to be done induced by receiving pick and replen- ishment orders 18 InsertExtractStoreRequests:Trips:RPSROAPOATAPPPSAPPSRepl. orderreceivedPick orderreceived7 Evaluation Framework This section describes the evaluation framework used to carry out the research in this paper. Two central concepts to the evaluation framework are the Rule Configuration (RC) and the Warehouse Scenario (WS). The RC specifies for each decision problem, which decision rule is used. The WS specifies the warehouse layout, number of robots, number of workstations, number of SKUs, whether or not return orders are part of the operations of the warehouse, and pick order size. During one simulation run the RC and WS do not change, so they can be seen as an input to a simulation run. The evaluation framework consists of two phases, one varying the RCs, the other varying the WSs. Phase 1 evaluates all 1620 possible RCs on one WS. For phase 1, we compare eight performance measures: (1) unit throughput rate, (2) pick order throughput rate, (3) order turnover time, (4) distance traveled per robot, (5) order offset, (6) fraction of orders that are late, (7) pile-on (8) the pick station idle time. Unit throughput rate is the number of picked units of all SKUs per hour. Pick order throughput rate is the number of pick orders fulfilled per hour. Order turnover time is the average time between submitting a pick order to the backlog and fulfilling it. Order offset is the average time between the due time and the completion time of the pick orders. Thus, a value smaller than zero shows how much in advance pick orders are completed. The rationale behind this is that follow-up processes at the distribution center are not deterministic, hence, pick orders completed earlier may improve the overall service level. The pick station idle time is measured as an average across all pick stations in the system. Phase 1 selects the RCs with the highest unit throughput rate. However, among these selected best RCs, the variety in the decision rules may be low. For a particular decision problem, all of the selected RCs may use the same decision rule. To ensure more diversity in the RCs in phase 2, we define 6 so- called "benchmark RCs", see Table 4. The benchmark RCs were chosen such, that all decision rules across all decision problems appear in at least one of the benchmark RCs. Each benchmark RC has been given a name that reflects a characteristic that the decision rules have most in common. Phase 2 evaluates Table 4: Benchmark RCs Benchmark RC Demand Speed Nearest Class Greedy Random POA ROA PPS RPS Due-Time Pod-Batch Demand Least-Demand Emptiest Fast-Lane Pod-Batch Lateness Nearest Random Nearest Age Class Pod-Match Pod-Batch Pile-on Random Random Random PSA Fixed Nearest Nearest Class Emptiest Station-Based Random Random FCFS Common-Lines Pod-Batch 19 the selected RCs from phase 1 and the benchmark RCs, while varying the warehouse scenarios. Since we are specifically interested in efficiency of RCs we neglect layout decisions for this work. Thus, we choose one specific layout, using the style described in Section 2. The concrete layout instance comprises 1149 pods and 1352 storage locations ( 85% filled) and is shown in Figure 6. When varying the number of pick stations during phase 2 we add workstations in the order given in Figure 6. Figure 6: Top view of the layout, including pick station indices, with the storage area in the middle, replenishment stations to the left, and pick stations to the right 7.1 Parameters In the following we describe the used parameters in more detail. The pa- rameters shared for both phases are outlined in Table 5. We set a continuous simulation horizon of 48 hours in order to decrease the impact of side effects like recurring replenishment overflows, which cause replenishment pauses described previously. Within a duration of 48 hours we observe sufficient repetitions of such patterns to achieve a reasonable mitigation of these side effects. Furthermore, for each RC and WS combination in phase 1 and in phase 2 we conduct 10 runs to lessen the effect of randomness. To keep the system under continuous pressure, like described above, we keep a constant pick and replenishment order backlog of 200 orders each. At simulation start inventory is generated until 70 % overall storage utilization to avoid cold starting the system. This is done using the same process used for generating replenishment orders during simulation and using assignment rules suiting the respective RPS rule in place. The storage capacity of a pod is set to 500 slots while the storage consumption of one SKU unit is drawn from a uniform distribution between 20 5π rad s with acceleration and deceleration rates of 0.5 m 2 and 8 slots, thus, a full pod contains 100 units in average. The popularity of the SKUs is determined by drawing a value from an exponential distribu- tion with parameter λ = 1 2 for each SKU to emulate a typical ABC curve in e-commerce. This popularity is the relative frequency parameter between all SKUs, thus, the frequency (if divided by the sum of all frequencies) is the probability of choosing a particular SKU when generating an order line for both replenishment and pick orders. One replenishment order restocks be- tween 4 and 12 units of one SKU following a uniform distribution. To emulate due times we distinguish between priority and normal orders that have to be completed in 30 minutes respectively 120 minutes. This reflects the need for preferring important orders. The movement behavior of the robots is emulated by using a maximum ve- locity of 1.5 m s2 . We set the rotational speed to 4 s , i.e., 2.5s for a full turn. Turning takes the same amount of time regardless of whether a robot is carrying a pod. The time for lifting and setting down a pod is set to 3s. This should reflect the capabilities of mobile robots used in similar industry applications reasonably close. For the actual pick operation of one unit at a pick station we assume a constant time of 8s. The complete time for handling one unit including additional operations, like putting the product unit in the correct pick order tote, is set to 15s. This distinction is considered to allow for an early release of the robot, such that no unnecessary robot waiting times are caused. This is not distinguished for replenishment operations, since we assume that a robot can only leave after fully completing the put operation to the pod. The time of a put operation of one replenishment order is set to 20s. The parameters in Table 5 are shared across all conducted experiments, while the parameters in Table 6 are depending on phase and scenario. For the first phase we assess all possible RCs for one fixed warehouse scenario. Note that the RPS rule Nearest and the ROA rule Pod-Batch both wait for the other one to decide first leading to no decision at all, hence, the combination of these rules is forbidden. For the fixed warehouse scenario we set the number of robots to 4 per pick station, i.e. 8 robots in the system at whole. Furthermore, we set the number of pick stations to 2, the number of SKUs to 1000 and exclude the processing of return orders. The order setting is set to Mixed. This means the number of lines per pick order and the number of units per order line are generated following truncated normal distributions with parameters shown in Table 6. This is done to resemble e-commerce pick order characteristics of generally small orders with occasional larger ones in between. Equation (8) shows that phase 1 has 1620 RCs, and since phase 1 has 1 WS and 10 runs are conducted per RC and WS combination, this results in 16200 simulation runs for phase 2. Phase 2 has 10 RCs (see Table 6) and Equation 21 Table 5: Parameters shared across all simulations Parameter Simulation Simulated duration of warehouse operations Number of simulation repetitions Size of pick order backlog Size of repl. order backlog Layout Value 48 hours 10 repetitions 200 pick orders 200 repl. orders 1149 pods, 1352 storage locations in 2 × 4 blocks, 12 aisles and 12 cross-aisles Orders uniform distribution between 4 and 12 units Number of units per repl. order 20 % Amount of priority orders in pick orders backlog submission time + 30 min. Priority pick order due time backlog submission time + 120 min. Normal pick order due time 60% of inventory capacity of the storage area Threshold when pick order generation starts Threshold when pick order generation stops 10% of inventory capacity of the storage area Threshold when repl. order generation starts 65% of inventory capacity of the storage area Threshold when repl. order generation stops 85% of inventory capacity of the storage area Inventory Initial inventory in the storage area 70% of the inventory capacity of the storage area 500 slots Exponential distribution, λ = 1 2 uniform distribution between 2 and 8 slots Space on a pod SKU frequency / popularity SKU size Robot movement Robot acceleration/deceleration Robot maximum velocity Time needed for a full turn of a robot Time needed for lifting and storing a pod Time needed for picking a unit Time needed for handling a unit at pick sta- tion Time needed for putting a repl. order on a pod Stations Repl. station capacity Pick station capacity 0.5 m s2 1.5 m s 2.5s 3s 8s 15s 20s two times pod capacity 8 pick orders 22 (8) shows that it has 360 WSs, which together with 10 runs per RC and WS combination leads to 36000 simulation runs for phase 2. {(roa, poa, rps, pps, psa) (roa = Pod-Batch) ∧ (rps = Nearest)}(cid:12)(cid:12) = 1620 #RC in phase 1 =(cid:12)(cid:12){ROA} × {P OA} × {RP S} × {P P S} × {P SA}\ #W S in phase 2 =(cid:12)(cid:12){number pick station} × {robots per pick station} ×{number of SKUs} × {return orders} × {pick order size}(cid:12)(cid:12) = (8) 6 × 5 × 2 × 2 × 3 = 360 (9) For phase 2 we limit the RCs to the 6 benchmark RCs and the 4 best ones from phase 1, i.e., the 4 RCs with highest throughput rate. Moreover, we vary the number of pick stations from 1 through 6 and the number of robots per pick station from 2 through 6. This leads to a range from 2 robots in the system to 36 robots across all WSs. In addition to WSs with 1000 SKU, we also assess WSs with 10000 SKUs stored in the system. For the order size we define two additional settings of small and large orders. For the Small pick order size, only single line / single unit pick orders are generated. For the Large pick order size, the distributions from the Mixed order setting are used but the min parameter for both is set to 2. Lastly, in WSs where we emulate the processing of return orders, 30 % of the generated replenishment orders are single unit. The total number of RC and WS combinations for the phase 2 is therefore 3600, which leads to 36000 simulation runs. 8 Computational Results This section shows the results from phase 1 and phase 2 of the evaluation framework. Throughout this section, the unit throughput rate is presented as a percentage of the upper bound on the unit throughput rate. The unit throughput rate is presented in this way to facilitate interpretation and com- parison of results across experiments. Moreover, the RMFS is supposed to have high pick rates as it eliminates the need of walking for the workers, while the robots are supposed to supply the pickers with a constant stream of pods to pick from. Presenting the unit throughput rate as a percentage shows clearly to what extent these aims are achieved. The upper bound is discussed in more detail in Appendix A. The length of the confidence intervals is always less than 1% of the mean, based on 10 runs per RC and WS combination, and therefore does not add much information. 23 Table 6: Varied parameters for phase 1 and 2 Phase 1 values Phase 2 values 1620 RCs 6 Benchmark RCs + 4 best RCs from phase 1 Parameter Rule configurations (RCs) Number of pick stations Robots per pick station Number of SKUs Return orders 2 4 1000 0 % 1, 2, 3, 4, 5, 6 2, 3, 4, 5, 6 1000, 10000 0 %, 30 % Small - line & unit distributions: min = 1, max = 1 min = 1, max = 1 Mixed - line & unit distributions: µ = 1, σ = 1, min = 1, max = 4 µ = 1, σ = 0.3, min = 1, max = 3 Large - line & unit distributions: µ = 1, σ = 1, min = 2, max = 4 µ = 1, σ = 0.3, min = 2, max = 3 10 360 3600 36000 Pick order size Mixed - line & unit distributions: µ = 1, σ = 1, min = 1, max = 4 µ = 1, σ = 0.3, min = 1, max = 3 # RC # WS # RC×WS # simulation runs 8.1 Phase 1 1620 1 1620 16200 The first phase aims to investigate throughput performance and the impact per decision problem of decision rules on throughput. Furthermore, we assess the behavior of the different output measures depending on decision rule selection. For this, Table 7 shows how across these simulations the seven previously introduced performance measures correlate with each other. At first, we can observe that as the unit throughput rate score improves, the other performance measures improve as well. As the unit throughput rate score increases, pick order throughput rate and pile-on increase as well, whereas the order turnover time, the distance that robots travel, the order offset, the fraction of orders that miss their due time, and the station idle time decreases. Although it is 24 Table 7: Correlations between the different performance measures for first phase e m i t r e v o n r u t r e d r O d e l e v a r t e c n a t s i D t u p h g u o r h t r e d r O e m i t e l d i n o i t a t S n o - e l i P t e s ff o r e d r O s r e d r o e t a L t u p h g u o r h t t i n σ U - Unit throughput Order throughput Order turnover time Distance traveled Order offset Late orders Pile-on Station idle time µ 0.556 241.963 0.187 2.438 0.450 3549.625 1220.445 -0.950 -0.950 122598.768 19433.860 -0.952 -0.952 0.880 -2565.458 1224.985 -0.950 -0.950 1.000 0.880 - - - - - 0.189 82.234 1.000 - - - - - - 0.115 -0.590 -0.591 0.685 0.549 0.684 1.450 0.899 0.899 -0.802 -0.796 -0.802 -0.448 - 0.186 -1.000 -1.000 0.950 0.952 0.950 0.591 -0.899 - - - - - - - - - - - - - - - - - - - - - - - not clear what the exact causal relationships are, the correlations suggest that pile-on and the distance traveled by the robots are the main drivers behind these improvements. With higher pile-on, more units are picked per pod, so order lines are fulfilled more quickly and fewer trips are needed to fulfill the pick orders. This also causes longer processing times for each pod at the pick station, which in turn increases the time for the next robot to queue and become ready at the station. In other words: a more continuous input of inventory at the pick station is achieved. Additionally, fewer trips for the pick process free up robots to do more replenishment tasks. With less distance traveled by the robots we expect pods to be presented at the pick stations more continuously. Similar to the pile-on this effect enables more continuous picking, which in turn increases the overall unit throughput rate. Both measures, pile- on and the traveled distance, are intermediate measures affected by the choice of strategy for the different decision problems, i.e., a better score in both helps decreasing the idle time at the stations, which in turn helps increasing the throughput. An increased throughput, in the constant pick order backlog setting of this work, also decreases the turnover time of pick orders and the due time offset. Only the number of orders being late is not strongly correlated with the two main throughput drivers. The two main throughput drivers can also be observed when looking at a scatter plot of all simulation runs of the first phase (see Figure 7). Here we can see the best results in unit throughput rate score are achieved with a high pile-on and less distance traveled per robot. The group of simulation runs with least distance traveled per bot and a pile- on around 4 are RCs involving the Nearest PPS rule, while the simulation runs with highest pile-on (greater 5) at the top of the plot are RCs involving the Demand PPS rule. In both groups we find runs with the highest unit throughput rate score, hence, a higher throughput is not only achieved by a 25 high pile-on. In particular within the top ten RCs in terms of unit throughput rate score the pile-on ranges between 3.84 and 6.36, while the distance traveled per bot ranges between 68.04 km to 80.36 km. Hence, pile-on and the traveled distance enable higher throughput, but may also compensate for each other. This is particularly interesting, because both come at operational costs. For traveled distance this is energy consumption and robot wear, while for pile-on it may be costs arising from potentially more complex replenishment processes. Furthermore, within both groups better results are obtained with RCs also involving the Pod-Match POA rule, which causes an additional boost in pile- on. In Figure 7 we also observe a 'cutoff' of simulation runs in the upper right and bottom left areas. This can be explained by the longer handling time at the station resulting from a higher pile-on. I.e., the longer a robot needs to wait at a station for the picking to finish the less it can travel in the meantime. Thus, rules increasing pile-on may help reducing the necessary travel distance, and by this also robot wear and energy consumption. Figure 7: Scatter plot for pile-on vs. traveled distance per robot colored by the achieved throughput rate score for all simulation runs of the first phase The pick order throughput rate is neglected completely in the remainder of this work, because it almost completely aligns with the unit throughput rate score. The reason for this is the constant backlog of 200 pick orders over 48 hours: with a pick order throughput rate of 241.963 completed orders per hour in average, omitting certain pick orders is almost impossible. Hence, we cannot observe a potential temporary throughput gain by preferring smaller or larger orders. In order to investigate the trade-off between picking many units and completing more pick orders an experiment with a fixed set of backlogged pick orders over a fixed period of time should be devised. For this, the possibly 26 1 2 3 4 5 6 7 60 70 80 90 100 110 120 130 140 150Pile-onDistance traveled per robot (in km)Simulation runs 20 30 40 50 60 70 80 90 100Unit throughput score (in %)tedious processing of leftover pick orders, which are presumably harder to pick quickly, needs to be investigated. We leave this work for future research. Table 8: Average unit throughput rates as percentages of the upper bound for all rules, together with the best / worst performance multiplier per decision problem Mult. (cid:0) best worst (cid:1) POA Common-Lines Due-Time Fast-Lane 41.93% 76.13% 50.93% FCFS 41.81% Pod-Match Random 81.18% 41.71% ROA PPS RPS PSA Random Pod-Batch 53.71% 57.99% Nearest Age Demand Lateness 52.70% 48.63% 62.16% Nearest Emptiest Least-Demand 58.42% 59.63% 57.71% Fixed Nearest 54.08% 58.79% 61.50% Class 56.16% Class 55.91% Random 47.56% Random Station-Based 53.60% 55.70% Pile-on Random 59.82% 48.88% 1.946 1.080 1.278 1.254 1.097 Table 8 shows for each decision problem the unit throughput rate score for each of the decision rules, averaged across all simulations in phase 1. We calculate the multiplier by dividing the highest unit throughput rate by the lowest. As the multiplier in unit throughput rates is rather large for the POA decision problem, system integrators and RMFS suppliers may benefit from carefully selecting a POA decision rule and from investigating better decision rules for this decision problem. The multiplier for the Replenishment Order Assignment is near 1, indicating that using a different decision rule does not offer much performance improvements. However, we note that we keep the sequence of incoming replenishment orders fixed at all times in this work, which limits improvement potential. Nevertheless, we expect limited degrees-of-freedom in replenishment operations to be more realistic, because the sequence will typically be a result of preceding operations or systems. Moreover, the limited number of replenishment stations diminishes the impact of ROA decision rules even more. Furthermore, the impact of the Pod Storage Assignment selection rule seems to be fairly low. This may be a reason of the quite small layout. We expect the impact of PSA decision rules to increase with the size of the instance layout, because the effect on the traveled distance would grow by a large amount. In the following we analyze the achieved throughput performance per decision rule. For this, Figure 8 shows the box-plots of unit throughput rate scores for each decision problem grouped per decision rule. The boundaries of the boxes are determined by the upper and lower quartile while the line in the middle indicates the median value. The whiskers extend from the boxes to the minimum and maximum values. The first observation is that throughput performance of the RMFS is most sensitive to the choice of POA decision rule among the defined decision rules. This aligns with the previously observed correlations, because the choice of POA immediately affects the pile-on, which 27 Figure 8: Unit throughput rate performance of all runs grouped per rule is identified as a major performance driver. The best performing POA strate- gies are FastLane and PodMatch, which both look at the incoming pods at a pick station when assigning new pick orders from the backlog. This suggests that a strategy aligning pick orders with the content of incoming pods seems most promising for throughput efficiency. This backs up the findings of [2]. Although the Common-Lines rule exploits a similar greedy strategy, it achieves substantially less throughput. Hence, only matching pick orders to each other but not to the content of the pods squanders throughput capabilities of the system. All other POA decision rules achieve similar throughput performance, since they do not consider order characteristics that would affect pile-on or traveled distance. When looking at the PPS rule box-plots the average best throughput perfor- mance with least variance is achieved by the Age, Nearest and Pile-on rules. All of them focus either on maximizing the pile-on or minimizing the trav- eled distance. Although the Age rule does only indirectly maximize pile-on, it achieves a higher average pile-on of 2.92 among all RCs containing it than the 28 20 %30 %40 %50 %60 %70 %80 %90 %100 %Common-LinesDue-TimeFast-LaneFCFSPod-MatchRandomunit throughput rate scorePOA rules20 %30 %40 %50 %60 %70 %80 %90 %100 %RandomPod-Batchunit throughput rate scoreROA rules20 %30 %40 %50 %60 %70 %80 %90 %100 %AgeDemandLatenessNearestPile-onRandomunit throughput rate scorePPS rules20 %30 %40 %50 %60 %70 %80 %90 %100 %ClassClosestEmptiestLeast-DemandRandomunit throughput rate scoreRPS rules20 %30 %40 %50 %60 %70 %80 %90 %100 %ClassFixedNearestRandomStation-Basedunit throughput rate scorePSA rulesactual Pile-on rule, which achieves an average pile-on of 2.79. The Demand rule has the highest spread across PPS rules with a very low median, but also provides some top performing RCs (see Table 9). This suggests that the throughput performance of the rule has a higher dependency on the selection of other rules. Although the variation among the ROA decision rules is small, we observe a slightly better throughput performance by the Pod-Batch rule. This is a reason of the smaller number of trips necessary when batching replenishment orders. Many of the top performing RCs contain the Emptiest or Closest RPS decision rule. The main reason for the good throughput performance again seems to rely on fewer and shorter trips. The Emptiest rule decreases the number of trips, because more replenishment orders are stored in pods at once until it is full. E.g., only 31.03 % pods need to be brought to replenishment stations in average when compared to the Random rule. The Closest rule benefits from a similar effect since the same (closest) pod is used for further replenishment or- ders even while it is already approaching. Furthermore, Closest decreases the distance per replenishment trip, because nearer pods are used. The Random rule performs worst for RPS. The main reason for this is that too many trips are caused by randomly selecting pods while only storing few replenishment orders per trip. Among the PSA decision rules we observe the best throughput performance for the Nearest strategy. This is again mainly caused by the shorter trips for the robots. When comparing the Nearest and the Station-based rule we see an overall benefit when ... shorter trips for replenishment operations too, and not using the robots working for replenishment operations to prepare pods for pick operations. However, this depends on the queue length at stations and the distribution of robots between replenishment and picking. I.e., if longer queue times are expected at replenishment stations, moving pods nearer to the pick stations when returning them to the inventory may improve overall throughput performance. The Fixed and Random decision rules differ little in their performance. The main reason for this is that the storage location per pod in the Fixed rule is randomly selected. Thus, leading to a very similar behavior. Due to the large sample sizes, the results of ANOVA and Tukey's range tests rejected the hypotheses that the means were equals at the 0.05 significance level within groups and pair-wise, with five exceptions. The null hypothesis of equal means was not rejected at the 0.05 significance level for POA rules FCFS and Due-Time, for Random and Due-Time, and for Random and FCFS. Fur- thermore, for PPS rules Random and Lateness the hypothesis of equal means could not be rejected, and for PSA rules Station-Based and Class. 29 Table 9: RCs with best throughput score selected from first phase (performance is unit throughput rate score) RC rank 1 2 3 4 PPS POA ROA RPS Pod-Match Pod-Batch Demand Emptiest Nearest Pod-Match Pod-Batch Demand Emptiest Station-Based Pod-Match Pod-Batch Nearest Emptiest Nearest Class Pod-Match Pod-Batch Demand Emptiest PSA performance 94.81 % 94.63 % 94.43 % 94.00 % 8.2 Phase 2 From the 1620 RCs in phase 1, the four with the highest unit throughput rate (see Table 9) together with the benchmark RCs form the set of ten RCs used in phase 2. The main purpose of phase 2 is to examine how well the RCs perform under different circumstances. In the following we analyze the results obtained for the 12 warehouse scenarios and 30 resource settings described before (see Section 7.1). Table 10: Best unit throughput rate score for all scenarios, robots per pick station and numbers of pick stations. Scenario abbreviations: [SKU count: 1000 (1K), 10000 (10K)]-[Order size: Small (S), Medium (M), Large (L)]- [Return orders: yes (R), no (N)] 4 5 6 3 2 1 Stations Robots 2 3 4 5 6 2 3 4 5 6 2 3 4 5 6 2 3 4 5 6 2 3 4 5 6 2 3 4 5 6 1K-S-N 44 82 91 97 97 59 89 94 97 98 64 90 95 97 98 60 87 93 97 98 57 87 93 97 98 59 87 94 97 98 1K-S-R 46 82 92 97 97 59 89 94 97 98 63 90 95 97 98 61 88 93 97 98 56 87 93 97 98 55 86 93 97 98 1K-M-N 45 83 92 97 98 60 90 95 98 98 64 90 95 98 98 60 88 93 98 98 57 88 94 98 98 59 88 94 98 98 1K-M-R 45 82 92 97 98 59 89 95 98 98 63 91 96 98 98 62 89 93 98 98 56 88 94 98 98 56 87 94 98 98 1K-L-N 54 83 93 99 99 66 90 97 99 99 68 91 96 99 99 67 88 94 99 99 63 86 94 98 99 63 87 94 99 99 1K-L-R 50 80 92 99 99 64 88 96 99 99 65 90 97 99 99 67 88 94 99 99 62 86 93 98 99 62 84 94 98 99 10K-S-N 21 39 55 68 78 27 44 59 72 81 28 46 61 73 80 29 47 59 69 77 30 47 57 68 77 31 45 58 68 76 10K-S-R 20 40 56 70 80 27 45 61 74 82 29 47 62 74 82 30 48 61 70 78 31 48 58 67 76 31 46 57 66 74 10K-M-N 23 41 58 71 81 28 46 61 75 84 29 48 64 76 83 30 49 61 71 80 31 48 60 71 80 32 47 60 71 79 10K-M-R 21 41 59 73 83 28 48 63 76 85 30 49 65 77 84 31 50 63 72 81 32 49 60 70 79 32 47 59 69 77 10K-L-N 36 63 84 94 98 44 71 89 96 99 46 73 87 94 98 47 69 83 92 97 46 67 84 91 97 45 68 83 91 97 10K-L-R 30 52 76 89 96 37 62 81 92 97 40 64 83 91 96 41 64 76 87 94 42 61 74 84 93 41 59 73 84 92 Table 10 shows the results, with the entries being the unit throughput rate as a percentage of the upper bound. In each cell the result of the best performing RC for the respective scenario and station / robot configuration is shown. The unit throughput rate scales well when adding more pick stations, the scaling is (almost) completely independent of the scenario characteristics. However, the necessary number of robots to achieve a given unit throughput rate greatly depends on the scenario characteristics, e.g., for more SKUs more robots are necessary to achieve a high unit throughput rate. The number of SKUs, does have a major impact on performance overall, where the main reason is that 30 pile-on is considerably lower for the 10000 SKU scenarios. A reason for this is the lower likeliness to have a pod with a good combination of SKUs matching the orders of the pick stations available. Thereby, if larger orders have to be processed with the system, this helps mitigating the negative effect of handling lots of SKUs. The main reason for this are the larger number of order lines ac- tive at a station when picking larger orders. I.e., more open order lines increase the likeliness of having a well matching pod available for the inventory required at a pick station. Processing return orders has an increased negative effect, if the order size of customer orders is large. However, in general, whether return orders are processed has a lesser effect on throughput performance than the other warehouse scenario variations. The reason behind this may be that even though approximately 19.76 % more time is spent on replenishment operations by the robots when compared to the scenarios without return order processing, replenishment operations are overall quick enough to mitigate the effect. Re- plenishment operations only consume 20.29 % out of the overall time consumed by the robots in average across all phase 2 simulation runs. Furthermore, we can conclude that with 1000 SKUs, the unit throughput rates are close to their theoretical maximum even with relatively few robots per stations. Table 11 shows the unit throughput rate score for the RCs for all combina- tions of number of robots (nr) and number of stations (ns), averaged across WSs and presented as whole percentages. From Table 11 we can see that the Ranked RCs from phase 1 perform similarly and better than the benchmark RCs. Among the benchmark RCs, the Greedy benchmark outperforms the oth- ers consistently across all settings and is the only one whose unit throughput rate scores approached those of the ranked RCs. 9 Conclusion In this work we studied the throughput performance of decision rules for mul- tiple decision problems occurring in the control of RMFS. By analyzing a total of eight output measures for a total of 1620 RCs, we found strong correla- tions between these. Most interestingly a high pile-on and a short distance traveled by the robots together almost immediately account for the success of a decision rule applied to RMFS. Hence, we propose using these two output measures as the key tactics when designing decision strategies for RMFS that aim to achieve high throughput. In the investigated high pressure situation further performance measures like the turnover time of pick orders were also highly correlated with the unit throughput rate, which is why we focused on the throughput itself as the main metric for a successful RMFS. Furthermore, we found that varying the decision rule used for solving the Pick Order Assignment affected the unit throughput rate the most. The average unit throughput rate was twice as high for the best decision rule as it was for the worst. This finding indicates that system engineers and warehouse operators should pay most attention to the Pick Order Assignment decision 31 Table 11: Unit throughput rate scores for the RCs in phase 2 (green ≡ best, red ≡ worst) Ranked RCs Benchmark RCs (ns , nr) 1 2 3 4 Demand Speed Nearest Class Greedy Random (1 , 2) (1 , 3) (1 , 4) (1 , 5) (1 , 6) (2 , 2) (2 , 3) (2 , 4) (2 , 5) (2 , 6) (3 , 2) (3 , 3) (3 , 4) (3 , 5) (3 , 6) (4 , 2) (4 , 3) (4 , 4) (4 , 5) (4 , 6) (5 , 2) (5 , 3) (5 , 4) (5 , 5) (5 , 6) (6 , 2) (6 , 3) (6 , 4) (6 , 5) (6 , 6) 30 28 32 28 60 59 61 58 76 75 76 74 86 86 86 84 91 91 91 90 42 41 43 39 68 67 68 65 81 80 80 78 88 88 88 86 92 92 92 91 45 43 46 42 70 69 70 66 81 81 81 79 88 88 88 86 92 92 92 90 45 44 47 42 68 67 69 65 78 78 78 76 86 86 86 84 91 90 90 89 43 42 45 40 67 66 68 64 78 76 77 75 85 84 85 83 90 90 90 88 44 42 45 41 66 64 66 63 77 76 77 75 85 84 85 83 90 89 89 88 18 34 49 62 73 25 40 55 67 77 27 43 57 68 77 29 44 56 66 74 29 44 55 63 72 30 43 53 62 70 11 23 23 34 45 11 23 29 37 47 15 23 31 41 50 14 23 34 43 53 14 24 35 45 54 15 25 35 45 53 25 42 55 66 74 32 47 59 69 76 32 48 60 69 76 34 49 59 68 74 35 48 58 66 73 35 48 58 65 71 34 57 72 82 87 42 64 76 84 89 44 65 77 84 89 45 64 74 82 87 43 63 73 81 86 43 62 73 80 85 14 24 37 47 58 18 29 40 51 62 19 30 42 52 62 20 31 42 51 60 20 31 41 49 58 20 31 40 48 55 23 40 57 70 79 28 47 63 74 83 32 50 65 75 83 34 51 64 74 81 34 51 63 72 80 35 51 62 71 79 32 problem. Moreover, the unit throughput rate score ranges from 25.24% for the worst RC assessed in phase 1 to 94.81% for the best scoring RC. Hence, the right combination of decision rules plays a crucial role when controlling an RMFS. We propose that future research may assess how to scale beyond the throughput performance of the merely simple decision rules investigated in this work. However, we observe some cross-dependencies between different strate- gies for the core decision problems featured in this paper, e.g., the Demand PPS rule is part of the best performing and the worst performing RC. Thus, an integrated and realistic evaluation or validation of new decision methods for RMFS is highly important, since dependencies exist and side-effects should not be neglected. Additionally, we found that the number of different SKUs in the system has a strong impact on the unit throughput rate. This finding is probably due to a decrease in pile-on for a higher number of SKUs. This effect is considerably less for larger orders, presumably because for larger pick orders pile-on tends to be higher. Having to process return orders seems to affect the unit throughput rate more, if the pick orders are large. Moreover, we found that the performance of the "greedy" benchmark consistently came close to the best ranked configurations of decision rules. This paper has studied solutions to several operational problems, which lead towards promising directions for future research. Each decision rule in this study has looked at an operational problem in isolation, but heuristics that try to integrate multiple operational problems and optimize these problems jointly could achieve substantial increases in order throughput or reductions in resources used. Investigating rules and heuristics that increase pile-on, i.e. the number of picks per handled pod, would also be of great use to practition- ers. While many decision rules and parameters were varied to deliver insightful re- sults we expect even more insight when varying the layout itself. For example, we expect a larger impact of the PSA rule selection when facing huge layout instances. This was not done in this work in order to keep a certain focus and to keep computational resource utilization for the conducted experiments tractable. RMFSs are a new category of automated systems and concepts spe- cific to RMFSs have not received much scholarly attention. An example would be cache zoning / priority zoning, that is the implementation of special zones near the workstations where pods are stored that will be needed in the near future, see also [6]. Another example would be a study of the automatic sorting of the system without explicit zones. Since pods can be relocated to another storage location each time they are transported to and from a workstation, the inventory can be sorted automatically to some degree during operations. It is not clear at which speed automatic sorting takes place or how much perfor- mance benefits from it. Automatic sorting is a unique feature of RMFSs, but as with so many other aspects of RMFSs, it remains to be explored. 33 Acknowledgements We would like to thank the Paderborn Center for Parallel Computing (PC2) for the use of their HPC systems for conducting the experiments. Marius Merschformann is funded by the International Graduate School - Dynamic Intelligent Systems, University of Paderborn. References [1] Michaela Beckschafer, Simon Malberg, Kevin Tierney, and Christoph Weskamp. Simulating storage policies for an automated grid-based ware- house system. In Tolga Bektas, Stefano Coniglio, Antonio Martinez- Sykora, and Stefan Voss, editors, Computational logistics, Lecture Notes in Computer Science, pages 468–482. Springer, Cham, 2017. [2] Nils Boysen, Dirk Briskorn, and Simon Emde. Parts-to-picker based order processing in a rack-moving mobile robots environment. European Journal of Operational Research, 262(2):550–562, 2017. [3] Business Wire. Amazon unveils its eighth generation fulfillment center. www.businesswire.com/multimedia/home/20141130005031/en, February 2015. [Accessed September 23rd, 2016]. [4] C. M. Chen, Y. Gong, M. B. M. de Koster, and J. A. E. E. van Nunen. A flexible evaluative framework for order picking systems. Production and Operations Management, 19(1):70–82, 2010. [5] J. J. Enright and P. R. Wurman. Optimization and coordinated autonomy in mobile fulfillment systems. Working paper, 2011. [6] M. Flipse. Altering and improving Kiva. Working Paper, 2011. [7] Christopher J. Hazard, Peter R. Wurman, and Raffaello D'Andrea. Al- phabet soup: A testbed for studying resource allocation in multi-vehicle systems. In Proceedings of AAAI Workshop on Auction Mechanisms for Robot Coordination, pages 23–30. Citeseer, 2006. [8] S. S. Heragu, X. Cai, A. Krishnamurthy, and C. J. Malmborg. Analytical models for analysis of automated warehouse material handling systems. International Journal of Production Research, 49(22):6833–6861, 2011. [9] T. Lamballais, M. Merschformann, D. Roy, L. Suhl, and M. B. M. De Koster. Optimal policies for resource reallocation in a robotic mo- bile fulfillment system. Working paper, 2017. [10] T. Lamballais, D. Roy, and M. B. M. de Koster. Estimating performance in a robotic mobile fulfillment system. European Journal of Operations Research, 256:976–990, 2017. 34 [11] T. Lamballais, D. Roy, and M. B. M. de Koster. Inventory allocation in robotic mobile fulfillment systems. Working Paper, available at SSRN, 2017. [12] M. Merschformann, L. Xie, and D. Erdmann. Path planning for robotic mobile fulfillment systems. Working paper, available at arXiv, https://arxiv.org/abs/1706.09347, 2017. [13] Marius Merschformann, Lin Xie, and Hanyi Li. RAWSim-O: A simula- tion framework for robotic mobile fulfillment systems. Working Paper, available at arXiv, https://arxiv.org/abs/1710.04726, 2017. [14] S. Nigam, D. Roy, M. B. M. de Koster, and I. J. B. F. Adan. Analysis of class-based storage strategies for the mobile shelf-based order pick system. In Progress in Material Handling Research: 2014, 2014. [15] Kees Jan Roodbergen, Iris F.A. Vis, and G. Don Taylor. Simultaneous International determination of warehouse layout and control policies. Journal of Production Research, 53(11):3306–3326, 2014. [16] D. Roy, A. Krishnamurthy, S. S. Heragu, and C. J. Malmborg. Per- formance analysis and design trade-offs in warehouses with autonomous vehicle technology. IIE Transactions, 44:1045–1060, 2012. [17] M. Wulfraat. bution center? http://www.mwpvl.com/html/kiva systems.html, 2012. for your distri- Is Kiva systems a good fit An unbiased distribution consultant evaluation. [18] P. R. Wurman, R. D'Andrea, and M. Mountz. Coordinating hundreds of cooperative, autonomous vehicles in warehouses. AI Magazine, 29(1):9– 19, 2008. [19] Bipan Zou, Yeming Gong, Xianhao Xu, and Zhe Yuan. Assignment rules International in robotic mobile fulfilment systems for online retailers. Journal of Production Research, 90:1–18, 2017. [20] Bipan Zou, Xianhao Xu, Yeming Yale Gong, and Ren´e de Koster. Eval- uating battery charging and swapping strategies in a robotic mobile ful- fillment system. European Journal of Operational Research, 2017. A Upper bound on the unit throughput rate In the following we introduce an upper bound for the number of units picked per hour. This can be done by considering the constant time for picking (T P ) and the constant time for handling (T O) a unit at a pick station. For an overview of all necessary times see Table 12. If a robot is queueing in the buffer of a pick station, it is assumed that it already turns the right pick face 35 Table 12: Times for determining the upper bound on unit throughput (all times in seconds) Symbol Explanation T P T O T DI T T O T DO Time for picking one unit from the pod (after which the robot can be released, if no further picks are necessary) Time for handling one unit at a station (including picking, putting, pack- ing, etc.) The time for the robot to move up within the queue to the pick station's waypoint The time for the robot to prepare for leaving the station (turning towards exit) The time for the robot to clear the station's waypoint (time to cover the minimal distance) of the pod towards the side where the picker will be. Since the robot is waiting in the queue, this happens in the best case without any additional loss of time. During the actual pick process, the robot is occupied for T P seconds. After this time the robot is allowed to leave the station while the overall handling time for one unit at a station of T O can be longer. (a) Sample times for both mentioned cases (b) Times of the different steps involved in picking units There are two cases to distinguish for obtaining a performance upper bound. First, if the time for picking a unit from the pod plus the time for moving up the next robot (i.e. the time to turn and drive away from the station and the time for the next robot to approach the station from the queue area) is smaller than the overall handling time of a unit at the station, there is a surplus of time available on the system side and the performance is limited by the handling time of the picker (see case 1 in Figure 9a). In the second case we face a longer time for moving up the next robot, hence, in this case we have a loss of time on the system's side and the system is limiting the throughput performance of the picker (see case 2 in Figure 9a). For the sake of clarity we define the time for moving up the next robot in queue as T M U := T DI + T T O + T DO. 36 Robot 1:Human:Robot 2:WaitTurnDriveTPTTOTDODriveTDIWaitTPHandlingTOHandlingTORobot 1:Human:Robot 2:WaitTurnDriveTPTTOTDODriveTDIWaitTPHandlingTOHandlingTOCase 1Case 2LossPlusTDITTOTDOTPTO U B := (cid:12)(cid:12)MO(cid:12)(cid:12) 3600 (cid:12)(cid:12)MO(cid:12)(cid:12) IPO T O 3600 T P + T M U − T O + IPOT O T P + T M U ≤ T O else (10) Considering both cases we can determine an upper bound on the unit through- put rate, i.e. the number of units picked per hour (see Equation 10). For the first case we only need to consider the unit handling time of the picker to determine the maximum throughput rate of one station and multiply it with the overall count of pick stations(cid:12)(cid:12)MO(cid:12)(cid:12). The second case is slightly more com- plicated, because we also need to consider the pile-on. In the denominator we first calculate the loss of time seen in Figure 9a and add it to the average handling time of a pod based on the estimated number of picks from it (IPO). We calculate how many pods are handled in one hour and multiply this by the IPO and the number of stations overall to get the overall upper bound. We recognize that this upper bound on the unit throughput rate relies on some heavy assumptions for real systems, but still propose it as a rule-of-thumb for practitioners, since it is useful for implementations where the time for moving the next robot, the handling times and the pile-on can be estimated. It is a natural limit of the system's performance that the system cannot exceed, even if the number of robots is more than sufficient to supply a continuous stream of pods and all rules are performing well. For this work, the upper bound is correct, because all mentioned times can be accurately determined or are constant and are not subject to random influence within the simulation. 37
1605.04360	1	1605	2016-05-14T01:20:25	Variational Inference with Agent-Based Models	[ "cs.MA" ]	In this paper, we develop a variational method to track and make predictions about a real-world system from continuous imperfect observations about this system, using an agent-based model that describes the system dynamics. By combining the power of big data with the power of model-thinking in the stochastic process framework, we can make many valuable predictions. We show how to track the spread of an epidemic at the individual level and how to make short-term predictions about traffic congestion. This method points to a new way to bring together modelers and data miners by turning the real world into a living lab.	cs.MA	cs	Variational Inference with Agent-Based Models Wen Dong Department of Computer Science and Engineering & Institute of Sustainable Transportation and Logistics State University of New York at Buffalo [email protected] 6 1 0 2 y a M 4 1 ] A M . s c [ 1 v 0 6 3 4 0 . 5 0 6 1 : v i X r a ABSTRACT In this paper, we develop a variational method to track and make predictions about a real-world system from con- tinuous imperfect observations about this system, using an agent-based model that describes the system dynamics. By combining the power of big data with the power of model- thinking in the stochastic process framework, we can make many valuable predictions. We show how to track the spread of an epidemic at the individual level and how to make short- term predictions about traffic congestion. This method points to a new way to bring together modelers and data miners by turning the real world into a living lab. Keywords Social simulation, interactive simulation, novel agent and multi-agent applications, epidemic dynamics, short term traf- fic forecasting, discrete event simulation, stochastic kinetic model, variational methods, expectation propagation, Bethe variational principle, Markov process. 1. INTRODUCTION Agent-based modeling has been employed by researchers in many disciplines to specify the elements of a complex sys- tem and their interactions, to check their understandings of a system, to conduct thought experiments and to inform de- sign and analysis [16, 47, 6, 21]. With the availability of big data in recent years [10, 30, 12], we hope to track and make predictions about a real world system from the data that represent the continuous observations of this system and an agent-based model that specifies how the system evolves, and consequently to turn our world into a living lab. In this paper, we identify the agent-based model as a discrete-event Markov process, and develop a variational inference method that searches the latent state trajectories of the elements of the system in the probability space that are most compat- ible with the noisy observations by minimizing the Bethe variational principle [4]. Data have traditionally been used by agent-based mod- elers to calibrate model parameters, drive model execution, and validate the model. The inquiry in this paper is instead about how continuous imperfect observations about a real- Appears in: Proceedings of the 15th International Conference on Autonomous Agents and Multiagent Systems (AAMAS 2016), J. Thangarajah, K. Tuyls, C. Jonker, S. Marsella (eds.), May 9 -- 13, 2016, Singapore. Copyright c(cid:13) 2016, International Foundation for Autonomous Agents and Multiagent Systems (www.ifaamas.org). All rights reserved. world system can help us make inferences about the system here and now. Instead of simulating traffic jams at rush hours using road network data and agent trips synthesized from census data [32, 42], we are more interested in predict- ing whether today's traffic jams will be formed earlier or last longer from the trajectories of probe vehicles, and how the knowledge about future traffic states will help drivers use the road network more efficiently. Instead of constructing the S- shaped curve of an infectious population from simulation [8, 31], we are more interested in who got a sniffle from his dynamic social network and how we can prevent epidemics from further spreading [29, 14, 18]. Instead of showing the emergence of cities and roads from how people explore and exploit resources [2, 44, 19], We are more interested in iden- tifying poverty and extracting census information from how people make phone calls [5, 10, 38]. Predictions with an agent-based model about real-world data are interpretable in terms of how agents interact with one another and change states, and are amenable to reason regarding non-recurrent scenarios. This transparency about the predictions is lack- ing in non-parametric approaches. Our approach is to identify an agent-based simulator as a Markov process, and to search in the probability space spec- ified by the simulator for agent behaviors and interactions that best match the continuous observations about our real- world system. The key observation behind this approach is that an agent-based simulator generates different sample paths with different probabilities -- it therefore defines a stochastic process with a probability measure assigned to the space of the sample paths that describe the interactions among the elements of the system. In this stochastic pro- cess, the system state as a function of time is composed of the states of its elements. This stochastic process is driven by a number of events that change the system state and happen with event rates that are functions of the current system state. A sample path of the stochastic process is de- fined by a sequence of events and the corresponding times when those events happened, from which we can unambigu- ously recover the system state as a function of time. An agent-based simulator therefore iteratively samples the next event according to event rates then changes the world state according to the sampled state starting from the initial state, until the required amount of simulated time has passed. To find out the maximum likelihood probability distribu- tion of the system state Xt for t1 ≤ t ≤ t2 from observed system state xt1 at time t1 and xt2 at time t2, we follow the forward-backward algorithm: we first let the probability mass diffuse from xt1 in the forward step according to how this system evolves from t1 to t2, and then we iteratively trace backwards from t2 to t1 in the backward step how the probability mass ended at system state xt2 instead of an- other state. After the forward step and the backward step, we get the probability distribution of system state Xt for t1 ≤ t ≤ t2 conditioned on both of its states xt1 and xt2 . The challenge in making probabilistic inferences about an agent-based model is that we have to deal with an exploding state space -- for just a simple task of tracking the binary states of 50 agents, we must cope with 250 combinatorial states because the agents interact with one another. And, of course, a real-world system is much larger. To cope with this exploding state space, we use mean field approximation: the probabilistic evolution of an agent state is determined by the mean field (average) effect of the states of the other agents. The variational framework for making inferences about stochastic processes was developed in the field of ma- chine learning [45] as minimizing Bethe variational principle [4] with applications to expectation propagation [33, 24] and loopy belief propagation [34]. This paper therefore advocates that we should combine the power of big data and the power of model-thinking in the stochastic process framework. Agent-based modeling is a physicist's approach for modeling human societies when data are unavailable and experiments impossible [16], and we believe that big data will transform agent-based model- ing from speculation into a physical science. This paper also offers a solution that fits to big time-series data any agent- based model defined by a production rule system based on mean-field approximation. Hence, this system brings to- gether modelers and data miners. We have benchmarked our solution on systems of hundreds of agents, and our bench- marking gives meaningful results. The rest of this paper is organized as follows. In Section 2 we introduce a probabilistic production (rule) system to describe the microscopic dynamics of a generative model, and identify the production system as a stochastic process. In Section 3 we derive a mean-field solution to the gener- ative model under the constraint of data. In other words, given a simulator and noisy observational data about a pro- cess generated by the simulator logic, our algorithm infers the probabilities on a per-agent basis of all possible out- comes. In Section 4 we give examples and benchmark this algorithm against other algorithms. We summarize what we have accomplished and offer our speculation about big data in Section 5. 2. STOCHASTIC PROCESS INDUCED BY AGENT-BASED MODELS In this section, we introduce the stochastic kinetic model described by the Gillespie algorithm [22] to make inferences about social dynamics from information about individuals in the social system. A "stochastic kinetic model" is a chemist's way of describing the temporal evolution of a system with M agent species driven by V events (or chemical reactions) parameterized by rate constants c = (c1, . . . , cV ). At any specific time t, the populations of the species are xt = (x(1) ). An event v happens with rate hv(xt, cv), changing the populations by ∆v. The V events are mutually independent. , . . . , x(M ) t t Gillespie algorithm 1. Initialize the system at time t = 0 with rate constants c1, . . . , cV and initialize the populations of the species as x(1), . . . x(M ). 2. Simulate the time τ to the next event according to exponential distribution τ ∼ Exponential(h0(x, c) = (cid:80)V v=1 hv(x, cv)). 3. Simulate the event v according to categorical distribu- tion v ∼ Categorical( h1 , . . . , hV h0 ). h0 4. Update and output time t ← t + τ and populations x ← x + ∆v. 5. Repeat steps 2-5 until the termination condition is sat- isfied. The stochastic kinetic model specified by the Gillespie al- gorithm assigns a probability measure to a sample path of the system induced by a sequence of events v1, . . . , vn, hap- pening between time 0 and time T , 0 = t0 < t1 < ··· < tn < T , which is P (v1, . . . , vn, t1, . . . , tn, x) n(cid:89) n(cid:89) i=1 = = hvi (xti−1 , cvi ) exp hvi (xti−1 , cvi ) exp (cid:32) (cid:18) − n(cid:88) i=1 − (cid:33) h0(xti−1 , c)(ti − ti−1) (1) (cid:19) d t h0(xt, c) . i=1 The stochastic kinetic model was designed to explain the macroscopic properties of a system via the microscopic in- teractions among particles in the system [47]. If we are able to observe particles, we expect not only to improve our esti- mation of the system properties but also to make inferences about the particles. When this model is applied to a so- cial system, the particles are the individuals in the system, and the capability to make inferences about these particles becomes even more important. An event in the Gillespie algorithm looks like the following: (1) v X(1)+···+α α (1) v X(1)+···+β (M ) v X(M ), (M ) v X(M )→β (cid:81)M hv (x,ck)=cv gv (x)=cv ∆v = (β(1) v − α(1) v ,··· , β(M ) m=1 g (m) v (x(m))=cv v − α(M ) v ). (cid:81)M m=1(x(m))α (2) (3) (m) v , v individuals of species 1, α(2) When α(1) individuals of species 2 ... meet, they trigger event v with rate constant cv, which result in β(1) individuals of species 2, and so on. The rate hv(x, cv) for this event to (m) v happen is rate constant cv times a total of(cid:81)M individuals of species 1, β(2) x(m)(cid:17)α (cid:16) v v v m=1 different ways for the individuals to meet. When the components of the agent-based model X (1), ··· , X (M ) lose meaning as the populations of agent species, we can find the probability distribution of a component state among a finite partition of the component's state space. When the event rates cannot be expressed as the multipli- cations of component contributions in the form of Eq. 3, we take Taylor expansions of the event rates around the mean value of the system state hv(x, cv) = ∂αhv(Ex, cv), (cid:88) (x − Ex)α α≥0 α! with each term being in the form of Eq. 3. α is a multi-index and E is the expectation operator. Although the stochastic kinetic model is a continuous time model, we work with a discrete time stochastic model in the rest of this paper, because our goal is to track stochastic kinetic dynamics from observations of populations or indi- viduals with countably many computational steps. There are two ways to turn a continuous stochastic process into a discrete one, both of which involve Jensen's uniformiza- tion/randomization method [23]. p(x0 → xτ ) =(cid:80)∞ The first method for discretizing a continuous time stochas- tic system is through approximating the continuous time process with a discrete time process on a countable set of equally spaced time points 0, τ, 2τ, . . . , with a time interval so small that the probability of more than one event happen- ing in the interval τ is negligible. This approximation works because the state transition kernel from time 0 to time τ is according to the uniformization method, where γ is a uniformization rate, I is the identity matrix and Q is the infinitesimal generator defined by hk, k = 1, . . . , V . With γ → ∞ and γτ = 1, we get a first-order approximation of the state transition kernel I + Q · τ . exp(−γτ ) (γτ )n (cid:17)n I + Q γ (cid:16) n=0 n! Specifically, let v1, . . . , vT be a sequence of events in the discrete time stochastic kinetic system, x1, . . . , xT be a se- quence of states (populations of species), and y1, . . . , yT be a set of observations about the populations. Our goal is to make inferences about {vt, xt : t = 1, . . . T} from {yt : t = 1, . . . , T} according to the following probabil- ity measure, where indicator function 1(xt − xt−1 = ∆vt ) is 1 if the previous state is xt−1 and the current state is xt = xt−1 + ∆vt , and 0 otherwise. P (xt, yt, vtxt−1), (4) P (v1,...,T , x1,...,T , y1,...,T ) = where P (xt, yt, vtxt−1) (cid:40) = P (vtxt−1)1(xt − xt−1 = ∆vt )P (ytxt), (5) ckτ · gk (xt−1) 1 −(cid:80) j cjτ gj (xt−1) if vt = k if vt = ∅ . (6) and P (vtxt−1) = The second way to discretize a continuous time stochas- tic system is by introducing a uniformization rate γ that is faster than all event rates in Q and inspecting a discrete time Markov chain defined by the state transition matrix I + Q λ , with the transitions happening at time t1, t2, . . . , sampled according to a uniform Poisson process with rate γ. This works because according to the uniformization method the uniformised continuous time process has the same probabil- ity measure as the original process. We employ a stochastic kinetic model to simplify the state space transition kernel for several reasons. First, the stochas- tic kinetic model already successfully describes the time evo- lution of reaction systems in many areas, including chem- istry and cell biology [1, 22]. It is therefore a more natural model for describing and tracking the spatio-temporal pro- cess driven by events. Second, the event based transition T(cid:89) t=1 kernel is more general and flexible -- we can define the num- ber of events based on the complexity of real transitions. 3. MAKING INFERENCES WITH AN AGENT-BASED MODEL In this section, we derive a mean-field solution to infer the probabilities on a per-agent basis of all possible paths of system evolution, given a simulator and noisy observational data about this system generated by the simulator logic. 3.1 Variational Inference Recall the forward-backward algorithm to make inferences with a state space model [39]. Let Xt be the hidden states and yt be the observations of a discrete-time state-space model (Kalman filter and hidden Markov model) identified by a transition probability P (Xt+1Xt) and an observation model P (YtXt), where t = 1,··· , T . The forward-backward algorithm for making inferences about hidden states Xt from observations yt is comprised of a forward/filtering sweep to compute the forward statistics α(xt) = P (xty1,··· , yT ) and a backward/smoothing sweep to estimate the one-slice statistics γ(yt) = P (xt y1,··· , yT ). From the forward statis- tics and the one-slice statistics we can extract the backward statistics β(xt) = γ(xt)/α(xt) and the two-slice statistics ξ(xt, xt+1) = α(xt)P (yt+1, xt+1 xt)β(xt+1)P (yt+1y1,··· ,t). Here we follow the tradition, use upper case letters for ran- dom variables and use lower case letters for the values of random variables. t t , X (2) ,··· , X (M ) The challenge with making inferences about a non-trivial agent-based model is that we have to search in a formidable state space -- Xt = (X (1) ), where the su- perscripts 1,··· , M represent the states of the interacting elements of the system. We therefore estimate the state dis- tributions of the hidden states in an amenable state space with mean field approximation γt(xt) =(cid:81) (cid:88) minimize over ξt(xt−1,t, vt) : (x(m) t ): (7) ξt(xt−1,t, vt) log ξt(xt−1,t, vt) m γ(m) t t P (xt, vt−1, ytxt−1) t;xt−1,t;vt (cid:89) −(cid:88) (cid:89) γ(m) t (x(m) t ) log γ(m) t (x(m) t ) t;xt m m ξt(xt−1, xt, vt) = γ(m) t (x(m) t ), for all t, m, x(m) t , (8) ξt(xt−1, xt, vt) = γ(m) t−1 (x(m) t−1), for all t, m, x(m) t−1, γ(m) t (x(m) ) = 1, for all t, m, x(m) t . subject to:(cid:88) (cid:88) (cid:88) xt: fixed x t (m) t vt;xt−1,t: fixed x (m) t vt;xt−1,t: fixed x (m) t−1 (cid:88) (9) (10) We apply the method of Lagrange multipliers to solve this optimization problem, which begins with forming the La- grange function to be optimized: t;xt−1,t;vt ξt(xt−1,t, vt) log −(cid:88) (cid:89) ξt(xt−1,t, vt) P (xt, vt−1, ytxt−1) (cid:89) γ(m) t (x(m) t ) log γ(m) t (x(m) t ) (11) t;xt m m (cid:88) (cid:88) (m) t t;m;x + + t;m;x (m) t−1 (cid:88) (cid:88) β(m) t,x (m) t ( ξt(xt−1, xt, vt) − γ(m) t (x(m) t )) vt;xt−1,t: fixed x (m) t α(m) t−1,x (m) t−1 ξt(xt−1, xt, vt) − γ(m) t−1 (x(m) t−1)). ( vt;xt−1,t: fixed x (m) t−1 t =i)) is associated with the marginal- t t t t (x (x (m) (m) (m) i β i α (m) t (m) t (m) t (x(m) (x(m) t,i ·1(x t,i ·1(x ) = α(m) )=exp((cid:80) )=exp((cid:80) Taking the derivative of the expression involving Lagrange ), we see that multipliers over ξt(xt−1, xt, vt) and γ(m) α (m) (m) ized forward probabilities, β t =i)) t is associated with the marginalized backward probabilities, with γ(m) ). The dual opti- mization problem is to find the marginal forward statistics α(m) (x(m) ) to maximize the approximate partition function given by Eq. 12, and the solution is the fixed point of Eq. 13, where nor- malization constant Zt = P (yty1,··· ,t−1): log P (y1,··· ,T )=(cid:80) ) and the marginal backward statistics β(m) t−1)P (xt,vtxt−1)(cid:81) log (cid:80) )β(m) (x(m) (x(m) (x(m) (m) t−1(x (cid:81) (m) t (m) t (m) (x α β ) t t t t t t t t m t xt−1,t m m P (cid:17)(cid:89) (cid:16) (cid:16) y(m) t α(m) t−1 x(m) 1 Zt x(m) t−1 ξt(xt−1, xt, vt) = P (xt, vtxt−1) (cid:17) ·(cid:89) where P (xt, vtxt−1) = (cid:17)·(cid:81) t−1))·(cid:81) ·(cid:89) ckτ(cid:81) (1−(cid:80) Marginalizing Eq. 13 over all chains X (m(cid:48)) (cid:16) k ckτ(cid:81) t −x m 1(x m 1(x (m) (m) m g k (cid:40) (m) (m) t−1=∆ k t −x (m) (m) t−1=0) (m) k (m) t−1 m g (m) (x m m x ) t (12) (13) (cid:17) (cid:16) β(m) t x(m) t , (14) vt = k (cid:54)= ∅ vt = ∅. for m(cid:48) (cid:54)= m and t, we find that the solution to the Bethe variational principle is the mean field approximation of the original dynamics with marginal two-slice statistic given by Eq. 15. From the mean field approximation, α(m) can be solved by forward-backward algorithm and fixed point iteration. t,i and β(m) t,i t t , vtx(m) P (x(m) t−1) t−1) · P (y(m) t x(m) t (15) )β(m) t (x(m) t ), (m) t −x (m) t−1=∆ (m) k ) 1(x (m) t −x (m(cid:48)) (x t )β t (m) t−1=0) (m(cid:48) ) (m(cid:48) ) k )g v (m) (16) t =k(cid:54)=∅ t = ∅, v(m) (m(cid:48)) t−1 ) (x ξt(x(m) t−1, x(m) t where P (x(m) t (m) k ckτ g 1−(cid:80) (cid:80) k g(m(cid:48)) k,t−1 = , vt) = 1 Zt · α(m) t−1(x(m) , vtx(m) t−1) ∝ t−1) (cid:81) (m(cid:48) ) k,t−1·1(x g t−1) (cid:81) m(cid:48)(cid:54)=m (m) k (m) (m) (x (x ckτ g α (m(cid:48) ) x t (cid:80) x t α (m(cid:48) ) (m(cid:48) ) k,t−1 g (m(cid:48) ) m(cid:48)(cid:54)=m (m(cid:48)) t t t k x (cid:80) (m(cid:48) ) (m(cid:48)) t−1 )P (y t−1 (x (m(cid:48) ) (m(cid:48)) t−1 ≡∆ −x t−1 )P (y(m(cid:48)) α(m(cid:48)) t−1 (x(m(cid:48)) (m(cid:48) ) (m(cid:48) ) t−1 ≡0 −x x t (m(cid:48) ) (m(cid:48)) t−1 (x t−1 )P (y (m(cid:48) ) −x t−1 ≡0 α(m(cid:48)) t−1 (x(m(cid:48)) (m(cid:48) ) (m(cid:48) ) −x t−1 ≡0 x t t−1 )P (y(m(cid:48)) j,t−1 + 1 −(cid:88) (cid:80) (cid:89) g(m) (m(cid:48)) (m(cid:48) ) cjτ x t t t g(m(cid:48)) k,t−1 = (cid:88) Zt = cjτ , , x(m(cid:48)) t )β(m(cid:48)) t (x(m(cid:48)) t ) (m(cid:48)) (m(cid:48) ) t (m(cid:48) ) k )g (m(cid:48)) t−1 ) (x (x )β t x(m(cid:48)) )β(m(cid:48)) t (x(m(cid:48)) t ) t (cid:89) g(m) j,t−1. j m j m The above is a factorized stochastic kinetic model. The , vt) in Eq. 15 marginal two-slice probability ξt(x(m) t−1, x(m) t t , vtx(m) takes the same form as the coupled two-slice probability ξt(xt−1, xt, vt) in Eq. 13. The marginal state transition ker- nel P (x(m) t−1) in Eq. 16 consists of choosing an event (or no event) vt according to event probability P (vtx(m) t−1) and changing the state x(m) in a deterministic way, similar to the joint state transition kernel P (xt, vtxt−1) in Eq. 14, except that we marginalize over all x(m(cid:48)) for m(cid:48) (cid:54)= m. t Hence the solution to the above Bethe variational princi- ple through Legendre-Fenchel transform [40] is one in which the interacting elements of the system evolve their states marginally according to the average effects of the other el- ements. As such, instead of searching the joint probability space of of (X1,··· , XT ), we search the marginal probability spaces of (X (m) ,··· , X (m) ). 1 T 3.2 Graphical Model Representation The stochastic kinetic model with its distinct graphical model structure is more suitable than traditional models for modeling complex interactions in social dynamics. To illus- trate this point, we compare its graphical model with the coupled hidden Markov model. A coupled hidden Markov model (CHMM, Figure 1(a)) combines a number of conventional hidden Markov models (HMMs) to model the dynamics of interacting processes [7, 35]. In CHMM the latent state of HMM at time t depends on the latent states of all HMMs at time t − 1. Traditional ways to simplify the state transition kernel include the fac- torial hidden Markov model which decouples the inter-chain probability dependence [28] and the hidden Markov decision tree which assumes fixed and sparse inter-chain probability dependence [20]. A stochastic kinetic model has a graphical model repre- sentation different from CHMM, as shown in Figure 1(b). First, we define a set of stochastic events to summarize the complex interactions and decouple direct dependencies be- tween nodes. Second, while the system can move from any state to any other state in a CHMM, in any infinitesimal time interval no more than one out of V possible events is happening in a stochastic kinetic model. Third, conditioned on the system state x(1) t−1 describing the popula- tions of species 1, . . . , M , the latent state at the next time step x(1) could be dependent. (Consider an event that changes population m1 and population m2 simultane- , . . . , x(M ) ously.) at time t are conditionally independent given the states t−1 at time t − 1. Thus the inference algorithms x(1) t−1, . . . , x(M ) of CHMM are not applicable for modeling the complex in- teractions in social dynamics driven by events. In contrast, in a CHMM the states x(1) t−1, . . . , x(M ) , . . . , x(M ) t t t t The factorial stochastic kinetic model has a graphical model similar to the stochastic kinetic model except that it fac- torizes stochastic events to individual species (Figure 1(c)). The new graphical model further simplifies the inference al- gorithm. 3.3 Parameter Learning In order to find the rate constants cv in a stochastic ki- netic model and in a factorial stochastic kinetic model, we maximize the expected log likelihood and the Bethe entropy approximation respectively over these rate constants. The likelihood of rate constants cv in a continuous time stochastic kinetic model with respect to a sample path iden- (a) CHMM (b) SKM (c) FSKM Figure 1: In comparison of coupled hidden markov model (a), a stochastic kinetic model (b) decouples inter-chain interactions with events v and allows factorization (c). tified by events v1,··· , vn, system states to xt1 ,··· , xtn and times t1,··· , tn is given in Eq. 1, and the event rates are given in Eq. 3. To find the maximum likelihood estimate of the rate constants, we set the partial derivatives of the log-likelihood over the rate constants to 0: log P = (cid:80) log P = i i log cvi +log gvi (xti−1 )−(cid:80) (cid:80) (cid:80) −(cid:88) j cj gj (xti−1 )·(ti−ti−1), gv(xti−1 ) · (ti − ti−1), (cid:80) (cid:80) (cid:80) ´ i gv(xti−1 )(ti − ti−1) i 1(vi = v) T 0 d t gv(x(t)) i 1(vi = v) i 1(vi = v) cv = . i ∂ ∂cv cv = ´ Hence, the maximum likelihood estimate of the rate con- stants cv is such that the numbers of events that are ex- pected to happen according to the event rates along the T sample path 0 dt cvgv(x(t)) match the numbers of events i 1(vi = v) that happened in the sample path. Indicator function 1(vi = v) takes value 1 if vi = v and 0 if vi (cid:54)= v. (cid:80) The likelihood of rate constants cv in a discrete time stochas- tic kinetic model with respect to a sample path identified by events v1,··· , vT , system states x1,··· , xT and observations y1,··· , yT is given in Eq. 4. To find the maximum likelihood estimate of the rate constants, we similarly set the partial derivatives of the log-likelihood over the rate constants to 0: log P = log p(ytxt) + log p(vtxt−1)1(xt−xt−1=∆vt ) (cid:88) (cid:80) ∂ ∂cv cv = log P = t (cid:80) 1−(cid:80) (cid:80) t t 1(vt = v) cv t 1(vt = v) τ gv (xt−1)·1(vt=∅) j cj τ gj (xt−1) (cid:88) −(cid:88) t t τ gv(xt−1)1(vt = ∅) j cjτ gj(xt−1) 1 −(cid:80) (cid:80) (cid:80) τ→0−→ ber of events that happened ((cid:80) j cjτ gj(xt−1)) and the fraction of null events ((cid:80)T (cid:80) extrapolated from the times of null events) with the num- t 1(vt = v)). As the in- terval τ approaches 0, the probability of a null event (1 − t=1 1(vt = ∅)/T ) both approach 1, and the maximum likelihood esti- mate of the rate constants in the discrete time stochastic kinetic model approaches the maximum likelihood estimate in the continuous time stochastic kinetic model. When the events v1,··· , vT and the system states x1,··· , xT are unobserved latent variables, we use the expectation max- imization (EM) algorithm to iteratively search for the rate constants that maximize the expected log likelihood over the probability distribution of the latent variables. EM is an it- erative method for finding the maximum likelihood estimate of the parameters in statistical models involving unobserved latent variables [11]. It alternates performing the expecta- tion (E) step, which constructs the expected log likelihood as a function of the parameters over the probability dis- tribution of the latent variables using the current estimate for the parameters, with the maximization (M) step, which computes the parameters to maximize the expected log like- lihood function constructed in the E step. The estimated parameters are used to determine the probability distribu- tion of the latent variables in the next E step. The expected log likelihood over the posterior probabil- ity of events v1,··· , vT and system states x1,··· , xT condi- tioned on the observations y1, . . . , yT takes the form in Eq. 17. Maximizing this expected log likelihood by setting its partial derivatives over the rate constants gives the updated estimate of rate constants in Eq. 18. Hence, the maximum likelihood estimate of the rate con- stants cv in a discrete-time stochastic kinetic model simi- larly matches the numbers of events that are expected to cv τ gv (xt−1)1(vt=∅) j cj τ gj (xt−1) , happen according to the event rates ((cid:80) 1−(cid:80) t ∂cv (cid:80) cv = t;xt−1 t ξ(vt = v) τ gv (xt−1)·ξ(xt−1,vt=∅) j cj τ gj (xt−1) t 1(vt = v) t τ gv(xt−1) . ∂E(log P ) E (log P ) = ξt(xt−1,t,vt;cold)·log P (xt,yt,vtxt−1;c) t;xt−1,t;vt ξ(vt=v) cv −(cid:88) t;xt−1 1 −(cid:80) τ gv (xt−1)ξ(xt−1,vt=∅) j cjτ gj(xt−1) (17) set = 0, (18) (cid:88) (cid:88) t = (cid:80) 1−(cid:80) x1,t-1x2,t-1x3,t-1x1,tx2,t-1x3,t-1y1,t-1y2,t-1y3,t-1y1,t-1y2,t-1y3,t-1t-1tHMM 1HMM 2HMM 3x1,t+1x2,t-1x3,t-1y1,t-1y2,t-1y3,t-1Time..................t+1x1,t-1x2,t-1x3,t-1x1,tx2,tx3,ty1,t-1y2,t-1y3,t-1y1,ty2,ty3,tt-1tHMM 1HMM 2HMM 3x1,t+1x2,t+1x3,t+1y1,t+1y2,t+1y3,t+1Time..................t+1vtvt+1x1,t-1x2,t-1x3,t-1x1,tx2,tx3,ty1,t-1y2,t-1y3,t-1y1,ty2,ty3,tt-1tHMM 1HMM 2HMM 3x1,t+1x2,t+1x3,t+1y1,t+1y2,t+1y3,t+1Time..................t+1v1,tv1,t+1v2,tv3,tv2,t+1v3,t+1(cid:80) ξ(vt = v) t τ γ(xt−1)gv (xt−1) . t;xt−1 (cid:80) τ→0−→ ) according to the event rates (cjτ gj(xt−1)) along the sample path with the expected num- As such, the rate constant cv for event v matches the expected number of times this event could have happened t cv τ gv (xt−1)ξ(xt−1,vt=∅) ((cid:80) 1−(cid:80) ber of times the events happened ((cid:80) αt−1(xt−1) =(cid:81) Using Bethe entropy approximation, t−1) and βt(xt) =(cid:81) t−1(x(m) j cj τ gj (xt−1) m α(m) ), and setting the discretization time interval τ to be small enough, the rate constants cv can be updated according to Eq. 19. t i ξ(vi = v)). m β(m) t (x(m) (cid:80) (cid:80) τ(cid:81) (cid:80) t m cv = ξ(vt = v) t (m) t−1 γ(m)(x (m) t−1)g x where ξ(vt = v) = (cid:80) j cold j τ(cid:81) m g , (x (m) v (m) t−1) (cid:81) m g(m) j,t−1+1−(cid:80) v,t−1 j cold (m) cold v j (19) τ(cid:81) m g (m) j,t−1 . Therefore, the rate constant for event v is the expected number of occurrences of this event summed over all times, divided by the total cross-section of this event also summed over all times. To summarize, we provide the variational agent-based in- ference algorithm below. Variational Inference with Gillespie Algorithm Given observations y(m) for t = 1, . . . , T and m = 1, . . . , M , and the stochastic kinetic model of a complex system de- fined by Eq. 4, find x(m) and rate constants ck for k = 1, . . . , V . , v(m) t t t • Latent state inference. , vtx(m) Iterate through the follow- ing forward pass and backward pass until convergence, where P (x(m) t−1) is given by Eq. 16. t -- Forward pass. For t = 1, . . . , T and m = 1, . . . , M , (x(m) t ) according to α(m) t−1(x(m) t−1)P (x(m) t , vtx(m) t−1)P (y(m) t x(m) t ). t (cid:88) update α(m) ) ← 1 Zt x (m) t−1,vt α(m) t (x(m) t -- Backward pass. For t = T, . . . , 1 and m = 1, . . . , M , update β(m) t−1 (x(m) t−1) according to β(m) t−1 (x(m) t−1) ← 1 Zt P (x(m) t , vtx(m) t−1)P (y(m) t x(m) t )β(m) t (x(m) t ). (cid:88) vt,x (m) t • Parameter estimation. Iterate through latent state in- ference (above) and rate constants estimate of ck ac- cording to Eq. 19, until convergence. 4. EXPERIMENTAL RESULTS In this section, we evaluate the performance of variational agent-based inference for two applications: epidemic dynam- ics and traffic dynamics. We selected these because they are important applications with significant practical value. 4.1 Epidemic Dynamics In this section, we infer the progression of common cold in a dynamic social network using an agent-based susceptible- infectious-susceptible (SIS) model at the individual level through a small numbr of volunteers who report their symp- toms, and estimate the total number of infectious individu- als. Being able to estimate the outbreak of an epidemic in advance and determine who has the highest probability of infection is important for health-care providers and health policy researchers who must optimize limited medical re- sources. Conventional agent-based epidemic simulators [17, 26, 41, 27, 15] lack the capability to infer epidemic spread- ing with symptoms observations in the social network, and thus the sample paths given by these simulators can differ significantly from the truth. In the SIS dynamics, each individual is either infectious (I) or susceptible (S), and the system has three events: a) an in infectious individual in the network infects a susceptible individual and turns that person infectious in the network with rate constant c1 (probability per unit time), b) an in- fectious individual recovers and becomes susceptible again with rate constant c2, and 3) a susceptible individual be- comes infectious by contacting an infectious individual from outside the system with rate constant c3. I + S → 2 × I, infection, rate constant = c1, I → S, recover, rate constant = c2, S → I, infection from outside, rate constant=c3. To model the SIS dynamics at the individual level with Gille- spie algorithm, we assign two "species" to each person p: I (p) ∈ {0, 1}, S(p) ∈ {0, 1} and I (p) + S(p) = 1. The proba- bility for a susceptible person p to become infectious through one unit time of contact with an infectious person q is thus h(x, c1) = c1 · s(p) · i(q) = c1. The mean field probability for the susceptible person p to become infectious is thus c1 · s(p) · EI (q) = c1 · s(p) (cid:88) (cid:88) EI (q), q∈neighbor of p q∈neighbor of p i.e., the average total number of infectious neighbors in the individual's social network times the probability of infection per infectious neighbor. If infection happens, we change S(p) from 1 to 0 and change I (p) from 0 to 1. We benchmark the performance of the variational agent- based inference algorithm using the Dartmouth College cam- pus data set [29]. This data set contains the locations of 13,888 on-campus WiFi users from April 2001 to June 2004. On top of this dynamic social network we synthesized epi- demic progression using the SIS model and set parameters such that an individual is on average infected twice per year and takes one week on average to recover. We ran- domly select 10% of individuals as volunteers who report their daily symptoms and from these we infer the daily in- fectious/susceptible state of the other 90% individuals. As far as we know, there is no real data set with both a large amount of sensor data and symptom reports; we hope our research encourages data collection and analysis in this di- rection. Figure 2(a) compares the receiver operating characteristic (ROC) curve of predicting whether an individual is infec- tious using either a variational agent-based inference algo- rithm or a support vector classifier [37]. The support vec- (and the parameters specifying the modeling details). From these data, the simulator will proceed to execute travel, to score daily itineraries, and to perturb daily itineraries in an attempt to improve them, and then repeat the three steps until equilibrium is reached. Simulation without continu- ous observtations about a real-world transportation network however doesn't tell us whether today's traffic jams will be formed earlier or last longer. Many existing algorithms to track and predict real-time traffic dynamics on the other hand -- vector ARIMA [3], state space models [46], neural networks [48], and Bayesian network models [25, 13] -- have difficulty in coping with noisy and missing data, with making predictions in non- recurrent scenarios, and with explaining predictions in terms of agent trips. According to Vlahogianni [43], the challenge in short-term traffic forecasting is not only to predict but also to explain phenomena at the city network level -- to fuse new data sources such as those from telematics units and to easily incorporate the effects of non-recurrent condi- tions. To join the event model of a discrete-event simulator with continued observations about real-world social systems, we make use of the fact that all discrete event simulators (at least, to the best of our knowledge) have a way to dump the events happening in a simulation run. As such, we can re- construct simulation runs according to the event sequences and so reconstruct the stochastic discrete-event model from simulation runs outside the simulator, instead of hacking the source code of a specific simulator over many man-months. MATSIM, for example, has about 140 thousand lines of code, and hacking its source code to make real-time infer- ences with real-world data wouldn't be easy. In this way, we dump four events: vehicle leaving a build- ing, vehicle entering a link, vehicle leaving a link and vehicle entering a building. From these four events, we have con- structed a data frame representing continued observations of the locations (the link or building) of all vehicles at equally spaced time steps. From the data frame we constructed a state transition matrix to represent vehicle dynamics (with each row and column representing a link/building), along with entries giving the state transition probabilities accord- ing to how long a vehicle stays at a link/building and how frequently a vehicle chooses the next link/building. By uni- formly sampling a given fraction of tracked vehicles, we have constructed an observation model that provides the proba- bility distribution of observed vehicles at a location given the total number of vehicles there. This system has only one event pi ◦ lj → pi ◦ lk, a vehicle i moving from link/building j to link/building k with rate constant plj ,lk , changing the location of vehicle from X (pi) t+1 = lk, chang- ing the number of vehicles on link li from X (li) to t − 1, and changing the number of vehicles on X (li) link lk from X (lk) t +1. But the infer- ence is much more complex due to the interactions between vehicles and links. = lj to X (pi) t+1 = x(li) to X (lk) t+1 = x(lk) t = x(li) t t t = x(lk) t We employ the mobility traces of more than 500 taxi cabs collected over 30 days in the San Francisco metropolitan area [36] to benchmark the advantages of considering possible future traffic conditions for individual transportation plan- ning. We extract the road network from OpenStreetMap, a collaborative project to create a free editable map of the world. We obtain population distribution and daily trip (a) Detecting infections (b) Estimating infectious population Figure 2: Statistical inference in Dartmouth data tor classifier (SVC) estimates the probability that an indi- vidual is infectious based on the percentage of his contacts reporting symptoms. Since only 10% of individuals are re- porting symptoms in the experiment setup, the likelihood that an individual is infected can be only roughly estimated by SVC. In addition, infectious individuals make different contributions to epidemic progression because infectious dis- eases from randomly infected individuals go first to the hubs of a social network then spread to the other nodes from those hubs [9], and SVC has difficulty in capturing such network- related features. Variational agent-based inference, on the other hand, can correctly predict 60% of infections with only a 20% false positive rate. Figure 2(b) compares the performance in estimating the number of infectious persons in the 90% of individuals who do not report their daily symptoms from daily symptoms reported by the 10%. A scaling-based method missed the rapid increase of infectious individuals in early September and overestimated the number of infectious individuals in mid-October and November. This occurred because not all infectious individuals contribute to epidemic progression the same way. 4.2 Traffic Dynamics In this experiment, we predict road traffic up to one hour ahead of time from a large set of tracked vehicle locations in conjunction with an agent-based transportation simulator called MATSIM [32]. While tracked vehicle locations from car telematics systems are already being exploited to provide drivers with real-time traffic information, the chaotic nature of transportation networks means that an incident at one location might affect the traffic condition of another location up to a hundred miles away. A decision made according to current travel times might therefore be suboptimal, and can in certain cases even lead to global system breakdown. Researchers employ a transportation simulator to explain the macroscopic phenomena of transportation dynamics by simulating how individuals travel in a real-world transporta- tion network. Such a simulator takes three primary compo- nents as its input: a road network like the one used in a GPS navigator, a population specification that lists the lo- cation and travel of individuals on a typical day synthesized from census data and trip surveys with the number of sim- ulated vehicles matching the number of vehicles in the real world, and a control file specifying how daily itineraries are scored and how individuals improve their daily itineraries 0.20.40.60.80.00.20.40.60.8False Positive RateTrue Positive RateABMSVM(%. reported symptoms)050100150timeinfectious populationSepOctNovDecgroundtruthABMscalingIn this paper, we have developed a variational method to make inferences about a real-world system from continuous imperfect observations about the system, using an agent- based model that describes the dynamics of this system. To demonstrate the value of combining the power of big data and the power of model-thinking in the stochastic process framework, we show how we can track epidemics at the in- dividual level from only a small number of volunteers who report their symptoms, and we make short-term predictions about road traffic from sporadically observed probe vehi- cles. This is just a taste of what this powerful combination of approaches can do, and we expect to see further applica- tions and theoretical development to test the bounds of this methodology. APPENDIX The duality between Eq. 7 and Eq. 12 is a duality between maximum relative entropy and maximum pseudo-likelihood. To get the dual form of the Bethe variational problem in Eq. 12, we set the derivatives of Eq. 11 over ξt(xt−1,tvt) and γ(m) t (x(m) ) to 0. t ∂L ∂ξt(xt−1,t,vt) =log ⇒ ξt(xt−1,t,vt)∝exp = log γ(m) ∂L (x(m) t ) t ∂γ(m) ⇒ γ(m) t (x(m) t ) ∝ exp α m t,x (m) m β (m) t ξt(xt−1,t,vt) −(cid:80) p(xt,vt,ytxt−1) exp (cid:32)(cid:80) − α(m) (m) t−1 ) + 1 − β(m) p(xt,vt,ytxt−1) +1−(cid:80) (cid:80) (cid:18) (m) t−1,x (x(m) (cid:19) (m) t t,x t,x m m α t t (m) t (m) t−1,x (m) t−1 set= 0 (cid:33) , (m) β t,x (m) t set = 0 α(m) t,x (m) t + β(m) t,x (m) t . get Eq. 12. After taking ξt(xt−1,t, vt) and γ(m) ) into Eq 7, we To derive Eq. 15, we marginalize Eq. 13 over all x(m(cid:48)) t−1,t (x(m) t t for m(cid:48) (cid:54)= m. t ) (cid:80) t−1, vt, x(m) ξ(x(m) = 1 Zt ∝ ckτ(cid:81) p(xt,vtxt−1)(cid:81) (m) (cid:80) t−1,t (m(cid:48)) t−1 g(m(cid:48) )(x xt−1,t,vt:fix x m(cid:48)(cid:54)=m α (m(cid:48)) t−1,t x ,vt (m(cid:48) ) t−1 )β m α(m)(x (m) t−1)β(m)(x (m) t )p(m)(y (m) t x (m) t ) (m(cid:48) ) t (m(cid:48) ) p(y t x (m(cid:48) ) t )1(∆x (m(cid:48) ) t (m(cid:48) ) k ) =∆ ·α (m) t−1(x (m) t−1)g(m)(x (m) (m) t−1)β t (x (m) t )p(y (m) t (m) t =∆ )1(∆x (m) t x if vt = k (cid:54)= ∅, (m(cid:48) ) (m(cid:48) ) x )1(∆x t t =0) (m) k ) (20) (cid:80) ∝ (cid:81) m(cid:48)(cid:54)=m α (m(cid:48) ) t−1,t x (m(cid:48) ) t−1 (x (m(cid:48) ) t−1 )β t (m(cid:48) ) (m(cid:48) ) t (x )p(y t (m(cid:48) ) (m) t−1)β (m) t (m) t ·α −(cid:80) k (x ckτ(cid:81) m(cid:48)(cid:54)=m (cid:80) α (m(cid:48)) t−1,t x (m) )p(y t (m) t (x (m(cid:48) ) t−1 g(m(cid:48) )(x x (m(cid:48) ) t−1 )β (m) t t (m(cid:48) ) (m) t =0) (m(cid:48) ) x t p(y t )1(∆x (m(cid:48) ) (m(cid:48) ) )1(∆x t =0) ·α (m) t−1(x (m) t−1)g(m)(x (m) (m) t−1)β t (x (m) t )p(y (m) t )1(∆x (m) t x if vt = ∅. (m) t =0) (21) After deviding Eq. 20 and Eq. 21 by (m(cid:48) ) t−1 (x (m(cid:48) ) t−1 )β t (m(cid:48) ) (m(cid:48) ) t (x (m(cid:48) ) )p(y t x (m(cid:48) ) t )1(∆x (m(cid:48) ) t =0) (cid:81) m(cid:48)(cid:54)=m (cid:80) α (m(cid:48) ) t−1,t x we get Eq. 15, g(m) k,t−1, g(m) k,t−1 and Zt. Figure 3: Considering future traffic reduces the rel- ative error of travel time versus actual travel time from 35% when future traffic is not considered to 29% when it is. statistics from the U.S. Census, and obtain the state transi- tion matrices from one link/building to another at equilib- rium from simulations. We map the latitudes and longitudes of tracked vehicle locations to links and buildings because our stochastic inference is at the link/building level. To benchmark how the estimation of travel time can be improved by considering future traffic conditions even though traffic conditions vary significantly, we extract the 20% of trips with the highest difference between actual travel time and estimated travel time according to traffic at the time of departure, then estimate the average travel time according to our probabilistic model. Such travel often occurs at the rising edge of overall traffic volume, during bad weather, and on less developed roads. Figure 3 compares these estimated travel times according to only traffic at the time of departure with the same data considering possible future traffic estimations through a ran- dom sample of 128 trips. The travel estimations that do not project future traffic exhibit on average a 35% relative error (in comparison with a 29% when projecting future traffic), and therefore occasionally differ significantly from the actual travel time. As such, we can combine the sporadically observed vehicle locations with the large compilation of typical trip plans to continuously estimate current and future traffic conditions. Starting from the number and behavior of tracked vehicles in a road link, we can determine the total number of vehicles in the link by scaling and estimating traffic conditions. If we trace the origins and destinations of the estimated number of vehicles through the factorial stochastic process model (filling any gaps with prior individual travel behaviors), we can extract information about the traffic at other road links. If we then iterate estimations between the traffic at links and the trip choices of simulated vehicles, we improve our estimation of both. 5. CONCLUSIONS AND DISCUSSIONS llllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllll102030405060102030405060actual leg time (minutes)estimated leg time (minutes)lnot considering future trafficconsidering future trafficREFERENCES [1] Adam Arkin, John Ross, and Harley H McAdams. Stochastic kinetic analysis of developmental pathway bifurcation in phage λ-infected escherichia coli cells. Genetics, 149(4):1633 -- 1648, 1998. [2] M. Batty. Cities and Complexity: Understanding Cities With Cellular Automata, Agent-Based Models, and Fractals. Mit Press, 2007. [3] Moshe Ben-Akiva, Michel Bierlaire, Haris Koutsopoulos, and Rabi Mishalani. Dynamit: a simulation-based system for traffic prediction. In DACCORS short term forecasting workshop, The Netherlands. Citeseer, 1998. [4] HA Bethe. Statistical theory of superlattices. In Proc. Roy. Soc. London A, volume 150, pages 552 -- 575, 1935. [5] Vincent D Blondel, Adeline Decuyper, and Gautier Krings. A survey of results on mobile phone datasets analysis. EPJ Data Science, 4(1):1 -- 55, 2015. [6] Andrei Borshchev. The Big Book of Simulation Modeling: Multimethod Modeling with AnyLogic 6. AnyLogic North America, 2013. [7] Matthew Brand, Nuria Oliver, and Alex Pentland. Coupled hidden markov models for complex action recognition. In Proc. of CVPR, pages 994 -- 999, 1997. [8] Claudio Castellano, Santo Fortunato, and Vittorio Loreto. Statistical physics of social dynamics. Reviews of Modern Physics, 81(2):591 -- 646, 2009. [9] Nicholas A. Christakis and James H. Fowler. The spread of obesity in a large social network over 32 years. The New England Journal of Medicine, 357:370 -- 379, 2007. [10] Yves-Alexandre de Montjoye, Zbigniew Smoreda, Romain Trinquart, Cezary Ziemlicki, and Vincent D Blondel. D4d-senegal: the second mobile phone data for development challenge. arXiv preprint arXiv:1407.4885, 2014. [11] Arthur P Dempster, Nan M Laird, and Donald B Rubin. Maximum likelihood from incomplete data via the em algorithm. Journal of the royal statistical society. Series B (methodological), pages 1 -- 38, 1977. [12] Wen Dong, Bruno Lepri, and Alex Sandy Pentland. Modeling the co-evolution of behaviors and social relationships using mobile phone data. In Proceedings of the 10th International Conference on Mobile and Ubiquitous Multimedia, pages 134 -- 143. ACM, 2011. [13] Wen Dong and Alex Pentland. A network analysis of road traffic with vehicle tracking data. In AAAI Spring Symposium: Human Behavior Modeling, pages 7 -- 12, 2009. [14] Wen Dong, Alex Pentland, and Katherine A. Heller. Graph-coupled hmms for modeling the spread of infection. In Nando de Freitas and Kevin P. Murphy, editors, UAI, pages 227 -- 236. AUAI Press, 2012. [15] Stefan B Edlund, Matthew A Davis, and James H Kaufman. The spatiotemporal epidemiological modeler. In Proceedings of the 1st ACM International Health Informatics Symposium, pages 817 -- 820. ACM, 2010. [16] Joshua M. M. Epstein. Generative Social Science: Studies in Agent-Based Computational Modeling (Princeton Studies in Complexity). Princeton University Press, 2007. [17] Stephen Eubank, Hasan Guclu, VS Anil Kumar, Madhav V Marathe, Aravind Srinivasan, Zoltan Toroczkai, and Nan Wang. Modelling disease outbreaks in realistic urban social networks. Nature, 429(6988):180 -- 184, 2004. [18] Kai Fan, Marisa Eisenberg, Alison Walsh, Allison Aiello, and Katherine Heller. Hierarchical graph-coupled hmms for heterogeneous personalized health data. In Proceedings of the 21th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, pages 239 -- 248. ACM, 2015. [19] Jay Wright Forrester and Jay W Forrester. Urban dynamics, volume 114. mIt press Cambridge, 1969. [20] Zoubin Ghahramani and Michael I Jordan. Factorial hidden markov models. Machine learning, 29(2-3):245 -- 273, 1997. [21] Nigel Gilbert. Agent-Based Models (Quantitative Applications in Social Sciences). Sage Publications, Inc., 2007. [22] Daniel T Gillespie. Stochastic simulation of chemical kinetics. Annu. Rev. Phys. Chem., 58:35 -- 55, 2007. [23] Winfried Karl Grassmann. Transient solutions in markovian queueing systems. Computers & Operations Research, 4:47 -- 53, 1977. [24] Tom Heskes and Onno Zoeter. Expectation propogation for approximate inference in dynamic bayesian networks. In Adnan Darwiche and Nir Friedman, editors, UAI, pages 216 -- 223. Morgan Kaufmann, 2002. [25] Eric J Horvitz, Johnson Apacible, Raman Sarin, and Lin Liao. Prediction, expectation, and surprise: Methods, designs, and study of a deployed traffic forecasting service. arXiv preprint arXiv:1207.1352, 2012. [26] Lars Hufnagel, Dirk Brockmann, and Theo Geisel. Forecast and control of epidemics in a globalized world. Proceedings of the National Academy of Sciences of the United States of America, 101(42):15124 -- 15129, 2004. [27] Lorenzo Isella, Juliette Stehl´e, Alain Barrat, Ciro Cattuto, Jean-Fran¸cois Pinton, and Wouter Van den Broeck. What's in a crowd? analysis of face-to-face behavioral networks. Journal of theoretical biology, 271(1):166 -- 180, 2011. [28] Michael I Jordan, Zoubin Ghahramani, and Lawrence K Saul. Hidden markov decision trees. In Proc. of NIPS, pages 501 -- 507, 1996. [29] David Kotz, Tristan Henderson, Ilya Abyzov, and Jihwang Yeo. CRAWDAD data set dartmouth/campus (v. 2007-02-08). Downloaded from http://crawdad.org/dartmouth/campus/, 2007. [30] Kalev Leetaru and Philip A Schrodt. Gdelt: Global data on events, location, and tone, 1979 -- 2012. In ISA Annual Convention, volume 2, page 4, 2013. [31] Thomas Liggett. Interacting particle systems, volume 276. Springer Science & Business Media, 2012. [32] MATSim development team (ed.). MATSIM-T: Aims, approach and implementation. Technical report, IVT, ETH Zurich, Zurich, 2007. [33] Thomas P. Minka. Expectation propagation for approximate bayesian inference. In Jack S. Breese and Daphne Koller, editors, UAI, pages 362 -- 369. Morgan Kaufmann, 2001. [34] Kevin P Murphy, Yair Weiss, and Michael I Jordan. Loopy belief propagation for approximate inference: An empirical study. In Proceedings of the Fifteenth conference on Uncertainty in artificial intelligence, pages 467 -- 475. Morgan Kaufmann Publishers Inc., 1999. A high-resolution human contact network for infectious disease transmission. Proceedings of the National Academy of Sciences, 107(51):22020 -- 22025, 2010. [42] Laron Smith, Richard Beckman, and Keith Baggerly. Transims: Transportation analysis and simulation system. Technical report, Los Alamos National Lab., NM (United States), 1995. [35] Wei Pan, Wen Dong, Manuel Cebrian, Taemie Kim, [43] Eleni I Vlahogianni, Matthew G Karlaftis, and John C James H Fowler, and Alex Sandy Pentland. Modeling dynamical influence in human interaction: Using data to make better inferences about influence within social systems. Signal Processing Magazine, IEEE, 29(2):77 -- 86, 2012. [36] Michal Piorkowski, Natasa Sarafijanovic-Djukic, and Matthias Grossglauser. CRAWDAD data set epfl/mobility (v. 2009-02-24). Downloaded from http://crawdad.org/epfl/mobility/, February 2009. [37] John Platt et al. Probabilistic outputs for support vector machines and comparisons to regularized likelihood methods. Advances in large margin classifiers, 10(3):61 -- 74, 1999. [38] Neeti Pokhriyal, Wen Dong, and Venu Govindaraju. Virtual networks and poverty analysis in senegal. arXiv preprint arXiv:1506.03401, 2015. Golias. Short-term traffic forecasting: Where we are and where we're going. Transportation Research Part C: Emerging Technologies, 43:3 -- 19, 2014. [44] Paul Waddell. Urbansim: Modeling urban development for land use, transportation, and environmental planning. Journal of the American Planning Association, 68(3):297 -- 314, 2002. [45] Martin J. Wainwright and Michael I. Jordan. Graphical models, exponential families, and variational inference. Foundations and Trends in Machine Learning, 1(1-2):1 -- 305, 2008. [46] Yibing Wang, Markos Papageorgiou, and Albert Messmer. Real-time freeway traffic state estimation based on extended kalman filter: Adaptive capabilities and real data testing. Transportation Research Part A: Policy and Practice, 42(10):1340 -- 1358, 2008. [39] Lawrence R Rabiner. A tutorial on hidden markov models and selected applications in speech recognition. Proceedings of the IEEE, 77(2):257 -- 286, 1989. [47] Darren James. Wilkinson. Stochastic modelling for systems biology. Taylor & Francis, Boca Raton, FL, 2006. [40] Ralph Tyrell Rockafellar. Convex analysis. Princeton [48] Hongbin Yin, S C Wong, Jianmin Xu, and CK Wong. university press, 2015. [41] Marcel Salath´e, Maria Kazandjieva, Jung Woo Lee, Philip Levis, Marcus W Feldman, and James H Jones. Urban traffic flow prediction using a fuzzy-neural approach. Transportation Research Part C: Emerging Technologies, 10(2):85 -- 98, 2002.
1402.3735	3	1402	2019-10-03T17:30:27	Decentralized Goal Assignment and Safe Trajectory Generation in Multi-Robot Networks via Multiple Lyapunov Functions	[ "cs.MA", "cs.RO", "eess.SY" ]	This paper considers the problem of decentralized goal assignment and trajectory generation for multi-robot networks when only local communication is available, and proposes an approach based on methods related to switched systems and set invariance. A family of Lyapunov-like functions is employed to encode the (local) decision making among candidate goal assignments, under which a group of connected agents chooses the assignment that results in the shortest total distance to the goals. An additional family of Lyapunov-like barrier functions is activated in the case when the optimal assignment may lead to colliding trajectories, maintaining thus system safety while preserving the convergence guarantees. The proposed switching strategies give rise to feedback control policies that are computationally efficient and scalable with the number of agents, and therefore suitable for applications including first-response deployment of robotic networks under limited information sharing. The efficacy of the proposed method is demonstrated via simulation results and experiments with six ground robots.	cs.MA	cs	Decentralized Goal Assignment and Safe Trajectory Generation in Multi-Robot Networks via Multiple Lyapunov Functions Dimitra Panagou, Matthew Turpin and Vijay Kumar 9 1 0 2 t c O 3 ] A M . s c [ 3 v 5 3 7 3 . 2 0 4 1 : v i X r a Abstract -- This paper considers the problem of decentralized goal assignment and trajectory generation for multi-robot net- works when only local communication is available, and proposes an approach based on methods related to switched systems and set invariance. A family of Lyapunov-like functions is employed to encode the (local) decision making among candidate goal assignments, under which a group of connected agents chooses the assignment that results in the shortest total distance to the goals. An additional family of Lyapunov-like barrier functions is activated in the case when the optimal assignment may lead to colliding trajectories, maintaining thus system safety while preserving the convergence guarantees. The proposed switching strategies give rise to feedback control policies that are com- putationally efficient and scalable with the number of agents, and therefore suitable for applications including first-response deployment of robotic networks under limited information shar- ing. The efficacy of the proposed method is demonstrated via simulation results and experiments with six ground robots. I. INTRODUCTION Task (target) assignment problems in multi-agent systems have received great interest within the robotics and controls communities in the past couple of years, in part because they encode the accomplishment of various objectives in applications such as surveillance, exploration and coverage [1]. A common thread towards solving assignment prob- lems is the development of algorithms that assign targets to agents by optimizing a predefined criterion [2] -- [6]. Standard goal assignment strategies are the greedy assignment [7], the iterative assignment [8], the multiple Traveling Salesman assignment [9], the MinPoS assignment [10] and market-based mechanisms [11] -- [13]. The main focus in this line of research is mostly related to the optimization aspects of the problem and the associated computational complexity, so that suboptimal solutions with certain performance guarantees are derived. A hybrid control solution to multi-robot goal assignment is given in [14], while a game-theoretic approach is presented in [15]. More recently, Dimitra Panagou is with the Department of Aerospace Engineering, Uni- versity of Michigan, Ann Arbor, MI, USA. [email protected] Matthew Turpin was with the GRASP Lab, School of Engineering and Ap- plied Science, University of Pennsylvania, Philadelphia, PA, USA when this work was done. He is now with Qualcomm. [email protected] is with the GRASP Lab, School of Engineering and Applied Science, University of Pennsylvania, Philadelphia, PA, USA. [email protected] Vijay Kumar We gratefully acknowledge the support of ARL grant W911NF-08-2- 0004, ONR grant N00014-09-1-1051, and TerraSwarm, one of six centers of STARnet, a Semiconductor Research Corporation program sponsored by MARCO and DARPA. the problem of assigning a group of mobile robots to an equal number of static targets under the assumption that the robots have limited range-based communication and target-sensing capabilities was studied in [16], and in [17], respectively. Nevertheless, such optimization-based techniques are often derived under relaxed assumptions, such as point or static agents, which are in principle not acceptable in realistic settings. For instance, first-response or search-and-rescue mis- sions using robotic agents (e.g., aerial robots) typically involve multiple tasks that, on one hand, may need to be performed as quickly as possible, yet on the other hand they can be accomplished only when information sharing is available, and under tight safety guarantees. In response to this challenge, in our earlier work [18], [19] we proposed a centralized approach on the Concurrent Assignment and Planning of Trajectories (CAPT) for multi-robot systems, that produces minimum distance traveled trajectories for non-point robots, that are ad- ditionally collision-free under mild assumptions on the initial locations of the robots and the locations of the goals. Hence we simultaneously addressed two challenges: (i) the combina- torially complex problem of finding a suitable assignment of robots to goal locations, and (ii) the generation of collision- free trajectories for every robot under given assumptions. More specifically, we considered unlabeled (interchangeable) robots and goal locations, and proposed algorithmic solutions for the centralized CAPT problem, as well as a decentralized implementation of the centralized algorithm that produces suboptimal solutions in terms of total distance traveled. It should be noted that although the combinations of N robots to N goals result in intractable enumeration for all but the smallest problems, fortunately, the O(N 3) Hungarian Algo- rithm from the Operations Research community [20] optimally solves this problem in a centralized manner, while extensions to distributed systems have been also developed [21]. An extensive treatment of assignment problems and algorithms can be found in [22]. Our simulated teams of N robots under the CAPT algorithm on a single computer take approximately 10−7N 3 seconds to compute the optimal assignment, i.e., a team of 100 robots takes about 0.1 seconds to plan its trajectories. This is efficient for the considered multi-robot applications. However, the decentralized implementation of the CAPT algorithm does not in general ensure safety for the multi-robot system, i.e., that the trajectories of the robots are collision-free. Let us note that some technical ingredients of the problem considered here might be applicable or useful to other types of decentralized/distributed multi-robot planning and control problems as well, such as, for instance, robotic team reconfig- uration [23], control of modular robots, or formation/flocking control. However, the main novel characteristics that differ- entiate our work from similar problems and approaches are: (i) the casting of both the assignment and safety problems for a dynamic multi-agent/multi-robot system into a switched systems stability problem, and (ii) the development of decen- tralized feedback control algorithms that generate convergent and collision-free trajectories with provable guarantees. The contributions of this work are described in more detail in the following section. A. Overview and Contributions In this paper we build upon our earlier work [18], [19] that considers a computationally efficient way of safely assigning goals in a multi-robot system in a centralized manner, and we focus on the problems of concurrently assigning goals and generating collision-free trajectories for the decentralized case, i.e., for when the robots have limited sensing and communication capabilities and hence exchange information only locally, i.e., when they become connected. We develop algorithms that concurrently address the problems of: (P1) Assigning goal locations to agents by minimizing a cost function expressing distance to the goals. (P2) Designing feedback control policies guaranteeing: (i) the convergence of the agents to their assigned goals, (ii) that the resulting trajectories are collision-free. The key specifications in the proposed formulation are: (S1) Agents and goal locations are interchangeable, which means that the mission is considered accomplished when each agent has converged to some goal location. (S2) Information exchange between a pair of agents is feasible and reliable only when they lie within a certain commu- nication range, which means that the decision making on the optimal goal assignment (in terms of distance to the goals) can be performed only locally, i.e., in a decentralized fashion. (S3) Agents are modeled as non-point robots. The adopted decentralized formulation and the underlying goal assignment and control design techniques in the current paper offer several improvements compared to the earlier work of the authors' in the conference paper [24], namely: The conservative conditions on the initial positions of the agents and of the goal locations in [24] are now relaxed via the design of vector-field controllers instead of gradient-based controllers. The proposed vector fields are always well-defined everywhere in each agent's state space, opposed to the barrier functions utilized in [24] that exhibit some pathological situations when the distances between neighbor agents and goal locations become smaller than a certain distance. This way, the overall switched system can be implemented by forcing the agents switch among two control policies (described below), instead of three in [24], where the third controller was taking care of the pathological situations of the barrier functions. In addition, the initial positions of the robots can be selected to be R apart, compared to the conference version that required them to be 2 apart. More specifically, the novelties of the proposed 2R approach are as follows: √ is, that Contribution 1: We consider the goal assignment and trajec- tory generation problem posed in [19] that seeks to minimize the distance traveled by all robots, and formulate it into a switched systems framework [25]; this provides a novel and natural way of encoding multiple tasks via Lyapunov-like functions, and obtaining guarantees on the accomplishment of the encoded control objectives. We build our approach based on multiple Lyapunov-like functions [26] and set invariance [27] techniques. We first encode the decision making on the optimal goal assignment (which in the sequel we call the Optimal Goal Assignment (OGA) policy) as a state-dependent switching logic among a family of candidate Lyapunov-like functions. Each Lyapunov-like function encodes a cost under a candidate goal assignment, the sum of squared distances to the goals. The switching logic dictates that, when a set of agents becomes connected at some time instant t, they decide to switch to the Lyapunov-like function of minimum value at time t. We show that this decision making gives rise to a Globally Asymptotically Stable (GAS) switched system which furthermore does not suffer from Zeno trajectories. Then, based on our recent work in [28], [29], we build an additional state-dependent switching logic which employs a family of Lyapunov-like barrier functions and vector fields encoding both inter-agent collision avoidance and convergence to the goal locations determined by the OGA policy. This control policy (in the sequel called the Collision Avoidance (CA) policy) provides sufficient conditions on determining whether the OGA policy is safe, and furthermore serves as a supervisor that takes action only when safety under the OGA policy is at stake. We show that the switching between the OGA policy and the CA policy results in safe and almost globally attractive trajectories for the multi-robot system, at the expense of possibly resorting to suboptimal paths in terms of total distance traveled; this situation appears only in the cases when the CA policy forces the agents to deviate from their optimal (straight line) paths to the goals, in order to maintain system safety. To the best of our knowledge, the proposed approach of encoding assignment and motion tasks via Lyapunov-like functions is a novel approach in multi- robot task and motion planning, and introduces a framework of correct-by-construction switched (hybrid) systems that satisfy the encoded tasks, along with the feedback control policies that accomplish the encoded tasks. Contribution 2: Part of this work appeared in [24]. Com- pared to the conference paper, here we develop a different conflict resolution and collision avoidance strategy that is based on directly defining feedback vector fields for each one of the agents, instead of barrier functions. The method generates safe reference directions and feedback controllers for the agents, and alleviates (i) the issues often raised in pure gradient-based solutions, where the behavior of the gradient- based controller is hard to tune, as well as (ii) the use of the third family of Lyapunov-like functions in the switching strategy in [24]. Furthermore, (iii) the new design provides safety and convergence guarantees for the decentralized multi- robot system under less restrictive assumptions on the spatial distributions of the agents and the goal locations, compared to the assumptions adopted in [19], [24]. Finally, (iv) we provide a robustness result on the safety of the resulting multi-robot trajectories in the case of communication failures, and (v) we implement and verify the effectiveness of the proposed algorithms experimentally with six ground robots. The paper is organized as follows: Section II presents the mathematical formulation of the problem, while the adopted notions and tools from switched systems theory are reviewed in Section III. The proposed goal assignment and collision avoidance policies are given in Sections IV and V, respec- tively, along with the mathematical proofs that verify their correctness. Simulation results to evaluate their efficacy are included in Section VI, along with a comparison with the results in the conference version [24], and a discussion on the communication requirements and planning times in the connected groups of agents. Experimental results with six ground robots are provided in Section VII. Section VIII summarizes our conclusions and thoughts on future research. II. PROBLEM FORMULATION We consider N agents i and equal number of goal locations Gi, i ∈ N = {1, 2, . . . , N}. The objective of each agent i is to safely converge to some goal Gi, while collaborating with its locally connected agents (see definition below) so that their total traveled distance is minimized. The motion of each agent i is governed by single integrator dynamics: where ri =(cid:2)xi ri = ui, (cid:3)T is the position vector of agent i with (1) yi respect to (w.r.t.) a global Cartesian coordinate frame G, and ui is its control vector comprising the velocities uxi, uyi w.r.t. the frame G. We denote the speed of agent i, i.e., the magnitude of the velocity vector ui, as ui = (cid:107)ui(cid:107) = (cid:113) yi. u2 xi + u2 We assume that each agent i: (i) has access to its position ri and velocities ui via onboard sensors, (ii) can reliably exchange information with any agent j (cid:54)= i which lies within its communication region Ci : {ri ∈ R2, rj ∈ R2 (cid:107)ri − rj(cid:107) ≤ Rc}, where Rc is the communication range. In other words, a pair of agents (i, j) is connected as long as the distance dij = (cid:107)ri − rj(cid:107) ≤ Rc. Remark 1: Note that maintaining connectivity for the entire network is not fundamental for the proposed assignment and collision avoidance strategies; in other words, maintaining a connected graph is neither necessary, nor a control objective in the proposed problem and approach. Remark 2: The consideration of equal number of robots and goal locations is mainly due to the fact that we do not consider the problem of globally assigning goals to robots at the initial time of the deployment, but rather we consider the problem of locally re-assigning goals to robots whenever/if the robots get connected along with safety and convergence guarantees. The initial goal assignment is random and hence, suboptimal in terms of total traveled distance in general. In the case of M goal locations and N robots with M > N, our approach would be initialized by randomly assigning a goal location per robot, and from there on the selected N goals would remain the same over the evolution of the system (i.e., the remaining M − N would never be selected by any robot). Similarly, in the case of M < N, only M robots would be randomly assigned a goal location each, while the remaining N−M robots would not be assigned any goal. The task is considered completed as long as each agent converges to some goal location, with goal locations being interchangeable -- each agent can converge to any goal loca- tion. This specification defines N! possible goal assignments k ∈ {1, 2, . . . , N!}. Each goal assignment can be encoded via an assignment matrix:  α11 ... αN 1  , . . . α1 N ... ... . . . αN N Ak = {αim} = with the rows representing agents and the columns repre- senting goal locations. For agent i ∈ {1, . . . , N} assigned to goal Gm, m ∈ {1, . . . , N}, one has αim = 1 and the remaining elements of the i-th row equal to zero. For instance, for 2 agents and 2 goal locations, the possible N! = 2 goal assignments are encoded via the assignment matrices: (cid:20)1 0 (cid:21) 0 1 (cid:20)0 1 (cid:21) 1 0 . A1 = , A2 = In this case, the matrix A1 encodes that agent 1 goes to G1 and agent 2 goes to G2, while the matrix A2 encodes that agent 1 goes to G2 and agent 2 goes to G1. (cid:3)T the position vectors of the goal locations Gm w.r.t. the global frame. Then, the assigned goal position for agent i under the Ak assignment matrix (or under the k-th goal assignment) can be expressed as: Denote rGm =(cid:2)xGm yGm N(cid:88) r(k) gi = α(k) im rGm , m=1 where α(k) im is the i-th row of Ak. The next lemma establishes that under any assignment, the robots will converge to the respective goal locations. This is essential for the switching control strategy developed later. Lemma 1: The position trajectories ri(t) of agent i are the k-th goal Globally Exponentially Stable (GES) w.r.t. assignment under the control law: (cid:16) (cid:17) ri − r(k) gi , (2) i = −λi u(k) where i ∈ {1, . . . , N}, λi > 0. Proof: Consider the candidate Lyapunov function V (k) i for each agent i under the k-th assignment as: (cid:13)(cid:13)(cid:13)ri − r(k) gi (cid:13)(cid:13)(cid:13)2 , V (k) i = which encodes the distance of agent i from its goal r(k) gi under the goal assignment k. The function V (k) is continu- ously differentiable, positive definite, and radially unbounded. Furthermore, the closed-loop system vector field (1) under the control law (2) is globally Lipschitz continuous. Given that i rGi = 0, one has that the time derivative of V (k) system trajectories under the control law (2) reads: (cid:13)(cid:13)(cid:13)ri − r(k) gi i (cid:13)(cid:13)(cid:13)2 ≤ −2λiV (k) i , i = −2λi V (k) yielding that the position trajectories ri(t) of each agent i are GES w.r.t. the assigned goal r(k) gi . Furthermore, by considering the Lyapunov function: N(cid:88) i=1 V (k) = V (k) i , (3) along the Σ along with (4) completely describes the trajectory of the switched system according to the following rule: (kn, tn) means that the system evolves according to: x(t) = fkn(x(t), t) for tn ≤ t < tn+1. Equivalently, for t ∈ [tn, tn+1) one has that the kn-th subsystem is active. We assume that the switching sequence is minimal in the sense that kn (cid:54)= kn+1. Now, for any subsystem k ∈ K, denote: Σ k = {t0k, t1k, . . . , tnk, tn+1k, . . . ,} encoding the sum of distances of the agents to their assigned goals, one has that (2) render its time derivative: the sequence of switching times when the k-th subsystem is switched on or switched off. Hence, V (k) i = I(Σ k) = [tnk, tn+1k] (cid:91) n∈N N(cid:88) i=1 V (k) = −2λiV (k) {λi} N(cid:88) i=1 {λi}V (k), i ≤ −2 min = −2 min i i is GES to zero. which implies that the sum of squared distances of agents i to their assigned goals r(k) gi Problem Description: Our goal is to develop decentralized algorithms that force Nc ≤ N connected agents to decide whether they keep their current goal assignment or swap goals. The criterion on the decision making is to choose the assignment k ∈ {1, . . . , Nc!} which minimizes the sum of total distance (i.e., results in the shortest paths) to the goal locations. In the sequel, we refer to this assignment as the optimal assignment. The challenge in the case of non- point robots is to ensure that the optimal assignment results not only in stable, but also in collision-free trajectories w.r.t. the assigned goal locations. We develop a switching policy between optimal (shortest) and safe (collision-free) trajectories with provable guarantees. This policy is realized via switching between the OGA and CA policies defined in the sequel, while the switched systems and control framework is employed to facilitate the stability and safety analysis for the resulting trajectories of the multi-robot system. We first review some basic results on switched systems theory. III. NOTIONS FROM SWITCHED SYSTEMS THEORY Following [26], let us consider the prototypical example of a switched system: k ∈ K = {1, 2, . . . , K}, x(t) = fk(x(t)), (4) where x(t) ∈ Rn, each fk is globally Lipschitz continuous and the k's are picked in such a way that there are finite switches in finite time. Consider a strictly increasing sequence of times: T = {t0, t1, . . . , tn, . . . ,}, n ∈ N, the set: I(T ) =(cid:83) n∈N[tn, tn+1] denoting the interval comple- tion of the sequence T , and the switching sequence: Σ = {x0; (k0, t0), (k1, t1), . . . , (kn, tn), . . .}, time, x0 is the initial state, N is where t0 is the initial the set of nonnegative integers, and the elements in the set {k0, k1, . . . , kn, . . .} take values in K. The switching sequence is the set of times that the k-th subsystem is active. Denote E(T ) = {t0, t2, t4, . . . ,} the even sequence of T ; then, E(Σ k) = {t0k, t2k, . . . , tnk, . . . n ∈ N} denotes the sequence of the switched on times of the k-th subsystem. In the sequel, the notation tnk will be equivalent to k(tn), denoting the time instance tn when the k-th subsystem becomes active. [26] A C 1 function V : Rn → R+ with Definition 1: V (0) = 0 is called a Lyapunov-like function for a vector field f (·) and the associated trajectory x(·) over a strictly increasing sequence of times T if: (i) (ii) V is monotonically non-increasing on E(T ). [26] If for each k ∈ K, Vk is a Lyapunov- like function for the k-th subsystem vector field fk(·) and the associated trajectory over Σ k, then the origin of the system is stable in the sense of Lyapunov. If in addition, for any tp, tq ∈ E(Σ k), where p < q, Vk(tq)(x(tq)) − Vk(tp)(x(tp)) ≤ −W (x(tp)) , V (x(t)) ≤ 0, ∀t ∈ I(T ), Theorem 1: (5) holds for some positive definite continuous function W , then the origin of the system is GAS [30]. The theorem essentially states that: Given a family of functions {Vk k ∈ K} such that the value of Vk decreases on the time interval when the k-th subsystem is active, if for every k the value of the function Vk at the beginning of such interval exceeds the value at the beginning of the next interval when the k-th system becomes active (see Fig. 1), then the switched system is GAS. A fundamental assumption of the theorem is the decreasing condition of V on E(Σ k). This is quite conservative and, in general, hard to check [32]. The concept of generalized Lyapunov-like functions relaxes this condition. [32] A C 0 function V : Rn → R+ with V (0) = 0 is called a generalized Lyapunov-like func- tion for a vector field f (·) and the associated trajectory x(·) over a strictly increasing sequence of times T = {t0, t1, . . . , tn, . . . ,}, n ∈ N, if there exists a function h : R+ → R+ satisfying h(0) = 0, such that: V (x(t)) ≤ h(V (x(tn))), for all t ∈ [tn, tn+1) and all n ∈ N. Definition 2: inter-agent distance happens to remain always greater than the communication range Rc, i.e., if dij(t) > Rc, ∀t ∈ [0,∞), we say that the agents are never involved in a meeting event, and as thus, no decision on the goal assignment needs to be made. Out of Lemma 1, the position trajectories ri(t), rj(t) are GES w.r.t. the assigned goal locations. Clearly, in such cases, no inter-agent collisions occur. Assume now that the agents initiate at t0 = 0 such that dij(t0) ≤ Rc, or that at some time instant td > 0 they lie within distance dij(td) ≤ Rc under some goal assignment k ∈ {1, 2} (again, not necessarily the optimal one). We say that the agents are involved in a meeting event and a decision regarding on the goal assignment has to be made. The proposed decision making strategy is defined as follows: The agents: Fig. 1. Branicky's "decreasing sequence" condition. Image from [31]. Theorem 2: [32] Consider the prototypical switched sys- tem (4) and assume that for each k ∈ K there exists a pos- itive definite generalized Lyapunov-like function Vk(x) w.r.t. fk(x(t)) and the associated trajectory x(t), t ∈ [tn, tn+1). Then: (i) The origin of the switched system is stable if and only if there exist class GK functions αk(·) satisfying Vk(x(t2qk)) − Vk(x(t0k)) ≤ αk((cid:107)x0(cid:107)), q ≥ 1, k = 1, 2, ..., K. (6) (ii) The origin of the switched system is asymptotically stable if and only if (6) holds and there exists k such that Vk(x(t2qk)) → 0 as q → ∞. The theorem essentially states that stability of the switched system is ensured as long as Vk(x(t2qk))−Vk(x(t0k)), which is the change of Vk between any "switched on" time t2qk and the first active time t0k, is bounded by a class GK function, regardless of where Vk(x(t0k)) is. Remark 3: The condition (6) is equivalently rewritten as: (cid:0)Vk(x(t2qk)) − Vk(x(t2(q−1)k))(cid:1) ≤ αk((cid:107)x0(cid:107)), ∞(cid:88) q=1 for any k ∈ K. Note that here Vk(x(t2qk))−Vk(x(t2(q−1)k)), q ≥ 1, stands for the change of Vk(x) at two adjacent "switched on" times. Thus, the condition means that Vk is allowed to grow on E(Σ k) but the total such change on any finite time interval should be bounded from above by a class GK function. Remark 4: Theorem 2 relaxes the decreasing requirement of Vk(x(t)) on the corresponding active intervals. Instead, we only need that Vk(x(t)) on an active interval does not exceed the value of some function of Vk at the "switched on" time. IV. THE OPTIMAL GOAL ASSIGNMENT POLICY A. The N = 2 Robot Case For simplicity, we first consider N = 2 agents i, j that need to safely move towards goal locations G1, G2, and build a switching logic that locally decides on the OGA. Assume that the agents initiate at t0 = 0 such that dij(t0) > Rc under a random assignment k ∈ {1, 2} (i.e., not necessarily the optimal one) and move under the control laws (2). If the 1) Exchange information on their current positions ri(td), rj(td) and goal locations rgi(td), rgj (td). 2) Compare the values of the Lyapunov functions V (k)(td) given out of (3) for each possible assignment k ∈ {1, 2}. 3) Move under: V (td) = min(cid:8)V (k)(td) : k ∈ {1, 2}(cid:9) . In other words: Without loss of generality, let k = 1 be the goal assignment before the decision making. The agents implement the following switching logic: • If V (1)(td) ≤ V (2)(td), then: V (td) = V (1)(td), i.e., they keep the same goal assignment, • If V (1)(td) > V (2)(td), then: V (td) = V (2)(td), i.e., they switch goal assignment. In the sequel, we refer to the decision making based on the logic described above as to the OGA policy. Note that in general for N > 2 agents, this policy results in locally optimal paths since it re-assigns goals among the connected agents only (which may be Nc < N in number), while it does not necessarily apply at the initial time t0, hence to the initial positions of the robots, but rather is triggered at any time td ≥ t0 that Nc ≤ N robots may happen to become connected. Problem 1: Establish that the OGA policy renders asymp- totically stable trajectories for the multi-robot system, i.e., that each agent does converge to some goal location. B. Stability Analysis on the Switched Multi-robot System We resort to control analysis tools for switched systems. The closed-loop dynamics of agent i read: ri(t) = u(k) i = −λi ri(t) − r(k) gi , (7) (cid:16) (cid:17) while similarly one may write the closed-loop dynamics for the trajectories rj(t) of agent j. The OGA policy gives rise to switched dynamics for each agent, in the sense that for k = 1 each agent moves under assignment matrix A1, while for k = 2 each agent moves under assignment matrix A2. Denote r =(cid:2)ri T(cid:3)T the state vector of the multi-robot (cid:34) (cid:35) (8) system, which is governed by the switched dynamics: r(t) = fk(r(t)), T rj where fk = , k ∈ K = {1, 2}. i u(k) u(k) j Assumption 1: We assume for now that there are only a finite number of switches per unit time. Denote n ∈ N and consider the sequence of switching times T = {t0, t1, t2, t3, . . . , tn, . . .} and the switching se- quence: Σ = {r0; (k0, t0), (k1, t1), . . . , (kn, tn), . . .}, kn ∈ K,∀n ∈ N. Assuming that the switching sequence is min- imal, imply a se- quential switching between the same subsystems1, it follows that kn (cid:54)= kn+1. Without loss of generality, assume that k0 = 1. Then the (minimal) switching sequence reads: Σ = {r0; (1, t0), (2, t1), (1, t3), (2, t4), . . .}. the switching signal does not that i.e., gi , r(k) gj . Theorem 3: The trajectories r(t) of the switched multi- robot system (8) are GAS w.r.t. the goal assignment k under the OGA policy. This implies that agents i, j converge to their goal locations r(k) Proof: We consider the candidate Lyapunov-like func- tions V (k), k ∈ {1, 2}, given out of (3), which encode the motion of the agents under the goal assignment k. Out of Lemma 1 one has that each individual k-th subsystem, i.e., the motion of the multi-robot system under the k-th assignment, is GES i.e., each V (k) is decreasing during the time intervals when the k-th subsystem is active. Furthermore, the OGA policy dictates that at the switching times {t0, t1, t2, t3, . . . ,} one has: V (1)(t0) > V (2)(t1) > V (1)(t2) > V (2)(t3) > . . . , i.e., the value of each Lyapunov-like function V (k) at the beginning of the time intervals when the k-th subsystem becomes active satisfies the decreasing condition. This proves that the switched system is stable. To draw conclusions on asymptotic stability, consider the time intervals [t2n, t2n+1), [t2n+1, t2n+2), n ∈ N, when the k = 1 and k = 2 subsystems are active, respectively. Then the following inequalities hold: V (1)(t0) > V (1)(t1), V (1)(t1) ≥ V (2)(t1), V (2)(t1) > V (2)(t2), V (2)(t2) ≥ V (1)(t2), V (1)(t2) > V (1)(t3), V (1)(t3) ≥ V (2)(t3), ... since the subsystem 1 is active on [t0, t1) and GES out of the OGA Policy since the subsystem 2 is active on [t1, t2) and GES out of the OGA Policy since the subsystem 1 is active on [t2, t3) and GES out of the OGA Policy which imply that: V (1)(t2n+2) < V (1)(t2n), ∀n, i.e.,: 0 < ρ1 < 1. Then: V (1)(t2n+2) = ρ1V (1)(t2n), V (1)(t2n+2) − V (1)(t2n) = −(1 − ρ1) V (1)(t2n) , (cid:124) (cid:123)(cid:122) (cid:125) W (r,r(1) g ) 1Note that this condition is met in our case since in the stability analysis we consider only those switching times when the goal assignment does change. (cid:16)(cid:107)r − r(1) g (cid:107)2(cid:17) where 1 − ρ1 > 0, and W (r, r(1) is a continuous, positive definite function w.r.t. the goals r(1) imposed by the k = 1 assignment. Similarly we obtain: V (2)(t2n+1) < V (2)(t2n+3), ∀n, which leads to g ) = g g V (2)(t2n+3) − V (2)(t2n+1) = −(1 − ρ2) V (2)(t2n+1) , (cid:124) (cid:123)(cid:122) W (r,r(2) g ) (cid:125) where 0 < ρ2 < 1. Therefore, the switched system is GAS. C. Avoiding Zeno Behavior In the sequel we establish that the switching policy does not result in Zeno trajectories [33], i.e., in trajectories that converge to some point or switching surface other than the desired equilibrium; such trajectories in practice would result in robots not converging to their respective goal locations. Given the sequence T = {t0, t1, . . . , tn, . . .}, n ∈ N, of switching times, a switched system is Zeno if there exists some finite time tZ such that: ∞(cid:88) n=0 lim n→∞ tn = (tn+1 − tn) = tZ. Zeno behavior means that the switching times have a finite accumulation point, i.e., that infinite amount of switchings occurs in a finite time interval. In general, the task of detecting possible Zeno trajectories and extending them beyond their accumulation points is far from trivial [25] and depends on the problem at hand. tn+1 − tn = 0 for all n > C. The definition above results in two qualitatively different types of Zeno behavior. A Zeno switched system is [33]: (i) Chattering Zeno if there exists a finite C ∈ N such that (ii) Genuinely Zeno if tn+1 − tn > 0 for all n ∈ N. The difference between these is prevalent especially in their detection and elimination. Chattering Zeno behavior results from the existence of a switching surface on which the vector fields "oppose" each other. This behavior can be eliminated by defining either (i) Filippov solutions which force the flow to "slide" along the switching surface; this results in a sliding mode behavior, or (ii) hysteresis switching, in order to, not only approximate a sliding mode, but also to maintain the property that two consecutive switching events are always separated by a time interval of positive length. The proposed OGA policy dictates that a switch among vector fields fk may occur when (cid:107)ri(t) − rj(t)(cid:107) ≤ Rc, i.e., when the multi-robot system trajectories r(t) hit the surface time instant when the system trajectories lie on the surface Sc(td) and the agents are involved in the decision making, t− d , t+ d the time instants before and after the decision making, Sc(t) := (cid:8)r ∈ R2 N (cid:107)ri(t) − rj(t)(cid:107) = Rc respectively, and k(cid:0)t− (cid:1) = k(cid:0)t+ k(cid:0)t− (cid:9). Denote td the (cid:1) the assignment before and (cid:1), k(cid:0)t+ (cid:1), in which case no switching occurs. after the decision making, respectively. The decision making on Sc(td) results in two different cases: 1) The agents decide to keep their goal assignment, i.e., d d d d 2) The agents decide to swap goals, i.e., k(cid:0)t− (cid:1) (cid:54)= k(cid:0)t+ d (cid:1) d and a switching occurs. Theorem 4: The switched multi-robot system (8) under the OGA policy does not suffer from Zeno points. d d d d yG2 yG1 (cid:1) = ∇Sc(¯r)fk (cid:0)t− Denote rG1 =(cid:2)xG1 (cid:0)t− (cid:0)t+ (cid:1) = fk (cid:1)}, and (cid:0)t+ (cid:1) = 0, where td is (cid:3)T , rG2 =(cid:2)xG2 (cid:3)T where Proof: We employ the results in [34], Theorem 2. Let us assume that the switched multi-robot system (8) has a Zeno point ¯r. Then it holds that: ¯r ∈ Sc(td), ¯r is an accumulation point of the set S = {r ∈ Sc(td) : fk satisfies: ∇Sc(¯r)fk a decision making and switching time instant. rG1 (cid:54)= rG2 the position vectors of the goal locations G1, G2, assigned to the agents i, j respectively, at time t− d . The agents implement the OGA policy at time td and decide to swap goals. (cid:21) (cid:20)−λi(ri − rG2 ) (cid:0)t+ (cid:1) = Then, the corresponding vector fields read: T(cid:3), where rij = (ri− rj). It is −λj(rj − rG1 ) (cid:1) ⇒ rG1 = (cid:0)t+ (cid:1) = fk (cid:0)t− (cid:0)t− while ∇Sc(td) =(cid:2)rij easy to verify that S = ∅, since fk rG2, a contradiction. Thus, the set of accumulation points of S is empty, which implies that no Zeno points ¯r are contained there, a contradiction. Thus, the OGA policy does not suffer from Zeno points. (cid:20)−λi(ri − rG1) (cid:21) −λj(rj − rG2) T −rij (cid:1) = , fk fk d d d d , This result establishes that the decision making under the OGA renders trajectories that do not accumulate on, i.e., always escape, the switching surface Sc(td); hence the robots do not get stuck away from their assigned goals. Finally, to maintain the property that there always will be a positive time interval τ > 0 between consecutive switching times, we implement a hysteresis-like switching logic; the idea is that after the (i, j) decision making at time td and while connected, the agents i, j do not get involved in a new (i, j) decision making, unless the agents have become disconnected first. Note that in addition, this hysteresis logic validates our Assumption 1 on finite amount of switches per unit time interval. D. The N > 2 Robot Case The results on the stability and non-Zeno behavior of the multi-robot system extend to the case of N > 2 agents. Con- sider N agents that need to converge to N goal locations. This gives rise to N! possible goal assignments k ∈ {1, 2, . . . , N!}, or in other words, N! candidate switched subsystems fk for the switched multi-robot system:  , fk = r1 ... rN u(k) ... u(k) N 1  , k ∈ K = {1, . . . , N!}. (9) r(t) = fk(r(t)), where: r = For N > 2 agents the decision making on the optimal goal assignment involves the Nc ≤ N connected agents at time td. The agents exchange information on their positions and goal locations at time td, compare the values of the Nc! Lyapunov functions V (k)(td), where k ∈ Kc(td) ⊆ K, and pick the goal assignment k which corresponds to min{V (k)(td)}. 1) Stability Analysis: Consider the sequence of switching times T = {t0, t1, . . . , tn, . . .} and the switching sequence Σ = {r0; (k0, t0), (k1, t1), . . . , (kn, tn), . . .}, kn ∈ K. Then, the reasoning and analysis followed in Theorem 3 apply to the N > 2 case as well. Theorem 5: The trajectories r(t) of the switched multi- robot system (9) are GAS w.r.t. the goal assignment k. Proof: We consider the candidate Lyapunov-like func- tions V (k), k ∈ {1, . . . , N!}, encoding the motion of the agents under goal assignment k. Out of Lemma 1, each individual k- th subsystem, i.e., the motion of the multi-robot system under the k-th assignment, is GES. This implies that each V (k) is decreasing on the time intervals that the k-th subsystem is active. Furthermore, the OGA policy dictates that at the switching times {t0, t1, t2, t3, . . . ,} one has: V (k(t0))(t0) > V (k(t1))(t1) > V (k(t2))(t2) > . . . , where k(tn) denotes the k-th subsystem that becomes active at time tn, n ∈ N; in other words, the value of each Lyapunov- like function V (k) at the beginning of the time intervals when the k-th subsystem becomes active satisfies the decreasing con- dition, i.e., the switched system is Lyapunov stable. To draw conclusions on the asymptotic stability, consider any pair of switching times tn1 < tn2 when the k-th subsystem becomes active, the corresponding time intervals [tn1, tm1 ), [tn2, tm2) while the k-th subsystem remains active, with tm1 < tn2, and follow the reasoning as in Theorem 3, i.e.,: V (k)(tn1) > V (k)(tm1 ), V (k)(tm1) ≥ V (l)(tm1), V (l)(tm1) ≥ V (k)(tn2), V (k)(tn2) > V (k)(tm2 ), since the subsystem k is active on [tn1, tm1) and GES out of the OGA Policy, l (cid:54)= k out of the OGA Policy since the subsystem k is active on [tn2 , tm2) and GES which imply that: V (k)(tm2) < V (k)(tm1), i.e., that: V (k)(tm2 ) = ρkV (k)(tm2), V (k)(tm2 ) − V (k)(tm1 ) = −(1 − ρk) V (k)(tm1 ), 0 < ρk < 1. Then: where 1−ρk > 0. Following the same reasoning as in Theorem 3, we have that the switched system is GAS. 2) Non-Zeno Behavior: Finally, following the same pattern as for the N = 2 case, one may verify that: Theorem 6: The switched multi-robot system (9) under the d (cid:0)t+ (cid:0)t− (cid:1) = fk (cid:1)}, where Sc(td) OGA policy does not suffer from Zeno behavior. Proof: Assume that the switched system (9) has a Zeno point ¯r. Then, ¯r ∈ Sc(td) and ¯r is an accumulation point of the set S = {r ∈ Sc(td) : fk is the switching surface at the decision and switching time td. The OGA policy dictates that at least one pairwise goal swap occurs at time td, since if the agents decided to keep the goals they had at time t− d , then td would not have been a switching time, a contradiction. Since at least one pair of agents (i, j) switches goal locations, denoted as rGi, rGj , the condition fk only when rGi = rGj , see the analysis in Theorem 4, i.e., (cid:1) on the switched vector fields holds true (cid:1) = fk (cid:0)t− (cid:0)t+ d d d a contradiction by construction. Thus, the set S = ∅, which furthermore implies that the set of its accumulation points is empty, implying that no Zeno points can be contained there. Thus, the switched multi-robot system (9) does not exhibit Zeno behavior. V. COLLISION AVOIDANCE POLICY The OGA policy does not ensure that inter-agent collisions, realized as keeping dij(t) ≥ ds, where ds is a safety distance, are avoided ∀t ∈ [0,∞) (see also [18], [19]). Problem 2: Establish (sufficient) conditions under which the OGA policy is collision-free. A. Detecting Conflicts We are referring to time t > td, i.e., after Nc ≤ N connected agents have decided on a goal assignment k based on the OGA policy, and move towards their goal locations r(k) gi . In the sequel we drop the notation ·(k), in the sense that the goal assignment k is kept fixed. We would like to identify a metric (a "supervisor") deter- mining whether the OGA policy results in collisions. Let us consider the collision avoidance constraint: cij(ri, rj) = (xi − xj)2 + (yi − yj)2 − ds (10) encoding that the inter-agent distance dij = (cid:107)ri − rj(cid:107) should always remain greater than a safety distance ds. 2 > 0, To facilitate the analysis using Lyapunov-like approaches, we encode the constraint (10) as a Lyapunov-like function. Recall that a barrier function [35] is a continuous function whose value on a point increases to infinity as the point approaches the boundary of the feasible region; therefore, a barrier function is used as a penalizing term for violations of constraints. The concept of Recentered Barrier Function (RBF) in particular was introduced in [36] in order to not only regulate the solution to lie in the interior of the constrained set, but also to ensure that, if the system converges, then it converges to a desired point. We define the logarithmic barrier R2 × R2 → R for the constraint (10) as: function bij(·) : T(cid:12)(cid:12)rgi bij(ri, rj) = − ln (cij(ri, rj)) , (11) which tends to +∞ as cij(·) → 0, i.e., as dij → ds. The RBF of (11) is defined as [36]: rij = bij(ri, rj) − bij(rgi, rj) − ∇bij (cid:3)T is the goal location for agent i, (ri − rgi), (12) bij(rgi, rj) is the value of (11) evaluated at rgi, ∇bij = is the gradient vector of the function bij(·), and ∇bij is the transpose of the gradient vector, evaluated at rgi. To ensure that we have a nonnegative function to be used as a Lyapunov-like function, we define [28]: wij(·) = (rij(ri, rj, rgi ))2 , where rgi = (cid:2)xgi (cid:105)T (cid:104) ∂bij T(cid:12)(cid:12)rgi ∂bij ∂yi (13) ygi ∂xi which now is a positive definite function. To furthermore encode that the position trajectories of agent i remain bounded in a prescribed region (for reasons that will be explained in the technical analysis later on), for each agent i we define the workspace constraint: ci0(ri, r0) = R2 0 − (xi − x0)2 − (yi − y0)2 > 0, which encodes that the position ri = (cid:2)xi r0 = (cid:2)x0 (cid:3)T of agent (cid:3)T and radius R0 > Rc. This can be thought i should always lie in the interior of a circle of center (14) as the workspace where the agents operate. Then, in the same reasoning followed before, we define the barrier function: y0 yi bi0(ri, r0) = − ln (ci0(ri, r0)) , (15) and its corresponding RBF: (ri − rgi), (16) ri0 = bi0(ri, r0) − bi0(rgi, r0) − ∇bi0 which vanishes only at the goal location rgi and tends to infinity as the position ri tends to the workspace boundary. To get a positive definite function we consider: wi0(·) = (ri0(ri, r0, rgi ))2 . (17) Therefore, an encoding that agent i stays ds away from all its neighbor agents j, while staying within the bounded workspace, can be now given by an approximation of the maximum function of the form: T(cid:12)(cid:12)rgi (wi0)δ + (cid:88) j∈Ni  1 δ wi = (wij)δ , (18) where δ ∈ [1,∞), and Ni ⊆ N is the set of connected (neighbor) agents j (cid:54)= i of agent i. The function wi vanishes at the goal location rgi and tends to +∞ as at least one of the terms wi0, wij tends to +∞, i.e., as at least one of the inter-agent distances dij tends to ds, or as agent i tends to the workspace boundary. Finally, to ensure that we have a Lyapunov-like function for agent i which uniformly attains its maximum value on constraints' boundaries we take: Wi = wi 1 + wi , (19) which is zero for wi = 0, i.e., at the goal position rgi of agent i, and equal to 1 as wi → ∞. For more details on the analytical construction the reader is referred to [28], [37]. The level sets of the Lyapunov-like function (19) can now be used as a sufficient criterion on determining whether the control inputs out of the OGA policy jeopardize safety. Let us consider the time derivative of (19): (cid:21) (cid:105)(cid:20) xi (cid:16) (cid:88) yi ∂Wi ∂yi j∈Ni + ∂xj j∈Ni (cid:18)(cid:104) ∂Wi (cid:88) (cid:17) , ζij (cid:44)(cid:104) ∂Wi , ζij T uj (cid:105)T (cid:105)(cid:20) xj (cid:21)(cid:19) yj ∂Wi ∂yj = (20) (cid:105)T (cid:104) ∂Wi ∂xi Wi = T ui + = ζi where ζi (cid:44)(cid:104) ∂Wi ∂xi ∂Wi ∂yi . This time derivative is defined almost everywhere, except for a set of measure zero corresponding to the case when the gradient term ∇bijrgi is undefined (namely, this corresponds to rj = rgi). Not surprisingly, the evolution of the time derivative (20) ∂Wi ∂yj ∂xj along the trajectories of agent i depends not only on its own motion (through ui), but also on the motion of its neighbor agents j (cid:54)= i through their velocities uj. This time derivative provides sufficient conditions on establishing safety under the OGA policy. To see how, let us consider the following Lemma. Lemma 2: If the control inputs ui, uj, where j ∈ Ni, out of the OGA policy satisfy the following condition for all i ∈ {1, . . . , N}: ζi T ui + ζij T uj < 0, (21) (cid:17) (cid:88) (cid:16) j∈Ni then the OGA policy is asymptotically stable w.r.t. the current goal assignment k and furthermore collision-free. Proof: To verify the argument, consider the properties of the Lyapunov-like function (19) and denote the constrained set for each agent i as Ki = {r ∈ R2 N cij(·) ≥ 0}, where j ∈ Ni ∪ {0}. The set Ki is by construction compact (i.e., closed and bounded), with the level sets of Wi being closed curves contained in the set Ki. Then, the condition (21) implies that the system trajectories ri(t) under the OGA control input ui evolve along the negative direction of the gradient vector ∇Wi = of Wi, i.e., always remain in the interior of the constrained set Ki; this further reads that the inter-agent distances dij(t) never become smaller than ds. (cid:104) ∂Wi ∂Wi ∂yi (cid:105) ∂xi Problem 3: Seek a policy that becomes active when the OGA policy is about to result in colliding trajectories, and ensures that the inter-agent distances dij remain greater than the critical distance ds. We develop the CA policy described in the following section. B. Resolving Conflicts (cid:88) (cid:16) (cid:17) Let us consider the switching surface: Qi = ζi . ζij T uj j∈Ni T ui + (22) Then for Qi < 0 one has out of Lemma 2 that the OGA policy is both GAS and collision-free. For Qi > 0 one has Wi > 0, which implies that the position trajectories ri(t) evolve along the positive direction of the gradient vector ∇Wi. This condition may jeopardize safety and dictates the definition of an additional policy to ensure that the trajectories ri(t) remain in the constrained set Ki. At this point, one may be tempted to use a CA control law for agent i along the gradient-descent direction of Wi, to render the trajectories of the multi-robot system collision-free in the set Qi > 0. However, such control laws typically grow unbounded for system trajectories approaching local minimum points, i.e., as (cid:107)∇Wi(cid:107) → 0. This may lead to implementation issues when it comes to realistic robotic agents and therefore is undesirable from a practical point-of-view. We thus emulate the gradient descent performance along the level sets of (19) with control inputs that are always well- defined. We define a vector field Fi : R2 → R2 to be used as a feedback motion plan for each agent i, in the spirit developed in our earlier work [29]. The analytical expression is given as: (23) (1 − σij)Fi (cid:88) (cid:89) σijFi Fi = g + oj, j∈N(cid:105) j∈N(cid:105) for ds ≤ dij ≤ Rs; for Rs < dij < Rc; for dij ≥ Rc; (24) (cid:3)T vector ygi σij = where: field Fi a dij 0, 3 + b dij 2 + c dij + d,  1, the attractive to the goal position rgi = (cid:2)xgi (cid:112)(x − xgi)2 + (y − ygi)2 (cid:112)(x − xgi)2 + (y − ygi)2 (cid:112)(xi − xj)2 + (yi − yj)2 (cid:112)(xi − xj)2 + (yi − yj)2 g is defined as: xg = − Fi yg = − Fi the repulsive vector field Fi defined as: x − xgi y − ygi xoj = yoj = Fi Fi , (25b) oj around each agent j (cid:54)= i is xi − xj yi − yj (26a) , , (26b) , (25a) Rs the radius of the circle where the repulsive vector field is active, the coefficients a, b, c and d have been computed as: a = − 2 (Rs − Rc)3 , c = − 6 RsRc (Rs − Rc)3 , b = d = 3(Rs + Rc) (Rs − Rc)3 , 2(3Rc − Rs) Rc (Rs − Rc)3 , so that (24) is a C1 function. The vector field (23) for each agent i has integral curves that are: 1) convergent to the goal location rgi as long as dij > Rc for all j (cid:54)= i, i.e., when no agent j lies in the communi- cation region Ci of agent i, and 2) repulsive w.r.t. agents j which lie within distance dij < Rs w.r.t. agent i, i.e., within the communication region Ci of agent i. x, Fi Most importantly, the integral curves of (23) are always well- defined within the constrained set Ki of agent i, in contrast to the gradient vector field out of the Lyapunov-like barrier function (19) that is undefined at the points where the function (19) is not defined. The vector field (23) by construction vanishes only at the goal location rgi. Thus, the vector field (23) dictates a safe reference direction φi = atan2(Fi y, Fi x) everywhere on Ki \ {rgi}, where Fi y are the components of (23). Furthermore, one can show ( [29], Theorem 5) that the vector field (23) is an almost globally convergent feedback motion plan for agent i, i.e., all agent trajectories along the integral curves of (23) converge to the goal location rgi, except for those starting from a set of initial conditions of measure zero. For more details, the reader is referred to [29]. Now, in the ideal case of a static environment, i.e., when agents j (cid:54)= i are not moving, the control policy on the linear speed ui of agent i can naturally be independent of the linear speeds uj of other agents j (cid:54)= i, in the sense that these are identically zero. In the practical case of dynamic environments however, i.e., when agents j (cid:54)= i are moving, the control policy on the magnitude ui of the linear velocity vector ui of each agent i should take into account the motion of neighbor agents Under the control law (27), agent i adjusts its linear speed ui according to the velocity profile shown in Fig. 2, given analytically out of (27), so that the distance dij w.r.t. the worst case neighbor j remains greater than ds. The worst case neighbor is the agent j ∈ {Ni Ji < 0} towards whom the rate of change of relative distance dij given by (29), due to the motion of agent i, is maximum. More specifically: The term j ∈ {Ni Ji < 0} describes the set of neighbor agents j of agent i towards whom agent i is moving [28]. Agent i computes safe velocities uij given out of (28) w.r.t. each neighbor in this set, and picks the minimum min{uij} of the safe velocities uij so that the first term in (29) is as small as possible. Now, the value of the safe velocity uij (28) when . Plugging this value dij = ds is equal to εiuisj = εi into (29) reads: T uj T ηi rij rij (cid:12)(cid:12)dij =ds dij = (εi − 1)uj rij T ηj ≥ 0. dij To see why this condition is true, recall that εi−1 < 0, uj > 0, T ηj ≤ 0: this is because agent j is either following a and rij vector field Fj that points away from agent i, or happens to move away from agent i in the first place. This implies that the inter-agent distance dij can not become less than ds, i.e., that collisions are avoided. C. Ensuring both Stability and Safety We now define the switching policy that collision-free motion for each agent i as follows: realizes the Let Nc ≤ N agents become connected at some time instant td ≥ 0 while moving under some goal assignment k ∈ {1, . . . , Nc!}. Define Ri(td) = cos(θi(td)) cos(φi(td)) the decision-making surface for agent i, where θi(td) is the direction of the velocity vector ui at time td given out of the OGA controller (2), and φi(td) is the direction of the vector field vector Fi at time td, given out of (23). Then: • If Ri(td) > 0, the trajectories under the OGA policy are safe, and agent i keeps moving under (2). • If Ri(td) ≤ 0, the trajectories out of the OGA policy might violate safety, and agent i moves under (27). The switching between the OGA policy and the CA policy across the switching surface Ri = 0 gives rise to closed-loop dynamics with discontinuous right-hand side for each agent i: ri(t) = v(p) i , (30) i where p ∈ P = {1, 2}, and v(p) is the closed-loop dynamics of agent i under the OGA policy (2) and the CA policy (27), respectively. More specifically, for p = 1, agent i moves under the OGA policy, while for p = 2, agent i moves under the CA policy. We formalize the switching logic between the resulting closed-loop subsystems via a hysteresis technique, similar to the one used in Section IV-C; i.e., after making the decision be- tween either subsystem p = 1 or subsystem p = 2, the agents do not get involved in a new decision making unless they become disconnected first. The hysteresis logic gives rise to If Jj (cid:44) rij Fig. 2. T ηi < 0, i.e., if agent i moves towards agent j, then agent i adjusts its linear speed according to the velocity profile shown here, given analytically by (27). (cid:18) (cid:19) as well, so that the inter-agent distances dij(t) are guaranteed to remain greater than the lower bound ds, ∀t ≥ 0. We use the following velocity coordination protocol to govern the speed ui for each agent i: Lemma 3: Let Nc ≤ N agents become connected at some time instant td ≥ 0 while moving under some goal assignment k ∈ {1, . . . , Nc!}. The motion of each agent i in the connected group is collision-free under the control law: ui = where: ηi = (cid:2)cos φi min j∈NiJi<0 {uij} ηi, (cid:3)T is the unit vector along the (27) sin φi direction of the vector field (23), rij = ri − rj, Ji = rij uisj = rij rij , 0 < εi < 1, and T uj T ηi T ui, uij = ui dij − ds Rc − ds + εi Rc − dij Rc − ds . (28) rij rij T uj T ηi (cid:124) (cid:123)(cid:122) (cid:125) uisj Remark 5: In effect, the control law (27) forces agent i to adjust its speed w.r.t. the worst-case neighbor j ∈ Ni, i.e., w.r.t. the neighbor j towards whom the rate of change of inter- agent distance dij is maximized via the motion of agent i. Proof: Our goal is to show that the constrained set Ki encoded via the outermost level set of the Lyapunov-like barrier (19) is rendered a positively invariant set under the control law (27). We employ the fundamental characterization of positively invariant sets [27] through the necessary and suf- ficient conditions of Nagumo's theorem on the boundary of the dt cij(ri, rj) ≥ constrained set Ki of each agent i, namely: d 0, ∀rj ∈ ∂Ki, ∀(i, j). Consider that at some time td ≥ 0, the distance dij(td) between a pair of connected agents (i, j) is dij(td) ≤ Rc. Consider the time derivative of the inter-agent distance, which after some calculations reads: T (ui − uj) (1) = dij rij d dt dij . (29) a switching sequence of times T = {τ1, τ2, τ3, . . . , τn, . . . ,}, where the time interval τn − τn−1 is finite ∀n ∈ N, and to the switching sequence: Σ(cid:63) = {ri(τ0); (1, τ0), (2, τ1), . . . , (1, τ2q), (2, τ2q+1), . . . ,}, where τ0 > td, q ∈ N, τ2q are the time instants when the subsystem p = 1 (OGA policy) is "switched on", τ2q+1 are the time instants when the subsystem p = 2 (CA policy) is "switched on". Problem 4: Establish that the proposed switching strategy Σ(cid:63) between the OGA policy and the CA policy renders the multi-robot trajectories asymptotically stable w.r.t. to the as- signed goals, and also collision-free. For the stability analysis of this switching strategy, it is inefficient to employ Branicky's decreasing condition on the "switched-on" time intervals of each individual subsystem; the main reason is that in principle the CA strategy might result in trajectories that either are not monotonically decreasing, or are even increasing with a bounded growth for a finite amount of switches. Therefore we resort to results on the stability of switched systems that remove this condition in order to draw conclusions on asymptotic stability. Theorem 7: The trajectories of the switched multi-robot system under the switching logic Σ(cid:63) are (i) collision-free, and (ii) (almost globally)2 asymptotically stable w.r.t. the assigned goals. Proof: The first argument is proved in Lemmas 2 and 3. The second argument can be verified by directly applying Theorem 2. To this end, let us first note that Lemma 1 implies that the function Vi for each agent i serves as generalized Lyapunov-like function for the subsystem p = 1. Furthermore, is a valid generalized Lyapunov-like function almost Wi everywhere for each agent i for the subsystem p = 2. To see why, consider that the vector field (23) is by construction transverse to the boundary of the constrained set Ki, and convergent to the goal location rgi, except for a set of initial conditions of measure zero [29]. Hence, during the time intervals when the subsystem p = 2 is active, the growth, if any, of the function Wi along the system trajectory is bounded. Furthermore, the stability condition (6) for each subsystem p is satisfied out of the boundedness of the solutions ri(t) within the constrained set Ki, which is by construction closed and bounded. Therefore, the switched system is stable. To establish (almost global) asymptotic stability, it should hold that for at least one of the individual subsystems p, the value of the corresponding generalized Lyapunov-like function decreases along the sequence of switching times. Let us assume that this is not the case; then this would imply that , p ∈ {1, 2}, oppose the closed-loop switched vector fields v(p) each other and cancel out on the switching surface Ri, forcing the system trajectories to get stuck there. In other words, this would imply that the switching policy Σ(cid:63) exhibits Zeno tra- jectories. A Zeno point ¯r ∈ Sc(td) would be an accumulation point of the set S = {r ∈ Sc(td) : v(p)(cid:0)t− (cid:1) = v(p)(cid:0)t+ (cid:1)}, i d d 2The almost global characterization is added for completeness due to the set-measure initial conditions from which at least one of the agents is forced to their respective goal under the CA policy. and satisfies: ∇Sc(¯r)v(p)(cid:0)t− (cid:1) = ∇Sc(¯r)v(p)(cid:0)t+ (cid:1) = 0, d d where td is a decision making and switching time instant. For each agent i, the first condition reads: −λi(¯ri − rgi) = uij(¯ri, ¯rj) ηi(¯ri, ¯rj), j ∈ Ni, (31) where the left-hand side is given out of (2) and the right-hand side is given out of (27), with ηi(¯ri, ¯rj) the unit vector along the direction φi of the vector field (23); recall that φi depends by construction on the states of the neighbor agents j ∈ Ni, hence the dependence on ¯rj in the right-hand side of (31). Geometrically, the condition (31) implies that the candidate set S of Zeno points lies on the line that connects the position of the agent with its goal location and is concurrently collinear with the vector field (23). The second set of conditions reads: ∇(cid:0)uij(¯ri, ¯rj)ηi(¯ri, ¯rj)(cid:1) = 0. (32a) (32b) The former one reduces to λi = 0, ∀i, which is a contradiction. Hence the candidate set of Zeno points is empty, i.e., the switching strategy between the OGA policy and the CA policy does not exhibit infinite switching between the OGA and the CA policies around an accumulation point. Therefore, the switched multi-robot system under the switching strategy Σ(cid:63) does not exhibit Zeno trajectories, and is almost globally asymptotically stable. −λi∇ (¯ri − rgi) = 0, D. Robustness against Communication Failures We note that establishing robustness against uncertainty (e.g., against sensor noise, communication dropouts and/or delays) is beyond the scope of the current paper. As a special case, we only consider how the case of failed communication can be handled within the centralized version of the proposed approach. Note that finding the set of initial conditions that result in collision-free trajectories for a given distribution of goals and for every possible decentralized assignment is essen- tially a backwards reachability problem, which is intractable for even low-dimensional systems, i.e., for teams of no more than 2-3 robots. However, the centralized case as treated in the earlier work of the authors [18] offers a (conservative, yet guaranteed) robustness measure for the case of failed communication in a decentralized setting; more specifically, we have the following corollary: Corollary 1: Let the multi-robot system under switched agent dynamics given by (30), and the switching sequence Σ(cid:63). If for all ri(0), rj(0), rGi, rGj , i, j ∈ {1, . . . , N}, it holds that (cid:107)ri(0) − rj(0)(cid:107) ≥ 2R 2, where R is the radius of the robots, then under failed communication the multi-robot trajectories are collision-free. 2, and (cid:107)rGi − rGj(cid:107) ≥ 2R √ √ Proof: The corollary states that if the initial robot po- 2 apart, then the result- sitions and goal locations are 2R ing multi-robot trajectories are collision-free, even when the agents' communication fail and no goal swap or collision avoidance controllers are implemented. Note that under failed communication, the affected agents will not be involved in any decision on re-assigning their goals (i.e., will keep their current goals), and will not update their control laws. Hence, whether the trajectories will be collision-free depends on the initial √ conditions (positions) of the robots and the spatial distribution of the goal locations. The analysis of the centralized version of the OGA policy in [18] proves that under the stated assump- tions, the OGA policy results in collision-free trajectories. This completes the proof. VI. SIMULATION RESULTS A. Simulations for N = 15 and N = 40 agents Simulation results are provided to evaluate the performance of the switched multi-robot system under the proposed switch- ing logic and control policies. In the interest of space, and given that the representation of large groups of robots results in rather congested figures, here we consider only two scenarios involving N = 15 and N = 40 agents, respectively. The initial positions and the assigned goal locations are selected in very close proximity so that the OGA policy on its own is not sufficient to ensure collision-free trajectories, see Fig. 3 and 4, respectively. The agents start from the initial con- ditions shown in Fig. 3(a), 4(a) towards goal locations marked with "x"; as discussed earlier, the initial goal assignment is not necessarily the optimal one. The communication ranges are denoted by the rings around each agent, and set equal to Rc = 4R, where R is the radius of the agents. The straight line paths from the starting locations to the initially assigned goals are depicted as the dotted lines, whereas the actual paths followed by the switching strategy are shown in Fig. 3(b) and 4(b), respectively. The final locations of the robots are shown in Fig. 3(c) and 4(c), respectively. As indicated by the colors of the initially assigned goals, the followed paths, and the final goal locations of the robots, the robots performed goal swaps during the duration of the simulation under the OGA policy, while they also moved under the CA policy during the path segments that are not straight. Figures 5 and 6, respectively, show the minimum distance between any two pairs of agents at each time instance of the simulations. As this distance is always greater than the safety distance ds, there is never a collision between any robots. Additional videos with teams of 6, 30, 80, 100, and 150 agents are available through the following links: https://www.dropbox.com/s/a5hykza2wla7x3n/6.mp4?dl=0 https://www.dropbox.com/s/1gazjlt7nell0mn/30.avi?dl=0 https://www.dropbox.com/s/48bnwsnlqibkgaj/80.avi?dl=0 https://www.dropbox.com/s/e6wljow9ft2ziap/100.avi?dl=0 https://www.dropbox.com/s/uif0yahjlhv0bsf/150.avi?dl=0 It is noteworthy that the simulation videos demonstrate the expected performance of the overall algorithm in terms of the convergence rate; more specifically, the slower convergence of the agents to their goal locations towards the end of the simulation, compared to their faster motion in the beginning of the simulation, is consistent with the exponential convergence rate of the OGA policy. 3 3Note however that the videos run at different frames-per-second or have deliberate pauses to explain the algorithm, and as thus the time lengths of the videos shall not be used for a direct comparison of the convergence time among the cases of different N. (a) Initial positions of the agents and the straight paths to the goals determined by the random initial goal assignment. (b) The actual paths to the goals under the proposed control strategy. (c) The final locations of the agents. Fig. 3. A simulation with N = 15 robots. The paths followed are straight when in the OGA segments, but are potentially curved when using the CA control law. -15-10-505101520-20-15-10-50510-10-5051015-15-10-505-15-10-5051015-15-10-505(a) Initial positions of the agents and the straight paths to the goals determined by the random initial goal assignment. Fig. 5. The minimum inter-agent distance between any two robots for a simulated trial with N = 15. Note that as this is always greater than the allowable safety distance, there were never any collisions between robots. (b) The actual paths to the goals under the proposed control strategy. (c) The final locations of the agents. Fig. 4. A simulation with N = 40 robots. The paths are straight during the OGA segments, but are potentially curved when using the CA control law. Fig. 6. The minimum inter-agent distance between any two robots for a simulated trial with N = 40. Note that as this is always greater than the allowable safety distance, there were never any collisions between robots. B. Comparison with earlier work in [24] √ As discussed earlier, the current approach dominates our earlier work [24] in the following aspects: (i) The initial positions of the robots and the goal locations can be close as ds = 2R, in contrast to [24] that requires the initial locations 2 apart for safety. (ii) The collision-free and goals to be 2R trajectories of the current approach are rendered via switching among two control policies, opposed to [24] that involves switching among three control policies. The approach here nevertheless does not consider the initially optimal assignment, and as thus, depending on the connected subcomponents at initial time and throughout the evolution of the system, might result in suboptimal paths in terms of total distance. For completeness, we include simulation results between the current approach and [24]; in the scenario illustrated in Fig. 7, the resulting paths are suboptimal; the total computational time -20-100102030-20-15-10-505101520-20-15-10-5051015202530-20-15-10-5051015-20-100102030-25-20-15-10-505101520Time [sec]00.10.20.30.40.50.60.70.8Minimum Pairwise Inter-agent Distance dij1.51.551.61.651.71.751.81.851.91.952Minimun Pairwise Interagent DistanceSafety DistanceTime [sec]00.10.20.30.40.50.60.70.8Minimum Pairwise Inter-Agent Distance dij1.51.551.61.651.71.751.81.851.9Minimum Pairwise Interagent DistanceSafety Distanceit follows that the number of re-plan operations, and hence the number of messages that are exchanged among the connected agents on deciding whether they swap goals, depends greatly on the communication range, the number of robots, and their initial distribution, i.e., the quality (in terms of sub-optimality) of the initial assignment. Note that for sufficiently large radius Rc, which results in all agents being connected at the same time, the centralized case is recovered, meaning that the agents decide among N! assignments. The time for a re-plan operation also directly depends on the number of agents in the connected component. Planning times on simulated teams of Nc robots using a single computer take approximately 10−7Nc 3 seconds, which reads that a component of 100 connected robots takes about 0.1 seconds to re-plan. The required computational time is typically sufficient for the proposed multi-robot applications. VII. EXPERIMENTAL RESULTS The proposed switching logic and control policies were also evaluated in experimental trials with six ground robots (AION R1 Rovers) in the Distributed Aerospace Systems and Control Lab, University of Michigan. The position of the robots was acquired via a VICON motion capture system. The Robot Operating System (ROS) was used to exchange data among the agents, as well as to convert the high-level control commands to low-level motor voltage commands that drive the robots' wheels. The control loop was closed at approximately every 30 msec. Since the robots are differentially-driven, hence closer modeled via the unicycle nonholonomic kinematics rather than single integrators, the proposed controllers were implemented by steering the robots' orientations along the direction of the velocity vector computed from either the control law (2) or the control law (27), and applying the corresponding linear velocity along this direction. This corresponds in part to the discrepancy in the resulting paths between the simulation results in Fig. 7(a) and experimental results as described below. The experiments demonstrate that the proposed switching logic and corresponding policies are applicable in practice, and produce results that are consistent with the expected theoretical performance. The controller parameters were set equal to: R = 0.35 m (robot radius), Rc = 1.1 m (communication radius), Rs = 0.9 m, ds = 0.7 m (minimum allowable distance). The video of experiment is available at https://www.dropbox.com/s/ujrtzn1aolc0yul/6-exp.mp4?dl=0. The derived paths of the robots under the proposed approach are shown in Fig. 8(a). Fig. 8(b) depicts the distance of each robot w.r.t. its corresponding goal over time, and furthermore illustrates the time instances where the OGA becomes active and invokes goal swaps along with concurrent reduction in the cost (distance-to-the-goal). The minimum pairwise distance among the robots over time is shown in Fig. 8(c), and demonstrates that the trajectories are always collision-free. VIII. CONCLUDING REMARKS We presented a switched systems approach on the decen- tralized concurrent goal assignment and trajectory generation for multi-robot networks which guarantees safety and global (a) Initial positions (red), goal locations (green) and the paths resulting from the current approach. (b) Initial positions (red), goal locations (green) and the paths resulting from the approach in [24]. Fig. 7. Comparison of the paths resulting from the current approach versus the approach in [24]. for the entire simulation in the former case is was 2.625 sec, while in the latter case was 1.375 sec. This demonstrates that even with suboptimal paths, the algorithm is computationally efficient for the considered applications. C. Notes on communication requirements and planning time The implementation of the proposed algorithm in a realistic setting requires the capability of exchanging messages (e.g., via wireless communication links) among the subset of agents that are considered connected at some time instance, i.e., among robots that lie within Rc distance and their neighbor agents. Once a connected group is formed, the OGA policy runs in a synchronous manner. Recall that in our approach the initial assignment is random, i.e., the focus of this work is not on determining the globally optimal assignment of the robots to the goals, but instead we focus on re-assigning goals among connected robots. Thus, since the decision-making mechanism (i.e., the OGA policy) on re-determining the goal assignment is triggered whenever (a subset of) agents become connected, stability to interchangeable goal locations. The proposed switching logic relies on multiple Lyapunov-like functions that encode goal swap among locally connected agents based on the total traveled distance, avoidance of inter-agent collisions, and convergence to the assigned goal locations. As such, the proposed methodology renders feedback control policies with local coordination only, and therefore is suitable for applica- tions such as in first-response deployment of robotic networks under limited information sharing. The proposed algorithms were evaluated via simulation results and experiments with six robots. Our future work will focus on agents with more complicated dynamics, as well as on the robustness of our algorithms under communication failures and uncertainty. REFERENCES [1] Z. Yan, N. Jouandeau, and A. A. Cherif, "A survey and analysis of multi- robot coordination," Int. Journal of Advanced Robotic Systems, vol. 10, pp. 1 -- 18, 2013. [2] S. Kloder and S. Hutchinson, "Path planning for permutation-invariant multirobot formations," IEEE Transactions on Robotics, vol. 22, no. 4, pp. 650 -- 665, Aug. 2006. [3] S. L. Smith and F. Bullo, "Monotonic target assignment for robotic networks," IEEE Trans. on Automatic Control, vol. 54, no. 9, pp. 2042 -- 2057, Sep. 2009. [4] L. Liu and D. Shell, "Multi-level partitioning and distribution of the task allocation," in assignment problem for large-scale multi-robot Proceedings of Robotics: Science and Systems, Los Angeles, CA, USA, June 2011. [5] N. Jouandeau and Z. Yan, "Improved trade-based multi-robot coor- dination," in Proc. of the 6th IEEE Joint International Information Technology and Artificial Intelligence Conference (ITAIC), vol. 1, Aug 2011, pp. 500 -- 503. [6] 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, Apr. 2016. [7] B. Yamauchi, "Decentralized coordination for multirobot exploration," Robotics and Autonomous Systems, vol. 29, pp. 111 -- 118, 1999. [8] B. Werger and M. Mataric, "Broadcast of local eligibility: behavior- based control for strongly cooperative robot teams," in Proc. of the 4th Int. Conference on Autonomous Agents, New York, 2000, pp. 21 -- 22. [9] J. Faigl, M. Kulich, and L. Preucil, "Goal assignment using distance cost in multi-robot exploration," in Proc. of the 2012 IEEE/RSJ Int. Conf. on Intelligent Robots and Systems, Vilamoura, Algarve, Portugal, Oct. 2012, pp. 3741 -- 3746. [10] A. Bautin, O. Simonin, and F. Charpillet, "MinPos: A novel frontier allocation algorithm for multi-robot exploration," ser. Lecture Notes in Computer Science; Intelligent Robotics and Applications, C.-Y. Su, S. Rakheja, and H. Liu, Eds. Springer, Heidelberg, 2012, vol. 7507, pp. 496 -- 508. [11] M. G. Lagoudakis, M. Berhault, S. Koenig, P. Keskinocak, and A. J. Kleywegt, "Simple auctions with performance guarantees for multi-robot task allocation," in Proc. of the 2002 IEEE/RSJ Int. Conf. on Intelligent Robots and Systems, Sendai, Japan, Sep. 2004, pp. 698 -- 705. [12] H.-L. Choi, L. Brunet, and J. P. How, "Consensus-based decentralized auctions for robust task allocation," IEEE Trans. on Robotics, vol. 25, no. 4, pp. 912 -- 926, Aug. 2009. [13] J. Hawley and Z. Butler, "Hierarchical distributed task allocation for multi-robot exploration," in Distributed Autonomous Robotic Systems, ser. Springer Tracts in Advanced Robotics, 2013, vol. 83, pp. 445 -- 458. [14] M. M. Zavlanos and G. J. Pappas, "Dynamic assignment in distributed motion planning with local coordination," IEEE Trans. on Robotics, vol. 24, no. 1, pp. 232 -- 242, Feb. 2008. [16] J. Yu, S. Chung, and P. Voulgaris, "Target assignment [15] G. Arslan, J. R. Marden, and J. S. Shamma, "Autonomous vehicle-target assignment: A game-theoretical formulation," ASME Journal of Dynamic Systems, Measurement, and Control, vol. 129, pp. 584 -- 596, Sep. 2007. in robotic networks: Distance optimality guarantees and hierarchical strategies," IEEE Trans. on Automatic Control, vol. 60, no. 2, pp. 327 -- 341, 2015. [17] X. Bai, W. Yan, M. Cao, and J. Huang, "Target assignment for robots constrained by limited communication range," 2017. [Online]. Available: https://arxiv.org/pdf/1702.04700.pdf (a) The resulting paths to the goals under the proposed control strategy. (b) The distances of the agents w.r.t. their goals over time, and the invoked goal swaps when the OGA becomes active. (c) The minimum clearance between any two robots over time. Fig. 8. Experimental results with N = 6 robots. [18] M. Turpin, N. Michael, and V. Kumar, "Trajectory planning and as- signment in multirobot systems," in Proc. of Int. Workshop on the Algorithmic Foundations of Robotics (WAFR), Boston, MA, Jun. 2012. [19] -- -- , "CAPT: Concurrent assignment and planning of trajectories for multiple robots," The Int. Journal of Robotics Research, vol. 33, no. 1, pp. 98 -- 112, 2014. [20] J. Munkres, "Algorithms for the assignment and transportation prob- lems," Journal of the Society for Industrial & Applied Mathematics, vol. 5, no. 1, pp. 32 -- 38, 1957. [21] S. Giordani, M. Lujak, and F. Martinelli, "A distributed algorithm for the multi-robot task allocation problem," in Trends in Applied Intelligent Systems. Springer, 2010, pp. 721 -- 730. [22] D. P. Bertsekas, Network Optimization: Continuous and Discrete Mod- els. Athena Scientific, 1998. [23] R. Fitch, A. Alempijevic, and T. Peynot, "Global reconfiguration of a team of networked mobile robots among obstacles," in Advances in Intelligent Systems and Computing, E. Menegatti, N. Michael, K. Berns, and H. Yamaguchi, Eds. Springer, 2016, pp. 639 -- 656. [24] D. Panagou, M. Turpin, and V. Kumar, "Decentralized goal assignment and trajectory generation in multi-robot networks: A multiple Lyapunov functions approach," in Proc. of the 2014 IEEE Int. Conf. on Robotics and Automation, Hong Kong, China, Jun. 2014, pp. 6757 -- 6762. [25] D. Liberzon, Switching in Systems and Control. Birkhauser, 2003. [26] M. S. Branicky, "Multiple Lyapunov functions and other analysis tools for switched and hybrid systems," IEEE Trans. on Automatic Control, vol. 43, no. 4, pp. 475 -- 482, Apr. 1998. [27] F. Blanchini and S. Miani, Set-Theoretic Methods in Control. Birkhauser, 2008. [28] D. Panagou, D. M. Stipanovi´c, and P. G. Voulgaris, "Distributed coordi- nation and control of multi-robot networks using Lyapunov-like barrier functions," IEEE Trans. on Automatic Control, vol. 61, no. 3, pp. 617 -- 632, 2016. [29] D. Panagou, "A distributed feedback motion planning protocol for multiple unicycle agents of different classes," IEEE Trans. on Automatic Control, vol. 62, no. 3, pp. 1178 -- 1193, 2017. [30] P. Peleties and R. DeCarlo, "Asymptotic stability of m-switched systems the American Control using Lyapunov-like functions," in Proc. of Conference, Boston, MA, Jun. 1991, pp. 1679 -- 1684. [31] J. Zhao and D. a general multiple Lyapunov function method." [Online]. Available: http://cecs.anu.edu.au/seminars/media/ switched systems: "Stability of J. Hill, [32] -- -- , "On stability, L2-gain and H∞ control for switched systems," Automatica, vol. 44, pp. 1220 -- 1232, 2008. [33] A. D. Ames and S. Sastry, "Characterization of Zeno behavior in hybrid systems using homological methods," in Proc. of the 2005 American Control Conference, Portland, OR, USA, Jun. 2005, pp. 1160 -- 1165. [34] F. Ceragioli, "Finite valued feedback laws and piecewise classical solutions," Nonlinear Analysis, vol. 65, no. 5, pp. 984 -- 998, 2006. [35] Y. Nesterov and A. Nemirovskii, Interior Point Polynomial Algorithms in Convex Programming. SIAM Studies in Applied Mathematics, 2001. [36] A. G. Wills and W. P. Heath, "A recentred barrier for constrained receding horizon control," in Proc. of the American Control Conference, Anchorage, AK, USA, May 2002, pp. 4177 -- 4182. [37] D. M. Stipanovi´c, C. J. Tomlin, and G. Leitmann, "Monotone ap- proximations of minimum and maximum functions and multi-objective problems," Applied Mathematics and Optimization, vol. 66, pp. 455 -- 473, 2012. Dimitra Panagou received the Diploma and PhD degrees in Mechanical Engineering from the Na- tional Technical University of Athens, Greece, in 2006 and 2012, respectively. Since September 2014 she has been an Assistant Professor with the De- partment of Aerospace Engineering, University of Michigan. Prior to joining the University of Michi- gan, Dr. Panagou was a postdoctoral research as- sociate with the Coordinated Science Laboratory, University of Illinois, Urbana-Champaign (2012- 2014), a visiting research scholar with the GRASP Lab, University of Pennsylvania (June 2013, fall 2010) and a visiting research scholar with the University of Delaware, Mechanical Engineering Department (spring 2009). Her research interests include the fields of planning, coor- dination, and distributed control and estimation for complex systems, with applications in unmanned aerial systems, robotic networks and autonomous multi-vehicle systems (ground, marine, aerial, space). She was a recipient of the NASA 2016 Early Career Faculty Award and of the Air Force Office of Science Research 2017 Young Investigator Award. Matthew Turpin received his BS degree in Mechan- ical Engineering from Northwestern University in 2009, and his PhD degree in Mechanical Engineer- ing from University of Pennsylvania in 2014. During his doctoral studies at the University of Pennsylvania he focused on formation control as well as on the coupling of task assignment and motion planning in multi-robot systems. He is currently a research sci- entist at Qualcomm Technologies Inc. where he has worked on research and development of software for multi-rotor MAVs, and currently works on planning and control for autonomous cars. Vijay Kumar is the Nemirovsky Family Dean of Penn Engineering with appointments in the De- partments of Mechanical Engineering and Applied Mechanics, Computer and Information Science, and Electrical and Systems Engineering at the Univer- sity of Pennsylvania. He served as the assistant director of robotics and cyber physical systems at the White House Office of Science and Technology Policy (2012-2014) and is the founder of Exyn Technologies, a company that develops solutions for autonomous flight. Dr. Kumar is a Fellow of the American Society of Mechanical Engineers (2003), a Fellow of the Institute of Electrical and Electronic Engineers (2005) and a member of the National Academy of Engineering (2013). Among his recent awards are the 2012 World Technology Network award, a 2013 Popular Science Breakthrough Award, a 2014 Engelberger Robotics Award, the 2017 IEEE Robotics and Automation Society George Saridis Leadership Award in Robotics and Automation and the 2017 ASME Robert E. Abbott Award. He has won best paper awards at DARS 2002, ICRA 2004, ICRA 2011, RSS 2011, RSS 2013, and BICT 2015 and has advised doctoral students who have won Best Student Paper Awards at ICRA 2008, RSS 2009, and DARS 2010.
cs/0604110	1	0604	2006-04-27T23:33:25	Modeling and Mathematical Analysis of Swarms of Microscopic Robots	[ "cs.MA", "cs.RO" ]	The biologically-inspired swarm paradigm is being used to design self-organizing systems of locally interacting artificial agents. A major difficulty in designing swarms with desired characteristics is understanding the causal relation between individual agent and collective behaviors. Mathematical analysis of swarm dynamics can address this difficulty to gain insight into system design. This paper proposes a framework for mathematical modeling of swarms of microscopic robots that may one day be useful in medical applications. While such devices do not yet exist, the modeling approach can be helpful in identifying various design trade-offs for the robots and be a useful guide for their eventual fabrication. Specifically, we examine microscopic robots that reside in a fluid, for example, a bloodstream, and are able to detect and respond to different chemicals. We present the general mathematical model of a scenario in which robots locate a chemical source. We solve the scenario in one-dimension and show how results can be used to evaluate certain design decisions.	cs.MA	cs	MODELING AND MATHEMATICAL ANALYSIS OF SWARMS OF MICROSCOPIC ROBOTS Aram Galstyan1, Tad Hogg2, Kristina Lerman1 1. USC Information Sciences Institute 4676 Admiralty Way Marina del Rey, CA 90292, USA 2. HP Labs 1501 Page Mill Road, MS 1U-19 Palo Alto, CA 94304, USA ABSTRACT The biologically-inspired swarm paradigm is being used to design self-organizing systems of locally interacting artifi- cial agents. A major difficulty in designing swarms with desired characteristics is understanding the causal relation between individual agent and collective behaviors. Math- ematical analysis of swarm dynamics can address this dif- ficulty to gain insight into system design. This paper pro- poses a framework for mathematical modeling of swarms of microscopic robots that may one day be useful in med- ical applications. While such devices do not yet exist, the modeling approach can be helpful in identifying various de- sign trade-offs for the robots and be a useful guide for their eventual fabrication. Specifically, we examine microscopic robots that reside in a fluid, for example, a bloodstream, and are able to detect and respond to different chemicals. We present the general mathematical model of a scenario in which robots locate a chemical source. We solve the sce- nario in one-dimension and show how results can be used to evaluate certain design decisions. 1. INTRODUCTION Swarm Intelligence is a relatively new concept for design- ing multi-agent systems [4]. Swarms are decentralized sys- tems composed of many simple agents with no central con- troller. Instead, swarms are self-organizing: constructive collective behavior emerges from local interactions among the agents and between agents and their environment. Self- organization is ubiquitous in nature: Aggregation into multi- cell organisms in Dictyostelium discoideum colony, trail for- mation in ants, hive building by bees, and mound construc- tion by termites, are a few examples. The success of these organisms has inspired the design of distributed problem- solving systems. A major difficulty in designing self-organizing swarms with desired characteristics is understanding the causal rela- tion between individual agent characteristics and collective behavior. One approach is experiments with real agents, e.g., robots, to evaluate behavior under real conditions. How- ever, fabricating enough agents to exhibit swarm behavior is often too expensive or not feasible with current technology. Simulations, such as sensor-based simulations of robots, are usually much faster and less costly than experiments. But these require good models of the robots and their environ- ments, which are often not yet available. Moreover, sim- ulations still require many repetitions to systematically ex- amine consequences of various choices for individual agent behaviors. Mathematical modeling and analysis offer an alternative to the time-consuming and costly experiments and simula- tions. Mathematical analysis can be used to study swarm dynamics, predict long-term behavior of even very large systems, and gain insight into system design: e.g., how indi- vidual agent characteristics affect the swarm behavior. Such analysis can suggest parameters that optimize group perfor- mance, prevent instabilities, etc. These particular choices could then be examined in more detail through simulations or experiments. This mathematical approach is particularly useful for preliminary evaluation of design trade-offs for robots that cannot yet be built and will operate in poorly characterized environments. As one such application, robots with sizes comparable to bacteria ("nanorobots") could provide many novel capabilities through their ability to sense and act in microscopic environments. Of particular interest are medi- cal applications, where robots and nanoscale-structured ma- terials inside the body could significantly improve disease diagnosis and treatment [9, 19, 20, 15]. Typically, each such robot will have dimensions of a few microns so could read- ily travel through even the tiniest blood vessels, will be able to communicate with other robots (e.g., using ultrasound) over distances of about 100 microns, and will consume in the range of 1 − 1000 picowatts. An example task for such robots operating in the blood- stream is to aggregate and form structures at locations with specific patterns of chemicals. Initially, as robots move pas- sively through the fluid, monitoring for the chemical pat- tern, they need to avoid disturbing the flow. Thus they will be at low density, i.e., with typical separation between neigh- boring robots many times larger than their size. On the other hand, to affect the environment, many robots will need to act together (e.g., to block flow in blood vessels feeding populations of cancer cells or to aid in repairing damaged tissues by forming suitable scaffolding). Thus the swarm must change to a higher density of robots, in spite of lim- ited communication ranges. In this denser configuration, communication among neighbors is easier, and simple local controls allow forming a variety of structures either static or dynamically responding to changing forces in the environ- ment, as proposed in the context of large self-reconfigurable robots [3, 24]. To quantify this example, a swarm could con- sist of around 1012 machines, vastly more than in swarms of larger devices. Even such a large number of machines would have a total mass of only about one gram, and a oc- cupy less than 0.1% of the typical blood volume of an adult, compared to the 20 − 40% occupied by cells. While fabricating such machines is beyond the ability of current technology, the rapid progress in engineering nano- scale devices should eventually enable their production. For example, ongoing development [12, 10, 18] of molecular- scale electronics, sensors and motors provides components for such robots, though these cannot yet be assembled into large numbers of complete systems. Programmable bacte- ria [25] give large numbers of functioning microscopic sys- tems, and separate DNA computers can respond to Boolean combinations of chemicals in their environment [2]. How- ever relying on biological organisms restricts the robots' material properties and, by using protein synthesis to per- form logic operations, limits the computational complexity and speed of control programs. Beyond the challenge of fabricating the robots is de- signing and analyzing controls suitable for their physical environments, tasks and capabilities, whose features differ considerably from those of larger robots. Fortunately, we can evaluate control methods prior to building the robots with theoretical studies [7, 9, 23] characterizing robot ca- pabilities and their task environments. Moreover, suitable control programs can compensate for some hardware lim- itations (e.g., limited available power or sensor accuracy). Thus studies of control approaches can identify useful tasks for robots with limited hardware capabilities, as may arise before the fabrication technology is fully mature. This paper extends the stochastic processes framework developed for modeling teams of larger robots and that has been applied successfully to a number of robotic applica- tions [17]. Our work complements other theoretical studies of microscopic robots [5, 23, 11] by providing a methodol- ogy for studying collective swarm behavior. It can also be applied to larger robots that implement chemotaxis as the basis for movement in an external field [6]. While compari- son with microscopic experiments must await the technical feasibility of fabricating microscopic robots, our mathemat- ical model can be used to evaluate various design trade-offs for the robot capabilities that can achieve the same swarm performance and can be a useful guide to eventual fabri- cation by allowing designers to focus on combinations of hardware capabilities easiest to fabricate. Below we present a general model of the dynamics of a swarm of microscopic robots that are interacting with an external field. In Section 3 we motivate the approach by describing a medically relevant scenario that considers a swarm of microscopic robots moving in a fluid to localize a chemical source. In Section 4 we solve a one-dimensional model and analyze different design choices. We conclude the paper with a discussion of the approach and results. 2. STOCHASTIC THEORY OF SPATIALLY-DEPENDENT PROCESSES Previously, we developed a framework based on stochas- tic Master Equation for mathematical analysis of homoge- neous robotic swarms [13, 16]. Specifically, consider an agent whose controller is an automaton consisting of L lo- cal states. We then assume the overall system is sufficiently well-described by an occupancy vector n = (n1, n2, ...nL), where nk is the fraction of agents in state k. That is, we need only consider the number of agents in each state rather than tracking their individual identities. In designing sys- tems where typical or average behavior is of primary inter- est, it is useful to characterize the system by a probability distribution P (n, t), which is the probability the system is in configuration n at time t, and evolves according to the Master Equation [14]. This distribution characterizes the behaviors of an ensemble or multiple independent instances of the swarm. When most such instances have nearly the same behavior,1 fluctuations among different instances can be neglected, and the average occupancies ¯nk provide a suf- ficient characterization of the system. The average fraction of agents in state k then evolves according to d¯nk dt = Xj wjk ¯nj − ¯nkXj wkj (1) 1Consider, for example, a swarm consisting of a large number of agents, each independently executing the same controller. Then, by the law of large numbers, behaviors of different instances of this swarm would be nearly identical. with wjk the transition rates between the states j and k. Note than in general wjk might depend on the fraction of agents in particular states, i.e., average occupancy vector ¯n. Eq. 1 governs the evolution of the collective behavior of the system. While this approach works well for many spatially uni- form systems [17], it is too coarse-grained for systems with a spatial correlation in agents' interactions. Thus, it is not sufficient to describe, for example, an ant-like swarm where agents interact through evolving chemical fields or micro- scopic robots monitoring chemicals released into a fluid. These situations require a generalization of the Master Equa- tion, in which each robot not only has a discrete controller state k but also a continuous coordinate x (i.e., its spatial location). As with the original formulation, we suppose the number of agents in each state is sufficient to determine the collective behavior of interest. Because x is a continuous variable, these counts become densities leading us to intro- duce ¯nk(x, t) as the average robot fraction density in state k at location x and time t. Thus a small volume ∆x around location x contains, on average, the fraction ¯nk(x, t)∆x of the robots in the system. Let us consider a system where agents interact with the environment through a certain external chemical field. Let us also assume that agents are able to interact through stig- mergy by releasing a special chemical into the environment that we call communicative signal. We denote ρ(x, t) and c(x, t) concentration of the chemical and communicative signal, respectively, at point x at time t. Then we can write down the generalized rate equation as follows: tion rates. Communication among nearby robots, e.g., with ultrasound, allows the robots to reduce noise in estimating chemical gradients and hence perform better than individ- ual robots or bacteria [1], but at a cost of additional power use for the communication. Robot motion, either moving passively with the fluid or using powered locomotion, e.g., to follow chemical concentration gradients, also contributes to the transitions. While Equation 2 is a general description of the overall system behavior, it is too complex in its present form to be useful. Fortunately, it can be simplified considerably into a more intuitive form by noting that in many physically real- istic situations agents' motion can be decoupled from state transitions, so that the transition rate can be represented as wjk = δjkWk(x, x′; ρ(x), ρ(x′), c(x), c(x′)) + δ(x − x′)wjk(ρ(x), c(x)) , (3) where δjk is Kroenecker's symbol2 and δ(x) is its continu- ous analogue δ -- function. In other words, during a transition between two discrete states we neglect the change in robot's position. In Eq. 3 Wk is an appropriately chosen kernel that describes agents' motion (as index k indicate, it can be dif- ferent for each state), while the second term describes tran- sition between discrete states. Equation 3 allows us to separate transition function into terms with purely spatial transitions and terms with purely state transitions. Indeed, using Eq. 3 we can decouple the agents' kinematics from the state transitions between dis- crete state and rewrite Eq. 2 as follows: ∂ ¯nk(x, t) ∂t = Z dx′ Xj − ¯nk(x, t)Z dx′ Xj wjk(x, x′; ρ, c)¯nj(x′, t) (2) wkj (x, x′; ρ, c) . ∂ ¯nk(x, t) ∂t wjk(ρ, c)¯nj (x, t) (4) = Lk ¯nk(x, t) + Xj − ¯nk(x, t)Xj wkj (ρ, c) Now the transition rates wjk depend not only the state in- dices j and k and occupation vector but also on the spatial coordinates and concentration of the chemical at the corre- sponding points. Note also that we have included the depen- dence of the transition rates on x and x′ explicitly to account for agents' kinematics even in the absence of chemical and communicative concentrations (e.g., to describe freely dif- fusing agent). The transition rates wjk(x, x′; ρ(x), c(x)) summarize the behaviors of the individual robots. For example, the robot's internal state could change when it detects a chem- ical concentration above a predetermined threshold. In ap- plications of microscopic robots, such detections will often involve a relatively small number of molecules (e.g., on the order of tens) so stochastic fluctuations in the number de- tected will be a significant source of sensor noise and can be included as a range of new states with nonzero transi- Here Lk is an operator (specified below) that describes the motion of agents in state k. The second and third terms in Eq. 4 describe agents state transitions. Note that now wjk(ρ, c) depends on spatial coordinates indirectly, through concentration ρ(x, t) and c(x, t). When the concentrations ρ and c are constants, Eq. 1 is recovered by integrating Eq. 4 over the spatial coordinates x, assuming that Lk preserves the number of agents in state k (e.g., no absorbing bound- aries) so that integral over the first term in Eq. 4 vanishes. To specify the operators Lk, we note that for the par- ticular environment we are interested in, (i.e., microscopic robots operating in a fluid) robots' motion can be described by a diffusion equation [14]. In this paper, we study chemo- tactic robots that respond to a chemical and signalling fields 2Kroenecker's symbol is defined as follows: δij = 1 if i = j and δij = 0, i 6= j. by propelling themselves in the direction of increasing con- centration. This capability is modeled after bacterial chemo- taxis which allows these single cell organisms to efficiently move towards food sources and away from noxious sources. Although in some cases the exact derivation from the mi- croscopic transition rates is feasible, if very involved (see, for example, [21, 8] for treatment of bacterial chemotaxis which can be treated as a biased random walk), chemotaxis in a chemical concentration field ρ(x, t) is usually intro- duced into the rate equations phenomenologically by pos- tulating a chemotactic velocity as VD = ηρ∇ρ(x, t), where ηρ is the so called chemotactic sensitivity (which may itself depend on ρ). One can then write for operators Lk Lk = Dk∇2 − v·∇−∇· [Vρ D(c,∇c)] (5) Here, Dk is the diffusion coefficient of agents in state k as- sumed to be a constant, v is the flow velocity, and Vρ D and D are the chemotaxis drift velocities of robots due to con- Vc centration gradients of the chemical and the communicative signal, respectively. D(ρ,∇ρ) + Vc To proceed further, we should also define how the chem- ical and concentration fields evolves in time. As an example relevant for microscopic robots, we consider the evolution of this fields in a moving fluid in which the robots operate. The evolutions of ρ(x, t) c(x, t) are governed by the diffu- sion equation: fusion coefficients. Another generalization would allow a variable number of agents in the system, in which case we would add a variable for the total number of agents in the system at time t to the variables introduced above describ- ing the fraction in various states. Equations 4 -- 7 together with appropriately chosen boun- dary conditions describe time evolution of the system. In the next section we use this formulation for a prototypical task: finding and localizing at the source of a chemical released into a flowing fluid. 3. TARGET LOCALIZATION WITH MICROSCOPIC ROBOTS Let us consider a D -- dimensional volume with multiple tar- gets that release certain chemical into the environment. The task of the microscopic swarm is to aggregate at these tar- gets in order to carry out some actions in the vicinity of the targets. This capability is fundamental to many medical applications envisioned for these microscopic robots. For example, the volume of fluid may be a blood vessel that has been damaged. Robots are required to aggregate at the in- jury site in order to assist in healing, forming clots, etc. We consider a simple robot controller that on a high level can be thought to consist of 3 discrete states described below: (6) State 1 (search): Do a biased random walk in the direction of the communicative signal concentration gradient. ∂ρ ∂t ∂c ∂t = Dρ∇2ρ − v · ∇ρ − γρρ + Qρ(x, t) = Dc∇2c − v · ∇c − γcc +Xk qk ¯nk(x, t) (7) In Eq. 6 the terms on the right describe, respectively, the diffusion of the chemical (with a diffusion constant Dρ), the advection of the chemical due to fluid motion with velocity v, the decay of the chemical at rate γρ, and its deposition by sources with intensity profile Qρ(x, t). Terms in Eq. 7 have similar meaning, except the deposition rates of sig- nalling chemical is proportional to the fraction of agents in state k, ¯nk(x, t) (note that, generally speaking, the coef- ficients qk themselves depend on the fraction of agents in state k). The parameters in this equation could, in general, depend on space and time, as well as the location of the robots (e.g., a sufficiently high concentration of the robots could significantly affect the fluid flow). For simplicity, we will treat them as constants. For microscopic robots, fluid motions will usually be at very low Reynolds number so the fluid flow will be laminar with the velocity v chang- ing smoothly with location. Viscous forces dominate the motions of such robots with requirements for locomotion mechanisms and power use quite different from experiences with larger robots [22]. Note that the system of Eqs. 4, 6 and 7 can be easily generalized to handle multiple chemicals with different dif- State 2 (communicate): Move towards the chemical source following the concentration gradient of the target chem- ical and release communicative signal to other robots. State 3 (disperse): Move away from the target in the direc- tion opposite to the target chemical's concentration gradient for some specified time τ. To fully specify a robot's behavior, we also have to de- scribe the transitions between these states. The robots start out in State 1, the searching for targets using random diffu- sive motion and following the gradient of the communica- tive signal. Once the concentration of the target chemical at a certain point in space is sufficiently high the robot at that point will switch to the State 2: it will start moving towards regions of high concentration (using biased diffu- sion or gradient following) while releasing a new chemical which acts as a communication signal to attract other robots. With some probability (that can be fixed, or dependent on the concentration of the robots at the source), robots in the State 2 will switch to State 3, where they will disperse from the source, moving in the direction opposite to the gradient. Finally, robots in the State 3 will switch to the searching state with probability 1/τ. The last behavior ensures that robots will not be stuck at local maxima of the chemical potential. Let denote by n1(x), n2(x), n3(x) the fraction of robots in each state at point x, with normalization condition Z dx(cid:0)n1(x) + n2(x) + n3(x)(cid:1) = 1. Let ρ(x) and c(x) be the concentrations of the chemical re- leased from the targets and the communicative signal re- leased by robots in State 2. We also denote by Vρ D and Vc D the robots' drift velocity in the concentration gradients of chemical (released by the targets) and communicative sig- nal (released by the robots), respectively. Then the set of equations describing the evolution of the system is as fol- lows: ∂n1 ∂t ∂n2 ∂t ∂n3 ∂t n3 τ = D1∇2n1 − v · ∇n1 − ∇ · [Vc − n1F (ρ) + = D2∇2n2 − v · ∇n2 − ∇ · [Vρ + n1F (ρ) − G(n2, ρ, c)n2 Dn1] Dn2] = D3∇2n3 − v · ∇n3 + ∇ · [Vρ + G(n2, ρ, c)n2 − n3 τ Dn3] (8) (9) (10) where F (ρ) is the concentration -- dependent transition rate from State 1 to State 2, G(n2; ρ; c) is the transition rate from State 2 to State 3, and 1/τ is the probability that a robot in State 3 will switch to State 1. Equations 8 -- 10 have a simple intuitive interpretation. The first two terms in Eq. 8 describe robots motion in State 1: diffusive searching and following communicative signal, if present. The third term describes the drift in the flow. The fourth term describes transitions to State 2 at the rate F (ρ), which depends on the concentration of the target field. The last term describes transition of robots from State 3 to State 1 after the robots have moved in the direction opposite to the concentration gradient for a period of time τ. G(n2, ρ, c) is the rate at which robots transition from State 2 to State 3, and it could in principle depend on the local concentrations of the gradients, as well as the number of agents present at the target site, for example, when presence of a certain min- imum number of robots is required for executing an action. We have to complement these three equations with two more to account for the evolution of chemicals ρ and c, that are obtained from Equations ??,?? and ?? as follows: ∂ρ ∂t ∂c ∂t = Dρ∇2ρ − v · ∇ρ + M Xi=1 Qiδ(x − xi) − γρρ (11) = Dc∇2c − v · ∇c + qcn2 − γcc (12) In Eq. 11 xi-s, i = 1, 2, ..M are the locations of the target sources, Qi is the intensity of source i, and γρ is the decay rate of the target chemical. Similarly, in Eq. 12 qc is the in- tensity of communication signal released by a robot in State 2, while γc is the decay rate of the signal. 4. RESULTS FOR A SIMPLIFIED 1D SCENARIO In this section we present results for a 1D geometry and a single target scenario. We consider the case when the liquid flow is very slow compared to other time scales so we can set v = 0. Also, since there is only one target, we neglect the third (dispersing) behavior so that two possible states are State 1 ("search") and State 2 ("communicate"). The target is located at x = 1 and serves as a point source for the chemical. We assume that the diffusion of the chemi- cal happens much faster compared to robots' diffusion, and it quickly reaches its steady state profile. Hence, the equa- tion for evolution of ρ(x, t) can be solved separately, with a solution ρ(x,∞) ≡ ρ(x) = Q0e−√γρ/Dρ(1−x), 0 ≤ x ≤ 1. (13) For the results presented here we used Q0 = 0.1, Dρ = 0.2 and γρ = 0.5. All robots start at State 1 and are initially localized at x = 0. We assume that a transition from State 1 to State 2 happens whenever a robot in State 1 detects the target's chemical above a certain threshold level ρ0, so that the tran- sition rate is F (ρ) = θ(ρ − ρ0), where θ(x) is the step function, θ(x) = 1 if x ≥ 0 and θ(x) = 0, x < 0. While in State 2, robots move in the chemical gradient with a con- stant drift velocity VD and release a communicative signal with intensity qc. To proceed further, we need to specify the dependence of the drift velocity in State 1 on the concentration of com- municative signal c. Again, we assume that once a robot detects communicative signal above certain threshold c0, it propels itself through the fluid in the direction of the gra- dient with a constant drift velocity VD. Then the dynamics of the system is described by the following system of equa- tions: ∂n1 ∂t ∂n2 ∂t ∂c ∂t = Dn = Dn = Dc ∂2n1 ∂x2 − VDθ(c − c0) ∂2n2 ∂x2 − VD ∂2c ∂x2 + qcn2 − γcc ∂n2 ∂x ∂n1 ∂x − F (ρ)n1 (14) (15) + F (ρ)n1 (16) To study the effect of different design parameters on ag- gregation behavior of the robots at the target, we solved the system Eq. 14 -- 16 numerically. We used the following pa- rameters (in dimensionless units): Dn = 0.01, Dc = 0.05, VD = 0.1, qc = 0.1, γc = 0.01. For the detection thresh- olds we used c0 = 0.001 and ρ0 = 0.01, the later assuring that that robots detect the chemical approximately midway in the interval [0, 1]. We used reflective boundary condi- tions for n1 and n2, ∂n1/∂x0,1 = ∂n2/∂x0,1 = 0, and absorbing boundary conditions for c, c(0) = c(1) = 0. In Fig. 1 we plot the spatia -- temporal evolution of robots' densities with and without communication. Clearly, the den- sity peak at x = 1 is stronger for the system with commu- nicative behavior. This suggests that communication indeed helps the robots to aggregate better. In addition, the ag- gregation process with communication happens faster than without communication. This is also shown in Fig. 2, where we plot the density of robots at x = 1 as a function of time for three different cases: free diffusion3 (VD = 0), gradient following without communication (VD 6= 0, qc = 0), and gradient following with communication (VD, qc 6= 0). As it can be seen from Fig. 2, the systems with gradient follow- ing and communicative behavior do demonstrate aggrega- tive behavior, and it is more pronounced for the system with communication. For instance, at time t = 10 the robot den- sity at x = 1 and with communication is more than 3 times higher than in the non -- communicating case. One of the design objectives is to have robots aggre- gate at the target fast enough, while at the same time not dissipating too much power due to the propelling. To ex- amine this tradeoff, let us consider the dependence of the aggregation time (defined as time needed for fraction n0 of robots to reach the vicinity of the target which we define as the interval [0.95, 1],) on the drift velocity VD. In Fig. 3 we plot aggregation time vs VD for three different values of n0. One observes that if increasing the drift velocity from VD = 0, the aggregation time decreases monotoni- cally, with a steeper decline for larger n0. However, it soon "saturates", so that increasing VD further has very small ef- fect on the aggregation time. This is because for large val- ues of VD/Dn, the aggregation time is mainly dominated by time required for robots to diffuse and detect chemical gradient, and increasing VD clearly does not have any effect on this time. Hence, depending on the desired number of robots in the vicinity of the target, as well as the required aggregation time, the best strategy for robots might be to have a moderate drift velocity. Note that this type of anal- ysis can be used to assess the energy -- efficiency of various behaviors since power required to propel a robot through a fluid with velocity VD scales with VD. . 3Note that the absence of aggregation for free diffusing robots is due to reflective boundary conditions at the source for n1 and n2. If one employs absorbing boundary conditions instead, robots will demonstrate aggrega- tive behavior even with free diffusion. n(x,t) 25 20 15 Time t 10 5 0.2 0 0 0.6 0.4 Distance x n(x,t) 1 0.8 (a) 12 10 8 6 4 2 0 30 12 8 4 0 30 20 10 Time t 0.2 0 0 0.6 0.4 Distance x 1 0.8 (b) Figure 1. Time evolution of robot densities without commu- nication (a) and with communication (b) 12 10 ) t , 1 = x ( n 8 6 4 2 0 0 free diffusion gradient following with communication 5 10 15 Time t 20 25 30 Figure 2. Time evolution of robot densities at x = 1 for three different strategies i e m T n o i t a g e r g g A 25 20 15 10 5 0 0 n =0.01 0 n =0.02 0 n =0.04 0 0.05 0.1 Drift Velocity V D 0.15 0.2 Figure 3. Aggregation time as a function of drift velocity, for three different values of n0. 5. DISCUSSION Microscopic robots face physical environments quite differ- ent from our experience with larger robots. In particularly, many interesting tasks for such robots will involve interac- tion with spatially varying fields, such as chemical concen- trations. In this paper, we showed how to incorporate such fields into a previously proposed approach relating individ- ual robot behaviors to that of the population as a whole. The analytic framework readily incorporates Brownian mo- tion and viscous fluid flow. The framework also applies to macroscopic robots programmed to emulate behavior of mi- croscopic organisms -- for example, chemotaxis [6]. As future work, it would be interesting to examine more realistic scenarios, involving three-dimensional fluid flow in vessels with quantitatively accurate physical parameters for the environment and robot sensors and actuators. We could also compare alternate approaches to various robot capabil- ities, e.g., acoustic vs. chemical communication. For in- stance, with more extensive use of communication, nearby robots could compare concentration measurements to bet- ter estimate concentration gradients, thereby introducing a spatial correlation in the state changes. Our approach ap- plies to these more complicated physical scenarios, but be- comes more computationally demanding to solve numeri- cally. Nevertheless, the computation cost involved to iden- tify useful design trade-offs through our approach will be less than that involved with detailed simulations (especially for tasks involving large populations of robots). Actual ex- periments with physical devices will be even more challeng- ing. Limitations of this approach are its underlying assump- tions, namely that to determine relevant collective swarm behaviors the occupancy numbers are a sufficient descrip- tion of the individual robots and fluctuations are small so the averages over many independent instances of the swarm are close to the actual observed behaviors in most of those instances. More broadly, this approach connects average, aggregate behavior with local robot controllers rather than providing specific details of individual robots. For applica- tions involving large populations of microscopic robots, the law of large numbers will ensure behaviors are usually close to average. 6. REFERENCES [1] J. P. Adler. Chemotaxis in bacteria. Science, 153:708 -- 716, 1966. [2] Yaakov Benenson, Binyamin Gil, Uri Ben-Dor, Rivka Adar, and Ehud Shapiro. An autonomous molecular computer for logical control of gene expression. Nature, 429:423 -- 429, 2004. [3] Hristo Bojinov, Arancha Casal, and Tad Hogg. Multiagent control of modular self-reconfigurable robots. Artificial In- telligence, 142:99 -- 120, 2002. Available as arxiv.org preprint cs.RO/0006030. [4] Eric Bonabeau, Marco Dorigo, and Guy Theraulaz. Swarm Intelligence: From Natural to Artificial Systems. Oxford University Press, New York, 1999. [5] Adriano Cavalcanti and Tad Hogg. Simulating nanorobots in fluids with low Reynolds number. In Proc. of the 11th Foresight Conference on Molecular Nanotechnology, Octo- ber 2003. [6] Amit Dhariwal, Gaurav S. Sukhatme, and Aristides A. G. Re- quicha. Bacterium-inspired robots for environmental moni- toring. In Proceedings of the 2004 IEEE Intl. Conference on Robotics and Automation, New Orleans, LA. IEEE Press, April 2004. [7] K. Eric Drexler. Nanosystems: Molecular Machinery, Man- ufacturing, and Computation. John Wiley, NY, 1992. [8] Radek Erban and Hans G. Othmer. From individual to col- lective behavior in bacterial chemotaxis. SIAM J. Applied Math, 65(2):361 -- 391, 2004. [9] Robert A. Freitas Jr. Nanomedicine, volume 1. des Bioscience, Georgetown, TX, 1999. www.nanomedicine.com. Lan- Available at [10] J. Fritz et al. Translating biomolecular recognition into nanomechanics. Science, 288:316 -- 318, 2000. [11] V. Gazi and K. M. Passino. Stability analysis of social forag- ing swarms. IEEE Trans. on Systems, Man and Cybernetics, B34:539 -- 557, 2004. [12] Joe Howard. Molecular motors: Structural adaptations to cellular functions. Nature, 389:561 -- 567, 1997. [13] Bernardo A. Huberman and Tad Hogg. The behavior of com- putational ecologies. In B. A. Huberman, editor, The Ecology of Computation, pages 77 -- 115, Amsterdam, 1988. Elsevier (North-Holland). [14] N. G. Van Kampen. Stochastic Processes in Physics and Chemistry. Elsevier Science, Amsterdam, revised and en- larged edition, 1992. [15] Balazs L. Keszler, Istvan J. Majoros, and James R. Baker Jr. Molecular engineering in nanotechnology: Structure and composition of multifunctional devices for medical applica- tion. In Proc. of the Ninth Foresight Conference on Molecu- lar Nanotechnology, 2001. [16] Kristina Lerman and Aram Galstyan. Two paradigms for the design of artificial collectives. In Kagan Tumer and David Wolpert, editors, Collectives and Design of Complex Sys- tems, pages 231 -- 256. Springer Verlag, New York, 2004. [17] Kristina Lerman, Alcherio Martinoli, and Aram Galstyan. A review of probabilistic macroscopic models for swarm robotic systems. In Sahin E. and Spears W., editors, Swarm Robotics Workshop: State-of-the-art Survey, number 3342 in LNCS, pages 143 -- 152. Springer-Verlag, Berlin Heidelberg, 2005. [18] Carlo Montemagno and George Bachand. Constructing nanomechanical devices powered by biomolecular motors. Nanotechnology, 10:225 -- 231, 1999. [19] Kelly Morris. Macrodoctor, come meet the nanodoctors. The Lancet, 357:778, March 10 2001. [20] NIH. National roadmap: http://nihroadmap.nih.gov/nanomedicine/index.asp. Nanomedicine, Institutes 2003. of Available Health at [21] Hans G. Othmer and Thomas Hillen. The diffusion limit of transport equations ii: Chemotaxis equations. SIAM J. Ap- plied Math, 62(4):1222 -- 1250, 2002. [22] E. M. Purcell. Life at low Reynolds number. American Jour- nal of Physics, 45:3 -- 11, 1977. [23] Aristides A. G. Requicha. Nanorobots, NEMS and nanoassembly. Proceedings of the IEEE, 91:1922 -- 1933, 2003. [24] B. Salemi, W.-M. Shen, and P. Will. Hormone controlled metamorphic robots. In Proc. of the Intl. Conf. on Robotics and Automation (ICRA2001), 2001. [25] Ron Weiss and Thomas F. Knight, Jr. Engineered communi- cations for microbial robotics. In Proc. of Sixth Intl. Meeting on DNA Based Computers (DNA6), 2000.
1611.08364	1	1611	2016-11-25T07:03:21	Robust Sequential Path Planning Under Disturbances and Adversarial Intruder	[ "cs.MA" ]	Provably safe and scalable multi-vehicle path planning is an important and urgent problem due to the expected increase of automation in civilian airspace in the near future. Although this problem has been studied in the past, there has not been a method that guarantees both goal satisfaction and safety for vehicles with general nonlinear dynamics while taking into account disturbances and potential adversarial agents, to the best of our knowledge. Hamilton-Jacobi (HJ) reachability is the ideal tool for guaranteeing goal satisfaction and safety under such scenarios, and has been successfully applied to many small-scale problems. However, a direct application of HJ reachability in most cases becomes intractable when there are more than two vehicles due to the exponentially scaling computational complexity with respect to system dimension. In this paper, we take advantage of the guarantees HJ reachability provides, and eliminate the computation burden by assigning a strict priority ordering to the vehicles under consideration. Under this sequential path planning (SPP) scheme, vehicles reserve "space-time" portions in the airspace, and the space-time portions guarantee dynamic feasibility, collision avoidance, and optimality of the paths given the priority ordering. With a computation complexity that scales quadratically when accounting for both disturbances and an intruder, and linearly when accounting for only disturbances, SPP can tractably solve the multi-vehicle path planning problem for vehicles with general nonlinear dynamics in a practical setting. We demonstrate our theory in representative simulations.	cs.MA	cs	Robust Sequential Path Planning Under Disturbances and Adversarial Intruder Mo Chen, Somil Bansal, Jaime F. Fisac, Claire J. Tomlin 6 1 0 2 v o N 5 2 ] A M . s c [ 1 v 4 6 3 8 0 . 1 1 6 1 : v i X r a Abstract-Provably safe and scalable multi-vehicle path plan- ning is an important and urgent problem due to the expected increase of automation in civilian airspace in the near future. Although this problem has been studied in the past, there has not been a method that guarantees both goal satisfaction and safety for vehicles with general nonlinear dynamics while taking into account disturbances and potential adversarial agents, to the best of our knowledge. Hamilton-Jacobi (HJ) reachability is the ideal tool for guaranteeing goal satisfaction and safety under such scenarios, and has been successfully applied to many small-scale problems. However, a direct application of HJ reachability in most cases becomes intractable when there are more than two vehicles due to the exponentially scaling computational complexity with respect to system dimension. In this paper, we take advantage of the guarantees HJ reachability provides, and eliminate the computation burden by assigning a strict priority ordering to the vehicles under consideration. Under this sequential path planning (SPP) scheme, vehicles reserve "space-time" portions in the airspace, and the space-time portions guarantee dynamic feasibility, collision avoidance, and optimality of the paths given the priority ordering. With a computation complexity that scales quadratically when accounting for both disturbances and an intruder, and linearly when accounting for only disturbances, SPP can tractably solve the multi-vehicle path planning problem for vehicles with general nonlinear dynamics in a practical setting. We demonstrate our theory in representative simulations. I. INTRODUCTION Recently, there has been an immense surge of interest in the use of unmanned aerial systems (UASs) in urban environ- ments. UASs have great potential in civil applications such as package delivery, aerial surveillance, disaster response, among many others [1]–[5]. Unlike previous uses of UASs for military purposes, civil applications will involve unmanned aerial vehi- cles (UAVs) flying in urban environments, potentially in close proximity of humans, other UAVs, and other important assets. As a result, government agencies such as the Federal Aviation Administration (FAA) and National Aeronautics and Space Administration (NASA) of the United States are urgently trying to develop new scalable ways to organize an airspace in which potentially thousands of UAVs can fly simultaneously in the same region [6], [7]. One essential problem that needs to be addressed is the safe multi-vehicle path planning problem: how a group of vehicles in the same vicinity can reach their destinations while avoiding This work has been supported in part by NSF under CPS:ActionWebs (CNS-931843), by ONR under the HUNT (N0014-08-0696) and SMARTS (N00014-09-1-1051) MURIs and by grant N00014-12-1-0609, by AFOSR under the CHASE MURI (FA9550-10-1-0567). The research of M. Chen and J. F. Fisac have received funding from the "NSERC" program and "la Caixa" Foundation, respectively. All authors are with the Department of Electrical Engineering and Com- puter Sciences, University of California, Berkeley. {mochen72, somil, jfisac, tomlin}@eecs.berkeley.edu situations which are considered dangerous, such as collisions. In many previous studies that address this problem, specific control strategies for the vehicles are assumed, and approaches such as those involving induced velocity obstacles [8]–[11] and involving virtual potential fields to maintain collision [12], [13] have been used. Other analyses of multi-vehicle systems include methods for real-time trajectory generation [14], for path planning for vehicles with linear dynamics in the presence of obstacles with known motion [15], and for cooperative path planning via waypoints which do not account for vehicle dynamics [16]. Other related work include those which consider only the collision avoidance problem without path planning. These results include those that assume the system has a linear model [17]–[19], rely on a linearization of the system model [20], [21], assume a simple positional state space [22], and many others [23]–[25]. However, the capability to flexibly plan provably safe and dynamically feasible trajectories without making strong assumptions on the vehicles' dynamics and other vehicles' motion is essential for dense groups of UAVs to safely fly in each other's vicinity. In addition, in a practical setting, any path planning scheme that also addresses collision avoidance must guarantee both goal satisfaction and safety of UAVs de- spite disturbances such as weather effects and communication faults [7]. Furthermore, unexpected scenarios such as UAV malfunctions or even UAVs with malicious intent need to be accounted for. The problem of trajectory planning and collision avoidance under disturbances in safety-critical systems has been stud- ied using Hamilton-Jacobi (HJ) reachability analysis, which provides guarantees on goal satisfaction and safety of optimal system trajectories [26]–[31]. Reachability-based methods are particular suitable in the context of UAVs because of the hard guarantees that are provided. In reachability analysis, one computes the reachable set, defined as the set of states from which the system can be driven to a target set. Many numerical tools are available for computing various definitions of reachable sets [32]–[35], and reachability analysis has been successfully used in applications involving systems with no more than two vehicles, such as pairwise collision avoidance [27], automated in-flight refueling [36], and many others [37], [38]. One of the main challenges of managing the next generation of airspace is the density of vehicles that needs to be accom- modated [7]. Such a large-scale system has a high-dimensional joint state space, making a direct application of dynamic programming-based approaches such as reachability analysis intractable. In particular, reachable set computations involve solving a HJ partial differential equation (PDE) or variational inequality (VI) on a grid representing a discretization of the state space, causing computational complexity to scale exponentially with system dimension. A. Contributions and Outline In this paper, we propose the sequential path planning (SPP) method to tackle the multi-vehicle path planning problem. Our approach is similar to the approaches of [39], [40], but provides hard guarantees on both the goal satisfaction and safety of all vehicles even in the presence of disturbances and a single intruder vehicle that could potentially be adversarial. In addition, our method scales only linearly with the number of vehicles when there is no intruder, and quadratically with the number of vehicles when there is a single intruder. On a high level, the SPP method assigns a strict priority ordering to the vehicles under consideration. Higher-priority vehicles plan their paths without taking into account the lower-priority vehicles. Lower-priority vehicles treat higher-priority vehicles as moving obstacles. Under this assumption, time-varying for- mulations of reachability [29], [31] can be used to obtain the optimal and provably safe paths for each vehicle, starting from the highest-priority vehicle. Thus, the curse of dimensionality is overcome for the multi-vehicle path planning problem at the cost of a mild structural assumption. In a sense, the SPP method reserves a portion of "space- time" in the airspace for each vehicle. The reserved space-time portion is recorded so that lower-priority vehicles can take it into account. Besides planning around the reserved space-time portions of higher-priority vehicles, no other communication between the vehicles is needed at execution time, even when disturbances and an intruder are present. In the absence of disturbances and intruders, and assuming each vehicle has perfect information about other vehicles' positions, each vehicle may plan and commit to an exact trajectory, with the reserved space-time being the collision set around the trajectory at every point in time. This basic concept of SPP is formally presented in Section IV. When the vehicles are affected by disturbances, exact tra- jectories cannot be known a priori, and thus the basic SPP algorithm cannot be directly applied. Fortunately, reachability analysis allows us to determine, at no additional computation cost, all possible states of each vehicle over time under the worst-case disturbance, given a control strategy. In addition, we can also determine suitable portions of space-time for each vehicle depending on the available information about the control strategies of higher-priority vehicles. SPP under disturbances and three different assumptions on the informa- tion available about the control strategy of other vehicles is formally presented in Section V. In scenarios where there could potentially be single, pos- sibly adversarial intruder in the airspace, each vehicle needs extra space around other vehicles in order to be able to perform avoidance maneuvers. Assuming the intruder may be present for some maximum duration, we use use reachability analysis to determine precisely the amount of space-time needed for each vehicle to be able to avoid the intruder under the presence of disturbances, making our proposed method sufficiently robust to most practical scenarios. SPP in the presence of a single intruder is formally presented in Section VI. II. PROBLEM FORMULATION Consider N vehicles which participate in the SPP pro- cess and denote these vehicles as the SPP vehicles Qi, i = 1, . . . , N. We assume their dynamics are given by i xi = fi(xi, ui, di), t ≤ tSTA ui ∈ Ui, di ∈ Di, i = 1 . . . , N (1) where xi ∈ Rni represents the state of vehicle Qi, ui ∈ Ui the control of Qi, and di ∈ Di the disturbance experienced by Qi. For convenience, we partition the state xi into the position component pi ∈ Rnp and the non-position component hi ∈ Rni−np: xi = (pi, hi). We assume that the control functions ui(·), di(·) are drawn from the set of measurable functions1. For convenience, we will use the sets Ui, Di to respectively denote the set of functions from which the control and disturbance functions ui(·), di(·) are drawn. : Rni × Ui × Di → Rni is uniformly continuous, bounded, and Lipschitz continuous in xi for fixed ui and di. With this assumption, given ui(·) ∈ Ui, di(·) ∈ Di, there exists a unique trajectory solving (1) [41]. In addition, we assume that the disturbances di(·) are drawn the set of non-anticipative strategies [27] Γ, defined as follows: We further assume that the flow field fi Γ := {N : Ui → Di : ui(r) = ui(r) a. e. r ∈ [t, s] ⇒ N [ui](r) = N [ui](r) a. e. r ∈ [t, s]} (2) i i Each vehicle Qi has initial state x0 i , and aims to reach its target Li by some scheduled time of arrival tSTA . The target in general represents some set of desirable states, for example the destination of Qi. In some situations, we may find that it is infeasible for Qi to get to Li at or before tSTA . Whenever unsure, we may first determine the earliest feasible tSTA as described in Section VI. On its way to Li, Qi must avoid a set of static obstacles Ostatic could be a tall building, a region around an airport, or any set of states that are forbidden for each SPP vehicle. ⊂ Rni. The interpretation of Ostatic In addition to the static obstacles, each vehicle Qi must also avoid the danger zones with respect to every other vehicle Qj, j (cid:54)= i. The danger zones in general can represent any joint configurations between Qi and Qj that are considered to be unsafe. In this paper, we define the danger zone of Qi with respect to Qj to be i i i Zij = {(xi, xj) : (cid:107)pi − pj(cid:107)2 ≤ Rc} (3) whose interpretation is that Qi and Qj are considered to be in an unsafe configuration when they are within a distance of Rc of each other. For concreteness, we will call Zij the collision set, and if (xi, xj) ∈ Zij, then Qi and Qj are said to have collided. Given the set of SPP vehicles, their targets Li, the static obstacles Ostatic , and the vehicles' danger zones with respect 1A function f : X → Y between two measurable spaces (X, ΣX ) and (Y, ΣY ) is said to be measurable if the preimage of a measurable set in Y is a measurable set in X, that is: ∀V ∈ ΣY , f−1(V ) ∈ ΣX , with ΣX , ΣY σ-algebras on X,Y . i to each other Zij, we would like, for each vehicle Qi, to synthesize a controller which guarantees that Qi reaches its target Li at or before the scheduled time of arrival tSTA , while avoiding the static obstacles Ostatic as well as the danger zones with respect to all other vehicles Zij, j (cid:54)= i. In addition, we would like to obtain the latest departure time tLDT such that Qi can still arrive at Li on time. i i i In general, the above optimal path planning problem must be solved in the joint space of all N SPP vehicles. How- ever, due to the high joint dimensionality, a direct dynamic programming-based solution is intractable. Therefore, we pro- pose to assign a priority to each vehicle, and perform SPP given the assigned priorities. Without loss of generality, let Qj have a higher priority than Qi if j < i. Under the SPP scheme, higher-priority vehicles can ignore the presence of lower-priority vehicles, and perform path planning without taking into account the lower-priority vehicles' danger zones. A lower-priority vehicle Qi, on the other hand, must ensure that it does not enter the danger zones of the higher-priority vehicles Qj, j < i; each higher-priority vehicle Qj induces a set of time-varying obstacles Oj i (t), which represents the possible states of Qi such that a collision between Qi and Qj could occur. is able to plan a trajectory that takes it to Li while avoiding the static obstacles Ostatic and the danger zones of higher- priority vehicles Qj, j < i, then the set of SPP vehicles Qi, i = 1, . . . , N would all be able to reach their targets safely. With the SPP scheme, the additional structure provided by the vehicle priorities allows us to reduce the complexity of the joint path planning problem. As we will see, under the SPP scheme, path planning can be done sequentially in descending order of vehicle priority in the state space of only a single vehicle. Thus, SPP provides a solution whose complexity scales linearly with the number of vehicles in the presence of disturbances, as opposed to exponentially with a direct application of dynamic programming approaches. In the presence of a single intruder, the computation complexity scaling becomes quadratic. is straight-forward to see that if each vehicle Qi It i In the following sections, we will explore SPP under dif- ferent assumptions. We begin with the basic SPP algorithm in which disturbances are ignored and perfect information of vehicles' positions is assumed. This simplification allows us to clearly establish the basic SPP algorithm. Next, we show how the basic SPP approach can be made robust to disturbances as well as an imperfect knowledge of other vehicles' positions. Finally, we further robustify the SPP approach by considering how the set of SPP vehicles may respond to the presence of an intruder vehicle which may be adversarial. All of our methods use time-varying reachability analysis to provide goal satisfaction and safety guarantees. III. TIME-VARYING REACHABILITY BACKGROUND We will be using reachability analysis to compute either a backward reachable set (BRS) V, a forward reachable set (FRS) W, or a sequence of BRSs and FRSs, given some target set L, time-varying obstacle G(t), and the Hamiltonian function H which captures the system dynamics as well as the roles of the control and disturbance. The BRS V in a time interval [t, tf ] or FRS W in a time interval [t0, t] will be denoted by V(t, tf ) W(t0, t) (backward reachable set) (forward reachable set) (4) In the SPP scheme, a lower-priority vehicle must avoid a set of moving obstacles on its way to the target. Several formulations of reachability are able to perform optimal path planning with hard guarantees on safety and performance under disturbances in such a scenario [29], [31]. For our application in SPP, we utilize the time-varying formulation in [31], which accounts for the time-varying nature of systems without requiring augmentation of the state space with the time variable. In the formulation in [31], a BRS is computed by solving the following final value double-obstacle HJ VI: max min{DtV (t, x) + H(t, x,∇V (t, x)), l(x) − V (t, x)}, (cid:110) (cid:111) − g(t, x) − V (t, x) t ≤ tf V (tf , x) = max{l(x),−g(tf , x)} = 0, (5) In a similar fashion, the FRS is computed by solving the following initial value HJ PDE: DtW (t, x)+H(t, x,∇W (t, x)) = 0, t ≥ t0 W (t0, x) = max{l(x),−g(t0, x)} (6) In both (5) and (6), the function l(x) is the implicit surface function representing the target set L = {x : l(x) ≤ 0}. Similarly, the function g(t, x) is the implicit surface function representing the time-varying obstacles G(t) = {x : g(t, x) ≤ 0}. The BRS V(t, tf ) and FRS W(t0, t) are given by V(t, tf ) = {x : V (t, x) ≤ 0} W(t0, t) = {x : W (t, x) ≤ 0} (7) Some of the reachability computations will not involve an obstacle set G(t), in which case we can simply set g(t, x) ≡ ∞ which effectively means that the outside maximum is ignored in (5). Also, note that unlike in (5), there is no inner mini- mization in (6). As we will see later, we will be using the BRS to determine all states that can reach some target set within the time horizon [t, tf ], whereas we will be using the FRS to determine where a vehicle could be at some particular time t. In addition, (6) has no outer maximum, since the FRSs that we will compute will not involve any obstacles. The Hamiltonian, H(t, x,∇V (t, x)), depends on the system dynamics, and the role of control and disturbance. Whenever H does not depend explicit on t, we will drop the argument. In addition, the Hamiltonian is an optimization that produces the optimal control u∗(t, x) and optimal disturbance d∗(t, x), once V is determined. For BRSs, whenever the existence of a control ("∃u") or disturbance is sought, the optimization is a minimum over the set of controls or disturbance. Whenever a BRS characterizes the behavior of the system for all controls ("∀u") or disturbances, the optimization is a maximum. We will introduce precise definitions of reachable sets, expressions for the Hamiltonian, expressions for the optimal controls as needed for the many different reachability calculations we use. IV. SPP WITHOUT DISTURBANCES AND WITH PERFECT INFORMATION In this section, we introduce the basic SPP algorithm assuming that there is no disturbance affecting the vehicles, and that each vehicle knows the exact position of higher- priority vehicles. Although in practice, such assumptions do not hold, the basic SPP algorithm can still serve as a useful approximation in certain situations. In addition, the description of the basic SPP algorithm will introduce the notation needed for describing the subsequent, more realistic versions of SPP. We also show simulation results for the basic SPP algorithm. The majority of the content in this section is taken from [42]. A. Theory Recall that the SPP vehicles Qi, i = 1, . . . , N, are each assigned a strict priority, with Qj having a higher priority than Qi if j < i. In the absence of disturbances, we can write the dynamics of the SPP vehicles as i i xi = fi(xi, ui), t ≤ tSTA ui ∈ Ui, i = 1 . . . , N and the obstacles Oj (8) In SPP, each vehicle Qi plans the path to its target set Li while avoiding static obstacles Ostatic i (t) induced by higher-priority vehicles Qj, j < i. Path planning is done sequentially starting from the first vehicle and proceeding in descending priority, Q1, Q2, . . . , QN so that each of the path planning problems can be done in the state space of only one vehicle. During its path planning process, Qi ignores the presence of lower-priority vehicles Qk, k > i, and induces the obstacles Oi k(t) for Qk, k > i. From the perspective of Qi, each of the higher-priority vehicles Qj, j < i induces a time-varying obstacle denoted Oj i (t) that Qi needs to avoid2. Therefore, each vehicle Qi must plan its path to Li while avoiding the union of all the induced obstacles as well as the static obstacles. Let Gi(t) be the union of all the obstacles that Qi must avoid on its way to Li: paths before Qi does. Thus, in the absence of disturbances, pj(t) is a priori known, and therefore Oj i (t), j < i are known, deterministic moving obstacles, which means that Gi(t) is also known and deterministic. Therefore, the path planning problem for Qi can be solved by first computing the BRS Vbasic ), defined as follows: (t, tSTA i i Vbasic i (t, tSTA i ) ={y : ∃ui(·) ∈ Ui, xi(·) satisfies (8), ∀s ∈ [t, tSTA ∃s ∈ [t, tSTA i ], xi(s) /∈ Gi(s), ], xi(s) ∈ Li, xi(t) = y} i (11) The BRS V(t, tSTA L = Li, G(t) = Gi(t), and the Hamiltonian i ) can be obtained by solving (5) with Hbasic (12) The optimal control for reaching Li while avoiding Gi(t) is (xi, λ) = min ui∈Ui i λ · fi(xi, ui) then given by λ · fi(xi, ui) ubasic i i i i i (t, xi) = arg min ui∈Ui (13) from which the trajectory xi(·) can be computed by integrating the system dynamics, which in this case are given by (8). In addition, the latest departure time tLDT can be obtained from the BRS V(t, tSTA i ∈ V(t, tSTA )}. In summary, the basic SPP algorithm is given as follows: i = arg supt{x0 ) as tLDT the case i = 1, Gi(t) = Ostatic Algorithm 1: Basic SPP algorithm: Suppose we are given i , vehicle dynamics (8), target sets Li, and initial conditions x0 static obstacles Ostatic , i = 1 . . . , N. For each i in ascending order starting from i = 1 (which corresponds to descending order of priority), 1) determine the total obstacle set Gi(t), given in (9). In 2) compute the BRS Vbasic latest departure time tLDT Vbasic ∀t; i (t, tSTA ) defined in (11). The is then given by arg supt{x0 i ∈ 3) determine the trajectory xi(·) using vehicle dynamics (8), with the optimal control ubasic 4) given xi(·), compute the induced obstacles Oi each k > i. In the absence of disturbances, Oi given by (10). (·) given by (13); k(t) for k(t) is (t, tSTA )}; i i i i i i B. Numerical Simulations Gi(t) = Ostatic i Oj i (t) (9) We now illustrate the basic SPP algorithm using a four- vehicle example. In this example, we will use the following dynamics for each vehicle: ∪ i−1(cid:91) j=1 With full position information of higher priority vehicles, the obstacle induced for Qi by Qj is simply Oj i (t) = {xi : (cid:107)pi − pj(t)(cid:107)2 ≤ Rc} (10) Each higher priority vehicle Qj plans its path while ignoring Qi. Since path planning is done sequentially in descending or- der or priority, the vehicles Qj, j < i would have planned their 2Note that the index k in Oi and the index j in Oj k denotes vehicles with lower priority than Qi, i (t) denotes vehicles with higher priority than Qi. px,i = vi cos θi py,i = vi sin θi θi = ωi ωi ≤ ¯ω (14) where xi = (px,i, py,i, θi) is the state of vehicle Qi, pi = (px,i, py,i) is the position, θi is the heading, vi is the speed, and ωi is the turn rate. In this example, we assume that the vehicles have constant speed vi = 1 ∀i, and that the control of each vehicle Qi is given by ui = ωi with ωi ≤ ¯ω = 1 ∀i. We have chosen these dynamics for clarity of illustration; the SPP algorithm can handle more general systems of the form in which the vehicles have different control bounds and dynamics. For this example, the target sets Li of the vehicles are circles of radius r in the position space; each vehicle is trying to reach some desired set of positions. In terms of the state space xi, the target sets are defined as Li = {xi : (cid:107)pi − ci(cid:107)2 ≤ r} (15) where ci are centers of the target circles. For the simulation of the basic SPP algorithm, we used r = 0.1. The vehicles have target centers ci, initial conditions x0 i , and scheduled times of arrivals tSTA as follows: i 1 = (−0.5, 0, 0), 2 = (0.5, 0, π), 3 = (−0.6, 0.6, 7π/4) , 4 = (0.6, 0.6, 5π/4) , c1 = (0.7, 0.2), x0 c2 = (−0.7, 0.2), x0 c3 = (0.7,−0.7), x0 c4 = (−0.7,−0.7), x0 tSTA 1 = 0 tSTA 2 = 0.2 tSTA 3 = 0.4 tSTA 4 = 0.6 (16) The setup for this example is shown in Fig. 1, which also shows the static obstacles as the black rectangles around the center of the domain. i i i = −1.48, tLDT The joint state space of this four-vehicle system is twelve- dimensional (12D), making the joint path planning and colli- sion avoidance problem intractable for direct analysis using HJ reachability. Therefore, we apply the SPP algorithm described in Algorithm 1 and repeatedly solve the double-obstacle HJ VI in (5) to obtain the optimal control for each vehicle to reach its target while avoiding higher-priority vehicles. In addition, due to the flexibility of the HJ VI with respect to time-varying systems, the different scheduled times of arrival tSTA can be trivially incorporated. Fig. 2, 3, and 4 show the simulation results. Since the state space of each vehicle is 3D, each of the BRSs Vbasic (t, tSTA ) is also 3D. To visualize the results, we slice the BRSs at the initial heading angles θ0 i . Fig. 2 shows the 2D BRS slices for each vehicle at its latest departure times tLDT 2 = = −1.44 determined from our −0.94, tLDT method. The obstacles in the domain Ostatic and the obstacles induced by higher-priority vehicles Oj i (t) inhibit the evolution of the BRSs, carving out thin "channels" that separate the BRSs into different "islands". One can see how these "chan- nels" and "islands" form by examining the time evolution of the BRS, shown in Fig. 3 for vehicle Q3. Finally, Fig. 4 shows the resulting trajectories of the four vehicles. Most interestingly, the subplot labeled t = −0.55 shows all four vehicles in close proximity without collision: each vehicle is outside of the danger zone of all other vehicles (although the danger zones may overlap). This close proximity is an indication of the optimality of the basic SPP algorithm given the assigned priority ordering. Since no disturbances are present, getting as close to other vehicles' danger zones as possible without entering the danger zones intuitively results in short transit times. 1 = −1.12, tLDT 4 3 i Fig. 1: Initial configuration of the four-vehicle example. The actual arrival times of vehicles Qi, i = 1, 2, 3, 4 are 0, 0.19, 0.34, 0.31, respectively. It is interesting to note that for some vehicles, the actual arrival times are earlier than the scheduled times of arrivals tSTA . This is because in order i to arrive at the target by tSTA , these vehicles must depart early enough to avoid major delays resulting from the induced obstacles of other vehicles; these delays would have led to a late arrival if vehicle Qi departed after tLDT . i i V. SPP WITH DISTURBANCES AND INCOMPLETE INFORMATION Disturbances and incomplete information significantly com- plicate the SPP scheme. The main difference is that the vehicle dynamics satisfy (1) as opposed to (8). Committing to exact trajectories is therefore no longer possible, since the disturbance di(·) is a priori unknown. Thus, the induced obstacles Oj i (t) are no longer just the danger zones centered around positions. A. Theory We present three methods to address the above issues. The methods differ in terms of control policy information that is known to a lower-priority vehicle, and have their relative advantages and disadvantages depending on the situation. The three methods are as follows: • Centralized control: A specific control strategy is en- forced upon a vehicle; this can be achieved, for example, by some central agent such as an air traffic controller. • Least restrictive control: A vehicle is required to arrive at its target on time, but has no other restrictions on its control policy. When the control policy of a vehicle is unknown, but its timely arrive at its target can be as- sumed, the least restrictive control can be safely assumed by lower-priority vehicles. • Robust trajectory tracking: A vehicle declares a nom- inal trajectory which can be robustly tracked under dis- turbances. In general, the above methods can be used in combina- tion in a single path planning problem, with each vehicle independently having different control policies. Lower-priority Initial Setup-0.500.5-0.8-0.6-0.4-0.200.20.40.6ObstacleTarget 1Target 2Target 3Target 4Position/Heading 1Danger zone 1Position/Heading 2Danger zone 2Position/Heading 3Danger zone 3Position/Heading 4Danger zone 4i Fig. 2: BRSs at t = tLDT for vehicles 1, 2, 3, 4, sliced at initial headings θ0 i . Black arrows indicate direction of obstacle motion. Due to the turn rate constraint, the presence of static obstacles Ostatic and time-varying obstacles induced by higher- priority vehicles Oj i (t) carve "channels" in the BRS, dividing it up into multiple "islands". i vehicles would then plan their paths while taking into account the control policy information known for each higher-priority vehicle. For clarity, we will present each method as if all vehicles are using the same method of path planning. In addition, for simplicity of explanation, we will assume that no static obstacles exist. In the situations where static obstacles do exist, the time-varying obstacles Gi(t) simply become the union of the induced obstacles Oj i (t) in (10) and the static obstacles. The material in this section is taken partially from [43]. 1) Centralized Control: The highest-priority vehicle Q1 first plans its path by computing the BRS (with i = 1) Vdstb ) ={y : ∃ui(·) ∈ Ui,∀di(·) ∈ Di, xi(·) satisfies (1), (t, tSTA i i ∀s ∈ [t, tSTA ∃s ∈ [t, tSTA i ], xi(s) /∈ Gi(s), xi(t) = y ], xi(s) ∈ Li} i (17) Since we have assumed no static obstacles exist, we have that for Q1,G1(s) = ∅ ∀s ≤ tSTA , and thus the above BRS is well-defined. This BRS can be computed by solving the HJ VI (5) with the following Hamiltonian: i Hdstb i (xi, λ) = min ui∈Ui max di∈Di λ · fi(xi, ui, di) (18) Fig. 3: Time evolution of the BRS for vehicle Q3, sliced at its initial heading θ0 4 . Black arrows indicate direction of obstacle motion. Top row: the BRS grows unobstructed by obstacles. Bottom row: the static obstacles Ostatic and the induced obstacles O1 3, carve out "channels" in the BRS. 3 = 7π 3,O2 i From the BRS, we can obtain the optimal control udstb i (t, xi) = arg min ui∈Ui max di∈Di λ · fi(xi, ui, di) (19) Here, as well as in the other two methods, the latest depar- ture time tLDT i is then given by arg supt x0 i ∈ Vdstb i (t, tSTA ). i If there is a central agent directly controlling each of the N vehicles, then the control law of each vehicle can be enforced. In this case, lower-priority vehicles can safely assume that higher-priority vehicles are applying the enforced control law. In particular, the optimal controller for getting to the target, udstb (t, xi), can be enforced. In this case, the dynamics of each i vehicle becomes xi = f cc di ∈ Di, i (t, xi, di) = fi(xi, udstb i i = 1, . . . , N, (t, xi), di) , tSTA ] t ∈ [tLDT i i (20) where ui no longer appears explicitly in the dynamics. From the perspective of a lower-priority vehicle Qi, a higher-priority vehicle Qj, j < i induces a time-varying obstacle that represents the positions that could possibly be within the collision radius Rc of Qj under the dynamics f cc j (t, xj, dj). Determining this obstacle involves computing Vehicle Q1, t=tLDT1=-1.12-1-0.500.51-1-0.500.51Vehicle Q2, t=tLDT2=-0.94-1-0.500.51-1-0.500.51Vehicle Q3, t=tLDT3=-1.48-1-0.500.51-1-0.500.51Vehicle Q4, t=tLDT4=-1.44-1-0.500.51-1-0.500.51ObstacleTargetsInitial pos. and headingReach-avoid setDanger zonest=0.40-1-0.500.51-1-0.500.51t=-0.27-1-0.500.51-1-0.500.51t=-0.88-1-0.500.51-1-0.500.51t=-1.48-1-0.500.51-1-0.500.51ObstacleTargetsInitial pos. and headingReach-avoid setDanger zonesi Fig. 4: The planned trajectories of the four vehicles. Top left: only vehicles Q3 (green) and Q4 (purple) have started moving, showing tLDT is not common across the vehicles. Top right: all vehicles have come within very close proximity, but none is in the danger zone of another. Bottom left: vehicle Q1 (blue) arrives at L1 at t = 0. Bottom right: all vehicles have reached their destination, some ahead of tSTA . i an FRS of Qj starting from3 xj(tLDT Wcc , t) is defined as follows: j ) = x0 j. The FRS where the set Pj(t) is the set of states in the FRS Wcc , t) projected onto the states representing position pj, and disre- garding the non-position dimensions hj: j (tLDT j Pj(t) = {pj : ∃hj, (pj, hj) ∈ Bj(t)}, Bj(t) = Wcc j (tLDT , t). j (24) (25) Finally, taking the union of the induced obstacles Oj i (t) as in (9) gives us the time-varying obstacles Gi(t) needed to define and determine the BRS Vdstb ) in (17). Repeating this process, all vehicles will be able to plan paths that guarantee the vehicles' timely and safe arrival. The centralized control algorithm can be summarized as follows: (t, tSTA i i conditions x0 obstacles Ostatic i , i = 1 . . . , N, for each i, ∀t; the case i = 1, Gi(t) = Ostatic i (t, tSTA ) defined in (17). The i ∈ is then given by arg supt x0 Algorithm 2: Centralized control algorithm: Given initial i , vehicle dynamics (1), target set Li, and static 1) determine the total obstacle set Gi(t), given in (9). In 2) compute the BRS Vdstb latest departure time tLDT Vdstb Vdstb (t, tSTA the FRS Wcc k > i. In the centralized control method, Oi computed using (23) where Pi(t) is given by (24). 4) finally, compute the induced obstacles Oi ) given by (19). Given udstb i (tLDT (t, xi) corresponding to (t, xi), compute 3) compute the optimal control udstb k(t) for each k(t) is , t) in (21); (t, tSTA ); i i i i i i i i i i i i (t, tSTA 2) Least Restrictive Control : Here, we again begin with the highest-priority vehicle Q1 planning its path by computing the BRS Vdstb ) in (17). However, if there is no centralized controller to enforce the control policy for higher-priority vehicles, weaker assumptions must be made by the lower- priority vehicles to ensure collision avoidance. One reasonable assumption is that all higher-priority vehicles follow the least restrictive control that would take them to their targets. This control would be given by (t, xj) in (19)} if xj(t) ∈ ∂Vdstb j (t, tSTA j ), (26) Such a controller allows each higher-priority vehicle to use any controller it desires, except when it is on the boundary of the BRS, ∂Vdstb ), in which case the optimal control udstb (t, xj) given by (19) must be used to get to the target j safely and on time. This assumption is the weakest assumption that could be made by lower-priority vehicles given that the higher-priority vehicles will get to their targets on time. (t, tSTA j j Suppose a lower-priority vehicle Qi assumes that higher- priority vehicles Qj, j < i use the least restrictive con- trol strategy ulrc j (t, xj) in (26). From the perspective of the lower-priority vehicle Qi, a higher-priority vehicle Qj could be in any state that is reachable from Qj's initial state j and from which the target Lj can be reached. xj(tLDT Mathematically, this is defined by the intersection of an FRS Wlrc j (tLDT j and the ) defined in (17) from the target set Lj, BRS Vdstb , t) from the initial state xj(tLDT (t, tSTA ) = x0 ) = x0 j j j j j j j (tLDT Wcc j (tLDT j , t) ={y : ∃dj(·) ∈ Dj, xj(·) satisfies (20), xj(tLDT j ) = x0 j , xj(t) = y}. (21) j (t, xj) ∈ ulrc This FRS can be computed using (6) with the Hamiltonian (cid:40){udstb j Uj otherwise Hcc The FRS Wcc j (t, xj, λ) = max dj∈Dj λ · f cc j (t, xj, dj) (22) j j j (tLDT , t) represents the set of possible states at time t of a higher-priority vehicle Qj given all possible disturbances dj(·) and given that Qj uses the feedback con- troller udstb (t, xj). In order for a lower-priority vehicle Qi to guarantee that it does not go within a distance of Rc to Qj, Qi must stay a distance of at least Rc away from the FRS Wcc , t) for all possible values of the non-position states hj. This gives the obstacle induced by a higher-priority vehicle Qj for a lower-priority vehicle Qi as follows: j (tLDT j Oj i (t) = {xi : ∃y ∈ Pj(t),(cid:107)pi − y(cid:107)2 ≤ Rc} (23) 3In practice, we define the target set to be a small region around the vehicle's initial state for computational reasons. t=-1.15-0.500.5-0.500.5t=-0.55-0.500.5-0.500.5t=0-0.500.5-0.500.5t=0.34-0.500.5-0.500.5ObstacleTargetsPositions, HeadingsTrajectoriesDanger Zonesj j (tLDT , t)∩Vdstb Wlrc ). In this situation, since Qj cannot be assumed to be using any particular feedback control, Wlrc , t) is defined as (t, tSTA j j j j (tLDT Wlrc j (tLDT j , t) ={y : ∃uj(·) ∈ Uj,∃dj(·) ∈ Dj, ) = x0 j , xj(·) satisfies (1), xj(tLDT xj(t) = y}. j (27) This FRS can be computed by solving (6) with the Hamil- tonian Hlrc j (xj, λ) = max uj∈Uj max dj∈Dj λ · fj(xj, uj, dj) (28) In turn, the obstacle induced by a higher-priority Qj for a lower-priority vehicle Qi is as follows: Oj i (t) = {xi : ∃y ∈ Pj(t),(cid:107)pi − y(cid:107)2 ≤ Rc}, where (29) Pj(t) = {pj : ∃hj, (pj, hj) ∈ Bj(t)}, Bj(t) = Wlrc (t, tSTA ). , t) ∩ Vdstb j (tLDT j j j (30) (31) The least restrictive control method can be summarized as follows: i initial conditions x0 static obstacles Ostatic , i = 1 . . . , N, for each i, ∀t; the case i = 1, Gi(t) = Ostatic i (t, tSTA ) defined in (17). The i ∈ is then given by arg supt x0 Algorithm 3: Least restrictive control algorithm: Given i , vehicle dynamics (1), target set Li, and 1) determine the total obstacle set Gi(t), given in (9). In 2) compute the BRS Vdstb latest departure time tLDT Vdstb (t, tSTA 3) compute the FRS Wlrc , t) and Vdstb ), compute the positions that the Qi could be in. The set of these positions is given by (30); k(t) for each k > i 4) compute the induced obstacles Oi i (t) in (27). Given Wlrc i (tLDT (t, tSTA ); i i i i i i i i using (29). Remark 1: The centralized control method described in the previous section can be thought of as the "most restrictive control" method, in which all vehicles must use the optimal controller at all times, while the least restrictive control method allows vehicles to use any suboptimal controller that allows them to arrive at the target on time. These two methods can be considered two extremes of a spectrum in which varying degrees of optimality is assumed for higher-priority vehicles. Vehicles can also choose a control strategy in the middle of the two extremes, and for example use a control within some range around the optimal control, or use the optimal control unless some condition is met. The induced obstacles and the BRS can then be similarly computed using the corresponding control strategy. 3) Robust Trajectory Tracking: Even though it is impossi- ble to commit to and track an exact trajectory in the presence of disturbances, it may still be possible to instead robustly track a feasible nominal trajectory with a bounded error at all times. If this can be done, then the tracking error bound can be used to determine the induced obstacles. Here, computation is done in two phases: the planning phase and the disturbance rejection phase. In the planning phase, we compute a nominal trajectory xr,j(·) that is feasible in the absence of disturbances. In the disturbance rejection phase, we compute a bound on the tracking error. It is important to note that the planning phase does not make full use of a vehicle's control authority, as some margin is needed to reject unexpected disturbances while tracking the nominal trajectory. Therefore, in this method, planning is done for a reduced control set U p ⊂ U. The resulting trajectory reference will not utilize the vehicle's full control capability; additional maneuverability is available at execution time to counteract external disturbances. In the disturbance rejection phase, we determine the error bound independently of the nominal trajectory. To compute this error bound, we find a robust controlled-invariant set in the joint state space of the vehicle and a tracking reference that may "maneuver" arbitrarily in the presence of an unknown bounded disturbance. Taking a worst-case approach, the track- ing reference can be viewed as a virtual evader vehicle that is optimally avoiding the actual vehicle to enlarge the tracking error. We therefore can model trajectory tracking as a pursuit- evasion game in which the actual vehicle is playing against the coordinated worst-case action of the virtual vehicle and the disturbance. Let xj and xr,j denote the states of the actual vehicle Qj and the virtual evader, respectively, and define the tracking error ej = xj − xr,j. When the error dynamics are independent of the absolute state as in (32) (and also (7) in [27]), we can obtain error dynamics of the form ej = fej (ej, uj, ur,j, dj), uj ∈ Uj, ur,j ∈ U p j , dj ∈ Dj, t ≤ 0 (32) To obtain bounds on the tracking error, we first conserva- tively estimate the error bound around any reference state xr,j, denoted Ej: Ej = {ej : (cid:107)pej(cid:107)2 ≤ REB}, (33) where pej denotes the position coordinates of ej and REB is a design parameter. We next solve a reachability problem with its complement E c j , the set of tracking errors violating the error bound, as the target in the space of the error dynamics. From E c j , we compute the following BRS: j (t, 0) ={y : ∀uj(·) ∈ Uj,∃ur,j(·) ∈ Up VEB ej(·) satisfies (32), ej(t) = y, ∃s ∈ [t, 0], ej(s) ∈ E c j}, j ,∃dj(·) ∈ Di, (34) where the Hamiltonian to compute the BRS is given by: λ · fej (ej, uj, ur,j, dj). HEB j (ej, λ) = max uj∈Uj (35) Letting t → −∞, we obtain the infinite-horizon control- invariant set Ωj := limt→−∞(cid:0)VEB j (t, 0)(cid:1)c. If Ωj is nonempty, then the tracking error ej at flight time is guaranteed to remain min j ,dj∈Dj ur∈U p 3) compute the BRS VEB 2) decide on a reduced control authority U p i for the plan- ning phase, and choose a parameter REB to conserva- tively bound the tracking error; i (t, 0) using (34) and make sure that Ωi (cid:54)= ∅. Given REB, the error bound on the tracking error is given by Ωi; 4) using Ωi, determine the augmented obstacle set Gi(t), given in (38); 5) compute the BRS Vrtt ) as described in (39) using the reduced target set Li, Gi(t) as obstacles, and the control authority U p is then given by arg supt x0 i . The latest departure time tLDT 6) compute the nominal trajectory xr,i(·) for Qi in the absence of disturbances, which can be obtained using the vehicle dynamics in (8) and the optimal control given in (41); computed using Ωi and xr,i(·) via (37). k(t) for each k > i can be 7) the induced obstacles Oi i ∈ Vrtt i (t, tSTA i (t, tSTA ); i i i B. Numerical Simulations We demonstrate our proposed methods for accounting for disturbances and incomplete information using a four-vehicle example. Each vehicle has the simple kinematics model in (14) but with disturbances added to the evolution of each state: px,i = vi cos θi + dx,i py,i = vi sin θi + dy,i θi = ωi + dθ,i, v ≤ vi ≤ ¯v,ωi ≤ ¯ω, (cid:107)(dx,i,dy,i)(cid:107)2 ≤ dr,dθ,i ≤ ¯dθ (42) where d = (dx,i, dy,i, dθ,i) represents Qi's disturbances in the three states. The control of Qi is ui = (vi, ωi), where vi is the speed of Qi and ωi is the turn rate; both controls have a lower and upper bound. For illustration purposes, we choose v = 0.5, ¯v = 1, ¯ω = 1; however, our method can easily handle the case in which these inputs differ across vehicles. The disturbance bounds are chosen as dr = 0.1, ¯dθ = 0.2, which correspond to a 10% uncertainty in the dynamics. within Ωj ⊆ Ej provided that the vehicle starts inside Ωj and subsequently applies the feedback control law κj(ej) = arg max uj∈Uj ur∈U p min j ,dj∈Dj λ · fej (ej, uj, ur,j, dj). (36) The induced obstacles by each higher-priority vehicle Qj can thus be obtained by: i (t) = {xi : ∃y ∈ Pj(t),(cid:107)pi − y(cid:107)2 ≤ Rc} Oj Pj(t) = {pj : ∃hj, (pj, hj) ∈ Bj(t)} Bj(t) = Ωj + xr,j(t), (37) where the "+" in (37) denotes the Minkowski sum4. Intu- itively, if Qj is tracking xr,j(t), then it will remain within the error bound Ωj around xr,j(t) ∀t. This is precisely the set Pj(t). The induced obstacles can then be obtained by augmenting a danger zone around this set. Finally, we can obtain the total obstacle set Gi(t) using (9). Since each vehicle Qj, j < i, can only be guaranteed to stay within Ωj, we must make sure during the path planning of Qi that at any given time, the error bounds of Qi and Qj, Ωi and Ωj, do not intersect. This can be done by augmenting the total obstacle set by Ωi: Gi(t) = Gi(t) + Ωi. (38) Finally, given Ωi, we can guarantee that Qi will reach its target Li if Ωi ⊆ Li; thus, in the path planning phase, we modify Li to be Li := {xi : Ωi + xi ⊆ Li}, and compute a BRS, with the control authority U p i , that contains the initial state of the vehicle. Mathematically, Vrtt i (t, tSTA i ) ={y : ∃ui(·) ∈ Up i , xi(·) satisfies (8), ], xi(s) /∈ Gi(t), ], xi(s) ∈ Li, xi(t) = y} ∀s ∈ [t, tSTA ∃s ∈ [t, tSTA i i (39) ) can be obtained by solving (5) using i (t, tSTA i The BRS Vrtt the Hamiltonian: Hrtt (40) The corresponding optimal control for reaching Li is given i (xi, λ) = min ui∈U p i λ · fi(xi, ui) by: λ · fi(xi, ui). i urtt i (t) = arg min ui∈U p (41) The nominal trajectory xr,i(·) can thus be obtained by using i (·) given by vehicle dynamics (8), with the optimal control urtt (41). From the resulting nominal trajectory xr,i(·), the overall control policy to reach Li can be obtained via (36). The robust trajectory tracking method can be summarized as follows: initial conditions x0 static obstacles Ostatic Algorithm 4: Robust trajectory tracking algorithm: Given i , vehicle dynamics (1), target sets Li, and 1) determine the total obstacle set Gi(t), given in (9). In , i = 1 . . . , N, for each i, i the case i = 1, Gi(t) = Ostatic ∀t; i 4The Minkowski sum of sets A and B is the set of all points that are the sum of any point in A and B. Fig. 5: Initial configuration of the four-vehicle example in the presence of disturbances. -0.500.5-0.500.5Initial SetupVehicle 1Vehicle 2Vehicle 3Vehicle 4For this example, we have chosen scheduled times of arrival i = 0 ∀i for simplicity. Each vehicle aims to get to a target tSTA set of the form (15) with target radius r = 0.1. The vehicles' i are given by (16). target centers ci and initial conditions x0 These parameters are the same as the example in Section IV-B, except that the tSTA values are the same for all vehicles, and that there are no static obstacles. The problem setup for this example is shown in Fig. 5. i With the above parameters, we obtain tLDT , i = 1, 2, 3, 4. Note that even though tSTA is assumed to be same for all vehicles in this example for simplicity, our method can easily handle the case in which tSTA is different for each vehicle as we have already shown in Section IV-B. i i i For each proposed method of computing induced obstacles, we show the vehicles' entire trajectories (colored dotted lines), and overlay their positions (colored asterisks) and headings (arrows) at a point in time in which they are in a relatively dense configuration. In all cases, the vehicles are able to avoid each other's danger zones (colored dashed circles) while getting to their target sets in minimum time. In addition, we show the evolution of the BRS over time for Q3 (green boundaries) as well as the obstacles induced by the higher- priority vehicles (black boundaries). Fig. 7: Evolution of the BRS and the obstacles induced by Q1 and Q2 for Q3 in the centralized control method. Since vehicles apply the optimal control at all times, the obstacle sizes are only slightly bigger than those in Fig. 2 and 3. Fig. 7 shows the evolution of the BRS for Q3 (green boundary), as well as the obstacles (black boundary) induced by the higher-priority vehicles Q1 and Q2. The locations of the induced obstacles at different time points include the actual positions of Q1 and Q2 at those times, and the sizes of obstacles remain relatively small. The tLDT values for the four vehicles (in order) in this case are −1.35,−1.37,−1.94 and −2.04, relatively close for vehicles pairs (Q1, Q2) and (Q3, Q4), because the obstacles generated by higher-priority vehicles are small and hence do not affect the tLDT of lower- priority vehicles significantly. i i 2) Least Restrictive Control: Fig. 8 shows the simulated trajectories in the situation where each vehicle assumes that higher-priority vehicles use the least restrictive control to reach their targets, as described in V-A2. Fig. 9 shows the BRS and induced obstacles for Q3. Q1 (red) takes a relatively straight path to reach its target. From the perspective of all other vehicles, large obstacles are induced by Q1, since lower-priority vehicles make the weak assumption that higher-priority vehicles are using the least restrictive control. Because the obstacles induced by higher-priority vehicles are so large, it is faster for lower- priority vehicles to wait until higher-priority vehicles pass by than to move around the higher-priority vehicles. As a result, the vehicles never form a dense configuration, and their trajectories are all relatively straight, indicating that they end up taking a short path to the target after higher-priority vehicles pass by. This is also indicated by the early tLDT i Fig. 6: Simulated trajectories in the centralized control method. Since the higher priority vehicles induce relatively small obstacles in this case, vehicles do not deviate much from a straight line trajectory towards their respective targets, and arrive at a dense configuration similar to that in Fig. 4. i 1) Centralized Control: Fig. 6 shows the simulated trajec- tories in the situation where a centralized controller enforces each vehicle to use the optimal controller udstb (t, xi) according to (19), as described in Section V-A1. In this case, vehicles appear to deviate slightly from a straight line trajectory to- wards their respective targets, just enough to avoid higher- priority vehicles. The deviation is small since the centralized controller is quite restrictive, making the possible positions of higher-priority vehicles cover a small area. In the dense configuration at t = −1.0, the vehicles are close to each other but still outside each other's danger zones. -0.500.5-0.500.5t = -1.0TargetsPositions, HeadingsDanger ZonesTrajectories-0.500.5-0.500.5t = -0.34-0.500.5-0.500.5t = -1.04-0.500.5-0.500.5t = -1.24-0.500.5-0.500.5t = -1.94TargetsInitial pos. and headingReachable setObstacleFig. 8: Simulated trajectories in the least restrictive control method. All vehicles start moving before Q1 starts, because the large obstacles make it optimal to wait until higher priority vehicles pass by, leading to earlier tLDT 's. i values for the four vehicles, −1.35,−1.97,−2.66 and −3.39, respectively. Compared to the centralized control method, tLDT 's are significantly earlier for all vehicles, except Q1, i the highest-priority vehicle, since it need not account for any moving obstacles. From Q3's (green) perspective, the large obstacles induced by Q1 and Q2 are shown in Fig. 9 as the black boundary. As the BRS (green boundary) evolves over time, its growth gets inhibited by the large obstacles for a long time, from t = −0.89 to t = −1.39. Eventually, the boundary of the BRS reaches the initial state of Q3 at t = tLDT 3 = −2.66. 3) Robust Trajectory Tracking: In the planning phase, we reduced the maximum turn rate of the vehicles from 1 to 0.6, and the speed range from [0.5, 1] to exactly 0.75 (constant speed). With these reduced control authorities, we determined from the disturbance rejection phase that a nominal trajectory from the planning phase can be robustly tracked within a distance of REB = 0.075. the tLDT In this case, Fig. 10 shows the vehicle trajectories in the situation where each vehicle robustly tracks a pre-specified trajectory and is guaranteed to stay inside a "bubble" around the trajectory. Fig. 11 shows the evolution of BRS and induced obstacles for vehicle Q3. The obstacles induced by other vehicles inhibit the evolution of the BRS, carving out thin channels, which can be seen at t = −2.59, that separate the BRS into different islands. values for the four vehicles are −1.61,−3.16,−3.57 and −2.47 respectively. In this method, vehicles use reduced control authority for path planning to- wards a reduced-size effective target set. As a result, higher- priority vehicles tend to have earlier tLDT 's compared to the other two methods, as evident from tLDT . Because of this "sacrifice" made by the higher-priority vehicles during the path planning phase, the tLDT's of lower-priority vehicles may be later compared to those in the other methods, as evident from tLDT will change for a vehicle compared to the other methods, as the conservative path planning leads to earlier tLDT 's for higher-priority vehicles . Overall, it is unclear how tLDT 4 i i 1 i i Fig. 9: Evolution of the BRS for Q3 in the least restrictive = −2.66, significantly control method. In this case, tLDT earlier than that in the centralized control method (−1.94), reflecting the impact of larger induced obstacles. 3 Fig. 10: Simulated trajectories for the robust trajectory tracking method. and later tLDT i 's for lower-priority vehicles. VI. SPP WITH AN INTRUDER In Section V, we made the basic SPP algorithm more robust by taking into account disturbances and considering situations in which vehicles may not have complete information about the control strategy of the other vehicles. However, if a vehicle not in the set of SPP vehicles enters the system, or even worse, if this vehicle is an adversarial intruder, the original plan can lead to vehicles entering into another vehicle's danger zone. -0.500.5-0.500.5t = -1.4TargetsPositions, HeadingsDanger ZonesTrajectories-0.500.5-0.500.5t = -0.39-0.500.5-0.500.5t = -0.89-0.500.5-0.500.5t = -1.39-0.500.5-0.500.5t = -2.66TargetsInitial pos. and headingReachable setObstacle-0.500.5-0.500.5t = -1.1TargetsPositions, HeadingsDanger ZonesTrajectoriesand that some fail-safe or enforcement mechanism exists to force the intruder out of the altitude level affecting the SPP vehicles. Note that we do not make any assumptions about t; however, we assume that once it appears, it stays for a maximum duration of tIAT. Assumption 2: The dynamics of the intruder are known and given by xI = fI (xI , uI , dI ). Assumption 2 is required for HJ reachability analysis. In situations where the dynamics of the intruder are not known exactly, a conservative model of the intruder may be used instead. Based on the above assumptions, we aim to design a control policy that ensures for each SPP vehicle separation with the intruder and with other SPP vehicles, and successful transit to the destination. However, depending on the initial state of the intruder, its control policy, and the disturbances in the dynamics of a vehicle and the intruder, a vehicle may arrive at different states after avoiding the intruder. Therefore, a control policy that ensures a successful transit to the destination needs to account for all such possible states, which is a path planning problem with multiple (infinite, to be precise) initial states and a single destination, and is hard to solve in general. Thus, we divide the intruder avoidance problem into two sub-problems. (i) We first design a control policy that ensures a successful transit to the destination if no intruder appears or successful avoidance of the intruder if it does appear. (ii) After the intruder is removed from the system at ¯t, we solve a new SPP problem (that is, we "re-plan") for vehicles which needed to avoid the intruder. In this case, the affected vehicles will re-plan as lowest-priority vehicles starting from the initial they happen to arrive at after avoiding the intruder. Suppose some vehicle Qi starts avoiding the intruder QI at some time t = t, and stops avoiding at t = ¯t. When t < t, Qi must plan its path taking into account the possibility that it may need to avoid an intruder QI. Since Qi may spend a duration of up to tIAT performing avoidance, its induced obstacles Oi k(t), k > i need to be computed in a way that reflects this possibility. The induced obstacles computation is discussed in Section VI-A1. We must also ensure that while avoiding the intruder, Qi does not collide with the total obstacle set Gi(t). This requires computing the augmented total obstacle Gi(t); the computation of Gi(t) and the controller that guarantees the avoidance of the augmented obstacles are discussed in Section VI-A2. In Section VI-A3, we describe how Qi can guarantee collision avoidance with the intruder. A pairwise collision avoidance problem such as this has been solved in isolation in [27]. Finally, when t > ¯t, Qi has already successfully avoided the intruder, but depending on the state it happens to arrive at after avoiding the intruder, it may need to re-plan its trajectory to reach the target safely. The re-planning process is discussed in Section VI-A4. 1) Induced Obstacle Computation: The goal of this section is to compute, for each lower-priority vehicle Qi, the time- varying obstacles induced by each higher-priority vehicle Qj, j < i, denoted by Oj i (t). As before, the total obstacle set Gi(t) can then be obtained using (9). To compute the obstacle Fig. 11: Evolution of the BRS for Q3 in the robust trajectory tracking method. As the BRS grows in time, the induced obstacles carve out a channel. Note that a smaller target set is used to compute the BRS to ensure that the vehicle reaches the target set by t = 0 for any allowed tracking error. If vehicles do not plan with an additional safety margin that takes a potential intruder into account, a vehicle trying to avoid the intruder may effectively become an intruder itself, leading to a domino effect. In this section, we propose a method to allow vehicles to avoid an intruder while maintaining the SPP structure. A. Theory In general, the effect of an intruder on the vehicles in structured flight can be entirely unpredictable, since the in- truder in principle could be adversarial in nature, and the number of intruders could be arbitrary. Therefore, for our analysis to produce reasonable results, two assumptions about the intruders must be made. Assumption 1: At most one intruder (denoted as QI) affects the SPP vehicles at any given time. The intruder is removed from the system after affecting the SPP vehicles after a duration of tIAT. The removal of the intruder can be done, for example, by forcing it out of the altitude range of the SPP vehicles. Let the time at which intruder appears in the system be t and the time at which it disappears be ¯t. Assumption 1 implies that ¯t ≤ t + tIAT. Thus, any vehicle Qi would need to avoid the intruder QI for a maximum duration of tIAT. This assumption can be valid in situations where intruders are rare, -0.500.5-0.500.5t = -0.50-0.500.5-0.500.5t = -2.09-0.500.5-0.500.5t = -2.59-0.500.5-0.500.5t = -3.57TargetsInitial pos. and headingReachable setObstaclethat Qi needs to avoid at time t, it is sufficient to consider the scenarios where t ∈ [t− tIAT, t]. This is because if t < t− tIAT, then the SPP vehicles would already be in the re-planning phase at time t and hence cannot be in conflict. Depending on the information known to a lower-priority vehicle Qi about Qj's control strategy, we can use one of the three methods described in Section V to compute the "base" obstacles Bj(t); these are the obstacles that would have been induced by Qj in the absence of an intruder. The base obstacles are respectively given by (25), (31) and (37) for the centralized control, least restrictive control and robust trajectory tracking methods. The induced obstacles, Oj i (t), are then given by the states that Qj can reach while avoiding the intruder, starting from some state in Bj(t), t ∈ [t − tIAT, t]. These states can be obtained by computing an FRS from the base obstacles. WO j (t − τ, t) ={y : ∃uj(·) ∈ Uj,∃dj(·) ∈ Dj, xj(·) satisfies (1), xj(t − τ ) ∈ Bj(t − τ ), xj(t) = y}. (43) WO j (t − τ, t) represents the set of all possible states that Qj can reach after a duration of τ starting from inside Bj(t− τ ). This FRS can be obtained by solving the HJ VI in (6) with the following Hamiltonian: HO j (xj, λ) = max uj∈Uj (44) Since τ ∈ [0, tIAT], the induced obstacles can be obtained as: λ · fj(xj, uj, dj). max dj∈Dj Pj(t) = {pj : ∃hj, (pj, hj) ∈ (cid:91) Oj i (t) = {xi : ∃y ∈ Pj(t),(cid:107)pi − y(cid:107)2 ≤ Rc} WO j (t − τ, t)} (45) τ∈[0,tIAT] Note that by the definition of base obstacles, Bj(t + τ2) ⊆ WBO j (t + τ1, t + τ2) ∀t, τ2 > τ1, where WBO j (t + τ1, t + τ2) denotes the FRS of Bj(t + τ1) computed for a duration of τ2 − τ1. Therefore, we have that WO j (t − tIAT, tIAT) ∀τ ∈ [0, tIAT]. Thus, Pj(t) in (45) can be equiva- lently written as j (t − τ, t) ⊆ WO Pj(t) = {pj : ∃hj, (pj, hj) ∈ WO j (t − tIAT, t)}. (46) 2) Augmented Obstacle Computation: We next need to en- sure that Qi doesn't collide with the obstacles Gi(·) computed in Section VI-A1 even when it is avoiding the intruder. In particular, we can compute a region around the obstacles Gi(·) such that for all disturbances, Qi can avoid colliding with obstacles for tIAT seconds regardless of its avoidance control, if Qi starts outside this region. Augmenting Gi(·) with this region gives us the augmented obstacles, Gi(·), that can then be used during the path planning of Qi to ensure collision avoidance with Gi(·). Suppose that the intruder appears in the system at some time time t = t − tIAT + τ, τ ∈ [0, tIAT]. In this case, we need to ensure that Qi does not collide with the obstacle Gi(t + τ ) at time t+τ, regardless of its control ui(s) and disturbance di(s) for the time interval s ∈ [t, t + τ ]. It is, therefore, sufficient to avoid the τ-horizon BRS of Gi(t + τ ) at time t. This argument applies for all τ ∈ [0, tIAT]. Mathematically, (cid:91) τ∈[0,tIAT] Gi(t) = VG i (t, t + τ ) (47) where VG backwards for τ seconds. Formally, i (t, t + τ ) represents BRS of Gi(t + τ ) computed i (t, t + τ ) ={y : ∃ui(·) ∈ Ui,∃di(·) ∈ Di, VG xi(·) satisfies (1), xi(t) = y, ∃s ∈ [t, t + τ ], xi(s) ∈ Gi(s)}. The Hamiltonian HG i to compute VG i (·) is given by: HG i (xi, λ) = min ui∈Ui min di∈Di λ · fi(xi, ui, di) (48) (49) Remark 2: Note that if we use the robust trajectory tracking method to compute the base obstacles, we would need to augment the obstacles in (47) by the error bound of Qi, Ωi, as discussed in section V-A3. Finally, we compute a BRS VAO ) for path planning that contains the initial state of Qi while avoiding these augmented obstacles: VAO (t, tSTA (t, tSTA ) ={y : ∃ui(·) ∈ Ui,∀di(·) ∈ Di, xi(·) satisfies (1),∀s ∈ [t, tSTA ∃s ∈ [t, tSTA ], xi(s) ∈ Li, xi(t) = y}. ], xi(s) /∈ Gi(s), i i i i i i The Hamiltonian HAO i (50) to compute BRS in (50) is given by: λ · fi(xi, ui, di) HAO i i max di∈Di (xi, λ) = min ui∈Ui Note that VAO (51) (·) ensures goal satisfaction for Qi in the absence of intruder. The goal satisfaction controller is given by: λ · fi(xi, ui, di) uAO i (t, xi) = arg min ui∈Ui max di∈Di Moreover, if Qi starts within VAO , it is guaranteed to avoid collision for a duration of tIAT, starting at any t < tSTA , irrespective of the control and disturbance applied during this time period. (52) i i 3) Optimal Avoidance Controller: First, we define relative dynamics of the intruder QI with state xI with respect to Qi with state xi. xI,i = xI − xi xI,i = fr(xI,i, ui, uI , di, dI ) (53) Given the relative dynamics, we compute the set of states from which the joint states of QI and Qi can enter danger zone ZiI despite the best efforts of Qi to avoid QI. This set of states is given by the BRS VCA(t, tIAT), t ∈ [0, tIAT]: VCA i (t, tIAT) ={y : ∀ui(·) ∈ Ui,∃uI (·) ∈ UI ,∃di(·) ∈ Di, ∃dI (·) ∈ DI , xI,i(·) satisfies (53), ∃s ∈ [t, tIAT], xI,i(s) ∈ LCA , xI,i(t) = y}, i (54) i = {xI,i : (cid:107)pI,i(cid:107)2 ≤ Rc}, and the Hamiltonian for where LCA computing this BRS is given by (cid:16) HCA i (xI,i, λ) = max ui∈Ui min uI∈UI ,di∈Di,dI∈DI (cid:17) λ · fr(xI,i, ui, uI , di, dI ) Once the value function V CA (t, xI,i) corresponding to the (t, tIAT) is computed, the optimal avoidance control i i can be obtained as: BRS VCA uCA i uCA i (t, xi, xI ) = arg max ui∈Ui (cid:16) (cid:17) i min uI∈UI ,di∈Di,dI∈DI λ · fr(xI,i, ui, uI , di, dI ) (55) Under normal circumstances when the intruder QI is far away, we have V CA (0, xI,i) > 0; as QI gets closer to Qi, (0, xI,i) decreases. If Qi applies the control uCA V CA i when i V CA (0, xI,i) = 0, then collision avoidance between Qi and i QI is guaranteed for a duration of tIAT under the worst-case intruder control strategy. In addition, obstacle augmentation (47) ensures that Qi does not collide with Gi(·) during the avoidance maneuver. The overall control policy for avoiding the intruder and collision with other vehicles is thus given by: (cid:26) uAO uA i (t) = i (t, xi) uCA i (t, xi, xI ) t ≤ t t ≤ t ≤ ¯t 4) Replanning after intruder avoidance: After the intruder disappears, goal satisfaction controllers which ensure that the vehicles reach their destinations can be obtained by solving an SPP problem as described in Section V, where the starting states of the vehicles are now given by the states they end up in, denoted x0 j, after avoiding the intruder. Let the optimal control policy corresponding to this goal satisfaction controller be denoted uL i (t, xi). The overall control policy that ensures intruder avoidance, collision avoidance with other vehicles, and successful transition to the destination is given by: (cid:26) uA u∗ i (t) = i (t, xi) uL i (t, xi) t ≤ ¯t t > ¯t Note that in order to re-plan using a SPP method, we need for all vehicles. This can be done to determine feasible tSTA by computing an FRS: i WRP i (¯t, t) ={y ∈ Rni : ∃ui(·) ∈ Ui,∀di(·) ∈ Di, xi(·) satisfies (1), xi(¯t) = x0 i , xi(t) = y,∀s ∈ [¯t, t], xi(s) /∈ GRP i represents the state of Qi at t = ¯t; GRP i (s)}, i (·) takes into where x0 account the fact that Qi needs to avoid higher-priority vehicles Qj, j < i and is defined in an way analogous to (9). (56) The FRS in (56) can be obtained by solving where W RP, gRP, lRP represent the FRS, obstacles during re- planning, and the initial state of Qi, respectively. The new tSTA of Qj is now given by the earliest time at which WRP j (¯t, t) j (¯t, t)∩Lj (cid:54)= intersects the target set Lj, tSTA ∅}. Intuitively, this means that there exists a control policy which will steer the vehicle to its destination by that time, despite the worst case disturbance it might experience. := arg inf t{WRP j Remark 3: Note that we only need to re-plan the trajectories of the vehicles that are affected by the intruder. In particular, if V CA(0, xI,i(t)) > 0 during the entire duration t ∈ [t, ¯t] for a vehicle, to apply any avoidance control, and hence re-planning would not be required for this vehicle. then the vehicle would need not Remark 4: In general, an intruder can be present in the system for much longer than tIAT, as long as it is not affecting the SPP vehicles. t thus really corresponds to the time an intruder starts affecting a SPP vehicle. Remark 5: Note that even though we have presented the analysis for one intruder, the proposed method can handle multiple intruders as long as only one intruder is present at any given time. We conclude this section with the overall SPP algorithm that takes into account an intruder that may appear for a duration of tIAT: , i = 1 . . . , N, for each i, Algorithm 5: Intruder Avoidance algorithm (offline plan- ning): Given initial conditions x0 i , vehicle dynamics (1), intruder dynamics in Assumption 2, target sets Li, and static obstacles Ostatic 1) determine the total obstacle set Gi(t), given in (9). In 2) compute the augmented obstacle set Gi(t) given by (47), 3) given Gi(t), compute the BRS VAO the case i = 1, Gi(t) = Ostatic where VG i (0, τ ) is given by (48); ) defined in (t, tSTA ∀t; i i i i (50); by computing VCA 4) the optimal control to avoid the intruder can be obtained 5) the induced obstacles Oi (t, tIAT) in (54) and using (55); k(t) for each k > i can be i computed using (45). Intruder Avoidance algorithm (online re-planning): For each vehicle i which performed avoidance in response to the intruder, 1) compute WRP 2) given tSTA i (¯t, t) using (56). The new tSTA for Qi is i , vehicle dynamics (1), target set Li, and , i = 1 . . . , N, use any of the three given by arg inf t{WRP static obstacles Ostatic SPP methods discussed in Section V for re-planning. j (¯t, t) ∩ Lj (cid:54)= ∅}; , x0 i i i B. Numerical Simulations (cid:110) (cid:111) max DtW RP(t, xi) + HRP i (t, xi,∇W RP(t, xi)), − gRP(t, xi) − W RP(t, xi) = 0 W RP(t, xi) = max{lRP(xi),−gRP(t, xi)} λ · fi(xi, ui, di) HRP i (xi, λ) = max ui∈Ui min di∈Di To illustrate that our SPP method is robust with respect to disturbances as well as a single intruder that is present for a duration of tIAT, we use a five-vehicle example in which one of the five vehicles is an intruder. We assume that each vehicle has the dynamics given in (42). For this example, we chose the parameters v = 0.1, ¯v = 1, ¯ω = 1, and disturbance bounds dr = 0.1, ¯dθ = 0.2, which correspond to a 10% uncertainty in the dynamics. (57) The vehicles' initial states, scheduled times of arrival, and target sets are the same as those described in Section V-B, except that in this example, we have increased the target radius to r = 0.15. For illustrate purposes, we have chosen to use the robust trajectory tracking method described in Section V-A3 for the base obstacles' computation, and hence each vehicle tracks a nominal trajectory. Fig. 12 shows the simulation at t = ¯t = −2.39, which corresponds to the time at which the intruder "disappears" from the domain. This time is chosen to maximally highlight the impact of the intruder. Here, the intruder is shown in black, and the SPP vehicles are shown in the other different colors. By the time t = −2.39, vehicle Q2 (red) and vehicle Q3 (green) have been avoiding the intruder for some time. This is evident from the amount of deviation between the actual positions of vehicles Q2 and Q3 (denoted by *) and their nominal positions (denoted by o) specified by the nominal trajectories they originally planned to track; these vehicles have abandoned nominal trajectory tracking in order to ensure safety with respect to the intruder. In contrast, Q4 (magenta), which has not needed to avoid the intruder, is tracking its nominal trajectory very closely (but not exactly, due to the presence of disturbances). The SPP vehicles are rather far apart because a large margin is needed to ensure that they maintain separation even when multiple vehicles need to avoid the intruder. In this example in particular, the lowest-priority vehicle Q4 needed to depart very early compared to Q2 and Q3 so that if an intruder were to arrive, Q4 does not impede the ability of the other vehicles to perform avoidance. The early departure of Q4 can be inferred from the fact that at t = −2.39, it is already nearly at its target. For the same reason, the highest-priority vehicle Q1 has not departed from its initial state yet, and thus is not shown at t = −2.39. Q2 and Q3 needed to depart very early compared to Q1 to ensure sufficient margin for avoidance maneuvers. Fig. 13 shows the nominal (black) and actual trajectories (red and green respectively) of vehicles Q2 (top subplot) and Q3 (bottom subplot). Specifically, the x and y positions over time are shown, and the black dotted vertical lines indicate the time interval in which the intruder is present. From Fig. 13, one can clearly see that before the intruder was present, both vehicles are able to track their nominal trajectories closely. When the intruder appears, the vehicles deviate from their nominal trajectories significantly. After the intruder disappears, both vehicles re-plan new trajectories, and at a later time, the resulting actual trajectories eventually arrive at the same location as the nominal trajectories. VII. CONCLUSIONS AND FUTURE WORK Guaranteed-safe multi-vehicle path planning is a challeng- ing problem, and previous analyses often either require strong assumptions on the motion of the vehicles or result in a large degree of conservatism. Optimal control and differential game techniques such as Hamilton-Jacobi (HJ) reachability are ideally suited for guaranteeing goal satisfaction and safety under disturbances, but become intractable for even a small number of vehicles. Fig. 12: The positions of the SPP vehicles and the intruder at t = −2.39, the end of the intruder's appearance. The red and green vehicles Q2, Q3 have not been tracking their nominal trajectories for a while, and have been avoiding the intruder instead. Thus, their positions are far away from their nominal trajectories, indicated by the small colored circles. Q4 has not needed to avoid the intruder, and tracks its nominal trajectory closely. The nominal trajectory of Q4 allows it to stay far enough away from other vehicles so that all vehicles can remain safe in the presence of the intruder. Fig. 13: The difference between the initially planned nominal trajectories and the actual trajectories for vehicles Q2 (top subplot) and Q3 (bottom subplot), which needed to perform avoidance with respect to the intruder during the time interval marked by the vertical black dotted lines. Before the intruder's presence, both vehicles track their nominal trajectories closely; however, both vehicles later deviate significantly from their nominal trajectories in order to avoid the intruder. After the intruder is gone, both vehicles replan their trajectories and arrive at their targets at a later time. Our robust sequential path planning (SPP) method assigns a strict priority ordering to vehicles to offer a tractable and practical approach to the multi-vehicle path planning problem. Under the proposed method, a portion of "space- time" is reserved for vehicles in the airspace in descending priority order to allow for dense vehicle configurations. Unlike previous priority-based methods, our approach accounts for disturbances and an adversarial intruder. SPP reduces the scaling of HJ reachability's computational complexity from exponential to linear with respect to the number of vehicles, while maintaining hard guarantees on goal satisfaction and safety under disturbances. In the presence of a single intruder vehicle, SPP still guarantees goal satisfaction and safety with -1-0.500.51-0.8-0.6-0.4-0.200.20.4t = -2.39TargetVehicle 2Vehicle 3Vehicle 4IntruderDanger zoneTrajectoryNominal Position-2.8-2.6-2.4-2.2-2-1.8-1.6-1.4-1.2-1-101Q2 PositionNominal x traj.Nominal y traj.Actual x traj.Actual y traj.Intruder time-3-2.5-2-1.5-1-0.5Time-101Q3 PositionNominal x traj.Nominal y traj.Actual x traj.Actual y traj.Intruder timea quadratically scaling computational complexity. In the future, we plan to investigate ways of guaranteeing a maximum number of vehicles that need to re-plan, combine reachability analysis with other path planning methods to improve computation speed, and to better understand the scenarios under which the SPP scheme is the most useful by running large-scale simulations. REFERENCES [1] B. P. Tice, "Unmanned Aerial Vehicles – The Force Multiplier of the 1990s," Airpower Journal, 1991. [2] W. DeBusk, "Unmanned Aerial Vehicle Systems for Disaster Relief: Tornado Alley," in AIAA Infotech@Aerospace 2010. American Institute of Aeronautics and Astronautics, Apr. 2010, pp. 2010–3506. [3] Amazon.com, Inc., "Amazon Prime Air," 2016. [Online]. Available: http://www.amazon.com/b?node=8037720011 [4] AUVSI News, "UAS Aid in South Carolina Tornado Investigation," 2016. [Online]. Available: http://www.auvsi.org/blogs/auvsi-news/2016/ 01/29/tornado [5] BBC Technology, "Google plans drone delivery service for 2017," 2016. [Online]. Available: http://www.bbc.com/news/technology-34704868 [6] Joint Planning and Development Office, "Unmanned Aircraft Systems (UAS) Comprehensive Plan," Federal Aviation Administration, Tech. Rep., 2014. [7] T. Prevot, J. Rios, P. Kopardekar, J. E. Robinson III, M. Johnson, and J. Jung, "UAS Traffic Management (UTM) Concept of Operations to Safely Enable Low Altitude Flight Operations," in 16th AIAA Aviation Technology, Integration, and Operations Conference. American Insti- tute of Aeronautics and Astronautics, Jun. 2016, pp. 1–16. [8] P. Fiorini and Z. Shiller, "Motion Planning in Dynamic Environments Using Velocity Obstacles," The International Journal of Robotics Re- search, vol. 17, no. 7, pp. 760–772, Jul. 1998. [9] G. Chasparis and J. Shamma, "Linear-programming-based multi-vehicle path planning with adversaries," in American Control Conference, Jun. 2005, pp. 1072–1077. [10] J. van den Berg, Ming Lin, and D. Manocha, "Reciprocal Velocity Ob- stacles for real-time multi-agent navigation," in International Conference on Robotics and Automation, May 2008, pp. 1928–1935. [11] A. Wu and J. P. How, "Guaranteed infinite horizon avoidance of unpredictable, dynamically constrained obstacles," Autonomous Robots, vol. 32, no. 3, pp. 227–242, Apr. 2012. [12] R. Olfati-Saber and R. M. Murray, "DISTRIBUTED COOPERA- TIVE CONTROL OF MULTIPLE VEHICLE FORMATIONS USING STRUCTURAL POTENTIAL FUNCTIONS," IFAC Proceedings Vol- umes, vol. 35, no. 1, pp. 495–500, 2002. [13] Y. L. Chuang, Y. R. Huang, M. R. D'Orsogna, and A. L. Bertozzi, "Multi-vehicle flocking: Scalability of cooperative control algorithms using pairwise potentials," in International Conference on Robotics and Automation, Apr. 2007, pp. 2292–2299. [14] Feng-Li Lian and R. Murray, "Real-time trajectory generation for the cooperative path planning of multi-vehicle systems," in Conference on Decision and Control, vol. 4, Dec. 2002, pp. 3766–3769. [15] A. Ahmadzadeh, N. Motee, A. Jadbabaie, and G. Pappas, "Multi-vehicle path planning in dynamically changing environments," in International Conference on Robotics and Automation, May 2009, pp. 2449–2454. [16] J. Bellingham, M. Tillerson, M. Alighanbari, and J. How, "Cooperative path planning for multiple UAVs in dynamic and uncertain environ- ments," in Conference on Decision and Control, vol. 3, Dec. 2002, pp. 2816–2822. [17] R. Beard and T. McLain, "Multiple UAV cooperative search under collision avoidance and limited range communication constraints," in Conference on Decision and Control, vol. 1, 2003, pp. 25–30. [18] T. Schouwenaars and E. Feron, "Decentralized Cooperative Trajectory Planning of Multiple Aircraft with Hard Safety Guarantees," in Guid- ance, Navigation, and Control Conference and Exhibit, Aug. 2004, pp. 2004–5141. [19] D. M. Stipanovic, P. F. Hokayem, M. W. Spong, and D. D. Siljak, "Cooperative Avoidance Control for Multiagent Systems," Journal of Dynamic Systems, Measurement, and Control, vol. 129, no. 5, p. 699, 2007. [20] M. Massink and N. De Francesco, "Modelling free flight with collision avoidance," in International Conference on Engineering of Complex Computer Systems, 2001, pp. 270–279. [21] M. Althoff and J. M. Dolan, "Set-based computation of vehicle behaviors for the online verification of autonomous vehicles," in International Conference on Intelligent Transportation Systems, Oct. 2011, pp. 1162– 1167. [22] Y. Lin and S. Saripalli, "Collision avoidance for UAVs using reachable sets," in International Conference on Unmanned Aircraft Systems, Jun. 2015, pp. 226–235. [23] E. Lalish, K. A. Morgansen, and T. Tsukamaki, "Decentralized reactive collision avoidance for multiple unicycle-type vehicles," in American Control Conference, Jun. 2008, pp. 5055–5061. [24] G. M. Hoffmann and C. J. Tomlin, "Decentralized cooperative collision avoidance for acceleration constrained vehicles," in Conference on Decision and Control, 2008, pp. 4357–4363. [25] M. Chen, C.-Y. Shih, and C. J. Tomlin, "Multi-Vehicle Collision Avoidance via Hamilton-Jacobi Reachability and Mixed Integer Pro- gramming," in Conference on Decision and Control (to appear), 2016. [26] E. N. Barron, "Differential games with maximum cost," Nonlinear Analysis, vol. 14, no. 11, pp. 971–989, Jun. 1990. [27] I. Mitchell, A. Bayen, and C. Tomlin, "A time-dependent Hamilton- Jacobi formulation of reachable sets for continuous dynamic games," Transactions on Automatic Control, vol. 50, no. 7, pp. 947–957, Jul. 2005. [28] O. Bokanowski, N. Forcadel, and H. Zidani, "Reachability and Minimal Times for State Constrained Nonlinear Problems without Any Controlla- bility Assumption," Journal on Control and Optimization, vol. 48, no. 7, pp. 4292–4316, Jan. 2010. [29] O. Bokanowski and H. Zidani, "MINIMAL TIME PROBLEMS WITH MOVING TARGETS AND OBSTACLES," IFAC Proceedings Volumes, vol. 44, no. 1, pp. 2589–2593, Jan. 2011. [30] K. Margellos and J. Lygeros, "HamiltonJacobi Formulation for ReachAvoid Differential Games," Transactions on Automatic Control, vol. 56, no. 8, pp. 1849–1861, Aug. 2011. [31] J. F. Fisac, M. , C. J. Tomlin, and S. S. Sastry, "Reach-avoid problems with time-varying dynamics, targets and constraints," in International Conference on Hybrid Systems Computation and Control, 2015, pp. 11– 20. [32] J. A. Sethian, "A fast marching level set method for monotonically advancing fronts." National Academy of Sciences, vol. 93, no. 4, pp. 1591–1595, Feb. 1996. [33] S. Osher and R. Fedkiw, Level Set Methods and Dynamic Implicit Surfaces. Springer-Verlag, 2006. [34] I. M. Mitchell, "Application of Level Set Methods to Control and Reach- ability Problems in Continuous and Hybrid Systems," Ph.D. dissertation, Stanford University, 2002. [35] "A toolbox of level set methods," Tech. Rep., 2007. [36] J. Ding, J. Sprinkle, S. S. Sastry, and C. J. Tomlin, "Reachability calculations for automated aerial refueling," in Conference on Decision and Control, 2008, pp. 3706–3712. [37] H. Huang, J. Ding, W. Zhang, and C. J. Tomlin, "A differential game approach to planning in adversarial scenarios: A case study on capture- the-flag," in International Conference on Robotics and Automation, 2011, pp. 1451–1456. [38] A. M. Bayen, I. M. Mitchell, M. K. Osihi, and C. J. Tomlin, "Aircraft Autolander Safety Analysis Through Optimal Control-Based Reach Set Computation," Journal of Guidance, Control, and Dynamics, vol. 30, no. 1, pp. 68–77, Jan. 2007. [39] M. Erdmann and T. Lozano-P´erez, "On multiple moving objects," Algorithmica, vol. 2, no. 1-4, pp. 477–521, Nov. 1987. [40] J. van den Berg and M. Overmars, "Prioritized motion planning for multiple robots," in IEEE/RSJ International Conference on Intelligent Robots and Systems, 2005, pp. 430–435. [41] N. L. Earl A. Coddington, Theory of ordinary differential equations. R.E. Krieger, 1955. [42] M. Chen, J. F. Fisac, S. Sastry, and C. J. Tomlin, "Safe sequential path planning of multi-vehicle systems via double-obstacle Hamilton- Jacobi-Isaacs variational inequality," in European Control Conference, Jul. 2015, pp. 3304–3309. [43] S. Bansal, M. Chen, J. F. Fisac, and C. J. Tomlin, "Safe Sequential Path Planning of Multi-Vehicle Systems Under Presence of Disturbances and Imperfect Information," in American Control Conference (Submitted), 2017.
cs/0403012	1	0403	2004-03-10T03:49:25	Distributed Control by Lagrangian Steepest Descent	[ "cs.MA", "cs.GT", "nlin.AO" ]	Often adaptive, distributed control can be viewed as an iterated game between independent players. The coupling between the players' mixed strategies, arising as the system evolves from one instant to the next, is determined by the system designer. Information theory tells us that the most likely joint strategy of the players, given a value of the expectation of the overall control objective function, is the minimizer of a Lagrangian function of the joint strategy. So the goal of the system designer is to speed evolution of the joint strategy to that Lagrangian minimizing point, lower the expectated value of the control objective function, and repeat. Here we elaborate the theory of algorithms that do this using local descent procedures, and that thereby achieve efficient, adaptive, distributed control.	cs.MA	cs	Distributed Control by Lagrangian Steepest Descent David H. Wolpert1, Stefan Bieniawski2, and Esfandiar Bandari2 1NASA Ames Research Center, Moffett Field, CA, 94035, and 2253 Durand, Dept. of Aeronautics Stanford, CA 94305, [email protected] {[email protected],[email protected]} Often adaptive, distributed control can be viewed as an iterated game between independent players. The coupling between the players' mixed strategies, arising as the system evolves from one instant to the next, is determined by the system designer. Information theory tells us that the most likely joint strategy of the players, given a value of the expectation of the overall control objective function, is the minimizer of a Lagrangian function of the joint strategy. So the goal of the system designer is to speed evolution of the joint strategy to that Lagrangian minimizing point, lower the expectated value of the control objective function, and repeat. Here we elaborate the theory of algorithms that do this using local descent procedures, and that thereby achieve efficient, adaptive, distributed control. PACS numbers: I. INTRODUCTION This paper considers the problem of adaptive dis- tributed control [1, 2, 3]. Typically in such problems, at each time t each control agent i sets its state xt i inde- pendently of the other agents, by sampling an associated distribution, qt i). Rather than directly via statisti- cal dependencies of the agents' states at the same time t, the coupling between the agents arises indirectly, through the stochastic joint evolution of their distributions {qt i } across time. i (xt More formally, let time be discrete, where at the be- ginning of each t every agent sets its state ("makes its move"), and then the rest of the system responds. Indi- cate the state of the entire system at time t as zt. (zt includes xt, as well as all stochastic elements not being di- rectly controlled.) So the joint distribution of the moves of the agents at any moment t is given by the product dis- i), and the state of the entire i (xt system, given joint move xt, is governed by P (zt xt). tribution qt(xt) = Qi qt Now in general the observations by agent i of aspects of the system's state at times previous to t will determine qt i . So qt i is statistically dependent on the previous states of the entire system, z{t′<t}. In other words, the agents can be viewed as players in a repeated game with Nature, each playing mixed strategies {qt i } at moment t [4, 5, 6, 7, 8]. Their interdependence arises through information sets and the like, in the usual way. From this perspective what the designer of a dis- tributed control system can specify is the stochastic laws governing the updating of the joint strategy. In other words, the designer wishes to impose a stochas- tic dynamics on a Multi-Agent System (MAS) that op- timizes an overall objective function of the state of the system in which the MAS is embedded, F (z).[32] For- mally, this means inducing a joint strategy q(x) with a good associated value of Eq(F ) = R dxq(x)E(F x) ≡ R dxq(x)G(x).[33] Once such a q is found, one can sample it to get a final x, and be assured that, on average, the associated F value is low. G is called the world utility. In this paper we elaborate a set of algorithms that iteratively update qt in such a manner. The algo- rithms presented here are based on using steepest descent techniques to minimize a G-parameterized Lagrangian, LG(q).[34] Because the descent is over Euclidean vectors q, these algorithms can be applied whether the xi are cat- egorical, continuous, time-extended, or a mixture of the three. So in particular, they provide a principled way to do "gradient descent over categorical variables". In the next section we first derive the Lagrangian LG(q) and discuss some of its properties. In the following section we show how to apply gradi- ent descent (and its embellishments) to optimization of the Lagrangian. If we view the agents as engaged in a team game, all having the same utility G, then this gra- dient descent is a distributed scheme for each agent to update its strategy, in a way that will steer the game to a bounded rational equilibrium [9, 14]. In this section we also consider second order methods. In contrast to gradient descent, in general any single ap- plication of Newton's method to update a product dis- tribution q will result in a new distribution p(q) that is not a product distribution. So we must instead solve for the product distribution q′(p) having minimal Kullback- Leibler distance to p. In this section we derive the rule for iterative updating of our distribution so as to move q in the direction of q′(p(q)). In practice any local descent scheme often requires Monte Carlo sampling to estimate terms in the gradient. To minimize the expected quadratic error of the estima- tion, typically the game is changed from being a team game. In other words, in general changing the agent's utilities gi to not all equal G will result in lower bias plus variance of the estimation of the gradient, and therefore will speed evolution to a good joint strategy. These and other techniques for shrinking bias plus variance are dis- cussed in [10, 11]. We end this section by mentioning some other tech- niques for improving the Monte Carlo sampling. These include data-aging, and techniques for managing the de- scent when it gets close to a border of the (simplex of) allowed q, Q. Most of these techniques introduced can be used even with schemes for minimizing LG(q) other than gradient descent. In the final section we introduce some alternatives to LG(q), designed to help speed convergence to a q with low E(G). Miscellaneous proofs can be found in the ap- pendix. The general mathematical framework for casting con- trol and optimization problems in terms of minimizing Lagangians of probability distributions is known as the theory of probability Lagangians. The precise version where the probability distributions are product distri- butions is known as "Product Distribution" (PD) the- ory. [11]. It has many deep connections to other fields, including bounded rational game theory and statistical physics [14]. As such it serves as a mathematical bridge connecting these disciplines. Some initial experimental results concerning the use of PD theory for distributed optimization and distributed control can be found in [15, 16, 17, 18, 19, 20]. See [10, 13, 21] for other uses and extensions of PD theory. II. PRODUCT DISTRIBUTION LAGRANGIANS A. The maxent Lagrangian Say the designer stipulates a particular desired value of E(G), γ. For simplicity, consider the case where the designer has no other knowledge concerning the system besides γ and the fact that the joint strategy is a product distribution. Then information theory tells us that the a priori most likely q consistent with that information is the one that maximizes entropy subject to that informa- tion [22, 23, 24].[35] In other words, of all distributions that agree with the designer's information, that distribu- tion is the "easiest" one to induce by random search. Given this, one can view the job of the designer of a distributed control system as an iterative equilibration process. In the first stage of each iteration the designer works to speed evolution of the joint strategy to the q with maximal entropy subject to a particular value of γ. Once we have found such a solution we can replace the constraint - replace the target value of E(G) - with a more difficult one, and then repeat the process, with another evolution of q [11]. Define the maxent Lagrangian by L(q) ≡ β(Eq(G) − γ) − S = β(Z dxq(x)G(x) − γ) − S(q), (1) S(q) is the Shannon entropy of where µ(x) , and for −R dxq(x)ln q(x) q, simplicity we here take the prior µ to be uniform.[36]. Writing it out, for a given γ, the associated most likely joint strategy is given by the q that minimizes L(q) over all those (q, β) such that 2 the Lagrange parameter β is at a critical point of L, i.e., such that ∂L ∂β = 0. Solving, we find that the qi are related to each other via a set of coupled Boltzmann equations (one for each agent i), qβ i (xi) ∝ e −βE q β (i) (Gxi) (2) where the overall proportionality constant for each i is set by normalization, the subscript qβ (i) on the expecta- tion value indicates that it is evaluated according to the distribution Qj6=i qj, and β is set to enforce the condition Eqβ (G) = γ. Following Nash, we can use Brouwer's fixed point theorem to establish that for any fixed β, there must exist at least one solution to this set of simultane- ous equations. If we evaluate E(G) at the solution qβ, we find that it is a declining function of β. So in following the iterative procedure of equilibrating and then lowering γ we we will raise β. Accordingly, we can avoid the steps of testing whether each successive constraint E(G) = γ is met, and simply monotonically increase β instead. This allows us to avoid ever explicitly specifying the values of γ. Simulated annealing is an example of doing this, where rather than work directly with q, one works with random samples of it formed via the Metropolis random walk algorithm [25, 26, 27, 28]. There is no a priori reason to use such an inefficient means of manipulating q however. Here we will work with q directly instead. This will result in an algorithm that is not simply "probabilistic" in the sense that the updating of its variables is stochastic (as in simulated annealing). Rather the very entity being updated is a probability distribution. B. Shape of the maxent Lagrangian Consider L as a function of q, with β and γ both treated as fixed parameters. (So in particular, Eq(g) need not equal γ.) First, say that q(i) is also held fixed, with only qi allowed to vary. This makes E(g) linear in qi. In addition, entropy is a concave function, and the unit simplex is a convex region. Accordingly, the Lagrangian of Eq. 1 has a unique local minimum over qi. So there is no issue of choosing among multiple minima when all of q(i) is fixed. Nor is there any problem of "getting trapped in a local minimum" in a computational search for that minimum. Indeed, in this situation we can just jump directly to that global optimum, via Eq. 2. Now introduce the shorthand for any function U (x), [U ]i,p(xi) ≡ Z dx(i)U (xi, x(i))p(x(i) xi). So [G]i,q(i) (xi) is agent i's "effective" cost function, ([G]i,q(i) ). This is Eq(i) (G xi). Consider the value Eqβ the value of E(G) at i's bounded rational equilibrium for the fixed q(i), i.e., it is the value at the minimum over qi of i L. View that value as a function of β. One can show that this is a decreasing function. In fact, its derivative just equals the negative of the variance of [G]i,q(i) (xi) evalu- ated under distribution qβ i (xi) (see appendix). Combin- ing this with the fact that E(G) is bounded below (for bounded G), establishes that the variance must go to zero for large enough β. So as β grows, qβ i (xi) → 0 for all xi that don't minimize Eq(i) (G xi). In other words, in that limit, qi becomes Nash-optimal. Next consider varying over all q ∈ Q, the space of all product distributions q. This is a convex space; if p ∈ Q and p′ ∈ Q, then so is any distribution on the line connecting p and p′. However over this space, the E(G) term in L is multilinear. So L is not a simple convex function of q. So we do not have guarantees of a single local minimum. The following lemma extends the technique of La- grange parameters to off-equilibrium points: Lemma 1: Consider the set of all vectors leading from x′ ∈ Rn that are, to first order, consistent with a set of constraints over Rn, {fi(x) = 0}. Of those vectors, the one giving the steepest ascent of a function V (x) is ~u = ∇V + Pi λi∇fi, up to an overall proportionality constant, where the λi enforce the first order consistency conditions, ~u · ∇fi = 0 ∀i. Now examine the derivatives of S(q) with respect to all components of q, i.e., the q-gradient of the entropy. At the border of Q, at least one of the ln(qi) terms in those derivatives will be negative infinite. Combined with Lemma 1, this can be used to establish that at the edge of Q, the steepest descent direction of any player's La- grangian points into the interior of Q (assuming finite β and {G}). (This is reflected in the equilibrium solutions Eq. 2.) Accordingly, whereas Nash equilibria can be on the edge of Q (e.g., for a pure strategy Nash equilibrium), in bounded rational games any equilibrium must lie in the interior of Q. In other words, any equilibrium (i.e., any local minimum) of a bounded rational game has non-zero probability for all joint moves. So just as when only vary- ing a single qi, we never have to consider extremal mixed strategies in searching for equilibria over all Q. We can use local descent schemes instead [15, 19, 29]. Lemma 1 can also be used to construct G with more than one solution to Eq. 2. One can also show that for every player i and any point q interior to Q, there are di- rections in Q along which i's Lagrangian is locally convex. Accordingly, no player's Lagrangian has a local maximum interior to Q. So if there are multiple local minima of i's Lagrangian, they are separated by saddle points across ridges. In addition, the uniform q is a solution to the set of coupled equations Eq. 2, but typically is not a local minimum, and therefore must be a saddle point. Say that we were not restricting ourselves to prod- uct distributions. So the Lagrangian becomes L(p) = β(Ep(G) − γ) − S(p), where p can now be any distribu- tion over x. There is only one local minimum over p of 3 this Lagrangian, the canonical ensemble: pβ(x) ∝ e−βG(x) In general pβ is not a product distribution. However we can ask what product distribution is closest to it. Now in general, the proper way to approximate a tar- get distribution p with a distribution from a subset C of the set of all distributions is to first specify a misfit measure saying how well each member of C approximates p, and then solve for the member with the smallest mis- fit. This is just as true when C is the set of all product distributions as when it is any other set. How best to measure distances between probability dis- tributions is a topic of ongoing controversy and research [30]. The most common way to do so is with the infi- nite limit log likelihood of data being generated by one distribution but misattributed to have come from the other. This is know as the Kullback-Leibler distance [22, 23, 31]: KL(p1 p2) ≡ S(p1 p2) − S(p1) (3) where S(p1 p2) ≡ −R dx p1(x)ln[ p2(x) µ(x) ] is known as the cross entropy from p1 to p2 (and as usual we im- plicitly choose uniform µ). The KL distance is always non-negative, and equals zero iff its two arguments are identical. As shorthand, define the "pq distance" as KL(p q), and the "qp distance" as KL(q p), where p is our tar- get distribution and q is a product distribution. Then it is straightforward to show that the qp distance from q to target distribution pβ is just the maxent Lagrangian L(q), up to irrelevant overall constants. In other words, the q minimizing the maxent Lagrangian is q with the minimal qp distance to the associated canonical ensem- ble. However the qp distance is the (infinite limit of the negative log of) the likelihood that distribution p would attribute to data generated by distribution q. It can be argued that a better measure of how well q approximates p would be based on the likelihood that q attributes to data generated by p. This is the pq distance; it gives a different Lagrangian from that of Eq. 1. Evaluating, up to an overall additive constant (of the canonical distribution's entropy), the pq distance is KL(p q) = −Xi Z dx p(x)ln[qi(xi)]. This is equivalent to a game where each coordinate i has the "Lagrangian" L∗ i (q) ≡ −Z dxi pi(xi)ln[qi(i)], (4) where pi(xi) is the marginal distributionR dx(i)p(x). The minimizer of this is just qi = pi ∀i, i.e., each qi is set to the associated marginal distribution of p. In the interests of space, the rest of this paper we restrict attention to the pq KL distance and associated maxent Lagrangian. III. DESCENT OF THE MAXENT LAGRANGIAN A. Gradient descent Consider the situation where each xi can take on a finite number of possible values, ξi. Say we are iter- atively evolving q to minimize L for some fixed β, and are currently at some point q ∈ Q. Using Lemma 1, we can evaluate the direction from q within Q that, to first order, will result in the largest drop in the value of L(q): ∂L(q) ∂qi(xi = j) = ui(j) −Xx′ i ui(x′ i)/ξi, (5) where ui(j) ≡ βE(G xi = j) + ln[qi(j)]. (Intuitively, i)/ξi is to keep the distribution in the set of all possible probability distribu- tions over x, P.) the reason for subtracting Px′ ui(x′ i Eq. 5 specifies the change that each agent should make to its distribution to have them jointly implement a step in steepest descent of the maxent Lagrangian. These up- dates are completely distributed, in the sense that each agent's update at time t is independent of any other agents' update at that time. Typically at any t each agent i knows qi(t) exactly, and therefore knows ln[qi(j)]. However often it will not know G and/or the q(i). In such cases it will not be able to evaluate the E(G xi = j) terms in Eq. 5 in closed form. One way to circumvent this problem is to have those expectation values be simultaneously estimated by all agents by repeated Monte Carlo sampling of q to pro- duce a set of (x, G(x)) pairs. Those pairs can then be used by each agent i to estimate the values E(G xi = j), and therefore how it should update its distribution. In the simplest version of such an update to q only occurs once every L time-steps. In this scheme only the samples (x, G(x)) formed within a block of L successive time-steps are used at the end of that block by the agents to update their distributions (according to Eq. 5). B. Higher order descent schemes In general, second order descent (e.g., Newton's method) of the maxent Lagrangian is non-trivial, due to coupling that arises between the agents and the re- quirement for associated matrix inversion. However re- call that one way to motivate the entropic product dis- tribution Lagrangian L(q) starts by saying that what we really want is the fully coupled canonical ensemble dis- tribution, pβ(x) ∝ exp(−βG(x)). L(q) then measures qp KL-distance to that desired distribution. From this perspective, any given iteration of second order descent of the maxent Lagrangian runs downhill on a quadratic approximation to a distribution, a distribution that is it- self a product distribution approximation to the ultimate distribution we want to minimize. 4 This suggests the alternative of making the approxi- mations in the opposite order. In this approach we first make a quadratic approximation (over the space of all p, not just all q) to the maxent Lagrangian, L(p). Via Newton's method this specifies a p∗ that minimizes that quadratic approximation. We can then find the product distribution that is nearest (in pq KL distance) to p∗. This scheme is called Nearest Newton descent. The gradient and Hessian of L(p) are given by ∂L ∂p(x) ∂2L ∂p(x)∂p(x′) p=p0 = βG(x) + 1 + ln(p0(x)) p=p0 = δx,x′ p0(x) where p0 is the current point. This Hessian is positive- definite (given that the current p is a member of P). By simple Lagrange parameters, the general solution is either on the border of P, or if in the interior is given by p∗(x) = −p0(x) (cid:2)βG(x) + ln(p(x)) + λ(cid:3) where λ is set by normalization. Solving, either p∗ is on the edge of the simplex, or p∗(x) p0(x) = 1 − S(p0) − ln(p0(x)) − β[G(x) − E(G)]. Note that the right-hand side is exactly the direction you should go using (simplex-constrained) gradient de- scent of L(p). So the direction to p∗ from p0 is given by the Hadamard product of p0 and the direction given by gradient descent. Now we can approximate p∗ with the product distri- bution having the minimal KL distance to it. In partic- ular, consider using pq KL distance rather than qp KL distance. Recall that for this kind of KL distance, the op- timal product distribution approximation to a joint dis- tribution is given by the product of the marginals of that joint distribution (see the discussion just below Eq. 4). Say that p0 is in the form of a product distribution, q0, i.e., that we are starting from a product distribution. Then calculating the marginals of the associated p∗ to get q∗ is trivial: q∗ i (j) q0 i (j) = 1 − S(q0 i ) − ln(q0 i (j)) − β[E(G xi = j) − E(G)] (6) Note that since the original quadratic approximation was over the full joint space, this formula automatically takes into account inter-agent couplings. In practice of course, it may make sense not to jump all the way from q0 to q∗, but only part-way there, to be conservative. (In fact, if q∗ isn't in the interior of the simplex, such partial jumping is necessary.) One potential guide to how far to i . Unlike the KL distances to the full joint Boltzmann distribution, we can readily calculate this KL distance. jump is the pq KL distance from p∗ to Qi q∗ The conditional expectations in Nearest Newton are the same as those in gradient descent. Accordingly, they too can be estimated via Monte Carlo sampling, if need be. It's also worth noting that Eq. 6 has the same form as one would get by evaluating the Hessian of the maxent Lagrangian, so long as one ignored inter-agent aspects of that Hessian. C. Practical issues In practice, the block-wise Monte Carlo sampling to estimate descent directions described above can be pro- hibitively slow. The estimates typically have high vari- ance, and therefore require large block size L to get a good descent direction. One set of ways to address this is to replace the team game with a non-team game, i.e., for each agent i have it estimate quantities E(gi xi = j) rather than E(G xi = j), where each private util- ity gi is chosen to have much lower variance than G [10, 11, 15].[37] Another useful technique is to allow samples from pre- ceding blocks to be re-used. One does this by first "ag- ing" that data to reflect the fact that it was formed under a different q(i) . For example, one can replace the empir- ical average for the most recent block k, Gi,j (k) ≡ PkL+L t=kL G(xt)δxt i,j PkL+L t=kL δxt i,j with a weighted average over the expected G's of all pre- ceding blocks, Pm Gi,j(m)e−κ(k−m) Pm e−κ(k−m) for some appropriate aging constant κ.[38] Typically such ageing allows L to be vastly reduced, and therefore the overall minimization of L to be greatly sped up. For such small L though, it may be that the most recent block has no samples of some move xi = j. This would mean that Gi,j(k) is undefined. One crude way to avoid such problems is to simply force a set of samples of each such move if they don't occur of their own accord, being careful to have the x(i) formed by sampling q(i) when forming those forced samples. Other useful techniques allow one to properly decrease the step size as one nears the border of Q. 5 IV. OTHER LAGRANGIANS FOR FINDING MINIMA OF G There are many alternative Lagrangians to the ones described above. The section focuses on such alterna- tive Lagrangians for the purpose of finding argminxG(x). Two classes of such Lagrangians are investigated: vari- ants of the Maxent Lagrangians, and variants of the two types of KL-distance Lagrangians. A. Maxent Lagrangians Say that after finding the q that minimizes the La- grangian, we IID sample that q, K times. We then take the sample that has the smallest G value as our guess for the x that minimizes G(x). For this to give a low x we don't need the mean of the distribution q(G) to be low - what we need is that the bottom tail of that distribu- tion is low. This suggests that in the E(G) term of the Maxent Lagrangian we replace q(x) → q(x) Θ[κ −R dx′ q(x′)Θ[G(x) − G(x′)]] κ where Θ is the Heaviside theta function. This new mul- tiplier of G is still a probability distribution over x. It equals 0 if G(x) is in the worst 1 − κ percentile (accord- ing to distribution q) of G values, and κ−1 otherwise. So under this replacement the E(G) term in the Lagrangian equals the average of G restricted to that lower κ'th per- centile. For κ = K −1, our new Lagrangian forces atten- tion in setting q on that outlier likely to come out of the K-fold sampling of q(G). One can use gradient descent and Monte Carlo sam- pling to minimize this Lagrangian, in the usual way. Note that the Monte Carlo process includes sampling the prob- ability distribution Θ[κ−R dx′ q(x′)Θ[G(x)−G(x′)]] as well as the qi. This means that only those points in the best κ'th percentile are kept, and used for all Monte Carlo estimates. This may cause greater noise in the Monte Carlo sampling than would be the case for κ = 1. κ As an example, say that for agent i, all of its moves have the same value of E(G xi), and similarly for agent j, and say that G is optimal if agents i and j both make move 0. Then if we modify the updating so that agent i only considers the best values that arose when it made move 0, and similarly for agent j, then both will be steered to prefer to make move 0 to their alternatives. This will cause them to coordinate their moves in an op- timal manner. A similar modification is to replace G with f (G) in the maxent Lagrangian, for some concave nowhere- decreasing function f (.). Intuitively, this will have the effect of coordinating the updates of the separate qi at the end of the block, in a way to help lower G. It does this by distorting G to accentuate those x's with good values. The price paid for this is that there will be more variance in the values of f (G) returned by the Monte Carlo sampling than those of G, in general. Note that if q is a local minimum of the Lagrangian for G, in general it will not be a local minimum for the Lagrangian of f (G) (the gradient will no longer be zero under that replacement, in general). So we can replace G with f (G) when we get stuck in a local minimum, and then return to G once q gets away from that local minimum. In this way we can break out of local minima, without facing the penalty of extra variance. Of course, none of these advantages in replacing G with f (G) hold for algorithms that directly search for an x giving a good G(x) value; x is a local minimum of G(x) ⇔ x is a local minimum of f (G(x)). An even simpler modification to the E(G) term than those considered above is to replace G(x) with Θ[G(x) − K]. Under this replacement the E(G) term becomes the probability that G(x) > K. So minimizing it will push q to x with lower G values. For this modified Lagrangian, the gradient descent update steps adds the following to each qi(xi): 6 low G, if β is large enough. This is why one way to min- imize G is to iteratively find a q that approximates the Boltzmann distribution, for higher and higher β. However there are other target distributions that are peaked about minimizers of G. In particular, given any distribution p′, the distribution θp(x) ≡ p(x)Θ[K − G(x)] R dx′ p(x′)Θ[K − G(x′)] is guaranteed to be more peaked about such minimizers than is p. So our minimization can be done by iterating the process of finding the q that best approximates θp and then setting p = q. This is analogous to the minimization algorithm considered in previous sections, which iterates the process of finding the q that best approximates the Boltzmann distribution and then increases β. For the choice of pq-KL distance as the approximation error, the q that best approximates θp is just the product of the marginal distributions of θp. So at the end of each iteration, we replace qi(xi) ← R dx′ (i)(x′ (i)q′ (i))Θ[K − G(xi, x′ R dx′ q′(x′)Θ[K − G(x′)] (i))] α(cid:2)βq(G < K xi) + ln(qi(xi)) βq(G < K x′ − Px′ i i) + ln(qi(x′ 1 i)) (cid:3). = = Px′ i q′(G < K xi) q′(G < K) q′(xi G < K) q′(xi) In gradient descent of the Maxent Lagrangian we must Monte Carlo estimate the expected value of a real number (G). In contrast, in gradient descent of this modified Lagrangian we Monte Carlo estimate the expected value of a single bit: whether G exceeds K. Accordingly, the noise in the Monte Carlo estimation for this modified Lagrangian is usually far smaller. In all these variants it may make sense to replace the Heaviside function with a logistic function or an exponen- tial. In addition, in all of them the annealing schedule for K can be set by periodically searching for the K that is (estimated to be) optimal, just as one searches for op- timal coordinate systems [11, 14]. Alternatively, a simple heuristic is to have K at the end of each block be set so that some pre-fixed percentage of the sampled points in the block go into our calculation of how to update q. Yet another possibility is to replace E(G) with the i.e., with the K such that κ'th percentile G value, R dx′ q(x′)Θ(G(x′) − K) = κ. (To evaluate the partial derivative of that k with respect a particular qi(xi) one must use implicit differentiation.) B. KL-based Lagrangians Both the qp-KL Lagrangian and pq-KL Lagrangians discussed above had the target distribution be a Boltz- mann distribution over G. For high enough β, such a distribution is peaked near argminxG(x). So sampling an accurate approximation to it should give an x with where q′ is the product distribution being replaced. The denominator term in this expression is known exactly to agent i, and the numerator can be Monte-Carlo estimated by that agent using only observed G values. So like gra- dient descent on the Maxent Lagrangian, this update rule is well-suited to a distributed implementation. Note that if we replace the Heaviside function in this algorithm with an exponential with exponent β, the up- date rule becomes qi(xi) ← E(e−βG xi) E(e−βG) . where both expectations are evaluated under q′, the dis- tribution that generated the Monte Carlo samples. It's interesting to compare this update rule with the parallel Brouwer update rule for the team game [11, 12, 19], to which it is very similar.[39] This update is guaranteed to optimize the associated Lagrangian, unlike the Brouwer update. On the other hand, since it is based on the pq- KL Lagrangian, as mentioned above there is no formal guarantee that this alternative to Brouwer updating will decrease E(G). This update rule is also very similar to the adaptive importance sampling of the original pq-KL approach dis- cussed in [11]. The difference is that in adaptive im- portance sampling the e−βG(x) terms get replaced by e−βG(x)/q′(x). Finally, consider using qp-KL distance to approximate q′(x) eβ(K−G(x)) R dx′ q′(x′)eβ(K−G(x′ )) rather than pq-KL distance. In the Lagrangian for that distance the q′ terms only con- tribute an overall additive constant. Aside from that con- stant, this qp-KL Lagrangian is identical to the Maxent Lagrangian. V. CONCLUSION Many problems in adaptive, distributed control can be cast as an iterated game. The coupling between the mixed strategies of the players arises as the system evolves from one instant to the next. This is what the system designer determines. Information theory tells us that the most likely joint strategy of the players, given a value of the expectation of the overall control objec- tive function, is the minimizer of a particular Lagrangian function of the joint strategy. So the goal of the sys- tem designer is to speed evolution of the joint strategy to that Lagrangian minimizing point, lower the expec- tated value of the control objective function, and repeat. Here we elaborate the theory of algorithms that do this using local descent procedures, and that thereby achieve efficient, adaptive, distributed control. VI. APPENDIX This appendix provides proofs absent from the main text. A. Derivation of Lemma 1 Proof: Consider the set of ~u such that the directional derivatives D~ufi evaluated at x′ all equal 0. These are the directions consistent with our constraints to first or- der. We need to find the one of those ~u such that D~uV evaluated at x′ is maximal. To simplify the analysis we introduce the constraint that ~u = 1. This means that the directional derivative D~uV for any function V is just ~u · ∇V . We then use La- grange parameters to solve our problem. Our constraints j = 1 and D~ufi(x′) = ~u · ∇fi(x′) = 0 ∀i. on ~u are Pj u2 Our objective function is D~uV (x′) = ~u · ∇V (x′). B. Proof of claims following Lemma 1 7 For generality, we provide the proofs in the general scenario where the private utilities gi may differ from one another. See the discussion in Section III C. i) Define fi(q) ≡ R dxiqi(xi), i.e., fi is the constraint forcing qi to be normalized. Now for any q that equals zero for some joint move there must be an i and an x′ i such that qi(x′ i) = 0. Plugging into Lemma 1, we can evaluate the component of the direction of steepest descent along the direction of player i's probability of making move x′ i: ∂L ∂qi(xi) + λ ∂fi ∂qi(xi) = βE(gi xi) + ln(qi(xi)) i [βE(gi x′′ R dx′′ −R dx′′ i ) + ln(qi(x′′ i 1 i ))] x′′ i ) + ln(qi(x′′ i such that qi(x′′ Since there must some x′′ i ) 6= 0, ∃xi such that βE(gi i )) is finite. Therefore our component is negative infinite. So L can be reduced by increasing qi(x′ i). Accordingly, no q having zero prob- ability for some joint move x can be a minimum of i's Lagrangian. ii) To construct a bounded rational game with multiple equilibria, note that at any (necessarily interior) local minimum q, for each i, Differentiating the Lagrangian gives 2λ0ui +Xi λi∇f = ∇V ∀i. with solution ui = ∇V −Pi λi∇f 2λ0 . λ0 enforces our constraint on ~u. Since we are only in- terested in specifying ~u up to a proportionality constant, we can set 2λ0 = 1. Redefining the Lagrange parameters by multiplying them by −1 then gives the result claimed. QED. βE(gi xi) + ln(qi(xi)) = βZ dx(i)gi(xi, x(i))Yj6=i qj(xj ) + ln(qi(xi)) must be independent of xi, by Lemma 1. So say there is a component-by-component bijection T (x) ≡ (T1(x1), T2(x2), . . .) that leaves all the {gj} unchanged, i.e., such that gj(x) = gj(T (x)) ∀x, j [40]. Define q′ by q′(x) = q(T (x)) ∀x. Then for any two values x1 i and x2 i , βEq′ (gi x1 i ) + ln(q′ i(x1 − βEq′ (gi x2 i )) i ) + ln(q′ i(x2 i )) = βZ dx(i)gi(x1 i , x(i))Yj6=i qj(T (xj)) + ln(qi(T (x1 i ))) − βZ dx(i)gi(x2 i , x(i))Yj6=i = βZ dx(i)gi(x1 i , T −1(x(i)))Yj6=i qj(T (xj))) + ln(qi(T (x2 i ))) qj(xj ) + ln(qi(T (x1 i ))) − βZ dx(i)gi(x2 i , T −1(x(i)))Yj6=i qj(xj)) + ln(qi(T (x2 i ))) 8 iii) To establish that at any q there is always a direction along which any player's Lagrangian is locally convex, fix all but two of the {qi}, q0 and q1, and fix both q0 and q1 for all but two of their respective possible values, which we can write as q0(0), q0(1), q1(0), and q1(1), respectively. So we can parameterize the set of q we're considering by two real numbers, x ≡ q0(0) and y ≡ q1(0). The 2 × 2 Hessian of L as a function of x and y is (cid:18) 1 x + 1 α a−x α b−y (cid:19) 1 y + 1 where a ≡ 1 − q0(0) − q0(1) and b ≡ 1 − q1(0) − q1(1), and x + 1 α is a function of gi and Qj6=0,1 qj. Defining s ≡ 1 b−y , the eigenvalues of that Hessian are and t ≡ 1 y + 1 a−x = βZ dx(i)gi(T (x1 i ), x(i)))Yj6=i qj(xj ) + ln(qi(T (x1 i ))) − βZ dx(i)gi(T (x2 i ), x(i)))Yj6=i qj(xj )) + ln(qi(T (x2 i ))) βEq(gi T (x1 = i )) + ln(qi(T (x1 i ))) − βEq(gi T (x2 i )) + ln(qi(T (x2 i ))) where the invariance of gi was used in the penultimate step. Since q is a local minimum though, this last differ- ence must equal 0. Therefore q′ is also a local minimum. Now choose the game so that ∀i, xi, T (xi) 6= xi. (Our congestion game example has this property.) Then the only way the transformation q → q(T ) can avoiding producing a new product distribution is if qi(xi) = qi(x′ i, i.e., q is uniform. Say the Hessians of the players' Lagrangians are not all positive definite at the uniform q. (For example have our congestion game be biased away from uniform multiplicities.) Then that q is not a local minimum of the Lagrangians. Therefore at a local minimum, q 6= q(T ). Accordingly, q and q(T ) are two distinct equilibria. i) ∀i, xi, x′ [1] D. Laughlin, M. Morari, and R. Braatz, "Robust perfor- mance of cross-directional control systems for web pro- cesses," Automatica, vol. 29, pp. 1394–1410, 1993. [2] J. Wolfe, D. Chichka, and J. Speyer, "Decentralized con- trollers for unmanned aerial vehicle formation flight," American Institute of Aeronautics and Astronautics, vol. 96, p. 3933, 1996. [3] M. Mesbai and F. Hadaegh, "Graphs, matrix inequalities, and switching for the formation flying control of multi- ple spacecraft," in Proceedings of the American Control Conference, San Diego, CA, 1999, pp. 4148–4152. [4] D. Fudenberg and J. Tirole, Game Theory. Cambridge, MA: MIT Press, 1991. [5] T. Basar and G. Olsder, Dynamic Noncooperative Game s + t ±p4α2 + (s − t)2 2 . The eigenvalue for the positive root is necessarily posi- tive. Therefore along the corresponding eigenvector, L is convex at q. QED. iv) There are several ways to show that the value of ([gi]i,q(i) ) must shrink as β grows. Here we do so by Eqβ evaluating the associated derivative with respect to β. i Define N (U ) ≡ R dy e−U(y), the normalization con- stant for the distribution proportional to e−U(y). View the xi-indexed vector qβ i as a function of β, gi and q(i). So we can somewhat inelegantly write E(gi) = Eqβ (gi). Then one can expand i (β,gi,q(i)),q(i) ∂E(gi) ∂β ∂2ln(N (β[gi]i,q(i) )) = − ∂β2 = −Var([gi]i,q(i) ) where the variance is over possible xi, sampled according to qβ i (xi). QED. Theory. Philadelphia, PA: Siam, 1999, second Edition. [6] M. Osborne and A. Rubenstein, A Course in Game The- ory. Cambridge, MA: MIT Press, 1994. [7] R. Aumann and S. Hart, Handbook of Game Theory with Economic Applications. North-Holland Press, 1992. [8] D. Fudenberg and D. K. Levine, The Theory of Learning in Games. Cambridge, MA: MIT Press, 1998. [9] D. H. Wolpert, "Bounded rationality game theory and information theory," 2004, submitted. [10] --, "Product distribution theory and collective intel- ligence," in Proceedings of WEHIA04, A. N. et al, Ed. Springer Verlag, 2004. a ensemble," [11] --, canonical 2003, "Factoring cond-mat/0307630. [12] S. Bieniawski and D. H. Wolpert, "Using product distri- butions for distributed optimization," in Proceedings of ICCS 04, 2004. [13] D. H. Wolpert and C. F. Lee, "Adaptive metropolis hast- ings sampling using product distributions," 2004, sub- mitted to ICCS04. [14] D. H. Wolpert, "Information theory - the bridge con- necting bounded rational game theory and statisti- cal physics," in Complex Engineering Systems, A. M. D. Braha and Y. Bar-Yam, Eds., 2004. [15] W. Macready, S. Bieniawski, and D. Wolpert, "Adaptive multi-agent systems for constrained optimization," 2004. [16] N. Antoine, S. Bieniawski, I. Kroo, and D. H. Wolpert, "Fleet assignment using collective intelligence," in Pro- ceedings of 42nd Aerospace Sciences Meeting, 2004, aIAA-2004-0622. [17] C. F. Lee and D. H. Wolpert, "Product distribution the- ory for control of multi-agent systems," 2004, submitted to AAMAS 04. [18] S. Airiau and D. H. Wolpert, "Product distribution the- ory and semi-coordinate transformations," 2004, submit- ted to AAMAS 04. [19] S. Bieniawski and D. H. Wolpert, "Adaptive, distributed control of constrained multi-agent systems," 2004, sub- mitted to AAMAS 04. [20] S. Bieniawski, D. H. Wolpert, and I. Kroo, "Discrete, continuous, and constrained optimization using collec- tives," in Proceedings of 10th AIAA/ISSMO Multidisci- plinary Analysis and Optimization Conference, Albany, New York, 2004, in press. [21] D. H. Wolpert, "Adaptive distributed control: beyond single-instant categorical variables," in Proceedings of MSRAS04, A. S. et al, Ed. Springer Verlag, 2004. [22] T. Cover and J. Thomas, Elements of Information The- ory. New York: Wiley-Interscience, 1991. [23] D. Mackay, Information theory, inference, and learning algorithms. Cambridge University Press, 2003. [24] E. T. Jaynes and G. L. Bretthorst, Probability Theory : The Logic of Science. Cambridge University Press, 2003. [25] S. Kirkpatrick, C. D. J. Gelatt, and M. P. Vecchi, "Opti- mization by simulated annealing," Science, vol. 220, pp. 671–680, May 1983. [26] R. Diekmann, R. Luling, and J. Simon, "Problem inde- pendent distributed simulated annealing and its applica- tions," in Applied Simulated Annealing. Springer, 1993, pp. 17–44. [27] O. Catoni, "Solving scheduling problems by simulated annealing," SIAM Journal on Control and Optimization, vol. 36, no. 5, pp. 1539–1575, 1998. [28] R. V. V. Vidal, Ed., Applied Simulated Annealing (Lec- ture Notes in Economics and Mathematical Systems). Springer, 1993. 9 [29] D. H. Wolpert and S. Bieniawski, "Theory of distributed control using product distributions," in Proceedings of CDC04, 2004. [30] D. H. Wolpert and W. Macready, "Metrics for sets of more than two points," in Proceedings of the Interna- tional Conference on Complex Systems, 2004, 2004, in press. [31] R. O. Duda, P. E. Hart, and D. G. Stork, Pattern Clas- sification (2nd ed.). Wiley and Sons, 2000. [32] Here we follow the convention that lower F is better. In addition, for simplicity we only consider objectives that depend on the state of the system at a single instant; it is straightforward to relax this restriction. [33] For simplicity, here we indicate integrals of any sort, in- cluding point sums for countable x, with the R symbol. [34] See [9, 10, 11, 12] for non-local techniques for finding qt, techniques that are related to ficticious play, and see [13] for techniques that exploit the Metropolis-Hastings algorithm. Other non-local techniques are related to im- portance sampling of integrals, and are briefly mentioned in [14]. [35] In light of how limited the information is here, the algo- rithms presented below are best-suited to "off the shelf" uses; incorporating more prior knowledge allows the al- gorithms to be tailored more tightly to a particular ap- plication. [36] Throughout this paper the terms in any Lagrangian that restrict distributions to the unit simplices are implicit. The other constraint needed for a Euclidean vector to be a valid probability distribution is that none of its com- ponents are negative. This will not need to be explicitly enforced in the Lagrangian here. [37] Formally, this means that each agent i has a separate Lagrangian, formed from Eq. 1 by substituting gi for G. The associaed joint solution q is then given by substitut- ing the appropriate gi for G in each instance of the cou- pled equations Eq. 2 (one instance for each i). See [14] for the relation of this to bounded rational game theory. [38] Not all preceding Gi,j (m) need to be stored to imple- ment this; exponential ageing can be done online using 3 variables per (i, j) pair. [39] That update is a variant of ficticious play, in which we simultaneously replace each qi(xi) with its ideal value if q(i) were to be held fixed, given by Eq. 2. [40] As an example, consider a congestion team game. In such a game all players have the same set of possible moves, and the shared utility G is a function only of the k- indexed bit string {N (x, k)}, where N (x, k) = 1 iff there is a move that is shared by exactly k of the players when the joint move is x. In this case T just permutes the set of possible moves in the same way for all players.
1901.02885	1	1901	2019-01-09T13:18:54	Swarm coordination of mini-UAVs for target search using imperfect sensors	[ "cs.MA" ]	Unmanned Aerial Vehicles (UAVs) have a great potential to support search tasks in unstructured environments. Small, lightweight, low speed and agile UAVs, such as multi-rotors platforms can incorporate many kinds of sensors that are suitable for detecting object of interests in cluttered outdoor areas. However, due to their limited endurance, moderate computing power, and imperfect sensing, mini-UAVs should be into groups using swarm coordination algorithms to perform tasks in a scalable, reliable and robust manner. In this paper a biologically-inspired mechanisms is adopted to coordinate drones performing target search with imperfect sensors. In essence, coordination can be achieved by combining stigmergic and flocking behaviors. Stigmergy occurs when a drone releases digital pheromone upon sensing of a potential target. Such pheromones can be aggregated and diffused between flocking drones, creating a spatiotemporal attractive potential field. Flocking occurs, as an emergent effect of alignment, separation and cohesion, where drones self organise with similar heading and dynamic arrangement as a group. The emergent coordination of drones relies on the alignment of stigmergy and flocking strategies. This paper reports on the design of the novel swarming algorithm, reviewing different strategies and measuring their performance on a number of synthetic and real-world scenarios.	cs.MA	cs	Draft version. Cite this work as: A.L. Alfeo, M.G.C.A. Cimino, N. De Francesco, A. Lazzeri, M. Lega, G. Vaglini, "Swarm coordination of mini-UAVs for target search using imperfect sensors", Intelligent Decision Technologies, IOS Press, Vol. 12, Issue 2, Pages 149-162, 2018 The final publication is available at IOS Press through http://dx.doi.org/10.3233/IDT-170317 Swarm coordination of mini-UAVs for target search using imperfect sensors Antonio L. Alfeo1, Mario G. C. A. Cimino1,, Nicoletta De Francesco1, Alessandro Lazzeri1, Massimiliano Lega2, Gigliola Vaglini1 1 Department of Information Engineering, University of Pisa, Largo L. Lazzarino 1, 56127 Pisa, Italy 2 Department of Engineering, University of Naples "Parthenope", 80143 Naples, Italy [email protected], [email protected], [email protected], [email protected], [email protected], [email protected] Corresponding author: Mario G. C. A. Cimino, [email protected] Tel: +39 050 2217 455; Fax: +39 050 2217 600. Abstract Unmanned Aerial Vehicles (UAVs) have a great potential to support search tasks in unstructured environments. Small, lightweight, low speed and agile UAVs, such as multirotors platforms can incorporate many kinds of sensors that are suitable for detecting object of interests in cluttered outdoor areas. However, due to their limited endurance, moderate computing power, and imperfect sensing, mini-UAVs should be into groups using swarm coordination algorithms to perform tasks in a scalable, reliable and robust manner. In this paper a biologically-inspired mechanisms is adopted to coordinate drones performing target search with imperfect sensors. In essence, coordination can be achieved by combining stigmergic and flocking behaviors. Stigmergy occurs when a drone releases digital pheromone upon sensing of a potential target. Such pheromones can be aggregated and diffused between flocking drones, creating a spatiotemporal attractive potential field. Flocking occurs, as an emergent effect of alignment, separation and cohesion, where drones self organise with similar heading and dynamic arrangement as a group. The emergent coordination of drones relies on the alignment of stigmergy and flocking strategies. This 2 paper reports on the design of the novel swarming algorithm, reviewing different strategies and measuring their performance on a number of synthetic and real-world scenarios. Keywords Swarm intelligence, mini-UAV, Stigmergy, Flocking, Target search 1 Introduction and Problem Statement In recent years, several research groups are working on new procedures and technologies to operate and monitor complex scenarios. Two specific areas include search and rescue and environmental monitoring. Both these topics require solutions to critical issues related to the mission requirements and the mission profile. The choice of a specific aerial platform for the monitoring of complex scenarios should be made by examining particular correspondence to the needs of the mission at the same time, and the multiplying effect of what is measurable by sensors positioned on the ground as fixed configuration. Advanced aerial platforms such as Unmanned Aerial Vehicles (UAVs), often called drones, are today the most frequent response to the needs of different missions. In particular, the specific category of mini- are generall have recently received a strong technological acceleration thanks to recent advances in miniaturization of battery, of communication, processing and sensing technology [1]. The remote/proximal sensing data obtained using mini-UAVs were validated in several environmental monitoring missions with complex scenarios as reported in previous research; these include: fusion of optical data with synthetic aperture radar data to detect environmental hazards [2,3], use of thermal imagery to monitor landfills [4], surface waters contamination [5] and to detect illegal dumping [6,7] and to identify other illegal activities [8]. In addition, remote sensing data can be strategically combined with other data layers in geographic information systems to monitor the vulnerability of cultural sites [9] and anticipate environmental violations [10,11]. 3 The use of a range of aerial platforms and advanced sensors to detect the illegal activities was validated in several real missions in Italy [12,13]. These are the first known use of these methods in both the fields of environmental research and law enforcement/environmental forensics. They also represent an example of collaboration between law enforcement and university teams on developing enhanced environmental protection methods. In the operational surveillance for successful identification and prosecution of environmental pollution culprits it is required an integrated system based on data from several sources. The surveillance service must also include geospatially tagged forensic data analysis (information arising from navigation/positional systems). The detection, identification and localisation of a target are key elements in all the above operations. Groups of mini-UAVs equipped with self-localisation and sensing capabilities offer new opportunities; indeed, groups of mini-UAVs can explore cluttered outdoor environments, where access to conventional platforms is inefficient, limited, impossible, or dangerous. In brief, the main motivations for adopting the UAV technology in the survey process are the following: reduction of risk of human falling, reduction of safety costs for plant stoppage, improved data density and quality due to a better proximity, accessibility to locations where people or vehicles have no access, faster and cheaper data acquisition due to the involvement of less workforce and equipment. The coordinated swarming drones could be also considered as single array of sensors configured to the measure of a host of environmental parameters. In search and rescue tasks, for example, a more effective appr to identify key locations as quick as possible. This exclusion process enables organisers to rescan the key locations that provided some circumstantial evidence. In this context, the quality of the sensing has also a direct impact on the overall mission performance [14]. 4 Therefore, an important aspect of the swarm coordination is the possibility to require a sufficient number of redundant samples of the target to reliably classify it A cooperative approach that exploits drones sensing, minimizes the error in target recognition [15]. In contrast, to use a unique drone implies costly structure and design, as well as vulnerability. Hence, a number of considerations support the use of coordinated swarming drones. An important requirement of the coordination strategy is to avoid centralized control approaches, leading to exponential increases in communication bandwidth and software complexity [16]. Swarm intelligence methodologies can be investigated to solve problems cooperatively while maintaining scalability. The main inspiration for swarming drones comes from the observation of social animals, such as insects, winged animals, and fish, that exhibit a collective intelligence which appears to achieve complex goal through simple rules and local interactions [17]. The main benefits of a swarm drones includes: robustness (for the ability to cope with the loss of individuals); scalability (due to the ability to perform well with different group size); and flexibility (thanks to the capability to manage a broad spectrum of different environments and tasks). To this aim, each individual of the swarm: acts with a certain level of autonomy; performs only local sensing and communication; operates without centralized control or global knowledge, and cooperates to achieve a global task [17]. In this paper, different coordination strategies are reviewed and tested empirically with both synthetic and real-world scenarios, with obstacles having irregular complex shapes. For this purpose, it is adopted a multi-agent simulation platform with the possibility of importing environments with obstacles and targets sampled from real landscapes. The paper is structured as follows. In Section 2 early requirements and coordination strategies are reported. Section 3 briefly characterizes the related work. In Section 4, the 5 analysis and the integration of the emergent schemes is covered. Section 5 reports on the design of the algorithm. Experimental studies are detailed in Section 6. Section 7 draws conclusions and future work. 2 Early requirements and coordination strategies From a structural standpoint, it is assumed that each mini-UAV is provided with the following capabilities: wireless communication capability for sending and receiving information from the ground station; self-location capability based on Global Position System (GPS) and inertial technology, returning the coordinates of its current location; one or more target sensing technology, capable of acquire data in the area over which it flies; processor with limited computing capability; obstacle avoidance capability, that is, locally managed detection and steering to avoid flying towards surrounding barriers and drones. Moreover, it is assumed that a certain level of uncertainty comes from noisy of faulty sensor measurements. Marker-based stigmergy is a fundamental swarm coordination mechanism, based on the release of information in the environment in the form of pheromones [18,19]. The pheromone is a volatile substance diffused locally and staying temporarily for other individuals that can properly react and modify their behavior [20]. Simulated (that is, digital) pheromones can be used to coordinate groups of drones for various tasks. In a distributed environment, a pheromone map of the search space can be maintained and made [21]. When the sensing system of a drone determines a potential target, it tries to trigger the cooperation of its swarm to achieve reliable sensing and target detection. Reliable sensing. Sensors on mini-UAV can generate faulty measurements for a number of reasons, such as power loss, software failure, small bias, miscalibration, slow-drifting, loss of accuracy, temporary freezing, to name a few [22]. In the literature, some 6 approaches to fault recognition assume the fault types can be described by a static parameterized model. If parameterized models for the fault types are available, a fault recognition algorithm can be applied. However, without health monitoring a static fault model is often not known [22]. In practice, there is likely to be some deviation between what the actual faults look like and what their models predict. This residual may be irrelevant for a single sensor system. However, in a conventional distributed approach even small residuals can have a significant impact on the overall effectiveness, due to the high number of occurrences potentially involved. In contrast, swarm systems could exhibit much higher levels of robustness, in the sense of tolerance to individual (or few) residual(s), than in conventional distributed systems. Nevertheless, a simplistic modelling approach may make incorrect assumptions, because the question of how many agents are needed to guarantee a required emergent behavior in a particular swarm and for a particular behavior is not straightforward [23]. This potential tolerance cannot be natively assumed without special analysis, design, and test, since swarm systems can exhibit a number of unexpected behaviors. Therefore, the proposed coordination algorithm needs to incorporate some mechanism able to exploits the inherent collective influence between measures, in order to verify its effectiveness under assumption of uncertainty in individual sensing. To this aim, this study tries to achieve a control on the number of redundant measures of the targets that are sufficient to ensure a sufficient level of reliability. Target detection. For a distributed target, the detection process is the identification of any parts of it, with sufficient detail to permit the intended action. For example: to detect a landmine means to find the location of it to avoid being maimed or killed. To detect radioactive substance means to trace perimeters were radioactivity levels are considered dangerous. To detect gas leak means to identify the area were natural gas seeping from 7 the ground implies fire and explosion hazards, and so on. The search problem is formulated by discretizing the environment into a set of cells. Each target is stationary and usually covers many cells. The objective is to determine in which cells the targets lie. Due to the distribution, the task requires that drones are dynamically arranged so as to be efficiently engaged when some member detects a part of the possible target. For this purpose, the drone releases a particular amount of pheromone on the cell of the sensed possible target, whose diffusion acts as an attractive potential on neighboring drones. To be attracted by pheromone trails, the available drones should be spatially organized into flocks. Flocking is a strategy to allow the self-organization of drones into a number of flocks. Flocking behavior is an emergent effect of individual rules based on alignment, separation and cohesion [24]. With alignment rules the drones tends to move in the same direction that nearby drones. With separation rules, the drone keeps a minimum distance able to provide the drone with flexibility when moving in the swarm, and for a better exploration. Finally, with cohesion rules the drone tends to move towards the swarm. As a result of flocking, each member of a flock has approximately the same heading of the other members, and attempts to remain in range with them. For this purpose, the structural dimensions of the pheromone should take into account the average size of a swarm (or vice versa). Otherwise, a highly diffused or poorly evaporated pheromone could attract disproportionate resources on a single target, thus interfering with the progressive development of the emergent behavior. In contrast, a poorly diffused or highly evaporated pheromone could not be sensed at all. As an effect of pheromone attraction, other drones can confirm the possible target through repeated sensing, and can surround the detected location in order to map the whole distribution. Thus a considerable amount of pheromone is aggregated for each possible target. Once a predefined number of drones confirmed the sensing of the possible target, it 8 is definitively considered to be a true target and then its sensing cannot activate additional pheromone. Since pheromone evaporates over time, after a certain time the pheromone intensity cannot be reinforced in a fully explored region, and in practice disappears. In the presented approach, stigmergy and flocking are two emergent behavioral patterns which should work in conjunction with other basic behavioral patterns of the drone, such as obstacle and boundary avoidance. The process of designing a combination strategy is bottom-up and consists in finding the right setting at the micro-level (agent-level) in order to obtain a coherent emergent behavior at macro-level (swarm-level) [25]. 3 Related Work The goal of this section is to briefly characterize the main approaches and results in the literature on stigmergic mechanisms coordinating swarms of small robots to perform target search or similar tasks. The published works in the field can be distinguished into three categories: using a physical substance as a pheromone, which is necessarily transmitted in an indirect way between robots, by means of the physical environment; using a digital pheromone, transmitted via direct communication between robots; using a digital pheromone, transmitted via an indirect communication between robots. The latter is the category of our approach. Kuyucu et al. in [26] use a swarm of robots releasing physical substance as a repulsive pheromone, for environment exploration. In particular, robots act combining three basic behaviors, with decreasing priority: wall avoiding, pheromone coordination, and random walk. Actually there are various approaches in the literature using physical pheromones, because they do not require a computational structure. Although real pheromones are not usable with aerial vehicle, they can be simulated. Thus, this type of research can be interesting to model new types of digital stigmergy. An example of stigmergic coordination between drones using direct communication is 9 presented by Dasgupta [27], where he focuses on automatic target recognition. Potential target are marked by drones, which also communicate the gossiped pheromone to nearby drones, with probability inversely proportional to the distance from the source. The proposed stigmergic schema employs also repulsive pheromone, as a negative feedback, when a predefined number of drones identify the same target. A disadvantage of such scheme is that the bandwidth required goes into an exponential explosion as the population grows. To avoid redundancy in target evaluation each UAV has to maintain in memory the state of each potential and confirmed target. In this way, the direct communication in the swarm should be strongly limited [28]. A swarm coordination schema with indirect coordination is proposed by Sauter et al. [18]. Here the coordination of a swarm of vehicles is based on digital pheromones maintained in an artificial space called pheromone map and composed by an arbitrary graph of place agents, that is, intermediate control nodes. There are two classes of agents which deposit, withdraw, and read pheromones, that is, walkers and avatars. A walker agent aims to make movements and action decisions, whereas avatars collect location information to make estimates when sensor information is not available. The schema has been applied to a range of scenarios, among which target acquisition. An important problem of this approach is that the exploration depends on the initial position of the swarm. This model does not consider complex targets but only simple targets without structure. To handle the unreliability in sensing, a certain number of drones must be attracted on a potential target. To achieve this goal a spatial organization of the available drones is required in order to sense the pheromone deposit released during a survey leaded by a peer of the same group. This result can be achieved keeping flocking formation. Flocking behavior is exhibited during the group flight. It is an emergent effect caused by the observance of three rules: preserving heading alignment with flock-mates, while 10 maintaining separation with respect to the nearest one and cohesion with the entire group, as described by Reynolds [24]. This flocking behavior formalization have been extensively used in swarming robots and drones coordination. Bouraqadi et al. [29] accomplish an unknown environment survey via a group of robots which has to stay close enough to maintaining the ability to communicate with each other. This objective is reached using Reynolds rules to organize the robots distribution and movements. Hauert et al. [30] apply flocking rules for the management of a drones swarm in order to keep an ad-hoc network during their flight and to coordinate their task. However, this application is based on the assumption of well-known search field, and then it is not applicable to unstructured environments, which is one of our requirements. 4 Behavioral specification of the proposed approach This section aims to characterize the emergent behavior of the coordination algorithm, via the integration of a variety of mechanisms. This purpose is achieved using the Tropos agent-oriented methodology [31]. Tropos is based on the notion of agent, which in this context is a drone, with related notions such as goals and plans. It allows a clear modeling of the operating environment and of the interactions that should occur between drones. Figure 1a shows a legend of the main concepts: actor, goal, plan, resource, capability, and social dependency between actors for goal achievement. Actors may be further specialized based on roles (circle with a bottom line) or agents (circle with upper line). A software agent represents a physical instance (human, hardware or software) of an actor that performs the assigned activities. A role represents a speci c function that, in different circumstances, may be played by the agents. Edges Edges between nodes form dependencies of the form: actor goal/task/resource actor . In additional to hard goals, soft goals are also used when having no clear-cut definition and/or criteria as to whether they are satisfied, for example for modeling goal/plan qualities and non-functional 11 requirements [32]. A detailed account of modeling activities can be found in [31]. Figure 1c represents a top view of the proposed algorithm. More specifically, on the bottom, a Physical Environment is a resource modeling the search field, which contains all the physical elements interacting with drones, whereas a Virtual Environment is a resource managing virtual pheromones and the targets (cells discovered or not). In the middle, Drone is the main actor, supporting the primary goal look for target, collectively attained by two levels of organization: the flock, that is, the organism consisting in locally coordinated drones, and the swarm, that is, the organism consisting in globally coordinated flocks. Conversely, a drone depends on the swarm for saving fly time, since the coordinated search is purposely organized to reduce the overall time. This purpose is based on the resource accomplishment time, managed via update plans of the virtual environment (on the bottom right of the figure): count target found and time unit. Other update plans of the virtual environment are diffuse and evaporate pheromones. The basic needs of a drone consist in the sensing procedures, carried out via both the physical and virtual environments, whereas the basic soft goals of a drone consist of: to cover the search space, collective flight, to point towards targets, and to follow obstacle-free paths. Such soft goals are attained via related roles (in ascending order of priority): obstacle avoider, tracker, flockmate, and explorer [33]. Above all, Fig. 1d shows how a drone reacts to local conditions. Each role is further detailed in Fig. 1b. Fig. 1e represents the obstacle avoider role, with the first priority level. At the second priority level, Fig. 1f represents the tracker role. Fig. 1b shows the third priority level, flockmate. Finally, Fig. 1g represents the minimum priority level, explorer. The above specification is a mixed actor-dependency model in which explained. As a result, the drone task and goals and its precedencies between roles have been detailed. The next Section focuses on the system design, to show how to implement and integrate the main functional and architectural components. 12 (a) (b) (c) (d) (f) Fig. 1. Behavioral representation of the proposed approach (e) (g) 5 Architectural and functional Design of the main subsystems This section is devoted to the modeling of environment and drones. 13 4.1 The design of the environment: pheromones and error dynamics It is assumed that the environment is constrained to a specific area. Without loss of generality, the area is discretized through a grid consisting of C2 cells, each identified by a pair (x,y) of coordinates, with x,y [ C]. The actual size of the area and the number of discretized squares depend on the specific application domain. Figure 2 shows a basic scenario of the pheromone dynamics, focused to the most significant stages of diffusion and evaporation. The levels of pheromone intensity are represented by different grey gradations: the darker the gradation is, the higher the intensity. (a) t = 1 (b) t = 2 (e) t = 20 (c) t = 3 (d) t = 5 (f) t = 25 Fig. 2. Basic scenario of pheromone dynamics: (a) releasing; (b) mainly diffusing; (c-d) diffusing and evaporating; (e-f) mainly evaporating. More specifically, in Fig. 2: (a) a single pheromone intensity I is released; (b) at the first steps, the pheromone is mainly diffusing (moving) to the nearby cells, with a constant diffusion rate [0,1] , StigDiffusion; (c-d) the pheromone is diffusing and evaporating; by evaporating, the pheromone decreases its intensity over time; it is ruled by the constant rate [0,1], StigEvaporation; (e-f) the pheromone is mainly evaporating. More formally, the pheromone intensity p released at the instant t on the cell (x,y) is then characterized by the following dynamics: 14 p x y , ( ) t (1 ) p x y , ( t 1) p x y , ( t 1, ) t d , x y ( t 1, ) t (1) where (1- ) px,y(t-1) represents the amount remaining after diffusion to nearby cells, px,y(t- 1,t) the additional deposits made within the interval (t-1,t], and dx,y(t-1,t) the input pheromone diffused from all the nearby cells. The latter can be formally calculated as: d , x y ( t 1, ) t 1 1 8 i j 1 1 ( , ) (0,0) i j p x i y j , ( t 1) (2) since each of the 8 neighbor cells propagates the portion of its pheromone to the cell (x,y) at each update cycle. The total amount in (1) is also multiplied by (StigEvaporation) to take the evaporation into account. The Environment supports also the management of the target detection with imperfect sensors. It is assumed that each target sensing can provide both false positive and false negative. However, this occurs, with a certain error probability of , only while checking the target cell or the cells adjacent to the target, as represented in Fig. 3. Fig. 3. Imperfect sensor model In essence, it is introduced the notion of degradation of the sensing quality as a function of the proximity to the target: as the proximity increases, the sensing may generate an altered measure resulting in a wrong detection. This assumption implies that the discontinuities represented in Fig.1 should be small with respect to the source signal. 4.2 The design of the drone behavior 15 The drone behavior is structured into a prioritized logic, where each priority level implements one basic behavior, or role. At each update cycle, or tick, the role assumed by the drone is a consequence of the environmental sensing. In descending order or priority, the roles are: obstacle avoider, tracker, flockmate and explorer. Figure 4 shows an overall representation of the drone behavior, using a UML (Unified Modeling Language) activity diagram. Here, every tick period, represented by the hourglass icon on bottom left, the environment updates his status, whereas the drone performs in parallel: the target detection, in which case it releases pheromone controlled by StigDiffusion and StigEvaporationRate parameters; the obstacle avoider. If a close object is detected, within the ObstacleVision radius, the drone points toward a free direction, when available, and moves forward. Otherwise, if there are no close objects detected, the drones play the tracker role: it tries to sense pheromone within the Olfaction radius and, if detected, points toward the pheromone peak. Alternatively, if pheromone is not detected, the drone plays the flockmate role: it tries to detect surrounding drones within the FlockVision radius, in order to point toward the flock. Finally, if there are no surrounding drones, as an explorer it performs a random turn within the WiggleVar angle, and then moves forward. Figure 5 represents a detailed modeling of the main procedures and roles played by a drone. Figure 5a models the basic drone behavior consisting in releasing attractive pheromone with StigIntensity intensity, upon target detection and moving forward according to a given velocity set to DroneVel. In Fig. 5b, Fig. 5c, and Fig. 5d the obstacle avoider, the tracker and the flockmate roles are modelled, respectively. Figure 6 shows the main procedures of behavior, the drone takes into account only drones within a FlockVision. Figure 6a represents the separation behavior: drones close to others have to separate for better [24]. In the flocking 16 exploration; thus, if a drone senses another drone closer than the MinimumSeparation, it turns by an angle MaxSeparateTurn. Figure 6b shows the alignment behavior: the drone calculates the average direction of the drones in the flock vision and turns by an angle MaxAlignTurn to conform its direction to the flock direction. Figure 6c illustrates the cohesion behavior: the drone calculates the barycenter of the drones in the Flock vision and turns by an angle MaxCohereTurn towards the barycenter. Fig. 4. Overall modeling of the drone behavior modularized in roles. 17 (a) Drone behavior (b) Obstacle avoider (c) Tracker (d) Flockmate Fig. 5. Detailed modeling of the main roles played by a drone (a) Separation (b) Alignment (f) Cohesion Fig. 6. Illustration of the procedures of the flocking behavior 6 Experimental studies 18 The proposed model has been implemented on NetLogo1, a leading simulation platform for swarm intelligence. The output of the system is the total time needed to find the 95% of targets. According to Fig. 4, the model requires 12 parameters, to be tuned via three-phases: early analysis, under the assumption of reliable sensing (that is, sensing error probability and sensing redundancy set to 0.1 and 1, respectively); parameter sensitivity analysis on representative scenarios, by evaluating the uncertainty in the output for each parameter; finally, accurate setting on each of the most sensitive parameters, via a bisection method to find the value minimizing the output. main structural and behavioral parameters of the model, with their range and their value set. Table 1 Structural and behavioral parameters. Name DroneVel WiggleVar ObstacleVision FlockVision MinimumSeparation MaxSeparateTurn MaxAlignTurn MaxCohereTurn Olfaction StigIntensity StigDiffusion StigEvaporation SensingError Redundancy Description (unit measure) Drone horizontal speed (m/s) Drone max rand- Drone object sensing distance (m) Flock visibility radius (m) Flock mobility distance (m) Fl Pheromone sensing distance (m) Pheromone release intensity Pheromone diffusion rate (%) Pheromone evaporation rate (%) Sensing error probability (%) Sensing Redundancy Range Set v. (0,15) 1 (0,180) 150 (0, 5) 2 [0, 50] 7 [0,5] 3 (0,180) 30 (0,180) 20 (0,180) 5 1 ) (0, 40K (0, ) [0,1] 0.85 [0,1] 0.05 (0, 100) [0.1,1] (0, {1,3,5} ) The algorithm has been tested on four different scenarios, such as Field, Dumps, Urban and Mines. The Field scenario is made by 5 groups of targets scattered over the area, with about 10 targets per group. There are no obstacles. Figure 7 shows a snapshot with the 19 spatial arrangement of flocks of different forms and sizes, together with four stigmergic formations on corresponding groups of targets. Here, it can be observed that stigmergic formations attracted flocks of drones. Fig. 7. A snapshot of the Field scenario with flocks and stigmergic formations An initial configuration of the Field scenario is shown in Fig. 8a. Here, 80 total drones (represented by triangular forms) are arranged into four dense flocks, placed at the antipodes of the area, whereas the targets are represented by clusters of black dots. The second scenario, called Dumps (Fig. 8b) represents a synthetic reconstruction of woodland with three abusive garbage aggregations, modelled by three groups of targets. Here, 30 targets and 100 trees are represented by gray and black dots, respectively. 80 total drones, arranged into 4 flocks, are initially placed at the antipodes of the area. The third scenario, 20 that is, Urban (Fig. 8c) is characterized by two cluster of 110 total targets placed on two sides of corresponding buildings. Overall, 7 total buildings are located. 40 drones, arranged into four flocks, are placed at the antipodes of the area, with no trees at all. Finally, the Urban Mines (Fig. 8d) scenario is derived from real-world examples of areas near Sarajevo, in Bosnia-Herzegovina, with landmine objects, selected from publicly available maps2. Recently, some authors actually proposed the use of mini-UAVfor detecting landmines [34]. Drones have been initially placed on the boundaries of the area. With respect to the map of the first three scenarios, whose area is 200 square meters, in the last scenario the area is 400 square meters. To carry out the experiments under the requirement of imperfect sensor model, a sensing error probability in the interval [0.1, 1] percent with uniform distribution has been added. Then, the system output has been evaluated by requiring a prefixed number of repeated measures of the targets in the termination criterion, that is, sensing redundancy values 3 and 5. To assess the effectiveness of the proposed approach, the performance of the model has been evaluated on three approaches: tigmergic Stigmergic and Flocking aproach 10 trials have been carried out. It has been determined that the resulting performance indicator samples are well- modeled by a normal distribution, using a graphical normality test. Hence, the 95% confidence intervals have been calculated. Table 2 summarizes, for each scenario, the that the use of stigmergy speeds up the target search process in any scenario. Moreover, results become even better in combination with flocking. It can be remarked that all scenarios have been processed by using a general purpose parameterization determined a 21 with reliable sensor model. Indeed, a parameterization ad initialization adapted to types of scenario might produce better results. Table 2. Characteristics and numerical results (mean confidence interval) of each scenario. Field Dumps Urban Urban Mines # targets # clusters # trees # buildings # drones 50 5 0 0 80 ) y c n a d n u d e r ( R (1) S (1) S+F (1) R (3) S (3) S+F (3) R (5) S (5) S+F (5) m h t i r o g l A 2,604 1,383 1,078 4,161 1,758 1,484 6,173 2,109 1,591 248 126 106 269 151 147 361 246 136 30 3 100 0 80 2,252 1,297 1,009 3,993 1,513 1,289 6,163 2,208 1,823 212 102 141 266 116 135 399 208 233 110 2 0 7 40 2,340 1,748 1,259 3,688 2,089 1,861 4,647 2,488 2,102 40 40 54 28 25 651 229 560 188 487 102 944 286 707 197 594 166 271 1,167 770 280 151 726 55 49 29 55 84 34 51 93 32 (a) (b) 22 (c) (d) Fig. 8. Models of three synthetic and one real-world scenarios: (a) Field; (b) Dumps; (c) Urban; (d) Urban Mines. To better highlight the scalability of our approach against redundancy, Fig. 9a-d shows the completion time for redundancy 1, 3 and 5, for each scenario. Here, it is apparent that Stigmergy introduces a significant improvement of trend over Random Fly, both alone and combined with flocking behavior. (a) (b) (d) 23 (c) Fig. 9. Completion time against redundancy, for each scenario and with different approaches: Random Fly (dotted line), Stigmergy (dashed), and Stigmergy + Flocking (solid). 7 Conclusions and Future Work In this paper, it is presented a novel swarm approach for coordinating mini-UAVs to perform target search with imperfect sensors. The approach uses a stigmergic behavior, consisting in the release of information in the environment in the form of attractive digital pheromones, in areas where potential targets are sensed. Moreover, the approach employs flocking behavior, resulting in a flexible arrangement of drones according to the stigmergic potential field. The paper illustrates the approach from the behavioral and architectural point of views, and then discusses the experimental studies. Results on synthetic and real- world scenarios prove the benefits of both stigmergy and flocking, in terms of tolerance to errors and scalability for increasing redundancy requirements. The overall mechanism can be better enabled if structural parameters are correctly tuned for the given scenario. Determining such correct parameters is not a simple task since different areas have different features. Thus, an appropriate tuning to adapt parameters to the specific search area is desirable to make the search more effective. For this purpose, to use a parameter optimization strategy is considered a key investigation activity for future work. Acknowledgements This work was carried out in the framework of the SCIADRO project, cofunded by the Tuscany Region (Italy) under the Regional Implementation Programme for Underutilized Areas Fund (PAR FAS 2007-2013) and the Research Facilitation Fund (FAR) of the Ministry of Education, University and Research (MIUR) References 24 1. K. Whitehead, C. H. Hugenholtz, S. Myshak, O. Brown, A. LeClair, A. Tamminga, T. E. Barchyn, B. Moorman, B. Eaton, Remote sensing of the environment with small unmanned aircraft systems (UASs), part 2: scientific and commercial applications, Journal of Unmanned Vehicle Systems, 02 (2014), 86-102. 2. A. Errico, C. V. Angelino, L. Cicala, G. Persechino, C. Ferrara, M. Lega, A. Vallario, C. Parente, G. Masi, R. Gaetano, G. Scarpa, D. Amitrano, G. Ruello, L. Verdoliva, G. Poggi, Detection of environmental hazards through the feature-based fusion of optical and SAR data: a case study in southern Italy, International Journal of Remote Sensing, 36 (2015), 3345-3367. 3. A. Errico, C. V. Angelino, L. Cicala, D. P. Podobinski, G. Persechino, C. Ferrara, M. Lega, A. Vallario, C. Parente, G. Masi, R. Gaetano, G. Scarpa, D. Amitrano, G. Ruello, L. Verdoliva, G. Poggi, SAR/multispectral image fusion for the detection of environmental hazards with a GIS, Proceedings of SPIE - The International Society for Optical Engineering, 9245 (2014), doi: 10.1117/12.2066476. 4. M. Lega, R. M. A. Napoli, A new approach to solid waste landfills aerial monitoring, WIT Transactions on Ecology and the Environment, 109 (2008), 193-199. 5. M. Lega, R. M. A. Napoli, Aerial infrared thermography in the surface waters contamination monitoring, Desalination and Water Treatment, 23 (2010), 141-151. 6. G. Persechino, P. Schiano, M. Lega, R. M. A. Napoli, C. Ferrara, J. Kosmatka, Aerospace-based support systems and interoperability: The solution to fight illegal dumping, WIT Transactions on Ecology and the Environment, 140 (2010), 203-214. 25 7. G. Persechino, M. Lega, G. Romano, F. Gargiulo, L. Cicala, IDES project: An advanced tool to investigate illegal dumping, WIT Transactions on Ecology and the Environment, 173 (2013), 603-614. 8. M. Lega, C. Ferrara, G. Persechino, P. Bishop, Remote sensing in environmental police investigations: Aerial platforms and an innovative application of thermography to detect several illegal activities, Environmental Monitoring and Assessment, 186 (2014), 8291-8301. 9. M. Lega, L. D'Antonio, R. M. A. Napoli, Cultural Heritage and Waste Heritage: Advanced techniques to preserve cultural heritage, exploring just in time the ruins produced by disasters and natural calamities, WIT Transactions on Ecology and the Environment, 140 (2010), 123-134. 10. M. Lega, G. Persechino, GIS and infrared aerial view: Advanced tools for the early detection of environmental violations, WIT Transactions on Ecology and the Environment, 180 (2014), 225-235. 11. F. Gargiulo, G. Persechino, M. Lega, A. Errico, IDES project: A new effective tool for safety and security in the environment, Lecture Notes in Computer Science (LNCS), 8286 (2013), 201-208. 12. M. Lega, D. Ceglie, G. Persechino, C. Ferrara, R. M. A. Napoli, Illegal dumping investigation: A new challenge for forensic environmental engineering, WIT Transactions on Ecology and the Environment, 163 (2012), 3-11. 13. M. Lega, J. Kosmatka, C. Ferrara, F. Russo, R. M. A. Napoli, G. Persechino, Using Advanced Aerial Platforms and Infrared Thermography to Track Environmental Contamination, Environmental Forensics, 13 (2012), 332-338. 26 14. L. F. Bertuccelli, J. P. How, Robust UAV search for environments with imprecise probability maps, Proceedings of the 44th IEEE Conference on Decision and Control, (2005), 5680-5685. 15. B. Bethke, M. Valenti, J. How, Cooperative vision based estimation and tracking using multiple UAVs, In Advances in Cooperative Control and Optimization, Springer, Berlin Heidelberg, (2007), 179-189. 16. R. McCune, R. Purta, M. Dobski, A. Jaworski, G. Madey, Y. Wei, A. Madey and M.B. Blake, Investigations of DDDAS for command and control of uav swarms with agent- based modeling, Proceedings of the 2013 Winter Simulation Conference: Simulation: Making Decisions in a Complex World, (2013), 1467-1478. 17. M. Brambilla, E. Ferrante, M. Birattari, M. Dorigo, Swarm robotics: a review from the swarm engineering perspective, Swarm Intelligence, 7 (2013), 1-41. 18. J. A. Sauter, R. Matthews, H. Van Dyke Parunak, S. A. Brueckner, Performance of digital pheromones for swarming vehicle control, Proceedings of the fourth international joint conference on Autonomous agents and multiagent systems, (2005), 903-910. 19. M.G.C.A. Cimino, A. Lazzeri, G. Vaglini, Combining stigmergic and flocking behaviors to coordinate swarms of drones performing target search, Proceedings of the Sixth International Conference on Information, Intelligence, Systems and Applications (2015), 1-6. 20. H. Van Dyke Parunak, M. Purcell, R. O'Connell, Digital pheromones for autonomous coordination of swarming UAV's, Proceedings of the 1st Technical Conference and Workshop on Unmanned Aerospace Vehicles, (2002), 48105-1579. 21. G. Ermacora, A. Toma, B. Bona, M. Chiaberge, M. Silvagni, A cloud robotics architecture for an emergency management and monitoring service in a smart city 27 environment, Proceedings of the International Conference of Intelligent Robots and Systems, (2013). 22. S. Reece, S. Roberts, C. Claxton, D. S. Nicholson, Multi-sensor fault recovery in the presence of known and unknown fault types, Proceedings of the 12th International Conference on Information Fusion, (2009). 23. A. F. T. Winfield, C. J. Harper and J. Nembrini, Towards the Application of Swarm Intelligence in Safety Critical Systems, Proceedings of the International Conference on System Safety, (2006), 89-95. 24. C. W. Reynolds, Flocks, herds and schools: A distributed behavioral model, Proceedings of the ACM Siggraph Conference on Computer Graphics, (1987), 25-34 . 25. C. Bernon, M. P. Gleizes, G. Picard, Enhancing self-organising emergent systems design with simulation, Engineering Societies in the Agents World VII, Springer Berlin Heidelberg, (2007), 284-299. 26. T. Kuyucu, I. Tanev, K. Shimohara, Superadditive effect of multi-robot coordination in the exploration of unknown environments via stigmergy, Neurocomputing, 148 (2015), 83-90. 27. P. Dasgupta, A multiagent swarming system for distributed automatic target recognition using unmanned aerial vehicles, IEEE Transactions on Systems, Man, and Cybernetics Part A: Systems and Humans, 38 (2008) 549-563. 28. Y. Tan, Z. Zhong-Yang, Research advance in swarm robotics, Defence Technology 9 (2013), 18-39. 29. N. Bouraqadi, A. Doniec, E. M. de Douai. Flocking-based multi-robot exploration. Proceedings of National conference on control architectures of robots, (2009). 30. S. Hauert, S. Leven, M. Varga, F. Ruini, A. Cangelosi, J.-C. Zufferey, D. Floreano, Reynolds flocking in reality with fixed-wing robots: communication range vs. 28 maximum turning rate, Proceedings of the International Conference on Intelligent Robots and Systems, (2011). 31. P. Bresciani, A. Perini, P. Giorgini, F. Giunchiglia, J. Mylopoulos, Tropos: An agent- oriented software development methodology, Autonomous Agents and Multi-Agent Systems, 8 (2004) , 203 236. 32. F. Pagliarecci, L. Penserini, L. Spalazzi, From a Goal-Oriented methodology to a BDI agent language: the case of Tropos and Alan, Proceedings of Workshop on Agents, Web Services and Ontologies Merging, (2007). 33. P. Giorgini, J. Mylopoulos, A. Perini, A. Susi, The Tropos Metamodel and its Use, Informatica, 29 (2005), 401.408. 34. J. Rodriguez, C. Castiblanco, I. Mondragon, J. Colorado, Low-cost quadrotor applied for visual detection of landmine-like objects, Proceedings of the International Conference on Unmanned Aircraft Systems, (2014).
1310.3781	2	1310	2019-07-15T22:16:25	An Agent-based Model of the Cognitive Mechanisms Underlying the Origins of Creative Cultural Evolution	[ "cs.MA", "cs.AI" ]	Human culture is uniquely cumulative and open-ended. Using a computational model of cultural evolution in which neural network based agents evolve ideas for actions through invention and imitation, we tested the hypothesis that this is due to the capacity for recursive recall. We compared runs in which agents were limited to single-step actions to runs in which they used recursive recall to chain simple actions into complex ones. Chaining resulted in higher cultural diversity, open-ended generation of novelty, and no ceiling on the mean fitness of actions. Both chaining and no-chaining runs exhibited convergence on optimal actions, but without chaining this set was static while with chaining it was ever-changing. Chaining increased the ability to capitalize on the capacity for learning. These findings show that the recursive recall hypothesis provides a computationally plausible explanation of why humans alone have evolved the cultural means to transform this planet.	cs.MA	cs	Gabora, L., & Saberi, M. (2011). An agent-based model of the cognitive mechanisms underlying the origins of creative cultural evolution. In A. Goel, F. Harrell, B. Magerko, & J. Prophet (Eds.), Proceedings of the 8th ACM Conference on Cognition & Creativity (pp. 299-306). New York: Association for Computing Machinery (ACM). An Agent-based Model of the Cognitive Mechanisms Underlying the Origins of Creative Cultural Evolution Liane Gabora Department of Psychology University of British Columbia Okanagan campus, 3333 University Way Kelowna BC, V1V 1V7, CANADA 250-807-9849 [email protected] Maryam Saberi School of Interactive Art and Technology Simon Fraser University Surrey BC, V3T 0A3, CANADA 778-868-5159 [email protected] ABSTRACT Human culture is uniquely cumulative and open-ended. Using a computational model of cultural evolution in which neural network based agents evolve ideas for actions through invention and imitation, we tested the hypothesis that this is due to the capacity for recursive recall. We compared runs in which agents were limited to single-step actions to runs in which they used recursive recall to chain simple actions into complex ones. Chaining resulted in higher cultural diversity, open-ended generation of novelty, and no ceiling on the mean fitness of actions. Both chaining and no-chaining runs exhibited convergence on optimal actions, but without chaining this set was static while with chaining it was ever-changing. Chaining increased the ability to capitalize on the capacity for learning. These findings show that the recursive recall hypothesis provides a computationally plausible explanation of why humans alone have evolved the cultural means to transform this planet. General Terms Algorithms, Experimentation, Theory, Verification ACM Classification Keywords Primary Classification: J.4 SOCIAL AND BEHAVIORAL SCIENCES: Subjects: Psychology and Sociology; Additional Classification: J.5 ARTS AND HUMANITIES I.2 ARTIFICIAL INTELLIGENCE I.2.6 Learning; I.2.11 Distributed Artificial Intelligence Author Keywords EVOC; Homo erectus; action; agent-based model; creativity; cultural diversity; cultural evolution; gesture; innovation; invention; mimetic; origin of culture; recursive recall, self-triggered recall and rehearsal loop INTRODUCTION How did humans come to be so creative? No other species has transformed this planet to anything near the degree that our species has. We now have the technology at our disposal to design computer models that enable us to investigate how our remarkable creativity came about. Humans are not only creative; we put our own spin on the inventions of others, such that new inventions build cumulatively on previous ones. This cumulative cultural change is referred to as the ratchet effect (Tomasello, Kruger, & Ratner, 1993), and it has been suggested that it is uniquely human (Donald, 1998). This paper uses an agent-based simulation to address the question of what enabled humans to be creative enough that cumulative, open-ended cultural evolution could take hold and flourish. Our approach is motivated by the view that there is a distinct difference between (1) culture as a system involving transmission of chance innovations, which is observed in several species, and (2) culture as a process involving the accumulation and differentiation of strategic innovations, which is unique to our species. Only (2) constitutes a genuine evolutionary process, for to achieve descent with adaptive modification it is necessary that change be cumulative. A POSSIBLE MECHANISM FOR THE ORIGIN OF CULTURAL EVOLUTION The minds of our earliest ancestors, Homo habilis, have been referred to as episodic because there is no evidence that their experience deviated from the present moment of concrete sensory perceptions (Donald, 1991). They were able to encode perceptions of events in memory, and recall them in the presence of a reminder or cue, but had little voluntary access to memories without environmental cues. They were therefore unable to voluntarily shape, modify, or practice skills and actions, and unable to invent or refine complex gestures or means of communicating. Homo erectus lived between approximately 1.8 and 0.3 million years ago. The size of the Homo erectus brain was approximately 1,000 cc, about 25% larger than that of Homo habilis, at least twice as large as the brains of living great apes, and 75% the cranial capacity of modern humans 1 of that indications exhibited many (Aiello, 1996; Ruff et al., 1997). This period is widely referred to as the beginnings of cumulative culture. Homo erectus enhanced intelligence, creativity, and ability to adapt to their environment, including sophisticated, task-specific stone handaxes, complex stable seasonal home bases, and long- distance hunting strategies involving large game, and migration out of Africa. This period marks the onset of the archaeological record and it is thought to be the beginnings of human culture. It is widely believed that this cultural transition reflects an underlying transition in cognitive or social abilities. Some have suggested that they owe their achievements to onset of theory of mind (Mithen, 1998) or the capacity to imitate (Dugatkin, 2001). However, there is evidence that other species possess theory of mind and the capacity to imitate (Heyes, 1998), yet do not compare to modern humans in intelligence and cultural complexity. Evolutionary psychologists have suggested the intelligence and cultural complexity of the Homo line can be attributed to the onset of massive modularity (Buss, 1999, 2004; Barkow, Cosmides, &Tooby, 1992). However, although the mind exhibits an intermediate degree of functional and anatomical modularity, neuroscience has not revealed vast numbers of hardwired, encapsulated, task- specific modules; indeed, the brain has been shown to be more highly subject to environmental influence than was previously believed (Buller, 2005; Byrne, 2000; Wexler, 2006). Donald (1991) proposed that with the enlarged cranial capacity of Homo erectus, the human mind underwent the first of three transitions by which it -- along with the cultural matrix in which it is embedded -- evolved from the ancestral, pre-human is characterized by a shift from an episodic to a mimetic mode of cognitive functioning, made possible by onset of the capacity for voluntary retrieval of stored memories, independent of environmental cues. Donald refers to this as a self-triggered recall and rehearsal loop. Self-triggered recall enabled information to be processed recursively, and reprocessed with to different contexts or perspectives. Self-triggered recall allowed our ancestor to access memories voluntarily and thereby act out1 events that occurred in the past or that might occur in the future. Thus not only could the mimetic mind temporarily escape the here and now, but by miming or gesture, it could communicate similar escapes in other minds. The capacity to mime thus ushered forth what is referred to as a mimetic form of cognition and brought about a transition to the mimetic stage of human culture. The self-triggered recall and rehearsal loop also enabled our ancestors to engage in a stream of thought. One thought or idea evokes another, revised version of it, which evokes yet another, and so forth condition. This transition respect 1 The term mimetic is derived from "mime," which means "to act out." recursively. In this way, attention is directed away from the external world toward one's internal model of it. Finally, self-triggered recall allowed for voluntary rehearsal and refinement of actions, enabling systematic evaluation and improvement of skills and motor acts. The recursive recall hypothesis is difficult to test directly, for if correct it would leave no detectable trace. It is, however, possible to computationally model how the onset of the capacity for recursive recall would affect the effectiveness, diversity, and open-endedness of ideas generated in an artificial society. THE MODEL We tested Donald's hypothesis using an agent-based computational model of culture referred to as 'EVOlution of Culture', abbreviated EVOC (Gabora, 2008b, 2008c; Gabora & Leijnen, 2009; Leijnen & Gabora, 2009). EVOC has been used to address such questions as (1) how does the presence of leaders or barriers to the diffusion of ideas affect cultural evolution dynamics, (2) what is the effect of the ratio of creators to imitators in a society, and (3) what is the impact of locally clustering the creators. It is an elaboration of Meme and Variations, or MAV (Gabora, 1994, 1995), the earliest computer program to model culture as an evolutionary process in its own right (as opposed to modeling the interplay of cultural and biological evolution).The approach was inspired by Holland's (1975) genetic algorithm, or GA. The GA is 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 satisfactory solution is found. The goal here was to distil the underlying logic of not biological evolution but cultural evolution, i.e. the process by which ideas adapt and build on one another in the minds of interacting individuals. In cultural evolution, the generation of novelty takes place through invention instead of through mutation and recombination as in biological evolution, and the differential replication of novelty takes place through imitation, instead of through reproduction with inheritance as in biological evolution (Gabora, 1996). EVOC (as did MAV) uses neural network based agents that could (1) invent new ideas by modifying previously learned ones, (2) evaluate ideas, (3) implement ideas as actions, and (4) imitate ideas implemented by neighbors. 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 actions. EVOC (like MAV) successfully models how 'descent with modification' occurs in a cultural context. The approach can thus be contrasted with computer models of how learning affects biological evolution (Best, 1999, 2006; Higgs, 2000; Hinton & Nowlan, 1987; Hutchins & Hazelhurst, 1991). EVOC consists of an artificial society of neural network based agents in a two-dimensional grid-cell world. It is written in Joone, an object oriented programming environment, using an open source neural network library individual 2 written in Java. We now summarize the key components of the agents and the world they inhabit. THE AGENTS Agents consist of (1) a neural network, which encodes ideas for actions and detects trends in what constitutes a fit action, and (2) a body, which implements actions. The core of an agent is an auto associative neural network, as shown in Figure 1. The network learns ideas for actions. An idea is a pattern consisting of six elements that dictate the placement of the six body parts. Learning and training of the neural network is as per Gabora (1995). Figure 1. The core of an agent is an autoassociative neural network composed of six input nodes and six corresponding output 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, MOVEMENT and OPPOSITE). Input nodes and output nodes are connected to hidden nodes of which they are instances (e.g. RIGHT ARM is connected to RIGHT.) Activation of any input node activates the MOVEMENT hidden node. Same-direction activation of symmetrical input nodes (e.g. positive activation -- which represents upward motion -- of both arms) activates the SYMMETRY node. Sequential opposite movement of an ARM activates the OPPOSITE node. This hidden node was newly added to this version of the program to implement the chaining of actions to form multi-step actions. The hidden nodes are used to bias invention using learned trends about what constitutes a fit action. In EVOC, the neural network can also be turned off to compare results with a data structure that cannot detect trends, and thus invents ideas merely at random. If the fitness of an action is evaluated to be higher than that of any action learned thus far, it is copied from the input/output nodes of the neural network that represent concepts of body parts to a six digit array that contains representations of actual body parts, referred to as the body. Since it is useful to know how many agents are doing essentially the same thing, when node activations are translated into limb movement they are threshold such that there are only three possibilities for each limb: stationary, up, or down. Six limbs with three possible positions each gives a total of 729 possible actions. Only the action that is currently implemented by an agent's body can be observed and imitated by other agents. THE ARTIFICIAL WORLD MAV allowed only worlds that were square and toroidal, or 'wrap-around' (such that agents at the left border that attempt to move further left appear on the right border). In EVOC the world can assume any shape, and be as sparsely or densely populated as required, with agents placed in any configuration. EVOC also allows for the creation of complete or semi-permeable permanent or eroding borders that decrease the probability of imitation along a frontier (although this was not used in the experiments reported here). INCORPORATION OF COGNITIVEPHENOMENA The neural network can be turned off to compare neural network results with those obtained using a database that cannot learn trends, and thus invents ideas at random. The following cultural evolution parameters can also be turned off or on (in some cases to varying degrees). Imitation Fit actions diffuse through the society when agents copy neighbors' actions. During imitation, the input is the action implemented by a neighbor. The process of finding a neighbor to imitate works through a form of lazy (non- greedy) search. An imitating agent randomly scans its neighbors, and adopts the first action that is fitter than the action it is currently implementing. If it does not find a neighbor that is executing a fitter action than its own current action, it continues to execute the current action. Invention Agents generate new actions by modifying their initial action or an action that has been invented previously or acquired through imitation. 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, MOVEMENT, and OPPOSITE hidden nodes. (Were this not the case there would be no benefit to using a neural network.) The extent to which the generation of novelty is biased by past experience is referred to as the learning rate. 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 probabilistic decision as to whether the position of that body part will change, and if it does, the direction of change is stochastically biased according to the learning rate. If the new the currently implemented idea, the agent learns and implements the action specified by that idea. (See Gabora, 1995 for further details.) idea has a higher fitness than 3 Learning Invention makes use of the ability to detect, learn, and respond adaptively to trends. Since a new action (or, in invention, a new idea for an action) is not learned unless it is fitter than the currently implemented action, new actions provide valuable information about what constitutes an effective idea. Thus knowledge acquired through the evaluation of previous actions is translated into educated guesses about what constitutes a successful action, and this is accomplished by updating the learning rate. There are two learnable trends. The first concerns the overall level of activity involved in successful actions. Learning this trend can bias the generation of new ideas in favor of either more movement or less. Each body part starts out at a stationary rest position, and with an equal probability of changing to movement in one direction or the other. If the fitter action codes for more movement, increase the probability of movement of each body part. Do the opposite if the fitter action codes for less movement. If movement is generally beneficial, increase the probability that new actions involve movement of more body parts. This trend is based on the assumption that movement in general (regardless of which particular body part is moving) can be beneficial or detrimental. This seems like a useful generalization since movement of any body part uses energy and increases the likelihood of being detected. It is implemented as follows: am1 = movement node activation for current action am2 = movement node activation for new action p(im)I = probability of increased movement at body part i p(dm)I = probability of decreased movement at body part i IF (am2>am1) THEN p(im)I = MAX(1.0, p(im)I + 0.1) ELSE IF (am2<am1) THEN p(im)I = MIN(0.0, p(im)I - 0.1) p(dm)I = 1 - p(im)i The second learnable trend is: if successful actions tend to be symmetrical (e.g. left arm moves to the right and right arm moves to the left), the probability increases that new actions are symmetrical. (See Gabora 1995 for details.) This generalization is biologically sensible, since many useful actions (e.g. walking) entail movement of limbs in opposite directions, while others (e.g. pushing) entail movement of limbs in the same direction. However, the reason for implementing a preference for symmetrical limb coordination is not to produce a biologically realistic model of motor control but to provide a systematic rationale for associating different actions with a graded distribution of fitness values. The preference for symmetrical limb coordination is implemented in a manner analogous to that of the first rule. Of particular relevance to the current investigation, if multi-step actions involving one limb moving sequentially one direction and then the other are beneficial, the probability of inventing such actions increases. In summary, each action is associated with a measure of its effectiveness, and generalizations about what seems to work and what does not are translated into guidelines for what action to take. Evaluation Before committing to implementing an idea as an action, agents can assess how successful the action would be if it were implemented. They evaluate the effectiveness of their actions according to how well they satisfy needs using a pre-defined equation that rewards actions that make use of trends detected by the SYMMETRY, MOVEMENT, and OPPOSITE hidden nodes. A successful action is one in which all body parts except the head are moving, and limb movement is asymmetrical. (Thus if the left arm is moving up, the right arm is moving down, and vice versa.) For the experiments reported here with chaining turned off, where total body movement, m, is calculated by adding the number of active body parts, and mh is head movement, the fitness, Fnc, is calculated as follows: sa = 1 if arms are moving symmetrically; 0 otherwise sl = 1 if legs are moving symmetrically; 0 otherwise mh = 1 if head is stationary; 0 otherwise Fnc = m+ 1.5(sa+ sl) + 2(1 -- mh) This gives a fitness function with a maximum value of ten, because while movement of the head contributes positively to the first term (which represents overall movement) it contributes negatively to the third term (which represents head stability). This corresponds to a relatively realistic action with respect to the need to attract the attention of, and maintain eye contact with others including a potential mate. Since there are two optimal movements for the hips, one optimal movement for the head (or in this case, a lack of movement), two optimal arm combinations, and two optimal leg combinations, there are eight different optimal actions. In previous versions of the program that did not allow for chained actions, once an agent converged on one of the eight optimal actions (out of the possible 729 actions), it implemented that action for the rest of the run. Therefore, once all agents had found one or another of the optimal actions, the mean fitness of actions across the society plateaus, and the set of implemented actions became fixed. Chaining Chaining, which is new to this version of the program, gives agents the opportunity to execute multi-step actions. For the experiments reported here with chaining turned on, if in the first step of an action an agent was moving at least one of its arms, it executes a second step, which again involves up to six body parts. If, in the first step, the agent moved one arm in one direction, and in the second step it 4 moved the same arm in the opposite direction, it has the opportunity to execute a three-step action. And so on. The agent is allowed to execute an arbitrarily long action so long as it continues to move the same arm in the opposite direction to the direction it moved previously. Once it does not do so, the chained action comes to an end. The longer it moves, the higher the fitness of this multi-step chained action. Where n is the number of chained actions, the fitness, Fc, is calculated as follows: implemented Fc = Fnc+ (n -- 1) The fitness function with chaining provides a simple means of simulating the capacity for recursive recall. We omit a detailed explanation of EVOC options that were not used in the experiments reported here, such as broadcasting, which allows the action of a leader to be visible to not just immediate neighbors, but all agents, thereby simulating the effects of media (Gabora, 2008a,b; Leijnen & Gabora, 2010). 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 the knowledge-based operators, and (3) implement a new action. To invent a new idea, for each node of the idea currently represented on the input/output layer of the neural network, the agent makes a probabilistic decision as to whether change will take place, and if it does, the direction of change is stochastically biased by the knowledge-based operators. If the new idea has a higher fitness than the currently learns and implements the action specified by that idea. To acquire an idea through imitation, an agent randomly chooses one of its neighbors, and evaluates the fitness of the action the neighbor is implementing. If its own action is fitter than that of the neighbor, it chooses another neighbor, until it has either observed all of its immediate neighbors, or found one with a fitter action. If no fitter action is found, the agent does nothing. Otherwise, the neighbor's action is copied to the input layer, learned, and implemented. Fitness of actions starts out low because initially all agents are immobile. Soon some agent invents an action that has a higher fitness than doing nothing, 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 initially increases due to the proliferation of new ideas, and then decreases as agents hone in on the fittest actions. THE GRAPHIC USER INTERFACE The graphic user interface (GUI) makes use of the open source charting project, JFreeChart, enabling variables to be user defined at run time, and results to become visible as the program runs. The topmost output panel, shown in Figure 2, is provided to give the reader a sense of how the program works. idea, the agent EXPERIMENTS In previous versions of EVOC, that did not allow chaining, there were only a fixed number of possible actions, and once all agents converged on optimal actions, cultural change ground to a halt. We hypothesized that if agents were able to chain actions together to form arbitrarily long actions, fitness would not stop increasing, and the agents would not converge on the same fixed set of actions. For these experiments the artificial world was a toroidal lattice with 100 nodes, each occupied by a single, stationary agent. We used a von Neumann neighborhood structure, (agents only interacted with their four adjacent neighbors). All graphs plot averages of ten runs, with an invention to imitation ratio of 1:1, and a rate of conceptual change of 1/6. this means approximately one body part movement per step of an action). All agents were stationary at the start of a run. there are six body parts (Since Figure 2. Output panel of GUI. At the upper left one specifies the Invention to Imitation Ratio. This is the probability that a given agent, on a given iteration, invents a new idea for an action, versus the probability that it imitates a neighbor's action. Below it is Rate of Conceptual Change, where one specifies the degree to which a newly invented idea differs from the one it was based on. Below that is Number of Agents, which allows the user to specify the size of the artificial society. Below that is where one specifies Number of Iterations, i.e. the duration of a run. Agents can be accessed individually by clicking the appropriate cell in the grid on the upper right. This enables one to see such details as the action currently implemented by that agent, or the fitness of that action. The graphs at the bottom plot the mean idea fitness and diversity of ideas. Tabs shown at the top give access to other output panels. Effect of Chaining on Fitness of Actions Figure 3 shows the results of comparing how the capacity to recursively chain actions together affects the mean fitness of actions across the artificial society over the duration of a run. 5 Figure 3. Mean fitness of actions in the artificial society with chaining versus without chaining. The capacity to chain together simple actions to form more complex ones increases the mean fitness of actions across the artificial society. This is most evident in the later phase of a run. Without chaining, agents converge on optimal actions, and the mean fitness of action reaches a plateau. With chaining, however, there is no ceiling on the mean fitness of actions. By the 100th iteration it reached almost 15, indicating a high incidence of chaining. Effect of Chaining on Diversity of Actions Figure 3 shows the results of comparing how the capacity to recursively chain actions together affects the mean fitness of actions across the artificial society over the duration of a run. Figure 4. Mean number of different actions in the artificial society with chaining (continuous line) versus without chaining (dashed line). Chaining also increases the diversity of actions. This is most evident in the early phase of a run before agents begin to converge on optimal actions, though it remains the case throughout the duration of a run. Interestingly, although in both cases there is convergence on optimal actions, without chained actions, this is a static set (thus mean fitness plateaus) whereas with chained actions the set of optimal actions is always changing, as increasingly fit actions are found (thus mean fitness keeps increasing). The Effect of Chaining on Learning Recall that agents have the capacity to learn trends from past experiences (using the knowledge-based operators), and thereby bias the generation of novelty in directions that have a greater than chance probability of being fruitful. Results obtained previously with this kind of computer model of cultural evolution showed that learning increases the mean fitness of actions throughout the duration of a run (Gabora, provides more opportunities to capitalize on the capacity to learn, we hypothesized that chaining would accentuate the impact of learning on the mean fitness of actions across the artificial society. Figure 5 compares the effect of chaining with knowledge-based operators turned on (agents are able to learn) versus when they are turned off (agents are unable to learn). chaining 1995). Since Figure 5. Mean fitness of actions in the artificial society with both chained actions and learning (top), with just chained actions (second from the top) with just learning (second from the bottom), and with neither chained actions nor learning (bottom). This hypothesis was confirmed. Since waving an arm back and forth increased activation of the OPPOSITE hidden node, agents learned to generate increasingly longer moves throughout the duration of a run. When agents are given the capacity to chain together simple actions to form more complex ones, the increase in mean fitness of actions throughout the duration of a run is more pronounced, indicating that the society as a whole benefits much more from learning. Discussion Many species possess what could be called culture, but human culture is unique in that ideas build on each other cumulatively; indeed culture is widely thought to be an evolutionary process (e.g. Bentley, Ormerod, & Batty, 2011; Cavalli-Sforza & Feldman, 1981; Gabora, 1996, 2008; Hartley, 2009;Mesoudi, Whiten & Laland, 2004, 2006; Whiten, Hinde, Laland, & Stringer, 2011). Using an agent based computer model of cultural evolution, we obtained support for the hypothesis that the onset of cumulative, open-ended cultural evolution can be attributed to the evolution of a self-triggered recall and rehearsal loop, enabling the recursive chaining of thoughts and actions. 6 the open-endedness of Our results suggest that it is the capacity for recursive recall that makes possible some computational models of language evolution (e.g. Kirby, 2001). Note that in the chaining versus no chaining conditions the size of the neural network is the same, but how it is used differs. This suggests that it was not larger brain size per se that initiated the onset of cumulative culture, but that larger brain size enabled episodes to be encoded in more detail, allowing more routes for reminding and recall, thereby facilitating the ability to recursively re- describe information stored in memory (Karmiloff-Smith, 1992), thereby tailor it to the situation at hand. By demonstrating that onset of the capacity for recursive recall constitutes a viable explanation for how cultural change became cumulative and open-ended, we believe that this research is a pivotal step forward in the development of a scientific theory of cultural evolution. Nevertheless, the model is simplified, and caution must be taken applying such results to the real world. The agents can be said to have the goal of optimizing the fitness of their actions, but they do not have values, norms, or social bonds remotely like ours. The goal behind a computational model such as EVOC, however, is to extract the skeleton of a process as complex as cultural evolution and investigate questions that are difficult to investigate by other means. To our knowledge, this is the only experimental test of, and tentative support for, the hypothesis that the capacity for self-triggered recall of thoughts and recursive execution of actions played a key role in ratcheting the kind of cumulative, open-ended novelty and cultural complexity that surrounds us. Despite that evolved cultural complexity essentially defines our humanness, we know of no other computational modeling support for this or any other competing hypothesis concerning how the transition from episodic to mimetic cognition came about. Future work will focus on simulating more complex forms of recursive recall that would require a larger neural network, and corroborating findings obtained with the model with archaeological and anthropological evidence. Another direction of future research involves using EVOC to investigate competing hypotheses concerning what cognitive mechanism could have given rise to the explosive cultural transition in the Middle-Upper Paleolithic (Klein, 1999; Leaky, 1984). ACKNOWLEDGMENTS We would like to acknowledge grants to the first author from the Social Sciences and Humanities Research Council of Canada and the Concerted Research Program of the Fund for Scientific Research, Belgium. REFERENCES 1. Aiello, L. C. Hominine pre-adaptations for language and cognition. In P. Mellars & K. Gibson (Eds.), Modeling the early human mind, Cambridge, UK: McDonald Institute Monographs (1996), 89 -- 99. 2. Barkow, J. H., Cosmides, L., & Tooby, J., Eds. The Adapted Mind: Evolutionary Psychology and the Generation of Culture. New York: Oxford University Press, 1992. 3. Bentley, R. A., Ormerod, P., & Batty, M. Evolving social influence in large populations. Behavioral Ecology and Sociobiology, 65 (2011), 537 -- 546. 4. Best, M. How culture can guide evolution: An inquiry into gene/meme enhancement and opposition. Adaptive Behavior, 7, 3 (1999), 289 -- 293. 5. Best, M. Adaptive value within natural language discourse. Interaction Studies, 7, 1(2006), 1-15. 6. Buller, D. J. Adapting minds. Cambridge, MA: MIT Press, 2005. 7. Buss, D. M. Evolutionary Psychology: The New Science of the Mind. Boston, MA: Pearson, 1999/2004. 8. Byrne, R. W., & Russon, A. E. Learning by imitation: A hierarchical approach. Behavioral and Brain Sciences, 21(1998), 667 -- 721. 9. Cavalli-Sforza, L. L., & Feldman, M. W. Cultural transmission and evolution: A quantitative approach. Princeton: Princeton University Press, 1981. 10. Cloak, F. T. Jr. Is a cultural ethology possible? Human Ecology, 3(1975), 161-182. 11. Donald, M. Hominid enculturation and cognitive evolution. In C. Renfrew & C. Scarre (Eds.), Cognition and Material Culture: The Archaeology of Symbolic Storage. McDonald Institute Monographs (1998), 7 -- 17. 12. Donald, M. Origins of the Modern Mind: Three Stages in the Evolution of Culture and Cognition. Cambridge, MA: Harvard University Press, 1991. 13. Dugatkin, L. A. Imitation Factor: Imitation in Animals and the Origin of Human Culture. New York: Free Press, 2001. 14. Gabora, L. A computer model of the evolution of culture. In R. Brooks & P. Maes (Eds.) Proceedings of the 4th International Conference on Artificial Life, July 4-6,Boston, MA, 1994. 15. Gabora, L. Meme and Variations: A computer model of cultural evolution. In L. Nadel & D. Stein (Eds.) 1993 Lectures in Complex Systems. Reading, MA: Addison- Wesley (1995), 471−486. 16. Gabora, L. A day in the life of a meme. Philosophica, 57, (1996), 901-938. 17. Gabora, L. 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, Sheridan Publishing (2008a), 1466-1471. 18. Gabora, L. Modeling cultural dynamics. In Proceedings of the Association for the Advancement of Artificial Intelligence (AAAI) Fall Symposium 1: Adaptive Agents in a Cultural Context, Menlo Park, CA: AAAI Press (2008b), 18-25. 7 19. Gabora, L., Leijnen, S., & von Ghyczy, T. The relationship between creativity, imitation, and cultural diversity. International Journal of Software and Informatics, (in press). 20. Hartley, J. From cultural studies to cultural science. Cult. Science, 2 (2009), 1-16. 21. Higgs, P. G. The mimetic transition: a simulation study of the evolution of learning by imitation. Proceedings of the Royal Society B: Biological Sciences, 267, 1450(2000), 1355 -- 1361. 22. Heyes, C. M. Theory of mind in nonhuman primates. Behavioral and Brain Sciences, 211(1998), 104 -- 134. 23. Hinton, G. E. & Nowlan, S. J. How learning can guide evolution. Complex Systems, 1(1987), 495 -- 502. 24. Holland, J. K. Adaptation in Natural and Artificial Systems. Ann Arbor, MI: University of Michigan Press, 1975. 25. Hutchins, E. & Hazelhurst, B. Learning in the cultural process. In Langton, C., Taylor, J., Farmer, D., &Rasmussen, S. (Eds.) Artificial Life II. Redwood City, CA: Addison-Wesley, 1991. 26. Kirby, S. Spontaneous evolution of linguistic Structure -- An iterated learning model of the emergence of regularity and irregularity. IEEE Transactions on Evolutionary Computation, 5, 2(2001), 102-110. 27. Leijnen, S., & Gabora, L. How creative should creators be to optimize the evolution of ideas? A computational model. Electronic Proceedings in Theoretical Computer Science, 9(2009), 108-119. 28. Mesoudi, A., Whiten, A., & Laland, K. Toward a unified science of cultural evolution. Evolution 58, 1(2004), 1-11. 29. Mesoudi, A., Whiten, A., & Laland, K. Toward a unified science of cultural evolution. Behavioral and Brain Sciences, 29(2006), 329-383. 30. Mithen, S. (Ed.). Creativity in Human Evolution and Prehistory. London, UK: Routledge, 1998. 31. Ruff, C., Trinkaus, E., & Holliday, T. Body mass and encephalization in Pleistocene Homo. Nature, 387 (1997), 173 -- 176. 32. Tomasello, M., Kruger, A. C., & Ratner, H. H. Cultural learning. Behavioral and Brain Sciences, 16(1993), 495 -- 552. 33. Wexler, B. Brain and Culture: Neurobiology, Ideology and Social Change. New York: Bradford Books, 2006. 34. Whiten, A., Hinde, R. A., Laland, K. N., Stringer, C. B. Culture evolves. Philosophical Transactions of the Royal Society B, 366(2011), 938 -- 948. 8
1905.09795	1	1905	2019-05-23T17:38:20	Nature-Inspired Computational Model of Population Desegregation under Group Leaders Influence	[ "cs.MA", "cs.SI" ]	This paper presents an agent-based model of population desegregation and provides a thorough analysis of the social behavior leading to it, namely, the contact hypothesis. Based on the parameters of frequency and intensity of influence of group leaders on the population, the proposed model is constituted by two layers: 1) a physical layer of the population that is influenced by and 2) a virtual layer of group leaders. The model of negotiation and survival of group leaders are governed by the nature-inspired evolutionary process of queen ants, also known as Foundress Dilemma. The motivation of using a virtual grouping concept (instead of taking a subset of population as the group leaders) is to stay focused on finding the conditions leading individuals in a society tolerating a significantly diversified (desegregated) neighborhood, rather than, indulging into complex details, which would be more relevant to studies targeting the evolution of societal group and leaders. A geographic information system-driven simulation is performed, which reveals that: 1) desegregation is directly proportional to the frequency of group leaders' contact with the population and 2) mostly, it remains ineffective with an increase in the intensity of group leaders' contact with the population. The mechanism of group selection (the conflict resolution model resolving the Foundress Dilemma) reveals an exciting result concerning negative influence of cooperative group leaders. Most of the time, desegregation decreases with increase in cooperative leaders (the leaders enforcing desegregation) when compared with fierce leaders (the leaders enforcing segregation).	cs.MA	cs	JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2017 1 Nature-Inspired Computational Model of Population Desegregation under Group Leaders Influence Kashif Zia, Dinesh Kumar Saini, Arshad Muhammad, and Alois Ferscha Member, IEEE 9 1 0 2 y a M 3 2 ] A M . s c [ 1 v 5 9 7 9 0 . 5 0 9 1 : v i X r a Abstract -- This paper presents an agent-based model of popula- tion desegregation and provides a thorough analysis of the social behavior leading to it, namely the contact hypothesis. Based on the parameters of frequency and intensity of influence of group leaders on the population, the proposed model is constituted by two layers: (i) a physical layer of the population that is influenced by (ii) a virtual layer of group leaders. The model of negotiation and survival of group leaders is governed by the nature-inspired evolutionary process of queen ants, also known as Foundress Dilemma. The motivation of using a virtual grouping concept (instead of taking a subset of population as the group leaders) is to stay focused on finding the conditions leading individuals in a society tolerating a significantly diversified (desegregated) neighborhood, rather than, indulging into complex details, which would be more relevant to studies targeting the evolution of societal group and leaders. A GIS-driven simulation is performed, which reveals that (i) desegregation is directly proportional to the frequency of group leaders' contact with the population, and (ii) mostly, it remains ineffective with an increase in the intensity of group leaders' contact with the population. The mechanism of group selection (the conflict resolution model resolving the Foundress Dilemma) reveals an exciting result concerning negative influence of cooperative group leaders. Most of the time, desegregation decreases with increase in cooperative leaders (the leaders enforcing desegregation) when compared with fierce leaders (the leaders enforcing segregation). Index Terms -- Population Segregation, Desegregation, Agent- based Model, Foundress Dilemma, Group Selection. I. INTRODUCTION According to Oxford dictionary, segregation is defined as "the action or state of setting someone or something apart from others". Historically, segregation is connected with the dynam- ics of biological, and more specifically human population. For a human population, segregation is a progressive outcome of population demography as a result of discriminatory settlement behavior of the individuals [1]. The discrimination is based on some bias, such as sex, race or color that influences people decision of where to live. It has been observed that such individualistic decisions have shaped the demographics of cities, sometimes resulting in exceptionally segregated settlements [2]. Settlements having extreme segregation are sometimes not desirable. For example, the countries that have experienced racial abuse and ethnic wars desire desegregation (anti-segregation) in its population [3]. Brewer [4] defined desegregation as: "extended contact between previously iso- lated social groups that are brought on either by acute or K. Zia is with the Faculty of Computing and Information Technology, Sohar University, Oman e-mail: [email protected]. D.K. Saini, and A. Muhammad are with the Faculty of Computing and Information Technology, Sohar University, Oman. Alois Ferscha is with the Institute of Pervasive Computing, Johannes Kepler University, Linz, Austria. Manuscript received April 19, 2005; revised August 26, 2015. gradual processes of change". However, this change cannot be brought about through legislation and laws [5]. It is established from the experience that such a change is inseparable without people themselves realize change [6] [7]. Human societies evolve with time may be due to the intention of people to mix-up and desegregate. Standard processes of evolution are randomness, agency, organization, and contingency [8]. Agent- based Modeling (ABM) provides an ideal platform to model and configure these processes [9], thus acting as an enabler not only to observe the evolutionary process itself but also, find answers of the interesting "what-if" questions. A model, whether agent-based or not is always based on a foundation. For social phenomena, such as segregation (desegregation), this can be based on a theory or empirical evidence. Since empirical evidence of individualistic behavior in the population is difficult to get and synthesize towards a population-level dynamics, the theory-driven foundations are often more appropriate. However, theories are intrinsically contradictory. Computer simulation provides means to evaluate the contradictory theories relating to the same phenomena and specify conditions in which one theory is applicable, while others are not. The rapidly growing field of Computational Social Science (CSS) [10] [11] is about understanding a society using the computer simulation. More recent tools, such as Cellular Au- tomata (CA) (from physics and mathematics), and Distributed Artificial Intelligence and agent technology (from computer science) have influenced this growth positively [12]. Due to its nature, the social simulation should be a theory-driven system with a focus on the explanation of the phenomena of interest (rather than the desire for specific outcomes) [13]. The novelty of social simulation is based on the observation that an over- whelmingly complex behavior emerges from relatively simple local activities [14]. ABM is a popular modeling technique used by CSS researchers due to its ability to reproduce the societal effects purely based on interactions at the local level [15]. CSS has a long history and a strong underpinning on the social theories. Schelling's seminal computational model of the population segregation [1] has been a reference point for diverse re- search streams [16], such as, population dynamics and land use [17] -- [21], social interaction dynamics [22], social and psychological behavior studies [23] [24], and business and market analysis [16]. Shelling's model suggests that even if the majority of people are not biased and tolerate a neighborhood (locality), which is not exactly of their liking, segregation is still a possibility on a global scale. More specifically, Schelling's model is successful in demonstrating that a predominantly tolerant population may also generate a JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2017 2 segregated settlement, thus, validating that what emerges at a global level may not be a true representation of the local parts. The strength of otherwise a basic neighborhood-based model was vested in agents' strict localized decisions producing an unexpected outcome at the population level. Schelling's model, therefore, was one of the first experimental evidence for the emergence of totally unpredictable global behavior, which was contradictory to the local rules producing it. Due to this reason, Schelling's model is often used to demonstrate the strengths of an agent-based modeling paradigm; a modeling technique specifically developed to analyze these situations. To model a socially-inspired model of desegregation, it is important to understand the human behavior and the social processes involved. Norman S. Miller and Marilynn B. Brewer defined a relevant hypothesis, that is the contact hypothesis as: "the idea that prejudice and hostility between members of segregated groups can be reduced by promoting the fre- quency and intensity of intergroup contact" [25]. However, there are other theories, which entirely contradict this theory. For example, it is also claimed in [26] that "contact tends to reproduce, rather than challenge, the inequitable racial structure." With the help of the model presented in this paper, the conditions in which the contact hypothesis is applicable are evidenced. Consequently, the conditions in which the contact hypothesis is not applicable are also evidenced, thus, favoring the contradictory hypothesis [26]. However, the modality/medium of intergroup contact is hard to resolve. There are numerous factors that can be considered, such as, group boundaries, the scope of influence, relative activism within a group, and the cognitive traits that would be responsible for bringing the change. This could lead to a complex mechanism of defining and evaluating the outcome of intergroup contact. However, in fact, the outcome of intergroup contact influences an individual in a population towards two possible states of minds; appreciation for the societal good (cooperation) or not (defection); provides the much-needed ground for simplification. Therefore, for the purpose of the model presented in this paper, the responsibility of intergroup contact is taken away from the real population and assigned to the artificial society of leaders. It is motivated by the argument that contact itself (without any external anti-segregation im- pact) is not enough to change the racial structure as suggested in [26]. This means that leaders' influence the population to segregate or desegregate based on their type. In this way, the notions of intensity and frequency of intergroup contact are equally replicated as intensity and frequency of leaders' influence on the population. Hence, the proposed model is based on two layers: a physical layer of real population of agents, which is influ- enced by a virtual layer of group leaders. The periodicity and intensity of the influence exerted by group leaders onto the real population transform it into a settlement, which is segregated or otherwise. Since the real population only decides to segregate or not to segregate, a simple model motivated by Schelling's computational model of segregation is sufficient [1]. If under this model, an agent does not make a move, even though a majority of its neighbors are not of its kind, it practically acts in favor of desegregation. The model of virtual layer's influence is then the model of interest and termed as the model of leaders influence. The model of leaders influence on the agents is partially based on the contact hypothesis [25]. How much these leaders influence the behavior of agents depends on the attributes related to the leaders, such as (i) capabilities, (ii) relationship, (iii) environment, and (iv) strategies of conflict-resolution (group dynamics). The capabilities are defined by frequency and intensity of interaction [25] of the leaders with the agents. The relationship between leaders and agents is realized by the presence and proximity constraints, which are in turn depen- dent on the density of leaders population (an agent must have one or more leaders in its proximity to get itself influenced). It is notable that the relationship constraints of the environment restrain the capabilities of the leaders. For this paper, the environment is spatial in nature. A CA based environment is used to define the interlayer relationship thus mechanizing the capabilities of interaction between leaders and agents. For other types of environments, this relationship would have been different. For example, if leaders are operating through social networks or mass media, their interaction mode and influence would have been connectivity and availability rather than proximity and presence. Lastly, the strategies of grouping and conflict resolution between the leaders are also modeled. A nature-inspired model of grouping, survival, and influence is used. Nature- inspired models provide basic mechanisms of inter-organism interaction leading to mutual survival, role assignment and cooperative problem-solving. For the model proposed in this paper, it makes modeling of leaders straight-forward in terms of: (i) interaction pattern and influence, (ii) relationship and dependability, and (iii) strategies of conflict-resolution. The process of negotiation and survival of the leaders is governed by the nature-inspired evolutionary model of queen ants be- havior termed as Foundress Dilemma. To the best of our knowledge, the model of such complexity and coverage is not reported previously. The rest of the paper is organized as follows. In section II, a review of related work is given. Section III presents an overview of the models. In section IV, a detailed description of agent-based models is given. Section V explains the simulation experiments and the analysis of the simulation results, followed by conclusions of the paper in section VI. II. RELATED WORK Nature-inspired computing is a computing paradigm in- spired by self-organizational behavior in naturally-appearing complex systems [28]. Originating from the contemporary ant colony [29], and particle swarm [30] optimization algorithms, more unusual behaviors have been studied more recently [31]. These include firefly [32], bat [33], and cuckoo search [34] algorithms. Nigel Gilbert and Klaus G. Troitzsch in their book proposed various social science computational models and simulation methods, which are helpful to understand the society [15]. Axelrod is one of the pioneers, who used an evolutionary approach to describe societal norms focusing on the stability and emergence [35]. A more general discussion on evolution- ary algorithms to understand the social processes as a whole JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2017 3 Fig. 1. A schematic representation of integrated model. The process of segregation works under the combined effect of the neighborhood of an agent and the influence exerted by the leaders. Formation of nests of leaders is governed by a model motivated from Foundress dilemma [27]. Each generation of leaders is placed on the map independent of previous placements (this guarantees influence distribution all across the agent population). The frequency of leaders' influence relates to how often they reproduce to produce new generations, whereas, the intensity of leaders' influence relates to how fierce they are (e.g. leader in Nest 2 is more fierce than other leaders, indicated by increased size). is given by Chattoe-Brown, et al. [36]. Similarly, Kenneth A. De Jong presented various dimensions of evolutionary computation in his book, which helps to understand the nature- and bio-inspired models of population behavior [37]. Cioffi, De Jong and Bassett in their paper described biologically- inspired approaches to understand complex social systems [38]. Their work is seminal, because it presents a combined approach of evolutionary approaches used with the agent- based modeling. An overview of altruism, cooperation, mutualism, strong reciprocity, and group selection in the social settings is pro- vided by West, et al. [39]. A framework for achieving group level goals with the formation of agents' coalitions is presented by Axtell [40]. Modeling and simulation work on segregation is widely available [41] -- [43]. Romans Pancs, and Nicolaas J Vriend [44] have adapted Schelling's segregation model to have a strict preference in favor of integration (desegregation). They found that even in such case, the best response dynamics leads to the segregation at a global level. Similar results are also presented by Junfu Zhang [45]. These are interesting results highlighting that population on its own cannot settle for desegregation; it requires external factors. Mostly, work on the population desegregation is primarily of empirical nature. For example, Smets, et al. [46] demon- strated the difficulty in quantifying the population behavior in varying degree of possible contacts. Sahasranaman studied evolutionary dynamics of neighborhood economic status in the cities, which also have a great impact on the desegregation of people in big cities [8]. Hatna, Erez and Besnion studied the Schelling model of ethnic residential dynamics, and their findings suggest an integrated-segregated dichotomy of the patterns in population behavior. Miles Hewstone and Alberto Voci [47] have studied the effectiveness of intergroup contact in reducing prejudice in a society. Contrarily, Jeffrey S Denis [26] concluded that "contact tends to reproduce, rather than challenge, the inequitable racial structure". It is contrary to the expected outcome (desegregation), when applying the contact theory [25]. Nevertheless, this is again an empirical study without involving any computational model. In this paper, the external driving factor for desegregation is an overlay virtual population representing the influence factors from the leaders. The grouping dynamics of these leaders is mediated by nature-inspired ant queen behavior [27]. Queen ants perform nests management following evolutionary mech- anism of group selection, which is essential for the survival of the entire population. Hence, using such mechanism to analyze population dynamics (including segregation / desegregation) is natural. Despite its relevance, to the best of our knowledge, no such work exists in the literature. This work is an effort to address this gap. III. MODEL OVERVIEW The model of population segregation is simple, motivated by Shelling's original publication [1]. Agents representing human population are of two ethnic types; expat and native. An agent has a desire to segregate if the majority of the agents in its neighborhood are not of its type. This contemporary model of segregation is extended to include the influence of leaders in the decision-making of the agents. Hence, an agent would only be happy, if it is happy with its surrounding as well as its compatibility with the type of influence exerted by the leaders in its surrounding. The model of leaders influence is motivated from [27] and assumes the presence of cognitive leaders (leaders influencing the opinion of the real population JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2017 4 it to segregate or otherwise). These leaders are virtual because they are not part of the real population. The leaders' behavior is motivated by ant queens' behavior reported in [27], briefly described in the following. Ants live in colonies. When young queens reach the age of reproduction, they leave their nest, and try to build their own nest; is a very challenging task for a queen. In many species of the ants, the queens are predominately fierce. They fight with other queens and independently build their colony. However, in some species of the ants, the queens, also of being fierce, are also cooperative and may live in a nest together. In this situation, one of them becomes a leader of the group and performs the duty of reproduction; while others sacrifice their natural instinct of reproduction in favor of the leader queen. In fact, it increases their chances of survival collectively, while adhering to the evolutionary phenomena of group selection. While involved in the process of group selection, the choice to fight or cooperate is defined as Foundress Dilemma [27]. Motivated from above, there are two types of leaders in the proposed model. Cooperative leaders are those, who promote a sense of cooperation (possibly leading to desegregation). Fierce leaders are those who promote a sense of non-cooperation (possibly leading to segregation). The Fierce leaders always fight with the cooperative leaders. The fierce leaders may fight with each other if one or both of them are fierce beyond a threshold. The model of leaders influence is comprised of five sequen- tial steps; named as reproduction, clustering, fighting, group competition, and influencing (the details are given in section IV). The agent-based model combines the grouping behavior, the dynamics of interaction of the group leaders and the real population, and neighborhood based segregation mechanism; with an underpinning on the environmental constraints. A schematic representation of the overall model is given in Fig. 1. As the population under the influence of a nest (or a leader) is mere of physical nature (agents in influence range of a leader), the outcome (the desegregation index) is entirely depended on the presence, evolution, frequency, and intensity of the influence. This helps to understand the conditions lead- ing to the population desegregation in terms of (i) spread of (de)segregation effort (relative number of cooperative vs. fierce leaders), (ii) evolving leaders' populations and distribution, (iii) intensity of (de)segregation effort (level of fierceness of fierce leaders), and (iv) the frequency of (de)segregation effort (speed of reproduction in the evolutionary process). IV. AGENT-BASED MODEL The agent-based model proposed in this paper models grouping dynamics, and interaction of the group leaders and the real population with an underpinning on the environmental constraints. Under these settings, the capabilities of group leaders (in terms of frequency and intensity of interaction) towards the agent's decision to stay or leave a place are evaluated. A. Simulation Space Both, the model of segregation and the model of leaders influence are simulated on the real map. The GIS maps of Fig. 2. Setting up a population of 5000 agents on the map of Sohar city. Agents in blue are expats and agents in red are natives. Two regions, one colored in red and the other colored in blue show the population of these regions segregated enough (segregation is more than segregation threshold of 40%) in favor of natives and expats, respectively. the city are converted into a CA-based structure. Each cell in the CA World belongs to a region/locality of the city. Each region of the city has a population, which is assigned an initial value equal to a fraction of (dependent on the area of the region) total agent population. Each member of this population represents a household and exists as an agent occupying a cell in the corresponding region. The type of each generated agent is set to either "expat" or "native", randomly. A region's regionType can also be one of "expat" or "native", if region's agent population is segregated enough (defined by prescribed segregation threshold), otherwise it would be "neutral". A setup representing these parameters is shown in Fig. 2. B. Model of Segregation An agent has a desire to segregate if the majority of agents in its neighborhood are not of its type. The definition of majority is based on the parameter PDTU (abbreviation for percentage difference to be unhappy). If an agent has the ratio of agents in its neighborhood not of its own type greater than PDTU, it becomes "unhappy" and try to move somewhere else. Otherwise, it is "happy" and remains where it is. An unhappy agent will move to a region whose regionType is equal to its own type or "neutral". Fig. 3 shows population after segregation model is applied with the parameters: agent population = 5000, segregation threshold = 60%, and PDTU = 40%. It is worth noting that the extension of segregation model so that it also incorporates the leaders' influence is given later in the sub-section "integrated model". JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2017 5 in isolation; there would be no clustering. Clusters act as nests, ready for consolidation and inducing influence on agent population. A scenario describing clustering is exhibited in Fig. 4. Individual Fight In a nest, the fight between the leaders to claim the nest would initiate after reproduction and clustering. A fierce leader would initiate a fight if prob init fight is greater than a random floating point number. For example, if the value of prob init fight is 0.5, there is an equal probability of fierce leader to initiate a fight or otherwise. If the fierce leader is set to initiate a fight, it will fight if there are more leaders in its nest. If the randomly chosen leader by the initiator is of type fierce as well, the probability for both of them winning the fight is equal. The leader losing the fight will die. If a fierce leader is imposing a fight on the cooperative leader, the probability of fierce leader dying is 60%, whereas, the probability of cooperative leader dying is 40%. All leaders in a nest perform the mechanism stated above until all other leaders have expired their turn or the leaders remaining are only cooperative, or there is no other leader (all died as a result of the fight). Group Competition The fights in the nests would follow group competition. If two nests are close enough (identified by radius competition), the survival function surv function from Bartz and Hoelldobler [48] would determine survival chances of the nest, which corresponds to the number of leaders living together in a nest. As a result of this competition, one of the nests (having chances of survival less than the other) would be destroyed along with the leaders in it. The cell representing a nest would first be assigned a value for the survival function surv function, calculated as [48]: surv function = −2.88+4.28×nest pop−0.377×nest pop2 (1) where nest pop is the number of leaders living together in a nest. Out of two competing nest, a nest having greater value of surv function would survive. Influence The last step how a nest of leaders would influence the population of agents. All the nests would generate an infec- tious pheromone in an infection range around them, which is the product of nest pop and radius competition. The infection type is one of the randomly chosen leaders in the nest. A cell under this influence range would update its infected with value with the type of influence that this nest has: "cooperation", "non-cooperation", or "null" (if not in infection range of any of the nest). D. Integrated Model Altogether, the conditions describing the state of an agent (being "happy" or "unhappy") in the context of segregation, not only depend on the population type in the neighborhood, Fig. 3. Population segregation after equilibrium is achieved at iteration 25. Some regions have attained segregation threshold (region as a whole having regionType "expat" [in blue] or "native" [in red]) while many regions are still below segregation threshold (with Type "neutral"). Even regions of regionType "neutral" achieve local segregation. C. Model of leaders Influence A small fraction of cooperative leaders are initially dispersed randomly across the map. Remaining are fierce leaders. These leaders are continuously created, and destroyed in each generation, hence evolving with time. Consequently, the model of segregation gets augmented to incorporate the influence induced by this model. The model of leaders influence operates in five sequential steps, described in the following. Reproduce At the start of each five-step execution cycle, all the co- operative, and fierce leaders reproduce themselves depending on the previous generation. The population of leaders in each generation is same, but the relation between the cooperative and fierce leaders may change due to mutation. A minimal mutation index equal to 0.01 would guarantee a change in the population distribution of leaders during evolution. Cluster After reproduction, the leaders if close enough (identified by cluster radius) would cluster, i.e. moving right adjacent (overlapped) to each other irrespective of their type. Hence, several nests would be formed, where each nest may contain more than one leaders of different types. During clustering, all leaders would choose a single cell in a specified range to cluster. Each leader on its turn chooses a cell in cluster radius, which has the maximum number of leaders placed on it and moves into that cell. If there are more than one options available, the leader can choose a cell randomly. If there is no such cell, the leader is placed JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2017 6 Fig. 4. Clustering of leaders. (a) A section of map with leaders (ant shape agents) placed randomly. (b) leaders cluster together creating nest (three of them are visible). (c) A closer view of one of the nest (the one on the right in (b)), with one cooperative and one fierce leader clustered. but also, the aura of influence induced into the cellular space around it. The index of interest, the Desegregation Index is calculated as follows. Each agent in the population calculates its own individual index of desegregation (IID), which is the ratio of the number of individuals NOT of its type, and the total number of neighbors in its neighborhood (8 adjacent cells of Moore's neighborhood [49]). For example, if an individual is of type "expat" and if 3 out of 5 of its neighbors are also expats, the value of IID would be equal to 0.4 (40% desegregation). Alternatively, if 2 out of 5 of its neighbors are also expats, the value of IID would be equal to 0.6 (60% desegregation). Such an agent will be happy in the second case given that PDTU is 50%. IID is combined with nest influence to evaluate the state of happiness of an agent X. That is: X = "happy(cid:48)(cid:48) "unhappy(cid:48)(cid:48) if cond1 is true if cond2 is true (2) (cid:40) where cond1 is true when IID >= PDTU AND infected with c = "cooperation" OR "null". This states the combined condition in which X's IID is high enough to suppress unhappiness and also, the nest influence supports desegregation ("cooperation") or there is no influence ("null" for no influencing nest in the surrounding); cond2 is true otherwise. V. SIMULATION AND RESULT ANALYSIS For simulation, Netlogo [50], a CA-based agent-based mod- eling tool is used. The simulation is executed for a real city (Sohar, Oman) using its GIS map. Several important cases across multiple settings of simulation parameters are analyzed. Parameters used in the simulation are as follows: • NOL: number of leaders constituting the virtual layer. The values are 25 and 50. • FC: fraction of leaders of type cooperative in the entire population of leaders (NOL). The values are 0.1, 0.2, and 0.5. • IR: leaders' influence rate, which represents how fre- quently leaders are influencing the population concerning TABLE I DESEGREGATION INDEX IR 5 25 50 FC 0.1 0.2 0.5 0.1 0.2 0.5 0.1 0.2 0.5 PIF=0.1 0.1287 0.1316 0.1172 0.0441 0.0441 0.0385 0.0432 0.0405 0.0402 NOL=25 PIF=0.2 0.1240 0.1199 0.1129 0.0458 0.0409 0.0385 0.0451 0.0353 0.0366 PIF=0.5 0.1156 0.1061 0.1008 0.0424 0.0369 0.0318 0.0394 0.0390 0.0349 PIF=0.1 0.1917 0.1908 0.1679 0.0631 0.0678 0.0531 0.0633 0.0636 0.0509 NOL=50 PIF=0.2 0.1701 0.1726 0.1595 0.0669 0.0569 0.0521 0.0521 0.0550 0.0509 PIF=0.5 0.1594 0.1611 0.1172 0.0536 0.0453 0.0469 0.0469 0.0532 0.0472 NOL = Number of Leaders PIF = Probability of Initiating Fight IR = Leader Influence Rate FC = Fraction of Cooperation in the population. TABLE II HAPPINESS INDEX IR 5 25 50 FC 0.1 0.2 0.5 0.1 0.2 0.5 0.1 0.2 0.5 PIF=0.1 0.6663 0.6936 0.6936 0.8925 0.8916 0.9063 0.8937 0.9017 0.9026 NOL=25 PIF=0.2 0.6774 0.6882 0.7065 0.8887 0.9003 0.9063 0.8891 0.9148 0.9112 PIF=0.5 0.7018 0.7245 0.7363 0.8958 0.9110 0.9216 0.9021 0.9058 0.9153 PIF=0.1 0.5006 0.5033 0.5638 0.8512 0.8423 0.8768 0.8520 0.8519 0.8823 NOL=50 PIF=0.2 0.5585 0.5553 0.5879 0.8414 0.8683 0.8808 0.8808 0.8698 0.8820 PIF=0.5 0.5869 0.5841 0.6901 0.8747 0.8956 0.8921 0.8921 0.8742 0.8894 NOL = Number of Leaders PIF = Probability of Initiating Fight IR = Leader Influence Rate FC = Fraction of Cooperation in the population. simulation time lag. Values are every 5th, 25th and 50th iteration. • PIF: probability of initiating the fight by a fierce leader, representing the intensity of influence of group leaders. Values are 10%, 20%, and 50%. The other parameters related to the model of influence are constants: • pmutation: mutation probability of each generation (set to 1%). • cluster-radius: radius of the cluster (set to 10 cells). JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2017 7 Fig. 5. Average Desegregation Index of Simulation runs corresponding to TABLE I. NOL = Number of Leaders, PIF = Probability of Initiating Fight, IR = Leader Influence Rate, FC = Fraction of Cooperation in the population. Fig. 6. Average Happiness Index of Simulation runs corresponding to TABLE II. NOL = Number of Leaders, PIF = Probability of Initiating Fight, IR = Leader Influence Rate, FC = Fraction of Cooperation in the population. • radius-competition: the radius of the group competition (set to 50 cells). The constant parameters related to the model of segregation (desegregation) are: • population size: set to 5000 agents. • segregation threshold: set to 40%. • PDTU: difference in population types in the neighbor- hood of an agent (the -ve percentage that makes it unhappy). set to 40%. It is worth noting that scenarios representing different values for constants of ranges, population, and thresholds would entirely change the simulation results. Therefore, these values are kept constant, and only those, which give a consistent meaning to phenomena of interest are varied. Table I and Table II outline the simulation results. To calculate the global Desegregation Index, IID of all the agents is accumulated and divided by the total number of agents. Similarly, Happiness Index, which is inversely proportional to IID, is an accumulated global value of agent happiness as formulated in the equation 2. Consequently, with an increase in the desegregation index, the population as a whole is expected to be less happy. This is exemplified by the value of the index in case of the model of segregation applied alone. Without vir- tual leaders, and their influence, the index is equal to 0.0152. In fact, this represents that the population is segregated, and index of desegregation is 0.0152. It understandably makes overall population very happy with an index equal to 0.9568. Analyzing Table I, and the corresponding graphs in Fig. 5, it is evident that with an increase in the number of leaders, the Desegregation Index increases, the population gets desegregated. It is seen that with more frequency of influence dissemination, the population gets desegregated as well. Overall, with an increase in the intensity of the in- fluence, the value of desegregation decreases. It is because the fierce leaders kill cooperative leaders, thus decreasing the corresponding influence on the population. The maximum that is, JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2017 8 Fig. 7. Comparison of desegregation index against simulation iterations between FC = 0.1 and FC = 0.5. Specific case of NOL = 50, IR = 5 and PIF = 0.1. [NOL = Number of Leaders, PIF = Probability of Initiating Fight, IR = Leader Influence Rate, FC = Fraction of Cooperation in the population] Fig. 8. Effect of Leaders' influence on the population represented by simulation screenshot at iteration 30 (left) and 31 (right) in case NOL = 50, IR = 5, PIF = 0.1, FC = 0.1. [NOL = Number of Leaders, PIF = Probability of Initiating Fight, IR = Leader Influence Rate, FC = Fraction of Cooperation in the population] desegregation is achieved when the intensity of the influence as well as the fraction of cooperative leaders are minimal. Happiness Index corresponds to Desegregation Index, and all values are consistent as shown in Table II, and corresponding graphs in Fig. 6. The maximum desegregation equal to 0.1917 is achieved when N OL = 50, IR = 5, P IF = 0.1, and F C = 0.1. However, there is a difference of almost 15%, when F C = 0.1 with F C = 0.5 is compared. A closer look at this trend exhibited in the graph in Fig. 7. The graph shows a regular pattern. Overall, the level of desegregation is highest when a new phase of influence dissemination is started (after each five simulation iterations, in this case, i.e. IR = 5), and decreases continuously afterwards unless it reaches a new peak at the wake of a new influence dissemination phase. In each such phase, the index is higher when FC is less, which means more desegregation. It means that more cooperative leaders act negatively for the desegregation. It is an interesting result, which is justified using simulation screenshots given in Fig. 8 and Fig. 9. Fig. 8 (left) shows the arrangement of the agents at iteration 30, one iteration before the leaders assert their influence. It is JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2017 9 Fig. 9. Effect of Leaders' influence on the population represented by simulation screenshot at iteration 30 (left) and 31 (right) in case NOL = 50, IR = 5, PIF = 0.1, FC = 0.5. [NOL = Number of Leaders, PIF = Probability of Initiating Fight, IR = Leader Influence Rate, FC = Fraction of Cooperation in the population] Fig. 10. Comparison of desegregation index (left) and happiness index (right) between IR = 5 and IR = 25 in case NOL = 50, FC = 0.1, PIF = 0.5. [NOL = Number of Leaders, PIF = Probability of Initiating Fight, IR = Leader Influence Rate, FC = Fraction of Cooperation in the population] Fig. 11. Simulation screenshot at iteration 25, just before invocation of periodic influence cycle for IR = 5 (left) and IR = 25 (right) representing the importance of more periodic group influence for population desegregation, in case NOL = 50, PIF = 0.1, and FC = 0.1. [NOL = Number of Leaders, PIF = Probability of Initiating Fight, IR = Leader Influence Rate, FC = Fraction of Cooperation in the population] JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2017 10 evident that there are few cooperative leaders. Consequently, the population is quite segregated (except for some isolated islands where cooperative leaders can assert their influence). At iteration 31 (right), the population gets quite desegregated due to the refreshed influence of the leaders. However, it does not happen at the fringes of the city due to the absence of cooperative leaders. These screenshots represent situations when F C = 0.1. A similar analysis is done when F C = 0.5 as shown in Fig. 9. There is not much difference in arrangement of the agents at iteration 40 (left), one iteration before the leaders assert their influence. However, at iteration 41, the reach of leaders at the fringes of the city (due to more number) is evident.Hence, this strange outcome relates to demographic features of the population, not the comparative number of leaders, and their influence. Another interesting result is about comparing the desegre- gation and happiness in case of varying influence periodicity. In the constant conditions of N OL = 50, P IF = 0.5, F C = 0.1, more periodic influence dissemination keeps the deseg- regation at a higher level. It is shown in the graph in Fig. 10 (left). At the initial stages of the simulation, both for IR = 5 and IR = 25, the Desegregation Index is similar (around 0.3), which drops for subsequent iterations following the similar trend. However, when the first period of refreshing of the influence happens (at iteration 5) for IR = 5, the Desegregation Index elevates from around 0.15, while for IR = 25, the index keeps dropping. In the subsequent periods, each separated by five timestamps, the Desegregation Index keeps fluctuating between a maximum of around 0.26 and a minimum of around 0.1. In case of IR = 25, the Deseg- regation Index almost drops to 0.1 and then gets a boost at iteration 26 when the first period of refreshing of the influence happens for IR = 25. After refresh, the peak value does not surpass 0.2. However, the real issue is the length of time for which the value of index remains between 0.1 and 0.0, which is almost 80% of the time. Opposite of Desegregation Index is the pattern of Happiness Index shown in the graph in Fig. 10 (right). The above pattern is evident from the simulation screenshots given in Fig. 11, evidencing the importance of increased periodicity of influence for the population desegregation. Overall, the simulation results reveal the following out- comes: • With the increase in the frequency of contact of the leaders with the population, desegregation increases. • With the increase in the intensity of contact of the leaders with the population, desegregation decreases. • Desegregation is only possible if the population gets in contact with the leaders. In the next section, a discussion on these outcomes is presented. VI. DISCUSSION AND CONCLUSION Contradictory views regarding the effect of contact on desegregation are available. On one extreme, the contact hy- pothesis [25] states that: "the idea that prejudice and hostility between members of segregated groups can be reduced by promoting the frequency and intensity of intergroup contact". Whereas, one the other extreme, it is stated that [26] "contact tends to reproduce, rather than challenge, the inequitable racial structure". However, both hypotheses do not take into account the contradictions and sustainability of influencing groups. Since, segregation is a result of internal bias of individuals, so must be the counter of it, i.e. desegregation. Social bias is not purely an internal factor. It consolidates with the passage of time, gained with experience and social influence. Hence, the aversion of bias induced by some imposing group should not be examined as a deterministic uni-dimensional factor acting against segregation. In this paper, a model based on group selection [27] is customized to integrate it with the model of group influence on the population. The model resolves the Foundress Dilemma for each generation of spatial group leaders and provides an evolutionary mechanism that may transform the virtual population of influence. The real population, in addition of having internal bias, gets influenced by an external non- deterministic bias, generated by virtual group leaders. The emerging nature of group influence integrated with population decision (to desegregate or not to) produces more knowledgeable outcomes, validating some aspects of the the- ories and hypothesis outlined above, while rejecting others. It can be argued that contact itself (without any external anti- segregation impact) is not enough to change racial structure (as suggested in [26]). However, nobody would have any such expectation, as this is against pure human nature. With external influence, the change in racial structure is possible, thus, providing a justification of keeping real population and leaders as two separate groups (layers). A part of contact hypothesis is validated, i.e. with the increase in the frequency of contact of leaders with the population, desegregation increases. However, with a strong assumption that the virtual population is analogous to the real population (predominately favoring segregation), the second part of the hypothesis is not validated. In fact, it was observed that with the increase in the intensity of contact of leaders with the population, desegregation decreases. More intense the viewpoint of virtual leaders (influence leaders) is, more intense would be its impact, practically, favoring segregation instead of desegregation. The simulation results reveal that it is also not true that decrease in the intensity of contact always increase the deseg- regation. A decrease in the intensity of contact combined with decrease in the number of cooperative leaders favoring deseg- regation results in best case with maximum desegregation. Future Work: Although the simulation results are validated against two contradictory theories of contact, validating them on real city population will be an interesting dimension to consider for the future work. REFERENCES [1] T. C. Schelling, "Dynamic models of segregation," Journal of mathe- matical sociology, vol. 1, no. 2, pp. 143 -- 186, 1971. [2] M. J. White, "Segregation and diversity measures in population distri- bution," Population index, pp. 198 -- 221, 1986. [3] O. D. L. Gona, Dawn of Desegregation: JA De Laine and Briggs V. Elliott. Univ of South Carolina Press, 2012. JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2017 11 [4] M. B. Brewer, "The social psychology of desegregation: An introduc- tion," Groups in Contact: The Psychology of Desegregation, p. 1, 2013. [5] A. J. Christopher, "The slow pace of desegregation in south african cities, 1996-2001," Urban studies, vol. 42, no. 12, pp. 2305 -- 2320, 2005. [6] S. Oldfield, "Urban networks, community organising and race: an analy- sis of racial integration in a desegregated south african neighbourhood," Geoforum, vol. 35, no. 2, pp. 189 -- 201, 2004. [7] G. Liu and W. L. Taylor, "School choice to achieve desegregation," Fordham L. Rev., vol. 74, p. 791, 2005. [8] S. F. I. R. T. C. C. Evolution, Website, 2017. [Online]. Available: https: //www.santafe.edu/research/themes/computational-cultural-evolution [9] M. W. Macy and R. Willer, "From factors to factors: computational sociology and agent-based modeling," Annual review of sociology, vol. 28, no. 1, pp. 143 -- 166, 2002. [10] G. N. Gilbert, Computational social science. Sage, 2010, vol. 21. [11] H. Wallach, "Computational social science," Computational Social Sci- ence, p. 307, 2016. [12] R. Conte, N. Gilbert, G. Bonelli, C. Cioffi-Revilla, G. Deffuant, J. Kertesz, V. Loreto, S. Moat, J.-P. Nadal, A. Sanchez et al., "Manifesto of computational social science," European Physical Journal-Special Topics, vol. 214, pp. p -- 325, 2012. [13] C. Cioffi-Revilla, "Computation and social science," in Introduction to computational social science. Springer, 2017, pp. 35 -- 102. [14] H. A. Simon, The sciences of the artificial. MIT press, 1996. [15] N. Gilbert and K. Troitzsch, Simulation for the social scientist. McGraw-Hill Education (UK), 2005. [16] D. Easley and J. Kleinberg, Networks, crowds, and markets: Reasoning about a highly connected world. Cambridge University Press, 2010. [17] J. Hwang and R. J. Sampson, "Divergent pathways of gentrification racial inequality and the social order of renewal in chicago neighbor- hoods," American Sociological Review, vol. 79, no. 4, pp. 726 -- 751, 2014. [18] P. Bayer, R. McMillan, A. Murphy, and C. Timmins, "A dynamic model of demand for houses and neighborhoods," Econometrica, vol. 84, no. 3, pp. 893 -- 942, 2016. [19] D. C. Parker, S. M. Manson, M. A. Janssen, M. J. Hoffmann, and P. Deadman, "Multi-agent systems for the simulation of land-use and land-cover change: a review," Annals of the association of American Geographers, vol. 93, no. 2, pp. 314 -- 337, 2003. [20] D. S. Massey and N. A. Denton, American apartheid: Segregation and the making of the underclass. Harvard University Press, 1993. [21] J. V. Henderson, Economic theory and the cities. Academic Press, 2014. [22] C. Castellano, S. Fortunato, and V. Loreto, "Statistical physics of social dynamics," Reviews of modern physics, vol. 81, no. 2, p. 591, 2009. [23] G. A. Akerlof, "A theory of social custom, of which unemployment may be one consequence," The quarterly journal of economics, vol. 94, no. 4, pp. 749 -- 775, 1980. [24] J. Crane, "The epidemic theory of ghettos and neighborhood effects on dropping out and teenage childbearing," American journal of Sociology, vol. 96, no. 5, pp. 1226 -- 1259, 1991. [25] N. S. Miller and M. B. Brewer, Groups in contact: The psychology of desegregation. Academic Press, 2013. [26] J. S. Denis, "Contact theory in a small-town settler-colonial context: The reproduction of laissez-faire racism in indigenous-white canadian relations," American Sociological Review, vol. 80, no. 1, pp. 218 -- 242, 2015. [27] Z. Shaffer, T. Sasaki, B. Haney, M. Janssen, S. C. Pratt, and J. H. Fewell, "The foundresss dilemma: group selection for cooperation among queens of the harvester ant, pogonomyrmex californicus," Scientific Reports, vol. 6, 2016. [28] J. Liu and K. C. Tsui, "Toward nature-inspired computing," Communi- cations of the ACM, vol. 49, no. 10, pp. 59 -- 64, 2006. [29] M. Dorigo, "Optimization, learning and natural algorithms," Ph. D. Thesis, Politecnico di Milano, Italy, 1992. [30] J. Kennedy, "Particle swarm optimization," in Encyclopedia of machine learning. Springer, 2011, pp. 760 -- 766. [31] I. Fister Jr, X.-S. Yang, I. Fister, J. Brest, and D. Fister, "A brief review of nature-inspired algorithms for optimization," arXiv preprint arXiv:1307.4186, 2013. [32] X.-S. Yang, Nature-inspired metaheuristic algorithms. Luniver press, 2010. [33] -- -- , "A new metaheuristic bat-inspired algorithm," in Nature inspired cooperative strategies for optimization (NICSO 2010). Springer, 2010, pp. 65 -- 74. [34] X.-S. Yang and S. Deb, "Cuckoo search via l´evy flights," in Nature & Biologically Inspired Computing, 2009. NaBIC 2009. World Congress on. IEEE, 2009, pp. 210 -- 214. [35] R. Axelrod, "An evolutionary approach to norms," American political science review, vol. 80, no. 4, pp. 1095 -- 1111, 1986. [36] E. Chattoe-Brown et al., "Just how (un) realistic are evolutionary algorithms as representations of social processes?" Journal of Artificial Societies and Social Simulation, vol. 1, no. 3, 1998. [37] K. D. Jong, "Evolutionary computation," Wiley Interdisciplinary Re- views: Computational Statistics, vol. 1, no. 1, pp. 52 -- 56, 2009. [38] C. Cioffi-Revilla, K. De Jong, and J. K. Bassett, "Evolutionary compu- tation and agent-based modeling: biologically-inspired approaches for understanding complex social systems," Computational and Mathemat- ical Organization Theory, vol. 18, no. 3, pp. 356 -- 373, 2012. [39] S. A. West, A. S. Griffin, and A. Gardner, "Social semantics: altruism, cooperation, mutualism, strong reciprocity and group selection," Journal of evolutionary biology, vol. 20, no. 2, pp. 415 -- 432, 2007. [40] R. L. Axtell, "Non-cooperative dynamics of multi-agent teams," in Proceedings of the first international joint conference on Autonomous agents and multiagent systems: part 3. ACM, 2002, pp. 1082 -- 1089. [41] W. A. Clark, "Residential preferences and neighborhood racial segrega- tion: A test of the schelling segregation model," Demography, vol. 28, no. 1, pp. 1 -- 19, 1991. [42] S. Grauwin, F. Goffette-Nagot, and P. Jensen, "Dynamic models of residential segregation: An analytical solution," Journal of Public Eco- nomics, vol. 96, no. 1, pp. 124 -- 141, 2012. [43] V. Cortez, P. Medina, E. Goles, R. Zarama, S. Rica et al., "Attractors, statistics and fluctuations of the dynamics of the schellings model for social segregation," Eur. Phys. J. B, vol. 88, p. 25, 2015. [44] R. Pancs and N. J. Vriend, "Schelling's spatial proximity model of segregation revisited," Journal of Public Economics, vol. 91, no. 1, pp. 1 -- 24, 2007. [45] J. Zhang, "Tipping and residential segregation: a unified schelling model," Journal of Regional Science, vol. 51, no. 1, pp. 167 -- 193, 2011. [46] P. Smets, N. Kreuk et al., "Together or separate in the neighbourhood? contacts between natives and turks in amsterdam," The Open Urban Studies Journal, vol. 1, pp. 35 -- 47, 2008. [47] M. Hewstone and A. Voci, "Diversity and integration: The role of inter- group contact in process of prejudice reduction and conflict resolution," Psicologia sociale, vol. 4, no. 1, pp. 9 -- 28, 2009. [48] S. H. Bartz and B. Holldobler, "Colony founding in myrmecocys- tus mimicus wheeler (hymenoptera: Formicidae) and the evolution of foundress associations," Behavioral ecology and sociobiology, vol. 10, no. 2, pp. 137 -- 147, 1982. [49] T. Kretz and M. Schreckenberg, "Moore and more and symmetry," in Springer, 2007, pp. 297 -- Pedestrian and evacuation dynamics 2005. 308. [50] S. Tisue and U. Wilensky, "Netlogo: A simple environment for modeling complexity," in International conference on complex systems, vol. 21. Boston, MA, 2004, pp. 16 -- 21. PLACE PHOTO HERE Kashif Zia Kashif Zia is currently working as Assistant Professor at Sohar University, Oman. He obtained his PhD from the Institute for Pervasive Computing, Johannes Kelper University in Linz, Austria. Zias research interests revolve around socio- technical systems and computational social science with particular focus on crowds dynamics and smart cities. His PhD work was related to large- scale agent-based modeling and analysis of socio- technical systems utilizing parallel and distributed simulation. JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2017 12 PLACE PHOTO HERE PLACE PHOTO HERE Dinesh Kumar Saini Dr. Dinesh Kumar Saini re- ceived a PhD degree in 2005. Dr. Saini hold the rank of an Associate Professor at Sohar University, Oman and an Adjunct Associate and Research Fel- low with the University of Queensland, Brisbane Australia. Research interest in Software Engineering, Cyber Defense and Mathematical Modeling, Content Management Systems, and Computational Social Science. Arshad Muhammad Arshad Muhammad received a PhD degree in 2010 Computer Science from Liv- erpool John Moores University, UK. From 2010 till 2013, worked as Assistant Professor with City Uni- versity, Pakistan. In 2013, he joined Sohar Univer- sity, Oman as Assistant Professor. He has more than 8 years of experience in research and academics. His areas of expertise are Peer-to-Peer networks, Networked Appliances, Agent-based modeling & simulation, IoT, and Service-Oriented Architectures. PLACE PHOTO HERE Alois Ferscha Alois Ferscha received a Mag. degree in 1984, and a PhD in business informatics in 1990, both from the University of Vienna, Austria. From 1986 through 2000 he was with the Department of Applied Computer Science at the University of Vienna at the levels of Assistant and Associate Professor. In 2000 he joined the University of Linz as full Professor where he is now head of the depart- ment for Pervasive Computing and the speaker of the JKU Pervasive Computing Initiative. Currently he is focused on pervasive and ubiquitous com- puting, networked embedded systems, embedded software systems, wireless communication, multiuser cooperation, distributed interaction and distributed interactive simulation.
1808.01169	1	1808	2018-08-03T12:03:12	Cities of the Future: Employing Wireless Sensor Networks for Efficient Decision Making in Complex Environments	[ "cs.MA", "cs.NI" ]	Decision making in large scale urban environments is critical for many applications involving continuous distribution of resources and utilization of infrastructure, such as ambient lighting control and traffic management. Traditional decision making methods involve extensive human participation, are expensive, and inefficient and unreliable for hard-to-predict situations. Modern technology, including ubiquitous data collection though sensors, automated analysis and prognosis, and online optimization, offers new capabilities for developing flexible, autonomous, scalable, efficient, and predictable control methods. This paper presents a new decision making concept in which a hierarchy of semantically more abstract models are utilized to perform online scalable and predictable control. The lower semantic levels perform localized decisions based on sampled data from the environment, while the higher semantic levels provide more global, time invariant results based on aggregated data from the lower levels. There is a continuous feedback between the levels of the semantic hierarchy, in which the upper levels set performance guaranteeing constraints for the lower levels, while the lower levels indicate whether these constraints are feasible or not. Even though the semantic hierarchy is not tied to a particular set of description models, the paper illustrates a hierarchy used for traffic management applications and composed of Finite State Machines, Conditional Task Graphs, Markov Decision Processes, and functional graphs. The paper also summarizes some of the main research problems that must be addressed as part of the proposed concept	cs.MA	cs	University at Stony Brook CEAS Technical Report Nr 831 Cities of the Future: Employing Wireless Sensor Networks for Efficient Decision Making in Complex Environments Alex Doboli, Daniel Curiac†, Dan Pescaru‡, Simona Doboli, Wendy Tang, Costantin Volosencu†, Michael Gilberti, Ovidiu Banias†, Codruta Istin‡ Department of Electrical and Computer Engineering State University of New York at Stony Brook, Stony Brook, NY 11794-2350 † Department of Automatics and Technical Informatics "Politehnica" University Timisoara, Timisoara, Romania ‡ Department of Computer Science "Politehnica" University Timisoara, Timisoara, Romania Department of Computer Science Hofstra University, Hempstead, NY 11549 April 1 2008 Cities of the Future: Employing Wireless Sensor Networks for Efficient Decision Making in Complex Environments Alex Doboli, Daniel Curiac†, Dan Pescaru‡, Simona Doboli, Wendy Tang, Costantin Volosencu†, Michael Gilberti, Ovidiu Banias†, Codruta Istin‡ Department of Electrical and Computer Engineering State University of New York at Stony Brook, Stony Brook, NY 11794-2350 Email: [email protected] † Department of Automatics and Technical Informatics "Politehnica" University Timisoara, Timisoara, Romania ‡ Department of Computer Science "Politehnica" University Timisoara, Timisoara, Romania Department of Computer Science Hofstra University, Hempstead, NY 11549 Abstract Decision making in large scale urban environments is critical for many applications involving continuous distribution of resources and utilization of infrastructure, such as ambient lighting control and traffic management. Traditional decision making methods involve extensive human participation, are expensive, and inefficient and unreliable for hard-to-predict situations. Modern technology, including ubiquitous data collection though sensors, automated analysis and prognosis, and online optimization, offers new capabilities for developing flexible, autonomous, scalable, efficient, and predictable control methods. This paper presents a new decision making concept in which a hierarchy of semantically more abstract models are utilized to perform online scalable and predictable control. The lower semantic levels perform localized decisions based on sampled data from the environment, while the higher semantic levels provide more global, time invariant results based on aggregated data from the lower levels. There is a continuous feedback between the levels of the semantic hierarchy, in which the upper levels set performance guaranteeing constraints for the lower levels, while the lower levels indicate whether these constraints are feasible or not. Even though the semantic hierarchy is not tied to a particular set of description models, the paper illustrates a hierarchy used for traffic management applications and composed of Finite State Machines, Conditional Task Graphs, Markov Decision Processes, and functional graphs. The paper also summarizes some of the main research problems that must be addressed as part of the proposed concept. 1. Introduction From a functional point of view, cities are complex systems in which resources and infrastructure are continuously distributed to activities of the local economy and inhabitants [CACCIT, NAE, Dineen (2000), Dion and Yagar (1996), Fehin (2004), Felici et al (2006), Greene (2007), Hull et al (2006), Korkmaz et al (2004), Licalzi and O'Connell (2005), Murty et al (2007)]. The distributed resources are electricity, water, heating gas, parking space, etc., but also commodities like dining services, lodging, hospitals, entertainment facilities, and so on. The city infrastructure includes roads, sewage, water pipes, power grid, etc., in general, any means that enables the continuous transferring of resources to satisfy emerging needs. In this context, a city must have sufficient decision making capabilities for efficiently and continuously allocating resources through the infrastructure to satisfy demands. Efficiency is defined with respect to many metrics like monetary cost, quality of services, response time, safety, overhead (related losses), etc. The quality of decision making represents a trade-off between the effectiveness and comprehensiveness of the decisions, and the related costs. Global decisions can be more efficient as they involve detailed analysis of large amounts of data. However, they are time and money-wise the most expensive as they require lengthy procedures executed by large personnel. In contrast, local decisions are much cheaper but they rely on small amounts of data, and use much simpler decision making procedures. In traditional urban set-ups, decision making remains arguably a lower efficiency process, which rarely relies on state-of-the-art technology for rapidly identifying the best options for a given situation. It has been recently suggested that modern technology, including ubiquitous data collection through sensors, automated analysis and prognosis, and online optimization can offer intriguing new capabilities for effective decision making in large urban environments while keeping the related costs low [Campbell et al (2006), Chakrabarty et al (2003), De et al (2005), Dubois et al (2003), Eisenman et al (2007), Greene (2007), Hull et al (2006), Lee et al (2006), Jain et al (2004), LeBrun et al (2005), Lee et al (2006), Murty et al (2007), Yoneki (2005)]. Popular applications like, large scale lighting control and traffic management, suggest that the main characteristics of decision making procedures include (i) flexibility, (ii) autonomy, (iii) scalability, (iv) efficiency, and (v) predictability. (i) Flexibility is the capability of a deployed system to address situations that are not predicted or addressed off-line. This is important due to the large variety of situation that occur in real-life. (ii) Autonomy refers to the capacity of a system to operate with minimum human intervention. This is critical as the cost of servicing decision making systems by humans might prohibit their deployment in large cities. Moreover, having humans tightly "embedded" in the decision making loop also increases the response time and can reduce the reliability due to inevitable human errors. (iii) Scalability characterizes the effectiveness of decision making methods with increasing amounts of demands, resources, and infrastructure. This is crucial for urban set-ups, which involve huge numbers of correlated activities, demands, resources, and infrastructure. (iv) For on-line and autonomous decision making systems, efficiency represents the quality of the decisions automatically selected during operation (online) as compared to the quality of offline decisions based on the entire state information being available. High efficiency is vital in real-life applications. (v) Predictability is the ability of devising automated procedures that estimate accurately and in real-time the quality of decision making alternatives. Predictability is a key enabling component for any decision making algorithm. In addition to the five criteria, another requirement refers to the capability of making decisions that comprehensively tackle different applications that are correlated or might become correlated in certain conditions. For example, it is likely that the illumination intensity control along roads must be correlated to the traffic level on the illuminated roads. Hence, the control procedures for street lighting control have to interact with the decision making procedures for traffic management, such as traffic signal control. This interdependency can be analyzed off-line and incorporated into the system before operation provided that there is a common model and specification formalism for the two separate decision making applications. A more complex situation occurs if correlations dynamically emerge only in certain conditions, which are hard to predict a-priori. Then, the decision making system ought to recognize any newly created dependencies, and provide in real-time a strategy for their common resolution. Traditional decision making systems incorporate mostly only static procedures, and thus are likely to fail in unpredicted situations. For example, the control of the heating gas pipes is often uncorrelated to the traffic flow. However, if gas leakages occur then the lighting intensity should be diminished to reduce the risk of explosions, which further requires the adjustment of the traffic signals, so that the traffic flow through that zone decreases. The identification of all correlations that can occur, while conceptually possible, is in practice infeasible due to the many situations and conditions that are possible in large cities. Moreover, it unnecessarily complicates the control procedures considering that many of the possible situations might actually not occur in practice. A number of recent research projects have focused on the topic of using state-of-the-art computing and communication technology, including wireless sensor networks, for various applications in large urban environments. The tackled applications include public transportation [Yoneki (2005)], traffic monitoring [Hull et al (2006), Lee et al (2006)], traffic light control [Cunningham et al], social networking and VoIP [Murty et al (2007)], data gathering [Campbell et al (2006)], communication [Kansal et al (2004)], just to mention a few. The present research focuses on efficient and comprehensive data acquisition, wireless routing protocols, programming paradigms, embedded hardware, and security. An intriguing opportunity is to device mechanisms that allow integrating and co-optimizing the correlated applications rather than the more traditional method of tackling them independently. This not only expands the optimization space, but also increases the reliability of decision making in new situations. Moreover, it is important to have a decision making approach that is simultaneously flexible, autonomous, scalable, efficient, and predictable. This is challenging as the performance of flexible and autonomous systems is hard to predict in the general case. Moreover, large systems are difficult to optimize through on-line decision making procedures. This paper proposes a new decision making paradigm for optimizing the continuous, real-time allocation of resources to satisfy demands in large urban environments. While the proposed paradigm is general, the paper refers to street lighting and traffic control as two illustrating case studies. The paradigm is structured as a semantic hierarchy in which different decision making models and strategies coexist and interact to produce flexible, autonomous, scalable, efficient, and predictable decisions. The hierarchal structure is scalable as only a reasonably large number of modules collaborate at each semantic level for reaching global decisions. Flexibility and autonomy is achieved by having reactive models at the bottom of the decision making hierarchy. Hence, unexpected sequences of events can be accommodated during operation. Efficiency and predictability results from having the reactive behavior constrained by the upper semantic levels, which use more deterministic models, such as data flow graphs, and Markov Decision Processes [Feinberg (1994), Feinberg (2002)]. The upper levels compute the limits within which the lower, reactive decision making modules must operate, so that the overall goals and constraints of the application are not violated. The lower semantic levels are in constant interaction with the environment and acquire data about real situations. This data is aggregated and propagated to the upper semantic levels, where it is employed for global decision making. Thus, the paradigm incorporates feedback mechanisms between the semantic levels, so that the upper levels constrain the lower levels while the latter provide feedback on the effectiveness of the constraints. Finally, the paper also introduces the defining elements of a specification formalism for integrating different but related applications. Various kinds of interactions between decision modules, such as collaboration, competition, guidance, and enabling interactions, are defined based on the goals, capabilities, inputs, and outputs of each module. This represents a uniform description style for different applications. While the proposed decision making paradigm is applicable to other theoretical formalisms too, this paper refers to a semantic hierarchy based on Finite State Machines (FSMs), Conditional Task Graphs (CTGs), Continuous Time Markov Decision Processes (CTMDPs), and Functional Graphs (FGs). The hierarchy is exemplified for coordinated traffic signal control, a main application in modern cities [Gartner (1983), Gartner et al (1992), Wahle and Schreckenberg (2001)]. It implements quantitative decision making in which decisions at successively higher semantic levels are used to cover increasingly broader geographical areas: (i) FSMs implement reactive control guided by signals coming from traffic sensors and cameras as well as neighboring FSMs, (ii) CTGs schedule the related activities over time for different traffic conditions, thus conduct optimization of signal sequences, (iii) CTMDP use macroscopic descriptions of the system to conduct scenarios-specific optimizations (hence, time is abstracted away from decision making), and (iv) FGs regulate the global allocation of resources and infrastructure to demands, hence taking out the impact of scenarios. The generality of the paradigm is motivated by summarizing a qualitative decision making process, in which the semantic hierarchy uses FSMs, fuzzy logic, and expert systems for decision making. The paper has the following structure. Section 2 introduces two applications (city lighting control and traffic management) to illustrate the importance of infrastructure management in cities. Section 3 defines the main characteristics of decision making in large urban environments, and Section 4 presents the proposed decision making approach. Section 5 introduces the specification model for integrated decision making, and Section 6 enumerates some of the main research issues related to the proposed concept. Finally, conclusions are offered. 2. Applications A myriad of applications can use the more efficient decision making capabilities that are enabled through ubiquitous data collection, automated prognosis, and intelligent advising. Applications differ depending on the geographical localization of decision making, and the kind of human interaction involved in the process. The first criterion defines not only the "globality" of the resulting decisions, but also it quantizes the hardness of the process, as selecting optimal decisions for smaller, geographically localized areas is arguably simpler than that for large, distributed areas. The second criterion centers on the human perspective to distinguish situations where humans are only passive participants to the process, and situations where humans participate to the decision making loop. In the first case, decision making actually results in an actuation activity, which produces a change in the way resources and infrastructure is used. In the latter case, the activity results in additional knowledge, which humans might use for taking more educated decisions. Two applications are discussed next: ambient lighting control and traffic management. A. Ambient lighting control Figure 1: Ambient lighting control system Ambient lighting control applications must automatically adjust the lighting intensity in public spaces so that optimal illumination is offered while minimizing the energy consumption and maximizing reliable operation. These systems must operate with no or little human intervention. They automatically turn on, off, and adjust the illumination intensity in public areas, streets, buildings, parks, and so on. The illumination intensity can be adjusted depending on the time of the day (e.g., dawn, dusk, etc.), weather conditions (sunny, cloudy, full moon), presence of obstacles, number of pedestrians, the nature of the activities carried on (e.g., walking, bicycle riding, etc.), and more. Other capabilities could be related to minimizing energy costs and improving safety, such as by switching to different energy providers during power outages, or selecting the providers that offer the cheapest energy. Similar applications include automated monitoring of the air quality and measuring the noise levels [Santini and Vitaletti (2007)]. Such applications not only reduce living cost and save energy, but also offer better quality-of-life in a city. The next paragraphs offer more insight into an ambient lighting control system (ALCS) that attempts to offer optimal illumination for a given energy budget. Figure 1 presents the ALCS structure. Optimal illumination is defined as a certain illumination intensity at the street level. Each lamp post is equipped with local sensors (e.g., luxmeters, motion detectors, orientation sensors, distance sensors, and other traffic related sensors), and a local controller to decide the illumination intensity. All local controllers of the lamp posts in a small zone (e.g., a street) are connected into a network to coordinate their decisions. Moreover, zones are linked together through wireless connections, and Lighting Coordination Unit (LCU) coordinates the different zones, if more global decisions ought to be made, such as to illuminate during events, holidays, etc. Moreover, if the energy cost budget is fixed, LCU divides the budget to the related sub-areas, so that the overall consumption stays within the budget. Sub-areas can renegotiate their assigned budget. LCUs also collect data for reports and statistics, and identify behavior patterns for different situations. To provide scalability and flexibility, ALCS implements a two-level control hierarchy, as in Figure 1: • Local controller: The controller decides the illumination intensity of a lamp depending on the conditions sensed through the sensors, and the interactions with the controllers of the neighboring lamp posts. For example, if one of the lamps ceases to operate, the illumination intensity of its neighbors is increased, so that they can cover the entire area. Similarly, lamps might be turned off to save energy, if the illumination level exceeds the needed values. Decision making implements a multi-mode control with specific control laws operating for each of the functioning modes. In a simpler case, decision making represents a reactive behavior in which the exceeding of pre-set threshold values (e.g., a neighboring lamp going off) results in switching to a new state in the decision making controller. For each of the states, the inputs from the sensors are used to control a simple "bang-bang" control algorithm, in which the lighting intensity increases if the sensed value is less than the optimal value, and decreases otherwise. Other control algorithms like PID control are also attractive. A complementary approach to local control implementation is using a fuzzy logic [Driankov et al (1996), Farinwata et al (2000), Passino and Yurkowich (1998)]. The fuzzy logic controller assesses qualitatively the inputs coming from sensors and neighboring controllers, and then selects the actuation outputs using the inputs in a set of predefined rules (which define the fuzzy logic controller). For example, if all inputs are large then a specific response is produced. Similar reactions correspond to all inputs being small, and so on for all qualitative combinations of input values. This approach can handle well the imprecision of the sensors outputs, as well as cases when the lighting intensity must be adjusted for hard-to-predict values, like the traffic volume. • Lighting Coordination Unit (LCU): The upper-level controller takes more long term decisions that relate to broader areas. The traditional approach implements a centralized controller for LCU. The decision making algorithm uses all requests from the zones to compute the energy resources allocated to each zone using optimization methods, like Linear Programming or heuristic techniques [Reeves et al (1993)]. While well understood, centralized control does not scale well for large systems as a large overhead is involved. Instead, decentralized control involves only a small number of neighboring zones, which decide resource allocation based on greedy minimization (maximization) of priorities [Sinopoli et al (2003)]. LCU also maintains connection with all related decision making systems within the city limit (like traffic management, disaster management, etc.), to responds efficiently to unscheduled events, e.g., traffic jams, high levels of environmental pollution, power breakdowns, and so on. More details on connecting the different decision making systems are given in Section 5. B. Traffic management Improving the efficiency of traffic is another important application of intelligent decision making in urban environments [Cunningham et al, Yoneki (2005), CALCCIT]. Poor traffic management results in long term waiting times, increased gas consumption, and pollution. Efficiency parameters might include the number of vehicles serviced in a given time interval, the time required to cover a certain trajectory, the pollution and noise levels due to traffic, the maximum number of vehicles that can be serviced without vehicle queuing as well as parameters that define special situations such as offering the shortest travel time to police cars, ambulances, fire-trucks, etc. Tasks related to traffic management include finding the optimal route for vehicles, identifying and warning about traffic "hot spots", re-routing in case of accidents and other unforeseen obstacles, redirecting traffic to avoid high pollution levels in a zone, and more. Decision making for traffic management must optimize the related goals (e.g., efficient traffic, reduced gas consumption, minimal noise levels, etc.) through either centralized or localized strategies [Cunningham et al]. Centralized algorithms express global goals and evaluate decision on a central processing unit, which also collects the data sampled by the individual sensing units. Centralized control, the traditional approach, offers good results for predictable traffic flows, however, its main limitations include low scalability and poor reliability. Localized algorithms rely on a small number of neighboring units for taking decisions, which only affect a small area. In addition, they include a mechanism for percolating the local interactions to the global area, so that large scale decision making is also possible. This approach is potentially more flexible in handling less predictable traffic situations, scalable, and reliable, nevertheless, its main challenge is the difficulty to correctly estimate global characteristics about functioning and performance. Similar to the previous example, the decision making approach proposed for traffic management is hierarchical, in which the lower control levels are localized methods and the upper levels are more global. The local controllers handle the street-level traffic in contrast to the global-level procedures, which identify strategies at the area level to guarantee the goals and performance constraints of the application. Compared to street lighting control, the decision making hierarchy is in this case higher as traffic management involves a wider geographical scale and operates in more diverse situations. Figure 2: Traffic management decison making system (Acknowledgement: The two sub-figures representing traffic situations are due to http://www.permisulauto.ro) Figure 2 depicts a three-level hierarchical system, even though more decision making levels can be added: • Intersection Traffic Unit (ITU): ITU is the controller of the traffic light that supervises each street intersection. ITU receives inputs from sensors, such as inductive loops embedded in the pavement, video cameras, etc., on the traffic intensity in each direction, the length of the vehicle queues, the time taken for a vehicle to move through the intersection, and more [Cunningham et al]. The ITU behavior can be extended with actions for preferred "vehicles", like bicycles or pedestrians, which ought to have higher priority to pass through the intersection. ITU might also include a "default" behavior which is executed whenever the controller fails to converge on a predictable operation. • Zone Traffic Coordination Unit (ZTCU): ZTCU coordinates the operation of ITUs in neighboring intersections to optimize the traffic through each of them by coordinating the decisions of the individual ITUs, redirecting the traffic flow, etc. ZTCU receives data from the participating ITUs, such as traffic characteristics, vehicle queues, moving obstacles, and so on. It might also receive inputs from the upper most Traffic Coordination Unit (see next item) about the targeted traffic efficiency goal, like avoiding long queues and minimizing the delay of vehicles. ZTCU computes optimal strategies for different traffic situations (patterns) in the intersections. • Traffic coordination unit (TCU): TCU operates at the global level by setting the parameters of the ZTCUs. Deterministic policies are used for decision making at this level. TCU also adapts the traffic flow to incorporate out-of-the-order, global situations, such as during the passing of a rapid intervention unit or fire truck that must traverse quickly through a large area. Sections 3 and 4 give more details on the methods proposed for each decision making level. 3. Proposed Decision Making Concept Decision making in complex urban environments must offer the following main capabilities: • Flexibility: During operation, the system must be flexible enough to accommodate new situations, objectives, and requirements, which are hard to predict off-line [Saleh et al (2002)]. • Scalability: Scalability is the property of the control algorithms to perform well for increasing numbers of service requests, resources, and infrastructure. There are many reports that indicate that decentralized methods, which use only information localized in time and space, handle better complex systems than centralized methods [Sinopoli et al (2003)]. • Autonomy: Decision making should minimize human intervention. This is important not only for the scalability requirement of decision making but also because of the higher costs and lesser, if human intervention is involved. • Efficiency: To be viable in real-life, decision making must be effective with respect to metrics, like short service time, high service quality, low costs, good reliability and safety, maximum number of satisfied services, and many more. Moreover, for flexible and autonomous decision making, efficiency also captures that quality of adapted, online decision making as compared to offline decision methods in situations when all state information being available. • Predictable: For autonomous and efficient operation, the characteristics of decision making ought to be predictable. For example, the resulting service time and costs of local decision making strategies should be predictable at a global level and over time. Also, control techniques should not lead to deadlock, thus the set goals should be reachable by the local strategy. Requirement 1 Feasible Design Space Current Operation Point Requirement 2 P1,U P1,L P2,L P2,U Figure 3: Definition of flexibility The five properties can be measured using the following metrics inspired by the similar metrics used in manufacturing engineering and parallel systems: • Flexibility: In [Kazmer et al (2003)], flexibility (in engineering design) is defined as the ratio of the feasible design space over the range of the specification. A similar definition can be used for decision making in urban environments. Figure 3 illustrates the definition for two performance requirements. The figure assumes that a specification range [Pi,L, Pi,U] is defined for the requirements P1 and P2. Moreover, some of the requirement values cannot be realized in practice, thus, a space of feasible design points is defined around the current functioning point. A larger feasible design space determines a higher flexibility of the system. The definition can be easily extended for more requirements. Scalability: According to [Jogalekar and Woodside (2000)], the scalability of a distributed systems with respect to attribute P reflects the amount of extra resources that ought to be spent for improving P's value from P1 to P2. If we denote Cost1 the resource cost for achieving performance P1, and Cost2 the resource cost for having performance P2, then scalability is the following ratio: • Scalabilit ( P,Py 2 1 ) = P 1 P 2 ⋅ Cost 2 Cost ⋅ 1 • Autonomy: Autonomy depends on the amount and quality of human input required to assure a system's correct functioning under a broad range of changing requirements, which act over a specified geographical area and over an interval of time. The following formula presents a qualitative expression of autonomy: Autonomy = T f ∫ T i ∫ V f V i ∫ PL PU ( Human effort ) dPerforman ce dArea dTime where performance P changes from PU to PL, the geographical area border is defined by Vi and Vf, and the time interval between moments is Ti and Tf. • Efficiency: Flexibility and autonomy are obtained at the expense of extra resources and infrastructure being spent in contrast to the optimal, offline solution. For example, in dusk conditions, an adaptive lighting system might be less efficient than a system specifically designed for such conditions. Nevertheless, it is not feasible to have a dedicated optimal design for each of the conditions that might appear in real-life. For performance P and range [PL,PH], efficiency is the average overhead spent by the adaptive system to achieve each of the performance values in the range: Efficiency ( P,P HL ) = PH ∫ PL ( P addaptive − P sin gle − value ) dP P H − P L • Predictability: Developing flexible and autonomous decision making strategies depends on having techniques that can estimate correctly the feasibility and quality of alternative decisions. Thus, the error between the estimated and the real values must stay within a tolerable limit. Figure 4: (a) The wide-scale decision making model and (b) the local control approach The proposed decision making model attempts to address the five main capabilities enumerated before. Figure 4 illustrates the concept. The approach realizes control through localized data acquisition and collaboration between the interrelated sub-systems. This is critical for having a scalable decision process. Depending on the goals, systems organize themselves into clusters. Each cluster might decide to elect virtual leaders used in collaborating with other clusters. Figure 4(a) shows decision making for three successively larger geographical parts: three local zones (shown in solid line), two larger areas (in dashed lines) and one broad region (in dotted line). Clusters correspond to increasingly higher semantic abstraction levels, each level using different formal representation models. For example, the lowest levels operate in terms of events and basic signals/data and the actions corresponding to them. This is typical to reactive behavior. The intermediate control levels reason in terms of optimizing the execution of multiple tasks over time but without analyzing the physical events and data. Finally, the upper levels consider global and qualitative decision making which are optimal across long period of times (hence, the temporal issues are eliminated from the model). Hence, the control algorithms for clusters covering broad regions remains scalable due to the more abstract information being used during analysis and decision making. Figure 4(b) illustrates the hierarchy of semantic abstraction levels. Each local decision making module has a layered structure to offer a gentle transition from reactive behavior to fully predictable functioning. • The physical layer is mostly concerned with the acquisition of physical signals (such as voltages coming from sensors) and the conversion of these signals into data and events. The physical layer utilizes reconfigurable hardware (e.g., reconfigurable analog and digital circuits) to optimize the efficiency of data acquisition and conversion. For example, faster hardware circuits are used if the speed requirements of data acquisition increase due to faster processes that have to be tracked, or more precise signals processing circuits are needed if the accuracy needs to be increased. • The events/data preparation layer aggregates data and classifies events. Multiple sensors might sense the same data or fragments of the same data. The redundant data can be used to improve the precision of sensing by eliminating noise. Also, comprehensive knowledge is assembled together from the data pieces sampled by each sensor. Event classification is important in situations where different event types are handled by different processing algorithms, like specific responses to different vehicle types (e.g., trucks, bicycles, passenger cars, etc.). • The reactive layer produces responses to the input data and events. • The planning layer controls the execution of tasks, so that the set goals are maximized and also the using of the infrastructure and resources is optimized. The planning layer guarantees that the constraints set for the system operation are met, including the deadlines and energy budgets of the individual tasks. Therefore, the behavior and performance of the planning layer must be fully predictable in contrast to the reactive layer, which is flexible but hard to estimate. • The strategic adaptation layer identifies repeatable patterns which can be used to optimize the functioning of the system with respect to goals, resources, and infrastructure. The identified patterns include sequences of input data, correlations between input data, correlations of events, optimal control sequences for certain conditions, and so on. including real-time constraints, bandwidth Section 4 offers more insight into the decision making components. 4. Efficient Decision Making Efficient and robust behavior in hard-to-predict environments does not only imply that the autonomous system can react to unexpected events and conditions but also is capable to meet all performance requirements, limitations, energy constraints, speed requirements, precision needs, safety, and so on. As we explained in the introductory sections, this is a hard problem, and there are no predictable control algorithms for integrated mixed-domain reactive behavior. We propose a predictable yet flexible decision making concept based on (i) a semantic modeling hierarchy to represent the performance needs and capabilities of a decision strategy at different levels of abstraction, and (ii) a joint top-down and bottom-up constraint transformation mechanism along the semantic hierarchy for guiding the dynamic reconfiguration process and signaling when a set of requirements is insatisfiable for the current set of resources. If the application requirements cannot be met then the autonomous decision making system can decide to shift some of the load to other systems (if they are available), or to trade-off the response quality, such as the response time and accuracy. Tackling the load shift problem without using a centralized control algorithm (which would pose serious scalability limitations) is beyond the scope of this work, but we plan to consider modern techniques for decentralized and local control when researching integrated dynamic reconfiguration, such as the algorithms in [Reynolds (1987), Brogan (1997), Ulam (2007), Brock (1999), Pereira (2003), Hsieh (2007), Sigurd (2006)]. The main features of the two components of the proposed predictive control mechanism are as follows: • Semantic modeling hierarchy: The semantic hierarchy is an abstraction model, which offers a smooth transition from the fully reactive behavior at the node level and the deterministic operation at the application level. The hierarchy includes several abstraction levels, so that each higher level is less flexible in tackling unknown conditions but is more predictable performance-wise. Figure 5 shows a semantic hierarchy with five layers, in which the two bottom layers represent reactive behavior, and the top-most layer offers a performance predictive description of the system as a task graph with data dependencies. • Top-down and bottom-up constraint transformation: A set of top-down and bottom-up constraints enforce the consistency between consecutive layers. The top-down constraints are defined by an upper level for the level immediately below, so that the behavior at the lower level remains within the "bounds" of the behavior at the upper level. The bottom-up constraints are defined from a lower level to the next upper level, and express the amount of performance "violation" (performance needs) that must be considered when re-computing the resource management and control algorithm at the next upper level. The adjustment (propagation) of top-down and bottom- up constraints is performed continuously at runtime. Hierarchical control of robots, a similar though less complex problem, is proposed in recent work. For example, a two level control hierarchy is proposed in [Pereira (2003)], where the lower level is a finite state machine guided by events and the upper level is a decentralized control method between an autonomous mobile system and its immediate neighbors. Local control is based on the potential field method. Similar approaches are discussed in [Ulam (2007), Ulam (2007b), Hsieh (2007)]. The proposed concept differs in several main issues from this work. First, the semantic modeling hierarchy includes more levels, including layers which use stochastic models, conditional task graphs, and task graphs with data dependencies. Second, the proposed hierarchy also aims at expressing end-to-end dynamic reconfiguration of the design and not only reactive functionality, like in the existing approaches. Hence, the proposed models will tackle a wider set of performance requirements, in contrast to existing approaches which mostly consider only convergence to the application goal and sporadically also timing constraints. Third, the predictive mechanism is based on a runtime propagation of top-down and bottom-up constraints and the re-computing of the corresponding control decisions. There is no such adaptation mechanism in the other methods. Existing techniques perform only an off-line static performance analysis, mostly related to the convergence and stability of the used control methods. The two components are discussed in the following two subsections. Function graph level: Probability State 2 Performance constraints Function 1 Probability Performance Probability Function 2 Function 3 Performance Performance Probability Function 4 Performance Markov Process level: p 12 p 21 State1 p 14 p 41 State 4 p 34 p 43 p 23 p 32 State3 Performance violations Performance violations COND Performance constraints Performance violations Performance constraints Conditional task graph level: Task1 Event 1 Event 3 Task 2 Task 3 Event 2 Task 2 Task 3 Reactive behavior level: Event 1 Task1 Event 3 PC Task 3 Task 2 Event 2 Runtime adaptation & reconfiguration Adaptive implementation level: SFG 1 Performance violations PC 1 1C C2 C3 SFG 3 SFG 2 4C && PC 2 Figure 5: Semantic modeling hierarchy for large-scale decision making 4.1 Modeling Hierarchy for Predictive Operation Figure 5 illustrates the suggested semantic modeling hierarchy for predictive decision making. The two bottom levels describe the reactive operation of the decision making system, and the three upper levels provide a smooth transition towards predictable control. The top most level defines the functionality as a task graph with a predictable and evaluatable performance. The five model layers are as follows: • Adaptive implementation level: This layer expresses the behavior of the dynamically modifiable system configuration. This layer defines the reactive adaptation of the implementation to dynamic conditions. Each node is a signal flow graph (SFG) defining a reconfigurable design (analog, hardware, and software). The transitions between SFGs are controlled by external events and other conditions set during operation. This level corresponds to the physical and events/data preparation layer in Figure 4(b). • Reactive behavior level: This layer defines the reactive functionality of a system as FSMs. Each node is a functional task (e.g., trajectory computing, data communication, sensing, etc.). State transitions are controlled by events that are set during operation. Each task is represented as an SFG graph at the adaptive implementation level. • Conditional task graph level: Conditional Task Graphs (CTGs) define various execution traces for the FSM at the reactive behavior level. A certain path of a CTG is executed depending on the runtime values of the conditions (events) labeling the graph arcs. The graph nodes are either individual tasks in the lower level FSM, or FSM portions, which are "lumped" together as a single node to reduce the size of the CTG. The conditions describing different paths are either single events in the FSM, or logic expressions based on the events. • Markov Process level: For each task (function), its execution traces (which is a complete path through the CTG) define a Markov Decision Process. This is suggested by experimental observation that the similar behavior of many computing systems can be approximated as a Markovian Process [Benini (2000), Qiu (2001), Kallakuri and Doboli (2007)]. Each trace is a state. Transition rates between states represent the probability of executing that trace in the CTG. • Function graph level: The application is described at the top-most level as a function graph in which a node is a specific function (task), and arcs define the order of performing the tasks. Nodes are unconnected, if no specific order is required. The performance of each task (e.g., its execution time) is described as a distribution that corresponds to the underlying Markov Decision Process. Please note that the system behavior is predictable at this level because the model can be deterministically evaluated. Area 1 ZTCU ITU TCU Area 2 (ATCU) (a) TCU ATCU ATCU ZTCU ZTCU ITU ITU ITU ITU ITU ITU (b) Figure 6: (a) Simple case study and (b) hierarchical decision making structure The hierarchical decision making in Figure 5 operates as follows. The two bottom most levels employ event driven decision making mechanisms, such as Finite State Machines. The reactive behavior level description switches from one task to another depending on the occurrence of events. For example, the system switches from Task1 to Task2, if Event1 occurs during operation. Each controlled unit executes its own controller. The adaptive implementation level refers to the capability of adapting online the implementation of a certain task, such as switching from a reconfigurable implementation to another under conditions Ci, which relate to implementation parameters, like the available energy, memory, communication bandwidth, etc. of a device. Several individual controllers might decide to collaborate by building collectively a shared description at the conditional task graph (CTG) level. The description defines how the individual controllers use jointly any shared resources to achieve common objectives, such as the access over time (schedule) of vehicles accessing intersections. The CTG description includes decisions specific to different conditions, such as various traffic loads. Please note that CTG decision making is at the level of small local areas rather than individual controllers. Next, the CTG descriptions of the correlated areas are used to build description covering broader areas, such as those at the Markov Decision Process level. Finally, the last level is at the function graph level, and is based on information collected from Markov Process level descriptions for several interrelated areas. This step distributes the overall goals into goals for the individual sub-systems using information from the Markov Decision Process level. The remaining part of the subsection presents the top-down bottom-up constraint transformation for predictive operation. The five consecutive layers in Figure 5 are related with each other through top-down and bottom-up constraints. The top-down constraints reflect the performance requirements that are imposed by the application, and which must be satisfied during operation. The bottom-up constraints (called performance violations in the figure) capture the amount of change in the control algorithms, if the current reactive implementation (at the bottom two levels) cannot operate in unpredictable conditions while the application requirements are satisfied. We will exemplify the concept of top-down and bottom-up constraints by looking at the bottom two levels in Figure 5. Depending on the performance requirements expressed at the reactive behavior level (e.g., the time constraint for task execution, precision, etc.), the control algorithm switches among the SFGs of the adaptive implementation level. If the constraint is not met then the control algorithm switches to the SFG for a more efficient implementation, and vice versa, if the required performance is exceeded. Similar control strategies have been recently proposed in [Lu (2001), Zhang (2002)a], but the methods are for software only and cannot handle mixed-domain constraints. If a performance constraint is violated then the control algorithm might decide to transition immediately to a safe implementation (as shown by the dashed line arc labeled as PC1 in the figure). For example, if the system is running out of time then it might decide to immediately switch to the most efficient state without going through any intermediate states. Similarly, the lower level will provide bottom-up feedback to adjust the transitions of the FSM at the reactive behavior level, in case the imposed performance constraints are violated. Such a "short cut" transition is indicated by the dashed line arc labeled as PC in the figure. Short cut transitions are also introduced by the performance constraints originated from the conditional task graph level. At the conditional task graph level, new nodes and edges are inserted into the graph for providing a predictive behavior. In the figure, the nodes are shaded and arcs are shown with dashed line. If performance constraints are not met then the resource management module might decide to perform a shaded node instead of the node corresponding to the functionality. The missed functionality might be allocated to another node, if there is a node available, or might be entirely skipped, if it does not violate the application requirements. The remaining of the section illustrates two possible instances of the proposed semantic decision making hierarchy: (A) traffic management based on FSM -- Conditional Task Graphs -- Markov Processes -- Function Graph hierarchy, and (B) light control using Fuzzy Logic -- Knowledge Based Rules. While the first hierarchy relies more on quantitative decisions, the second semantic hierarchy is more geared towards knowledge based (qualitative) decisions. 4.2 Traffic Management using FSM-CTGs-Markov Descision Processes-Function Graphs hierarchy Figure 6 shows the simple traffic management case study that has been explored and the corresponding semantic hierarchy. Figure 6(a) presents a street network, in which arcs indicate the flow of traffic. The goal of decision making is to maximize the traffic flow that passes in a given time T through the area. Traffic lights (shown as black bubbles) are positioned at each intersection, and each is controlled by its local ITU (Intersection Traffic Unit, see Section 2). Neighboring ITUs form a ZTCU (Zone Traffic Coordination Unit, see Section 2) that coordinates their operation. Moreover, ZTCUs of an area are interconnected and coordinated by an ATCU (Area Traffic Control Unit). Finally, ATCU are supervised by TCU (Traffic Coordination Unit). Figure 6(b) illustrates the resulting semantic hierarchy. ITUs, ZTCUs, and ATCUs can be statically defined, or can emerge dynamically during operation based on the actual traffic patterns and the best way of coordinating the traffic lights. We have assumed static definitions in this case study. Figure 7 illustrates the two lower levels of the semantic hierarchy: the FSM level and the Conditional Task Graph (CTG) level. The figure also shows the interaction between the two levels. • Each of the ITUs in the figure implements a FSM with three states (red, yellow, and green). Assuming cyclic traffic signals, arcs indicate state transitions which are taken after the FSM spends ∆Ti,j time units in the current state (the time is called split time). The time intervals are specific to each ITU and each transition, and are fixed by the ZTCU for optimizing the traffic flow. The schedule of the traffic signal repeats over a period T (called cycle time). Neighboring ITUs communicate with each other to decide the delay between the monitored intersections (the delay is called offset). • The ZTCU behavior is expressed as a Conditional Task Graph (CTG), which indicates the activities performed by two cars arriving at ITU1 and ITU2 in the figure. Activities include passing through the intersection, moving to the next intersection, and so on. Each of the activities is characterized by an execution time Tex k, which is estimated based on the data collected locally from the sensors at the traffic lights. ZTCUs set the cycle and split times of each intersection, and also find the initial delay values. Execution times differ for different traffic scenarios, e.g., light traffic (indicated as branch L) and heavy traffic (shown as branch T). The semantics of the application states that one and only one of the branches Li and Hi is executed for a traversal of the graph. This semantics defines a conditional (multimode) behavior. ∆ T i,j ZTCU T 0 ITU 1 ITU 2 ITU 3 ITU 1 Red ∆ T ∆ T 2,1 1,1 ∆ T 3,1 Yellow Green ∆ T 4,1 ITU 2 Red ∆ T 1,2 2,2 ∆ T ∆ T 3,2 Yellow ITU 3 Red ∆ T ∆ T 2,3 1,3 ∆ T 3,3 Yellow Green ∆ T 4,2 Green ∆ T 4,3 ex i T T 1 T 7 L T 2 T 3 H T 4 T 5 L T 8 T 9 H T 10 T 11 T 6 T 12 T 13 L T 14 T 15 H T 16 T 17 T18 T 0 T 1 ex T 1 H T 3 ex T 3 H L ex 2T' 2T' ex 4 T T 4 ex T' 4 T' 4 L T 2 ex T2 T10 ∆T i,j T 6 ex T 6 H L Out 9 T 5 ex T 5 H L ex 7 T T 7 ex T' 7 T' 7 ex 8 T T 8 ex T' 8 T' 8 In 3 ITU 1 ex 6 T In 6 ITU 2 ex 3 T In 1 ex 1 T ex 2 ex 4 T T ex 5 T ex 7 T ex 8 T ex 9 T T 9 ex 9 T Figure 7: FSM -- Conditional Task Graph hierarchy Figure 7(b) presents the concrete CTG for a zone with two consecutive intersections controlled by the traffic signals ITU1 and ITU2. The graph presents the activities occurring during one cycle of the entire area, even though the cycle might include several cycles of the individual traffic signals. The process is repetitive after each area cycle. In this example, we assumed single-lane roads, which means that two cars coming through the intersections ITU1 or ITU2 cannot move simultaneously. Hence, certain road sections (shown as dark areas in the figure) act as shared resources for two or more cars, while other roads are dedicated resources. The model defines tasks T1 and T2 as being the travel of a vehicle entering the area from the left, tasks T3 and T4 as representing the moving of a vehicle coming from the top-left road, tasks T6 and T7 represent a car moving from top right, and so on. The resulting CTG is illustrated in the figure. Each task has two attributes: Ni is the number of vehicles that traverses the section during one cycle, and ex is the time required for the vehicles to pass through the section. Obviously, the two attributes are Ti related to each other. In the graph, the tasks sharing the same resources (the road sections are depicted in dark) are marked with dashed and dotted lines, respectively. In addition, the dummy task ∆∆∆∆T12 represents the offset time between the two traffic signals. The ZTCU for the area computes online the optimal scheduling of the tasks Ti while tackling requirements such as, (i) to maximize the total number of vehicles passing though the zone, (ii) to minimize the time taken to cars to move through the zone, (iii) to minimize the fuel consumption in a certain area, (iv) to minimize the amount of polluting gases that are generated, and so on. To avoid continuous recomputing of schedules and to improve the flexibility of the approach, the ZTCU scheduler calculates traffic signal schedules for different traffic conditions, which determine the "behavior" (and thus the mathematical models) for the task attributes, such as parameters Ni and Ti Please note that each computed schedule corresponds to traffic conditions described by conditional behavior and the qualitative parameters Li and Hi in the figure. For example, if the traffic load is low on the segment 1 then the branch labeled as L is taken after the task T1 is completed, hence task T2 is performed. If the traffic load is high then the branch labeled as H is selected, and the task T2' is pursued. Tasks T2 and T2' can never be performed simultaneously as they correspond to mutually excluding ex. conditions (e.g., the traffic intensity cannot simultaneously be low and high). This generates eight different situations for the case study in Figure 7(b), and eight different optimal scheduling situations for the zone, such as the schedule S(L,L,L) with the total time Ttot (L,L,L) for the situation in which the traffic density is light through all three intersections. The actual conditions, which are labeled as L and H, can be different for different traffic lights, and can involve more than two situations. The concrete task scheduling for CTGs can be computed with algorithms similar to the one in [Eles et al (2000)]. Figure 8(a) illustrates the interaction between the local ITUs at each traffic signal and the CTG scheduling at the ZTCU for the interacting ITUs. The scheduling table computed online by the ZTCU is shown in the figure. The table includes the scheduling solutions for different scenarios that might arise in traffic. Please note that the actual semantics of the conditions Li and Hi (e.g., what traffic conditions actually define ex. More them) are refined during execution depending on the variation of the parameters Ni and Ti conditions allow more precise tuning of the schedules, however, the computational complexity of calculating the optimized schedules is also higher [Eles et al (2000)]. The table presents the schedules for two situations: one in which traffic is light along all directions, and the second in which the traffic along direction two is more intense while the others remain low. The traffic schedules are computed so that the total number of cars passing through the zone is maximized while the timing along different directions is kept within fixed deadlines (e.g., to avoid large delays along certain directions due to enabling heavy traffic along other directions). The boxes in gray indicate some of the idle times along a direction. ITU 1: ZTCU: Scenario 2 L 1 L 2 L 3 L 4 . . . L 1 H 2 L 3 L 4 . . . Scenario 1 After ∆ T x Yellow Scenario 1 Red Green After ∆ T y Scenario 2 6T 8T 9T''' T 1 2T 5T' 7T' 9T' 3T T 4 5T'' 7T'' 9T'' Time 6T 8T 9T''' 3T T 4 5T'' 7T'' 9T'' T 1 2T 5T' 7T' 9T' p (x,y,z)−>(u,v,w) ZTCU ATCU Conditional Task Graph level S (L,L,L) T(L,L,L) ... S (H,H,H) T(H,H,H) p (L,L,L)−>(L,L,H) TCU S (L,L,L) S (L,L,H) p (L,L,H)−>(L,L,L) ... Traverse Area 1 T area1 S (L,H,L) S (L,H,H) Traverse Area 2 T area2 T area i Figure 8: FSM - Conditional Task Graph -- Markov Process -- Functional Graph hierarchy The information in the CTG scheduling table is mapped down onto the FSMs of each individual traffic signal. For example, for Scenario 1 (in which traffic is low along all directions), the time required for performing task T3 plus the following idle time generate the timing constraint ∆∆∆∆Tx, which states that after that amount of time the FSM must be in state Green, thus allowing traffic to progress along direction two. Similarly, for Scenario 2, the time for task T1 plus the following idle time define the constraint ∆∆∆∆Ty, which states that the traffic signal ought to be in state Red after ∆∆∆∆Ty, which allows traffic to progress along direction one. As long as each traffic signal meets its local timing constraints (∆∆∆∆Tx and ∆∆∆∆Ty), the behavior at the ZTCU level still meets the traffic schedules specified in the ZTCU table. Please note that the local FSMs can switch more often, such as if no vehicle is traveling in one direction and a single car is approaching from another direction, but for "pools" of vehicles the signaling must follow the timing constraints of the ZTCU table. Figure 8(b) presents the upper levels of the hierarchy, including the Markov Process (ATCU) and the Functional Graph (TCU) levels, and their connection to the lower Conditional Graph level (ZTCU). For an estimated number of vehicles that pass through the area, the ZTCU has the optimal activity schedules for each estimated traffic scenarios. If the timing delay Tarea_i of a ZTCU schedule exceeds the necessary timing constraint set for the area traversal, then the traffic light controllers are adjusted so that the new traffic scenarios has a schedule that meets the constraint. Another action set by the ATCU is to distribute the traffic flow among the two areas so that the set timing constraints are met and the total number of cars passing through the area is maximized. This decision making is achieved by setting up a Markov process for the area in which the states correspond to the traffic scenarios of the ZTCU (eight scenarios in our case study). The transition rates between the states are the probabilities p(x,y,z)->(u,v,w) of shifting from one scenario to another. ATCU maintains a macro-level perspective of the traffic flow in which the moving of vehicles appears like a continuum flow without losses (the number of vehicles that enters a zone is equal to the number of vehicles that exists). For the example in Figure 7, In1+In3+In6=Out9 (indexes reflect the corresponding street segment). The following paragraphs introduce briefly the concept of Continuous-Time Markov Decision Processes, and explain how this formalism is used to model the higher levels of the ATCU decision making. A Continuous-Time Markov Decision Process (CTMDP) is defined as the following sixth-tuple: }r,k,q),(A,A,I{ Where I is a finite state space, A is a finite set of actions, A() is the function indicating the actions associated to any state in I, k is the number of criteria, rk(i,a) is the reward rate for criterion k, and q(i,j,a) is the transition rate from state i to state j, if action a is conducted. For ATCU decision making, the different traffic scenarios define the states of a CTMDP, such one state corresponds to low traffic conditions through each intersection of the zone. Hence, one state corresponds to one column in the CTG scheduling table. If the CTMDP corresponds to multiple scheduling tables then each column generates a different state in the table. The action set A represents the alternative routes that are available for moving inside a zone. The action set correlates to the available traffic scenarios and also the traffic ex and Ni. Rewards rk are the number of cars passing through a certain zone in a time equal parameters Ti to the time modeled through the scheduling table of the CTG. Thus, parameter k is one in this case. Decision making is expressed as the following optimization problem: a,ix)a,i(r Maximize ∑∑ ⋅ Such that ∑ x)a,i(q a,j ∑∑ ∑∑ x a,i 0 − ∑∑ = 1 x a,i ≥ x)a,i(r k a,i ⋅ x)a,j,i(q ⋅ = 0 a,i ≥ c k Qualitative Decision Making using Fuzzy Logic. An alternative to the previous approach can be based on hierarchical qualitative decision making, such as fuzzy logic and expert systems. Fuzzy logic computes the control outputs using a set of predefined rules that operate on a qualitative assessment of the inputs. The qualitative assessment (through membership functions) is critical as it decides the meaning of the inputs for the controller. Obviously, in a dynamic system, the meaning might change over time to reflect the insight gained during operation. Hence, the decision making process can be organized as a two level strategy, in which the upper level component uses learned patterns and dependencies to set the characteristics for decision making of the lower level, fuzzy logic controller. Figure 10 shows the block diagram of the two-level hierarchical control system. The fuzzy logic controller ZLCU receives aggregated information from (i) the illumination sensors, aggregated by block AG-il (e.g., the intensity of the ambient light i), and (ii) the traffic sensors, aggregated by block AG-t (e.g., the traffic "density" d). ZLCU computes the command variable u (e.g., for the actuators of the ambient illumination system). The fuzzy controller is supervised by the upper level control system LCU that sets parameter a, which is used by the fuzzy logic controller to qualitatively assess the two inputs. Figure 10: (a) Hierarchical qualitative control scheme and (b) fuzzy logic controller Figure 10(b) presents the block diagram of the fuzzy logic controller ZLCU. The received inputs are fuzzified in the input interface block Fuz. Then, an inference based on the max-mi method is done on the fuzzyfied inputs using rule BR. The output information is defuzzified with the centroid method to produce the output u. The fuzzy variables ai, ad and au can take three qualitative values: small (S), medium (M), and big (B). The parameters of these membership functions on the universes of discourse are indexed with the indexes m for the medium value, M for the maximum, value and Ml for the limited maximum value. These parameters are modified by the high level component LCU. Samples of membership functions are given in Figure 11(a). LCU transmits the following set of values to ZLCU: a =      a im a a dm um a iM a a dM uM a iMl a a dMl uMl      (1) A simple rule base BR of the fuzzy controller is presented in Table 1. Table 1: Simple rule base u i B M S S S S M d M S M B B S M B Figure 4 presents the surface of the fuzzy controller u = f ,( di ) for a particular value set of the parameters (0.5; 1; 1.2) on the universe of discourse [0; 1.2] for all three fuzzy variables. Changes in the membership function parameters and the rule base will change the above surface. Figure 11: (a) Membership functions and (b) surface of the fuzzy controller 5. Specification for Integrated Decision Making The paradigm discussed in Section 4 ought to support interdependent applications of different types, as the global optimization goal might not involve a certain application. For example, increasing the traffic flow during night time is linked to the illumination control as optimal illumination might allow higher vehicle speeds while still maintaining safety. This requires that the decision making (DM) modules across the entire semantic hierarchy utilize a uniform specification model, which is independent of the application type and the control strategy utilized by each individual DM module. The specification model must also allow easy decision making across the DM modules for different applications. The specification model proposed for decision making in large scale urban environments is based on a goal-oriented description paradigm, in which each DM module has well-defined goals. The goals are set by the application, such as providing the required illumination intensity, or maximizing the traffic flow through a region. In addition, each DM module has limited set of capabilities, such as the maximum illumination it can provide, the highest traffic flow on a route, communication distance and bandwidth, local energy resources, and so on. Module inputs are either connected to the sensors or connections from DM modules positioned on the same or higher semantic levels. DM outputs are either actuation and control signals or connections to modules on the same or lower semantic levels. Each module conducts internally a decision making process using models similar to those detailed in Section 4, however these models are not accessible outside a DM module. Figure 12 illustrates the goal-based specification concept. Area Zone Integration Region DM Goals: Capabilities: Inputs: Outputs: Internal DM: Region Area DM Goals: Capabilities: Inputs: Outputs: Internal DM: Local DM Goals: Capabilities: Inputs: Outputs: Internal DM: Local DM Goals: Capabilities: Inputs: Outputs: Internal DM: Local DM Goals: Capabilities: Inputs: Outputs: Internal DM: Region DM Goals: Capabilities: Inputs: Outputs: Internal DM: Area DM Goals: Capabilities: Inputs: Outputs: Internal DM: Figure 12: Uniform specification approach for integrated decision making Different kinds of interactions can be set up between the participating DM modules: • Collaborative interactions: Collaborative interactions are set up between DM modules on the same semantic level and which have non-conflicting (non-competing) goals, such as the individual traffic light controllers which must all offer optimal lighting intensity of an area. The DM modules interact with each other through inputs and outputs, e.g., one DM module produces outputs that serve as inputs to another module. The goals and capabilities of DM modules are not affected by these kinds of interactions. • Competing interactions: Such interactions are set up between DM modules from the same semantic level but with competing goals, like maximizing the illumination of a certain zone but also minimizing the energy consumption of the zone (the latter can be set by the power grid application). Similarly to collaborative interactions, DM modules interact through their inputs and outputs but cannot change their goals or capabilities components. • Guiding interactions: Guided interactions are between DM modules at consecutively higher levels in the semantic hierarchy. The modules at the upper levels generate outputs that are used to set the goals of the modules on the lower levels. This way the first modules are steering the goals and hence the behavior of the latter modules. For example, the DM modules at the area level set constraints, which then become part of the goals section of the related traffic signal controllers. • Enabling interactions: Enabling interactions are information transfers from the lower semantic levels to the upper levels. A lower DM module transmits information about its capabilities to an upper DM module, so that the second uses this knowledge during a decision making process that might affect the goals set for the lower module through guided interactions. Enabling interactions is the information link through which the upper decision making modules acquire knowledge about the actions that are ongoing in the real world. These concepts are illustrated in Figure 13 which shows the interactions and integration of the decision making mechanisms for different infrastructure management subsystems in a city, such as street lighting control, traffic control, pollution monitoring, and disaster management. 6. Future Research Directions Several main research problems emerge related to the proposed decision making approach for large scale urban environments. The identified problems can be grouped into three categories, and are described next. The starting point in addressing the problems represents the developing of more orthogonal applications, like heat, water, gas, and power distribution, fast response in emergency situations, energy harvesting on large areas, environmental monitoring (including noise levels and pollution), social networking, and many more. Developing a suitable testbed for simulating these applications is also very important. A. Exploring the formal model: semantic hierarchy and specification. An important goal is to study the full potential for decision making of the proposed approach based on semantic hierarchy. The paper details a hierarchy based on Finite State machines, Conditional task Graphs, Markov Decision Processes, and Functional Graphs, but other quantitative or qualitative models can be incorporated into the hierarchy too. Some of the formalisms that we plan to research include Petri Nets, Signal Flow Graphs, Timed Automata, Hybrid Automata, Flow based models, linear and nonlinear programming, and likely other too. Also, it is important to understand the interactions between the mathematical formalisms, including techniques for propagating constraints top-down from the more abstract to the more precise levels, and also bottom-up for specifying feasibility constraints at the higher levels. Moreover, it is important to understand how the propagated constraints relate to the mathematical formalisms used in the hierarchy, as the same constraints might be utilized for different specification formalisms. The scalability of the hierarchical model for different formalisms must be characterized too. Finally, the current formalism targets only static sensing nodes but it is important to extend it for mobile nodes too. Figure 13: Integrated decision making in a city Another important issue relates to lay out the mathematical foundation of collaborative interactions between decision making systems. This is important for understanding what other collaboration types are possible, if entities interact though goals, capabilities, inputs, and outputs. Moreover, the specific manifestation of collaborative interactions for different semantic models needs to be addressed too. B. Online decision making algorithms. A set of specific online decision making algorithms ought to be developed for the semantic hierarchy model. This includes algorithms for forming clusters based on application goals, electing virtual leaders for each cluster, and instantiating the virtual hierarchy at the level of each embedded sensing node. This activity also implies identifying the best collaboration model among related decision modules, including finding the correlations that emerge dynamically during operation. Also, the implications of the semantic and goal based operation mode on the underlying wireless and wired communication protocols and routing algorithms needs to be explored too. Other important research problems refer to the actual optimization algorithms that are used for online optimization of the goals expressed in different semantic models. This includes online algorithms for resource allocation and scheduling of data flow graphs and Conditional Task Graphs, and solving equations expressed as Markov Decision Processes or flow models. Moreover, fast algorithms are needed for integrating data from the low levels into constraints for the upper semantic levels, and also for propagating top-down constraints form the upper levels to the constituent decision making entities. These algorithms ought to tackle a variety of trade-offs, like cost, speed, reliability, energy consumption, etc. Also, the importance of these constraints might also change dynamically during execution, and thus decision making should be capable to respond quickly to changes. A possibility involves developing algorithms that can adapt to new requirements by incrementally changing the current solution. Considering that the adapted solutions are not optimal, it is important to understand and estimate the performance loss of an incrementally adapted solution as compared to the optimized solution. C. Performance prediction for reliable operation. A challenging objective is to provide the means for accurately estimating in real-time the performance of alternative decisions for new situations. This problem is essential for devising any efficient and reliable control algorithm. For an alternative, estimations should describe performance attributes, like quality (e.g., illumination intensity for lighting control), response time, precision, energy consumption, cost, etc. The predicted performance should include both continuous valued/time and discrete valued/time performance. Attributes like energy consumption, illumination intensity, response time, and travel time are examples of continuous valued-continuous time performance. Examples of discrete-valued attributes are the number of cars passing through a zone, the bit precision of a decision, etc. Performance estimation should rely at a minimum on simulation, instead macromodeling techniques ought to be developed based on linear and nonlinear interpolation, piecewise linear modeling, neural networks, wavelet functions, and so on. Also, it is important to detect and extract online any new patterns and correlations that emerge during operation, and use this information to further optimize the response of the decision making systems. 7. Current Information Systems for Large Urban Environments A number of recent research projects focus on developing wireless sensor networks for applications in urban environments. The monitoring of the noise level in a city and building the related noise maps is presented in [Santini (2007)]. The TIME (Transport Information Monitoring Environment) project [Yoneli (2005)] targets the improving of the traffic efficiency in Cambridge UK by monitoring and processing information collected in real-time about traffic. The U-City approach [Licalzi (2005)] attempts to connect the information systems in residential, governmental, commercial, and healthcare areas. The CarTel project [Hul (2006)] focuses on distributed mobile sensor networks for monitoring traffic. The approach is based on opportunistic data muling through mobile sensors, and offers delay-tolerant pipes and delay-tolerant stack for data transport. The concept of adaptive traffic-light control through Collaborative Reinforcement Learning is introduced in [(Cunningham)]. Several other approaches are detailed next. The CitySense concept [Murty (2007)] is wireless sensor network based infrastructure for developing application for large urban environments. Possible applications for CitySense include serving web pages, VoIP, social networking, and monitoring the networks deployed in a city. CitySense assumes static deployment of powerful computing platforms equipped with radios using the 802.11 wireless protocol to transmit over a mesh network. Hence, each node has the capabilities to perform locally complex computations and to communicate over multiple channels at high data rates. CitySense also relies on moving vehicles to provide connectivity to the areas which are hard to communicate with through radios. It uses predictions about the trajectory of the vehicle to prepare in advance the data to be communicated to the vehicle. Also this information is used to decide the aggregation points of the collected information. The MobEyes system [Lee (2006)] has been developed for monitoring and data acquisition in urban environments. The system uses moving vehicles such as cars to transport necessary information to/from areas in which wireless communication cannot be offered. The communicated information represents a summary of the entire information harvested by the nodes during operation. The single-hop communication protocol (called MDHP) transmits the missing data summaries to a vehicle that enters the communication range of the sensor. The summaries are stored locally by the sensor or collected from its neighbors. Moving police agents collect the important summaries. The protocol is scalable, and requires acceptable large latency overhead but cannot tackle real-time constraints. The MetroSense approach [Campbell (2006)] also tackles the problem of data gathering in urban settings, and exploits mobile vehicles for improving the covered area. This is very important for keeping the cost of the network down and also to offer connectivity if of remote zones. The authors propose an opportunistic delegation model in which tasks are temporarily re-assigned to available nodes. Similar to MobEyes, vehicles are used to carry information to remote areas. The authors propose four protocols for moving data through vehicles: lazy uploading, lazy tasking, direction-based muling, and adaptive multihop. In lazy uploading, a point-to-point connection is used to transmit data to the vehicle as long as it stays in the zone covered by the radio. The next available vehicle continues the data transfer. A similar strategy is used in lazy tasking to assign tasks (through reprogramming) to new processing devices. Direction-based muling uses predictions on vehicle moving to transfer data to the vehicles that are more likely to move towards the target of the data communication. In adaptive multihop, the characteristics of the communication, such as the task boundary, change dynamically depending on static threshold values, like the load threshold. The concept is exemplified in [Eisenman (2007)] for the practical case of data collection by a network of bicycles to improve the quality of bicycle riding. [Kansal (2004)] propose a networking infrastructure which uses autonomous mobile wireless routers for improving the energy consumption, accuracy, data rate of communication over large areas. The infrastructure also gives connectivity to remote areas. The method computes and adjusts the trajectories for the mobile routers depending on the nature of their dynamics, e.g., random, predictable, and controlled movement. A set of adaptive control procedures are suggested to adjust the speed of the mobile routers depending on the speed and latency of data collection, and the connectivity of the network. In deciding the dynamics, the procedures exploit the correlations between the sensing speed, communication distance, data rate and energy consumption. 8. Conclusions Cities are complex systems in which resources and infrastructure are continuously allocated to economy and inhabitants. A city must have sufficient decision making systems for continuously dispatching resources like water, electricity, heat, traffic, etc. in flexible, autonomous, scalable, efficient, and predictable fashion. This task is challenging as the quality of decision making is a trade-off between effectiveness, comprehensiveness, and related costs. Global decisions can be more efficient, however, they are time consuming and financially expensive. In contrast, local decisions are cheap but offer mostly local decisions. In addition, the integrated nature of modern applications requires control strategies that can comprehensively tackle correlated applications or which might become correlated in certain conditions. Intriguing new control solutions can be devised based on modern technology involving ubiquitous data collection through sensors, automated analysis and prognosis, and online optimization. This enables global decision making over large areas while keeping the costs low. Also, new mechanisms can be devised to integrate and co-optimize the correlated applications. This expands the optimization space, and increases the reliability of decision making in new situations. This paper proposes a new decision making paradigm for optimizing the continuous, real-time allocation of resources to satisfy demands in large cities. The paper refers to street lighting and traffic control as case studies to illustrate the paradigm. The proposed concept is structured as a semantic hierarchy in which different decision making models and strategies coexist and interact to produce flexible, autonomous, scalable, efficient, and predictable decisions. For scalability, only a reasonably large number of modules collaborate with each other to reach global decisions. Flexibility and autonomy results by using reactive models for the bottom decision making levels. Efficiency and predictability is due to the reactive behavior being constrained by the upper semantic levels, which use more deterministic models. The upper levels compute the limits within which reactive decision making must operate, so that the overall goals and constraints are not violated. The data sampled from the environment is aggregated and propagated to the upper semantic levels, where it is used for global decision making. Finally, the paper discusses the defining elements of the specification formalism for integrating different but related applications. Interactions between decision modules, such as collaboration, competition, guidance, and enabling interactions, are defined based on the goals, capabilities, inputs, and outputs of each module. While the proposed decision making paradigm is applicable to other theoretical formalisms too, this paper refers to a semantic hierarchy based on Finite State Machines (FSMs), Conditional Task Graphs (CTGs), Continuous Time Markov Decision Processes (CTMDPs), and Functional Graphs (FGs). The hierarchy is exemplified for coordinated traffic signal control, a main application in modern cities. Decisions at successively higher semantic levels are used to cover increasingly broader geographical areas: FSMs implement reactive control guided by signals coming from sensors as well as neighboring FSMs, CTGs schedule the related activities over time for different traffic conditions, CTMDP use macroscopic descriptions of the system to conduct scenarios-specific optimizations, and FGs regulate the global allocation of resources and infrastructure to global demands. Three main research problems emerge related to the proposed decision making approach. First, it is important to study semantic hierarchies based also on other formalisms like Petri Nets, Signal Flow Graphs, Timed Automata, Hybrid Automata, Flow based models, linear and nonlinear programming, and so on. Moreover, it is needed to devise the theoretical foundations for propagating constraints top-down from the more abstract to the more precise levels, and also bottom-up for specifying feasibility constraints at the higher levels. Second, efficient online decision making algorithms ought to be developed for the semantic hierarchy model, e.g., algorithms for forming goal oriented clusters, electing virtual leaders, and mapping the virtual hierarchy onto each embedded node. Also, new algorithms need to be designed for resource allocation and scheduling at different levels of the model hierarchy while tackling trade-offs, like cost, speed, reliability, energy consumption, etc. As the importance of these constraints might change during execution, the algorithms should incrementally change the current solution to adapt it to the new requirements. Third, procedures for accurate, real-time performance estimation ought to be devised. The tackled performance should include both continuous valued/time and discrete valued/time performance. Moreover, performance estimation should rely mostly on macromodeling, so that it can be performed in real-time. Other tasks relate to detecting and extracting online any emerging patterns and correlations, and using this information to further optimize decision making. References [Albus (2000)] J. Albus, "Features of Intelligence Required by Unmanned Ground Vehicles", NIST Workshop on Performance Metrics for Intelligent Systems, 2000. [Balarin et al (1997)] F. Balarin, L. Lavagno, P. Murthy, A. Sangiovanni-Vincentelli, "Scheduling for Embedded Real-Time Systems", IEEE Design & Test of Computers, Jan-March 1998, pp. 71-82. [Benett et al (1997)] F. Benett, D. Clarke, J. B. Evans, A. Hopper, A. Jones, D. Leask, "PicoNet: Embedded Mobile Networking", IEEE Personal Communication Magazine, 4(5): 8-15, Oct. 1997. [Bonnet et al (2000)] P. Bonnet, J. Gehrke, P. Seshadri, "Querying the Physical World", IEEE Personal Communications, Vol. 7, October 2000, pp. 10-15. [Borgatti et al (2003)] M. Borgatti, F. Lertora, B. Foret, L. Cali, "A Reconfigurable System Featuring Dynamically Extensible Embedded Microprocessor, FPGA, and Customizable I/O", IEEE Journal of Solid- State Circuits, Vol. 38, No. 3, pp. 521-529, March 2003. [Brooks et al (2003)] R. Brooks, P. Ramanathan, A. Sayeed, "Distributed Target Classification and Tracking in Sensor Networks", Proceedings of IEEE, Vol. 91, No. 8, pp. 1163-1171, August 2003. [Buck et al (1994)] J. Buck, S. Ha, E. Lee, D. Messerschmitt, "Ptolemy: A Framework for Simulation and Prototyping Heterogeneous Systems", International Journal of Computer Simulation, 4: 155-182, April 1994. [Byers and Nasser (2000)] J. Byers, G. Nasser, "Utility-Based Decision-Making in Wireless Sensor Networks", First Annual Workshop on Mobile and Ad Hoc Networking and Computing, MobiHoc 2000, 2000, pp. 143-144. in available a Next-Generation Urban Traffic Control Environment", [Campbell et al (2006)] A. Campbell, S. Eisenman, N. Lane, E. Miluzzo, R. Peterson, "People-Centric Urban Sensing", Proc. of 2nd Annual International Workshop on Wireless Internet, 2006. [Chakrabarty et al (2003)] A. Chakrabarty, A. Sabharwal, B. Aazhang, "Using Predictable Observer Mobility for Power Efficient Design of a Sensor Network", Proc. of Second International Workshop on Information processing in Sensor Networks, 2003. [Chong and Kumar (2003)] C.-Y. Chong, S. P. Kumar, "Sensor Networks: Evolution, Opportunities, and Challenges", Proceedings of the IEEE, Vol. 91, No. 8, pp. 1247-1256, August 2003. [Cohen et al (2004)] I. Cohen, M. Goldszmidt, T. Kelly, J. Symons, J. Chase, "Correlating instrumentation to system states: A building block for automated diagnosis and control", Proc. 6th Symposium on Operating Systems Design and Implementation, 2004. [Cunningham et al] R. Cunningham, J. Dowling, A. Harrington, V. Reynolds, R. Meier, V. Cahill, "Self- Optimization at http://www.ercim.org/publication/Ercim_News/enw64/cunningham.html. [De et al (2005)] P. De, A. Raniwala, S. Sharma, T. Chiueh, "MiNT: A Miniaturized Network Testbed for Mobile Wireless Research", Proc. of IEEE Infocom, 2005. [De et al (2006)] P. De, R. Krishnan, A. Raniwala, K. Tatavarthi, N. Syed, J. Modi, T. Chiueh, "MiNT-m: An Autonomous Mobile Wireless Experimentation Platform", Proceedings of Mobisys, 2006. [Dineen (2000)] M. Dineen, "Real-Time Display of Dublin Traffic Information on the Web", MS Thesis, University of Dublin, 2000. [Dion and Yagar (1996)] F. Dion, S. Yagar, "Real-Time Control of Signalized Networks -- Different Approaches for Different Needs", Eight International Conference on Road Traffic Monitoring and Control, IEE, pp. 56-60, 1996. [Driankov et al (1996)] D. Driankov, H. Hellendorn, M. Reinfrank, L. Ljung, R. Palm, B. Graham, A. Ollero, "An Introduction to Fuzzy Control", Springer, 1996. [Dubois et al (2003)] H. Dubois-Ferriere, M. Grossglauser, M. Vetterli, "Age Matters: Efficient Route Discovery in Mobile Ad Hoc Networks using Encounter Ages", Proc. of MobiHoc, 2003. [Eisenman et al (2007)] S. Eisenman, N. Lane, E. Miluzzo, R. Peterson, G. Ahn, A. Campbell, "The BikeNet Mobile Sensing System for Cyclist Experience Mapping", Proc. of Sensys 2007, 2007. [Eles et al (2000)] P. Eles, A. Doboli, P. Pop, Z. Peng, "Scheduling with Bus Access Optimization for Distributed Embedded Systems", IEEE Transaction on VLSI, Vol. 8, No. 5, pp. 472-491, October 2000. [Estrin et al (1999)] D. Estrin, R. Govindan, J. Heidemann, and S. Kumar, "Next Century Challenges: Scalable Coordination in Sensor Networks", Proc. Int. Conf. Mobile Computing and Networking (MOBICOM), 1999, pp. 263-270. [Farinwata et al (2000)] S. Farinwata, D. Filev, R. Langari, "Fuzzy Control: Synthesis and Analysis", Wiley, 2000. [Fehin (2004)] K. Fehin, "Adaptive Traffic Signals. Are we missing the boat?", ITE District 6 2004 Annual Meeting, available online at http://www.dksassociates.com/paper_signals.asp. [Feinberg (1994)] E. Feinberg, "Constrained Semi-Markov Decision Processes with Average Rewards", ZOR- Mathematical Methods of Operational Research, Vol.39, pp. 257-288,1994. [Feinberg (2002)] E. Feinberg, "Optimal Control of Average Reward Constrained Continuous Time Finite Markov Decision Processes", Proc. of IEEE Conference on Decision and Control, pp. 3805-3810, 2002. [Felici et al (2006)] G. Felici, G. Rinaldi, A. Sforza, K. Truemper, "Intelligent Traffic Control: A Logic Programming Approach and a Real Application", Transportation Research Part C: Emerging Technologies, Volume 14, Issue 3, June 2006, Pages 175-189. [Gajski et al (1994)] D. Gajski, F. Vahid, S. Narayan, J. Gong, "Specification and Design of Embedded Systems", Prentice Hall, 1994. [Gartner (1983)] N. Gartner, "A Demand Responsive Strategy for Traffic Signal Control", Transportation Research Record, No. 906, 1983. [Gartner et al (1992)] N. Gartner, C. Messer, A. Rathi, "Traffic Flow Theory: A State of the Art Report -- Revised Monograph on Traffic Flow Theory", U.S. Department of Transportation, Transportation Research Board, Washington D.C, 1992. [Girault et al (1999)] A. Girault, B. Lee, E. Lee, "Hierarchical Finite State Machines with Multiple Concurrency Models", IEEE Transactions on CAD of Integrated Circuits and Systems, Vol. 18, No. 6, pp. 742-760, June 1999. [Greene (2007)] K. Greene, "A Wireless Sensor City", Technology Review Published by MIT, April 13 2007. [Guibas (2002)] L. Guibas, "Sensing, Tracking and Reasoning with Relations", IEEE Signal Processing Magazine, March 2002, pp. 73-85. [Helbing et al (2007)] D. Helbing, S. Lammer, J.-P. Lebacque, "Self-Organized Control of Irregular or Perturbed Network Traffic", Optimal Control and Dynamic Games, Vol. 7, pp. 239-274, Springer, 2007. [Hull et al (2006)] B. Hull, V. Bychkovsky, K. Chen, M. Goraczko, A. Miu, E. Shih, Y. Zhang, H. Balakrishnan, S. Madden, "CarTel: A Distributed Mobile Sensor Computing System", Proc. of ACM SenSys, 2006. [Imielinski and Goel (2000)] T. Imielinski, S. Goel, "DataSpace - Querying and Monitoring Deeply Networked Collections in Physical Spaces", IEEE Personal Communications Magazine, August 2000, pp. 44-51. [Ismail et al (1995)] T.-B. Ismail et al, "Synthesis Steps and Design Models for Codesign", Computer Magazine, February 1995. [Jones et al (2001)] C. Jones, K. Sivalingam, P. Agrawal, J. Chgen, "A Survey of Energy Efficient Network Protocols for Wireless Networks'', Wireless Networks, Vol. 7, 2001, pp. 343-358. [Jain et al (2004)] S. Jain, R. Shah, W. Brunette, G. Boriello, S. Roy, "Exploiting Mobility for Energy Efficient Data Collection in Sensor Networks", Proc. of IEEE Workshop on Modeling and Optimization in Moble Ad Hoc and Wireless Networks, 2004. [Jogalekar and Woodside (2000)] P. Jogalekar and M. Woodside, "Evaluating the Scalability of Distributed Systems", IEEE Transactions on Parallel and Distributed Systems, Vol. 11, No. 6, pp. 589-603, June 2000. [Kazmer et al (2003)] D. Kazmer, D. Hatch, L. Zhu, C. Roser, D. Kapoor, "Definition and Application of a Process Flexibility Index", Journal of Manufacturing Science and Engineering, Vol. 125, pp. 164-173, February 2003. [Licalzi and O'Connell (2005)] P. Licalzi O'Connell, "Korea's High-Tech Utopia, Where Everything Is Observed", The New York Times, October 5 2005. [Kallakuri and Doboli (2007)] S. Kallakuri, A. Doboli, "Customization of Arbitration Policies and Buffer Space Distribution using Continuous Time Markov Decision Processes", IEEE Transactions on VLSI Systems, 2007. [Kansal et al (2004)] A. Kansal, A. Somasundara, S. Jea, M. Srivastava, D. Estrin, "Intelligent Fluid Infrastructure for Embedded Networks", Proc. of MobiSys'04, 2004. [Klemmer et al (2002)] S. Klemmer, S. Waterson, K. Whitehouse, "Towards a Location-Based Context- Aware Sensor Infrastructure", IEEE Signal Processing Magazine, March 2002. [Korkmaz et al (2004)] G. Korkmaz, E. Ekici, F. Ozguner, U. Ozguner, "Urban Multihop Broadcast Protocols for Inter-Vehicle Communication Systems", Proc. of VANET'04, 2004. [Kostrzewski et al (2001)] A. Kostrzewski, S. Ro, T. Jannson, "Visual Sensor Network Based on Video Streaming and IP-Transparency", Proc. of SPIE, Vol. 4396, 2001, pp. 74-8 [LeBrun et al (2005)] J. LeBrun, C.-N. Chuah, D. Ghosal, M. Zhang, "Knowledge-based Opportunistic Forwarding in Vehicular Wireless Ad Hoc Networks", Proc. of IEEE VTC, 2005. [Lee and Sangiovanni-Vincentelli (1998)] E. Lee, A. Sangiovanni-Vincentelli, "A Framework for Comparing Models of Computation", IEEE Transactions on CAD of Integrated Circuits and Systems, Vol. 17, No. 12, pp. 1217-1229. December 1998. [Lee et al (2006)] U. Lee, E. Magistretti, B. Zhou, M. Gerla, P. Bellavista, A. Corradi, "MobEyes: Smart Mobs for Urban Monitoring with Vehicular Sensor Networks", IEEE Wireless Communications, Vol. 13, No. 5, October 2006, pp. 52-57. [Liao et al (1997)] S. Liao, S. Tijang, R. Gupta, "An Efficient Implementation of Reactivity for Modeling Hardware in the Scenic Design Environment", Proc. of Design Automation Conference, 1997. [Meystel (2000)] A. Meystel, "Measuring Performance of Systems with Autonomy: metrics for Intelligence of Constructed Systems", NIST Workshop on Performance Metrics for Intelligent Systems, 2000. [Murty et al (2007)] R. Murty, A. Gosain, M. Tierney, A. Brody, A. Fahad, J. Bers, M. Welsh, "CitySense: A Vision for an Urban-Scale Wireless Networking Testbed", Harvard Technical Report TR-13-07, 2007. [Ng et al (2005)] C. Ng, P. Buanadonna, B. Chun, A. Snoeren, A. Vahdat, "Addressing Strategic Behavior in a Deployed Microeconomic Resource Allocator", Proc. of SIGCOMM'05, 2005. [Passino and Yurkovich (1998)] K. Passino, S. Yurkovich, "Fuzzy Control", Addison Wesley Longman, 1998. [Porche (1996)] I. Porche, "A Decentralized Scheme for Real-Time Optimization of Traffic Signals", Proc. of IEEE International Conference on Control Applications, pp. 582-589, 1996. [Pottie and Kaiser (2000)] G. J. Pottie, W. J. Kaiser, "Wireless Integrated Network Sensors", Communications of the ACM, Vol. 43, No. 5, 2000, pp. 51-58. [Qiu et al (2001)] Q. Qiu, Q. Wu, M. Pedram, "Stochastic Modeling of a Power Managed System Construction and Optimization", IEEE Transactions on CADICS, Vol. 20, No. 9, pp. 120-121, 2001. [Raychaudhuri et al (2005)] D. Raychaudhuri, I. Seskar, M. Ott, S. Ganu, K. Ramachandran, H. Kremo, R. Siracusa, H. Liu, M. Singh, "Overview of the ORBIT Radio Grid Testbed for Evaluation of Next- Generation Wireless Network Protocols", Proc. of WCNC'05, 2005. [Reeves et al (1993)] C. Reeves et al, "Modern Heuristic Techniques for Combinatorial Problems", J. Wiley, 1993. [Richter et al (2006)] S. Richter, D. Aberdeen, J. Yu, "Natural Actor-Critic for Road Traffic Optimisation", Proc. of NIPS, 2006. [Saleh et al (2002)] Saleh et al, "Space System Flexibility Provided by on-Orbit Sensing: Part II", Journal of Spacecraft and Rockets, Vol. 39, No. 4, pp. 561-570, July -- August 2002. [Santini and Vitaletti (2007)] S. Santini, A. Vitaletti, "Wireless Sensor Networks for Environmental Noise Monitoring", GI/ITG Workshop Wireless Sensor Networks, Aachen, 2007. [Shah et al (2003)] R. Shah, S. Roy, S. Jain, W. Brunette, "DataMULEs: Modeling a Three Tiered Architecture for Sparse Sensor Networks", Proc. of First International Workshop on Sensor Network Protocols and Applications, 2003. [Shen et al (2001)] C.-C. Shen, C. Srisathapornphat, C. Jaikaeo, "Sensor Information Networking Architecture and Applications", IEEE Personal Communications, August 2001, pp. 52-59. [Sinopoli et al (2003)] B. Sinopoli, C. Sharp, L. Schenato, S. Schaffert, S. Sastry, "Distributed Control Applications Within Sensor Networks", Proceedings of IEEE, Vol. 91, No. 8, pp. 1235-1246, August 2003. [Small and Haas (2003)] T. Small, Z. Haas, "The Shared Wireless Infostation Model: a New and Ad hoc Networkign Paradigm", Proceeding of ACM Mobihoc'03, 2003. [Tomlinson and Bull] A. Tomlinson and L. Bull, "Towards Distributed Adaptive Control for Road Traffic Junction Signals Using Learning Classifier Systems", available at citeseer.ist.psu.edu/601649.html. [Wahle and Schreckenberg (2001)] J. Wahle, M. Schreckenberg, "A Multi-Agent System for On-line Simulations based on Real World Data", Proc. of Hawaii International Conference on System Science, 2001. [Wang et al (2001)] I.-J. Wang, S. Jones, "The Scalability of a Class of Wireless Sensor Networks", Modeling and Design of Wireless Networks, E. Chong, editor, Proceedings of SPIE, Vol. 4531, 2001, pp. 103-113. [Warneke et al (2001)] B. Warneke, B. Liebowitz, K. Pister, "Smart Dust: Communicating with a Cubic Millimeter Computer", IEEE Computer, 2001, pp. 2-9. [Weiser (2002)] M. Weiser, "The Computer for the 21st Century", IEEE Pervasive Computing, Vol. 1, No. 1, pp. 15-25, January-March 2002. [Welsh and Bers] M. Welsh, J. Bers, "CitySense. An Open, City-Wide Wireless Sensor Network", presentation. [Weng and Doboli (2004)] Y. Weng, A. Doboli, "Smart Sensor Architecture Customized for Image Processing Applications", Proc. of 10th IEEE Real-Time and Embedded Technology and Applications Symposium, 2004. [Werner et al (2005)] G. Werner-Allen, P. Swieskowski, M. Welsh, "MoteLab: A Wireless Sensor Network Testbed", Proc. of International Symposium on Information Processing in Sensor Networks, 2005, pp.483-488. [Wildstrom et al (2005)] J. Wildstrom, P. Stone, E. Witchel, R. Mooney, M. Dahlin, "Towards Self- Configuring Hardware for Distributed Computer Systems", Proc. of International Conference on Autonomic Computing, 2005, pp. 241-249. [Wiering (2000)] M. Wiering, "Multi-Agent Reinforcement Learning for Traffic Light Control", Proceedings of ICML, 2000. [Wiering et al (2004)] M. Wiering, J. van Veenen, J. Vreeken, A. Koopman, "Intelligent Traffic Light Control", Technical report UU-CS-2004-029, Institute of Information and Computing Sciences, Utrecht University, 2004. [William (1999)] B. William, "Modeling and Forecasting Vehicular Traffic Flow as a Seasonal Stochastic Time Series Process", Doctoral Dissertation, University of Virginia, Charlottesville, 1999. [Wolshon and Taylor (1999)] B. Wolson, W. Taylor, "Analysis of Intersection Delay under Real-Time Adaptive Signal Control", Transportation Research part C, Pergamon, pp. 53-72, 1999. [Yoneki (2005)] E. Yoneki, "Evolution of Ubiquitous Computing with Sensor Networks in Urban Environments", Proc. of UbiComp'05, 2005. [Zhao et al (2002)] F. Zhao, J. Shin, J. Reich, "Information-Driven Dynamic Sensor Collaboration for Tracking Application", IEEE Signal Processing Magazine, March 2002. [Zhao and Guibas (2004)] F. Zhao, L. Guibas, "Wireless Sensor Networks", Morgan Kaufmann, 2004. [CALCCIT] /Traffic_Surveillance/road-based/in-road/loop_summary.html. [NAE] Restore and improve urban infrastructure, http://www.engineeringchallenges.org /cms/8996/9136.aspx http://www.calccit.org/itsdecision/serv_and_tech
1903.02868	1	1903	2019-03-07T12:23:15	Convergence of Multi-Agent Learning with a Finite Step Size in General-Sum Games	[ "cs.MA", "cs.LG" ]	Learning in a multi-agent system is challenging because agents are simultaneously learning and the environment is not stationary, undermining convergence guarantees. To address this challenge, this paper presents a new gradient-based learning algorithm, called Gradient Ascent with Shrinking Policy Prediction (GA-SPP), which augments the basic gradient ascent approach with the concept of shrinking policy prediction. The key idea behind this algorithm is that an agent adjusts its strategy in response to the forecasted strategy of the other agent, instead of its current one. GA-SPP is shown formally to have Nash convergence in larger settings than existing gradient-based multi-agent learning methods. Furthermore, unlike existing gradient-based methods, GA-SPP's theoretical guarantees do not assume the learning rate to be infinitesimal.	cs.MA	cs	Convergence of Multi-Agent Learning with a Finite Step Size in General-Sum Games Xinliang Song Tsinghua University Beijing, China [email protected] Tonghan Wang Tsinghua University Beijing, China [email protected] Chongjie Zhang Tsinghua University Beijing, China [email protected] 9 1 0 2 r a M 7 ] A M . s c [ 1 v 8 6 8 2 0 . 3 0 9 1 : v i X r a ABSTRACT Learning in a multi-agent system is challenging because agents are simultaneously learning and the environment is not stationary, undermining convergence guarantees. To address this challenge, this paper presents a new gradient-based learning algorithm, called Gradient Ascent with Shrinking Policy Prediction (GA-SPP), which augments the basic gradient ascent approach with the concept of shrinking policy prediction. The key idea behind this algorithm is that an agent adjusts its strategy in response to the forecasted strat- egy of the other agent, instead of its current one. GA-SPP is shown formally to have Nash convergence in larger settings than existing gradient-based multi-agent learning methods. Furthermore, unlike existing gradient-based methods, GA-SPP's theoretical guarantees do not assume the learning rate to be infinitesimal. KEYWORDS Multi-Agent Learning; Nash Equilibrium; Convergence; Finite Step Size ACM Reference Format: Xinliang Song, Tonghan Wang, and Chongjie Zhang. 2019. Convergence of Multi-Agent Learning with a Finite Step Size in General-Sum Games. In Proc. of the 18th International Conference on Autonomous Agents and Multiagent Systems (AAMAS 2019), Montreal, Canada, May 13 -- 17, 2019, IFAAMAS, 9 pages. 1 INTRODUCTION Multi-agent learning (MAL) is concerned with a set of agents that learn to maximize their expected rewards. There are a number of important applications that involve MAL, including competitive settings such as self-play in AlphaZero [23] and generative ad- versarial networks in deep learning [15, 20], cooperative settings such as when learning to communicate [13, 25] and multiplayer game [14], or some mix of the two [17, 26]. Although promising em- pirical results, establishing a theoretical guarantee of convergence for MAL, especially for gradient-based methods, is fundamentally challenging because of its non-stationary environment. Recent multi-agent learning (MAL) algorithms [7, 9 -- 11, 19, 22] with satisfactory empirical results are proposed, but most of them do not provide theoretical analyses of convergence. There are only a few worksthat provide theoretical results before them. Singh et al. [24] first consider the theoretical convergence of gradient-based methods in MAL. After that, several variants [1, 4, 5, 24, 27] are proposed and they provide theoretical convergence in general-sum Proc. of the 18th International Conference on Autonomous Agents and Multiagent Systems (AAMAS 2019), N. Agmon, M. E. Taylor, E. Elkind, M. Veloso (eds.), May 13 -- 17, 2019, Montreal, Canada. © 2019 International Foundation for Autonomous Agents and Multiagent Systems (www.ifaamas.org). All rights reserved. games, but theoretical guarantees are restricted in 2-agent, 2-action games and they assume that the learning rate is infinitesimal, which is not practical. Some other online learning algorithms [8, 12, 16] have also been proposed with theoretical guarantees, but just for specific settings, such as congestion games and potential games. In this paper, we propose a new multi-agent learning algorithm that augments a basic gradient ascent algorithm with shrinking policy prediction, called Gradient Ascent with Shrinking Policy Prediction (GA-SPP). The key idea behind this algorithm is that an agent adjusts its strategy in response to the forecasted strategy of the other agent, instead of its current one. This paper makes three major novelties. First, to our best knowledge, GA-SPP is the first gradient-ascent MAL algorithm with a finite learning rate that provides convergence guarantee in general-sum games. Second, GA- SPP provides convergence guarantee in larger games than existing gradient-ascent MAL algorithms, which include m × n positive semi-definite games, a class of 2 × n general-sum games, and 2 × 2 general-sum games. Finally, GA-SPP guarantees to converge to a Nash Equilibrium when converging in any m×n general-sum game. Although GA-SPP shares some similar ideas about using policy prediction with IGA-PP [27] and the extra-gradient method [2], it has several major differences from them. For example, apart from using a finite step size, another significant difference between GA- SPP and IGA-PP is that forecasted strategies of the opponent are projected to the valid probability space. This improvement enables GA-SPP's Nash convergence when converging, which does not hold for IGA-PP. In contrast to the extra-gradient approach, GA- SPP uses shrinking prediction lengths which can be different from the policy update rate. This improvement makes GA-SPP not only more flexible in practice but also stronger in terms of theoretical guarantees. Like IGA-PP, we assume that agents know the other agent's strategy and its current strategy gradient, but we do not require the learning rate to be infinitesimal. Even though GA-SPP needs some restricted assumptions, it pushes forward the state of the art of MAL with theoretical analysis. We expect that our work can shed a light for theoretical understanding of dynamics and complexity of MAL problems and like IGA-PP and WoLF-IGA [6], can encourage broadly applicable multi-agent reinforcement learning algorithms. Our proposed learning algorithm also provides a different approach for computing Nash Equilibiria of subsets of larger games, other than well-established offline algorithms [18, 21], whose computa- tion complexity increases sharply with the number of actions. Notation We use following notations in this paper: ∆ denotes the valid strategy space (i.e., a simplex). Π∆: ℜn → ∆ denotes the convex projection to the valid space, Π∆[x] = arдminz ∈∆∥x − z∥. P∆(x, v) denotes the projection of a vector v on x ∈ ∆, P∆(x, v) = lim η→0 Π∆[x + ηv] − x η . (cid:19) (cid:18) v1 v2 (v1; v2) denotes , where v1, v2 are column vectors. 2 GRADIENT ASCENT We begin with a brief overview of normal-form games and then review the basic gradient ascent algorithm. 2.1 Normal-Form Games A 2-agent, m × n -action, general-sum normal-form game is defined by a pair of matrices  r11 ... rm1 R =  ... ... ... r1n ... rmn  c11 ... cm1  ... ... ... c1n ... cmn and C = specifying the payoffs for the row agent and the column agent, respectively. The agents simultaneously select an action from their available set, and the joint action of the agents determines their payoffs according to their payoff matrices. If the row agent selects action a ∈ {1, ..., m} and the column agent selects action b ∈ {1, ..., n}, respectively, then the row agent receives a payoff rab and the column agent receives a payoff cab. The agents can choose actions stochastically based on some prob- ability distribution over their available actions. This distribution is said to be a mixed strategy. Let αi ∈ [0, 1] denote the probability of choosing the i-th action by the row agent and βj ∈ [0, 1] denote the probability of choosing the j-th action by the column agent, where βj ≤ 1. We use ∆1 to denote a m-1 dimensional simplex and ∆2 to denote a n-1 dimensional simplex. This (m − 1) × (n − 1) representation is equivalent to the m × n representation, and, following the previous work on gradient-based methods, we choose the former one. Let i ∈ {1, ..., m − 1}, j ∈ {1, ..., n − 1},m−1 αi ≤ 1,n−1 1 1 α = [α1 ... αm−1]T β = [β1 ... βn−1]T , , em−1 = [1 ... 1]T en−1 = [1 ... 1]T , , where the dimension of em−1 is m − 1, the dimension of en−1 is n − 1. Then α ∈ ∆1, β ∈ ∆2. With a joint strategy (α , β), the row agent's and column agent's expected payoffs are R(β; 1 − e C(β; 1 − e Vr(α , β) = (α; 1 − e Vc(α , β) = (α; 1 − e m−1α)T T m−1α)T T n−1β), T n−1β). T (1) A joint strategy (α∗, β∗) is called a Nash equilibrium if for any mixed strategy α of the row agent, Vr(α∗, β∗) ≥ Vr(α , β∗), and for any mixed strategy β of the column agent, Vc(α∗, β∗) ≥ Vc(α∗, β). It is well-known that every game has at least one Nash equilibrium. 2.2 Learning using Gradient Ascent in Iterated Games In an iterated normal-form game, agents repeatedly play the same game. Each agent seeks to maximize its expected payoff in response to the strategy of the other agent. Using the basic gradient ascent algorithm, a agent can increase its expected payoff by updating its strategy with a step size along the gradient of the current strategy. The gradient is computed as the partial derivative of the agent's expected payoff with respect to its strategy: ∂α Vr(α , β) = ∂βVc(α , β) = where Im−1 is (m-1)-order identity matrix and In−1 is (n-1)-order identity matrix. If (αk , βk) are the strategies on the k-th iteration and both = (Im−1 em−1)R(β; 1 − e = (In−1 en−1)C T(α; 1 − e ∂Vr(α , β) ∂Vc(α , β) n−1β), T m−1α), T agents use gradient ascent, then the new strategies will be: (2) ∂α ∂β αk +1 = Π∆1[αk βk +1 = Π∆2[βk + η∂α Vr(αk , βk)], + η∂βVc(αk , βk)], (3) where η is the gradient step size. If the updates move the strategies out of the valid probability space, the function Π∆ will project it back. Singh et al. [24] analyzed the gradient ascent algorithm by exam- ining the dynamics of the strategies in the case of an infinitesimal step size (limη→0). This algorithm is called Infinitesimal Gradient Ascent (IGA). IGA cannot converge in some 2-agent 2-action zero- sum game. GIGA-WoLF and IGA-PP extended IGA and provide theoretical guarantee of Nash equilibrium in 2-agent 2-action game through similar methods. However, these algorithms require a in- finitesimal step size, which is not practical. We will describe a new gradient ascent algorithm that enables the agents' strategies to converge to a Nash equilibrium with a finite step size in a larger game setting. 3 GRADIENT ASCENT WITH SHRINKING POLICY PREDICTION (GA-SPP) As shown in Eq. 3, the gradient used by IGA to adjust the strategy is based on current strategies. Suppose that an agent can estimate the change direction of the opponent's strategy, i.e., its strategy derivative, in addition to its current strategy. Then the agent can forecast the opponent's strategy and adjust its own strategy in response to the forecasted strategy. With this idea, we design a gradient ascent algorithm with shrinking policy prediction (GA- SPP). Its updating rule consists of three steps. In Step 1, the new derivative terms with γ serve as a short-term prediction of the opponent's strategy. If the opponent's forecasted strategy is out of boundary of simplex, it will be projected back to the valid space. In Step 2, agents update their strategies on the basis of the fore- casted strategy of its opponent. In Step 3, agents terminate or adjust their prediction lengths. If predicted strategies are equal to the current strategies, the algo- rithm will terminate. Step 3 can make sure GA-SPP only converges to Nash equilibrium (NE) instead of other points. Because when (αk +1, βk +1) = (αk , βk), GA-SPP will stop, if there is no Step 3, then (α k +1, βk +1) (cid:44) (αk , βk) may happen. In this situation, GA- SPP may converge to a non-NE point. We will prove this property of GA-SPP in Proposition 1. Algorithm 1: Updating rule of GA-SPP 1 α k +1 = Π∆1[αk + γk ∂α Vr(αk , βk)]; βk +1 = Π∆2[βk + γk ∂βVc(αk , βk)]; + η∂α Vr(αk , βk +1)]; 2 αk +1 = Π∆1[αk + η∂βVc(α k +1, βk)];; βk +1 = Π∆2[βk 3 if (α k +1, βk +1) == (αk , βk) then else terminate; if (αk +1, βk +1) == (αk , βk) & (α k +1, βk +1) (cid:44) (αk , βk) then γk +1 = µkγk , (0 < µk < 1), back to (1); γk +1 = γk, back to (1); else end end The prediction length γk and gradient step size η will affect the convergence of the GA-SPP algorithm. With a too large prediction length, the gradient computed with the forecasted strategy will deviate too much from the gradient computed with the opponent's current strategy. As a result, the agent may adjust its strategy in the improper direction and cause their strategies to fail to converge. Following conditions ensure that γ and η are appropriate: Condition 1: Condition 2: Condition 3: γ0 > 0, η > 0 2 4γ 0 δr δc < 1 1 η, γ0 < δr +δc where δr = rmax − rmin, δc = cmax − cmin, rmax and cmax is the maximum reward for the row and column agent, rmin and cmin is the minimum reward for the row and column agent. Condition 3 makes sure that the theoretical guarantee of Nash convergence in the game settings analyzed in Section 4. In exper- iments, the algorithm can still work in some other games if we choose larger prediction length or let agents have different predic- tion lengths. 3.1 Analysis of GA-SPP In this section, we will show that if agents' strategies converge by following GA-SPP, then they must converge to a Nash equi- librium, which is described by Proposition 1. Using this proposi- tion, we will then prove the Nash convergence of GA-SPP in three classes of games: m× n positive semi-definite games, a class of 2× n general-sum games, and 2 × 2 general-sum games, respectively, in the following sections. Before proving Proposition 1, we will first show that if the pro- jected gradients of a strategy pair are zero, then this strategy must be a Nash equilibrium, which is described by Lemma 3.1. For brevity, let ∂α denotes ∂α Vr(α , β), ∂β denotes ∂βVr(α , β). Lemma 3.1. In (m×n)-action games, if the projected partial deriva- tives at a strategy pair (α∗, β∗) are zero, that is, P∆1(α∗, ∂α ∗) = 0 and P∆2(β∗, ∂β∗) = 0, then (α∗, β∗) is a Nash equilibrium. Proof. Assume that (α∗, β∗) is not a Nash equilibrium. Then at least one agent, for example, the column agent, can increase its expected payoff by changing its strategy unilaterally. Assume that the improved point is (α∗, β). Because of the convexity of the strategy space ∆2 and the linear dependence of Vc(α , β) on β, then, for any ϵ > 0, (α∗, (1 − ϵ)β∗ + ϵβ) must also be an improved point, which implies that the projected gradient of β at (α∗, β∗) is not zero. By contradiction, (α∗, β∗) is a Nash equilibrium. □ Proposition 1. In 2-agent, m × n games, if two agents follow GA-SPP with appropriate γ , η (satisfying Condition 1, 2, and 3) and GA-SPP converges, then (α∗, β∗) is a Nash equilibrium. Here is a proof sketch (the detailed formal proof is described in supplementary material1). According to Step 3 in the algorithm 1, if the strategy pair trajectory converges at (α ∗, β∗), then (α ∗, β∗) = (α k +1, β k +1) = (αk , βk)or (α ∗, β∗) = limk→∞(α k +1, β k +1) = limk→∞(αk , βk). For both cases, we can have α ∗ = Π∆1[α ∗ + η ∂α∗] and β∗ = Π∆2[β∗ + η ∂β∗]. From here, we can show that, for any arbitrary small ϵ > 0, α ∗ = Π∆1[α ∗+ϵ ∂α∗]and β∗ = Π∆2[β∗+ϵ ∂β∗], which imply P∆1(α ∗, ∂α∗) = 0 and P∆2(β∗, ∂β∗) = 0. Then according to Lemma 3.1, (α ∗, β∗) is a Nash equilibrium. 4 CONVERGENCE OF GA-SPP We will show the Nash convergence of GA-SPP in three classes of games in this section. 4.1 m × n Positive Semi-Definite Games A function Φ(v, w) is called a positive semi-definite function if it obeys the inequality defined in [3]: Φ(w, w) − Φ(w, v) − Φ(v, w) + Φ(v, v) ≥ 0. (4) To facilitate the proof, we define the normalized value function for a game: Φ(v, w) = Vr(α 1, β2) + Vc(α 2, β1), where v = (α 1, β1) ∈ {∆1 × ∆2}, w = (α 2, β2) ∈ {∆1 × ∆2}. (5) Definition 1. A 2-agent m×n game is called positive semi-definite (PSD) game if its normalized value function obeys Φ(w, w) − Φ(w, v) − Φ(v, w) + Φ(v, v) ≥ 0. It means that for a PSD game, its payoff matrices satisfies Vr(α 1, β1) + Vc(α 1, β1) + Vr(α 2, β2) + Vc(α 2, β2) ≥Vr(α 1, β2) + Vc(α 1, β2) + Vr(α 2, β1) + Vc(α 2, β1) ∀α 1, α 2 ∈ ∆1, ∀β1, β2 ∈ ∆2. (6) (7) Zero-sum games are a subset of PSD games, because their value functions satisfy Vr(α , β) + Vc(α , β) = 0, then both sides of inequality 7 are equal to zero. 1https://drive.google.com/file/d/1TZeRf0xp4g4wg-JX7zA9TjqC2S619pAp/view?usp= sharing For a PSD game, if (α∗, β∗) is a Nash equilibrium and v∗ = (α∗, β∗), then its normalized function obeys ⟨∇2Φ(w, w), w − v∗⟩ ≥ 0 ∀w ∈ {∆1 × ∆2}. (8) In the proof of Theorem 1, we will use this inequality. theorem 1. If, in a 2-agent, m × n iterated positive semi-definite norm-form game, both agents follow the GA-SPP algorithm (with Condition 1, 2, and 3), then their strategies will converge to a Nash equilibrium. Proof. Motivated by [2], our proof will use some variational inequalities techniques. (9) From the first and second step of GA-SPP (Algorithm 1), we have estimates α k +1 − αk +1 ≤ γk ∂α Vr(αk , βk) − η∂α Vr(αk , βk +1), βk +1 − βk +1 ≤ γk ∂βVc(αk , βk) − η∂βVc(α k +1, βk). We present the first and second step of GA-SPP in the form of variational inequalities: ⟨α k +1 − αk − γk ∂α Vr(αk , βk), z1 − α k +1⟩ ≥ 0 ∀z1 ∈ ∆1, ⟨βk +1 − βk − γk ∂βVc(αk , βk), z2 − βk +1⟩ ≥ 0 ∀z2 ∈ ∆2; (10) ⟨αk +1 − αk − η∂α Vr(αk , βk +1), z1 − αk +1⟩ ≥ 0 ∀z1 ∈ ∆1, ⟨βk +1 − βk − η∂βVc(α k +1, βk), z2 − βk +1⟩ ≥ 0 ∀z2 ∈ ∆2. (11) . Put z1 = α∗, z2 = β∗ in Eq. 11, then set z1 = αk +1, z2 = βk +1 in Eq. 10, and take into account of Eq. 9, we can get (the detailed computation is listed in our supplementary material) (cid:18) α 1 Let v = (cid:19) β1 ⟨v k +1 − v +η⟨∇2Φ(v 2∥∇2Φ(v +h k , v∗ − v k +1, v k , v k +1⟩ + ⟨v k +1), v∗ − v k) − ∇2Φ(v k +1 − v k +1⟩ k +1, v k +1)∥2 ≥ 0, k +1 − v k +1⟩ k , v (12) where h = max{γ0, η}. By means of identity, the first two scalar products in Eq. 12 can be rewritten as k − v∗∥2 − 1 2 ∥v k +1∥2 − 1 2 ∥v 1 2 ∥v k +1 in Eq. 8, then the third term in Eq.12 is non- positive. For the last term of Eq. 12, if ∇2Φ(v k) satisfies the Lipschitz condition with constant L, then following estimate is correct k +1 − v∗∥2− k ∥2 k +1 − v 1 2 ∥v k +1 − v Set w = v k , v (13) . ∇2Φ(v k) − ∇2Φ(v k , v k +1, v k +1) ≤ Lv k +1 − v k . (14) 2 Now put Eq. 13 and Eq. 14 in Eq. 12, we can yield k +1∥2+ k +1 − v∗∥2 + ∥v ∥v (1 − 2h k +1 − v Note that ∇2Φ(v 2)∥v k) = ∂α Vr(α , β) + ∂βVc(α , β). According to Eq. 2, ∂α Vr(α , β) is a function of β, ∂βVc(α , β) is a function of α. The maximum value of 2-norm of ∂α is not greater than 2/2, and not greater than δc . So the Lipschitz δr k +1 − v k ∥2 ≤ ∥v 2/2 for 2-norm of ∂β k − v∗∥2 k , v (15) L . lows K K constant L ≤ δr√2 + δc√2 √2 2 and 1 − 2h 1 . According to Condition 3, h = max{γ0, η} < > 0. Sum up inequality Eq. 15 , so hL < L 2 2 δc +δr from k = 0 to k = K, we get ∥vK +1 − v∗∥2 + K 2) K k =0 k =0 ∥v k +1 − v ∥v k +1 − v k +1∥2+ (16) k ∥2 ≤ ∥v0 − v∗∥2 . (1 − 2h 2 L From the gained inequality (Eq. 16) the bound of trajectory fol- ∥vK +1 − v∗∥2 ≤ ∥v0 − v∗∥2 , (17) and the series are convergent v ∥v ∥v k =0 k ∥2 < ∞, < ∞. k +1 − v k +1 − v k +1 − v k +1−v k =0 k +1∥2 = 0, limk→∞ ∥v k +1∥2 k +1 − As a result, limk→∞ ∥v k ∥2 = 0, so limk→∞ ∥v k ∥2 = 0. It implies limk→∞ ∥αk +1− αk ∥2 = 0 and limk→∞ ∥βk +1 − βk ∥2 = 0. So GA-SPP can converge. With Proposition 1, GA-SPP must converge to a Nash equilibrium. Therefore, proof of Theorem 1 is completed. □ theorem 2. If, in a 2-agent, m × n iterated positive semi-definite norm-form game, one agent follows the GA-SPP algorithm (with Condition 1, 2, and 3), another agent uses GA, then their strategies will converge to a Nash equilibrium. The proof of this theorem is omitted, which is similar to that of Theorem 1. 4.2 A Subclass of 2 × n General-Sum Games In this section, we will show that GA-SPP converges to a Nash equilibrium in a subclass of 2-agent 2×n general games (Theorem 3). A 2-agent, 2 × n, general-sum normal-form game's payoff matri- ces can be written as (cid:21) (cid:20)c11 c21 ... ... r1n r2n , C = ... ... c1n c2n . (cid:20)r11 r21 R = Let (cid:21) r1 = [r11 ... r1,n−1]T, c1 = [c11 ... c1,n−1]T, r2 = [r21 ... r2,n−1]T, c2 = [c21 ... c2,n−1]T. Then agents' expected payoffs (Eq. 1) are Vr(α, β) = (α βT)r1 + r1n(α(1 − βTen−1)) + (1 − α)βTr2 + r2n((1 − α)(1 − βTen−1)), Vc(α, β) = (α βT )c1 + c1n(α(1 − βTen−1)) + (1 − α)βTc2 + c2n((1 − α)(1 − βTen−1)). The gradients (Eq. 2) can be written as ∂αVr(α, β) = ∂βVc(α, β) = ∂Vr(α, β) ∂Vc(α, β) ∂α ∂β = βTur + br , = αuc + bc , (18) (19) where br = r1n − r2n, bc = c2 − c2nen−1, ur = r1 − r2 − br en−1, and uc = c1 − c2 − (c1n − c2n)en−1. theorem 3. If, in a 2-agent, 2× n, norm-form game, if there exists a δ > 0 such that the payoff matrices obey ur + δuc = 0, (20) and both agents follow the GA-SPP algorithm (with Condition 1, 2, and 3), then their strategies will converge to a Nash equilibrium. Proof. For a 2-agent 2×n game, if we put Eq. 18 into Definition 1, then we derive ur +uc = 0. It shows that 2×n games in Theorem 3 with δ = 1 are PSD games. First we consider 2 × n positive semi-definite games (ur + uc = 0). According to Theorem 1, in this particular case, GA-SPP can converge to a Nash Equilibrium. It means the following iteration can converge: αk +1 = Π∆1[αk + γk(β T kur1 βk +1 = Π∆2[βk − γk(αkur1 T αk +1 = Π∆1[αk + η(β kur1 βk +1 = Π∆2[βk − η(αkur1 + br1)], + bc1)]; + br1)], + bc1)]. (21) For brevity, we omit step 3 of GA-SPP. For a 2 × n, norm-form game that obeys Eq. 23, we have ur2 √ + δ β. If α and β follows GA-SPP, then , y = δuc2 the update rule of x and y is = 0. Let x = α√ δ xk +1 = Π∆x [xk + γk(y yk +1 = Π∆y xk +1 = Π∆x [xk + η(y T kur2 [yk − γk(αkur2 T kur2 [yk − η(xkur2 + + + )], br2√ √ δ δbc2)]; + )], br2√ √ δ δbc2)]. yk +1 = Π∆y (22) Comparing Eq. 22 with Eq. 21, (x, y) can be viewed as a strategy pair of another 2 × n PSD game following GA-SPP. Notice that the proof of Theorem 1 only requires that the valid space is a bounded convex set. Therefore, if (α, β) follows GA-SPP, (x, y) can converge, then (x, y) can still converge in 2 × n, norm-form game can converge. With Proposition 1, we finish the proof of Theorem 3. □ theorem 4. If, in a 2-agent, 2× n, norm-form game, if there exists δ > 0, and the payoff matrices obey and one agent follow the GA-SPP algorithm (with Condition 1, 2, and 3), another agent uses GA, then their strategies will converge to a Nash equilibrium. The proof of this theorem is omitted, which is similar to that of Theorem 3. ur + δuc = 0, (23) λ1 = γkur uc + 4.3 2 × 2 General-Sum Games In this section, we will prove the Nash convergence of GA-SPP in 2 × 2 general-sum games. theorem 5. If, in a 2-agent, 2 × 2, iterated general-sum game, both agents follow the GA-SPP algorithm (with Condition 1, 2, and 3), then their strategies will converge to a Nash equilibrium. Proof. With Proposition 1, in order to prove Theorem 5, we just need to prove the convergence of GA-SPP in 2 × 2 games, which is accomplished by Lemma 4.1, 4.2, and 4.3. □ Next, we will analyze the structure of 2 × 2 games firstly, and then show the convergence in different cases respectively. In a 2-agent, 2-action game, the reward functions (Eq. 1) can be written as Vr(α, β) = r11(α β) + r12(α(1 − β)) + r21((1 − α)β) Vc(α, β) = c11(α β) + c12(α(1 − β)) + c21((1 − α)β) + r22((1 − α)(1 − β)), + c22((1 − α)(1 − β)). And the gradient function (Eq. 2) can be written as ∂αVr(α, β) = ∂βVc(α, β) = ∂Vr(α, β) ∂Vc(α, β) ∂α ∂β = ur β + br , = uc α + bc , where ur = r11 +r22−r12−r21, br = r12−r22, uc = c11 +c22−c12−c21, and bc = c21 − c22. We have ur ≤ 2δr , uc ≤ 2δc. We can formulate the first two update rules of GA-SPP (1): αk +1 = Π∆[αk + η∂αk Vr(αk , Π∆[βk + γ ∂βk , βk])], βk +1 = Π∆[βk + η∂βk Vc(βk , Π∆[αk + γ ∂αk , αk])], where ∆ = ∆1 = ∆2 = [0, 1]. To prove the Nash convergence of GA-SPP, we will examine the dynamics of the strategy pair following GA-SPP. In a 2-agent, 2-action, general-sum game, (α, β) can be viewed as a point in R2 constrained to lie in the unit space. According to Eq. 24, if (αk , βk) is a unconstrained point, then value of (αk +1, βk +1) is (24) (cid:21)(cid:20)αk (cid:21) (cid:20)γkur uc (cid:20)γkur bc + br uc γkucbr + bc ur (cid:21) . = η +η γkur uc βk (25) (cid:21) (cid:20)αk +1 βk +1 (cid:21) (cid:20)αk βk − We denote the 2 × 2 matrix in Eq. 25 as U. If the matrix U is invertible, in the unconstrained condition, there exists and only exists one point so that the left hand side of Eq. 25 is zero. We call this point the center (or origin) and denote it as (αc , βc). The eigenvalues of U is given by √ ur uc and λ2 = γkur uc − √ (26) k δr δc < 1) and ur ≤ 2δr , uc ≤ 2δc, 2 ur uc . According to Condition 2 (4γ then γ 2 kur uc < 1. There are three cases of U: • Case 1: ur uc = 0, i.e., U is not invertible; • Case 2: ur uc < 0, i.e., having two imaginary conjugate eigen- values with negative real; • Case 3: ur uc > 0, i.e., having two real eigenvalues. If U is invertible, then ur uc (cid:44) 0. Let x = α + bc uc Eq. 25 can be reformulated as: , y = β + br ur , (cid:21) (cid:20)xk +1 yk +1 (cid:21) (cid:20)xk yk − (cid:20)γur uc uc = η (cid:21)(cid:20)xk (cid:21) yk ur γur uc (27) By setting the left hand side of Eq. 27 to zero, we can get an equation, the only solution of which is x = 0, y = 0. Now we give the proof of Lemma 4.3. Proof. From Eq. 26, real eigenvalues imply ur uc > 0. So without the loss of generality, we assume that ur > 0 and uc > 0 (the analysis for the case ur < 0 and uc < 0 is analogous and thus omitted). Proof of Lemma 4.3 depends on the location of (αc , βc), which has three possibilities: (1) both αc and βc are in the valid probability range [0,1], (2) only one of αc , βc is in the valid probability range [0,1], (3) neither αc nor βc is in the valid probability range [0,1]. Proofs of convergence in these three cases are given in Property 4.1, 4.2, and 4.3, respectively. □ Notice that U has two real eigenvalues: λ1 > 0 and λ2 < 0 and two nonparallel eigenvectors. The central point (x = 0, y = 0) with two eigenvectors (v1 = [√ uc]) can form a new 2D coordinate system. The basic idea of proof is to analyze coordinates of the strategy pair update trajectory in the (cid:114)ur new coordinate system. To be brief, we introduce two functions to compute it instead of converting coordinate. uc], v2 = [√ √ (cid:114)ur ur , −√ ur , y, G = x − (28) F = x + uc y. uc Property 4.1. If U has real eigenvalues, both of αc , βc are in the valid probability range([0, 1]), GA-SPP leads the strategy pair trajectory to converge to a NE. Proof. As shown in Fig. 2(a), the initial point will affect the Nash convergence result because there are three Nash equilibrium points. The first case is F0 = 0. Then the strategy pair point will keep staying in the line F = 0 while the absolute value of G decreases, which means that the point moves to the center point (F = 0, G = 0), i.e., (α∗, β∗). For example, if P2 is the initial point, the point will travel along the line F = 0 and moves to (F = 0, G = 0). We can compute Fk +1 and Gk +1 by Fk +1 = (1 + ηλ1)Fk . Gk +1 = (1 + ηλ2)Gk , (29) According to Condition 1, 2 and 3, 0 < (1 + ηλ2) < 1, so the GA-SPP will converge to (F = 0, G = 0), i.e., (x = 0, y = 0). Another case is when F0 > 0, from Fig. 2(a), we can tell that the strategy pair point first touches the boundary xmax (α = 1) or ymax (β = 1) after finite iteration steps, after then it travels along the boundary and moves to (xmax , ymax). For example, if P1 is the initial point, the point will touch the boundary α = 1 (i.e., xmax ), then it travels along α = 1 and moves to (xmax , ymax). Without no more than one exceptional case, we can derive Fk +1 > Fk in each iteration. From the monotone bounded theorem, the GA-SPP will converge to Fmax , i.e., (xmax , ymax). Figure 1: Strategy updating directions of the GA-SPP. a) when U has real eigenvalues and b) when U has imaginary eigenvalues with negative real part. To prove Theorem 5, we only need to show that GA-SPP always leads the strategy pair to converge in these three cases. Lemma 4.1. If, in a 2-agent, 2-action, iterated general-sum game, U is not invertible, for any initial strategy pair, GA-SPP leads the strategy pair trajectory to converge to a Nash equilibrium (NE) with finite steps. Proof. From Eq. 26, if U is not invertible, then ur uc = 0. Assume uc = 0 (the analysis for the case ur = 0 is analogous and thus omitted for brevity). According to Eq. 25, βk +1 = Π∆[βk + ηbc]. Because ηbc is con- stant and β ∈ [0, 1], strategy β will no longer change after finite steps. We denote this value by β∗. Then, αk +1 = Π∆[αk + η(ur β∗ + br)]. Because η(ur β∗ + br) is a constant and α ∈ [0, 1], after a certain number (finite) of steps, strategy α will also stop chang- ing. We denote this value by α∗. So in this case, the strategy pair will converge to (α∗, β∗), and with Proposition 1, this is a Nash equilibrium. Note that the index of ηk was omitted in the proof, which is because the situation of (αk +1, βk +1) = (αk , βk) & (αk +1, βk +1) (cid:44) (αk , βk) did not occur in this case. Lemma 4.3 also has this property. □ Lemma 4.2. If, in a 2-agent, 2-action, iterated general-sum game, U has two imaginary conjugate eigenvalues with negative real, for any initial strategy pair, GA-SPP leads the strategy pair trajectory to converge to a NE. Proof. Since ur uc < 0, there exists a δ such that ur + δuc = 0. This is a 2-dimensional situation of Theorem 3. So GA-SPP can converge to a NE. □ In the rest of this section, we will introduce Lemma 4.3 and the basic idea of proof. For the detailed mathematical proof, we refer readers to the supplementary material. Lemma 4.3. If, in a 2-agent, 2-action, iterated general-sum game, U has real eigenvalues, for any initial strategy pair, GA-SPP leads the strategy pair trajectory to converge to a point that is a NE. Before proof, we first introduce some variables to simplify the expressions. (a) (b) (c) Situation when F0 < 0 is similar to that when F0 > 0, so we omit Figure 2: Updating directions of strategy pair in Lemma 4.3. (a) Both αc and βc are in the valid probability range, F(P1) > 0, F(P2) = 0; (b) only one of αc and βc is in the valid range, F(P3) = 0, F(P4) < 0; (c) neither αc and βc is in the valid range. (b) Chicken Swerve (-2,-2) (-1,1) Table 1: Benchmark games Straight (1,-1) (-1,-1) Swerve Straight (a) Prisoners' Dilemma Silent Betray Silent (-1,-1) (0,-3) Betray (-3,0) (-2,-2) it for brevity. In all, GA-SPP can converge for any initial strategy pair. □ Property 4.2. When U has real eigenvalues and only one of αc and βc is in the valid probability range, GA-SPP leads the strategy pair trajectory to converge to a NE. Proof of Property 4.2 can be classified into 4 cases. Without loss of generality, we just consider one of them, where βc < 0 and αc ∈ [0, 1]. As shown in Fig. 2(b), we can also divide the proof into 3 parts: F0 > 0, F0 < 0, and F0 = 0. If the point is in the part F > 0, according to Property 4.1, the algorithm will converge to (xmax , ymax). If F0 ≤ 0, we can see that the point will first touch a boundary of the valid probability space, after that it will move into the part F > 0. For example, if P3 (F = 0) is the initial point, the point will travel along line F = 0 until it hits the boundary, then it will be projected to the subspace where F > 0. If P4 (F < 0) is the initial point, it will touch the boundary β = 0 (ymin) and then y remains ymin while x increases until it move into the subspace where F > 0. Property 4.3. If U has real eigenvalues and neither αc nor βc is in the valid probability range [0, 1], GA-SPP leads the strategy pair trajectory to converge to a NE. Updating directions of strategy pair is shown in Fig. 2(c) for this case. For the detailed proof, please refer to supplementary material. theorem 6. If, in a 2-agent, 2 × 2, iterated general-sum game, one agent follow the GA-SPP algorithm (with Condition 1, 2, and 3), another agent uses GA, then their strategies will converge to a NE The proof is omitted, which is similar to that of Theorem 5. 5 EXPERIMENTAL ANALYSIS IN NORMAL-FORM GAMES In this section, we will illustrate GA-SPP in games with experiments and compare GA-SPP with IGA-PP and GIGA-WoLF, both of which have theoretical guarantees, in some larger games. 5.1 Benchmark games We first illustrate the results of GA-SPP on four representative benchmark games presented in Tab. 1. GA-SPP converges to NE in all of these games (Fig. 3). (c) Battle of Sexes Opera Football Opera (3,2) (0,0) Football (1,1) (2,3) (d) Rock-Paper-Scissors R (0,0) (1,-1) (-1,1) P (-1,1) (0,0) (1,-1) S (1,-1) (-1,1) (0,0) R P S Figure 3: Action probabilities of row agent following GA- SPP in four benchmark games. Parameters: η = 0.001, γ = 0.1. Initial polices: (0.7, 0.3) and (0.3, 0.7). 5.2 Games beyond theoretical settings We also evaluate GA-SPP in Shapley's game and a 2 × 3 game, presented in Tab. 2. Although the theoretical analyses of GA-SPP have not covered these games, empirical results show that it still converge. We now compare GA-SPP, GIGA-WoLF, and IGA-PP in these two games. Fig. 4 shows the row player's action probabilities over time if both players follow GA-SPP, GIGA-WoLF, or IGA-PP in Shapley's game respectively. GIGA-WoLF fails to converge in this non-zero sum game, but GA-SPP and IGA-PP can converge to a Nash equilibrium. 00.20.40.60.8100.10.20.30.40.50.60.70.80.91 ⇐ G=0 ⇐ F=0 (α,β)P1P200.20.40.60.8100.10.20.30.40.50.60.70.80.91 ⇐ F=0P4P300.20.40.60.8100.10.20.30.40.50.60.70.80.91 ⇐ G=0P502004006008001000iteration steps00.10.20.30.40.50.60.70.80.91Prisoners' DilemmaPr(Silent) of row playerPr(Betray) of row player050100150200250300iteration steps00.10.20.30.40.50.60.70.80.91ChickenPr(Swerve) of row playerPr(Straight) of row player0100200300400500iteration steps00.10.20.30.40.50.60.70.80.91Battels of the SexesPr(Opera) of row playerPr(Football) of row player0200040006000800010000iteration steps00.10.20.30.40.50.60.70.80.91Rock-Paper-ScissorsPr(Rock) of row playerPr(Paper) of row playerPr(Scissors) of row playerFigure 4: Comparison between GA-SPP, GIGA-WoLF, and IGA-PP in Shapley's game. GIGA-WoLF cannot converge while GA- SPP and IGA-PP converge to NE. GA-SPP has a slighter oscillation. Parameters: γ = 3, η = 0.001. Initial polices: (0.1, 0.8, 0.1) and (0.8, 0.1, 0.1). Table 2: Games with larger settings (b) A 2x3 Game C1 (3,3) (2,2) C2 (0,5) (1,1) C3 (1,-2) (-1,0) (a) Shapley's Game C1 (0,0) (0,1) (1,0) C2 (1,0) (0,0) (0,1) C3 (0,1) (1,0) (0,0) R1 R2 R1 R2 R3 Figure 6: Following GA-SPP, action probabilities of agents fail to converge in three player matching pennies. Parameters: η = 0.001, γ = 0.3. Initial polices: (0.1, 0.9), (0.4, 0.6) and (0.7, 0.3). 5.3 Problem games Although GA-SPP has better performance than other MAL algo- rithms, the convergence of GA-SPP is not perfect. As shown in Fig. 6, in the three player matching pennies, GA-SPP cannot converge with a constant prediction length. This failed case show the difficulties of MAL work and indicate that gradient method may not be the ideal way to handle a complex game. Because dynamic of gradient method in such game is not linearly, the chaotic phenomenon will occur. We may need different approaches to deal with such problems. In order to make MARL work effectively in more cases, it is important to analyze and solve these problems. 6 CONCLUSION This paper introduced a new gradient-based multi-agent learning algorithm, called gradient-ascent with shrinking policy prediction (GA-SPP). We proved Nash convergence of GA-SPP with a finite step size in three classes of general-sum games: m × n positive semi-definite games, a subclass of 2 × n general-sum games, and 2 × 2 general-sum games, respectively, which provide a stronger theoretical guarantee than existing gradient-based MAL algorithms. We also empirically verified the strong convergence property of GA-SPP with example games. In future work, we aim to relax as- sumptions of GA-SPP and extend it to stochastic games where each agent only has observations of their in-game payoffs and has no gradient information about other agents. Figure 5: Comparison between GA-SPP, and IGA-PP in a 2×3 game under different prediction lengths. IGA-PP's conver- gence to Nash Equilibrium is affected by prediction length, while GA-SPP can always converge to NE. Parameters: η = 0.001, γ = 0.01 in upper and γ = 0.1 in lower. Initial polices: (0.8, 0.2) and (0.1, 0.8, 0.1). Fig. 5 shows results of GA-SPP and IGA-PP in a 2×3 game under different prediction lengths. Although IGA-PP can converge, it does not converge to a Nash equilibrium. On the contrary, the strategies lead by GA-SPP successfully converge to Nash equilibrium under different prediction lengths. The essential reason is that GA-SPP projects the predicted strategies to a valid space at every step. By examining with different learning rates, we observe that GA- SPP often converges faster than GIGA-WoLF. A possible explanation is introduced in [27]. We do not show these results for sake of space. 05001000150020002500300035004000iteration steps00.10.20.30.40.50.60.70.80.91GA-SPP in Shapley's GamePr of row agent selecting 1st actionPr of row agent selecting 2nd actionPr of row agent selecting 3rd action0246810iteration steps×10400.10.20.30.40.50.60.70.80.91GIGA-WoLF in Shapley's GamePr of row player selecting 1st actionPr of row player selecting 2nd actionPr of row player selecting 3rd action05001000150020002500300035004000iteration steps00.10.20.30.40.50.60.70.80.91IGA-PP in Shapley's GamePr of row agent selecting 1st actionPr of row agent selecting 2nd actionPr of row agent selecting 3rd action00.511.522.53iteration steps×10500.10.20.30.40.50.60.70.80.91GA-SPP in game (h)Pr of row player selecting 1st actionPr of column player selecting 1st actionPr of column player selecting 2nd action0246810iteration steps×10400.10.20.30.40.50.60.70.80.91IGA-PP in game (h)Pr of row player selecting 1st actionPr of column player selecting 1st actionPr of column player selceting 2nd action00.511.522.53iteration steps×10400.10.20.30.40.50.60.70.80.9GA-SPP in game (h)Pr of row player selecting 1st actionPr of column player selecting 1st actionPr of column player selecting 2nd action0200040006000800010000iteration steps00.10.20.30.40.50.60.70.80.91IGA-PP in game (h)Pr of row player selecting 1st actionPr of column player selecting 1st actionPr of column player selceting 2nd action0500100015002000iteration steps00.20.40.60.81GA-SPP in Three Player Matching PenniesPr of fisrt agent selecting 1st actionPr of second agent selecting 1st actionPr of third agent selecting 1st actionREFERENCES [1] Sherief Abdallah and Victor R. Lesser. 2008. A multiagent reinforcement learning algorithm with non-linear dynamics. Journal of Artificial Intelligence Research 33, 1 (2008), 521 -- 549. [2] Anatoly Antipin. 2003. Extragradient approach to solution of two person non-zero sum games. Optimization And Optimal Control. 1 -- 28 pages. [3] A. S. Antipin. 1995. The convergence of proximal methods to fixed points of extremal mappings and estimates of their rate of convergence. Computational Mathematics and Mathematical Physics 35, 5 (1995), 539 -- 551. [4] Bikramjit Banerjee and Jing Peng. 2007. Generalized multiagent learning with performance bound. Autonomous Agents and Multi-Agent Systems 15, 3 (2007), 281 -- 312. [5] Michael Bowling. 2005. Convergence and no-regret in multiagent learning. In [6] Michael Bowling and Manuela Veloso. 2001. Convergence of gradient dynamics Advances in neural information processing systems. 209 -- 216. with a variable learning rate. In ICML. 27 -- 34. [7] Branislav BoÅąanskÃ¡, Viliam LisÃ¡, Marc Lanctot, JiÅŹÃŋ ÄŇermÃąk, and Mark H.M. Winands. 2016. Algorithms for computing strategies in two-player simultaneous move games. Artificial Intelligence 237 (2016), 1 -- 40. [8] Johanne Cohen, Amélie Héliou, and Panayotis Mertikopoulos. 2017. Learning with bandit feedback in potential games. In Proceedings of the 31th International Conference on Neural Information Processing Systems. [9] Jacob W. Crandall. 2014. Towards minimizing disappointment in repeated games. Journal of Artificial Intelligence Research 49, 1 (2014), 111 -- 142. [10] Jacob W. Crandall and Michael A. Goodrich. 2011. Learning to compete, coordi- nate, and cooperate in repeated games using reinforcement learning. Machine Learning 82, 3 (2011), 281 -- 314. [11] Steven Damer and Maria L. Gini. 2017. Safely Using Predictions in General-Sum Normal Form Games. adaptive agents and multi agents systems (2017), 924 -- 932. [12] Constantinos Daskalakis, Alan Deckelbaum, and Anthony Kim. 2011. Near- optimal no-regret algorithms for zero-sum games. In Proceedings of the twenty- second annual ACM-SIAM symposium on Discrete Algorithms. Society for Indus- trial and Applied Mathematics, 235 -- 254. [13] Jakob Foerster, Ioannis Alexandros Assael, Nando de Freitas, and Shimon White- son. 2016. Learning to communicate with deep multi-agent reinforcement learn- ing. In Advances in Neural Information Processing Systems. 2137 -- 2145. [14] Jakob Foerster, Gregory Farquhar, Triantafyllos Afouras, Nantas Nardelli, and Shimon Whiteson. 2017. Counterfactual multi-agent policy gradients. arXiv preprint arXiv:1705.08926 (2017). [15] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. 2014. Generative adversarial nets. In Advances in neural information processing systems. 2672 -- 2680. [16] Walid Krichene, Benjamin Drighès, and Alexandre M Bayen. 2015. Online learn- ing of nash equilibria in congestion games. SIAM Journal on Control and Opti- mization 53, 2 (2015), 1056 -- 1081. [17] Joel Z Leibo, Vinicius Zambaldi, Marc Lanctot, Janusz Marecki, and Thore Graepel. 2017. Multi-agent reinforcement learning in sequential social dilemmas. In Proceedings of the 16th Conference on Autonomous Agents and MultiAgent Systems. International Foundation for Autonomous Agents and Multiagent Systems, 464 -- 473. [18] C. E. Lemke and J. T. Howson. 1964. Equilibrium Points of Bimatrix Games. J. Soc. Indust. Appl. Math. 12, 2 (1964), 413 -- 423. [19] Reshef Meir, Maria Polukarov, Jeffrey S. Rosenschein, and Nicholas R. Jennings. 2017. Iterative voting and acyclic games. Artificial Intelligence 252 (2017), 100 -- 122. [20] Luke Metz, Ben Poole, David Pfau, and Jascha Sohl-Dickstein. 2016. Unrolled Generative Adversarial Networks. CoRR abs/1611.02163 (2016). arXiv:1611.02163 http://arxiv.org/abs/1611.02163 [21] Ryan Porter, Eugene Nudelman, and Yoav Shoham. 2004. Simple search methods for finding a Nash equilibrium. In National Conference on Artifical Intelligence. [22] H. L. Prasad, L A Prashanth, and Shalabh Bhatnagar. 2015. Two-Timescale Algorithms for Learning Nash Equilibria in General-Sum Stochastic Games. adaptive agents and multi-agents systems (2015), 1371 -- 1379. [23] David Silver, Thomas Hubert, Julian Schrittwieser, Ioannis Antonoglou, Matthew Lai, Arthur Guez, Marc Lanctot, Laurent Sifre, Dharshan Kumaran, Thore Graepel, et al. 2017. Mastering chess and shogi by self-play with a general reinforcement learning algorithm. arXiv preprint arXiv:1712.01815 (2017). [24] Satinder P. Singh, Michael J. Kearns, and Yishay Mansour. 2000. Nash Conver- gence of Gradient Dynamics in General-Sum Games. In UAI '00: Proceedings of the 16th Conference in Uncertainty in Artificial Intelligence, Stanford University, Stanford, California, USA, June 30 - July 3, 2000. 541 -- 548. [25] Sainbayar Sukhbaatar, Rob Fergus, et al. 2016. Learning multiagent communica- tion with backpropagation. In Advances in Neural Information Processing Systems. 2244 -- 2252. [26] Ardi Tampuu, Tambet Matiisen, Dorian Kodelja, Ilya Kuzovkin, Kristjan Kor- jus, Juhan Aru, Jaan Aru, and Raul Vicente. 2017. Multiagent cooperation and competition with deep reinforcement learning. PloS one 12, 4 (2017), e0172395. [27] Chongjie Zhang and Victor R Lesser. 2010. Multi-Agent Learning with Policy Prediction.. In AAAI.
1803.09792	1	1803	2018-03-26T18:54:40	Min-Max Tours for Task Allocation to Heterogeneous Agents	[ "cs.MA", "cs.RO", "eess.SY" ]	We consider a scenario consisting of a set of heterogeneous mobile agents located at a depot, and a set of tasks dispersed over a geographic area. The agents are partitioned into different types. The tasks are partitioned into specialized tasks that can only be done by agents of a certain type, and generic tasks that can be done by any agent. The distances between each pair of tasks are specified, and satisfy the triangle inequality. Given this scenario, we address the problem of allocating these tasks among the available agents (subject to type compatibility constraints) while minimizing the maximum cost to tour the allocation by any agent and return to the depot. This problem is NP-hard, and we give a three phase algorithm to solve this problem that provides 5-factor approximation, regardless of the total number of agents and the number of agents of each type. We also show that in the special case where there is only one agent of each type, the algorithm has an approximation factor of 4.	cs.MA	cs	Min-Max Tours for Task Allocation to Heterogeneous Agents Amritha Prasad∗, Han-Lim Choi† and Shreyas Sundaram∗∗ Abstract We consider a scenario consisting of a set of heterogeneous mobile agents located at a depot, and a set of tasks dispersed over a geographic area. The agents are partitioned into different types. The tasks are partitioned into specialized tasks that can only be done by agents of a certain type, and generic tasks that can be done by any agent. The distances between each pair of tasks are specified, and satisfy the triangle inequality. Given this scenario, we address the problem of allocating these tasks among the available agents (subject to type compatibility constraints) while minimizing the maximum cost to tour the allocation by any agent and return to the depot. This problem is NP-hard, and we give a three phase algorithm to solve this problem that provides 5-factor approximation, regardless of the total number of agents and the number of agents of each type. We also show that in the special case where there is only one agent of each type, the algorithm has an approximation factor of 4. 1 Introduction Multi-robot systems will play a large role in a variety of modern and future applications including exploration, surveillance, search and rescue operations, cooperative control and operations in haz- ardous environments. In order to effectively utilize these multi-robot systems in such applications, it is necessary to allocate an appropriate set of tasks to each robot or agent in the system. Such problems have been widely considered in the literature [1–5], most typically for the case where all agents are the same. However, future multi-robot systems are also projected to have a large amount of diversity in terms of the capabilities of the agents, and the applications will consist of tasks that can only be done by agents that possess certain capabilities [6–10]. We address this problem in this paper, namely allocating tasks efficiently to heterogeneous agents while meeting task-agent compatibility constraints. Gerkey and Mataric [11] gives a taxonomy of task allocation in multi-robot systems. Prorok et. al. [6] distributes a swarm of heterogeneous robots (of different types, with each type having different traits) among a set of tasks that require specialized capabilities in order to be completed. They optimize the transition rates for each type of robot so that the desired trait distribution is reached, but do not consider travel time between tasks. In this paper, we focus on task allocation to heterogeneous robots (of different types, with different capabilities) such that the task-agent compatibility constraints are satisfied and the maximum tour cost incurred by any agent is mini- mized. The existing literature has several works on task allocation to agents where the total tour cost (i.e., sum of tour costs incurred by all agents) is minimized. Bektas [12] gives an overview on ∗∗Amritha Prasad and Shreyas Sundaram are with the School of Electrical and Computer Engineering at Purdue University. Email: {prasad20,sundara2}@purdue.edu †Han-Lim Choi is with the Department of Aerospace Engineering at Korea Advanced Institute of Science and Tech- nology. Email: [email protected] the work on the Traveling Salesperson Problem (TSP) with multiple (homogeneous) salespersons, where the sum of the cost of tours by all salespersons is minimized. Frieze [13] extends Christofides' 3 2 -approximation algorithm [14] for the single Traveling Salesperson Problem to the k-person Trav- eling Salesperson Problem (where all salespersons are identical and start at the same node). Xu and Rodrigues [15] provide a 3 2 -approximation algorithm for multiple TSP with a fixed number of depots. Malik, Rathinam and Darbha [16] provide a 2-approximation algorithm for a generalized multiple depot, multiple TSP with symmetric travel costs. Yadlapalli and Rathinam [17] provide a 3-approximation algorithm for a case where two heterogeneous vehicles with associated travel costs start from distinct initial locations and are jointly required to visit a set of targets. Bae and Rathinam [18] provide a 2-approximation algorithm for a special case of the above problem where the travel cost for one vehicle is at most that of the other. As opposed to the above works that minimize the sum of the costs of each agent, the following works focus on minimizing the maximum tour cost incurred by any agent in a group of homogeneous agents. Frederickson, Hecht and Kim [19] give tour splitting heuristics for the k-person variants of the Traveling Salesperson Problem. Given a tour constructed (using an F -approximation algorithm) on a set of nodes, they provide a 1 + F − 1 k factor algorithm that outputs k subtours such that the cost of the largest subtour is minimized. Yu and Liu [20] consider networks with multiple depots (or start nodes) and provide a 6-approximation factor algorithm. Even et. al. [21] and Arkin et. al. [22] provide a 4-approximation algorithm for a problem known as the min-max tree cover problem. This was improved by Khani and Salavatipour [23] who proposed a 3-approximation algorithm. The recent work by Sadeghi and Smith [4] give a decentralized auction based multi- robot task allocation where the objective is to minimize the total time taken to perform all tasks. They consider a case of heterogeneous robots, where the heterogeneity is captured by constraints on the motion of the robots. The above works focus either on minimizing the total cost incurred by all agents or on mini- mizing the maximum cost incurred by any agent in a set of agents, where all agents have the same functionality. In this paper, we combine the idea of heterogeneity in agent functionality with that of minimizing the maximum cost incurred by any agent. Specifically, consider a scenario where a set of tasks at different locations need to be executed; however, not all tasks can be done by all agents and certain task-agent compatibility constraints must be satisfied. Agents are partitioned into different types based on functionality of the agents. Tasks are partitioned into sets of type- specific tasks and generic tasks, where type-specific tasks can only be performed by agents of a given type and generic tasks can be performed by any agent. To capture this scenario, we present the Heterogeneous Task Allocation Problem (HTAP) which aims to allocate a set of tasks among heterogeneous agents such that the maximum time to complete tasks by any agent is minimized. This is an important metric, especially when the tasks to be performed are time critical or when the quality of service is characterized by maximum delays. We find tours for each agents that start and end at a start node (e.g., home base), such that the task- agent compatibility constraints are met, and approximately minimize the maximum cost incurred by any agent to complete its tour. The rest of this paper is organized as follows. In the next section, we state our problem formulation. We then present a naive algorithm and show that the approximation factor of this algorithm, although bounded, increases linearly with the number of agents. This motivates the need to develop better algorithms that perform well as the number of agents increase. We present two such algorithms in the subsequent section. We first develop an algorithm called CycleSplit. We leverage insights learned from CycleSplit and use it as a baseline to develop an alternate algorithm which we call HeteroMinMaxSplit. We show that these are 5-approximation algorithms for HTAP. We also show that in the special case where each of the heterogeneous agents are distinct, the proposed algorithms haves an approximation factor of 4. 2 Problem Statement let mi be the number of agents of type i, with Pm Consider a set of tasks T that are required to be completed by a set of k heterogeneous agents A = {A1, A2,..., Ak}. Each agent is one of m types. Let f : {1, . . . , k} → {1, . . . , m} be a function that takes an agent number as input and outputs the type of that agent. For each i ∈ {1, 2, . . . , m}, i=1 mi = k. Let T be composed of two broad classes of tasks: type-specific tasks and generic tasks. Type-specific tasks can be performed only by a specific type of agent, whereas generic tasks can be performed by any agent. Let T0 denote the set of generic tasks and Ti, 1 ≤ i ≤ m, denote the set of type-specific tasks that can be performed by agents of type i. Thus, T = T0 ∪ (∪m i=1Ti). Let all agents start at a node vs, called the start node (or home base). Consider a complete graph G(V, E) with vertex set V = T ∪ {vs} and edge set E = {(u, v) : u, v ∈ V, u 6= v}. Let each edge e = (u, v) ∈ E have a weight d(u, v) given by the distance between the nodes u and v. Let the direct travel cost between two nodes be the weight of the edge connecting the two nodes in G. We assume the distances satisfy the triangle inequality. The cost of executing a task is assumed to be very small compared to travel costs and is hence neglected. A tour on a subset of nodes V ′ ⊆ V is a closed path from vs through all nodes in V ′ (in some order) ending at vs. All tours start and end at the start vertex vs. The cost of the tour is defined as the sum of weights of all edges on that tour. Let C ∗(V ′) denote the tour cost of the optimal tour on set V ′. The objective of the allocation problem is to partition the set of tasks T among the agents subject to the task-agent compatibility constraints, such that the maximum cost among all agents to tour their allocated tasks is minimized. This is framed as the Heterogeneous Task Allocation Problem given below. Heterogeneous Task Allocation Problem (HTAP): min S1,S2,...,Sk⊆T subject to ∪k max 1≤j≤k C ∗(Sj) j=1 Sj = T, Sj = Vj ∪ Rj, ∀j ∈ {1, 2, . . . , k}, Vj ⊆ Tf (j), Rj ⊆ T0. In the above formulation, Sj is the task set allocated to agent Aj, for 1 ≤ j ≤ k. The first constraint implies that each task must be executed by some agent. The remaining conditions state that the task set allocated to each agent Aj is a union of type-specific tasks Vj (which is a subset of Tf (j), where f (j) is the type of agent Aj) and generic tasks Rj (which is a subset of generic tasks T0). Remark 1. Although we do not explicitly specify the condition that a task cannot be performed by more than one agent, minimization of the objective function implicitly ensures that the optimal solution does not have any node appearing in multiple subsets due to the triangle inequality. Any instance of the Traveling Salesperson Problem (TSP) can be trivially reduced to an instance of the Heterogeneous Task Allocation Problem (HTAP) by setting the number of agents k = 1 and all tasks to be generic. Thus, the HTAP is trivially NP-Hard, and has no polynomial time solution unless P = NP. Hence, in the rest of this paper we develop approximation algorithms for HTAP. An α-approximation algorithm for a problem is an algorithm that efficiently finds a solution within a factor α of the optimum solution for every instance of that problem. 3 A Naive Approximation Algorithm for the Heterogeneous Task Allocation Problem We start with the following simple algorithm to solve HTAP. In this naive algorithm, we allocate all tasks in T to a single agent A1. To every other agent, we allocate the type-specific tasks associated with that agent. Agent A1 computes a tour using some approximation algorithm (e.g., Christofides' algorithm) on set T (which includes type-specific tasks of all agents and all the generic tasks). Each other agent Aj, j 6= 1, computes a tour on its type-specific tasks using some approximation algorithm (e.g., Christofides' algorithm). Algorithm 1 describes this naive allocation. Algorithm 1 NaiveAllocation Algorithm 1: procedure NaiveAllocation(A, T, G) 2: 3: 4: Allocate type-specific tasks to each agent of that type. Allocate all generic tasks to agent A1. Compute tour (starting and ending at node vs) for the set of tasks allocated to each agent using an approximation algorithm. Return tours for each agent. 5: 6: end procedure We provide the approximation factor of this algorithm in the following theorem. Theorem 1. Suppose the algorithm that is used to compute the tour for each agent in line 4 of NaiveAllocation algorithm has approximation factor α. Then, NaiveAllocation is an αk approximation algorithm for the Heterogeneous Task Allocation Problem (HTAP). Proof. For 1 ≤ j ≤ k, let S∗ allocation for HTAP. Note that S∗ given by S∗ j where R∗ optimal allocation policy. Note that R∗ j ⊆ Tf (j) ∪ R∗ j denote the set of tasks allocated to agent Aj under the optimal j ⊆ T ∀j, S∗ j is j is the subset of tasks in T0 allocated to agent Aj under the i = ∅ if j 6= i and ∪k j = T . The set S∗ j ∩ S∗ j=1S∗ j ⊆ T0 ∀j, R∗ j ∩ R∗ i = ∅ for j 6= i and ∪k j=1R∗ j = T0. For 1 ≤ j ≤ k, let Sj denote the set of tasks allocated to agent Aj by NaiveAllocation. Note that Sj ⊆ T ∀j, S1 = T and Sj = Tf (j), j 6= 1. Let C ∗(·) denote the cost to optimally tour a given set of tasks and let CN A(·) denote the tour costs returned by the NaiveAllocation algorithm. The approximation ratio R for the NaiveAllocation algorithm is given by R = max 1≤j≤k max 1≤j≤k CN A(Sj) C ∗(S∗ j ) = CN A(S1) C ∗(S∗ j ) max 1≤j≤k = CN A(T ) max 1≤j≤k C ∗(S∗ j ) ≤ αC ∗(T ) max 1≤j≤k C ∗(S∗ j ) ≤ α(Pk max 1≤j≤k j=1 C ∗(S∗ j )) C ∗(S∗ j ) ≤ αk. The inequality CN A(T ) ≤ αC ∗(T ) follows from the fact that an α-approximation algorithm is used j=1 C ∗(Sj) as Sj, 1 ≤ j ≤ k, to compute the tour in line 4 of NaiveAllocation. Also, C ∗(T ) ≤Pk form a partition of T and the triangle inequality holds. The previous theorem shows that the approximation factor of even a naive algorithm as the one described by Algorithm 1 is bounded. However, the bound grows linearly with the number of agents k. This motivates us to look for more efficient algorithms to solve the task allocation problem at hand, in particular, algorithms that perform well as the number of agents become large. We provide two constant factor algorithms in the following section. 4 Constant Factor Approximation Algorithms for the Heteroge- neous Task Allocation Problem 4.1 Cycle Splitting Approach Consider an instance of HTAP. In order to find an allocation of tasks to agents, we must allocate type-specific tasks among agents of the required type and allocate generic tasks among all agents. We approach the problem by handling these two allocations separately. Given a set of tasks Ti and a number k, let TaskSplitter be any algorithm that splits the set of tasks Ti into k sub-tours {Ti1, Ti2, ..., Tik} within some factor β of the optimal split (in the min-max sense). For example, Frederickson, Hecht and Kim [19] give a tour splitting heuristic that takes a tour on a set of locations to be visited (first node of which is set as the start node), and a positive integer k as input and gives a set of k subtours (starting and ending at the start node) as the output. The cost of these subtours are within a factor of 1 + F − 1/k of the optimal min-max cost, where F is the approximation factor of the algorithm used to generate the initial tour on all input locations. Consider the following algorithm to allocate a given set of tasks T to a group of k heterogeneous agents. Algorithm 2 Min-Max Tour by CycleSplit Algorithm 1: procedure CycleSplit(A, T, G) 2: for each agent type i, 1 ≤ i ≤ m do Find Christofides' tour (starting and ending at vs) on the set of type-specific tasks of type i. Use TaskSplitter to get mi subtours (starting and ending at vs) on set Ti. Allocate one subtour to each agent of type i. end for Find Christofides' tour (starting and ending at vs) on the set of generic tasks T0. Use TaskSplitter to split tour on T0 into k subtours (starting and ending at vs), denoted by {R1, R2, . . . , Rk}. Allocate subtour Rj to agent Aj, 1 ≤ j ≤ k. Combine the type-specific task subtour and generic task subtour allocated to each agent. Return a tour for each agent. 3: 4: 5: 6: 7: 8: 9: 10: 11: 12: end procedure Theorem 2. CycleSplit is a 2β-approximation algorithm for the Heterogeneous Task Allocation Problem (HTAP), where β is the approximation factor of the algorithm TaskSplitter used in steps 4 and 8. Proof. For 1 ≤ j ≤ k, let S∗ HTAP. Then, S∗ allocated to agent Aj and R∗ j be the allocation of tasks to agent Aj under an optimal algorithm for is the subset of type-specific tasks j is the subset of generic tasks assigned to agent Aj. Let Sj = Vj ∪ Rj j , where V ∗ j j can be expressed as S∗ j = V ∗ j ∪ R∗ be the allocation to agent Aj by the CycleSplit algorithm, where Vj is the subset of type-specific tasks allocated to agent Aj and Rj is the subset of generic tasks allocated to agent Aj. Let C ∗(·) denote the cost to optimally tour a given set of tasks starting and ending from the node vs. Let CT S(Rj) and CT S(Vj) denote the cost of the subtours Rj and Vj returned by the TaskSplitter algorithm in steps 4 and 8 respectively. Let CCS(Sj) denote the cost of a tour on Sj returned by the CycleSplit algorithm. Thus, the approximation factor of the CycleSplit algorithm is given by R = ≤ ≤ max 1≤j≤k max 1≤j≤k max 1≤j≤k max 1≤j≤k max 1≤j≤k max 1≤j≤k CCS(Sj) C ∗(S∗ j ) CT S(Vj) C ∗(S∗ j ) CT S(Vj) C ∗(V ∗ j ) + + max 1≤j≤k max 1≤j≤k max 1≤j≤k max 1≤j≤k max 1≤j≤k ≤ {CT S(Vj) + CT S(Rj)} max 1≤j≤k C ∗(S∗ j ) CT S(Rj) C ∗(S∗ j ) CT S(Rj) , C ∗(R∗ j ) (1) where we use the facts that Sj = Vj ∪ Rj, S∗ j = V ∗ j ∪ R∗ j and that the triangle inequality holds. 1, V ′ 2, . . . , V ′ Consider a case with the same set of type-specific tasks as above, but with no generic tasks. Let V ′ k denote the set of tasks allocated to agents under an optimal allocation (in the sense of minimizing the maximum cost to tour any subset). Since the CycleSplit algorithm allocates type-specific tasks independent of generic tasks, the allocation under this algorithm will be V1, V2, . . . , Vk. Since the set of tasks with only type-specific tasks is a subset of the total set of tasks (including generic tasks), min-max cost for case with only type-specific tasks cannot exceed the min-max cost for the set of total tasks. Thus, max 1≤j≤k CT S(Vj) ≤ β max 1≤j≤k C ∗(V ′ j ). (2) The inequality in the equation above follows from the fact that TaskSplitter algorithm splits a tour into subtours that are within a factor β of the optimal min-max cost. By the same argument as before, min-max cost to optimally tour {V ′ j }, 1 ≤ j ≤ k cannot exceed the min-max cost to optimally tour {V ∗ j }, 1 ≤ j ≤ k, by the optimality of the partition {V ′ j }, 1 ≤ j ≤ k. Thus, max 1≤j≤k C ∗(V ′ j ) ≤ max 1≤j≤k C ∗(V ∗ j ). Combining equations (2) and (3), we get max 1≤j≤k CT S(Vj) ≤ max 1≤j≤k βC ∗(V ∗ j ). (3) (4) Using a similar procedure as before, this time comparing the allocation of a case with no type specific tasks against the set of all tasks, we get, max 1≤j≤k CT S(Rj) ≤ max 1≤j≤k βC ∗(R∗ j ). Substituting equations (4) and (5) in equation (1), we get R ≤ 2β. (5) (6) Frederickson, Hecht and Kim [19] provide an algorithm SPLITOUR1 that splits a tour (starting and ending at a start-node) on a set of nodes into k subtours each starting and ending at the start node. This algorithm has an approximation factor 1 + F − 1 k , where F is the approximation factor of the algorithm used for constructing the initial tour. Corollary 1. CycleSplit algorithm using the SPLITOUR algorithm from [19] for task splitting is a 5 − 2 k approximation algorithm for the Heterogeneous Task Allocation Problem (HTAP). Proof. The algorithm SPLITOUR is a 5 algorithm into k subtours (since Christofides provides an F = 3 tour). Equation (1) can be written as 2 − 1 k factor algorithm to split a tour using Christofides' 2 factor approximation for the initial R ≤ 5 2 − 1 max 1≤i≤m mi! + 5 2 − 1 k!. We can upper bound max 1≤i≤m mi with k. Thus, R ≤ 5 − 2 k . (7) (8) Corollary 2. In the special instance of HTAP where there is exactly one agent of each type (i.e., the heterogeneous agents are distinct), CycleSplit is a 4 − 1 k approximation factor algorithm. Proof. In this special case, mi = 1, 1 ≤ i ≤ m. Thus max 1≤i≤m mi = 1. So, equation (7) can be simplified as R ≤ 5 2 − 1! + 5 2 − 1 k! = 4 − 1 k . Remark 2. From Corollary 1, we see that regardless of the value of k, CycleSplit is a 5- approximation algorithm for HTAP. Furthermore, from Corollary 2, we can see that for the special instance of HTAP where all heterogeneous agents are distinct, CycleSplit is a 4-approximation algorithm. Note that the CycleSplit algorithm while using SPLITOUR as TaskSplitter in lines 4 and 8, splits a TSP tour on the set of tasks into k subtours evenly and allocates each subtour to one of the agents. While this approach works well when all agents have same the capacity (or balanced initial load), it may not perform as well when different agents have different capacity or initial load. In a case where agents have type-specific tasks that they must perform, each agent may have different task loads prior to allocation of generic tasks. In order to optimize the min-max tour cost, it is advantageous to take this into consideration. This motivates the formulation of an algorithm that takes into consideration the existing "load" of each agent from type-specific tasks when allocating generic tasks to agents. To address this, we propose a modified algorithm in the following section. 1Frederickson, Hecht and Kim call the algorithm k-SPLITOUR in their paper; however, we refer to it as SPLI- TOUR in order to avoid confusing k with the number of agents as we use it in this paper. 4.2 Min-Max Splitting Approach for Heterogeneous Agents Based on the intuition gained from the CycleSplit algorithm, we propose a modified algorithm called HeteroMinMaxSplit. This algorithm considers the fact that agents may have different "loads" after the allocation of type-specific tasks. Specifically, instead of splitting a tour on the set of generic tasks into nearly equal segments of a given size, the generic tasks are allocated to agents while taking into consideration the cost incurred by the agent to tour its specific tasks. In the HeteroMinMaxSplit algorithm, we allocate tasks to agents in three phases. • Phase 1: Type-specific task allocation • Phase 2: Generic task allocation (accounting for Phase 1 allocation) • Phase 3: Rebalancing tasks within agents of each type. In Phase 1, type-specific tasks are allocated among agents of the associated type as in Cy- cleSplit. The generic tasks are allocated among all agents in Phase 2. Unlike in CycleSplit, however, the allocation during Phase 2 tries to balance the total cost incurred by agents by taking into account the existing allocation to agents after Phase 1. Phase 3 is a post-processing step aimed to further balance the load among agents. After Phase 2, all tasks allocated to agents of the same type can be done by all agents in that type. Thus, we pool all tasks allocated to agents of the type i into a set T ′ i into mi sub-tours (Algorithm SPLITOUR may be used for this purpose [19]). Balancing the load in this fashion can lower the min-max tour cost by distributing the load more evenly among agents, as we will show later. i and re-allocate them to agents of type i by splitting the Christofides' tour on T ′ We present the algorithm in two steps. We first give an algorithm HeteroSplit that takes an additional input λ, which we take to be a rational number. This algorithm finds tours (if they exist) for each agent such that cost for each agent to complete its tour is less than λ. The HeteroMinMaxSplit algorithm then performs a binary search on λ, running HeteroSplit in each iteration, to find the best set of tours for the agents. Algorithm 3 Heterogeneous Task Split within bound λ ⊲ Let λ be a rational parameter that denotes the desired upper bound for the tour length for any agent 1: procedure HeteroSplit(A, T, G, λ) 2: 3: 4: 5: 6: 7: 8: 9: ⊲ Phase 1: Type-specific task allocation for each agent type i, 1 ≤ i ≤ m do Find Christofides' tour (starting and ending at vs) on the set of type-specific tasks of type i. Use TaskSplitter to split the tour on Ti into mi subtours (starting and ending at vs). Allocate one subtour to each agent of type i. end for ⊲ Phase 2: Generic task allocation Mark all agents as free. Remove from G, all vertices in Ti, 1 ≤ i ≤ m and all edges incident on these vertices. Denote the resulting graph by G′. Mark all tasks in graph G′ as unallocated. Find Christofides' tour starting and ending at vs on nodes in G′. Denote the tour by H. Consider the next unallocated task, say t, along H starting from vs. For each free agent j, find cost to tour (starting and ending at vs) all tasks allocated to it along with t using Christofides' algorithm. Select agent with the minimum cost provided the minimum cost is less than λ. If no free agent can add the task to its tour without exceeding cost λ, return failure. Allocate t to the selected free agent. Keep allocating unallocated tasks along H to the selected agent as long as adding the task does not cause the agent's tour cost to exceed λ. Remove the tasks allocated in the previous step from the set of unallocated tasks and mark agent as busy. Go to step 10 if the set of unallocated tasks is non-empty and free agents remain. ⊲ Phase 3: Rebalancing within agent types for each agent type i, 1 ≤ i ≤ m do Deallocate tasks from all mi agents of type i. Let T ′ from agents of type i. Find Christofides' tour on T ′ Run TaskSplitter on H ′ to split it into mi subtours. Allocate one subtour to each agent of type i. i to get tour H ′. i denote the set of tasks deallocated 10: 11: 12: 13: 14: 15: 16: 17: 18: 19: end for Return tours for each agent. 20: 21: end procedure A binary search can be performed to find the optimal value for λ to be provided as a parameter to HeteroSplit. The value cannot be less than than twice the largest edge from the start node d(vs, t), where d(vs, t) is the distance from the start node vs to to any task location, i.e., 2 max t∈T the location of a task t. Also, λ cannot exceed max 1≤i≤k C(Ti) + C(T0), where C(·) is the cost to do a Christofides' tour on the given set of tasks. Hence, we can perform a binary search within this window to find the optimal value of λ. Algorithm 4 Min-Max Tour by HeteroMinMaxSplit Algorithm 1: procedure HeteroMinMaxSplit(A, T, G) 2: Do binary search in the interval [2 max t∈T d(vs, t), max 1≤i≤k C(Ti) + C(T0)] to find smallest value of λ for which HeteroSplit(A, T, G, λ) returns a set of valid tours. Return the set of tours returned by HeteroSplit(A, T, G, λ). 3: 4: end procedure 4.3 Bound on Approximation Ratio of HeteroMinMaxSplit Algorithm For the purpose of applying SPLITOUR from Frederickson, Kim and Hecht's work, we now provide a brief sketch of the algorithm SPLITOUR [19]. It takes as input a set of n nodes labeled v1 through vn (with the initial node as the start node) and a positive number k. First, the algorithm constructs a Christofides' tour on all the nodes. Let L be the cost of this tour and let cmax d(v1, vn). be the maximum direct distance of any node from the start node, i.e., cmax = max 1<i≤n For 1 ≤ j < k, it finds the largest vertex vp(j) along the tour such that the cost to traverse the tour from the start node to vp(j) does not exceed j k (L − 2cmax) + cmax. It returns k subtours (v1, . . . , vp(1), v1), (v1, vp(1)+1, . . . , v1), . . ., (v1, vp(k−1)+1, . . . , vn, v1), each of which has cost less than 1 k (L − 2cmax) + 2cmax. Frederickson, Hecht and Kim [19] show that if the length of the subtours do not exceed 1 ( 5 2 − 1 k ) of the optimal min-max cost. k (L − 2cmax) + 2cmax, then the min-max cost of the subtours is no worse than a factor The following proposition leverages the splitting heuristic summarized above to establish results on the performance of our algorithms. Theorem 3. HeteroMinMaxSplit is a (5 − 2 Task Allocation Problem (HTAP). k )-approximation algorithm for the Heterogeneous Proof. Let λ1 be the maximum cost among all the subtours returned by running SPLITOUR on Christofides' tours on type-specific tasks after line 6 of the CycleSplit algorithm, i.e., λ1 = CT S(Vj), where Vj denotes the set of type-specific tasks allocated to agent Aj and CT S(·) max 1≤j≤k denotes the cost associated with the subtour returned by SPLITOUR. Note that λ1 is also the maximum cost after Phase 1 of the HeteroSplit algorithm. Set λ to be λ1 + 1 Running SPLITOUR on a Christofides' tour on the set of generic tasks T0 to split it into k subtours will returns subtours of length (or cost) no more than 1 k (L − 2cmax) + 2cmax, where L is the length (or cost) of the initial Christofides' tour on T0, and cmax is the maximum direct distance of any node in T0 from the start node. No matter how each of these subtours are assigned to agents (such that each agent gets one subtour), the total cost is no more than λ1 + 1 k (L − 2cmax) + 2cmax. k (L − 2cmax) + 2cmax. Consider HeteroSplit with this value of λ. In Phase 1, HeteroSplit allocates type-specific tasks to agents, same as steps 2-6 of CycleSplit. The maximum cost of any agent's tour after Phase 1 is λ1. In Phase 2, the HeteroSplit algorithm computes a Christofides' tour on the set of generic tasks. The algorithm selects an agent that has minimum cost to complete its current allocated tasks in addition to the next task along the Christofides' tour on generic tasks. The algorithm then allocates tasks along the tour to the selected agent as long as the cost does not exceed λ. Regardless of which agent gets selected first, the set of tasks that are allocated to the selected agent in HeteroSplit will contain the tasks in the first subtour generated by SPLITOUR. The allocation by HeteroSplit to the first selected agent may contain more tasks than the first subtour of SPLITOUR, but not less. Thus, after allocating tasks to the first agent, the number of tasks left to be allocated to the remaining k − 1 agents in HeteroSplit is no more than the number of tasks left in the k − 1 subtours to be allocated to the remaining k − 1 agents in SPLITOUR. In SPLITOUR, after the allocation of the first subtour to the an agent, it is guaranteed that the remaining k − 1 subtours can be allocated to the remaining k − 1 agents, since the allocation is possible regardless of the order in which agents are selected. Given that the set of tasks left for allocation in HeteroSplit is no bigger than the set of tasks left to be allocated in SPLITOUR, it is guaranteed that the remaining tasks can be allocated to other k − 1 agents in HeteroSplit algorithm for this value of λ. Thus, by inducting on the number of agents to which tasks have been allocated along the initial tour, the HeteroSplit algorithm is guaranteed to return a feasible solution to HTAP for λ = λ1 + 1 k (L − 2cmax) + 2cmax. Since HeteroSplit checks tour cost against λ before further allocation of tasks to an agent, it guarantees that the tour costs for all agents is no more than λ. Thus, the approximation factor is given by, R ≤ ≤ max 1≤j≤k max 1≤j≤k max 1≤j≤k λ C ∗(S∗ j ) = λ1 C ∗(S∗ j ) max 1≤j≤k + 1 k (L − 2cmax) + 2cmax max 1≤j≤k C ∗(S∗ j ) CT S(Vj) C ∗(V ∗ j ) + 1 k (L − 2cmax) + 2cmax max 1≤j≤k C ∗(R∗ j ) (9) 1 max 1≤i≤m mi! + 5 2 − 1 k! ≤ 5 2 ≤ 5 − − 2 k , (10) (11) j are subsets of j . Equation (10) is obtained from the results of Frederickson, Hecht and Kim [19], summarized where equation (9) is given by the triangle inequality and the fact that V ∗ S∗ prior to this theorem. j and R∗ The HeteroMinMaxSplit algorithm performs a binary search over λ and is thus guaranteed k approximation to find the λ used above by HeteroSplit. Thus, HeteroMinMaxSplit is a 5− 2 algorithm for HTAP. Corollary 3. HeteroMinMaxSplit is a (4− 1 of HTAP where all agents are distinct. k )-approximation algorithm for the special instance Proof. In this special instance, mi = 1, for 1 ≤ i ≤ m. Substituting in equation (10), we get the approximation factor R ≤ 4 − 1 k . From Theorem 3, we see that HeteroMinMaxSplit is a 5-approximation algorithm regardless of the value of k. In this special instance where all heterogeneous agents are distinct, we see from Corollary 3 that HeteroMinMaxSplit is a 4-approximation algorithm regardless of the value of k. 4.4 Examples Illustrating HeteroMinMaxSplit and CycleSplit Algorithms HeteroMinMaxSplit takes into consideration the "load" from type-specific tasks while allocating generic tasks and can outperform CycleSplit when the type-specific tasks are not uniformly distributed among agents. We illustrate this through the following examples. Example 1. Consider the scenario where three agents are collectively required to do nine tasks. Agent A1 is a type 1 agent, A2 is a type 2 agent and A3 is a type 3 agent. All agents are located at node vs. The start node vs is located at distance d′ > 2 from nodes V1 and V2 as shown in Figure 1. Nodes A, B and C are located at unit distance from node V1, and node V2 is located at a distance d > 3 (such that the triangle inequality is satisfied) from node C as shown in Figure 1. Distances between nodes not directly connected are defined as the sum of distances along the shortest path traversed to reach the node (the graph can be considered as a road network that the agents can traverse). vs d′ d V2 t9 d′ C t5 t6 A t1 t2 t7 t8 V1 1 1 B 1 t3 t4 Figure 1: Task locations in Example 1 Tasks in set T0 = {t1, t2, t3, t4, t5, t6} are generic tasks. Tasks {t1, t2} are located at node A, {t3, t4} are located at node B, and {t5, t6} are located at node C. Agent A1 has one type-specific task t7, i.e. T1 = {t7}. Agent A2 has one type-specific task t8, i.e. T2 = {t8}. Tasks t7 and t8 are located at node V1. Agent A3 has one type-specific task t9, i.e., T3 = {t9} and t9 is located at node V2. Both CycleSplit and HeteroMinMaxSplit algorithms first allocate the type-specific tasks to the respective agents. Thus, agent A1 is allocated task t7, agent A2 is allocated task t8 and agent A3 is allocated task t9. CycleSplit algorithm splits the generic tasks at nodes A, B and C to the three agents. Thus, agent A1 gets allocated tasks {t1, t2}, agent A2 gets allocated tasks {t3, t4}, and agent A3 gets allocated tasks {t5, t6}. The min-max cost is thus max{2d′ + 2, 2d′ + 2, 2d′ + d + 1} = 2d′ + d + 1. The HeteroMinMaxSplit algorithm allocates tasks in Phase 2 based on the allocation done in Phase 1. Thus, for λ = 2d′ + 4 < 2d′ + d + 1, HeteroSplit allocates all tasks at the nodes A, B and C to agents A1 and A2 (which already travel to node V1 to do their type-specific tasks). Thus, tasks {t1, t2} located at node A are allocated to agent A1, and tasks {t3, t4, t5, t6} located at nodes B and C are allocated to agent A2. Since only one agent of each type is present, agents retain their allocation after Phase 3. The min-max cost is thus max{2d′ + 2, 2d′ + 4, 2d′} = 2d′ + 4. For large values of d′ and d, HeteroMinMaxSplit gives a factor of 2 over CycleSplit on the min-max tour cost. Note that in the above example, Phase 3 does not improve the min-max cost. We illustrate the benefit of Phase 3 (rebalancing "load" between agents of a given type) of the HeteroMinMaxS- plit algorithm in the following example. Example 2. Consider a scenario with two type 1 agents A1 and A2 located at the start node vs. Let type-specific tasks for type 1 agents T1 = {t1, t2} be located at node A, which is at a distance of 1 unit from vs as shown in Figure 2. Generic tasks T0 = {t3, t4} are located at node B which is at a distance of 1 from vs as shown in Figure 2. A 1 vs B 1 t1 t2 t3 t4 Figure 2: Task locations in Example 2 In Phase 1, one type-specific task gets allocated to each of the agents, say t1 gets allocated to agent A1 and t2 gets allocated to agent A2. In Phase 2, one generic task gets allocated to each of the agents, say t3 gets allocated to agent A1 and t4 gets allocated to agent A2. So at the end of Phase 2, both agents need to visit nodes A and B to complete their allocated tasks. In Phase 3, all tasks allocated in the previous phases to agents of the same type are deallocated and redistributed amongst them. In this phase, tasks {t1, t2, t3, t4} are pooled and the TaskSplitter algorithm SPLITOUR is run on this set of tasks. Thus, one agent, say A1, gets tasks at node A and the other agents A2 gets tasks at node B. In this example, Phase 3 reduces the tour cost from 4 to 2 for both the agents, thus reducing the min-max cost by a factor of 2. Example 3. Consider two agents: A1 of type 1 and A2 of type 2. Task t1 is of type 1 and must be performed by agent A1, and tasks t2, t3 are generic tasks that can be done by either of the agents, i.e., T1 = {t1}, T2 = ∅, T0 = {t2, t3}. Task t1 is located at node A. Tasks t2 and t3 are located at node B. Agents start from start node vs. A t1 vs 1 B 1 t2 t3 Figure 3: Task locations in Example 3 CycleSplit algorithm allocates type-specific task t1 to agent A1 and splits generic tasks {t2, t3} among agents A1 and A2. Hence, A1 is allocated tasks {t1, t2} and A2 is allocated {t3}. The tour costs for agents A1 and A2 are 4 and 2 respectively, and thus, the min-max tour cost is max{4, 2} = 4. HeteroMinMaxSplit algorithm allocates type-specific task t1 to agent A1 in Phase 1. In Phase 2, generic tasks {t2, t3} are split among agents A1 and A2 based on remaining capacity. For λ = 2, HeteroSplit allocates {t1} to agent A1 and {t2, t3} to agent A2. The min-max tour cost for the tour returned by HeteroMinMaxSplit is thus max{2, 2} = 2. In this example, HeteroMinMaxSplit algorithm gives an improvement by a factor of 2 over CycleSplit. The following example extends it to a general case with k agents. Example 4. Consider the following generalization of the previous example for k agents. Agents {A1, A2, . . . , Ak} are located at start node vs. Tasks Ti = {ti} located at node Vi has to performed by Ai for i = 1, 2, . . . , k − 1, and tasks T0 = {tk, tk+1, . . . , t2k−1} located at node Vk are generic tasks and can be completed by any agent. Let d(·, ·) denote the distance function. The distance between the start node vs and any other node is 1 unit, i.e., d(vs, Vi) = 1, for 1 ≤ i ≤ k. The distance between two nodes Vi and Vj is given by d(Vi, Vj) = d(Vi, vs)+d(vs, Vj) = 2 if i 6= j and zero if i = j. In this case, CycleSplit algorithm allocates tasks {ti, tk+i−1} for agents Ai, 1 ≤ i ≤ k − 1 and task {t2k−1} to agent Ak. This allocation has a min-max tour cost of 4. HeteroMinMaxSplit allocates task {ti} to agent Ai for 1 ≤ i ≤ k − 1 and tasks {tk, tk+1, . . . , t2k−1} to agent Ak. This allocation brings down the min-max tour cost to 2. 5 Summary In this work, we considered the Heterogeneous Task Allocation Problem (HTAP) where we aim to allocate tasks to heterogeneous agents subject to agent-task compatibility while minimizing the min-max tour cost. We provide two approximation algorithms to solve the Heterogeneous Task Allocation Problem (HTAP). We first propose a 2β approximation algorithm CycleSplit, where β is the approximation factor of the algorithm used by CycleSplit to split tours. We then use the CycleSplit algorithm to develop the HeteroMinMaxSplit algorithm which has an approximation ratio 5 − 2 k , where k is the number of agents available. The HeteroMinMaxSplit algorithm works in three phases and allocates tasks to agents in a "balanced" way. An interesting extension to the work would be to study the case of allocation when tasks can be done by different subsets of agents (as opposed to agents of only a certain type). References [1] K. Sundar and S. Rathinam, "An exact algorithm for a heterogeneous, multiple depot, multiple traveling salesman problem," in Unmanned Aircraft Systems (ICUAS), 2015 International Conference on. IEEE, 2015, pp. 366–371. [2] H.-L. Choi, L. Brunet, and J. P. How, "Consensus-based decentralized auctions for robust task allocation," IEEE Transactions on Robotics, vol. 25, no. 4, pp. 912–926, 2009. [3] L. Bertuccelli, H.-L. Choi, P. Cho, and J. How, "Real-time multi-uav task assignment in dy- namic and uncertain environments," in AIAA Guidance, Navigation, and Control Conference, 2009, p. 5776. [4] A. Sadeghi and S. L. Smith, "Heterogeneous task allocation and sequencing via decentralized large neighborhood search," Unmanned Systems, vol. 5, no. 02, pp. 79–95, 2017. [5] F. Bullo, E. Frazzoli, M. Pavone, K. Savla, and S. L. Smith, "Dynamic vehicle routing for robotic systems," Proceedings of the IEEE, vol. 99, no. 9, pp. 1482–1504, 2011. [6] A. Prorok, M. A. Hsieh, and V. Kumar, "Adaptive distribution of a swarm of heterogeneous robots," Acta Polytechnica, vol. 56, no. 1, pp. 67–75, 2016. [7] E. G. Jones, B. Browning, M. B. Dias, B. Argall, M. Veloso, and A. Stentz, "Dynamically formed heterogeneous robot teams performing tightly-coordinated tasks," in Robotics and Au- tomation, 2006. ICRA 2006. Proceedings 2006 IEEE International Conference on. IEEE, 2006, pp. 570–575. [8] T. H. Labella, M. Dorigo, and J.-L. Deneubourg, "Division of labor in a group of robots inspired by ants' foraging behavior," ACM Transactions on Autonomous and Adaptive Systems (TAAS), vol. 1, no. 1, pp. 4–25, 2006. [9] R. Kilgore, K. Harper, C. Nehme, and M. Cummings, "Mission planning and monitoring for heterogeneous unmanned vehicle teams: A human-centered perspective," in AIAA Infotech@ Aerospace 2007 Conference and Exhibit, 2007, p. 2788. [10] Office of Naval Research, "Autonomous systems innovation summit final report," Office of Naval Research, Tech. Rep., 2008. [11] B. P. Gerkey and M. J. Matari´c, "A formal analysis and taxonomy of task allocation in multi- robot systems," The International Journal of Robotics Research, vol. 23, no. 9, pp. 939–954, 2004. [12] T. Bektas, "The multiple traveling salesman problem: an overview of formulations and solution procedures," Omega, vol. 34, no. 3, pp. 209 – 219, 2006. [13] A. M. Frieze, "An extension of Christofides heuristic to the k-person travelling salesman prob- lem," Discrete Applied Mathematics, vol. 6, no. 1, pp. 79–83, 1983. [14] N. Christofides, "Worst-case analysis of a new heuristic for the travelling salesman problem," DTIC Document, Tech. Rep., 1976. [15] Z. Xu and B. Rodrigues, "A 3/2-approximation algorithm for the multiple TSP with a fixed number of depots," INFORMS Journal on Computing, vol. 27, no. 4, pp. 636–645, 2015. [16] W. Malik, S. Rathinam, and S. Darbha, "An approximation algorithm for a symmetric gen- eralized multiple depot, multiple travelling salesman problem," Operations Research Letters, vol. 35, no. 6, pp. 747–753, 2007. [17] S. Yadlapalli, S. Rathinam, and S. Darbha, "3-approximation algorithm for a two depot, heterogeneous traveling salesman problem," Optimization Letters, vol. 6, no. 1, pp. 141–152, 2012. [18] J. Bae and S. Rathinam, "A primal-dual approximation algorithm for a two depot heteroge- neous traveling salesman problem," Optimization Letters, vol. 10, no. 6, pp. 1269–1285, 2016. [19] G. N. Frederickson, M. S. Hecht, and C. E. Kim, "Approximation algorithms for some routing IEEE, problems," in Foundations of Computer Science, 1976., 17th Annual Symposium on. 1976, pp. 216–227. [20] W. Yu and Z. Liu, "Improved approximation algorithms for min-max and minimum vehicle Springer, routing problems," in International Computing and Combinatorics Conference. 2015, pp. 147–158. [21] G. Even, N. Garg, J. Konemann, R. Ravi, and A. Sinha, "Min–max tree covers of graphs," Operations Research Letters, vol. 32, no. 4, pp. 309–315, 2004. [22] E. M. Arkin, R. Hassin, and A. Levin, "Approximations for minimum and min-max vehicle routing problems," Journal of Algorithms, vol. 59, no. 1, pp. 1–18, 2006. [23] M. R. Khani and M. R. Salavatipour, "Improved approximation algorithms for the min-max tree cover and bounded tree cover problems," Algorithmica, vol. 69, no. 2, pp. 443–460, 2014.
1207.2697	1	1207	2012-07-11T16:22:08	Genetic agent approach for improving on-the-fly web map generalization	[ "cs.MA", "cs.CG", "cs.NE" ]	The utilization of web mapping becomes increasingly important in the domain of cartography. Users want access to spatial data on the web specific to their needs. For this reason, different approaches were appeared for generating on-the-fly the maps demanded by users, but those not suffice for guide a flexible and efficient process. Thus, new approach must be developed for improving this process according to the user needs. This work focuses on defining a new strategy which improves on-the-fly map generalization process and resolves the spatial conflicts. This approach uses the multiple representation and cartographic generalization. The map generalization process is based on the implementation of multi- agent system where each agent was equipped with a genetic patrimony.	cs.MA	cs	International Journal of Information Technology Convergence and Services (IJITCS) Vol.2, No.3, June 2012 GENETIC AGENT APPROACH FOR IMPROVING ON- THE-FLY WEB MAP GENERALIZATION Brahim lejdel1 and Okba kazar2 1Computer science departement, University of el-Oued, Algeria. [email protected] 2Computer science department, University of Biskra, Algeria. [email protected] ABSTRACT The utilization of web mapping becomes increasingly important in the domain of cartography. Users want access to spatial data on the web specific to their needs. For this reason, different approaches were appeared for generating on-the-fly the maps demanded by users, but those not suffice for guide a flexible and efficient process. Thus, new approach must be developed for improving this process according to the user needs. This work focuses on defining a new strategy which improves on-the-fly map generalization process and resolves the spatial conflicts. This approach uses the multiple representation and cartographic generalization. The map generalization process is based on the implementation of multi- agent system where each agent was equipped with a genetic patrimony. KEYWORDS On-the-fly maps generalization, genetic agent, spatial constraints, genetic algorithm, multiple representation. 1. INTRODUCTION The web mapping has known great growth in parallel of the rapid development of the internet. To provide on-the-fly web mapping to the user, the process of on-the-fly map generalization must rely on fast, effective, and powerful methods. An principals challenge of such on-the-fly maps generalization is to offer the user a spatial data in real time and in height quality , it must allow also to solve spatial conflicts that may appear between objects especially due to lack of space on display screens [1]. Many research works have proposed solutions to on-the-fly map generalization, an important work that was done by [1], he proposed a new approach based on the implementation of a multi- agent system for the generation of maps on the fly and the resolution of spatial conflicts. This approach is based on the use of multiple representation and cartographic generalization. Also, to improve the process of on-the-fly map generalization, another approach was proposed in [2] which based on a new concept called SGO (Self-generalizing object). In the same context and for reducing the spatial conflicts in the map, a good method was proposed in [3] which based on the genetic algorithm. The work presented in this paper comes under the research favouring the use of agents were equipped with genetic patrimony; it based on the approach presented in [4]. It uses the genetic agent in order to generate data on arbitrary scales thanks to an on-the-fly map generalization process. The proposed approach exclusively aims to solve the following problems related to on-the-fly web mapping applications: DOI : 10.5121/ijitcs.2012.2301 01 International Journal of Information Technology Convergence and Services (IJITCS) Vol.2, No.3, June 2012 - How can we adapt the contents of maps to users' needs and improve the time of generalization process of spatial data, which improve in result the transfer time of data map? - How can we solve spatial conflicts in order to improve the quality of map? The paper is structured as follows. In the part 2, we present on-the-fly map generalization process and its problems. Part 3 briefly reminds the approaches of on-the-fly map generalization. Part 4 presents the proposed approach for improving on-the-fly web mapping generalization. In part 5, we present the experimentation and some results. Finally, part 6 concludes and cites some perspectives of this work. 2. ON-THE-FLY WEB MAP GENERALIZATION The on-the-fly web map generalization is defined as; the creation in real-time and according to the user's request, of a cartographic product appropriate to its scale and purpose, from a largest- scale database. The main characteristics of on-the-fly web mapping are [1]: - Required maps must be generated in real-time. - Generation of a temporary and reduced scale dataset for visualization purposes from the database [5] in order to use the computer's memory efficiently [6]. - A real-time map generation process has to take into account users' preferences and contexts. - A real-time map generation process must adapt maps' contents to display space and resolution of display media as well as to the contextual use of these maps. - The scale and theme of the map are not predefined [6]. - There is no way to verify the quality of the final map that will be sent to the user. The main problems linked to on-the-fly map generalization are the time of delivering the cartographic data and its quality. The generalization process time is a crucial factor to provide a user cartographic data. The waiting time must be compatible with Newell's cognitive band, which is less than 10 seconds [7]. Also, in order to produce maps suited to a user's requests, on- the-fly map generalization must be flexible enough to take into account the level of detail, the kind of the map…etc. 3. APPROACHES OF ON-THE-FLY MAP GENERALIZATION There are three fundamental approaches to providing on the fly map generalization: 3.1 Generalization-oriented approaches This approach based on map generalization which is known to be a complex and time consuming process. In order to accelerate this process, certain authors propose methods based upon pre-computed attributes [1]. Cartographic generalization operators have to be applied to spatial objects on-the-fly. The generalization operators are principally the selection, the simplification and the displacement. The generalization oriented approach is very flexible [2]. However, it is not widely used because of the time it takes to provide requested maps. Furthermore, due to its complexity, generalization process cannot be carried out by simply applying generalization algorithms sequentially without taking into account the objects' spatial neighbourhood. 2 International Journal of Information Technology Convergence and Services (IJITCS) Vol.2, No.3, June 2012 3.2 Representation-oriented approaches Currently, it is the ideal solution to allow users to get data at the desired level of abstraction compared by the previous approach. This approach proposes to store several pre-defined representations of a given object (usually at different scales) within the same database [2]. The simplest representations are usually obtained from the manual or semi-automatic generalizations of the most detailed representations. However, in terms of personalization, multi-representations are extremely limited because all scales are predefined. Other important problems related to multiple representations are the difficulty to create necessary map scales [2]. All these problems limit the effective use of multi-representations for on-the-fly map generalization. 3.3 Hybrid approaches Hybrid approaches take advantage of the flexibility of generalization oriented approaches and the suitability of representation-oriented approaches to generate maps in real-time by combining their use [1]. Several authors proposed an approach based on this hybrid approach, such as [1], [2] and [6]. Its advantage is that it reduces the effort needed for generalization process and improves the quality of the result because smaller the difference between the initial map scale and the desired one, easier the generalization process. However, to be truly efficient, this method must rely on a database that includes several scales, leading to the typical problems associated with multiple representations [2]. To improve this third approach, it is necessary to develop new methods that minimize, as much as possible, the problems associated with automatic generalization and multiple representations and resolve all the spatial conflicts. 4. PROPODED APPRAOCH This work is based on the approach proposed in [4]. We combine genetic agent, map generalization process and multiple representations approach for improving the delivery time of map and resolving spatial conflicts to increase the quality of result map. This approach aims exclusively to improve the map generalization process. The spatial objects are modelled as agent. This agent is a concept of artificial intelligent. Each agent is equipped with genetic patrimony. Thus, genetic agent has some knowledge of its internal state, and some sensory information concerning environmental context, which permit it to decide what action (or action sequence) executed in order to achieve its goals. In this context, the role of genetic agent is identifying the best sequence of generalization's algorithms with good parameters that allow to perform the best map generalization process. In order to implement this process, each agent is able to identify and assess its internal constraints, and applies generalisation algorithms to itself in order to satisfy as well as possible these constraints. They are divided into two kinds [8]: -Internal constraints: The constraints relate an isolated object. In this context, we take account for building agent; constraints of size, granularity, squareness, preservation of shape, etc. -Relationalships constraints: They are the constraints which involve more than one spatial object, the constraint that prevents symbols from overlapping each others (, e.g the symbol of a road should not overlap with that of a house), or the constraint that requires aligned buildings to remain aligned. In this section, the genetic agent architecture is presented. We describe also the structure of plan executed by the genetic agent, its different modules and the cycle life of genetic agent. 3 International Journal of Information Technology Convergence and Services (IJITCS) Vol.2, No.3, June 2012 4.1.Genetic Agent architecture As previously mentioned, this research work based upon the use of hybrid approach which allows to improve the generalization process of spatial data described in [4]. This approach consists to model the geographic objects contained in databases (roads, buildings, etc.) into the decisional entities of the generalization process. In this approach, every geographic object becomes a software agent whose goal is to satisfy its constraints as much as possible. Each geographic agent has three main components: -Patrimony genetic. -An optimizer: using the genetic algorithm to find the best solution according to the internal constraints or relationships ones. -Ability to communicate with neighbouring agents. Environnement Perception of environment Message received from others agents Creation of population Optimizer Evaluation of selection Optimizat ion Control Module Selection Crossover, mutation N o Stop ? Yes Evaluation Best action Execution Message sent to others agents Figure 1. Genetic agent architecture 4 International Journal of Information Technology Convergence and Services (IJITCS) Vol.2, No.3, June 2012 4.2 Plan Structure of genetic agent The genetic agent is an agent which has a patrimony genetic. In this context, we consider two main kinds of genetic agent; road's genetic agent and building's genetic agent. In the following, we represent the gene of these two kinds: 4.2.1 Plan of Road gene An agent road can be characterized by its identified, and a set of algorithms that can be applied on it, to perform its generalization. These algorithms can take parameters. The coding of gene is made by multiple forms; we use character code for coding the identifiers, the binary form to implement the application or not of such algorithms (0 to say that the algorithm is not applied, 1 for say that the algorithm is applied) and the real forms to encode the parameters of the algorithms. Thus, we can represent agent's plan of roads by the figure 2: P1 DP -- 1 -- 10-6 Accor-1 Schemat-0 Figure 2. Example of road's plan 4.2.2 Plan of Building gene The agent Building can be characterized by its identifier and by its generalization's algorithms: simplification algorithm and displacement one... etc. The same type of coding is used to encode this type of gene. So, the following figure represents an example of building's plan: P2 Simp -- 1 ∆X(displacement) ∆Y(displacement) Figure 3. Example of building's Plan 4.3 Genetic agent Modules The main role of a genetic agent is to generalize its self, in order to adapt it to the level of detail requested by the user. For this reason, the genetic agent is responsible for the satisfaction of its constraints. It must be collaborate with the other agent for avoid a constraints violation. It applies the proper solution composed of a sequence of generalization's algorithms, which is generated by its optimizer. The architecture of genetic agent is composed of two main modules. The one is called generalization module which perform the map generalization process and the other is the optimization control module which controls the data flow and the condition of optimization, such as the number of agent in conflicts and the elapsed time of generalization process. 4.3.1 Map Generalization module This module carries out the map's generalization process; it applies the solution defined by the optimizer. The optimizer execute genetic algorithm to define the best chromosome. It is represented by sequence of algorithms and their good parameters, depending on the environment condition and the genetic agent state. For finding the solution, the optimizer follows the classical steps of a genetic algorithm are selection, crossover and mutation [4]. The figure 4 represents a solution generated by the optimizer, such as, each two chromosomes' parents together to make two chromosomes' child, in result. The solution is refined gradually over the iterations until convergence to a solution that approaches the optimal solution, a certain degree of imperfection is acceptable. 5 International Journal of Information Technology Convergence and Services (IJITCS) Vol.2, No.3, June 2012 Figure 4. Solution generated by the optimizer The solution evaluated by a fitness, we use in this work a fitness based on certain measures that assess the quality of spatial data, is restricted to a set of measures: • OS (object shape): calculates the loss of the object's shape during processing. For buildings, we compute the differentiation in the surface S and for roads, using measures of McMaster [9]: OS = ∑ (McMaster ∆+∆ S ) (1) A minimum number is a good solution. • NC: number of objects in conflict, a minimum number is a good solution. • DP: sums the normalised, absolute, distance of each object has been displaced from its starting position. DB = n ∑ 1 ( 2 idx + dy ) 2 i (2) Also, a minimum number is a good solution. So, the general function is: f = NC + DP + OS (3) The best solution is one that has the smallest value of general function. After each iteration (selection, crossover, mutation), agents in conflicts exchange the messages for calculate the value of general fitness which allow to select the bests chromosomes, i.e. chromosomes with minimum fitness. After this step, it retains only the most relevant solutions. 6 International Journal of Information Technology Convergence and Services (IJITCS) Vol.2, No.3, June 2012 4.3.2 Optimization Control module This module controls principally the time constraint and the quality of the map. It controls generalization's time for not exceeds the maximum limits and the number of spatial conflicts. The optimizer stops the execution of the genetic algorithm for three reasons, when it achieve certain fitness, when it achieve a set number of iterations or when it have passed a certain running time. In the two latter cases, we choose the solution that has the best fitness. The optimization control module assures a compromise between the time of delivering spatial data and its quality. Spatial BD Generalization module Control optimization module Internet Genetic Agent I N T E R F A C E Figure 5. System architecture It exists also an interface between the user and the map generalization module; it allows user to transmit its requests. The request transmitted carries important information for the researched data, such as identification of the zone, the kind of map and its level of detail. The level of detail is arbitrary and not predefined. This module permits also to drawing the final generalized data. In this approach, genetic agents negotiate with each other agent, via message that are used as input in genetic algorithm, these data allow agent to solve various conflicts at once and prevent new conflicts from appearing and achieve a best generalization process. 4.4 Cycle life of Genetic Agent In the beginning of map generalization process, the geographic agent has evaluated its internal constraints and relationalship ones, in its agents group. In our multi-agent system, the agents aim to satisfy the offensive constraints without violating one of defensive constraints [10]. Fitness is utilised for measuring the satisfaction of the constraints (calculate the degree of constraints satisfaction). A list of possible plans is attached which represent the populations of the genetic algorithm. A plan proposes a sequence of generalization's algorithms and its required parameters which improves the agent's state. The agents active and run theirs genetic algorithms in parallel for reducing the time of map generalization process. The process of improving an agent's state represented in the following sequence: constraints evaluation, proposing plans which represent the population of genetic, the execution of the genetic algorithm by the optimizer. The generalization process is controlled by the optimization control module for selecting the best plan, the previously steps are the same generic behaviour of all genetic agents, the figure 5 represent the life cycle of agent. 7 International Journal of Information Technology Convergence and Services (IJITCS) Vol.2, No.3, June 2012 Control optimization module Best plan Begin Evaluate constraints Optimizer Update Data base End Propose plans 5.IMPLEMENTATION AND RESULTS Figure 6. Genetic agent cycle life We use in this work the Jade platform [11], JADE is a software framework, fully implemented in Java that simplifies the implementation of multi-agent systems through a middleware. JADE implements FIPA's (Foundation of Intelligent Physical Agents) specifications [2]. The algorithms of generalization and the genetic algorithm are implemented in Java. The agents execute their genetic algorithms in parallel to decrease the response time of the map generalization process. We implement also the various steps of the genetic algorithm; selection, crossover, mutation, and the fitness function for evaluate the solution that was composed of a sequence of generalization's algorithms with good parametric values to improve the map generalization process. Initially, for avoid the complexity; we have used a limited number of objects. We used data in a Shapefile format (ESRI) from the Habitats' direction of el-oued (Algeria). These data are on a scale of 1:1000 and cover a little part of Choot city. Two classes of objects, namely "buildings" and "roads", were used. The scale requested by the user is 1:1500. For drawing the generalized object we use a visualization tool which is based upon the GeoTools library [12]. We present in figure 7 the initial results of our approach. 8 International Journal of Information Technology Convergence and Services (IJITCS) Vol.2, No.3, June 2012 6.CONCLUSION Figure 7. Initial results In this approach, each type of map features (buildings, roads) is modelled as agent; these agents are equipped with patrimony genetic which allows improving the map generalization process. Multi-agent systems and genetic algorithms provide various advantages to real-time map generalization and quality of generalized data such as, the autonomy and flexibility which improve the personalization of maps at the time of their creations, finding the best solution and resolve all the appearance spatial conflicts. Moreover, the genetic agent has equipped with sensor to check its environment, a communicator to send requests or responses, and the ability to move itself into a new position. Also, to decrease the time of on-the-fly map generalization process, the agents execute theirs genetic algorithms, in parallel and at the same time of the transfer of generalized data. The initial experiments have shown us the potential advantages of using collaborative agent. Each agent is equipped with patrimony genetic that enable it to generate maps in real-time and with height quality. Thus, each agent can: - Define the optimal or near optimal actions (sequence of algorithms) of generalization and it can generate its self. - Adapt its generalization with the other geographic agents. - Collaborate with the others agent for accelerating the time of map generalization process and resolves all the spatial conflicts. This research work opens various directions of researches, such as: - Improving the performance of our approach with the utilization of other methods of optimization that allow the agent to define the optimal solution. - Customizing the map according to user's needs. In another work, we will model the spatial constraints in the form of ontology for different kinds of map. For example, one can have ontology for constraints semantic of tourist maps. Thus, when using the system, the most appropriate ontology can be selected according to the kind of map that will be generated. 9 International Journal of Information Technology Convergence and Services (IJITCS) Vol.2, No.3, June 2012 REFERENCES [1] Nafaa jabeur (2006), A multi-agent system for on-the-fly web map generation and spatial conflict resolution, University of Laval, Quebec. [2] Mamane Nouri SABO (2007), Généralisation et des patrons géométriques pour la création des objets auto-generalisants (sgo) afin d'améliorer la généralisation cartographique à la volée, Université de laval , Quebec. [3] Wilson, I.D. & Ware, M. (1995), Reducing Graphic Conflict In Scale Reduced Maps Using A Genetic Algorithm, ICA Map Generalization Workshop, Paris, 2003. [4] Lejdel brahim & kazar okba (2012), Genetic agent for optimizing generalization process of spatial data, International Journal of Digital Information and Wireless Communications (IJDIWC), Vol. 1, No. 3, pp-729-737. [5] Van Oosterom & V. Schenkelaars. (1995), The development of an interactive multi-scale GIS, International Journal of Geographical Information Systems, Vol. 9(5):489-507. [6] Cecconi, A. (2003), Integration of cartographic generalization and multi-scale databases for enhanced web mapping, PhD. Thesis, University of Zurich. [7] A. Newell. (1990), Unified Theories of Cognition, Harvard University Press, Cambridge MA, p549. [8] Lejdel brahim & Mrs K. amieur & Zaia alimazighi, (2009), modelling and managing multiples representations of spatial data by a hybrid approach (application them: road network), proceeding of International conference on information & communication systems, pp219-226. [9] McMaster, R.B. (1983), Mathematical measures for the evaluation of simplified lines on maps, PhD thesis, University of Kansas., USA. [10] Bader M. (2001), Energy Minimizing Methods for Feature Displacement in Map Generalization, PhD thesis, Department of Geography, University of Zurich. [11] JADE (2006), Java Agent Development Framework, http://jade.tilab.com/. [12] GeoTools (2006), the open source java GIS toolkit, http://docs.geotools.org/. 10
1807.03646	1	1807	2018-07-10T13:48:29	Ontology-based multi-agent system to support business users and management	[ "cs.MA", "cs.CY" ]	For some decision processes a significant added value is achieved when enterprises' internal Data Warehouse (DW) can be integrated and combined with external data gained from web sites of competitors and other relevant Web sources. In this paper we discuss the agent-based integration approach using ontologies (DSS-MAS). In this approach data from internal DW and external sources are scanned by coordinated group of agents, while semantically integrated and relevant data is reported to business users according to business rules. After data from internal DW, Web sources and business rules are acquired, agents using these data and rules can infer new knowledge and therefore facilitate decision making process. Knowledge represented in enterprises' ontologies is acquired from business users without extensive technical knowledge using user friendly user interface based on constraints and predefined templates. The approach presented in the paper was verified using the case study from the domain of mobile communications with the emphasis on supply and demand of mobile phones.	cs.MA	cs	Ontology-based multi-agent system to support business users and management Dejan Lavbič, Olegas Vasilecas and Rok Rupnik Dejan Lavbič, Olegas Vasilecas and Rok Rupnik. 2010. Ontology-based multi-agent system to support business users and management, Technological and Economic Development of Economy (TEDE), 16(2), pp. 327 - 347. Abstract For some decision processes a significant added value is achieved when enterprises' internal Data Warehouse (DW) can be integrated and combined with external data gained from web sites of competitors and other relevant Web sources. In this paper we discuss the agent-based integration approach using ontologies (DSS- MAS). In this approach data from internal DW and external sources are scanned by coordinated group of agents, while semantically integrated and relevant data is reported to business users according to business rules. After data from internal DW, Web sources and business rules are acquired, agents using these data and rules can infer new knowledge and therefore facilitate decision making process. Knowledge represented in enterprises' ontologies is acquired from business users without extensive technical knowledge using user friendly user interface based on constraints and predefined templates. The approach presented in the paper was verified using the case study from the domain of mobile communications with the emphasis on supply and demand of mobile phones. Keywords Decision support, agent, multi-agent system, ontology, data warehouse, information retrieval, business rules, business process management 1 Introduction There is a growing recognition in the business community about the importance of knowledge as a critical resource for enterprises. The purpose of knowledge management is to help enterprises create, derive, share and use knowledge more effectively to achieve better decisions, to increase of competitiveness and to decrease the number of errors. In order to run business effectively an enterprise needs more and more information about competitors, partners, customers, and also employees as well as information about market conditions, future trends, government policies and much more. There are several products and technologies available on the market that support advanced Business Process Management (BPM) (Trkman et al., 2007) and advanced decision support. Enterprises expect these applications to support wide range of functionalities - analysing customer profiles, building and analysing business strategies, developing customer-specific products, carrying out targeted marketing and predicting sales trends. Amount of documents in the Web, enterprise data repositories, and public document management systems with documents are rapidly growing. This huge amount of data is managed in some extent, but knowledge workers, managers, and executives still have to spend much of their working time reading dozens of various types of electronic documents spread over several sources in process of making decisions. There is just too much information to digest in a daily life. The tremendous amount of documents that is still growing has far exceeded the human ability for comprehension without intelligent tools. Different applications within information systems (IS) that support wide range of functionalities need to be integrated in order to provide the appropriate level of information support. One of 1 the prominent approaches for IS integration is the use of ontologies and Multi-Agent Systems (Fuentes et al., 2006; Soo et al., 2006; Dzemydiene and Tankeleviciene, 2009). The approach presented in this paper is targeted towards using ontologies for several tasks, where emphasis is on using business rules (BR) approach for interoperability between business user and IS. By introduction of BR approach business users do not have to be fully familiar with the technology to manipulate the common understanding of a problem domain in a form of ontology and therefore enabling agents to execute defined analyses models. The use of ontologies in Multi-Agent System (MAS) environment enables agents to share a common set of concepts about contexts, user profiles, products and other domain elements while interacting with each other. Agents can exploit the existing reasoning mechanisms to infer derived contexts from known contexts, to make decisions and to adapt to the environment, current status, and personal setting of the user. The purpose of this paper is to present the approach of integration of several information resources for Decision Support in Enterprises using agent-oriented approach based on ontologies. The goal of our research is to minimize the gap between business users and agents as special type of application systems that perform tasks in their behalf. The intention was to apply BR approach for ontology manipulation in MAS. Ontology used in our Multi-Agent System for Decision Support in Enterprises (DSS-MAS) was divided into task and domain ontologies while business users were enabled to manipulate them directly in a user friendly environment without requirement of detailed technical knowledge. The remainder of this paper is structured as follows. First we present some background in the following section 2 with emphasis on agents, ontologies and related work with clear definition of the problem and solution proposal. Next, in section 3, we introduce our case study of integrated Multi-Agent environment from the domain of mobile communications with emphasis on architecture and the roles of agents and ontologies. The case study is focused in one of the mobile operators and furthermore oriented to supply and demand of mobile phones. After presentation of system architecture and decomposition of ontology of every agent from DSS-MAS will be presented in detail. Details of case study implementation will be given in section 4.3. Finally the last section 5 presents conclusions. 2 Some background on decision support, multi-agent systems and ontologies Decision support systems (DSS) have evolved significantly and there have been many influences from technological and organizational developments (Shim et al., 2002). DSS once utilized more limited database, modelling, and user interface functionality, but technological innovations enabled more powerful DSS function- ality. DSS once supported individual decision makers, but later DSS technologies are applied to workgroups or teams, especially virtual teams. The advent of the Web has enabled inter-organizational DSS and has given rise to numerous new applications of existing technology as well as many new decision support technologies themselves. Internet facilitates access to data, information and knowledge sources, but at the same time, it threatens to cognitively overload the decision makers. Authors in (Vahidov and Kersten, 2004) claim that internet technologies require a new type of decision support that provides tighter integration and higher degree of direct interaction with the problem domain. Based on that they propose a generic architecture for dynamic and highly complex electronic environments where DSS's should be situated in the problem domain. Chen et al. (2001) conducted an interesting research about integrated interactive environment for knowledge discovery from heterogeneous data resources. Their work is grounded on acquiring, collecting, and extracting relevant information from multiple data sources, and then forming meaningful knowledge patterns. The proposed system employs common DW and OLAP1 techniques to form integrated data repository and generate database queries over large data collections from various distinct data resources. Multi-Agent Systems (MAS) offer a new dimension for cooperation and coordination in an enterprise. The MAS paradigm provides a suitable architecture for a design and implementation of integrated IS, especially DSS. With agent-based technology a support for complex IS development is introduced by natural decomposition, abstraction and flexibility of management for organisational structure changes (Kishore et al., 1OnLine Analytical Processing (OLAP) 2 2006). The MAS consists of a collection of autonomous agents that have their own goals and actions and can interact and collaborate through communication means. In a MAS environment, agents work collectively to solve specific enterprises' problems. MAS provide an effective platform for coordination and cooperation among multiple functional units in an enterprise. The research on agents and MAS has been on the rise over the last two decades. The stream of research on IS and enterprise integration (Lei et al., 2002; Kang and Han, 2003; Tewari et al., 2003) makes the MAS paradigm appropriate platform for integrative decision support within IS. Similarities between the agent in the MAS paradigm and the human actor in business organisations in terms of their characteristics and coordination lead us to a conceptualisation where agents in MAS are used to represent actors in human organizations. Several approaches (Tewari et al., 2003; Rivest et al., 2005; Kishore et al., 2006; Soo et al., 2006) deal with agent support for integration and decision support. Research in (Kishore et al., 2006) has shown that MAS paradigm provides an excellent approach for modelling and implementing integrated business IS. Authors within that research proposed a conceptual framework for MAS based integrative business IS. Some promising results were also found in (Soo et al., 2006), where authors propose a cooperative MAS platform to support the invention process based on the patent document analysis. The platform allows the invention process to be carried out through the cooperation and coordination among software agents delegated by the various domain experts in the complex industrial R&D environment. Today, semantic technologies based on ontologies and inference are considered as a promising means towards the development of the Semantic Web (Davies et al., 2006). In the field of Computer Science and Information Technology (IT) in general ontology has become popular as a paradigm for knowledge representation in Artificial Intelligence (AI), by providing a methodology for easier development of interoperable and reusable knowledge bases (KB). The most popular definition, from an AI perspective, is given in (Gruber, 1993) as follows: "An ontology is an explicit specification of a conceptualization", where a conceptualization is abstracted view of the world that we wish to represent for some purpose. Ontologies can be considered as conceptual schemata, intended to represent knowledge in the most formal and reusable way possible. Formal ontologies are represented in logical formalisms, such as OWL, which allow automatic inferencing over them. An important role of ontologies is to serve as schemata or intelligent views over information resources. Thus they can be used for indexing, querying, and reference purposes over non-ontological datasets and systems, such as databases, document and catalogue management systems. Because ontological languages have a formal semantics, ontologies allow a wider interpretation of data that is inference of facts which are not explicitly stated. In this way, they can improve the interoperability of the conceptualization behind them, their coverage of arbitrary datasets. Ontology can formally be defined as specific sort of knowledge base and can be characterized as comprising a 4-tuple (Davies et al., 2006) O = hC, R, I, Ai (1) Where C is set of classes representing concepts we wish to reason about in the given domain (Offer, Finding, Phone, Customer, etc.). R is set of relations holding between those classes (Message hasRecipient Actor). I is a set of instances, where each instance can be an instance of one or more classes and can be linked to other instances by relations (Nokia is A PhoneBrand; Finding309 hasValue 11,23). A is a set of axioms (If a new customer buys is A Nokia E72, promotional discount of 10% should be offered). It is widely recommended that knowledge bases, containing concrete data (instance data or ABox) are always encoded with respect to ontologies, which encapsulate a general conceptual model of some domain knowledge, thus allowing easier sharing and reuse of KBs. Typically, ontologies designed to serve as schema for KBs do not contain instance definitions, but there is no formal restriction in this direction. Drawing the borderline between the ontology (i.e. the conceptual and consensual part of the knowledge) and the rest of the data, represented in the same formal language, is not always trivial task. In our approach we include instances as part of ontologies because instances we define are a matter of conceptualization and consensus and are not only descriptions, crafted for some purpose. Related work on using ontologies in information systems for decision support is extensive. Regarding the domain of DW and OLAP analyses research has dealt with Document Warehousing (Tseng and Chou, 2006) where extensive semantic information about the documents is available but still not fully employed as in traditional DW. The use of ontologies was found useful as a common interpretation basis for data and metadata. Furthermore research has extended to Web DW (Marotta et al., 2002) with the emphasis on 3 managing the volatile and dynamic nature of Web sources. Utilization of ontologies is also addressed in Information Retrieval (IR) where it has been used for fuzzy tagging of data from the Web (Buche et al., 2006; Macias and Castells, 2007), query construction tool in semi-automatic ontology mapping (Suomela and Kekalainen, 2006) and semantic based retrieval of information from the World Wide Web (Shan et al., 2003; Garces et al., 2006). Use of ontologies in Data Mining (DM) has also been considered in (Bernaras et al., 1996; Zhou et al., 2002; Singh et al., 2003; Cao et al., 2004) where ontology was used for representation of context awareness and handling semantics inconsistencies. Ontologies have been widely used for data, application and information integration in the context of domain knowledge representation (Qiu, 2006). Jovanović and Gašević (2005) concludes that the need for knowledge sharing and interoperable KBs exists and the key element for achieving interoperability are domain ontologies. In that approach XSLT2 transformation is used to enable knowledge interoperability. Authors in (Vasilecas and Bugaite, 2006) use ontologies for ontology based IS development and they address the problem of automation of information processing rules. There are also other approaches (Fuentes et al., 2006; Orgun and Vu, 2006; Clark and McCabe, 2007) that use ontologies as knowledge representation mechanisms. Authors in (Clark and McCabe, 2007) use a formal ontology as a constraining framework for the belief store of a rational agent. The static beliefs of the agent are the axioms of the ontology, while dynamic beliefs are the descriptions of the individuals that are instances of the ontology classes. Another work presented in (Fuentes et al., 2006) also uses heterogeneous domain ontology for location based IS in a MAS framework with the emphasis on context-aware MAS. They propose a global ontology to let agents work with heterogeneous domains using a wireless network and the intention is to provide customization about different environment services based on user location and profile. 2.1 Problem and proposal for solution The review of related work presented in this section pointed out that modern DSS's changed quite substantially especially with the advent of the Web and availability of extensive information in online repositories. For managing complexity and integration issues with decision support many approaches relied on MAS paradigm and used ontologies as knowledge representation mechanism. The existing approaches mainly focused on either supporting existing business processes or improving decision support at some level of detail or integration of several structured resources to achieve better decision support. To our knowledge none of the approaches addressed the problem of enriching data from internal data sources with unstructured data found on internet. The interactivity of reviewed solutions is also limited; meaning that business users are usually limited to small set of parameters they can define to alter default behaviour of the system. These user requirements are usually entered directly into the system and no abstraction layers are provided as in business rules management systems (BRMS) to enable users without technical skills to manipulate the content. This paper introduces a novel approach in integration of unstructured information found in the Web with information available in several internal data sources (e.g. database, DW, ERP3, etc.). The MAS paradigm with agents was used for implementation purposes, mainly because related work pointed out that it is a very appropriate solution for integration of business IS. One of the reasons to choose agents is also modelling notion where business users and agents are modelled in a very similar manner. Problem of interaction between human actors and computer programs is also addressed by introduction of ontologies as knowledge representation mechanism. The approach presented in this paper is targeted towards using ontologies for several tasks, where emphasis is on using BR approach to ensure interoperability between business user and IS. Ontology is used not only for every agent to represent the interpretation of a problem domain but also for communication between agents and business users. The use of ontologies in MAS enables agents to share a common set of facts used in user profiles, product descriptions and other domain elements, while interacting with each other. With exploiting reasoning mechanisms new findings can be derived from initially known facts and improve the KB by extending it with new knowledge. To simplify this communication template system based on BR was introduced to enable manipulation of knowledge within the system by users with less technical skills and to control behaviour of individual agent. The approach will be further explained in the following section. The case study presented in this paper is from the domain of mobile telecommunications. 2XSLT (XSL Transformations) is a declarative, XML-based language used for the transformation of XML documents into other XML documents 3Enterprise Resource Planning (ERP) 4 It is presented in detail (with the impact it has on improving decision support within enterprise) in section 4.3. In the domain of mobile communications that was used for case study we had to define several tasks in DSS-MAS, needed for decision support - OLAP analyses, DM, IR, context and profile definition, notification, etc. 3 DSS-MAS 3.1 DSS-MAS architecture DSS-MAS that we propose in this paper is introduced in Figure 1. The case study presented in this paper is from the domain of mobile telecommunications and is based on business environment and information resources from one of the mobile operators. DSS-MAS is situated in the environment of several existing systems, like Data Mining Decision Support System (DMDSS), or DW and various resources available outside of an enterprise on the World Wide Web. Global goal that agents in DSS-MAS should achieve is to support decision making process while using existing systems for business analysis and employing information from environment where enterprise resides. To support this goal DSS-MAS includes several agent roles that are as following: Data Mining Agent (DMA), OLAP Agent (OLAPA), Information Retrieval Agent (IRA), Knowledge Discovery Agent (KDA), Notifying Agent (NA) and Mobile Agent (MA). The agents in DSS-MAS have both reactive and proactive characteristics. Reactive are mainly due to responding to the environment according to the model defined in the KB they use. Proactive are due to their ability to learn from the environment and change the initially defined KB to for example improve performance. Ontologies are used as a main interconnection object for domain knowledge representation, agent-to-agent communication and most important for agent-to-business user communication. An important element of an environment is the World Wide Web, where agents acquire information for the purpose of decision making. Retrieved information is saved in a KB and available for further employment for DM and DW analyses. All information gathered from internal and external sources is considered by KDA, where inference over several task ontologies used by individual agents (DMA, OLAPA, IRA, etc.) is performed. Moreover the sub goal of DSS-MAS is delivering of the right information at the right time and to the right users. The system needs to be context aware and to consider the relevant features of the business, i.e. context information such as time, location, and user preferences (Liao et al., 2005). Business users in DSS-MAS are able to employ agents to perform tasks on their behalf. For example, managers in enterprises have to request reports from their systems - OLAP or from transactional databases, and managers have to review reports every appointed period of time (day, week, month, etc.). This task of information acquisition is predecessor for decision making and is more or less straightforward - business user sends a request for analyses and reviews the content according to some Key Performance Indicators (KPI). KPI is simply a measure of performance and is commonly used in enterprises to evaluate how successful they are. In DSS-MAS tasks like this are automated and user participation is reduced as much as possible. An initial analysis model (e.g. OLAP or DM) has to be captured in the ontology by business users, while execution and optimisation is left for agents. Business users first define initial parameters for analyses to be performed, while agents perform these analyses and recommend improvements. When some action is required from business user, he is notified and has the ability to act or change rules of agent's execution. To enable these functionalities we introduce ontologies as a mediation mechanism for knowledge exchange between actors (agents and business users) that cooperate in DSS-MAS. The following section will present the structure and organization of ontologies we have used for the case study. 3.2 The role of ontology According to Guarino (1998), ontology can be structured into different sub-ontologies - upper ontology, domain ontology, task ontology and the application ontology. Following similar guidelines we have defined upper ontology named Common ontology and combined domain and task ontologies in Notifying ontology, Information retrieval ontology, Data Mining and Warehousing ontology (see Figure 2). The 5 Figure 1: Architecture of MAS used for Decision Support in Enterprises 6 proposed clustering of ontologies is based on the common understanding of the problem domain being defined in Common ontology. Every agent has its own interpretation of a KB, which is a specialization of a Common ontology with detail definition of knowledge required by individual agent. Common ontology is limited to abstract concepts and it covers reusable dimensions, which are primarily used by KDA. Task ontologies specify concepts of notification, IR, DM and DW. Mobile communications in our case is the domain of all task ontologies and the emphasis is on supply and demand of mobile phones. As already mentioned, we have used the knowledge management approach in our research where every agent has knowledge about its own problem domain. In this case whenever new facts about the common knowledge are discovered, which might be of interest for other agents, they are updated to the common ontology. The role of ontology in our approach is therefore twofold: • knowledge representation mechanism used by agents and • common understanding of problem domain used for communication between business users and agents by utilizing business rules manipulation with introduced templates (see section 4.2). Figure 2 shows an excerpt from intersection of several ontologies used in our case study. This part of ontology clearly defines the common elements being used for communication between agents and business users (domain specific elements such as phones, new phones and customers, all described with domain specific characteristics). A part of OLAP elements needed for conducting OLAP analyses is also presented. Ontology also presents notification with taxonomy of various warning levels and business users classification by organizational unit and decision making level. 3.3 The role of agents Our case study uses domain of mobile telecommunications as a platform where we focus on the sales of mobile phones and their accessories. Manipulation with internal data storage is handled by two types of agents - OLAPA and DMA. They both have distinct tasks but still share common goal – periodically or on demand autonomously execute analyses models. Business users at first define these models and describe them with all required parameters (e.g. search for anomalies in sales of Nokia phones in last month period). The information about the execution is stored in the ontology (based on business user preferences) or is requested by another agent of the system. Business user preferences in this context define the execution parameters about the analysis, for example the period at which the analysis is performed (e.g. perform analysis every other day at 13:00). OLAPA has on firsthand straightforward task of performing OLAP analyses on behalf of an agent or a business user and reporting its findings back to the requesting entity and all other entities that should be informed, according to the business policy. Nevertheless OLAPA does much more - after each execution it prepares the report for business user based on findings – movements and KPIs. If certain finding is substantially different from finding obtained in previous case further analysis is performed to discover the reason of change by drilling down (more detailed) or up (less detailed) the hierarchies and levels. The knowledge is acquired in ontology. Business users can change the behaviour of agents by changing the ontology using graphical user interface. This interface incorporates all logical restrictions defined in ontology and does not allow users to enter unacceptable values and the most important is that it does not require technical knowledge from users. Previous our experiences have shown that business users have great difficulties especially with setting the parameters required to run DM and DW analyses models. So, user interface really has to be friendly and intuitive. In our approach this issue was solved by introducing the architecture depicted in Figure 3 and using templates as further discussed in section 4.3. Nowadays Web-based information retrieval systems are widely distributed and deeply analysed from different points of view. The main objective of all such systems is to help users to retrieve information they really need (obviously as quickly as it is possible) (Garces et al., 2006). While the techniques regarding DW, multi-dimensional models, OLAP, or even ad-hoc reports have served enterprises well, they do not completely address the full scope of existing problems. It is believed that, for the business intelligence (BI) of an enterprise, only about 20% of information can be extracted from formatted data stored in relational databases (Tseng and Chou, 2006). The remaining 80% of information is hidden in unstructured or semi-structured documents. This is because the most prevalent medium for expressing information and knowledge is text. 7 Figure 2: Excerpt from intersection of several ontologies used in our case study 8 Figure 3: Prototype of selected case study For instance, market survey reports, project status reports, meeting records, customer complaints, e-mails, patent application sheets, and advertisements of competitors now are all recorded in documents. For that reason in DSS-MAS we introduced IRA for retrieval of data mainly from the World Wide Web. The tasks that IRA performs in presented case study can be grouped into three categories: • identification of new online shops, • analysis of mobile phones presented online and • extending Data Warehouse with information found online. First two tasks are concerned about the supply of mobile phones at various online shops worldwide. Identifi- cation of new online shops is conducted with web crawling and the use of several existing services on the Internet, such as Google, Google Product Search and Bing. Not only these internet resources are managed through ontology, but also rules for text extraction are defined as rules which make all domain knowledge available in IR ontology and not encoded in agent itself. More details about implementation of DSS-MAS case study can be found in section 4.3. Furthermore every shop found online is analysed to identify unique patterns for searching phones. Search patterns include guidelines for agents performing search at various web pages. They are based on XQuery4 and regular expressions. To search for phones at Google Products Search, the following URL search pattern http://www.google.com/products?q=((\w+\s*)+) is used, accompanied with additional information for web scrapping of required information (e.g. price, availability etc.). Using these search patterns IRA is searching through online shops and determines phones with their market prices and stores this information into IR ontology to be available for further knowledge inference by KDA. Information of found phones is used to determine new market trends, enable price comparison between competitors, facilitate possible inclusion in enterprise's sales program, etc. One of the tasks that IRA also performs is extending DW with information found online. While business users perform OLAP analyses, they deal with only internal information about the business, but in process of decision making other resources also have to be examined, e.g. news about the suppliers and competitors, opinions about certain products and organisations, etc. IRA therefore scans the DW dimension data (through hierarchies and levels) from DW dimensional schema and uses this information for searching several internet resources (news archives, forums, stock changes, Google trends, etc.). When users review OLAP reports these data from the Internet is also displayed according to their restrictions in dimensions. For example, when business users are making decision whether to increase support to Nokia or Sony Ericsson phones it only has reports about sales of selected brand names from their market program. Using our approach the user is provided with additional data that is found online and what will make decision better founded. By this integration of internal and external information users have integrated data source available that they can query from single location. 4XQuery is a query language that is designed to query collections of structured and semi-structured data 9 KDA is an important element of DSS-MAS since it consolidates all findings from IR, DM and DW and furthermore delivers derived findings to NA. To employ inference capabilities over several ontologies the enterprises' BR are essential. While business concepts are captured in ontology, these concepts further have to be restricted to define specific meaning. Generally BR are prepared by business users and also some parts of BR in enterprises tend to change frequently; therefore we introduced architecture for BR management (see Figure 3 and further discussion in section 4.3). Findings of KDA are presented as instances of Domain-specific-element and Findings classes (see ontology in Figure 2). As it can be seen from Figure 1 NA represents an interface to DSS-MAS for all external applications and business users. The main role of NA is the information dissemination by simply delivering the right information at the right time to the right users. While in vast majority of today's applications users have to request the information using so called "pull model" in our approach we implemented the "push model", where information is proactively delivered by agents to the user without a specific request. This is achieved by making system context aware and considering the relevant features of the business, i.e. context information such as time, location, position in the organisational hierarchy, etc. All knowledge about notification is defined in Notifying ontology, where every user has his own context defined and the position within organisation across two dimensions - organisational unit (e.g. Marketing, Sales, Human resources, etc.) and decision making level (e.g. Chief Executive Officer (CEO), Chief Information Officer (CIO), Chief Financial Officer (CFO), Chief Marketing Officer (CMO), Chief Analytics Officer (CAO), etc.). According to that position rules for delivery of several message types are defined. These message types range from Notification to Warning and Critical alert. Each message also addresses the domain of specific organisational unit, e.g. when a new mobile phone is found online at competitor's website, CMO and CAO have to be notified. Organisational structure, as part of Notifying ontology, also defines that both CMO and CAO are inferior to CEO therefore he is also notified, but only in a case of a Critical alert. According to the business user profile, notification can be sent using several technologies from Windows Alert, e-mail, Really Simple Syndication (RSS), Short Message Service (SMS), etc. These notification types are also ordered by priority for each business user and according to this type the content is also adapted. Mobile agent is an example of an application that can reside on a mobile device (e.g. Personal Digital Assistant (PDA), mobile phone, etc.) and uses resources of DSS-MAS through NA. The typical use case includes sending mobile agent across network to DSS-MAS, where all needed information according to owner context is collected and then the mobile agent is returned back to originating location on a mobile device and presents the collected data to business user. When the process of acquiring data is in progress, business user does not have to be connected to the network, he can just wait offline until mobile agent is ready to return with the findings. In the following section details about the case study implementation will be presented with technologies used, templates for business rules acquisition and presentation of one specific scenario from case study. 4 Case study implementation and discussion 4.1 Technology The selected language for ontology presentation is OWL DL (Russomanno and Kothari, 2004), since it offers the highest level of semantic expressiveness for selected case study and is one of the most widely used and standardised ontology language nowadays that has extensive support in different ontology manipulation tools. Besides OWL logical restrictions, Semantic Web Rule Language (SWRL) rules were also used due to its human readable syntax and support for business rules oriented approach to knowledge management (Horrocks et al., 2005). SWRL rules are stored as OWL individuals and are described by OWL classes contained in the SWRL ontology. The use of SWRL enables storing schema, individuals and rules in a single component, which makes management much easier. SWRL rule form in a combination with templates that is introduced in the following subsection b is very suitable for knowledge formalization by business users that do not have extensive technical knowledge. 10 The ontology manipulation interface for business users is based on Protégé Ontology Editor and Knowledge Acquisition System (Stanford Medical Informatics, 2006a) and SWRL Tab (Stanford Medical Informatics, 2006b) for Protégé. It enables entering OWL individuals and SWRL rules where a step further is made towards using templates for entering information (see Figure 3). At the execution level KAON2 inference engine is used to enable inference capabilities. Due to limitation of SHIQ(D) subset of OWL-DL and DL-safe subset of SWRL language, before inference is conducted, semantic validation takes place to ensure that all preconditions are met. We selected FIPA5 compliant MAS platform JADE6 in DSS-MAS because it offers broad range of functionalities and is most widely used platform. This is due to very good support and availability of agent framework, where a lot of common agents' tasks are already implemented (i.e. agent communication at the syntax level, agent management, migration of agents, etc.). For Mobile Agent implementation an add-on JADE-LEAP7 was used to support the mobility of agents. 4.2 Mediation with BR templates Using templates with ontology, business logic is excluded from the actual software code whereas the majority of data for templates is acquired from ontology axioms and natural language descriptions in ontology, while other templates are prepared by users with technical knowledge. The main goal of using mediation with BR templates is to enable acquiring knowledge from actual knowledge holders i.e. business users and enable transformation of this high-level knowledge into information system level, where this data together with concepts from business vocabulary can be directly used for inference purposes and bring added value without any further programming by technically educated users. When acquiring new knowledge into the system from business users, the process always starts with focusing on concepts of business vocabulary that are persisted in a form of ontology. Users can freely traverse through this information space, select concepts and further manipulate all related information within the selected context. Altering and adding new information is all time limited to formal definition of concepts that is defined in ontology. For easier manipulation business user is aided with template and business vocabulary, so BR building process is simplified as it will be presented in detail in the following section. Example 4.1 presents a BR template that is used for definition of aggregation of findings or domain specific elements. The user interface that is available is directly linked to ontology, where constraints on classes, properties and individuals are considered in realtime. This approach allows to minimize the risk of entering wrong constraints. The DSS-MAS system supports entering of new statements in several forms from simple IF-THEN form to decision table or decision tree. Example 4.1 (BR template for general finding definition). IF Condition = {∃x : x ∈ Domain specific element ∪ Finding} ∧ Condition ≥ 1 THEN Result = {∃y : y ∈ Finding} ∧ Result ≥ 1 The following Example 4.2 represents a BR that states: If there exist two consequent increases of sold phones of the same phone brand and a new phone of this phone brand was found online within last 2 weeks, then offer a promotion discount of 10% on this new phone to all new customers. 5Foundation for Intelligent Physical Agents (FIPA) 6Java Agent DEvelopment Framework (JADE) 7Java Agent Development Environment-Lightweight Extensible Agent Platform (JADE-LEAP) 11 Example 4.2 (Example of a rule, developed by using template). IF First finding is Increase (Finding) which { is related to first amount sold which is Measure (OLAP element) AND is related to first date which is Dimension (OLAP element) AND is related to first phone which is Phone (Domain specific element) which { has characteristic brand which is Phone brand (Domain specific characteristic) } AND Second finding is Increase (Finding) which { } } } is related to second amount sold which is Measure (OLAP element) AND is related to second date which is Dimension (OLAP element) which { is greater than first date } AND is related to second phone which is Phone (Domain specific element) which { has characteristic brand which is Phone brand (Domain specific characteristic) } AND Found phone is New phone (Domain specific element) which { has characteristic brand which is Phone brand (Domain specific element) AND has date of appearance found date which is Dimension (OLAP element) which { is greater than now - 14 days } AND New customer is New customer (Domain specific element) THEN } Promotion discount is Discount price (Finding) which { is related to new customer AND is related to found phone AND has value "10" AND has unit "%" When constraint presented in Example 4.2 is transformed to execution form at information system level, standardized SWRL and OWL languages are used to enable reusability (see Figure 4). By this transformation a rule is produced that can be directly used in the inference engine to produce results in a form of inferred triples that are presented to the user. 4.3 Case study One of the most common cases of using DSS-MAS system is combing information found online with BI reports (DW, DM, IR, etc.) developed on internal data in enterprise DW. Figure 5 presents one of this scenario. Scenario presented at Figure 5 is triggered by results of IRA activity, when three new mobile phones: Apple iPhone 3GS, Nokia E72 and Sony Ericsson Xperia 1 are found by IRA at online mobile shops. According 12 Figure 4: Constraint presented in ontology in SWRL and OWL syntax to the execution policy from Common ontology, OLAPA is notified with a request to rebuild all DW reports where brands of identified phones can be found in dimension elements. After running OLAP on Sales schema with constrains of Nokia brand in Phone dimension and last year in Date dimension OLAPA creates a report as depicted in Example 4.3. Example 4.3 (Business Intelligence findings). [Phone] by brand [Nokia], by date [Q1, 2010] → [Amount sold], [risen by 11, 23%] [Phone] by brand [Nokia], by date [Last month] → [Amount sold], [risen by 5, 87%] At the information system level the first finding is represented as an excerpt from ontology and is depicted in Figure 6. The fields that appear in the report are all instances of Domain specific element, OLAP element and Finding from ontology (see Figure 2). After these findings have been asserted, KDA will be executed to derive new knowledge. Based on these new facts represented at Example 4.3 and enterprise business rules (see Example 4.2), the KDA produced results represented in Example 4.4, by using inference engine knowledge is asserted in ontology. Example 4.4 (Derived finding). [New customer] , [Nokia E72] → [Promotion discount], [10%] After consolidation of all new findings KDA sends message to NA with request to forward notifications to appropriate users. The result of triggered activity of NA is the list of business users that have to be notified about this event. The list shows that in this case CMO and CEO have to be notified whereas their context has to be considered. According to CMO's preferences an e-mail is sent with the following content presented in Example 4.5. Example 4.5 (Report of findings with explanation). FACTS New [Phone][Nokia E72] is available on the market. [Amount sold] of [Phone] by brand [Nokia] and by date [Q1, 2010] have [risen by 11, 23%] and by date [last month] have [risen by 5, 87%]. CONCLUSION [Phone][Nokia E72] should be offered to [new customer] and offered at [promotion discount] of [10%]. 13 Figure 5: Case study of using DSS-MAS in mobile phone domain Figure 6: Example of representing the finding in ontology 14 The CEO uses a Mobile Agent on his mobile device and is also notified by a truncated message with new finding, while explanation is available upon request. 5 Conclusion In this paper we have discussed DSS-MAS where internal and external data is integrated using agent-oriented approach and ontologies as a common understanding of a problem domain and for communication between business users and agents. Agents were used due to their mentalistic notions for modelling, similarities between the agent in the MAS paradigm and the human actor in business organisations, and also possibilities for the use of ontologies as means of agents' internal knowledge base representation. The external information from the Web was integrated by IRA agent with the data in organisation's DW and after applying BR new knowledge was derived by employing agents' inference capabilities. Tasks like information retrieval from competitors, creating and reviewing OLAP reports are autonomously performed by agents, while business users have control over their execution through manipulation of knowledge base. The research also has emphasized agent-to-business user communication, trying to minimize that gap. This was accomplished by introducing different views on ontologies for business user and agent. While agents deal with formal description of business concepts, logical constraints and rules, business user has a simplified view of formal description of knowledge. User is able to manipulate with ontology through templates, where little technical knowledge is required. The mediation mechanism transforms these business level concepts into formal specification at the level of information system. Presented approach was verified and implemented using a case study from the domain of mobile telecommu- nications, where the aim was to provide the knowledge worker an intelligent analysis platform that enhances decision making process. The application domain was reduced to its sub domain dedicated for supply and analysis of demand of mobile phones in one of the mobile operators. DW system is constructed from several heterogeneous data sources where majority of those sources are internal to the enterprise. Our approach added information found on the Web (i.e. competitors' offers, stock rates, etc.) to these internal data sources and improved the decision support process within the enterprise. The proposed approach also addressed business users and their communication with the system which was simplified by using templates to define some business requirements that were transformed into analyses models (OLAP, DW, etc.), automatically performed by agents which reported results back to users in charge. The case study presented in the paper was implemented in Java and using mainly open source technologies. References Bernaras, A., Laresgoiti, I., and Corera, J. (1996). Building and reusing ontologies for electrical network applications. In European Conference on Artificial Intelligence (ECAI '96). Buche, P., Dibie-Barthelemy, J., Haemmerle, O., and Hignette, G. (2006). Fuzzy semantic tagging and flexible querying of XML documents extracted from the Web. Journal of Intelligent Information Systems, 26(1):25–40. Cao, L. B., Chao, L., Dan, L., and Li, L. (2004). Ontology services-based information integration in mining telecom business intelligence. Pricai 2004: Trends in Artificial Intelligence, Proceedings, 3157:85–94. Chen, M., Zhu, Q. M., and Chen, Z. X. (2001). An integrated interactive environment for knowledge discovery from heterogeneous data resources. Information and Software Technology, 43(8):487–496. Clark, K. L. and McCabe, F. G. (2007). Ontology schema for an agent belief store. International Journal of Human-Computer Studies, 65(7):640–658. Davies, J., Studer, R., and Warren, P. (2006). Semantic Web technologies - trends and research in ontology- based systems. John Wiley & Sons, Chichester, England. 15 Dzemydiene, D. and Tankeleviciene, L. (2009). Multi-layered knowledge-based architecture of the adaptable distance learning system. Technological and Economic development of Economy, 15(2):229–244. Fuentes, V., Carbo, J., and Molina, J. M. (2006). Heterogeneous domain ontology for location based information system in a multi-agent framework. Intelligent Data Engineering and Automated Learning - Ideal 2006, Proceedings, 4224:1199–1206. Garces, P. J., Olivas, J. A., and Romero, F. P. (2006). Concept-matching IR systems versus word-matching information retrieval systems: Considering fuzzy interrelations for indexing Web pages. Journal of the American Society for Information Science and Technology, 57(4):564–576. Gruber, T. R. (1993). A Translation Approach to Portable Ontology Specifications. Knowledge Acquisition, 5(2):199–220. Guarino, N. (1998). Formal Ontology and Information Systems. In FOIS'98, pages 3–15. IOS Press. Horrocks, I., Patel-Schneider, P. F., Bechhofer, S., and Tsarkov, D. (2005). OWL rules: A proposal and prototype implementation. Journal of Web Semantics, 3(1):23–40. Jovanović, J. and Gašević, D. (2005). Achieving knowledge interoperability: An XML/XSLT approach. Expert Systems with Applications, 29(3):535–553. Kang, N. and Han, S. (2003). Agent-based e-marketplace system for more fair and efficient transaction. Decision Support Systems, 34(2):157–165. Kishore, R., Zhang, H., and Ramesh, R. (2006). Enterprise integration using the agent paradigm: foundations of multi-agent-based integrative business information systems. Decision Support Systems, 42(1):48–78. Lei, Y., Motta, E., and Dominque, J. (2002). IIPS: An Intelligent information presentation system. In 7th International Conference on Intelligent User Interfaces. Liao, L. J., Xu, K. Q., and Liao, S. S. Y. (2005). Constructing intelligent and open mobile commerce using a semantic web approach. Journal of Information Science, 31(5):407–419. Macias, J. A. and Castells, P. (2007). Providing end-user facilities to simplify ontology-driven web application authoring. Interacting with Computers, 19(4):563–585. Marotta, A., Motz, R., and Ruggia, R. (2002). Managing source schema evolution in Web Warehouses. journal of the Brazilian Computer Society, 8(2):20–31. Orgun, B. and Vu, J. (2006). HL7 ontology and mobile agents for interoperability in heterogeneous medical information systems. Computers in Biology and Medicine, 36(7-8):817–836. Qiu, R. G. (2006). Towards ontology-driven knowledge synthesis for heterogeneous information systems. Journal of Intelligent Manufacturing, 17(1):99–109. Rivest, S., Bedard, Y., Proulx, M. J., Nadeau, M., Hubert, F., and Pastor, J. (2005). SOLAP technology: Merging business intelligence with geospatial technology for interactive spatio-temporal exploration and analysis of data. ISPRS Journal of Photogrammetry and Remote Sensing, 60(1):17–33. Russomanno, D. J. and Kothari, C. R. (2004). Expressing inter-link constraints in OWL knowledge bases. Expert Systems, 21(4):217–228. Shan, F., Jun, M., Chao, T., and Xu, L. D. (2003). A programmable agent for knowledge discovery on the web. Expert Systems, 20(2):79–85. Shim, J. P., Warkentin, M., Courtney, J. F., Power, D. J., Sharda, R., and Carlsson, C. (2002). Past, present, and future of decision support technology. Decision Support Systems, 33(2):111–126. Singh, S., Vajirkar, P., and Lee, Y. (2003). Context-based data mining using ontologies. Conceptual Modeling - ER 2003, Proceedings, Lecture Notes in Computer Science, 2813:405–418. 16 Soo, V. W., Lin, S. Y., Yang, S. Y., Lin, S. N., and Cheng, S. L. (2006). A cooperative multi-agent platform for invention based on patent document analysis and ontology. Expert Systems with Applications, 31(4):766–775. Stanford Medical Informatics (2006a). Protégé. Stanford Medical Informatics (2006b). SWRLTab. Suomela, S. and Kekalainen, J. (2006). User evaluation of ontology as query construction tool. Information Retrieval, 9(4):455–475. Tewari, G., Youll, J., and Maes, P. (2003). Personalized location-based brokering using an agent-based intermediary architecture. Decision Support Systems, 34(2):127–137. Trkman, P., Štemberger, M. I., Jaklič, J., and Groznik, A. (2007). Process approach to supply chain integration. Supply chain management - An International journal, 12(2):116–128. Tseng, F. S. C. and Chou, A. Y. H. (2006). The concept of document warehousing for multi-dimensional modeling of textual-based business intelligence. Decision Support Systems, 42(2):727–744. Vahidov, R. and Kersten, G. E. (2004). Decision station: situating decision support systems. Decision Support Systems, 38(2):283–303. Vasilecas, O. and Bugaite, D. (2006). Ontology-based information systems development: The problem of automation of information processing rules. In Advances in Information Systems, Proceedings, volume 4243 of Lecture Notes in Computer Science, pages 187–196. Zhou, L., Booker, Q. E., and Zhang, D. (2002). ROD - Toward Rapid Ontology Development for Underdevel- oped Domains. In 35th Hawaii International Conference on System Sciences, pages 957–965. 17
1906.06863	2	1906	2019-06-18T13:01:39	A Generic Approach for Accelerating Belief Propagation based DCOP Algorithms via A Branch-and-Bound Technique	[ "cs.MA" ]	Belief propagation approaches, such as Max-Sum and its variants, are a kind of important methods to solve large-scale Distributed Constraint Optimization Problems (DCOPs). However, for problems with n-ary constraints, these algorithms face a huge challenge since their computational complexity scales exponentially with the number of variables a function holds. In this paper, we present a generic and easy-to-use method based on a branch-and-bound technique to solve the issue, called Function Decomposing and State Pruning (FDSP). We theoretically prove that FDSP can provide monotonically non-increasing upper bounds and speed up belief propagation based DCOP algorithms without an effect on solution quality. Also, our empirically evaluation indicates that FDSP can reduce 97\% of the search space at least and effectively accelerate Max-Sum, compared with the state-of-the-art.	cs.MA	cs	A Generic Approach to Accelerating Belief Propagation based Incomplete Algorithms for DCOPs via A Branch-and-Bound Technique Ziyu Chen,1Xingqiong Jiang,1Yanchen Deng2,∗ and Dingding Chen1 and Zhongshi He1 College of Computer Science, Chongqing University, Chongqing, China 1 {chenziyu, jxq, dingding, zshe}@cqu.edu.cn, 2 [email protected] 9 1 0 2 n u J 8 1 ] A M . s c [ 2 v 3 6 8 6 0 . 6 0 9 1 : v i X r a Abstract Belief propagation approaches, such as Max-Sum and its variants, are important methods to solve large-scale Dis- tributed Constraint Optimization Problems (DCOPs). How- ever, for problems with n-ary constraints, these algorithms face a huge challenge since their computational complexity scales exponentially with the number of variables a func- tion holds. In this paper, we present a generic and easy-to- use method based on a branch-and-bound technique to solve the issue, called Function Decomposing and State Pruning (FDSP). We theoretically prove that FDSP can provide mono- tonically non-increasing upper bounds and speed up belief propagation based incomplete DCOP algorithms without an effect on solution quality. Also, our empirically evaluation indicates that FDSP can reduce 97% of the search space at least and effectively accelerate Max-Sum, compared with the state-of-the-art. Introduction Distributed Constraint Optimization Problems (DCOPs) which require agents to coordinate their decisions to op- timize a global objective, are a fundamental framework for modeling multi-agent coordination in multi-agent sys- tems (Hirayama and Yokoo 1997). Thus, DCOPs are widely deployed in some real world coordination tasks including meeting scheduling (Enembreck and Barths 2012), sensor networks (Farinelli, Rogers, and Jennings 2014), power net- works (Fioretto et al. 2017), etc. Algorithms for DCOPs can be classified into two cat- egories: complete and incomplete, according to whether they guarantee to find the optimal solution. Complete al- gorithms (Hirayama and Yokoo 1997; Modi et al. 2005; Petcu and Faltings 2005; Yeoh, Felner, and Koenig 2008; Vinyals, Rodriguez-Aguilar, and Cerquides 2009; Gersh- man, Meisels, and Zivan 2009) can get the optimal solu- tions but incur exponential communication or computation overheads since DCOPs are NP-hard. In contrast, incom- plete algorithms (Maheswaran, Pearce, and Tambe 2004; Arshad and Silaghi 2004; Zhang et al. 2005; Ottens, Dim- itrakakis, and Faltings 2012; Nguyen, Yeoh, and Lau 2013; ∗Corresponding author. Copyright c(cid:13) 2019, Association for the Advancement of Artificial Intelligence (www.aaai.org). All rights reserved. Okamoto, Zivan, and Nahon 2016) trade accuracy for com- putation time and memory so that they can be applied to large-scale problems. As a kind of incomplete algorithms based on belief propagation, Max-Sum (Farinelli et al. 2008) and its variants (Rogers et al. 2011; Zivan and Peled 2012; Chen, Deng, and Wu 2017) have drawn a lot of attention since they can easily be deployed to any DCOP setting. Moreover, they can explicitly handle n-ary constraints and more variables per agent (Cerquides et al. 2014). In more detail, agents in Max-Sum propagate and accumulate beliefs through the whole factor graph. And each agent only holds its belief about the utility for each possible assignment and continuously updates its belief based on the messages re- ceived from its neighbors. In spite of many advantages of belief propagation ap- proaches, they suffer from a huge challenge in scalability. Specifically, they perform maximization operations repeat- edly to locally accumulate beliefs for the involved variables, given the local utility function and a set of incoming mes- sages. The computation complexity of this step grows expo- nentially as the number of constraint arities. In other words, when a constraint function holds n variables and the do- main size of each variable is d, Max-Sum needs to perform dn maximization operations to yield the best assignment for each variable. To address the issue, two kinds of methods were proposed to improve the scalability of belief propagation approaches. The first kind is the algorithms based on a branch-and-bound technique including BnB-MS (Stranders et al. 2009) and BnB-FMS (Macarthur et al. 2011) which both consist of a preprocessing phase and a pruning phase. In the preprocess- ing phase, the two algorithms use localized message-passing to simplify DCOPs. Specifically, BnB-MS reduces the num- ber of moves that each agent needs to consider in coordinat- ing mobile sensors while BnB-FMS removes tasks that an agent should never perform in dynamic task allocations. In the pruning phase, both algorithms reduce the search space using a branch-and-bound technique to speed up maximiza- tion operations. Unfortunately, these algorithms require to exchange a lot of messages in their preprocessing phases. Moreover, the bounds in these algorithms are computed by either brute force or domain-specific knowledge, which lim- its their applicability. The second kind of approaches is sorting based, such as 2 n−1 G-FBP(Kim and Lesser 2013) and GDP(Khan et al. 2018), which is applicable to all DCOP settings. G-FBP uses par- tially sorted lists to adapt FBP. Specifically, it selects and sorts the top cd values of the search space, presuming that the maximum value can be found from the selected range. Here, c is a constant. However, G-FBP will incur additional computation once the maximum value cannot be found within the selected range. Different from G-FBP, the main idea of GDP is to explore only the rows that can cover the differences between the sum of the maximal utility of each message and the message utility corresponding to the assignment that produces the largest local utility. Thus, GDP needs to sort the local utilities of all function-nodes inde- pendently by each assignment of each variable in the pre- processing phase and Vi is the sorted result of each assign- ment i. Then, GDP returns a pruned range [p, q] or [p, q) according to whether q == p − t, where p = max(Vi), q = maxc{c ∈ Vi : c ≤ (p − t)} and t = m − b. Here, m is the summation of the maximum value for each received mes- sage from other variable-nodes and b is the summation for the corresponding values of p from the incoming messages of a function-node. However, GDP needs additional time to perform sorting operations in the preprocessing phase. More importantly, GDP is an one-shot pruning procedure that can- not use the learned experience from the assignment combi- nations explored to dynamically prune the search space. Given the background, we devote to develop a generic and fast method for belief propagation based on a branch-and- bound technique, called Function Decomposing and State Pruning (FDSP). In more detail, we propose a domain- independent approach based on dynamic programming to effectively evaluate the upper bound of a given partial as- signment, which overcomes the aforementioned drawbacks of BnB-MS and BnB-FMS. We further enforce the upper bounds by exploiting the fact that the assignment of the tar- get variable is given. Finally, we prune the search space whenever the upper bound of a partial assignment is no greater than the best lower bound explored so far. The exper- imental results show the effectiveness of FDSP which can prune at least 97% of the search space when solving com- plex problems. Background Distributed Constraint Optimization Problems A distributed constraint optimization problem can be repre- sented by a tuple (cid:104)A, X, D, F(cid:105) such that: • A = {a1, a2, . . . , ah} is a set of agents. • X = {x1, x2, . . . , xq} is a set of variables. • D = {D1, D2, . . . , Dq} is a set of finite and discrete do- mains, variable xi taking an assignment value in Di. • F = {F1, F2, . . . , Fr} is a set of constraints, where each constraint Fk : xk → R+ denotes how much utility is assigned to each possible combination of assignments of the involved variables xk ⊆ X. Thus, a constraint function Fk(xk) denotes the utility for each possible assignment combination of the variables in xk, Figure 1: A DCOP instance using a factor graph n = xk represents the arity of Fk, and d = Di denotes the domain size of variable xi. Note that the variables in xk are ordered according to their own indices, where a variable xk,i is ordered before a variable xk,j if i < j. Given this, the goal for a DCOP is to find the joint variable assignment X∗ such that a given global objective function is maximal. Generally, the objective function is described as the sum over F : (cid:88) X∗ = arg max X Fk(xk)∈F,xk⊆X Fk(xk) Max-Sum Algorithm As a belief propagation approach, Max-Sum is a message- passing inference-based algorithm operating on factor graphs which comprise variable-nodes and function-nodes. Function-nodes which represent the constraints in a DCOP are connected to variable-nodes they depend on, while variable-nodes which represent the variables in a DCOP are connected to function-nodes they are involved in. As shown in Fig. 1, F1 and F2 are two function-nodes, x1, x2, x3, x4 and x5 are variable-nodes, where x1, x2, x3 and x4 are con- nected to F2, and x1, x4 and x5 are connected to F1. Here, F1 (i.e., F1(x1), where x1 = {x1, x4, x5}) is a 3-ary con- straint and F2 (i.e., F2(x2), where x2 = {x1, x2, x3, x4}) is a 4-ary one. In Max-Sum, beliefs are propagated and accumulated through the whole factor graph via the messages exchanged between variable-nodes and function-nodes. The message from a variable-node xi to a function-node Fk (xk), called the query message. It is defined by (cid:88) (1) Fj∈Ni\Fk RFj→xi (xi) Qxi→Fk (xi) = αik + where αik is a scalar set such that(cid:80) (cid:32) Fk (xk) + (cid:80) Qxi→Fk (xi) = 0, xi ∈ xk and Ni\Fk is a set of neighbors of xi except the target function-node Fk. The response message sent from a (cid:33) function-node Fk (xk) to a variable-node xi is given by Qxj→Fk (xj) (2) RFk→xi(xi) = max xk\xi xi xj∈xk\xi When a variable-node xi makes its decision, it first ac- cumulates the belief for each possible assignment from all messages it receives. Then, it selects a value to maximize the total utilities. The procedure can be formalized by: (cid:88) Fk∈Ni x∗ i = arg max xi RFk→xi(xi) (3) x5F1x4x1F2x3x2the base case t = (20+17+10)−(9+17+8) = 13. Hence, p − t = 0. Accordingly, GDP returns a fixed pruned range [13, 1] for value G. Obviously, the pruned range contains the entire search space of value G, and cannot be reduced in the subsequent search process. Under such circumstances, we propose a generic, fast and easy-to-use approach based on a branch-and-bound tech- nique, called FDSP that can use the learned experience from the combinations explored to dynamically prune the search space. FDSP FDSP generally consists of two components: estimation to provide upper bounds and pruning to reduce the search space. To provide the optimal upper bound for a partial assignment, the estimation must return the upper bounds for both the local function and the incoming messages. FDSP computes the function estimations in the preprocess- ing phase, called Function Decomposing (FD), while the message estimations are (re)constructed once the messages have changed. Pruning is implemented by a procedure called State Pruning (SP) which is based on a branch and bound technique. That is, the algorithm does not expand any partial assignment whose upper bound is less than the known lower bound. FDSP can be easily applied to any belief propagation based incomplete algorithms to deal with DCOPs with n-ary constraints. Function Decomposing serves in a preprocessing phase to compute the function estimation for each variable of a function-node Fk(xk). Given a partial assignment P Axk,i to variables {xk,w ∈ xk1 ≤ w ≤ i}, the upper bound of the xk,1 local function is maximization of Fk(xk) over the remaining unassigned variables. That is F unEstxk,i (P Axk,i xk,1 , z) (4) Here, F unEstxk,i (·) is the uninformed function estimation for the i-th variable xk,i in xk, which provides optimistic up- per bounds on the utilities of the subsequent search spaces of P Axk,i xk,1. Macarthur et al. (2011) tried to compute the esti- mation by using brute force, which incurs exponential oper- ations for each partial assignment. They also suggested that the domain characteristics can be used to compute the esti- mation, which has a limited generalization and cannot guar- antee the tightness. In contrast, our proposed FD is an one- shot preprocessing procedure that uses dynamic program- ming to compute the estimation, which can significantly re- duce the computation efforts. Specifically, the estimations are computed recursively according to Eq. (5). z={xk,jj>i} Fk(P Axk,i xk,1 ) = max (cid:40)Fk (xk) F unEstxk,i = max xk,i+1 F unEstxk,i+1 i = n otherwise (5) That is, the estimation for a variable is maximization of the one for the next variable. Particularly, the estimation for the last variable is the function itself. Note that, compared to the exponential overhead for each partial assignment in BnB- MS, our proposed FD only requires O(di+1) operations to (a) (b) Figure 2: The utility matrix and message exchange for F2 In this paper, we use the term "target variable" to denote the destination variable-node of the outgoing message being computed, and Mxk,i represents the message from variable xk,i. Proposed Method Motivation The maximization operation in Eq. (2) is the most com- putationally expensive operation in Max-Sum. For a n-ary DCOP, the complexity of computing the response message is O(dn). Take Fig. 2 as an example. Assume that each vari- able takes a value in {R, G} and the function-node F2 has received the messages Mx1 = {9, 20}, Mx2 = {17, 11} and Mx3 = {8, 10} from x1, x2 and x3, respectively. Then, F2 requires dn = 24 = 16 operations to generate the mes- sage RF2→x4 (x4) since its domain size d = 2 and arity n = 4. Obviously, the complexity of this step grows ex- ponentially as d and n scale up. Therefore, this is a huge challenge for scalability of belief propagation algorithms. As mentioned earlier, some efforts have been made to op- timize this maximization operation. Nevertheless, the im- proved algorithms based on a branch-and-bound technique including BnB-MS and BnB-FMS require a number of mes- sages to be passed in the preprocessing phase. And, these algorithms were proposed for the exact application, which makes it difficult to directly solve general DCOPs. Be- sides, the algorithms based on sorting, such as G-FBP and GDP, suffer from some drawbacks although they are generic. Specifically, G-FBP cannot guarantee that the maximum value can be found in the selected range, which can lead to a complete traverse to all the possible combinations in the worst case, while GDP requires sorting for each value in the domain of each variable in the preprocessing phase, which makes its use prohibitively expensive. Additionally, GDP cannot use the learned knowledge from the combi- nations explored to dynamically prune the search space. In other words, GDP is actually an one-shot pruning method. Taking Fig. 2 for example, according to GDP, the local util- ities of F2 are sorted independently by each value of the do- main. When computing the pruned range of value G, there are VG = {13, 9, 8, 7, 5, 5, 4, 1}, p = max(VG) = 13 and x1x2x3x4F2RRRR4RRRG13RRGR26RRGG5RGRR1RGRG5RGGR2RGGG4GRRR10GRRG7GRGR6GRGG9GGRR11GGRG8GGGR12GGGG1......x4x1F2x3x2RF2→x4(x4){9,20}{8,10}{17,11}Figure 3: The sketch of function decomposing compute the function estimation for each variable xk,i in the preprocessing phase. In fact, the uninformed function estimation F unEstxk,i for xk,i could provide a tighter upper bound if we know the assignment of a variable xk,j such that j > i. In this way, we can compute a tight upper bound even if there are many unassigned variables between the last assigned variable and the target variable in a partial assignment. By considering all the possible assignments of each variable ordered after xk,i, we further reinforce the upper bound and propose the informed function estimations. Eq. (6) gives the formal def- inition to the informed function estimation for xk,i in terms of xk,j where j > i. (cid:40)F unEstxk,j (vk,j) max xk,i+1 F unEstxk,j=vk,j xk,i+1 i = j − 1 otherwise (6) F unEstxk,j=vk,j xk,i = Similar to the uninformed function estimations, the in- formed function estimations are computed in a recursive fashion by maximizing the estimation for the next variable. And the estimation for the last variable before xk,j is the corresponding uninformed function estimation with a given assignment xk,j = vk,j. Fig. 3 gives the sketch of FD. The procedure begins with computing the uninformed function estimation for each vari- able in Fk according to Eq. (5) from the last one to the first one (line 1-3). Then, for every possible assignment of each variable, we compute the informed function estimation for each variable whose index is smaller than the current vari- able according to Eq. (6) (line 4-6). Taking Fig. 2 for example, we can compute the unin- formed function estimations for variable x1, x2, x3 and x4 as follows: F unEstx4 = F2, F unEstx3 = maxx4 F unEstx4 F unEstx2 = maxx3 F unEstx3 F unEstx1 = maxx2 F unEstx2 These estimations can provide the upper bounds for the par- tial assignments with respect to their variables. Besides, the informed function estimations in terms of x4 = R are com- puted as follows: F unEstx4=R F unEstx4=R F unEstx4=R State Pruning is geared towards speeding up the com- putation of the messages from function-nodes to variable- nodes by branch and bound. That is, when the upper bound = F unEstx4(x4 = R) = maxx3 F unEstx4=R = maxx2 F unEstx4=R x3 x2 x3 x2 x1 Figure 4: The sketch of state pruning of a partial assignment is no greater than the lower bound, the search space corresponding to the partial assignment will be discarded. Fig. 4 gives the pseudo code of SP. The algorithm begins with calculating the message esti- mation for each variable xk,i ∈ xk\xk,t, which gives the maximal message utility with respect to all non-target vari- ables after it given these variables unassigned, according to Eq. (7) (line 7-15). M sgEstxk,i = max(Mxk,j) (7) (cid:88) j>i∧j(cid:54)=t Here, xk,t is the target variable and M sgEstxk,i denotes the upper bounds on the received messages from the vari- ables after xk,i except xk,t. In order to reduce the unneces- sary computation, Fk recomputes the message estimations for each variable ordered before xk,i only when the mes- sage from xk,i changes. Besides, instead of directly comput- ing message estimations according to Eq. (7), Fk further re- duces the computation efforts by recursively backing up the maximal message utilities from the last non-target variable Algorithm1:FunctionDecomposing(FD)foreachfunction-nodeFk(xk)FunctionDecompose(Fk(xk))1n←xk2fori←nto1do3computeFunEstxk,ibyEq.(5)4forj←nto2do5foreachvk,j∈Dxk,j∧i<jdo6computeFunEstxk,j=vk,jxk,ibyEq.(6)Algorithm2:StatePruning(SP)foreachfunction-nodeFk(xk)FunctionFDSP(Fk(xk),xk,t,Mxk\xk,t)7n←xk8Result←∅9lb←−∞10msgChanged←false11fori←nto1do12ifMxk,ihaschangedthen13msgChanged←true14ifmsgChanged=truethen15∀xk,i∈xk\xk,t,computeMsgEstxk,ibyEquation716start←117ifxk,start=xk,tthen18start←219foreachvk,t∈Dxk,tdo20Assignxk,t←vk,t21util∗←FDSPRec(Fk(xk),xk,start,xk,t,Mxk\xk,t,0)22Result←Result∪{(vk,t,util∗)}23returnResultFunctionFDSPRec(Fk(xk),xk,i,xk,t,Mxk\xk,t,msgUtil)24next←i+125ifxk,next=xk,tthen26next←next+127util∗←−∞28foreachvk,i∈Dxk,ido29Assignxk,i←vk,i30msgUtilxk,i=msgUtil+Mxk,i(vk,i)31computeubxk,ibyEq.(8)32ifubxk,i>lbthen33ifxk,next(cid:54)=nullthen34retUtil←FDSPRec(Fk(xk),xk,next,xk,t,Mxk\xk,t,msgUtilxk,i)35util∗←max(util∗,retUtil)36else37retUtil←Fk(Assign)+msgUtilxk,i38util∗←max(util∗,retUtil)39lb←max(lb,util∗)40returnutil∗to the first one. That is, the message estimation of a variable is the sum of the maximal message utility and the message estimation of the non-target variable next to it. Consider the function-node F2 in Fig. 2(b). When we are computing the message for x4, the message estimations for x1, x2 and x3 are computed as follows: M sgEstx3 = 0 M sgEstx2 = M sgEstx3 + maxMx3 = 10 M sgEstx1 = M sgEstx2 + maxMx2 = 10 + 17 = 27 Then, Fk computes the maximum utility util∗ of each as- signment of the target variable xk,t in Dxk,t (line 16-22). Specifically, Fk assigns assignment vk,t to xk,t according to the order of values in Dxk,t (line 19). Thus, the current partial joint state Assign = {∅, . . . , vk,t, . . . ,∅}, where ∅ represents an unassigned variable (line 20). After that, FD- SPRec is called for the variable xk,start which is the first unassigned variable to recursively expand the search space (line 21). Finally, Fk stores util∗ to Result when util∗ for the current assignment vk,t is returned (line 22). The proce- dure (line 20 - 22) repeats until all the assignments of xk,t have been visited. In FDSPRec, Fk first finds xk,next that is the unassigned variable next to xk,i (line 24-26). Note that xk,t is an as- signed variable. Then, Fk decides to expand the search space or update the maximum utility and the lower bound (line 28- 39). In more detail, Fk expands the search space by append- ing the assignment vk,i of xk,i to the partial joint state (line 29). And then, Fk computes the utilities contributed (i.e., msgU tilxk,i) by the incoming messages with respect to the current Assign by summing the accumulated msgU til with the entry in terms of Mxk,i and assignment vk,i (line 30). Then, Fk computes the current upper bound ubxk,i accord- ing to Eq. (8) (line 31). (cid:26)msgU tilxk,i + M sgEstxk,i + F unEstxk,i msgU tilxk,i + M sgEstxk,i + F unEstxk,t=vk,t xk,i (cid:0)Assignxk,i (cid:1) (cid:0)Assignxk,i xk,1 xk,1 (cid:1) i < t i > t ubxk,i = (8) Specifically, if i > t(i.e., xk,i is after xk,t), which means the assignments of all variables before xk,t have been given, the current upper bound of the local function is provided by the uninformed function estimation of variable xk,i. Otherwise, the upper bound is computed by the informed function esti- mation. In other words, the informed function estimation is used to compute a tight upper bound whenever it is applica- ble. Next, Fk decides whether to expand the search space ac- cording to the lower bound (line 32-39). If the upper bound ubxk,i is greater than the current lower bound lb and xk,i is not the last variable, the algorithm proceeds by calling the recursive function FDSPRec to expand the search space (line 33-35). Otherwise, the search space corresponding to Assign can be discarded. If xk,i is the last non-target vari- able, i.e., the search space has been fully expanded, Fk com- putes the current utility retU til of the complete assignment by adding the local utility Fk(Assign) and msgU tilxk,i. Then, Fk updates the maximum utility util∗ and the lower bound lb (line 36-39). Finally, when all the assignments of xk,i have been visited, the algorithm returns util∗ (line 40). Figure 5: Calculating RF2→x4(x4 = R) using FDSP Fig. 5 shows an example for calculating the message from function F2 to variable x4 when x4 = R (i.e., RF2→x4(x4 = R)) in Fig. 2, where the numbers with circles represent the trace of SP. Since x4 is fixed to assignment R, F2 needs to compute the maximum utility util∗ by extending the partial assignment Assign = {∅,∅,∅, R}. Firstly, F2 visits the first assignment R of x1 and computes ubx1 = 9 + 27 + 26 = 62 by Eq. (8). Then, it expands Assign = {R,∅,∅, R} by vis- iting the first assignment R of x2 since ubx1 > lb(= −∞). Similarly, F2 expands Assign = {R, R,∅, R}. At this point, since ubx3 > lb and Assign is fully assigned, F2 computes the utility retU til of Assign = {R, R, R, R}: retU til = F2(Assign) + msgU tilx3 = 4 + (9 + 17 + 8) = 38, and updates util∗ = 38 and lb = 38. Next, F2 vis- its the next assignment G of x3. Similarly, F2 computes the current upper bound ubx3 = 62 and the current utility retU til = 62 corresponding to Assign = {R, R, G, R}, and updates lb and util∗. After that, F2 visits the second assignment G of x2 since all assignments of x3 have been exhausted. And, F2 computes ubx2 = 32 which is less than lb, so Assign = {R, G,∅, R} is discarded. Similarly, Assign = {G,∅,∅, R} is also discarded. Finally, F2 finds the maximum utility util∗ = 62, i.e., RF2→x4 (x4 = R) = 62. As seen from the example, FDSP can prune at least 75% of the search space during computing the message from the function-node F2 to variable-node x4, where d = 2 and n = 4. Theoretical Analysis In this section, we will theoretically prove that FDSP can speed up belief propagation based incomplete algorithms without an effect on solution quality, i.e., FDSP can pro- vide monotonically non-increasing upper bounds and never prunes the optimal assignment with the maximum utility util∗. Lemma 1. For a function-node Fk(xk) and a given partial assignment P A with (xk,t = vk,t) in which (xk,i = vk,i) is the last non-target entry, the upper bound of any direct subsequent partial assignment P A(cid:48) = P A ∪ (xk,j = vk,j) is at least as low as the one of P A, where xk,j is the variable next to xk,i such that j (cid:54)= t. Proof. Recall that the upper bound of a given partial as- signment is computed according to either the uninformed function estimation or the informed function estimation, de- pending on the index of the target variable. Thus, three cases {∅,∅,∅,R}{R,∅,∅,R}{R,R,∅,R}{R,R,R,R}{R,R,G,R}[−∞,62]1[62,59]6[−∞,62]2[62,32]5[−∞,38]3[38,62]4need to be discussed: 1) all the upper bounds are computed according to the uninformed function estimations; 2) the up- per bound of P A is computed according to the uninformed function estimation, while the one of P A(cid:48) is computed ac- cording to the informed function estimation; 3) all the up- per bounds are computed according to the informed func- tion estimations. Here, we only give the prove for case 2) (i.e., i + 1 = t, t + 1 = j) due to the limited space. Similar analysis can be applied to case 1) and 3). ubxk,i (P A) = F unEst xk,t=vk,t xk,i (P A) + msgU tilxk,i + M sgEstxk,i l≤i (cid:88) Mxk,l (vk,l) + (cid:88) Mxk,l (vk,l) + (cid:88) l≤i l>i∧l(cid:54)=t (cid:88) max(Mxk,l ) (cid:88) (cid:88) max(Mxk,l ) max(Mxk,l ) l>i∧l(cid:54)=t Mxk,l (vk,l) + l≤j∧l(cid:54)=t l>j = F unEstxk,t (P A) + (cid:48) ≥ F unEstxk,j (P A (cid:48) ≥ F unEstxk,j (P A ) + ) + (cid:48) = ubxk,j (P A ) the second step holds since i = t − 1. Thus, Here, according to Eq. (6) we have F unEst (P A) = F unEstxk,t (P A). and the step = maxxk,j F unEstxk,j (P A, xk,j) ≥ F unEstxk,j (P A(cid:48))(Eq. (5)) and max(Mxk,j ) ≥ Mxk,j (vk,j), respectively. step since F unEstxk,t (P A) xk,t=vk,t xk,i third Besides, hold fourth the Thus the lemma is proved. Theorem 1. FDSP does not affect the optimality of Eq. (2). Proof. Prove by contradiction. For a function-node Fk(xk), assume that the optimal assignment of Eq. (2) is Assign∗, and the corresponding utility value is val(Assign∗). As- sume that FDSP has missed that assignment. Thus, there must exist a partial assignment P A ⊂ Assign∗ such that ub(P A) < lb ≤ val(Assign∗). According to Lemma 1, the upper bound is monotonically non-increasing, i.e., ub(P A) ≥ ub(Assign∗). Note that (cid:88) ∗ ) = F unEstxk,n (Assign ) + msgU tilxk,n + M sgEstxk,n ∗ ubxk,n (Assign ∗ = Fk(Assign ) + Mxk,l (vk,l ) l≤n∧l(cid:54)=t ∗ = val(Assign ) Here, n = xk. Thus, we have ub(P A) ≥ ub(Assign∗) = val(Assign∗), which is contradict to the assumption. There- fore, the upper bound of a partial assignment cannot be less than the value of any subsequent full assignment and the op- timality is hereby guaranteed. Complexity Analysis Each variable xk,i needs to compute and stores an unin- formed function estimation and d(n − i) informed function estimations in the preprocessing phase. Thus, the time and space of each variable require O([1 + d(n − i)]di+1) and O([1 + d(n − i)]di), respectively, where the value of i be- comes smaller as FD performs. Thus, FDSP in the prepro- cessing phase needs a small overhead. Besides, since each function-node needs to explore the search space with respect to the target variable, the time complexity in the worst case is O(dn). However, with SP, only the small search space needs to be explored. Therefore, the overall overhead is small. For this point, our empirical evaluation also verifies that FDSP only requires little time to run. Empirical Evaluation We empirically evaluate the performances of FDSP and GDP which are both applied to Max-Sum on four config- urations of n-ary random DCOPs. Since BnB-MS and BnB- FMS are not generic algorithms and G-FBP is inferior to GDP (Khan et al. 2018), we do not include them for com- parison. The complexity of a n-ary DCOP can be quantified by the number of function-nodes, the average/maximal ar- ity and the domain size (Kim and Lesser 2013; Khan et al. 2018). In addition to these parameters, we also find the num- ber of variable-nodes can affect the complexity. Intuitively, given the number of function-nodes and the average arity per function-node, the graph density is actually determined by the number of function-nodes. Therefore, we introduce a new parameter called variable tightness (denoted as var T ) to depict the complexity from another perspective, which is defined as follows. var T = 1 − number of variable-nodes total number of arities It can be concluded that given the function-node number and total arity number, the number of variable-nodes decreases as var T increases, which will generate a denser and more complex problem since each variable-node has to connect more function-nodes. For each configuration other than the first one, we gener- ate sparse factor graphs and dense factor graphs by randomly selecting var T from [0.1, 0.5] and (0.5, 0.9], respectively. In the first configuration, we set the number of function-nodes to 100 and the minimal arity to 2, and uniformly select the costs, the domain size and the maximal arity from [1, 100], [2, 10] and [2, 7], respectively. And, var T varies from 0.1 to 0.9. In the second one, we vary the maximal arity from 2 to 7. In the third configuration, we vary the number of function-nodes from 10 to 100. In the last one, we set the number of function-nodes to 50 and vary the domain size from 2 to 7. Also, we benchmark Max-Sum ADVP+FDSP to demonstrate the generalization of FDSP. To guarantee Max-Sum ADVP to converge, we alternate its directions ev- ery 100 iterations. All the omitted parameters except var T in each configuration are the same as the ones in the first configuration. For each of the setting, we generate 25 ran- dom instances and the results are averaged over all instances. The algorithms terminate after 200 iterations for each in- stance. Fig. 6 gives the comparison under different var T. It can be observed that FDSP clearly outperforms GDP under dif- ferent var T, and the gap is widen as var T grows. Con- cretely, FDSP can prune at least 97% of the search space while GDP only prunes at most 87% of the search space when computing Eq. (2). Moreover, FDSP performs simi- larly as var T grows, which indicates FDSP is less sensitive to the complexity of problems. On the other hand, the perfor- mance of GDP decreases as var T increases, and GDP per- Figure 6: Performance comparison on different var T Figure 8: Performance comparison on different function- node numbers Figure 7: Performance comparison on different arities Figure 9: Runtime on different domain sizes forms poorly when solving the problems with var T>= 0.5. This is because the sum of the difference between the maxi- mal value and the value corresponding to the maximal local utility in each message will increase when the graph density increases as var T grows. As a result, GDP provides a large pruned range so as to prune only a small proportion of the search space. Fig. 7 shows the performance comparison on different maximal arities. It can be concluded that FDSP outperforms GDP in both sparse and dense factor graphs, especially in dense factor graphs. FDSP prunes around 60%-99% of the search space in both sparse and dense factor graphs, while GDP can only prune at most 80% and 36%, respectively. That is because FDSP provides tighter bounds to make Max- Sum explore fewer combinations. Fig. 8 presents the results under different number of function-nodes. Similar to the first configuration, FDSP prunes at least 97% of the search space in both sparse and dense factor graphs, while GDP can only prune at most 88% and 55% of the search space in sparse and dense factor graphs, respectively. This is because GDP is an one-shot pruning procedure and cannot use the learned experience from the assignment combinations explored to dynamically prune the search space. Fig. 9 gives the runtime under different domain sizes. It can be seen that our FDSP exhibits great superiority over GDP and Max-Sum ADVP when solving the problems with large domain sizes, which indicates that FDSP can scale up well and only requires few computation efforts. GDP would perform even worse in practice since the runtime presented in Fig. 9 actually does not take sorting, which is quite expen- sive when the domain size is large, into consideration. Be- sides, one can easily observe that Max-Sum ADVP+FDSP is superior to Max-Sum ADVP when solving the problems with large domain sizes in sparse and dense factor graphs, which indicates that FDSP can also effectively accelerate the variants of Max-Sum. Conclusion In this paper, we propose FDSP, a generic, fast and easy-to- use method based branch and bound, which significantly ac- celerates belief propagation based incomplete DCOP algo- rithms. Specifically, we first propose function decomposing (FD) to effectively compute the function estimation, which dramatically reduces the overheads in computing an upper bound of a partial assignment. Then, we further present state pruning (SP) based on branch and bound to reduce the search space. Besides, we theoretically prove that our bounds are monotonically non-increasing during the search process and FDSP never prunes the assignment with the maximum utility. Our experimental results clearly show that FDSP can prune around 97%-99% of the search space and only requires little time, especially for the large and complex problems. 0.10.20.30.40.50.60.70.80.9var_T0102030405060708090100% of Search Space PrunedFDSPGDP234567Maximal Arity0102030405060708090100% of Search Space PrunedFDSP-Sparse SettingsGDP-Sparse SettingsFDSP-Dense SettingsGDP-Dense Settings102030405060708090100Function-node Number0102030405060708090100% of Search Space PrunedFDSP-Sparse SettingsGDP-Sparse SettingsFDSP-Dense SettingsGDP-Dense Settings234567Domain Size101102103RunTime(s)FDSP-Sparse SettingsGDP-Sparse SettingsMax-Sum_ADVP-Sparse SettingsMax-Sum_ADVP+FDSP-Sparse SettingsFDSP-Dense SettingsGDP-Dense SettingsMax-Sum_ADVP-Dense SettingsMax-Sum_ADVP+FDSP-Dense SettingsAcknowledgment This research is funded by Chongqing Research Pro- gram of Basic Research and Frontier Technology (No. cstc2017jcyjAX0030), Fundamental Research Funds for the Central Universities (No. 2018CDXYJSJ0026) and Grad- uate Research and Innovation Foundation of Chongqing, China (Grant No. CYS18047). References J.; Farinelli, [Arshad and Silaghi 2004] Arshad, M., and Silaghi, M. C. 2004. Distributed simulated annealing. Distributed Con- straint Problem Solving and Reasoning in Multi-Agent Sys- tems 112. A.; [Cerquides et al. 2014] Cerquides, 2014. A tutorial Meseguer, P.; and Ramchurn, S. D. on optimization for multi-agent systems. Computer Journal 57(6):799 -- 824. [Chen, Deng, and Wu 2017] Chen, Z.; Deng, Y.; and Wu, T. 2017. An iterative refined max-sum ad algorithm via single- side value propagation and local search. In Proc. of the 16th Conference on AAMAS, 195 -- 202. [Enembreck and Barths 2012] Enembreck, F., and Barths, J. P. A. 2012. Distributed constraint optimization with mulbs: A case study on collaborative meeting scheduling. Journal of Network & Computer Applications 35(1):164 -- 175. [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 Proc. of the 7th Conference on AAMAS, 639 -- 646. [Farinelli, Rogers, and Jennings 2014] Farinelli, A.; Rogers, A.; and Jennings, N. R. 2014. Agent-based decentralised coordination for sensor networks using the max-sum al- Autonomous Agents and Multi-Agent Systems gorithm. 28(3):337 -- 380. [Fioretto et al. 2017] Fioretto, F.; Yeoh, W.; Pontelli, E.; Ma, Y.; and Ranade, S. J. 2017. A distributed constraint opti- mization (dcop) approach to the economic dispatch with de- mand response. In Proc. of the 16th Conference on AAMAS, 999 -- 1007. [Gershman, Meisels, and Zivan 2009] Gershman, A.; Meisels, A.; and Zivan, R. 2009. Asynchronous for- ward bounding for distributed cops. Journal of Artificial Intelligence Research 34:61 -- 88. [Hirayama and Yokoo 1997] Hirayama, K., and Yokoo, M. 1997. Distributed partial constraint satisfaction problem. In International Conference on Principles and Practice of Constraint Programming, 222 -- 236. Springer. [Khan et al. 2018] Khan, M.; Tran-Thanh, L.; Jennings, N.; et al. 2018. A generic domain pruning technique for gdl- based dcop algorithms in cooperative multi-agent systems. In Proc. of the 17th Conference on AAMAS, 1595 -- 1603. [Kim and Lesser 2013] Kim, Y., and Lesser, V. 2013. Im- proved max-sum algorithm for dcop with n-ary constraints. In International Conference on Autonomous Agents and Multi-Agent Systems, 191 -- 198. [Macarthur et al. 2011] Macarthur, K. S.; Stranders, R.; Ramchurn, S. D.; and Jennings, N. R. 2011. A distributed anytime algorithm for dynamic task allocation in multi- In AAAI Conference on Artificial Intelli- agent systems. gence, 701 -- 706. [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 Proceedings of the International Conference on Parallel and Distributed Computing Systems (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. [Nguyen, Yeoh, and Lau 2013] Nguyen, D. T.; Yeoh, W.; and Lau, H. C. 2013. Distributed gibbs: A memory-bounded sampling-based dcop algorithm. In Proc. of the 12th Con- ference on AAMAS, 167 -- 174. [Okamoto, Zivan, and Nahon 2016] Okamoto, S.; Zivan, R.; and Nahon, A. 2016. Distributed breakout: beyond satisfac- tion. In Proc. of the 25th IJCAI, 447 -- 453. [Ottens, Dimitrakakis, and Faltings 2012] Ottens, B.; Dimi- trakakis, C.; and Faltings, B. 2012. Duct: An upper confi- dence bound approach to distributed constraint optimization In Twenty-Sixth AAAI Conference on Artificial problems. Intelligence, 528 -- 534. [Petcu and Faltings 2005] Petcu, A., and Faltings, B. 2005. A scalable method for multiagent constraint optimization. In Proc. of the 19th IJCAI, 266 -- 271. [Rogers et al. 2011] Rogers, A.; Farinelli, A.; Stranders, R.; and Jennings, N. R. 2011. Bounded approximate decen- tralised coordination via the max-sum algorithm. Artificial Intelligence 175(2):730 -- 759. [Stranders et al. 2009] Stranders, R.; Farinelli, A.; Rogers, A.; and Jennings, N. R. 2009. Decentralised coordination of mobile sensors using the max-sum algorithm. In IJCAI 2009, Proceedings of the International Joint Conference on Artificial Intelligence, Pasadena, California, Usa, July, 299 -- 304. [Vinyals, Rodriguez-Aguilar, and Cerquides 2009] Vinyals, M.; Rodriguez-Aguilar, J. A.; and Cerquides, J. 2009. Gen- eralizing dpop: Action-gdl, a new complete algorithm for dcops. In Proceedings of The 8th International Conference on Autonomous Agents and Multiagent Systems-Volume 2, 1239 -- 1240. International Foundation for Autonomous Agents and Multiagent Systems. [Yeoh, Felner, and Koenig 2008] Yeoh, W.; Felner, A.; and Koenig, S. 2008. Bnb-adopt: an asynchronous branch-and- bound dcop algorithm. In International Joint Conference on Autonomous Agents and Multiagent Systems, 591 -- 598. [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. [Zivan and Peled 2012] Zivan, R., and Peled, H. 2012. Max/min-sum distributed constraint optimization through value propagation on an alternating dag. In Proceedings of the 11th International Conference on Autonomous Agents and Multiagent Systems-Volume 1, 265 -- 272. International Foundation for Autonomous Agents and Multiagent Sys- tems.
1904.00820	2	1904	2019-05-28T18:34:59	A Game-Theoretic Framework for Resource Sharing in Clouds	[ "cs.MA", "cs.NI" ]	Providing resources to different users or applications is fundamental to cloud computing. This is a challenging problem as a cloud service provider may have insufficient resources to satisfy all user requests. Furthermore, allocating available resources optimally to different applications is also challenging. Resource sharing among different cloud service providers can improve resource availability and resource utilization as certain cloud service providers may have free resources available that can be ``rented'' by other service providers. However, different cloud service providers can have different objectives or \emph{utilities}. Therefore, there is a need for a framework that can share and allocate resources in an efficient and effective way, while taking into account the objectives of various service providers that results in a \emph{multi-objective optimization} problem. In this paper, we present a \emph{Cooperative Game Theory} (CGT) based framework for resource sharing and allocation among different service providers with varying objectives that form a coalition. We show that the resource sharing problem can be modeled as an $N-$player \emph{canonical} cooperative game with \emph{non-transferable utility} (NTU) and prove that the game is convex for monotonic non-decreasing utilities. We propose an $\mathcal{O}({N})$ algorithm that provides an allocation from the \emph{core}, hence guaranteeing \emph{Pareto optimality}. We evaluate the performance of our proposed resource sharing framework in a number of simulation settings and show that our proposed framework improves user satisfaction and utility of service providers.	cs.MA	cs	A Game-Theoretic Framework for Resource Sharing in Clouds 1 Faheem Zafari∗, Kin K. Leung∗, Don Towsley†, Prithwish Basu(cid:5), and Ananthram Swami‡ ∗Imperial College London, †University of Massachusetts Amherst, (cid:5)BBN Technologies, ∗{faheem16, kin.leung}@imperial.ac.uk, †[email protected], (cid:5)[email protected], ‡U.S. Army Research Laboratory ‡[email protected] 9 1 0 2 y a M 8 2 ] A M . s c [ 2 v 0 2 8 0 0 . 4 0 9 1 : v i X r a Abstract Providing resources to different users or applications is fundamental to cloud computing. This is a challenging problem as a cloud service provider may have insufficient resources to satisfy all user requests. Furthermore, allocating available resources optimally to different applications is also challenging. Resource sharing among different cloud service providers can improve resource availability and resource utilization as certain cloud service providers may have free resources available that can be "rented" by other service providers. However, different cloud service providers can have different objectives or utilities. Therefore, there is a need for a framework that can share and allocate resources in an efficient and effective way, while taking into account the objectives of various service providers that results in a multi-objective optimization problem. In this paper, we present a Cooperative Game Theory (CGT) based framework for resource sharing and allocation among different service providers with varying objectives that form a coalition. We show that the resource sharing problem can be modeled as an N−player canonical cooperative game with non-transferable utility (NTU) and prove that the game is convex for monotonic non-decreasing utilities. We propose an O(N ) algorithm that provides an allocation from the core, hence guaranteeing Pareto optimality. We evaluate the performance of our proposed resource sharing framework in a number of simulation settings and show that our proposed framework improves user satisfaction and utility of service providers. I. INTRODUCTION With the wide-scale proliferation of cloud computing, a large number of individuals and businesses have transitioned towards the use of cloud computing platforms [1]. Users (applications)1 request resources from a cloud service provider for a particular task and service provider needs to provide requested resources based on a Service Level Agreement (SLA). The goal for any service provider is to satisfy as many applications as possible by a timely provision of resources, resulting in an increase in user satisfaction as well as utility of service providers. However, the provider often lacks sufficient available resources and thus allocation of resources to different users is not straightforward. Optimal allocation of resources for cloud computing is a fundamental research problem [2] -- [4]. Due to the recent advent of paradigms such as Big Data [5], Internet of Things (IoT) [6], and Machine to Machine (M2M) communication, the problem is further exacerbated as the demand for limited resources has increased further. Even when resources are available, allocating them optimally to different applications requires efficient algorithms. Hence, it is important to ascertain the availability of resources and allocate them optimally to different applications. To ensure the availability of resources, numerous solutions have been proposed in the literature ranging from buying more resources to meet peak demands, to creating shared resource pools among geographically co-located service providers. However, different service providers may have different objectives and existing resource pool based mechanisms do not consider utilities of different service providers. Ideally, a resource sharing mechanism should consider objectives of different service providers, which results in a multi-objective optimization (MOO) problem. Furthermore, while resource sharing or creation of a resource pool may tackle the resource availability problem, it does not address resource allocation problem, i.e., optimally allocating the available resources to different users. Therefore, there is a need for a framework that not only ascertains resource availability by enabling resource sharing among different service providers, but also optimally allocates resources to users, while accounting for the service providers' objectives. We present a cooperative game theory based framework that can be used for resource sharing and allocation in cloud computing. We allow resource sharing among different cloud service providers, while considering the objectives (utilities) of cloud service providers. Rather than considering a specific resource, we consider a wide range of resources that are present across different cloud service providers. We show that different resource providers can benefit by creating a coalition among This work was supported by the U.S. Army Research Laboratory and the U.K. Ministry of Defence under Agreement Number W911NF-16-3-0001. The views and conclusions contained in this document are those of the authors and should not be interpreted as representing the official policies, either expressed or implied, of the U.S. Army Research Laboratory, the U.S. Government, the U.K. Ministry of Defence or the U.K. Government. The U.S. and U.K. Governments are authorized to reproduce and distribute reprints for Government purposes notwithstanding any copy-right notation hereon. Faheem Zafari also acknowledges the financial support by EPSRC Centre for Doctoral Training in High Performance Embedded and Distributed Systems (HiPEDS, Grant Reference EP/L016796/1), and Department of Electrical and Electronics Engineering, Imperial College London. 1Throughout the paper, we use the term "users" and "applications" interchangeably. 2 service providers and then allocating resources to users. We also show that the coalition of service providers is stable, where no service provider will incur any loss (when compared with working alone). Also, we obtain a Pareto optimal (details in Section II) solution to our multi-objective resource sharing problem. The main contributions of this paper are: 1) We propose a cooperative game theory (CGT) based multi-objective resource sharing and allocation framework for cloud computing that guarantees Pareto optimality and consider different objectives of the service providers that share their resources. 2) We show that the resource sharing and allocation problem can be modeled as a canonical game with non-transferable utility. Furthermore, when service provider objectives are represented by monotone non-decreasing utilities, we show that our canonical game is convex. Due to the convexity of game, the core is non-empty and the grand coalition formed by all service providers is stable. 3) We address the problem of obtaining an allocation from the core by proposing a computationally efficient O(N ) algorithm that is guaranteed to provide an allocation from the core. 4) We evaluate the performance of our proposed framework through extensive experiments and show that resource sharing can improve utility of all service providers and also increase user satisfaction. The paper is further structured as: Section II provides a primer on multi-objective optimization, cooperative game theory, canonical games, and the core. Section III presents the system model used in this paper. It also presents the resource sharing and allocation problem along with theoretical properties of the cooperative game used to obtain the Pareto optimal solution. Section IV presents our experimental results. Section V provides a review of different solutions proposed in literature for resource allocation, while Section VI concludes the paper. In this section, we present a discussion on multi-objective optimization, cooperative game theory, and the core. II. PRELIMINARIES A. Multi-Objective Optimization For m inequality constraints and p equality constraints, MOO identifies a vector x∗ = [x∗ 1, x∗ 2,··· , x∗ t ]T that optimizes a (1) (2) vector function such that ¯f (x) = [f1(x), f2(x),··· , fN (x)]T gi(x) ≥ 0, i = 1, 2,··· , m, hi(x) = 0 i = 1, 2,··· , p. where x = [x1, x2,··· , xt]T is a vector of t decision variables and the feasible set is denoted by F . The fundamental difference between a single objective optimization (SOO) and MOO is that MOO involves a vector of objective functions rather than a single objective function. Therefore, in MOO, the optimal solution is not a single point but a frontier of solutions known as Pareto frontier or Pareto boundary (see [7] for details). Some basic definitions related to MOO are: Definition 1. Pareto Optimality: For any maximization problem, x∗ is P areto optimal if the following holds for every x ∈ F , (3) ¯f (x∗) ≥ ¯f (x). where ¯f (x) = [f1(x), f2(x),··· , fN (x)]T and ¯f (x∗) = [f1(x∗), f2(x∗),··· , fN (x∗)]T . Figure 1 shows a MOO problem (minimization problem) with two objective functions f1 and f2. The boundary Pareto frontier as it consists of all the Pareto optimal solutions. ←→ ab is the B. Cooperative Game Theory Cooperative game theory provides a set of analytical tools that assists in understanding the behavior of rational players in a cooperative setting [8]. Players can have agreements among themselves that affect the strategies as well as obtained utilities of game players. Coalition games are one of the basic types of cooperative games that deal with the formation of coalitions, namely groups of two or more cooperating players. Formally, Definition 2. Coalition Games [8]: Any coalition game with non-transferable utility (discussed below) can be represented by the pair (N ,V) where N is the set of players that play the game, while V is a set of payoff vectors such that [9]: 1) V(S) is a closed and convex subset of RS. 2) V(S) is comprehensive, i.e., if we are given payoffs x ∈ V(S) and y ∈ RS where y ≤ x, then y ∈ V(S). In other words, if the members of coalition S can achieve a payoff allocation x, then the players can change their strategies to achieve an allocation y. 3 Fig. 1. MOO with two objective functions 3) The set {xx ∈ V(S) and xn ≥ zn,∀n ∈ S}, with zn = max{yny ∈ V({n})} ≤ ∞ ∀n ∈ N is a bounded subset of RS. In other words, the set of vectors in V(S) for a coalition S where the coalition members receive a payoff at least as good as working alone (non-cooperatively) is a bounded set. Definition 3. Value of a coalition: The sum of all players' payoff from a particular pay-off vector for any coalition is known as the value of a coalition. It is worth mentioning that V is the set of payoff vectors, while v is the sum of payoffs that all players get in a particular payoff vector (coalition). Definition 4. Non-Transferable Utility (NTU) [8]: If the total utility of any coalition cannot be assigned a single real number or if there is a rigid restriction on utility distribution among players, then the game has a non-transferable utility. Definition 5. Characteristic function [8]: The characteristic function for any coalition game with NTU is a function that assigns a set of payoff vectors, V(S) ⊆ RS, where each element of the payoff vector xn represents a payoff that player n ∈ S obtains, depending on the selected strategy, within the coalition S. Definition 6. Characteristic form Coalition Games [8]: A coalition game is said to be of characteristic form, if the value of coalition S ⊆ N depends only on the members of coalition. Definition 7. Superadditivity of NTU games [9]: A canonical game with NTU is said to be superadditive if the following property is satisfied. v(S1 ∪ S2) ⊃ {x ∈ RS1∪S2(xn)n∈S1 ∈ v(S1), (xm)m∈S2 ∈ v(S2)} ∀S1 ⊂ N , S2 ⊂ N , S1 ∩ S2 = ∅. (4) i.e., if any two disjoint coalitions S1 and S2 form a large coalition S1 ∪ S2, then the coalition S1 ∪ S2 can always give its members the payoff that they would have received in the disjoint coalition S1 and S2. Definition 8. Canonical Game: A coalition game is canonical if it is superadditive and in characteristic form. The core is a widely used solution concept for canonical games as discussed below. C. Core We first define some terms related to the core [8], [9]. Definition 9. Group Rational: A payoff vector x ∈ RN is group-rational if(cid:80) Definition 10. Individually Rational: A payoff vector x ∈ RN is individually-rational if every player can obtain a payoff no less than acting alone, i.e., xn ≥ v({n}),∀n ∈ N . Definition 11. Imputation: A payoff vector that is both individually and group rational is an imputation. Definition 12. Grand Coalition: The coalition formed by all game players in N is the grand coalition. Based on the above definitions, we can now define the core of an NTU canonical coalition game. n∈N xn = v(N ). 4 Fig. 2. Our system model with multiple cloud service providers Definition 13. Core [8]: For any NTU canonical game (N ,V), the core is the set of imputations in which no coalition S ⊂ N has any incentive to reject the proposed payoff allocation and deviate from the grand coalition to form a coalition S instead. This can be mathematically expressed as CN T U = {x ∈ V(N )∀S, (cid:64)y ∈ V(S), such that yn > xn,∀n ∈ S}. (5) Remark 1. Any payoff allocation from the core is Pareto-optimal as evident from definition of the core. Furthermore, the grand coalition formed is stable, i.e., no two players will have an incentive to leave the grand coalition to form a smaller coalition. However, the core is not always guaranteed to exist. Even if the core exists, it may be very large as it a convex set. Therefore, finding a suitable allocation from the core is challenging. III. SYSTEM MODEL type k available at service provider n. The vector C = {(cid:80) Let N = {1, 2,··· , N} be the set of all the cloud service providers that act as players in our game. We assume that each player has a set of K = {1, 2,··· , K} different types of resources such as communication, computation and storage resources. The nth service provider represents its available resources as C n = {C (n) is the amount of resources of K } represents all available resources at different service providers. Each service provider n has a set of native applications Mn = {1, 2,··· , Mn} that ask for resources. The set of all applications that ask for resources from the set of service providers (coalition of service providers) is given by M = M1 ∪ M2 ··· ∪ MN , where Mi ∩ Mj = ∅, ∀i (cid:54)= j, i.e., each application originally asks only one service provider for resources. Every player n ∈ N has a request (requirement) matrix R(n). ,··· ,(cid:80) K }. C (n) ,··· , C (n) ,(cid:80) n∈N C (n) n∈N C (n) n∈N C (n) k 2 1 1 where the ith row corresponds to the ith application while columns represent different resources, i.e., r(n) is the amount of ij jth resource that application i ∈ Mn requests. Request matrix R is an augmentation of all the request matrices (from all the service providers) as given below. R(n) = = (cid:34) r(n) (cid:35) 1. . . r(n) Mn (cid:34) r1. (cid:35) . . rM R = = (cid:35) (cid:34) r(n) ··· ··· . . . . . . 11 . . . Mn1 ··· ··· r(n) r(n) r(n) 1K . . . MnK (cid:35) (cid:34) r11 ··· ··· r1K . . . . . . . . . rM 1 ··· ··· rMK . . . where rij is the amount of jth resource that application i ∈ M requests. The resource sharing and allocation framework, based on R and C, has to make an allocation decision X that optimizes the utilities un(X ) of all the service providers n ∈ N and satisfy user requests as well. The allocation decision X is a vector given by X = {X (1), X (2),··· , X (N )} that indicates how much of each resource k ∈ K is allocated to application i at service provider n ∈ N . The matrix X (n) is an augmentation of different matrices {X (n) }, where X (n) is the allocation decision at service provider n for the applications i originally belonging to service provider i∈Mj j ∈ N\n. X (n) is the allocation decision at service provider n for applications i belonging to service provider n, i.e., the original applications of the service provider n. Mathematically, ,··· , X (n) ,··· , X (n) i∈MN i∈Mn i∈Mn i∈M1 (cid:34) x(n) (cid:35) 1. . . x(n) M (cid:35) (cid:34) x(n) 11 ··· ··· x(n) 1K . . . . . . . . . . . . M 1 ··· ··· x(n) x(n) MK X (n) = = where x(n) ik is the amount of resource k ∈ K belonging to player n ∈ N that is allocated to application i ∈ M. (6) (7) (8) A. Optimization Problem 5 In this section, we first present the resource allocation problem for a single service provider (no resource sharing with other service providers). Then we present the MOO problem for resource sharing and allocation. For a single cloud n ∈ N (not part of a coalition), the allocation decision matrix X (n) SO, where SO stands for single objective, is given by: X (n) SO = The optimization problem is: (cid:34) x(n) (cid:34) x(n) (cid:35) 1. . . x(n) Mn = ··· ··· x(n) 1K . . . . . . 11 . . . Mn1 ··· ··· x(n) x(n) . . . MnK (cid:35) . (cid:88) SO ) + i∈Mn ik ≤ C (n) x(n) k (cid:88) x(n) ik r(n) k∈K ik ∀k ∈ K, un(X (n) max X (n) SO s.t. (cid:88) ∀n ∈ N , ∀i ∈ Mn, i ik ∀ i ∈ Mn, k ∈ K, ik ≤ r(n) x(n) ik ≥ 0 ∀ i ∈ M, k ∈ K. x(n) (cid:80) resources only to its own applications.(cid:80) The goal of this single objective optimization problem for every service provider is to maximize its utility by allocating available ik captures the satisfaction of resource requests by applications. Next, we present the resource allocation problem for different service providers that share their resources with each other. In the resource sharing case, each service provider aims to maximize sum of utilities obtained by a) allocating resources to its native ), ∀j ∈ N\n; applications, i.e., un(X (n) and c) satisfying requests of applications. This leads to: ); b) allocating resources to applications of other service providers, i.e., un k∈K x(n) ik /r(n) j (X (n) i∈Mn i∈Mn i∈Mj (9) (10a) (10b) (10c) (10d) (cid:80) (cid:88) (cid:88) i∈Mn k∈K ik l∈N x(l) r(n) ik (cid:19) maxX ) + ζn (cid:88) wnun(X (n) (cid:18) s.t. (cid:88) (cid:88) j∈N ik ≥ 0 ∀ i ∈ M, k ∈ K, n ∈ N . x(n) i∈Mn ik ≤ C (n) x(n) ik ≤ r(n) x(j) j∈N\n ik k i ∀k ∈ K,∀n ∈ N ,∀i ∈ M, ∀ i ∈ M, k ∈ K, n ∈ N , j (X (n) un i∈Mj ) + ∀n ∈ N , (11a) (11b) (11c) (11d) The first constraint in both optimization problems, (10) and (11), indicates that the allocated resources cannot be more than the capacity. The second constraint in (10c) and (11c) indicates, that the allocated resources should not be more than the required resources. The last constraint, (10d) and (11d), indicates that allocation decision cannot be negative. j (X (n) un i∈Mj ) is the utility that domain n receives for sharing its resources with domain j. wn is the weight associated with the utility un and ζn is the weight affiliated with all un j . In this paper, we assume that utility of the service providers is a non-decreasing monotone function. We also assume that a subset N1 ⊂ N of service providers has resource deficit, while a subset N2 ⊂ N has a resource surplus. B. Game Theoretic Solution As mentioned earlier, we model each cloud service provider as a player in our game to obtain the solution for the MOO problem in (11). Let N be the set of players that can play the resource sharing and allocation game. The value of coalition for the game players S ⊆ N is given by (12), where FS is the feasible set. (cid:18) (cid:88) n∈SX∈FS (cid:88) j∈S, j(cid:54)=n (cid:80) (cid:88) (cid:88) i∈Mn k∈K ik l∈S x(l) r(n) ik (cid:19) . v(S) = wnun(X ) + ζn j (X ) + un (12) Theorem 1. Resource allocation and sharing problem (with multiple objectives) for the aforementioned system model can be modeled as a canonical cooperative game with NTU. Proof. To prove this, we need to show that the characteristic function of resource sharing and allocation problem satisfies the following two conditions: • Characteristic form of payoff: As the utility function in a resource sharing and allocation problem only relies on the service providers that are part of the coalition, the game or payoff function is of characteristic form. = wnun(X ) + ζn (cid:80) n ∈ S2, X ∈ FS2 (cid:18) (cid:88) (cid:88) (cid:88) (cid:88) k∈K i∈Mt ζt n∈S1, n(cid:54)=t l∈(S2∪t) x(l) ik j (X ) + un (cid:18) j∈S2∪t, (cid:88) (cid:32) (cid:88) (cid:80) j(cid:54)=n − r(t) ik (cid:88) (cid:88) n ∈ S1, X ∈ FS1 l∈(S1∪t) x(l) ik i∈Mt k∈K r(t) ik n(X ) + ut (cid:18) (cid:88) wnun(X ) + ζn j∈{(S2∪t)\(S1∪t)}, n ∈ S2\S1, X ∈ FS2\FS1 = ζt (cid:88) n∈S2, n(cid:54)=t (cid:33) (cid:88) (cid:88) (cid:88) j(cid:54)=n n∈S1, n(cid:54)=t n∈S2, n(cid:54)=t (cid:80) l∈{(S2∪t)\(S1∪t)} x(l) ik r(n) ik (cid:88) k∈K (cid:19) + wtu(S2) t (X )+ (cid:88) j (X ) + un (cid:80) (cid:88) (cid:88) i∈Mn i∈Mt k∈K n(X ) − wtu(S1) ut t (X ) − ζt n(X ) + ut l∈{(S2∪t)\(S1∪t)} x(l) ik (cid:80) (cid:19) l∈(S2∪t) x(l) ik (cid:88) (cid:88) i∈Mn k∈K r(n) ik (cid:88) j∈S1∪t, j(cid:54)=n (cid:88) n∈S2, n(cid:54)=t n(X )+ ut (cid:19) + wtu(S2) t (X ) + ζt (cid:80) (cid:88) (cid:88) i∈Mn k∈K r(n) ik wnun(X ) + ζn j (X ) + un l∈(S1∪t) x(l) ik + wtu(S1) t 6 (X )+ = v(S2) − v(S1) + wtu(S2) t (X ) + ζt ≥ v(S2) − v(S1) n(X ) − wtu(S1) ut t (X ) − ζt r(t) ik (cid:88) n∈S1, n(cid:54)=t (cid:88) (cid:88) i∈Mt k∈K n(X ) + ut (cid:80) l∈{(S2∪t)\(S1∪t)} x(l) ik r(t) ik (14) • Superadditivity: For any S1, S2 ⊆ N where S1 ∩ S2 = ∅, S1, S2 ⊂ (S1 ∪ S2). The proof follows from definition of monotone utilities. Definition 14 (Convex Games). A coalition game is said to be convex if and only if for every player n ∈ N , the marginal contribution of the player is non-decreasing with respect to (W.R.T.) set inclusion. Mathematically, for S1 ⊆ S2 ⊆ N\{n} v(S1 ∪ {n}) − v(S1) ≤ v(S2 ∪ {n}) − v(S2). (13) Theorem 2. Our canonical game is convex. Proof. Let us consider two coalition S1 and S2, where S1 ⊆ S2 ⊆ N\{t}, t ∈ N , and FS1 and FS2 are the feasible sets for S1 and S2 respectively that it achieves by allocating resources to its own applications. We calculate v(S2 ∪{t})− v(S1 ∪{t}) in Equation (14), where u(S1) l(X ) ≥ are the utilities of player t in S1 and S2 respectively. l(X )−wtu(S1) The last inequality in (14) follows from the fact that wtu(S2) 0 (monotonicity of the utility). Remark 2. The core of any convex game (N ,V) is non-empty and large [10]. Remark 3. Our canonical cooperative game (N ,V) with NTU can be used to obtain the Pareto-optimal solutions for the multi-objective optimization problem given in (11). n∈S1,n(cid:54)=t ut n∈S2,n(cid:54)=t ut (X )+ζt (X )−ζt and u(S2) (cid:80) (cid:80) t t t t While the existence of core, i.e., core being non-empty, guarantees the grand coalition is stable, finding a suitable allocation from the core is challenging particularly when the core is large (our game). Algorithm 1 provides an allocation from the core2, by solving N + 1 optimization problems. Theorem 3. The allocation decision obtained using Algorithm 1 lies in the core. Proof. To prove the theorem, we need to show that the allocation decision obtained using Algorithm 1: a) is individually rational; b) is group rational; and c) no players have the incentive to leave the grand coalition and form another sub-coalition S ⊂ N . Individual Rationality: For each player n ∈ N , the solution obtained using (15) is individual rational due to the constraint Group Rationality: The value of the grand coalition v{N} as per Equation (15a) is the sum of utilities of all players that in (15b). Hence the solution obtained as a result of Algorithm 1 is individually rational. they achieve from the pay-off vector V(N ) ⊆ RN. Hence the allocation obtained from Algorithm 1 is group rational. 2Provided that the aforementioned assumptions hold. Algorithm 1 Game-theoretic Pareto optimal allocation 7 Input: R, C, and vector of utility functions of all players u Output: X , u(X ) Step 1: u(X ) ←0, X ←0 Step 2: for n ∈ N do v({n}) ←Optimal objective function value in (10) Step 3: X ←Optimal allocation decision from (15) u(X ) ←Payoff vector from (15) TABLE I SIMULATION NETWORK SETTINGS. Setting 1 2 3 4 Parameters N = 3, Mn = 3, ∀n ∈ N N = 3, Mn = 20, ∀n ∈ N N = 6, Mn = 6, ∀n ∈ N N = 6, Mn = 20, ∀n ∈ N Furthermore, due to super-additivity of the game and monotonic non-decreasing nature of the utilities, no subgroup of players have an incentive to form a smaller coalition. Hence Algorithm 1 provides a solution from the core. In algorithm 1, we first solve the single objective optimization problem (10) for all players n ∈ N . We then solve the problem in (15) that provides the allocation from the core. v({n}) in (15b) is the payoff a player n receives when working alone. (cid:18) (cid:88) (cid:18) n∈N maxX wnun(X (n) i∈Mn ) + ζn j (X (n) un i∈Mj ) + j∈N\n s.t. constraints in (11b)-(11d), (cid:88) (cid:88) j∈N\n (cid:19) , (cid:80) (cid:88) (cid:88) i∈Mn (cid:88) (cid:80) k∈K (cid:88) ik l∈N x(l) r(n) ik i∈Mn k∈K ik l∈N x(l) r(n) ik (cid:19) (15a) (15b) wnun(X (n) i∈Mn ) + ζn j (X (n) un i∈Mj ) + ≥ v({n}),∀n ∈ N . In the next section, we evaluate the performance of our resource sharing and allocation framework. IV. EXPERIMENTAL RESULTS We evaluate the performance of proposed resource sharing and allocation framework for a number of settings as shown in Table I. Each player has three different types of resources (K = 3), i.e., storage, communication and computation. Without loss of generality, the model can be extended to include other type of resources/parameters. We used linear and sigmoidal utilities (see (16)) for all the players. However, the results hold for any monotone non-decreasing utility. (cid:88) (cid:18) K(cid:88) i∈Mn k=1 un(X ) = (cid:19) 1 1 + e−µ(x(n) ik −r(n) ik ) ∀n ∈ N . (16) µ is chosen to be either 0.01 or 0.1 to capture the requirements of different applications. The request matrices R(n),∀n ∈ N and the capacity vectors C (n),∀n ∈ N are randomly generated for each setting within a pre-specified range3. To show the advantage of resource sharing, we allocate larger capacities to certain players that, for improving their utilities, share the available resources with other domains and assist other players in meeting demand. Since all the service providers do not differentiate between their native and applications of other service providers, we set wn and ζn, ∀n ∈ N to 1. The simulations were run in Matlab R2016b on a Core-i7 processor with 16 GB RAM. To solve the optimization problems, we used the OPTI-toolbox. Below, we provide detailed experimental results. A. Verification of game-theoretic properties In Table II, we present results for a 3-player 20-application game that verify different game theoretic properties such as individually rationality, group rationality, super additivity and show that the obtained allocation is from the core. The pay-off all players receive in the grand coalition, i.e., {1, 2, 3} is at least as good as players 1, 2 and 3 working alone. This shows that the solution obtained using Algorithm 1 for the grand coalition is individually rational. Similarly, the value of coalition 3The larger the number of applications in our simulation settings, the larger is range for random number generation. 8 PLAYER PAYOFF IN DIFFERENT COALITIONS FOR A 3 PLAYER - 20 APPLICATION GAME WITH µ = 0.01 TABLE II Coalition {1} {2} {3} {1, 2} {1, 3} {2, 3} {1, 2, 3} Player 1 584.40 0 0 584.40 584.40 0 584.40 Player 2 0.00 90 0 227.15 0 118.91 205.11 Player 3 0.00 0 90 0 229.84 119.50 202.30 Value of coalition 584.40 90 90 811.55 814.24 238.41 991.81 Fig. 3. Player utilities working alone and in the grand coalition (GC) for different simulation settings with µ = 0.01 and µ = 0.1. is the sum of pay-off all players receive is the value of coalition, hence our solution is group rational. Furthermore, with an increase in the coalition size, value of coalition also increases. Hence, the grand coalition has the largest value, which shows the superadditive nature of the game. Also, no set of players has any incentive to divert from the grand coalition and form a smaller coalition. Hence, the grand coalition is stable and the allocation we obtain using Algorithm 1 is from the core. Similar results were seen for other settings given in Table I. However, we do not include them here due to space constraints. B. Efficacy of the resource sharing framework 1) Player utilities: Figure 3 shows the utility of game players in different simulation settings. It is evident from the figure that the utility of players in grand coalition (GC) in all simulation settings either remains as good as working alone (very small number of cases) or drastically improves with our game theoretic resource sharing framework. This shows that cloud service providers can greatly benefit by employing a resource sharing scheme. The average improvement in the utility of players seen in our simulations is almost 207%. However, it is worth mentioning that the improvement greatly depends on parameters such as resource surplus, deficit, and the utilities used. 2) User satisfaction: Satisfying user requests is important for cloud service providers as it is mandated by the SLA. Furthermore, it helps in retaining their users (customers) resulting in long-term profits. Figure 4 shows the average user request satisfaction in different player-application settings. For a 3− player game with 3 and 20 applications respectively, player 1 has a resource deficit and is not able to satisfy all of its user's request when operating alone. Players 2 and 3 have a resource surplus so they can satisfy their users even working alone. When all the three players work in a coalition, the request satisfaction for all three players is 100%. Similarly, for a 6−player game with 6 and 20 applications respectively, we see that players who could not satisfy their applications alone are able to satisfy all their user requests by joining the coalition and "renting" resources from the players that have a resource surplus. It is worth mentioning that resource sharing can drastically improve user satisfaction. However, the extent of improvement depends on the nature of the utility function, capacity and resource requests. 3) Resource utilization: Figure 5 shows the resource utilization at different players in various simulation settings. For a 3−player game with 3 and 20 applications respectively, we see that the resource utilization for player 2 and 3 (players with 3 Applications, =0.01AloneGC050100150UtilityPlayer 1Player 2Player 33 Applications, =0.1AloneGC050100150UtilityPlayer 1Player 2Player 320 Applications, =0.01AloneGC0200400600UtilityPlayer 1Player 2Player 320 Applications, =0.1AloneGC0200400600UtilityPlayer 1Player 2Player 36 Applications, =0.01AloneGC050100150UtilityPlayer 1Player 2Player 3Player 4Player 5Player 66 Applications, =0.1AloneGC050100Utility20 Applications, =0.01AloneGC0100200300400Utility20 Applications, =0.1AloneGC0100200300400Utility9 Fig. 4. Improvement in user request satisfaction for various player-application settings and µ = 0.01. Fig. 5. Resource utilization for various player-application settings and µ = 0.01. resource surplus) increases in the grand coalition when compared with working alone. This is because these players rent out their resources to player 1 that may need them resulting in an increased utility. Player 1 with resource deficit had 100% resource utilization rate working alone. However, resource utilization for player 1 in grand coalition either may remain the same or reduce, since its users' requests may be satisfied by other players. Similar results are seen for a 6−player game with 6 and 20 applications as evident from the figure. It is evident from our results that the use of resource sharing and allocation is beneficial for cloud service providers. By allowing different cloud service providers to cooperate, the utility of service providers increases and resources are utilized in an optimal manner. All the service providers in the game have the incentive to work together and use their resource capacities in the best possible way. V. RELATED WORK There have been a number of solutions proposed in literature related to resource availability and resource allocation in cloud computing. Alicherry et al. [11] presented a 2-approximation algorithm for network resource allocation in a distributed cloud environment. Ergu et al. [12] proposed a model for resource allocation in cloud computing environment. Tsai et al. [13] proposed an improved differential evolution algorithm for resource allocation in cloud that combines differential evolution algorithm with Taguchi method. Pawar et al. [14] proposed a dynamic resource allocation that considers different SLA parameters and premptable task execution. Wei et al. [15] proposed a non-cooperative game theory based approach for resource allocation in the cloud. The objective of the algorithm is to maximize fairness among different users. Duan et al. [16] modeled scheduling in hybrid clouds as a sequential cooperative game. The authors proposed a storage and communication aware algorithm that jointly optimizes the execution time and economic cost of scheduling Bag-of-Tasks work flows. Shi et al. [17] proposed a Shapley value based mechanism for on-demand bandwidth allocation between data centers. They focus on network bandwidth and do not take computing and storage resources into account. Shi et al. [1] also proposed an online auction mechanism for 3 Player - 3 ApplicationAloneGrand Coalition020%40%60%80%100%User SatisfactionPlayer 1Player 2Player 33 Player - 20 ApplicationAloneGrand Coalition020%40%60%80%100%User Satisfaction Player 1Player 2Player 36 Player - 6 ApplicationAloneGrand Coalition020%40%60%80%100%User SatisfactionPlayer 1Player 2Player 3Player 4Player 5Player 66 Player - 20 ApplicationAloneGrand Coalition020%40%60%80%100%User SatisfactionAloneGrand Coalition020%40%60%80%100%Resource Utilization3 Player - 3 ApplicationPlayer 1Player 2Player 3AloneGrand Coalition020%40%60%80%100%Resource Utilization 3 Player - 20 ApplicationPlayer 1Player 2Player 3AloneGrand Coalition020%40%60%80%100%Resource Utilization6 Player - 6 ApplicationPlayer 1Player 2Player 3Player 4Player 5Player 6AloneGrand Coalition020%40%60%80%100%Resource Utilization6 Player - 20 Application10 dynamically providing virtual clusters in geo-distributed clouds. Zafari et al. [18] modeled resource sharing among mobile edge clouds as a canonical game with transferable utility (TU). However, our work in this paper differed from [1], [11], [15] -- [17]. We considered the multi-objective nature of the resource sharing problem and allowed different cloud service providers to share resources and improve their utilities, while satisfying the requests of different users. Our work is closest to [18]. However, authors in [18] considered the case where each service provider first allocates resources to its own applications and shares the remaining resources with the applications of other service providers. They only considered the utility of the service provider and did not consider the user satisfaction. Our framework guarantees Pareto optimality like [18], however, it is more generic than [18] as TU games can be modeled as a special case of NTU [9]. VI. CONCLUSIONS In this paper, we proposed a cooperative game theory based framework for resource sharing and allocation among cloud service providers. We showed that for a monotonic non-decreasing utility, resource sharing among multiple cloud service providers can be modeled as a canonical game with non-transferable utility. We prove that the game is convex, hence the core of the game is non-empty. We proposed an O(N ) algorithm that provides allocation decision from the core. Simulation results showed that our proposed framework improves utility of cloud service providers. Furthermore, request satisfaction of users also improved. REFERENCES [1] W. Shi, C. Wu, and Z. Li, "An Online Auction Mechanism for Dynamic Virtual Cluster Provisioning in Geo-Distributed Clouds," IEEE Transactions on Parallel and Distributed Systems, vol. 28, no. 3, pp. 677 -- 688, 2017. [2] Z. Xiao, W. Song, Q. Chen et al., "Dynamic resource allocation using virtual machines for cloud computing environment." IEEE Trans. Parallel Distrib. Syst., vol. 24, no. 6, pp. 1107 -- 1117, 2013. [3] W.-Y. Lin, G.-Y. Lin, and H.-Y. Wei, "Dynamic Auction Mechanism for Cloud Resource Allocation," in Proceedings of the 2010 10th IEEE/ACM International Conference on Cluster, Cloud and Grid Computing. IEEE Computer Society, 2010, pp. 591 -- 592. [4] A. Beloglazov, J. Abawajy, and R. Buyya, "Energy-Aware Resource Allocation Heuristics for Efficient Management of Data Centers for Cloud Computing," Future generation computer systems, vol. 28, no. 5, pp. 755 -- 768, 2012. [5] V. Mayer-Schonberger and K. Cukier, "Big Data: A Revolution That Will Transform How We Live, Work, and Think," 2013. [6] L. Atzori, A. Iera, and G. Morabito, "The Internet of Things: A Survey," Computer networks, vol. 54, no. 15, pp. 2787 -- 2805, 2010. [7] J.-H. Cho, Y. Wang, R. Chen, K. S. Chan, and A. Swami, "A Survey on Modeling and Optimizing Multi-Objective Systems," IEEE Communications Surveys & Tutorials, vol. 19, no. 3, pp. 1867 -- 1901, 2017. [8] Z. Han, Game Theory in Wireless and Communication Networks: Theory, Models, and Applications. Cambridge University Press, 2012. [9] R. B. Myerson, Game theory. Harvard university press, 2013. [10] W. W. Sharkey, "Cooperative Games with Large Cores," International Journal of Game Theory, vol. 11, no. 3-4, pp. 175 -- 182, 1982. [11] M. Alicherry and T. Lakshman, "Network Aware Resource Allocation in Distributed Clouds," in Infocom, 2012 proceedings IEEE. IEEE, 2012, pp. [12] D. Ergu, G. Kou, Y. Peng, Y. Shi, and Y. Shi, "The Analytic Hierarchy Process: Task Scheduling and Resource Allocation in Cloud Computing Environment," The Journal of Supercomputing, vol. 64, no. 3, pp. 835 -- 848, 2013. [13] J.-T. Tsai, J.-C. Fang, and J.-H. Chou, "Optimized Task Scheduling and Resource Allocation on Cloud Computing Environment using Improved Differential Evolution Algorithm," Computers & Operations Research, vol. 40, no. 12, pp. 3045 -- 3055, 2013. [14] C. S. Pawar and R. B. Wagh, "Priority Based Dynamic Resource Allocation in Cloud Computing with Modified Waiting Queue," in Intelligent Systems and Signal Processing (ISSP), 2013 International Conference on. IEEE, 2013, pp. 311 -- 316. [15] G. Wei, A. V. Vasilakos, Y. Zheng, and N. Xiong, "A Game-Theoretic Method of Fair Resource Allocation for Cloud Computing Services," The journal 963 -- 971. [16] R. Duan, R. Prodan, and X. Li, "Multi-Objective Game Theoretic Scheduling of Bag-of-Tasks Workflows on Hybrid Clouds," IEEE Transactions on [17] W. Shi, C. Wu, and Z. Li, "A Shapley-Value Mechanism for Bandwidth on Demand between Datacenters," IEEE Transactions on Cloud Computing, of supercomputing, vol. 54, no. 2, pp. 252 -- 269, 2010. Cloud Computing, vol. 2, no. 1, pp. 29 -- 42, 2014. vol. 6, no. 1, pp. 19 -- 32, 2018. [18] F. Zafari, J. Li, K. K. Leung, D. Towsley, and A. Swami, "A Game-Theoretic Approach to Multi-Objective Resource Sharing and Allocation in Mobile Edge," in Proceedings of the 2018 on Technologies for the Wireless Edge Workshop. ACM, 2018, pp. 9 -- 13.
1711.08615	1	1711	2017-11-23T08:36:17	Controlling Elections through Social Influence	[ "cs.MA", "cs.SI" ]	Election control considers the problem of an adversary who attempts to tamper with a voting process, in order to either ensure that their favored candidate wins (constructive control) or another candidate loses (destructive control). As online social networks have become significant sources of information for potential voters, a new tool in an attacker's arsenal is to effect control by harnessing social influence, for example, by spreading fake news and other forms of misinformation through online social media. We consider the computational problem of election control via social influence, studying the conditions under which finding good adversarial strategies is computationally feasible. We consider two objectives for the adversary in both the constructive and destructive control settings: probability and margin of victory (POV and MOV, respectively). We present several strong negative results, showing, for example, that the problem of maximizing POV is inapproximable for any constant factor. On the other hand, we present approximation algorithms which provide somewhat weaker approximation guarantees, such as bicriteria approximations for the POV objective and constant-factor approximations for MOV. Finally, we present mixed integer programming formulations for these problems. Experimental results show that our approximation algorithms often find near-optimal control strategies, indicating that election control through social influence is a salient threat to election integrity.	cs.MA	cs	Controlling Elections through Social Influence Department of Computer Science and Center for Artificial Intelligence in Society Bryan Wilder University of Southern California [email protected] Yevgeniy Vorobeychik Department of Electrical Engineering and Computer Science Vanderbilt University [email protected] Abstract Election control considers the problem of an adversary who attempts to tamper with a voting process, in order to either ensure that their favored candidate wins (constructive control) or another candidate loses (destructive control). As online social networks have become significant sources of information for potential voters, a new tool in an attacker's arsenal is to effect control by harnessing social influence, for example, by spreading fake news and other forms of misinformation through online social media. We consider the computational problem of election control via social influence, studying the conditions under which finding good adversarial strategies is computationally feasible. We consider two objectives for the adversary in both the constructive and destructive control settings: probability and margin of victory (POV and MOV, respec- tively). We present several strong negative results, showing, for example, that the problem of maximizing POV is inapproximable for any constant factor. On the other hand, we present approximation algorithms which provide somewhat weaker approximation guarantees, such as bicriteria approximations for the POV objective and constant- factor approximations for MOV. Finally, we present mixed integer programming formulations for these problems. Experimental results show that our approximation algorithms often find near-optimal control strategies, indicating that election control through social influence is a salient threat to election integrity. 1 Introduction The integrity of elections is crucial to the functioning of democratic institutions. As a result, a large body of work has focused on the robustness of elections to various forms of control, where a malicious party attempts to manipulate election results, for example, by bribing voters and adding or removing votes. While it is important to understand the vulnerability of elections to such control, there are many countries where blatant tampering is (fortunately) uncommon. For instance, outright voter fraud is very rare in the United States federal and state elections [1, 12]. However, more subtle forms of election control may attempt to subvert legitimate information channels towards malicious means. For example, political advertising and news (in the editorial form) are common legitimate means for convincing prospective voters. Such communication, when sufficiently transparent, is often critical to the effective functioning of democracy, and can exert considerable influence on voting behavior [23, 15, 5]. Malicious control over information promulgated through these channels can thus have considerable impact, but is also difficult to achieve due to the relative transparency of traditional media sources. The increasing importance of social media, such as Facebook and Twitter, for propagating information, including about political issues [11, 42, 27], is a game changer. Both the decentralized nature of information sources on social media, and their associated lack of transparency, present malicious parties with an unprecedented opportunity to influence a democratic political process. Recent evidence of deliberate election tampering in the 2016 US presidential election through fake news-deliberately falsified news content-suggests that this issue is a major concern for election integerity for years to come [2]. For example, it is estimated that the typical American adult saw at least one fake news story during the election cycle [2], and such stories have been shown to impact voters' judgment [39]. 1 Motivated by these concerns, we initiate the first algorithmic study of the problem of election control through social influence. In our setting, there is a social network of voters who elect a single winner by plurality vote. An outside party may select a subset of nodes as seed nodes for a news story or advertisement. Each of these seed nodes shares the story with their friends. Each friend has some probability of being influenced in their voting preferences, as well as sharing the story further. The question is whether, given a limited budget, the attacker can influence enough voters to ensure that a target candidate wins or loses the election. This problem is closely related to influence maximization, which has been studied primarily in the context of viral advertising. There, the objective is simply to maximize the expected number of people who receive a message. While influence maximization admits a simple (1 − 1/e)-approximation algorithm, election control through social influence presents a number of new algorithmic challenges. We study both constructive and destructive control for two different objectives: 1) maximizing the expected margin by which a target candidate wins (loses) the election (margin of victory, or MOV), and 2) maximizing the probability that a target candidate wins (loses) election (probability of victory, or POV). Summary of main results: We provide a mix of negative (hardness and inapproximability) and positive (algorithmic) results for the problem of election control through social influence. Our main contributions are the following: • We show that the MOV objective in the two candidate case is monotone submodular and hence admits a (1−1/e) greedy approximation algorithm. • We prove that the POV objective is hard to approximate to within any multiplicative factor for both constructive and destructive control, even in elections with only two candidates. • We provide a bicriteria approximation algorithm for the POV objective in the two-candidate case. In fact, our algorithm applies to the more general problem of maximizing the probability that a submodular function exceeds a given threshold value and may be of general interest. • In the multicandidate case, we provide algorithms which achieve similar guarantees as the two-candidate case up to the loss of a constant factor (independent of the number of candidates). Such guarantees hold for both constructive and destructive control, for both the MOV and POV objectives. • We give mixed integer linear programming (MILP) formulations for all of the above settings which can be used to find optimal solutions. • We experimentally compare our approximation algorithms to the optimal strategies produced by the MILP. Despite formal hardness results, our approximation algorithms often find near-optimal solutions, particularly for the MOV objective. This suggests that computational hardness may not always be a practical barrier to controlling elections via social influence. 2 Related work Our work is closely related to two research areas: election control and influence maximization. These bodies of work are separate: to our knowledge there is no previous work which considers election control using social influence. The computational study of election control was started by Bartholdi et al. [3], who considered constructive control. The destructive control setting was introduced by Hemaspaandra et al. [26]. A large body of work has studied election control under different settings and voting rules [14, 18] including bribing voters [20, 22, 4, 19, 44], adding or deleting voters [17, 32, 21, 31], and adding or deleting candidates [7, 21, 31]. Another topic is strategic behavior on the part of the voters themselves [34, 37, 38]. The main difference between our work and previous work on election control is that we introduce and analyze social influence as a novel mechanism for both constructive and destructive control. There is a large, parallel literature devoted to influence maximization in social networks. This line of work was in- troduced by Kempe et al. [28] who introduced influence maximization in the independent cascade model and proposed a greedy algorithm based on submodularity. Since then, a number of newer algorithms have been proposed, mostly attempting to scale up the greedy algorithm to very large graphs [9, 13, 41, 24]. Subsequent work has also introduced several related settings, e.g., continuous time dynamics [40, 16], bandit settings where dynamics are learned over time [10, 30, 6], or robust problems where influence probabilities are uncertain [25, 8, 33, 43]. None of this work considers using social influence to control an election, and our setting brings a range of new technical challenges. Almost all 2 work on influence maximization is founded on submodularity of the objective function. However, even though we use the same model of influence spread, objectives related to election control often violate submodularity, and we need to develop new algorithmic techniques. We mention here work by Krause et al. [29] on robust submodular optimization. For optimizing the POV objective, we use a similar form of surrogate objective. However, our objective is to maximize the probability of a desired outcome, not the worst-case value, so both our final algorithm and analysis are novel. ci votes. 3 Problem formulation We consider an election with candidates C = {c∗, c1, c2, ...c(cid:96)}. c∗ is a special target candidate, and the objective of the election control problem is to make c∗ either win the election (constructive control) or lose (destructive control). The voters are the nodes of a graph G = (V, E). Each voter v has an ordering πv over the candidates and casts a vote for πv(1), i.e., their first ranked candidate. We assume that voters do not behave strategically. The winner is decided via the plurality rule (the candidate with the most votes wins the election). If there is a tie, we say that the attacker = {v ∈ V : πv(j) = ci} be the set of voters who fails. This tie-breaking does not impact any of our results. Let V j rank candidate ci in place j. Initially, ci has V 1 ci Social influence: There is an attacker who wishes to change the results of the election by spreading messages which cause voters to change their ordering over candidates. In constructive control, the attacker can spread a message which causes any voter v who becomes influenced to promote c∗ by one place in πv (exchanging c∗ with the candidate previously ranked above them). If πv(c∗) = 1, the message has no effect on v, but v may still decide to share the message with their neighbors. In destructive control, a voter who is influenced demotes c∗ by one place in πv. Influence spreads via the independent cascade model (ICM), the most common model in the influence maximization literature. Each edge (u, v) ∈ E has a propagation probability pu,v. If u is influenced, it makes one attempt to influence each neighbor v. Each attempt succeeds independently with probability pu,v. The attacker may select a set of k seed nodes who are influenced at the start of the process. The diffusion then proceeds in discrete time steps until no new activations are made. We also introduce a useful alternate view of the ICM, the live-graph model. We can equivalently see the ICM as removing each edge (u, v) from the graph with probability 1 − pu,v. A node is influenced if it is reachable from any seed node via the edges that remain. Call any specific setting of present/absent edges a scenario y, with induced graph Gy. Let m = 2E be the total number of scenarios. Let f (S, y) denote the number of nodes which are reachable from any seed node in S on graph Gy. The expected number of nodes influenced under the ICM is just f (S) = Ey [f (S, y)]. Similarly, the probability that the number of influenced nodes exceeds any threshold value ∆ is Pry [f (S, y) ≥ ∆]. At times, we will want to specifically reason about the probability some subset of V is influenced. For any A ⊆ V , let f (S, y, A) denote the number of nodes in A reachable from S in scenario y. Analogously, f (S, A) = Ey[f (S, y, A)]. We remark that such functions can be evaluated up to arbitrary precision by averaging over random samples for y. For simplicity, we ignore such issues here since they are well understood [41, 9, 13]. Objectives: We now formally introduce our two objectives, starting with the simpler two-candidate case. In a two-candidate election, constructive and destructive control are clearly equivalent since maximizing the probability that c∗ loses is the same as maximizing the probability that the other candidate wins (and vice versa). Hence, we study only constructive control without loss of generality. In the margin of victory (MOV) objective, we want to maximize the expected number of votes by which c∗ wins the election. We define our objective as the change in the expected margin: (cid:2)f (S, y, V 2 c∗ )(cid:3) . MOV(S) = 2 E y The factor 2 is present since reaching a voter in V 2 c∗ both adds a vote for c∗ and removes a vote for the opponent. We study the expected change in the margin (not the margin itself) so that approximation ratios are well defined even when the margin is negative. c∗(cid:1) + 1 be the number of voters that c∗ needs to reach in order to win the election. The POV In the probability of victory (POV) objective, we want to maximize the probability that c∗ wins the election. Let (cid:0)V 1 −V 1 c1 ∆ = 1 2 objective is (cid:2)f (S, y, V 2 c∗ ) ≥ ∆(cid:3) POV(S) = Pr y 3 which is just the probability that at least ∆ of the voters who have c∗ in second place are reached. In the multicandidate case, constructive and destructive control are no longer equivalent. Further, the impact of messages is more complex than before. E.g., in constructive control not only does c∗ gain a vote, but another candidates loses a vote; we need to keep track of the number of votes lost by each other candidate. We start out by defining functions which give the change in the margin between c∗ and another candidate ci in a given scenario y when seed set S is chosen. Let χ(v, S, y) be 1 if node v is reachable from seed set S in the graph Gy. The change in margin (in constructive and destructive control, respectively) is given by gC(S, y, ci) = gD(S, y, ci) = χ(v, S, y) + 2 χ(v, S, y) + 2 χ(v, S, y) χ(v, S, y) (cid:88) (cid:88) v∈V 2 c∗\V 1 ci v∈V 1 c∗\V 2 ci (cid:88) (cid:88) v∈V 2 c∗∩V 1 ci v∈V 1 c∗∩V 2 ci which gives value 2 for every node that is flipped from ci to c∗ (or vice versa) and hence count double towards the margin, and 1 for other nodes. Based on this, we now give expressions for the change in margin given any fixed scenario y and seed set S. We start with constructive control. Note that before any intervention, the margin is just maxciV 1 c∗. Hence, the change in margin is −gC(S, y, cj) − V 1 −V 1 cj ci c∗. Afterwards, the margin is maxcjV 1 (cid:21) c∗ −V 1 mC(S, y) = V 1 (cid:20) max ci (cid:20) (cid:18) ci = min cj gC(S, y, cj) + max ci (cid:21) c∗ −gC(S, y, cj) − V 1 − max cj V 1 cj −(cid:12)(cid:12)(cid:12)V 1 cj (cid:12)(cid:12)(cid:12)(cid:19) V 1 ci . That is, the change in margin is driven by candidate with largest starting vote (V 1 ) and smallest loss in vote cj (gC(S, y, cj)). Now considering all scenarios y, the constructive control objectives are MOVC(S) = E y [mC(S, y)] POVC(S) = Pr y [mC(S, y) ≥ ∆C] . where ∆C = maxciV 1 ci −V 1 c∗+1 is the necessary change in margin for c∗ to win. For destructive control, we can similarly write the change in margin and corresponding objectives as (cid:0)gD(S, y, ci) + V 1 (cid:1) − max ci cj [mD(S, y)] y POVD(S) = Pr y mD(S, y) = max MOVD(S) = E ci cj V 1 [mD(S, y) ≥ ∆D] . where ∆D = V 1 c∗− maxciV 1 ci +1. 4 Elections with two candidates We start with elections with only two candidates. Recall that in this setting, constructive and destructive control are equivalent, so our results are stated only for constructive control (trying to ensure c∗ wins the election). In order to state our algorithmic results, we first introduce some background on submodular optimization and influence maximization. A set function f : V → R is submodular if for all A ⊆ B ⊆ V and all x (cid:54)∈ B, f (A∪{x})−f (A) ≥ f (B∪{x})−f (B). Intuitively, submodularity formalizes the property of diminishing returns. The function f (S) which gives the expected number of nodes reached by S under the independent cascade model is known to be monotone submodular. It is well known that whenever f is a monotone submodular function, the greedy algorithm gives a (1 − 1/e)-approximation to the problem maxS≤k f (S). It is natural to hope that submodularity would transfer to our election control objectives MOV and POV. Our first result is that submodularity does in fact hold for the MOV objective with two candidates. Previous results for influence maximization do not directly apply because the MOV objective only counts nodes who have c∗ in second place. Nevertheless, similar reasoning applies. 4 Algorithm 1 Algorithms for MOV objective 1: function GREEDY(h, k) 2: 3: 4: 5: S ← ∅ while S< K do v ← arg maxv∈V \S h(S ∪ {v}) − h(S) S ← S ∪ {v} 6: 7: function MOVCONSTRUCTIVE(k) 8: 9: 10: function MOVDESTRUCTIVE(k) 11: 12: return S h(S) :− Ey return GREEDY(h, k) h(S) :− Ey return GREEDY(h, k) (cid:2)f (S, y, V 2 c∗ )(cid:3) (cid:2)f (S, y, V 1 c∗ )(cid:3) Algorithm 2 Algorithms for POV objective 1: function ENUMERATETHRESHOLD(h, ∆, k) 2: 3: 4: h(cid:48)(S) :− Ey [min (β, h(S, y))] Sβ = GREEDY(h(cid:48), k) for β = ∆...n do return arg maxSβ ,β=∆....n Pry [h(Sβ, y) ≥ ∆] 5: 6: function POVCONSTRUCTIVE(k) 7: 8: 9: function POVDESTRUCTIVE(k) 10: //recall mC(S, y) = Ey return ENUMERATETHRESHOLD(mC, ∆C, k) return ENUMERATETHRESHOLD(mD, ∆D, k) (cid:2)f (S, y, V 2 c∗ )(cid:3) for 2-candidate case Theorem 1. In an election with two candidates, MOV is a monotone submodular function. Proof. We first fix a particular scenario y and show that the function f (·, y, V 2 c∗ ) is submodular. This suffices to show that Ey[f (·, y, V 2 c∗ )] is submodular since a nonnegative linear combination of submodular functions remains submodular. Monotonicity is clear since adding additional seeds to A can only make more nodes reachable. To show submodularity, we can write the marginal gain as (cid:88) v∈V 2 c∗ f (A ∪ {x}, y, V 2 c∗ ) − f (A, y, V 2 c∗ ) = (1 − χ(v, A, y))χ(v,{x}, y). Compare the above expression for a set A and any B ⊇ A. For any single node v, χ(v, B, y) = 1 whenever c∗ )− c∗ ) ≥ χ(v, A, y) = 1. Hence, the term in the above summation for each node v can only be smaller for f (B ∪{x}, y, V 2 c∗ ) − f (A, y, V 2 f (B, y, V 2 f (B ∪ {x}, y, V 2 c∗ ). We conclude that f (A ∪ {x}, y, V 2 c∗ ) and submodularity now follows by taking the expectation over y. c∗ ) than for f (A ∪ {x}, y, V 2 c∗ ) − f (A, y, V 2 c∗ ) − f (B, y, V 2 Hence, we can apply a greedy algorithm to MOV to obtain a (1 − 1/e)-approximation (MOVCONSTRUCTIVE in Algorithm 1). Moreover, this ratio is tight since MOV contains regular influence maximization as a special case when all nodes have c∗ in second place (V 2 c∗ = V ). It is NP-hard to approximate influence maximization with ratio better than 1 − 1/e [28]. Hence, two-candidate MOV is computationally intractable with respect to exact optimization but high quality and efficient approximation algorithms exist. We now turn to the POV objective, where we maximize the probability that c∗ wins the election. It is natural to think that submodularity may also carry over to this setting. However, this is not the case; we can provide a simple 2 − k + 1 have c∗ as their first counterexample where POV violates submodularity. Consider n isolated nodes, where n choice. Hence, to win the election it is necessary and sufficient to influence k nodes, which can be accomplished by any choice of k seeds from among those with c∗ as their second choice. Fix a seed set B containing k − 1 of these nodes and consider any A ⊂ B (that is, A is strictly smaller). We have POV(B) = POV(A) = 0. By adding a node 5 v ∈ V 2 c∗ \ B to B, we have POV(B ∪ {v}) − POV(B) = 1. However, since A< k − 1, POV(A ∪ {v}) = 0 and hence POV(A∪{v})− POV(A) = 0. This contradicts the definition of submodularity. Essentially, the POV objective displays a sharp threshold behavior, where additional seed nodes have no value until we are close to winning. This behavior in fact translates into the following strong hardness result: Theorem 2. It is NP-hard to compute an α-approximation to the problem maxS≤k POV(S) for any α > 0, even for two candidates and even when the instance is deterministic. Proof. We consider a deterministic objective: the ICM with all propagation probabilities either 0 or 1. Without loss of generality, we have only a single scenario and will drop the dependence on y in f. Suppose that we have an α-approximation for POV-maximization. We show how we can use this algorithm to optimally solve the influence maximization problem (i.e., maximizing f (·, V )), which is known to NP-hard since it includes maximum coverage as a special case. Let OP TIM be the optimal value of the influence maximization problem and OP TP OV (∆) be the optimal value for Problem 1 with the given threshold. Fix any ∆ > 0. If ∆ ≤ OP TIM , then there is a set S with POV(S) = 1 and hence OP TP OV (∆) = 1. Otherwise, there is no set with value ∆ and OP TP OV (∆) = 0. Since the objective to the POV problem is either 0 or 1, any α-approximation algorithm for it must return 1 whenever OP TP OV (∆) = 1. Now, we can just enumerate over ∆ = 1...n, where n is the number of nodes in the graph. At each value of ∆, we ask the α-approximation algorithm to solve the POV maximization problem with that value of ∆. We return the solution corresponding to the highest value of ∆ for which we can find a set with f (S, V ) ≥ ∆. By the above, this set is an optimal solution to the influence maximization problem. (cid:2)f (S, y, V 2 c∗ ) ≥ ∆(cid:3) We remark that since this hardness result has broader implications. Recall that POV(S) = Pry c∗ ) is a submodular function. Therefore, the inapproximability result in Theorem where by Theorem 1, f (S, y, V 2 2 shows that it is in general hard to approximate the problem of of maximizing the probability that a submodular function exceeds a given threshold value. This is a natural objective in other domains, e.g. for a risk-averse decision maker who wants to control the probability of a bad outcome. We pair this hardness result with a positive algorithmic result regarding bicriteria approximations. A bicriteria approximation algorithm gives up solution quality in more than one dimension, and is of interest when hardness results preclude the usual notion of approximation (as for our problem). We provide an algorithm which has a solution quality guarantee whenever the election is winnable by a "large margin". That is, there is a seed set with high probability of greatly exceeding ∆ votes. Our algorithm will attempt to maximize the probability of exceeding exactly ∆ votes, but has a guarantee relative to the optimal value for threshold 1 α ∆ for some α < 1. That is, it is only compared to the optimal value of a harder problem. Our algorithm is a greedy strategy based on the surrogate function h(S) = Ey (cid:2)min{β, f (S, y, V 2 c∗ )}(cid:3), where β is a chosen threshold value. The intuition is to replace the sharp discontinuity of the original POV objective by a surrogate which interpolates smoothly up to the threshold β. However, we do not give any "credit" for nodes reached beyond β since (unlike in the MOV case) we only care about crossing the threshold. It is easy to see that the minimum of a submodular function and a constant is itself submodular [29]. Hence, h is submodular and amenable to greedy optimization. POVCONSTRUCTIVE (Algorithm 2) iterates over a series of possible thresholds for h, optimizes each one greedily, and outputs the best of the resulting seed sets. Specifically, it tries every value of β from ∆...n. For each β, it finds a seed set Sβ by greedily optimizing Ey the Sβ which maximizes Pry objective (Line 5). c∗ )}(cid:3) (Algorithm 2, Lines 3-4). Then, it outputs c∗ ) ≥ ∆(cid:3), i.e., the one which has the best probability of exceeding the true (cid:2)min{β, f (S, y, V 2 (cid:2)f (Sβ, y, V 2 The reason that we need to enumerate over values for β, instead of just using the true threshold ∆, is that optimizing the surrogate h might result in a solution which has value below β in every scenario. However, we can show that if OP T (β) is high, there must be many scenarios where Sβ has value close to β (a notion formalized in our proof). Hence, if we try a sufficiently large β > ∆ and OP T (β) is still high, there must be many scenarios with value at least ∆. This is formalized in the following guarantee: Theorem 3. In an election with two candidates, POVCONSTRUCTIVE produces a solution S such that (cid:0)1 − 1 e (cid:1) OP TP OV α ∆(cid:1) − α (cid:0) 1 1 − α [f (S, y, V 2 c∗ ) ≥ ∆] ≥ max 0<α<1 Pr y 6 β . We will prove that Pry[f (Sβ, y, V 2 Proof. POVCONSTRUCTIVE enumerates over values of α by trying thresholds β = ∆...n. Fix a specific β and set . This suffices the prove the theorem α = ∆ because we output the best of the Sβ. A minor point is that the theorem takes the max over 0 < α < 1, while we try only the discrete points α = ∆ A = {y : f (Sβ, y, V 2 c∗ ) ≥ ∆] ≥ (1− 1 n . However, these are equivalent because f (S, y, V 2 We divide the set of scenarios into those where the Sβ has value at least αβ and those where it has less value. Let c∗ ) is always integral. ∆+1 , ..., ∆ e )OP TP OV ( 1 α ∆)−α ∆ , ∆ 1−α c∗ ) < αβ}. We have c∗ ) ≥ αβ} and B = {y : f (Sβ, y, V 2 c∗ )} + (cid:18) min{β, f (Sβ, y, V 2 (cid:88) (cid:19) (cid:88) (cid:18) (cid:19) 1 m 1 − 1 e min{β, f (Sβ, y, V 2 c∗ )} y∈B βm · OP TP OV (β) 1 m ≥ y∈A 1 − 1 e h(S) ≥ max S≤k where the first inequality follows from submodularity and the second follows since the solution attaining value OP TP OV (β) for the POV maximization problem is a feasible solution to the problem maxS≤k h(S) which has value at least βm · OP TP OV (β). We are interested in the minimum possible size of A given that the total value is lower bounded as above. By inspection, A is minimized when min{β, f (S, y)} = β for each y ∈ A and f (S, y) = αβ for each y (cid:54)∈ A. In this case, we have (cid:18) 1 − A (cid:19) m (cid:18) A m + α (cid:19) ≥ 1 − 1 e OP TP OV (β) and hence A m ≥ = 1 m (cid:0)1 − 1 e m(cid:88) (cid:1) OP TP OV i=1 1 − α 1[f (Sβ, y, V 2 c∗ ) ≥ αβ] = Pr y (cid:0) 1 α ∆(cid:1) − α (cid:2)f (Sβ, y, V 2 c∗ ) ≥ ∆(cid:3) which completes the proof. Theorem 3 in fact applies to the general problem of maximizing the probability that a submodular function exceeds a threshold value (complementing our hardness result in Theorem 2). As discussed above, this may be of interest independently of election control. 5 Multiple candidates We now consider election control with more than two candidates. There is a target candidate c∗ and other candidates c1...c(cid:96). Note that constructive and destructive control are distinct in this setting. We will give algorithms for both cases for both the MOV and POV objectives. The problem becomes significantly harder in the multicandidate setting because we must now reason simultane- ously about several objectives – whether each alternate candidate ci will accumulate more votes that c∗. We demon- strate that, up to the loss of a constant in the approximation ratio, it suffices to concentrate only on the number of votes gained or lost by c∗ (not the margin against each ci individually). This concept yields (bicriteria) approximation algorithms for each setting along the lines of the two-candidate case. We start out with the MOVC objective (constructive control for the margin of victory), since the idea is simpler to illustrate in this case. The basic intuition is that the change in margin between candidate ci and c∗ can be re-expressed as follows: gC(S, y, ci) = χ(v, S, y) (cid:88) = f(cid:0)S, y, V 2 c∗\V 1 v∈V 2 ci c∗ (cid:88) c∗∩V 1 v∈V 2 c∗ ∩ V 1 ci ci (cid:1) χ(v, S, y) + 2 (cid:1) + f(cid:0)S, y, V 2 7 Now, we can express the final margin in scenario y as mC(S, y) = f(cid:0)S, y, V 2 (cid:1) + min c∗ cj (cid:18) f (cid:16) S, y, V 2 c∗ ∩ V 1 cj (cid:17) V 1 ci + max ci −(cid:12)(cid:12)(cid:12)V 1 cj (cid:12)(cid:12)(cid:12)(cid:19) ci ci ci ci = cj n1− V 1 ci min cj (cid:19) (cid:18) max S≤k −V 1 −V 1 where the first term is common to all candidates and reflects the total number of voters who switch to c∗ and the min term selects the cj who has the most remaining votes. In general, this second term can be very difficult to approximate because it is the minimum of submodular functions, which is not in general submodular (or even approximable [29]). We might hope that there is some special structure to the election control problem, but this is not the case: Theorem 4. For any > 0, it is NP-hard to compute any Ω(cid:0) 1 (cid:1)-approximation to the problem (cid:1) + max 0 for all cj. Thus, the problem is just maxS≤k mincj f(cid:0)S, y, V 2 Kempe [25] show that it is NP-hard to compute Ω(cid:0) 1 (cid:1). We reduce from the robust influence (cid:1)-approximation to the problem maxS≤k mini=1...r fi(S). maximization (RIM) problem [25, 8]. In RIM, we are given a set of objectives f1...fr, each of which represent expected influence spread in an instance of the independent cascade model on a common underlying graph G. He and Proof. We will consider instances where all of the cj start with an equal number of votes and so maxciV 1 f(cid:0)S, y, V 2 c∗ ∩ V 1 c∗ ∩ V 1 Their proof holds when each fi is deterministic (assigns probability 0 or 1 to each edge), so we will assume that the instance is in this form. Let Gi be a graph in which each edge of G assigned probability 0 by fi has been removed. We create a graph G(cid:48) as follows. G(cid:48) contains each Gi as a disconnected subgraph. For every v ∈ G, we add a vertex v(cid:48) to G(cid:48). v(cid:48) has an outgoing edge to the copy of node v in each of the Gi subgraphs. Each such edge has propagation probability 1. There is a target candidate c∗ and r additional candidates c1...cr. Each of the v(cid:48) nodes that were added has c∗ as their first choice. Each node in subgraph Gi has ci as their first choice and c∗ as their second choice. generality, we will assume that this algorithm only selects nodes from the v(cid:48) (since if a seed set contains the copy of v in any subgraph, we can only obtain greater influence spread by exchanging it for v(cid:48)). Note that, for any such set of seed nodes, fi(S) = f (S, V 2 ) Thus, if S is an α-approximate solution for our problem, it is also an Suppose that we have an α-approximation algorithm for our problem for some α = Ω(cid:0) 1 (cid:1)-approximate solution to the RIM problem. Ω(cid:0) 1 n1− Therefore, we should not hope for any algorithm which can closely approximate the entirety of the objective; the min component is too difficult to handle. However, we can leverage the fact that the first term, f (S, y, V 2 c∗ ), is easy to optimize because it is just a submodular function. Hence, the objective is the sum of an easy term and a hard term. Importantly, we can show that optimizing just the easy term (which is what MOVCONSTRUCTIVE does) is sufficient to obtain a constant factor approximation. (cid:1). Without loss of c∗ ∩ V 1 (cid:1)- approximation to the MOVC problem with any number of (cid:0)1 − 1 n1− n1− cj ci Theorem 5. MOVCONSTRUCTIVE obtains a 1 3 candidates. e Proof. Let c(S, y) = arg minci f (V 2 mC. Let S∗ be an optimal seed set. Note that for all scenarios y, seed sets S, and candidates ci, f (S, y, V 2 f (S, y, V 2 be the candidate achieving the minimum in the definition of c∗ ) ≥ ). Hence, we have c∗ ∩ V 1 ) − V 1 ci ci c∗ ∩ V 1 ci (cid:104) f(cid:0)S∗, y, V 2 c∗ (cid:1)(cid:105) ≥ 1 (cid:104) f(cid:0)S∗, y, V 1 c∗ E y 3 E y (cid:16) (cid:1) + f (cid:16) (cid:17) (cid:17)(cid:105) S∗, y, V 1 c∗ ∩ V 1 c(S∗,y) + f S∗, y, V 1 c∗ ∩ V 1 c(S,y) Note that Ey[f (·, y, V 2 Let S be the resulting seed set. We have c∗ )] is a monotone submodular function, which MOVCONSTRUCTIVE greedily maximizes. 8 (cid:17)(cid:105) c∗ ∩ V 1 c(S,y) (cid:16) (cid:1) + f (cid:104) E y ≥ E y c∗ c∗ S, y, V 2 1 − 1 e (cid:16) f(cid:0)S, y, V 2 (cid:1) + f (cid:1)(cid:105) (cid:104) f(cid:0)S, y, V 2 (cid:19) (cid:18) (cid:104) f(cid:0)S∗, y, V 2 (cid:16) (cid:104) (cid:1) + min f(cid:0)S, y, V 2 (cid:104) (cid:16) f(cid:0)S, y, V 2 (cid:1) + f (cid:18) (cid:19) (cid:104) f(cid:0)S∗, y, V 1 E y c∗ f 1 − 1 e E y c∗ c∗ ≥ 1 3 S∗, y, V 2 c∗ ∩ V 1 which allows us to bound the margin of victory relative to S∗ as V 1 V 1 (cid:17) (cid:17)(cid:105) c∗ ∩ V 1 S, y, V 2 (cid:16) c∗ ∩ V 1 S, y, V 2 + max = E c(S,y) c∗ cj cj ci ci ci y y MOVC(S) = E ci + max c∗ ∩ V 1 (cid:1) + f ≥ 1 3 (cid:105) c(S,y)(cid:105) (cid:104)V 1 (cid:16) cj −V 1 − E y (cid:17) (cid:17) + f (cid:16) S∗, y, V 2 c∗ ∩ V 1 c(S,y) (cid:17)(cid:105) c(S∗,y) S∗, y, V 2 c(S∗,y) + f S∗, y, V 2 c∗ ∩ V 1 c(S,y) (cid:17)(cid:105) c(S,y)(cid:105) (cid:104)V 1 + max ci V 1 ci − E y and some additional algebra (deferred to the appendix) yields (cid:18) (cid:19)(cid:32) (cid:104) (cid:16) (cid:17) MOVC(S) ≥ 1 3 1 − 1 e MOVC(S∗) + E f S∗, y, V 2 y c(S,y) Now by definition of c(S∗, y), f (S∗, y, V 2 c(S∗,y)) − V 1 c(S∗,y)≤ f (S∗, y, V 2 c(S,y) and so This yields c(S,y)(cid:105)(cid:33) . + V 1 c(S∗,y)−V 1 c(S,y)) − V 1 c∗ ∩ V 1 (cid:17) c∗ ∩ V 1 (cid:16) (cid:17) − f (cid:104) (cid:16) (cid:16) f c∗ ∩ V 1 (cid:16) (cid:19)(cid:32) c(S,y)≥ f (cid:18) (cid:16) (cid:19)(cid:32) (cid:19) + f MOVC(S∗). c(S∗,y)−V 1 V 1 S∗, y, V 2 c∗ ∩ V 1 c(S∗,y) S∗, y, V 2 c∗ ∩ V 1 c(S,y) MOVC(S) ≥ 1 3 1 − 1 e MOVC(S∗) + E S∗, y, V 2 c∗ ∩ V 1 c(S,y) y (cid:17) − f (cid:16) (cid:104) S∗, y, V 2 c∗ ∩ V 1 c(S∗,y) S∗, y, V 2 c∗ ∩ V 1 MOVC(S∗) + E f y S∗, y, V 2 c∗ ∩ V 1 c(S∗,y) (cid:17) (cid:17)(cid:105)(cid:33) c(S,y) (cid:17)(cid:105)(cid:33) (cid:18) (cid:18) = 1 3 ≥ 1 3 1 − 1 e 1 − 1 e We also have a corresponding result for the destructive control case. Here, we can rewrite the change in mar- . MOVDESTRUCTIVE greedily ) + V 1 (cid:16) (cid:2)f (S, y, V 1 c∗ )(cid:3), which we show is a good surrogate for Ey [mD(S, y)]. (cid:1)-approximation to the multicandidate MOVD problem. (cid:0)1 − 1 − maxcjV 1 c∗ ∩ V 2 f (S, y, V 1 (cid:17) cj ci ci e gin as mD(S, y) = f (S, y, V 1 optimizes the submodular function Ey Theorem 6. MOVDESTRUCTIVE obtains a 1 2 c∗ ) + maxci The proof, which is similar to that of Theorem 5, can be found in the appendix. Now, we extend these ideas to obtain similar guarantees for the POVC and POVD objectives. Starting with POVC, recall that our objective is to maximize Pry [mC(S, y) ≥ ∆C], the probability that the change in margin exceeds the number of votes needed to win. We will prove a guarantee for the same algorithm POVCONSTRUCTIVE as from the two-candidate case. Recall that POVCONSTRUCTIVE optimizes the surrogate Ey enumerating over possible values of the threshold β. We have the following bicriteria approximation guarantee: (cid:2)min(cid:0)β, f(cid:0)S, y, V 2 (cid:1)(cid:1)(cid:3), c∗ 9 Theorem 7. Let OP T (∆) denote the optimal value of the problem maxS≤k Pry [mC(S, y) ≥ ∆]. Let S be the set produced by POVCONSTRUCTIVE. We have 3e−1 OP T(cid:0) 1 e−1 α ∆C 1 − α (cid:1) − α POVC(S) ≥ max 0<α<1 (cid:80) y min(cid:0)β, f (S, y, V 2 (cid:80) The proof can be found in the appendix. The main difference from the two candidate case is that we do not directly y min (β, mC(S, y)) since it may no longer be submodular. Instead, we greedily optimize the submodu- optimize 1 m lar surrogate function 1 y min (β, mC(S, y)). m From there, the same argument as in Theorem 3 extends to the multicandidate case. Analogous reasoning also yields a bicriteria guarantee for destructive control: Theorem 8. Let OP T (∆) denote the optimal value of the problem maxS≤k Pry [mD(S, y) ≥ ∆]. Let S be the set produced by POVDESTRUCTIVE. We have c∗ )(cid:1) and show that this surrogate approximates 1 (cid:80) m POVD(S) ≥ max 0<α<1 e−1 3e−1 OP T ( 1 1 − α α ∆) − α 6 Exact solutions via mixed-integer programming Thus far, we have only considered approximation algorithms for election control, motivated by computational hardness results for exact optimization. Now we give mixed-integer linear programming (MILP) formulations to find exact so- lutions. This serves two purposes. First, it allows us to study the effectiveness of election control for problem instances with are within the range of state of the art MILP solvers. Second, we can determine the empirical effectiveness of the approximation algorithms proposed in earlier sections. an objective over all sv ∈ {0, 1}V which satisfy(cid:80) There are two principal difficulties in obtaining MILP formulations. First, the objective is stochastic, ranging over an exponential number of scenarios. Second, even for a single fixed scenario, the number of nodes reached by a seed set is a nonlinear function. We first show how to linearize the problem when have only a single scenario y. Recall that y corresponds to a sampled graph Gy, where every edge e is removed independently with probability 1 − pe. Our MILP will have a binary variable sv ∈ {0, 1} for each node v ∈ V , where sv = 1 indicates that v is a seed node. We will maximize v∈V sv ≤ k (at most k nodes are seeded). The challenge is to v, v ∈ V be a binary variable indicating whether embed the nonlinear objective into the constraints of the MILP. Let xy v to be 1 only if v truly is reachable in Gy from some node with v is influenced in scenario y. We must constrain xy sv = 1. To accomplish this, let R(v, y) be the set of nodes which have a directed path to v in scenario y. R(v, y) does not depend on the decision variables s and can easily precomputed. Using this set, we constrain the x variables as: v ≤ (cid:88) xy u∈R(v,y) su ∀v ∈ V. Now we deal with stochasticity in the objective using sample average approximation. We first sample a set of v for each sampled scenario. Finally, we average over the scenarios Gyi, i = 1...r, maintaining a separate copy xyi variables in each scenario to obtain the final objective. 6.1 Formulations Using these components, we now give concrete formulations for each of the problem instances that we consider. We will assume that scenarios y1...yr have been sampled, where r is a tunable parameter trading off computational cost and sampling error. 10 6.1.1 Constructive control We create a variable gC(yi, cj) for each scenario yi and candidate cj which represents the change in the margin between cj and c∗ in scenario yi. Using these variables, we set a variable mC(yi) for each scenario yi which represents the overall change in margin. These variables are set using the constraints gC(yi, cj) ≤ (cid:88) (cid:88) xyi v + v∈V 2 c∗ xyi v v∈V 2 c∗∩V 1 −V 1 cj cj V 1 ci ∀i, j mC(yi) ≤ gC(yi, cj) + max ci Using these variables, we have the following MILP to maximize the MOV: r(cid:88) i=1 mC(yi) max s,x,gC ,mC 1 r sv ≤ k (cid:88) v∈V The next formulation maximizes the POV: r(cid:88) 1 r max s,x,gC ,mC ,u ui − M (1 − ui) + max i=1 V 1 ui ∈ {0, 1} ci ci sv ≤ k, (cid:88) v∈V −V 1 c∗+1 − mC(yi) ≤ 0 i = 1...r Here, ui is a binary variable representing whether c∗ wins the election in scenario yi, while M is a large number. 6.1.2 Destructive control Now, we use an analogous set of constraints to set variables gD(yi, cj) and mD(yi): gD(yi, cj) ≤ (cid:88) (cid:88) − M (1 − zj i ≥ 1 ∀i, zj xyi v + (cid:88) (cid:18) i ∈ {0, 1} ∀i, j zj v∈V 1 v∈V 1 c∗ i ) + mD(yi) − xyi v cj c∗∩V 2 gD(yi, cj) + V 1 cj (cid:19) ≤ 0 ∀i, j − max V 1 ck k j The second and third constraints use a new set of binary variables zj mD(yi) is at most the change in margin between cj and c∗. The constraint(cid:80) i = 1 indicates that in scenario yi, i ≥ 1 requires that mD(yi) must be bounded by one such value, and so can be at most the maximum margin. With these variables in place, the MOV and POV MILPs are analogous to those for constructive control. i , where zj j zj 7 Experiments We now present experimental results comparing the performance of our approximation algorithms to the solutions found via mixed integer programming. We show results on four datasets. First, netscience, a collaboration network of researchers in network science, with 1461 nodes [36]. Second, facebook, the subgraph centered on 10 Facebook 11 Table 1: Percent of MILP value obtained by approximation algorithm. netscience facebook polblogs irvine k = 25 50 100 25 50 100 25 50 100 25 50 100 Constructive C= 2 C= 5 C= 10 Destructive C= 2 C= 5 C= 10 99.5 82.8 80.9 99.9 73.8 75.9 99.3 90.1 89.1 99.6 73.2 87.2 100. 91.5 98.3 99.9 83.3 97.2 100 90.8 80.9 100 87.7 81.8 100 90.9 83.7 100 79.7 85.0 100 90.9 88.6 100 81.6 89.0 100 97.4 98.7 100 100 98.9 100 97.8 99.2 100 97.8 99.3 99.4 99.5 99.6 99.4 99.3 99.6 100 97.7 96.9 100 100 100 100 95.4 97.4 99.8 99.0 97.3 99.4 96.8 98.9 98.8 100 99.0 Figure 1: Probability of victory. Top: constructive. Bottom: destructive. Left: netscience. Right: polblogs. users, with 2888 nodes [35]. Third, polblogs, a network of links between political blogs, with 1224 nodes [36]. Fourth, irvine, a graph representing instant messages exchanged between students at U.C. Irvine, with 1889 nodes [35]. We select these datasets because they represent the kinds of social and communication networks on which political messages (such as fake news) spread. We also note that our approximation algorithms can easily be scaled to much larger networks since we can apply the same techniques developed in the influence maximization literature [41, 13]. However, our focus here is to characterize the performance of our algorithms in comparison to the optimal solution, so we select datasets which are feasible for mixed integer programming. For each network, we randomly generated 30 sets of voter preferences. We start out with the MOV objective. Table 1 shows the percentage of the MILP's value which is obtained by our approximation algorithms (MOVCONSTRUCTIVE and MOVDESTRUCTIVE respectively), averaging over the 30 instances on each network with propagation probability p = 0.1. We vary the number of seed nodes k and the number of candidates C. We see that the approximation algorithms perform well across all settings, obtaining expected change in margin at least 73% of that of the MILP. The approximation algorithms fare particularly well for 2-candidate elections, obtaining nearly 100% of the optimal value on all networks. The empirical approximation ratio degrades as the number of candidates grows, particularly when the budget k is small. We conclude that our approximation algorithms are highly effective for election control via the MOV in both constructive and destructive control, particularly in realistic settings with a moderate number of candidates. 12 406080100C0.00.20.40.60.81.0POVApprox.MILP406080100C0.00.20.40.60.81.0POVApprox.MILP6080100120D0.00.20.40.60.81.0POVApprox.MILP406080100120D0.00.20.40.60.81.0POVApprox.MILPWe now turn to the POV objectives. We show results for k = 50,C= 5, comparing our bicriteria approxima- tion algorithms POVCONSTRUCTIVE and POVDESTRUCTIVE to the corresponding MILP formulations. To keep the experiments timely, we ran each approximation algorithm for 150 random values of the threshold β instead of enu- merating over all (empirically, this resulted in very similar solution quality). Figure 1 shows the results on netscience and polblogs for constructive and destructive control. Results for facebook and irvine can be found in the appendix. The x axis shows the starting margin (∆C or ∆D) in each randomly generated instance while the y axis shows the probability of victory obtained. The instances fall into three groups. First, when the margin is small, both the approx- imation algorithms and the MILP have a high POV. Second, when the margin is large, both have a small POV. Third are intermediate points where the approximation algorithm and MILP strongly diverge. Averaged over all random instances, the approximation algorithm obtains 40-60% of the MILP's value (depending on the network). However, there are instances among the intermediate cases where, e.g., the approximation algorithm obtains a POV of 0.1%, but the MILP finds a solution with POV 99%. We conclude that election control for the POV objective can be very computationally difficult in narrowly winnable elections. This dovetails with our theoretical results, which show that the problem is inapproximable in general, but bicriteria guarantees are possible when the election is winnable by a large margin. 8 Conclusion Fake news and other targeted misinformation are an increasingly prevalent way of interfering with democratic elec- tions. We introduce and study the problem of election control through social influence, providing algorithms and hardness results for maximizing both the margin and probability of victory for an attacker in both constructive and destructive control. Our results indicate that social influence is a salient threat to election integrity, particularly in the MOV case where we provide high-quality approximation algorithms. Maximizing the probability of victory is manageable in easier instances, but difficult both theoretically and empirically in narrow races. References [1] John S Ahlquist, Kenneth R Mayer, and Simon Jackman. Alien abduction and voter impersonation in the 2012 us general election: Evidence from a survey list experiment. Election Law Journal, 13(4):460–475, 2014. [2] Hunt Allcott, Matthew Gentzkow, et al. Social media and fake news in the 2016 election. Journal of Economic Perspectives, 31(2):211–236, 2017. [3] John J Bartholdi, Craig A Tovey, and Michael A Trick. How hard is it to control an election? Mathematical and Computer Modelling, 16(8-9):27–40, 1992. [4] Dorothea Baumeister, G´abor Erd´elyi, Olivia J Erd´elyi, and Jorg Rothe. Complexity of manipulation and bribery in judgment aggregation for uniform premise-based quota rules. Mathematical Social Sciences, 76:19–30, 2015. [5] Ted Brader. Striking a responsive chord: How political ads motivate and persuade voters by appealing to emotions. American Journal of Political Science, 49(2):388–405, 2005. [6] Alexandra Carpentier and Michal Valko. Revealing graph bandits for maximizing local influence. In Artificial Intelli- gence and Statistics, pages 10–18, 2016. [7] Jiehua Chen, Piotr Faliszewski, Rolf Niedermeier, and Nimrod Talmon. Elections with few voters: Candidate control can be easy. In AAAI, volume 15, pages 2045–2051, 2015. [8] Wei Chen, Tian Lin, Zihan Tan, Mingfei Zhao, and Xuren Zhou. Robust influence maximization. arXiv preprint arXiv:1601.06551, 2016. [9] Wei Chen, Chi Wang, and Yajun Wang. Scalable influence maximization for prevalent viral marketing in large-scale social networks. In Proceedings of the 16th ACM SIGKDD international conference on Knowledge discovery and data mining, pages 1029–1038. ACM, 2010. [10] Wei Chen, Yajun Wang, and Yang Yuan. Combinatorial multi-armed bandit: General framework and applications. In International Conference on Machine Learning, pages 151–159, 2013. [11] Feng Chi and Nathan Yang. Twitter adoption in congress. Review of Network Economics, 10(1), 2011. 13 [12] Ray Christensen and Thomas J Schultz. Identifying election fraud using orphan and low propensity voters. American Politics Research, 42(2):311–337, 2014. [13] Edith Cohen, Daniel Delling, Thomas Pajor, and Renato F Werneck. Sketch-based influence maximization and com- putation: Scaling up with guarantees. In Proceedings of the 23rd ACM International Conference on Conference on Information and Knowledge Management, pages 629–638. ACM, 2014. [14] Vincent Conitzer, Tuomas Sandholm, and J´erome Lang. When are elections with few candidates hard to manipulate? Journal of the ACM (JACM), 54(3):14, 2007. [15] Stefano DellaVigna and Ethan Kaplan. The fox news effect: Media bias and voting. The Quarterly Journal of Eco- nomics, 122(3):1187–1234, 2007. [16] Nan Du, Le Song, Manuel Gomez Rodriguez, and Hongyuan Zha. Scalable influence estimation in continuous-time diffusion networks. In Advances in neural information processing systems, pages 3147–3155, 2013. [17] G´abor Erd´elyi, Edith Hemaspaandra, and Lane A Hemaspaandra. More natural models of electoral control by partition. In International Conference on Algorithmic DecisionTheory, pages 396–413. Springer, 2015. [18] G´abor Erd´elyi, Markus Nowak, and Jorg Rothe. Sincere-strategy preference-based approval voting fully resists con- structive control and broadly resists destructive control. Mathematical Logic Quarterly, 55(4):425–443, 2009. [19] G´abor Erd´elyi, Christian Reger, and Yongjie Yang. The complexity of bribery and control in group identification. In Proceedings of the 16th Conference on Autonomous Agents and MultiAgent Systems, pages 1142–1150. International Foundation for Autonomous Agents and Multiagent Systems, 2017. [20] Piotr Faliszewski, Edith Hemaspaandra, and Lane A Hemaspaandra. How hard is bribery in elections? Journal of Artificial Intelligence Research, 35:485–532, 2009. [21] Piotr Faliszewski, Edith Hemaspaandra, and Lane A Hemaspaandra. Multimode control attacks on elections. Journal of Artificial Intelligence Research, 40(1):305–351, 2011. [22] Piotr Faliszewski, Edith Hemaspaandra, Lane A Hemaspaandra, and Jorg Rothe. Llull and copeland voting computa- tionally resist bribery and constructive control. Journal of Artificial Intelligence Research, 35:275–341, 2009. [23] Alan S Gerber, Dean Karlan, and Daniel Bergan. Does the media matter? a field experiment measuring the effect of newspapers on voting behavior and political opinions. American Economic Journal: Applied Economics, 1(2):35–52, 2009. [24] Amit Goyal, Wei Lu, and Laks VS Lakshmanan. Celf++: optimizing the greedy algorithm for influence maximization in social networks. In Proceedings of the 20th international conference companion on World wide web, pages 47–48. ACM, 2011. [25] Xinran He and David Kempe. Robust influence maximization. In Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, San Francisco, CA, USA, August 13-17, 2016, pages 885–894, 2016. [26] Edith Hemaspaandra, Lane A Hemaspaandra, and Jorg Rothe. Anyone but him: The complexity of precluding an alternative. Artificial Intelligence, 171(5-6):255–285, 2007. [27] Jesse Holcomb, Jeffrey Gottfried, and Amy Mitchell. News use across social media platforms. Pew Research Journal- ism Project, 2013. [28] David Kempe, Jon Kleinberg, and ´Eva Tardos. Maximizing the spread of influence through a social network. In Proceedings of the ninth ACM SIGKDD international conference on Knowledge discovery and data mining, pages 137–146. ACM, 2003. [29] Andreas Krause, H Brendan McMahan, Carlos Guestrin, and Anupam Gupta. Robust submodular observation selec- tion. Journal of Machine Learning Research, 9(Dec):2761–2801, 2008. [30] Siyu Lei, Silviu Maniu, Luyi Mo, Reynold Cheng, and Pierre Senellart. Online influence maximization. In Proceedings of the 21th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, pages 645–654. ACM, 2015. [31] Hong Liu, Haodi Feng, Daming Zhu, and Junfeng Luan. Parameterized computational complexity of control problems in voting systems. Theoretical Computer Science, 410(27-29):2746–2753, 2009. 14 [32] Andrea Loreggia, Nina Narodytska, Francesca Rossi, K Brent Venable, and Toby Walsh. Controlling elections by replacing candidates or votes. In Proceedings of the 2015 International Conference on Autonomous Agents and Multi- agent Systems, pages 1737–1738. International Foundation for Autonomous Agents and Multiagent Systems, 2015. [33] Meghna Lowalekar, Pradeep Varakantham, and Akshat Kumar. Robust influence maximization. In Proceedings of the 2016 International Conference on Autonomous Agents & Multiagent Systems, pages 1395–1396. International Foundation for Autonomous Agents and Multiagent Systems, 2016. [34] Reshef Meir, Ariel D Procaccia, Jeffrey S Rosenschein, and Aviv Zohar. Complexity of strategic behavior in multi- winner elections. J. Artif. Intell. Res.(JAIR), 33:149–178, 2008. [35] KONECT network dataset collection. http://konect.uni-koblenz.de/networks/ego-facebook, 2017. [36] Mark Newman. Network data. http://www-personal.umich.edu/mejn/netdata/, 2017. [37] Svetlana Obraztsova and Edith Elkind. Optimal manipulation of voting rules. In Proceedings of the 11th International Conference on Autonomous Agents and Multiagent Systems-Volume 2, pages 619–626. International Foundation for Autonomous Agents and Multiagent Systems, 2012. [38] Svetlana Obraztsova, Yair Zick, and Edith Elkind. On manipulation in multi-winner elections based on scoring rules. In Proceedings of the 2013 international conference on Autonomous agents and multi-agent systems, pages 359–366. International Foundation for Autonomous Agents and Multiagent Systems, 2013. [39] Gordon Pennycook, Tyrone D Cannon, and David G Rand. Prior exposure increases perceived accuracy of fake news. 2017. [40] Manuel Gomez Rodriguez and Bernhard Scholkopf. Influence maximization in continuous time diffusion networks. arXiv preprint arXiv:1205.1682, 2012. [41] Youze Tang, Xiaokui Xiao, and Yanchen Shi. Influence maximization: Near-optimal time complexity meets practical efficiency. In Proceedings of the 2014 ACM SIGMOD international conference on Management of data, pages 75–86. ACM, 2014. [42] Sunil Wattal, David Schuff, Munir Mandviwalla, and Christine B Williams. Web 2.0 and politics: the 2008 us presi- dential election and an e-politics research agenda. MIS quarterly, pages 669–688, 2010. [43] Bryan Wilder, Amulya Yadav, Nicole Immorlica, Eric Rice, and Milind Tambe. Uncharted but not uninfluenced: Influence maximization with an uncertain network. In Proceedings of the 16th Conference on Autonomous Agents and MultiAgent Systems, pages 1305–1313. International Foundation for Autonomous Agents and Multiagent Systems, 2017. [44] Yongjie Yang, Yash Raj Shrestha, and Jiong Guo. How hard is bribery with distance restrictions? In ECAI, pages 363–371, 2016. A Proofs We now provide proofs that were deferred from the main text. We start out with with the full proofs for the MOVC and MOVD objectives: Theorem 5: MOVCONSTRUCTIVE obtains a 1 3 candidates. Proof. Let c(S, y) = arg minci f (V 2 mC. Let S∗ be an optimal seed set. Note that for all scenarios y, seed sets S, and candidates ci, f (S, y, V 2 f (S, y, V 2 (cid:1)-approximation to the MOVC problem with any number of be the candidate achieving the minimum in the definition of c∗ ) ≥ (cid:0)1 − 1 ). Hence, we have c∗ ∩ V 1 ci ) − V 1 ci e c∗ ∩ V 1 ci (cid:104) f(cid:0)S∗, y, V 2 c∗ (cid:1)(cid:105) ≥ 1 (cid:104) f(cid:0)S∗, y, V 1 c∗ E y (cid:16) (cid:1) + f (cid:17)(cid:105) E y (cid:16) 3 + f S∗, y, V 1 c∗ ∩ V 1 c(S,y) (cid:17) S∗, y, V 1 c∗ ∩ V 1 c(S∗,y) 15 c∗ )] is a monotone submodular function, which MOVCONSTRUCTIVE greedily maximizes. (cid:16) (cid:1) + f S∗, y, V 2 c∗ ∩ V 1 c(S∗,y) (cid:17) c∗ Let S be the resulting seed set. We have Note that Ey[f (·, y, V 2 (cid:104) f(cid:0)S, y, V 2 c∗ (cid:1) + f (cid:16) E y and so MOVC(S) = E S, y, V 2 c∗ ∩ V 1 c(S,y) S, y, V 2 y ci ci f cj c∗ + f c(S,y) c(S,y) + max ≥ 1 3 V 1 V 1 1 − 1 e S∗, y, V 2 (cid:1)(cid:105) (cid:104) (cid:17)(cid:105) ≥ E f(cid:0)S, y, V 2 (cid:18) (cid:19) (cid:104) f(cid:0)S∗, y, V 2 (cid:16) (cid:17)(cid:105) E y c∗ ∩ V 1 (cid:17) (cid:17)(cid:105) c∗ ∩ V 1 S, y, V 2 (cid:16) c∗ ∩ V 1 (cid:1) + f (cid:17)(cid:105) (cid:16) (cid:1) + f (cid:17) (cid:16) (cid:1) + f (cid:1) + min (cid:16) (cid:104) + max c∗ ∩ V 1 S∗, y, V 2 V 1 − E c∗ ∩ V 1 S∗, y, V 2 V 1 −V 1 c∗ ∩ V 1 c(S,y)+V 1 f (S∗, y, V 2 −V 1 V 1 S∗, y, V 2 −V 1 c(S,y)) + max c(S∗,y) c(S∗,y) + max + max + max V 1 (cid:16) c(S,y) c(S,y) c∗ c∗ cj ci ci ci ci ci ci ci ci y ci ci cj −V 1 − E (cid:105) (cid:104)V 1 c(S,y)(cid:105) (cid:17) (cid:104)V 1 c(S,y)(cid:105) (cid:17) c(S,y)(cid:105) (cid:17) y c(S∗,y)(cid:105) (cid:17) c(S∗,y)(cid:105) c(S∗,y) c(S∗,y)−V 1 c∗ ∩ V 1 ) − V 1 c(S,y)+V 1 (cid:17) cj cj + V 1 c(S,y) cj c∗ c∗ c∗ E y + f 1 − 1 e 1 − 1 e c∗ ∩ V 1 S∗, y, V 2 (cid:16) (cid:104) (cid:1) + min f(cid:0)S, y, V 2 (cid:16) (cid:104) (cid:1) + f f(cid:0)S, y, V 2 (cid:18) (cid:19) (cid:104) f(cid:0)S∗, y, V 1 (cid:16) (cid:19) (cid:18) (cid:104) f(cid:0)S∗, y, V 1 (cid:16) (cid:19) (cid:18) (cid:104) c∗ ∩ V 1 f(cid:0)S∗, y, V 1 (cid:16) (cid:17) E y (cid:18) (cid:19) (cid:104) c∗ ∩ V 1 f(cid:0)S∗, y, V 1 E y (cid:19)(cid:32) (cid:18) 1 − 1 e + f (S∗, y, V 2 1 − 1 e S∗, y, V 1 S∗, y, V 1 c∗ ∩ V 1 c(S,y) + f E y c∗ 1 − 1 e MOVC(S∗) + E f S∗, y, V 2 y y = E y ≥ 1 3 ≥ 1 3 = 1 3 + f = = 1 3 1 3 Now by definition of c(S∗, y), f (S∗, y, V 2 c(S∗,y)) − V 1 c(S∗,y)≤ f (S∗, y, V 2 c(S,y) and so V 1 c(S∗,y)−V 1 S∗, y, V 2 c∗ ∩ V 1 c(S∗,y) S∗, y, V 2 This yields c(S,y)(cid:105)(cid:33) . c(S∗,y)−V 1 c(S,y)) − V 1 c∗ ∩ V 1 (cid:17) c(S,y) c∗ ∩ V 1 (cid:17) + (cid:17)(cid:105)(cid:33) (cid:17)(cid:105)(cid:33) c∗ ∩ V 1 (cid:16) (cid:17) − f (cid:104) (cid:16) (cid:17) − f (cid:16) (cid:104) (cid:16) f y y MOVC(S) ≥ 1 3 MOVC(S∗) + E S∗, y, V 2 c∗ ∩ V 1 c(S,y) S∗, y, V 2 c∗ ∩ V 1 c(S∗,y) f S∗, y, V 2 c∗ ∩ V 1 c(S,y) MOVC(S∗) + E f S∗, y, V 2 c∗ ∩ V 1 c(S∗,y) c∗ ∩ V 1 (cid:16) (cid:19)(cid:32) c(S,y)≥ f (cid:18) 1 − 1 e (cid:16) (cid:19)(cid:32) (cid:19) (cid:18) (cid:18) = 1 3 ≥ 1 3 1 − 1 e 1 − 1 e MOVC(S∗). 16 (cid:0)1 − 1 e (cid:1)-approximation to the multicandidate MOVD problem. be the candidate achieving the maximum in the definition y 2 c∗ c∗ ci ci c∗ c∗ ≥ E y E y (cid:104) (cid:18) S, y, V 1 (cid:1) + f (cid:1)(cid:105) (cid:17) ) + V 1 MOVD(S) = E Proof. Now let c(S, y) = arg maxci f (V 1 of mD. Let S∗ be an optimal seed set. Similarly to before, we have MOVDESTRUCTIVE greedily maximizes Ey Theorem 6: MOVDESTRUCTIVE obtains a 1 2 c∗ ∩ V 2 (cid:1)(cid:105) ≥ 1 (cid:104) f(cid:0)S∗, y, V 1 f(cid:0)S∗, y, V 1 (cid:104) f(cid:0)S∗, y, V 1 (cid:104) (cid:16) f(cid:0)S, y, V 1 (cid:1) + f (cid:19) (cid:16) (cid:104) (cid:1)(cid:105) (cid:104) c∗ ∩ V 2 f(cid:0)S∗, y, V 1 (cid:19) (cid:18) (cid:16) (cid:104) + E (cid:1) + f f(cid:0)S∗, y, V 1 (cid:17) (cid:104) (cid:16) (cid:19) (cid:18) (cid:16) (cid:104) (cid:1) + f f(cid:0)S∗, y, V 1 (cid:16) (cid:17) (cid:19) (cid:18) (cid:104) (cid:16) (cid:1) + f f(cid:0)S∗, y, V 1 (cid:16) (cid:17) (cid:19)(cid:104) (cid:18) (cid:16) (cid:104) 1 − 1 e + E c∗ ∩ V 2 S, y, V 1 c∗ ∩ V 2 MOVD(S∗) + E c∗ ∩ V 2 1 − 1 e 1 − 1 e 1 − 1 e ≥ 1 2 ≥ 1 2 ≥ 1 2 S, y, V 1 S, y, V 1 S, y, V 1 y E y c(S,y) c(S,y) c(S,y) c(S,y) + f + f E y E y E y c∗ c∗ c∗ c∗ S∗, y, V 1 + V 1 S∗, y, V 1 + V 1 S∗, y, V 1 + V 1 f ≥ 1 2 1 − 1 e f y f S, y, V 1 y Now using the definition of c(S, y), we have that f (S, y, V 1 V 1 c(S∗,y). This yields (cid:16) S∗, y, V 1 c∗ ∩ V 2 c(S∗,y) (cid:17)(cid:105) . . Call the resulting seed set S. We have (cid:105) (cid:105) + V 1 c(S,y)− max V 1 ci ci V 1 ci ci ci ci c(S,y) (cid:105) V 1 + V 1 (cid:17) c(S,y)− max c∗ ∩ V 2 (cid:17)(cid:105) c∗ ∩ V 2 c(S∗,y) (cid:17) c(S,y)− max (cid:105) c∗ ∩ V 2 (cid:17) c(S,y)− max c∗ ∩ V 2 c(S,y)+V 1 c∗ ∩ V 2 c(S∗,y) V 1 c(S∗,y) (cid:17) c(S∗,y)−V 1 + V 1 c(S,y)) + V 1 (cid:16) c(S,y) ci ci c∗ ∩ V 2 (cid:17) − f (cid:16) (cid:104) (cid:17) (cid:17)(cid:105)(cid:105) (cid:105) c(S∗,y)(cid:105)(cid:105) V 1 ci . c(S∗,y)− max c(S,y)−V 1 ci c(S,y)≥ f (S, y, V 1 c∗ ∩ V 2 c(S∗,y)) + V 1 c(S,y)−V 1 S, y, V 1 c∗ ∩ V 2 c(S∗,y) S, y, V 1 c∗ ∩ V 2 c(S,y) and so we have MOVD(S) ≥ 1 2 ≥ 1 2 MOVD(S∗) + E y f S, y, V 1 c∗ ∩ V 2 c(S∗,y) MOVD(S∗) (cid:16) (cid:19)(cid:104) (cid:19) c(S∗,y)≥ f (cid:18) (cid:18) 1 − 1 e 1 − 1 e We now prove corresponding bicriteria guarantees for the POV objectives. Theorem 7: Let OP T (∆) denote the optimal value of the problem maxS≤k Pry [mC(S, y) ≥ ∆]. Let S be the set produced by POVCONSTRUCTIVE. We have 3e−1 OP T(cid:0) 1 e−1 α ∆C 1 − α (cid:1) − α POVC(S) ≥ max 0<α<1 17 (cid:80) y min(cid:0)β, f (S, y, V 2 (cid:80) (cid:80) Proof. The main difference from the two candidate case is that 1 y min (β, mC(S, y)) is no longer submodular m since mC need not be a submodular function. However, the proof of Theorem 4.3 only uses submodularity in establish- ing an approximation guarantee for greedy optimization of the surrogate. In the multicandidate case, we will greedily optimize 1 m y min (β, mC(S∗, y)) optimizes 1 m for some constant factor γ, then the same argument as in Theorem 4.3 extends to the multicandidate case. Fix any particular value of β. We establish a constant factor approximation as follows: c∗ )(cid:1), which is submodular. Let Sβ be the resulting seed set and S∗ be a set that (cid:80) y min (β, mC(Sβ, y)) ≥ γ 1 y min (β, mC(S, y)). If we can prove that 1 (cid:80) m m (cid:88) y 1 m min (β, mC(S∗, y)) = 1 m ≤ 1 m (cid:88) (cid:88) y y Via the definition of c(S∗, y), we have that V 1 This yields (cid:88) y 1 m (cid:17) (cid:17)−f C(S∗,y)(cid:17) (cid:17) c(S∗,y) . + V 1 c(Sβ ,y)−V 1 c∗ ∩ V 1 S∗, y, V 2 (cid:17)(cid:17) c(Sβ ,y) min (β, mC(Sβ, y) + mC(S∗, y) − mC(Sβ, y)) (cid:16) β, mC(Sβ, y) + f(cid:0)S∗, y, V 2 (cid:17) − f (cid:16) (cid:16) c∗ min + f (cid:1) (cid:1) − f(cid:0)Sβ, y, V 2 (cid:16) c∗ c∗ c(Sβ ,y) c(Sβ ,y) (cid:16) Sβ, y, V 2 c∗ ∩ V 1 S∗, y, V 2 S∗, y, V 2 c(S∗,y) c(S∗,y)≤ f c∗ ∩ V 1 c∗ ∩ V 1 (cid:16) c∗ ∩ V 1 S∗, y, V 2 c(Sβ ,y)−V 1 (cid:16) (cid:16) (cid:16) (cid:16) (cid:16) (cid:19) 1 (cid:1) + f β, mC(Sβ, y) + f(cid:0)S∗, y, V 2 (cid:1)(cid:17) β, mC(Sβ, y) + 2f(cid:0)S∗, y, V 2 (cid:16) (cid:88) β, f(cid:0)S∗, y, V 2 (cid:16) (cid:88) (cid:16) (cid:88) (cid:17) y 2e e − 1 2e e − 1 (cid:17) (cid:17) (cid:17) (cid:16) c∗ (cid:1)(cid:17) β, f(cid:0)Sβ, y, V 2 (cid:17) β, mC(Sβ, y) β, mC(Sβ, y) β, mC(Sβ, y) β, mC(Sβ, y) (cid:88) 1 m 2 m 1 m min min min y y + + + c∗ c∗ min min min min min min β, mC(Sβ, y) y (cid:88) (cid:88) (cid:88) (cid:88) (cid:88) y y y y (cid:1)(cid:17) ≤ 1 + 2e e − 1 m y min (β, mC(S∗, y)) ≤ 1 m ≤ 1 m ≤ 1 m ≤ 1 m ≤ 1 m (cid:18) and now the conclusion follows by applying the same argument as in Theorem 3. Theorem 8: Let OP T (∆) denote the optimal value of the problem maxS≤k Pry [mD(S, y) ≥ ∆]. Let S be the set produced by POVDESTRUCTIVE. We have POVD(S) ≥ max 0<α<1 e−1 3e−1 OP T ( 1 1 − α α ∆) − α Proof. Applying the reasoning as in Theorem 7, we have (cid:88) y 1 m min (∆, mD(S∗, y)) = (cid:88) (cid:88) y y 1 m ≤ 1 m (cid:16) min (∆, mC(S, y) + mD(S∗, y) − mD(S, y)) ∆, mD(S, y) + f(cid:0)S∗, y, V 1 c∗ (cid:1) − f(cid:0)S, y, V 1 c∗ (cid:16) min (cid:17) − f (cid:16) (cid:17) − f c∗ ∩ V 2 (cid:16) S, y, V 1 c∗ ∩ V 2 (cid:16) S, y, V 1 c∗ ∩ V 2 c(S,y) c∗ ∩ V 2 c(S∗,y) (cid:17) + V 1 (cid:1) C(S∗,y)−V 1 (cid:17) C(S,y)(cid:17) . + f S∗, y, V 1 c(S∗,y) S, y, V 1 c(S,y) The definition of c(S, y) implies that V 1 c(S∗,y)−V 1 c(S,y) ≤ f 18 so we have 1 m (cid:88) y min (∆, mD(S∗, y)) ≤ 1 m ≤ 1 m (cid:16) (cid:16) min min (cid:88) (cid:88) y y (cid:1) + f ∆, mD(S, y) + f(cid:0)S∗, y, V 1 (cid:1)(cid:17) ∆, mD(S, y) + 2f(cid:0)S∗, y, V 1 c∗ c∗ (cid:16) (cid:17)(cid:17) S∗, y, V 1 c∗ ∩ V 2 c(S∗,y) and now the theorem follows from the same argument as in Theorem 7. B Additional experimental results Figure 2: Probability of victory in constructive control. Left: irvine. Right: facebook Figure 3: Probability of victory in destructive control. Left: irvine. Right: facebook. On irvine, the MILP was termi- nated after 24 hours, and had not found competitive solutions with the approximation algorithm on the intermediate margin instances by that time. 19 6080100120C0.00.20.40.60.81.0POVApprox.MILP255075100125C0.00.20.40.60.81.0Prob. successApprox.MILP75100125150175200D0.00.20.40.60.81.0POVApprox.MILP20406080100120D0.00.20.40.60.81.0Prob. successApprox.MILP
1209.0308	1	1209	2012-09-03T11:20:57	Optimizing Supply Chain Management using Gravitational Search Algorithm and Multi Agent System	[ "cs.MA", "cs.AI" ]	Supply chain management is a very dynamic operation research problem where one has to quickly adapt according to the changes perceived in environment in order to maximize the benefit or minimize the loss. Therefore we require a system which changes as per the changing requirements. Multi agent system technology in recent times has emerged as a possible way of efficient solution implementation for many such complex problems. Our research here focuses on building a Multi Agent System (MAS), which implements a modified version of Gravitational Search swarm intelligence Algorithm (GSA) to find out an optimal strategy in managing the demand supply chain. We target the grains distribution system among various centers of Food Corporation of India (FCI) as application domain. We assume centers with larger stocks as objects of greater mass and vice versa. Applying Newtonian law of gravity as suggested in GSA, larger objects attract objects of smaller mass towards itself, creating a virtual grain supply source. As heavier object sheds its mass by supplying some to the one in demand, it loses its gravitational pull and thus keeps the whole system of supply chain per-fectly in balance. The multi agent system helps in continuous updation of the whole system with the help of autonomous agents which react to the change in environment and act accordingly. This model also reduces the communication bottleneck to greater extents.	cs.MA	cs	Optimizing Supply Chain Management using Gravitational Search Algorithm and Multi Agent System Muneendra Ojha Department of Information Technology, Indian Institute of Information Technology Deoghat, Jhalwa, Allahabad 211012, UP, India [email protected] Abstract Supply chain management is a very dynamic operation research prob- lem where one has to quickly adapt according to the changes perceived in env i- ronment in order to maximize the benefit or minimize the loss. Therefore we re- quire a system which changes as per the changing requirements. Multi agent system technology in recent times has emerged as a possible way of efficient solu- tion implementation for many such complex problems. Our research here focuses on building a Multi Agent System (MAS), which implements a modified version of Gravitational Search swarm intelligence Algorithm (GSA) to find out an opt i- mal strategy in managing the demand supply chain. We target the grains distribu- tion system among various centers of Food Corporation of India (FCI) as applica- tion domain. We assume centers with larger stocks as objects of greater mass and vice versa. Applying Newtonian law of gravity as suggested in GSA, larger ob- jects attract objects of smaller mass towards itself, creating a vir tual grain supply source. As heavier object sheds its mass by supplying some to the one in demand, it loses its gravitational pull and thus keeps the whole system of supply chain pe r- fectly in balance. The multi agent system helps in continuous updation of the whole system with the help of autonomous agents which react to the change in e n- vironment and act accordingly. This model also reduces the communication bot- tleneck to greater extents. Keywords Multi Agent Systems, Supply Chain management, Particle Swarm Optimization, Gravitational Search Algorithm, Food Corporation of India. 1 Introduction Supply chain management can be defined as “a goal-oriented network of pro- cesses and stockpoints used to deliver goods and services to customers” [1]. Sci- ence of supply chains deals with an array of suppliers, plants, warehouses, cus- tomers, transportation networks and information systems that make up actual 2 supply chains. But to study it we must break the overall complexity to a level which is relevant to the task at hand. We must define the application domain and then address the issues which are most important. In the above definition proces s- es represent activities involved in producing and distributing goods or services. On the other hand stockpoints in this definition refers to the locations in supply chain where inventories are held. Finally processes and stockpoints are connected by a network describing various ways in which goods can be supplied through the chain. Supply chain has four levels of decision classification hierarchy - strategic, tactical, operational, and real time level, where strategic level is the most im- portant among these, dealing with the resource assignment, goods movement, lo- cation of facilities to make it cost effective etc [2]. Success rate at other levels de- pend heavily on the decisions taken at this levels. However, the decision making is a difficult process as it involves the resolution of complex resource allocation and scheduling problem involving selection of different available stakeholders. Ac- cording to the findings of Beamon [3], and Amiri [4], finding the best solution for supply chain management is NP-hard problem, so it must be strategically dealt with, on case to case basis developing efficient methodology that can find out the optimal or near-optimal solution in minimum computational time. Over the last decades, research community has shown greater interest in ana- lyzing natural phenomena and mapping their models to find out optimized sol u- tion for high-dimensional search space cases. Scientists have adopted various me- ta-heuristic methods like Genetic Algorithm [5], Simulated Annealing [6], Ant Colony Search Algorithm [7], Particle Swarm Optimization [8] etc. and shown that these algorithms are able to find efficient solution of complex computational problems like pattern recognition [9], image processing [10], filter modeling [11] etc. As like any other heuristic method, these algorithms do not promise to provide the most accurate and best solution for the problems but try to give the most opt i- mized one, again on the case to case basis. So applying a new optimization algo- rithm on an untested domain is always an open problem. In this paper we discuss the results of applying a relatively new swarm optim i- zation algorithm, the Gravitational Search Algorithm (GSA) [12] in supply chain management domain. We have modeled the software application using multi agent system where intelligent agents autonomously take decision according to the changing environment and adapt to the perceived needs. In the following section we describe the relevance of GSA and Multi agent system in finding the solution of supply chain problem. Later we explain the problem s tructure with a mathemat- ical model for the proposed clustering algorithm. Finally we end with implementa- tion details and further scope of the research work. 2 Combining Multi Agent Technology with Swarm Optimization 3 Particle swarm optimization (PSO) is an evolutionary computational technique used for optimization motivated by the social behavior of individuals in large groups in nature [13]. Scientists and technicians have used different approaches to understand how such algorithms works and applied variations to optimize a large number of real life problems. They have been able to solve a variety of engineer- ing problems, but fine tuning the PSO parameters as well as finding general crite- ria applied to all problems remains a major challenge. Thus we focus on a particu- lar problem and try to find the best algorithm which is well suited to its requirements. 2.1 Why GSA is relevant in supply chain management problem? Swarm based algorithms use a collection of agents like ants or honey bees and perform parallel search with multiple starting points. Each member carries out its intended operation with a target to maximize respective utility function, and shares its information with others. These operations are normally very simple, but their collective result produces astonishing results. This phenomenon in which local i n- teraction between agents produce a global result facilitating the system to solve a problem without involving any central authority is known as swarm intelligence [14]. Gravitational Search Algorithm (GSA) is also a particle swarm optimization algorithm which works on the Newtonian laws of Gravity: “Every particle in the universe attracts every other particle with a force that is directly proportional to the product of their masses and inversely proportional to the square of the distance between them” [12]. It has found a good number of interested audiences albeit be- ing relatively new. Just like other swarm intelligence techniques, it also models the laws of nature to find out an optimized solution to a non linear and dynamic problem domain. In the GS algorithm, each entity is considered an object whose performance is measured by the mass it carries. All objects attract each other by gravitational force which causes global movement towards objects of heavier mass. As heavier masses move slower than lighter ones, this creates virtual clu s- ters representing better solutions. For the case of supply chain management, we assume all stakeholders including producer, consumer, supplier, warehouses, stra- tegic partners, vendors, and haulers etc, as objects with some respective masses. We divide these stakeholder in two groups i.e. suppliers and customers. One can be supplier at some point of time and customer at other. Although we do not con- sider the demand part of the SCM process, if we focus on calculating the total cost of moving goods from a supplier to a customer, we can simulate a major part of the whole process of supply chain management. Cost of moving might include purchase price, cost of quality, inventory losses, transportation costs and any other 4 factor which directly or indirectly affects the amount of money the customer has to shed. Thus suppliers act as objects of heavier masses or stars and customers im- itate objects with lower masses, creating planetary systems revolving around stars. 2.2 Multi Agent System representing the complete supply chain A rational agent is defined as an agent, which has clear preferences, models un- certainty via expected values, and always chooses to perform the action that r e- sults in the optimal outcome for itself from among all feasible actions [15]. A co- ordinated group of rational agents form the basis of multi agent system. Multi agent system is based on the assumption that agents coordinate among each other with the vision of achieving a common global goal. Referring the goal to be global is important here as individual agents have only local knowledge and thus local goals. But the agent system as a whole strives for one par ticular target to achieve. So whatever be the individual motives of agents, it should add an extra step to- wards the combined target of system and this is the raison d‟être of these agents [16]. When an event happens each agent cooperates to organize a collaborative group to solve the problem. Agents communicate among the nearest of neighbors thus eliminating number of communications and form virtual clusters. In supp ly chain management (SCM) also, more and more consumers collect towards a prominent supplier thus forming some kind of demand supply cluster, where d e- mands of consumers are fulfilled by nearest of supplier. Thus MAS presents an in- tuitive method of demand supply chain implementation in which each agent repre- sents one or other real life entity participating in the whole process. Many preceding studies have considered multi -agent system as the most suited way to solve complicated problems under the diverse environmental dynamisms [17][18]. A number of researchers have also successfully applied the multi agent technology for manufacturing, supply chain management, planning, scheduling, and control [19]. Their studies mainly focused on making better manufacturing system using an agent technology for automation and efficiency in the process planning and scheduling. Also, there are representative studies that have used MAS in developing intelligent manufacturing supply chain management systems e.g. the AARIA (Autonomous Agents for Rock Island Arsenal) project of Intell i- gent Automation company [20], ABCDE (Agent-Based Concurrent Design Envi- ronment) system developed by KSI (Knowledge Science Institute) of University of Calgary [21], a virtual manufacturing agent made by LIPS institute of the Uni- versity of Texas at Austin [22] and MASCOT (Multi-Agent Supply Chain cOor- dination Tool) agent of Intelligent Coordination and Logistics team of Carnegie - Mellon University [23]. In the SCM problem, GSA with multi agents system is intuitively the most suited method of implementation. Every participating entity either a consumer or supplier is represented by an agent. Every agent carries some mass. All heavier agents are supplier and lighter ones are consumers. As the problem under consid- eration is the distribution among centers of Food Corporation of India (FCI), mass determines the total amount of food grains one center has in its stock. A center can be at the same time supplier for one type of agricultural produce (e.g. wheat) and consumer for some other type of grains (e.g. rice). Thus effective mass and grav i- tational pull is context dependent. 5 3 FCI distribution system and methodology Food Corporation of India (FCI) was established to fulfill the objectives of pr o- curement of agricultural products, its storage, movement, export if asked, disposal in case of excess amount, and quality control. FCI has five zones with each zone having its member state representing region. Food Corporation of India North East South Zone North Zone East Zone West Zone Delhi Punjab Haryana UP Uttarakhand J&K HP Rajasthan Fig. 1. Hierarchical structure of FCI collection and distribution centers. North zone itself has 54 district centers in total involved in the procurement and other food management activities. For example, North zone of FCI has eight regions namely Delhi, Haryana, HP, J&K, Punjab, Rajasthan, UP and Uttarakhand regions as depicted in the Figure 1, which in turn have 54 district centers. Punjab & Haryana Regions are the major surplus states as far as production/procurement o f food grains is concerned. More than 90% stocks of these supplier states are moved to other parts of the country [24]. Apart from FCI the other major agencies involved in storage and warehous- ing of food grains in public sector are Central Warehousing Corporation (CWC) and 17 State Warehousing Corporations (SWCs). For procuring food grains from farmers, FCI creates nodal centers where it purchases at a price fixed by the Gov- ernment. It is then sent to respective region‟s storage centers or warehouses. Every warehouse has a minimum requirement as well as maximum capacity. With the broad requirements clear about FCI and its supporting agencies, we can model the whole distribution system as MAS. An agent in our system repre- sents a district center of FCI. As there are 54 district centers in North Zone, so there will be 54 agents corresponding to them. Each agent thus would have a min- imum requirement, i.e. a minimum amount of particular food grain, which it has to 6 maintain as reserve at any given instance. This reserve is decided on many factors like, Government of India policy on food security, type of food, population of the region, requirements for the region in near future, rate at which food is being con- sumed in public distribution system, state rules etc. Any amount of food above this reserve level is assumed to be surplus which can be sent to other agents in the system if required. Thus we can clearly differentiate between supplier and con- sumer at any point of time using following definition: Supplier: An agent which has food grains in excess of its set reserve amount and any point of time. Consumer: An agent which has food grains below the decided reserve at an in- stance of time. Farmer FCI purchase centers Procurement ensures minimum selling prices to the farmers and al- so effective market intervention to keep prices under check meanwhile adding the food security of the country. Storage centers of FCI, CWC and SWC Private contrac- tors and Storage centers FCI has 30.52 million tonnes (owned & hired) of storage capac i- ty in over 1820 godowns all over India Food grains are redistributed from surplus states to deficit states. FCI moves about 250 Lakh tonnes of food grains over an average distance of 1500 Kms. Stocks procured in the markets and purchase centers is first collected in the nearest depot and from there dispatched to the recipient States within a limited time. Public Distribution System It‟s the responsibility of state gov- ernments to distribute food to the consumers through the network of Fair Price Shops (FPSs) Fig. 2. Broad functionality and objectives of FCI and its supporting agencies like Central Ware- housing Center, State Warehousing Center and Public Distribution Centers. Source: Food Corpo- ration of India [24] As this definition is dependent on time and type of food grain, any agent can be a supplier for one type and consumer for other type of food grains at the same point of time. Two reasons can be visualized for automatic redistribution of food grains between centers: first, when a center needs it because it has receded below its reserve or when a center has accrued more than what it can efficiently manage. Both of these cases prompt their representative agents to autonomously act in case such scenario is perceived in the environment. Though FCI has its structural lim i- tations and boundaries, we assume it to be free from any such regulations. Al t- hough we do suggest a priority queue for suppliers in which higher priority must be given to consumers of the same zone or state rather than a consumer of diffe r- ent zone. Figure 2 gives a broad idea of the levels at which different agencies work and how food grains purchased from farmer reaches household. 7 4 The mathematical model Each participating physical entity in this supply chain system is represented by an agent. There are agents for zonal offices, regional states, district centers, nodal centers, storage warehouses, and purchase centers. Out of these, any agent can be supplier or consumer depending on the resources it owns. A consumer is consum- er till it has requires food grains to achieve a particular target level referred to as set reserved limit. It is decided by an array of parameters namely, type of food grain, near future requirement prediction, number of Public Distribution Centers (PDCs) it caters, geographical area this agent covers, positional importance, and minimum food reserves it has to maintain (decided as a policy matter by gover n- ment, hardcoded but can be changed if needed). In GSA, initially the system is considered a randomly distributed isolated space of masses (agents). Each agent carries some mass and in gravitational system, eve- ry object gets attracted by gravitational pull of another object. In our system we do not assume an agent to move. Rather we assume that whichever supplier exerts greatest force of attraction on consumer will be the first one to attempt in fulfilling its requirements. At a time „t‟ we define force acting on a consumer mass „Mc‟ from a supplier mass „Ms‟ to be ( ) ( ) ( ) ( ) (1) Where,  Fij(t) = Force of attraction by a supplier and consumer at time t.  Mc(t) = Mass of Consumer at time t = Amount it can supply = Reserve limit – Current Stock  Ms(t) = Mass of Supplier at time t = Amount it needs = Reserve limit – Current Stock  Gc(t) = Urgency factor, initially set randomly between 0 to 1. Increas- es with time.  R = Possible cost incurred in supplying food grains per unit to the consumer. 8 An agent can be attracted by another agent by a positive force of attraction. Put- ting a minus sign in equation 1 ensures that no two consumers or suppliers are a t- tracted towards each other. Unlike the traditional gravitational search algorithm proposed by [12] we have made some changes in our algorithm. We consider the attracting force to be one dimensional only. We were thus able to eliminate the possibility in which multip le smaller suppliers could collaborate and exert greater force than a single large sup- plier. In this case combined force could take a consumer to cluster of smaller sup- pliers but do not consider the cost incurred in multiple smaller transportations. We have also set R to be the cost incurred per unit where, (2) As cost increases attraction between supplier and consumer decreases though it never becomes zero as is the case in gravitational forces of nature as well. Gravita- tional constant Gc is initially randomly set as the urgency factor defining how ur- gent is the need of this particular customer. As food is supplied to it and need d e- creases, value of Gc also changes with time. Thus it is a function of initial value G0 and time t: ( ) ( ) (3) Masses of individual agents are also dependent on time t. As supply starts, the total available stock of both consumer and supplier change. This changes the r e- spective masses and thus mutual force of attraction is also decreased. If there are N agents, position of an agent is defined in 2-dimensional space as: ) ( (4) When agents attract each other they tend to move towards each other in the vi r- tual space. This movement is not reflected in the real scenario nor has any physi- cal significance. Maximum closeness between two agents during clustering phase is restricted to sum of their radii. 5 Implementation and result discussion Application simulation is divided in two phases. First one is the clustering phase when we form virtual clusters of supplier consumer agents depicting some kind of planetary system in which supplier is the star and consumer are the planets revolving around it. As the clustering phase starts, agents get attracted towards those of greater mass and their respective Euclidean distance decreases. When clustering ends each supplier is attached with many consumers as shown in Figure 3. If the supplier has enough capability to fulfill demands of more than one co n- 9 sumer it must decide which consumer to cater first. Thus every supplier creates a priority queue on the basis of first hierarchical boundaries of FCI and then in the decreasing order of consumers‟ urgency factor. Hierarchical boundaries are im- portant here as we assume an agent will cater to the needs of its state prior to send- ing grains to neighboring state. Providing autonomy to the agent system, we a s- sume that if a consumer finds a supplier „A‟ in neighboring state providing cheaper food than agent „B‟ in the same state it will approach „A‟ r ather than „B‟ as long as both states are in same zone of FCI. A A B B Fig. 3. Initially all agents are scattered in the search space and are attracted by more than one supplier. But they move closer to the one which has greater force of attraction forming virtual clusters. Actual supply of goods can only be realized physically and once it is approved and done, agents‟ information must be updated manually. As the system itself no- tifies, how much food is to be provided by a supplier to the consumer; it can a u- tomatically update its inventory after approval . Once the inventory is updated, ef- fective mass of both supplier and consumer changes. Agents that were customer for some time but their needs have been satiated will move out of the SCM pro- cess and similarly an agent that was a supplier for some time but its excess mass is already shed, will also exit the system. This behavior is perfectly represented by multi agent system as MAS gives autonomy and intelligence to its agents to pa r- ticipate in the process as long as required or leave the environment whenever deemed necessary. There is no need of a centralized command to dictate the centers which must be catered. Agents automatically form collaborative supplier-consumer clusters and cater to the needs of each other thus removing the bottleneck of central node. Number of consumers and suppliers as a given time can be generated using Pois- son distribution. As soon as a center becomes consumer the representative agent passes on its identification and requirement message to its immediate neighbors which in turn pass it on to their respective neighbors and so on. With the one hop communication channel, agents do not need to broadcast requirement to whole network thus reducing burden on the network. All the supplier agents respond 10 back with their identification and availability. When consumer receives response of a supplier it refers to the database in order to fetch supplier information like l o- cation, state, zone, modes of transportation, price of grain and any other co n- straints based on which the final cost is calculated. It calculates the attraction force using equation 1 and updates this agent as the global best. A consumer also maintains list of suppliers for as soon as new supplier comes up, the queue is r e- shuffled and new global best is updated if required. In case one supplier is not able to fulfill the requirement, consumer will approach the next in queue without pro- nouncing its need again on the network. 6 Conclusion The current system can be implemented for very large scale real life demand supply chain management. This system can be applied to observe the pattern of product usage, to find out high demand zones, or the consumption rate of an area. It can also draw out the reliability factor of a supplier and cost to benefit ratio for a product. This system can also help organizations in evaluating needs of customer and adapt accordingly. References [1] Hopp, W.J.: Supply Chain Science. McGraw Hill, New York (2006) [2] Poirier, C., Quinn, F.: How are we doing? A survey of supply chain progress, Supply Chain Management Review, 24–31 (2004). [3] Beamon, B.M.: Supply chain design and analysis: models and methods. International Journal of Production Economics 71(1-3), 145–155 (1998). [4] Amiri, A.: Designing a distribution network in a supply chain system: formulation and efficient solution procedure. European Journal of Operation research 171(2), 567–576 (2006). [5] Tang, K.S., Man, K.F., Kwong, S., He, Q.: Genetic algorithms and their applications, IEEE Signal Processing Magazine, 13(6) 22–37 (1996). [6] Kirkpatrick, S., Gelatto, C.D., Vecchi, M.P.: Optimization by simulated annealing, Science 220, 671–680 (1983). [7] Dorigo, M., Maniezzo, V., Colorni, A.: The ant system: optimization by a colony of cooperating agents, IEEE Transactions on Systems, Man, and Cybernetics – Part B 26 (1), 29–41 (1996). [8] Kennedy, J., Eberhart, R.C.: Particle swarm optimization, in Proceedings of IEEE In- ternational Conference on Neural Networks, vol. 4, pp. 1942–1948 (1995). [9] Tan, X., Bhanu, B.: Fingerprint matching by genetic algorithms, Pattern Recognition 39, 465–477 (2006). 11 [10] Nezamabadi-pour, H., Saryazdi, S., Rashedi, E.: Edge detection using ant algorithms, Soft Computing 10, 623–628 (2006). [11] Lin, Y.L., Chang, W.D., Hsieh, J.G.: A particle swarm optimization approach to non- linear rational filter modeling, Expert Systems with Applications 34, 1194–1199 (2008). [12] Rashedi, E., Nezamabadi-pour, H., Saryazdi, S.: GSA: a gravitational search algo- rithm. Information Science, 179(13): 2232–2248 (2009). [13] Fernández-Martínez, J.L. and García-Gonzalo, E.: What Makes Particle Swarm Opti- mization a Very Interesting and Powerful Algorithm? In B.K. Panigrahi, Y. Shi, and M.-H. Lim (Eds.): Handbook of Swarm Intelligence, ALO 8, pp. 37–65 (2011). [14] Tarasewich, P., McMullen, P.R.: Swarm intelligence: power in numbers, Communica- tion of ACM 45 62–67 (2002). [15] Russell, S. J., Norvig, P.: Artificial Intelligence: A Modern Approach (2nd ed.),Prentice Hall, New Jersey, ISBN 0-13-790395-2 (2003) [16] Nwana, H. S., Lee, L. C., & Jennings, N. R.: Coordination in software agent systems. The British Telecom Technical Journal, 14(4), 79-88 (1996). [17] Choi, H. S., Kim, H. S., Park, B. J., & Park, Y. S. : Multi-agent based integration scheduling system under supply chain management environment. Lecture Notes in Ar- tificial Intelligence, 3029, 249-263 (2004). [18] Julka, N., Karimi, I., & Srinivasan, R.: Agent-based supply chain management-2: A refinery application. Computers and Chemical Engineering, 26, 1771 -1781 (2002). [19] Shen, W., Ulieru, M., Norrie, D. H., & Kremer, R. : Implementing the Internet enabled supply chain through a collaborative agent system in Proceedings of Workshop on Agent Based Decision Support for Managing the Internet Enabled Supply-Chain, Se- attle. 55-62 (1999) [20] Baker, A. D., Parunak, H. V. D., & Erol, K.: Manufacturing over the Internet and into your living room: Perspectives from the AARIA project (Tech. Rep. TR208-08-97). ECECS Dept (1997). [21] Balasubramanian, S., & Norrie, D. H.: A multi-agent intelligent design system inte- grating manufacturing and ship-floor control in the Proceedings of the First Interna- tional Conference on Multi-Agent Systems. San Francisco: The AAAI press/The MIT Press. (1995) [22] Chuter, C. J., Ramaswamy, S., & Baber, K. S.: A virtual environment for construction and analysis of manufacturing prototypes. Retrieved from http://ksi.cpsc.ucalgaly.ca/ projects/mediator (1995). [23] Norman, M. S., David, W. H., Dag, K., & Allen, T. : MASCOT: an agent-based archi- tecture for coordinated mixed-initiative supply chain planning and scheduling. Pro- ceedings of the Third International Conference on Autonomous Agent (Agents ‟99), Seattle, WA. (1999). [24] Food Corporation of India. URI: http://fciweb.nic.in
0802.4450	1	0802	2008-02-29T19:05:31	A Study On Distributed Model Predictive Consensus	[ "cs.MA" ]	We investigate convergence properties of a proposed distributed model predictive control (DMPC) scheme, where agents negotiate to compute an optimal consensus point using an incremental subgradient method based on primal decomposition as described in Johansson et al. [2006, 2007]. The objective of the distributed control strategy is to agree upon and achieve an optimal common output value for a group of agents in the presence of constraints on the agent dynamics using local predictive controllers. Stability analysis using a receding horizon implementation of the distributed optimal consensus scheme is performed. Conditions are given under which convergence can be obtained even if the negotiations do not reach full consensus.	cs.MA	cs	1 A Study On Distributed Model Predictive Consensus Tam´as Keviczky∗ and Karl Henrik Johansson Abstract We investigate convergence properties of a proposed distributed model predictive control (DMPC) scheme, where agents negotiate to compute an optimal consensus point using an incremental subgradient method based on primal decomposition as described in [1], [2]. The objective of the distributed control strategy is to agree upon and achieve an optimal common output value for a group of agents in the presence of constraints on the agent dynamics using local predictive controllers. Stability analysis using a receding horizon implementation of the distributed optimal consensus scheme is performed. Conditions are given under which convergence can be obtained even if the negotiations do not reach full consensus. I. INTRODUCTION Engineered systems are becoming increasingly complex and larger in size, which presents a need for the distribution of decision-making processes that interact with or are part of these large- scale technologies and applications. An important problem that arises among such distributed decision-making systems (often called agents), is related to consensus-seeking and rendezvous, which has received a high level of interest in the recent literature [3]. The consensus-seeking and rendezvous problem consists of designing distributed control strategies such that the state or output of a group of agents asymptotically converge to a common value, a consensus point, which is agreed upon either a priori or on-the-fly using some negotiation scheme. In this paper, we assume that a consensus point is not fixed in advance, but is rather determined by an T. Keviczky is with the Delft Center for Systems and Control, Delft University of Technology, 2628 CD, Delft, The Netherlands, [email protected] K.H. Johansson is with the ACCESS Linneaus Center, School of Electrical Engineering, Royal Institute of Technology (KTH), 100 44 Stockholm, Sweden, [email protected] May 9, 2014 Technical Report 2 optimal control problem. We focus on the combination of model predictive controllers and subgradient-based negotiation of optimal consensus (along the lines of the work in [2]), and in- vestigate conditions for asymptotic convergence of such distributed control schemes. We propose an algorithm for distributed model predictive consensus, which guarantees convergence under reasonable assumptions given a sufficient number of subgradient iterations can be performed without interruption. We will model agents as constrained linear dynamical systems and build on the decentralized negotiation algorithm described in [2] to compute exactly or at least approach the optimal consensus point. This negotiation algorithm relies on primal decomposition of the optimal consensus and control problem and makes use of distributed implementation of an incremental subgradient method. Each agent performs individual planning of its trajectory and negotiates with neighbors to find an optimal or near optimal consensus point, before applying a control signal. The paper is structured as follows. Section II introduces the optimal consensus problem and some basic notation and assumptions. The decentralized negotiation scheme of [2] is summa- rized in Section III along with a decentralized receding horizon implementation of the optimal consensus problem. Stability of the proposed decentralized negotiation and control scheme is studied in Section IV for both converged and interrupted negotiations. Finally, Section V presents a numerical simulation example, which illustrates the approach for an aerial refueling scenario, and Section VI the conclusions. II. PROBLEM FORMULATION Consider N > 1 dynamic agents whose dynamics are described by the following discrete-time state equations xi t+1 = Aixi t + Biui t, yi t = C ixi t, (1) for i = 1, . . . , N, where Ai ∈ Rni and inputs of each agent are constrained to lie in polyhedral sets ×ni, Bi ∈ Rni ×mi and C i ∈ Rp×ni. We assume that the states xi t ∈ X i, ui t ∈ U i, t ≥ 0. (2) May 9, 2014 Technical Report Definition 1: [2] The dynamic agents described by (1) reach consensus at time T if yi T +k = θ, ∀k ≥ 0, i = 1, . . . , N, ui T +k = ui T , ∀k ≥ 0, i = 1, . . . , N, 3 (3) where θ lies in a compact and convex set Θ ⊂ Rp. In this paper, the consensus point θ is a vector that specifies, for example, the position and velocity the agents shall converge to. Our objective is to find a consensus point θ ∈ Θ ⊂ Rp and a sequence of inputs ui T −1, t ∈ U i for all t = 1, . . . , T − 1, such that all agent outputs are equal at with i = 1, . . . , N and ui 0, . . . , ui time T : yi T = θ, i = 1, . . . , N. (4) We will also require each agent to be at an equilibrium at time T and denote the state and e(θ)). The control equilibrium pairs of the i-th agent corresponding to a θ value with (xi e(θ), ui set of equilibria for each agent i = 1, . . . , N thus will be a function of θ on the domain Θ: We assume that the following cost function is associated with the i-th system: where Qi ∈ Rni deviations from the equilibrium states corresponding to the consensus point and the use of control ×mi are positive definite symmetric matrices (i.e., we penalize ×ni and Ri ∈ Rmi effort). Assumption 1: Each agent dynamics (Ai, Bi) is controllable and systems (Ai, (Qi) 1 2 ) are observable. We then formulate the following finite-time optimal control problem at time t based on [2]: k,t, k = 0, . . . , T − 1, for all Problem 1: Let T > 0 be fixed. Determine control vectors ui May 9, 2014 Technical Report e(θ), ui E i(θ) =(cid:0)xi e(θ)(cid:1) , u ∈ Rmi =nx ∈ Rni V i(cid:0)xi k, ui x = Aix + Biu, C ix = θo . e(θ)(cid:1)⊺ Qi(cid:0)xi e(θ)(cid:1)⊺ Ri(cid:0)ui e(θ)(cid:1) e(θ)(cid:1) , k − xi k − ui k − xi k, θ(cid:1) =(cid:0)xi +(cid:0)ui k − ui (5) (6) i = 1, . . . , N and the consensus point θt, which solve the following optimization problem: min Ut,θt N Xi=1 T −1 Xk=0 k,t, θt(cid:1) k,t, ui V i(cid:0)xi subj. to xi k+1,t = Aixi k,t + Biui k,t, yi k,t = C ixi k,t, xi k,t ∈ X i, k = 1, . . . , T, k,t ∈ U i, ui k = 0, . . . , T − 1, yi T,t = θt, xi T,t = xi e(θt), 0,t = xi xi t, i = 1, . . . , N, θt ∈ Θ, 4 (7a) (7b) (7c) (7d) (7e) (7f) (7g) k,t, . . . , uN time t + k obtained by starting from the state xi where Ut , [u0,t, . . . , uT −1,t] ∈ RT Pi mi with uk,t , [u1 tion vector containing control inputs, xi k,t], denotes part of the optimiza- k,t denotes the state vector of the i-th agent predicted at t and applying to system (1) the input sequence k−1,t. The full optimization vector consists of the vector Ut defined above and the 0,t, . . . , ui ui consensus variable θt. The subscript t will be significant later in Section III, when this problem will be solved repeatedly in a receding horizon fashion. By implementing the solution to Problem 1, agents reach consensus at time T in the sense of Definition 1. We will make the following assumptions on the feasibility of reaching the consensus point by all agents: Assumption 2: The rendezvous time horizon T is large enough so that all θt in the set Θ are feasible, i.e., reachable consensus equilibrium points for all agents. Assumption 3: For all θt ∈ Θ and i = 1, . . . , N, there exists a sequence ui 0, . . . , ui T −1 in the relative interior of U i such that yi This means that it should be possible to reach θt without saturating the control signal (not necessarily in an optimal way). T = θ. May 9, 2014 Technical Report The solution of Problem 1 was distributed among the agents in [2] by using primal decom- position in combination with an incremental subgradient method [4]. First, a multiparametric solution of the individual optimization problems was defined as 5 qi(xi t, θt) = min Ut T −1 Xk=0 k,t, ui V i(cid:0)xi k,t, θt(cid:1) subj. to (7a) − (7g), k = 1, . . . , T − 1. The optimal consensus problem in (7) can then be written as N q ∗(xt) = min θt qi(xi t, θt) Xi=1 subj. to θt ∈ Θ. The set of optimal consensus points is defined as Θ∗ t =(θt ∈ Θ (cid:12)(cid:12)(cid:12)(cid:12)(cid:12) N Xi=1 qi(xi t, θt) = q ∗(xt)) . (8) (9) (10) It can be established that the cost function qi(·) defined in (8) is a convex function and a subgradient gi for qi(·) at θt is given by the Lagrange multipliers corresponding to the terminal point constraint. A principal method for solving problem (8) is the subgradient method θt(k + 1) = PΘ"θt(k) − α(k) gi(k)# (11) N Xi=1 where gi(k) is a subgradient of qi at θt(k), α(k) is a positive stepsize, and PΘ denotes projection on the set Θ ⊂ Rp. In the following, we will consider the incremental subgradient method proposed in [5]. It is similar to the standard subgradient method (11), the main difference being that at each iteration k, θt is changed incrementally, through a sequence of N steps. Each step is a subgradient iteration for a single component function qi, and there is one step per component function. Thus, an iteration can be viewed as a cycle of N subiterations. If θt(k) is the vector obtained after k cycles, the vector θt(k + 1) obtained after one more cycle is where ϑN t (k) is obtained after the N steps θt(k + 1) = ϑN t (k), ϑi t(k) = PΘ(cid:2)ϑi−1 gi(k) ∈ ∂qi(xi t (k) − α(k)gi(k)(cid:3) , (k)), t, ϑi−1 t i = 1, . . . , N, (12) (13) May 9, 2014 Technical Report starting with ϑ0 t (k) = θt(k), 6 (14) t, ϑi−1 t (k)) denotes the subdifferential (set of all subgradients) of qi at the point where ∂qi(xi ϑi−1 t (k). The updates described by (13) are referred to as the subiterations of the k-th cycle. We will make the following assumptions, which will allow us to formulate well-posed prob- lems and characterize the number of subgradient iterations needed for convergence to a certain tolerance. Assumption 4 (Existence of an Optimal Solution): The optimal solution set Θ∗ t is nonempty. Assumption 5 (Subgradient Boundedness): There exists a scalar β such that kgik ≤ β, (15) ∀gi ∈ ∂qi(cid:0)xi t, θt(k)(cid:1) ∪ ∂qi(cid:0)xi i = 1, . . . , N, k ≥ 0, t, ϑi−1 t (k)(cid:1) , where N is the number of subiterations in each cycle. Definition 2: We will denote the Euclidean distance from a point z to the set Θ∗ Since we assume that the set Θ is compact, Assumptions 4 and 5 are automatically satisfied. t ). t by dist(z, Θ∗ Definition 3: A function γ(·), defined on nonnegative reals, is a K function if it is continuous, strictly increasing with γ(0) = 0. In the next section, we briefly describe the agreement mechanism of [2] and propose a closed- loop feedback control policy, which can be used in a receding horizon fashion, interleaved with subgradient-based negotiation of optimal consensus point updates. III. DECENTRALIZED NEGOTIATION AND RECEDING HORIZON IMPLEMENTATION SCHEME The optimal consensus point θ∗ t can be computed in a distributed way using the incremental subgradient method described in (12)-(14). Reference [2] describes an algorithm, where an estimate of the optimal consensus point is passed around between agents. Upon receiving an estimate from its neighbor, an agent solves the optimization problem (8) to evaluate its cost of reaching the suggested consensus point and to compute an associated subgradient (Lagrange multiplier of terminal point constraint). The agent then performs a subiteration by updating the consensus estimate according to (13) and passing the estimate to the next agent. Each agent only computes a subgradient with respect to its own part of the objective function and not the global May 9, 2014 Technical Report 7 objective function. The convergence of the incremental subgradient algorithm is guaranteed if the agents can be organized into a cycle graph (for more details see [2]). Remark 1: Besides some technical assumptions given in [1], the primal decomposition scheme and convergence to the optimal solution of (7) using sequential local subgradient iterations is possible due to decoupled and independently constrained agent dynamics. Furthermore, the overall objective function is decomposable into a sum of terms that share only a single coupling variable, θt. Thus fixing a θt value in the cost and constraints separates the optimal control problem into local ones. The control solution U ∗ t provides an open-loop control strategy for finite-time optimal consensus. However, this solution is sensitive to t corresponding to a negotiated optimal consensus point θ∗ model mismatch and disturbances, which suggests considering a receding horizon implementation and repeated solution of the finite-time optimal consensus problem due to its feedback nature. Our goal in such an approach is to guarantee constraint fulfillment and asymptotic convergence to a consensus point by repeatedly solving optimal consensus problems and implementing the first sample of the control solution. More formally, let U ∗ the first sample of U ∗ t 0,t, . . . , u∗ t = [u∗ is applied to the collection of agents: T −1,t] and θ∗ t be an optimal solution of (7) at time t. Then, ut = u∗ 0,t. (16) The optimization (7) is repeated at time t + 1, based on the new state xt+1. Remark 2: Stability of such a combination of DMPC and incremental subgradient methods is not a trivial question, especially since the terminal constraint value in the receding horizon scheme based on (7) is an optimization variable as well. The main point of the following investigation is to rule out a scenario where repeatedly solving and implementing the first step of a finite-time optimal control solution with changing terminal constraint value eventually results in divergence or lack of stability. Compared to the work in [1], this question arises because we are no longer considering only the open-loop implementation of a control sequence that terminates with the t ), based on new t ) at time T , but one that is updated every time step (along with θ∗ value ue(θ∗ measurements in a receding horizon fashion. May 9, 2014 Technical Report 8 IV. STABILITY ANALYSIS In this section we will be primarily interested in establishing conditions for asymptotic con- vergence of the combined DMPC and consensus algorithm to the set of equilibria defined as E =(cid:0)E i(θ), . . . , E N (θ), θ(cid:1) , θ ∈ Θ. (17) A. Fully Converged Negotiations For now, we will assume that in each implementation cycle (i.e., at sampling time t), the t have converged before the implemen- tation of the corresponding control actions. In other words, the optimal solution of problem (7) distributed negotiations on the optimal consensus value θ∗ is attained by every agent in each time step by means of the distributed consensus algorithm of [1]. This allows us to consider the overall system as a whole for stability analysis, using the following aggregate dynamics xt+1 = Axt + But, yt = Cxt, (18) where A = diag(Ai) ∈ RPi ni RpN ×Pi ni. The states and inputs of the overall system are constrained by ×Pi ni, B = diag(Bi) ∈ RPi ni ×Pi mi and C = diag(C i) ∈ U i, t ≥ 0, (19) xt ∈ X =Yi X i, ut ∈ U =Yi where the symbol Q denotes the standard Cartesian product of sets. Note that according to (4), consensus for the aggregate system dynamics means yT = CxT = 1N ⊗ θ∗ t . Stability analysis in this case pertains to the study of the receding horizon control scheme given in (7) and (16) with a terminal point constraint to one of its optimization variables θt. This will be performed next. The set of states at time k feasible for Problem 1 is given by Xk = {x ∃u ∈ U such that Ax + Bu ∈ Xk+1} ∩ X , with (20) XT −1 = {x ∃u ∈ U and θ ∈ Θ such that x = Ax + Bu and C(Ax + Bu) = 1N ⊗ θ} ∩ X . May 9, 2014 Technical Report Denote with c(xt) = u∗ 0,t, 9 (21) the control law obtained by applying the receding horizon control policy in (7) and (16) with cost t ]. Consider the aggregate function (6) for each agent, when the current state is xt = [x1 dynamical model (18) and denote with t , . . . , xN xt+1 = Axt + Bc(xt), (22) the closed-loop dynamics of the entire system. In the following theorem, we state sufficient conditions for the asymptotic convergence of the closed-loop system to the set of equilibria E. Theorem 1: Assume that (A0) Qi ≻ 0, Ri ≻ 0 for all i = 1, . . . , N. (A1) For all θt ∈ Θ there exists a unique equilibrium xi e(θt) ∈ X i, ui e(θt) ∈ U i for all i = 1, . . . , N such that xi e = Aixi e + Biui e and C ixi e = θt. (A2) The state and input constraint sets X i and U i contain all xi e and ui e equilibrium pairs in their interior, respectively, for all i = 1, . . . , N. Then, the closed-loop system (22) asymptotically converges to the set of equilibria E with domain of attraction X0. Proof: We introduce the following notation: T −1 and J i(cid:0)xi t, U i t , θt(cid:1) = J(xt, Ut, θt) = Xk=0 Xi=1 N V i(cid:0)xi J i(cid:0)xi t, U i k,t, ui k,t, θt(cid:1) t , θt(cid:1) . (23) (24) The optimal value function obtained from solving problem (7) at time t will thus be denoted as J ∗(xt, U ∗ t , θ∗ t ). We will show first that the optimal value function J ∗(xt, U ∗ t , θ∗ loop trajectories of the overall system at each time step J ∗(xt+1, U ∗ the assumptions of the theorem hold. t ) decreases along the closed- t ), if t+1, θ∗ t+1) ≤ J ∗(xt, U ∗ t , θ∗ Let the initial state at time t be xt = x0,t ∈ X0 and let U ∗ optimizers of problem (7). Denote with x∗ t = [x0,t, x∗ 1,t, . . . , x∗ trajectory, with 1N ⊗ θ∗ t = Cx∗ T,t. Let xt+1 = x∗ 1,t = Ax0,t + Bu∗ 0,t, . . . , u∗ T −1,t] and θ∗ t be the t = [u∗ T,t] the corresponding optimal state 0,t and consider problem (7) for May 9, 2014 Technical Report 10 1,t, . . . , u∗ time t + 1. We will construct an upper bound for J ∗(xt+1, U ∗ Ut+1 = [u∗ xt+1, xt+1 = [x∗ t + 1 if and only if v ∈ U keeps C(Ax∗ t+1). Consider the sequence T −1,t, v] and the corresponding state trajectory resulting from the initial state T,t + Bv]. The input Ut+1 will be feasible for the problem at T,t + Bv) equal to some 1N ⊗ θ with θ ∈ Θ at step T T,t + Bv) = 1N ⊗ θ. Such v exists by hypothesis (A1). Since x∗ T,t is an equilibrium of the system, this also allows us to choose a feasible v, for which in fact of the prediction, i.e., C(Ax∗ 1,t, . . . , x∗ T,t, Ax∗ t+1, θ∗ C(Ax∗ T,t + Bv) = 1N ⊗ θ∗ t . This is accomplished by noticing that x∗ T,t = xe(θ∗ t ) and selecting v = ue(θ∗ t ). (25) J(xt+1, Ut+1, θ∗ generated by U ∗ t ) will be an upper bound for the optimal J ∗(xt+1, U ∗ t+1). Since trajectories t and Ut+1 overlap (except for the first and last sampling intervals), it is immediate t+1, θ∗ to show that J ∗(xt+1, U ∗ t+1, θ∗ t+1) ≤J(xt+1, Ut+1, θ∗ t ) =J ∗(xt, U ∗ t , θ∗ t ) − (x0,t − xe(θ∗ t ))⊺Q(x0,t − xe(θ∗ t )) (26) − (u∗ 0,t − ue(θ∗ t ))⊺R(u∗ 0,t − ue(θ∗ t )) + ((Ax∗ T,t + Bv) − xe(θ∗ t ))⊺Q((Ax∗ T,t + Bv) − xe(θ∗ t )) + (v − ue(θ∗ t ))⊺R(v − ue(θ∗ t )), where Q = diag(Qi) ∈ RPi ni value given in (25) leads to Ax∗ ×Pi ni, R = diag(Ri) ∈ RPi mi ×Pi mi. Choosing the particular v T,t + Bv − xe(θ∗ t ) = 0, so equation (26) becomes J ∗(xt+1, U ∗ t+1, θ∗ t+1) − J ∗(xt, U ∗ t ) t , θ∗ ≤ − (x0,t − xe(θ∗ t ))⊺Q(x0,t − xe(θ∗ t )) (27) − (u∗ 0,t − ue(θ∗ t ))⊺R(u∗ 0,t − ue(θ∗ t )) ≤ − γ(k(xt − xe(θ), ut − ue(θ))k), ∀xt ∈ Xt. where γ is a class K function. This inequality along with hypothesis (A0) on the matrices Q t ) decreases along the state trajectories of the closed-loop system t ) → and R ensure that J ∗(xt, U ∗ (22) for any xt ∈ Xt. Since J ∗(xt, U ∗ J ⋆ as t → ∞, where J ⋆ is a nonnegative constant. We conclude that J ∗(xt+1, U ∗ t , it follows that J ∗(xt, U ∗ t ) ≥ 0 for all xt, U ∗ t , θ∗ t , θ∗ t , θ∗ t , θ∗ t+1, θ∗ t+1) − May 9, 2014 Technical Report 11 t , θ∗ t ) → 0 as t → ∞ and this implies that γ(k(xt − xe(θ), ut − ue(θ))k) → 0. From J ∗(xt, U ∗ γ(·) being a K function, it follows that xt − xe(θ), ut − ue(θ) → 0 as t → ∞. ✷ B. Interrupted Negotiations In case the distributed negotiation process is interrupted (e.g., due to execution time constraints) implemented, the θi or otherwise allowed to run only for a finite number of iterations before the control inputs are t . This means that individual agents will issue control commands that will guide them to possibly close but different t values do not converge to a common optimal value θ∗ terminal consensus points. In such a situation, we desire to find conditions under which repeated negotiation and implementation of intermediate consensus results will still allow asymptotic convergence to a common consensus point for each agent. We propose an algorithm that fulfills the above objective if the subgradient iterations in subsequent time steps approach the optimal consensus point to an increasingly more accurate level and at the same time the local MPC solutions satisfy an improvement property along the closed-loop evolution of the agents’ dynamics. The first requirement ensures that the mismatch t values diminishes as t → ∞. The second requirement is analogous to the standard suboptimal MPC scheme in [6], where it is established that feasibility between different interrupted θi of such an improvement constraint implies stability of the receding horizon control scheme. by agent i in the subgradient negotiation process of time instance t by θi In the following, we will denote the last (i.e., implemented) final consensus point reached t. This intermediate consensus point is not optimal for the global optimization problem (7), but due to Assumptions 2-3 it is certainly feasible for the following local problem: min θi t qi(cid:0)xi t, θi t(cid:1) subj. to θi t ∈ Θ. (28a) (28b) Distinguishing between the local θi t variables allows the original global optimization problem (9) to be restated as N qi(xi t, θi t) min θt Xi=1 subj. to θ1 t = · · · = θN t ∈ Θ. (29) Technical Report May 9, 2014 As opposed to the fully converged subgradient scheme in Section IV-A, the θi t variables do not converge to the globally optimal one, thus we cannot rely on optimality of the MPC scheme to prove global convergence. Instead, an improvement property as shown in [6], which is required for asymptotic convergence to the set of equilibria will be formulated as N 12 Xi=1 (cid:0)J i(cid:0)xi t+1, U i t+1, θi t+1(cid:1) − J i(cid:0)xi t, U i t , θi t(cid:1)(cid:1) (30) ≤ − γ(k(xt − xe(θ), ut − ue(θ))k), where γ is a class K function. A feasible sequence for such a constraint always exists based on Assumptions 2-3 and the earlier developments in Section IV-A. general different. Thus, if the agents’ initial states are close to their local xi optimum, since the common terminal point constraint is missing and the local θi∗ The value function improvement property in (30) is not sufficient for convergence to the global t values are in t ) equilibria, which are significantly different from each other, then any subgradient-based or other adjustment of the t value would necessarily result local terminal point constraints towards the globally optimal θ∗ e(θi∗ in both local and global cost increase. which ensures that the θi This suggests that an additional requirement besides the cost improvement property is needed, t values will also converge over time to a common θt. This can be accomplished by requiring that in each iteration the subgradient-based negotiation scheme is executed at least until kθi t − θ∗ t k ≤ εt ∀i = 1, . . . , N, where the approximation bound is updated for instance according to εt ≤ ǫ 1 t , ǫ > 0. (31) (32) In order to have some information about the required number of incremental subgradient iterations that guarantee fulfillment of constraints (30) and (31), we will make use of the following result from [7]. It can be shown that under a strong convexity type assumption, the incremental subgradient method defined earlier in (12)-(14) with an appropriately chosen stepsize α(k) has a sublinear convergence rate: Proposition 1: [7] Let Assumptions 4 and 5 hold, and assume that there exists a positive scalar µ such that May 9, 2014 q(xt, θt) − q ∗(xt) ≥ µ (dist(θt, Θ∗ t ))2 , ∀θt ∈ Θ. (33) Technical Report Then for the sequence {θt(k)} generated by the incremental subgradient method with the stepsize of α(k) = 1 2µ k+1, µ > 0, we have 1 13 (dist(θt(k + 1), Θ∗ t ))2 ≤ . (34) 1 + ln(k + 1) k + 1 N 2β2 4µ2 In the following, we describe a scheme, which allows the two conditions (30) and (31) to be tested based on the cyclic communication scheme underlying the subgradient-based negotiation. They execute at least the number of iterations dictated by the optimal θ∗ In Algorithm 1, agents perform cyclic iterations of the subgradient (SG) method (12)-(14). t approximation require- ment in (31). Satisfaction of the test (31) is signaled by a flag fSG. If needed, agents continue with subgradient iterations until the global cost improvement property in (30) is satisfied. This is signaled by flag fDMPC. In order to accomplish this, agents pass along besides their current subiterate ϑi t(k) of the consensus point in iteration k at sampling time t, the two binary variables (flags) fDMPC and fSG corresponding to tests (30) and (31), and two vectors of dimension N: Jcurr and Jprev. These vectors contain the individual cost values associated with the current and previous sampling time, respectively. Jprev has values corresponding to the cost of using the final ϑi t−1(k) consensus points for implementation during the previous sampling time t − 1. The current cost Jcurr gets filled up cyclically using the most recent subiterate ϑi t(k) for each agent. When an agent computes its own consensus point subiterate, it calculates the corresponding local cost value and checks the sum of previous and current cost values for each agent to decide whether the improvement property (30) is satisfied. If it is, then it sets a flag fDMPC, which indicates that the improvement property (30) is fulfilled and every other agent should enter in an implementation phase, provided that condition (31) is also satisfied. The message reaches all other agents eventually as they pass along this information in a cyclic pattern. If property (30) is not satisfied, then it puts its current cost value entry in the vector Jcurr and passes it on to the next agent. Theorem 2: Under the assumptions of Section II, Algorithm 1 converges asymptotically to the set of equilibria E. Proof: The main idea of the proof follows along the lines of Theorem 1, except for two crucial points. A feasible sequence for the improvement constraint (30) always exists based on Assumptions 2-3 and the developments in Section IV-A. This improvement property guarantees May 9, 2014 Technical Report 14 that even with interrupted negotiations, the distributed MPC problem converges asymptotically to t are different in this case). However, these terminal points are guaranteed to form a single consensus point, attained asymptotically by the repeated some set of different terminal points (since θi application of the iterative subgradient method, due to (31) and the compactness of Θ. ✷ Remark 3: Although Algorithm 1 guarantees global convergence, it requires an increasing number of subgradient iterations in subsequent time steps in order to approach the optimal value with a decreasing tolerance. The requirements (30) and (31) are only sufficient conditions and thus might be somewhat conservative. Decreasing the initial stepsize of the subgradient iterations may solve this problem. The increasing number of subgradient iterations can also be alleviated in practice in the following way: Once εt gets small enough or another condition indicating closeness to the global consensus point is satisfied, the θ consensus point can be fixed for all agents and the scheme could proceed with a pure decentralized MPC scheme. This would ensure convergence due to the result shown in Section IV-A. V. NUMERICAL EXAMPLE OF AN AERIAL REFUELING SCENARIO This numerical example considers a simplified aerial refueling scenario with three aircraft, which illustrates the importance of negotiating an optimal consensus point and the difference from standard rendezvous problems. The scheme involves the linearized longitudinal dynamics of three aircraft representing the tanker and two smaller aircraft to be serviced, respectively. The simplified objective is to control all three aircraft from their initial status to the same altitude and airspeed, which will allow the refueling process to be initiated. The consensus variable θt = [∆h ∆V ]⊺ will thus be a vector of dimension two, with entries of altitude and true airspeed deviations from the trim values, respectively. The optimal rendezvous altitude and airspeed will be determined using the distributed (cyclic) negotiation process based on local subgradient iterations. Once the corresponding control actions are implemented, the negotiations start again in the subsequent sampling time from the new initial states. This receding horizon procedure controls all three aircraft to a common altitude and airspeed. The optimal choice of the rendezvous parameters depend on the dynamics and constraints of the individual aircraft, along with the weighting matrices of the local cost functions, which penalize deviations in altitude, velocity and the use of control actions. The tanker is represented by a Boeing 747-100/200 aircraft model based on [8]. The two May 9, 2014 Technical Report 15 serviced vehicles are modeled as F-16 aircraft from [9], [10]. The models were all discretized with a sampling time of 0.05 s. The cost function of the tanker penalizes altitude-changes heavily, and the cost functions associated with the two fighters are based on their health or fuel level. The linearized longitudinal models of all aircraft involved correspond to a straight and level flight condition at 4000 m altitude and 184 m/s true airspeed. The control inputs of the linearized Boeing 747-100/200 aircraft represent deviations from the trim values of thrust and elevator deflection, denoted by uB747 = [δB747 δe]⊺, respectively. The F-16 aircraft models are equipped with an inner-loop pitch rate controller so the control inputs of their linearized models are deviation from trim throttle and desired pitch rate command uF16 = [δF16 th The constraints on the aircraft inputs are defined as the following δqcmd]⊺, respectively. th   −50000 lb −10 deg −1000 lb −100 deg/s   δB747 th δe δF16 th δqcmd       +150000 lb +10 deg +5000 lb +100 deg/s   ≤ ≤ . (35) The initial altitudes of the two F-16s are chosen as ±30.48 m around the trim altitude. The tanker’s initial altitude w.r.t. the trim value is −10 m. All aircraft are initialized as flying straight and level with the same trim velocity. If we simply considered the average value of the initial altitudes as a rendezvous point, then the aircraft should approach approximately −3 m. Due to the different dynamics and cost functions of the aircraft representing fuel cost and vehicle health priorities, this average value is not optimal. We consider the following weighting matrices in the local cost functions QF16 QB747 =  1 =  2 =  QF16 0 1 0 0 1  , RB747 =  10 1 =  , RF16  10 2 =  , RF16  0 10 0 0.4 0 10−7 0 10−5 0 0 10−5 0 0 104  , 0.5  , 0.1  . 0 (36) (37) (38) These weights intend to mimic a situation, where the F-16 indexed as #1 penalizes altitude changes and pitch control more heavily due to a restriction posed by limited fuel supply or elevator control authority. May 9, 2014 Technical Report 16 Simulation results of the centralized solution to the optimal finite-time rendezvous problem with time horizon of T = 100 steps (5 seconds) are shown in Figure 1. The optimal rendezvous point is determined to be at 21.2340 m away from the trim altitude and the common terminal velocity at straight and level flight equilibrium is −1.1933 m/s away from the initial trim velocity. Notice that the rendezvous altitude is quite close to F-16 #1 as an indication of its limited maneuvering capabilities represented in its local cost function. The problem can be solved in a distributed fashion using the incremental subgradient method described in [2] and implemented through a cyclic sequential update. Since the optimal ren- dezvous point and the corresponding control solutions converge to the centralized one asymp- totically, only convergence of θi t(k) are shown in Figure 3. In order to have a receding horizon implementation of the incremental subgradient method proposed in Algorithm 1, the maximum number of major subgradient iterations were limited to 15 before applying the local control solutions at each new MPC update. The limited number point θi of subgradient iterations lead to a mismatch between local versions of the common rendezvous t. Still, the MPC implementation of the finite-time optimal rendezvous problem with local rendezvous point mismatch stabilizes the system and provides acceptable performance besides providing feedback action. The resulting trajectories are shown in Figure 2 until sampling time t = 100. The convergence of θB747 one horizontal axis represents the subgradient iterations k (limited at 15) and the other horizontal (k) in each MPC update period is shown in Figure 4, where t axis corresponds to the MPC sampling time index t. Notice that besides the improvement made in each MPC sampling period, there is also a trend indicating that the local θB747 (k) value t (k) and approaches it more and more closely in subsequent MPC t starts from a closer point to θ∗ updates. VI. CONCLUSIONS We have introduced a distributed model predictive control (DMPC) framework, where the control objective is to agree upon and achieve an optimal consensus point for constrained dynamic agents. The negotiation scheme makes use of the cyclic incremental subgradient algorithm described in [2]. Convergence properties of the combined DMPC / incremental subgradient approach were analyzed and a sufficient minimum number of subgradient iterations were es- tablished. An algorithm was proposed that ensures convergence of the decentralized scheme. May 9, 2014 Technical Report 17 F16 #1 B747 F16 #2 20 40 60 80 100 t ] m [ h ∆ 40 30 20 10 0 −10 −20 −30 −40 0 Fig. 1. Centralized solution of the finite-time optimal rendezvous problem in an aerial refueling scenario. ] m [ h ∆ 40 30 20 10 0 −10 −20 −30 −40 0 F16 #1 B747 F16 #2 20 40 60 80 100 t Fig. 2. MPC implementation of the finite-time optimal rendezvous problem with subgradient iterations limited at 15. May 9, 2014 Technical Report 18 F16 #1 B747 F16 #2 100 k ∗ θ − ) k ( i θ k k ∗ θ k 10−1 0 10 20 30 k 40 50 60 Fig. 3. Convergence of local estimates to the optimal rendezvous point using the incremental subgradient algorithm. k ∗ t θ − ) k ( 7 4 7 B θ k t 100 k ∗ t θ k 10−2 0 5 Subgradient iteration number 10 k 60 80 15 100 0 20 40 Sampling time of MPC implementation t Fig. 4. Convergence of local estimates to the optimal rendezvous point during each MPC update. May 9, 2014 Technical Report 19 Besides the aerial refueling numerical example shown in Section V, the approach proposed in this paper can be used in other distributed “synchronization” problems as well, where agents with constrained dynamics have to agree upon and achieve simultaneously an “optimal” consensus value, which is not known a priori. Our current work considers schemes that relax the cyclic, sequential communication requirement and rely on parallel, localized iterations. VII. ACKNOWLEDGEMENTS The authors wish to thank Richard Murray for helpful discussions. The work of Tam´as Keviczky was partially funded by the Boeing Corporation while he was a postdoctoral scholar at the Control and Dynamical Systems, California Institute of Technology. The work by Karl H. Johansson was partially funded by the Swedish Research Council, the Swedish Strategic Research Foundation, and the European Commission through the NoE HYCON. It was partially done during a stimulating sabbatical visit at Caltech, which is gratefully acknowledged. REFERENCES [1] B. Johansson, A. Speranzon, M. Johansson, and K. H. Johansson, “Distributed model predictive consensus,” in Proc. 17th Int’l Symposium on Mathematical Theory of Networks and Systems (MTNS), Kyoto, Japan, July 2006, pp. 2438–2444. [2] ——, “On decentralized negotiation of optimal consensus,” Automatica, 2007, doi: 10.1016/j.automatica.2007.09.003. [3] R. Olfati-Saber, J. A. Fax, and R. M. Murray, “Consensus and cooperation in networked multi-agent systems,” Proceedings of the IEEE, vol. 95, no. 1, pp. 215–233, 2007. [4] D. P. Bertsekas, A. Nedi´c, and A. E. Ozdaglar, Convex Analysis and Optimization. Athena Scientific, 2003. [5] A. Nedi´c and D. Bertsekas, “Incremental subgradient methods for nondifferentiable optimization,” SIAM J. on Optimization, vol. 12, pp. 109–138, Mar. 2001. [6] P. O. M. Scokaert, D. Q. Mayne, and J. B. Rawlings, “Suboptimal model predictive control (feasibility implies stability),” IEEE Trans. on Automatic Control, vol. 44, no. 3, pp. 648–654, Mar. 1999. [7] A. Nedi´c and D. Bertsekas, “Convergence rate of incremental subgradient algorithms,” in Stochastic Optimization: Algorithms and Applications, S. Uryasev and P. M. Pardalos, Eds. Kluwer Academic Publishers, Dordrecht, Netherlands, 2001, pp. 223–264. [8] A. Marcos, S. Ganguli, and G. J. Balas, “An application of h∞ fault detection and isolation to a transport aircraft,” Control Engineering Practice, vol. 13, pp. 105–119, 2005. [9] B. L. Stevens and F. L. Lewis, Aircraft Control and Simulation. Wiley-Interscience, 1992. [10] T. Keviczky and G. J. Balas, “Receding horizon control of an F-16 aircraft: a comparative study,” Control Engineering Practice, vol. 14, no. 9, pp. 1023–1033, Sept. 2006. May 9, 2014 Technical Report 20 /* M is large number */ Algorithm 1: Cyclic incremental DMPC algorithm 1 2 3 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 29 30 31 32 33 34 35 36 37 Initialize β, µ, ǫ, θi 0; fDMPC, fSG ←− false; k, t ←− 0; Jcurr ∈ RN ←− 0; Jprev ∈ RN ←− −M · 1; loop Measure states xi t; repeat 1 α(k) ←− 1 k+1; 2µ t (k) ←− θt(k); ϑ0 for i = 1 to N do if fDMPC ∧ fSG then J i t (xi prev ←− J i Implement ui∗ t(k) ←− ϑi ϑi t, U i 0,t(ϑi t(k − 1) t(k − 1)); t , ϑi t(k − 1)); else t, ϑi−1 t (k)); Compute a gi(k) ∈ ∂qi(xi ϑi t t (xi t, U i curr ←− J i J i i=1(J i curr − J i Set fDMPC true; t(k) ←− PΘ(cid:2)ϑi−1 if PN (k) − α(k)gi(k)(cid:3); t , ϑi prev) ≤ 0 then t(k)); else Set fDMPC false; end end end θt(k + 1) ←− ϑN if 1+ln(k+1) t (k); N 2β 2 4µ2 ≤ ǫ k+1 Set fSG true; t then else Set fSG false; end k ←− k + 1; until new measurement is available ; t ←− t + 1; k ←− 0; end loop May 9, 2014 Technical Report
1910.05092	1	1910	2019-10-11T11:41:03	Modeling Cyber-Physical Human Systems via an Interplay Between Reinforcement Learning and Game Theory	[ "cs.MA", "cs.LG" ]	Predicting the outcomes of cyber-physical systems with multiple human interactions is a challenging problem. This article reviews a game theoretical approach to address this issue, where reinforcement learning is employed to predict the time-extended interaction dynamics. We explain that the most attractive feature of the method is proposing a computationally feasible approach to simultaneously model multiple humans as decision makers, instead of determining the decision dynamics of the intelligent agent of interest and forcing the others to obey certain kinematic and dynamic constraints imposed by the environment. We present two recent exploitations of the method to model 1) unmanned aircraft integration into the National Airspace System and 2) highway traffic. We conclude the article by providing ongoing and future work about employing, improving and validating the method. We also provide related open problems and research opportunities.	cs.MA	cs	Modeling Cyber-Physical Human Systems via an Interplay Between Reinforcement Learning and Game Theory Mert Albabaa, Yildiray Yildiza aDepartment of Mechanical Engineering, Bilkent University, Turkey Abstract Predicting the outcomes of cyber-physical systems with multiple human interactions is a challenging problem. This article reviews a game theoretical approach to address this issue, where reinforcement learning is employed to predict the time-extended interaction dynamics. We explain that the most attractive feature of the method is proposing a computationally feasible approach to simultaneously model multiple humans as decision makers, instead of determining the decision dynamics of the intelligent agent of interest and forcing the others to obey certain kinematic and dynamic constraints imposed by the environment. We present two recent exploitations of the method to model 1) unmanned aircraft integration into the National Airspace System and 2) highway traffic. We conclude the article by providing ongoing and future work about employing, improving and validating the method. We also provide related open problems and research opportunities. Keywords: Cyber-physical human systems, Game theory, Reinforcement learning, Model validation 1. Introduction In a 2006 NASA report, allocation of tasks, as well as switching, between humans and automation is stated as one of the highest priority research needs for a successful next gen- eration airspace development, where several new automation components are expected to be introduced to cope with the inevitable increase in traffic density (Sheridan et al., 2006). A 2012 U.S. Department of Defense report (Murphy and Shields, 2012) declares that the tax- onomy established by the "levels of autonomy" creates "a focus on machines, rather than on the human-machine system," which in turn "has led to designs that provide specific functions rather than overall resilient capability". The same report suggests that human-system col- laboration should be the defining theme for the design and operation of autonomous systems. In a 2017 review article (Lamnabhi-Lagarrigue et al., 2017), where an in-depth analysis is provided about the current and future roles of the systems and control field, the question "how to optimally conjugate automated systems with the interplay of humans?" is posed as one of the grand challenges. It is therefore clearly understood by engineers and scientists that considering humans as integral parts of complex physical systems that embody commu- nication, computation, control and networking technologies, can potentially accelerate the Preprint submitted to Annual Reviews in Control October 14, 2019 advancement of technology that can address pressing human needs. This requires a new perspective that considers the human, the physical plant and enabling cyber-technologies as a single system, namely a cyber-physical and human system (CPHS). This is in contrast to imagining the human solely as a user who is isolated from the technology. Adopting the CPHS framework brings its own challenges, especially when it comes to obtaining predictive models. Apart from the intricacies of cyber-physical system (CPS) modeling, the human element is usually the most demanding component of a CPHS in terms of forecasting the future behavior. The difficulty intensifies when the system contains more than one human, which requires factoring in multiple human-human and human-automation interactions. The focus of this paper is on the latter type, where human interactions are an inseparable feature of the system. To simplify the exposition in this paper, we use "human-machine interactions", "human-autonomy interactions" and "human interactions" interchangeably, although a more careful use of the language that pays attention to nuances is possible. To obtain realistic models of CPHS, therefore, we need to concentrate on modeling methods that give us the ability to include typical human characteristics. For example, human interactions are generally not deterministic in nature, which can be captured in CPHS models by utilizing a probabilistic modeling framework. Furthermore, before taking an action, a human generally contemplates other intelligent agents' (such as other humans or automation) possible actions and then tries to choose a move that will increase the chances of obtaining the best outcome reflecting his or her preferences. A representational CPHS model needs to incorporate this "strategic behavior" (Camerer, 2011) of humans. Finally, as much as we want to believe otherwise, humans do not always, if at all, act in an optimal fashion. Our cognitive capabilities and computational powers are not always ample enough to provide the best response to a given situation. This final point is important for distinguishing CPHS models from autonomy models, where, based on available information, an algorithm can possibly be designed to react in the most appropriate way. Models that comprise these three attributes of human reactions, namely, being probabilistic, strategic and non-optimal, have a higher chance of success in terms of representing real-life behavior, for the cases where multiple humans are involved. A solution for CPHS modeling that addresses human interactions is proposed by the introduction of the "Semi Network-Form Games" (SNFG) formalism (Lee and Wolpert, 2011; Lee et al., 2013). SNFG merges three modeling tools: Bayesian networks, game theory (GT) and reinforcement learning (RL). While the Bayesian networks form the probabilistic foundation of the method, game theory provides the required mechanism to produce strategic behavior expected from human interactions, and reinforcement learning enables obtaining time-extended scenarios where humans take successive actions. Non-optimal behavior, which is thought to be a typical human trait, emerges naturally in SNFG with the type of the exploited game theoretical solution technique and reinforcement learning. The first research result that exploits the SNFG idea of merging GT and RL to create a CPHS modeling framework for a realistic engineering system, including more than 2 humans, has appeared in Musavi et al. (2016), where the problem of integrating unmanned aircraft systems (UAS) into the national airspace (NAS) is carefully studied. The results reported in 2 Musavi et al. (2016) are supported by extensive simulation capabilities, where the interactions of 180 manned aircraft pilots are modeled. This study builds upon earlier initial attempts to model smaller airspace scenarios with human interactions (Yildiz et al., 2012, 2013, 2014). Simultaneous modeling of a large number of decision making agents (pilots, in this case) in a complex scenario is difficult due to the computational cost. A common approach in the literature is to model a single decision maker, whose actions are of interest, and force the rest to obey certain kinematic and dynamic constraints, to obtain a reasonable model behavior. However, although this approach presents some insight into the dynamics of the system, it is limiting due to grossly simplifying the real-life human interactions. The study conducted by Musavi et al. (2016) provided, for the first time in the literature, probabilistic outcomes of UAS integration scenarios, where each of the 180 aircraft pilots are modeled using the game theoretical decision making process, simultaneously. In this research, Bayesian Networks are not utilized and the probabilistic decision making is obtained through employing a stochastic reinforcement learning algorithm proposed in Jaakkola et al. (1994). Recently, this work is extended for the cases where the aircraft can move both horizontally and vertically in a 3-dimensional airspace (Musavi et al., 2018). This extension required a dramatically larger observation space for the pilots which ruled out the possibility of using exact RL methods. To address this issue, Neural-Fitted Q-iteration (Riedmiller, 2005; Riedmiller et al., 2007; Gabel et al., 2011) is integrated into the game theoretical framework, which uses neural networks (NN) for compactly estimating the exact state-action values needed by the RL algorithm. Another study, exploiting the same approach, conducted by Li et al. (2018) achieved a similar result in the automotive domain, by creating a modeling framework for road traffic consisting of 50 manned vehicles and an autonomous car. This result was a continuation of leading studies conducted by Oyler et al. (2016) and Li et al. (2016). Similar to Musavi et al. (2016) and Musavi et al. (2018), this contribution is also the first in the automotive literature where a large number of decision making drivers are simultaneously modeled using a game theoretical modeling approach. Recently, an extended version of this work, which covers a larger class of interaction scenarios, with the help of a road traffic simulation on a 5-lane highway, is presented in Albaba et al. (2019), where validation studies by processing real traffic data, which is provided in Colyar and Halkias (2007), is conducted. In this paper, we first present the basic components of the modeling approach discussed above and then provide the examples of CPHS framework creation, using GT and RL, for engineering systems that can be employed in predicting the outcomes of having several humans, automation and physical systems interact with each other in extended periods of time. These ideas exist in the literature in a fragmented manner, and by elucidating them in an aggregated form here, we provide a concise single source. The deliberations in this article on employing game theory and reinforcement learning for building CPHS modeling frameworks should benefit control practitioners whose goal is to obtain models of engineering systems where humans are active players. Finally, we discuss ongoing and future work about the topic, together with open problems that may provide several different research opportunities for the CPHS community. 3 The organization of the paper is as follows: In Section 2, we review the existing work. In Section 3, we explain the basic building blocks of the game theoretical model. In Section 4, we show how these blocks are combined together to form the overall modeling approach. In Sections 5 and 6, we present the exploitation of the approach to create models of two different engineering systems containing multiple human interactions. We provide a computational complexity analysis in Section 7. In Section 8 and 9 we discuss ongoing and future work, and related open problems, respectively. Finally, we provide a summary of the article in Section 10. 2. Review of existing work In this section, we review research activities looking for the answer to this question: "How can we predict the outcomes of engineering scenarios involving a cyber-physical human system with not one but several human elements?". The literature revolving around this salient question, either partially or completely, covers a wide range of engineering realms, which is hard to exhaustively examine in the limits of this article. Since, up until now, there has been two main studies addressing this question by exploiting the game theoretical modeling approach elaborated in this paper, we will focus on areas that are in the scope of these two research efforts: Unmanned aircraft system integration into the national airspace and road transportation. 2.1. Unmanned aircraft systems integration into the national airspace Although unmanned aircraft systems (UAS) has been attracting increasingly more at- tention, we have not yet witnessed the maturation of the civil markets. One of the main reasons for this underutilized potential can be attributed to the lack of regular access to the National Airspace System (NAS) (Dalamagkidis et al., 2008). Due to the well-justified risk- averse nature of the aviation industry, advances in developing rules and procedures for UAS integration into NAS are progressing relatively slowly, which results in UAS flying mainly in restricted airspace. Until it is clearly assured that UAS will not pose a danger to the existing air traffic and thus their integration is proven to be safe, routine access to NAS will not be realized (FAA, 2013; European RPAS Steering Group, 2013). There are many studies that address the problem of UAS integration into the airspace in terms of providing methods and tools to ensure safety. Ding et al. (2016) proposes an autonomous decision making system for UAVs for determining a safe landing site in the case of an anomaly. For UAV platoons, controller design frameworks are presented in Chen et al. (2015, 2016, 2017b), where col- lisions are eliminated and target states are reached. In Chen et al. (2017a), collision-free trajectories are designed for large-scale multi-UAV systems in the presence of disturbances in vehicle dynamics, and the method is demonstrated using up to 200 UAVs, in simulation environment. UAS integration still remains to be a challenge (Melnyk, 2019) and since we don't have the necessary experience to evaluate the effects of integration and there is not enough data yet, the only way to predict the outcomes of adding UAS into the airspace is conducting careful simulations (DeGarmo, 2004). To obtain reliable simulation results, a high-fidelity 4 model of the airspace in the presence of manned and unmanned aircraft together with their interactions need to be obtained. A typical approach in the airspace traffic modeling literature includes the assumption of pilots always following an ideal behavior pattern without any deviations. This is an unreal- istic assumption since, as discussed earlier, a representative model should allow probabilistic human behavior. For example, it is well-documented that pilots may ignore controller's commands or do not obey traffic collision avoidance system (TCAS) resolution advisories during emergency situations (Pritchett, 2010). In addition, it is reported that only 13% of pilot reactions agreed with the pilot model, which predicts a deterministic behavior, used for establishing TCAS algorithms (Kuchar and Drumm, 2007; Lee and Wolpert, 2011). One of the main issues that needs to be solved for a safe integration is the development of a dependable sense-and-avoid (SAA) technology for the unmanned aircraft systems. It is not possible to mature this technology without testing its performance through simulations with reliable pilot models (Maki et al., 2012). There exist several SAA methods introduced in the literature, with validations conducted via simulations and experiments. Kuchar et al. (2004) carefully analyzed the utilization of TCAS as the SAA logic, and performed simulation tests using the aircraft encounter model developed by Kochenderfer et al. (2008). During the tests, pilot reactions were assumed to be known beforehand, based on the respective motions of the conflicting aircraft. In another study, Perez-Batlle et al. (2012) suggested maneuvers to solve UAS separation conflicts and tested these suggestions with simulations that contain pilots following the recommended maneuvers without any error. The perfor- mance of the SAA algorithm proposed in Florent et al. (2010) was assessed by simulations and experiments. For these assessments, the manned aircraft in separation conflict with the UAS were assumed to continue their motion unaffected, while the UAS were implementing the proposed SAA technique. Another example where predefined pilot action models are employed for evaluating various SAA algorithms can be found in Billingsley (2006). The modeling framework for UAS integration into NAS that is discussed in this article differentiates itself from the above mentioned environments by providing a platform where pilot actions are not pre-determined but obtained through a decision making process by satisfying a utility function, or a "happiness function", that reflects pilot preferences. Fur- thermore, with the proposed approach, several pilot-pilot and pilot-UAS interactions (180 of them) can be modeled in time-extended scenarios, in a probabilistic manner, with the help of the convergence of game theory and reinforcement learning. The details of this framework are provided in Section 5. 2.2. Road transportation We have reliable physical models of road vehicles that have high predictive power. How- ever, the modeling problem quickly becomes difficult to handle when the task is modeling the vehicle together with the driver operating it. Furthermore, if the desired outcome is a model for traffic containing several vehicles, both manned and unmanned, obtaining answers turns out to be a real challenge. The main obstacle in this matter in question is the lack of accurate human interaction models. 5 Valid human interaction models in road traffic may prove themselves useful for two main tasks: Creating accurate traffic simulators that can be used for initial testing and tuning of autonomous car control algorithms, and designing autonomous vehicle control systems based on human way of driving, which can improve the passengers' comfort by making them feel as if a human driver is in control (Carvalho et al., 2015). There are several successful driver modeling studies in the literature. Real traffic data is employed to obtain Hidden Markov Model based driver models in Lefevre et al. (2014) and Lefevre et al. (2015). A semiautonomous vehicle control architecture is proposed in Vasudevan et al. (2012) and Shia et al. (2014), where the driver model is obtained, through k-means clustering, and used to inform the controller that produces corrections for driver inputs. Logical, if-then-else commands form the driver decisions in a modeling framework created by Salvucci et al. (2001). A multi-agent simulator is employed to model lane changing in Hidas (2002). Lane changing behavior is modeled also in Kumar et al. (2013), where Bayesian filters and support vector machines are utilized to predict driver intent. There is another body of work on modeling drivers as controllers in a closed loop control system, examples of which can be found in Hess and Modjtahedzadeh (1990); Sharp et al. (2000); Treiber et al. (2000); Salvucci and Gray (2004) and Ungoren and Peng (2005). A more recent example of feedback controller type driver modeling can be seen in Wakitani et al. (2018). The proposed modeling framework for road traffic in this article has the following dis- tinctions compared to earlier work: 1) The driver interaction models are scalable and a traffic scenario consisting of several vehicles interacting with each other can be modeled using a game theoretical approach; 2) driver behavior is obtained through employing a de- cision making process, instead of assuming a preset driver action model that is a function of time or states; 3) driver models are simultaneously strategic, meaning that they consider other intelligent agents (drivers, unmanned vehicles) possible actions and produce a response accordingly, based on a utility function representing their priorities. This is in contrast to modeling a single driver as a decision maker while assigning pre-determined trajectory pro- files for the others. All the listed advantages are realized through the game theoretical modeling approach discussed in this article. There are other game theoretical driving mod- eling approaches reported in the literature such as the ones proposed by Yoo and Langari (2012) and Yoo and Langari (2013), where driver interactions are successfully modeled. On the other hand, unlike the method discussed in this article, these studies do not consider a time-extended scenario. Dextreit and Kolmanovsky (2014) also consider a game theoretical approach to model the interactions between the driver and the ego vehicle's power train. By penalizing the vehicle-driver system features of fuel consumption, emissions, battery state and operating conditions, this approach is demonstrated, via experiments, to perform better than the baseline controller, in terms of these system features, while still providing good drivability. However, the utilized game theoretical approach, namely Stackelberg solution, quickly becomes computationally intractable as the number of intelligent agents in the game increases, unlike the hierarchical game theoretical approach used in the method proposed in this article. The details of the proposed method are provided in Section 6. 6 3. The basics In this section, we present the fundamental building blocks of the game theoretical mod- eling framework discussed in this article. These blocks are game theory and reinforcement learning. Below, we explain these pieces with a relatively narrow scope, using a semi-formal language to allow easy access, and with enough details that will enable understanding of the basic ideas necessary to grasp the following sections. There are several sources in the literature that can be used for formal introductions to these topics, such as Fudenberg and Tirole (1991) and Camerer (2011) for game theory, and Wiering and van Otterlo (2012) and Sutton and Barto (2018) for reinforcement learning. 3.1. Game theory Game theory studies the interactions between strategic agents. A strategic agent is one that considers other agents' possible actions and their effects on the game while making his or her own decisions. The theory makes predictions about the outcomes of these interactions using precise mathematics. Players in a game theoretical setting refer to the entities who can effect the game by their moves (or actions, or decisions). Strategy of a player defines the procedure based on which a player chooses his or her actions. A solution concept is a well established set of rules that are used to predict how a game will unfold. A Nash equilibrium is a solution concept, defined similarly to the equilibrium in system dynamics: When players have no incentive to deviate from their selected actions, the game is said to be in Nash equilibrium. This means that in Nash equilibrium, players choose their best actions against each others' actions. A typical example where Nash equilibrium can be observed is a game called the Prisoner's Dilemma. In this game, there are two prisoners, Prisoner A and Prisoner B. They are put in separate rooms so that they can not communicate. Both of them are provided with the following information: If Prisoner A confesses the crime, he will be released provided that Prisoner B denies the crime, which will cause Prisoner B serve 10 years in prison. Similarly, if Prisoner B confesses, he will be released provided that Prisoner A denies the crime, which will put Prisoner A in prison for 10 years. If they both deny, each will serve 3 years. If they both confess, each will serve 5 years in prison. We can represent this game in matrix form as shown in Figure 1, where players' payoffs amount to the negative of the years to be served in prison. It is seen that, although resulting in a low overall payoff, (Confess, Confess) choice is the only Nash equilibrium since no player wants to change their move once they are in this state. It is important to note that Nash equilibrium is not always unique and there may be more than one Nash equilibrium depending on the game. There are other equilibrium concepts such as quantal response equilibrium where instead of giving the best response to other players' actions, the players choose a probability dis- tribution over their action space where actions with higher expected payoffs have higher probability of being played. Not all solution concepts predict an equilibrium. For example, level-k thinking is a non- equilibrium game theoretical model of strategic interactions, which assigns different levels of reasoning for players (Stahl and Wilson, 1995; Costa-Gomes et al., 2009). In this model, the 7 Figure 1: Prisoner's Dilemma lowest level of reasoning is level-0, which represents non-strategic thinking, simply meaning that the players who reason at this level have a strategy that does not take into account other players' possible actions. A level-1 player, on the other hand, takes the best action assuming that his or her opponents are level-0 players. Similarly, a level-k player responds best to his belief that the other players are reasoning at level-(k-1). Therefore, this model assumes an iterated best response (Crawford, 2008). Experimental results presented in Camerer (2011) corroborate the predictions of this solution concept with varying success. An elementary example for the level-k reasoning model can be given considering two people, named Diana and Ritchie, walking towards each other, along a collision path, in a university corridor. If Ritchie decides to continue walking without considering Diana's possible actions, Ritchie can be considered as a level-0, non-strategic thinker. On the other hand, if Diana believes that Ritchie is a level-0 thinker and therefore decides that the best action is stepping right, then Diana can be modeled as a level-1 player. Although there exists experimental evidence for level-k predictions, this simple example presents a difficulty in this approach: The players' beliefs of others may be wrong. Another issue is that even if the players correctly estimate their opponents' levels, the behavior patterns coded by different levels are sensitive to the selection of the level-0 algorithm. That's why level-0 is often referred as the anchoring level. Although these "problems" are real, they may actually be considered as the strengths of this method when it comes to modeling humans. As discussed in the Introduction section, in CPHS with multiple humans, to obtain reasonable predictions, we need models that do not foresee optimal behavior all the time. Another perspective is that level-k thinking represents the interactions between intelligent agents that do not have a long interaction history, therefore it is actually expected that their initial assumptions about each others' strategies may not fully reflect reality. This also helps obtain non-optimal human reactions that are providing best responses to their beliefs about the outside world. 3.2. Reinforcement learning Reinforcement learning (RL) can be defined as a mathematical representation of learning through reward and punishment. To clarify this definition, and to be able explain the RL algorithm used in the CPHS modeling framework discussed in this paper, we first need to identify main elements of RL and make certain definitions that hold true for almost all RL methods. In RL, there exists an agent capable of exerting actions that can change the state 8 ConfessDenyConfess-5, -50, -10Deny-10, 0-3, -3PrisonerBPrisonerAof the environment where the agent operates. The RL problem can be defined as finding the optimal set of action sequence for an agent to achieve a given goal defined as a function of the environment states, through interaction with the said environment. For example, the problem may be making a mobile robot (agent) to go from point A to point B (goal) in a 10 by 10 grid-world with obstacles (environment) by deciding whether to move left, right, forward or backward (actions), in every step of the way. In this scenario, the state can be defined as the grid location the robot is occupying. RL uses the idea of a reward function to describe the preferences of the agent (or the designer of the agent) while its learning to achieve a predetermined goal. In the mobile robot example given above, the reward can simply be defined as zero if the robot is at the goal state of point B, and a fixed negative number, otherwise. A policy is defined as a probabilistic map from states to actions. The task of the RL algorithm is to find a policy that will make the agent maximize a cumulative discounted reward, during the time of its operation. One way to express the cumulative reward is given as C = γtrt, (1) where γ is the discount factor and r is the reward obtained in every step t. There are various RL techniques proposed to discover action sequences that will attain the goal of maximizing (1). Almost all of these different methods are based on estimating value functions, which can be regarded as the value of being in a certain state, based on the policy being implemented. A value function can be given as V π(s) = Eπ γkrt+kst = s (2) where π represents the implemented policy and s represents state. A similar function, the estimation of which also characterizes the RL method, stands for the value of taking a certain action a, in a given state s. This function is defined as Qπ(s, a) = Eπ γkrt+kst = s, at = a . (3) k=0 The aim of RL is to find the optimum policy π∗ that will maximize (optimize) the value function. The optimal value function is written as V π∗ and all its state values are larger than or equal to that of all other value functions that are created by policies different than the (s) ≥ V π(s),∀π,∀s. Similarly, we can write optimal policy. This can be represented as V π∗ (s, a) ≥ Qπ(s, a),∀π,∀(s, a), for the optimal action value function. Once the optimal Qπ∗ action value function is found, the policy π∗(s) = arg max a Qπ∗ (s, a) (4) can be used to select the best action in each state. The answer to the question "how to find the optimal value function" determines the type of RL algorithm. The process of finding the 9 ∞(cid:88) t=0 (cid:40) ∞(cid:88) k=0 (cid:40) ∞(cid:88) (cid:41) (cid:41) optimal policy is sometimes called training. Figure 2 depicts the general training process of RL. In the figure, the agent observes the states and produces an action based on the observed states. This action influences the environment and results in a new set of states. These new states are evaluated by the reward function and a reward signal is formed. The agent uses this signal to update the policy that is being trained and the next cycle starts with the new action. Figure 2: Reinforcement learning process In the following subsections, we explain one of the most basic RL algorithms, Q-learning, and then present two other RL methods that are utilized in the game theoretic modeling framework elaborated in this article. Remark 1. The RL algorithms used in this paper are used to obtain human response poli- cies, which can then be used to quantitatively analyze scenarios where humans are involved, in the simulation environment. Therefore, there is no physical interaction with the environment during learning. 3.2.1. Q-learning One of the RL methods that played a significant role for the success of RL is Q-learning (Watkins, 1989). In Q-learning, an incremental estimate of the optimal action value function is realized using the update rule Qk+1(st, at) = Qk(st, at) + α rt + γ max a Qk(st+1, a) − Qk(st, at) , (5) (cid:18) (cid:19) where α is the step size and γ is the forgetting factor. It is noted that Q-learning is indifferent to the policy that the agent uses during training to explore the environment, which means moving from one state to the other . This type of RL algorithms, where exploration and value function updates (or policy updates) are independent, are called off policy methods. It can be shown that the learned action value function Q converges to the optimal action-value function Q∗ with probability 1, if all state-action pairs (s, a) continue to be visited during training, and if the number of these visits converges to infinity. Convergence is achieved exactly if 10 AgentEnvironmentactionstatesReward Functionrewardk=1α2 k=1αk = ∞, Σ∞ the step size parameter obeys a variant of the stochastic approximation conditions, given k ≤ ∞. For a constant step size parameter, convergence is achieved as Σ∞ in the mean if 0 ≤ α ≤ 1. It is noted that continuously visiting state-action pairs is a requirement for any algorithm that is developed for the general case, where a model for the environment is not provided, to achieve convergence to the optimal solution.(Sutton and Barto, 1998). The discount factor γ determines how valuable the future rewards are. For instance, when γ is 0, agent only tries to maximize the immediate rewards. As this value approaches to 1, the agent becomes more far-sighted. 3.2.2. Neural fitted Q-learning Instead of keeping a table of Q values, some RL methods use compact structures that provide approximate Q values. This approach is especially useful when the state space is very large. Neural Networks (NN) are one of the effective tools used to store the Q-values in a compact manner, thanks to their universal approximation property. Unlike the conventional Q-learning method explained above, the state-action values are not stored in a table but instead, computed as the output of a function constructed by the specific NN structure: If a state-action pair is fed to the NN as the input, the corresponding approximate Q-value can be obtained as the NN output. The training of the NN can be achieved by first defining an error function reflecting the difference between the current and the target Q-values, and then minimizing this function by back-propagation. Although calculating the Q-values using a NN provides an effective method, it can fail, either completely or by requiring impractical convergence times, due to the global representation mechanism (Riedmiller, 2005): During the training process, NN weights are updated after the introduction of each individual state- action pair, which also effects the Q-values of other pairs. This may nullify the previous training gains in other regions. On the other hand, the global representation enables the generalization power of NNs by assigning similar Q-values to similar state-action pairs, and thus eliminating the need to train the NN for every possible pair. Therefore, a method is needed that can both exploit this property and eliminate its detrimental effects. Neural Fitted Q-learning, proposed by Riedmiller (2005), achieves to both employ the generalization power of NNs and prevent its potentially harmful effects by storing previous experiences in the form of 3-tuples, (s, a, s(cid:48)), in which s is the original state, a is the action taken and s(cid:48) is the reached state, and reusing these experiences whenever an update is performed after the introduction of a new data point. Calling a collection of these experiences as set E, the NFQ method is given in Algorithm 1. In the implementation of NFQ learning, it is advised that instead of a random collection of experiences, greedy search, using available Q-values, and random exploration are used together. NFQ contains two types of hyper-parameters: parameter of the Q-learning, the forgetting factor γ, and the parameters used for NN training. The selection of the Q-learning hyper- parameter γ is explained in the previous section. For NN training, the resilient propagation algorithm proposed in Riedmiller and Braun (1993), which works well for batch learning, is suggested since NFQ is also based on batch learning. It is discussed in Riedmiller and Braun (1993) that different values of hyperparameters do not effect the performance of the algorithm dramatically and therefore some predefined values suggested in the paper can be 11 Algorithm 1 NFQ Algorithm 1: k = 0 2: Initialize the neural network 3: while k < N do 4: Generate experience set G = {(inputi, targeti), i = 1, 2, ..., #E} where inputi = si, ai, where si is the state and ai is the action of ith experience, targeti = ri + γ mina Qk(s(cid:48)i, a), where ri is the transition cost and γ mina Qk(s(cid:48)i, a) i=1(Qk(si, ai)−targeti)2, where n refers to the experience is the weighted expected maximum path reward for the next state s(cid:48)i. Calculate the batch error as Σn set size. Train the network to minimize the batch error, using resilient back-propagation and obtain Qk+1 . k+ = 1 5: 6: 7: 8: 9: 10: end while used for most of the problems. In many NN training software packages (Abadi et al., 2015; Paszke et al., 2017) these values are already set as default so no further tuning is necessary. 3.2.3. Jaakkola reinforcement learning An agent being trained by RL uses the available information from the environment. This information is generally called the "state" of the environment. When the state of the environ- ment contains all relevant information about the current and the past interaction dynamics between the agent and the environment, this state is said to have the Markov property (Sut- ton and Barto, 2018). A learning task involving interactions with an environment that has Markov property is called a Markov Decision Process (MDP). More specifically, defining the probability of transitioning from state "s" to state "s(cid:48)" and obtaining a reward "r", given an action "a" as P (s(cid:48), rs, a), if this probability depends only on "s" and "a" but not on earlier states and actions, this learning task is called an MDP. In almost all RL methods that have convergence guarantees, the underlying dynamics is assumed to be an MDP. In the aerospace and automotive application scenarios that are investigated in this paper, although the underlying dynamics are MDPs, the agents can realistically observe only a portion of the available states. Therefore, from the agents' point of view, the tasks are Partially Observable Markov Decision Processes (POMDP). Jaakkola reinforcement learning algorithm (Jaakkola et al., 1994) is developed specifically for systems that can be modeled as POMDPs and therefore is a suitable RL method to be employed in the learning tasks that are discussed in this work. In Jaakkola algorithm, along with Q-function, the value function, V , is also used. At the beginning of the Jaakkola Algorithm, Q values are set to zero for each state-action pair. Moreover, for each state, probability distribution of actions is set to a uniform distribution. Then, for each iteration 12 (s, a, s(cid:48)), Q and V values are updated according to following equations: (cid:18) (cid:18) (cid:18) (cid:18) (cid:19) (cid:19) (cid:19) (cid:19) 1 − χt(s, a) Kt(s, a) 1 − χt(s) Kt(a) 1 − χt(s, a) Kt(s, a) 1 − χt(s) Kt(s) βt(s, a) = βt(s) = Qt(s, a) = Vt(s) = γtβt−1(s, a) + χt(s, a) Kt(s, a) γtβt−1(s) + χt(s) Kt(s) Qt−1(s, a) + βt(s, a)(Rt − R) Vt−1(s) + βt(s)(Rt − R) (6) where, s is the state, a is the action and t is the time step. Moreover, χt(s, a) (χt(s)) is equal to 1 if the given state-action pair (state) is visited, and 0 otherwise; Kt(m, a) (Kt(s)) is the number of times the state-action pair (state) is visited; Rt is the reward in time step t; R is average reward and γt is the discount factor. After the calculation of Q and V functions, Jaakkola algorithm updates its trained policy π(as) using the update rule π(as) = (1 − )π(as) + π1(as), (7) where is the update rate, and π1(as), the policy that the trained policy is being changed towards, is a greedy-policy based on the calculated Q(s,a) values. In other words, π1(as)=1 if the action "a" has the highest Q-value in a given state "s". It can be shown (Jaakkola et al., 1994) that this policy update always increases the average reward, unless the condition [Q(s, a) − V (s)] > 0 max a (8) is satisfied. The algorithm increases the average reward until the condition (8) is false, which constitutes a local maximum. The Jaakkola algorithm consists of two hyper-parameters: the discount factor γ and the update rate . For convergence guarantees, γ should initially be selected as a number between zero and one, and should be scheduled in such a way that it converges to 1 in the limit. , on the other hand, should satisfy 0 ≤ ≤ 1. 4. Bringing the pieces together: An interplay between game theory and rein- forcement learning A schematic of the cyber-physical human system (CPHS) we are interested in, involving multiple human interactions, is depicted in Figure 3. This is a simplified diagram of the overall system, showing how humans' actions change the cyber-physical system states, which are observed by humans who produce the next set of actions and initiate the next cycle accordingly. It is noted that humans' actions are also affecting each other through the closed loop nature of the information flow. Another important take-away from this figure is that 13 human observations are not necessarily the same and an individual human agent does not have full state information. Observation i blocks, where i = 1, 2, ..., n, represent the human observation process. This process is limited and imperfect, therefore each human receives a noisy subset of the whole system state. Human goal blocks represent what each of the human agent is trying to accomplish, which can be driving from point A to point B in a road traffic scenario, or protecting his or her own aircraft from a hacker attack in a cyber-security scenario. Figure 3: Cyber-physical human system with multiple humans Obtaining a model for the system shown in Figure 3 requires considering the interaction of human actions. Humans are strategic thinkers and therefore if we want to build a model that represents reality, to the best of our ability, we need to find a way to have agent models that consider other agents' possible actions before making a move. As discussed in Section 3.1, this is what game theory is all about: Modeling the interactions between strategic agents. Therefore, we may utilize game theory to solve this problem. However, there is one complication in this approach. This is a system where the interaction may last for long periods of time, and therefore obtaining an equilibrium solution can be computationally intractable. If the number of agents get larger, as in several real world applications, even for short periods of interactions, the computational cost grows rapidly. One way around this problem is using the non-equilibrium game theoretical solution concept, level-k thinking, explained in Section 3.1, where human actions are predicted in an iterated manner, instead of being evaluated at the same time. This means that once the anchoring level, level-0, is selected, a level-1 human agent's behavior can be identified as the best response to all the other human actions that are determined by the level-0 policy. Similarly, once level-1 behavior is identified, all the agents in the system are assigned the level-1 policy except the 14 HumansCyber-Physical Systemhuman goal 1statesAgent naction 1action 2action nObservation 1Agent 2Agent 1Physical SystemsAutomationComputationCommunicationNetworkinghuman goal 2Observation 2human goalnObservationnone whose level-2 behavior is to be found. Therefore, to predict the policy of a level-k agent, all the rest of the agents' policies are set to level-(k-1), which effectively make them a part of the environment whose dynamics are known, and level-k policy is determined as the best response to the rest of the level-(k-1) policy actions (see Figure 4). This isolates the level-k policy as the single policy that needs to be computed. It is noted that in Figure 4, the actions of the level-(k-1) agents are not shown and the states still belong to the CPS. These details are omitted for brevity. Figure 4: Obtaining level-k policy Once the difficulty of high computational cost due to multiple decision makers is solved using level-k thinking, the problem reduces to estimating the optimal action sequence of an intelligent agent in a given environment. This problem can be stated as a reinforcement learning (RL) problem by properly defining the states, actions, the environment and the reward function. (Compare Figure 2 and Figure 4.) Therefore, using one of the suitable RL methods described in Section 3.2, modeling of the CPHS with multiple human interactions can be completed. The type of the preferred RL algorithm depends heavily on the observation and action spaces of the agents, which in turn depend on the engineering domain of the CPHS operation. The general algorithm used in obtaining the agent policies, demonstrating the interplay between the game theory and RL, is provided in Algorithm 2, where k is the maximum desired level. It is noted that, a level-k agent can be made to best-respond to all lower levels, k=0, 1, .. (k-1), instead of only best-responding to level-(k-1), which may or may not be desired depending on the application, by including all lower levels in the agent's policy space. In the following sections, based on Musavi et al. (2016) and Albaba et al. (2019), we ex- plain how this interplay between RL and game theory is used to create modeling frameworks for two different engineering realms, where multiple (180 in one case and 125 in the other) human interactions are involved. 15 HumanEnvironmenthuman goalstatesactionObservationAgent level-kPhysical SystemsAutomationComputationCommunicationNetworkingAgent level-(k-1)Agent level-(k-1)Algorithm 2 Interplay between RL and Game Theory 1: Set i = 0 2: while i < k do 3: 4: Load the level-i policy Set cognition levels of all players in the environment other than the learning agent to level-i, i.e. set policies of players to level-i policy Place the learning agent in the initialized environment, in which all players are level-i Start the training of the learning agent using a reinforcement learning method, through which agent learns how to best respond to level-i players 5: 6: 7: Once the training is completed, learning agent becomes a level-(i+1) player 8: 9: 10: end while Save the policy of the learning agent as level-(i+1) policy i += 1 Remark 2. The methods discussed in this paper, which are used to model cyber-physical human systems, are used solely for modeling purposes. Therefore, they are not meant to be used in physical motion systems. Remark 3. There exist several successful cyber-physical system models in the literature. The bottleneck in driving models for cyber-physical human systems, is computing the multi- agent, multi-move decision making dynamics, which corresponds to obtaining the models in the green "Humans" block in Figure 3. Therefore, in this paper, to emphasize the power of the discussed game theoretical modeling approach in predicting human responses, we avoided employing complicated vehicle models. Remark 4. The cyber-physical human systems modeling method explained in this paper merges reinforcement learning and game theory. It is noted that the level-k approach is not the only game theoretical method that can be employed here. Theoretically, other game theo- retical methods can also be used. However, as explained in this section, thanks to the hierar- chical modeling approach inherent in the level-k reasoning, modeling multi-move, multi-agent scenarios with simultaneous decision makers can be handled in a computationally tractable manner, which makes the level-k method a suitable candidate for the scenarios investigated in this paper. 5. Hybrid airspace modeling With hybrid airspace we refer to an airspace where manned and unmanned aircraft coex- ist. As discussed at length in Section 2.1, obtaining hybrid airspace models is a necessity for successful integration of unmanned aircraft systems (UAS) into the National Airspace Sys- tem (NAS). In this section, we present how the game theoretic modeling framework discussed here is used to realize the modeling of these systems. 16 Figure 5: A hybrid airspace scenario where red squares represent manned aircraft and the cyan square is used to indicate an unmanned aircraft (UA). The yellow circles in the 600 km x 300 km airspace are the waypoints assigned to the UA. (Musavi et al. (2016), reprinted with permission of the American Institute of Aeronautics and Astronautics, Inc.) Figure 5 shows a hybrid airspace scenario where a UAS (cyan) is assigned to follow certain waypoints (yellow) in a crowded airspace filled with manned aircraft (red). We are interested in how the overall system will evolve in time. To be able to predict the possible outcomes of this scenario, we need to create a model that will capture the reaction dynamics of manned aircraft pilots in cases of separation conflicts. As discussed earlier, the game theoretical modeling approach produces policies, which are probabilistic maps from observations to actions, to represent human reactions. Therefore, to obtain these policies, we first need to clearly define the observation and action spaces and represent them in a way that is meaningful for the RL algorithm. Once these spaces are explicitly defined, pilot goals and preferences need to be expressed in form of a reward function. Furthermore, to train pilot policies in a simulation environment, we need to realize the motions of the aircraft, manned and unmanned, using their physical models. Finally, sense and avoid algorithms need to be integrated to UAS dynamics for collision avoidance. Below, we explain how these pieces are obtained and then assembled to create the overall hybrid airspace model. Furthermore, we discuss the validation studies conducted using data. 5.1. Observation and action spaces Self-separation concept, where the pilots (and crew) are responsible for keeping a safe distance from encountered traffic, is a procedure that is being explored for Next Generation (NextGen) Air Transportation System (Wing et al., 2013). One of the technologies that can make this possible is Automatic Dependent Surveillance Broadcast (ADSB), which provides 17 state information of the surrounding traffic, with a precision better than the radar (Kacem et al., 2018). In the above scenario, we assume that aircraft are equipped with this technology. To factor in cognitive limitations, the observation space is set as a pie shaped region formed by two circles sharing a common center, which is depicted in Figure 6, on the left. The inner circle radius is taken as 1 nmi and the outer circle radius is taken as 5 nmi, reflecting the standard separation requirements for manned aircraft (Perez-Batlle et al., 2012). These circles are then divided into 3 slices resulting in 6 different observation regions. If an intruder approaches one of these regions, the region that is being approached is coded with number 1, 2, 3 or 4, depending on the approach angle, while the rest obtains a zero value. For example, in Figure 6, Aircraft B is approaching Aircraft A's observation space with a 120 degree angle which makes the corresponding entry for the approached region take the value of 2, while the rest are assigned 0. This is shown under the state components section on the rightmost side of the figure. The information of the best immediate actions that will make the aircraft approach its predetermined trajectory (Best Trajectory Action, BTA) and to its predetermined destination (Best Destination Action, BDA), together with the pilot's Previous Action (PA) are also coded into the state components. The pilot action space consists of 3 actions: 45 degrees left, straight and 45 degrees right, and these actions are coded as 0, 1 and 2, respectively, in the state components. It is noted that these actions go through aircraft dynamics to produce the actual aircraft motion. During training, the state components entries are continuously updated, based on the explanations given above, and used in the RL algorithm. With the aforementioned observation and action spaces, the RL algorithm need to assign Q values to 56 × 33 × 3 = 1, 265, 625 state-action pairs. Remark 5. The actions described in this section are not the actions of the aircraft but the decision inputs to the real continuous aircraft dynamics, whose outputs are the aircraft actions. The aircraft dynamics is explained in Section 5.3. 5.2. Reward function For the RL algorithm to evaluate the desirability of selecting a certain action given a state, a reward function that specifies pilot preferences and goals is required. For the UAS integration into NAS scenario, the reward function r = −ω1C − ω2S − ω3I + ω4D + ω5P − ω6E (9) is used, where C and S represent the number of aircraft (manned or unmanned) occupying a space within the collision and separation regions of the ego aircraft, respectively. The definition of these regions can be found at Planning et al. (2007) and Perez-Batlle et al. (2012). I is binary and gets the values of 1 or 0 depending on whether the ego aircraft is approaching or distancing from the intruder. D and P are the degrees of approach (or distancing) by the ego aircraft to its destination and to its defined trajectory, respectively, normalized by the distance covered in one time step. To accommodate the tendency of humans to minimize energy consumption as much as possible, E is introduced to represent effort, taking the values of 1 or 0, depending on whether the pilot takes a new action or not, respectively. 18 Figure 6: Observation space. (Musavi et al. (2016), reprinted with permission of the American Institute of Aeronautics and Astronautics, Inc.) Remark 6. The selection of the reward function plays a crucial role in the performance of the reinforcement learning algorithm. A formal procedure for the determination of a proper reward function is an open area of research. However, addressing context-dependent problems that reflect the agent's preferences, while keeping the function simple is the general approach for determining the reward functions. One problem that needs to be solved in this section is avoiding collisions and separation violations, which is addressed by the terms C, S and I in (9). Another problem is reaching a given destination while following a predetermined trajectory, which is addressed by the terms D and P . Finally, the problem of energy con- servation is addressed by the term E. An alternative to (9) could be a reward function that uses continuous variables for all the terms. This would increase the resolution but could unnecessarily complicate the function. In certain problems such as learning a policy from an expert, the reward function selection can be done using a more systematic approach by solving an inverse reinforcement learning problem (Sutton and Barto, 2018). 5.3. Physical models of manned and unmanned aircraft As explained above, aircraft are controlled by pilots' commands of heading angle changes (±45 degrees). Using the standard turn rate of 3 deg/s angular velocity (Nancy, 2016), we model the aircraft turning motion with a first order dynamics having a time constant of 10 seconds. The related differential equation is Ψ = −0.1(Ψ − Ψd), (10) 19 Agent AAgent B000020121Observation Space & Approach AngleBDABTAPA120°Starting PointDestinationState Componentswhere Ψ and Ψd represent the current and the desired heading angles, respectively. The manned aircraft are assumed to be flying in en route phase with a constant velocity v, having the x and y coordinate components and vx = v sin Ψ vy = v cos Ψ. (11) (12) Unmanned aircraft are assumed to fly autonomously while avoiding collisions using an on- board sense and avoid (SAA) algorithm. This algorithm commands velocity vector changes if an intruder is detected. Using a 1 second time constant (Mujumdar and Padhi, 2011) and a first order dynamics, the velocity vector dynamics are modeled as (cid:126)v = −((cid:126)v − (cid:126)vd), (13) where (cid:126)vd is the desired velocity vector. Two SAA algorithms are employed in the simula- tion environment with different velocity vectoring properties. Both of them first detects a probable conflict by forecasting the future trajectories of both the ego and intruder aircraft, and checking whether the minimum calculated distance, R, between them is smaller than a predetermined threshold value. If a conflict is detected, a desired velocity vector command, (cid:126)vd, is produced. One of the SAA algorithms (SAA1) proposed by Fasano et al. (2008) issues the velocity command (cid:126)vd = (cid:126)vei cos(η − ξ) sin(ξ) sin(η) (cid:126)vei (cid:126)vei − sin(η − ξ) (cid:126)r (cid:126)r + (cid:126)vi , (14) (cid:32) (cid:18) (cid:19) (cid:33) in cases of conflict, where the ego unmanned aircraft velocity command and intruder aircraft velocity are represented by (cid:126)vd and (cid:126)vi, respectively. Similarly, (cid:126)r and (cid:126)vei refer, respectively, to the relative position and velocity between these two aircraft, and η is the angle between these two vectors. Angle ξ is determined as ξ = sin−1(R/(cid:126)r). The other SAA algorithm (SAA2) used in simulations provides the velocity command (cid:126)vd = −(cid:126)ve( (cid:126)r0·(cid:126)vei (cid:12)(cid:12)(cid:12)(cid:12) − (cid:126)ve( (cid:126)r0·(cid:126)vei (cid:126)vei ) − (R − (cid:126)rm) (cid:126)rm(cid:126)rm (cid:126)vei ) − (R − (cid:126)rm) (cid:126)rm(cid:126)rm (cid:12)(cid:12)(cid:12)(cid:12) (15) to resolve conflicts, where (cid:126)r0 and (cid:126)rm stand for the initial and minimum relative positions between the ego unmanned aircraft and the intruder (Mujumdar and Padhi, 2011). 5.4. Simulation results After the necessary pieces explained above are set to create the overall hybrid airspace model, pilot reactions are obtained using the interplay, explained in Section 4, between the level-k game theoretical approach (Section 3.1) and Jaakkola reinforcement learning method 20 Algorithm 3 Interplay between RL and Game Theory in NAS 1: i = 0 2: while i < k (k is the maximum cognition level) do 3: 4: Load the level-i policy Set the policies of all the pilots in the scenario, other than the ego pilot (the pilot being trained), to level-i Start the training of the ego pilot using reinforcement learning, through which the pilot learns how to best respond to level-i pilots 5: 6: Once the training is completed, the ego pilot becomes a level-(i + 1) pilot 7: 8: 9: end while Save the policy of the ego pilot as level-(i + 1) policy i+ = 1 (Section 3.2.3). The algorithm that details this process is given in Algorithm 3. Once the pilot reaction dynamics are created, quantitative analyses on the integration scenario are conducted using Monte Carlo simulations. Since one of the bottlenecks of UAS integration into NAS is the maturation of SAA algorithms, we present how the modeling method dis- cussed in this article can be used to conduct comparative quantitative analysis on various aspects of different SAA methods. Specifically, we analyze the effect of distance horizon and time horizon variables used in the SAA development. In this article, we define the former as the scan radius of the algorithm, and the latter as the amount of projection time used to detect a conflict. Figure 7 and Figure 8 depict the effect of varying distance and time horizons on tra- jectory deviations, flight times and number of separation violations, for SAA1 and SAA2, respectively. A few conclusions, not in increasing or decreasing importance, can be drawn from these results. First, although time horizon makes a significant effect on both safety (separation violations) and performance (trajectory deviations and flight times) measures for SAA1, the unmanned aircraft equipped with SAA2 is not affected as much with the vari- ation of this parameter. Second, the effect of the distance horizon on separation violation numbers levels out quickly for SAA1, while SAA2 keeps showing lesser and lesser violation rates, although SAA1 provides a safer traffic in the same parameter range. Third, UAS trajectory deviations are generally smaller when SAA2 is employed. Fourth, an interesting but expected phenomenon is observed: the trajectory deviations of manned and unmanned aircraft have negative correlation. As the UAS trajectory deviations increase, showing that SAA is doing more work for keeping a safe distance, manned aircraft pilots spend less effort and have less trajectory deviations. All these results obtained in a simulation environment with real-time decision making pilots can be used in testing and tuning of SAA algorithms, as well as making a quantitative comparison between different approaches. 21 (a) manned aircraft trajectory deviation (b) UAS trajectory deviation (c) UAS flight time (d) number of separation violations Figure 7: When SAA1 is employed in HAS, effects of varying time and distance horizons on separation viola- tions, flight times and trajectory deviations are presented. (Musavi et al. (2016), reprinted with permission of the American Institute of Aeronautics and Astronautics, Inc.) 22 . (a) manned aircraft trajectory deviation (b) UAS trajectory deviation (c) UAS flight time (d) number of separation violations Figure 8: Safety vs. performance in HAS, when SAA2 is employed is depicted by showing the changes in trajectory deviations, flight times and separation violations with the changes in distance and time horizons. (Musavi et al. (2016), reprinted with permission of the American Institute of Aeronautics and Astronautics, Inc.) 23 5.5. Validation Although we do not have enough, if at all, UAS integration data to test against the predictions of the game theoretical modeling approach, the resulting models created by the method need to be validated using other means, to achieve at least a minimum level of credibility. Several validation methods are introduced in the literature, such as face validity, comparison with validated models, historical data validation, parameter sensitivity analysis and predictive validation (MITRE Corp., 2014). Among these, the ones that are based on data are the most effective validation methods. It is noted that when the data becomes available, one needs to test if relevant statistics between the data and the model are matching. For example, average aircraft trajectory deviation and the average number of separation violations can be compared. Furthermore, during individual encounters, pilot decisions and minimum distances between UAS and manned aircraft, predicted by the model and obtained from the data can be investigated to see whether or not they show similar characteristics. It is noted that the game theoretical modeling approach has enough degrees of freedom, such as reward function terms and weights, to be able to be modified, in case some discrepancies between data and the model are observed when the data becomes available in the future. This is an important feature that needs to be found in predictive models (Law, 2008). Since currently data is not available, we discuss two methods that can be used without data: face validation and comparison with a validated model. Face validation evaluates two aspects of the model: the logic behind the modeling method, and input-output relationships of the model. Two main approaches used in the proposed model, namely reinforcement learning and game theory, are well-established fields that showed promise in modeling real-life behavior. Experimental backing of the level-k game theory is also provided in relevant references discussed in the review of existing work sections in this article. Also, the logic behind the selection of reward function terms is explained in earlier sections. We also discussed above the input-output relationships when we explained the effects of different parameters on the airspace scenario with figures, and showed that the results are reasonable. Sample trajectories of a data-validated model developed by the Lincoln Laboratory (Kochenderfer et al., 2008) is available online in two text files that are open to public: cor ac1.txt and cor ac2.txt. A comparison of one of the available trajectories (encounter 3) with the game theoretical model prediction is provided in Figure 9. In the figure, start- ing points of aircraft are indicated by thick dots. It is seen that pilot decisions, as well as minimum distance between aircraft during the encounter are similar. Further trajectory comparisons between the game theoretical model and this data-validated model can be found in Musavi et al. (2016). 6. Road traffic modeling Similar to the unmanned airspace system (UAS) integration study, to obtain the model of the road traffic, we need the physical models of the cars, driver observation and action spaces, and a reward function reflecting the goals and preferences of the drivers. Below, we 24 (a) Trajectories created by the validated model. (b) Trajectories created by the proposed model. (c) Separation distances for each model. Figure 9: Comparison of the trajectories created by the validated model and the game theoretical modeling approach for sample encounter number 3. (Musavi et al. (2016), reprinted with permission of the American Institute of Aeronautics and Astronautics, Inc.) explain these components and also provide validation of the resultant overall model with traffic data. 6.1. Driver observation and action spaces Several different traffic scenarios can be modeled via the proposed approach. To be able to test the results with data, we developed the model of a 5-lane highway, similar to the US101 Hollywood Freeway, whose raw data is available at Colyar and Halkias (2007). In this scenario, the human drivers are assumed to be observing (or being able to process from all available data), his or her immediate neighboring lane cars (front left, front right, rear left, rear right) and the car in front. Figure 10 presents a snapshot of a typical ego vehicle (red) motion, where the driver can observe the surrounding 5 cars. As in the UAS integration scenario, the observation space is quantized and the distances are coded as nominal, close and far. It is noted that this quantization introduces noise and uncertainty to driver obser- vations since instead of the exact location of neighboring cars, only a certain region occupied by the car is known. The importance of introducing noise and uncertainty to the human observations are discussed in Section 4 and represented by Observation blocks in Figure 3. To determine reasonable values for quantization, the distance distribution between cars are processed from the raw data provided in Colyar and Halkias (2007) and plotted in Figure 11. Based on this distribution, nominal range is defined to be between 11 m - 27 m, close is 25 Figure 10: Ego vehicle and surrounding traffic Figure 11: Distribution of distances to car in front. defined as smaller than 11 m and far is used for distances larger than 27 m. It is noted that nominal region consists of approximately half of the area under the curve depicted in the figure. Once the positions of the surrounding cars are defined, their relative motions against the ego car are expressed as stable, approaching and distancing. As a result, the observation space of the ego car consists of quantized positions and relative motions of the surrounding cars. As we use traffic data to obtain a meaningful observation space, part of the action space is also formed by considering the acceleration distribution, the plot of which is given in Figure 12, obtained by processing the same traffic data. According to this acceleration distribution, acceleration inputs of the drivers are categorized into 5 separate continuous sub-distributions: 1) Maintain, a normal distribution with zero mean and 0.075 standard deviation, 2) accelerate a uniform distribution between 0.5 m/s2 - 2.5 m/s2, 3) decelerate a uniform distribution between -0.5 m/s2 and -2.5 m/s2, 4) hard accelerate, a half normal 26 Figure 12: Acceleration distribution distribution with a 3.5 m/s2 mean and 0.3 m/s2 standard deviation, 5) hard decelerate, a half normal distribution with -3.5 m/s2 mean and 0.3 m/s2 standard deviation. Drivers sample these distributions to create an acceleration action, if they choose to. Two other action choices for the drivers are move to the left lane and move to the right lane, during both of which the velocity is assumed to remain constant. 6.2. Reward function The driver reward function used in the reinforcement learning (RL) algorithm is r = ω1C + ω2S + ω3D + ω4E, (16) where C is binary, taking values of -1 and 0, depending on whether a collision occurred or not; S is the normalized deviation of the ego vehicle speed from the mean speed of the traffic; D is related to the distance between the ego vehicle and the car in front, and takes the values of -1, 0 or 1 depending on whether the distance is close, nominal or far, respectively. This term reflects the driver preference of having as much headway as possible. E is the effort variable taking the value of 0 if the action is maintain; a value of -0.25, if the action is accelerate or decelerate; and -0.5 if it's hard accelerate or hard decelerate. E receives -1 if the driver changes lane. Remark 7. The terms of the reward function provided in (16) are determined based on the specific problems that need to be solved to obtain driver dynamics. The solution of these problems reflects the driver's preferences. Specific issues to be solved can be listed as col- lision avoidance (term C), performance (term S), safety (term D) and comfort or energy conservation (term E). Instead of discrete valued terms used in (16), continuously vary- ing terms could be employed. However, while this would reflect the driver preferences in a finer manner, it could introduce unnecessary complexity. A more systematic selection of the reward function can be conducted for problems such as imitation learning, where inverse reinforcement learning methods are applied (Sutton and Barto, 2018). 27 6.3. Physical models For the modeled traffic scenario, simple kinematic models are used to obtain vehicle motion. Once the driver command is received, velocity and position dynamics are obtained as x(t + ∆t) = x(t) + vx(t) ∗ ∆t + y(t + ∆t) = y(t) + vy(t) ∗ ∆t vx(t + ∆t) = vx(t) + a(t) ∗ ∆t, 1 2 a(t)∆t2 (17) where vx and vy are the velocity components in the x and y directions, x and y coordinates are the same as given in Figure 10, and ∆t is the simulation time step. It is noted that the dynamics are not discrete but continuous, and have to be approximated due to computer implementation. 6.4. Training, average rewards and entropy Driver policies with various levels are trained using the Jaakkola reinforcement learning algorithm explained in Section 3.2.3. During training, up to 125 vehicles are used in the 5-lane traffic scenario. Figures 13-15 show the time evolution of average rewards during the training of level-1, level-2 and level-3 policies, together with the average entropy of the probability distribution over actions. The entropy of a probability distribution is calculated as E = − n(cid:88) pi log(pi), (18) i=1 where n is the number of probability values pi. Since the drivers have 7 actions, we have n = 7. The entropy is highest when the distribution is uniform, which is the case in the beginning of training, and drops as the training progresses and the distribution gets away from uniform. However, as seen from the figures, although the average reward converges relatively fast, entropy continues to drop at a very slow rate. The reason for this is that there are several states with lower probability of being visited during training and after a certain driving pattern emerges, the effect of these rarely visited states becomes very small. This can also be observed from Figure 16, where the entropies of the two frequently visited states are shown to converge to very small values. 28 (a) Average reward per episode in level- 1 training (b) Entropy per episode in level-1 train- ing Figure 13: Level-1 training (a) Average reward per episode in level- 2 training (b) Entropy per episode in level-2 train- ing Figure 14: Level-2 training (a) Average reward per episode in level- 3 training (b) Entropy per episode in level-3 train- ing Figure 15: Level-3 training 29 (a) Entropy per visit of a state (b) Entropy per visit of a different state Figure 16: Entropy per visit plots of two randomly selected states. 6.5. Data validation The driver policies obtained using the game theoretical framework are compared with real traffic data provided by Colyar and Halkias (2007). The data provides acceleration values of the drivers at each measurement instant. This data is first processed to obtain the states the drivers are in and then to find the frequency of taken actions at each state. These state- action frequency distributions form the real driver policies. To compare the distributions modeled using the proposed game theoretical approach and the distributions obtained from data, we use the Kolmogorov-Smirnov (KS) Test for Discontinuous Distributions (Conover, 1972). In this test, if the unknown discrete probability distribution function is F (x) and the hypothesized distribution is H(x), the null hypothesis H0 is defined as Three test statistics used in the test are H0 : F (x) = H(x) f or all x. and D = supxHc(x) − Sn(x), D− = supx(Hc(x) − Sn(x)) D+ = supx(Sn(x) − Hc(x)), (19) (20) (21) (22) where Sn(x) and Hc(x) are the cumulative distribution functions (CDF) of the observed data and of the hypothesized distribution (model), respectively. The observed values of the test statistics D, D− and D+ are defined as d, d− and d+, respectively. The goal of the KS test is to calculate the probability of observing at least the value d for the test statistics D, which can be stated as P (D ≥ d), given that the null hypothesis is true. This is achieved by first calculating P (D+ ≥ d) and P (D− ≥ d), and then obtaining P (D ≥ d) as P (D ≥ d) = P (D+ ≥ d+) + P (D− ≥ d−). (23) 30 The details of obtaining P (D+ ≥ d) and P (D ≥ d−) are omitted here for brevity and can be found in (Conover, 1972). Once P (D ≥ d) is calculated using (23), the null hypothesis is rejected if P (D ≥ d) ≤ 0.05. Since this test provides meaningful results for distributions with non-zero entries, action probabilities that are lower than 0.01 are set to 0.01 with normalization, for both the real data and the driver model. Data validation is conducted individually for each driver: For the driver of interest, first, the action probability distributions for each visited state are computed. Second, these dis- tributions are compared with derived driver policies using the proposed modeling framework via KS test. Specifically, the distributions from the data are compared with level-1, level-2 and level-3 policies. Finally, the percentage of the states that can be successfully modeled are reported. This procedure is repeated for every driver whose traffic data is available at Colyar and Halkias (2007). The states that are visited less than a certain threshold number, nlimit, during actual driving by the drivers and during training of the policies are not taken into account. The results are reported for different values of nlimit. The algorithm used for this validation procedure is provided in Algorithm 4, where nstate is the number of visited states by the driver, nV driver is the number of times the driver of interest visited the state being evaluated, nV model is the number of times the state is visited during the training of the driver model, ncomp is the number of states that are used in the comparison and nsuccess is the number of states for which the null hypothesis is not rejected. Algorithm 4 Procedure of Comparison between one driver-one policy - Kolmogorov Smirnov 1: for i = 1 to nstate do 2: 3: 4: if nV driver ≥ nlimit and nV model ≥ nlimit then ncomp+ = 1 Set pi to the probability mass function given by the driver model for the state being evaluated. Set ki to the probability mass function calculated from the driver data for the state being evaluated. Set Hc to the cumulative distribution function obtained from pi. Set Sn to the cumulative distribution function obtained from ki. Test the null hypothesis using KS test. if Null hypothesis is not rejected then 6: 7: 8: 9: 10: 11: end if 12: 13: end for 14: Set the percentage of the successfully modeled states to nsuccess/ncomp. nsuccess+ = 1 5: end if Data validations using Algorithm 4 are conducted for nlimit values of 0, 3 and 5, for each individual driver and the results are reported below. 31 6.5.1. Validation with nlimit = 1 Figure 17 shows the percentage of successfully modeled states for each driver, using level-1, level-2 and level-3 policies as the driver models. Each vertical thin line represents an individual driver and the x-axis shows the driver labels. It is seen that there are more than two thousand drivers in the real traffic data. The y-axis shows the percentage of successfully modeled states. The colors blue, red and yellow represent the successfully modeled states by level-1, level-2 and level-3 policies, respectively. Figure 18 also shows the successfully modeled states, but this time using a uniform distribution (UD) model for the drivers. Finally, Figure 19 shows the difference between the level-k models (combined) and the UD model, in terms of modeling percentages. It is seen that the level-k models cumulatively perform better than the UD model. It is noted that although the level-k models perform better, the percentage of states that are modeled using the UD model are high. The main reason for this is the small nlimit, which is 1 in this case. For the state that is visited only 1 time by the actual driver, the KS test does not reject the hypothesis that the action distribution over this state, obtained from the data, is sampled from a UD, since the sample size is not large enough to arrive at a conclusion. Figure 17: Percentages of successfully modeled states for each driver, using level-k models, when nlimit = 1. 32 Figure 18: Percentages of successfully modeled states for each driver, using the uniform distribution model, when nlimit = 1 Figure 19: Difference in successfully modeled state percentages, for each driver, using level-k (combined) and UD models, when nlimit = 1. Figure 20 shows the number of drivers, on the y-axis, and the percentage of the success- fully modeled states by the level-k models (top) and the UD model (bottom). For example, it shown that around 80% of the total visited states of 300 drivers are successfully modeled by the level-k models, while only a handful of drivers' states are successfully modeled up to 33 60% by the same models. The more the distribution of bars on these figures are grouped on the right, the better, since it shows that successfully modeled state percentages are higher. Figure 20: Distribution of modeled percentages for combined policies and dumb policy, when nlimit = 3. 6.5.2. Validation with nlimit = 3 Figure 21 and 22 show the percentages of the successfully modeled states by the level- k models and the UD model, respectively. As Figure 23 demonstrates, with an increased threshold (nlimit = 3) the better performance of level-k models, compared to the UD model, becomes more prominent . This is due to the increased power of the KS test with the elimination of rarely visited states by the drivers. 34 Figure 21: Percentages of successfully modeled states for each driver, using level-k models, when nlimit = 3. Figure 22: Percentages of successfully modeled states for each driver, using the uniform distribution model, when nlimit = 3. 35 Figure 23: Difference in successfully modeled state percentages, for each driver, using level-k (combined) and UD models, when nlimit = 3. Figure 24 also demonstrates the increased performance difference between the level-k and the UD models, in favor of the level-k, by showing the distribution of the number of drivers over successfully modeled state percentages. As more of the rarely visited states are removed from the comparison, the KS test is able to reject the null hypothesis more frequently for the UD models. Figure 24: Distribution of modeled percentages for combined policies and dumb policy, when nlimit = 3. 36 6.5.3. Validation with nlimit = 5 When even more of the rarely visited states are removed from the test by increasing the threshold further to nlimit = 5, the same trend in increased performance difference between the level-k models and the UD model continues, as seen in Figures 25, 26, 27 and 28. Figure 25: Percentages of successfully modeled states for each driver, using level-k models, when nlimit = 5. Figure 26: Percentages of successfully modeled states for each driver, using the uniform distribution model, when nlimit = 5. 37 Figure 27: Difference in successfully modeled state percentages, for each driver, using level-k (combined) and UD models, when nlimit = 5. Figure 28: Distribution of modeled percentages for combined policies and dumb policy, when nlimit = 5. To summarize, KS test results demonstrate that with varying levels of success, the ex- ploited game theoretical modeling framework provides driver models whose predictive power can be validated with real traffic data. Since it is hard to find similar driver models in the literature, with probability distributions over actions, the results are compared with a uniform distribution model. For all three nlimit values, level-k policies (combined) performed better than the UD model. 38 7. Computational Complexity In both of the application areas, hybrid airspace modeling and road traffic modeling, dis- cussed in the previous sections, the exploited game theoretic framework employed Jaakkola reinforcement learning method explained in Section 3.2.3. In this section, a computational cost analysis is provided for this method. Specifically, two questions are answered: 1) How do the action and observation space sizes affect the computational complexity? 2) How do the complexities of the physical motion models affect the computational complexity? In the analysis, the dimensions of the state and action spaces are taken as S and A, respectively. It is assumed that a) every state is visited K times during one sweep, and b) the value functions are updated after each visit. For each state-action pair, (si, ai), the functions given below are calculated: 1 − χt(si, ai) Kt(si, ai) γtβt−1(si, ai) + χt(si, ai) Kt(si, ai) γtβt−1(si) + χt(si) Kt(si) Qt−1(si, ai) + βt(si, ai)(Rt − R) Vt−1(si) + βt(si)(Rt − R) (cid:18) (cid:18) (cid:18) (cid:18) (cid:19) (cid:19) (cid:19) (cid:19) 1 − χt(si) Kt(ai) 1 − χt(si, ai) Kt(si, ai) 1 − χt(si) Kt(si) βt(si, ai) = βt(si) = Qt(si, ai) = Vt(si) = f or each action ak (cid:54)= ai f or each state sj (cid:54)= si and action al βt(si, ak) = γtβt−1(si, ak) Qt(si, ak) = Qt−1(si, ak) + βt(si, ak)(Rt − R) βt(sj, al) = γtβt−1(sj, al) βt(sj) = γtβt−1(sj) Qt(sj, al) = Qt−1(sj, al) + βt(sj, al)(Rt − R) Vt(sj) = Vt−1(sj) + βt(sj)(Rt − R) f or each state sx and action ay π(aysx) = (1 − )π(aysx) + π1(aysx). (24) Therefore, for each state visit, ((24 + (A− 1)∗ (4)) + (S − 1)∗ (A)∗ (8) + S ∗ A∗ 4) operations are required. Visiting all of the states, K times each, the total number of operations become, K ∗ S ∗ ((24 + (A − 1) ∗ (4)) + (S − 1) ∗ (A) ∗ (8) + S ∗ A ∗ 4), which can be expressed more compactly as R = caS2A + cbSA + ccS, (25) where R is the number of required operations, and ca, cb and cc are constants. Therefore, the number of total operations can be given as O(AS2). 39 Using more complex vehicle dynamics, for example increasing the number of differential equations of the vehicle dynamics by c, will result in a computational cost that is c times of the initial cost. Since this a constant effect, the total number of operations can still be expressed by O(AS2). 8. Ongoing and Future work In this section, we explain the current and future work about the game theoretical mod- eling method explained in this paper. Part of these studies are already formulated but not included in this article for a concise and clear exposition. Below, we briefly mention a few that have the potential to improve the existing work in a meaningful manner. 8.1. 3D hybrid airspace Unmanned aircraft systems (UAS) integration into National Airspace System (NAS) studies presented in this article use a 2-dimensional (2D) geometry for aircraft motion. Although it is demonstrated that the proposed framework can be used to provide significant qualitative analysis power in UAS integration scenarios, the study is still limited and need to be extended to 3D airspace. Preliminary studies are conducted in this direction, details of which can be found in Musavi et al. (2018). One important distinction in this study is the need for an approximate reinforcement learning algorithm to handle the dramatically increased state space due to the 3D geometry. In the 2D case, the observation space consists of 9 variables. 6 of these variables can take 5 different values, while the remaining 3 variables can take 3 different values. Therefore, the size of the observation space is 56 × 33 = 421875. This means that 421875 rows are required in the Q-table to represent each state. To store the values required during training, 16 columns are required. These columns are: state id, state value - V, state visit count, state beta, action 1 probability, action 1 Q value, action 1 count, action 1 beta, action 2 probability, action 2 Q value, action 2 count, action 2 beta, action 3 probability, action 3 Q value, action 3 count and action 3 beta. Hence, 421875 × 16 = 6750000 values are stored in the Q-table as double. Since each double requires 8 bytes, 54 MB memory is required to store the Q table. This is not a significant amount of memory, which can be handled by any modern computer without any problem. However, when the geometry changes from 2D to 3D, the memory requirement becomes infeasible, even if we keep the action space the same. Adding 6 more observation states with 5 possible values for each, the dimension of the observation space becomes 512 × 33 = 6591796875, which translates into a requirement of 800GB of memory to store the Q table. This calculation shows the necessity to use an approximate reinforcement learning method that eliminates the need for storing a Q table. The Neural Fitted Q-learning method, explained in Section 3.2.2 is currently being tested for this task. 8.2. Large scale cyber-security scenarios One ongoing study is creating a model of a large scale cyber-attack scenario, where multiple attackers try to hack into a cyber-physical system and several defenders try to 40 keep the system safe. Reliable predictions of attacker-defender dynamics is valuable since they help design systems resilient to cyber-attacks. A two-person model of a cyber-attack scenario of a smart grid system is already conducted by Backhaus et al. (2013). For a larger scenario, similar to the 3D airspace case, problems of increased state space should be solved together with integrating fast optimization algorithms to optimize the system design. Furthermore, reward function design for multiple attackers and defenders is a challenge especially if coordination within the attackers or defenders is envisioned. 8.3. Data validation Although data validation studies are presented in earlier sections, they are still at their initial stages due to several reasons. First, a lot more data is required for a reliable validation. For example, in UAS integration studies, validation is conducted using a data-validated model of manned aircraft encounters. We hope to obtain UAS and manned aircraft encounter data in the near future as the technology advances. Furthermore, in the traffic scenario, we used US101 data for validation but more road data is expected to be collected to ensure that the proposed model has the capability to model a large variation of highway configurations. In addition, as we use more data, we plan to fit the model parameters to certain amount of data and then validate with independent traffic data. It is noted that parameter fitting to data is not straightforward in the proposed method since reinforcement learning is involved at various reasoning levels. Second, new statistical goodness of fit methods need to be implemented to validate the model's power of prediction. To the best of our knowledge, no goodness of fit methods are implemented for either the UAS integration scenarios or road traffic scenarios where multiple actions are involved. In our ongoing work we are using Chi-square goodness of fit test and Kolmogorov goodness of fit test to validate the method. 9. Open problems and research opportunities As mentioned in Section 8.1, one of the limitations of the level-k thinking solution concept is that an agent's assumption about other agents' levels does not change during the game, and one solution to this may be the dynamic level-k approach. However, this is only a partial solution since the level types still remain unchanged. For example, if Agent A, after watching Agent B for a few time steps, decides that his or her assumption that Agent B is a level-0 player is wrong, then Agent A needs to update his or her assumption to another level. However, the set of levels he or she can choose from are fixed: Agent A can modify his assumption and assume that Agent B is a level-1 player (instead of level-0), therefore the best response to this is producing the actions of a level-2 player. Agent-A can only change his or her assumption to level-k, where k=1,2..., n, where all levels are already trained and determined. A better, but more computationally expensive approach would be the following: After watching Agent B for a few time steps, Agent A can update his or her assumption by using, for example, a Bayesian update and come up a with a policy that is not in the pre-trained set of levels. After this update, Agent A runs a reinforcement learning algorithm online to determine the best response to Agent B's newly updated policy. To achieve this, we need to find new software and hardware solutions to handle the computational demand. 41 Another open problem from a control engineering point of view is the problem of stability. Stability in RL is already being studied by the control community as an open problem (Bu¸soniu et al., 2018). When used in collaboration with game theory, specifically level-k thinking, the problem becomes even harder to solve, and thus presents a rewarding research direction. 10. Summary In this article, we reviewed a modeling approach where reinforcement learning and game theory work in tandem to predict cyber-physical human system (CPHS) behavior. Starting from the basic building blocks, we explained the method in detail and then presented two cases where models are created for unmanned aircraft systems (UAS) integration into Na- tional Airspace (NAS) and highway traffic scenarios. In both cases, validation studies were discussed using different methods, including using real world data. Finally, we presented ongoing and future works, together with related open problems, which can serve as fruitful research directions. Acknowledgements This effort was sponsored by Turkish Academy of Sciences under the Young Scientist Award Programme. References References Abadi, M., Agarwal, A., Barham, P., Brevdo, E., Chen, Z., Citro, C., Corrado, G.S., Davis, A., Dean, J., Devin, M., Ghemawat, S., Goodfellow, I., Harp, A., Irving, G., Isard, M., Jia, Y., Jozefowicz, R., Kaiser, L., Kudlur, M., Levenberg, J., Man´e, D., Monga, R., Moore, S., Murray, D., Olah, C., Schuster, M., Shlens, J., Steiner, B., Sutskever, I., Talwar, K., Tucker, P., Vanhoucke, V., Vasudevan, V., Vi´egas, F., Vinyals, O., Warden, P., Wattenberg, M., Wicke, M., Yu, Y., and Zheng, X. (2015). TensorFlow: Large- scale machine learning on heterogeneous systems. Available: https://www.tensorflow.org/ [retrieved 30 September 2019]. Albaba, M., Yildiz, Y., Li, N., Kolmanovsky, I., and Girard, A. (2019). Stochastic driver modeling and validation with traffic data. In Proc. Amer. Control Conf. Philadelphia, PA. Backhaus, S., Bent, R., Bono, J., Lee, R., Tracey, B., Wolpert, D., Xie, D., and Yildiz, Y. (2013). Cyber-physical security: A game theory model of humans interacting over control systems. IEEE Transactions on Smart Grid, 4(4), 2320 -- 2327. Billingsley, T.B. (2006). Safety analysis of TCAS on Global Hawk using airspace encounter models. Ph.D. thesis, Massachusetts Institute of Technology. 42 Bu¸soniu, L., de Bruin, T., Toli´c, D., Kober, J., and Palunko, I. (2018). Reinforcement learning for control: Performance, stability, and deep approximators. Annual Reviews in Control. Camerer, C.F. (2011). Behavioral game theory: Experiments in strategic interaction. Prince- ton University Press. Carvalho, A., Lefevre, S., Schildbach, G., Kong, J., and Borrelli, F. (2015). Automated driving: The role of forecasts and uncertainty-a control perspective. European Journal of Control, 24, 14 -- 32. Chen, M., Bansal, S., Tanabe, K., and Tomlin, C.J. (2017a). Provably safe and robust drone routing via sequential path planning: A case study in san francisco and the bay area. arXiv preprint arXiv:1705.04585. Chen, M., Hu, Q., Fisac, J.F., Akametalu, K., Mackin, C., and Tomlin, C.J. (2017b). Reachability-based safety and goal satisfaction of unmanned aerial platoons on air high- ways. Journal of Guidance, Control, and Dynamics, 40(6), 1360 -- 1373. Chen, M., Hu, Q., Mackin, C., Fisac, J.F., and Tomlin, C.J. (2015). Safe platooning of unmanned aerial vehicles via reachability. In 2015 54th IEEE Conference on Decision and Control (CDC), 4695 -- 4701. IEEE. Chen, M., Shih, J.C., and Tomlin, C.J. (2016). Multi-vehicle collision avoidance via hamilton- jacobi reachability and mixed integer programming. In 2016 IEEE 55th Conference on Decision and Control (CDC), 1695 -- 1700. IEEE. Colyar, J. J. and Halkias, Techni- cal Report FHWA-HRT-07-030, Federal Highway Administration. Available: https://www.fhwa.dot.gov/publications/research/operations/07030/index.cfm [retrieved 9 February 2019]. (2007). Us highway 101 dataset. Conover, W.J. (1972). A kolmogorov goodness-of-fit test for discontinuous distributions. Journal of the American Statistical Association, 67(339), 591 -- 596. Costa-Gomes, M.A., Crawford, V.P., and Iriberri, N. (2009). Comparing models of strategic thinking in Van Huyck, Battalio, and Beil's coordination games. Journal of the European Economic Association, 7(2-3), 365 -- 376. Crawford, V.P. (2008). Modeling behavior in novel strategic situations via level-k thinking. In North American Economic Science Association Meetings. Dalamagkidis, K., P., V.K., and A., P.L. (2008). On unmanned aircraft systems issues, challenges and operational restrictions preventing integration into the national airspace system. Progress in Aerospace Sciences, 44(7), 503 -- 519. 43 DeGarmo, M.T. (2004). Issues concerning integration of unmanned aerial vehicles in civil airspace. Technical report, MITRE Corporation, Center for Advanced Aviation System Development. Available: https://www.mitre.org/sites/default/files/pdf/04 1232.pdf [re- trieved 9 February 2019]. Dextreit, C. and Kolmanovsky, I.V. (2014). Game theory controller for hybrid electric vehicles. IEEE Transactions on Control Systems Technology, 22(2), 652 -- 663. Ding, J., Tomlin, C.J., Hook, L.R., and Fuller, J. (2016). Initial designs for an automatic forced landing system for safer inclusion of small unmanned air vehicles into the national airspace. In 2016 IEEE/AIAA 35th Digital Avionics Systems Conference (DASC), 1 -- 12. IEEE. Fasano, G., Accardo, D., Moccia, A., Carbone, C., Ciniglio, U., Corraro, F., and Luongo, S. (2008). Multi-sensor-based fully autonomous non-cooperative collision avoidance system for unmanned air vehicles. Journal of aerospace computing, information, and communi- cation, 5(10), 338 -- 360. Florent, M., Schultz, R.R., and Wang, Z. (2010). Unmanned aircraft systems sense and avoid flight testing utilizing ads-b transceiver. In Proceedings of Infotech@ Aerospace. Fudenberg, D. and Tirole, J. (1991). Game theory, 1991. Cambridge, Massachusetts, 393(12), 80. Gabel, T., Lutz, C., and Riedmiller, M. (2011). Improved neural fitted q iteration applied to a novel computer gaming and learning benchmark. In Adaptive Dynamic Programming And Reinforcement Learning (ADPRL), 2011 IEEE Symposium on, 279 -- 286. IEEE. Hess, R.A. and Modjtahedzadeh, A. (1990). A control theoretic model of driver steering behavior. IEEE Control Systems Magazine, 10(5), 3 -- 8. Hidas, P. (2002). Modelling lane changing and merging in microscopic traffic simulation. Transportation Research Part C: Emerging Technologies, 10(5), 351 -- 371. Jaakkola, T., Satinder, P.S., and Jordan., I. (1994). Reinforcement learning algorithm for partially observable markov decision problems. Advances in Neural Information Processing Systems 7: Proceedings of the 1994 Conference. Kacem, T., Wijesekera, D., and Costa, P. (2018). Ads-bsec: A holistic framework to secure ads-b. IEEE Transactions on Intelligent Vehicles, 3(4), 511 -- 521. Kochenderfer, M., Espindle, L., Kuchar, J., and Griffith, J.D. (2008). Correlated encounter model for cooperative aircraft in the national airspace system version 1.0. Project Report ATC-344, Lincoln Laboratory. 44 Kuchar, J.K., Andrews, J., Drumm, A., Hall, T., Heinz, V., Thompson, S., and Welch, J. (2004). A safety analysis process for the traffic alert and collision avoidance system (tcas) and see-and-avoid systems on remotely piloted vehicles. In Proceedings of AIAA 3rd Unmanned Unlimited Technical Conference, Workshop and Exhibit. Kuchar, J. and Drumm, A.C. (2007). The traffic alert and collision avoidance system. Lincoln Laboratory Journal, 16(2), 277. Kumar, P., Perrollaz, M., Lefevre, S., and Laugier, C. (2013). Learning-based approach for online lane change intention prediction. In IEEE Intelligent Vehicles Symposium, 797 -- 802. Lamnabhi-Lagarrigue, F., Annaswamy, A., Engell, S., Isaksson, A., Khargonekar, P., Murray, R.M., Nijmeijer, H., Samad, T., Tilbury, D., and den Hof, P.V. (2017). Systems & control for the future of humanity, research agenda: Current and future roles, impact and grand challenges. Annual Reviews in Control, 43, 1 -- 64. Law, A.M. (2008). How to build valid and credible simulation models. In Simulation Con- ference, 2008. WSC 2008. Winter, 39 -- 47. IEEE. Lee, R. and Wolpert, D. (2011). Chapter: Game theoretic modeling of pilot behavior during mid-air encounters. in Decision making with multiple imperfect decision makers. Intelli- gent Systems Reference Library Series. Springer. Lee, R., Wolpert, D.H., Bono, J., Backhaus, S., Bent, R., and Tracey, B. (2013). Counter- factual reinforcement learning: How to model decision-makers that anticipate the future. In Decision Making and Imperfection, 101 -- 128. Springer. Lefevre, S., Carvalho, A., and Borrelli, F. (2015). Autonomous car following: A learning- based approach. In IEEE Intelligent Vehicles Symposium, 920 -- 926. Lefevre, S., Gao, Y., Vasquez, D., Tseng, E., Bajcsy, R., and Borrelli, F. (2014). Lane keeping assistance with learning-based driver model and model predictive control. In In Proceedings of the 12th international symposium on advanced vehicle control. Li, N., Oyler, D., Zhang, M., Yildiz, Y., Girard, A., and Kolmanovsky, I. (2016). Hierarchical reasoning game theory based approach for evaluation and testing of autonomous vehicle control systems. In Proc. Conference on Decision and Control, 727 -- 733. Li, N., Oyler, D.W., Zhang, M., Yildiz, Y., Kolmanovsky, I., and Girard, A.R. (2018). Game theoretic modeling of driver and vehicle interactions for verification and validation of autonomous vehicle control systems. IEEE Transactions on control systems technology, 26(5), 1782 -- 1797. Maki, D., Parry, C., Noth, K., Molinario, M., and Miraflor, R. (2012). Dynamic protection zone alerting and pilot maneuver logic for ground based sense and avoid of unmanned aircraft systems. In Proceedings of Infotech@ Aerospace. 45 European RPAS Steering Group (2013). Roadmap for the integration of civil rpas into the european aviation system. Technical report, European Commission. Available: http://ec.europa.eu/enterprise/sectors/aerospace/files/rpas-roadmap en.pdf [retrieved 9 February 2019]. FAA (2013). Integration of civil unmanned aircraft systems (uas) in the national airspace system (nas) roadmap. Technical report, U.S. Department of Transportation. Avail- able: http://www.faa.gov/about/initiatives/uas/media/uas roadmap 2013.pdf [retrieved 9 February 2019]. MITRE Corp. (2014). Systems Engineering Guide: Verification and Validation of Simulation Models. Melnyk, R. (2019). A demonstration of reliability and certification standards for unmanned aircraft system control links. In AIAA Scitech 2019 Forum, 1785. Mujumdar, A. and Padhi, R. (2011). Reactive collision avoidance of using nonlinear geo- metric and differential geometric guidance. Journal of guidance, control, and dynamics, 34(1), 303 -- 311. Murphy, R. and Shields, J. (2012). The role of autonomy in dod systems. Task force report, U.S. Department of Defense, Defense Science Board, Washington, D.C., USA. Musavi, N., Onural, D., Gunes, K., and Yildiz, Y. (2016). Unmanned aircraft systems airspace integration: A game theoretical framework for concept evaluations. Journal of Guidance, Control, and Dynamics, 40(1), 96 -- 109. Musavi, N., Manzoor, A., and Yildiz, Y. (2018). A 3d game theoretical framework for the evaluation of unmanned aircraft systems airspace integration concepts. arXiv preprint arXiv:1802.07218. Nancy, A. (2016). Pilots handbook of aeronautical knowledge. Washington: US Department of Transportation, Federal Aviation Administration, Flight Standards Service. Oyler, D., Yildiz, Y., Girard, A., and Kolmanovsky, I. (2016). A game theoretical model of traffic with multiple interacting drivers for use in autonomous vehicle development. In Proc. Amer. Control Conf. Boston, MA. Paszke, A., Gross, S., Chintala, S., Chanan, G., Yang, E., DeVito, Z., Lin, Z., Desmaison, In NIPS A., Antiga, L., and Lerer, A. (2017). Automatic differentiation in PyTorch. Autodiff Workshop. Perez-Batlle, M., Pastor, E., Royo, P., Prats, X., and Barrado, C. (2012). A taxonomy of uas separation maneuvers and their automated execution. In Proceedings of the 2nd International Conference on Application and Theory of Automation in Command and Control Systems, 1 -- 11. 46 Planning, J. et al. (2007). Concept of operations for the next generation air transportation system. Technical report, Citeseer. Pritchett, A.R. (2010). The system safety perspective. In Human Factors in Aviation (Second Edition), 65 -- 94. Elsevier. Riedmiller, M. (2005). Neural fitted q iteration -- first experiences with a data efficient neural reinforcement learning method. In European Conference on Machine Learning, 317 -- 328. Springer. Riedmiller, M. and Braun, H. (1993). A direct adaptive method for faster backpropagation learning: The rprop algorithm. In Proceedings of the IEEE international conference on neural networks, volume 1993, 586 -- 591. San Francisco. Riedmiller, M., Montemerlo, M., and Dahlkamp, H. (2007). Learning to drive a real car in 20 minutes. In Frontiers in the Convergence of Bioscience and Information Technologies, 2007. FBIT 2007, 645 -- 650. IEEE. Salvucci, D., Boer, E., and Liu., A. (2001). Toward an integrated model of driver behavior in cognitive architecture. Transportation Research Record: Journal of the Transportation Research Board, 1779, 9 -- 16. Salvucci, D.D. and Gray, R. (2004). A two-point visual control model of steering. Perception, 33(10), 1233 -- 1248. Sharp, R.S., Casanova, D., and Symonds, P. (2000). A mathematical model for driver steering control, with design, tuning and performance results. Vehicle System Dynamics, 33(5), 289 -- 326. Sheridan, T.B., Corker, K.M., and Nadler, E.D. (2006). Final report and recommendations for research on human-automation interaction in the next generation air transportation system. Technical Report (DOT-VNTSC-NASA- 06-05), U.S. Department of Transporta- tion, Research and Innovative Technology Administration, Cambridge, MA, USA. Shia, V., Gao, Y., Vasudevan, R., Campbell, K.D., Lin, T., Borrelli, F., and Bajcsy, R. (2014). Semiautonomous vehicular control using driver modeling. IEEE Transactions on Intelligent Transportation Systems, 15(6), 2696 -- 2709. Stahl, D. and Wilson, P. (1995). On players' models of other players: Theory and experi- mental evidence. Games and Economic Behavior, 10(1), 218 -- 254. Sutton, R.S. and Barto, A.G. (1998). Reinforcement learning: An introduction. MIT press, Cambridge. Sutton, R.S. and Barto, A.G. (2018). Reinforcement learning: An introduction. MIT press. 47 Treiber, M., Hennecke, A., and Helbing, D. (2000). Congested traffic states in empirical observations and microscopic simulations. Physical Review E, 62(2), 1805 -- 1824. Ungoren, A.Y. and Peng, H. (2005). An adaptive lateral preview driver model. Vehicle System Dynamics, 43(4), 245 -- 259. Vasudevan, R., Shia, V., Gao, Y., Cervera-Navarro, R., Bajcsy, R., and Borrelli, F. (2012). In Proc. Amer. Control Safe semi-autonomous control with enhanced driver modeling. Conf., 2896 -- 2903. Wakitani, S., Yamauchi, Y., Kinoshita, T., Yamamoto, T., Miyakoshi, M., Harada, S., and Yano, Y. (2018). Design of a vehicle driver model based on database-driven control approach. In 2018 IEEE Conference on Control Technology and Applications (CCTA), 840 -- 845. IEEE. Watkins, C.J.C.H. (1989). Learning from delayed rewards. Ph.D. thesis, King's College, Cambridge. Wiering, M. and van Otterlo, M. (eds.) (2012). Reinforcement Learning, State-of-the-art. Springer. Wing, D., Prevot, T., Morey, S., Lewis, T., Martin, L., Johnson, S., Cabrall, C., Como, S., Homola, J., Sheth-Chandra, M., Mercer, J., Morey, S., and Va, H. (2013). Pilot and controller evaluations of separation function allocation in air traffic management. In 10th USA/Europe Air Traffic Management Research and Development Seminar. Yildiz, Y., Agogino, A., and Brat, G. (2013). Predicting pilot behavior in medium scale scenarios using game theory and reinforcement learning. In AIAA Modeling and Simulation Technologies (MST) Conference, 4908. Yildiz, Y., Agogino, A., and Brat, G. (2014). Predicting pilot behavior in medium-scale scenarios using game theory and reinforcement learning. Journal of Guidance, Control, and Dynamics, 37(4), 1335 -- 1343. Yildiz, Y., Lee, R., and Brat, G. (2012). Using game theoretic models to predict pilot behavior in nextgen merging and landing scenario. In Proc. AIAA Modeling and Simulation Technologies Conference, AIAA 2012-4487. Minneapolis, Minnesota. Yoo, J.H. and Langari, R. (2012). Stackelberg game based model of highway driving. In Proc. ASME Dynamic Systems and Control Conference joint with JSME Motion and Vibration Conference. Fort Lauderdale, Florida. Yoo, J.H. and Langari, R. (2013). A stackelberg game theoretic driver model for merging. In Proc. ASME Dynamic Systems and Control Conference. Palo Alto, California. 48
1508.02677	1	1508	2015-08-11T18:24:58	Call Graph Profiling for Multi Agent Systems	[ "cs.MA", "cs.SE" ]	The design, implementation and testing of Multi Agent Systems is typically a very complex task. While a number of specialist agent programming languages and toolkits have been created to aid in the development of such systems, the provision of associated development tools still lags behind those available for other programming paradigms. This includes tools such as debuggers and profilers to help analyse system behaviour, performance and efficiency. AgentSpotter is a profiling tool designed specifically to operate on the concepts of agent-oriented programming. This paper extends previous work on AgentSpotter by discussing its Call Graph View, which presents system performance information, with reference to the communication between the agents in the system. This is aimed at aiding developers in examining the effect that agent communication has on the processing requirements of the system.	cs.MA	cs	Call Graph Profiling for Multi Agent Systems Dinh Doan Van Bien, David Lillis and Rem W. Collier School of Computer Science and Informatics [email protected], {david.lillis, rem.collier}@ucd.ie University College Dublin Abstract. The design, implementation and testing of Multi Agent Sys- tems is typically a very complex task. While a number of specialist agent programming languages and toolkits have been created to aid in the de- velopment of such systems, the provision of associated development tools still lags behind those available for other programming paradigms. This includes tools such as debuggers and profilers to help analyse system be- haviour, performance and efficiency. AgentSpotter is a profiling tool de- signed specifically to operate on the concepts of agent-oriented program- ming. This paper extends previous work on AgentSpotter by discussing its Call Graph View, which presents system performance information, with reference to the communication between the agents in the system. This is aimed at aiding developers in examining the effect that agent communication has on the processing requirements of the system. 1 Introduction By its nature, a Multi Agent System (MAS) is a complex system consisting of loosely-coupled autonomous software entities that are required to communicate with one another in order to achieve individual or system objectives. To facili- tate the development of such systems, a number of agent-oriented programming languages and MAS toolkits have been developed by a variety of researchers [1]. However, the availability of ancillary tools to aid with debugging and profiling is limited, particularly when compared with the available tools for other program- ming paradigms and languages. Profiling is a performance analysis technique that is based on the notion that in a program, only a few places, called bottlenecks or hot spots, can account for the majority of the execution time of a program. Hence, by fixing only these sections of the code, the performance of a program can be substantially improved. Profiling was introduced almost 40 years ago by Donald E. Knuth in his empirical study of FORTRAN programs [2], and has since been successfully adapted to a variety of different languages, platforms and software architectures, including large distributed systems. AgentSpotter is a profiling tool designed specifically for MASs. Its aim is to map traditional concepts of profiling to those of agent oriented software engi- neering so as to facilitate the compilation of profiling data that is specifically relevant to MAS programming. This data can then be presented in an intuitive, 5 1 0 2 g u A 1 1 ] A M . s c [ 1 v 7 7 6 2 0 . 8 0 5 1 : v i X r a visual fashion in order to aid multi agent developers in improving the perfor- mance of their systems. Previous work outlined the AgentSpotter architecture and its Space-Time Diagram view, which was aimed at identifying events in a system that impact performance [3]. This paper continues this work by introducing AgentSpotter's Call Graph View, which attempts to contextualise performance information by linking it to causal events, such as message passing. Section 2 provides a brief discussion of some related tools that have been de- veloped for debugging and profiling MASs. In Section 3, we give a brief overview of the AgentSpotter agent profiling application. Following this, in Section 4 we introduce the concept of a call graph, and analyse how the traditional concept of a call graph can be applied to a MAS. Section 5 presents the concrete imple- mentation of an agent call graph within the AgentSpotter profiling tool, followed by a discussion of the proposed approach in Section 7. Finally, we conclude and outline ideas for future work in Section 8. 2 Related Work The work presented in this paper draws from two principal research areas. Firstly, in order to provide a profiling tool for MASs, it is necessary to examine the con- cepts and features of existing profiling tools for other programming paradigms, such as object-oriented programming. It is also necessary to explore the available programming tools aimed at aiding the debugging and profiling of MASs. Initially proposed by Knuth, the key motivating factor behind profiling tools is his observation that "less than 4% of a program accounts for more than half of its running time" [2]. By identifying and improving code that represents a per- formance bottleneck, software developers can greatly improve the overall per- formance of their programs. An important motivator for the use of specialist profilers to identify these bottlenecks is the frequent tendency of developers' mental maps of their programming not to match the reality of how their pro- grams behave. Thus, a profiler may identify areas of concern that a programmer may not have considered. In the context of more traditional, non-MAS, programming, developers gen- erally have access to long-established and widely-accepted profiling tools such as gprof [4] or performance analysis APIs such as the Java Virtual Machine Tool Interface (JVMTI) [5] or ATOM [6]. However, those developing MASs do not tend to have access to such well-established tools. One MAS framework that does provide the ability to glean data about sys- tem performance is Cougaar [7]. This provides access to data on historical per- formance data, event detection, monitoring of ACL messages and a number of other services. The LS/TS agent platform provides an administrator tool that records some high-level system monitoring information [8]. The main limitation of these systems is the lack of post-processing of the raw performance data in order to produce meaningful synthetic indicators like a profiler would do. Besides performance analysis, most agent frameworks provide a debugging tool similar to the Agent Factory Debugger [9], which provides information about the mental state and communication from the viewpoint of individual agents. A different type of debugging tool is the Agent Viewer that is provided in the Brahms toolkit [10], which displays agent timelines so as to understand when agents' actions are taken. As the work in this paper also requires the monitoring of inter-agent com- munication (see Section 4), it is also important to acknowledge the availability of existing communication analysis tools for MAS platforms. A number of such tools have been developed for a variety of agent frameworks and toolkits to aid developers in understanding the interaction between agents in their systems. An early example of such a toolkit is Zeus [11], which contains a "society tool" that visualises the interaction between agents, so as to help in understanding the topology of the social contacts within the MAS. This type of tool also aids in debugging MASs, since developers can ensure that the expected communication and collaboration between agents is indeed taking place. In the JADE agent development framework, a Sniffer Agent is a FIPA- compliant agent that monitors messages created with an Agent Communication Language (ACL) passed between agents and presents these in a simple graphical interface [12]. A more sophisticated tool, called ACLAnalyser, provides more de- tailed information on agent communication [13]. Again, the principal aim of this is to aid in debugging errors in MASs that relate to coordination or cooperation. 3 AgentSpotter AgentSpotter is a profiling tool designed specifically for gathering and displaying profiling information on MASs [3]. Figure 1 illustrates the abstract architecture of the system, designed to maximise compatibility with any type of agent plat- form. The AgentSpotter Service runs within the Run-Time Environment of an Agent Platform, gathering data about the agents themselves (actions performed, messages exchanged), along with system data such as CPU and memory usage. This is the only platform-specific portion of the system that must be ported in order to allow AgentSpotter to be run on different agent platforms. The data gathered is logged into a Snapshot File, which allows it to be accessed and anal- ysed offline, once the system has finished running. AgentSpotter Station is a visual application that provides a number of visual- isations on various aspects of system performance, in order to help programmers to identify performance bottlenecks in their code. The outputs utilised in this paper are gleaned from running a concrete imple- mentation of the AgentSpotter service within the Agent Factory framework [14]. Agent Factory is a modular and extensible framework that provides comprehen- sive support for the development and deployment of agent-oriented applications. A more detailed description of this implementation and the data gathered by AgentSpotter can be found in [3]. Fig. 1. AgentSpotter Architecture Previously, it was shown how AgentSpotter was used to map traditional profiling concepts on to agent-oriented concepts. This focused on two types of visualisation: -- Flat Profile: provides data on such things as agent activity, messages and reasoning/action duration in a tabular form. -- Space-Time Diagram: provides a navigable visualisation representing the data from the flat profile in a more intuitive manner. The focus of this paper is on an agent-oriented call graph. Whereas a space- time diagram can aid in identifying the timing and extent of actions executed by agents, a call-graph is traditionally intended to also indicate the reasons why particular actions were undertaken at particular times. 4 Call Graph Concept 4.1 Traditional Call Graph The concept of a call graph was introduced in 1982 in the "gprof" profiling tool [4]. This is an improvement on the popular "prof" UNIX profiling tool. In additional to summarising the time spent in different functions, it also recursively presents all the call stacks annotated with the time spent in the various functions that are called. Another name for the call graph is "hierarchical profile", which conveys the idea that gprof provides information to aid in understanding the impact of one function in relation to all the functions that have called it. Although the textual output of gprof is very dense and requires some practice to understand, the user interfaces of modern profiler have made call graphs more tractable by presenting them as tree view controls that can be interactively explored. Figure 2 shows an example of a typical call graph. This relates to a fictional C language program that is designed to remove duplicate lines from a text file. In this tree view, the root node is the main function, which represents 100% of the total execution time of the program (including the execution time of Run-Time EnvironmentPro(cid:127)led ApplicationAgentSpotter ServicePro(cid:127)lerSystem MonitorsAGENT PLATFORMAGENTSPOTTER STATION (visualisation application)Snapshot File (.aspot)Session Summary TablesSpace-Time DiagramCall Graph Tree ViewSessionsAgentsEventsMessagesmessagesactivityeventseventsdatadataqueriesQUERY ENGINESession summary queryAgent activity queryMessages sent/received queryCall graph querySystem activity queryFig. 2. Call Graph Tree View of a fictional C program that removes duplicate lines from a file its child nodes). Each node represents a function within the program, with the child nodes representing functions that are called within the parent function. The percentages represent the cumulative proportion of the program's execution time that is attributable to a node and its children. In the example, the main function calls process file, which is then represented as a sub-tree with leaves representing its own calls to the bubble sort, load file and remove duplicates functions. The key benefit of the call graph tree view is the extended context it gives to performance information. For instance, this simple example reveals that the program spends 90% of its time processing a file. The tree shows that one of the top-level function process file's callees, the bubble sort operation, accounts for 80% of its caller time. A flat profile would have shown the time for these functions separately without explicitly showing the hierarchical link between them. 4.2 Agent-Oriented Call Graph Model When constructing a flat profile for a MAS, it was necessary to map a number of concepts relating to traditional programming to equivalent concepts in the do- main of agent-oriented programming [3]. A similar mapping must be performed in order to allow for the development of an agent-oriented call graph. The central measure used in the traditional call graph is the function execu- tion time. Each node represents a function, which can take the action of calling other functions as part of its execution. The consequence of this action is that some amount of time is spent executing the child function. Thus we can say that the impact of calling a function is that this additional processing time has been incurred. In many MASs, agents tend to perform actions as a reaction to the receipt of ACL messages from other agents in the system. Thus in the same way the impact of a functional call in a traditional system is the execution time of that function, within a MAS, the impact of a message can be related to the addi- tional processing that must be undertaken in order to react to the information contained therein, formulate a response or perform a requested task. Because of this mapping, we introduce, as a first simplified approach, the agent message impact measure to be used as an equivalent to the function processing time used in traditional profiling. It is important to acknowledge that not all MASs are reactive in nature and other events (such as those arising from an agent's envi- ronment) that are not represented by this measure. Further discussion can be found in Section 7. The quantification of such a measure is a difficult task, given the data typi- cally available from MASs. One potential measurement for TMα,B, the impact of a message Mα sent from an agent A to an agent B and received at time stamp α is to use the total amount of computation time used by the agent B until agent B receives a message MΩ from another agent X at time stamp Ω ≥ α. Let b be the duration of an activity by agent B at time stamp t where α ≤ t ≤ Ω. The impact of message Mα on agent B, TMα,B, is then given by the recurrent equation: Ω(cid:88) TMα,B = bt (1) t=α Fig. 3. Agent Message Impact Concept Diagram In Figure 3 we have tried to summarise this concept in a graphical form. The diagram clearly shows that the three activity stars that lie between α and Ω Agent AAgent BAgent X{message Mα at time αtimemessage MΩ at time Ωimpact of MαLegend:Message eventAgent activityAgent life line{impact of MΩαΩβmake up the total impact of Mα on agent B. Note that the outgoing message at time stamp β does not break the computation sequence. It is now easy to determine the total impact Tx,y of all the messages sent by a given agent x to another agent y. Let M be the total number of messages sent, 1 ≤ m ≤ M a single message impact identifier, αm the reception time stamp of message m from x to y, and Ωm, where αm ≤ Ωm, the next reception time stamp message coming right after m from any other source. The total impact Tx,y is then given by the equation: Tx,y = bt (2) By applying the equations recursively, we can compute the total impact Tx of an agent x on N other agents numbered 1 ≤ a ≤ N as follows: M(cid:88) Ωm(cid:88) m=1 t=αm N(cid:88) Ma(cid:88) Ωm(cid:88) a=1 m=1 t=αm K(cid:88) Nk(cid:88) Ma(cid:88) Ωm(cid:88) k=1 a=1 m=1 t=αm Tx = bt (3) Finally, the total impact TS of all the K agents numbered 1 ≤ k ≤ K of a session S is given by the equation: TS = bt (4) It must be noted that the total activity time AS of the session S is given by the equation: AS = TS + K(cid:88) αk0−1(cid:88) bt (5) k=1 t=αS where αS is the first recorded time stamp in session S and αk0 the time stamp of the very first message received by agent k. To put it differently, the total impact for each agent can be computed only after it has received its first message. This proposed method of calculating agent message impact is imperfect, and superior metrics are likely to be developed in the future. However, it does pro- vide useful information for the debugging and development of MASs. Both the drawbacks and benefits of this approach are outlined in more detail in Section 7. 5 Call Graph Visualisation Specification The conceptual model we have presented deals with the session level, the emitter agent level, the receiver agent level and the message level. The graphical trans- lation of the model, outlined In Figure 4, should be a tree view representing the levels we have previously enumerated plus an additional level for the FIPA ACL message content. A message content is defined as a performative plus an expression e.g. "request:doSomeThing(123)". This additional level should give developers necessary contextual information for the messages. It is important to note that this fixed-depth call graph tree represents a divergence from tradi- tional call graphs, whose depth is dictated by the depth of the deepest function call stack. Fig. 4. Call Graph Tree View levels The session at the root of the tree should add up to 100% of all emitter agents' impact as defined by Equation 4. Then at each level, each node should recursively total the impact of its child nodes down to the message leaf nodes. These leaf nodes simply report their impact as defined by Equation 1. More precisely, at each level, for each node, the following values should be displayed: -- Label: informative text associated with the node. The structure of the label depends on the level as follows: • session: "capture date, time - duration"; • emitter agent: "from: agent id"; • sender agent: "to: agent id"; • FIPA ACL: "performative: contents"; • message: "sent: time stamp rec: time stamp". -- Total impact time: sum of impact times of all the current node's children. -- % parent time: percentage of the current node total impact time divided by the node's parent total impact time. -- % session time: percentage of the current node total impact time divided by the session total impact time. Ideally, developers should be able to order the intermediary tree levels differ- ently so as to produce different call graph interpretations. For example, moving the FIPA level right above the emitter agent level would list for each FIPA ACL entry their total impact for all the emitter/receiver pairs. 5.1 User Interface Despite having a fixed depth, a call graph tree view could potentially be very wide at the leaf level for sessions that produce thousands of messages. Therefore, to help developers navigate easily through the tree, AgentSpotter Station offers an advanced tree navigation user interface that expands only that part of the tree that is currently explored so as to reduce the visual clutter. The currently explored part of the tree is highlighted in a different colour to give the developer some visual feedback. Moreover, to speed up the retrieval of information on the system, a search fea- ture allows developers to enter a keyword (e.g. an agent name or a performative). Doing so has the effect of highlighting in a special colour all the visible nodes that contain the specified keyword, significantly improving the visual retrieval speed of a node. Finally, developers can zoom and pan around the tree view to locate items even more quickly. 5.2 Implementation Fig. 5. Call Graph Tree View Screen Shot A sample screen shot of the visualisation of the call graph can be seen in Figure 5. In this figure, the element numbered 1 on the screen shot is the tree root, i.e. the session level which represents 100% of the cumulative recorded activity time. The tree root is highlighted in blue because it is the current tree selection in this specific example. As such, it determines the branch that is expanded, as stated in Section 5.1, so as to reduce the visual clutter. In order to provide a sufficient level of detail, all the children and grandchildren of a selected node are visible. Consequently, when the tree root is selected, only the first two subsequent levels are expanded, that is the emitter agent level and the receiver agent level. Hence, selecting an emitter agent node should make the FIPA ACL message level visible, and so on. As an illustration, the call graph numbered 3 shown in Figure 5 screen shot, has an agent receiver node selected; as a result, this branch is fully expanded down to the message impact level. The element numbered 2 is a text area used to enter a search keyword. The number of nodes matching the keyword is displayed and all the matching nodes that are visible are highlighted in pink. For instance, the element numbered 3 is one of the nine nodes containing the "explorer" keyword and so is highlighted in the screenshot. In a large expanded tree, this highlighting greatly adds to the visual effect and consequently to the navigability of the tree. The bottom-most highlighted node in the tree represents a message sent from the "botagent3" agent to the agent named "explorer". Clicking on this node would cause the subtree rooted at that node to be expanded so as to examine the content and timing of that message. The visualisation is completely interactive and can be controlled using the mouse or the keyboard. Possible interactions include panning, scrolling, expand tree branches, zooming in and out. One other important feature is the ability to alter the hierarchy of the nodes. Whereas the recursive nature of function calls means that these are inherently inflexible in the tree hierarchy they create, the nature of message-passing is a different situation. The hierarchy above places the sender of each message in a higher position in the hierarchy than the recipient. This means that the cumu- lative performance data for the higher-level nodes represents the overall impact of all messages sent by a particular agent to other agents. However, this may not encapsulate the information that a developer requires at a particular point in time. Changing the hierarchy to place the recipient agent above the sender changes the focus of the cumulative performance data. In this case, the figures represent the contribution to overall running time of a particular agent, based on the message that it receives from any and all sources. This may potentially identify entire individual agents as bottlenecks. This may be because the sys- tem's load is imbalanced, meaning that one agent may bear an inequitable share of the processing burden. Alternatively, in a distributed MAS, an agent may simply reside on a machine with inferior hardware resources. By exploiting the flexible nature of this hierarchy, users of the call graph tree view can alter the data being presented to better fit their needs. 6 Evaluation To demonstrate the effectiveness of the call graph as a profiling tool, a sim- ple benchmark application was developed. This consists of two types of agents. Overseer agents request worker agents to perform small, medium or large tasks. If a worker agent has recently been overloaded, it may refuse to execute the Fig. 6. Benchmark Call Graph Tree View for master1 and master2 required task. Occasionally, overseer agents will delegate the assignment of tasks to a worker agent, in which case the worker agent becomes an overseer agent for a brief period. A flat profile and space-time diagram for this benchmark system is contained in [3]. Figure 6 shows a portion of the call graph tree view for a run of this application. Here, the names of overseer agents begin with "master", whereas the names of worker agents begin with "agent". The benchmark application profile (displayed in Figure 6 reveals that over- seer agents master1 and master2 do not have the same impact on performance. Intuitively, one would expect each overseer agent to have an equal impact. How- ever, in reality, we can see that the impact of messages sent by the "master2" agent accounts for only 20.4% of the overall session running time. Studying the call graph in more details helps in explaining this imbalance, by studying the effects of the messages with the content "pleaseDoThing(20)" that were sent by both master1 and master2 to agent001. These are emphasised in Figure 6 by means of the red rectangles. In each case, the parameter passed as part of the a "pleaseDoThing" request is related to the amount of work that the agent is being requested to perform. The call graph shows that some requests from master2 have a 0.0 impact which in practice means they were ignored (no actions took place as a result of receiving those messages). In other words, when master1 sends a request to an agent, and immediately afterwards that master2 sends the same request to the agent, the overloaded agent simply refuses to execute the request. These "please- DoThing(20)" messages sent by master1 are reasonably consistent in terms of their impact, are never refused and account for a total of 8.8% of the total ses- sion running time. In contrast, only a single such request sent by master2 was honoured by agent001. This action accounted for a mere 0.5% of the session running time. It is important to note that the greater impact of master1's messages does not necessarily constitute a bottleneck, merely an imbalance in the system. This type of analysis would motivate the use of the space-time diagram to examine the timing of the messages in question, so as to further find why messages from master2 are more likely to be ignored by the worker agent. A bottleneck would be identified by comparing the impact of different mes- sages being sent (rather than the same message being sent by different agents). For instance, it is notable that the session impact percentages for "pleaseDoTh- ing(1)" messages sent by master1 to agent001 are far lower than for "please- DoThing(20)". In this simple benchmark application, this is an unsurprising result, as the increased workload is explained by the messages themselves, with the latter message requesting more processing to be undertaken by the former. However, figures such as these would indicate a bottleneck if the results are un- expected (i.e. where high-impact messages are not intended to trigger high-cost actions on the part of the message recipients) and so would motivate a closer examination of the longer-running actions to increase efficiency. It may be possible to make such a deduction from viewing the underlying agent code itself, however the use of the call graph makes this far more easily apparent without the need for detailed examination of the code. This also means that testers that are not necessarily familiar with the code (or even perhaps testers who do not understand the programming language used) can identify bottlenecks and behavioural anomalies for developers to address. 7 Discussion The proposed metric for measuring the agent message impact outlined in Sec- tion 4.2 has a number of drawbacks. It operates on the naive assumption that the actions of an agent are directly related to the messages received by it. The impact of a message on an agent is thus taken as the sum of the execution times of all actions undertaken by the agent between the receipt of that message and the receipt of the next message. This means that the proposed metric may not be appropriate to certain types of MAS, where agent action is not intended to occur as a result of communication with other agents. The principal drawback with such an approach is that there is no provable causal link between the receipt of messages and the execution of actions. Agents may decide to act for reasons other than the receipt of ACL messages. For instance, a perceptor may have detected changes in the environment that may require some reaction. Also, when actions are executed as a direct consequence of the receipt of an ACL communication, there is no guarantee that all of the relevant actions have been performed prior to the receipt of the next message. Thus the agent message impact arising from the receipt of a single message may not be particularly informative. Ideally, the best method of measuring the impact of the receipt of an ACL message would be to track the internal reasoning process of the agent, so as to identify those actions that are performed as a direct result of the receipt of a message and take only these into account when calculating the message impact. This is, however, a particularly difficult task, as the reasoning used by agents is extremely platform-dependent and would require a substantial amount of work to be performed in order to port AgentSpotter to other agent platforms and frameworks. This contravenes one of the fundamental aims of AgentSpotter, which is intended to be as platform-agnostic as is practicable. A further difficulty arises when considering the fact that the receipt of mes- sages is not the only event that may cause an agent to undertake some pro- cessing. An agent may also perceive events occurring in its environment, which may prompt some action being taken. Again, the mechanism by which agents perceive such events will vary between agent platforms. Furthermore, toolkits such as CArtAgO [15] or SoSAA [16], which can be used for environment ab- straction, will raise events in different ways, thus adding to the complexity and framework-dependence of the task. In contrast, FIPA-compliant ACL messages will be standardised across toolkits and platforms. Even if we are to settle for a framework-specific profiling system, the task of identifying direct causal links between events is also non-trivial. Whereas some agents may contain straightforward agent code that reacts to the receipt of a message or an environmental percept by always invoking a particular action, this is unlikely to always be the case. Events may instead lead to a refinement of an agent's goals, or even more subtly, an alteration of its current belief state, which in turn may result in goal refinements. Goals may be adopted based on the entire belief set, making it difficult to ascertain for certain whether the belief triggered by the message was a cause of the change in the agent's goals or a merely coincidental occurrence. Even when goals have been adopted, a plan selection algorithm is typically used to decide upon the best path to take towards satisfying those goals. Again, this is a potentially difficult process to trace reliably. Although the proposed measure does have drawbacks and is somewhat sim- plistic, it is also important to highlight the benefits of such a measure. Whereas a high impact measurement for a single message may not be indicative of a major system bottleneck (and may indeed be merely coincidental), consistently high impact measures for similar types of messages are far more likely to be a result of a causal link between the receipt of the message and the processing that follows. Thus by providing an approximation of the communication dynamics within a MAS, valuable insights may be gained. It is this type of analysis that makes the call graph a useful tool in identifying situations that result in a high processing load and thus aid in helping developers concentrate on the appropriate portions of the code base to improve system efficiency. 8 Conclusions and Future Work We have proposed a new visualisation, the call graph tree view, in order to pro- vide detailed information about the performance impact of agents interactions. After discussing the concept of call graph in a traditional programming context, we have then mapped it into an agent-oriented concept based on the idea that when an agent sends an ACL message to another, its impact on the amount of processing the recipient perform can be measured and used to identify sys- tem bottlenecks, load imbalances and efficiency issues. Although the proposed measure is not optimal, it does provide users with data that is appropriate and useful in the context of a profiler application. We have then extended this notion to a tree model with multiple levels: session, message emitter, message receiver, message. Finally, we have described the advanced user interface that allows de- velopers and testers to interact with this model in the form of a zoomable and searchable tree view. For further development of the call graph view of the AgentSpotter applica- tion, there are two principal areas for improvement. Firstly, as we have acknowl- edged in Section 7, the current measure for gauging the processing impact of a message being passed between agents is not an ideal one. We intend to inves- tigate other possible measures that will include a stronger causal link between the receipt of a message and the resulting processing activity. In doing so, the other AgentSpotter views (flat profile and space-time diagram) will be utilised to ensure that any proposed measures reflect the reality of the system's execution as closely as possible. The second significant area of future work is in the area of agent conversation protocols. The work presented in this paper considers each ACL message to be entirely independent of all other messages. The reality of agent communication is somewhat different. In the agent architecture presented as the benchmark ap- plication in Section 6, some overseer agents request that worker agents perform certain tasks. In our simple application, this is done by means of a single message containing the work request being sent to the worker. In reality, a more complex conversation would be used. The initial request for a task to be performed may be answered with an acceptance or rejection of the task being assigned, followed perhaps by the communication of the result of the task. Clearly, an agent accept- ing and performing a task will consume more processing resources than when it is rejected. However, in the existing model, both scenarios will be grouped together, under the initial message requesting action. Such behaviour may mask inefficiencies in the processing code by including the low-cost rejection actions in its session percentages. By introducing an additional conversation level into the tree, these situations can be separated, meaning that actions will be grouped according to entire agent transactions rather than single messages. References 1. Bordini, R., Braubach, L., Dastani, M., Seghrouchni, A., Gomez-Sanz, J., Leite, J., O'Hare, G., Pokahr, A., Ricci, A.: A survey of programming languages and platforms for multi-agent systems. Informatica 30(1) (2006) 33 -- 44 2. Knuth, D.E.: An empirical study of FORTRAN programs. j-SPE 1(2) (April/June 1971) 105 -- 133 3. Doan Van Bien, D., Lillis, D., Collier, R.: Space-time diagram generation for profiling multi agent systems. In: Proceedings of PROMAS'09, Budapest, Hungary (May 11-12th 2009) 4. Graham, S.L., Kessler, P.B., Mckusick, M.K.: Gprof: A call graph execution pro- filer. SIGPLAN Not. 17(6) (1982) 120 -- 126 5. Sun Microsystems, Inc.: JVM Tool Interface (JVMTI), Version 1.0. Web pages at http://java.sun.com/j2se/1.5.0/docs/guide/jvmti/ (accessed August 4th, 2008) (2004) 6. Srivastava, A., Eustace, A.: Atom: a system for building customized program analysis tools. In: PLDI '94: Proceedings of the ACM SIGPLAN 1994 conference on Programming language design and implementation, New York, NY, USA, ACM (1994) 196 -- 205 7. Helsinger, A., Thome, M., Wright, T., Technol, B., Cambridge, M.: Cougaar: a scalable, distributed multi-agent architecture. In: Systems, Man and Cybernetics, 2004 IEEE International Conference on. Volume 2. (2004) 8. Rimassa, G., Calisti, M., Kernland, M.E.: Living Systems R(cid:13)Technology Suite. Whitestein Series in Software Agent Technologies and Autonomic Computing. In: Software Agent-Based Applications, Platforms and Development Kits. Birkhauser Basel (2005) 73 -- 93 9. Collier, R.: Debugging Agents in Agent Factory. Lecture Notes in Computer Science 4411 (2007) 229 10. Seah, C., Sierhuis, M., Clancey, W., Cognition, M.: Multi-agent modeling and simulation approach for design and analysis of MER mission operations. In: Pro- ceedings of 2005 International conference on human-computer interface advances for modeling and simulation (SIMCHI'05). (2005) 73 -- 78 11. Nwana, H., Ndumu, D., Lee, L.: ZEUS: An advanced tool-kit for engineering distributed multi-agent systems. Applied AI 13(1) (1998) 2 12. Bellifemine, F., Caire, G., Greenwood, D., Corporation, E.: Developing multi-agent systems with JADE. Springer (2007) 13. Botıa, J., Hernansaez, J., Skarmeta, F.: Towards an Approach for Debugging MAS Through the Analysis of ACL Messages . Computer Systems Science and Engineering 20 (2005) 14. Collier, R., O'Hare, G., Lowen, T., Rooney, C.: Beyond Prototyping in the Factory of Agents. Multi-Agent Systems and Application III: 3rd International Central and Eastern European Conference on Multi-Agent Systems, Ceemas 2003, Prague, Czech Republic, June 16-18, 2003: Proceedings (2003) 15. Ricci, A., Viroli, M., Omicini, A.: CArtAgO: A framework for prototyping artifact- based environments in MAS. Lecture Notes in Computer Science 4389 (2007) 67 16. Dragone, M., Lillis, D., Collier, R.W., O'Hare, G.M.P.: SoSAA: A Framework for Integrating Agents & Components. In: Proceedings of the 24th Annual Symposium on Applied Computing (ACM SAC 2009), Special Track on Agent-Oriented Pro- gramming, Systems, Languages, and Applications, Honolulu, Hawaii, USA (March 2009)
1807.06613	3	1807	2019-06-06T10:27:01	Deep Reinforcement Learning for Swarm Systems	[ "cs.MA", "cs.AI", "cs.LG", "eess.SY", "stat.ML" ]	Recently, deep reinforcement learning (RL) methods have been applied successfully to multi-agent scenarios. Typically, these methods rely on a concatenation of agent states to represent the information content required for decentralized decision making. However, concatenation scales poorly to swarm systems with a large number of homogeneous agents as it does not exploit the fundamental properties inherent to these systems: (i) the agents in the swarm are interchangeable and (ii) the exact number of agents in the swarm is irrelevant. Therefore, we propose a new state representation for deep multi-agent RL based on mean embeddings of distributions. We treat the agents as samples of a distribution and use the empirical mean embedding as input for a decentralized policy. We define different feature spaces of the mean embedding using histograms, radial basis functions and a neural network learned end-to-end. We evaluate the representation on two well known problems from the swarm literature (rendezvous and pursuit evasion), in a globally and locally observable setup. For the local setup we furthermore introduce simple communication protocols. Of all approaches, the mean embedding representation using neural network features enables the richest information exchange between neighboring agents facilitating the development of more complex collective strategies.	cs.MA	cs	Journal of Machine Learning Research 20 (2019) 1-31 Submitted 7/18; Revised 1/19; Published 2/19 Deep Reinforcement Learning for Swarm Systems Maximilian Hüttenrauch L-CAS University of Lincoln LN6 7TS Lincoln, UK Adrian Šošić Bioinspired Communication Systems Technische Universität Darmstadt 64283 Darmstadt, Germany Gerhard Neumann L-CAS University of Lincoln LN6 7TS Lincoln, UK Editor: George Konidaris [email protected] [email protected] [email protected] Abstract Recently, deep reinforcement learning (RL) methods have been applied successfully to multi-agent scenarios. Typically, the observation vector for decentralized decision making is represented by a concatenation of the (local) information an agent gathers about other agents. However, concatenation scales poorly to swarm systems with a large number of homogeneous agents as it does not exploit the fundamental properties inherent to these systems: (i) the agents in the swarm are interchangeable and (ii) the exact number of agents in the swarm is irrelevant. Therefore, we propose a new state representation for deep multi-agent RL based on mean embeddings of distributions, where we treat the agents as samples and use the empirical mean embedding as input for a decentralized policy. We define different feature spaces of the mean embedding using histograms, radial basis functions and neural networks trained end-to-end. We evaluate the representation on two well-known problems from the swarm literature -- rendezvous and pursuit evasion -- in a globally and locally observable setup. For the local setup we furthermore introduce simple communication protocols. Of all approaches, the mean embedding representation using neural network features enables the richest information exchange between neighboring agents, facilitating the development of complex collective strategies. Keywords: deep reinforcement learning, swarm systems, mean embeddings, neural net- works, multi-agent systems 1. Introduction In swarm systems, many identical agents interact with each other to achieve a common goal. Typically, each agent in a swarm has limited capabilities in terms of sensing and manipulation so that the considered tasks need to be solved collectively by multiple agents. A promising application where intelligent swarm systems take a prominent role is swarm robotics (Bayındır, 2016). Robot swarms are formed by a large number of cheap and easy c(cid:13)2019 Maximilian Hüttenrauch and Adrian Šošić and Gerhard Neumann. License: CC-BY 4.0, see https://creativecommons.org/licenses/by/4.0/. Attribution requirements are provided at http://jmlr.org/papers/v20/18-476.html. Hüttenrauch, Šošić and Neumann to manufacture robots that can be useful in a variety of situations and tasks, such as search and rescue missions or exploration scenarios. A swarm of robots is inherently redundant towards loss of individual robots since usually none of the robots plays a specific role in the execution of the task. Because of this property, swarm-based missions are often favorable over single-robot missions (or, let alone, human missions) in hazardous environments. Be- havior of natural swarms, such as foraging, formation control, collective manipulation, or the localization of a common 'food' source can be adapted to aid in these missions (Bayındır, 2016). Another field of application is routing in wireless sensor networks (Saleem et al., 2011) since each sensor in the network can be treated as an agent in a swarm. A common method to obtain control strategies for swarm systems is to apply opti- mization-based approaches using a model of the agents or a graph abstraction of the swarm (Lin et al., 2004; Jadbabaie et al., 2003). Optimization-based approaches allow to compute optimal control policies for tasks that can be well modeled, such as rendezvous or consensus problems (Lin et al., 2007) and formation control (Ranjbar-Sahraei et al., 2012), or to learn pursuit strategies to capture an evader (Zhou et al., 2016). Yet, these approaches typically use simplified models of the agents / the task and often rely on unrealistic assumptions, such as operating in a connected graph (Dimarogonas and Kyriakopoulos, 2007) or having full observability of the system state (Zhou et al., 2016). Rule-based approaches use heuristics inspired by natural swarm systems, such as ants or bees (Handl and Meyer, 2007). Yet, while the resulting heuristics are often simple and can lead to complex swarm behavior, the obtained rules are difficult to adapt, even if the underlying task changes only slightly. Recently, deep reinforcement learning (RL) strategies have become popular to solve multi-agent coordination problems. In RL, tasks are specified indirectly through a cost function, which is typically easier than defining a model of the task directly or a finding a heuristic for the controller. Having defined a cost function, the RL algorithm aims to find a policy that minimizes the expected cost. Applying deep reinforcement learning within the swarm setting, however, is challenging due to the large number of agents that need to be considered. Compared to single-agent learning, where the agent is confronted only with observations about its own state, each agent in a swarm can make observations of several other agents populating the environment and thus needs to process an entire set of information that is potentially varying in size. Accordingly, two main challenges can be identified in the swarm setting: 1. High state and observation dimensionality, caused by large system sizes. 2. Changing size of the available information set, either due to addition or removal of agents, or because the number of observed neighbors changes over time. Most current multi-agent deep reinforcement learning methods either concatenate the in- formation received from different agents (Lowe et al., 2017) or encode it in a multi-channel image, where the image channels contain different features based on a local view of an agent (Sunehag et al., 2017; Zheng et al., 2017). However, both types of methods bare major draw- backs. Since neural network policies assume a fixed input dimensionality, a concatenation of observations is unsuitable in the case changing agent numbers. Furthermore, a concate- nation disregards the inherent permutation invariance of identical agents in a swarm system and scales poorly to large system sizes. Top-down image based representations alleviate the 2 Deep Reinforcement Learning for Swarm Systems issue of permutation invariance, however, the information obtained from neighboring agents is of mostly spatial nature. While additional information can be captured by adding more image channels, the dimensionality of the representation increases linearly with each feature. Furthermore, the discretization into pixels has limited accuracy due to quantization errors. In this paper, we exploit the homogeneity of swarm systems and treat the state infor- mation perceived from neighboring agents as samples of a random variable. Based on this model, we then use mean feature embeddings (MFE) (Smola et al., 2007) to encode the current distribution of the agents. Each agent gets a local view of this distribution, where the information obtained from the neighbors is encoded in the mean embedding. Due to the sample-based view of the collected state information, we achieve a permutation invari- ant representation that is furthermore invariant to the number of agents in the swarm / the number of perceived neighbors. Mean feature embeddings have so far been used mainly for kernel-based feature repre- sentations (Gretton et al., 2012), but they can be also applied to histograms or radial basis function (RBF) networks. The resulting models are closely related to the "invariant model" formulated by Zaheer et al. (2017). However, compared to the summation approach de- scribed in their paper, the averaging of feature activations proposed in our approach yields the desired invariance with respect to the observed agent number mentioned above. To the best of our knowledge, we are the first to use mean embeddings inside a deep reinforcement learning framework for swarm systems where both the feature space of the mean embedding as well as the policy are learned end-to-end. We test our state representation on various rendezvous and pursuit evasion problems using Trust Region Policy Optimization (TRPO) (Schulman et al., 2015) as the underlying deep RL algorithm. In the rendezvous problem, the agents need to find a collective strategy that allows them to meet at some arbitrary location. In the pursuit evasion domain, a group of agents collectively tries to capture one or multiple evaders. Policies are learned in a centralized-learning / decentralized-execution fashion fashion, meaning that during learning data from all agents is collected centrally and used to optimize the parameters as if there was only one agent. Nonetheless, each agent only has access to its own perception of the global system state to generate actions from the policy function. We compare our representation to several deep RL baselines as well as to optimization-based solutions, if available. Herein, we perform our experiments both in settings with global ob- servability (i.e., all agents are neighbors) and in settings with local observability (i.e., agents are only locally connected). In the latter setting, we also evaluate different communication protocols (Hüttenrauch et al., 2018) that allow the agents to transmit additional informa- tion about their local graph structure. For example, an agent might transmit the number of neighbors within its current neighborhood. Previously, such additional information could not be encoded efficiently due to the poor scalability of the histogram-based approaches. Our results show that agents using our representation can learn faster and obtain poli- cies of higher quality, suggesting that the representation as mean embedding is an efficient encoding of the global state configuration for swarm-based systems. Moreover, mean em- beddings are simple to implement inside existing neural network architectures and can be 3 Hüttenrauch, Šošić and Neumann applied to any deep RL algorithm, which makes the approach applicable in a wide variety of scenarios. The source code to reproduce our results can be found online.1 2. Related Work The main contribution of this work lies in the development of a compact representation of state information in swarm systems, which can easily be used within deep multi-agent reinforcement learning (MARL) settings that contain homogeneous agent groups. In fact, our work is mostly orthogonal to other research in the field of MARL and the presented ideas can be incorporated into most existing approaches. To provide an overview, we begin with a brief survey of algorithms used in (deep) MARL, we revisit the basics of mean embedding theory, and we summarize some classic approaches to swarm control for the rendezvous and pursuit evasion task. 2.1 Deep RL Recently, there has been increasing interest in deep reinforcement learning for swarms and multi-agent systems in general. For example, Zheng et al. (2017) provide a many-agent reinforcement learning platform based on a multi-channel image state representation, which uses Deep Q-Networks (DQN) (Mnih et al., 2015) to learn decentralized control strategies in large grid worlds with discrete actions. Gupta et al. (2017) show a comparison of centralized, concurrent and parameter sharing approaches to cooperative deep MARL, using TRPO (Schulman et al., 2015), DDPG (Lillicrap et al., 2015) and DQN. They evaluate each method on three tasks, one of which is a pursuit task in a grid world using bitmap-like images as state representation. A variant of DDPG for multiple agents in Markov games using a centralized action-value function is provided by Lowe et al. (2017). The authors evaluate the method on tasks like cooperative communication, navigation and others. The downside of a centralized action-value function is that the input space grows linearly with the number of agents, and hence, their approach scales poorly to large system sizes. A more scalable approach is presented by Yang et al. (2018). Employing mean field theory, the interactions within the population of agents are approximated by the interaction of a single agent with the average effect from the overall population, which has the effect that the action-value function input space stays constant. Experiments are conducted on a Gaussian squeeze problem, an Ising model, and a mixed cooperative-competitive battle game. Yet, the paper does not address the state representation problem for swarm systems. Omidshafiei et al. (2017) investigate hysteretic Q-learning (Matignon et al., 2007) and distillation (Rusu et al., 2015). They use deep recurrent Q-networks (Hausknecht and Stone, 2015) to solve single and multi-task Dec-POMDP problems. Following this work, Palmer et al. (2017) add leniency (Panait et al., 2006) to the hysteretic approach to prevent "relative overgeneralization" of agents. The approach is evaluated on a coordinated multi-agent object transportation problem in a grid world with stochastic rewards. Sunehag et al. (2017) tackle the "lazy agent" problem in cooperative MARL with a single team reward by training each agent with a learned additive decomposition of a value function based on the team reward. Experiments show an increase in performance on cooperative 1. https://github.com/LCAS/deep_rl_for_swarms 4 Deep Reinforcement Learning for Swarm Systems two-player games in a grid world. Rashid et al. (2018) further develop the idea with the insight that a full factorization of the value function is not necessary. Instead, they introduce a monotonicity constraint on the relationship between the global value function and each local value function. Results are presented on the StarCraft micro management domain. Finally, Grover et al. (2018) show a framework to model agent behavior as a represen- tation learning problem. They learn an encoder-decoder embedding of agent policies via imitation learning based on interactions and evaluate it on a cooperative particle world (Mordatch and Abbeel, 2018) and a competitive two-agent robo sumo environment (Al- Shedivat et al., 2018). The design of the policy function in the approach of Mordatch and Abbeel (2018) is similar to ours but the model uses a softmax pooling layer. However, instead of applying (model-free) reinforcement learning to optimize the parameters of the policy function, they build an end-to-end differentiable model of all agent and environment state dynamics and calculate the gradient of the return with respect to the parameters via backpropagation. An application related to our approach can be found in the work by Gebhardt et al. (2018), where the authors use mean embeddings to learn a centralized controller for ob- ject manipulation with robot swarms. Here, the key idea is to directly embed the swarm configuration into a reproducing kernel Hilbert space, whereas our approach is based on embedding the agent's local view. Furthermore, using kernel-based feature spaces for the mean embedding scales poorly in the number of samples and in the dimensionality of the embedded information. 2.2 Optimization-Based Approaches for Swarm Systems To provide a concise summary of the most relevant related work, we concentrate on opti- mization-based approaches that derive decentralized control strategies for the rendezvous and pursuit evasion problem considered in this paper. Ji and Egerstedt (2007) derive a control mechanism preserving the connectedness of a group of agents with limited commu- nication abilities for the rendezvous and formation control problem. The method focuses on high-level control with single integrator linear state manipulation and provides no rules for agents that are not part of the agent graph. Similarly, Gennaro and Jadbabaie (2006) present a decentralized algorithm to maximize the connectivity (characterized by an expo- nential model) of a multi-agent system. The algorithm is based on the minimization of the second smallest eigenvalue of the Laplacian of the proximity graph. An approach providing a decentralized control strategy for the rendezvous problem for nonholonomic agents can be found in the work by Dimarogonas and Kyriakopoulos (2007). Using tools from nons- mooth Lyapunov theory and graph theory, the stability of the overall system is examined. A control strategy for the pursuit evasion problem with multiple pursuers and single evader that we investigate in more detail later in this paper was proposed Zhou et al. (2016). The authors derive decentralized control policies for the pursuers and the evader based on the minimization of Voronoi partitions. Again, the control mechanism is for high-level linear state manipulation. Furthermore, the method assumes visibility of the evader at all times. A survey on pursuit evasion in mobile robotics in general is provided by Chung et al. (2011). 5 Hüttenrauch, Šošić and Neumann 2.3 Analytic Approaches Another line of work concerned with the curse of dimensionality can be found in the area of multi-player reach-avoid games. Chen et al. (2017), for example, look at pairwise interactions between agents. This way, they are able to use the Hamilton-Jacobian-Isaacs approach to solve a partial differential equation in the joint state space of the players. Similar work can be found in (Chen et al., 2014a,b; Zhou et al., 2012). 3. Background In this section, we give a short overview of Trust Region Policy Optimization and mean embeddings of distributions. 3.1 Trust Region Policy Optimization Trust Region Policy Optimization is an algorithm to optimize control policies in single-agent reinforcement learning problems (Schulman et al., 2015). These problems are formulated as Markov decision processes (MDPs), which can be compactly written as a tuple (cid:104)S,A, P, R(cid:105). In an MDP, an agent chooses an action a ∈ A according to some policy π(a s) based on its current state s ∈ S and progresses to state s(cid:48) ∈ S according to the transition dynamics P , i.e., s(cid:48) ∼ P (s(cid:48) s, a). After each step, the agent receives a reward r = R(s, a), provided by the reward function R, which judges the quality of its decision. The goal of the agent is to find a policy that maximizes the cumulative reward achieved over a certain period of time. In TRPO, the policy is parametrized by a parameter vector θ containing the weights and biases of a neural network. In the following, we denote this parametrized policy as πθ. The reinforcement learning objective is expressed as finding a new policy that maximizes the Aπold(s, a) , where Aπold(s, a) = Qπold(s, a)− V πold(s). Herein, the state-action value function Qπold(s, a) is typically estimated via trajectory rollouts, while for the value function V πold(s) linear or neural network baselines are used that are fitted to the Monte-Carlo returns, resulting in an estimate A(s, a) for the advantage function. The objective is to be maximized subject to a fixed constraint on the Kullback-Leibler (KL) divergence of the policy before and after the parameter update, which ensures that the updates to the policy parameters θ are bounded, in order to avoid divergence of the learning process. The overall optimization problem is summarized as expected advantage function of the current policy πold, i.e., J TRPO = E(cid:104) πθ (cid:105) πθold (cid:21) (cid:20) πθ E θ maximize subject to E[DKL(πθoldπθ)] ≤ δ. A(s, a) πθold The problem is approximately solved using conjugate gradient optimization, after linearizing the objective and quadratizing the constraint. 3.2 Mean Embeddings Our work is inspired by the idea of embedding distributions into reproducing kernel Hilbert spaces (Smola et al., 2007) from where we borrow the concept of mean embeddings. A probability distribution P (X) can be represented as an element in a reproducing kernel 6 Deep Reinforcement Learning for Swarm Systems Hilbert space by its expected feature map (i.e., the mean embedding), µX = EX [φ(X)], where φ(x) is a (possibly infinite dimensional) feature mapping. Given a set of observations {x1, . . . , xm}, drawn i.i.d. from P (X), the empirical estimate of the expected feature map is given by m(cid:88) i=1 µX = 1 m φ(xi). Using characteristic kernel functions k(x, x(cid:48)) = (cid:104)φ(x), φ(x(cid:48))(cid:105), such as Gaussian RBF or Laplace kernels, mean embeddings can be used, for example, in two-sample tests (Gretton et al., 2012) and independence tests (Gretton et al., 2008). A characteristic kernel is required to uniquely identify a distribution based on its mean embedding. However, this assumption can be relaxed to using finite feature spaces if the objective is merely to extract relevant information from a distribution such as, in our case, the information needed for the policy of the agents. 4. Deep Reinforcement Learning for Swarms The reinforcement learning algorithm presented in the last section has been originally de- signed for single-agent learning. In order to apply this algorithm to the swarm setup, we switch to a different problem domain and show the implications on the learning algorithm. Policies in this context are then optimized in a centralized -- learning / decentralized -- execution fashion. 4.1 Problem Domain The problem domain for our swarm system is best described as a swarm MDP environment (Šošić et al., 2017). The swarm MDP can be regarded as a special case of a decentralized partially observable Markov decision process (Dec-POMDP) (Bernstein et al., 2002) and is constructed in two steps. First, an agent prototype is defined as a tuple A = (cid:104)S,O,A, π(cid:105), determining the local properties of an agent in the system. Herein, S denotes the set of the agent's local states, O is the set of possible local observations, A is the set of actions available to the agent, and π : O × A → [0, 1] is the agent's stochastic control policy. Based on this definition, the swarm MDP is constructed as (cid:104)N, A, P, O, R(cid:105), where N is the number of agents in the system and A is the aforementioned agent prototype. The coupling of the agents is specified through a global state transition model P : S N×S N×AN → [0,∞) and an observation model O : S N ×{1, . . . , N} → O, which determines the local observation oi ∈ O for agent i at a given swarm state s ∈ S N, i.e., oi = O(s, i). Finally, R : S N × AN → R is the global reward function, which encodes the cooperative task for the swarm by providing an instantaneous reward feedback R(s, a) according to the current swarm state s and the corresponding joint action assignment a ∈ AN of the agents. The specific state dynamics and observation models considered in this paper are described in Section 5. The model encodes two important properties of swarm networks: First, all agents in the system are assumed to be identical, and accordingly, they are all assigned the same 7 Hüttenrauch, Šošić and Neumann decentralized policy π. This is an immediate consequence of the two-step construction of the model, which implies that all agents share the same internal architecture. Second, the agents are only partially informed about the global system state, as prescribed by the observation model O. Note that both the transition model and the observation model are assumed to be invariant to permutations of the agents in order to ensure the homogeneity of the system. For details, see (Šošić et al., 2017). 4.2 Local Observation Models The local observation oi introduced in the last section is a combination of observations oi loc an agent makes about local properties (like the agent's current velocity or its distance to a wall) and observations Oi of other agents. In order to describe the observation model used for the agents, we use an interaction graph representation of the swarm. This graph is given by nodes V = {v1, v2, . . . , vN} corresponding to the agents in the swarm and an edge set E ⊂ V × V , which we assume contains unordered pairs of the form {vi, vj} indicating that agents i and j are neighbors. The interaction graph is denoted as G = (V, E). If both the set of nodes and the set of edges are not changing, we call G a static interaction graph; if either of the set undergoes changes, we instead refer to G as a dynamic interaction graph. The set of neighbors of agent i in the graph G is given by NG(i) = {j {vi, vj} ∈ E}. information agent i receives from all neighbors is given by the set Oi =(cid:8)oi,j j ∈ NG(i)(cid:9). Within this neighborhood, agent i can sense local information about other agents, for exam- ple distance or bearing to each neighbor. We denote the information agent i receives from agent j as oi,j = f (si, sj), which is a function of the local states of agent i and agent j. The observation oi,j is available for agent i only if j ∈ NG(i). Hence, the complete state As the observations of other agents are summarized in form of sets {Oi}, we require an efficient encoding that can be used as input to a neural network policy. In particular, it must meet the following two properties: • The encoding needs to be invariant to the indexing of the agents, respecting the unorderedness of the elements in the observation set. Only by exploiting the system's inherent homogeneity we can escape the curse of dimensionality. • The encoding must be applicable to varying set sizes because the local graph structure might change dynamically. Even if each agent can observe the entire system at all times, the encoding should be applicable for different swarm sizes. 4.3 Local Communication Models In addition to perceiving local state information of neighboring agents, the agents can also communicate information about the interaction graph G (Hüttenrauch et al., 2018). For example, agent j can transmit the number of perceived neighbors to agent i. Furthermore, the agents can also perform more complex operations on their local neighborhood graph. For example, they could compute the shortest distance to a target point (such as an evader) that is perceived by at least one agent within their local sub-graph. Hence, by using local 8 Deep Reinforcement Learning for Swarm Systems communication protocols, observation oi,j can contain information about both, the local states si and sj as well as the graph G, i.e., oi,j = f (si, sj,G). 4.4 Mean Embeddings as State Representations for Swarms In the simplest case, the local observation oi,j that agent i receives of agent j is composed of the distance and the bearing angle of agent i to agent j. However, oi,j can also contain more complex information, such as relative velocities or orientations. A straightforward way to represent the information set Oi is to concatenate the local quantities {oi,j}j into a single observation vector. However, as mentioned before, this representation has various drawbacks as it ignores the permutation invariance inherent to a homogeneous agent network. Further- more, it grows linearly with the number of agents in the swarm and is, therefore, limited to a fixed number of neighbors when used in combination with neural network policies. To resolve these issues, we treat the elements in the information set Oi as samples from a distribution that characterizes the current swarm configuration, i.e., oi,j ∼ pi(· s). We can now use an empirical encoding of this distribution in order to achieve permutation invariance of the elements of Oi as well as flexibility to the size of Oi. As highlighted in Section 3.2, a simple way is to use a mean feature embedding, i.e., (cid:88) oi,j∈Oi µOi = 1 Oi φ(oi,j), where φ defines the feature space of the mean embedding. The input dimensionality to the policy is given by the dimensionality of the feature space of the mean embedding and, hence, it does not depend on the size of the information set Oi any more. This allows us to use the embedding µOi as input to a neural network used in deep RL. In the following sections, we describe different feature spaces that can be used for the mean embedding. Figure 1 illustrates the resulting policy architectures with further details given in Appendix F. 4.4.1 Neural Network Feature Embeddings In line with the deep RL paradigm, we propose to use a neural network as feature map- ping φNN whose parameters are determined by the reinforcement learning algorithm. Using a neural network to define the feature space allows us to handle high dimensional obser- vations, which is not feasible with traditional approaches such as histograms (Hüttenrauch et al., 2018). In our experiments, a rather shallow architecture with one layer of RELU units already performed very well, but deeper architectures could be used for more complex applications. To the best of our knowledge, we present the first approach for using neural networks to define the feature space of a mean embedding. 4.4.2 Histograms An alternative feature space are provided by histograms, which can be related to image- like representations. In this approach, we discretize the space of certain features, such as the distance and bearing to other agents, into a fixed number of bins. This way, we can collect information about neighboring agents in the form of a fixed-size multi-dimensional histogram. Herein, the histogram bins define a feature mapping φHIST using a one-hot- coding for each observed agent. A detailed description of this approach can be found in 9 Hüttenrauch, Šošić and Neumann (b) RBF and histogram em- bedding policy (c) policy net- work for concate- nation (a) neural network embedding policy network Figure 1: Illustration of (a) the neural network mean embedding policy, (b) the network ar- chitecture used for the RBF and histogram representation, and (c) for the simple concatenation of observations. The numbers inside the boxes denote the dimen- sionalities of the hidden layers. The color coding in (a) highlights which layers share the same weights. The plus sign denotes the mean of the feature activations. our previous work (Hüttenrauch et al., 2018). While the approach works well in discrete environments where each cell is only occupied by a single agent, the representation can lead to blurring effects between agents in the continuous case. Moreover, the histogram approach does not scale well with the dimensionality of the feature space. 4.4.3 Radial Basis Functions A specific problem of the histogram approach is the hard assignment of agents into bins, which results in abrupt changes in the observation space when a neighboring agent moves from one bin to another. A more fine-grained representation can be achieved by using RBF networks with a fixed number of basis functions evenly distributed over the observation space. The resulting feature mapping φRBF is then defined by the activations of each basis function and can be seen as a "soft-assigned" histogram. However, both representations (histogram and RBF) suffer from the curse of dimensionality, as the number of required basis functions typically increases exponentially with the number of dimensions of the observation vector. 4.5 Other Representation Techniques Inspired by the work of Mordatch and Abbeel (2018), we also investigate a policy function that uses a softmax pooling layer instead of the mean embedding. The elements of the pooling layer ψ = [ψ1, . . . , ψK] are given by (cid:80) ψk = oi,j∈Oi exp(cid:0)αφk(oi,j)(cid:1) φk(oi,j) (cid:80) oi,j∈Oi exp (αφk(oi,j)) 10 neuralnetworkembedding64oi,164oi,2φNNOi64oi,N...64actionoiloc64φRBF/φHIST64actionoiloc64concat(Oi)64actionoilocDeep Reinforcement Learning for Swarm Systems for each feature dimension of φ = [φ1, . . . , φK] with a temperature parameter α. Note that the representation becomes identical to our mean embedding for α = 0, while setting α (cid:29) 1 results in max-pooling and α (cid:28) −1 corresponds to min-pooling. In our experiments, we choose α = 1 as a trade-off between a mean embedding and max-pooling and additionally study the performance of max-pooling over each individual feature dimension. 4.6 Adaption of TRPO to the Homogeneous Swarm Setup Gupta et al. (2017) present a parameter-sharing variant of TRPO that can be used in a multi- agent setup. During the learning phase, the algorithm collects experiences made by all agents and uses these experiences to optimize one policy with a single set of parameters θ. Since, in the swarm setup, we assume homogeneous agents that are potentially indistinguishable to each other, we omit the agent index introduced by Gupta et al. (2017). The optimization problem is expressed using advantage values based on all agents' observations. During execution, however, each agent has only access to its own perception. Hence, the terminology of centralized -- learning / decentralized -- execution is chosen. During the trajectory roll-outs, we use a sub-sampling strategy to achieve a trade-off between the number of samples and the variability in advantage values seen by the learning algorithm. Our implementation is based on the OpenAI baselines version of TRPO with 10 MPI workers, where each worker samples 2048 time steps, resulting in 2048N samples. Subsequently, we randomly choose the data of 8 agents, yielding 2048 × 10 × 8 = 163840 samples per TRPO iteration. The chosen number of samples worked well throughout our experiments and was not extensively tuned. 5. Experimental Results Our experiments are designed to study the use of mean embeddings in a cooperative swarm setting. The three main aspects are: 1. How do the different mean embeddings (neural networks, histograms and RBF repre- sentation) compare when provided with the same state information content? 2. How does the mean embedding using neural networks perform when provided with additional state information while keeping the dimensionality of the feature space constant? 3. How does the mean embedding of neural network features compare against other pool- ing techniques? of K trajectories are sampled, each yielding a return of Gk = (cid:80)T In this section, we first introduce the swarm model used for our experiments and present the results of different evaluations afterwards. During a policy update, a fixed number t=1 r(t). The results are k=1 Gk. Videos demonstrating (cid:80)K presented in terms of the average return, denoted as ¯G = 1 K the agents' behavior in the different tasks can be found online.2 2. http://computational-learning.net/deep_rl_for_swarms 11 Hüttenrauch, Šošić and Neumann 5.1 Swarm Models Our agents are modeled as unicycles (a commonly used agent model in mobile robotics; see, for example, Egerstedt and Hu, 2001), where the control parameters either manipulate the linear and angular velocities v and ω (single integrator dynamics) or the corresponding accelerations v and ω (double integrator dynamics). In the single integrator case, the state of an agent is defined by its location x = (x, y) and orientation φ. In case of double integrator dynamics, the agent is additionally characterized by its current velocities. The exact state definition and kinematic models can be found in Appendix A. Note that these agent models are more complex than what is typically considered in optimization-based approaches, which mostly assume single integrator dynamics directly on x. Depending on the task, we either opt for a closed state space where the limits act as walls, or a periodic toroidal state space where agents exceeding the boundaries reappear on the opposite side of the space. Either way, the state is bounded by xmax = ymax = 100. We study two different observation scenarios for the agents, i.e., global observability and local observability. In the case of global observability, all agents are neighbors, i.e. NG(i) = {j ∈ {1, . . . , N} i (cid:54)= j}, (cid:112) which corresponds to a fully connected static interaction graph. For the local observabil- ity case, we use ∆-disk proximity graphs, where edges are formed if the distance di,j = (xi − xj)2 + (yi − yj)2 between agents i and j is less than a pre-defined cut-off distance dc for communication, resulting in a dynamic interaction graph. The neighborhood set of the graph is then defined as NG(i) = {j ∈ {1, . . . , N} i (cid:54)= j, di,j ≤ dc}. For a detailed description of all observational features available to the agents in the tasks, see Appendices B and C. 5.2 Rendezvous In the rendezvous problem, the goal is to minimize the distances between all agents. The reason why we choose this experiment is because a simple optimization-based baseline con- troller can be defined by the consensus protocol, xi = − (cid:88) j∈N (i) (xi − xj), where xi = (xi, yi) denotes the location of agent i. To make the solution compatible to the double integrator agent model, we make use of a PD-controller (see Appendix A for details). The reward function for the problem can be found in Appendix E.1. We evaluate different observation vectors oi,j which are fed into the policy. To compare the histogram and RBF embedding with the proposed neural network approach, we restrict the basic observation model (see below) to a set of two features: the distance di,j between two agents and the corresponding bearing φi,j. This restriction allows for a comparison to the optimization-based consensus protocol, which is based on displacements (an equivalent formulation of distance and bearing). To show that the neural network embeddings can 12 Deep Reinforcement Learning for Swarm Systems Figure 2: Illustration of two neighboring agents facing the direction of their velocity vectors νi and νj, along with the observed quantities, shown with respect to agent i. The observed quantities are the bearing φi,j to agent j, agent j's relative orientation θi,j to agent i, their distance di,j and a relative velocity vector ∆νi,j = νi− νj. In this trivial example, agent i's observed neighborhood size as well as the neighborhood size communicated by agent j are N (i) = N (j) = 1. be used with more informative observations, we further introduce an extended set and a communication (comm) set. These sets may include relative orientations θi,j or relative velocities ∆νi,j (depending on the agent dynamics), as well as the own neighborhood size and those of the neighbors. An illustration of these quantities can be found in Figure 2. 5.2.1 Global Observability First, we study the rendezvous problem with 20 agents in the global observability setting with double integrator dynamics to illustrate the algorithm's ability to handle complex dynamics. To this end, we compare the performances of policies using histogram, RBF and neural network embeddings on the basic set, as well as neural network embeddings on the extended set. The observations oi,j in the basic set comprise the distance di,j and bearing φi,j. In the extended set, which is processed only via neural network embeddings, we additionally add neighboring agents' relative orientations θi,j and velocities ∆νi,j. The local properties oi loc consist of a shortest distance and orientation to the closest boundary, i.e., di wall. The sets are summarized as follows: wall and φi Basic : Extended : oi,j = {di,j, φi,j} oi,j = {di,j, φi,j, θi,j, ∆νi,j} loc = {di oi loc = {di oi wall, φi wall, φi wall} wall}. The results are shown in Figure 3a. On first sight, they reveal that all shown methods even- tually find a successful strategy, with the histogram approach showing worst performance. Upon a closer look, it can be seen that the best solutions are found with the neural network embedding, in which case the learning algorithm also converges faster, demonstrating that this form of embedding serves as a suitable representation for deep RL. However, there are two important things to note: • The differences between the approaches seem to be small due to the wide range of obtained reward values, but the NN+ method brings in fact a significant performance 13 agentiagentj∆νi,jdi,jνiνjφi,jθi,jHüttenrauch, Šošić and Neumann (a) 20 agents with global observability (b) 20 agents with local observability Figure 3: Learning curves for the rendezvous task with different observation models. The curves show the median of the average return ¯G based on the top five trials on a log scale. Legend: NN++: neural network mean embedding of comm set, NN+: neural network mean embedding of extended set, NN: neural network embedding of basic set, RBF: radial basis function embedding of basic set, HIST: histogram embedding of basic set, CONCAT+: simple concatenation of extended set. gain. Compared to the NN and RBF embedding, the performance of the learned NN+ policy is ∼10% better in terms of the average return of an episode (Figure 3a) and almost twice as good (∼4 × 10−2 versus ∼8 × 10−2) in terms of the mean distance between agents at the steady state solution after around 200 time steps (Figure 5a). Furthermore, the NN+ embedding reaches the mean distance achieved by the NN and RBF embeddings roughly 20 to 30 time steps earlier, which corresponds to an improvement of ∼25%. • Although the performance gain of NN+ can be partly explained by the use of the extended feature set, experiments with the same feature set using the histogram / RBF approach did not succeed to find solutions to the rendezvous problem; hence, the corresponding results are omitted. The reason is that the dimensionality of the input space scales exponentially for the histogram / RBF approach while only linearly for the neural network embedding, which results in a more compact feature representation that keeps the learning problem tractable. Together, these two observations suggest that the neural network embedding provides a suitable learning architecture for deep RL, whereas the histogram / RBF approach is only suited for low-dimensional spaces. Figure 4 shows a visualization of a policy using the neural network mean embedding of the extended set. After random initialization, the agents' locations quickly converge to a single point. Figure 5 shows performance evaluations of the best policies found with each of the mean embedding approaches. We plot the evolution of the mean distance between all agents over 1000 episodes with equal starting conditions. We also include the performance of the PD-controller defined in Appendix A. It can be seen in Figures 5a and 5c that the policies using the neural network embeddings decrease the mean distance most quickly and also find 14 0100200300400500-500-100-50-10TRPOiterationsaveragereturn¯GNN+NNRBFHISTCONCAT+0100200300400500-1000-500-100-50TRPOiterationsaveragereturn¯GNN++NN+RBFDeep Reinforcement Learning for Swarm Systems (a) snapshots (b) full episode Figure 4: Visualization of a learned policy for the pursuit evasion task. The policy is learned and executed by 10 agents using a neural network mean embedding of the extended set. Pursuers are illustrated in blue, the evader is highlighted in red. Visualization of a learned policy for the rendezvous task. The policy is learned and executed by 20 agents using a neural network mean embedding of the extended set. the best steady-state solutions among all learning approaches. While the optimization-based solution (PD) eventually drives the mean distance to zero, a small error remains for the learning-based approaches. However, the learned policies are faster in reducing the distance and therefore show a better average reward. Although the optimization-based policy is guaranteed to find an optimal stationary solution, the approach is build for simpler dynamics and hence performs suboptimally in the considered scenario. Note, that the controller gains for this approach have been tuned manually to maximize performance. In order to show the generalization abilities of the embeddings, we finally evaluate the obtained policies (except for the concatenation) with 100 agents. The results are displayed in Figure 5b. Again, the neural network embedding of the extended set is quickest in reducing the inter-agent distances, resulting in the best overall performance. 5.2.2 Local Observability The local observability case is studied with 20 agents and a communication cut-off distance of dc = 40. Due to the increased difficulty of the task, we resort to single integrator dynamics for this experiment. Again, we evaluate the basic and the extended set, which in this case contains the single integrator state information. Accordingly, we remove the relative velocities from the information sets. Moreover, we employ a local communication strategy that transmits the number of observed neighbors as additional information. Note that this information can be used by the agents to estimate in which direction the center of mass of the swarm is located. 15 t=0t=30t=50t=90Hüttenrauch, Šošić and Neumann While the received neighborhood sizes {N (j)}j∈N (i) are treated as part of agent i's local observation of the swarm, the own perceived neighborhood size N (i) is considered as part of the local features oi loc. The observation models for the local observability case are thus summarized as: Basic : Extended : Comm : oi,j = {di,j, φi,j} oi,j = {di,j, φi,j, θi,j} oi,j = {di,j, φi,j, θi,j, N (j)} loc = {di oi loc = {di oi loc = {di oi wall, φi wall, φi wall, φi wall} wall} wall, N (i)}. For the experiment, we limit our comparison to RBF embeddings (which showed best performance among all non-neural-network solutions) of the basic set and neural network embeddings of the extended set and the comm set. The results are illustrated in Figure 3b, which shows that the neural network embeddings lead to a quicker learning progress. Fur- thermore, by introducing the comm model, a higher return is achieved. Compared to the global observability case, however, the learning process exhibits an increased variance caused by the information loss in the reward signal (see Appendix E). Figure 5c illustrates the performances of the learned policies. Again, the neural network embedding is quicker in reducing the inter-agent distances and converges to better steady- state solutions. In order to test the efficacy of the communication protocol, we further evaluate the learned policies with 10 agents. The results are displayed in Figure 5d. As expected, the performance decreases due to the lower chance of agents seeing each other but we still notice a benefit caused by the communication. 5.3 Pursuit Evasion with a Single Evader Our implementation of the pursuit evasion scenario is based on the work by Zhou et al. (2016), from which we adopt the evader strategy. The strategy is based on Voronoi regions, which the pursuers try to minimize and the evader tries to maximize. While the original paper considers a closed world, we change the world type from closed to periodic, thereby making it impossible to trap the evader in a corner. In order to encourage a higher level of coordination between the agents, we set the evader's maximum velocity to twice the pursuers' maximum velocity. An episode ends once the evader is caught, i.e., if the distance of the closest pursuer is below a certain threshold. In all our experiments, the evader policy is fixed and not part of the learning process. The reward function for the problem is based on the shortest distance of the closest pursuer and can be found in Appendix E.2. 5.3.1 Global Observability Again, we study the global observability case with ten agents. Since the pursuit of an evader is a more challenging task already, we reduce the movement complexity to single integrator dynamics. The basic and extended set are equal to those in the rendezvous experiment with single integrator dynamics, with additional information about the evader in the local properties oi loc. In here, we add the distance di,e and bearing φi,e of agent i to the evader e. Accordingly, the observation sets are given as: Basic : Extended : oi,j = {di,j, φi,j} oi,j = {di,j, φi,j, θi,j} 16 loc = {di oi loc = {di oi wall, φi wall, φi wall, di,e, φi,e} wall, di,e, φi,e}. Deep Reinforcement Learning for Swarm Systems (a) 20 agents (global observability) (b) 100 agents (global observability) (c) 20 agents (local observability) (d) 10 agents (local observability) Figure 5: Comparison of the mean distance between agents in the rendezvous experiment achieved by the best learned policies and the consensus protocol. In (a) and (b), the policy is learned with 20 agents and executed by 20 and 100 agents, respectively. In (c) and (d), the policy is learned with 20 agents and executed by 20 and 10 agents. Results are averaged over 1000 episodes with identical starting conditions. The results in Figure 6a reveal that successful strategies can be obtained with all meth- ods. However, this time, a clear advantage can be seen for the policies using neural network mean embeddings of the extended set, both in terms of behavior quality and in the number of samples necessary to find the solution. Figure 7 illustrates the strategy that such a policy exerts. After random initialization, the agents first spread in a way that leaves no possibility for the evader to increase its Voronoi region, thereby keeping the evader almost on the same spot. Once this configuration is reached, they surround the evader in a circular pattern and start to reduce the distance until one pursuer successfully reaches the distance threshold. To investigate the performance of the best mean embedding policies (learned with 10 agents), we estimate the corresponding probabilities that the evader is caught within a cer- tain time frame. For the sake of completeness, we also include the method proposed by Zhou 17 010020030040050010−210−1100101102tmeandistanceNN+NNRBFHISTCONCAT+Consensus010020030040050010−210−1100101102tmeandistanceNN+NNRBFHISTConsensus02004006008001,00010−1100101102tmeandistanceNN++NN+RBF02004006008001,00010−1100101102tmeandistanceNN++NN+RBFHüttenrauch, Šošić and Neumann (a) 10 agents with global observability (b) 20 agents with local observability Figure 6: Learning curves for the pursuit evasion task with different observation models. The curves show the median of the average return ¯G based on the top five trials on a log scale. Legend: NN++: neural network mean embedding of comm set, NN+: neural network mean embedding of extended set, RBF: radial basis function embedding of basic set, HIST: histogram embedding of basic set, CONCAT+: concatenation of extended set. (a) snapshots (b) full episode Figure 7: Visualization of a learned policy for the pursuit evasion task. The policy is learned and executed by 10 agents using a neural network mean embedding of the extended set. Pursuers are illustrated in blue, the evader is highlighted in red. 18 0100200300400500-100-50-10TRPOiterationsaveragereturn¯GNN+NNRBFHISTCONCAT+02004006008001,000-1000-500-100-50TRPOiterationsaveragereturn¯GNN++NN+RBFt=0t=100t=110t=121Deep Reinforcement Learning for Swarm Systems et al. (2016), which was originally not designed for a setup with a faster evader, though. The results are plotted in Figure 8 as the fraction of episodes ending at the respective time instant, averaged over 1000 episodes. The plot in Figure 8b reveals that the evader may be caught using all presented methods if the policies are executed for long time periods. As already indicated by the learning curves, using a neural network mean embedding represen- tation yields the quickest capture among all methods. The additional information in the extended set further increases performance. Next, we examine the generalization abilities of the learned policies, this time on scenarios with 5, 20 and 50 agents (Figures 8a,c,d). Increasing the amount of agents leads to a quicker capture for all methods; however, the best performance is still shown by the agents executing a neural network policy based on embeddings of the extended set. Interestingly, when using fewer agents than in the original setup (Figure 8a), all methods struggle to capture the evader. After inspection of the behavior, we found that the strategy of establishing a circle around the evader causes too large gaps between the agents through which the evader can escape. 5.3.2 Local Observability The local observability case is studied with 20 agents and a communication cut-off distance of dc = 40. Additionally, we introduce an observation radius do = 20 within which the pursuers can observe the distance and bearing to the evader. We reuse the basic and extended set from last section and modify the comm set to include the shortest path information of other agents in the neighborhood of agent i to the evader. This way, each agent i can compute a shortest path to the evader over a graph of connected agents, such that the path P = (v1, v2, . . . , vM ) minimizes the sum di,e m=1 dm,m+1 where v1 represents agent i and vM is the evader. The observation sets are given as: min =(cid:80)M−1 Basic : Extended : Comm : oi,j = {di,j, φi,j} oi,j = {di,j, φi,j, θi,j} oi,j = {di,j, φi,j, θi,j, dj,e min} loc = {di oi loc = {di oi loc = {di oi wall, φi wall, φi wall, φi wall, di,e, φi,e} wall, di,e, φi,e} wall, di,e, φi,e, di,e min}. Note that in this case the distance and bearing to an evader are only available if di,e ≤ do. Furthermore, the correct shortest path is only available if an agent and the evader are in the same sub-graph, otherwise, a pre-defined value is fed into the policy. Again, we limit the comparison for the local observability case to the more promising methods of neural network and RBF mean embeddings. The results in Figure 6b show that the performance gain of the neural network mean embeddings is even more noticeable than in the global observability case, with a clear advantage in the presence of the local commu- nication protocols. The inspection of the termination probabilities in Figure 9 confirms that the neural network mean embedding results in a significantly improved policy. 5.4 Pursuit Evasion with Multiple Evaders Lastly, we study a pursuit evasion scenario with multiple evaders, i.e., we assume that agent i receives observation samples {oi,e} from several evaders, which are processed using a second mean embedding to account for the variable set size. Where in the previous experiment 19 Hüttenrauch, Šošić and Neumann (a) 5 agents (b) 10 agents (c) 20 agents (d) 50 agents Figure 8: Performance comparison of the best learned policies and the optimization ap- proach minimizing Voronoi regions in the pursuit evasion task with global ob- servability. The curves show the probability that the evader is caught after t time steps. All policies are learned with 10 agents but executed with different agent numbers, as indicated below each subfigure. Results are averaged over 1000 episodes with identical starting conditions. the agents had precise information about the evader in terms of distance and bearing, they now have to extract this information from the respective embedding. An additional level of difficulty results from the fact that the reward function no longer provides any guidance in terms of the distances to the evaders since it only counts the number of evaders caught in each time step (see Appendix E.3 for details). We study a scenario with 50 pursuers and 5 evaders using the global observability setup in Section 5.3.1, except that we respawn caught evaders to a new random location instead of terminating the episode. The observation sets, containing the same type of information but arranged according to the inputs of the neural networks, are designed as follows: wall} wall}. oi,j = {di,j, φi,j} oi,j = {di,j, φi,j, θi,j} oi,e = {di,e, φi,e} oi,e = {di,e, φi,e} Basic : Extended : loc = {di oi loc = {di oi wall, φi wall, φi 20 02004006008001,0000.00.20.40.60.81.0tp(evadercaught)NN+NNRBFHISTVoronoi01002003000.00.20.40.60.81.0tp(evadercaught)NN+NNRBFHISTVoronoiConcat0501001500.00.20.40.60.81.0tp(evadercaught)NN+NNRBFHISTVoronoi0204060801001200.00.20.40.60.81.0tp(evadercaught)NN+NNRBFHISTVoronoiDeep Reinforcement Learning for Swarm Systems Figure 9: Performance comparison of the best policies in the pursuit evasion task with local observability. The curves show the probability that the evader is caught after t time steps. All policies are learned and executed by 20 agents. Results are averaged over 1000 episodes with identical starting conditions. Figure 10a shows the learning curves for policies with neural network and RBF mean embeddings and for the concatenation approach. The return directly relates to the number of evaders caught during an episode. Again, the neural network mean embedding performs significantly better than the RBF embedding. The curves clearly indicate the positive influ- ence of the additional information present in the extended set. With this amount of agents, the dimensionality of the concatenation has increased to a point where learning is no longer feasible. 5.5 Evaluation of Pooling Functions Figure 11 shows learning curves of policies based on mean embeddings, softmax pooling, and max-pooling (as described in Section 4.5) of features of the extended set for the rendezvous and pursuit evasion task with global observability. In the rendezvous task (Figure 11a), all pooling techniques eventually manage to find a good solution. Policies using neural network mean embedding, however, on average converge more quickly while policies using max-pooling show slightly worse performance. Given its reduced computational complexity compared to the softmax-pooling, the mean embedding provides the most effective approach among all proposed architectures. When examining the results of the pursuit evasion task (Figure 11b), we find that the algorithm produces two distinct solutions. A sub-optimal one, which is only able to circle the evader but is unable to catch it (a catch is realized if the distance of the closest pursuer to the evader is below a certain threshold), and a solution which additionally catches the evader after a short period of time. Therefore, we not only report the performance of the top 5 trials out of 16, but also provide the number of times the algorithm was able to discover the better of the two solution (Table 1). Once that the algorithm finds a good solution, the mean embedding and softmax solutions perform comparably well but the max-pooling approach shows a significantly worse performance. More importantly, however, the algorithm was 21 01002003004000.00.20.40.60.81.0tp(evadercaught)NN++NN+RBFHüttenrauch, Šošić and Neumann (a) Evaluation of different observation models and embedding techniques. (b) Comparison of the proposed mean embedding policy to policies with statis- tics of the observations as input. Figure 10: Learning curves for 50 agent pursuit evasion with 5 evaders. The curves show the median of the average return ¯G based on the top five trials. Legend: NN+ 2x: two neural network mean embeddings of the extended set, RBF 2x: two radial basis function mean embeddings of the basic set, concat: simple concatenation of extended set. MSSK+, MSS+, MS+ and M+: Combinations of mean, standard deviation, skew and kurtosis of the features in the extended set. (a) Rendezvous with 20 agents and global observability. (b) Pursuit evasion with 10 pursuers and global observability. Figure 11: Learning curves of different embedding and pooling architectures based on the extended set. The curves show the median of the average return ¯G based on the top five trials on a log scale. Legend: MEAN: neural network mean embedding, SM: softmax feature pooling, MAX: max feature pooling. able to find the good solution more often using the mean embedding than using the other pooling approaches. 22 010020030040050002040TRPOiterationsaveragereturn¯GNN+2xRBF2xconcat010020030040050002040TRPOiterationsaveragereturn¯GNN+MSSK+MSS+MS+M+050100150200250-100-50TRPOiterationsaveragereturn¯GMEANSMMAX0100200300400500-100-50-10TRPOiterationsaveragereturn¯GMEANSMMAXDeep Reinforcement Learning for Swarm Systems mean 10/16 sm max 6/16 4/16 Table 1: Number of times the algorithm discovered policies that led to a successful catch. 5.6 Comparison to Moment-Based Representations Finally, we compare the mean embedding of neural network features to a representation using statistics of the input. Figure 10b shows an evaluation on the pursuit evasion task with 50 agents and 5 evaders. Here, we use combinations of the empirical mean, standard deviation, skew and kurtosis of each feature of the extended set as the input to a policy function. The plot reveals that neural network mean embeddings can capture more relevant information about the characteristics of the distribution of agents than simple statistics of the elements in the extended set. Similar results were obtained for the other tasks although the differences in performance were less pronounced. 5.7 Computational Complexity Unlike in classical optimization-based control, where the controller is derived from an as- sumed dynamics model, model-free reinforcement learning methods like TRPO find their control policies through interaction with the environment, without requiring explicit knowl- edge of the underlying system dynamics. While this comes at the cost of an additional exploration phase, learning-based approaches typically offer an increased flexibility in that the same control architecture can adapt to different tasks and environments, without being affected by potential model mismatches. More importantly, considering the final learned policy from a computational perspective, the synthesis of the control signal involves no additional conceptual steps compared to an optimization-based approach. While a typical experiment with 20 agents in our setup takes between four and six hours of training on a machine with ten cores (sampling trajectories in parallel), a forward pass through the trained neural network to compute the instantaneous control signal takes only about 1 ms, which enables an execution in real time. Furthermore, all control strategies learned through our framework are decentralized, which allows an arbitrary system size scaling in a real swarm network, where the required computations are naturally distributed over all agents. When learning new policies, the memory requirements scale O(N (N − 1)) with the number of agents (assuming global observability) since we need to store the local views of all agents. However, decentralized execution after the policy is learned scales linearly in N per agent. An incremental online computation of the mean can be chosen if memory restrictions exist (Finch, 2009). For comparison, the complexity of calculating Voronoi regions for the pursuit evasion policy scales O(N log N ) with the number of agents (Aurenhammer, 1991). Concerning the system sizes considered in our experiments, the resulting computation time of both policy types is in the same order of magnitude during task execution. 23 Hüttenrauch, Šošić and Neumann 6. Conclusion In this paper, we proposed the use of mean feature embeddings as state representations to overcome two major problems in deep reinforcement learning for swarms: the high and possibly changing dimensionality of information perceived by each agent. We introduced three different approaches to realize such embeddings -- two manually designed approaches based on histograms / radial basis functions and an end-to-end learned neural network fea- ture representation. We evaluated the approaches on different variations of the rendezvous and pursuit evasion problem and compared their performances to that of a naive feature concatenation method and classical approaches found in the literature. Our evaluation re- vealed that learning embeddings end-to-end using neural network features scales well with increasing agent numbers, leads to better performing policies, and often results in faster con- vergence compared to all other approaches. As expected, the naive concatenation approach fails for larger system sizes. Acknowledgments The research leading to these results has received funding from EPSRC under grant agree- ment EP/R02572X/1 (National Center for Nuclear Robotics). Calculations for this research were conducted on the Lichtenberg high performance computer of the TU Darmstadt. Appendix A. Agent Kinematics In the single integrator case, the state of an agent is given by si = [xi, yi, φi] ∈ S = {[x, y, φ] ∈ R3 : 0 ≤ x ≤ xmax, 0 ≤ y ≤ ymax, 0 ≤ φ < 2π}, and the linear velocity v and angular velocity ω can be directly controlled by the agent. The kinematic model is given by x = v cos φ y = v sin φ φ = ω. In the double integrator case, the state is given by si = [xi, yi, φi, vi, ωi] ∈ S = {[x, y, φ, v, ω] ∈ R5 : 0 ≤ x ≤ xmax, 0 ≤ y ≤ ymax, 0 ≤ φ < 2π,v ≤ vmax,ω ≤ ωmax} and the agent can only indirectly change its velocity by acceleration. With the control inputs av and aω, the model is then given by v = av ω = aω x = v cos φ y = v sin φ φ = ω. For the experiments, we use finite differences to model the system in discrete time. 24 Deep Reinforcement Learning for Swarm Systems Appendix B. Observation Model Irrespective of the task, an agent i can sense the following properties about other agents j ∈ N (i) within its neighborhood: di,j φi,j = arctan( yj − yi xj − xi ) − φi yi − yj xi − xj ) − φj θi,j = arctan( ∆vi,j = vi(cid:2)cos φi, sin φi(cid:3) − vj(cid:2)cos φj, sin φj(cid:3) distance to neighboring agents bearing to neighboring agents relative orientation relative velocity Furthermore, each agent has access to the following local properties: wall = min(xi − xmin, yi − ymin, xmax − xi, ymax − yi) di wall = ϕi φi vi, ωi wall − φi distance to closest wall orientation to closest wall own velocity where ϕi wall denotes the absolute bearing of agent i to the closest wall segment. Appendix C. Task Specific Communication Protocols In the rendezvous task, agent i additionally can sense information about neighborhood sizes: N (i) N (j) : j ∈ N (i) own neighborhood size neighborhood size of neighbor j In pursuit evasion, we additionally have one or multiple evaders with states se = [xe, ye] ∈ {[x, y] ∈ R2 : 0 ≤ x ≤ xmax, 0 ≤ y ≤ ymax}. Agents can sense the distance and bearing to an evader, given that the evader is within an observation distance do: (cid:113) di,e = (xi − xe)2 + (yi − ye)2 if di,e ≤ do φi,e = arctan( ye − yi xe − xi ) − φi if di,e ≤ do distance to evader bearing to evader (cid:80)M−1 Furthermore, we assume that each agent i can compute a shortest path to the evader over a graph of connected agents, such that the path P = (v1, v2 . . . , vM ) minimizes the sum m=1 dm,m+1 where v1 is agent i and vM is the evader. Appendix D. Controller for Double Integrator Dynamics We use a simple PD-controller to transform the consensus protocol with high-level direct state manipulation to the unicycle model with double integrator dynamics. It is given by 25 Hüttenrauch, Šošić and Neumann av = K1(vd − v) aω = K2(φd − φ) + D2(ωd − ω) vd = (cid:107) x(cid:107) y x ) φd = arctan( ωd = 0, where the parameters K1, K2 and D2 are tuned manually to give good performance on the problem. Appendix E. Reward Functions E.1 Rendezvous The reward function is defined in terms of the inter-agent distances {di,j} as R(s, a) = α min(di,j, dc) + β(cid:107)a(cid:107), N(cid:88) N(cid:88) i=1 j=i+1 α = −(cid:16) N (N−1) where in the global observability case we set the cut-off distance dc = max(xmax, ymax) to the maximum possible inter-agent distance in the respective environment. The factor serves as a reward normalization factor and β = −1 × 10−3 controls (cid:17)−1 dc 2 how strongly high action outputs of the policy are penalized. E.2 Pursuit Evasion For the case of a single evader, the pursuit evasion objective may be expressed in terms of the distance to the closest pursuer. More specifically, the reward function is given as R(s, a) = − 1 do min(dmin, do), where dmin = min(d1,e, . . . , dN,e). For the global observability case, we set do to the maxi- mum possible distance of di,e . E.3 Pursuit Evasion with Multiple Evaders In the case of multiple evaders, we use a sparser reward function that counts how many evaders are caught per time step, with no additional guidance of inter-agent distances. An evader e is assumed to be caught if the closest pursuer's distance dmin,e = min(d1,e, . . . , dN,e) is closer than a threshold distance dt = 3. The reward function is given by R(s, a) = 1[0,dt](dmin,e), E(cid:88) e=1 26 Deep Reinforcement Learning for Swarm Systems (a) Activation functions evaluation. (b) Layer size evaluation. Figure 12: Learning curves for 20 agent rendezvous with (a) different activation functions for the mean embedding and (b) different layer numbers and sizes using a RELU activation function. The curves show the median of the average return ¯G based on the top five trials. where E is the number of evaders and 1[a,b](x) = is the indicator function. Appendix F. Policy Architectures (cid:40) 1 if x ∈ [a, b] 0 else This section briefly summarizes the chosen policy architectures. Illustrations can be found in Figure 1. F.1 Neural Network Embedding Policy We evaluated different layer sizes and activation functions on the rendezvous problem and show the results in Figure 12. In all other experiments, the neural network mean feature embedding for agent i, given by φNN(Oi) = 1 Oi φ(oi,j), (cid:88) oi,j∈Oi is realized as the empirical mean of the outputs of a single layer feed-forward neural network, φ(oi,j) = h(W oi,j + b), with 64 neurons and a RELU non-linearity h. 27 0100200300400500−18−16−14−12TRPOiterationsaveragereturn¯Grelutanhlinear0100200300400−18−16−14−12TRPOiterationsaveragereturn¯G[32][64][128][32,32][64,64]Hüttenrauch, Šošić and Neumann F.2 Histogram Embedding Policy The histogram embedding is achieved with a two-dimensional histogram over the distance and bearing space to other agents. We use eight evenly spaced bins for each feature, resulting in a 64 dimensional feature vector. F.3 RBF Embedding Policy The RBF embedding is given by a vector φRBF(Oi) =(cid:2)ψ1(Oi), . . . , ψM 2(Oi)(cid:3) of M 2 contri- butions from M = 8 radial basis functions whose center points are evenly distributed in the distance and bearing space. With oi,j = [di,j, φi,j], µm = [µd,m, µφ,m], and σ = [σd, σφ] its components are given by (cid:88) oi,j∈Oi ψm(Oi) = ρm(oi,j), where we choose (cid:34) (cid:32) − 1 2 ρm(oi,j) = exp (di,j − µd,m)2 σ2 d + (φi,j − µφ,m)2 σ2 φ (cid:35)(cid:33) . The policy network structure used for both, the histogram and the RBF representations, is illustrated in Figure 1b. F.4 Concatenation Policy For the concatenation method, we first concatenate agent i's neighborhood observations contained in the set Oi and process them with one hidden layer of 64 neurons and a RELU non-linearity. The resulting feature vector is then concatenated with the local properties oi loc and fed into a second layer of same size. Finally, the output of the second layer is mapped to the action. The corresponding policy network structure can be seen in Figure 1c. References Maruan Al-Shedivat, Trapit Bansal, Yura Burda, Ilya Sutskever, Igor Mordatch, and Pieter Abbeel. Continuous adaptation via meta-learning in nonstationary and competitive envi- ronments. In International Conference on Learning Representations, 2018. Franz Aurenhammer. Voronoi diagrams -- a survey of a fundamental geometric data struc- ture. ACM Computing Surveys (CSUR), 23(3):345 -- 405, 1991. Levent Bayındır. A review of swarm robotics tasks. Neurocomputing, 172:292 -- 321, 2016. Daniel S. Bernstein, Robert Givan, Neil Immerman, and Shlomo Zilberstein. The com- plexity of decentralized control of Markov decision processes. Mathematics of Operations Research, 27(4):819 -- 840, 2002. Mo Chen, Zhengyuan Zhou, and Claire J Tomlin. Multiplayer reach-avoid games via low dimensional solutions and maximum matching. In American Control Conference (ACC), pages 1444 -- 1449, 2014a. 28 Deep Reinforcement Learning for Swarm Systems Mo Chen, Zhengyuan Zhou, and Claire J Tomlin. A path defense approach to the multiplayer In IEEE 53rd Annual Conference on Decision and Control (CDC), reach-avoid game. pages 2420 -- 2426, 2014b. Mo Chen, Zhengyuan Zhou, and Claire J Tomlin. Multiplayer reach-avoid games via pairwise outcomes. IEEE Transactions on Automatic Control, 62(3):1451 -- 1457, 2017. Timothy H. Chung, Geoffrey A. Hollinger, and Volkan Isler. Search and pursuit-evasion in mobile robotics. Autonomous Robots, 31(4):299, 2011. D. V. Dimarogonas and K. J. Kyriakopoulos. On the rendezvous problem for multiple nonholonomic agents. IEEE Transactions on Automatic Control, 52(5):916 -- 922, 2007. M. Egerstedt and Xiaoming Hu. Formation constrained multi-agent control. IEEE Trans- actions on Robotics and Automation, 17(6):947 -- 951, 2001. Tony Finch. Incremental calculation of weighted mean and variance. University of Cam- bridge, 4:11 -- 5, 2009. G. H. W. Gebhardt, K. Daun, M. Schnaubelt, and G. Neumann. Learning robust policies for object manipulation with robot swarms. In IEEE International Conference on Robotics and Automation, 2018. M. C. De Gennaro and A. Jadbabaie. Decentralized control of connectivity for multi-agent systems. In IEEE Conference on Decision and Control, pages 3628 -- 3633, 2006. Arthur Gretton, Kenji Fukumizu, Choon H Teo, Le Song, Bernhard Schölkopf, and Alex J In Advances in Neural Information Smola. A kernel statistical test of independence. Processing Systems, pages 585 -- 592, 2008. Arthur Gretton, Karsten M. Borgwardt, Malte J. Rasch, Bernhard Schölkopf, and Alexander Smola. A kernel two-sample test. Journal of Machine Learning Research, 13(Mar):723 -- 773, 2012. Aditya Grover, Maruan Al-Shedivat, Jayesh K Gupta, Yura Burda, and Harrison Edwards. Learning policy representations in multiagent systems. arXiv:1806.06464, 2018. Jayesh K. Gupta, Maxim Egorov, and Mykel Kochenderfer. Cooperative multi-agent control using deep reinforcement learning. In International Conference on Autonomous Agents and Multiagent Systems, pages 66 -- 83, 2017. Julia Handl and Bernd Meyer. Ant-based and swarm-based clustering. Swarm Intelligence, 1(2):95 -- 113, 2007. Matthew Hausknecht and Peter Stone. Deep recurrent Q-learning for partially observable MDPs. In AAAI Fall Symposium Series, 2015. Maximilian Hüttenrauch, Adrian Šošić, and Gerhard Neumann. Local communication pro- tocols for learning complex swarm behaviors with deep reinforcement learning. In Inter- national Conference on Swarm Intelligence, 2018. 29 Hüttenrauch, Šošić and Neumann A. Jadbabaie, Jie Lin, and A. S. Morse. Coordination of groups of mobile autonomous agents using nearest neighbor rules. IEEE Transactions on Automatic Control, 48(6):988 -- 1001, 2003. M. Ji and M. Egerstedt. Distributed coordination control of multiagent systems while pre- serving connectedness. IEEE Transactions on Robotics, 23(4):693 -- 703, 2007. 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. arXiv:1509.02971, 2015. J. Lin, A. Morse, and B. Anderson. The multi-agent rendezvous problem. Part 2: The asynchronous case. SIAM Journal on Control and Optimization, 46(6):2120 -- 2147, 2007. Zhiyun Lin, M. Broucke, and B. Francis. Local control strategies for groups of mobile autonomous agents. IEEE Transactions on Automatic Control, 49(4):622 -- 629, 2004. Ryan Lowe, YI WU, Aviv Tamar, Jean Harb, OpenAI Pieter Abbeel, and Igor Mordatch. Multi-agent actor-critic for mixed cooperative-competitive environments. In Advances in Neural Information Processing Systems, pages 6379 -- 6390, 2017. L. Matignon, G. J. Laurent, and N. L. Fort-Piat. Hysteretic Q-learning: An algorithm for decentralized reinforcement learning in cooperative multi-agent teams. In IEEE/RSJ International Conference on Intelligent Robots and Systems, pages 64 -- 69, 2007. 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, 518:529 -- 533, 2015. Igor Mordatch and Pieter Abbeel. Emergence of grounded compositional language in multi- agent populations. In AAAI Conference on Artificial Intelligence, 2018. Shayegan Omidshafiei, Jason Pazis, Christopher Amato, Jonathan P. How, and John Vian. Deep decentralized multi-task multi-agent reinforcement learning under partial observ- ability. In International Conference on Machine Learning, pages 2681 -- 2690, 2017. Gregory Palmer, Karl Tuyls, Daan Bloembergen, and Rahul Savani. Lenient multi-agent deep reinforcement learning. arXiv:1707.04402, 2017. Liviu Panait, Keith Sullivan, and Sean Luke. Lenient learners in cooperative multiagent sys- tems. In International Joint Conference on Autonomous Agents and Multiagent Systems, pages 801 -- 803, 2006. B. Ranjbar-Sahraei, F. Shabaninia, A. Nemati, and S. D. Stan. A novel robust decentralized adaptive fuzzy control for swarm formation of multiagent systems. IEEE Transactions on Industrial Electronics, 59(8):3124 -- 3134, 2012. 30 Deep Reinforcement Learning for Swarm Systems Tabish Rashid, Mikayel Samvelyan, Christian Schroeder de Witt, Gregory Farquhar, Jakob Foerster, and Shimon Whiteson. QMIX: Monotonic value function factorisation for deep multi-agent reinforcement learning. arXiv:1803.11485, 2018. Andrei A Rusu, Sergio Gomez Colmenarejo, Caglar Gulcehre, Guillaume Desjardins, James Kirkpatrick, Razvan Pascanu, Volodymyr Mnih, Koray Kavukcuoglu, and Raia Hadsell. Policy distillation. arXiv:1511.06295, 2015. Muhammad Saleem, Gianni A Di Caro, and Muddassar Farooq. Swarm intelligence based routing protocol for wireless sensor networks: Survey and future directions. Information Sciences, 181(20):4597 -- 4624, 2011. John Schulman, Sergey Levine, Pieter Abbeel, Michael Jordan, and Philipp Moritz. Trust region policy optimization. In International Conference on Machine Learning, pages 1889 -- 1897, 2015. Alex Smola, Arthur Gretton, Le Song, and Bernhard Schölkopf. A Hilbert space embedding In International Conference on Algorithmic Learning Theory, pages for distributions. 13 -- 31, 2007. Adrian Šošić, Wasiur R KhudaBukhsh, Abdelhak M Zoubir, and Heinz Koeppl. Inverse reinforcement learning in swarm systems. In International Conference on Autonomous Agents and Multiagent Systems, pages 1413 -- 1421, 2017. Peter Sunehag, Guy Lever, Audrunas Gruslys, Wojciech Marian Czarnecki, Vinicius Zam- baldi, Max Jaderberg, Marc Lanctot, Nicolas Sonnerat, Joel Z Leibo, Karl Tuyls, and Thore Graepel. Value-decomposition networks for cooperative multi-agent learning. arXiv:1706.05296, 2017. Yaodong Yang, Rui Luo, Minne Li, Ming Zhou, Weinan Zhang, and Jun Wang. Mean field multi-agent reinforcement learning. arXiv:1802.05438, 2018. Manzil Zaheer, Satwik Kottur, Siamak Ravanbakhsh, Barnabas Poczos, Ruslan R Salakhut- dinov, and Alexander J Smola. Deep sets. In Advances in Neural Information Processing Systems, pages 3391 -- 3401, 2017. Lianmin Zheng, Jiacheng Yang, Han Cai, Weinan Zhang, Jun Wang, and Yong Yu. MA- gent: A many-agent reinforcement learning platform for artificial collective intelligence. arXiv:1712.00600, 2017. Zhengyuan Zhou, Ryo Takei, Haomiao Huang, and Claire J Tomlin. A general, open-loop In IEEE 51st Annual Conference on Decision and formulation for reach-avoid games. Control (CDC), pages 6501 -- 6506, 2012. Zhengyuan Zhou, Wei Zhang, Jerry Ding, Haomiao Huang, Dušan M Stipanović, and Claire J Tomlin. Cooperative pursuit with Voronoi partitions. Automatica, 72:64 -- 72, 2016. 31
1707.06492	1	1707	2017-07-20T13:11:16	Effects of Network Structure on the Performance of a Modeled Traffic Network under Drivers' Bounded Rationality	[ "cs.MA", "nlin.AO" ]	We propose a minority route choice game to investigate the effect of the network structure on traffic network performance under the assumption of drivers' bounded rationality. We investigate ring-and-hub topologies to capture the nature of traffic networks in cities, and employ a minority game-based inductive learning process to model the characteristic behavior under the route choice scenario. Through numerical experiments, we find that topological changes in traffic networks induce a phase transition from an uncongested phase to a congested phase. Understanding this phase transition is helpful in planning new traffic networks.	cs.MA	cs	Effects of Network Structure on the Performance of a Modeled Traffic Network under Drivers' Bounded Rationality Toru Fujinoa, Yu Chena aGraduate School of Frontier Sciences, University of Tokyo, 5-1-5, Kashiwanoha, Kashiwa, Chiba 277-8561, Japan Abstract We propose a minority route choice game to investigate the effect of the network structure on traffic network performance under the assumption of drivers' bounded rationality. We investigate ring-and-hub topologies to cap- ture the nature of traffic networks in cities, and employ a minority game- based inductive learning process to model the characteristic behavior under the route choice scenario. Through numerical experiments, we find that topological changes in traffic networks induce a phase transition from an un- congested phase to a congested phase. Understanding this phase transition is helpful in planning new traffic networks. Keywords: Agent-based simulation, Traffic networks, Route-choice problems, Bounded rationality 1. Introduction There are various types of traffic networks, e.g., road traffic networks, railway networks, and aviation networks. These underlying infrastructures support numerous social and economic activities. Hence, understanding the complex behavior and performance of traffic systems based on such networks has attracted considerable attention across a wide range of disciplines, in- cluding traffic engineering [1, 2, 3], economics [4], physics [5], and computer science [6]. Email addresses: [email protected] (Toru Fujino), [email protected] (Yu Chen) Preprint submitted to Physica A May 25, 2021 The main difficulty in analyzing traffic network performance lies in devi- ations from the user equilibrium (UE), in which drivers in traffic networks are assumed to be completely rational and have perfect information about their route choice behavior. In real situations, the above two assumptions of UE are invalid. It is known from the Selten's study [4] that UE can not be reached even in an idealized system, where drivers have the same origin and destination pair and only two routes to choose from. Laboratory exper- iments in their study demonstrate that the number of drivers on each route fluctuates until the end of the experiment, even though the mean number of drivers on each route becomes close to the equilibrium. Agent-based models (ABMs) provide a good framework for dealing with the bounded rationality of drivers on abstract networks. In ABMs, drivers are represented by agents with limited intelligence. A number of studies have considered such a framework. Nakayama et al. [1] built a driver model with cognition and learning under the route-choice scenario, and they found that the network flow does not necessarily converge to UE, but may reach the deluded equilibrium, which can be derived from the drivers' false perception of the environment. Gourley et al. [7] used a Minority Game (MG) to for- mulate the agents' route choice behavior between a shortcut with variable congestion costs and a detour with a constant cost. One factor that has a nontrivial influence on the traffic network perfor- mance is the variation in traffic network structures. In the real world, there are numerous types of structured road networks, which makes it more difficult for drivers to select an optimal route among the alternatives. Many studies from this viewpoint have focused on the UE solution [8, 9, 10, 11, 12, 13]. These influential aspects of traffic network analysis, however, have not been combined in a unified framework, and so these analyses have not been comprehensive. Although previous studies on ABMs have incorporated the simple aspects of the drivers bounded rationality, another important aspect of bounded rationality, such as preference heterogeneity, has not yet been considered, and variations in network structures have not been investigated. In the previous researches on UE, the effects of network structures have been examined comprehensively, but these studies have not dealt with the drivers' bounded rationality. In this paper, we aim to overcome the deficiencies in existing approaches by building a new ABM. To deal with drivers bounded rationality and pref- erence heterogeneity, we borrow the strategy of MG and Market directed resource allocation game (MDRAG) [14]. To simulate variations in network 2 structures, we employ Gourleys ring-and-hub topology [7]. In addition, we explore the optimal network structure under these conditions for implication to plan new traffic networks. 2. Model 2.1. Overview of the simulation model The complexity of the route choice problem can be understood by the following scenario similarly explained by Gourley [7]. When you are the only traveler on a road traffic network, it is not difficult to determine the quickest route to get to your destination. However, in most cases, a large number of travelers must simultaneously choose a route along a common network to reach their respective destinations. This condition makes the traffic situation complicated. If too many travelers select the same shorter (or quicker) route, the route tends to become extremely congested. In such cases, making a detour may be a better choice. At the system level, if the travelers were perfectly rational and had perfect information, it would be possible to attain the UE state, wherein no driver could reduce the transportation cost by switching to a different route. However, as drivers are boundedly rational in reality, i.e., traffic information is incomplete and their capacity to evaluate routes is limited, it is impossible for them to make perfect decisions. In addition, each driver has a different preference for choosing his or her route, and this heterogeneity leads to complicated situations. Therefore, UE can rarely be attained in real situations. To tackle this complexity and achieve the research objectives, we con- struct a new agent-based model, based on the Gourley's model [7], with the extension of consideration of network topological changes. Additionally, we introduce preference heterogeneity [14] to Gourley's model in an attempt to capture the route choice behavior more realistically. Thus, the proposed model consists of two parts, a network model, and an agent's decision-making model. 2.2. Ring-and-hub network topology To express the route choice scenario which I described above, we employ a ring-and-hub network shown in Fig. 1. This network consists of N peripheral nodes, which are mutually connected by their nearest neighbors, and a central hub node, which is connected to only λ ≤ N peripheral nodes. Each node is assigned exactly one agent, so there are N agents on the network. Each 3 agent departs from her origin node and later arrives at her destination node, which is randomly assigned at the begining of the simulation. There are two possible routes by which each agent can reach their destination (indicated by blue and red lines in Fig. 1). One is an outside route that does not pass through the central hub, whereas the other is an inside route that passes through the central hub. The outside route can be considered as a detour, and the inside route is a shortcut. Fig. 1: Ring-and-hub network. The outside route is depicted by the blue line, and the inside route is depicted by the red line. In this modeled traffic network, the transportation costs for the outside and inside routes of agent n = 1, . . . , N at each time step t = 1, . . . , T are represented by C (n) out (t) is simply the number of links along the shortest peripheral path connecting the origin and the destination. C (n) in (t) depends on the number of agents in the central hub. The cost function is given as follows: in (t), respectively. C (n) out (t) and C (n) C (n) out (t) = d(O,D), C (n) in (t) = (cid:40) d(O,Hin) + α · d(Hin,Hout) + d(Hout,D) d(O,Hin) + β · d(Hin,Hout) + d(Hout,D) if Nin (t) ≤ L if Nin (t) > L. (1) (2) Here, α is a noncongestion coefficient and β is a congestion coefficient; L is the central hub capacity, which represents the limited number of agents who 4 ODHinHoutCnoutODHinHoutCnincan be handled by the hub without congestion, and Nin (t) is the number of central hub users at time step t. d(·,·) is the length of the shortest path on the peripheral path between nodes. O and D are the origin and desti- nation nodes, respectively. Hin and Hout are the interchange nodes that the agent can use to pass through the hub node along the inside route. In this experiment, we set α and β to 1 2, respectively. Thus, hub users may reach their destination at a lower cost under noncongestion, or at a higher cost under congestion. An example is illustrated in Fig. 1. In this example case, C (n) in (t) will be 11.5 if the hub node is not congested and 24.5 if the hub node is congested. There may be several possible in- side routes in this example, but the agents only consider the least-cost inside route among the candidates. out (t) will be 18; C (n) 2 and 3 To express various network structures parametrically, we vary the number of links connected to the hub node, λ, while the symmetry of the network remains unchanged, as illustrated in Fig. 2. Note that it is impractical to consider all topological combinations of ring-and-hub networks. In addition, considering that traffic networks are artificially designed, it is not unreason- able to consider only symmetric cases. Fig. 2: Hub links are symmetrically added to the existing network one at a time 2.3. Minority route-choice game model This route choice game on the ring-and-hub network can be interpreted as a situation whereby a traveler considers whether to use a freeway to reach the destination or not. The hub node corresponds to the freeway, and the peripheral nodes correspond to some sites with or without an entrance of the freeway. If the driver's origin is far from the entrance to the freeway, using a local road would be more efficient, as there would be a high cost associated 5 λ=2λ=3λ=4λ=Nwith reaching a nearby interchange. If the driver's origin is close to an interchange, he/she would use the freeway in order to get to the destination earlier. In this case, however, the driver must also consider the possibility of traffic congestion on the freeway. In this experiment, agents follow the Minority route choice game (MRCG), which is an inductive reasoning rule to select the routes to reach their re- spective destinations. This inductive reasoning rule has a common structure with strategies in the MG case [15, 16], because route-choice problems are very similar to the choice of minority [17, 7]. MRCG agents can choose from S strategies. Each strategy s suggests an action s(µ) according to the past hub states in the form of an M -bit history string µ. At each time step t, the hub state h (t) is recorded as either 0 (uncongested) or 1 (congested). Hence, the total number of possible history strings is P = 2M , and there are 2P possible strategies. A binary number representing an action is assigned to each history string corresponding to an agent's choice of the outside route (0) or the inside route (1). The probability of assigning each action is de- termined by an integer parameter K = 0, . . . , P [14], which represents the tendency of the strategy to have an action 0, where action 0 is employed with probability K P . If K is small, the strategy is likely to suggest action 0 persistently, and vice versa. 2 = 2M−1, where each In a homogeneous version of MRCG, we set K = P strategy has an equal likelihood of action 0 or 1 for each history string. In the heterogeneous version of MRCG, we select an integer K uniformly from 0, . . . , P for each strategy. An example for M = 3, K = 4 is listed in Table. 1. P and action 1 is employed with probability P−K Among the S strategies, each agent n selects the strategy with the best score. At each time step t, the score of each strategy Us (t) is updated based on whether the strategy achieves a lower-cost route or not. If a strategy results in a lower cost, its score is increased. If a strategy leads to a higher cost, its score is decreased. Note that this update is performed regardless of whether or not the strategy is executed at that time step. The update equation is as follows: (cid:16) (cid:17) Us (t + 1) = Us (t) + sign where sign is the sign function. out (t) − C (n) C (n) in (t) (2s (µ) − 1) (3) The entire structure of the simulation model is illustrated in Fig. 3. 6 History string Action 000 001 010 011 100 101 110 111 0 1 1 1 0 1 1 0 Table 1: An example of the homogeneous MRCG strategy for M = 3, K = 4. In the action column, 1 denotes the inside route and 0 denotes the outside route. For instance, given the history string 010, this strategy suggests action 1, denoting the inside route, whereas the history string 111 leads to action 0. 2.4. Evaluation of the network performance To assess the performance of the modeled traffic network, we employ the following basic measures. Average cost The average cost is calculated by averaging costs over all agents and time steps: T(cid:88) N(cid:88) t=1 n=1 (cid:104)C(cid:105) = 1 T N C (n) (t) , (4) where C (n)(t) is the cost of agent n at time step t. Depending on the route chosen by the agent, C (n)(t) can be either C (n) out(t) or C (n) in (t). Congestion ratio The congestion ratio is the ratio of the congested state to the uncongested state: h (t) . (5) (cid:88) t=1 r = 1 T Average number of hub users The average number of hub users is the 7 Fig. 3: Simulation model outline average number of hub users over all time steps: Nin (t) . (6) T(cid:88) t=1 Nin = 1 T (cid:118)(cid:117)(cid:117)(cid:116) 1 T T(cid:88) t=1 Standard deviation of Nin (t) The standard deviation of Nin (t) is cal- culated over all time steps: σNin = (Nin − Nin (t))2. (7) Number of potential hub users Potential hub users are those agents who incur lower costs by taking the inside route in the uncongested state 8 Network modelAgent modelupdates scoresselects strategytakes actioncalculates cost𝑁"#≤𝐿non-congestioncongestionTrueFalseupdates hub stateupdates historytakes actioncalculates costupdates history・・・AgentAgentEnd of time stepStart of time stepupdates scoresselects strategythan by taking the outside route; i.e., C (n) under noncongestion. The number of potential hub users is defined as: in (th (t) = 0) < C (n) out (th (t) = 0) (cid:17) in (th (t) = 0) (8) (cid:16) N(cid:88) n=1 Np = H out (th (t) = 0) − C (n) C (n) where H (·) is the Heaviside step function: (cid:40) H (x) = 0 if x ≤ 0 1 if x > 0. (9) (10) (11) (12) Specifically, potential agents have incentive to use the central hub since they can reduce the cost by using it. 3. Network structure vs network performance 3.1. Solution of Nash equilibrium To assess the effects of agents' bounded rationality, our simulation results are compared with the Nash Equilibrium (NE) solutions, as NE is easier to derive in this model than UE. Under NE, each agent selects the optimal route, which implies that no agent can reduce the travel cost by switching route. Therefore, if Np ≤ L, all potential agents would use the hub, and if Np > L, the hub will contain exactly L potential agents. Note that the hub will never be congested under NE. Multiple solutions for NE can be derived, but for the comparison with the simulation results, we only derive the NE solution with the best allocation (the lowest cost) and that with the worst allocation (the highest cost). out (th (t) = 0) When the hub node is uncongested, the difference between C (n) and C (n) in (th (t) = 0) is: l(n) = C (n) out (th (t) = 0) − C (n) in (th (t) = 0) . For all agents, the differences in cost form the array (l(1), l(2),··· , l(n),··· , l(N )), which can be sorted in ascending order as follows: (l(1(cid:48)), l(2(cid:48)),··· , l(n(cid:48)),··· , l(N(cid:48))). 9 In this array, entries 1(cid:48) to N(cid:48) p + 1 to N(cid:48) are non-potential agents. Therefore, the average cost for the NE with the best allocation, (cid:104)C(cid:105)N E , and that for the NE with the worst allocation, (cid:104)C(cid:105)N E p are potential agents, and entries N(cid:48) b w , are: (cid:33) in (th (t) = 0) + C (n(cid:48)) out (th (t) = 0) C (n(cid:48)) (13) n(cid:48)≤min{Np,L} (cid:32) (cid:88) (cid:32) (cid:88) n(cid:48)≤Np−L (cid:104)C(cid:105)N E b = 1 N (cid:104)C(cid:105)N E w = 1 N min{Np,L}<n(cid:48) (cid:88) (cid:88) (cid:88) Np<n(cid:48) out (th (t) = 0) + C (n(cid:48)) + in (th (t) = 0) C (n(cid:48)) (cid:33) Np−L<n(cid:48)≤Np out (th (t) = 0) C (n(cid:48)) (14) Note that the first term in Eq. (14) disappears when Np ≤ L. 3.2. Baseline results of MRCG Fig. 4: Average cost for homogeneous agents Fig. 5: Congestion ratio for homogeneous agents For the baseline case, we set the network scale to N = 100 and the hub capacity to L = 80. The number of hub links was varied from 2 to N , and the memory length was set to M = 2, 4, 8. We fixed S = 8 because M is the dominant parameter in MG and it would be less meaningful to change S 10 102030405060708090100λ051015202530354045›CfiHomo-MRCG (S = 8, M = 2)Homo-MRCG (S = 8, M = 4)Homo-MRCG (S = 8, M = 8)RandomNash equil. (best)Nash equil. (worst)102030405060708090100λ0.00.10.20.30.40.50.6rHomo-MRCG (S = 8, M = 2)Homo-MRCG (S = 8, M = 4)Homo-MRCG (S = 8, M = 8)Fig. 6: Number of hub users for homogeneous agents Fig. 7: Standard deviation of the number of hub users for homogeneous agents and M simultaneously. Here, K is fixed to 2M−1 to represent homogeneous strategies. In each set of parameters, we computed the average results over 1000 simulations. Each simulation was carried out for 1000 time-steps, and the first half of the simulation is for reaching the stable state, and the second half of the simulation is for getting the statistical results. To evaluate the adaptivity of MG agents, we conducted simulations with random agents, i.e., routes were chosen at random. From Fig. 4, we can see that the simulation results behave similarly to the NE solutions in the range 2 ≤ λ ≤ 9. In this range, agents can adapt themselves to the environment, and the performance is much better than in the case of random agents. For λ ≥ 10, however, the homogeneous MRCG agents deviate from the NE solution. This deviation is caused by growth in the congestion ratio at this point, as can be seen in Fig. 5. This congestion can not be seen under the assumption of the Nash equilibrium. In the range of λ ≤ 9, it is found that since the number of potential agents (purple dots in Fig. 6), who can reduce their cost by using the central hub, is less than the hub capacity, the network does not suffer from congestion. As more hub links are added to the network, Np increases, and then Nin (t) increases as well. This leads to the hub congestion. We can regard this change as a phase transition from the uncongested phase to the congested phase. It is important to investigate how different values of M influence the network performances. In the congested phase, Fig. 4 indicates that, as the memory length increases, the average cost decreases and becomes close to 11 102030405060708090100λ020406080100NinHomo-MRCG (S = 8, M = 2)Homo-MRCG (S = 8, M = 4)Homo-MRCG (S = 8, M = 8)NpoL102030405060708090100λ051015202530σNinHomo-MRCG (S = 8, M = 2)Homo-MRCG (S = 8, M = 4)Homo-MRCG (S = 8, M = 8)Randomthe worst case of Nash equilibrium. In addition, r decreases, enhancing the network performance. The network stability should also be investigated in terms of the standard deviation of the number of hub users. From Fig. 7, we can see the minimum of σNin at λ = 4 in the uncongested phase. After this critical point, the value of σNin increases as more hub links are added to the network for M = 2 agents until λ = N . In contrast, for large memory agents, σNin reaches the nearly stable state at λ = 15. In the uncongested phase, the stability of homogeneous MRCG agents is better than that of random agents, and those with shorter memory lengths lend better stability to the system. In this region, it can be inferred that the agents are divided into two groups, one comprising hub users and one comprising non-hub users, and σNin is better than that of random agents thanks to this division. In the congested phase, however, the stability becomes worse than in the random agents. The increase in memory length positively affects stability, especially in the range λ ≥ 70. Overall, in terms of stability, networks with fewer hub links perform better than those with a large number of hub links. Fig. 8: Congestion ratio for different scale networks We also conducted numerical experiments for different network scales, with N = 20, 40, 60, 80, S = 8, M = 2, and L N = 0.8. For the different networks, we can also observe a phase transition in Fig. 8. Thus, this phase transition can be considered a fundamental property of agents with bounded rationality. 12 0.10.20.30.40.50.60.70.80.91.0λN0.000.020.040.060.080.10rHomo-MRCG (N = 20)Homo-MRCG (N = 40)Homo-MRCG (N = 60)Homo-MRCG (N = 80)Homo-MRCG (N = 100)Fig. 9: Average cost for heterogeneous agents Fig. 10: Congestion ratio for heterogeneous agents 3.3. Effects of preference heterogeneity on the network performance Next, we investigated the effect of preference heterogeneity on network performance. To this effect, we introduced preference heterogeneity into the agents strategies. The value of K, which determines the action distribution 2 = 2M−1 in the homogeneous case, but for of strategies, was set to K = P the heterogeneous version of MRCG, we set K by selecting integer values uniformly from 0, . . . , P . The other parameter settings for the simulations were the same as for homogeneous MRCG. Similarly to the homogeneous MRCG case, we can observe a phase tran- sition occurring at around λ = 9 in Fig. 9. In the uncongested phase, we can see the identical behavior of (cid:104)C(cid:105) as discussed in 3.2, giving a good match with the NE solution. In the congested phase, the average cost becomes higher than in the homogeneous case. The cost is also higher than in the case of random agents. This phenomenon is because r is high even for the large memory capacity (see Fig. 10), unlike the cases of homogeneous MRCG. In contrast to the cases of homogeneous MRCG, Nin approaches to the hub capacity as the memory length increases seen in Fig. 11. Regarding network stability, we can also see the minimum of σNin in the uncongested phase. As M increases, the stability of the network is enhanced, as shown in Fig. 12. In the congested phase, similar to the homogeneous MRCG scenario, the increase in M acts to reduce the value of σN in. Unlike the homogeneous MRCG, however, σNin for M = 8 agents is always below that for random agents. 13 102030405060708090100λ051015202530354045›CfiHetero-MRCG (S = 8, M = 2)Hetero-MRCG (S = 8, M = 4)Hetero-MRCG (S = 8, M = 8)RandomNash equil. (best)Nash equil. (worst)102030405060708090100λ0.00.10.20.30.40.50.6rHetero-MRCG (S = 8, M = 2)Hetero-MRCG (S = 8, M = 4)Hetero-MRCG (S = 8, M = 8)Fig. 11: Number of hub users for heteroge- neous agents Fig. 12: Standard deviation of the number of hub users for heterogeneous agents 3.4. Implications of results for traffic network planning Fig. 13: Optimal number of hub links In reality, there are numerous road networks that accommodate different numbers of drivers without congestion, as they have different road capacities. To construct well-designed networks, it is important to know the optimal network structure according to the different hub-node capacities. For this purpose, we conducted a numerical experiment by changing the hub capacity ratio L N = 0.2 because the simulation result for this case does not exhibit any phase transition. We set N = 100, S = 8, M = 2 for this experiment. N from 0.3 to 0.9. We omit the results for L 14 102030405060708090100λ020406080100NinHetero-MRCG (S = 8, M = 2)Hetero-MRCG (S = 8, M = 4)Hetero-MRCG (S = 8, M = 8)NpoL102030405060708090100λ051015202530σNinHetero-MRCG (S = 8, M = 2)Hetero-MRCG (S = 8, M = 4)Hetero-MRCG (S = 8, M = 8)Random0.30.40.50.60.70.80.9LN0.02.55.07.510.012.515.017.520.0CoptimalHomo-MRCGHetero-MRCGFig. 13 shows the optimal value of λ, the number of hub links that min- imizes (cid:104)C(cid:105), for different hub capacities. We can see that the optimal λ increases nonlinearly as the hub capacity is augmented. However, the opti- mal λ value remains small compared to the network scale (N = 100), even for a large hub capacity, L N = 0.9, the optimal value of λ is less than or equal to 15. This suggests that traffic networks with ring-and-hub structures should have less than or equal to 15% hub links. N = 0.9. For example, when L 4. Conclusion In this study, we built an ABM to investigate the effects of network structures on network performance under the assumption that drivers were operating under the bounded rationality. Through numerical experiments, we found that networks undergo a critical phase transition from the uncon- gested phase to the congested phase, corresponding to the particular number of hub links. The network performance is maximized around this transition point. While the system becomes similar to the NE case in the uncongested phase, the performance of the network in the congested phase declines and is worse than its corresponding NE performance. In addition, we examined how agents' preference heterogeneity affects network performance. We found that agents' preference heterogeneity enhances the system stability in terms of hub usage for large memory agents, but increases the transportation cost. Moreover, we conducted simulations to examine the performances under the different hub node capacities. The simulation results demonstrate that the optimal value of λ increases nonlinearly, but remains much smaller than the whole network size N . Overall, our results show the importance of considering the bounded ra- tionality and preference heterogeneity of drivers in the modeling and research of traffic network problems, and partially answer the question of how to build a well-structured traffic network in reality. Concretely, our simulation results imply that traffic networks with much access to a freeway may easily reach a state of congestion. In contrast, limiting access to freeways could mitigate traffic congestion and enable the Nash equilibrium to be reached. The simulation model in this study, employing a ring-and-hub topology, focuses on a specific situation, and only symmetric network topologies have been considered. In future works, we would like to generalize this research topic to obtain a universal perspective and apply the proposed analyses to 15 practical problems. In addition, a validation of the proposed model should be conducted. References [1] S. Nakayama, R. Kitamura, S. Fujii, Drivers' learning and network be- havior: Dynamic analysis of the driver-network system as a complex system, Transportation Reserach Record 1676 (1999) 30 -- 36. [2] S. Nakayama, , R. Kitamura, Route choice model with inductive learn- ing, Transportation Research Record 1725 (2000) 63 -- 70. [3] S. Nakayama, R. Kitamura, S. Fujii, Drivers' route choice rules and network behavior: Do drivers become rational and homogeneous through learning?, Transportation Research Record 1752 (2001) 62 -- 68. [4] R. Selten, T. Chmura, T. Pitz, S. Kube, M. Schreckenberg, Commuters route choice behavior, Games and Economic Behavior 58 (2007) 394 -- 406. [5] T. Ezaki, R. Nishi, K. Nishinari, Taming macroscopic jamming in trans- portation networks, Journal of Statistical Mechanics: Theory and Ex- periment 2015 (6) (2015) P06013. [6] T. Roughgarden, Selfish Routing and the Price of Anarchy, MIT Press, 2005. [7] S. Gourley, N. F. Johnson, Effects of decision-making on the transport costs across complex networks, Physica A 363 (2006) 82 -- 88. [8] J. Wu, Z. Gao, H. Sun, H. Huang, Congestion in different topologies of traffic networks, Europhysics Letters 74 (3) (2006) 560 -- 566. [9] W. Jian-Jun, G. Zi-You, S. Hui-jun, Optimal traffic networks topology: A complex networks perspective, Physica A 387 (2008) 1025 -- 1032. [10] H. Youn, M. T. Gastner, H. Jeong, Price of anarchy in transportation networks: Efficiency and optimality control, Physical Review Letters 101 (2008) 128701. [11] Z. Xiao-Mei, G. Zi-You, Topological effects on the performance of trans- portation networks, Chinese Physics Letters 24 (1) (2007) 283 -- 286. 16 [12] J.-F. Zheng, Z.-H. Zhu, H.-M. Du, Z.-Y. Gao, Congestion and efficiency in complex traffic networks, International Journal of Modern Physics C 24 (10) (2013) 1350072. [13] Z.-H. Zhu, J.-F. Zheng, Z.-Y. Gao, H.-M. Du, Properties of volume- capacity ratio in congested complex networks, Physica A 400 (2014) 200 -- 206. [14] W. Wang, Y. Chen, J. Huang, Heterogeneous preferences, decision- making capacity, and phase transitions in a complex adaptive system, Proceeding of the National Academy of Science 106 (21) (2009) 8423 -- 8428. [15] W. B. Arthur, Inductive reasoning and bounded rationality, American Economic Review (Papers and Proceedings) 84 (2) (1994) 406 -- 411. [16] D. Challet, Y.-C. Zhang, Emergence of cooperation and organization in an evolutionary gaem, Physica A 246 (1997) 407 -- 418. [17] T. Chmura, T. Pitz, Successful strategies in repeated minority games, Physica A 363 (2006) 477 -- 480. 17
1804.06011	1	1804	2018-04-17T01:42:44	God Save the Queen	[ "cs.MA" ]	Queen Daniela of Sardinia is asleep at the center of a round room at the top of the tower in her castle. She is accompanied by her faithful servant, Eva. Suddenly, they are awakened by cries of "Fire". The room is pitch black and they are disoriented. There is exactly one exit from the room somewhere along its boundary. They must find it as quickly as possible in order to save the life of the queen. It is known that with two people searching while moving at maximum speed 1 anywhere in the room, the room can be evacuated (i.e., with both people exiting) in $1 + \frac{2\pi}{3} + \sqrt{3} \approx 4.8264$ time units and this is optimal~[Czyzowicz et al., DISC'14], assuming that the first person to find the exit can directly guide the other person to the exit using her voice. Somewhat surprisingly, in this paper we show that if the goal is to save the queen (possibly leaving Eva behind to die in the fire) there is a slightly better strategy. We prove that this "priority" version of evacuation can be solved in time at most $4.81854$. Furthermore, we show that any strategy for saving the queen requires time at least $3 + \pi/6 + \sqrt{3}/2 \approx 4.3896$ in the worst case. If one or both of the queen's other servants (Biddy and/or Lili) are with her, we show that the time bounds can be improved to $3.8327$ for two servants, and $3.3738$ for three servants. Finally we show lower bounds for these cases of $3.6307$ (two servants) and $3.2017$ (three servants). The case of $n\geq 4$ is the subject of an independent study by Queen Daniela's Royal Scientific Team.	cs.MA	cs	God Save the Queen ∗ Konstantinos Georgiou† ¶ Danny Krizanc∗∗ Jurek Czyzowicz∗ ¶ Evangelos Kranakis§ ¶ Ryan Killick§ (cid:107) Lata Narayanan††¶ Jaroslav Opatrny†† Sunil Shende‡‡ April 18, 2018 Abstract Queen Daniela of Sardinia is asleep at the center of a round room at the top of the tower in her castle. She is accompanied by her faithful servant, Eva. Suddenly, they are awakened by cries of "Fire". The room is pitch black and they are disoriented. There is exactly one exit from the room somewhere along its boundary. They must find it as quickly as possible in order to save the life of the queen. It is known that with two people searching while moving at maximum speed 1 anywhere in the room, the room can be evacuated (i.e., 3 ≈ 4.8264 time units and this is with both people exiting) in 1 + 2π optimal [10], assuming that the first person to find the exit can directly guide the other person to the exit using her voice. Somewhat surprisingly, in this paper we show that if the goal is to save the queen (possibly leaving Eva behind to die in the fire) there is a slightly better strategy. We prove that 3 + √ Toronto, ON, of Mathematics, Ryerson University, of Computer Science, [email protected] ryankillick,[email protected] ¶Research supported in part by NSERC of Canada. (cid:107)Research supported in part by the Ontario Graduate Scholarship (OGS) Program. ∗Universite du Qubec en Outaouais, Gatineau, Qu´ebec, Canada, [email protected] †Dept. Canada, §School Canada, ∗∗Department of Mathematics & Comp. Sci., Wesleyan University, Middletown, CT, USA, [email protected] ††Department of Comp. Sci. ‡‡Department USA, ∗This is the full version of the paper with the same title which will appear in the proceedings of the 9th International Conference on Fun with Algorithms, (FUN'18), June 13–15, 2018, La Maddalena, Maddalena Islands, Italy. Computer [email protected] and Software Eng., Concordia University, Montreal, Qu´ebec, Canada, lata,[email protected] Carleton University, Ottawa ON, of Science, Rutgers University, Camden, 1 √ this "priority" version of evacuation can be solved in time at most 4.81854. Furthermore, we show that any strategy for saving the queen requires time at 3/2 ≈ 4.3896 in the worst case. If one or both of the queen's least 3+π/6+ other servants (Biddy and/or Lili) are with her, we show that the time bounds can be improved to 3.8327 for two servants, and 3.3738 for three servants. Finally we show lower bounds for these cases of 3.6307 (two servants) and 3.2017 (three servants). The case of n ≥ 4 is the subject of an independent study by Queen Daniela's Royal Scientific Team [12]. 1 Introduction In traditional search, a group of searchers (modeled as mobile autonomous agents or robots) may collaboratively search for an exit (or target) placed within a given search domain [1, 2, 17]. Although the searchers may have differing capabilities (communication, perception, mobility, memory) search algorithms, previously em- ployed, generally make no distinction between them as they usually play identical roles throughout the execution of the search algorithm and with respect to the ter- mination time (with the exception of faulty robots, which also do not contribute to searching). In this work we are motivated by real-life safeguarding-type situations where a number of agents have the exclusive role to facilitate the execution of the task by a distinguished entity. More particularly, we introduce and study Priority Evacuation, a new form of search , under the wireless communication model, in which the search time of the algorithm is measured by the time it takes a special searcher, called the queen, to reach the exit. The remaining searchers in the group, called servants, are participating in the search but are not required to exit. 1.1 Problem Definition of Priority Evacuation (PEn) A target (exit) is hidden in an unknown location on the unit circle. The exit can be located by any of the n + 1 robots (searchers) that walks over it (n = 1, 2, 3). Robots share the same coordinate system, start from the center of the circle, and have maximum speed 1. Among them there is a distinguished robot, called the queen, and the remaining n robots are referred to as servants. All servants are known to the queen by their identities. Robots may run asymmetric algorithms, and can communicate their findings wirelessly and instantaneously (each message is composed by an identity and a location). Only the queen is required to be able to receive messages. Feasible solutions to this problem are evacuation algorithms, i.e. robots' movements (trajectories) that guarantee the finding of the hidden exit. The cost of an evacuation algorithm is the evacuation time of the queen, i.e., the worst case total time until the queen reaches the exit. None of the n servants needs 2 to evacuate. 1.2 Related work Related to our work is linear search which refers to search in an infinite line. There have been several interesting studies attempting to optimize the search time which were initiated with the influential works of Bellman [7] and Beck [6]. A long list of results followed for numerous variants of the problem, citing which is outside the scope of this work. For a comprehensive study of seminal search-type problems see [2, 3]. The problem of searching in the plane by one or more searchers, has been con- sidered by [4, 5]. The unit disk model considered in our present paper is a form of two-dimensional search that was initiated in the work of [10]. In this paper the authors obtained evacuation algorithms in the wireless and face-to-face communi- cation models both for a small number of robots as well optimal asymptotic results for a large number of robots. Additional evacuation algorithms in the face-to-face communication model were subsequently analyzed for two robots in [13] and later in [8]. Other variations of the problem include the case of more than one exit, see [9] and [16], triangular and square domains in [14], robots with different moving speeds [15], and evacuation in the presence of crash or byzantine faulty robots [11]. Notably, all relevant previous work in search-type problems considered the objective of minimizing the time it takes either by the first or the last agent to reach the hidden target. In contrast, this paper considers an evacuation (search- type) problem where the completion time is defined with respect to a distinguished mobile agent, the queen, while the remaining n servants are not required to evac- uate. Our current focus is to design efficient algorithms for n = 1, 2, 3 servants, as well as give strong lower bounds. Notably, the algorithms we propose signifi- cantly improve upon evacuation costs induced by naive trajectories, and in fact the trajectories we propose are non-trivial. Our main contribution concerns priority evacuation for each of the cases of n = 1, 2, 3 servants, all of which require spe- cial treatment. Moreover, all our algorithms are characterized by the fact that the queen does contribute effectively to the search of the hidden item. In sharp con- trast, the independent and concurrent work of [12] studies the same problem for n ≥ 4 servants where the queen never contributes to the search. More importantly, the proposed algorithms of [12] admit a unified description and analysis that does not intersect with the current work. 3 1.3 Our Results & Paper Organization Section 2 introduces necessary notation and terminology and discusses preliminar- ies. Section 3 is devoted to upper bounds for PEn for n = 1, 2, 3 servants (see Subsections 3.1, 3.2, and 3.3, respectively). All our upper bounds are achieved by fixing optimal parameters for families of parameterized algorithms. In Section 4 we derive lower bounds for PEn, n = 1, 2, 3. An interesting corollary of our pos- itive results is that priority evacuation with n = 1, 2, 3 servants (i.e. with n + 1 searchers) can be performed strictly faster than ordinary evacuation with n + 1 robots where all robots have to evacuate. Indeed, an argument found in [10] can be adjusted to show that the evacuation problem with n + 1 robots cannot be solved faster than 1 + 4π 3. Surprisingly, when one needs to evacuate only one designated robot, the task can provably (due to our upper bounds) be executed faster. All our results, together with the comparison to the lower bounds of [10], are summarized in Table 1. We conclude the paper in Section 5 with a discussion of open problems. 3(n+1) + √ Table 1: Upper bounds (UB) and lower bounds (LB) for priority evacuation. # of Servants UB for PEn n = 1 n = 2 n = 3 4.8185 (Theorem 3.1) 3.8327 (Theorem 3.3) 3.3738 (Theorem 3.7) LB for PEn 4.3896 (Theorem 4.1) 3.6307 (Theorem 4.6) 3.2017 (Theorem 4.10) LB for Ordinary Evacuation 4.826445 (see [10]) 4.128314 (see [10]) 3.779248 (see [10]) 2 Notation and Preliminaries We use n to denote the number of servants, and we set [n] = {1, . . . , n}. Queen and servant i will be denoted by Q and Si, respectively, where i ∈ [n]. We assume that all robots start from the origin O = (0, 0) of a unit circle in R2. As usual, points in A ∈ R2 will be treated, when it is convenient, as vectors from O to A, and kAk will denote the euclidean norm of that vector. 2.1 Problem Reformulation & Solutions' Description Robots' trajectories will be defined by parametric functions F(t) = (f(t), g(t)), where f, g : R 7→ R are continuous and piecewise differentiable. In particular, 4 search algorithms for all robots will be given by trajectories Sn :=nQ(t),{Si(t)}i∈[n] o , where Q(t),Si(t) will denote the position of Q and Si, respectively, at time t ≥ 0. Definition 2.1 (Feasible Trajectories). We say that trajectories Sn are feasible for PEn if: (a) Q(0) = Si(0) = O, for all i ∈ [n], (b) Q(t),{Si(t)}i∈[n] induce speed-1 trajectories for Q,{Si}i∈[n] respectively, and (c) there is some time t0 ≥ 1, such that each point of the unit circle is visited (searched) by at least one robot in the time window [0, t0]. We refer to the smallest such t0 as the search time of the circle. Note that feasible trajectories do indeed correspond to robots' movements for PEn in which, eventually the entire circle is searched, and hence the search time is bounded. We will describe all our search/evacuation algorithms as feasible trajec- tories, and we will assume that once the target is reported, Q will go directly to the location of the exit. For feasible trajectories Sn with search time t0, and for any trajectory F(t) (either of the queen or of a servant), we denote by I(F) the subinterval of [0, t0] that contains all x ∈ [0, t0] such that kF(x)k = 1 (i.e. the robot is on the the circle) and no other robot has been to F(x) before. Since robots start from the origin, it is immediate that I(F) ⊆ [1, t0]. With this notation in mind, note that the exit can be discovered by some robot F, say at time x, only if x ∈ I(F). In this case, the finding is instantaneously reported, so Q goes directly to the exit, moving along the corresponding line segment between her current position Q(x) and the reported position of the exit F(x). Hence, the total time that Q needs to evacuate equals x + kQ(x) − F(x)k . Therefore, the evacuation time of feasible trajectories Sn to PEn is given by ex- pression maxF∈Sn sup x∈I(F) {x + kQ(x) − F(x)k} . Notice that for "non-degenerate" search algorithms for which the last point on the circle is not searched by Q alone, the previous maximum can be simply computed over the servants, i.e the evacuation cost will be max i∈[n] sup x∈I(Si) {x + kQ(x) − Si(x)k} . (1) 5 In other words, we can restate PEn as the problem of determining feasible trajec- tories Sn so as to minimize (1). 2.2 Useful Trajectories' Components Feasible trajectories induce, by definition, robots that are moving at (maximum) speed 1. The speed restriction will be ensured by the next condition. Lemma 2.2. An object following trajectory F(t) = (f(t), g(t)) has unit speed if and only if (cid:0)f0(t)(cid:1)2 +(cid:0)g0(t)(cid:1)2 = 1, ∀t ≥ 0. Proof: For any t ≥ 0, the velocity of F is given by F0(t) = (df(t)/dt, dg(t)/dt), and its speed is calculated as kF0(t)k. Robots' trajectories will be composed by piecewise smooth parametric func- tions. In order to describe them, we introduce some further notation. For any θ ∈ R, we introduce abbreviation Cθ for point {cos (θ) , sin (θ)}. Next we introduce parametric equations for moving along the perimeter of a unit circle (Lemma 2.3), and along a line segment (Lemma 2.4). Lemma 2.3. Let b ∈ [0, 2π) and σ ∈ {−1, 1}. The trajectory of an object moving at speed 1 on the perimeter of a unit circle with initial location Cb is given by the parametric equation C(b, σt) := (cos (σt + b) , sin (σt + b)). If σ = 1 the movement is counter-clockwise (ccw), and clockwise (cw) otherwise. Proof: Clearly, C(b, 0) = Cb. Also, it is easy to see that kC(b, t)k = 1, i.e. the (cid:18) d object is moving on the perimeter of the unit circle. Lastly, (cid:18) d (cid:19)2 (cid:19)2 cos (σt + b) + sin (σt + b) = σ2 (− sin (σt + b))2+σ2 (cos (σt + b))2 = 1, dt dt so the claim follows by Lemma 2.2. Lemma 2.4. Consider distinct points A = (a1, a2), B = (b1, b2) in R2. The trajectory of a speed 1 object moving along the line passing through A, B and with initial position A is given by the parametric equation (cid:18) b1 − a1 kA − Bk t + a1, (cid:19) . b2 − a2 kA − Bk t + a2 L(A, B, t) := 6 Proof: It is immediate that the parametric equation corresponds to a line. Also, it is easy to see that L(A, B, 0) = A and L(A, B,kA − Bk) = B, i.e. the object starts from A, and eventually visits B. As for the object's speed, we calculate (cid:19)2 (cid:18) b1 − a1 kA − Bk (cid:18) b2 − a2 kA − Bk (cid:19)2 = 1 + = (cid:18) d dt (cid:18) b1 − a1 kA − Bk t + a1 (cid:19)(cid:19)2 (cid:18) d dt (cid:18) b2 − a2 kA − Bk t + a2 + (cid:19)(cid:19)2 so, by Lemma 2.2, the speed is indeed 1. Robots trajectories will be described in phases. In each phase, robot, say F, will be moving between two explicit points, and the corresponding trajectory F(t) will be implied by the previous description, using most of the times Lemma 2.3 and Lemma 2.4. We will summarize the details in tables of the following format. Robot F # Description Trajectory Duration 0 1 ... F(t) F(t) t0 t1 ... Phase 0 will usually correspond to the deployment of F from the origin to some point of the circle. Also, for each phase we will summarize it's duration. With that in mind, trajectory F(t) during phase i, with duration ti, will be valid for all t ≥ 0 with t − (t0 + t1 + . . . ti−1) ≤ ti. jectories. For any ρ ∈ [0, 1] and θ ∈ [0, 2π), we introduce notation Lastly, the following abbreviation will be useful for the exposition of the tra- K(θ, ρ) := (1 − ρ)Cπ−θ + ρC−θ. In other words, K(θ, ρ) is a convex combination of antipodal points Cπ−θ, C−θ of the unit circle, i.e. it lies on the diameter of the unit circle passing through these two points. Moreover, it is easy to see that kCπ−θ − K(θ, ρ)k = 2ρ, and hence kK(θ, ρ) − C−θk = 2 − 2ρ. As it will be handy later, we also introduce abbreviation AK(θ, ρ) := kCπ − K(θ, ρ)k . The choice of the abbreviation is clear, if the reader denotes Cπ = (−1, 0) by A. 7 Figure 1: An illustration of trajectories S(t),Q(t), and their critical angles at some fixed time τ, with S(τ) = S,Q(τ) = Q,S0(τ) = u,Q0(τ) = v. 2.3 Critical Angles The following definition introduces a key concept. In what follows, abstract tra- jectories will be assumed to be continuous and differentiable, which in particular implies that corresponding velocities are continuous. Definition 2.5 (Critical Angle). Let S(t) ∈ R2 denote the trajectory of a speed-1 object, where t ≥ 0. For some point Q ∈ R2, we define the (S, Q)-critical angle −−−−→S(τ)Q, at time t = τ to be the angle between the velocity vector S0(τ) and vector i.e. the vector from S(τ) to Q. We make the following critical observation, see also Figure 1. Theorem 2.6. Consider trajectories S(t),Q(t) of two speed-1 objects S,Q, where t ≥ 0. Let also φ, θ denote the (S,Q(t))-critical angle and the (Q,S(t))-critical angle at time t, respectively. Then t + kQ(t) − S(t)k is strictly increasing if cos (φ) + cos (θ) < 1, strictly decreasing if cos (φ) + cos (θ) > 1, and constant otherwise. Theorem 2.6 is an immediate corollary of the following lemma. Lemma 2.7. Consider trajectories S(t),Q(t) and their critical angles π, θ, as in the statement of Theorem 2.6. Then kQ(t) − S(t)k = cos (φ) + cos (θ) . d dt Proof: For any fixed t, let d denote D(t), and S, Q denote points S(t),Q(t), re- spectively. Denote also by u, v the velocities of S,Q at time t, respectively, i.e. u = S0(t), v = Q0(t). See also Figure 1. 8 𝑆𝑆(𝑡𝑡)𝑄𝑄(𝑡𝑡)𝑆𝑆𝑄𝑄𝑢𝑢𝑣𝑣𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑆𝑆𝑆𝑆𝑢𝑢𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑝𝑆𝑆𝑆𝑆𝑣𝑣ϕ𝜃𝜃With that notation, observe that (cid:13)(cid:13)(cid:13)−→ (cid:13)(cid:13)(cid:13) = d. Since kuk = kvk = 1, we see that SQ projSQu = cos (φ) d −→ SQ and projSQv = cos (θ) d −→ QS. Now consider two imaginary objects S,Q, with corresponding velocities S0(t) = projSQu and Q0(t) = projSQv. It is immediate that kQ(t) − S(t)k = In particular, projSQu − projSQv is the projection of the relative velocities of S,Q on the line segment connecting S(t),Q(t). As such, the distance between S,Q changes at a rate determined by velocity (cid:13)(cid:13)(cid:13)Q(t) − S(t) (cid:13)(cid:13)(cid:13). projSQu − projSQv = cos (φ) + cos (θ) (cid:13)(cid:13)(cid:13)projSQu − projSQv (cid:13)(cid:13)(cid:13) = cos (φ) + cos (θ). Moreover, projSQu, projSQv where are antiparallel iff and only if cos (φ) , cos (θ) > 0, in which case the two objects come closer to each other. d −→ SQ, 3 Upper Bounds 3.1 Evacuation Algorithm for PE1 This subsection is devoted in proving the following. Theorem 3.1. Consider the real function f(x) = x+sin (x), and denote by α0 > 0 the solution to equation f(f(α − sin (α))) = sin (α) , with α0 ≈ 1.14193. Then PE1 can be solved in time 1 + π − α0 + 2 sin (α0) ≈ 4.81854. The value of α0 is well defined in the statement of Theorem 3.1. Indeed, by letting g(x) = f(f(x − sin (x))) − sin (x), we observe that g is continuous, while g(1) ≈ −0.213934 and g(π/2) ≈ 1.00729, hence there exists α0 ∈ (1, π/2) with g(α0) = 0. In order to prove Theorem 3.1, and given parameters α, β, we introduce the family of trajectories SEARCH1(α, β), see also Figure 2. 9 Figure 2: Algorithm SEARCH1(α, β) depicted for the optimal parameters of the algorithm. In all subsequent figures, as well as here, the orange points on the perimeter of the disc correspond to the worst adversarial placements of the trea- sure, which due to our optimality conditions induce the same evacuation cost. The orange points in Q's trajectories correspond to the Q's positioning when the trea- sures are reported, in the worst cost induced cases. The green dashed line depict Q's trajectory after Q abandons her trajectory and moves toward the reported exit following a straight line. Algorithm SEARCH1(α, β) # Description Robot Q 0 Move to point Cπ 1 2 Move to point C−α+β, 3 0 Move to point Cπ 1 S1 Search circle cw till point Cβ−α Search circle ccw till point C−α Search circle cw till point C−α Trajectory L(O, Cπ, t) C(π, t − 1) L(C−α, C−α+β, t − (1 + π − α)) C(β − α, 1 + π − α + 2 sin (β/2) − t) L(O, Cπ, t) C(π,−t + 1) Duration 1 π − α 2 sin (β/2) β 1 π + α − β Partitioning the circle clockwise, we see that the arc with endpoints Cπ, Cπ+α−β is searched by S1, while the remaining of the circle is searched by Q. Therefore, robots' tra- jectories in SEARCH1(α, β) are feasible, and it is also easy to see that they are continuous as well. The search time equals 1 + π + max{α − β, 2 sin (β/2) + β − α}, as well as I(Q) = [1, 1+π−α]∪[1+π−α+2 sin (β/2) , 1+π−α+2 sin (β/2)+β], I(S1) = [1, 1+π+α−β]. An illustration of the above trajectories for certain values of α, β can be seen in Figure 2. First we make some observations pertaining to the monotonicity of the evacuation cost. Lemma 3.2. Assuming that α > π/3 and that cos (α)+cos (α − β/2) > 1, the evacuation cost of SEARCH1(α, β) is monotonically increasing if the exit is found by S1 during Q's phase 1 and monotonically decreasing if the exit is found by S1 during Q's phase 2. 10 𝜋𝜋−𝛼𝛼𝛽𝛽𝐶𝐶−𝛼𝛼𝐶𝐶−𝛼𝛼+𝛽𝛽𝐶𝐶𝛼𝛼SEARCH1𝛼𝛼,𝛽𝛽𝒬𝒬𝑡𝑡𝒮𝒮1𝑡𝑡Proof: Suppose that the exit is found by S1 during Q's phase 1, i.e. at time x after robots start searching for the first time, where 0 ≤ x ≤ π − α. It is easy to see that the critical angles between Q,S1 are both equal to π − x. But then 2 cos (π − x) ≥ 2 cos (α) > 2 cos (π/3) = 1. Hence, by Theorem 2.6, the evacuation cost is decreasing in this case. Now suppose that the exit is found by S1 during Q's phase 2, i.e. at time x after Q starts moving along the chord with endpoints C−α, C−α+β, where 0 ≤ x ≤ 2 sin (β/2). If φx, θx denote the S1,Q critical angles, then it is easy to see that φ0 = cos (α) and that θ0 = α − β/2. Since cos (φ0) + cos (θ0) > 1, Theorem 2.6 implies that the evacuation cost is initially decreasing in this phase. For the remaining of Q's phase 2, it is easy to see that both φx, θx are decreasing in x, hence cos (φx) + cos (θx) is increasing in x, hence, the evacuation cost will remain decreasing in this phase. Now we can prove Theorem 3.1 by fixing certain values for parameters α, β of SEARCH1(α, β). In particular, we set α0 as in the statement of Theorem 3.1, and β0 = 2f(α0 − sin (α0)) ≈ 0.925793. The trajectories of the robots, for the exact same values of the parameters, can be seen in Figure 2. Proof: [Proof of Theorem 3.1] Let f, α0 be as in the statement of Theorem, and set β0 = 2f(α0 − sin (α0)) ≈ 0.925793. We argue that the worst evacuation time of SEARCH1(α0, β0) is 1 + π − α0 + 2 sin (α0). Note that for the given values of the parameters, we have that α0 > π/3, that α0 − sin (β0/2) ≤ β0, and that cos (α0) + cos (α0 − β0/2) > 1. First we observe that if the exit if found by Q, then the worst case evacuation time E0(α0, β0) is incurred when the exit is found just before Q stops searching, that is E0(α0, β0) = 1 + π − α0 + 2 sin (β0/2) + β0. Next we examine some cases as to when the exit is found by S1. If the exit is found by S1 during the 1st phase of Q, then the evacuation time is, due to Lemma 3.2, given as {x + kQ(x) − S1(x)k} = 1 + π − α0 + 2 sin (α0) . E1(α0, β0) = sup 1≤x≤1+π−α0 Recall that cos (α0) + cos (α0 − β0/2) > 1, and so, again by Lemma 3.2 we may omit the case that the exit is found by S1 while Q is at phase 2. The end of Q's phase 2 happens at time τ := 1 + π − α0 + 2 sin (β0/2), when have that Q(τ) = C−α+β, and S1(τ) = Cα−2 sin(β0/2), and both robots are intending to search ccw. Condition α0 − sin (β0/2) ≤ β0 says that S1 will finish searching prior to Q, and this happens when S1 reaches point C−α+β. During this phase, the distance between Q,S1 stays invariant and equal to 2α0 − β0 − 2 sin (β0/2). We conclude that the cost in this case would be E2(α0, β0) = 1 + π + α0 − β0 + 2 sin (α0 − β0/2 − sin (β0/2)) . Then, we argue that that the choice of α0, β0 guarantees that E0(α0, β0) = E1(α0, β0) = E2(α0, β0), as wanted. Indeed, E0(α0, β0) = E1(α0, β0) implies that sin (β0/2) + β0/2 = sin (α0). But then, we can rewrite E2(α0, β0) as E2(α0, β0) = 1 + π + α0 − β0 + 2 sin (α0 − sin (α0)) . 11 Equating the last expression with E1(α0, β0) implies that β0/2 = α0 − sin (α0) + sin (α0 − sin (α0)) = f(α0 − sin (α0)). Substituting twice β0/2 in the already derived condition sin (β0/2) + β0/2 = sin (α0) implies that f(f(α − sin (α0))) = sin (α0) . Figure 2 depicts the worst placements of the exit, along with the trajectories of the queen (in dashed green lines) after the exit is reported. It should be stressed that Q's Phases 2,3 are essential for achieving the promised bound. Indeed, had we chosen α = β = 0, the worst case evacuation time would have been {1 + x + 2 sin (x)} . sup 1≤x≤1+π {x + kQ(x) − S1(x)k} = sup 0≤x≤π The maximum is attained at x0 = 2π/3 (and indeed, both critical angles in this case are 3 ≈ 4.82645. The π/3 and in particular 2 cos (π/3) = 1), inducing cost 1 + 2π/3 + latter is the cost of the evacuation algorithm for two robots without priority of [10]. √ 3.2 Evacuation Algorithm for PE2 In this subsection we prove the following theorem. Theorem 3.3. PE2 can be solved in time 3.8327. Given parameters α, ρ, we introduce the family of trajectories SEARCH2(α, ρ), see also Figure 3. Figure 3: Algorithm SEARCH2(α, β) depicted for the optimal parameters of the algorithm. 12 𝐶𝐶𝜋𝜋−𝛼𝛼𝐶𝐶−𝛼𝛼/2𝐾𝐾𝛼𝛼/2,𝜌𝜌𝛼𝛼𝛼𝛼/2SEARCH2𝛼𝛼,𝛽𝛽𝒬𝒬𝑡𝑡𝒮𝒮1𝑡𝑡𝒮𝒮2𝑡𝑡Search the circle ccw till point Cπ Algorithm SEARCH2(α, ρ) # Description Robot Q 0 Move to point Cπ−α 1 2 Move to point K(α/2, ρ) 3 Move to point C−α/2 0 Move to point Cπ−α 1 0 Move to point Cπ 1 S1 S2 Search the circle cw till point C−α/2 Search the circle cw till point C−α/2 1 α Duration Trajectory L(O, Cπ−α, t) C(π − α, t − 1) L(Cπ, K(α/2, ρ), t − (1 + α)) AK(α/2, ρ) L(K(α/2, ρ), C−α/2) L(O, Cπ−α) C(π − α,−t + 1) L(O, Cπ) C(π, t − 1) 2 − 2ρ π − α/2 π − α/2 1 1 Notice that, by definition of SEARCH2(α, ρ), robots' trajectories are continuous and feasible, meaning that the entire circle is eventually searched. Indeed, partitioning the circle clockwise, we see that: the arc with endpoints Cπ, Cπ−α is searched by Q, the arc with endpoints Cπ−α, C−α/2 is searched by S1, and the arc with endpoints C−α/2, Cπ is searched by S2. It is immediate from the description of the trajectories that the search time is 1 + π − α/2. Moreover I(Q) = [1, 1 + α], I(S1) = I(S2) = [1, 1 + π − α/2]. An illustration of the above trajectories for certain values of α, ρ can be seen in Figure 3. Now we make some observations, in order to calculate the worst case evacuation time. Lemma 3.4. Suppose that π − α/2 ≥ α + AK(α/2, ρ) + 2 − 2ρ. Then kQ(x) − S1(t)k is continuous and differentiable in the time intervals I1, I2, I3 of Q's phases 1,2,3, respec- tively. Moreover, the worst case evacuation time of SEARCH2(α, ρ) can be computed as  max 1 + α + 2 sin (α) , supt∈I2 {t + kQ(t) − S1(t)k} supt∈I3 {t + kQ(t) − S1(t)k} 1 + π − α/2  where I2 = [1+ α, 1+ α+ AK(α/2, ρ)], I3 = [1+ α+ AK(α/2, ρ), 3−2ρ+ α+ AK(α/2, ρ)]. Proof: Note that the line passing through O and C−α/2, call it , has the property that each point of it, including K(α/2, ρ) is equidistant from S1,S2. Moreover, in the time window [1 + α, 1 + α + AK(α/2, ρ)] that only S1,S2 are searching, Q stays below line . At time 1 + α + AK(α/2, ρ), Q is, by construction, equidistant from S1,S2, a property that is preserved for the remaining of the execution of the algorithm. As a result, the evacuation time of SEARCH2(α, ρ) is given by sup1≤t≤1+π−α/2{t + kQ(t) − S1(t)k}. Now note that condition π−α/2 ≥ α+AK(α/2, ρ)+2−2ρ guarantees that Q reaches point C−α/2 no later than S1. Moreover, in each time interval I1, I2, I3, Q's trajectory is differentiable (and so is S1's trajectory). 13 Now Theorem 3.3 can be proven by fixing parameters α, ρ for SEARCH2(α, ρ), in particular, α = 0.6361, ρ = 0.7944. Notably, the performance of SEARCH2(α, ρ) is provably improvable (slightly) using a technique we will describe in the next section. Proof: [Proof of Theorem 3.3] We choose α = 0.6361, ρ = 0.7944. The trajectories of Figure 3 correspond exactly to those values. The time that Q needs to reach C−α/2 equals 1 + α + AK(α/2, ρ) + 2 − 2ρ = 3.6174, while the time that S1,S2 reach the same point is 1 + π − α/2 = 3.82354. Hence, Lemma 3.4 applies. The worst case evacuation time during phase 1 is 1 + a + 2 sin (α) = 2.82423. The worst case evacuation time after Q reaches C−α/2, equals 1+ π− α/2 = 3.82354. Hence, it remains to compute the maxima of t + kQ(t) − S1(t)k in the two intervals I2, I3, which can be done numerically using the trajectories of SEARCH2(α, ρ), since expression is dif- ferentiable in each of the intervals. To that end, when t ∈ I2 = [1.6361, 3.2062] we have that Q(t) = (0.9931t − 2.62481, 0.191866 − 0.11727t) S1(t) = (cos (3.50549 − t) , sin (3.50549 − t)) , so that t + kQ(t) − S1(t)k becomes t+ When t ∈ I3 = [3.2062, 3.6174] we have that (− sin (3.50549 − t) − 0.11727t + 0.191866)2 + (− cos (3.50549 − t) + 0.9931t − 2.62481)2 q q Q(t) = (0.949847t − 2.48613, 0.818501 − 0.312715t) while S1's trajectory equation remains unchanged, so that t + kQ(t) − S1(t)k becomes t+ (− sin (3.50549 − t) − 0.312715t + 0.818501)2 + (− cos (3.50549 − t) + 0.949847t − 2.48613)2 In particular, it follows that sup t∈I2 {t + kQ(t) − S1(t)k} ≈ sup t∈I3 {t + kQ(t) − S1(t)k} ≈ 3.8327 with corresponding maximizers (with approximate values) τ2 = 3.10066 and τ3 = 3.32114, respectively. Figure 3 also depicts the locations of the optimizers, i.e the worst case lo- cations on the circle for the exit to be found, along with the corresponding evacuation trajectory in dashed green colour. 3.3 Evacuation Algorithm for PE3 3.3.1 A Simple Algorithm In this section we prove the following preliminary theorem (to be improved in Section 3.3.2). Theorem 3.5. PE3 can be solved in time 3.37882. 14 Figure 4: Algorithm SEARCH3(α, β, ρ) depicted for the optimal parameters of the algorithm. Given parameters α, β, ρ, we introduce the family of trajectories SEARCH3(α, β, ρ), corresponding to robots Q,S1,S2,S3, see also Figure 4. Algorithm SEARCH3(α, β, ρ) Robot Q S1 S2 S3 Search the circle ccw till point Cπ 2 , ρ) # Description 0 Move to point Cπ−α 1 2 Move to point K( α+β 3 Move to point C− α+β 0 Move to point Cπ−α−β 1 0 Move to point Cπ 1 0 Move to point Cπ−α−β 1 2 2 2 Search the circle cw till point C− α+β Search the circle ccw till point C− α+β Search the circle ccw till point C−α Duration 1 α 2 , ρ), t − (1 + α)) AK( α+β 2 , ρ) 2 − 2ρ ) Trajectory L(O, Cπ−α, t) C(π − α, t − 1) L(Cπ, K( α+β L(K( α+β 2 , ρ), C− α+β L(O, Cπ−α−β) C(π − α − β,−t + 1) L(O, Cπ) C(π, t − 1) L(O, Cπ−α−β) C(π − α − β,−t + 1) 2 1 π − α+β 2 π − α+β 2 1 1 β As before, it is immediate that, in SEARCH3(α, β, ρ), robots' trajectories are contin- uous and feasible, meaning that the entire circle is eventually searched. In particular, the arc with endpoints Cπ, Cπ−α is searched by Q, the arc with endpoints Cπ−α−β, C− α+β is searched by S1, the arc with endpoints C−π, C− α+β is searched by S2, and the arc with endpoints Cπ−α, Cπ−α−β is searched by S3. Also, the search time is 1 + π − α+β 2 , and 2 2 I(Q) = [1, 1 + α], I(S1) = I(S2) = [1, 1 + π − α + β 2 ], I(S3) = [1, 1 + β]. An illustration of the above trajectories for certain values of α, β, ρ can be seen in Figure 4. 15 𝐶𝐶𝜋𝜋−𝛼𝛼𝐶𝐶−𝛼𝛼−𝛽𝛽/2𝐾𝐾𝛼𝛼+𝛽𝛽2,𝜌𝜌𝛼𝛼(𝛼𝛼+𝛽𝛽)/2𝐶𝐶𝜋𝜋−𝛼𝛼−𝛽𝛽𝛽𝛽SEARCH3𝛼𝛼,𝛽𝛽,𝜌𝜌𝒬𝒬𝑡𝑡𝒮𝒮1𝑡𝑡𝒮𝒮2𝑡𝑡𝒮𝒮3𝑡𝑡  Before we prove Theorem 3.5, we need to make some observation, in order to calculate the worst case evacuation time. 2 ≥ α + Lemma 3.6. Suppose that α ≤ β, α + AK( α+β 2 , ρ) + 2 − 2ρ. Then the following functions are continuous and differentiable AK( α+β in each associated time intervals: kQ(x) − S3(t)k in I1 = {t ≥ 0 : α ≤ t − 1 ≤ β}, 2 , ρ)} and in I3 = {t ≥ kQ(x) − S1(t)k in I2 = {t ≥ 0 : 2 , ρ) ≤ 2 − 2ρ}. Moreover, the worst case evacuation time of 0 : SEARCH3(α, β, ρ) can be computed as 2 , ρ) ≥ β, and π − α+β t − 1 − α ≤ AK( α+β t − 1 − α − AK( α+β max supt∈I1 {t + kQ(t) − S3(t)k} supt∈I2 {t + kQ(t) − S1(t)k} supt∈I3 {t + kQ(t) − S1(t)k} 1 + π − α+β 2 Proof: Conditions α ≤ β and α + AK( α+β than S3, and that when S3 stops searching Q is still in her phase 2, respectively. 2 , ρ) ≥ β mean that Q stops searching no later The line passing through O and C−(α+β)/2, call it , has the property that each point 2 , ρ) is equidistant from S1,S2. Moreover, while S1,S2 are search- of it, including K( α+β 2 , ρ), Q is, by construction, ing, Q never goes above line . At time 1 + α + AK( α+β equidistant from S1,S2, a property that is preserved for the remaining of the execution of the algorithm. As a result, S2 can be ignored in the performance analysis, and when it comes to the case that S1 finds the exit, the evacuation cost is given by the supremum of t + kQ(t) − S1(t)k in the time interval I2 or in the interval I3. Note that in both intervals, the evacuation cost is continuous and differentiable, by construction. If the exit is reported by S3 then the evacuation cost is t + kQ(t) − S3(t)k for t ∈ [1, 1 + β]. However, it is easy to see that the cost is strictly increasing for all t ∈ [1, 1 + α] (in fact it is linear). Since the evacuation cost is also continuous, we may restrict the analysis in interval I1. 2 , ρ) + 2 − 2ρ implies that S1,S2 reach point C−(α+β)/2 no earlier than Q. Hence Q waits at C−(α+β)/2 till the search of the circle is over, which can be easily seen to induce the worse evacuation time after Q reaches C−(α+β)/2. Lastly, observe that π − α+β 2 ≥ α + AK( α+β Next, we prove Theorem 3.5 by fixing parameters α, β, ρ for SEARCH3(α, β, ρ). Proof: [Proof of Theorem 3.5] We choose α = 0.26738, β = 1.2949, ρ = 0.70685. The trajectories of Figure 4 correspond exactly to those values. The time that Q needs to reach 2 , ρ) + 2− 2ρ = 3.17984, while the time that S1,S2 reach equals 1 + α + AK( α+β C− α+β 2 the same point is 1 + π − α+β 2 = 3.36045. Hence, Lemma 3.6 applies. From the above, it is immediate that the worst evacuation time after Q reaches C−(α+β)/2 equals 1 + π − α+β 2 = 3.36045. Hence, it remains to compute the maxima of t + kQ(t) − S3(t)k in interval I1, and of t + kQ(t) − S1(t)k in intervals I2, I3. To that end, when t ∈ I1 = [1.26738, 2.2949] we have that Q(t) = (−2.23643 + 0.97558t, 0.278372 − 0.219643t) S3(t) = (cos(t + 0.579313), sin(t + 0.579313)) , 16 so that t + kQ(t) − S3(t)k becomes t+p(−0.219643t − sin(t + 0.579313) + 0.278372)2 + (0.97558t − cos(t + 0.579313) − 2.23643)2 in which case {t + kQ(t) − S3(t)k} = 1 + β + kQ(1 + β) − S3(1 + β)k ≈ 3.37882 sup t∈I1 When t ∈ I2 = [1.26738, 2.59354], Q's trajectory is the same as in I1 and S1(t) = (cos(2.57931 − t), sin(2.57931 − t)) , so that t + kQ(t) − S1(t)k becomes t+p(− sin(2.57931 − t) − 0.219643t + 0.278372)2 + (− cos(2.57931 − t) + 0.97558t − 2.23643)2. When t ∈ I3 = [2.59354, 3.17984], S1's trajectory is the same as in I2 and Q(t) = (−1.54793 + 0.710111t, 1.5348 − 0.704089t) , so that t + kQ(t) − S1(t)k becomes t+p(sin(2.57931 − t) + 0.704089t − 1.5348)2 + (cos(2.57931 − t) − 0.710111t + 1.54793)2. Numerically sup t∈I2 sup t∈I3 {t + kQ(t) − S1(t)k} = τ2 + kQ(τ2) − S1(τ2)k ≈ 3.37882 {t + kQ(t) − S1(t)k} = τ3 + kQ(τ3) − S1(τ3)k ≈ 3.37882 where τ2 ≈ 2.34029 and τ3 ≈ 2.84758. 3.3.2 Improved Search Algorithm In this section we improve the upper bound of Theorem 3.5 by 0.00495 additive term. Theorem 3.7. PE3 can be solved in time 3.37387. The main idea can be described, at a high level, as a cost preservation technique. By the analysis of Algorithm SEARCH3(α, β, ρ) for the value of parameters of α, β, ρ as in the proof of Theorem 3.5, we know that there are is a critical time window [τ2, τ3] so that the total evacuation time is the same if the exit is found by S1 either at time τ2 or τ3, and strictly less for time moments strictly in-between. In fact, during time [τ2, 1 + α + AK( α+β 2 , ρ)] Q is executing phase 2, and in the time window [1 + α + AK( α+β 2 , ρ), τ3] Q is executing phase 3 of SEARCH3(α, β, ρ). From the above, it is immediate that we can lower Q's speed in the time window [τ2, τ3] so that the evacuation time remains unchanged no matter when S1 finds the exit in 17 the same time interval (notably, S3 has finished searching prior to τ2 and kQ(t) − S1k ≥ kQ(t) − S2k). But this also implies that we must be able to maintain the evacuation time even if we preserve speed 1 for Q, that will in turn allow us to twist parameters α, β, ρ, hopefully improving the worst case evacuation time. We show this improvement is possible by using the following technical observation Theorem 3.8. Consider point Q = (q1, q2) ∈ R2. Let S(t) be the trajectory of an object S moving at speed 1, where t ≥ 0, and denote by φ the (S, Q)-critical angle at time t = 0. Assuming that cos (φ) ≥ 0, then there is some τ > 0, and a trajectory Q(t) = (f(t), g(t)) of a speed-1 object, where t ≥ 0, so that t + kQ(t) − S(t)k remains constant, for all t ∈ [0, τ]. Moreover, Q(t) can be determined by solving the system of differential equations (f0(t))2 + (g0(t))2 = 1 t + kQ(t) − S(t)k = kS(0) − Qk (f(0), g(0)) = (q1, q2). (2) (3) (4) Proof: An object with trajectory (f(t), g(t)) satisfying (2) and (4) has speed 1 (by Lemma 2.2), and starts from point Q = (q1, q2). We need to examine whether we can choose f, g so as to satisfy (3). By Lemma 2.7, such a trajectory Q(t) exists exactly when we can guarantee that cos (φ) + cos (θ) = 1 over time t. When t = 0 we are given that cos (φ) > 0, hence there exists θ satisfying cos (φ) + cos (θ) = 1. This uniquely determines the velocity of Q at t = 0. By continuity of the velocities, there must exist a τ > 0 such that cos (φ)+cos (θ) = 1 admits a solution for θ also as φ changes over time t ∈ [0, τ], in which time window the cosine of the (S,Q(t))-critical angle at time t remains non-negative. increasing at t = τ, i.e. that S does not move away from point Q. Note that condition cos (φ) ≥ 0 of Theorem 3.8 translates to that kS(t) − Qk is not Now fix parameters α, β, ρ together with the trajectories of S1,S2,S3 as in the descrip- tion of Algorithm SEARCH3(α, β, ρ). The description of our new algorithm N-SEARCH3(α, β, ρ) will be complete once we fix a new trajectory for Q. Naming specific values for parameters α, β, ρ will eventually prove Theorem 3.7. In order to do so, we introduce some further notation and conditions, denoted below by (Conditions i-iv), that we later make sure are satisfied. Consider Q's trajectory as in SEARCH3(α, β, ρ). Let τ0 denote a local maximum of t + kQ(t) − S1(t)k as it reads for t ≥ 0 with t − 1 − α ≤ AK( α+β expression is differentiable by Lemma 3.6), i.e. 2 , ρ) (recall that in this time window, τ0 − 1 − α ≤ AK( α + β 2 Set Q = Q(τ0), and assume that , ρ) (Condition i) "The cosine of the (S, Q)-critical angle at time τ0 is non-negative." (Condition ii) 18 Then obtain from Theorem 3.8 trajectory (f(t), g(t)) that has the property that it preserves τ0 + kQ(τ0) − S1(τ0)k in the time window [τ0, τ0]. Assume also that "There is time τ1 ≤ τ0 such that point K1 := (f(τ1), g(τ1)) is equidistant from S1(τ1),S2(τ1)," (Condition iii) for the first time after time τ0, such that τ1 ≤ 1 + π − α + β 2 . (Condition iv) Then consider the following modification of SEARCH3(α, β, ρ), where the trajectories of S1,S2,S3 remain unchanged, see also Figure 5. Figure 5: Algorithm SEARCH3(α, β, ρ) depicted for the optimal parameters of the algorithm. Algorithm N-SEARCH3(α, β, ρ) Robot Q # Description 0 Move to point Cπ−α 1 2 Move toward point K( α+β 3 4 Move to point C− α+β Search the circle ccw till point Cπ 2 , ρ) Preserve τ0 + kQ(τ0) − S1(τ0)k 2 Trajectory L(O, Cπ−α, t) C(π − α, t − 1) L(Cπ, K( α+β (f(t), g(t)) L(K1, C− α+β ) 2 2 , ρ), t − (1 + α)) Duration 1 α τ0 − 1 − α τ1 − τ0 (cid:13)(cid:13)(cid:13)K1 − C− α+β 2 (cid:13)(cid:13)(cid:13) Note that in phase 2, Q is not reaching (necessarily) point K rather it moves toward it 2 . Trajectories of S1,S2,S3 are for a certain duration. The search time is still 1 + π − α+β continuous as before, and I(S1) = I(S2) = [1, 1 + π − α + β 2 as well as I(Q) = [1, 1 + α]. ], I(S3) = [1, 1 + β], 19 𝐶𝐶𝜋𝜋−𝛼𝛼𝐶𝐶−𝛼𝛼−𝛽𝛽/2𝐾𝐾𝛼𝛼+𝛽𝛽2,𝜌𝜌𝛼𝛼(𝛼𝛼+𝛽𝛽)/2𝐶𝐶𝜋𝜋−𝛼𝛼−𝛽𝛽𝛽𝛽N-SEARCH3𝛼𝛼,𝛽𝛽,𝜌𝜌𝒬𝒬𝑡𝑡𝒮𝒮1𝑡𝑡𝒮𝒮2𝑡𝑡𝒮𝒮3𝑡𝑡Condition i makes sure that while Q is at phase 2, and before it reaches K( α+β 2 , ρ), there is a time moment τ0 when the rate of change of t + kQ(t) − S1(t)k is 0. Together with condition ii, this implies that Theorem 3.8 applies. In fact, for the corresponding critical angles φ, θ between S1,Q at time τ0, we have that cos (φ) + cos (θ) = 1 by con- struction. Hence trajectory (f(t), g(t)) of phase 3 is well defined, and indeed, Q jumps from phase 2 to phase 3 while Q is still moving toward point K. Notably, Q's trajectory is even differentiable at t = τ0 (but not necessarily at t = τ1). Then, Condition iii says that Q eventually will enter phase 4, and that this will happen before S1,S2 finish the explo- ration of the circle. Overall, we conclude that in N-SEARCH3(α, ρ), robots' trajectories are continuous and feasible. An illustration of the above trajectories for certain values of α, β, ρ can be seen in Figure 5. Now we make some observations, in order to calculate the worst case evacuation time. Lemma 3.9. Suppose that α ≤ β, 1 + β ≤ τ0, and 1 + π − α+β as well as Conditions i-iv are satisfied. Then the following functions are continuous and differentiable in each associated time intervals: kQ(x) − S3(t)k in I1 = {t ≥ 0 : α ≤ t − 1 ≤ β}, kQ(x) − S1(t)k in I2 = {t ≥ 0 : 1 + α ≤ t ≤ τ0 and in . Moreover, the worst case evacuation time of t ≥ 0 : t − τ1 ≤(cid:13)(cid:13)(cid:13)K1 − C− α+β I3 =n 2 ≥ τ1 + (cid:13)(cid:13)(cid:13)o 2 N-SEARCH3(α, β, ρ) can be computed as (cid:13)(cid:13)(cid:13)K1 − C− α+β 2 (cid:13)(cid:13)(cid:13)  max supt∈I1 {t + kQ(t) − S3(t)k} supt∈I2 {t + kQ(t) − S1(t)k} supt∈I3 {t + kQ(t) − S1(t)k} 1 + π − α+β 2  Proof: Conditions α ≤ β and 1 + β ≤ τ0 mean that Q stops searching no later than S3, and that when Q enters phase 3 after S3 is done searching, respectively. The line passing through O and C−(α+β)/2, call it , has the property that each point of 2 , ρ) is equidistant from S1,S2. Moreover, while S1,S2 are searching, it, including K( α+β Q never goes above line . Also, while Q is executing phase 3, Q remains equidistant from S1,S2 and this is preserved for the remainder of the execution of the algorithm. As a result, S2 can be ignored in the performance analysis, and when it comes to the case that S1 finds the exit, the evacuation cost is given by the supremum of t + kQ(t) − S1(t)k in the time interval I2 or in the interval I3. Note that in both intervals, the evacuation cost is continuous and differentiable, by construction. If the exit is reported by S3 then the evacuation cost is t + kQ(t) − S3(t)k for t ∈ [1, 1 + β]. However, it is easy to see that the cost is strictly increasing for all t ∈ [1, 1 + α] (in fact it is linear). Since the evacuation cost is also continuous, we may restrict the analysis in interval I1. (cid:13)(cid:13)(cid:13)K1 − C− α+β (cid:13)(cid:13)(cid:13) implies that S1,S2 reach point C−(α+β)/2 no earlier than Q. Hence Q waits at C−(α+β)/2 till the search of the circle is over, which can be easily seen to induce the worse evacuation time after Q reaches C−(α+β)/2. 2 Lastly, observe that 1 + π − α+β 2 ≥ τ1 + Next we prove Theorem 3.7 by fixing parameters α, β, ρ for N-SEARCH3(α, β, ρ). 20 The time that Q needs to reach C− α+β 2 equals 1 + τ1 + Proof: [Proof of Theorem 3.7] We choose α = 0.27764, β = 1.29839, ρ = 0.68648. The trajectories of Figure 4 correspond exactly to those values. For these values we see that 2 , ρ) = 1.29041, while τ0− α−1 = 1.04877. Hence the transition between phase AK( α+β 1 and phase 2 of Q is well defined. (cid:13)(cid:13)(cid:13)K1 − C− α+β 2 (cid:13)(cid:13)(cid:13) = 3.18073, while the time that S1,S2 reach the same point is 1 + π − α+β 2 = 3.35358. Therefore we may attempt to solve numerically the differential equation of Theorem 3.8. It turns out that for the resulting trajectory (f(t), g(t), and for τ1 = 2.89288, point (f(τ1), g(τ1) is equidistant from S1,S2. Moreover, Q enters phase 4 at time τ1 = 2.89288, prior to 1 + π − α+β From the above, it is immediate that the worst evacuation time after Q reaches C−(α+β)/2 equals 1 + π − α+β 2 = 3.35358. Hence, it remains to compute the maxima of t + kQ(t) − S3(t)k in interval I1, and of t + kQ(t) − S1(t)k in intervals I2, I3. 2 . Hence, Conditions i-iv are all met, as well as Lemma 3.9 applies. To that end, when t ∈ I1 = [1.27764, 2.29839] we have that Q(t) = (0.978782t − 2.25053, 0.261795 − 0.204905t) S3(t) = (cos(t + 0.565563), sin(t + 0.565563)) , so that t + kQ(t) − S3(t)k becomes t+p(−0.204905t − sin(t + 0.565563) + 0.261795)2 + (0.978782t − cos(t + 0.565563) − 2.25053)2 in which case {t + kQ(t) − S3(t)k} = 1 + β + kQ(1 + β) − S3(1 + β)k ≈ 3.37387 sup t∈I1 When t ∈ I2 = [1.27764, 2.32641], Q's trajectory is the same as in I1 and S1(t) = (cos(2.56556 − t), sin(2.56556 − t)) , so that t + kQ(t) − S1(t)k becomes t+p(− sin(2.56556 − t) − 0.204905t + 0.261795)2 + (− cos(2.56556 − t) + 0.978782t − 2.25053)2. When t ∈ I3 = [2.89288, 3.18073], S1's trajectory is the same as in I2 and Q(t) = (0.705254t − 1.53797, 1.54604 − 0.708955t0.706399t − 1.53762, 1.5407 − 0.707814t) , so that t + kQ(t) − S1(t)k becomes t+p(− sin(2.56556 − t) − 0.708955t + 1.54604)2 + (− cos(2.56556 − t) + 0.705254t − 1.53797)2. Numerically sup t∈I2 {t + kQ(t) − S1(t)k} = τ0 + kQ(τ0) − S1(τ0)k = τ1 + kQ(τ1) − S1(τ1)k {t + kQ(t) − S1(t)k} ≈ 3.37387. = sup t∈I3 The reader may also consult Figure 5. 21 Figure 6: (Left) The queen must be in region R at time f(s3). Here s3 = E and q3 = F . 4 Lower Bounds In this section we derive lower bounds for evacuation. In Section 4.1 we treat the case of n = 1 (see Theorem 4.1) and in Section 4.2 we treat the case of n = 2 and 3 (see Theorem 4.3). 4.1 Lower Bound for PE1 We will derive the lower bound using an adversarial argument placing the exit at an un- known vertex of a regular hexagon. 3/2 ≈ 4.3896 Theorem 4.1. The worst-case evacuation time for PE1 is at least 3+π/6+ Proof: At time 1 + π/6, at most π/3 of the perimeter of the circle can have been explored by the queen and servant. Thus, there is a regular hexagon, none of whose vertices have been explored. If the exit is at one of these vertices, by Theorem 4.2, it takes 2 + 3/2 for the queen to evacuate. The total time is 1 + π/6 + 2 + 3/2. √ √ √ Next we proceed to provide a lower bound on a unit-side hexagon. Label the vertices of the hexagon V as A, . . . , F as shown in Figure 6. Fix an evacuation algorithm A. For any vertex v of the hexagon, we call f(v) the time of first visit of the vertex v by either the servant or the queen, according to algorithm A. We call q(v) the time that the queen gets to the vertex v. Clearly, q(v) ≥ f(v), and if the queen arrives at the vertex no later than the servant, q(v) = f(v). Theorem 4.2. For any algorithm A, the evacuation time for the queen when the exit is at one of the vertices of the hexagon is maxv∈V {q(v)} ≥ 2 + Proof: Suppose there is an algorithm in which the queen can always evacuate in time 3/2. Consider the trajectories of the servant and the queen. If either the queen < 2 + 3/2. √ √ 22 BCDEFA2+p32−f(s3)R√ √ 3/2 − f(s3) ≤ √ or the servant are the first to visit 4 vertices, then for the fourth such vertex v, we have f(v) ≥ 3, a contradiction. Therefore, the queen is the first to visit three vertices, and the servant is the first to visit three vertices. We denote the three vertices visited first by the servant as s1, s2, s3 (in the order they are visited) and the three vertices visited first by the queen as q1, q2, q3, and note that they are all distinct. Notice that neither s3 nor q3 can be visited before time 2, that is, f(s3), f(q3) ≥ 2. If f(q3) ≤ f(s3), then we place the exit at s3, and the queen needs time at least 1 to get to s3, which implies that T ≥ q(s3) ≥ f(q3) + 1 ≥ 3, a contradiction. We conclude that at time f(s3), the queen is yet to visit q3. Since the exit can be at either s3 or q3, at time 3/2 from both s3 and q3. f(s3), the queen must be at distance < 2 + Assume without loss of generality that s3 = E (see Figure 6). Since A, B, D are all 3 from E, we conclude that q3 is either C or F . Assume without at distance at least loss of generality that q3 = F . Let R denote the lens-shaped region that is at distance 3/2 − f(s3) from both E and F . Recall that at time f(s3), the queen must be < 2 + √ inside the region R. Notice that if f(s3) ≥ 1.5 + 3/2, the region R is empty, yielding a √ contradiction. So it must be that 2 ≤ f(s3) < 1.5 + Clearly s2 6= F since q3 = F . If s2 6= C, then f(s3) ≥ √ We now work backwards to deduce the trajectories of the servant and the queen. 3/2, a contradiction. Therefore, s2 = C. By the same reasoning, s1 = A. Therefore, the queen is the first to visit D and B. If q1 = D and q2 = B, we place the exit at E; since f(q2) ≥ 1 and dist(B, E) = 2, we have T ≥ q(E) ≥ 3, a contradiction. Thus, q2 = D and q1 = B. Consider the location of the queen at time 1. If she is at distance ≥ 1 + 3/2 from C at time 1, then if the exit is at C, q(C) ≥ 2 + 3/2. So at time 1, the queen must be at 3/2 from C and consequently she is at distance ≥ 1−√ distance < 1 + 3/2 from vertex D. Therefore f(q2) = f(D) ≥ 2 − √ 3/2. Also, f(D) < 1.5 since if the queen reaches √ 3/2 > f(s3). D at or after time 1.5, she cannot reach the region R before time 1.5 + So f(D) ≤ f(s3). If the exit is at E = s3, the queen cannot reach the exit before time √ f(D) + dist(D, E) ≥ 2 − √ 3, concluding the proof by contradiction. 3 + 1 > 1.5 + 3/2 + 3 = 2 + √ √ √ √ √ 3/2. √ We remark that the above bound is optimal, and is achieved by the algorithm depicted in Figure 7. 4.2 Lower Bounds for PE2 and PE3 - Proof Outline In the case of n = 2 and n = 3 the proof is rather technical. Next we present a high level outline as to why the lower bounds hold. Theorem 4.3. The worst-case evacuation time for PE2 is at least 3.6307 and for PE3 at least 3.2017. Throughout this section we will use T to refer to the evacuation time of an arbitrary algorithm and use U to refer to the unit circle which must be evacuated. The main thrust of the proof relies on a simple idea – the queen should aid in the exploration of U. This is immediately evident for the particular case of n = 2 since, if the queen does not explore, it will take time at least 1 + π for the servants to search all of 23 Figure 7: Blue trajectory: servant and red trajectory: queen. At point H, if the queen hears of an exit at E, she goes there, otherwise she goes to F . U and we already have an upper bound smaller than this (Theorem 3.3). Thus, a general overview of the proof is as follows: we show that in order to evacuate in time T the queen must explore some minimum length of the perimeter of U. We will then demonstrate that the queen is not able to explore this minimum amount in any algorithm with evacuation time smaller than what is given in Theorem 4.3. To be concrete, consider the case of n = 2 and assume that we have an algorithm with evacuation time T < 1 + π. Then, in order for the robots to have explored all of U in time T , the queen must explore a subset of the perimeter of total length at least 2(1 + π − T ). Intuitively, this minimum length of perimeter will increase in size as T decreases. Now consider that it is not possible for the queen to always remain on the perimeter (indeed, in each of the algorithms presented, the queen leaves the perimeter). To see why this is consider that, in any algorithm with evacuation time T , it must be the case that all unexplored points of U are located a distance no more than T − t from the queen at all times t ≤ T . If the queen is on the perimeter at any time t satisfying T − t ≤ 2, then, there will be some arc θ(t,T ) ⊂ U (see Lemma 4.4) such that all points of θ(t,T ) are at a distance at least T − t from the queen. Thus, if the queen is to be on the perimeter at the time t we can conclude that all of the arc θ(t,T ) must have already been discovered. However, we will find (see Lemma 4.5) that θ(t,T ) will often grow at a rate much larger than the robots can collectively explore and at some point the queen will have to leave the perimeter. In fact, there will be an interval of time during which it is not possible for the queen to be exploring and this in turn implies that there is a maximum amount of perimeter that can be explored by the queen. Intuitively, the maximum length of perimeter that can be explored by the queen will decrease as T decreases. The lower bound will result by balancing the minimum amount of perimeter the queen needs to search and the maximum amount of perimeter that the queen is able to search. The above argument will need a slight modification in the case of n = 3. In this case we will show that there is some critical time t∗ before which the queen must have explored 24 some minimum amount of perimeter. Again, the lower bound follows by balancing the maximum amount of perimeter the queen can explore by the time t∗ and the minimum amount of perimeter the queen needs to explore before the time t∗. 4.3 Lower Bounds for PE2 and PE3 - Proof Details In this section we present the complete details of the proofs for the lower bounds in the cases n = 2 and n = 3. Throughout this section we will use T to refer to the evacuation time of an arbitrary algorithm and use U to refer to the unit circle which must be evacuated. The idea of the proofs are to bound the amount of perimeter the queen can search for a given evacuation time T and then show that the queen must search a minimum amount of the perimeter in order to achieve the evacuation time T . The lower bounds result by balancing the minimum amount of perimeter the queen must search with the maximum amount of perimeter the queen can search. We begin with two lemmas which will be used for both the n = 2 and n = 3 bounds. Their necessity will become apparent shortly. Lemma 4.4. Consider any r < 2 and a point P ∈ U. Define the circle DP as the disk centered on P with radius r. Then the subset of the perimeter of U which is not contained (cid:1). in DP has length θ = 4 cos−1(cid:0) r to this figure, one can easily observe that r = 2 sin(cid:0) π θ we find that θ = 4 cos−1(cid:0) r (cid:1). 2 2 Proof: Without loss of generality assume that the point P is located at (−1, 0). Since r < 2 the disks U and DP will intersect at two boundary points A and B between which the distance along the perimeter of U is θ. This situation is depicted in Figure 8. Referring (cid:1) = 2 cos(cid:0) θ (cid:1). Rearranging for 2 − θ 4 4 Figure 8: Setup for the proof of Lemma 4.4. The boundary of the disk DP is indicated in blue. The arc of U which is excluded from DP is highlighted in red and has length θ. 25 Lemma 4.5. Consider the function θ(t,T ) = 4 cos−1(cid:0)T −t (cid:1) with T > 0. Then dθ dt > 3 for t satisfying T − 2 < t < T − 2 2 dt > 2 √ 5. 3 for all t satisfying T − 2 < t < T and dθ Furthermore, dθ Proof: The rate of change of θ(t,T ) with t is given by dT < −2 for all T − 2 < t < T . 4p4 − (T − t)2 . dθ dt = From this relation it is simple to confirm that dθ for T − 2 < t < T − 2 3 function θ(t,T ) that dθ dT < −2 for all T − 2 < t < T . √ dr > 3 5. It should also be obvious by the symmetry of T and t in the dt > 2 for T −2 < t < T and that dθ(r) 4.3.1 Lower bound for n = 2 We begin with the main result of the section. Theorem 4.6. For n = 2 and any algorithm the queen cannot be evacuated in time less than T2 which is the solution to the equations τ = T2 − 2 cos (cid:19) (cid:18) τ − 1 2 (cid:18)2t∗ + τ 2(T2 + 1) − 3 4 4 (cid:19) . t∗ = 1 T2 = t∗ + 2 cos Solving these equations numerically gives τ ≈ 1.7815, t∗ ≈ 2.3154, and T2 ≈ 3.6307. We will see that the queen cannot be located on the perimeter of the circle during the interval of time (τ, t∗) and thus τ − 1 represents the maximum amount of perimeter that can be explored by the queen before the time t∗. The time t∗ is chosen such that for all T < T2 a solution to the equations in Theorem 4.6 do not exist, and, as such, τ − 1 will represent the maximum length of the perimeter that can be explored by the queen. In the following lemma we show that the queen must explore a length of the perimeter greater than τ − 1 in order to evacuate in time less than T2. Lemma 4.7. For n = 2 and any evacuation algorithm with T < 1 + π, the queen must explore a subset of the perimeter of length y ≥ 2(1 + π − T ). In particular, if T < T2, we need y > 2(1 + π − T2) ≈ 1.0217. Proof: take time 1 + 2π−y 2 circle in time T and therefore 1 + 2π−y T < T2 ≈ 3.6307 we need y > 1.0217. If the queen explores a subset of the perimeter of length y then the robots will to explore the circle. The robots need to at least explore the entire 2 ≤ T , or, equivalently, y ≥ 2(1 + π − T2). For We will now show that the maximum length of perimeter the queen can explore is less than τ − 1 if T < T2. This will be the goal of the next two lemmas. 26 dα > 0. to T gives us 2(t − 1) + 1 dT ≥ 1 satisfying 0 < α ≤ t and t satisfying 1 < t ≤ T . Then dt then dt Lemma 4.8. Consider the equation T = t + 2 cos(cid:0) 1 Proof: Implicitly differentiating the equation T = t+2 cos(cid:0) 1 4 α(cid:1) . Implicitly differentiating the equation T = t + 2 cos(cid:0) 1 4 α(cid:1) 4 α(cid:1) . 1 − sin(cid:0) 1 sin(cid:0) 1 1 − sin(cid:0) 1 Since the sine function ranges from −1 to 1 we can easily see that dt 2(t − 1) + 1 2(t − 1) + 1 = 1 2 · dt dT = α gives us dt dα 1 We can esily see that the denominator of dt provided that the numerator is positive. This clearly occurs for 1 t < 1 + 2π − α 2 . 2(t − 1) + 1 2(t − 1) + 1 dα will never be negative and thus dt 2(t − 1) + 1 dα > 0 4 α < π or 2 , and, if 0 < t < 1+2π− α 2 α(cid:1) with T > 0, α 4 α(cid:1) with respect 2 2(t − 1) + 1 dT ≥ 1 2. 4 α(cid:1) with respect to Lemma 4.9. Define τ as in Theorem 4.6. Then, for n = 2 and any evacuation algorithm with T < T2, the queen cannot explore a subset of the perimeter with length y > τ − 1. Proof: We start with an observation: if the queen is to evacuate in time T , then, at any time t < T , all points of U that are a distance greater than T − t from the queen must be explored by a robot. If the queen is located on the perimeter at the time t > T − 2 then by Lemma 4.4 there is an arc of length θ(t,T ) = 4 cos−1 (cid:18)T − t (cid:19) 2 all points of which lie a distance greater than T − t from the queen (as an abuse of notation we will refer to the arc with length θ(t,T ) as θ(t,T )). Thus, in order for the queen to be on the perimeter at the time t, the arc θ(t,T ) must be explored. As we have 3 robots in total the maximum length of θ(t,T ) that can be explored at any time t is 3(t− 1). However, we claim that the queen cannot have explored any of θ(t,T ) if the time t satisfies t < 1 2(T +1). Indeed, observe that the endpoints of θ(t,T ) lie a distance T − t away from the queen (by definition) and the queen – who took a unit of time to reach the perimeter – could have explored a point on the perimeter at most a distance t − 1 from her current position. Thus, 2(T + 1), the queen cannot have explored any if t − 1 < T − t, or, alternatively, t < 1 of the arc θ(t,T ). We must therefore have θ(t,T ) ≤ 2(t − 1) for times t that satisfy t < 1 3 > 3 and thus we can assume that T > 3. We make the following claim: if T < T2 then the smallest time t0 > 0 solving 2(T + 1). We note that, if this is the θ(t0,T ) = 2(t0 − 1) satisfies dθ case, the queen will have to leave the perimeter at the time t0 (since she has not explored any of the arc θ(t,T ) and, immediately after the time t0, θ(t,T ) will be too large to have been explored by the servants alone). 2(T + 1). We note that there is a trivial lower bound of 1 + 2π > 2 and t0 < 1 (cid:12)(cid:12)t=t0 dt 27 dt 2 (cid:19) 2(t0 − 1) to get We first show that t0 < 1 2(T + 1). To this end we rearrange the equation θ(t0) = t0 = T − 2 cos (cid:18) t0 − 1 (cid:12)(cid:12)t=τ ≈ 5.2511 > 2 and τ < 1 which is the definition of τ in Theorem 4.6 (in the case that T = T2). One can easily confirm that in the case of T = T2 we have dθ 2(T + 1). Now observe that θ(t,T ) is a decreasing function of T and this implies that for T < T2 we have θ(τ,T ) > θ(τ,T2). We can therefore conclude that the time t0 must occur earlier 2(T + 1). Since than the time τ. We note that τ < 2 and, since T ≥ 3, we have τ < 1 2(T + 1). t0 < τ we can conclude that t0 < 1 The second part of the claim follows directly from Lemma 4.5 where we show that dt > 2 for all t satisfying T − 2 < t < T . As the queen must leave the perimeter at the time t0 < τ, by Lemma 4.7, we can say that the queen must be able to return to the perimeter and explore before the algorithm terminates. Thus, consider the smallest time t1 > t0 at which the queen may return to the perimeter. In order for the queen to be on the perimeter we will still need the arc θ(t,T ) to be completely explored. However, in this case it may be possible that t1 ≥ 1 2(T +1) and as such the queen could have explored at most a length t0−1 of θ(t,T ) at the time t1. We can therefore conclude that t1 will satisfy θ(t1) = 2(t1 − 1) + y with y = 0 if t1 < 1 2(T + 1), and y ≤ t0 − 1 if t1 ≥ 1 2(T + 1). Writing the equation θ(t1) = 2(t1 − 1) + y in full and rearranging we find that dθ t1 = T − 2 cos (cid:19) (cid:18)1 2(t1 − 1) + 1 4 y 2(T + 1) separately. 2(T + 1) and t1 ≥ 1 . We will now consider the cases t1 < 1 2(T + 1) Case 1: t1 < 1 In this case t1 can be observed to satisfy the same equation as t0. We claim that this is dt > 2 and the arc θ(t,T ) will not possible if t1 > t0. Indeed, by Lemma 4.5 we have dθ always grow at a rate larger than the servants alone can explore. Thus, a solution to the equation θ(t1) = 2(t1 − 1) with t1 > t0 does not exist. This implies that the queen can explore a maximum subset of the perimeter of total length t0−1 < τ −1 if t1 < 1 2(T +1). 2(T + 1) Case 2: t1 ≥ 1 In this case t1 satisfies (cid:18)1 2(t1 − 1) + 1 4 y (cid:19) . t1 = T − 2 cos dy > 0 (see Lemma 4.8) we will show that, even when Although it can be confirmed that dt1 (cid:19) 2(T +1). Thus we assume t1 is as large as possible (i.e. y = t0 −1), we cannot have t1 ≥ 1 that t1 satisfies 4(t0 − 1) (cid:18)1 2(t1 − 1) + 1 t1 = T − 2 cos . 28 Now write t1 = t1(T ) as a function of T and note that, by Lemma 4.8, we have dt1 2. Using this we can say that t1(T2) − t1(T ) > 1 1 have t1(T2) = 1 T ). Rearranging this inequality gives us t1(T ) < 1 assumption that t1 ≥ 1 the proof. dT > 2(T2 − T ). By definition of T2 we 2(T2 − 2(T + 1) which contradicts with our 2(T + 1). This concludes 2(T + 1) and we must conclude that t1 < 1 2(T2 + 1) and we can therefore write 1 2(T2 + 1) − t1(T ) > 1 At this point the proof of Theorem 4.6 is rather trivial. Proof: (Theorem 4.6) Assume that we have an algorithm with evacuation time T < T2. Then, by Lemma 4.7, the queen must explore a subset of the perimeter of length at least y > 1.0217. However, by Lemma 4.9, the queen can only explore a subset of the perimeter of length y < τ − 1 ≈ 0.7815 if T < T2. It is therefore not possible for the queen to evacuate in time less than T2. 4.3.2 Lower bound for n = 3 The main result of this section is given below: Theorem 4.10. For n = 3 and any algorithm the queen cannot be evacuated in time less than T3 which is the solution to the equations (cid:19) (cid:18)3 4(τ − 1) (cid:18)−2 (cid:19) − (τ − 1) 3 cos−1 (cid:18)3(t∗ − 1) + (τ − 1) (cid:19) 3 3 2 τ = T3 − 2 cos t∗ = 1 + 2 T3 = t∗ + sin Solving these equations numerically gives τ ≈ 1.2319, t∗ ≈ 2.4564, and T3 ≈ 3.2017. As before, τ represents the beginning of an interval of time during which the queen cannot be located on the perimeter. In this case, however, t∗ is not the first time at which it is possible for the queen to return to the perimeter. Instead it represents a particularly critical time of any algorithm with n = 3 at which the evacuation time is maximized (although it will happen that t∗ occurs before the queen can return to the perimeter). We will show that the queen must explore a subset of the perimeter with total length more than τ − 1 before the time t∗ in order to evacuate in time less than T2. We begin with a lemma that was first introduced in [10]: Lemma 4.11. Consider a perimeter of a disk whose subset of total length u+ > 0 has not been explored for some > 0 and π ≥ u > 0. Then there exist two unexplored boundary points between which the distance along the perimeter is at least u. This next lemma is used to determine the critical time t∗. 29 2 (cid:17) n ≤ t ≤ 1 + 2π−y n , it takes time at least T = t + sin(cid:16) n(t−1)+y Lemma 4.12. Consider an evacuation algorithm with n servants and assume that at the time t the queen has explored a total subset of the perimeter of length y. Then, for x and y satisfying 1 + π−y to evacuate the queen. Proof: Consider an algorithm with evacuation time T and with n servants. Then, at the time t, the total length of perimeter that the robots have explored is at most n(t− 1) + y ≥ π (since each robot may search at a maximum speed of one, the queen has explored a subset of length y, and the robots need at least a unit of time to reach the perimeter). Thus, by Lemma 4.11, there exists two unexplored boundary points between which the distance along the perimeter is at least 2π − n(t − 1) − y − for any > 0. The chord and an adversary may place the exit at either endpoint of this chord. The queen will therefore take at least more time to evacuate and the total evacuation time will be at . As this is true for any > 0 taking the limit → 0 we connecting these points has length at least 2 sin(cid:16) sin(cid:16) least t + sin(cid:16) obtain T ≥ t + sin(cid:16) (cid:17) = t + sin(cid:16) n(t−1)+y 2 π − n(t−1)+y 2 π − n(t−1)+y π − n(t−1)+y π − n(t−1)+y − 2 − 2 (cid:17) . (cid:17) (cid:17) (cid:17) 2 − 2 2 2 In the next two lemmas we show that in order to evacuate in time T < T2 the queen must explore a length of the perimeter greater than τ − 1 and then demonstrate that this is not possible. Lemma 4.13. Define τ and t∗ as in Theorem 4.10. Then, for n = 3 and any evacuation algorithm with T < T3, the queen must explore a subset of U with total length y > τ − 1 before the time t∗. Proof: Consider an algorithm with evacuation time T < T3. We make the assumption that the queen has only explored a subset of total length y < τ − 1 at the time t∗ and show that this leads to a contradiction. for all y satisfying 0 ≤ y ≤ τ − 1 Observe that t∗ satisfies 1 + π−y 3 ≤ t∗ ≤ 1 + 2π−y 3 and thus, by Lemma 4.12, we can write (cid:18)3(t∗ − 1) + y (cid:18)3(t∗ − 1) + y 2 (cid:19) (cid:19) . . 2 30 T ≥ t∗ + sin Since T < T3 we also have Since T3 = t∗ + sin(cid:16) 3(t∗−1)+(τ−1) (cid:17) T3 > t∗ + sin 2 (cid:18)3(t∗ − 1) + (τ − 1) (cid:19) we further have > sin sin (cid:18)3(t∗ − 1) + y (cid:19) . 2 2 3 we know that sin(cid:16) 3(t∗−1)+y 2 Finally, since t∗ ≥ 1 + π−y argument and thus we get (cid:17) is a decreasing function of its 3(t∗ − 1) + (τ − 1) 2 3(t∗ − 1) + y 2 < which implies that y > τ − 1 which contradicts with our assumption that y < τ − 1. all of the arc θ(t,T ) = 4 cos−1(cid:0)T −t (cid:1) must be explored. Since we have 4 robots in total, Lemma 4.14. Define τ and t∗ as in Theorem 4.10. Then, for n = 3 and any evacuation algorithm with T < T3, the queen cannot explore a subset of the perimeter with length y > τ − 1 before the time t∗. Proof: As was the case for n = 2, if the queen is to be on the perimeter at the time t then the maximum length of arc that can be explored at any time t is 4(t − 1). However, we can again say that the queen cannot search any of the arc θ(t) if t ≤ 1 2(T + 1). We must therefore have θ(t,T ) ≤ 3(t − 1) for times t that satisfy t < 1 2(T + 1). Assume first that T ≥ 3. We make the following claim: if 3 ≤ T < T3 then the smallest time t0 > 0 solving θ(t0,T ) = 3(t0−1) satisfies dθ 2(T +1). If this is the case the queen will have to leave the perimeter at the time t0. 2(T + 1). Let us rearrange the equation θ(t0,T ) = 3(t0 − 1) to get We first demonstrate that t0 < 1 > 3 and t0 < 1 (cid:12)(cid:12)t=t0 dt 2 2 3 31 t0 = T − 2 cos (cid:18)3 (cid:19) 4(t0 − 1) (cid:12)(cid:12)t=τ > 3 and τ < 1 dt √ which is the definition of τ in Theorem 4.10 (in the case that T = T3). One can easily confirm that in the case of T = T3, both dθ 2(T + 1). Now observe that θ(t,T ) is a decreasing function of T and this implies that for T < T3 we have θ(τ,T ) > θ(τ,T3). The time t0 must therefore occur earlier than the time τ. We note that τ < 2 and, since we are assuming that T ≥ 3, we have τ < 1 2(T + 1). Since t0 < τ we can finally conclude that t0 < 1 The second part of the claim follows from Lemma 4.5 if we can show that t0 < T − √ 5. We note that T ≥ 3 and thus T − 2 5 ≥ 1.5093. Since τ ≈ 1.2319 and t0 < τ we 2 3 3 can clearly see that t0 < T − 2 3 If T < 3 then it should be obvious that the queen cannot even be at the perimeter at the time t = 1. Thus, in this case, we take t0 = 1. 2(T + 1). √ Since the queen must leave the perimeter at the time t0 < τ, by Lemma 4.13, we know that the queen must be able to return to the perimeter and explore before the time t∗. We claim that this is not possible. Indeed, observe that the queen cannot return to the perimeter until the earliest time t > t0 at which θ(t) = 3(t−1)+y (where we have set y < τ−1 as the length of the arc θ(t) explored by the queen). Thus, in order for the queen to have returned to the perimeter before the time t∗ we must have θ(t∗) ≤ 3(t − 1) + y. However, since (cid:1) > 4 cos−1(cid:0)T3−t∗ T < T3 we have θ(t∗) = 4 cos−1(cid:0)T −t∗ (cid:17) = sin(cid:0)cos−1(cid:0)−2 sin(cid:16) 3(t∗−1)+(τ−1) (cid:1)(cid:1) = (cid:1). We note that T3 − t∗ = (cid:17) ≈ q 5 9 and thus θ(t∗) > 4 cos−1(cid:16)√ 4.7556. Since τ ≈ 1.2319, and t∗ ≈ 2.4564 we have 3(t∗−1)+ y ≤ 3(t∗−1)+(τ −1) ≈ 5 6 5. 2 2 4.6010. We can therefore see that it is not the case that θ(t∗) ≤ 3(t − 1) + y and thus the queen cannot have returned to the perimeter before the time t∗. We can finally conclude that the queen can only explore a subset of the perimeter of length t0 − 1 < τ − 1 before the time t∗. At this point the proof of Theorem 4.10 is trivial. Proof: (Theorem 4.10) Assume we have an algorithm with evacuation time T < T3. Then, by Lemma 4.13, the queen must explore a subset of the perimeter of length at least τ − 1 by the time t∗. However, by Lemma 4.14, the queen can only explore a subset of the perimeter of length y < τ − 1 if T < T3. We must therefore conclude that it is not possible for the queen to evacuate in time less than T3. 5 Conclusion We considered an evacuation problem concerning priority searching on the perimeter of a unit disk where only one robot (the queen) needs to find the exit. In addition to the queen, there are n ≤ 3 other robots (servants) aiding the queen by contributing to the exploration of the disk but which do not need to evacuate. We proposed evacuation algorithms and studied non-trivial tradeoffs on the queen evacuation time depending on the number n of servants. In addition to analyzing tradeoffs and improving the bounds obtained for the wireless communication model, an interesting open problem would be to investigate other models with limited communication range, e.g., face-to-face. References [1] R. Ahlswede and I. Wegener. Search problems. Wiley-Interscience, 1987. [2] S. Alpern and S. Gal. The theory of search games and rendezvous, volume 55. Kluwer Academic Publishers, 2002. [3] Steve Alpern, Robbert Fokkink, Leszek Gasieniec, Roy Lindelauf, and V.S. Subrah- manian, editors. Ten Open Problems in Rendezvous Search, pages 223–230. Springer NY, New York, NY, 2013. [4] R. Baeza Yates, J. Culberson, and G. Rawlins. Searching in the plane. Information and Computation, 106(2):234–252, 1993. [5] R. Baeza-Yates and R. Schott. Parallel searching in the plane. Computational Geom- etry, 5(3):143–154, 1995. [6] A. Beck. On the linear search problem. 1964. Israel J. of Mathematics, 2(4):221–228, [7] R. Bellman. An optimal search. SIAM Review, 5(3):274–274, 1963. 32 [8] S. Brandt, F. Laufenberg, Y. Lv, D. Stolz, and R. Wattenhofer. Collaboration without communication: Evacuating two robots from a disk. In Proceedings of Algorithms and Complexity - 10th International Conference, CIAC 2017, Athens, Greece, May 24-26, 2017, pages 104–115, 2017. [9] J. Czyzowicz, S. Dobrev, K. Georgiou, E. Kranakis, and F. MacQuarrie. Evacuating two robots from multiple unknown exits in a circle. Theor. Comput. Sci., 709:20–30, 2018. [10] J. Czyzowicz, L. Gasieniec, T. Gorry, E. Kranakis, R. Martin, and D. Pajak. Evacu- ating robots from an unknown exit located on the perimeter of a disc. In Proceedings DISC, Austin, Texas, pages 122–136. Springer, 2014. [11] J. Czyzowicz, K. Georgiou, M. Godon, E. Kranakis, D. Krizanc, W. Rytter, and M. Wlodarczyk. Evacuation from a disc in the presence of a faulty robot. In Pro- ceedings SIROCCO 2017, 19-22 June 2017, Porquerolles, France, pages 158–173, 2018. [12] J. Czyzowicz, K. Georgiou, R. Killick, E. Kranakis, D. Krizanc, L. Narayanan, J. Opatrny, and S. Shende. Priority evacuation from a disk using mobile robots, 2018, Submitted. [13] J. Czyzowicz, K. Georgiou, E. Kranakis, L. Narayanan, J. Opatrny, and B. Vogten- huber. Evacuating robots from a disk using face-to-face communication (extended abstract). In Proceedings of Algorithms and Complexity, CIAC 2015, Paris, France, May 20-22, 2015, pages 140–152, 2015. [14] J. Czyzowicz, E. Kranakis, D. Krizanc, L. Narayanan, J. Opatrny, and S. Shende. Wireless autonomous robot evacuation from equilateral triangles and squares. In Pro- ceedings of Ad-hoc, Mobile, and Wireless Networks, ADHOC-NOW, Athens, Greece, June 29 - July 1, 2015, pages 181–194, 2015. [15] I. Lamprou, R. Martin, and S. Schewe. Fast two-robot disk evacuation with wireless communication. In Proceedings DISC, Paris, France, pages 1–15, 2016. [16] D. Pattanayak, H. Ramesh, P.S. Mandal, and S. Schmid. Evacuating two robots from two unknown exits on the perimeter of a disk with wireless communication. In Proceedings of the 19th International Conference on Distributed Computing and Networking, ICDCN 2018, Varanasi, India, January 4-7, 2018, pages 20:1–20:4, 2018. [17] L. Stone. Theory of optimal search. Academic Press New York, 1975. 33
1801.00216	4	1801	2019-12-08T11:20:20	Emotion-Based Crowd Simulation Model Based on Physical Strength Consumption for Emergency Scenarios	[ "cs.MA" ]	Increasing attention is being given to the modeling and simulation of traffic flow and crowd movement, two phenomena that both deal with interactions between pedestrians and cars in many situations. In particular, crowd simulation is important for understanding mobility and transportation patterns. In this paper, we propose an emotion-based crowd simulation model integrating physical strength consumption. Inspired by the theory of "the devoted actor," the movements of each individual in our model are determined by modeling the influence of physical strength consumption and the emotion of panic. In particular, human physical strength consumption is computed using a physics-based numerical method. Inspired by the James-Lange theory, panic levels are estimated by means of an enhanced emotional contagion model that leverages the inherent relationship between physical strength consumption and panic. To the best of our knowledge, our model is the first method integrating physical strength consumption into an emotion-based crowd simulation model by exploiting the relationship between physical strength consumption and emotion. We highlight the performance on different scenarios and compare the resulting behaviors with real-world video sequences. Our approach can reliably predict changes in physical strength consumption and panic levels of individuals in an emergency situation.	cs.MA	cs	JOURNAL OF LATEX CLASS FILES, VOL. X, NO. X, OCTOBER 2018 1 Emotion-Based Crowd Simulation Model Based on Physical Strength Consumption for Emergency Scenarios Mingliang Xu, Chaochao Li, Pei Lv, Wei Chen, Zhigang Deng, Bing Zhou and Dinesh Manocha Abstract -- Increasing attention is being given to the modeling and simulation of traffic flow and crowd movement, two phenomena that both deal with interactions between pedestrians and cars in many situations. In particular, crowd simulation is important for understanding mobility and transportation patterns. In this paper, we propose an emotion-based crowd simulation model integrating physical strength consumption. Inspired by the theory of "the devoted actor," the movements of each individual in our model are determined by modeling the influence of physical strength consumption and the emotion of panic. In particular, human physical strength consumption is computed using a physics-based numerical method. Inspired by the James-Lange theory, panic levels are estimated by means of an enhanced emotional contagion model that leverages the inherent relationship between physical strength consumption and panic. To the best of our knowledge, our model is the first method integrating physical strength consumption into an emotion-based crowd simulation model by exploiting the relationship between physical strength consumption and emotion. We highlight the performance on different scenarios and compare the resulting behaviors with real-world video sequences. Our approach can reliably predict changes in physical strength consumption and panic levels of individuals in an emergency situation. Index Terms -- Pedestrian traffic simulation, crowd simulation, emotional contagion, James-Lange theory ! 1 INTRODUCTION E FFICIENT and accurate crowd simulation is useful for intelligent transportation systems since it can help im- prove emergency planning and prevent congestion in transit hubs such as train stations and airports [1]. One can also an- alyze human mobility through the trajectories obtained by crowd simulation models to get more knowledge of pedes- trian mobility behaviors in both qualitative and quantitative ways. Because of various complex factors, it is challenging to model realistic crowd behaviors in emergency scenarios. At a broad level, crowd behavior in emergencies is governed by panic and physical strength consumption [2]. The main purpose of crowd simulation algorithms is to model the movements (in terms of speed and direction) of individuals in a crowd [3]. We basically deal with two aspects of human motivations: physiological and psycho- logical factors. Physical strength consumption and emotion [4] are two representative physiological and psychological factors, respectively. Both have a great influence on individ- ual movements. These two factors influence each other and 310058. • Mingliang Xu, Chaochao Li, Pei Lv, and Bing Zhou are with Center for Interdisciplinary Information Science Research, Zhengzhou University, 450000. • Wei Chen is with State Key Lab of CAD&CG, Zhejiang University, • Zhigang Deng is with Department of Computer Science, University of • Dinesh Manocha is with Department of Computer Science and Electrical & Computer Engineering, University of Maryland, College Park, MD, USA. E-mail: {iexumingliang, ielvpei, iebzhou} @zzu.edu.cn; [email protected]; [email protected]; [email protected]; [email protected] Houston, Houston, TX, 77204-3010. evolve dynamically. It is important to describe the inher- ent relationship between these two factors, which is more obvious in emergency or evacuation situations [5]. Many approaches incorporate emotions of individuals in crowd simulations, making it one of the most commonly used psychological factors [6]. Panic can prevent an individual from taking proper actions in emergency situations [4]. Re- searchers have observed that external dangers can directly cause changes in panic levels in an individual, thereby further determining his or her movements [7]. We mainly focus on the emotion of panic in emergency situations. Most of the previous studies don't consider the effect of physical strength consumption on panic [8]. Physical strength is a person's or animal's ability to exert force on physical objects using muscles [9]. Physical strength consumption is defined as the energy expenditure [10] of a human, which directly affects that human's moving speed [11]. However, it is difficult to describe the inherent relationship between phys- ical strength consumption and panic and to then combine these factors to determine the movement of each individual [5]. Therefore, an emotion-based crowd simulation model integrating physical strength consumption is challenging due to the following reasons: (1) It is difficult to model the physical strength con- sumption of an individual in a crowd accurately [12]. This task involves considering many factors that are needed to quantify the influence of physical strength consumption on crowd movement [13]. (2) Accurately modeling an individual's panic level in a crowd is difficult because of its constant and dynamic changes [14]. Various factors such as physical strength con- sumption and individual movement affect panic levels. Inspired by the theory of "the devoted actor" [2], which JOURNAL OF LATEX CLASS FILES, VOL. X, NO. X, OCTOBER 2018 2 2.1 Simple crowd simulation models In this subsection, we summarize representative crowd simulation models that do not consider psychological or physiological factors [1], [15], [16], [17]. In the real world, many environmental factors influence an individual's movement, i.e. scene layout, moving pedes- trians, and stationary groups [18], [19], [20]. During the evacuation of a crowd, the behavioral choice of an individ- ual is highly dependent on the moving directions of nearby individuals, the hazard location, and obstacles [21]. Mous- said et al. [22] propose a cognitive science approach based on behavioral heuristics. Guided by visual information, pedestrians apply two simple cognitive procedures to adapt their moving directions and speed. Zhou et al. [23] propose a fuzzy logic approach to model and simulate pedestrian dynamic behaviors, which are based on human experiences and human knowledge, and perceptual information ob- tained from interactions with the surrounding environment. Zhou et al. [24] focus on the role of leaders who can guide the movements of passengers during the evacuation. Cassol et al. [25] focus on global path planning with the main goal of identifying the best evacuation routes for a specific population when leaving a certain building. To realize better behavioral choices, most approaches calculate the position of each individual at the next time step to obtain a conflict- free moving path in a global scenario [26]. However, these approaches are not applicable to highly complex scenes with dense crowds. Other approaches use local obstacle avoidance methods. Namely, once the movement state of an individual is determined, the movement states of other individuals are updated by using local collision avoidance techniques [27]. In practice, these approaches face many difficulties in terms of accurately controlling the individual movements. Researchers in this field are increasingly focusing on in- tegrating global path planning and local obstacle avoid- ance [28], [29]. Weiss et al. [30] model collision avoidance constraints both in terms of short and long-term ranges to deal with sparse and dense crowds. In [31], intergroup- and intragroup-level techniques are presented to perform coher- ent and collision-free navigation using reciprocal collision avoidance. Mutual information about the dynamic crowd is used to guide agents' movements by combining both macroscopic and microscopic controls [32]. By constructing a visual tree, the shortest path without collisions is obtained in [33]. In addition, in [34], [35], [36], and [37], path planning and navigation algorithms are described for crowd simula- tion in complex contexts. Furthermore, in [38], an effective long-range collision avoidance algorithm is proposed. In contrast to these works, our model enhances the tradi- tional social force model to avoid collisions with surround- ing individuals and obstacles by combining panic and phys- ical strength consumption calculations. Traditional crowd simulation models are not concerned with this approach. In our model, we mainly deal with moving directions and moving speeds, which are largely influenced by panic and physical strength consumption during a relatively short period of time. Fig. 1: The relationships among physical strength consump- tion, panic, and movement of individuals in a crowd. Phys- ical strength consumption is calculated according to the actual speed, mass, and moving time of individuals. The panic levels are determined by the location and physical strength consumption of the individual, and they further affect the individual's desired speed. Moreover, each indi- vidual's current panic level will affect his or her moving direction by changing the acceleration based on the inferred force. shows that an individual's physiological state has an effect on his or her psychological state, we propose the first (to the best of our knowledge) emotion-based crowd simulation model based on physical strength consumption (illustrated in Figure 1). The main contributions of our work include: • We introduce a physical strength consumption calcu- lation method based on how individuals work under the laws of physics and quantitively characterize their dynamic changes during the crowd movement. • We present a comprehensive emotion calculation method for physical strength consumption based on the James-Lange theory. Our new proposed model is used to derive the relationship between physical strength consumption and panic and examines how both of them govern the movement. The rest of this paper is organized as follows. Back- ground and related work are reviewed in Section 2. The definition of our proposed crowd simulation model is in- troduced in Section 3. Different benchmark scenarios and results are presented in Section 4. 2 RELATED WORK In this section, we provide a brief overview of prior work on crowd simulation. We divide the summaries based on whether the works involve physical, psychological, or phys- iological factors. JOURNAL OF LATEX CLASS FILES, VOL. X, NO. X, OCTOBER 2018 2.2 Crowd simulation with psychological factors The psychological state of an individual plays a vital role in his or her decision-making process [39], [40], [41]. Stress and panic are typical psychological factors and have a great influence on the movement of individuals in a crowd. In this subsection, we introduce representative works on them. In [14], authors focus on stress, which is defined as any change caused by interactions between the environment and individuals. Generally, stress is caused by a discrepancy be- tween environmental demands and the abilities of individ- uals. Stress can have positive effects on individual behavior. In emergency or evacuation situations, stress improves the performance of individuals [14]. It can be chronic and long- term [14]. However, stress and panic are inherently different. Panic is short-term and changeable [42] and usually leads to negative effects on individuals [43]. One of the most disastrous forms of collective human behavior is the kind of crowd stampede induced by panic, often leading to fatalities as people are crushed or trampled [7]. An individual's stress and panic are mirrored by others and they are disseminated within the crowd [6]. There are two separate lines of emotional contagion research: epidemiological-based and thermodynamics-based. The epidemiological SIR model [44] divides the individ- uals in a crowd into three categories: infected, susceptible, and recovered. At first this model is used to simulate the spread of rumor [45]. Then the epidemiological SIR model is used to describe emotion propagation. In [6], the epidemio- logical SIR model is combined with the OCEAN personality model [46]. The phenomenon of emotional contagion occurs more obviously in a panicked crowd. In [47], the cellular automata model is combined with the SIR model (CA-SIRS) to describe emotional contagion in an emergency situation. In [48], a qualitatively simulated approach is proposed to model emotional contagion process in a large-scale emer- gency evacuation siutation, which confirms that the effec- tiveness of rescue guidance is influenced by the leading emotion of the crowd. There is another kind of emotional contagion models based on thermodynamics [49]. Bosse et al. define emotional contagion within groups based on a multi-agent approach. They focuses mostly on emotions of groups rather than those of single individuals. Neto et al. [40] improve this model adapting it into BioCrowds and coping with emotional contagion within different groups of agents. Some researchers combine these two kinds of emotional contagion models to describe dynamic emotion propagation from the perspective of social psychology [50]. Because panic has a great influence on individual move- ment and often leads to serious consequences, we focus on panic in emergency situations. Inspired by the James-Lange theory in biological psychology, we improve the Durupinar model [6] by considering the influence of physical strength consumption on panic levels. In contrast to previous meth- ods considering only panic, we further demonstrate the relationship between physical strength consumption and panic. 2.3 Crowd simulation with physiological factors To complete a comprehensive analysis of crowd movement, we must consider not only psychological factors, but also 3 physiological factors of individuals as other important fac- tors in determining the crowd movement [11]. Physical strength is one of the most important physiolog- ical parameters that affects individual movement. Bruneau et al. [51] apply the principle of minimum energy (PME) on groups of different sizes and densities. In [10], [52], some physiological indicators (such as physical strength consumption and heart rate) are described. Furthermore, the relationship between physical strength consumption and heart rate is revealed, which is also a method for predicting physical strength consumption based on heart rate during moderate and vigorous exercise. Work in [11] shows that the relationship between physical strength consumption and speed is nonlinear. In [13], researchers investigate how the cumulative consumption of physical strength affects the evacuation time of individuals. Guy et al. [53] propose the principle of least effort (PLE) to compute the physical strength consumption required by various movements. Fur- thermore, Guy et al. [54] propose a less energy-consuming, conflict-free crowd movement method based on the criterion of minimal physical strength consumption [53]. These ap- proaches are focused on the relationship between physical strength and other physiological parameters (heart rate and oxygen uptake, for example) or individual movement. In [12], the authors choose four other basic physiological char- acteristics, including gender, age, health, and body shape, and map them to a navigation method. Inspired by prior approaches, we focus on physical strength consumption, which is a very important physio- logical factor. Physical strength consumption is central to research in human biology and biological anthropology [55] and is closely related to a variety of factors such as heart rate, oxygen consumption, etc. [52]. It directly affects the moving speed of an individual [13]. Other physiological factors (such as gender, age, health, and body shape) can influence movement through physical strength consump- tion. We analyze the relationship between physical strength consumption and panic. We also describe the effects of physical strength consumption on the physical movements of individuals. 3 OUR MODEL Our emotion-based crowd simulation model comprehen- sively considers physical strength consumption in emergen- cies to influence crowd movement. The flowchart of our model is presented in Figure 2. Strenuous movements are often observed in individuals in emergency or evacuation situations, and the relationship among them is more obvi- ous in such situations [56]. Therefore, we mainly focus on simulating crowd movements in such emergency situations. Our model consists of three important components: physical strength consumption, panic, and individual move- ment. Human physical strength consumption is computed with a physics-based method (Section 3.2). Panic levels are determined through an enhanced emotional contagion model that leverages the inherent relationship between physical strength consumption and panic (Section 3.3). Our model computes the movement of an individual by mod- eling the physical influence of strength consumption and panic (Section 3.4). JOURNAL OF LATEX CLASS FILES, VOL. X, NO. X, OCTOBER 2018 4 Fig. 2: The flowchart of our model. (a) Changes in the external environment can cause emotional fluctuations. For example, a hazard occurs and the red area represents the range of influence of the hazard. (b) The emotional changes of one individual are calculated according to the direct impact of the hazard and emotional contagion of his or her neighbors (Section 3.3). (c) The desired speed and direction of each individual are calculated based on an updated panic assessment (Section 3.4). (d) Using the limit of physical strength consumption, the actual speed is determined [11] (Section 3.4). (e) The calculation of physical strength consumption affected by the actual speed (Section 3.2). In contrast, the cumulative physical strength consumption also determines the actual maximal speed of the individual at the next timestep (Section 3.4). The current physical strength consumption reflects the emotional experience of an individual (Section 3.3). (f) The position of the individual is updated according to its actual speed. If the individual is panicked, we return to step (b); otherwise, the flowchart ends (Section 3.3). · di represents the distance of the individual at time t. F x i work done by the individual along the horizontal direction. F y is the pulling force of the individual along the vertical i is the rising direction. This force overcomes gravity. hi i · hi represents the height of the individual at time t, and F y work done by the individual along the vertical direction. is defined as According to the laws of physics, F x i follows: F x i = fi + i − vx (vx τ i−1)m 3.1 Symbols and notations For convenience, the important parameters and their de- scriptions used in our model are listed in Table 1. 3.2 Physical strength consumption calculation Physical strength consumption is one of the most commonly used physiological indicators and is closely related to indi- vidual movement. It is defined by the following equation: P (t) = Phor (t) + Pver (t) (1) where P (t) denotes the total physical strength consumption at time t and Phor (t), Pver (t) denote the physical strength consumption along the horizontal and the vertical direc- tions, respectively. They are defined as follows: t(cid:88) t(cid:88) i=1 Phor (t) = Pver (t) = · di F x i i · hi F y (2) (3) F x is the driving force of the individual along the horizontal i direction. This force overcomes friction. di is the moving i=1 A diagram of the physical strength consumption calculation is shown in Figure 3. The friction fi is defined in Equation 5, ki is defined in Equation 6, and ti is defined in Equation 7 according to [57], [58]. µ is the friction factor, which is related to the shoes and the ground. In our implementation, µ=0.58 is adopted, which is also recommended in [59]. vi is the current velocity magnitude, vmin is the minimal velocity magnitude, and vmax is the maximal velocity magnitude. fi = ti · µ · mg · ki (5) ki = 1.5 + 0.5 · ti = 0.6 − 0.2 · vi − vmin vmax − vmin vi − vmin vmax − vmin (4) (6) (7) JOURNAL OF LATEX CLASS FILES, VOL. X, NO. X, OCTOBER 2018 TABLE 1: The parameters used in our model. Notation Description P (t) Phor (t) Pver (t) F x F y E Eo Ep Eh o Ec o Vi (t) V s i (L, t) Physical strength consumption at time t Physical strength consumption along the horizontal direction at time t Physical strength consumption along the vertical direction at time t Driving force of individual along the horizontal direction Pulling force of individual along the vertical direc- tion Panic emotion Emotional cognitive component Emotional experience component The emotion affected by hazards Emotional contagion Moving direction of the individual i at time t Safety evacuation direction of the individual i at location L and at time t V round i vdesired i vactual i i i of the individual i con- of the individual i is lim- (t) Combined moving directions of individuals who are in the perceived range of the individual i at time t The desired speed vdesired siders only the emotion factor. The actual speed vactual ited by his own physical strength consumption. Maximum speed vp according to current physical strength consumption Maximum speed that the individual i can run Speed of the individual i in the normal case (emo- tion value is equal to zero) The radius of perceived range The number of individuals in a scene The radius of an individual vp vMAX i vNOR i PR Num Ra 5 Fig. 3: Schematic of the physical strength consumption cal- culation. vx i is the velocity component of an individual in the horizontal direction at time i, and the length of each time step is τ. The horizontal speed of the individual changes from vx i in time interval τ. i−1 to vx level of panic. The panic level E consists of two components. The first is the emotional cognitive component Eo, which relates to the hazard and encompasses emotional contagion. The second is the emotional experience component Ep, which is calculated using physical strength consumption and heart rate. Therefore, the final emotion value is defined as follows: E = w · Eo + (1 − w) · Ep (11) where w is a weighting parameter, and 0 < w < 1. 3.3.1 The emotional cognitive component In this section, we present the calculation method of Eo. Eo consists of three terms: effect from hazard Eh o , emotional contagion Ec o, and emotional attenuation Ed o . oEh Effect from hazard Eh oEh perceive a hazard, they may become panicked. Eh as follows [60]: o : When individuals are able to o is defined n−1(cid:88)  1√ 0 2π·rs Eh o (L, t) = Γs (L, t) s=0 − (L−Ls)2 2rs2 e if(cid:107) L − Ls(cid:107) < rs and t ∈ U otherwise (12) (13) where ki is the coefficient of the weight, ti is the time of the individual's foot touching the ground, ki ∝ vi, ti ∝−1 vi, fi ∝ ki, and fi ∝ ti. If one stands with both feet on a force plate, ti = ki = 1. The physical strength consumption in the horizontal direction is defined by: (cid:110)(cid:16) i )2 −(cid:0)vx i−1 (vx (cid:1)2(cid:17) t(cid:80) i=1 (cid:0)vx (cid:111) (cid:1) τ m + ti · µ · mg · ki i + vx i−1 2 · Phor (t) = 1 According to the laws of physics, F y i following equation: (8) is defined by the Γs (L, t) = F y i = mg + i − vy (vy τ i−1)m (9) i is the velocity component in the vertical direction where vy at time i. The physical strength consumption in the vertical direc- tion is defined by: 2 · Pver (t) = 1 (cid:110)(cid:16) i )2 −(cid:0)vy (vy i−1 (cid:1)2(cid:17) m +(cid:0)vy t(cid:80) i=1 (cid:111) (cid:1) mgτ i + vy i−1 (10) 3.3 Panic calculation considering physical strength consumption This section presents the calculation method for the panic level of an individual. E ∈ [0, 1] indicates the approximate where L is the location of an individual, Ls is the location of a hazard, rs is the radius of the influence range of the hazard, and U is the duration of the hazard. oEc oEc Effect from emotional contagion Ec o : There are two kinds of representative models of emotional contagion: the Neto model [40] and the Durupinar model [6]. They use fundamentally different mechanisms, but both can generate good results. However, the Neto model defines too many parameters for each pairwise interaction [61] and it is hard to compute these parameters automatically. Moreover, per- sonality is also a very important, long-term, stable psycho- logical factor and it is vital for simulating heterogeneous crowd behavior [6]. The Neto model simplifies the person- ality factor while the Durupinar model pays more attention JOURNAL OF LATEX CLASS FILES, VOL. X, NO. X, OCTOBER 2018 to that factor and is effective at capturing the differences between individuals. Personality is an important part of our model. We consider the effect of personality on panic. Ac- cording to the above analysis, the Durupinar model is more suitable. In the virtual scenario of Section 4.1, we implement a comparable experiment to verify our motivation. Next, we present the emotional contagion method in our model. During evacuation, individuals can be in one of two states: susceptible or infected. When the panic level of an individual exceeds a certain threshold T1, the individual will be infected. If the intensity of an individual's panic surpasses another threshold T2, then the individual can spread the panic to his or her neighbors. In a general case, T1 < T2. T1 and T2 are correlated with individual person- alities. Here we represent the personalities of individuals using the OCEAN personality model [6]. The personality of an individual is represented by a five-dimensional vector < O, C, E, A, N >. Each factor is randomly distributed with a Gaussian distribution N (0, 0.25) [6]. T1 ∝−1 N, T1 ∝ C [46]. T2 ∝−1 E [6], [46]. T1 and T2 are defined by the following: T1 = α · C − β · N + γ (14) where α = 0.1, β = 0.1, and γ = 0.15. T2 = δ − ξ · E (15) where δ = 0.35, and ξ = 0.1. These parameters are deter- mined according to the methods in [62], [63]. Within the perceived range, when a susceptible individ- ual i sees an expressive individual j (the panic value is higher than threshold T2), i gets exposed by receiving a random dose di from a specified probability distribution multiplied by the panic intensity of j. The dose values di are randomly distributed with a Gaussian distribution N (0.1, 0.01). We denote the panic value of individual j at time t(cid:48) as ej (t(cid:48)). The panic value of individual i due to emotional contagion is defined in Equation 16 [6]. t(cid:80) (cid:80) Ec i,o (L, t) = t(cid:48)=0 ∀jj∈Visibility(i)∧j is expressive di (t(cid:48)) ej (t(cid:48)) (16) oEd Effect from emotional attenuation Ed oEd o : Emotional at- tenuation is defined in Equation 17 [60]. o (L, t) = Eo (Lpre, t − 1) · η(t) Ed (17) where Ed o (L, t) is an emotion decay function and η(t) is the decay rate. η(t) is positively related to the individual personality factor N [6]. Inspired by [6], [13], it is defined as follows:  0 η (t) = eβ2 (t−t2)−eβ2(t−1−t2) where β2 = 0.1, η ∝ N, and α = 0.1. 1+eβ2 (t−t2) t < t1 + α · N t ≥ t1 (18) The change of the emotional cognitive component ∆Eo (L, t) is defined in Equation 19 [60]. The Eo is defined in Equation 20 [60]. ∆Eo (L, t) = Eh o (L, t) + Ec o (L, t) Eo (L, t) = Eo (Lpre, t − 1) + ∆Eo (L, t) o (L, t) − Ed (19) (20) 6 3.3.2 The emotional experience component In this section, we present the calculation method of Ep. Individual emotions undergo three stages: cognition, action, and experience. First, an event occurs, and the individual perceives the current scene (emotional cognitive stage). Sub- sequently, the individual acts in a way that corresponds with physiological changes (action stage). Finally, the individual has the emotional experience (experience stage) [5]. Under emergency situations, once a hazard occurs, the individuals around it immediately take different actions, requiring physical strength consumption. Physical strength consumption in one minute is chosen as the measure of physiological changes. The current heart rate is calculated using physical strength consumption. Then, the increment of the emotional experience value is calculated based on the heart rate increment. Thereafter, the current emotional ex- perience value Ep is obtained. The details of the calculation method are as follows. Equation 21 describes the relationship between physical strength consumption in a minute (KJ/min) and heart rate (beat/min) when individuals experience panic and attempt to escape from the hazard [52]. According to Equation 21, we can calculate the current heart rate (HR) based on physical strength consumption in a minute ((cid:52)P ). (cid:26)87.3306 + 1.5850∆P (t) − 0.3151weight − 0.3197age 45.6221 + 2.2361∆P (t) + 0.2824weight − 0.1655age gender = 1 gender = 0 (21) where gender=1 for males and 0 for females, age (year) ∈ [19,45], weight (kg) ∈ [47,116], ∆P (t) = P (t) − P (t − τ ), and τ = 60s . HR(t) = Furthermore, according to [64], heart rate (HR) and intensity of anxiety or fear (emotional experience) are pos- itively correlated. In [64], the heart rate per minute is recorded before and after an electric shock, and emotional experience is reported once per minute. (cid:52)Ep and (cid:52)HR are the increments of emotional experience and heart rate, respectively, compared with the values when individuals are not panicked. The (cid:52)HR is defined as follows: (cid:52)HR(t) = HR(t) − HR(0) (22) where HR(t) is the heart rate at time t and HR(0) is the heart rate when individuals are not panicked. relationship between (cid:52)HR and (cid:52)Ep. Using a linear curve fitting method, we can obtain the (cid:52) Ep(t) = 0.03669 · (cid:52)HR(t) − 0.0724 Ep is defined in Equation 24 and Ep (0) = 0. Ep (t) = Ep (t − 1) + ∆Ep (t) (23) (24) 3.4 Individual movement model Based on the results of physical strength consumption and panic, the movement of each individual can be determined accurately through moving direction and moving speed. 3.4.1 Moving direction When a hazard occurs, individuals who can perceive the hazard directly will be panicked and calculate their own safety evacuation directions V s (t) is the i (L, t) [60]. V round i JOURNAL OF LATEX CLASS FILES, VOL. X, NO. X, OCTOBER 2018 7 combined moving directions of individuals who are in the perceived range of the individual i. Γs (L, t) · → LsL if(cid:107) L − Ls(cid:107) < rs and t ∈ U n−1(cid:80) → s=0 V → V s i (L, t) = otherwise (cid:80) ∀jj∈Visibility(i)∧j is expressive Vj (t) (25) (26) V round i (t) = Finally, the moving direction Vi (t) of actual velocity of an individual who directly perceives the hazard is defined as follows: Vi (t) = E · V s i (L, t) + (1 − E) · V round i (t) (27) where E is the panic emotion value. The moving direction of an individual i is influenced by panic level, safety evacua- tion direction, and other neighboring panicked individuals. Individual i can perceive the hazard indirectly through the surrounding panicked individuals. The individual i moves in the direction of Vi (t), as shown in Equation 28. V old (t) is the moving direction of the individual i at the i last moment when he is not panicked. The more panicked the individual is, the more easily he moves with other neigh- boring panicked individuals. Nonetheless, if the individual i is not panicked, he or she still moves in his or her original direction. Vi (t) = (1 − E) · V old i (t) + E · V round i (t) (28) 3.4.2 Moving speed In a panic situation, the speed of an individual i is expressed by the following equation [7]: i = (1 − E) · vNOR i + E · vMAX vdesired i where vdesired is the speed considering only the emotion factor, and 0 ≤ E ≤ 1. The speed of an individual in the normal case (the panic value is equal to zero) is vNOR , and the maximal speed is vMAX . The more panicked an individual is, the faster his or her speed. (29) i i i However, an individual is limited by his or her own physical strength consumption. In some cases, the moving speed of an individual cannot reach the desired speed due to the maximum limit dictated by current physical strength consumption. The actual speed cannot exceed the maximal speed vp. = min(cid:0)vdesired , vp(cid:1) (30) vactual i i The dependence of the decay rate [13] and maximal speed on physical strength consumption is presented in Table 2. The actual speed can be calculated using Equation 31. = min(cid:0)(1 − E) · vNOR · ξ(cid:1) i + E · vMAX i , vMAX i (31) vactual i TABLE 2: Dependence of speed decay rate and maximal- limit speed on physical strength consumption. As the physi- cal strength consumption increases, the maximal-limit speed decreases. Physical strength consumption p (J) Decay rate ξ (%) Maximal-limit speed vp (m/s) 0.0000 -- 20154.0000 20154.0000 -- 40279.6713 40279.6713 -- 81121.0042 81121.0042 -- 166258.8920 166258.8920 -- 181569.6090 181569.6090 -- 196355.1760 100.0000 99.8500 89.4200 75.8000 69.8200 65.7200 vMAX i · 0.9985 · 0.8942 · 0.7580 · 0.6982 · 0.6572 vMAX i vMAX i vMAX i vMAX i vMAX i video begins with initial normal behavior and ends with se- quences of abnormal behavior. In addition, some real-world scenes (Figures 7, 9, and 10) are chosen from real emergency incidents to evaluate our model. The simulation results show that our proposed method can generate realistic group behaviors. It can also reliably predict the changes of physical strength consumption and panic levels of a crowd in an emergency. We also use our proposed model in different virtual scenarios, such as a subway station and a crosswalk. These scenes have dense crowds and the probability of hazard occurrence in these scenarios is high. We simulate the crowd movement in these scenarios after one hazard. We have implemented the proposed model using Visual C++ to simulate crowd movements in emergencies. The Unity3D game engine has been used to visualize our crowd simulation results. The computing platform corresponds a PC with a quadcore 2.50 GHz CPU,16 GB memory, and an Nvidia GeForce GTX 1080 Ti graphics card. The parameter values in different scenarios used in the simulation are listed in Table 3. The mass of each individual is set to 60kg on average and the radius to 0.3m (in Table 3) [66]. Each factor of the vector < O, C, E, A, N > is randomly distributed with a Gaussian distribution N (0, 0.25) [6]. In most scenes, the dose values di are randomly distributed with a Gaussian distribution N (0.1, 0.01) [6]. In the Neto model, = 0.5, δ = 0.5, η = 0.5, and β = 1 [40]. The parameter values are obtained by comparing the simulations with real-world videos. A combination of genetic and greedy strategies are used to sample plausible parameters for our model, maximizing the match of the simulation algorithm to real data [63]. 4.1 Comparisons To validate our approach, we compare the simulation results obtained by different methods with real-world crowd evac- uation videos. The trend in the simulation results obtained by our model is that they are more similar to real-world videos than results from other approaches. 4 EXPERIMENTS Our proposed algorithm is used to simulate crowd move- ment in various scenarios and we demonstrate the benefits of it in these different scenarios. We evaluate our model on the public UMN dataset [65] (Figures 4). The dataset comprises the videos of 11 different scenarios of an es- cape event in 3 different indoor and outdoor scenes. Each 4.1.1 Comparisons with scenarios from public UMN dataset Comparisons between real scenes (chosen from the public UMN dataset [65]) and the corresponding simulation results are presented in Figure 4. We take two different real-world scenarios as examples, and detailed results can be seen in the supplementary video. Our model is compared with two JOURNAL OF LATEX CLASS FILES, VOL. X, NO. X, OCTOBER 2018 TABLE 3: List of parameter values used in our simulations di vNOR vMAX N(0.1,0.01) N(0.1,0.01) - N(0.1,0.01) N(0.1,0.01) - N(0.1,0.01) 2 2 2 2 2 2 2 0.8 0.8 0.8 0.8 0.8 0.8 0.8 size of scene 230111 230111 230111 25.653.5 25.653.5 25.653.5 600600 N(0.4,0.01) 2/2.5 0.8/1.2 600600 N(0.1,0.01) [2,4.5] [0.8,1.2] 600600 - - 0.5 - - 0.5 - - - δ - - 0.5 - - 0.5 - - - η - - 0.5 - - 0.5 - - - - [2,4.5] [0.8,1.2] 600600 0.5 0.5 0.5 N(0.1,0.01) [2,4.5] [0.8,1.2] 600600 - - - - [2,4.5] [0.8,1.2] 600600 0.5 0.5 0.5 10 10 10 10 10 10 8 15 10 10 10 10 8 β - - 1 - - 1 - - - 1 - 1 Scenarios Grass Grass Grass Room Room Room phone explosion British Parliament building Virtual scenario Virtual scenario Virtual scenario Virtual Scenario Model Ours Durupinar Neto Ours Durupinar Neto Ours Num Ra T1 T2 PR 16 16 16 19 19 19 152 0.3 0.3 0.3 0.3 0.3 0.3 0.3 α 0.1 - - 0.1 - - 0.1 β 0.1 - - 0.1 - - 0.1 γ 0.15 - - C N N(0,0.25) N(0,0.25) - - - - δ 0.35 - - 0.15 N(0,0.25) N(0,0.25) 0.35 - - - - - - - - 0.15 N(0,0.25) N(0,0.25) 0.35 ξ 0.1 - - 0.1 - - 0.1 E N(0,0.25) - - N(0,0.25) - - N(0,0.25) Ours 37 0.3 0.1 0.1 0.15 N(0,0.25) N(0,0.25) 0.35 0.1 N(0,0.25) Ours Cube-Neto Durupinar Neto 300 300 300 300 0.3 0.3 0.3 0.3 0.1 0.1 0.15 N(0,0.25) N(0,0.25) 0.35 0.1 N(0,0.25) - - - - - - - - - - - - - - - - - - - - - - - - (a) (e) (b) (c) (d) (f) (g) (h) Fig. 4: Comparisons between real scenes and simulation results by different models: (a) and (e) are real-world videos, (b) and (f) are simulated by our model, (c) and (g) are simulated by the Durupinar model, (d) and (h) are simulated by the Neto model. Each row represents one scene. (a) The red ellipse is Individual No. 1 and the yellow one is Individual No. 2. Individual No. 1 gets closer to Individual No. 2. (b) As the speed is influenced by physical strength consumption, simulating the situation where Individual No. 1 gets closer to Individual No. 2 is easier using our model. In (g) and (h), the simulation trajectories of different individuals in the red circle by the Durupinar and Neto model are similar and individuals easily get together, which is different from the real-world video. other representative emotion models: the Durupinar model [6] and the Neto model [40]. We annotate the trajectories of all the individuals in the real-world video using the video annotation tool in [67] and assign initial movement states to these models. Therefore, we can predict the trajectories of these individuals and compare them with the actual ones. In the Grass scenario, Individual No. 1 moves faster than Individual No. 2, and Individual No. 1 moves closer to Individual No. 2 (Figure 4a). The simulation result obtained by our model in the Grass scenario is more realistic than those obtained by the Durupinar and Neto models because the speed is influenced by physical strength consumption in our model. If an individual has consumed more physical strength than other individuals, his moving speed decreases and other individuals move faster than he does. Thus, simulating the situation is easier when one individual gets closer to another individual. In the Room scenario, some individuals are marked with red circles in the simulation results obtained by the Durupinar and Neto models (Figures 4g and 4h). The mov- ing directions and moving speeds of these individuals are almost the same. The simulation result by our model con- JOURNAL OF LATEX CLASS FILES, VOL. X, NO. X, OCTOBER 2018 TABLE 4: Entropy metric and spatial distance for different simulation algorithms on scenarios of Grass and Room. A lower value implies higher similarity with respect to the real-world crowd videos. Scenario Model Entropy metric Spatial distance Grass Room Ours Durupinar Neto Ours Durupinar Neto 3.386800 3.429000 3.409700 5.393900 5.463200 5.493300 1.042990 1.105531 1.046776 1.481327 1.484492 1.527570 9 forms to the real-world video. This is because the emotion mechanism of our model changes the moving directions of individuals and drives them to move away from the hazard. Meanwhile, the physical strength consumption influences the individual's speed. We use the entropy metric [68] to evaluate the trajectories of different simulation algorithms on different scenarios. En- tropy metric is used to measure the similarity between real- world data and simulation results. A lower value of the en- tropy metric means a smaller error and better similarity with the real-world data. Its calculation method is described as follows. The real-world crowd state is denoted as (x1 . . . xt), which includes the positions of all the agents at different timesteps. (y1 . . . yt) is the corresponding calculation result of our model. M is the estimated error variance. (xk[j] − yk[j])(xk[j] − yk[j])T (32) t(cid:80) n(cid:80) t·n · M = 1 k=0 j=1 where t is the total number of timesteps and n is the number of agents in the scenario. The Entropy metric is given by: e(µ) = 1 2 n log((2πe)d det(M )) (33) where d is the dimension of the state of a single agent. In this paper, we mainly discuss the 2D locations of agents. So, d = 2 in this paper. For each scenario, a user study is performed. There are 39 participants (51.28% female, 66.67% in the age group of 20-30) in this study and participants are asked to compare the movement states in the original video clips with the movement states in the crowd simulation results (Figure 5). These similarity scores are computed from the user studies. A score of 1 indicates most dissimilar and a score of 5 indicates most similar movement. Higher values indicate greater similarity. We also calculate average spatial distance between the simulated results and the ground truth over all the timesteps and individuals. Tables 4 and Figure 5 show that the simulated moving trends of our model are closer to those in the real-world videos than the results of other models. A rational approach is to combine physical strength consumption and panic to determine the movement of each individual. 4.1.2 Comparisons with real-world emergency scenarios In this subsection, we compare the simulation results ob- tained by different methods with real-world videos, partic- ularly real emergency incidents. Fig. 5: Comparison of similarity scores for movement states (higher values indicate greater similarity). We compare the movement states in the original videos with those in crowd simulation results achieved by different algorithms. (a) (c) (e) (b) (d) (f) Fig. 6: Crowd simulation results generated by different models in the Square scenario at the 98th frame: (a) Real- world scenario, (b) Neto model, (c) Neto-PS model, (d) Du- rupinar model, (e) Durupinar-PS model, and (f) our model. The Durupinar and Neto models don't consider physical strength consumption. The mechanism of physical strength consumption in this paper is integrated into these two models, which are denoted as Durupinar-PS and Neto-PS. In the real-world Square scenario, we compare our simulation results with the Durupinar, Durupinar-PS, Neto, and Neto- PS models (Figure 6). We also compare our results in a real- world emergency scenario, which is related to terrorist at- tacks on Kenya's shopping mall (Figure 7). More details can be seen in the supplementary video. Figure 8 shows the av- erage speeds of all the individuals at different timesteps for JOURNAL OF LATEX CLASS FILES, VOL. X, NO. X, OCTOBER 2018 10 (a) (c) (e) (b) (d) (f) Fig. 7: Crowd simulation results by different models in the scenario of terrorist attacks on Kenya's shopping mall at the 214th frame: (a) Real-world scenario, (b) Neto model, (c) Neto-PS model, (d) Durupinar model, (e) Durupinar-PS model, and (f) our model. these different models. We find that the simulation results of the Durupinar-PS and Neto-PS are closer to the movements and behaviors in the real videos, than those of the original Durupinar and Neto models. These comparisons validate that our proposed mechanism of physical strength con- sumption (physical strength consumption calculation and the effect of physical strength consumption on emotion) can enhance the performance of existing crowd simulation models that are only based on emotion. Our model describes emotional changes in a comprehensive manner. Based on the James-Lange theory, we describe three stages individual emotions undergo (Section 3.3.2) and combine emotional contagion with the effect of physical strength consumption on panic. The comparisons with Durupinar-PS and Neto- PS models show that our model integrating emotion and physical strength consumption is better than other emotion- based crowd simulation models. One piece of real-world video including both crowd and vehicles is chosen to simulate by our method. In this real sce- nario, we compare our simulation result with the Durupinar, Durupinar-PS, Neto, and Neto-PS models. In this paper, we mainly focus on emotions of crowds, especially panic emo- tions in emergencies. In particular, the drivers can express their panic through vehicles. Moreover, in emergencies, the drivers may not follow the traffic rules. In emergency scenarios, some behaviors of vehicles, such as sudden ac- celeration, intuitively demonstrate the drivers' panic. These behaviors can also cause the surrounding pedestrians to be panicked and thus act indirectly as emotional contagions. In common cases, when pedestrians are walking in front of (a) (b) Fig. 8: The average speeds of all the individuals at differ- ent timesteps for these different models in (a) the Square scenario and (b) the scenario of terrorist attacks on Kenya's shopping mall. The average speeds of our model at different timesteps are closer to the real scene than those of other models. The average speeds of the Neto-PS and Durupinar- PS models at different timesteps are closer to the real-world scenarios, than those of the original Neto and Durupinar models. vehicles, vehicles will slow down, change direction, or stop to avoid pedestrians. Vehicles and surrounding pedestrians influence each other in such traffic. Based on above analysis, in our model we treat vehicles as one kind of special large- sized agents with full physical strength and high moving speed. The radius of the vehicles is set to 3. The compar- isons show that our simulation result conforms to the real- world video, and can enhance the performance of existing crowd simulations under the complex scenarios including both crowd and vehicles. More details can be seen in the supplementary video. Table 5 shows the values of entropy metric and spatial distance for the above three scenarios. Comparing with the Durupinar, Durupinar-PS, Neto, Neto-PS models, our model can generated more similar simulation results with real- world scenarios. We also take two real emergency incidents as exam- ples to verify our proposed crowd simulation method. Our crowd simulation results of the scene after the mobile phone explosion on the subway in the Shanghai Metro Line 8 are presented in Figures 10a and 10b. Crowd simulation by our JOURNAL OF LATEX CLASS FILES, VOL. X, NO. X, OCTOBER 2018 11 (a) (c) (e) (b) (d) (f) Fig. 9: Crowd simulation results by different models in the scenario including both crowd and vehicles at the 88th frame: (a) Real-world scenario, (b) Neto model, (c) Neto-PS model, (d) Durupinar model, (e) Durupinar-PS model, and (f) our model. TABLE 5: Entropy metric and spatial distance for different simulation algorithms on scenarios of Square, Crowd and vehicles, and Terrorist attacks on shopping mall. A lower value implies higher similarity with respect to the real- world data. Scenario Model Entropy metric Spatial distance Square Ours Durupinar Durupinar-PS Neto Neto-PS Ours Durupinar Crowd and vehicles Durupinar-PS Neto Neto-PS Ours Durupinar Durupinar-PS Neto Neto-PS Terrorist attacks on shopping mall 1.278777 3.292511 1.883470 3.415817 1.568437 3.740438 5.851273 5.785572 5.984731 5.768604 1.060493 5.782041 3.809553 5.783371 3.230872 0.210871 0.316492 0.238043 0.327066 0.220118 0.816685 2.402975 2.067541 2.346846 2.070267 0.225969 0.761416 0.405739 0.755286 0.382926 (a) (c) (b) (d) Fig. 10: Comparisons between real-world videos and sim- ulation results by our approach. (a) The mobile phone explosion incident on the subway in the Shanghai Metro Line 8; (c) the shooting incident at the British Parliament building on March 22, 2017; (b,d) our corresponding simu- lation results. At first, the individuals in the red circle aren't panicked. Because of emotional contagion, they are then influenced by the panicked crowd around them, become panicked, and run away from the hazards. TABLE 6: Entropy metric and spatial distance for different simulation algorithms on scenarios of Phone explosion and Shooting at British Parliament. Scenario Model Entropy metric Spatial distance Phone explosion Shooting at British Parliament Ours Durupinar Neto Ours Durupinar Neto 1.565650 4.826474 2.857793 0.842232 2.246331 1.626734 1.137298 1.523882 1.495380 0.749183 0.784255 0.771396 model of the shooting at the British Parliament building on March 22, 2017 is presented in Figures 10c and 10d. We show the spread of panic in both scenarios. The color of the cylinders represents the emotional intensity of the individuals. These two real-world videos have poor quality images. Even using manual methods, it is still difficult to track the positions of people in each frame accurately. Due to these objective limitations, we cannot measure the similarity between the simulated trajectories and the ground truth for these scenes in a direct way. Therefore, we use the dominant path and entropy metric to quantitively evaluate our simulation results. The dominant path is defined based on collectiveness of crowd movements and it can be treated as the movement trend of the crowd [69], [70]. We also calcu- late the spatial distance between the simulated trajectories and the ground truth for the scenarios. From Table 6 and Figure 11, we can see that both the overall moving trend and the process of emotional contagion are similar to what is found in the recorded real-world crowd video clips. 4.1.3 Comparisons in virtual scenarios In the virtual scenario, we compare our simulation results with those of the Durupinar [6], Neto [40], and Ours-Neto JOURNAL OF LATEX CLASS FILES, VOL. X, NO. X, OCTOBER 2018 12 in our simulation result are more variable than those shown in the Ours-Neto model. Therefore, the simulation result by our model is more suitable for emergency situations than other models. 4.2 Application of our model in various virtual scenar- ios Fig. 11: Comparison of similarity scores for movement states and the process of emotional contagion (higher values indicate greater similarity). A user study is performed and participants are asked to compare the movement states and processes of emotional contagion in the original videos with those in crowd simulation results achieved by different algorithms. (a) (c) (b) (d) Fig. 12: Crowd simulation results by different models in the virtual scenario at the 1000th frame: (a) Durupinar model, (b) Neto model, (c) Ours-Neto model, and (d) our model. models. The parameter values we used are described in Ta- ble 3. In Figure 12a (the simulation result by the Durupinar model), the speeds of individuals are variable and their locations are scattered. Because of different thresholds and personality mechanisms, the Durupinar model can simulate heterogeneous crowd behaviors. However, there are too many individuals who are not affected by the panicked crowd and this result is unreasonable. In Figure 12b, in- dividuals move much slower than the individuals in the simulation results by other models. The reason is that the emotion calculated by the Neto model is much smaller. Moreover, the individual movement is too regular, which is unsuitable for emergency situations. In Figure 12c (simula- tion result by our model with the same emotional contagion method as the Neto model) and Figure 12d (our simulation result), most of the individuals are affected by the hazard and run away from it. Because of physical strength con- sumption and personality factors, the speeds of individuals (a) (b) Fig. 13: Crowd simulation results at a subway station: (a) the higher level of the subway station, (b) the lower level of the subway station. After the hazard occurs, the emotional contagion in our model begins to work. Although the direct impact of the hazard is limited, the hazardous area grows through emotional contagion among individuals and the number of individuals who run away from the hazard increases. Our model can be applied in different virtual scenarios. Subway stations and crosswalks are crowded and the prob- ability of hazard occurrence in these scenarios is very high. We simulate a hazard occurring in these scenarios and three examples are shown. Figure 13(a) shows crowd simulation at the higher level of the subway station. Figure 13(b) shows crowd simulation at the lower level of the subway station. Figure 14 shows crowd simulation at a crosswalk. We show each step of the process: hazard occurring, individuals run- ning away from the hazard, emotional contagion spreading, and moving speed attenuating. More details can be seen in the supplementary video. Our simulation results provide information about decision-making to deal with emergency situations. The heat maps of panic in the virtual scene are presented in Figure 15. Although the direct impact of the hazard is limited, the panic area grows through the emotional con- tagion mechanism in our model. When individuals are far from the hazard, panic attenuates. As accidents may happen JOURNAL OF LATEX CLASS FILES, VOL. X, NO. X, OCTOBER 2018 13 panicked than those in the Durupinar model simulation re- sults. The intensity of the panic calculated by the Durupinar model is lower than that calculated by our model. The reason is that our model considers not only emotional con- tagion among individuals, but also the impact of physical strength consumption on panic levels. Our model represents a comprehensive description of individual panic levels and is more conducive to the spread of panic than the Durupinar model. Therefore, the simulation results by our model are more reasonable for emergency situations. 5 CONCLUSION AND LIMITATIONS In contrast to traditional emotion-based crowd simulation models, we integrate physical strength consumption into our model. We not only present a panic level calculation, but also delineate the effect of physical strength consumption on panic. Finally, both physical strength consumption and panic determine the movement of each individual. Our pro- posed model is verified by simulations, and it is compared with real-world videos and previous approaches. Results have shown that our proposed model can reliably generate realistic group behaviors. It can also predict the changes of physical strength consumption and panic of a crowd in an emergency situation. However, our model has several limitations. Although our model can generate realistic crowd movements, the panic levels and physical strength consumption of the crowd in an emergency scene cannot be obtained directly. Our model can only infer them during the simulation. Thus, the initial state of our model is difficult to determine and it is usually time consuming to do so. In the future, we plan to use the latest wearable equipment to collect these data and provide a new method that can quickly and accurately determine the initial state. Furthermore, at present, our model mainly focuses on emergency scenarios. In future work, we want to extend our model to a variety of general situations. We also plan to develop a coupled framework for crowd and traffic flow to generate realistic mixed traffic simulations and model the influence of these two flows on each other. REFERENCES [1] Y. Ma, T. Lin, Z. Cao, C. Li, F. Wang, and W. Chen, "Mobility viewer: An eulerian approach for studying urban crowd flow," IEEE Transactions on Intelligent Transportation Systems, vol. 17, no. 9, pp. 2627 -- 2636, 2016. ´A. G´omez, L. L´opez-Rodr´ıguez, H. Sheikh, J. Ginges, L. Wilson, H. Waziri, A. V´azquez, R. Davis, and S. Atran, "The devoted ac- tor's will to fight and the spiritual dimension of human conflict," Nature Human Behaviour, vol. 1, pp. 673 -- 679, 2017. [2] [4] [3] M. Xu, Y. Wu, P. Lv, H. Jiang, M. Luo, and Y. Ye, "miSFM: On combination of mutual information and social force model towards simulating crowd evacuation," Neurocomputing, vol. 168, pp. 529 -- 537, 2015. J. Chun, H. Lee, Y. S. Park, W. Park, J. Park, S. H. Han, S. Choi, and G. J. Kim, "Real-time classification of fear/panic emotion based on physiological signals." in Proc. The Eighth Pan-Pacific Conference on Occupational Ergonomics, 2007, pp. 1 -- 9. J. W. Kalat, Biological Psychology (10th ed.). Cengage Learning, 2007. [6] F. Durupınar, U. G ud ukbay, A. Aman, and N. I. Badler, "Psy- chological parameters for crowd simulation: From audiences to mobs," IEEE Transactions on Visualization and Computer Graphics, vol. 22, no. 9, pp. 2145 -- 2159, 2016. [5] Fig. 14: Crowd simulation result at a crosswalk. At the lower left corner, a car explodes. Then individuals run away from the hazard. randomly in public places, we can take preventive action in advance and reduce loss by accurately predicting the panic area. (a) (c) (b) (d) Fig. 15: Panic heat maps of the virtual scene: (a) heat map at the 13th frame, (b) heat map at the 29th frame, (c) heat map at the 57th frame, (d) heat map at the 120th frame. The red area is more panicked than the green area in the heat map. The deeper the color, the more panicked the area. (a) (b) Fig. 16: The heat maps of panic at the 185th frame of the crosswalk scenario: (a) heat map of the crowd simulation generated by our model and (b) heat map of the crowd sim- ulation generated using the Durupinar model. The red area is more panicked than the green area in the heat map. The deeper the color, the more panicked the area. We highlight the same area of the two simulation results. The individuals in our model simulation result are more panicked than the individuals in the Durupinar model simulation results. The panic heat maps generated from crowd simulations in the crosswalk scenario by our model and those generated using the Durupinar model are presented in Figure 16. The individuals in the simulation results by our model are more JOURNAL OF LATEX CLASS FILES, VOL. X, NO. X, OCTOBER 2018 14 [7] D. Helbing, I. Farkas, and T. Vicsek, "Simulating dynamical fea- tures of escape panic," Nature, vol. 407, no. 6803, pp. 487 -- 490, 2000. [8] Y. Li, M. Christie, O. Siret, and R. Kulpa, "Cloning crowd mo- tions," in Proc. ACM SIGGRAPH/Eurographics Symp. Comput. Ani- mation, 2012, pp. 201 -- 210. J. M. Frederic P. Miller, Agnes F. Vandome, Physical Strength. Alphascript Publishing, 2011. [9] [10] W. R. Leonard, "Measuring human energy expenditure: What have we learned from the flex-heart rate method?" American Journal of Human Biology, vol. 15, no. 4, pp. 479 -- 489, 2003. [11] D. M. Buchner, E. B. Larson, E. H. Wagner, T. D. Koepsell, and B. J. De Lateur, "Evidence for a non-linear relationship between leg strength and gait speed," Age and Ageing, vol. 25, no. 5, pp. 386 -- 391, 1996. [12] L. Zheng, D. Qin, Y. Cheng, L. Wang, and L. Li, "Simulating heterogeneous crowds from a physiological perspective," Neuro- computing, vol. 172, no. C, pp. 180 -- 188, 2012. [13] J. Koo, B. I. Kim, and Y. S. Kim, "Estimating the effects of mental disorientation and physical fatigue in a semi-panic evacuation," Expert Systems with Applications, vol. 41, no. 5, pp. 2379 -- 2390, 2014. [14] S. Kim, S. J. Guy, D. Manocha, and M. C. Lin, "Interactive simulation of dynamic crowd behaviors using general adaptation syndrome theory," in Proc. Symp. Interactive 3D Graphics and Games, 2012, pp. 55 -- 62. [15] S. Kim, A. Bera, A. Best, R. Chabra, and D. Manocha, "Interactive and adaptive data-driven crowd simulation," in Proc. IEEE Virtual Reality Conf., Mar. 2016, pp. 29 -- 38. [16] H. L. Kang, M. G. Choi, Q. Hong, and J. Lee, "Group behavior from video:a data-driven approach to crowd simulation," in Proc. ACM SIGGRAPH/Eurographics Symp. Comput. Animation, 2007, pp. 109 -- 118. [17] I. Karamouzas, N. Sohre, R. Narain, and S. J. Guy, "Implicit crowds: optimization integrator for robust crowd simulation," ACM Transactions on Graphics, vol. 36, no. 4, pp. 1 -- 13, 2017. [18] S. Yi, H. Li, and X. Wang, "Pedestrian behavior modeling from stationary crowds with applications to intelligent surveillance," IEEE Transactions on Image Processing, vol. 25, no. 9, pp. 4354 -- 4368, 2016. [19] T. Feng, L. F. Yu, S. K. Yeung, K. K. Yin, and K. Zhou, "Crowd- driven mid-scale layout design," ACM Transactions on Graphics, vol. 35, no. 4, pp. 132 -- 145, 2016. [20] M. S. Kaiser, K. T. Lwin, M. Mahmud, D. Hajializadeh, T. Chaip- imonplin, A. Sarhan, and M. A. Hossain, "Advances in crowd analysis for urban applications through urban event detection," IEEE Transactions on Intelligent Transportation Systems, vol. PP, no. 99, pp. 1 -- 21, 2017. [21] N. M. Joseph, N. M. Joseph, N. M. Joseph, N. M. Joseph, N. M. Joseph, and N. M. Joseph, "Zootopia crowd pipeline," in Proc. ACM SIGGRAPH, 2016, pp. 59:1 -- 59:2. [22] M. Moussaid, D. Helbing, and G. Theraulaz, "How simple rules determine pedestrian behavior and crowd disasters," Proceedings of the National Academy of Sciences of the United States of America, vol. 108, pp. 6884 -- 8, 04 2011. [23] M. Zhou, H. Dong, F. Y. Wang, Q. Wang, and X. Yang, "Modeling and simulation of pedestrian dynamical behavior based on a fuzzy logic approach," Information Sciences, vol. 360, pp. 112 -- 130, 2016. [24] M. Zhou, H. Dong, Y. Zhao, P. A. Ioannou, and F. Y. Wang, "Opti- mization of crowd evacuation with leaders in urban rail transit stations," IEEE Transactions on Intelligent Transportation Systems, vol. PP, no. 99, pp. 1 -- 12, 2019. [25] V. J. Cassol, E. S. Testa, C. R. Jung, M. Usman, P. Faloutsos, G. Berseth, M. Kapadia, N. I. Badler, and S. R. Musse, "Evalu- ating and optimizing evacuation plans for crowd egress," IEEE Computer Graphics & Applications, vol. 37, no. 4, pp. 60 -- 71, 2017. [26] M. Zucker, J. Kuffner, and M. Branicky, "Multipartite RRTs for rapid replanning in dynamic environments," in Proc. IEEE Conf. Robotics and Automation, 2007, pp. 1603 -- 1609. [27] B. Kluge and E. Prassler, "Reflective navigation: Individual be- haviors and group behaviors," in Proc. IEEE Conf. Robotics and Automation, 2004, pp. 4172 -- 4177. [28] H. Wang, J. Ondrej, and C. OSullivan, "Trending paths: A new semantic-level metric for comparing simulated and real crowd data," IEEE Transactions on Visualization and Computer Graphics, vol. 23, no. 5, pp. 1454 -- 1464, 2017. [29] M. Zhou, H. Dong, D. Wen, X. Yao, and X. Sun, "Modeling of crowd evacuation with assailants via a fuzzy logic approach," IEEE Transactions on Intelligent Transportation Systems, vol. 17, no. 9, pp. 2395 -- 2407, 2016. [30] T. Weiss, A. Litteneker, C. Jiang, and D. Terzopoulos, "Position- based multi-agent dynamics for real-time crowd simulation," in Proc. ACM SIGGRAPH/Eurographics Symp. Comput. Animation, 2017, pp. 27:1 -- 27:2. [31] L. He, J. Pan, S. Narang, and D. Manocha, "Dynamic group behaviors for interactive crowd simulation," in Proc. ACM SIG- GRAPH/Eurographics Symp. Comput. Animation, 2016, pp. 139 -- 147. [32] M. Xu, Y. Wu, Y. Ye, I. Farkas, H. Jiang, and Z. Deng, "Collective crowd formation transform with mutual informationcbased run- time feedback," Computer Graphics Forum, vol. 34, no. 1, pp. 60 -- 73, 2015. [33] F. Belkhouche, "Reactive path planning in a dynamic environ- ment," IEEE Transactions on Robotics, vol. 25, no. 4, pp. 902 -- 911, 2009. [34] S. Patil, J. Van Den Berg, S. Curtis, M. C. Lin, and D. Manocha, "Directing crowd simulations using navigation fields," IEEE Trans- actions on Visualization and Computer Graphics, vol. 17, no. 2, pp. 244 -- 254, 2011. [35] R. Kulpa, A. Olivierxs, J. Ondrej, and J. Pettr´e, "Imperceptible relaxation of collision avoidance constraints in virtual crowds," ACM Transactions on Graphics, vol. 30, no. 6, pp. 138:1 -- 138:10, 2011. J. Guy, M. C. Lin, and D. Manocha, "Interactive navigation of heterogeneous agents us- ing adaptive roadmaps," IEEE Transactions on Visualization and Computer Graphics, vol. 15, no. 1, pp. 34 -- 48, 2009. [36] R. Gayle, A. Sud, E. Andersen, S. [37] A. Sud, E. Andersen, S. Curtis, M. C. Lin, and D. Manocha, "Real-time path planning in dynamic virtual environments using multiagent navigation graphs," IEEE Transactions on Visualization and Computer Graphics, vol. 14, no. 3, pp. 526 -- 538, 2008. [38] 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. [39] M. Xu, Z. Pan, M. Zhang, P. Lv, P. Zhu, Y. Ye, and W. Song, "Character behavior planning and visual simulation in virtual 3d space," IEEE Multimedia, vol. 20, no. 1, pp. 49 -- 59, 2013. [40] A. B. F. Neto, C. Pelachaud, and S. R. Musse, "Giving emotional contagion ability to virtual agents in crowds," in Proc. Intelligent Virtual Agents, 2017, pp. 63 -- 72. [41] H. Jiang, Z. Deng, M. Xu, X. He, T. Mao, and Z. Wang, "An emotion evolution based model for collective behavior simulation," in Proc. Symp. Interactive 3D Graphics and Games, 2018, pp. 10:1 -- 10:6. [42] A. Ortony, G. L. Clore, and A. Collins, "The cognitive structure of emotions." Contemporary Sociology, vol. 18, no. 6, pp. 2147 -- 2153, 1988. [43] L. Berkowitz, "On the formation and regulation of anger and aggression: A cognitive-neoassociationistic analysis." American Psychologist, vol. 45, no. 4, p. 494, 1990. [44] W. O. Kermack and A. G. McKendrick, "Contributions to the mathematical theory of epidemics. III. further studies of the prob- lem of endemicity," Proc. the Royal Society of London. Series A, vol. 141, no. 843, pp. 94 -- 122, 1933. [45] L. Zhao, H. Cui, X. Qiu, X. Wang, and J. Wang, "SIR rumor spreading model in the new media age," Physica A: Statistical Mechanics and its Applications, vol. 392, no. 4, pp. 995 -- 1003, 2013. [46] F. Durupinar, N. Pelechano, J. Allbeck, U. Gudukbay, and N. I. Badler, "How the ocean personality model affects the perception of crowds," IEEE Computer Graphics and Applications, vol. 31, no. 3, pp. 22 -- 31, 2011. [47] L. Fu, W. Song, W. Lv, and S. Lo, "Simulation of emotional conta- gion using modified SIR model: a cellular automaton approach," Physica A: Statistical Mechanics and its Applications, vol. 405, pp. 380 -- 391, 2014. [48] J. H. Wang, S. M. Lo, J. H. Sun, Q. S. Wang, and H. L. Mu, "Qual- itative simulation of the panic spread in large-scale evacuation," Simulation, vol. 88, no. 12, pp. 1465 -- 1474, 2012. [49] T. Bosse, R. Duell, Z. A. Memon, J. Treur, and C. N. V. D. Wal, "A multi-agent model for mutual absorption of emotions," in Proc. 23th EuropeanConf. Mod. and Sim., 2009. [50] J. Tsai, E. Bowring, S. Marsella, and M. Tambe, "Empirical eval- uation of computational fear contagion models in crowd disper- sions," Autonomous Agents and Multi-agent Systems, vol. 27, no. 2, pp. 200 -- 217, 2013. [51] J. Bruneau, A.-H. Olivier, and J. Pettre, "Going through, going around: A study on individual avoidance of groups," IEEE Trans- JOURNAL OF LATEX CLASS FILES, VOL. X, NO. X, OCTOBER 2018 15 actions on Visualization and Computer Graphics, vol. 21, no. 4, pp. 520 -- 528, 2015. [52] L. Keytel, J. Goedecke, T. Noakes, H. Hiiloskorpi, R. Laukkanen, L. van der Merwe, and E. Lambert, "Prediction of energy expen- diture from heart rate monitoring during submaximal exercise," Journal of Sports Sciences, vol. 23, no. 3, pp. 289 -- 297, 2005. [53] S. J. Guy, J. Chhugani, S. Curtis, P. Dubey, M. Lin, and D. Manocha, "PLEdestrians: a least-effort approach to crowd simulation," in Proc. ACM SIGGRAPH/Eurographics Symp. Comput. Animation, 2010, pp. 119 -- 128. [54] S. J. Guy, J. Chhugani, C. Kim, N. Satish, M. Lin, D. Manocha, and P. Dubey, "Clearpath: highly parallel collision avoidance for multi- agent simulation," in Proc. ACM SIGGRAPH/Eurographics Symp. Comput. Animation, 2009, pp. 177 -- 187. [55] S. J. Ulijaszek, "Human energetics in biological anthropology," American Journal of Human Biology, vol. 8, no. 5, pp. 682 -- 683, 2010. [56] J. E. Almeida, R. J. Rossetti, F. Aguiar, and E. Oliveira, "Chapter 8 - crowd simulation applied to emergency and evacuation scenar- ios," in Advances in Artificial Transportation Systems and Simulation, R. J. Rossetti and R. Liu, Eds. Boston: Academic Press, 2015, pp. 149 -- 161. [57] J. Nilsson and A. Thorstensson, "Ground reaction forces at differ- ent speeds of human walking and running," Acta Physiologica, vol. 136, no. 2, pp. 217 -- 227, 1989. [58] R. Cross, "Standing, walking, running, and jumping on a force plate," American Journal of Physics, vol. 67, no. 4, pp. 304 -- 309, 1999. [59] J. M. Burnfield, Y. J. Tsai, and C. M. Powers, "Comparison of utilized coefficient of friction during different walking tasks in persons with and without a disability," Gait & Posture, vol. 22, no. 1, pp. 82 -- 88, 2005. [60] M. Xu, X. Xie, P. Lv, J. Niu, H. Wang, C. Li, R. Zhu, Z. Deng, and B. Zhou, "Crowd behavior simulation with emotional contagion in unexpected multi-hazard situations," IEEE Transactions on Systems, Man, and Cybernetics: Systems, pp. 1 -- 15, 2019. [61] J. Tsai, E. Bowring, S. Marsella, and M. Tambe, "Empirical eval- uation of computational emotional contagion models," in Proc. Intelligent Virtual Agents, 2011, pp. 384 -- 397. [62] A. E. Basak, U. Gudukbay, and F. Durupinar, "Using real life incidents for realistic virtual crowds with data-driven emotion contagion," Computers & Graphics, vol. 72, 2018. [63] D. Wolinski, S. J. Guy, A. H. Olivier, M. Lin, and D. Manocha, "Parameter estimation and comparative evaluation of crowd sim- ulations," Computer Graphics Forum, vol. 33, no. 2, pp. 303 -- 312, 2014. [64] H. M. Petry and O. Desiderato, "Changes in heart rate, muscle ac- tivity, and anxiety level following shock threat," Psychophysiology, vol. 15, no. 5, pp. 398 -- 402, 1978. [65] R. Mehran, A. Oyama, and M. Shah, "Abnormal crowd behavior detection using social force model," in Proc. IEEE Conf. Comput. Vis. Pattern Recognit, 2009, pp. 935 -- 942. [66] M. Zhou, H. Dong, Y. Zhao, P. A. Ioannou, and F. Y. Wang, "Opti- mization of crowd evacuation with leaders in urban rail transit stations," IEEE Transactions on Intelligent Transportation Systems, vol. PP, no. 99, pp. 1 -- 12, 2019. [67] C. Vondrick, D. Patterson, and D. Ramanan, "Efficiently scaling up crowdsourced video annotation," Int. J. Comput. Vision, vol. 101, no. 1, pp. 184 -- 204, 2013. [68] S. J. Guy, J. V. D. Berg, W. Liu, R. Lau, M. C. Lin, and D. Manocha, "A statistical similarity measure for aggregate crowd dynamics," ACM Transactions on Graphics, vol. 31, no. 6, pp. 1 -- 11, 2012. [69] M. Xu, C. Li, P. Lv, N. Lin, R. Hou, and B. Zhou, "An efficient method of crowd aggregation computation in public areas," IEEE Transactions on Circuits & Systems for Video Technology, vol. PP, no. 99, pp. 1 -- 12, 2017. [70] C. Li, P. Lv, D. Manocha, H. Wang, Y. Li, B. Zhou, and M. Xu, "Acsee: Antagonistic crowd simulation model with emotional contagion and evolutionary game theory," IEEE Transactions on Affective Computing, pp. 1 -- 1, 2019. Mingliang Xu received the Ph.D. degree in com- puter science and technology from the State Key Laboratory of CAD&CG, Zhejiang Univer- sity, Hangzhou, China. He is a Full Professor with the School of Information Engineering, Zhengzhou University, Zhengzhou, China, where he is currently the Director of the Center for Interdisciplinary Infor- mation Science Research. He was with the De- partment of Information Science, National Natu- ral Science Foundation of China, Beijing, China, from 2015 to 2016. His current research interests include computer graphics, multimedia, and artificial intelligence. He has authored over 60 journal and conference papers in the above areas, including the ACM Transactions on Graphics, ACM Transactions on Intelligent Systems and Technology, IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE, IEEE TRANSACTIONS ON IMAGE PRO- CESSING, IEEE TRANSACTIONS ON CYBERNETICS, IEEE TRANS- ACTIONS ON CIRCUITS AND SYSTEMS FOR VIDEO TECHNOLOGY, ACM SIGGRAPH (Asia), ACM MM, and ICCV. Dr. Xu is the Vice General Secretary of ACM SIGAI China. Chaochao Li received the B.S. degree in com- puter science and technology and the mas- ter's degree in computer application technol- ogy from the School of Information Engineer- ing, Zhengzhou University, Zhengzhou, China, where he is currently pursuing the Ph.D. degree. His current research interests include computer graphics and computer vision. Pei Lv received the Ph.D. degree from the State Key Laboratory of CAD&CG, Zhejiang Univer- sity, Hangzhou, China, in 2013. He is currently an Associate Professor with the School of Information Engineering, Zhengzhou University, Zhengzhou, China. His current re- search interests include video analysis and crowd simulation. He has authored over 20 jour- nal and conference papers in the above ar- eas, including the IEEE TRANSACTIONS ON IMAGE PROCESSING, IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS FOR VIDEO TECHNOLOGY, and ACM Multimedia. Wei Chen is a professor at the State Key Lab of CAD&CG, Zhejiang University. His research interests are visualization and visual analysis, and he has published more than 30 IEEE/ACM Transactions and IEEE VIS papers. He ac- tively served as guest or associate editors of IEEE Transactions on Visualization and Com- puter Graphics, IEEE Transactions on Intelligent Transportation Systems, and Journal of Visu- alization. For more information, please refer to http://www.cad.zju.edu.cn/home/chenwei/. 16 JOURNAL OF LATEX CLASS FILES, VOL. X, NO. X, OCTOBER 2018 Zhigang Deng is a Full Professor of Computer Science at University of Houston (UH). He is also the Director of Graduate Studies at UH's Com- puter Science Department and the Founding Di- rector of the UH Computer Graphics and Inter- active Media (CGIM) Lab. He earned a Ph.D. in Computer Science at the Integrated Media Sys- tem Center (NSF ERC) and the Department of Computer Science at the University of Southern California in 2006. He completed a B.S. degree in Mathematics from Xiamen University (China), and an M.S. in Computer Science from Peking University (China).His interests are in the broad areas of computer graphics, computer ani- mation, human computer interaction, virtual human modeling and ani- mation, and visual computing for biomedical/healthcare informatics. He is the recipient of ACM ICMI Ten Year Technical Impact Award, UH Teaching Excellence Award, Google Faculty Research Award, UHCS Faculty Academic Excellence Award, and NSFC Overseas and Hong Kong/Macau Young Scholars Collaborative Research Award. Besides being the CASA 2014 conference general co-chair and the SCA 2015 conference general co-chair, he currently serves as an Associate Editor of several journals including Computer Graphics Forum and Computer Animation and Virtual Worlds Journal. Bing Zhou received B.S. and M.S. degrees from Xi'an Jiaotong University, Xi'an, China, in 1986 and 1989, respectively, and a Ph.D. from Bei- hang University, Beijing, China, in 2003, all in computer science. He is currently a Professor with the School of Information Engineering, Zhengzhou University, Zhengzhou, China. His current research inter- ests include video processing and understand- ing, surveillance, computer vision, and multime- dia applications. Dinesh Manocha is the Paul Chrisman Iribe Chair in Computer Science & Electrical and Computer Engineering at the University of Mary- land College Park. He is also the Phi Delta Theta/Matthew Mason Distinguished Professor Emeritus of Computer Science at the Univer- sity of North Carolina - Chapel Hill. He has won many awards, including Alfred P. Sloan Re- search Fellow, the NSF Career Award, the ONR Young Investigator Award, and the Hettleman Prize for scholarly achievement. His research in- terests include multi-agent simulation, virtual environments, physically- based modeling, and robotics. His group has developed a number of packages for multi-agent simulation, crowd simulation, and physics- based simulation that have been used by hundreds of thousands of users and licensed to more than 60 commercial vendors. He has published more than 500 papers and supervised more than 35 PhD dissertations. He is an inventor of 9 patents, several of which have been licensed to industry. His work has been covered by the New York Times, NPR, Boston Globe, Washington Post, ZDNet, as well as DARPA Legacy Press Release. He is a Fellow of AAAI, AAAS, ACM, and IEEE and also received the Distinguished Alumni Award from IIT Delhi. He was a co-founder of Impulsonic, a developer of physics-based audio simulation technologies, which was acquired by Valve Inc. See http://www.cs.umd.edu/∼dm
0902.4098	1	0902	2009-02-24T09:35:26	Coordination in multiagent systems and Laplacian spectra of digraphs	[ "cs.MA", "cs.DM", "math.CO", "math.OC" ]	Constructing and studying distributed control systems requires the analysis of the Laplacian spectra and the forest structure of directed graphs. In this paper, we present some basic results of this analysis partially obtained by the present authors. We also discuss the application of these results to decentralized control and touch upon some problems of spectral graph theory.	cs.MA	cs	ISSN 0005-1179, Automation and Remote Control, 2009, Vol. 70, No. 3, p. 128. c(cid:13) Pleiades Publishing, Ltd., 2009. Original Russian Text c(cid:13) P.Yu. Chebotarev and R.P. Agaev, 2009, published in Avtomatika i Telemekhanika, 2009, No. 3, pp. 136 -- 151. DETERMINATE SYSTEMS Coordination in multiagent systems and Laplacian spectra of digraphs P. Yu. Chebotarev and R. P. Agaev Trapeznikov Institute of Control Sciences, Moscow, Russia Received July 21, 2008 Abstract -- Constructing and studying distributed control systems requires the analysis of the Laplacian spectra and the forest structure of directed graphs. In this paper, we present some basic results of this analysis partially obtained by the present authors. We also discuss the application of these results published earlier to decentralized control and touch upon some problems of spectral graph theory. PACS numbers: 87.19.lr, 02.10.Yn 1. INTRODUCTION In the cooperative control of distributed multiagent systems, the generation of control actions is decentralized. The actions result from negotiations between agents. As a metaphor, we can remember [1] that the musicians of "Persimfans"1, which existed in Moscow between 1922 and 1932, performed extremely complex musical compositions without a conductor. The string section formed a full circle (partly with their backs to the audience), while the wind section was situated inside of that circle. Every musician not only heard, but was also able to see the others. This way the magic chemistry among the performers, harmony and dynamic coordination between all the participants brought such a synchronization, that it became the substitute of the role of the conductor. Amazingly, one of the Persimfans' distinguishing features was their ability, according to the testimony of the extremely demanding and tough critics, to maintain a particularly subtle and profoundly individual approach to the interpretation of musical pieces, normally unthinkable without the help of the conductor. The theory of decentralized control has a long history. In addition to the works discussed in [1], one can remember the theory of statistical consensus by DeGroot [2], methods of step- by-step coordination of expert judgements, such as Delphi methodology developed at the RAND Corporation in the 1950s [3], works on distributed networked computations and distributed decision- making [4, 5] and, of course, modeling the collective behavior of animals (see, e. g., [6]). Starting from 2003 (approximately) we can observe an avalanche of publications on decentralized control connected with the application of spectral graph theory in this field. In this paper, we discuss continuous and discrete models of distributed coordination which can be considered as basic models of decentralized control. It is shown how the recent advantages in the algebraic graph theory (including some results published by the present authors before the "boom" imploded in 2003) can be applied in this area. 1 Abbreviation for Perviy Simfonicheskiy Ansambl' (First Symphony Ensemble). 128 MULTIAGENT SYSTEMS AND LAPLACIAN SPECTRA 2. A CONTINUOUS MODEL OF DISTRIBUTED CONSENSUS Consider the basic continuous distributed consensus algorithm: xi(t) = − nXj=1 aij(t) (xi(t) − xj(t)) , i = 1, . . . , n. 129 (1) Here, n is the number of agents, xi(t) is an information state (characteristic, parameter, etc.) of the ith agent, aij(t) ≥ 0 is the weight with which agent i takes into account the discrepancy in the information state with agent j. The information states can be, among others, positions (if the agents need to rendezvous in space), velocities (if they execute decentralized formation maneuvers), arrival times (if these must be synchronized), and so on. Decentralized control usually requires solving more complex problems than simply reaching a consensus. For example, if moving in formation is considered, then a typical task is moving along a prescribed course and in a prescribed and fixed configuration. During a violent maneuver the configuration can be altered, but after the maneuver it must be restored. Alteration and restoration of a prescribed geometric shape are also typical when a "flock" of moving physical objects encounters an obstacle or a hazard. It should be noted that reaching consensus is an important element of control strategy in all such cases. Actually, it is a key element, because it usually determines the stability of the system, its controllability, etc. That is why the analysis and synthesis of consensus algorithms, such as (1), is a necessary component of solving various problems of decentralized control. In this paper we focus on graph theoretic results underlying the analysis of distributed consensus algorithms. Let Γ(t) be the communication digraph associated with the consensus model (1). The vertices of Γ(t) are identified with the agents, and Γ(t) has an arc from vertex j to vertex i (denoted by (j, i) or j → i) if and only if aij(t) 6= 0. The presence of this arc in Γ(t) means that agent i coordinates its information state with that of j. The weight of the (j, i) arc is aij(t). In matrix form, model (1) can be written as where x(t) = (x1(t), . . . , xn(t))T and the matrix L(t) = [ℓij(t)]n×n is defined as follows: x(t) = −L(t)x(t), (2) (3) ℓij(t) = −aij(t), j 6= i, aik(t), j = i. Xk6=i L(t) is the Kirchhoff matrix of the communication digraph Γ(t). Sometimes, instead of Γ(t), the digraph Γ′(t) is considered such that j 6= i and aij(t) 6= 0 result in the presence of the i → j arc (instead of j → i in Γ(t)). With respect to Γ′(t), the matrix L(t) defined above is the Laplacian matrix. The classes of Kirchhoff matrices and Laplacian matrices coincide; they only differ in the way they are assigned to digraphs. We will call L(t) the Laplacian matrix of the algorithm (1). The process (1) is said to be convergent if for any initial conditions xi(0), i = 1, . . . , n, and every i, j = 1, . . . , n, it holds that xi(t) − xj(t) → 0 as t → ∞. The convergence properties of the algorithm (1) are determined by the spectral properties of the matrix L(t). 3. SIMPLE PROPERTIES OF THE LAPLACIAN MATRICES By definition, L(t) has zero row sums, consequently, it is singular and the vector 1 = [1, . . . , 1]T belongs to its kernel. Since the off-diagonal entries of L(t) are nonpositive, the diagonal entries AUTOMATION AND REMOTE CONTROL Vol. 70 No. 3 2009 130 CHEBOTAREV AND AGAEV being nonnegative, and L(t) has a weakly dominant diagonal, by Gersgorin's, theorem the real parts of all nonzero eigenvalues of L(t) are strictly positive. That is why all nonzero eigenvalues of the matrix (−1)·L(t) in (2) have strictly negative real parts. The Laplacian matrix of an undirected graph is symmetric, positive semidefinite, and its spec- trum is real and nonnegative. Moreover, 0 is a simple eigenvalue of this matrix if and only if the corresponding graph is connected (see, e. g., [7]). If 0 = λ1 ≤ λ2 ≤ . . . ≤ λn are the Laplacian eigenvalues of an undirected graph, then λ2 is referred to as the algebraic connectivity of the graph. This concept was introduced by M. Fiedler, and it is widely used in both theoretical studies and applications. The Laplacian matrices of undirected graphs have been much studied. They date back to the famous matrix tree theorem by Kirchhoff; A.K. Kelmans pioneered their systematic investigation in the 1960s and 1970s; in the 1990s, the results of their study were presented in a series of reviews by R. Merris ([7] is one of them) and B. Mohar; then the monographs [8, 9] were published. The Laplacian matrices of directed graphs and relationships between their properties and the properties of the corresponding digraphs are still very poorly studied. To some extent, this is due to the fact that the mathematical problems which involve the complex spectra of digraph Laplacians are much more difficult than the corresponding problems regarding the real spectra of ordinary graphs. At the same time, the need for such studies is well recognized. Distributed control is one of the applications where this need is particularly urgent. 4. THE CONVERGENCE OF CONSENSUS ALGORITHMS Assume that the Laplacian matrix L(t) is kept constant: L(t) = L. It is not difficult to formulate a matrix-theoretic necessary and sufficient condition of convergence for the algorithm (2). As already mentioned, the vector 1 = [1, . . . , 1]T belongs to the kernel of L. If 0 is a simple eigenvalue of L, then x(t) → ¯x 1, where ¯x is a scalar, consequently, xi(t) − xj(t) → 0 as t → ∞ for all i, j = 1, . . . , n, and the coordination trajectories converge. As shown in [10, Theorem 4], the zero eigenvalue of L is always semisimple, i. e., its algebraic and geometric multiplicities coincide. Hence, if this eigenvalue is not simple, then the kernel of L is not one-dimensional, and the convergence is violated. Thus, the convergence analysis of algorithm (2) reduces to the determination of conditions under which 0 is a simple eigenvalue of L. 5. RANKS OF LAPLACIAN MATRICES AND A CONVERGENCE CRITERION FOR DISTRIBUTED CONSENSUS ALGORITHMS The rank of the digraph Laplacian was studied in [11]. Recall some graph theory notation. A digraph is called strongly connected (or strong) if it con- tains directed paths from every vertex into every other vertex. Every maximal (by inclusion) strong subgraph of a given digraph is called its strong component or bicomponent. A basis bicomponent of a digraph is a bicomponent such that the digraph does not have any arcs flowing into this bi- component from outside. It is easy to verify that every vertex of a digraph is reachable by a path from at least one basis bicomponent. If a basis bicomponent consists of a single vertex, we call it an undominated vertex of the digraph. A digraph is weakly connected if the graph obtained from it by replacing all arcs with undirected edges is connected. Weak components of a digraph are the maximal by inclusion weakly connected subgraphs of this digraph. A digraph is unilaterally connected if for any vertices i and j 6= i, it contains either a directed path from i into j or a directed path from j into i (or both). Unilateral components of a digraph are its maximal by inclusion uni- laterally connected subgraphs. It should be noted that the relation of unilateral connectivity need AUTOMATION AND REMOTE CONTROL Vol. 70 No. 3 2009 MULTIAGENT SYSTEMS AND LAPLACIAN SPECTRA 131 not be transitive, so it does not induce a decomposition of the vertex set into equivalence classes. Therefore, as distinct from the strong and weak components, unilateral components can overlap. A subgraph of a digraph is spanning if the vertex sets of the graph and subgraph coincide. A diverging tree is a rooted directed tree containing directed paths from the root into all other vertices. A diverging forest is a rooted directed forest all of whose weak components are diverging trees. For an arbitrary digraph Γ, consider its spanning diverging forests. Such forests are also called out-forests of Γ. A spanning diverging forest F of Γ is a maximum out-forest if Γ has no out-forest with the number of arcs greater than in F . It is easily seen that every maximum out-forest has the minimum possible number of weak components (diverging trees); this number will be called the out-forest dimension of Γ and denoted by d. The number of arcs in any maximum out-forest is obviously n − d, where n is the number of vertices in Γ. The following results were obtained in [11]. Proposition 1. Let L be the Kirchhoff matrix of a digraph Γ. Then rank L = n − d, where n is the number of vertices in Γ and d is the out-forest dimension of Γ. Proposition 2. The out-forest dimension of a digraph is equal to its number of basis bicompo- nents. Proposition 3. The out-forest dimension of a strong digraph is unity. Proposition 4. The out-forest dimension of a digraph is no less than its number of weak com- ponents and does not exceed the number of its strong components and the number of its unilateral components. Zero is a simple eigenvalue of L if and only if rank L = n − 1. That is why the above propositions imply Corollary 1. Corollary 1. Let L be the Kirchhoff matrix of a digraph Γ. Then 0 is a simple eigenvalue of L if and only if Γ has a spanning diverging tree or, equivalently, Γ has only one basis bicomponent. By Propositions 1 and 3, the premise of Corollary 1 is satisfied, for instance, for any strong digraph. Later on, the first statement of Corollary 1 was obtained in [12 -- 16]. Corollary 1 provides a convergence criterion for the consensus algorithm (2). Theorem 1. The consensus algorithm (2) with a stable Laplacian matrix L(t) = L converges to a vector with equal components for any vector of initial conditions x(0) if and only if the cor- responding communication digraph Γ has a spanning diverging tree or, equivalently, has a unique basis bicomponent. A more general problem is describing the whole domain in the space of initial conditions be- longing to which guarantees the fulfillment of the xi(t) − xj(t) → 0, i, j = 1, . . . , n, condition, provided that convergence does not generally hold. This problem reduces to the analysis of the kernel of L. 6. THE KERNEL AND EIGENPROJECTION OF THE DIGRAPH LAPLACIAN The eigenprojection 2 corresponding to eigenvalue 0 or, for short, simply eigenprojection of a square matrix A is a projection (i. e., an idempotent matrix) Z such that R(Z) = N (Aν) and 2 It is also called principal idempotent. AUTOMATION AND REMOTE CONTROL Vol. 70 No. 3 2009 132 CHEBOTAREV AND AGAEV N (Z) = R(Aν), where R(A) and N (A) are the range and the kernel (null space) of A, respectively, and ν = ind A is the index of A, i. e., the minimum k ∈ {0, 1, . . . } such that rank Ak+1 = rank Ak. In other words, the eigenprojection of A is the projection on N (Aν ) along R(Aν ). The eigenprojection is unique, as an idempotent matrix is uniquely determined by its range and kernel. A number of equivalent definitions of eigenprojection can be found in [17]. In [11] we considered the normalized matrix of maximum out-forests of a digraph denoted by ¯J (the definition of this matrix is given below in Section 8) and proved the following properties of ¯J. Proposition 5. Let L, d and ¯J be the Kirchhoff matrix, the out-forest dimension and the nor- ¯J 2 = ¯J; malized matrix of maximum out-forests of digraph Γ, respectively. Then rank ¯J = d; L ¯J = ¯JL = 0. By the second statement of Proposition 5, ¯J is a projection. By the third statement, R( ¯J) ⊆ N (L) and R(L) ⊆ N ( ¯J ). Using the relationship between the ranks of ¯J and L (rank ¯J = d and rank L = n − d by Proposition 1) and the semisimplicity of 0 as the eigenvalue of L, we obtain that the above inclusions can be replaced with equalities. Finally, the semisimplicity of 0 as the eigenvalue of L implies that ind L = 1. Thus, the following two propositions hold true. Proposition 6. Let L and ¯J be the Kirchhoff matrix and the normalized matrix of maximum out-forests of Γ, respectively. Then ¯J is the eigenprojection of L. Proposition 7. The linear spans of the columns and the rows of ¯J coincide with the kernel and the left null space of L, respectively. By virtue of Propositions 6 and 7, matrix ¯J is useful for the analysis of continuous distributed consensus algorithms. In Section 7, we will see that ¯J is also applicable to the study of iterative consensus algorithms. 7. ITERATIVE CONSENSUS ALGORITHMS Consider the finite-difference counterpart of the continuous consensus algorithm (1) with con- stant coefficients aij: xi(k + 1) = xi(k) − ε nXj=1 aij (xi(k) − xj(k)), i = 1, . . . , n, where k is the discrete time and ε > 0 is the step size. Rewrite algorithm (4) in matrix form: where x(k + 1) = P x(k), P = I − εL. (4) (5) (6) If the step size ε is small enough, then P is a row stochastic matrix. This follows from the fact that L has nonpositive off-diagonal entries and zero row sums. The corresponding condition of the "smallness" of ε is as follows [11]: 0 < ε ≤max i Xj6=i aij −1 . (7) AUTOMATION AND REMOTE CONTROL Vol. 70 No. 3 2009 MULTIAGENT SYSTEMS AND LAPLACIAN SPECTRA 133 The matrix P = exp(−εL) corresponds to the continuous consensus algorithm (2), and (6) can be considered as the expansion of exp(−εL) to the linear term. Matrix (6) is sometimes called the Perron matrix with parameter ε of digraph Γ. From (5), for any natural m one has x(m) = P mx(0), (8) therefore, the properties of the process (5) are determined by the properties of the sequence {P, P 2, . . . , P m, . . . }. From the theory of Markov chains it is known that this sequence need not converge (the necessary and sufficient condition of its convergence is the aperiodicity of the chain), but the Ces`aro limit (also called the time average limit) P ∞ = lim m→∞ 1 m mXi=1 P i, (9) always exists and coincides with the limit of the sequence {P, P 2, . . . , P m, . . . } in case the latter converges. Otherwise, if the chain is periodic with period s, then where P (1), . . . , P (s) are the limits of its converging subsequences: P (i) = lim j→∞ P js+i. P ∞ = 1 s(cid:16)P (1) + . . . + P (s)(cid:17) It was shown in [18] that P ∞ is the eigenprojection of I − P . Hence, by (6), P ∞ is the eigenpro- jection of L as well. This fact along with the uniqueness of the eigenprojection and Proposition 6 imply Proposition 8. Proposition 8. Let a row stochastic matrix P be connected with the Kirchhoff matrix L of a digraph Γ by equation (6) with ε > 0. Then the matrix P ∞ defined by (9) coincides with the normalized matrix ¯J of maximum out-forests of Γ. Thus, the matrix P ∞, which determines the asymptotic behavior of the consensus algorithm (4), is equal to the normalized matrix of maximum out-forests of the communication digraph that corresponds to this algorithm. The matrix ¯J is thereby important for the analysis of iterative consensus processes. Indeed, according to (8), to know the average asymptotic state of the algorithm (4), it suffices to consider the product P ∞x(0) = ¯J x(0). To compute the matrix ¯J, one can use the algorithm proposed in [19, 20], which reduces to min{n − d − 1, 0} multiplications of matrices of order n. It should be remarked that Proposition 8 coincides with the Markov chain tree theorem first obtained by Wentzell and Freidlin [21] and rediscovered by Leighton and Rivest [22]. Consider the convergence of the consensus algorithm (4). Since the spectral radius of P is 1, the convergence is violated only if P has an eigenvector not proportional to 1 = [1, . . . , 1]T and corresponding to an eigenvalue of modulus 1. The subspace of invariant vectors of P coincides with the kernel of L, hence it is one-dimensional if and only if the premise of Theorem 1 holds. As for the complex eigenvalues of modulus 1, P can have them only if ε coincides with the right endpoint of the interval (7). Indeed, in the opposite case, the increase of ε up to the endpoint of the interval would have resulted in the appearance of an eigenvalue with the modulus greater than 1, in contradiction with the stochasticity of P (see also [23]). Consequently, the convergence of consensus algorithm (4) is guaranteed by the fulfillment of the premise of Theorem 1 together with the strict form of inequality (7). Otherwise, if ε coincides with the right endpoint of the interval (7) and the Markov chain determined by the matrix P = I − εL is periodic, then the convergence is violated. AUTOMATION AND REMOTE CONTROL Vol. 70 No. 3 2009 134 CHEBOTAREV AND AGAEV 8. THE NORMALIZED MATRIX OF MAXIMUM OUT-FORESTS Can the entries of the matrix ¯J, which proves to be useful for the analysis of consensus algorithms, be interpreted in terms of the communication digraph Γ ? Such a connection is specified by the very definition of ¯J : every entry ¯Jij of ¯J is defined [11] as the ratio of the total weight3 of Γ's maximum out-forests that have vertex i belonging to a tree diverging from j to the total weight of all maximum out-forests in Γ. The following theorems [11] summarize the properties of ¯J. Theorem 2. Suppose that V (Γ) = {1, . . . , n} and E(Γ) are the vertex set and arc set of a digraph Γ, K is the vertex set of some basis bicomponent of Γ, K + is the set of vertices reachable by paths from K and unreachable from the other basis bicomponents of Γ, fK is the union of the vertex sets of all basis bicomponents, d is the out-forest dimension of Γ, ¯J is the normalized matrix of maximum out-forests of Γ, T is the set of spanning diverging trees in the restriction ΓK of Γ to K, Tj is the subset of T consisting of the trees that diverge from j, Γ−K is the spanning subgraph of Γ with the edge set E(Γ) \ E(ΓK ), PK→i is the set of maximum out-forests of Γ−K in which i is reachable from some vertex that belongs to K, ε(·) designates the weight of a set of subgraphs (equal to the sum of the weights of its elements), σn−d is the weight of the set of maximum out-forests of Γ. Then the following statements are true: ¯Jij = ¯Jjj = ε(Tj)/ε(T); of ¯J . 1. ¯J is a row stochastic matrix : ¯Jij ≥ 0, and ¯Jik = 1 i, j ∈ V (Γ); ¯Jjj = 1. In particular, if j is an undominated vertex, then ¯Jjj = 1; 3. Let j ∈ K. For any i ∈ V (Γ), ¯Jij = ε(Tj) ε(PK→i)/σn−d. Moreover, if i ∈ K +, then nPk=1 2. ¯Jij 6= 0 ⇔ (j ∈ fK and i is reachable from j in Γ); 4. Pj∈K 5. If j1, j2 ∈ K, then ¯J·j2 = (cid:0)ε(Tj2)/ε(Tj1 )(cid:1) ¯J·j1, where ¯J·j1 and ¯J·j2 are the j1 and j2 columns iji of maximum out-forests of the digraph ΓK coincides with 1. The normalized matrix ¯JK =h ¯J K Corollary 2 (from the statement 3 of Theorem 2). the principal submatrix of ¯J corresponding to the basis bicomponent K. 2. If i ∈ K + and j ∈ K +\K, then ¯J is preserved under the variations of the weight of (i, j). The following Theorem 3 is concerned with the comparison of the entries of ¯J. Theorem 3. In the notation of Theorem 2, let K(k) be a basis bicomponent of Γ that contains 1. ¯Jii ≥ ¯Jji; a vertex k ∈fK. For all i, j ∈ V (Γ) we have: 2. If ¯Jii > ¯Jji, then i ∈fK, j /∈ K +(i), therefore, Γ contain no paths from j into i; 3. If ¯Jii > ¯Jji > 0, then j /∈fK, consequently, j is not the root in any maximum out-forest of Γ; 4. If ¯Jij > 0, then ¯Jii = ¯Jji. As mentioned above, these properties are useful for the analysis of the consensus algorithms (2) and (5) as well as generalizations of these algorithms, in particular, when the convergence is not guaranteed by the structure of the communication digraph. Some of the foregoing results regarding the kernel of the digraph Laplacian were also obtained in [24]. 3 The weight of a digraph is the product of the weights of all its arcs. AUTOMATION AND REMOTE CONTROL Vol. 70 No. 3 2009 MULTIAGENT SYSTEMS AND LAPLACIAN SPECTRA 135 9. OTHER PROBLEMS OF DECENTRALIZED CONTROL The distributed consensus algorithm (2) is the simplest algorithm of decentralized control. Yet this algorithm is a basic one, because the linear operator of agreement (−1)·L(t) it contains is an indispensable constituent of more complex algorithms as well. Therefore, the convergence properties of these more complex algorithms are in many respects determined by the spectral properties of the corresponding Laplacian matrices and by the forest structure of the related communication digraphs. In this section, we briefly survey some modifications of the consensus algorithm (2) and some other related problems of decentralized control. First of all, it should be noted that the model (2) does not generally imply that the commu- nication digraph Γ(t) is fixed. The rejection of this assumption makes the model more realistic. Indeed, in many applications the agents primarily communicate with the closest neighbors, but those can be acquired or lost during the motion. To preserve the relative simplicity of the model, some restrictions on the structure of the communication digraph are applied. Usually, either it is supposed to be piecewise constant and a set of possible switchings is specified or the weights aij(t) are defined with a permissible variation. Sometimes fluctuation of the weights is described by a probabilistic model. As a result, the analysis, in most cases, reduces to solving a series of prob- lems with stable communication digraphs and combining the partial solutions. In some cases, the analysis involves the study of infinite products of stochastic matrices taken from a specific set [25]. Besides that, the model becomes more realistic if it allows communication delays. These delays can be associated with either information transmission or with processing messages after receipt. In this case, the model of the form (1) contains terms such as xj(t − δij ), where δij is a delay typical of the pair of agents (i, j). If the delays are the same for all pairs of agents, then their presence does not essentially alter the properties of the coordination trajectories [26]. More general cases were considered in [27]. In many technical applications it is desirable to maximize the rate of convergence. An approach to solving such problems was developed in [28]. It is noteworthy that in case of undirected graphs, the problem of the best convergence is connected [29] with the maximization of the algebraic connectivity (the second smallest eigenvalue of the Laplacian matrix already mentioned in Section 3) of the communication graph. The algebraic connectivity itself can serve as a good measure of the convergence rate of the algorithm [30]. For the case of directed graphs, different generalizations of the algebraic connectivity were proposed, including the second smallest real part of the Laplacian eigenvalues [31], the second smallest modulus of the Laplacian eigenvalues, and the second smallest eigenvalue of the symmetric part (L+LT)/2 of the Laplacian matrix [23, 32]. However, in the general case of digraphs, the problem of optimizing the convergence rate is still insufficiently studied. Not even the problem of localizing the spectra of nonsymmetric Laplacian matrices (see Section 10) has been completely investigated. Let us turn to some extensions of the model (2) that have a more complex structure. For the analysis of synchronization in systems of nonlinear oscillators and for solving some other problems, the following class of models was considered: xi = f (xi) − γ nXj=1 aij(t) (xi − xj), i = 1, . . . , n, γ > 0. (10) It has been shown in [33] that in case of undirected communication graphs, the corresponding synchronization process is essentially determined by the algebraic connectivity of the graph. AUTOMATION AND REMOTE CONTROL Vol. 70 No. 3 2009 136 CHEBOTAREV AND AGAEV The consensus algorithm for a double-integrator dynamics has the form: xi = − nXj=1 aij(t) ((xi − xj) + γ( xi − xj)) , i = 1, . . . , n, (11) where γ > 0 controls velocity matching. It was shown in [34] that Γ is of considerable importance here, however, for the algorithm to converge, as well as for the basic model (Theorem 1), the communication digraph must have a spanning diverging tree. More general differential models of the second order were used to describe the motion of the complexes (flocks, schools, troop, swarms, etc.) of physical objects. "Flocking" requires both maintaining the pattern formed by the objects' positions (up to translation and/or rotation) and formation maneuvering [31, 35]. In this case, the existence of a spanning diverging tree in the com- munication digraph is also a necessary condition of the effective control for many typical problems. 10. LOCALIZING THE SPECTRA OF NONSYMMETRIC LAPLACIAN MATRICES For the analysis of models generalizing simple consensus models (1) or (4) it is necessary to know the whole Laplacian spectrum of the communication digraph rather then the sole multiplicity of its zero eigenvalue. Another important but rather difficult problem is recovering connections between the Laplacian spectrum of a digraph and the topological properties of the same digraph, including the "degree of cyclicity" and many others. The problem of localizing the spectra of Laplacian matrices was partially solved in [10]. Let us consider some results of this paper. A real square matrix of order n will be called a standardized Laplacian matrix if (1) its row sums are equal to 0 and (2) its off-diagonal elements are nonpositive and do not exceed 1/n in absolute value. This standardization enables one to consider results on the spectra of Laplacian matrices the class Gb of weighted digraphs with positive arc weights not exceeding b > 0 is considered and L(Γ) is the Laplacian matrix of a weighted digraph Γ ∈ Gb on n vertices, then the standardized with various n in the same framework. Standardized Laplacian matrices will be denoted by eL. If Laplacian matrix associated with Γ in this class is, by definition, eL(Γ) = (nb)−1L(Γ). Let J ∈ Rn×n be the matrix with all entries 1/n; K = I − J. Then K is the standardized Laplacian matrix of the complete digraph with all arc weights b in the class Gb. Define the matrices and P = eL + J eLc = K −eL. (12) (13) By (12) and (13), P is a stochastic matrix, eLc being the standardized Laplacian matrix of the complementary weighted digraph Γc in which (b − εij) is the weight of arc (i, j), j 6= i, provided that εij is the weight of this arc in Γ. If εij = b, then Γc has no (i, j) arc and vice versa: if Γ has no (i, j) arc, then the weight of (i, j) in Γc is b. It follows from (12), (13) and the definition of K that Consider the results [10] connecting the spectra of eL, P and eLc. AUTOMATION AND REMOTE CONTROL Vol. 70 No. 3 P = I −eLc. (14) 2009 x =(cid:18)I − J 1 − λ(cid:19) v fP (λ) = λ − 1 λ (λ) = (−1)n−1 λ feL(λ), 1 − λ feLc feL(1 − λ). (15) (16) (17) MULTIAGENT SYSTEMS AND LAPLACIAN SPECTRA 137 respectively. Then for λ /∈ {0, 1}, the following statements are equivalent:4 Theorem 4. Let eL be a standardized Laplacian matrix; let P and eLc be defined by (12) and (13), (a) λ ∈ speL; (b) λ ∈ sp P ; (c) 1 − λ ∈ speLc, and these eigenvalues have the same geometric multiplicity. Furthermore, v is an eigenvector of eL corresponding to λ /∈ {0, 1} if and only if the vector 5 is an eigenvector of P corresponding to λ and an eigenvector of eLc corresponding to 1 − λ. Theorem 5. Let feL(λ), fP (λ) and feLc (λ) be the characteristic polynomials of eL, P and eLc, respectively. Then for all λ 6∈ {0, 1}, Theorem 6. For any standardized Laplacian matrix eL and the corresponding stochastic matrix P defined by (12), eL and P are semiconvergent. For completeness, the following theorem contains some results mentioned above. Theorem 7. Suppose that d and dc are the in-forest dimensions of the digraph whose standard- and these eigenvalues are semisimple; ized Laplacian matrix is eL and the complementary digraph, respectively; mA(λ) is the algebraic multiplicity of λ ∈ sp A; VA(λ) is the eigenspace of A corresponding to λ. Then meL(1) = dc − 1; (ii) mP (0) = d − 1, mP (1) = dc; (0) = dc, (i) meL(0) = d, (iii) meLc (iv) if v ∈ VeL(0) and Kv 6= 0, then Kv ∈ VP (0) = VeLc (0) and Kx 6= 0, then Kx ∈ VeL(1). Let eΛn be the class of standardized Laplacian matrices of order n. We now turn to the problem of localizing the spectra of the matrices eL ∈ eΛn. The following results were obtained in [10]. (1) = d − 1, meLc if x ∈ VP (1) = VeLc Theorem 8. All eigenvalues of standardized Laplacian matrices of order n belong to the meet of : • two closed disks, one centered at 1/n, the other centered at 1 − 1/n, each having radius 1 − 1/n, • two closed smaller angles, one bounded with the two half-lines drawn from 1 through e−2πi/n and e2πi/n, the other bounded with the half-lines drawn from 0 through e−(π2−πn)i and e(π2−πn)i, and (1); • the band ℑ(z) ≤ 1 2n cot π 2n . Remark. Using the theorem by Dmitriev and Dynkin [36] on the spectra of stochastic matrices, it n if and only if Γ is a Hamiltonian n sin 2π n . can be shown that speL contains an eigenvalue with argument π cycle on n vertices. In this case, such an eigenvalue λ is unique, λ ≤ 2 n , and ℑ(λ) ≤ 1 n sin π 2 − π 4 spA is the spectrum of A. 5 For simplicity, we sometimes write A/α instead of (1/α)A, where A is a matrix and α 6= 0 is a complex number. AUTOMATION AND REMOTE CONTROL Vol. 70 No. 3 2009 138 CHEBOTAREV AND AGAEV ℑ(λ) ✻ π n 1 2n cot π 2n 1 n sin 2π n 0 1 n 0.5 n−1 n 1 ℜ(λ) ✲ Fig. 1. The domain which, according to Theorem 8, contains the spectrum of each standardized Laplacian matrix of order n (hatched); in this figure, n = 7. The components of any corresponding eigenvector are the vertices of a regular polygon. Similarly, speL contains an eigenvalue that belongs to the segment [1, e2πi/n] if and only if the complementary digraph Γc is a Hamiltonian cycle on n vertices. As above, such an eigenvalue λ′ is unique and ℑ(λ′) ≤ 1 n sin 2π n . This remark and Theorem 8 are illustrated by Fig. 1. Let λk(n) = n−1(cid:16)k − e−2πi/n − . . . − e−2kπi/n(cid:17) = n−1 k − sin kπ n sin π n e−(k+1)πi/n! , k = 1, . . . , n − 1; (18) by S(n) we denote the closed convex polygon with vertices λ0(n) = 0, λ1(n), . . . , λn−2(n), λn−1(n) = 1, λn−2(n), . . . , λ1(n). (19) Theorem 9. Every point of the polygon S(n) is an eigenvalue of some standardized Laplacian matrix eL ∈ eΛn. Let h(n) = sup{ℑ(λ) : λ is an eigenvalue of some eL ∈ eΛn}. By Theorem 8, for all n = 2, 3, . . . it holds that h(n) ≤ 1 2n cot π 2n . Theorem 10. If n is odd, then h(n) = 1 2n cot π , 1 2n moreover, h(n) = ℑ(cid:16)λ(n−1)/2(n)(cid:17) , where λ(n−1)/2(n) is defined by (18). AUTOMATION AND REMOTE CONTROL Vol. 70 No. 3 2009 MULTIAGENT SYSTEMS AND LAPLACIAN SPECTRA 139 Proposition 9. If n > 2 is even, then max 0≤k≤2n−1 ℑ(λk(n)) = ℑ(λn/2(n)) = 1 n cot π n < 1 2n cot π 2n . Corollary (from Theorems 8 and 9 and Proposition 9). lim n→∞ h(n) = 1/π. The following conjecture is yet unproved. Conjecture. All eigenvalues of standardized Laplacian matrices of order n belong to the poly- gon S(n) whose vertices are defined by (19). Note that the vertices λk(n) and λk(n) of the polygon S(n) belong to the spectrum of the (kI − C − C 2 − . . . − C k), where C = [cuv] is the matrix of standardized Laplacian matrix eLk(n) = a cyclic permutation of order n with entries cuv =( 1, v − u ∈ {1, n − 1}, standardized Laplacian matrices are eLk(n) belong to the class of balances digraphs, which are relevant to the synthesis of multiagent control protocols (see, e. g., [23]). In balanced digraphs, the total weight of arcs converging to any vertex is equal to the total weight of arcs diverging from it. Besides that, for these digraphs, the total weight of converging arcs is the same for all vertices, i. e., they are regular. The digraphs whose 1 n 0, otherwise. standardized Laplacian matrix where S = [sij] is the standardized Laplacian matrix of a digraph obtained from the digraph with (I −C n/2) by decreasing the weights of some arcs. It turns out that If n is even and k < n/2, then of special interest are the matrices of the form eLk(n) + S, the maximum imaginary part of the eigenvalues of eLk(n) + S does not exceed such a part for eLk(n), and the spectrum of eLk(n) + S is invariant under any change of nonzero sij and sji that preserves sij + sji. In other words, the spectra of the standardized Laplacian matrices eLk(n) + S depend on the total weights of opposite "diagonal" arcs of the digraph rather than on their individual weights. 1 n Theorem 11. The boundary of the polygon S(n) with vertices (19) converges, as n → ∞, to the curve made up by the parts of two cycloids whose parametric equations are: z(τ ) = x(τ ) + i y(τ ) and z(τ ) = x(τ ) − i y(τ ), where τ ∈ [0, 2π] and x(τ ) = (2π)−1(τ − sin τ ), y(τ ) = (2π)−1(1 − cos τ ). Fig. 2 shows the polygons S(n) at n = 4 and n = 5 as well as the limit curve whose equation is given by Theorem 11. As mentioned above, the results on localizing the spectra of Laplacian matrices given in this section are necessary for the analysis and synthesis of control algorithms that contain reaching consensus as one of their elements (see, e. g., [31]). 11. CONCLUSION Recently, decentralized control of multiagent systems has become one of the most popular and rapidly evolving branches of control theory. Dozens of research groups have published many hun- dreds of papers and since these groups are working concurrently and are studying similar mod- els, their results substantially overlap. To get familiar with this trend, the recent surveys and AUTOMATION AND REMOTE CONTROL Vol. 70 No. 3 2009 140 CHEBOTAREV AND AGAEV ✻ℑ(λ) 1/π n = 4 0 0.5 ✲ℜ(λ) 1 n = 5 Fig. 2. The polygon of the spectra of standardized digraph Laplacians at n = 4 and n = 5 and the limit oval specified by Theorem 11. monographs [23, 32, 37, 38] can be recommended. The methods of algebraic graph theory play an increasingly important part in the studies of the last decade dealing with multiagent systems. More specifically, the subject of these investigations is the relationship between the topological properties of digraphs that represent information interchange among agents and the spectral properties of the corresponding Laplacian matrices. At the same time, the Laplacian theory of directed graphs6 is still insufficiently developed; the need for new strong results in this field is very acute. One more thesis accepted by most researchers is that for a new stage of its development, this discipline re- quires more experimental studies7 and more practical applications of the theoretical results. These will reveal to what extent the proposed algorithms are robust to perturbations and discrepancies between theory and real behavior. REFERENCES 1. Varshavsky, V. and Pospelov, D. Puppets without Strings, Moscow: Mir Publishers, 1988. 2. DeGroot, M.H., Reaching a Consensus, J. Amer. Statist. Associat., 1974, vol. 69, no. 345, pp. 118 -- 121. 3. Dalkey, N.C. and Helmer, O., An Experimental Application of the Delphi Method to the Use of Experts, Management Sci., 1963, vol. 9, pp. 458 -- 467. 4. Borkar V., and Varaiya, P., Asymptotic Agreement in Distributed Estimation, IEEE Trans. Automat. Control., 1982, vol. 27, pp. 650 -- 655. 5. Tsitsiklis, J.N. and Athens, M., Convergence and Asymptotic Agreement in Distributed Decision Prob- lems, IEEE Trans. Automat. Control., 1984, vol. 29, pp. 690 -- 696. 6. Reynolds, C.W., Flocks, Herds, and Schools: A Distributed Behavioral Model, Comput. Graphics., 1987, vol. 21, no. 4, pp. 25 -- 34. 7. Merris, R., Laplacian Matrices of Graphs: A survey, Linear Algebra Appl., 1994, vol. 197 -- 198, pp. 143 -- 176. 8. Chung, F.R.K., Spectral Graph Theory, Providence, RI: Amer. Math. Soc., 1997. 9. Godsil, C. and Royle, G., Algebraic Graph Theory, New York: Springer-Verlag, 2001. 10. Agaev, R. and Chebotarev, P., On the Spectra of Nonsymmetric Laplacian Matrices, Linear Algebra Appl., 2005, vol. 399, pp. 157 -- 168. 1 6 In [31] this area was called an "unexpected new mathematical territory." 7 They are still scarce; see, e. g., [39, 40]. AUTOMATION AND REMOTE CONTROL Vol. 70 No. 3 2009 MULTIAGENT SYSTEMS AND LAPLACIAN SPECTRA 141 11. Agaev, R.P. and Chebotarev, P.Yu. The Matrix of Maximum Out Forests of a Digraph and Its Appli- cations, Automat. Remote Control, 2000, vol. 61, no. 9, part 1, pp. 1424 -- 1450. 12. Jadbabaie, A., Lin, J., and Morse, A.S., Coordination of Groups of Mobile Autonomous Agents Using Nearest Neighbor Rules, IEEE Trans. Automat. Control, 2003, vol. 48, no. 6, pp. 988 -- 1001. 13. Ren, W., Beard, R.W., and McLain, T.W., Coordination Variables and Consensus Building in Multiple Vehicle Systems, in Cooperative Control: A Post-Workshop Volume 2003, Block Island Workshop on Cooperative Control., Kumar, V., Leonard, N.E., and Morse, A.S., Ed., New York: Springer-Verlag, 2005, pp. 171 -- 188. 14. Caughman, J.S., Lafferriere, G., Veerman, J.J.P, and Williams, A., Decentralized Control of Vehicle Formations, Syst. Control Lett., 2005, vol. 54, no. 9, pp. 899 -- 910. 15. Lin, Z., Francis, B., and Maggiore, M., Necessary and Sufficient Graphical Conditions for Formation Control of Unicycles, IEEE Trans. Automat. Control, 2005, vol. 50, no. 1, pp. 121 -- 127. 16. Moreau, L., Stability of Multiagent Systems with Time-Dependent Communication Links, IEEE Trans. Automat. Control, 2005, vol. 50, no. 2, pp. 169 -- 182. 17. Agaev, R.P. and Chebotarev, P.Yu., On Determining the Eigenprojection and Components of a Matrix, Automat. Remote Control, 2002, vol. 63, no. 10, pp. 1537 -- 1545. 18. Rothblum, U.G., Computation of the Eigenprojection of a Nonnegative Matrix at Its Spectral Radius, Math. Program. Study, 1976, vol. 6, pp. 188 -- 201. 19. Agaev, R.P. and Chebotarev, P.Yu., Spanning Forests of a Digraph and Their Applications, Automat. Remote Control, 2001, vol. 62, no. 3, pp. 443 -- 466. 20. Chebotarev, P. and Agaev, R., Forest Matrices Around the Laplacian Matrix, Linear Algebra Appl., 2002, vol. 356, pp. 253 -- 274. 21. Wentzell, A.D. and Freidlin, M.I., On Small Random Perturbations of Dynamical Systems, Russian Mathematical Surveys, 1970, vol. 25, no. 1, pp. 1 -- 55. 22. Leighton, T. and Rivest, R.L., The Markov Chain Tree Theorem, Computer Science Techn. Rep. MIT/LCS/TM -- 249, Laboratory of Computer Science, MIT, Cambridge, Mass., 1983. 23. Olfati-Saber, R., Fax, J.A., and Murray, R.M., Consensus and Cooperation in Networked Multi-Agent Systems, Proc. IEEE, 2007, vol. 95, no. 1, pp. 215 -- 233. 24. Caughman, J.S. and Veerman, J.J.P., Kernels of Directed Graph Laplacians, Electron. J. Combinat., 2006, vol. 13, no. 1 -- R39. 25. Ren, W., Beard, R.W., and Kingston, D.B., Multi-Agent Kalman Consensus With Relative Uncertainty, in Proc. Amer. Control Conf., Portland, OR, 2005, pp. 1865 -- 1870. 26. Moreau, L., Stability of Continuous-Time Distributed Consensus Algorithms, in Proc. IEEE Conf. De- cision Control, Paradise Island, Bahamas, 2004, pp. 3998 -- 4003. 27. Fang, L., and Antsaklis, P.J., On Communication Requirements for Multiagent Consensus Seeking, in Networked Embedded Sensing and Control, Lecture Notes in Control Inform. Sci., vol. 331, Antsaklis, P.J. and Tabuada, P., Ed., Berlin: Springer-Verlag, 2006, pp. 53 -- 67. 28. Xiao, L. and Boyd, S., Fast Linear Iterations for Distributed Averaging, Syst. Control Lett., 2004, vol. 53, pp. 65 -- 78. 29. Kim, Y. and Mesbahi, M., On Maximizing the Second Smallest Eigenvalue of a State-Dependent Graph Laplacian, IEEE Trans. Automat. Control, 2006, vol. 51, pp. 116 -- 120. 30. Olfati-Saber, R. and Murray, R.M., Consensus Problems in Networks of Agents with Switching Topology and Time-Delays, IEEE Trans. Automat. Control, 2004, vol. 49, pp. 1520 -- 1533. 31. Veerman, J.J.P., Lafferriere, G., Caughman, J.S., and Williams, Flocks and Formations, J. Statist. Physics, 2005, vol. 121, no. 5 -- 6, pp. 901 -- 936. 32. Wu, C.W., Synchronization in Complex Networks of Nonlinear Dynamical Systems, World Scientific, 2007. 33. Preciado, V.M. and Verghese, G.C., Synchronization in Generalized Erdos-R´enyi Networks of Nonlinear Oscillators, in Proc. IEEE Conf. Decision Control, Eur. Control Conf., Seville, Spain, 2005, pp. 4628 -- 4633. AUTOMATION AND REMOTE CONTROL Vol. 70 No. 3 2009 142 CHEBOTAREV AND AGAEV 34. Ren, W. and Atkins, E., Second-Order Consensus Protocols in Multiple Vehicle Systems with Local Interactions, in Proc. AIAA Guidance, Navigation, Control Conf., San Francisco, CA, 2005. Paper AIAA -- 2005 -- 6238. 35. Lawton, J.R., Beard, R.W., and Young, B., A Decentralized Approach to Formation Maneuvers, IEEE Trans. Robot. Automat., 2003, vol. 19, no. 6, pp. 933 -- 941. 36. Dmitriev, N. and Dynkin, E., Characteristic Roots of Stochastic Matrices, Izv. Akad. Nauk SSSR (Bull. Acad. Sci. USSR), Mat., 1946, vol. 10, pp. 167 -- 184. Translated in Eleven Papers Translated from the Russian, Bernik, V.I., Ed., Amer. Math. Soc., 1988. 37. Ren, W., Beard, R.W., and Atkins, E.M., Information Consensus in Multivehicle Cooperative Control, IEEE Control Syst. Magazin, 2007, vol. 27, no. 2, pp. 71 -- 82. 38. Ren, W. and Beard, R.W., Distributed Consensus in Multi-Vehicle Cooperative Control, London: Springer-Verlag, 2008. 39. Regmi, A., Sandoval, R., Byrne, R., et al., Experimental Implementation of Flocking Algorithms in Wheeled Mobile Robots, in Proc. Amer. Control Conf., Portland, OR, 2005, pp. 4917 -- 4922. 40. Marshall, J.A., Fung, T., Broucke, M.E., et al., Experiments in Multirobot Coordination, Robot. Au- tonom. Syst., 2006, vol. 54, no. 3, pp. 265 -- 275. This paper was recommended for publication by B.T. Polyak, a member of the Editorial Board AUTOMATION AND REMOTE CONTROL Vol. 70 No. 3 2009
1911.12716	2	1911	2019-12-20T13:26:56	HS-CAI: A Hybrid DCOP Algorithm via Combining Search with Context-based Inference	[ "cs.MA" ]	Search and inference are two main strategies for optimally solving Distributed Constraint Optimization Problems (DCOPs). Recently, several algorithms were proposed to combine their advantages. Unfortunately, such algorithms only use an approximated inference as a one-shot preprocessing phase to construct the initial lower bounds which lead to inefficient pruning under the limited memory budget. On the other hand, iterative inference algorithms (e.g., MB-DPOP) perform a context-based complete inference for all possible contexts but suffer from tremendous traffic overheads. In this paper, $(i)$ hybridizing search with context-based inference, we propose a complete algorithm for DCOPs, named {HS-CAI} where the inference utilizes the contexts derived from the search process to establish tight lower bounds while the search uses such bounds for efficient pruning and thereby reduces contexts for the inference. Furthermore, $(ii)$ we introduce a context evaluation mechanism to select the context patterns for the inference to further reduce the overheads incurred by iterative inferences. Finally, $(iii)$ we prove the correctness of our algorithm and the experimental results demonstrate its superiority over the state-of-the-art.	cs.MA	cs	HS-CAI: A Hybrid DCOP Algorithm via Combining Search with Context-based Inference Dingding Chen,1 Yanchen Deng,2 Ziyu Chen,1,∗ Wenxing Zhang,1 Zhongshi He1 1College of Computer Science, Chongqing University, Chongqing, China 2School of Computer Science and Engineering, Nanyang Technological University, Singapore {dingding,chenziyu,zshe}@cqu.edu.cn, [email protected], [email protected] 9 1 0 2 c e D 0 2 ] A M . s c [ 2 v 6 1 7 2 1 . 1 1 9 1 : v i X r a Abstract Search and inference are two main strategies for opti- mally solving Distributed Constraint Optimization Problems (DCOPs). Recently, several algorithms were proposed to combine their advantages. Unfortunately, such algorithms only use an approximated inference as a one-shot preprocess- ing phase to construct the initial lower bounds which lead to inefficient pruning under the limited memory budget. On the other hand, iterative inference algorithms (e.g., MB-DPOP) perform a context-based complete inference for all possible contexts but suffer from tremendous traffic overheads. In this paper, (i) hybridizing search with context-based inference, we propose a complete algorithm for DCOPs, named HS- CAI where the inference utilizes the contexts derived from the search process to establish tight lower bounds while the search uses such bounds for efficient pruning and thereby reduces contexts for the inference. Furthermore, (ii) we in- troduce a context evaluation mechanism to select the context patterns for the inference to further reduce the overheads in- curred by iterative inferences. Finally, (iii) we prove the cor- rectness of our algorithm and the experimental results demon- strate its superiority over the state-of-the-art. Introduction Distributed Constraint Optimization Problems (DCOPs) (Hirayama and Yokoo 1997; Fioretto, Pontelli, and Yeoh 2018) are an elegant model for representing Multi-Agent Systems (MAS) where agents coordinate with each other to optimize a global objective. Due to their ability to capture essential MAS aspects, DCOPs can formalize various appli- cations in the real world such as sensor network (Farinelli, Rogers, and Jennings 2014), task scheduling (Maheswaran et al. 2004; Fioretto, Yeoh, and Pontelli 2017), smart grid (Fioretto et al. 2017) and so on. Incomplete algorithms for DCOPs (Zhang et al. 2005; Maheswaran, Pearce, and Tambe 2006; Farinelli et al. 2008; Ottens, Dimitrakakis, and Faltings 2017) aim to rapidly find solutions at the cost of sacrificing optimality. On the contrary, complete algorithms guarantee the optimal solu- tion and can be generally classified into inference-based and search-based algorithms. DPOP (Petcu and Faltings ∗Corresponding author. 2005b) and Action GDL (Vinyals, Rodriguez-Aguilar, and Cerquides 2009) are typical inference-based complete al- gorithms which employ a dynamic programming paradigm to solve DCOPs. However, they require a linear number of messages of exponential size with respect to the induced width. Accordingly, ODPOP (Petcu and Faltings 2006) and MB-DPOP (Petcu and Faltings 2007) were proposed to trade the message number for smaller memory consumption by propagating the dimension-limited utilities with the corre- sponding contexts iteratively. That is, they iteratively per- form a context-based inference to solve DCOPs optimally when the memory budget is limited. Search-based complete algorithms like SBB (Hirayama and Yokoo 1997), AFB (Gershman, Meisels, and Zivan 2009), PT-FB (Litov and Meisels 2017), ADOPT (Modi et al. 2005) and its variants (Yeoh, Felner, and Koenig 2010; Gutierrez, Meseguer, and Yeoh 2011) perform distributed backtrack searches to exhaust the search space. They have a linear size of messages but an exponential number of mes- sages. Furthermore, these algorithms only use local knowl- edge to update the lower bounds, which exerts a trivial ef- fect on pruning and makes them infeasible for solving large- scale problems. Then, PT-ISABB (Deng et al. 2019), DJAO (Kim and Lesser 2014) and ADPOT-BDP (Atlas, Warner, and Decker 2008) came out to attempt to hybridize search with inference, where an approximated inference is used to construct the initial lower bounds for the search pro- cess. More specifically, PT-ISABB and ADOPT-BDP use ADPOP (Petcu and Faltings 2005a) to establish the lower bounds, while DJAO employs a function filtering technique (Brito and Meseguer 2010) to get them. Here, ADPOP is an approximate version of DPOP by dropping the exceeding dimensions to ensure that each message size is below the memory limit. However, given the limited memory budget, the lower bounds obtained in a one-shot preprocessing phase are still inefficient for pruning since such bounds cannot be tightened by considering the running contexts. That is, the existing hybrid algorithms use only a context-free approxi- mated inference as a one-shot phase to construct the initial lower bounds. In this paper, we investigate the possibility of combining search with context-based inference to solve DCOPs effi- is the assignments to all the variables such that the total cost is minimized. That is, X∗ = arg min fij(xi = di, xj = dj) (cid:88) di∈Di,dj∈Dj fij∈F (a) Constraint graph (b) Constraint matrix (c) Pseudo tree Figure 1: An example of a DCOP and its pseudo tree ciently. Specifically, our main contributions are listed as fol- lows: • We propose a novel complete DCOP algorithm by hy- bridizing search with context-based inference, called HS- CAI where the search adopts a tree-based SBB to find the optimal solution and provides contexts for the inference, while the inference iteratively performs utility propaga- tion for these contexts to construct the tight lower bounds to speed up the search process. • We introduce a context evaluation mechanism to extract the context patterns for the inference from the contexts derived from the search process so as to further reduce the number of context-based inferences. • We theoretically show the completeness of HS-CAI and prove that the lower bounds produced by the context- based inference are at least as tight as the ones estab- lished by the context-free approximated inference under the same memory budget. Moreover, the experimental results demonstrate HS-CAI outperforms state-of-the-art complete DCOP algorithms. Background In this section, we expound the preliminaries including DCOPs, pseudo tree, MB-DPOP and ODPOP. Distributed Constraint Optimization Problems A distributed constraint optimization problem (Modi et al. 2005) can be defined by a tuple (cid:104)A, X, D, F(cid:105) where • A = {a1, a2, . . . , an} is a set of agents. • X = {x1, x2, . . . , xm} is a set of variables. • D = {D1, D2, . . . , Dm} is a set of finite, discrete do- mains. Each variable xi takes a value in Di. • F = {f1, f2, . . . , fq} is a set of constraint functions. Each function fi : Di1 × ··· × Dik → R ≥0 specifies the non- negative cost for each combination of xi1,··· , xik. For the sake of simplicity, we assume that each agent holds exactly one variable (and thus the term agent and vari- able could be used interchangeably) and all constraints are binary (i.e., fij : Di × Dj → R ≥0). A solution to a DCOP A DCOP can be represented by a constraint graph where a vertex denotes a variable and an edge denotes a constraint. Fig. 1 (a) presents a DCOP with five variables and seven constraints. For simplicity, the domain size of each variable is four and all constraints are identical as shown in Fig. 1(b). Pseudo Tree A depth first search (Freuder and Quinn 1985; Dechter, Co- hen, and others 2003) arrangement of a constraint graph is a pseudo tree with the property that different branches are independent, and categorizes its constraints into tree edges and pseudo edges (i.e., non-tree edges). Thus, the neighbors of agent ai can be classified into its parent P (ai), children C(ai), pseudo parents P P (ai) and pseudo children P C(ai) based on their positions in the pseudo tree and the type edges they connect with ai. For clarity, we denote all the (pseudo) parents of ai as AP (ai) = P P (ai)∪{P (ai)} and the set of ancestors who share constraints with ai and its descendants as its separators Sep(ai) (Petcu and Faltings 2005b). Fig. 1(c) presents a possible pseudo tree deriving from Fig. 1(a). MB-DPOP and ODPOP MB-DPOP and ODPOP apply an iterative context-based utility propagation to aggregate the optimal global utility. Specifically, MB-DPOP first uses a cycle-cuts idea (Dechter, Cohen, and others 2003) on a pseudo tree to determine cycle-cut variables and groups these variables into clusters. Within the cluster, MB-DPOP preforms a bounded-memory exploration; anywhere else, the utility propagation from DPOP applies. Specifically, agents in each cluster propa- gate memory-bounded utilities for all the contexts of cycle- cut variables. As for ODPOP, each agent adopts an incre- mental and best-first fashion to propagate the context-based utility. Specifically, an agent repeatedly asks its children for their suggested context-based utilities until it can calculate a suggested utility for its parent during the utility propaga- tion phase. The phase finishes after the root agent receiving enough utilities to determine the optimal global utility. Proposed Method In this section, we present a novel complete DCOP algo- rithm which utilizes both search and inference interleaved. Motivation It can be concluded that search can exploit bounds to prune the solution space but the pruning efficiency is closely re- lated to the tightness of bounds. However, most of search- based complete algorithms can only use local knowledge to compute the initial lower bounds, which leads to inefficient pruning. On the other hand, inference-based complete algo- rithms can aggregate the global utility promptly, but their memory consumption is exponential in the induced width. Therefore, it is natural to combine both search and inference xixj012306198137862942532754 by performing memory-bounded inferences to construct ef- ficient lower bounds for search. Unfortunately, the existing hybrid algorithms only perform an approximated inference to construct one-shot bounds in the preprocessing phase, which would lead to inefficient pruning given a limited memory budget. In fact, the bounds can be tightened by the context-based inference. That is, instead of dropping a set of dimensions, named approximated dimensions, to stay be- low the memory budget, we explicitly consider the running- context assignments to a subset of approximated dimensions (i.e., the context patterns) and compute tight bounds w.r.t. the context patterns. Here, we denote an assigned subset of approximated dimensions as decimated dimensions. More specifically, we aim to combine the advantages of search and context-based inference to optimally solve DCOPs. Different from the existing hybrid methods, we compute tight bounds for the running contexts by perform- ing the context-based inference for the context patterns cho- sen from the contexts derived from the search process. Proposed Algorithm Now, we present our proposed algorithm which consists of a preprocessing phase and a hybrid phase. Preprocessing Phase performs a bottom-up dimension and utility propagation to accumulate the approximated di- mensions and establish the initial lower bounds based on these propagated utilities for search. Accordingly, we em- ploy a tailored version of ADPOP with the limit k to prop- agate the approximated dimensions and incomplete utilities (we omit the pseudo code of this phase due to the limited space). Particularly, during the propagation process, each agent ai selects the dimensions Si of its highest ancestors to approximate to make the dimension size of each out- going utility below k. Then, the approximated dimensions SListi (i.e., the dimensions approximated by ai and its de- scendants) and utility preU tili P (ai) sent from ai to its parent can be computed as follows: SListi = Si ∪ ( SListc i ) (1) ∪ ac∈C(ai) preU tili P (ai) = min Si∪{xi}(localU tili ⊗ ( ⊗ ac∈C(ai) preU tilc i )) (2) i and preU tilc Here, SListc i are the approximated dimen- sions and utility received from its child ac ∈ C(ai), respec- tively. localU tili denotes the combination of the constraints between ai and its (pseudo) parents, i.e., localU tili = ⊗ ai∈AP (ai) fij (3) Taking Fig. 1(c) for example, if we set k = 1, the dimen- sions dropped by a4 are {x1, x2}. Thus, the approximated dimensions and utility sent from a4 to a3 are SList4 3 = {x1, x2} and preU til4 (f41 ⊗ f42 ⊗ f43), re- spectively. Hybrid Phase consists of the search part and context- based inference part. The search part uses a variant of SBB on a pseudo tree (i.e., a simple version of NCBB (Chechetka 3 = min {x1,x2,x4} and Sycara 2006)) to expand any feasible partial assign- ments and provides contexts for the inference part. By using such contexts, the inference part propagates context-based utilities iteratively to produce tight lower bounds for the search part. Traditionally, the context-based inference is implemented by considering the assignments to the approximated dimen- sions (that is, the decimated dimensions are equal to the ap- proximated dimensions). For example, MB-DPOP performs an iterative context-based inference by considering each as- signment combination of cycle-cut variables. However, the approach is not a good choice for our case due to the fol- lowing facts. First, the number of the assignments of cycle- cut variables is exponential, which would incur unacceptable traffic overheads. Moreover, the propagated utilities w.r.t. a specific combination may go out of date since another infer- ence is required as long as the assignments change, which would be very common in the search process. Therefore, it's unnecessary to perform inference for each assignment com- bination of the approximated dimensions. Therefore, we consider to reduce the number of context- based inferences by making the propagated context-based utilities compatible with more contexts. Specifically, for agent ai, we consider the decimated dimensions P Listi ⊆ SListi. Then, the specific assignments to P Listi serve as a context pattern and the propagated utilities will cover all the partial assignments with the same context pattern. That is, ai's context pattern can be defined by: ctxti = {(xj, Cpai(xj))∀xj ∈ P Listi} (4) where Cpai refers to ai's received current partial assign- ment, and Cpai(xj) is the current assignment of xj. Taking agent a3 in Fig. 1(c) as an example, given the limit k = 1, we have SList3 = {x1, x2}. Assume that P List3 = {x1}. Thus, only the assignment of x1 will be considered and x4 (i.e., SList3\P List3) will be still dropped during the context-based inference part. Further, as- sume that ctxt3 = {(x1, 0)}. Then, ctxt3 covers four con- texts ({(x1, 0), (x2, 0)},{(x1, 0), (x2, 1)},{(x1, 0), (x2, 2)} and {(x1, 0), (x2, 3)}). Therefore, a3 only need to perform inference for ctxt3 rather than the four contexts above in this case. Selecting a context pattern is challenging as it offers a trade-off between the tightness of lower bounds and the number of compatible partial assignments. Thus, a good context pattern should comply with the following require- ments. First, the good context pattern should be compati- ble with more contexts so as to avoid unnecessary context- based inference. In other words, these assignments are not likely to change in a short term. Second, the context pattern should result in tight lower bounds. Therefore, we propose a context evaluation mechanism to select the context pat- tern according to the frequency of an assignment of each dimension in the approximated dimensions. In more detail, we consider the context pattern consisting of the assign- ments whose frequency is greater than a specified threshold t. Given t, we have P Listi = {xjCnti((cid:104)xj, Cpai(xj)(cid:105)) > t,∀xj ∈ SListi} where Cnti((cid:104)xj, Cpai(xj)(cid:105)) refers to the frequency of Cpai(xj) for xj. With an appropriate t, the as- After the preprocessing phase (line 1), the root agent starts the context evaluation by initializing evali with true (line 2- 3). Besides, it also starts the search part via CPA messages with its first assignment to its children (line 4-5). Upon re- ceipt of a CPA message, ai first holds the previous Cpai, and stores the received Cpai and evali (line 6). Afterwards, it initializes the lower bounds for each child ac ∈ C(ai) ac- cording to its received utilities (line 7, 23-27). Concretely, the lower bound for ac is established by the context-based utility compatible with Cpai received from ac (line 24-25). Otherwise, the bound is computed by the utility received from ac in the preprocessing phase (line 26-27). Next, ai updates Cnti based on Cpai and the previous one (line 8, 28-32). Specifically, for each dimension in SListi, ai clears its counter if its assignment differs from its previous one (line 29-30). Otherwise, ai increases that counter (line 31- 32). Then, ai finds a context pattern ctxti if the pattern for the context-based inference has not been determined (line 9- 10). After finding one, it allocates ctxti and sends the allo- cated patterns via CTXT messages to its children who need to execute the context-based inference (i.e., inf erChildi) (line 11-15, 33-35). Here, ctxtc i , the allocated pattern for ac, is a partition of ctxti based on ac's approximated dimen- sions SListc i (line 33). When receiving a CTXT message, ai allocates the re- ceived pattern if there is any child who needs to perform the context-based inference (line 17, 33-35). Otherwise, it sends its context-based utility to its parent (line 18-19, 36-40). Here, its context-based utility is computed by the following steps. Firstly, it joins its local utility with the context-based utilities from inf erChildi and the utilities from the other children (line 36). Next, it applies ctxti to fix the values of the partial dimensions in Si so as to improve the complete- ness of the utility (line 38). Finally, it drops the dimensions of the utility to stay below the limit k (line 39). After all the context-based utilities from inf erChildi have arrived, ai sends its context-based utility to its parent if it is not the starter of the context-based inference (line 21-22). Considering a3 in Fig. 1(c), assume that the con- text pattern has not been determined and Cnt3 = {((cid:104)x1, 0(cid:105), 1), ((cid:104)x2, 0(cid:105), 1)}. Given t = 1, we have Cnt3 = {((cid:104)x1, 0(cid:105), 2), ((cid:104)x2, 1(cid:105), 1)} and ctxt3 = {(x1, 0)} after a3 receives a CPA with {(x1, 0), (x2, 1)}. Since the approxi- mated dimensions for its child a4 are {x1, x2}, a3 sends a CTXT message with the context pattern {(x1, 0)} to a4. When receiving the pattern {(x1, 0)}, a4 sends the context-based utility ctxtU til4 (f41(x1 = 0) ⊗ 3 = min {x2,x4} f42 ⊗ f43) to a3. Then, a3 uses ctxtU til4 3 to compute the lower bound for a4 after receiving the CPA message with {(x1, 0), (x2, 2)} or {(x1, 0), (x2, 3)}. Theoretical Results In this section, we first prove the effectiveness of the context- based inference on HS-CAI, and further establish the com- pleteness of HS-CAI. Finally, we give the complexity anal- ysis of the proposed method. signments in the context pattern could not change in a short term. On the other hand, if a partial assignment is hard for pruning, then there would be more assignments included in the context pattern, which guarantees the lower bound tight- ness. In addition, we introduce the variable evali to ensure that the descendants of ai perform an inference only for a context pattern at a time. Once ai has found or received a context pattern, evali is set to false to stop the context evaluation. And when the context pattern is incompatible with Cpai, evali is set to true to indicate that ai can find a new context pattern. Next, we will detail how to implement the context eval- uation mechanism and context-based inference part. Algo- rithm 1 presents the sketch of these procedures. We ignore the search part since it is an analogy to the tree-based branch and bound search algorithm PT-ISABB. Algorithm1:HybridphaseforagentaiWhenInitialize():1performADPOPasapreprocessingphase2ifaiistherootthen3evali←true4di←thefirstelementinDi,Cpa←(xi,di)5sendCPA(Cpa,evali)toac∈C(ai)WhenreceivedCPA(Cpa,eval)fromP(ai):6preCpai←Cpai,Cpai←Cpa,evali←eval7InitBounds()8UpdateSCounter()9ifevalithen10ctxti←{(xj,Cpai(xj))Cnti(hxj,Cpai(xj)i)>t,∀xj∈SListi}11ifctxti>0then12AllocateContext()13evali←flase14ctxti←∅15//markaiasthestarterofthecontext-basedinferencepart16//originalCPAmessagehandlerWhenreceivedCTXT(ctxti)fromP(ai):17AllocateContext()18ifinferChildi=0then19SendCtxtUtil()WhenreceivedCTXTUTIL(ctxtUtilc)fromac∈inferChildi:20ctxtUtilci←ctxtUtilc21ifreceivedallCTXTUTILfrominferChildiandctxti>0then22SendCtxtUtil()FunctionInitBounds():23foreachac∈C(ai)do24ifreceivedaCTXTUTILthatiscompatiblewithCpaifromacthen25lbci(di)←ctxtUtilci(xi=di,Cpai(Sep(ac)))26else27lbci(di)←preUtilci(xi=di,Cpai(Sep(ac)))FunctionUpdateSCounter():28foreachxj∈SListi∧(xj,dj)∈Cpaido29ifpreCpai(xj)6=Cpai(xj)then30Cnti(hxj,Cpai(xj)i)←131else32Cnti(hxj,Cpai(xj)i)←Cnti(hxj,Cpai(xj)i)+1FunctionAllocateContext():33ctxtci←{(xj,dj)(xj,dj)∈ctxti,∀xj∈SListci},∀ac∈C(ai)34inferChildi←{acctxtci>0,∀ac∈C(ai)}35sendCTXT(ctxtci)toac,∀ac∈inferChildiFunctionSendCtxtUtil():36F←localUtili∪{ctxtUtilci∀ac∈inferChildi}∪{preUtilci∀ac∈C(ai)\inferChildi}37S0i←{xj∀xj∈Si∧(xj,dj)/∈ctxti}38join←⊗f∈Ff(ctxti)39ctxtUtili←minS0i∪{xi}join40sendCTXTUTIL(ctxtUtili)toP(ai)Lower Bound Tightness Lemma 1. For a given Cpai, the lower bound lbc i (di) of ac ∈ C(ai) for di produced by the context-based utility (ctxtU tilc i ) is at least as tight as the one estab- lished by the utility (preU tilc i (P A) ≥ i (P A), where P A = Cpai(Sep(ac)) ∪ {(xi, di)}. preU tilc Proof. Directly from the pseudo code, S(cid:48)c, the dimensions dropped by ac in the context-based inference part (line 37), can be defined by: i ). That is, ctxtU tilc S(cid:48)c = {xj(xj, dj) /∈ ctxtc,∀xj ∈ Sc} where ctxtc is the context pattern for ac's context-based in- ference, and Sc is the dimensions dropped by ac in the pre- processing phase. Since ai has received ctxtU tilc i from ac, we have ctxtc > 0 (line 34-35, 20-21). Thus, S(cid:48)c ⊂ Sc is established. Next, we will prove Lemma 1 by induction. Base case. ai's children are leaf agents. For each child i (P A) = ( min localU tilc)(P A) S(cid:48)c∪{xc} ctxtU tilc ac ∈ C(ai), we have (cid:88) (cid:88) = min xc ( xj∈AP (ac )\S(cid:48)c ( xj∈AP (ac )\Sc = min xc (cid:88) + xj∈Sc\S(cid:48)c (cid:88) ≥ min xc ( xj∈AP (ac )\Sc = ( min Sc∪{xc} (cid:88) (cid:88) xj∈S(cid:48)c xj∈Sc (cid:88) fcj (xc, dj ) + fcj (xc, dj ) + fcj (xc, xj )) min xj fcj (xc, xj ) min xj (fcj (xc, dj ) − min xj fcj (xc, xj ))) fcj (xc, dj ) + fcj (xc, xj )) min xj xj∈Sc localU tilc)(P A) = preU tilc i (P A) where dj is the assignment of xj in P A. The equation in the third to the fourth step holds since S(cid:48)c ⊂ Sc. Thus, we have proved the basis. Inductive hypothesis. Assume that the lemma holds for all ai's children. Next, we are going to show the lemma holds for ai as well. For each ac ∈ C(ai), we have ctxtU tilc i (P A) = ( min (localU tilc + S(cid:48)c∪{xc} ctxtU tilc(cid:48) c a c(cid:48) ∈inf erChildc (cid:88) (cid:88) (cid:88) + a + a c(cid:48) ∈C(ac )\inf erChildc (localU tilc + ≥ ( min Sc∪{xc} c(cid:48) ∈C(ac )\inf erChildc (localU tilc + ≥ ( min Sc∪{xc} preU tilc(cid:48) c ))(P A) (cid:88) ctxtU tilc(cid:48) c a c(cid:48) ∈inf erChildc c ))(P A) preU tilc(cid:48) (cid:88) preU tilc(cid:48) c ))(P A) a c(cid:48) ∈C(ac ) = preU tilc i (P A) where inf erChildc are ac's children who need to perform the context-based inference. Thus, Lemma 1 is proved. Correctness Lemma 2. Given the optimal solution X∗, costc(X∗), the cost to the sub-tree rooted at ac ∈ C(ai) is no less than the lower bound lbc lbc i (X∗(xi)). i (X∗(xi)). That is, costc(X∗) ≥ Proof. Since we have proved the lower bounds constructed by the context-based inference part are at least as tight as the ones established by the preprocessing phase in Lemma 1, to prove the lemma, it is sufficient to show that costc(X∗) ≥ ctxtU tilc i (X∗). Next, we will prove Lemma 2 by induction as well. Base case. ai's children are leaf agents. For each child ac ∈ C(ai), we have (cid:88) costc(X∗) = fcj (d∗c , d∗j ) aj∈AP (ac ) = localU tilc(X∗) ≥ ( min S(cid:48)c∪{xc} = ctxtU tilc i (X∗) localU tilc)(X∗) where d∗l is the assignment of xl in X∗, and S(cid:48)c is the di- mensions dropped by ac in the context-based inference part. Thus, the basis is proved. Inductive hypothesis. Assume the lemma holds for all ac ∈ C(ai). Next, we will prove the lemma also holds for ai. For each child ac ∈ C(ai), we have fcj (d∗c , d∗j ) + cost(Spa∗c ) = (cid:88) (cid:88) costc(cid:48) (X∗) aj∈AP (ac ) = localU tilc(X∗) + ≥ localU tilc(X∗) + (cid:88) = (localU tilc + a (cid:88) (cid:88) c(cid:48) ∈C(ac ) a c(cid:48) ∈C(ac ) a c(cid:48) ∈C(ac ) costc(cid:48) (X∗) ctxtU tilc(cid:48) c (X∗) c(cid:48) ∈C(ac ) a (localU tilc + ctxtU tilc(cid:48) c )(X∗) (cid:88) (cid:88) c(cid:48) ∈C(ac ) a a ctxtU tilc(cid:48) c ))(X∗) ctxtU tilc(cid:48) c (localU tilc + (cid:88) c(cid:48) ∈inf erChildc c ))(X∗) preU tilc(cid:48) ≥ ( min S(cid:48)c∪{xc} ≥ ( min S(cid:48)c∪{xc} + a c(cid:48) ∈C(ac )\inf erChildc = ctxtU tilc i (X∗) Thus, the lemma is proved. Theorem 1. HS-CAI is complete. Proof. Immediately from Lemma 2, the optimal solution will not be pruned in HS-CAI. Furthermore, it has been proved that each agent will not receive two identical Cpas in the search part from PT-ISABB (Deng et al. 2019), and the termination of HS-CAI relies on the search part. Thus, HS-CAI is complete. Complexity When it performs a context-based inference, ai needs to store the context-based utilities and utilities received from all the overall space complexity is max) where dmax = maxaj∈Sep(ai) Dj, and O(C(ai)dk k is the maximum dimension limit. Since a CTXTUTIL message only contains a context-based utility, its size is its children. Thus, each experiment, we generate 50 random instances and re- port the average of over all instances. Parameter Tuning Firstly, we aim to examine the effect of different ρ on the performance of HS-CAI to verify the effectiveness of the context evaluation mechanism. Specifically, we consider the DCOPs with 22 agents and the domain size of 3. The graph density varies from 0.2 to 0.6 and ρ varies from 0.05 to 0.65. Here, we do not show the experiment results of ρ greater than 0.65 since the larger ρ leads to the exact same results as ρ with 0.65. Fig. 2 presents the network load of HS-CAI(-M) and HS-CAI with different ρ. The average induced widths in this experiment are 8 ∼ 16. It can be seen from the fig- ure that HS-CAI requires much less network load than HS- CAI(-M). That is because HS-CAI performs inference only for the context patterns selected by the context evaluation mechanism rather than all the contexts as HS-CAI(-M) does. Besides, given the memory budget limit k, it can be ob- served that HS-CAI does not decrease all the time with the increase of ρ. This is due to the fact that increasing ρ which leads to large t can decrease the number of context-based in- ferences but also loose the tightness of lower bounds to some degree. Exactly as mentioned above, the context pattern se- lection offers a trade-off between the tightness of lower bounds and the number of compatible partial assignments. Moreover, it can be seen that the best value of ρ is close to 0.25 in HS-CAI(k = 6) while the one in HS-CAI(k = 10) is near to 0.45. Thus, we choose ρ to 0.25 for HS-CAI(k = 6) and 0.45 for HS-CAI(k = 10) in the following comparison experiments. Performance Comparisons Fig. 3 gives the experimental results under different agent numbers on the sparse configuration where we consider the graph density to 0.25, the domain size to 3 and vary the agent number from 22 to 32. Here, the average induced widths are 9 ∼ 17. It can be seen from Fig. 3(a) and (b) that although the hybrid complete algorithms (e.g., HS-AI and HS-CAI) and PT-FB all use the search strategy to find the optimal so- lution, HS-AI and HS-CAI are superior to PT-FB in terms of the network load and message number. This is because the lower bounds in PT-FB cannot result in effective pruning by only considering the constraints related to the assigned agents. Also, given a fixed k, HS-CAI requires fewer mes- sages than HS-AI since the lower bounds produced by the context-based inference are tighter than the ones established by the context-free approximated inference. Besides, it can be seen that HS-CAI(k = 6) can solve larger problems than HS-AI(k = 6) and the inference-based complete algorithms like DPOP and MB-DPOP, which demonstrates the superior- ity of hybridizing search with context-based inference when the memory budget is relatively low. Although inference requires larger messages than search, it can be observed from Fig. 3(b) that HS-CAI incurs less network load than HS-AI, which indicates that HS-CAI can find the optimal solution with fewer messages owing to the effective pruning and context evaluation mechanism. More- over, we can see from Fig. 3(c) that when solving problems (a) k = 6 (b) k = 10 Figure 2: Network load of varying ρ on different densities max). For a CPA message, it is composed of the assign- O(dk ment of each agent and a context evaluation flag. Thus, the size of a CPA message is O(A). Other messages like CTXT only carry an assignment combination of the approximated dimensions, it only requires O(A) space. The preprocessing phase in HS-CAI only requires A− 1 messages, since only the utility propagation are performed. For the search part and context-based inference part in the hybrid phase, the message number of the search part grows exponentially to the agent number with the same as the search-based complete algorithms. And the message num- ber of the context-based inference part is proportional to the number of the context patterns selected by the context eval- uation mechanism. Empirical Evaluation In this section, we first investigate the effect of the parameter t in the context evaluation mechanism on HS-CAI. Then, we present the experimental comparisons of HS-CAI with state- of-the-art complete DCOP algorithms. Configuration and Metrics We empirically evaluate the performance of HS-CAI and state-of-the-art complete DCOP algorithms including PT- FB, DPOP and MB-DPOP on random DCOPs. Besides, we consider HS-CAI without the context-based inference part as HS-AI and HS-CAI without the context evaluation mech- anism as HS-CAI(-M). Here, HS-AI is actually a variant of PT-ISABB in DCOP settings. All evaluated algorithms are implemented in DCOPSovler1, the DCOP simulator de- veloped by ourselves. Besides, we consider the parameter t in HS-CAI related to both h and dmax, where dmax = maxai∈ADi and h is the height of a pseudo tree. There- fore, we set t = (dmax)ρh. Moreover, we choose k = 6 and k = 10 as the low and high memory budget for MB-DPOP, HS-AI, HS-CAI(-M) and HS-CAI. In our experiments, we use the message number and network load (i.e., the size of total information exchanged) to measure the communication overheads, and the NCLOs (Netzer, Grubshtein, and Meisels 2012) to measure the hardware-independent runtime where the logical operations in the inference and the search are accesses to utilities and constraint checks, respectively. For 1 https://github.com/czy920/DCOPSovler 0.20.30.40.50.6Density105106107108Network Load(KB)(cid:1)(cid:1)(cid:2)(cid:3)(cid:2)(cid:4)(cid:1)(cid:1)(cid:2)(cid:3)(cid:5)(cid:4)(cid:1)(cid:1)(cid:2)(cid:3)(cid:6)(cid:4)(cid:1)(cid:1)(cid:2)(cid:3)(cid:7)(cid:4)(cid:8)(cid:9)0.20.30.40.50.6Density105106107108Network Load(KB)(cid:1)(cid:1)(cid:2)(cid:3)(cid:2)(cid:4)(cid:1)(cid:1)(cid:2)(cid:3)(cid:5)(cid:4)(cid:1)(cid:1)(cid:2)(cid:3)(cid:6)(cid:4)(cid:1)(cid:1)(cid:2)(cid:3)(cid:7)(cid:4)(cid:8)(cid:9)(a) Message Number (b) Network Load (c) NCLOs Figure 3: Performance comparison under different agents on sparse configuration (a) Message Number (b) Network Load (c) NCLOs Figure 4: Performance comparison under different agents on dense configuration with the agent number to 28, HS-CAI(k = 6) requires fewer NCLOs than HS-AI(k = 6) in spite of the exponential com- putation overheads incurred by the iterative inferences. This is because that HS-CAI(k = 6) can provide tight lower bounds to speed up the search so as to greatly reduce the constraint checks when solving the large scale problems un- der the limited memory budget. Besides, we consider the DCOPs with the domain size of 3 and graph density of 0.6 as the dense configuration. The agent number varies from 14 to 24 and the average in- duced widths are 8 ∼ 18. Fig. 4 presents the performance comparison. It can be seen from Fig. 4(a) that DPOP and MB-DPOP cannot solve the problems with the agent num- ber greater than 20 due to the large induced widths. Further- more, since the inference-based complete algorithms have to perform inference on the entire solution space, these al- gorithms require much more NCLOs than the other com- petitors as Fig. 4(c) shows. Additionally, although they both perform the context-based inference, it can be seen from Fig. 4(b) and (c) that HS-CAI exhibits great superiority over MB-DPOP in terms of the network load and NCLOs. That is because HS-CAI only performs inference for the con- text patterns extracted by the context evaluation mechanism, while MB-DPOP needs to iteratively perform inference for all the contexts of cycle-cut variables. As for HS-CAI with different k, it can be seen from Fig. 4(a) and (c) that HS- CAI(k = 10) requires fewer messages but more NCLOs than HS-CAI(k = 6). That is because HS-CAI(k = 10) can produce tighter lower bounds but will incur more com- putation overheads than HS-CAI(k = 6). Conclusion By analyzing the feasibility of hybridizing search and infer- ence, we propose a complete DCOP algorithm, named HS- CAI which combines search with context-based inference for the first time. Different from the existing hybrid complete algorithms, HS-CAI constructs tight lower bounds to speed up the search by executing context-based inference itera- tively. Meanwhile, HS-CAI only needs to perform inference for a part of the contexts obtained from the search process by means of a context evaluation mechanism, which reduces the huge traffic overheads incurred by iterative context-based in- ferences. We theoretically prove that the context-based in- ference can produce tighter lower bounds compared to the context-free approximated inference under the same mem- ory budget. Moreover, the experimental results show that HS-CAI can find the optimal solution faster with less traf- fic overheads than the state-of-the-art. In the future, we will devote to further accelerating the search process by arranging the search space with the infer- ence results. In addition, we will also work for reducing the overheads caused by a context-based utility propagation. Acknowledgments This work is funded by the Chongqing Research Pro- gram of Basic Research and Frontier Technology (No.:cstc2017jcyjAX0030), Fundamental Research Funds for (No.:2019CDXYJSJ0021) and Graduate Research and Innovation Foundation of Chongqing (No.:CYS17023). the Central Universities 222426283032Agent number103104105106107Message Number(cid:1)(cid:2)(cid:3)(cid:2)(cid:4)(cid:5)(cid:6)(cid:1)(cid:2)(cid:3)(cid:2)(cid:7)(cid:1)(cid:8)(cid:9)(cid:10)(cid:4)(cid:5)(cid:6)(cid:1)(cid:2)(cid:3)(cid:2)(cid:7)(cid:1)(cid:8)(cid:11)(cid:12)(cid:10)(cid:2)(cid:13)(cid:6)(cid:14)(cid:5)(cid:15)(cid:16)(cid:6)(cid:17)(cid:18)(cid:7)(cid:1)(cid:8)(cid:9)(cid:10)(cid:15)(cid:16)(cid:6)(cid:19)(cid:17)(cid:18)(cid:7)(cid:1)(cid:8)(cid:9)(cid:10)(cid:15)(cid:16)(cid:6)(cid:17)(cid:18)(cid:7)(cid:1)(cid:8)(cid:11)(cid:12)(cid:10)(cid:15)(cid:16)(cid:6)(cid:19)(cid:17)(cid:18)(cid:7)(cid:1)(cid:8)(cid:11)(cid:12)(cid:10)222426283032Agent number104105106107Network Load (KB)222426283032Agent number105106107108NCLOs141618202224Agent Number103104105106107Message Number(cid:1)(cid:2)(cid:3)(cid:2)(cid:4)(cid:5)(cid:6)(cid:1)(cid:2)(cid:3)(cid:2)(cid:7)(cid:1)(cid:8)(cid:9)(cid:10)(cid:4)(cid:5)(cid:6)(cid:1)(cid:2)(cid:3)(cid:2)(cid:7)(cid:1)(cid:8)(cid:11)(cid:12)(cid:10)(cid:2)(cid:13)(cid:6)(cid:14)(cid:5)(cid:15)(cid:16)(cid:6)(cid:17)(cid:18)(cid:7)(cid:1)(cid:8)(cid:9)(cid:10)(cid:15)(cid:16)(cid:6)(cid:19)(cid:17)(cid:18)(cid:7)(cid:1)(cid:8)(cid:9)(cid:10)(cid:15)(cid:16)(cid:6)(cid:17)(cid:18)(cid:7)(cid:1)(cid:8)(cid:11)(cid:12)(cid:10)(cid:15)(cid:16)(cid:6)(cid:19)(cid:17)(cid:18)(cid:7)(cid:1)(cid:8)(cid:11)(cid:12)(cid:10)141618202224Agent Number104105106107Network Load (KB)141618202224Agent Number105106107108NCLOsReferences [Atlas, Warner, and Decker 2008] Atlas, J.; Warner, M.; and Decker, K. 2008. A memory bounded hybrid approach to distributed constraint optimization. In Proceedings 10th In- ternational Workshop on DCR, 37 -- 51. [Brito and Meseguer 2010] Brito, I., and Meseguer, P. 2010. Improving DPOP with function filtering. In Proceedings of the 9th AAMAS, 141 -- 148. [Chechetka and Sycara 2006] Chechetka, A., and Sycara, K. 2006. No-commitment branch and bound search for dis- tributed constraint optimization. In Proceedings of the 5th AAMAS, 1427 -- 1429. [Dechter, Cohen, and others 2003] Dechter, R.; Cohen, D.; et al. 2003. Constraint processing. Morgan Kaufmann. [Deng et al. 2019] Deng, Y.; Chen, Z.; Chen, D.; Jiang, X.; and Li, Q. 2019. PT-ISABB: A hybrid tree-based com- plete algorithm to solve asymmetric distributed constraint optimization problems. In Proceedings of the 18th AAMAS, 1506 -- 1514. [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 AAMAS, volume 2, 639 -- 646. [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. [Fioretto et al. 2017] Fioretto, F.; Yeoh, W.; Pontelli, E.; Ma, Y.; and Ranade, S. J. 2017. A distributed constraint opti- mization (DCOP) approach to the economic dispatch with demand response. In Proceedings of the 16th AAMAS, 999 -- 1007. [Fioretto, Pontelli, and Yeoh 2018] Fioretto, F.; Pontelli, E.; and Yeoh, W. 2018. Distributed constraint optimization problems and applications: A survey. Journal of Artificial Intelligence Research 61:623 -- 698. [Fioretto, Yeoh, and Pontelli 2017] Fioretto, F.; Yeoh, W.; and Pontelli, E. 2017. A multiagent system approach to scheduling devices in smart homes. In Proceedings of the 16th AAMAS, 981 -- 989. [Freuder and Quinn 1985] Freuder, E. C., and Quinn, M. J. 1985. Taking advantage of stable sets of variables in con- straint satisfaction problems. In Proceedings of the 9th IJ- CAI, volume 85, 1076 -- 1078. [Gershman, Meisels, and Zivan 2009] Gershman, A.; Meisels, A.; and Zivan, R. 2009. Asynchronous forward bounding for distributed COPs. Journal of Artificial Intelligence Research 34:61 -- 88. [Gutierrez, Meseguer, and Yeoh 2011] Gutierrez, P.; Meseguer, P.; and Yeoh, W. 2011. Generalizing ADOPT and BnB-ADOPT. In Proceedings of the 22nd IJCAI, 554 -- 559. [Hirayama and Yokoo 1997] Hirayama, K., and Yokoo, M. 1997. Distributed partial constraint satisfaction problem. In International Conference on Principles and Practice of Constraint Programming, 222 -- 236. [Kim and Lesser 2014] Kim, Y., and Lesser, V. 2014. DJAO: a communication-constrained DCOP algorithm that com- bines features of ADOPT and Action-GDL. In Proceedings of the 28th AAAI, 2680 -- 2687. [Litov and Meisels 2017] Litov, O., and Meisels, A. 2017. Forward bounding on pseudo-trees for DCOPs and AD- COPs. Artificial Intelligence 252:83 -- 99. [Maheswaran et al. 2004] Maheswaran, R. T.; Tambe, M.; Bowring, E.; Pearce, J. P.; and Varakantham, P. 2004. Tak- ing DCOP to the real world: Efficient complete solutions for distributed multi-event scheduling. In Proceedings of the 3rd AAMAS, volume 1, 310 -- 317. [Maheswaran, Pearce, and Tambe 2006] Maheswaran, R. T.; Pearce, J. P.; and Tambe, M. 2006. A family of graphical- game-based algorithms for distributed constraint optimiza- tion problems. In Coordination of large-scale multiagent systems. Springer. 127 -- 146. [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. [Netzer, Grubshtein, and Meisels 2012] Netzer, A.; Grub- shtein, A.; and Meisels, A. 2012. Concurrent forward bounding for distributed constraint optimization problems. Artificial Intelligence 193:186 -- 216. [Ottens, Dimitrakakis, and Faltings 2017] Ottens, B.; Dimi- trakakis, C.; and Faltings, B. 2017. DUCT: An upper con- fidence bound approach to distributed constraint optimiza- tion problems. ACM Transactions on Intelligent Systems and Technologyc 8(5):69. [Petcu and Faltings 2005a] Petcu, A., and Faltings, B. 2005a. Approximations in distributed optimization. In Interna- tional Conference on Principles and Practice of Constraint Programming, 802 -- 806. [Petcu and Faltings 2005b] Petcu, A., 2005b. optimization. In Proceedings of the 19th IJCAI, 266 -- 271. [Petcu and Faltings 2006] Petcu, A., and Faltings, B. 2006. ODPOP: An algorithm for open/distributed constraint opti- mization. In Proceedings of the 21st AAAI, 703 -- 708. [Petcu and Faltings 2007] Petcu, A., and Faltings, B. 2007. MB-DPOP: A new memory-bounded algorithm for dis- tributed optimization. In Proceedings of the 20th IJCAI, 1452 -- 1457. [Vinyals, Rodriguez-Aguilar, and Cerquides 2009] Vinyals, M.; Rodriguez-Aguilar, J. A.; and Cerquides, J. 2009. Generalizing DPOP: DPOP, a new complete algorithm for DCOPs. In Proceedings of the 8th AAMAS, 1239 -- 1240. [Yeoh, Felner, and Koenig 2010] Yeoh, W.; Felner, A.; and Koenig, S. 2010. BnB-ADOPT: An asynchronous branch- and-bound DCOP algorithm. Journal of Artificial Intelli- gence Research 38:85 -- 133. [Zhang et al. 2005] Zhang, W.; Wang, G.; Xing, Z.; and Wit- tenburg, L. 2005. Distributed stochastic search and dis- and Faltings, B. A scalable method for multiagent constraint tributed breakout: properties, comparison and applications to constraint optimization problems in sensor networks. Ar- tificial Intelligence 161(1-2):55 -- 87.
1801.00464	2	1801	2018-01-25T14:08:35	Comparative Analysis of Human Movement Prediction: Space Syntax and Inverse Reinforcement Learning	[ "cs.MA" ]	Space syntax matrix has been the main approach for human movement prediction in the urban environment. An alternative, relatively new methodology is an agent-based pedestrian model constructed using machine learning techniques. Even though both approaches have been studied intensively, the quantitative comparison between them has not been conducted. In this paper, comparative analysis of space syntax metrics and maximum entropy inverse reinforcement learning (MEIRL) is performed. The experimental result on trajectory data of artificially generated pedestrian agents shows that MEIRL outperforms space syntax matrix. The possibilities for combining two methods are drawn out as conclusions, and the relative challenges with the data collection are highlighted.	cs.MA	cs	Comparative Analysis of Human Movement Prediction: Space Syntax and Inverse Reinforcement Learning Soma Suzuki Center for Advanced Spatial Analysis University College London Abstract Space syntax matrix has been the main approach for human movement prediction in the urban environment. An alternative, relatively new methodology is an agent- based pedestrian model constructed using machine learning techniques. Even though both approaches have been studied intensively, the quantitative comparison between them has not been conducted. In this paper, comparative analysis of space syntax metrics and maximum entropy inverse reinforcement is performed. The experimental result on trajectory data of artificially generated pedestrian agents shows that MEIRL outperforms space syntax matrix. The possibilities for combining two methods are drawn out as conclusions, and the relative challenges with the data collection are highlighted. learning (MEIRL) I. Introduction The problem of predicting human movement in urban environments features prominently in urban planning as the acquired model is considered as the artificial laboratory that enables us to test ideas and hypothesis that are not easy to explore in the real world (Torrens 2015). Historical approach to infer pedestrian movement is a methodology developed in space syntax community, in which city is represented as a connectivity graph so as to explain human movement rate in each space (Hillier 1993). Even though it turned out that some defined metrics are correlated to human movement rates in the city, it is most unlikely that humans move merely based on the number of links each space has. An alternative promising approach to predict pedestrian movement is to generate human-like agent through learning humans' preferences on road choice given observational trajectory data. This approach is known as Inverse Reinforcement Learning (IRL) in Markov Decision Process (MDP) in machine learning community (Ng and Russell 2000). IRL has been already studied intensively for predicting human movement (Ziebart the valid However, comparison between IRL and space syntax matrix has not been performed. This paper conducts a comparative analysis of two methods for predicting human movement in the urban environment. In section II, space syntax approach for modeling human movement is explained. Next, section III introduces Maximum Entropy Inverse Reinforcement Learning (MEIRL) to generate human-like agents. Then, the quantitative comparison between space syntax matrix and MEIRL is conducted in section IV. Finally, 2008). the possibility of combining space syntax matrix and MEIRL and the future work is discussed as conclusion in section V. II. Space Syntax Metrics First developed by Hillier in University College London (Hillier 1984), space syntax has grown into an independent research area. Primarily, space syntax is a methodology for investigating spatial complexities in an attempt to identify its particular structure that resides at the level of the entire configuration. The methodology is based on the theory that the form-function relation in cities reflects the structural properties of its configuration (Hillier 1998). Through enormous empirical study, it turns out that space syntax metrics can capture human movement (Penn 1998) and has been widely adopted on human movement prediction. In space syntax approach, the city is first represented as connectivity graph, consisting of nodes representing spaces and edges if the corresponding spaces are intersected as shown in Figure 1. Then, individual spaces are ranked based on space syntax matrix for inferring human movement rates. Figure 1: A fictive city (a) and connectivity graph (b) Among various space syntax metrics, researchers in space syntax community claim that pedestrian movement is most well- described by local integration. As illustrated below, local integration is a normalized closeness centrality initially developed in social network analysis (Scott 2000). Space Syntax Metrics: Local Integration Local integration is based on the concept of depth, which assesses how far it is from a node to other nodes within 2 steps. For any particular node in the connectivity graph, the distance (or step) between point and point j denotes the length of the shortest path linking them, if any; otherwise = ∞. For all points, the following set of axioms holds: = 0 only if i = j , with dij dij • • • dij dij ≥ 0 dij = dji dij + djk ≥ dik To obtain local integration, closeness centrality of a node needs to be calculated first. Closeness centrality measures the centrality of a node in a network, which is obtained as the reciprocal of the sum of the number of the steps to other nodes. Thus closeness centrality of node i within two steps is calculated as follows: Ci = 1 ∑k j dij (i f dij ≤ 2) where k represents the number of nodes in the connectivity graph. For instance, node 1 has two nodes in step one and 1 node in step two, thus the sum 1 / (2 × 1 + 1 × 2) indicates the closeness centrality of node 1 considering two steps. Then, the reciprocal of closeness centrality divided by the number of nodes involved minus one leads to local mean depth MD. Finally, by standardizing the variation of mean depth MD between zero and one, relative asymmetry RA is obtained as the local integration of a node (Hillier 1984). MDi = Ci k − 1 2(MDi − 1) R Ai = k − 2 Thereafter individual space is ranked according to the relative asymmetry RA to predict pedestrian movement rate in each space. Even though why the closeness centrality or local integration can be adopted to infer human movement is not clearly justified, a vast amount of empirical data proves that local integration works very well for predicting pedestrian movement in the city. III. Maximum Entropy IRL The alternative approach to predict pedestrian movement is to learn human preference and generate human-like agent. In this section, Maximum Entropy Inverse Reinforcement Learning (MEIRL) is introduced as the methodology for learning pedestrians' preferences. Inverse Reinforcement Learning (IRL) is a problem in Markov Decision Process (MDP), that is, the problem of extracting a reward function given observed, optimal behavior (Ng and Russell 2000). Markov Decision Process A finite MDP is a tuple (S, A, where , γ, R ), Psa S is a finite set of N states { }. sn A is a set of K actions { a1, a2, …, }. ak s1, s2, …, is state transition probabilities upon Psa taking action a in state s. γ ∈ R : 𝕊 × 𝔸 → ℝ that depends on state and action. [0,1] is the discount factor. is the reward function π* π* : a1, a2, 𝕊 → 𝔸 sn Psa [0,1] and a policy π : The classical problem of MDP is to find the optimal policy such that expected reward is maximized. Inverse Reinforcement Learning Conversely, Inverse Reinforcement Learning (IRL) problem consists of finding the reward function from an observed policy. More specifically, given a finite space S = { s1, s2, …, }, set of actions A = { …, }, ak , a discount factor transition probabilities , the goal of γ ∈ 𝕊 → 𝔸 IRL is to find a reward function such that π is an optimal policy . The acquired reward function explains the intrinsic preference of policy demonstrator, or expert agents. Therefore, learning reward function through IRL enables us to generate artificial agents that behave in such a way that pedestrians "would" do in the city. Training data for IRL is a set of m independent trajectories sampled from the expert's (pedestrians) policy Learning optimal policy from experts' demonstrations is called imitation learning. In the imitation learning setting, the targeted reward function is approximated based on path feature counts, , which is the Fζ = ∑ Sj∈ζ sum of the state features along the path. Thus, that linearly maps given the reward weight θ each state feature to a state reward value, the total reward value of a trajectory is: D = {ζ1, ζ2, . . . , ζm} . πE Fsj Fsj Rθ(Fζ) = θTFζ = ∑ sj∈ζ θTFsj The problem of imitation learning, therefore, consists of approximating the reward weight by matching feature counts between expert θ agent's policy and the learner agent's policy. Maximum Entropy IRL To obtain the distribution of path that performs one-to-one correspondence to feature counts, the principle of maximum entropy (Jaynes 1957) is employed. In other words, the adopted distribution of path does not exhibit any additional preferences for some paths beyond feature counts (Ziebart 2008). The probability of a path to be ζi chosen by an agent under reward weight is θ given as follows: P(ζiθ) = 1 Z(θ) = m ∑ ζ∈D Z(θ) exp(θTFζi) exp(θTFζ) θ * Z(θ) is a where the partition function normalized constant. Estimating optimal reward weight implies maximizing the likelihood of the observed data under the D maximum entropy probability distribution illustrated above. The optimal reward weight is then obtained using the gradient of the θ * (log) likelihood. θ* = arg max θ L(θ) L(θ) = 1 M = 1 M = 1 M m ∑ ζ∈D m ∑ ζ∈D m ∑ ζ∈D logP(ζ θ) log 1 Z(θ) exp(θTFζ) θTFζ − logZ(θ) = 1 M m ∑ ζ∈D θTFζ − log m ∑ ζ∈D exp(θTFζ) m Fζ Thus, the gradient of the log-likelihood is ∇θ L(θ) = 1 ∑ M ζ∈D ∑m ζ∈D exp(θTFζ) ∑m ζ∈D exp(θTFζ) m dθTFζ dθ − m P(ζ θ)Fζ ∑ ζ∈D P(sθ)Fs = 1 ∑ Fζ − M ζ∈D = F − ∑ F s∈ζ where the is the expected state feature count of an expert agent. Therefore, the gradient is obtained as the difference between the expected state feature counts of expert agents and the expected state feature counts of learner agents, which can be expressed in terms of expected state visitation frequencies P(sθ) . A pseudocode is presented below. Pseudocode Maximum Entropy IRL θ ← pr ng() and gather D for n = 1 : N do Rθ ← RewardUpdate(θ) πθ ← PolicyUpdate(Rθ) P(sθ) ← FrequencyCompute(πθ) ∇θ L(θ) ← GradCompute(P(sθ)) θ ← ParamUpdate(∇θ L(θ)) selected area. The brief process flow of API is shown in Figure 2. First, geographic data on OSM is downloaded in the selected area by using R package Osmar as illustrated in Figure 3. Then, as only the street data is needed for the trace generation, the street network is extracted from the downloaded geographical data as shown in Figure 4. It is done by selecting data that is tagged as "highway" which is the main key used for identifying any kind of road, street or path on OpenStreetMap. Next, the highway-osmar object is converted into a graph object in order to compute the shortest path. In the graph representation, each node represents an intersection and edge corresponds to a road as shown in Figure 5. Finally, using R package Igraph, the shortest path between randomly specified starting nodes and goal nodes are computed based on Dijkstra's shortest path algorithm. 400 trajectories dataset are generated with the random start and goal point as illustrated in Figure 6: 300 training set and 100 testing set. The constructed connectivity graph is also served to define Markov Decision Process for MEIRL methodology. With the optimal reward function parameter , generated agents are assumed to possess θ* the preference similar to real pedestrians. Therefore, the state visitation frequencies under the optimal reward function P(sθ) parameter is used to rank individual spaces for human movement prediction. θ IV. Experiment While In this section, the comparison between space syntax matrix and MEIRL is conducted given the trajectory dataset generated by using an open source vendor Open Street Map (OSM). Dataset To acquire plausible model, the vast amount of pedestrian trajectory, or training data is essential. In this experiment, the artificially generated pedestrians are adopted as collecting dataset pertaining to individual walking trajectory is very challenging (Torrens 2011). the current pervasiveness of mobile devices enables the collection of mobile location data at the unprecedented scale and granularity, the challenge invariably remains in associating movement path with the reason "why" people moved as the recorded trajectory. To illustrate, the GPS data of those who are commuting to her office and those who are on the way to a nightclub cannot be used together as the training data to model pedestrian movement: the former will most likely choose the shortest path and the latter will likely be a random walk. With the purpose of facilitating machine learning approach toward human movement prediction in the city, the API is developed so as to be a pipeline for generating ready-to-be-used training dataset. Assuming that people walking along the shortest route are moving in order merely to reach the goal point, this API generates the shortest path from an arbitrary start point to a goal point in the Figure 3: Selecting area in Open Street Map Figure 4: Extracted street network Figure 6: Generated training data. Different colors are introduced only for distinguishing each path. Finally, MEIRL is conducted. The evolution of feature counts difference during the training is presented in Figure 7. In the experiment, the learning rate is set to a large value and gradually reduce as optimization progress. MEIRL performed well, showing quick convergence. Thereafter, the correlation between space syntax matrix and the test data and the correlation between the state visitation frequencies by MEIRL and the test data are compared. Difference between the expected state feature counts e c n e r e f f i d s t n u o c e r u t a e F Figure 5: Connectivity graph Figure 7: Evolution of feature counts difference As a result, MEIRL showed stronger correlation 0.49 than 0.45 of space syntax matrix, meaning state visitation frequencies by MEIRL captures human movement in the city better than space syntax matrix. V. Conclusion The MEIRL approach successfully performed more accurate prediction than space syntax matrix. However, I believe space syntax will be invariably useful in human movement prediction, employed together with presented agent-based approach. A major challenge in applying reinforcement learning in multi- agent model is how to manage the explosive computational cost as the state-action space grows exponentially with the number of agents and the learning becomes prohibitively slow. To alleviate this drawback, as known as structural reinforcement learning it is possible to reduce the size of the state-action space by supplying the model with the partial, but fundamental pedestrian movement in the form of space syntax graph. Future work may involve employing actual raw GPS data with the development of pipeline where the pedestrian trace data is plausibly associated with the label that explains the motivation of the corresponding movement. in virtual Reference Torrens, P. M. (2015) Slipstreaming human geosimulation geographic environments. Ann. GIS 2015, 325–344. Hillier, B., & Penn, A. & Hanson, J. & Grajewski, T. & Xu, J. (1993) Natural movement: configuration and attraction in urban pedestrian movement. Environment and Planning B, 20, 29-66. Ng, A. Y., & Russell, S. J. (2000) Algorithms for inverse reinforcement learning. Icml, 663– 670. Ziebart, B. D., & Maas, A. L., & Bagnell, J. A., & Dey, A. K. (2008) Maximum entropy learning. inverse reinforcement AAAI Conference on Artificial Intelligence. Hillier, B., & Hanson, J. (1984) The Social Logic of Space. Cambridge University Press. Hillier, B. (1998) A note on the intuiting of forms: Three issues in the theory of design. Environment and Planning B: Planning and Design (25th anniversary issue), 37-40. Penn, A., & Hillier, B., & Banister, D., & Xu, J. (1998), Configurational modeling of urban movement networks, Environment and Planning B: planning and design, 25, 59-84. Scott, J. (2000), Social Network Analysis: a handbook,2, Sage Publications Ltd: London. Ng, A. Y., & Russell, S. J. (2000) Algorithms for inverse reinforcement learning. Icml, 663– 670. Lin, Z., & Feygin, S. A. (2016) Simulating Human-like Navigation in Urban Transportation Environments with Multi- agent Deep (Inverse) Reinforcement Learning. Torrens, P., & Li, X., & Griffin, W. A. (2011) Building agent-based walking models by machine-learning on diverse databases of space-time trajectory samples. Trans. Geogr. Inf. Sci. 15, 67–94.
1708.09630	4	1708	2018-04-09T02:25:13	Resilient Autonomous Control of Distributed Multi-agent Systems in Contested Environments	[ "cs.MA", "cs.LG", "eess.SY" ]	An autonomous and resilient controller is proposed for leader-follower multi-agent systems under uncertainties and cyber-physical attacks. The leader is assumed non-autonomous with a nonzero control input, which allows changing the team behavior or mission in response to environmental changes. A resilient learning-based control protocol is presented to find optimal solutions to the synchronization problem in the presence of attacks and system dynamic uncertainties. An observer-based distributed H_infinity controller is first designed to prevent propagating the effects of attacks on sensors and actuators throughout the network, as well as to attenuate the effect of these attacks on the compromised agent itself. Non-homogeneous game algebraic Riccati equations are derived to solve the H_infinity optimal synchronization problem and off-policy reinforcement learning is utilized to learn their solution without requiring any knowledge of the agent's dynamics. A trust-confidence based distributed control protocol is then proposed to mitigate attacks that hijack the entire node and attacks on communication links. A confidence value is defined for each agent based solely on its local evidence. The proposed resilient reinforcement learning algorithm employs the confidence value of each agent to indicate the trustworthiness of its own information and broadcast it to its neighbors to put weights on the data they receive from it during and after learning. If the confidence value of an agent is low, it employs a trust mechanism to identify compromised agents and remove the data it receives from them from the learning process. Simulation results are provided to show the effectiveness of the proposed approach.	cs.MA	cs	Resilient Autonomous Control of Distributed Multi-agent Systems in Contested Environments Rohollah Moghadam, Student Member, IEEE, Hamidreza Modares, Member, IEEE 1 8 1 0 2 r p A 9 ] A M . s c [ 4 v 0 3 6 9 0 . 8 0 7 1 : v i X r a Abstract-An autonomous and resilient controller is proposed for leader-follower multi-agent systems under uncertainties and cyber-physical attacks. The leader is assumed non-autonomous with a nonzero control input, which allows changing the team behavior or mission in response to environmental changes. A resilient learning-based control protocol is presented to find optimal solutions to the synchronization problem in the presence of attacks and system dynamic uncertainties. An observer-based distributed H∞ controller is first designed to prevent propagating the effects of attacks on sensors and actuators throughout the network, as well as to attenuate the effect of these attacks on the compromised agent itself. Non-homogeneous game algebraic Riccati equations are derived to solve the H∞ optimal synchro- nization problem and off-policy reinforcement learning is utilized to learn their solution without requiring any knowledge of the agent's dynamics. A trust-confidence based distributed control protocol is then proposed to mitigate attacks that hijack the entire node and attacks on communication links. A confidence value is defined for each agent based solely on its local evidence. The proposed resilient reinforcement learning algorithm employs the confidence value of each agent to indicate the trustworthiness of its own information and broadcast it to its neighbors to put weights on the data they receive from it during and after learning. If the confidence value of an agent is low, it employs a trust mechanism to identify compromised agents and remove the data it receives from them from the learning process. Simulation results are provided to show the effectiveness of the proposed approach. Index Terms-Autonomous Controller, Resilient Controller, Reinforcement Learning, Multi-agent System, Distributed Con- trol, H∞ control. I. Introduction The term autonomous is a combination of the Greek words auto (self) and nomous (law, rule) [1]. In system theory, an autonomous system is a system that is self-governing and does not explicitly depend on the independent variable. If the independent variable is time, these systems are also called time-invariant systems. On the other hand, in control theory, an autonomous control is a self-governing control system in the sense that it acts independently and does not rely on prior knowledge of the system dynamics and human intervention. An autonomous controller should be able to learn from what it perceives to compensate for partial or incorrect prior knowledge. Autonomous control design for a multi-agent system (MAS) has gained significant interest due to applications in a variety Rohollah Moghadam and Hamidreza Modares are with the Department of Electrical and Computer Engineering, Missouri University of Science and Technology, Emerson Hall, 301 W 16th St, Rolla, MO 65409 USA (e-mails: [email protected], [email protected]). Manuscript received December 30, 2017; of disciplines from robot swarms to power systems and wire- less sensor networks. In a distributed MAS, decisions are made locally using only agents' available information. This provides scalability, flexibility and avoids a single point of failure [2]– [5]. However, designing autonomous controllers for MASs is challenging and requires learning from experience. Moreover, distributed MASs are prone to cyber-physical attacks. Due to their networked nature, attacks can escalate into disastrous consequences and significantly degrade the performance of the entire network [6]. In a contested environment with adversarial inputs, corrupted data communicated by a single compromised agent can be propagated to the entire network through its neighbors. This corrupted data will be used by autonomous agents for learning which misleads the entire network and, consequently, causes no emergent behavior or an emergent misbehavior. The main bottleneck in deploying successful distributed MASs is designing secure control protocols that can learn about system uncertainties while showing some level of functionality in the presence of cyber-physical attacks. Reinforcement learning (RL) [7]–[9], inspired by learning mechanisms observed in mammals, has been successfully used to learn optimal solutions online in single agent systems for both regulation and tracking control problems [10]–[16] and recently for MASs [17]–[19]. Existing RL-based controllers for leader-follower MASs assume that the leader is passive and without any control input. In this case, the leader is not able to react to environmental or mission changes by replanning its trajectories. On the other hand, existing active leader controllers (e.g., [20]–[22]) are not autonomous as they require having complete knowledge of the leader and agent's dynamics. Moreover, these approaches are generally far from optimal and only take into account the stability, which is the bare minimum requirement. Finally, existing learning-based RL solutions to MASs are not resilient against cyber-physical attacks. Resilient control protocols for MASs have been designed in the literature [23]–[35] to mitigate attacks. Most of the existing approaches either use the discrepancy among the state of agents and their neighbors to detect and mitigate attacks, or use an exact model of agents to predict expected normal behavior and, thus, detect an abnormality caused by attacks. However, we will show that a stealthy attack on one agent can cause an emergent misbehavior in the network with no discrepancy between agent's states and, therefore, the former approaches cannot mitigate these type of attacks. Moreover, this discrepancy could be a result of a legitimate change in the state of the leader. Blindly rejecting the agent's neighbor information can harm the network connectivity and the convergence of the network. On the other hand, model- based approaches require a complete knowledge of the agent's dynamics, which may not be available in many practical applications and avoid the design of autonomous controllers. H∞ control protocols have also been proposed to attenuate the effect of disturbances in MASs [36]–[38]. However, as shown in this paper, standard H∞ control protocols can be misled and become entirely ineffective by a stealthy attack. To the author's knowledge, designing an autonomous and resilient controller that does not require any knowledge of the agent's dynamic and can survive against cyber-physical attacks has not been investigated yet. This paper presents an autonomous and resilient dis- tributed control protocol for leader-follower MASs with a non- autonomous leader. To alleviate the effects of attacks on the MASs, a distributed observer-based control protocol is first developed to prevent corrupted sensory data caused by attacks on sensors and actuators from propagating across the network. To this end, only the leader communicates its actual sensory information and other agents estimate the leaders' state using a distributed observer. To further improve resiliency, distributed H∞ control protocols are designed to attenuate the effect of the attacks on the compromised agent itself. Non-homogeneous game algebraic Riccati equations (ARE) are derived for solv- ing the optimal H∞ synchronization problem for each agent. An off-policy RL algorithm is developed to learn the solutions of the non-homogeneous game ARE without requiring the complete knowledge of the agent's dynamics. To avoid the usage of corrupted data coming from compromised neighbors during and after learning, a trust-confidence based control protocol is developed for attacks on communication links and attacks that hijack the entire node. A confidence value is defined for each agent based solely on its local evidence. Then, each agent communicates its confidence value with neighbors to indicate the trustworthiness of its own informa- tion. Moreover, a trust value is defined for each neighbor to determine the significance of the incoming information. The agent incorporates these trust values along with the confidence values received from neighbors in its update law to eventually isolate the compromised agent. II. Preliminary (cid:105) (cid:104) ai j matrix, with di =(cid:80) In this section, a background of the graph theory is pro- vided. A directed graph G consists of a pair (V,E) in which V = {v1,··· , vN} is a set of nodes and E ⊆ V × V is a set of edges. The adjacency matrix is defined as A = , with ai j > 0 if (v j, vi) ∈ E, and ai j = 0 otherwise. The set of nodes vi with edges incoming to node v j is called the neighbors of node vi, namely Ni = {v j : (v j, vi) ∈ E}. The graph Laplacian matrix is defined as L = D − A, where D = diag(di) is the in-degree j∈Ni ai j as the weighted in-degree of node vi. A (directed) tree is a connected digraph that in-degree of every node is one, except the root node. A directed graph has a spanning tree if there exists a directed tree that connects all nodes of the graph. A leader can be pinned to multiple nodes, resulting in a diagonal pinning matrix G = diag (bi) ∈ RN×N with the pinning gain bi > 0 when the node has access to the 2 leader node and bi = 0, otherwise. 1N is the N-vector of ones and Im(R) denotes the range space of R. λmin(A) denotes the minimum eigenvalue of matrix A. (cid:107).(cid:107) denotes the Euclidean norm for vectors or the induced 2-norm for matrices. The notation A ⊗ B is Kronecker product of matrices A and B. Assumption 1. The communication graph has a spanning tree, and the leader is pinned to at least one root node. III. Standard Synchronization Control Protocols and Their Vulnerability to attacks In this section, the standard synchronization control protocol for MASs is reviewed and its vulnerability to attacks is examined. Consider N agents with identical dynamics given by xi = Axi + Bui + Dωi (1) where xi(t) ∈ Rn and ui(t) ∈ Rm are the state and control input of agent i, respectively. ωi(t) ∈ Rd denotes the attack signal injected into agent i. A, B, and D are the drift, input, and attack dynamics, respectively. Assumption 2. The pair (A, B) is stabilizable. Let the leader dynamics be non-autonomous, i.e., the control input of the leader is a nonzero signal, and is given by (2) where ζ0(t) ∈ Rn and v0(t) ∈ Rm denote the state and input of the leader, respectively. A and B are the same as other agents. ζ0 = Aζ0 + Bv0 Assumption 3. The control input v0 is given and bounded, i.e., there exists a positive constant vm such that (cid:107)v0(cid:107) ≤ vm. Define the tracking error for agent i as (3) Define the local neighborhood tracking error ei ∈ Rn for εi = xi − ζ0 agent i as [39] ei = ai j(x j − xi) + bi(ζ0 − xi) (4) (cid:88) j∈Ni where bi (cid:62) 0 is the pinning gain, and bi > 0 for at least one root node i. The standard distributed tracking control protocol is then given by [40] ui = c1Kei + c2h(Kei) (5) where c1 and c2 are positive scalar coupling gains, and K is a design matrix gain. h(.) is a nonlinear function defined for x ∈ Rn such that  x(cid:107)x(cid:107) , 0, h(x) = i f (cid:107)x(cid:107) (cid:44) 0 i f (cid:107)x(cid:107) = 0 It can be seen that the following condition is required to assure synchronization (6) (7) εi → 0 ⇔ xi → ζ0 Theorem 1. [40] Consider the agent dynamics (1)-(2) with ωi(t) = 0. Suppose Assumptions 1 and 3 hold. Then, agents synchronize to the leader, i.e., εi → 0 ∀i ∈ N under the controller (5) with c1 ≥ (1/λmin(L + G)), c2 ≥ vm and K = −BT P−1, where λmin is the minimum eigenvalue of (L+G) and P > 0 is a solution to the linear matrix inequality AP + PAT − 2BBT < 0. Definition 1. In a graph, agent i is reachable from agent j if there exists a directed path of any length from node j to node i. Definition 2: An agent is called a disrupted/compromised agent, if it is directly under attack. Otherwise, it is called an intact agent. It is shown in [40] for the proof of Theorem 1 that in the absence of attack, if the controller is designed to make the local neighborhood tracking error for all agents go to zero, the synchronization is guaranteed. In the following theorem, however, it is shown that even though the local neighborhood tracking (4) goes to zero for all agents, in the presence of a specific designed attack, it does not guarantee synchronization for intact agents that have a path to a compromised agent. Note that the leader is assumed to be a trusted agent with more advanced sensors and with higher security. Note also that the leader does not receive any information from other agents, which makes it secure against attacks on other agents and the communication network. Lemma 1 [41]. Let Σ be a diagonal matrix with at least one nonzero positive element, and L be the Laplacian matrix. Then, (L + Σ) is a nonsingular M-matrix. Theorem 2. Consider the MAS (1)-(2) with the control pro- tocol (5). Assume that agent i is under an attack ωi that is generated by ωi = Γωi , where the eigenvalues of Γ are a subset of the eigenvalues of the agent's dynamic A. Then, intact agents that are reachable from agent i do not synchronize to the leader, while their local neighborhood tracking error (4) is zero. Proof. See Appendix A. Remark 1. H∞ is one of the most common and effective approaches to attenuate disturbances. However, Theorem 2 implies that the standard H∞ controllers for MASs that use the exchange of relative information can be bypassed by the attacker. This is because although the goal of the H∞ is to attenuate the effect of adversarial input on the local neighborhood tracking error, Theorem 2 shows that the at- tacker can make the local neighborhood tracking error go to zero, while agents are far from synchronization. Therefore, a different controller framework and H∞ controller is presented in this paper that guarantees attenuating attacks on sensors and actuators of a compromised agent. (cid:3) IV. The Proposed Attack Mitigation Approach In this section, the proposed resilient control approach is presented. First, a distributed observer-based H∞ control protocol is developed to not only prevent attacks on physical components, i.e., attacks on sensors and actuators (we call them Type 1 attacks), from propagating throughout the net- work but also attenuate their effect on the compromised agent. Then, a trust-confidence based control protocol is examined 3 to identify and isolate neighbors that are compromised by attacks on the communication network or attacks that take over the control of a compromised agent (we call them Type 2 attacks). Figure 1 shows the structure of the proposed control framework. Fig. 1: The proposed autonomous and secure distributed H∞ controller for multi-agent systems. The observer and H∞ controller allow mitigation of Type 1 attacks and designing autonomous controllers for agents. The trust-confidence mechanism mitigates Type 2 attacks. A. Overall structure of the proposed approach We now formulate a resilient observer-based H∞ distributed control protocol for MAS (1)-(2) in the presence of attacks. In the proposed approach, only the leader communicates its actual sensory information and followers do not exchange their actual state information. This stops propagating Type 1 attacks from a compromised agent to others. To this end, the followers estimate the leader's state using a distributed observer and communicate this estimation to their neighbors to achieve consensus on the leader state. The distributed observer is designed as ∀i = 1, . . . , N (8)  ri = Ari + Bυi (cid:88) υi = cFηi + ρh(Fηi) ηi = j∈Ni where h(.) is defined in (6) and ηi is a revisited local neigh- borhood observer tracking error for agent i defined by C j(t)Ti j(t)ai j(r j − ri) + bi(ζ0 − ri) (9) where 0 ≤ C j(t) ≤ 1 is the confidence of agent j and 0 ≤ Ti j(t) ≤ 1 is the trust value of agent i to its neighbor j. The confidence and trust values along with the design parameters c, ρ and F in (8) are designed in subsection B to mitigate Type 2 attacks, i.e., to identify and remove entirely compromised agents or attacks on communication network, and thus guarantee that ri → ζ0 for all intact agents, regardless of attacks. To further increase resiliency at the local level and attenuate the effects of Type 1 attacks on the compromised agent itself, the control input ui in (1) is designed as a function of ri and xi in subsection C (see Theorem 4) to guarantee that the following bounded L2-gain condition is satisfied for the agent i (cid:82) ∞ (cid:82) ∞ 0 e−αit(cid:107)zi(t)(cid:107)2dt 0 e−αit(cid:107)ωi(t)(cid:107)2dt ≤ γ2 i (10) with zi defined as the controlled or performance output and is obtained by (cid:107)zi(cid:107)2 = (xi − ζ0)T Qi(xi − ζ0) + uT i Riui (11) 4 where αi and γi represent the discount factor and the attenua- tion level of the attack ωi, respectively, and the weight matrices Qi and Ri are symmetric positive definite. If condition (10) is satisfied, then, the H∞ norm of Tωi,zi, i.e., the transfer function from the attack ωi to the performance output zi, is less than or equal to γi. Note also that ωi(t) does not need to be a bounded energy signal because of the discount factor αi. The problem formulation can now be given as follows. Problem 1. (Resilient H∞ Synchronization Problem) Con- sider N agents defined in (1)-(2) with the distributed observer given by (8)-(9). Design the control protocol ui = fi(xi, ri, vi) in (1) along with C j(t), Ti j(t), c, ρ and F in (8)-(9) such that 1) The bounded L2-gain condition (10) is satisfied when 2) The ωi (cid:44) 0. (cid:107)xi(t) − ζ0(t)(cid:107) → 0, i = 1, . . . , N when ωi = 0. synchronization problem is solved, i.e., B. The proposed distributed observer design The distributed observer (8)-(9) only communicates the observer state ri, which cannot be affected by Type 1 attacks on physical components. A trust-confidence mechanism is designed in the following to mitigate Type 2 attacks. To this end, a confidence value is defined for each agent to indicate the trustworthiness of its own observer information. Agents communicate their confidence value with their neighbors to alert them to put less weight on the information they are receiving from them, depending on how low their level of confidence is. This slows down the propagation of Type 2 attacks. If an agent is not confident about its own observer information, then assigns a trust value to its neighbors and incorporates these trust values along with the confidence values received from neighbors in its update law to determine the significance of the incoming information. Figure 2 shows the block diagram of the proposed distributed monitor. it Note also that it is assumed that the attacker designs its signal based on Theorem 2 to deceive intact agents, so that they cannot monitor any anomaly by examining their local neighborhood tracking error. This is considered the worst attack scenario. If, however, the attacker does not satisfy the conditions of Theorem 2, then, intact agents can easily detect attacks by using KullbackLeibler divergence criteria to check discrepancy between the normal statistical properties of the local neighborhood tracking error and its actual ones [42]. 1) Confidence value: A confidence value is defined for each agent which shows the validity of its information. To proceed, define oi = (cid:107)ηi(cid:107) Fig. 2: The distributed monitor structure. for agent i . In contrast to oi, si does not converge to zero if agent i is in the path of an attacker. This is because si = 0 requires r j = ri = ζ0, ∀ j ∈ Ni, which indicates that agent i and its neighbors are synchronized and, therefore, are not in the path of an attacker. In the absence of attack, oi and si converge to zero and have the same behavior. Therefore, by comparing oi and si one can detect whether or not the agent is in the path of a compromised agent. The confidence value Ci(t) in (9) for an intact agent i is defined as (cid:90) t 0 Ci(t) = β e−β(t−τ)qi(τ)dτ with qi(t) = ∆i ∆i + (cid:107)si(t) − oi(t)(cid:107) (14) (15) where β > 0 is a discount factor used to determine how much we value the current experience with regard to the past experiences. ∆i is a threshold value to account for factors other than attacks, i.e., channel fading and disturbance. If agent i is not in the path of any compromised agent, (cid:107)si − oi(cid:107) is zero almost all the time and, consequently, Ci is almost one. On the other hand, if agent i is affected by an attacker, then (cid:107)si − oi(cid:107) (cid:29) ∆i, and Ci is less than one and its value depends on how close the agent is to the source of the attack. Equation (14) can be implemented by the differential equation Ci(t) + βCi(t) = βqi(t). The worst case scenario is assumed in which a disrupted agent broadcasts the confidence value 1 to its neighbors to fool them. 2) Trust value: The trust value is defined to determine the importance of the incoming information of each agent's neighbor. To calculate the trust value of agent i to agent j , j we first measure the difference between the state of agent and the average of the state of all neighbors of agent i using di j(t) = κ (cid:90) t (cid:13)(cid:13)(cid:13)r j(t) − hi(t) (cid:13)(cid:13)(cid:13) + θi θi 0 e−κ(t−τ)li j(t)dτ (16) r j (17) (cid:88) j∈Ni , hi = 1 Ni with li j(t) = for agent i where ηi is defined in (9). Based on Theorem 2, ηi and, consequently, oi converges to zero for intact agents. Now, define (cid:88) j∈Ni si = ai j (cid:13)(cid:13)(cid:13)r j − ri (cid:13)(cid:13)(cid:13) + bi (cid:107)ζ0 − ri(cid:107) where hi denotes the average value of the neighbors of agent i and Ni is the number of neighbors of agent i . The discount factor κ > 0 determines how much we value the current expe- rience of interaction with regard to the past experiences. θi is a threshold value to take into account factors other than attacks. (12) (13) (cid:16) (18) Ci(t), di j(t) Ti j(t) = max (cid:13)(cid:13)(cid:13) ∀ j ∈ Ni Equation (16) can be implemented as di j(t) + κdi j(t) = κli j(t). Now, we define the trust value of agent i to its neighbor j given as Ti j(t) in (9) as (cid:17) Ti j can also be normalized to satisfy (cid:80) (cid:13)(cid:13)(cid:13)r j − hi j∈Ni Ti j(t) = 1. If there is no attack and the network is also synchronized, then, is one ∀ j ∈ Ni. Moreover, when there is no attack and agent i receives considerably different values from its neighbors before synchronization, e.g. as a result of a change in the state of the leader, since Ci is close to one as there is no attack, Ti j is almost one ∀ j ∈ Ni. On the other hand, if agent i is affected by an attack, then Ci is small and the trust of agent i to agent j depends on is zero, and, consequently, Ti j (cid:13)(cid:13)(cid:13) , j ∈ Ni. (cid:13)(cid:13)(cid:13)r j − hi It is shown in Theorem 3 that the proposed observer-based distributed control protocol (8)-(9) guarantees synchronization of intact agents, if the following assumption is satisfied. Assumption 4. The network connectivity is at least (2 f + 1), i.e., at least half of the neighbors of each agent are intact [43]. Define H = (L + G) (19) where L is the graph Laplacian matrix and G is the diagonal pinning matrix.Then, based on Lemma 1, H is a non-singular M-matrix. The following lemmas are used in the proof of Theorem 3. Lemma 2 [44]. Let P > 0 and Assumption 2 be satisfied. Then, the solution P to the following ARE equation AT P + PA + I − PBBT P = 0, (20) is positive definite. Lemma 3 [45]. Let Assumption 1 be satisfied. Then, there exists a positive vector φ = [φ1, . . . , φN]T such that (ΦH + H T Φ) > 0 where Φ = diag{φ1, . . . , φN}, φ = (H T )−11N, and H is defined in (19). Theorem 3. Let Assumption 4 be satisfied. Consider the dynamic observer defined in (8)-(9) with C j(t) given by (14) and Ti j(t) defined in (18). Let F = −BT P, c (cid:62) φmax to the algebraic Riccati equation (20) and φmax = max Then, ri → ζ0 for all intact agents. i=1,...,N Proof. See Appendix B. (cid:14)λmin(ΦH + H T Φ) and ρ ≥ vmax, where P is the solution (φi). (cid:3) Remark 2. Note that for Type 1 attacks, the proposed control framework does not impose any constraints on the number of neighbors or the total number of agents under attacks. Note also that in contrast to existing mitigation approaches, we do not discard the neighbor's information for an agent based solely on the difference between their values. Therefore, when the discrepancy between agents is because of a legitimate change in the leader, the confidence and trust values for each agent become 1 and, consequently, all agents synchronize to the leader. 5 The following subsection shows how to design a resilient observer-based H∞ distributed controller. Non-homogeneous game AREs are derived for solving the optimal H∞ synchro- nization problem. C. The proposed resilient controller It was shown in Theorem 3 that ri → ζ0 for all intact agents regardless of attacks. Similar to [46], one can show that if ui(t) in (1) is designed to guarantee xi → ri, using the separation principle, one can guarantee xi → ζ0. Therefore, in the following, the control input ui is designed to solve Problem 1 with ζ0 replaced with ri. as i(t) = xi(t) − ri(t). The dynamic of the error becomes Define the error between the state of agent i and its observer i(t) = Ai(t) + Bui(t) + Dωi(t) − Bυi(t) Define the augmented system state as Xi(t) = [i(t)T ri(t)T ]T ∈ R2n Using (21) and (8) together yields the augmented system as Xi = T Xi + B1ui + D1ωi + E1υi with T = (cid:34) A 0 0 A (cid:35) (cid:35) (cid:34) B 0 , B1 = , D1 = (21) (22) (cid:35) (23) , (24) (cid:35) (cid:34) D 0 (cid:34) −B (cid:35) B 0 0 E1 = (cid:34) Q1i 0 The performance output (11) in terms of the augmented state (22) (while ζ0 is replaced with ri) becomes (cid:107)Zi(cid:107)2 = XT i QiXi + uT i Riui, Qi = Using (8) and (22), the control protocol for the augmented system can be written as the following non-homogeneous control input ui = KiXi + gi (25) Note that gi is used to compensate the non-homogeneous term υi in the augmented system (23). With the aid of (10) (while ζ0 is replaced with ri), define the discounted performance function in terms of the augmented system (23) as i QiXi + uT XT i Riui − γi 2ωi T ωi (26) dτ The value function for linear systems is quadratic with the form as Vi(Xi(t)) = Xi(t)T PiXi(t) + 2XT i Πi + Γi (27) i QiXi + uT i Riui − γ2 and the corresponding Hamiltonian function becomes i ωi − αiVi + H (Xi, ui, ωi) (cid:44) XT Remark 3. It is assumed here that the full state of agents is available for measurement. However, if not available, the proposed design procedure can be extended for the case of dynamics controllers in which the states of agents are esti- mated using a local observer. This is because local observers dVi dt i ωT (28) (cid:90) ∞ J(Xi, ui, ωi) = e−αi(τ−t)(cid:104) t (cid:105) can estimate agents' state without any exchange of information with their neighbors. On the other hand, if the entire agent is compromised and its state observer is manipulated, its neighbors detect it and discard its information using the proposed trust-confidence mechanism. Theorem 4. (Non-homogenous game ARE) The optimal solu- tion for the discounted performance function (26) is i = −Ri u∗ −1BT 1 (PiXi + Πi) (29) the following non-  where Pi and Πi are the solution of homogeneous game ARE PiT + T T Pi − αiPi − PiB1Ri Πi = Γi = αiΓi + ΠT −1BT 1 Pi + 1 − T T − 1 Πi − 1 αiI2n + PiB1Ri i B1Ri −1BT −1BT γ2 i ΠT i D1DT 1 (cid:18) 1 γ2 i 1 γ2 i PiD1DT 1 1 Pi + Qi = 0 PiD1DT (cid:19) Πi − 2υT Πi − PiE1υi Πi i ET 1 (30) (cid:3) Proof. See Appendix C. The following theorem shows that the control protocol (29) along with (30) solve the H∞ synchronization problem. Theorem 5. Consider the MAS (1)-(2) with the observer (8)- (9). Let the control input ui be given as (29)-(30). Then, Problem 1 is solved. Proof. See Appendix D. (cid:3) V. Model-Free Resilient Off-Policy RL for Solving Optimal Synchronization for Intact Agents In this section, an RL algorithm is proposed to solve Problem 1 on-line without requiring any knowledge of the agents' dynamics. The off-policy RL allows separating the behavior policy from the target policy for both control input and attack. In order to find the optimal control (29) without the requirement of the knowledge of the system dynamics, the off-policy RL algorithm [47] is used in this subsection. Off-policy algorithm has two separate stages. In the first stage, an admissible policy is applied to the system and the system information is recorded over the time interval T. Then, in the second stage, without requiring any knowledge of the system dynamics, the information gathered in stage 1 is repeatedly used to find a sequence of updated policies uk i and ω∗ i . To this end, the augmented system dynamics (23) is first written as i converging to u∗ i and ωk ui − uk Xi = T kXi + B1 + D1 (31) i where Xi is defined in (22) and uk i denote the control and disturbance target policies in iteration k to be updated. Now, using (27), one has + E1υi i (cid:17) (cid:16) ωi − ωk i and ωk (cid:16) (cid:17) Vk i (Xi(t)) = Xi(t)T Pk (32) In this case, the control policy and worst case attack signal i Xi(t) + 2XT + Γk i i Πk i 6 (cid:17) − XT (cid:16) Xi(t)T Pk i ) − 2(uk+1 i i (ωk+1 i ) + 2γ2 i Taking time derivative of Vk i along the augmented system dynamic (32) yields (34) i QiXi i + Γk i i )T Ri(uk i (ωk i )T (ωk i Xi(t) + 2XT i Πk )T Ri(ui − uk i ) )T (ωi − ωk i ) Vk i (Xi(t)) = αi −(uk +γ2 The following off-policy integral RL Bellman equation is derived by multiplying both sides of (34) by e−αi(τ−t) and integrating e−αiT Vk = − (cid:19) i )T (ωk i (ωk dτ i ) )T Ri(ui − uk dτ i ) (35) The off-policy RL algorithm presented by iterating on (35) to solve the non-homogeneous game ARE, is listed in Algorithm 1. (cid:90) t+T (cid:18) i (Xi(t + T)) − Vk i (Xi(t)) (cid:90) t+T XT i QiXi + (uk i ) − γ2 i ) − 2(uk+1 )T (ωi − ωk i )T Ri(uk e−αi(τ−t) e−αi(τ−t) i (ωk+1 2γ2 (cid:19) (cid:18) + t i t i The following theorem shows that using the proposed control framework, the learning mechanism is resilient against both Types 1 and 2 attacks. Theorem 6. Consider the MAS (1)-(2) under the control protocol (29) with the observer dynamics given by (8)-(9). Let the off-policy Algorithm 1 be used to learn the gains in (29). Then, Problem 1 is solved, if Assumptions 1 to 4 are satisfied. Proof. Similar to [47], one can show that the off-policy RL Algorithm 1 solves Problem 1 in an optimal manner, as long as ri → ζ0. That is, for intact agents xi → ζ0 and for a compromised agent the L2 condition (10) is satisfied. This boils down the proof to show that ri → ζ0 even if the system is under attack. On the other hand, Theorem 3 shows that ri → ζ0 regardless of attacks. This completes the proof. (cid:3) Remark 4. In the proposed Algorithm 1, steps 4 to 9 for finding the trust and confidence values are continually em- ployed even after learning. However, once the optimal gain is found, the learning steps 10-12 are skipped, because the gains required for the control policy are computed and there is no need for further computation, unless another learning phase is initiated by any change in the agent dynamics. One might argue that the off-policy Algorithm 1 requires to measure the attack signal ωi(t) which is restrictive. However, the off-policy Algorithm can learn about the worst-case attack signal using actual measurable disturbances (either applied intentionally or coming from nature), instead of measuring attack signals. Once it learned the worst-case scenario, it can attenuate attacks without measuring them. VI. Simulation Results In this section, an example is provided to verify the effec- tiveness of the proposed control protocol. The communication graph is given in Fig. 3. Consider 5 agents with dynamics as (cid:35) (cid:34) 0 −4 1 0 (cid:35) (cid:34) 1 0 (cid:35) (cid:34) 1 0 xi + ui + ωi (36) (56) can be written as uk+1 i ωk+1 i = −Ri 1 γ2 i = −1BT DT 1 (Pk i Xi + Πk 1 (Pk i ) i Xi + Πk i ) (33) xi = Algorithm 1 Autonomous Resilient Control Protocol (ARCP) for agent i 1: Procedure ARCP 2: Apply an admissible behavior policy to the agent 3: For i=1:N Use r j, 4: Compute Ci as 5: j ∈ Ni to compute si and oi (cid:90) t ∆ie−β(t−τ) Ci(t) = β ∆i + (cid:107)si(τ) − oi(τ)(cid:107) dτ 0 di j(t) = κ (cid:90) t Communicate Ci with agents j ∈ Ni For each agent j ∈ Ni , compute (cid:80) θie−κ(t−τ) (cid:17) (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)r j(τ) − 1Ni (cid:16) Use C j, j ∈ Ni and Ti j to find Ti j(t) = max Ci(t), di j(t) θi + 0 j∈Ni (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) dτ r j ηi = C j(t)Ti j(t)ai j(r j − ri) + bi(ζ0 − ri) (cid:88) j∈Ni 6: 7: 8: 9: 10: 11: End for 12: Given uk Compute the observer state ri using update law (8) Collect enough samples Xi = [Xi; Xi(iTs)] and ui at different sampling interval Ts (cid:18) i (Xi(t)) = Bellman equation e−αiT Vk − i and collected information solve the following (cid:90) t+T i (Xi(t + T)) − Vk (cid:90) t+T e−αi(τ−t) (cid:90) t+T i )T Ri(uk )T (ωi − ωk )T Ri(ui − uk t +2γ2 i −2 e−αi(τ−t)(ωk+1 e−αi(τ−t)(uk+1 XT i QiXi + (uk i ) − γ2 i )dτ i (ωk i )T (ωk i ) (cid:19) dτ i t i )dτ i t 13: Once the learning is done, replace the behavior policy with the control solution found using RL Fig. 3: The communication graph among the agents. The leader dynamics is given by (cid:35) (cid:34) 0 −4 1 0 x0 = x0 + (cid:35)(cid:16) (cid:34) 1 0 (cid:17) 4e−0.15t sin(2t) (37) The design parameters are Q1i = 100I2, Ri = 1, αi = 0.1 and, γi = 10 for all agents. Now, assume that Agent 2 is 7 (a) (b) Fig. 4: Agents state, when Agent 2 is under the Type 1 attack (38). (a) The standard control protocol (5) is used. (b) The local neighborhood tracking error (4) for each agent. affected by a Type 1 attack with the attack signal ω2 given as (cid:40) 10 sin(2t) ω2 = 0 t ≥ 10 otherwise (38) Fig. 5: The state of all agents when Agent 2 is under Type 1 attack and the observer defined in (8) is used for each agent. The agents' states are shown in Fig. 4. It is observed from Fig. 4a that when the standard distributed controller (5) is used, before attack all agents synchronize to the leader. However, after the attack, Agents 4 and 5, which have a path to the compromised Agent 2, do not synchronize to the leader. One can see from Fig. 4b that, as stated in Theorem 2, the local neighborhood tracking error (4) converges to zero for all intact agents except the compromised agent. The performance of the observer-based H∞ controller (29) in the presence of Type 1 attack (38) is shown in Fig. 5. One can see that the compromised agent is the only agent that does not follow the leader. Moreover, the H∞ controller attenuates the effect of the attack on the disrupted agent, which can be seen by comparing the deviation level of the compromised agent state from its desired value in Figs. 4a and 5. Now, assume that Agent 2 is under Type 2 attack. In this case, the attack signal (38) is applied to the observer of Agent 2. However, Assumption 4 is not satisfied. The result is shown in Fig. 6. It can be seen from Fig. 6a that using the trust-confidence mechanism, only the compromised agent and its direct neighbor do not synchronize to the leader. The confidence value of the agents is shown in Fig. 6b. One can see that Agents 4 and 5 are not confident about their own observer information, since they are in the path of the compromised agent. To satisfy Assumption 4, it is considered that 2 incoming links from Agent 5 and Agent 1 are connected to Agent 4. Fig. 7 shows the agents output and confidence 05101520Time(s)-10-50510LeaderAgent1DisruptedAgent2Agent3Agent4Agent505101520Time(s)-6-4-2024e1e2e3e4e505101520Time(s)-10-50510LeaderAgent1DisruptedAgent2Agent3Agent4Agent5non-root nodes. Then, it can be partitioned as  0 −∆ ¯L =  ∈ R(N+1)×(N+1) 0 . . . 0 L f 8 (39) where ∆ ∈ RN×1 denotes a vector whose i-th element is nonzero and indicates that the follower i is connected to the leader. L f ∈ RN×N indicates the interaction between the leader and the followers. Without loss of generality, assume D = B. Using (6), the global dynamic of the MAS (1) under attack with the control input (5) in terms of the Laplacian matrix (39) and after some manipulations, one has x =  B  v0 + (IN+1 ⊗ B) (cid:19), (cid:13)(cid:13)(cid:13)(cid:13)) ¯L ⊗ K + ¯ω (cid:46)(cid:13)(cid:13)(cid:13)(cid:13)(cid:16) (cid:17)  v0  B (cid:16)−c1 ¯L ⊗ K + ¯ω (cid:17), 0 ( ¯L ⊗ K 0 x (IN+1 ⊗ A) x + (cid:18)−(c1 + c2 (IN+1 ⊗ A) x + + (IN+1 ⊗ B) (cid:13)(cid:13)(cid:13)(cid:13)(cid:16) (cid:13)(cid:13)(cid:13)(cid:13)(cid:16) i f i f ( ¯L ⊗ K ( ¯L ⊗ K (cid:17) (cid:17) (cid:13)(cid:13)(cid:13)(cid:13) (cid:44) 0 (cid:13)(cid:13)(cid:13)(cid:13) = 0 x x s , xT 0 , xT 1 , . . . , xT 1 , . . . , ωT N]T and ¯ω = [0, ωT (40) where x = [ζT N]T , and 0 in ¯ω indicates that the leader is a trusted node and is not under attack. It can be seen that agents reach a steady state, i.e., xi → Axi +[BT 0]T v0, if the last terms of (40) tend to zero, i.e., ¯ω ∈ Im(¯c( ¯L ⊗ K)), where ¯c = (c1 + c2/(cid:107)( ¯L ⊗ K)x(cid:107)) or ¯c = c1. Otherwise, since the attack signal has common eigenvalues with the agent dynamics, the agents' states go to infinity. In the latter case, the local neighborhood tracking error goes to zero as xi → ∞. To prove the former case, we first show that ¯ω ∈ Im(¯c( ¯L ⊗ K)), if the attack signal is designed as given in the statement of Theorem. Note that ¯ω ∈ Im(¯c( ¯L ⊗ K)), if there exists a nonzero vector xs such that Define xss = [ζT ¯c( ¯L ⊗ K)xss = ¯ω (41) s ]T , where ζs and xs denote the steady states of the leader and followers, respectively. Since the leader is not under attack, using (39) and (41), one has ζs = Aζ0 +Bv0 and for the followers, − ¯c(∆ ⊗ K)ζs + ¯c(L f ⊗ K)xs = ω (42) Based on Assumption 1, the followers have at least one incoming link from the leader. On the other hand, L f captures the interaction among all followers, as well as the incoming link from the leader. The former is a positive semi-definite Laplacian matrix and the latter is a diagonal matrix with at least one nonzero positive element added to it. Therefore, as stated in Lemma 1, L f is nonsingular and, thus, the solution to (42) becomes xs = (¯c(L f ⊗ K))−1(ω + ¯c(∆ ⊗ K)ζs) (43) Since the eigenvalues of the attack signal are assumed a subset of the eigenvalues of the agent dynamics A, for every ¯ω there exists a nonzero vector xss such that (41) holds. Therefore, ¯ω ∈ Im(¯c( ¯L⊗K)). Now, using (39), the global form of the state neighborhood tracking error (4) can be written as (44) e = −( ¯L ⊗ In)x (a) (b) Fig. 6: Agent 2 is under Type 2 attack but Assumption 4 does not hold. (a) Agents state when the control protocol (29) with the observer (9) is used. (b) Confidence values of all agents. (a) (b) Fig. 7: Agent 2 is under Type 2 attack and Assumption 4 holds. (a) Agents state when the control protocol (29) with the observer (9) is used. (b) Confidence values of all agents. value when Assumption 4 is satisfied. It can be seen from Fig. 7a that only the compromised agent does not synchronize to the leader and all intact agents converge to the leader. The confidence value of the agents is shown in Fig. 7b. One can see that Agent 4 is not confident about its own observer information since it is the only immediate neighbor of the compromised agent. VII. Conclusion A resilient autonomous control framework is proposed for a leader-follower MAS with an active leader. It is first shown that existing standard synchronization control protocols are prone to attacks. Then, a resilient learning-based control proto- col is presented to find optimal solutions to the synchronization problem in the presence of attacks and system dynamic uncertainties. A distributed observer-based H∞ controller is first designed to prevent propagating the effects of attacks on sensors and actuators throughout the network, as well as attenuating the effect of these attacks on the compromised agent itself. Non-homogeneous game algebraic Riccati equa- tions are derived to solve the H∞ optimal synchronization problem. Off-policy reinforcement learning is utilized to learn their solution without requiring any knowledge of the agent's dynamics. Then, a trust-confidence based distributed control protocol is proposed to mitigate attacks that hijack the entire node and attacks on communication links. It is shown that the proposed RL-based H∞ control protocol is resilient against attacks. Appendix A Proof of Theorem 2 Let ¯L be the graph Laplacian matrix of the entire network, in which the leader is the only root node and N followers are 05101520Time(s)-10-5051015LeaderAgent1DisruptedAgent2Agent3Agent4Agent505101520Time(s)00.20.40.60.811.2C1C2C3C4C505101520Time(s)-20-1001020LeaderAgent1DisruptedAgent2Agent3Agent4Agent505101520Time(s)00.20.40.60.811.2C1C2C3C4C5Since (41) is satisfied, one has ¯c( ¯L ⊗ K)xs = ¯ω ⇒ ¯c(IN ⊗ K)e = − ¯ω or, equivalently ¯cKei = −ωi (45) (46) Equation (46) implies that the local neighborhood tracking error is zero, i.e., e j = 0 for intact agents that are not directly under attacks, i.e., ω j = 0∀ j (cid:44) i. This completes the proof. Appendix B Proof of Theorem 3 Type 1 attacks cannot affect the observer state since the observer cannot be physically affected by an attacker. On the other hand for Type 2 attacks, based on Assumption 4, the total number of compromised agents is assumed less than half of the network connectivity, i.e., 2 f + 1. Therefore, even if f neighbors of an intact agent are attacked and collude to send the same value to misguide it, there still exists f + 1 intact neighbors that communicate values different than the compromised ones. Thus, r j (cid:44) ri for some j ∈ Ni and, therefore, although oi in (12) is zero based on Theorem 2, si in (13) is nonzero and, consequently, its confidence value Ci in (14) will decrease and the attack will be detected. Moreover, since at least half of its neighbors are intact, it can update its trust values to remove the compromised neighbors. On the other hand, the entire network is still connected to the agent under attack and, therefore, the graph is still connected with the intact agents. Therefore, there exists a spanning tree in the graph associated with all intact agents. Let LI(t) be the graph of remaining intact agents with ¯ai j = C j(t)Ti j(t)ai j as its weights. Since LI(t) has a spanning tree, based on the above discussion, the inequality defined in Lemma 2 is still satisfied. The global form of the observer (8) can be written as r = (IN ⊗ A)r + (c(LI(t) + G) ⊗ BF)η + (ρIN ⊗ B)Ψ (47) where r = [rT N]T denotes the global state vector of the agents observer and Ψ = [h(Fη1), . . . , h(FηN)]T . The global form of (9) is 1 , . . . , rT 0 ) η = ((LI(t) + G) ⊗ In)(r − ζ 1 , . . . , ηT N]T and ζ (48) = (1N ⊗ In)ζ0. Using (47) and where η = [ηT defining HI(t) = LI(t) + G, the dynamic of the tracking error (48) becomes η = (IN ⊗ A)η + (cHI(t) ⊗ BF)η + (HI(t) ⊗ B)(ρΨ − v0) (49) = 1N ⊗ v0. Consider the Lyapunov candidate function with v0 as 0 V(η) = ηT (Φ ⊗ P)η where Φ is defined in Lemma 3 and is a positive definite matrix. Using (49), the time derivative of (50) yields V = 2ηT (Φ ⊗ PA)η + 2ηT (cΦHI(t) ⊗ PBF)η (cid:124)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:123)(cid:122)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:125) (cid:124)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:123)(cid:122)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:125) + 2ηT (ΦHI(t) ⊗ PB)(ρΨ − v0) (51) V1 V2 (50) (cid:46) By applying the stationary conditions [50] as ∂Hi ∂ui = 0, ∂ωi = 0, the optimal control and the worst case attack ∂Hi signal can be written as (cid:46) Substituting F = −BT P, V1 can be expressed as 9 Based on Lemma 3, (ΦHI(t) + H T V1 = ηT (Φ⊗(PA+AT P))η−cηT ((ΦHI(t)+H T I (t)Φ)⊗PBBT P)η (52) I (t)Φ) > 0 and the inequality −c(ΦHI(t)+H T I (t)Φ) ≤ −cλmin(ΦHI(t) + H T I (t)Φ)Φ φmax holds [48]. Using the Kro- necker property A ⊗ (B + C) = A ⊗ B + A ⊗ C and ARE (20), (52) becomes (cid:46) I (t)Φ) ≤ −cλmin(ΦHI(t)+H T I (t)Φ) ⊗ PBBT P)η V1 = ηT (Φ ⊗ (PA + AT P))η − cηT ((ΦHI(t) + H T (cid:32) ≤ −ηT (Φ ⊗ In)η λmin(ΦHI(t) + H T (cid:33) I (t)Φ) 1 − c + ηT (Φ ⊗ PBBT P)η φmax V2 = (53) I (t)Φ), which is satisfied if the condition of c in the statement of the theorem is held. Now, for V2 one has Therefore, V1 ≤ 0 if c ≥ φmax N(cid:88) N(cid:88) (cid:14)λmin(ΦHI(t) + H T (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)BT P N(cid:88) N(cid:88) j PB(ρψi−v0) ≤ −(ρ−vM) This implies that V2 ≤ 0, if ρ ≥ vM. Therefore, (54) V = V1 + V2 ≤ 0, if the conditions defined in the statement of the theorem are satisfied. Since V(t) ≤ 0, V(t) is bounded. Note that V(t) ≡ 0 implies that η = 0. Hence, using LaSalles invariance principle [49], it follows that the observer state ri asymptotically converges to the leader state ζ0. αi jηT j (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) αi jηT i=1 i=1 j=1 j=1 Appendix C Proof of Theorem 4 Using the value function (27) for the left-hand side of (26) and differentiating along with the augmented system (23) gives the following Bellman equation (cid:16) i Pi(T Xi + B1ui + D1ωi + E1υi) i Πi + Γi H (Xi, ui, ωi) = 2XT +2(T Xi + B1ui + D1ωi + E1υi)T Πi + 2XT −αi +XT Xi(t)T PiXi(t) + 2XT i Riui − γ2 i QiXi + uT i Πi + Γi i ωT i ωi = 0 (cid:17) (55) i = −Ri u∗ 1 ∗ i = ω γ2 i 1 (PiXi + Πi) −1BT DT 1 (PiXi + Πi) (56) Substituting (56) into Bellman equation (55) results in the following tracking game ARE H(cid:0)Xi, u∗ ∗ i , ω i (cid:16) (cid:1) = XT i  Xi + 2XT PiT + T T Pi − αiPi − PiB1Ri −1BT (cid:16) Πi − PiB1Ri  1 Pi + Qi Πi + PiE1υi − αiΠi + T T Πi 1 i −1BT 1 P i Πi+ i B1R−1 i BT 1 Πi + 1 γ2 i ΠT i D1DT 1 Πi + 2υT i ET 1 Πi (57) Since (57) is satisfied for all Xi, this can occur if, and only if, (30) holds. This completes the proof. PiD1DT 1 γ2 i 1 γ2 i + + + PiD1DT 1 Γi − αiΓi − ΠT  = 0 Appendix D Proof of Theorem 5 i ωT i QiXi + uT i Riui − γ2 i ωi − αiV∗ tion V∗ become The Hamiltonian function (28) for the optimal value func- i , and any control policy ui and disturbance policy ωi H(V∗ i , ui, di) = XT +V∗T iX (T Xi + B1ui + D1ωi + E1υi) i , u∗ i , ui, di) = H(V∗ Using (57), one has H(V∗ i ) + i )T (ωi − ω∗ i ) + γ2(ωi − ω∗ i )T Ri(ui − u∗ (ui − u∗ i ) and based i , ω∗ i , u∗ (28), H(V∗ 0. i ) on Hamiltonian equation i Riui − γ2 i ωi − αiV∗ Then (58) gives XT i ωT i QiXi + uT + i ) − i )T Ri(ui − u∗ iX (T Xi + B1ui + D1ωi + E1υi) = −(ui − u∗ V∗T γ2(ωi − ω∗ i )T (ωi − ω∗ i ). Now, using the optimal control policy ui = u∗ i , ω∗ (58) = i yields i i i Riui − γ2 i ωi − αiV∗ XT i QiXi + uT i ωT i +V∗T iX (T Xi + B1ui + D1ωi + E1υi) i ) ≤ 0 = −γ2(ωi − ω i )T (ωi − ω ∗ ∗ (59) (cid:16) Multiplying both sides of (59) by e−αit, and defining V∗ V∗T iX (T Xi + B1ui + D1ωi + E1υi) as the derivative of V∗ (cid:17) the trajectories of the closed-loop system, gives i Riui + γ2 d/dt Integrating both sides of (60) and using the fact that V∗ i ωT i ωi 0, for every T > 0 and every ωi ∈ L2[0,∞), one has i QiXi − uT e−αitV∗ i i = i along (60) i (.) ≥ (cid:17) ≤ e−αit(cid:16)−XT (cid:90) T e−αiτ(cid:16) (cid:90) T 0 e−αiτγ2 ≤ 0 (cid:17) i QiXi + u∗T XT i Riu∗ i ωidτ + V∗ i ωT i i (X(0)) dτ (61) Using the separation principle for the combination of the observer and the controller completes the proof of Part 1 of Problem 1. The proof of Part 2 is similar to [47] and, therefore, is omitted. References [1] P. J. Antsaklis, K. M. Passino, and S. J. Wang, "An introduction to autonomous control systems," IEEE Control Systems, vol. 11, pp. 5–13, June 1991. [2] 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, pp. 215–233, Jan 2007. 10 [3] A. Jadbabaie, J. Lin, and A. S. Morse, "Coordination of groups of mobile autonomous agents using nearest neighbor rules," IEEE Transactions on Automatic Control, vol. 48, pp. 988–1001, June 2003. [4] J. Fax and R. Murray, "Information flow and cooperative control of vehicle formations," IEEE Transactions on Automatic Control, vol. 49, no. 9, pp. 1465–1476, 2004. cited By 2455. [5] W. Ren, R. W. Beard, and E. M. Atkins, "Information consensus in multivehicle cooperative control," IEEE Control Systems, vol. 27, pp. 71–82, April 2007. [6] R. Moghadam and H. Modares, "An internal model principle for the attacker in distributed control systems," in 2017 IEEE 56th Annual Conference on Decision and Control (CDC), pp. 6604–6609, Dec 2017. [7] R. S. Sutton and A. G. Barto, Reinforcement learning: An introduction, vol. 1(1). MIT press Cambridge, 1998. [8] W. B. Powell, Approximate Dynamic Programming: Solving the curses of dimensionality, vol. 703. John Wiley & Sons, 2007. [9] F. L. Lewis, D. Vrabie, and K. G. Vamvoudakis, "Reinforcement learning and feedback control: Using natural decision methods to design optimal adaptive controllers," IEEE Control Systems, vol. 32, no. 6, pp. 76–105, 2012. [10] D. Liu, D. Wang, and H. Li, "Decentralized stabilization for a class of continuous-time nonlinear interconnected systems using online learning optimal control approach," IEEE Transactions on Neural Networks and Learning Systems, vol. 25, pp. 418–428, Feb 2014. [11] Y. Jiang and Z.-P. Jiang, "Computational adaptive optimal control for continuous-time linear systems with completely unknown dynamics," Automatica, vol. 48, no. 10, pp. 2699 – 2704, 2012. [12] B. Luo, H. N. Wu, and T. Huang, "Off-policy reinforcement learning for h∞ control design," IEEE Transactions on Cybernetics, vol. 45, pp. 65– 76, Jan 2015. [13] R. Kamalapurkar, B. Reish, G. Chowdhary, and W. E. Dixon, "Con- current learning for parameter estimation using dynamic state-derivative estimators," IEEE Transactions on Automatic Control, vol. 62, pp. 3594– 3601, July 2017. [14] Z. Ni, H. He, and J. Wen, "Adaptive learning in tracking control based on the dual critic network design," IEEE Transactions on Neural Networks and Learning Systems, vol. 24, pp. 913–928, June 2013. [15] B. Kiumarsi and F. L. Lewis, "Actor-critic-based optimal tracking for partially unknown nonlinear discrete-time systems," IEEE Transactions on Neural Networks and Learning Systems, vol. 26, no. 1, pp. 140–151, 2015. [16] K. G. Vamvoudakis and F. L. Lewis, "Online solution of nonlinear two-player zero-sum games using synchronous policy iteration," Inter- national Journal of Robust and Nonlinear Control, vol. 22, no. 13, pp. 1460–1483, 2012. [17] M. I. Abouheaf, F. L. Lewis, K. G. Vamvoudakis, S. Haesaert, and R. Babuska, "Multi-agent discrete-time graphical games and reinforce- ment learning solutions," Automatica, vol. 50, no. 12, pp. 3038 – 3053, 2014. [18] L. Busoniu, R. Babu ska, and B. D. Schutter, "A comprehensive survey of multiagent reinforcement learning," IEEE Transactions on Systems, Man, and Cybernetics, Part C (Applications and Reviews), vol. 38, pp. 156–172, March 2008. [19] S. Zuo, Y. Song, F. L. Lewis, and A. Davoudi, "Optimal robust output containment of unknown heterogeneous multiagent system using off-policy reinforcement learning," IEEE Transactions on Cybernetics, vol. PP, no. 99, pp. 1–11, 2017. [20] Y. Hong, J. Hu, and L. Gao, "Tracking control for multi-agent consensus with an active leader and variable topology," Automatica, vol. 42, no. 7, pp. 1177 – 1182, 2006. [21] Y. Yang, H. Modares, D. C. Wunsch, and Y. Yin, "Optimal containment control of unknown heterogeneous systems with active leaders," IEEE Transactions on Control Systems Technology, vol. PP, no. 99, pp. 1–9, 2018. [22] Z. Li, X. Liu, W. Ren, and L. Xie, "Distributed tracking control for linear multiagent systems with a leader of bounded unknown input," IEEE Transactions on Automatic Control, vol. 58, pp. 518–523, Feb 2013. [23] F. Pasqualetti, A. Bicchi, and F. Bullo, "Consensus computation in unreliable networks: A system theoretic approach," IEEE Transactions on Automatic Control, vol. 57, pp. 90–104, Jan 2012. [24] H. J. LeBlanc, H. Zhang, X. Koutsoukos, and S. Sundaram, "Resilient asymptotic consensus in robust networks," IEEE Journal on Selected Areas in Communications, vol. 31, pp. 766–781, April 2013. [25] H. Zhang and S. Sundaram, "Robustness of information diffusion algorithms to locally bounded adversaries," in 2012 American Control Conference (ACC), pp. 5855–5861, June 2012. 11 [48] Y. Wan, J. Cao, G. Chen, and W. Huang, "Distributed observer-based cyber-security control of complex dynamical networks," IEEE Transac- tions on Circuits and Systems I: Regular Papers, vol. 64, pp. 2966–2975, Nov 2017. [49] A. Isidori, Nonlinear Control Systems. Communications and Control & Sons, 2012. Engineering, Springer London, 1995. [50] F. L. Lewis, D. Veabie, and V. L. Syrmos, Optimal control. John Wiley [26] A. Teixeira, H. Sandberg, and K. H. Johansson, "Networked control systems under cyber attacks with applications to power networks," in Proceedings of the 2010 American Control Conference, pp. 3690–3696, June 2010. [27] S. Amin, X. Litrico, S. S. Sastry, and A. M. Bayen, "Cyber security of water scada systems-part ii: Attack detection using enhanced hydro- dynamic models," IEEE Transactions on Control Systems Technology, vol. 21, pp. 1679–1693, Sept 2013. [28] S. Weerakkody, X. Liu, S. H. Son, and B. Sinopoli, "A graph-theoretic characterization of perfect attackability for secure design of distributed control systems," IEEE Transactions on Control of Network Systems, vol. 4, pp. 60–70, March 2017. [29] Z. Feng and G. Hu, "Distributed secure average consensus for linear multi-agent systems under dos attacks," in 2017 American Control Conference (ACC), pp. 2261–2266, May 2017. [30] A. Khazraei, H. Kebriaei, and F. R. Salmasi, "Replay attack detection in a multi agent system using stability analysis and loss effective watermarking," in 2017 American Control Conference (ACC), pp. 4778– 4783, May 2017. [31] S. Sundaram and C. N. Hadjicostis, "Distributed function calculation via linear iterative strategies in the presence of malicious agents," IEEE Transactions on Automatic Control, vol. 56, pp. 1495–1508, July 2011. [32] W. Zeng and M. Y. Chow, "Resilient distributed control in the presence of misbehaving agents in networked control systems," IEEE Transac- tions on Cybernetics, vol. 44, pp. 2038–2049, Nov 2014. [33] H. J. LeBlanc, H. Zhang, X. Koutsoukos, and S. Sundaram, "Resilient asymptotic consensus in robust networks," IEEE Journal on Selected Areas in Communications, vol. 31, pp. 766–781, April 2013. [34] S. Abhinav, H. Modares, F. L. Lewis, F. Ferrese, and A. Davoudi, "Synchrony in networked microgrids under attacks," IEEE Transactions on Smart Grid, vol. PP, no. 99, pp. 1–1, 2017. [35] H. Modares, R. Moghadam, F. Lewis, and A. Davoudi, "Static output- feedback synchronization of multi-agent systems: A secure and unified approach," IET Control Theory & Applications, February 2018. [36] Y. Cao and W. Ren, "Optimal linear-consensus algorithms: An lqr perspective," IEEE Transactions on Systems, Man, and Cybernetics, Part B (Cybernetics), vol. 40, pp. 819–830, June 2010. [37] J. Li, H. Modares, T. Chai, F. L. Lewis, and L. Xie, "Off-policy rein- forcement learning for synchronization in multiagent graphical games," IEEE Transactions on Neural Networks and Learning Systems, vol. 28, pp. 2434–2445, Oct 2017. [38] Q. Jiao, H. Modares, S. Xu, F. L. Lewis, and K. G. Vamvoudakis, "Multi- agent zero-sum differential graphical games for disturbance rejection in distributed control," Automatica, vol. 69, no. Supplement C, pp. 24 – 34, 2016. [39] H. Zhang, F. L. Lewis, and A. Das, "Optimal design for synchronization of cooperative systems: State feedback, observer and output feedback," IEEE Transactions on Automatic Control, vol. 56, pp. 1948–1952, Aug 2011. [40] Z. Li and Z. Duan, Cooperative Control of Multi-Agent Systems: A Consensus Region Approach. Automation and Control Engineering, Taylor & Francis, 2014. [41] F. Lewis, H. Zhang, K. Hengster-Movric, and A. Das, Cooperative Con- trol of Multi-Agent Systems: Optimal and Adaptive Design Approaches. Communications and Control Engineering, Springer London, 2013. [42] M. Basseville and I. Nikiforov, Detection of abrupt changes: theory and application. Prentice-Hall information and system sciences series, Prentice Hall, 1993. [43] F. Pasqualetti, F. Drfler, and F. Bullo, "Attack detection and identification in cyber-physical systems," IEEE Transactions on Automatic Control, vol. 58, pp. 2715–2729, Nov 2013. [44] H. Zhang, F. L. Lewis, and A. Das, "Optimal design for synchronization of cooperative systems: State feedback, observer and output feedback," IEEE Transactions on Automatic Control, vol. 56, pp. 1948–1952, Aug 2011. [45] Z. Li, G. Wen, Z. Duan, and W. Ren, "Designing fully distributed consensus protocols for linear multi-agent systems with directed graphs," IEEE Transactions on Automatic Control, vol. 60, pp. 1152–1157, April 2015. [46] H. Modares, S. P. Nageshrao, G. A. D. Lopes, R. Babuka, and F. L. Lewis, "Optimal model-free output synchronization of heterogeneous systems using off-policy reinforcement learning," Automatica, vol. 71, pp. 334 – 341, 2016. [47] H. Modares, F. L. Lewis, and Z. P. Jiang, "H∞ tracking control of com- pletely unknown continuous-time systems via off-policy reinforcement learning," IEEE Transactions on Neural Networks and Learning Systems, vol. 26, pp. 2550–2562, Oct 2015.
1311.6149	1	1311	2013-11-24T18:08:56	Agent Approach in Support of Enterprise Application Integration	[ "cs.MA" ]	The present approach highlights the synergies between application integration and interaction protocols. Since both fields have advanced in different directions, a number of important technical problems can be addressed by their proper synthesis. In our previous work, we proposed a methodological approach based on Interaction Protocols for Enterprise Applica tion Integration (EAI). This approach permits to specify MAS (Multi-Agent System) interaction protocols, verify their behavior and use them to integrate multiple business applications. The result of the proposed approach is a validated interaction protocol. Based on this protocol, we define in this paper, an agent- based architecture for the EAI. It includes all the concepts nec- essary to support communication and coordination mechanisms such as inter-agent and agent-Web services communication.	cs.MA	cs	International Journal of Computer Science and Telecommunications [Volume 4, Issue 1, January 2013] 47 Agent Approach in Support of Enterprise Application Integration ISSN 2047-3338 Djamel Benmerzoug Faculty of NTIC, University of Constantine 2 benme rzougd j@yahoo .fr Abstract— The present approach highlights the synergies between application integration and interaction protocols. Since both fields have advanced in different directions, a number of important technical problems can be addressed by their proper synthesis. In our previous work [7][6][5], we proposed a methodological approach based on Interaction Protocols for Enterprise Applica tion Integration (EAI). This approach permits to specify MAS (Multi-Agent System) interaction protocols, verify their behavior and use them to integrate multiple business applications . The result of the proposed approach is a validated interaction protocol. Based on this protocol, we define in this paper, an agent- based architecture for the EAI. It includes all the concepts nec- essary to support communication and coordination mechanisms such as inter-agent and agent-Web services communication. Index Terms– Multi-Agent System, Interaction Protocol, Enterprise Application Integration I . INTRODUCT ION T HE penetration of Internet and the World Wide Web in accordance with new technological advances urged companies to seize the opportunities offered by electronic commerce and electronic business. Especial promising are the opportunities for cooperation among several geographically distributed orga- nizations based on open networks and innovative technologies. information and communication Indeed, several companies have showed many interests to the integration of these tech- niques. Interaction Protocols (IP), provide a middle ground between rigid automation and flexible manual execution. Using proto- cols as a mechanism to structure communications, agents can be less sophisticated. Protocol designers analyze and design protocols for desirable properties. Agents publicly declare the protocols in which they can participate making it easier to find agents with whom to interoperate. The importance of interaction is increasing as more and more monolithic systems are broken down into smaller ser- vices. The importance of interaction has been stressed by many authors [23][7][12]. Moreover, interaction is also considered a key issue in various industry standard[22]. In previous work [6][7], we described the use of IP to define and manage collaborative processes in B2B (Business to Business) relationships where the autonomy of participants is preserved. We showed the practicability of our approach by embedding it in a Web services language for specifying pro- tocols, which conducive to reuse, refinement and aggregation of our business protocols. We also elaborated translation rules from interaction pro- tocols notations used in our approach into Colored Petri Nets (CPN). These rules are implemented in IP2CPN 1 [7]: the tool we developed to automatically generate Petri nets from protocols specifications. Resulting Petri nets can be analyzed with dedicated tools to detect errors as early as possible. In this paper, we describe how the Multi-agent System (MAS) use the verified and the validated IP to establish the EAI. The proposed architecture based on the notion of agent. It includes all the concepts necessary to ensure all the phases of the EAI life-cycle. Our suggestion consists in the addition of a specific agent between the MAS application and its IP parts conceived as Web services. In fact, we defined two types of agent, namely, the Enterprise Agent representing an individual enterprise and the Integrator Agent. The remainder of the paper is organized as follows: Section 2 discusses some related work. Section 3 overviews our previ- ous work : The use of IP to define and manage collaborative processes in B2B relationships. Section 4 and 5 give details about the proposed agent-based architecture. In section 6, we study the different aspects related to the communication in our architecture. Section 7 gives some implementation aspects. Finally, concluding remarks are given in section 8. I I . R E L AT E D W O R K EAI is an active research domain. The community is still debating the issues of enterprises collaboration at the business process level. In fact, several areas of research are relevant to our work. We discuss some of them briefly and highlight the differences. a) Web services composition.: BPEL4WS [17] is a language designed to specify the static composition of Web services. However, it mixes interaction activities and business logic making it unsuitable for reuse [13]. In contrast to our approach, BPEL4WS elements are only used to specify messages exchanges between the different business partners. Afterwards, this specification is used by agents to enact the integration of business processes at run time. The Symphony project [10] has developed an algorithm for analyzing a composite service specification for data and control dependences and partitioning it into a set of smaller components. These components are then distributed to differ- ent locations and, when deployed, cooperatively deliver the same semantics as the original workflow. Symphony does not provide any support for failures arising from workflow 1 IP2CPN: Interaction Protocols To Colored Petri Nets Djamel Benmerzoug 48 mismatches since it assumes that the distributed processes will be derived from a single complete BPEL process. Other approaches try to combine Web services with multi- agent systems to ensure composition and interoperability. For example [8] summarize the relationship between agents and Web services with the aphorism Adaptive Workflow Engines = Web Services + Agents: namely, Web services provide the computational resources and agents provide the coordination framework. They propose the use of the BPEL4WS language as a specification language for expressing the initial social order of the multi-agent system. [8] does not provide any design issues to ensure the correctness of their interaction protocols. b) Interaction protocols modelling.: Interaction proto- cols are specific, often standard, constraints on the behaviors of the autonomous agents in a multiagent system. Protocols are essential to the functioning of open systems, such as those that arise in most interesting Web applications. The Dooley graph [21] is an alternative formalism for visu- alizing agent inter-relationships within a conversation. Object- oriented methods like UML [15] offer a way to reduce the gap between users and analyst when considering message transfers, yet they only address the dynamic behavior of individual objects and are informal. The state transition diagram has been extensively used for IP specification due to its clarity. The weakness is that it does not reflect the asynchronous character of the underlying communication [19]. Furthermore, it is not easy to represent integration of protocols. To ensure the correctness of IP specification at design time, [16] propose a translation rules for the conversation of an interaction protocol given in AUML to Petri net. Unfortu- nately, no automatic procedures were provided that guide the conversation of an interaction protocol given in AUML to Petri net representations. c) Discussion.: As shown before, several important strands of research of relevance to interaction protocols have been carried out before. However, the synthesis, enhancement, and application of these techniques on business process inte- gration is a novel contribution of this paper. Given the autonomy and heterogeneity of business partners, we believe that the most natural way to treat interactions is as communications. An interaction protocol involves the exchange of messages with a streamlined set of tokens. Indeed, the use of IP to define business process integration enables a greater autonomy of companies because each company hides its internal activities, services and decisions required to support public processes. In fact, our approach allows us to provide a clear separation of the inter-enterprise collaboration management and local business process one. This dichotomy makes full use of existing workflow system components, to support both public processes and private business processes. Formalisation and verification of autonomous and heteroge- neous collaborative business processes is a key prerequisite for the effective business application. As asserted by [20], mini- mal specifications based on external behavior will maximize interoperability. The present approach shows a design issues of how IP can be specified and verified. Another advantage of our approach is the integration com- pleteness property inherent from our BPEL4WS specification. It means that the IP is itself published and accessed as a Web service that can participate in other application integration. Particular protocols may then be selected and composed to support a desired business application. I I I . AN OV ERV IEW O F T H E P R O P O S E D A P P RO A C H In recent years, EAI modelling and reengineering have been longstanding activities in many companies. Most internal processes have been streamlined and optimized, whereas the external processes have only recently become the focus of business analysts and IT middleware providers. The static integration of inter-enterprise processes as common in past years can no longer meet the new requirements of customer orientation, flexibility and dynamics of cooperation [2]. In [4] we have developed an agent-based method for de- veloping cooperative enterprises information systems. This method permits to explicitly map the business process into software agents. In [5], we have described the use of IP to define and manage public processes in B2B relationships. The use of IP to define public processes enables a greater autonomy of companies. In this way, the IP provide a high abstraction level in the modelling of public processes. In other hand, the B2B integration scenarios typically in - volve distributed business processes that are autonomous to some degree, hence the importance of IP-based modelling. IP are a useful way for structuring communicative interaction among business partners, by organizing messages into relevant contexts and providing a common guide to the all parts. Formally an IP is defined as follow: Definition : An Interaction Protocol is a quadruplet: IP = <ID, R, M, fM >, where: • I D is the identifier of the interaction protocol • R = r1 , r2 , . . . , rn (n > 1) is a set of Roles (priva te business process or Web Services) • M is a set of non-empty primitive (or/and) complex messages, where: – A Primitive Message (P M ) corresponds to the sim- ple message, it is defined as follow: P M = <Sende r, Receiver, CA, Option>, where: ∗ Sender, Receiver ∈ R ∗ CA ∈ FIPA ACL Communicative Act (such as: cfp, inform, . . . ) ∗ Option: contain additional information (Syn- chronous / Asynchronous message, constraints on message,. . . ) – A Complex Message (CM) is built from simple r (primitive) ones by means of operators: CM = P M1 opP M2 . . . opP Mm , whe re: ∗ m > 1, op ∈ {XOR, OR, AND}, and International Journal of Computer Science and Telecommunications [Volume 4, Issue 1, January 2013] 49 ∗ ∀ i ∈ [1,m[, P Mi .S ender = P Mi+1 .S ender , P Mi .S ender ∈ R. • fM : a flow relation defined as : fM ⊆ (RxR), where (RxR) is a Cartesian product (r1 , r2 ) ∈ (RxR), for r1 , r2 ∈ R Figure 1 shows our approach for the treatment of business process integration based on interaction protocols. In this approach, we find two levels: • Design level: Formalisation and verification of interaction protocols. • Operational level: Management of interaction protocols at run-time. Interaction Protocol AUML interaction Diagram BPEL4W S Specification Design Level Operational Level 4 uses Integrator Agent 1 3: Feedback Veri fication (CPN-AMI tool) 2 CPN Model Enterprises Agents Fig. 1. The prop osed appro a ch A. Design level: Formalisation and Verification of Interaction Protocols Developing effective protocols to be executed by au- tonomous partners is challenging. Similar to protocols in traditional systems, IP in open and web-based settings need to be specified rigorously so that business partners can interact successfully. For this reason, developed a method for IP design and verification. This method (see figure 1) uses different models and languages. IP based on the use of Our method motivates AUML/BPEL4WS for application integration modelling, where pre- and post-conditions, rules, guards are specified in OCL2 . AUML (Agent UML) notation [1] [3] is a UML profile dedicated to agents trying to simplify the transition from soft- ware engineering to multi-agent system engineering. In other hand, BPEL4WS [17] (Business Process Execution Language for Web Services) is a de facto standard for describing Web 2 OCL: Object Constraint Language. (www.omg.org/cgi-bin) services composition. In our context, BPEL4WS was used as a specification language for expressing the interaction protocols of the multi-agent system [5]. When protocols are employed in open environments, such as the internet, they must be executed by agents that behave more or less autonomously and whose internal designs are not known. In such settings, therefore, there is a risk that the participating agents may fail to comply with the protocol [18]. Without a rigorous means to verify compliance, the very idea of protocols for interoperation is subverted. The use of formal methods is important because ensuring the correctness of complex protocols is seldom possible via other design approaches. High -level Petri nets are a suitable formal method for the design of IP because of their ability to express concurrency, non-determinism and system concepts at different levels of abstraction. CPN are a popular form of high-level Petri nets that have extensive tool support for the design of systems, including protocols. For this reason, we use Colored Petri Net (CPN) as target notation. CPN allows analysis to check for properties in IP. The semantics of the IP notations used in our approach and its application are described on the basis of translation rules into CPN. Consequently, lifelines, messages, constraints, splitting/merging paths, interaction terminations, and other IP construction elements are translated into a Petri net. After- wards, the resulting Petri net specification can be analysed by dedicated tools to detect errors as early as possible. In a previous work [6], we elaborated translation rules from interaction protocols notations used in our approach into Col- ored Petri nets. These rules are implemented in IP2CPN: the tool we developed to automatically generate the Petri net from protocols specification. The resulting Petri net specification can be analyzed by CPN-AMI tool [11] to detect errors as early as possible. B. Operational Level: Management of interaction protocols at run-time As we already have said before, the BPEL4WS process specification is considered as a specifying language for ex- pressing the interaction protocol of the multi-agents system. In this the rest of this paper we briefly describe how the MAS use the verified and the validated BPEL4WS specification to establish the EAI. Our suggestion consists in the addition of a specific agent between the MAS application and its IP parts conceived as Web services. IV. AN AG EN T-BAS ED AR CH IT ECT UR E F OR EN T ER PR I SE A P P L I C AT I O N I N T E G R AT I O N The proposed architecture based on the notion of agent. It includes all the concepts necessary to ensure all the phases of the Enterprise Application Integration life-cycle. As shown in Fig. 2, our suggestion consists in the addition of a spe- cific agent between the MAS application and its IP parts conceived as Web services. In fact, we defined two types of agent, namely, the Enterprise Agent representing an individual enterprise and the Integrator Agent. Djamel Benmerzoug 50 Integrato r Agent Web Se rv ices Know ledge base (Inetr act io n t races , Web ser v ices compos it io n …) req uest Interac tio n Ma na ger Inter act io n Bu i lder Inter act io n Co ntro ler IP as B PE L4WS Sp ec ifi catio n UDDI Publis h / request res ult Se rvi ces Manage r Commun ica tion Bus r res ult Agen t 1 Agen t 2 …………….. Agen t n Communication Module Planification & Coordination Execution Module G lobal Know ledge Ind iv idua l Knowledge Fig. 3. The stru cture of the Enterprise Agent Enterprise 1 Enterprise 2 …………….. Enterprise n constructed by the agent. B. Description of the Enterprise Agent Fig. 2. The prop osed ar chit e ctur e The main roles of the Integrator agent are the creation, monitoring, and control of IP life cycle. Its architecture fea- tures two modules: an interaction manager and a service one. The interaction manager contains operational knowledge (e.g., Interactions states). It also provides operations for monitoring interaction (i.e., creating and deleting instances). The service manager provides methods for receiving service requests, tracing service executions, and communicating with service requesters in accordance with IP definition (e.g., sending a no- tification informing the requester that deadline for cancelling an operation is passed). A. Description of the Integrator Agent As shown in Fig. 2, the Integrator agent is composed of four main modules. • The knowledge module includes a database and a set of actions. The role of this module is to store information that is necessary for the success of the interaction. Such information is traces of interaction used during the phases of the current states and coordination tasks running. This information is useful for future use of the protocol. • The Interaction Manager module allows the processing of messages depending on the specification of the inter- action protocols (the BPEL4WS specification) and the knowledge parts. • The services management module provides the services necessary for locating and invoking services. • The communication module allows the transport of mes- sages. In our work, we use the FIPA-ACL [14] as the language of communication between the different agents. This module has the role to structure the messages We consider the Enterprise Agent as an entity including a communication module, a planning and coordination module, and an execution module (Fig. 3). • The Communication Module: contains all the processes required to handle the messages: reception, filtering, and translation of incoming messages, and formulation and sending of the outgoing messages. • The Planning and Co-ordination Module: is responsible for managing the cooperation and formulating the offers for achieving sub-goals announced by the Integrator agent. • The Execution Module: this module contains the infor- mation about the internal resources of the individual enterprise, which makes possible the performance of local tasks that are assigned to the enterprise. It establishes the correspondence between the sub-goal assigned to the agent and the internal resources of the enterprise capable of achieving this sub-goal. • The Global Knowledge Module: contains information related to the organisational and operational rules defined by the enterprises. • The Individual Knowledge Module contains information about the agent itself: its capacities and skills, and the current state and workload, i.e. for each skill, indicators are assigned to determine availability, as well as the cost of such skill. V. RO LE S AND B EH AV IOR S O F AG EN T S IP must specify the semantics of the messages exchanged. It also describes the reactions of an agent in a given message. In our case, two main roles involved in the interaction protocols: 1) The Integrator agent supports data extraction from the IP specification and message exchange between agents involved in the interaction. International Journal of Computer Science and Telecommunications [Volume 4, Issue 1, January 2013] 51 knowledges IP Integrator Agent m 0 m 0 m 0 m 0 m 0 Enterp rises Agents Fig. 4. Agents behaviors 2) The role of participants’ agents is to coordinate to satisfy all the requests of the Integrator agent. The interaction protocol that we specified in this paper describes the rules used by the participants’ agents for a correct coordinate. The behavior of agents, as shown in Figure 4, can then be summarized as follows: 1) The Integrator agent (the initiator of the IP) announces the beginning of the integration process and the allo- cation of tasks. For each task to be performed, the Integrator agent announces a specification that includes a description, and a list of requirements and constraints for this task. This information is encapsulated in a message, which represents the initiator message (m0 in Fig. 4). 2) When participants’ agents are identified, the allocation phase of sub goals for participating agents begins. Each participant (Agent) performs the part that has been assigned. The composition of individual results leads to the realization of the overall goal. 3) The Integrator agent ensures the coordination between the different enterprises agents in accordance with the rules of interaction defined in the BPEL4WS specifica- tion. In the context of this work we will not discuss the aspect of local coordination, which may be different from one company to another. To ensure the successful completion of the interaction, the Integrator agent must know how many responses should be ex- pect from the participating agents. The analysis of interaction rules defined in the BPEL4WS specification and the semantics of the ACL allow when if others messages may be received or not. V I . COMM UN IC AT I ON MOD E S This integration process must keep as much as possible the autonomy of architecture core based on agents. Indeed, The agents are coordinated with the Integrator agent and the exchange of messages to enact the EAI. In this architecture, the following communications pathways exist: • agent to agent communication occurs via FIPAs Agent Communication Language (ACL) and is facilitate by a FIPA compliant Agent Management System. kn owl edge bas es Interact ion Manager 1 5 Services Manager 6 4 4 6 WS1 WSn 2 3 UDDI Fig. 5. Agent-Web Ser vi ces Commu ni ca tio n • agent to Web service communication is accomplished via SOAP messages. • agent to BPEL4WS dataspace communication utilizes appropriate protocols/interfaces provided by the datas- pace. The dataspace is used to store BPEL4WS process variables, which maintain the state of the IP. A. Inter-Agent Communication FIPA [14] supplies a standard syntax for messages. These messages are based on the theory of the act of speech, which is the result of the linguistic analysis of human communication. The basis of this theory is to produce an action from the language. In the FIPA-ACL, no specific language for the description of the contents of messages is imposed. Several languages can be used for the description of the contents of the exchanged messages such as KIF (Knowledge Interchange Format), Semantic Language (SL), Prologue and XML (eXten- sible Mark-up Language). XML will be used for the descrip- tion, the specification and the interpretation of the contents of messages exchanged. So, the messages exchanged among the agents are described in FIPA-ACL/XML. The use of XML for the contents of communications among agents permits the display of messages in a Web browser and facilitates the integration with existing Web-based applications. B. Agent - Web service Communication This type of communication is ensured by the Services Manager module of the Integrator agent. This module provides the necessary services for locating and invoking services. Its behavior is presented as follows (see Fig. 5): 1) When the Integrator agent detects a need for which it does not have the capacity , it uses the functionality of Djamel Benmerzoug 52 pub l ic cla ss Integrator exten d s Ag en t { pro te cted vo id se tup () { add Be ha viour(n e w S imp le Beh a vio ur(th is) { import java. uti l.; import jade. cor e. ; import jade. cor e. behavi our.; import j ade.l ang. acl. AC LMessage.; // Tra itemen t .. . cl ass Communication extends C ycl icB ehavi our the Service Manager module to find Web services that could satisfy his need. 2) The Service Manager module requests the UDDI regis try 3) The Service Manager module obtains from the registry a list of descriptions of Web services that matching the search criteria. 4) The Service Manager module starts the invocation of services found to obtain their attributes. 5) The Integrator agent requests the service manager mod- ule to invoke the Web service selected. 6) The Service Manager module sends a parallel invocation message to the selected service and cancel messages to other services not concerned. V I I . SOME I M P L E M E N TAT I O N A S P E C T S The platform chosen for the implementation is Java Agen t } } public class EntrepriseAgent extend s Ag en t { cla ss re ce vo ir e x ten d s Sim p le Beha v iour { //Tra itemen t .. . } pub l ic re ce vo ir( Ag en t a) { su per(a) ; { publ ic C ommuni cati on (A gent a) {super( a );} Vect or goal s = new V ect o r(); publ ic void acti on() { // wait f or message AC LMessage r eceived = m yA gent.r ecei ve(); if(AC LMessage == nul l) { bl ock ();} el se { // message i nt er pr et ati on … … … … … … … … … … // Start P lanR etri eval B ehavi our addBehaviour(new PlanRetrieval(this,goals)); } } } - b - Development framework (JADE) [9]. JADE is a software de- velopment framework, fully implemented in Java, which aims at the development of multi-agent systems and applications that comply with FIPA standard. To achieve such a goal, JADE offers the following list of features to the agent programmer: } pro te cted vo id se tup() { rece vo ir m ybeh a vio ur = ne w rece vo ir( th is); add Be ha viour(m yb eha v iour) ; }} • Distributed agent platform. The agent platform can be distributed on several hosts, each one of them executes one Java Virtual Machine. • FIPA-Compliant agent platform, which includes the Agent Management System, the Directory Facilitator and the Agent Communication Channel. • Efficient transport of ACL messages between agents. A. Using JADE to Develop the Proposed Architecture Creating a JADE agent is as simple as defining a class extending the Jade.core.Agent class and implementing the setup() method. The setup() method is intended to include agent initializations. The actual job an agent has to do is presented as JADE behaviors. The Fig. 6 represents the imple- mentation of the Integrator and the Enterprise agent classes. These classes are extensions of the basic Agent class defined in JADE. The communication module presented in our architecture is derived from JADE class CyclicBehaviour and will therefore run continuously. If no messages have arrived, the behavior will block and restart after a new message has arrived. If a message has arrived, the perception interface has to interpret this message into a set of goals, starts the execution module and resumes waiting for incoming messages. The communi- cation module behavior code is shown in Fig. 6. The Interaction Manager Module is derived from the class SimpleBehaviour. This module takes as input a BPEL4WS specification and builds the appropriate interaction protocol. In this context, an interaction protocol is implemented as an FSMBehaviour behavior. The FSMBehaviour class is a CompositeBehaviour that executes its children (sub behaviour that are going to be used by FSMBehaviours) according to a Finite State Machine defined by the user. In fact, each sub- behaviour is dedicated to an interaction that represents the set of messages exchanged to satisfy a given set of requirements. - a - Fig. 6. a) Agent Cla ss Extension, b) Partial specification of the Communi- cation module V I I I . C O N C L U S I O N In this paper, we have proposed an agent-based architecture for enterprise application integration. The multi-agent provides high level communication languages that preserve the entire richness of the communications. Thus, the FIPA-ACL is used in our approach to formulate messages exchanged between the different agents constituting the system. The content of these m es sag es is described with XML. The use of the JADE platform has allowed us to implement the Integrator Agent and the Enterprise Agents and thus, to simulate the interaction between these entities. R E F E R E N C E S [1] B. Bauer, F. Bergenti, P. Massonet, and J. Odell. Agents and the UML: A Unified Notation for Agents and Multi-agent Systems. In Agent-Oriented Software Engineer ing II, Second Internationa l Workshop, AOSE’01, volume 2222 of LNCS, pages 148–150. Springer, 2001. [2] J. Barjis. Collaborat ive, Participative and Interactive Enterprise Model- ing. In ICEIS’09, volume 24 of Lecture Notes in Business Information Processing, pages 651–662. Springer, 2009. [3] Bernhard Bauer and James Odell. UML 2.0 and ag en ts : how to build agent-ba sed systems with the new UML standard. Eng. Appl. of AI, 18(2):141–157, 2005. [4] Djamel Benmerzoug, Mahmoud Bou faida , and Zizette Bou faida. From the Analysis of Cooperation Within O rganizational Environments to the Design of C oop er at iv e Information Systems: An Agent-Ba sed Approach. In OTM Workshops, volume 3292 of LNCS, pages 495–506, Larnaca, Chypre, October 20 04. S prin ger. International Journal of Computer Science and Telecommunications [Volume 4, Issue 1, January 2013] 53 [5] Djamel Benmerzoug, Mahmoud Bou faida , and Fabrice Kordon. A Specification and Validation Approach for Business Process Integration based on Web Services and Agents. In Proceedings of the 5th Interna - tional Work shop on Modelling, Simulation, Verification and Validation of Enterprise Information Systems, MSVVEIS-2007, In conjunction with ICEIS 2007, pages 163–168. NSTIIC press, 2007. [6] Djamel Benmerzoug, Fabrice Kordon, and Mahmoud Boufaida. A Petri- Net based Formalisation of Interaction Protocols applied to Business Process Integration. In Advances in Enterprise Engineer ing 4th International Work shop on Enterprise & I , Organizational Modeling and Simulation (EOMAS’08), volume 10 of LNBIP, pages 78–92, Montpel- lier, F ranc e, June 2008. Springer. [7] Djamel Benmerzoug, Fabrice Kordon, and Mahmoud Bou faida. For- malisation and Verification of Interaction Protocols for Bu siness Process Integration: a Petri net Approach. International Journa l of Simulation and Process Modelling, 4(3–4):195–204, 2008. [8] P.A. Buhler and J .M. Vida l. Towards adaptive work flow enactment u sing multiagent systems. Interna tiona l Journal On Information Technology and Management, 6:61–87, 2005. [9] C. Huang C. Trappey, A. Trappey and C . Ku. The design of a JADE - based autonomou s work flow management system for collaborative SoC design. Expert Syst. Appl., 36(2):2659–2669, 2009. [10] Chandra S. Mann V. Cha fle, G. and M. Nanda. Decentralized orchestra- tion of composite web services. In The Alternate Track on Web Services at the 13th International World Wide Web Conference, pages 134–143, 2004. [11] CPN -AMI:. http://move.lip6.fr/software/cpnami/. [12] Gero Decker and Mathia s Weske. Interaction-centr ic modeling of process choreographies. Journal of Information Systems, 36(2):292–312, 2011. [13] Nirmit Desai, Amit K. Chopra, and Munindar P. S ing h. Amoeba: A methodology for modeling and evolving cross-organizational bu siness processes. ACM Trans. Softw. Eng. Methodol., 19(2):1– 45, October 2009. [14] FIPA. Foundation for Intelligent Physical Agents: Communicative Act Library Specification, 1997. [15] G. Booch, J . Rumbaugh, and I. Jacobson. The Unified Modelling Language for Object-Oriented Development, Document Set Version 1.0. Rational Software Corporation, Santa Clara, 2002. [16] H. Mazouzi, A. Fallah-Seghrouchni, and S. Hadd ad . Open Protocol De- sign for Complex Interactions in Multi-Agent Systems. In AAM A S ’02 : Proceedings of joint conference on international first the Autonomou s agents and multiagent systems, pages 517–526, New York, NY, USA, 2002. ACM. [17] SAP Siebel Systems IBM, Microsoft. Business process execution language for web services version 1.1. Technical report, 2003. [18] M. Venkatraman and P.S Munindar. Verifying compliance with com- mitment protocols. Int. Journal of Autonomous Agents and Multi- Agent Sys tems, 2(3):217 – 236, 1999. [19] F. Mar tial. Coordinating Plans of Autonomous Agents, volume 610. Spr inger Heidelberg, 1992. [20] Mike P. Papazoglou, Klaus Pohl, Michael Parkin, and Andrea s Metzger, editors. Service Research Cha llenges and Solutions for the Future Internet - S-Cube - Towards Engineering, Managing and Adapting Service-Ba sed Systems, volume 6500 of Lecture Notes in Computer Science. Springer, 2010. [21] H. V. Parunak. Visualizing Agent Conversations: Using Enhanced Dooley Graphs for Agent Design and Analysis, 1996. [22] Wil M. P. van der Aalst, Arjan J. Mooij, Christian Stahl, and Karsten Wolf. Service interaction: Patterns, formalization, and analysis. In Marco Bernardo, Luca Padovani, and Gianluigi Zavattaro, editors, Formal Methods for Web Services, 9th International School on Formal Methods for the Design of Computer, Communication, and Software Sys tems , Lecture Notes in Computer Science, pages 42–88. Spr inger, 2009. [23] Johannes Maria Zaha, Marlon Duma s, Arthur H. M. ter Hofstede, Alistair P. Barros, and Gero Decker. Service interaction modeling: Bridging global and local views. In Tenth IEEE International Enterprise Distr ibuted Object Computing Conference (EDOC 2006), pages 45–55. IEEE Computer Society, 2006.
1112.0404	2	1112	2012-02-06T10:36:03	A Cyclic Representation of Discrete Coordination Procedures	[ "cs.MA", "cs.DM", "eess.SY", "math.OC" ]	We show that any discrete opinion pooling procedure with positive weights can be asymptotically approximated by DeGroot's procedure whose communication digraph is a Hamiltonian cycle with loops. In this cycle, the weight of each arc (which is not a loop) is inversely proportional to the influence of the agent the arc leads to.	cs.MA	cs	ISSN 0005-1179, Automation and Remote Control, 2012, Vol. 73, No. 1, p. 161. c(cid:13) Pleiades Publishing, Ltd., 2012. Original Russian Text c(cid:13) R.P. Agaev and P.Yu. Chebotarev, 2012, published in Avtomatika i Telemekhanika, 2012, No. 1, pp. 178 -- 183. MULTI-AGENT SYSTEMS A Cyclic Representation of Discrete Coordination Procedures1 R. P. Agaev and P. Yu. Chebotarev Trapeznikov Institute of Control Sciences, Russian Academy of Sciences, Moscow, Russia Received May 20, 2011 Abstract -- We show that any discrete opinion pooling procedure with positive weights can be asymp- totically approximated by DeGroot's procedure whose communication digraph is a Hamiltonian cycle with loops. In this cycle, the weight of each arc (which is not a loop) is inversely proportional to the influence of the agent the arc leads to. 1. INTRODUCTION One of the first discrete models of reaching consensus (decentralized coordination) was proposed by n)T is the vector of initial opinions of the members of a group DeGroot [1]. Suppose that s(0) = (s0 and s(k) = (sk n)T is the vector of their opinions after the kth step of coordination. In accordance with DeGroot's model, s(k) = P s(k − 1), k = 1, 2, . . . , where P is a row stochastic influence matrix whose entry pij specifies the degree of influence of agent j on the opinion of agent i. Thereby, 1, . . . , sk 1, . . . , s0 s(k) = P ks(0), k = 1, 2, . . . . (1) A consensus is [asymptotically] reached if limk→∞ sk i = ¯s for some ¯s ∈ R and all i ∈ {1, . . . , n}. De- Groot states that a consensus is reached for any initial opinions if and only if the matrix P ∞ = limk→∞ P k exists and all rows of P ∞ are identical, which is equivalent to the regularity1 of P . If P is not regular, then the opinions do not generally tend to agreement. Yet, a consensus can be reached if the vector of initial opinions belongs to a certain subspace. In [2], we give a characterization of this subspace and propose the method of orthogonal projection which generalizes DeGroot's method. It is shown that in the method of orthogonal projection, as well as in DeGroot's method, no nonbasic agent can affect the final result. In this paper, we show that the result of any DeGroot's procedure with a strong communication digraph as well as the result of any procedure of orthogonal projection without nonbasic agents can be represented by DeGroot's procedure whose communication digraph is a Hamiltonian cycle with loops. In this cycle, the weight of each arc (which is not a loop) is inversely proportional to the influence of the agent the arc leads to. 2. BASIC CONCEPTS AND RESULTS With a stochastic influence matrix P in DeGroot's model we associate the communication digraph Γ with vertex set V (Γ) = {1, . . . , n}. Γ has a (j, i) arc with weight wji = pij whenever pij > 0 (i.e., whenever agent j influences agent i). Thus, arcs in Γ are oriented in the direction of influence; the weight of an arc is the power of influence. 1 This work was supported in part by the Russian Foundation for Basic Research, project no. 09-07-00371 and the program of RAS Presidium "Mathematical Theory of Control." 1 A stochastic matrix is said to be regular if it has no eigenvalues of modulus 1 except for the simple eigenvalue 1. 161 162 AGAEV, CHEBOTAREV The Kirchhoff matrix (see [3, 4]) L = L(Γ) = (ℓij) of digraph Γ is defined as follows: if j 6= i, then wki, i = 1, . . . , n. I will denote ℓij = −wji whenever Γ has the (j, i) arc and ℓij = 0 otherwise; ℓii = Pk6=i the identity matrix of appropriate dimension. By virtue of the above definitions, for the digraph Γ associated with P we have L(Γ) = I − P. (2) Any maximal by inclusion strong (i.e., with mutually reachable vertices) subgraph of a digraph is called a strong component (or a bicomponent) of this digraph. A basic bicomponent is a bicomponent such that the digraph has no arcs coming into this bicomponent from outside. Vertices belonging and not belonging to basic bicomponents are called basic and nonbasic, respectively. Similarly, we call an agent basic/nonbasic whenever the vertex representing this agent is basic/nonbasic. Let b and ν be the number of basic vertices and the number of basic bicomponents in Γ, respectively. If a consensus in DeGroot's model is reached and vertex j is nonbasic, then, as stated in [1], column j in the limiting matrix P ∞ is zero and the initial opinion of agent j does not affect the resulting opinion. Now we present the results from algebraic graph theory used in this paper. A longer list of results useful for decentralized control is given in [2]. If the sequence of powers P k of a stochastic matrix P has a limit P ∞, then P ∞ = J, (3) where J is the normalized matrix of maximum out-forests of the corresponding weighted digraph Γ (a corollary of the matrix tree theorem for Markov chains [5]). P ∞ is the eigenprojection corresponding to 0 (principal idempotent) of L and rank P ∞ = ν; rank L = n − ν, (4) where ν is the number of basic bicomponents in Γ [6, Proposition 11]. By (4) dim N (L) = ν, where dim N (L) is the dimension of the kernel (the nullity) of L. Finally, by [4, Proposition 12], ind L = 1, where ind L (the index of L) is the order of the largest Jordan block of L corresponding to the zero eigenvalue. This implies that mL(0) = ν, (5) where mL(0) is the multiplicity of 0 as an eigenvalue of L. 3. CONVERGENCE IN DEGROOT'S MODEL AND PROPERTIES OF THE COMMUNICATION DIGRAPH As noted above, DeGroot's method with matrix P leads to a consensus for any initial opinions if and only if there exists a limiting matrix P ∞ = limk→∞ P k and all its rows are identical. The equality of the rows of P ∞ implies that P ∞ = 1πT for some probability vector (the components are non-negative and sum to 1) π, where 1 = (1, . . . , 1)T. In this case, the consensus ¯s is expressed by the inner product of the vectors π and s(0): s(∞) = P ∞s(0) = 1πTs(0) = 1¯s, (6) where s(∞) is the resulting vector of opinions, π is the final weight distribution of the DeGroot algorithm, and ¯s = πTs(0) is the consensus. A probability vector π is called a stationary vector of a stochastic matrix P if it is a left eigenvector of P corresponding to the eigenvalue 1: πTP = πT. Obviously, this condition is satisfied for the vector π in the representation P ∞ = 1πT of P ∞, provided that the convergence of DeGroot's method is guaranteed by the regularity of P. AUTOMATION AND REMOTE CONTROL Vol. 73 No. 1 2012 A CYCLIC REPRESENTATION OF DISCRETE COORDINATION PROCEDURES 163 By Theorem 3 in [1], if for any vector of initial opinions s(0), DeGroot's method converges to the consensus πTs(0), then π is a unique stationary vector of P . Let us formulate a criterion of convergence in DeGroot's model in terms of the communication di- graph Γ. The equality of the rows of P ∞ is equivalent to rank P ∞ = 1. Therefore, owing to (4), when the sequence P k converges, consensus is reached for any initial opinions if and only if the communication digraph Γ corresponding to P has a single basic bicomponent (ν = 1). Consequently, provided that the sequence P k converges, ν = 1 is equivalent to the regularity of P . In turn, by (5), this is the case if and only if 0 is a simple eigenvalue of L. Finally, ν = 1 if and only if Γ has a spanning out-tree (also called arborescence and branching) [6, ij ) = J = ( Jij) is the normalized matrix of spanning out-trees: Proposition 6]. In this case (see (3)), P ∞ = (p∞ ij = πj = Jij = p∞ tj t , i, j = 1, . . . , n, (7) where tj is the total weight2 of Γ's spanning out-trees rooted at j and t is the total weight of all spanning out-trees of Γ. A survey of some results on DeGroot's iterative pooling model and its generalizations can be found in [7 -- 9]. Note that one of the new applications of DeGroot's model is information control in social networks [10]. 4. REPRESENTING CONSENSUS PROCEDURES BY WEIGHTED HAMILTONIAN CYCLES WITH LOOPS As shown in [2, Section 6], the final result of the method of orthogonal projection can be represented by a weight vector α: the inner product of α and the vector of initial opinions gives the consensus: ¯s = αTs(0). This is analogous to the representation of the result of a convergent DeGroot's method: P ∞ = 1πT and ¯s = πTs(0) (see (6)). On the other hand, given a probability vector π, it is easy to construct a weighted communication digraph generating π as the final weight distribution (the stationary vector of P ) of DeGroot's method. In this section, we show that all positive weight distributions on the set of agents' opinions are generated by a rather narrow class of digraphs, namely, Hamiltonian cycles of the form n → (n − 1) → · · · → 2 → 1 → n with loops, where the vertices are denoted by 1, . . . , n. By (3) P ∞ = J, where J is the normalized matrix of maximum out-forests of the weighted digraph Γ corresponding to P. Thus, given π, a necessary and sufficient condition of the fulfilment of P ∞ = 1πT is J = 1πT. (8) A Hamiltonian cycle is a strong digraph, so its maximum out-forests are precisely spanning out-trees (see (7)). Recall that the normalized matrix of spanning out-trees (which in this case coincides with J) is the matrix whose (i, j)-entry is tj/t, i, j = 1, . . . , n. Adding loops does not change the set of out-trees. Thus, to solve the problem, i.e., to implement a given probability vector π > 0 as the final weight distribution of DeGroot's method with a communication digraph in the form if a Hamiltonian cycle with loops, it is sufficient to construct a weighted cycle whose vector (t1, . . . , tn)T is proportional to π. Indeed, in this case (t1/t, . . . , tn/t)T coincides with π, which guarantees (8). Such a cycle can be constructed by means of the following lemma. Lemma 1. For any positive vector q = (q1, . . . , qn)T, there exists a unique weighted Hamiltonian cycle of the form n → (n − 1) → · · · → 2 → 1 → n whose vector (t1, . . . , tn)T of total weights of out-trees coincides with q. The weight of the arc entering vertex k in this cycle is q−1 i=1 qi, k = 1, . . . , n. k · n−1qQn 2 The weight of an out-tree (and, more generally, of a digraph) is the product of the weights of all its arcs. AUTOMATION AND REMOTE CONTROL Vol. 73 No. 1 2012 164 AGAEV, CHEBOTAREV Fig. 1. The proof of Lemma 1 is given in the Appendix. Now we apply Lemma 1 to solve our problem. Proposition 1. For any positive probability vector π = (π1, . . . , πn)T, there exists a family of weighted Hamiltonian cycles of the form n → (n − 1) → · · · → 1 → n with loops such that using each of them as the communication digraph in DeGroot's method implements the consensus ¯s = πTs(0) for any vector of initial opinions s(0). The weight of the arc entering vertex k in such a cycle is proportional to π−1 k . As stated above, any communication digraph whose vector (t1, . . . , tn) of the weights of out-trees is proportional to π implements π as the final weight distribution of DeGroot's method. Therefore, to prove Proposition 1, it is sufficient to observe that by Lemma 1, a Hamiltonian cycle with loops and any given order of visiting vertices can be taken as such a digraph. The only thing that should be taken care of is that this cycle must be a communication digraph, i.e., the matrix I − L (see (2)) corresponding to it must be stochastic. Since L has zero row sums and nonpositive off-diagonal entries, I − L is stochastic if and only if the diagonal entries of L are less than or equal to 1. For a Hamiltonian cycle with loops, this condition is satisfied if and only if all its arc weights do not exceed 1. Example 1. Consider the influence matrix P =   0.9 0.1 0 0 0.75 0.25 0.3 0.25 0.2 0.15 0 0 0.1 0.35 0 0.65   and the corresponding communication digraph (Fig. 1a, which does not show loops for simplicity). Proposition 1 enables one to construct another communication digraph, in the form of a Hamiltonian cycle with loops, that shares the final consensus obtained through DeGroot's method with the given digraph for any initial opinions. Observe that P is regular and its stationary vector is: π = 1 101 (45, 36, 10, 10)T . As stated in Section 3, H = P ∞ = J . Proposition 1 P ∞ = 1πT. Now we construct a Hamiltonian cycle H with loops such that P ∞ implies that the vector of arc weights indexed by the vertices of H can be x = 101β( 1 10 )T, where 0 < β ≤ 10 18 , 1, 1)T. In this communication digraph (which is shown in Fig. 1b), loops with weights 7 18 must be attached to vertices 1 and 2, respectively. It follows from Proposition 1 that if the weights of arcs entering each vertex are preserved, then the order of vertices in the cycle can be arbitrary. In particular, the cycle in Fig. 1c is also suitable. 101 we obtain x = ( 2 101 . Taking β = 10 9 and 13 45 , 1 36 , 1 10 , 1 9 , 5 Thus, for each cycle with loops of this type, DeGroot's method leads to the same final consensus procedure (determined by the weight distribution) as for the original communication digraph. Since in Proposition 1, πi is the weight of the original opinion of ith agent in the final consensus (the "influence" of the ith agent), it is worth noting that for communication digraphs in the form of AUTOMATION AND REMOTE CONTROL Vol. 73 No. 1 2012 A CYCLIC REPRESENTATION OF DISCRETE COORDINATION PROCEDURES 165 Hamiltonian cycles with loops, this weight is inversely proportional to the weight of the arc entering vertex i. Thus, in the upshot, the most powerful agent is the one least subject to the influence of the previous agent in the cycle rather than the one maximally affecting the next agent. Finally, Proposition 1 enables one, for any given coordination procedure based on the method of orthogonal projection [2], to construct a DeGroot algorithm (with the communication digraph in the form of a Hamiltonian cycle) that leads to the same consensus for any initial opinions. 5. CONCLUSION It is shown that any convergent discrete iterative pooling procedure taking into account the opinions of all agents with positive weights can be approximated (in terms of achieving the same end result) by DeGroot's procedure whose communication digraph is a Hamiltonian cycle with loops. The weight of the arc entering vertex i in this cycle is inversely proportional to the influence of agent i, while the order of visiting vertices can be arbitrary. APPENDIX Proof of Lemma 1. For any vertex in a Hamiltonian cycle, there is exactly one tree outgoing from this vertex. That tree includes all arcs of the cycle except for the arc entering this vertex. Hence the elements of the vector (t1, . . . , tn)T of total weights of out-trees are given by the equations tk = Y j6=k xj = x−1 k n Y j=1 xj, k = 1, . . . , n, (A.1) where xj is the weight of the arc entering vertex j. Consequently, xk = t−1 k n Y j=1 xj = t−1 k n−1vuut n Y i=1 Y j6=i xj = n−1qQn i=1 ti tk , k = 1, . . . , n. Thus, for the fulfillment of tk = qk (k = 1, . . . , n) it is necessary that xk = n−1qQn i=1 qi qk , k = 1, . . . , n. (A.2) The sufficiency of (A.2) is verified by substituting (A.2) into (A.1). The lemma is proved. ⊓⊔ REFERENCES 1. DeGroot, M.H., Reaching a Consensus, J. Amer. Statist. Associat., 1974, vol. 69, pp. 118 -- 121. 2. Agaev, R.P. and Chebotarev, P.Yu., The Projection Method for Reaching Consensus and the Regularized Power Limit of a Stochastic Matrix, Automat. Remote Control, 2011, vol. 72, no. 12, pp. 2458 -- 2476. 3. Tutte, W.T., Graph Theory. Reading, MA: Addison-Wesley, 1984. 4. Chebotarev, P. and Agaev, R., Forest Matrices Around the Laplacian Matrix, Linear Algebra Appl., 2002, vol. 356, pp. 253 -- 274. 5. Wentzell, A.D. and Freidlin, M.I., On Small Random Perturbations of Dynamical Systems, Russian Math. Surveys, 1970, vol. 25, no. 1, pp. 1 -- 55. 6. Agaev, R.P. and Chebotarev, P.Yu., The Matrix of Maximum Out Forests of a Digraph and Its Applications, Automat. Remote Control, 2000, vol. 61, no. 9, pp. 1424 -- 1450. AUTOMATION AND REMOTE CONTROL Vol. 73 No. 1 2012 166 AGAEV, CHEBOTAREV 7. Jackson, M.O., Social and Economic Networks. Princeton, NJ: Princeton Univ. Press, 2008. 8. Agaev, R.P. and Chebotarev, P.Yu., Convergence and Stability in Consensus and Coordination Problems (A Survey of Basic Results), UBS, 2010, vol. 30, no. 1, pp. 470 -- 505. In Russian. 9. Gilardoni, G.L. and Clayton, M.K., On Reaching a Consensus Using DeGroot's Iterative Pooling, Ann. Statist., 1993, vol. 21, pp. 391 -- 401. 10. Barabanov, I.N., Korgin, N.A., Novikov, D.A. and Chkhartishvili, A.G., Dynamic Models of Informational Control in Social Networks, Automat. Remote Control, 2010, vol. 71, no. 11, pp. 2417 -- 2426. AUTOMATION AND REMOTE CONTROL Vol. 73 No. 1 2012
cs/0505045	1	0505	2005-05-17T13:06:41	A T Step Ahead Optimal Target Detection Algorithm for a Multi Sensor Surveillance System	[ "cs.MA", "cs.RO" ]	This paper presents a methodology for optimal target detection in a multi sensor surveillance system. The system consists of mobile sensors that guard a rectangular surveillance zone crisscrossed by moving targets. Targets percolate the surveillance zone in a poisson fashion with uniform velocities. Under these statistics this paper computes a motion strategy for a sensor that maximizes target detections for the next T time steps. A coordination mechanism between sensors ensures that overlapping areas between sensors is reduced. This coordination mechanism is interleaved with the motion strategy computation to reduce detections of the same target by more than one sensor. To avoid an exhaustive search in the joint space of all sensors the coordination mechanism constraints the search by assigning priorities to the sensors. A comparison of this methodology with other multi target tracking schemes verifies its efficacy in maximizing detections. A tabulation of these comparisons is reported in results section of the paper	cs.MA	cs	Abstract – This paper presents a methodology for optimal target detection in a multi sensor surveillance system. The system consists of mobile sensors that guard a rectangular surveillance zone crisscrossed by moving targets. Targets penetrate the surveillance zone with poisson rates at uniform velocities. Under these conditions we present a motion strategy computation for each sensor such that it maximizes target detection for the next T time-steps. A coordination mechanism among sensors ensures that overlapping and overlooked regions of observation among sensors are minimized. This coordination mechanism is interleaved with the motion strategy computation to reduce detections of the same target by more than one sensor for the same time-step. To avoid an exhaustive search in the joint space of all the sensors the coordination mechanism constrains the search by assigning priorities to the sensors and thereby arbitrating among sensory tasks. A comparison of this methodology with other multi target tracking schemes verifies its efficacy in maximizing detections. A tabulation of these comparisons is reported in the results section of this paper. “Sample” and “time-step” are used equivalently and interchangeably in this paper. Index Terms- multi sensor system, optimal target detection, sensor network, sensor surveillance, multi-agent system I INTRODUCTION A T-Step Ahead Optimal Target Detection Algorithm for a Multi Sensor Surveillance System Henry Hexmoor and Shravan Sogani K Madhava Krishna International Institute of Information Technology CSCE Department, University of Arkansas Fayetteville, AR 72701 Hyderabad – 500 019, INDIA [email protected], [email protected] [email protected] strategy of a sensor that maximizes the number of detections for the next T time steps. The surveillance zone is discretized into a lattice of cells. Each cell stores information about the expected number of targets to be seen at that cell position purely based on target arrival statistics. This is done offline at all the cell locations across the entire surveillance zone. Then during the real-time phase of the algorithm an estimate of the number of detections based on combination of current target information and target arrival statistics is computed. These values are computed not across the entire lattice but for cells that are at a level T from current positions of all the sensors. Cells that are at a level one from the sensor’s current location store detection values for the next T samples, while those that are at level T from the sensor store information only for time step T. By level one we mean the immediate set of eight neighbors in the lattice from the current location of the sensor and level T indicates cells on lattice that are T neighbors away in an eight-connected sense. A depth limited search, with the depth limited by T, maximizing the number of detections performed on the lattice from the sensor’s current position yields the motion strategy for the sensor. Computing the best strategy from its individual perspective can result in motions for each sensor with overlapping regions of observation. A priority based coordination mechanism is interleaved with the motion strategy computation to reduce detections of the same target by more than one sensor that further enhances the detection performance for the entire system. To avoid an exhaustive search in the joint space of all the sensors the coordination mechanism constrains the search by assigning priorities to the sensors. Hence while optimal from a single sensor’s perspective the algorithm is not optimal across all sensors – a choice that one is reasonable to assume in light of prevailing real time demands. The details of this method are presented in section 3 of this paper. Extensive simulation results confirm the efficacy of this strategy in the form of tabular comparisons presented in section 4. The proposed approach is particularly suitable for guarding large open areas that are crisscrossed by moving targets when the number of sensors at disposal for monitoring them is limited. Due to limited number of sensors as well as to glean characteristics of targets over several observations the sensors are allowed to be move. Such surveillance systems find utilities in many security, surveillance and reconnaissance applications. We present a framework for maximizing target detections in a surveillance system consisting of multiple mobile sensors (i.e., robots). The sensors collectively guard a rectangular surveillance zone crisscrossed by moving targets. A-priori information about targets is the knowledge of their poisson rates of entry. Instantaneous or current knowledge about all targets within the sensing range of a sensor is available to the sensor in the form of the target’ velocity information and motion direction. While current information is used in deterministically computing the best places to be at in the subsequent time steps, it does not take into consideration the changes in the environment that occur due to consistent flow of target traffic into and out of the surveillance zone. Knowledge of the statistics of target entry and exit on the other hand are used to compute the statistically optimal places to be at in the future from a given location, but this does not consider the current information gleaned through sensor observations. Hence a combination of both current target information and target statistics is used to compute a motion The rest of the paper is organized as follows. Section 2 places the current work in the context of similar works found in the literature. Section 3 presents the formulation of the methodology and section 4 attests to the efficacy of the methodology in our simulation work. Section 5 concludes and provides further scope for this work. II. BACKGROUND REVIEW Many reconnaissance and surveillance, security, applications require distributed autonomous observation of movements of targets navigating in a bounded area of interest. Multi-sensor surveillance finds applications such as in border patrol, guarding of secured areas, search and rescue and warehouse surveillance [1, 2]. It involves detection of multiple intrusions and/or tracking through coordination between the sensors. Detection and target tracking has been researched from multiple viewpoints. Some efforts have focused on the problem of identifying targets from a given set of data through particle filters [3], and probabilistic methods [4]. The problem of data association or assigning sensor measurements to the corresponding targets were tackled by Joint Probabilistic Data Association Filters by the same researchers such as in [3]. Kluge and others [5] use dynamic timestamps for tracking multiple targets. Krishna and Kalra [6] presented clustering based approaches for target detection and further extended it to tracking and avoidance. The focus of these approaches has been on building reliable estimators for predicting target trajectories that is different from the objective of this effort to maximize target detections. In the context of distributed task allocation and sensor coordination Parker proposed a scheme for delegating and withdrawing robots to and from targets through the ALLIANCE architecture [7]. The protocol for allocation was one based on “impatience” of the robot towards a target while the withdrawal was based on “acquiescence”. Jung and Sukhatme [8] present a strategy for tracking multiple intruders through a distributed mobile sensor network and a strategy for maximizing sensor coverage[8, 9]. Lesser’s group have made significant advances to the area of distributed sensor networks [10] and sensor management [11]. In [12] Parker presents a scheme called A-CMOMMT where the goal is to maximize the number of targets observed over a time interval of length T based on the same philosophy of behavior-based control as in [7]. The authors of this paper present their scheme for resource allocation and coordination in a distributed sensor system through a set of fuzzy rules in [13] and further compare various resource allocation strategies in terms of their detection performance in [14]. The author of [15] has looked at the problem of static placement of sensors in known polygonal environments and [16] describes a distributed sensor approach to target tracking using fixed sensor locations. The current approach is disparate from those of [15,16] in that in the current scheme the sensors are mobile. Among the approaches that we have encountered the closest to this effort are [12] and [8]. In [12] a behavior-based approach, A-CMOMMT, is compared with three other heuristic approaches where the sensor’s motion strategy is arbitrary or random in the first, stationary (the sensor does not move) in the second and based on local force control in the third. However the approaches including A-CMOMMT are heuristic in that while effective they do not guarantee that the motion strategy generated is actually the best possible strategy for those T time steps based on the gathered data. In [8] a motion strategy for tracking multiple targets based on density estimates is presented. The robot attempts to maximize target detections by maintaining itself at a particular distance from the center of gravity of currently observed targets. It is nonetheless not clear if this could again be the best strategy. In contrast, our approach presents a scheme that moves the sensor to provably the optimal regions in the surveillance region in order to maximize detections over the next T samples. In other words, through a depth limited search the sensor comes up with the best possible motion strategy that results in maximizing detections over next T time steps among the various possible strategies from the current sensor location. This is combined with the coordination mechanism to reduce overlapping areas of observation positions this method as novel and different from other approaches the authors have come across in literature. III. THE METHODOLOGY A. Description of Surveillance Zone, Sensors and Targets: We consider the surveillance system depicted in figure 1. The sides of the outer rectangle or the biggest rectangle in figure 1 form the boundary of the surveillance zone – the area enclosed by it is the area of interest where sensors attempt to optimize their rates of detection. The shaded circles are effective sensor ranges in their starting positions. The field of vision (FOV) of a sensor is 360 degrees. The squares with thick boundaries in figure 1 with the sensors at their center are the inscribed squares of the circular FOV of a sensor. In other words, the diameter of the FOV of a sensor is the diagonal of one such inscribed square. Purely from the point of view of facilitating easier computations the sensor considers only those targets that lie within its inscribed square as targets within its FOV. It needs to be emphasized that this simplification does not have any bearing on the overall philosophy of this approach. In the results section the efficacy of this method is verified by uniformly applying this same condition across all other approaches that are compared and the extension to a case of circular FOV is merely one of more involved but computable computations. The entire surveillance region is discretized into a lattice of cells. The cells are represented as the small squares inside the FOV of the leftmost and topmost sensor. The dashed lines along the length and breadth indicate that the cells proceed to fill the entire surveillance zone. At each of the cell locations various aspects of target statistics are computed that are described later. The crosses outside the surveillance zone are the source points from where targets emanate as per Poisson statistics. Targets percolate from each of those sources into that horizontal or vertical half-plane that contains the surveillance zone. Therefore, all targets coming from a particular source will be contained with an angular span of πradians. Furthermore, the following assumptions are made for sensors and targets • A sensor can detect all targets within its FOV or occlusion relations are not considered. • The takeoff angle of a target from its source point is uniformly distributed in [ ]π,0 • All targets move with the same uniform velocity within the surveillance zone along linear trajectories, which can be ascertained by the sensor. The last assumption allows that the statistical values of various parameters computed at every cell in the lattice to take a unique value rather than a probability distribution. In case of a distribution the expected values of the parameters need to be made use of. Figure 1: The surveillance zone is represented by the outer rectangle. The sensors by shaded circles, the crosses outside are the target sources and the small squares within the FOV of top and left sensor are the discretized cells. The dashed lines along the length and breadth indicate that the cells proceed to fill the entire surveillance zone. B Problem Statement and Approach The problem attacked in this paper is stated as follows. Given: SN : The set of all sensors in the system, • { }s , where si denotes the sensor with sn s N s ,...,1,0= S label i , ordered in a sequence. Hence sn the label of the last enumerated sensor is also the number of sensors in the system, which is a constant. tIN , : The set of all targets in the system at time t, where, { }tI i N in i tIn , is the number of targets , and ,...,1,0= tI , , in the system at time step t that varies at every time step and hence dynamic. • J g = • tm , is maximized. In following cost function :,tmg A binary variable that takes the value 1 if a target sn sensors at t. im is observed by any one of the Objective: To develop an algorithm such that the n T tI , ∑ ∑ t m 1 1 = = other words, the number of detections of the targets present in the system at every time step is maximized over T time steps. Approach: The methodology is delineated as a sequence of steps below and some of the steps are elaborated therein and some are explained subsequently. Due to brevity of space, some of the computations mentioned in these steps that peripheral to the core algorithm are not detailed. 1) At each cell, Pi , in the lattice the following are computed and stored. 1a). Piλ : The expected rate of target entry into the FOV of a sensor centered at Pi . If there are Q target sources each Q θ j ∑ = λλ Pi π j 1 = emanating targets at rate λthen , where the ratio the rate at which the sensor at Pi sees targets denotes the expected fraction of the total number of θj π targets that would enter the FOV of the sensor at Pi from the jth of the Q target sources. In figure 1 Q=8. We also denote by θ j jPi =, λ λ π from source j. In other words, among the targets taking off at any angle within [ ]π,0 , from source j only those that takeoff within the angular span [ ]j would enter the FOV of , θδδ + j j the sensor, where jδ is the smallest angle formed by the segment connecting a vertex of the FOV square and source j is the largest angle. and j θδ + j PieT 1b) : The expected escape time for a target through the field of vision of the sensor centered at Pi. The escape time is essentially the time for which a target would be in the FOV before it escapes from a sensor. 1c) d eT : The number of targets expected to be λ= Pi Pi Pi detected by a sensor at Pi for any time step. 1d) jPi ,η : The normalization constant at cell Pi due to target source j. This constant renders the summation of the probability distribution function (pdf) of the escape time of a target through the FOV of a sensor at Pi due to source j to go te , to unity. In other words if is the random variable that jPi measures the escape time of a target for a sensor at cell Pi )jPi ( due to source j and tef jPi ,η sees to that be its pdf then , t b ( ) ∫ tef t a integral are the minimum and maximum possible escape , where the lower and upped bounds of the jPi , = 1 )jPi ( tef times. It is to be noted that in general is not the , same in [ ]b a t t and needs to be broken into sub functions for , respective time intervals for which it has the same closed form representation. The details of how these sub probability distribution functions are computed is not given here for conciseness. 2) For every sensor sj in the system located at a particular cell location Pj the path it takes for the next T time steps is computed such that the number of target detections is maximized based on steps 3 and 4. 3) For sj currently at Pj compute for all cells, Pk , that are at level n, n = T from Pj , the expected number of } ,...,2,1,0{ tk dn , , where the first index, k, in the superscript detections, refers to the cell Pk and the second index t denotes the time step for which it was computed. The computation is based on what sj currently observes at Pj and the target arrival statistics, elements of which are computed in step 1. This computation is carried out at the current cell Pj and those cells one time-step away for the next T time steps. This is done because the sensor can choose to remain at the current cell itself for the next T time steps or migrate to a neighboring cell and stay there for next T-1 steps or come back to the starting cell for any of the future T steps depending on whichever of those various possible strategies yields maximum detections. For cells at depth two this computation is carried T-1 times, corresponding to time steps starting from 2 and ending at T with T inclusive. For a cell at depth T the computation is done once for time step T alone. This results in a tree where each cell at time step t is connected to its eight 1+t . Since the cell that nearest cells as well as itself at connects to itself does for a different time step the two cells though at the same location are different due to difference in time as their information content varies with every time step. Figure 2 shows an instance of the tree, where a cell location i tj labeled as connects to its eight neighbors as well as ,( ), itself at t+1, where j i represents the index of the cell in the ) ,( two dimensional lattice and t its index in time. Thus every node in the tree except the leaf node is connected to nine nodes and the depth of the tree is T. 4) On this nine connected space-time tree perform a depth- tk dn , limited search to maximize the detections with as the objective value of a link that connects a current cell node Pj at t-1 to another node Pk for the next time step t. 5) The paths of the sensors are ordered based on the number of detections. The sensor path that has the highest number of detections has the highest priority. Ties are broken randomly. 6) For overlapping FOV between sensors at certain locations in the paths the number of detections is reduced at the corresponding cell for the corresponding time-step for the sensor with lower priority. The reduction is computed as tk tk tk , , , . The fraction f is proportional to the n n nf = − d d d area of overlaps given by the area of overlapping FOV divided by the area of the FOV. 7) The optimal path for the sensor with the lower priority is tk dn , values through depth recomputed based on the updated limited search. For a sensor with priority p, we check for overlaps with all those sensor trajectories with priorities 1, 2, …, p-1. 8) The steps two to seven are repeated until the end of simulation. Step 1 of the algorithm is performed only once at the start of the simulation across all the cells and is essentially an offline step unless the target statistics changes dynamically. Figure 2: An instance of the nine connected spacetime tree, where every node except the leaves connects to its eight neighbors as well as itself at the next time step. tk dn , B Computing tk dn , Computation of varies if it is being carried out for the same cell location for a future time or for a different cell location. For the case of computing at the same cell the procedure is as follows: Given that the sensor at cell Pi currently or at t=0 observes Pin number of targets then the number of targets it is likely to see T time-steps into the future is given by: Q T ( ∑ ∑ teP λ λ jPi Pi , j t 1 1 = = tT −< k +− n Ti , d = n jPi , Pi T − ) (1) te Here have the same connotations as , , λ λ Pi jPi jPi , , discussed in step 1 and Q is the number of target sources as before. The first term on the right hand side or RHS of equation (1) is the deterministic part that is computed purely based on what the sensor senses now. It merely states that out particle seen currently k of them would disappear by T of Pin time steps. The second term onwards on the RHS denote the TPiλ statistical counterpart. It says that are likely to enter the FOV in T time steps and of which the number in the third term are likely to leave the FOV by T. The third term containing the ( )tT teP denotes summation is explained as follows. −<, jPi the probability that a target from source j that entered the FOV t would have escaped the FOV by T . Hence at ( )tT λ teP is the estimate of the number of −<, jPi jPi , λ is the particles that entered at t and escaped by T , since jiP , estimate of number of particles entering for every t from j into the FOV of the sensor at Pi . The summation over t signifies jPi , Pi ( teP that estimate of the number of particles that would have left by T needs to be done in correspondence with the time at which they entered between now and T. The summation over Q te , signifies that the pdf of is not the same for every source jPi ( )tT j at that cell location. The teP is given below by −<, jPi its pdf with notations from step 1.  0  1   η  jPi ,  t tT 1 ; >−  b The number of candidates likely to be seen for a different cell Pk at T given that Pin number of targets then the number of targets are currently seen at Pi is given by:  Q T ( ∑ ∑  teP λκ λ Pi jPi ,   j t 1 1 = = ) tT =−< tT −< tT ≤−≤ k +− n Tk , d tT <− t T ∫ t a = n T − t b . jPi , tef ( jPi , ); t a ; a ) Pkd ( (2). 1 κ−+ Equation (2) has similar connotations as (1) except for the appearance of κthat represents the fraction of area common to FOV erected at Pi and Pk and Pkd is as discussed in step 1. For that fraction of the FOV at Pk that is not visible from Pi , Pkd , which is the expected number of detections at cell Pk at any instant is made use of. And since ( )κ−1 is the fraction of FOV at Pk that is not visible Pkd is multiplied by that fraction. This makes use of the assumption that the distribution of targets is uniform across an area. In our plots of Pkd done for the surveillance zone variations between cells are steep only towards the edges of the zone. A more rigorous computation involves evaluating the pdf for the overlapped and non-overlapped areas separately. This is being avoided since (2) is computed during the online phase of the algorithm. This assumption is also invoked in step 6 while reducing the number of detections due to overlapping FOV. C The Coordination Phase: The coordination phase prevents paths of sensors to come close to one another to avoid overlapping FOV. This is done tk dn , by imposing penalties and reducing values of and recomputing paths for sensors with lower priorities as described in steps 5, 6 and 7. The path of the sensor with highest detections is fixed. The path with the second highest recomputed if there are overlapping areas of observation with the first. The recomputed path constitutes a motion strategy that is optimal in terms of detections under the constraint that the path of highest sensor is fixed. Similarly the path of the sensor with the least priority sensor when recomputed is optimal under the constraint that the paths of the sensors with higher priority are fixed. Hence at the coordination phase the optimality of the algorithm is not complete. A fully complete  )   algorithm would involve a search in the joint space of all sensors that is combinatorially hard. The depth limited search 9 3 = over T=3 time steps involves a search over paths for 729 reaching an optimal motion strategy. Hence a search in the n joint space for n sensors involves a search over 729 paths, that is computationally infeasible for online applications. The coordination phase resembles the decoupled priority based approaches to multi robot path planning [17,18]. IV SIMULATION RESULTS In this section we report results obtained through simulations on our environment developed using Borland’s JBuilder IDE. The value of the number of time steps, T, used for these simulations is 3, carried out on a P4 workstation with clock speed of 1.8 GHz. Figure 3 shows a snapshot of the environment with 10 sensors. The current position of the sensor is shown through the bigger shaded circle, while the targets through smaller ones. The traces of sensor movements are also shown. Target traces are not shown to avoid cluttering. Figure 3: A simulation system with ten sensors and several targets. The robots are shown through larger circles and targets through smaller ones. The traces of sensors are also shown. Table 1 compares the performance of the system in the presence and absence of coordination. The experiments done with sensor coordination are labeled as 1a, 2a, …, while those without coordination are labeled as 1b, 2b,… . Each experiment lasted for 150 time steps in total. The first column of the table denotes the index of the experiment; the second denotes the number of sensors in the system, the third the average detections or the number of targets detected by at- least one sensor per sample. The fourth signifies the average fraction of targets detected per sample, which is the number of targets detected per sample by one or more sensors divided by the number of targets in the surveillance zone per sample. Columns 5, 6 and 7 represent the number of targets detected exactly by one, two and three sensors per sample. The last column denotes the target velocity. Detections by more than three sensors are not tabulated for they generally assume insignificant values. Of particular interest in comparing the two schemes are the average fraction of target detections as well as the number of targets detected by exactly one, two and three sensors. It is expected that the average fraction of detections as well as the number of targets detected by exactly one sensor to be more for the method with coordination incorporated. Simultaneously the number of targets detected by two or more sensors is anticipated to be higher for the method that does not use the coordination phase. As the overlapping areas of FOV are higher in the absence of coordination the number of targets detected by two or more sensors is also higher. This in turn allows more targets to go undetected and the average fraction of detections to be lesser. The abbreviations used in the column headings are explained in the table caption. EN NS AD ZD AF D1S D2S D3S TV 10 0 2 29 0.6 22 32 10 1a 10 0.8 6.8 19.5 0.52 24.9 27.1 10 1b 22.2 25.7 0.46 22 0.2 0 10 2a 6 10 0 3.2 18.4 0.44 26.8 21.5 2b 6 15 0 0.3 20 0.49 21.1 20.4 10 3a 15 0 3.5 13.8 0.41 24.1 17.3 10 3b 27.2 20.2 0.57 26.5 0.7 0 10 4a 6 10 0 1.3 25.7 0.56 20.4 26.9 6 4b 15 0 0.4 19.6 0.47 22.5 20 10 5a 18.7 23.8 0.44 16.4 1.5 0.9 15 5b 10 Table 1: Comparison between motion strategies with and without sensor coordination. Experiments with labels ending in ‘a’ or the first line in every row have coordination incorporated while those ending in ‘b’ or the second line in each row are uncoordinated. Abreviations are as follows: EN = experiment number, NS = number of sensors, AD = Average Detections per sample, ZD = Zero Detections per sample, AF = Average Fraction per sample, D1S = Detections by exactly 1 sensor, D2S = Detections by exactly 2 sensors, D3S = Detections by exactly 3 sensors, TV = Targets’ Velocity As seen from table 1 the performance of the coordinated scheme with decoupled optimization is better for all experiments with differing NS and TV values than the method sans coordination. For example the number of targets detected by exactly one sensor per sample, the D1S column, is significantly higher in a number of experiments for the coordinated vis-à-vis the uncoordinated scheme. It is higher by 10, 4 and 6 detections per sample in experiments 1, 2 and 3. Also the number of targets detected by two sensors is significantly lesser across all experiments for the coordinated method. In experiments 1 and 5 the number of targets detected by three sensors is almost one per sample for the method lacking coordination. Correspondingly average fraction of targets detected per sample for the coordinated method is higher in all the experiments. The results of the table tally with the expectations and the underlying reasons for these expectations mentioned earlier. In table 2 we compare the current strategy with four of our previous methods of target detection and pursuit [14]. The first column signifies the experiments. Indexes 1a, 2a, etcetera correspond to the current method. Those with indices ‘b’, ‘c’, ‘d’ and ‘e’ correspond to coordinated-distracted, coordinated - dedicated, local-distracted and local-dedicated methods of resource allocation of sensors to targets for target detection [14]. Table 2 shows that across all experiments the current scheme surpasses all others in terms of average fraction of targets detected as well as other criteria such as D1S, D2S, AD and AF. It is to be noted that a difference of 0.06 in average fraction per sample between two methods corresponds to a difference of 4 or 5 target detections per sample (there are 70 to 80 targets in the zone per time-step on an average), which in turn corresponds to a difference of 600 detections in 150 time-steps of simulation. EN NS AD ZD AF D1S D2S D3S TV 10 0 0.2 22 0.46 25.7 22.3 6 1a 1b 6 20.3 27.7 0.42 17.9 2.3 0.1 10 10 0.8 4.4 13.1 0.37 30.5 18.3 6 1c 10 0 2.3 18.7 0.43 27 21 6 1d 10 0.5 4.6 13.1 0.37 30.5 18.3 6 1e 15 0.0 0.5 21.0 0.46 20.2 21.5 10 2a 2b 10 19.0 26.1 0.40 17.5 1.6 0.0 15 15 0.02 1.6 18.0 0.42 25.8 19.7 10 2c 15 0.02 2.2 17.0 0.41 25.4 19.2 10 2d 15 0.0 1.2 18.8 0.42 25.8 20.0 10 2e 10 0 0.7 26.5 0.57 20.2 27.2 6 3a 3b 6 21.2 26.1 0.44 20.2 1.0 0 10 10 0 1.7 19.8 0.45 25.8 21.5 6 3c 10 0 1.0 20.9 0.46 25.4 21.9 6 3d 10 0 1.7 19.8 0.45 25.7 21.5 6 3e 10 0.0 0.7 39.8 0.40 66.0 40.4 10 4a 4b 10 30.0 68.9 0.32 26.3 3.5 0.2 10 10 0.0 1.9 29.0 0.31 74.4 30.9 10 4c 10 0.8 3.5 27.4 0.31 73.7 31.6 10 4d 1.8 4e 10 31.6 73.8 0.30 29.8 0.0 10 0.7 3.4 3.4 3.7 3.2 0.0 0.2 0.04 0.04 0.0 30.0 21.4 25.9 24.8 25.5 29.3 17.7 22.5 21.0 22.4 57.4 64.9 60.3 61.4 60.7 0.35 0.24 0.30 0.28 0.29 Table 2: Comparison of current method (label a) with four previous methods called coordinated-distracted (b), coordinated-dedicated (c), local-distracted (d) and local-dedicated (e). The abreviations are same as in table 1. The comparisons reported in the table were done by introducing targets according to Poisson rates for one of the methods for an experiment that we call as the base method. For the remaining methods of that experiment the targets were introduced in the same fashion as the base method. The value of λwas 0.3 and Q=8 in these simulations. V CONCLUSIONS AND SCOPE A method for motion strategy computation of a sensor that maximizes the number of target detections for the next T- 5a 5b 5c 5d 5e 6 6 6 6 6 15 15 15 15 15 [10] B Horling, R Vincent, R Miller, J Shen, R Becker, K Rawlins, and V Lesser, “Distributed Sensor Network for Real Time Tracking”, In Proceedings of the 5th International Conference on Autonomous Agents:. 417-424, 2001. [11] B Horling, R Miller, M Sims, and V Lesser, “Using and Maintaining Organization in a Large-Scale Distributed Sensor Network”, In Proceedings of the Workshop on Autonomy, Delegation, and Control, (AAMAS 2003). [12] L Parker, “Distributed algorithms for multi-robot observation of multiple moving targets”, Autonomous Robots, 12, 3 (2002) [13] K Madhava Krishna, H Hexmoor, S Pasupuleti and S Chellappa, “A Surveillance System Based on Multiple Mobile Sensors”, Proceedings of FLAIRS, 2004, Special Track in Sensor Fusion. [14] K Madhava Krishna and Henry Hexmoor, “Resource Allocation for a Multi-sensor Surveillance System”, Workshop on Strategies Collaborative Agents, www.cs.unb.ca/~ghorbani/cola/2004/cola04.pdf, 29-39. [15] A. J. Briggs, “Efficient geometric algorithms for robot sensing and control”, Ph.D. thesis, Cornell, 1995 [16] E H Durfee, V R Lesser and D D Corkill, “Coherent cooperation among communicating problem solvers”, IEEE Trans. on Computers, C-36:1275- 1291, 1987. [17]. K. Azarm and G. Schmidt. “Conflict-Free Motion of Multiple Mobile Robots Based on Decentralized Motion Planning and Negotiation”, Proc. IEEE Int. Conf. on Robotics and Automation, p. 3526-3533, 1997. [18] M Bennewitz, W Burgard and S Thrun, “Finding and optimizing solvable priority schemes for decoupled path planning techniques for teams of mobile robots”, Robotics and Autonomous Systems, 41 (2), pp. 89-99, 2002. time steps is presented. The method makes use of a-priori known statistics of target arrivals along with detections reported for the current sample to estimate the number of detections on a lattice of cells. To reduce overlaps a coordination mechanism is specified that performs a constrained search, where the constraints are in the form of priorities assigned to sensors. The absence of an exhaustive search in the joint space due to the constraints renders the algorithm sub-optimal from a multi sensor perspective although optimal with respect to a single sensor in presence of those constraints. The tabulations presented in the results section vindicate the performance of the current approach in comparison to previous approaches for target detection and pursuit. The performance enhancement due to the coordination phase that reduces overlaps in FOV between sensors is also tabulated in section IV. As with any optimization algorithm the tradeoff is between the completeness of the optimization method vis-à- vis real-time requirements. In this aspect considering real time requirements a greedy decoupled optimization is resorted to at the multi sensor level. The performance is also dependent on the resolution of the discretization or the distance between the cells in the lattice. Future scope of this effort lies in identifying possible geometric properties of the surveillance area and nature of sensor paths to reduce the search space in the joint space of sensors that would provide for a real-time multi-sensor optimally detecting algorithm. Other areas of interest involve modifying the framework to provide for probability distributions for statistical parameters computed at the cells that allow targets to take multiple velocities. We also intend to compare the current method with other known methods [12] and extend the current method to include obstacles of simple geometry within surveillance area. REFERENCES: [1] S. A. Stoerer, P. E. Rybski, M. D. Erickson, M. Wyman, M. Gini, D. F. Hougen, N. Papanikolopoulos, “A Robot Team for Exploration and Surveillance: Design and Architecture,” Proceedings of the International Conference on Intelligent Autonomous Systems 6, pp. 767-774, Venice, Italy, July 2000. [2] H R Everett, G A Gilbraeth and R T Laird, “Coordinated control of multiple security robots”, Proceedings of SPIE Mobile Robots, 292-305, 1993 [3] D Schulz; W Burgard, D Fox and A Cremers, “Tracking multiple moving targets with a mobile robot using particle filters and statistical data association”, IEEE International Conference on Robotics and Automation, 1165-1170, 2001 [4] D Schulz and W Burgard, “Probabilistic state estimation of dynamic objects with a moving mobile robot”, Robotics and Autonomous Systems, 2001. [5] B. Kluge, C Kohler and E Prassler, “Fast and robust tracking of multiple objects through a laser range finder”, IEEE International Conference on Robotics and Automation, 1165-1170, 2001 [6] K Madhava Krishna and P K Kalra, “Detection tracking and avoidance of multiple dynamic objects”, Journal of Intelligent and Robotic Systems, 33(4): 371-408, 2002 [7] L Parker, “Cooperative robotics for multi-target observation”, Intelligent Automation and Soft Computing, 5[1]:5-19, 1999 [8] B Jung and G.S. Sukhatme, “Multi-Target Tracking using a Mobile Sensor Network”, Proc., IEEE Intl. Conf. On Robotics and Automation, 2002 [9] S Poduri and G S Sukhatme, “Constrained Coverage for mobile sensor networks”, IEEE ICRA, 165-171, 2004
1811.05931	2	1811	2019-03-11T13:04:28	Evolving intrinsic motivations for altruistic behavior	[ "cs.MA" ]	Multi-agent cooperation is an important feature of the natural world. Many tasks involve individual incentives that are misaligned with the common good, yet a wide range of organisms from bacteria to insects and humans are able to overcome their differences and collaborate. Therefore, the emergence of cooperative behavior amongst self-interested individuals is an important question for the fields of multi-agent reinforcement learning (MARL) and evolutionary theory. Here, we study a particular class of multi-agent problems called intertemporal social dilemmas (ISDs), where the conflict between the individual and the group is particularly sharp. By combining MARL with appropriately structured natural selection, we demonstrate that individual inductive biases for cooperation can be learned in a model-free way. To achieve this, we introduce an innovative modular architecture for deep reinforcement learning agents which supports multi-level selection. We present results in two challenging environments, and interpret these in the context of cultural and ecological evolution.	cs.MA	cs	Evolving Intrinsic Motivations for Altruistic Behavior Jane X. Wang, Edward Hughes, Chrisantha Fernando, Wojciech M. Czarnecki, Edgar A. Duéñez-Guzmán, & Joel Z. Leibo {wangjane,edwardhughes,chrisantha,lejlot,duenez,jzl}@google.com DeepMind London 9 1 0 2 r a M 1 1 ] A M . s c [ 2 v 1 3 9 5 0 . 1 1 8 1 : v i X r a ABSTRACT Multi-agent cooperation is an important feature of the natural world. Many tasks involve individual incentives that are misaligned with the common good, yet a wide range of organisms from bac- teria to insects and humans are able to overcome their differences and collaborate. Therefore, the emergence of cooperative behav- ior amongst self-interested individuals is an important question for the fields of multi-agent reinforcement learning (MARL) and evolutionary theory. Here, we study a particular class of multi- agent problems called intertemporal social dilemmas (ISDs), where the conflict between the individual and the group is particularly sharp. By combining MARL with appropriately structured natu- ral selection, we demonstrate that individual inductive biases for cooperation can be learned in a model-free way. To achieve this, we introduce an innovative modular architecture for deep rein- forcement learning agents which supports multi-level selection. We present results in two challenging environments, and interpret these in the context of cultural and ecological evolution. KEYWORDS multi-agent; evolution; altruism; social dilemmas ACM Reference Format: Jane X. Wang, Edward Hughes, Chrisantha Fernando, Wojciech M. Czar- necki, Edgar A. Duéñez-Guzmán, & Joel Z. Leibo. 2019. Evolving Intrinsic Motivations for Altruistic Behavior. In Proc. of the 18th International Confer- ence on Autonomous Agents and Multiagent Systems (AAMAS 2019), Montreal, Canada, May 13 -- 17, 2019, IFAAMAS, 10 pages. 1 INTRODUCTION Nature shows a substantial amount of cooperation at all scales, from microscopic interactions of genomes and bacteria to species- wide societies of insects and humans [39]. This is in spite of natural selection pushing for short-term individual selfish interests [9]. In its purest form, altruism can be favored by selection when coop- erating individuals preferentially interact with other cooperators, thus realising the rewards of cooperation without being exploited by defectors [10, 14, 22, 23, 51]. However, many other possibili- ties exist, including kin selection, reciprocity and group selection [43, 44, 54, 56, 57, 60]. Lately the emergence of cooperation among self-interested agents has become an important topic in multi-agent deep reinforcement learning (MARL). [35] and [28] formalize the problem domain as an intertemporal social dilemma (ISD), which generalizes matrix game Proc. of the 18th International Conference on Autonomous Agents and Multiagent Systems (AAMAS 2019), N. Agmon, M. E. Taylor, E. Elkind, M. Veloso (eds.), May 13 -- 17, 2019, Montreal, Canada. © 2019 International Foundation for Autonomous Agents and Multiagent Systems (www.ifaamas.org). All rights reserved. social dilemmas to Markov settings. Social dilemmas are character- ized by a trade-off between collective welfare and individual utility. As predicted by evolutionary theory, self-interested reinforcement- learning agents are typically unable to achieve the collectively op- timal outcome, converging instead to defecting strategies [35, 48]. The goal is to find multi-agent training regimes in which individuals resolve social dilemmas, i.e., cooperation emerges. Previous work has found several solutions, belonging to three broad categories: 1) opponent modelling [16, 34], 2) long-term planning using perfect knowledge of the game's rules [36, 49] and 3) a specific intrinsic motivation function drawn from behavioral economics [28]. These hand-crafted approaches run at odds with more recent end-to-end model-free learning algorithms, which have been shown to have a greater ability to generalize (e.g. [11]). We propose that evolution can be applied to remove the hand-crafting of intrinsic motivation, similar to other applications of evolution in deep learning. Evolution has been used to optimize single-agent hyperparame- ters [30], implement black-box optimization [59], and to evolve neu- roarchitectures [41, 55], regularization [4], loss functions [27, 29], behavioral diversity [7], and entire reward functions [52, 53]. These principles tend to be driven by single-agent search and optimization or competitive multi-agent tasks. Therefore there is no guarantee of success when applying them in the ISD setting. More closely related to our domain are evolutionary simulations of predator- prey dynamics [61], which used enforced subpopulations to evolve populations of neurons which are sampled to form the hidden layer of a neural network.1 To address the specific challenges of ISDs, the system we propose distinguishes between optimization processes that unfold over two distinct time-scales: (1) the fast time-scale of learning and (2) the slow time-scale of evolution [similar to 26]. In the former, individual agents repeatedly participate in an intertemporal social dilemma using a fixed intrinsic motivation. In the latter, that motivation is itself subject to natural selection in a population. We model this intrinsic motivation as an additional additive term in the reward of each agent [6]. We implement the intrinsic reward function as a two- layer fully-connected feed-forward neural network, whose weights define the genotype for evolution. We propose that evolution can help mitigate this intertemporal dilemma by bridging between these two timescales via an intrinsic reward function. Evolutionary theory predicts that evolving individual intrinsic reward weights across a population who interact uniformly at ran- dom does not lead to altruistic behavior [1]. Thus, to achieve our goal, we must structure the evolutionary dynamics [43]. We first implement a "Greenbeard" strategy [10, 31] in which agents choose 1See also [50] and [47] for reviews of other evolutionary approaches to cooperative multi-agent problems. interaction partners based on an honest, real-time signal of cooper- ativeness. We term this process assortative matchmaking. Although there is ecological evidence of assortative matchmaking [33], it cannot explain cooperation in all taxa [17, 18, 25]. Moreover it isn't a general method for multi-agent reinforcement learning, since honest signals of cooperativeness are not normally observable in the ISD models typically studied in deep reinforcement learning. To address the limitations of the assortative matchmaking ap- proach, we introduce an alternative modular training scheme loosely inspired by ideas from the theory of multi-level (group) selection [25, 60], which we term shared reward network evolution. Here, agents are composed of two neural network modules: a policy net- work and a reward network. On the fast timescale of reinforcement learning, the policy network is trained using the modified rewards specified by the reward network. On the slow timescale of evo- lution, the policy network and reward network modules evolve separately from one another. In each episode every agent has a distinct policy network but the same reward network. As before, the fitness for the policy network is the individual's reward. In contrast, the fitness for the reward network is the collective return for the entire group of co-players. In terms of multi-level selection theory, the policy networks are the lower level units of evolution and the reward networks are the higher level units. Evolving the two modules separately in this manner prevents evolved reward networks from overfitting to specific policies. This evolutionary paradigm not only resolves difficult ISDs without handcrafting but also points to a potential mechanism for the evolutionary origin of social inductive biases. The paper is structured as follows. In Section 2, we define our problem domain, and describe in detail our agent architecture and training methods. In Section 3, we present results from our experi- ments and further analyses of agent policies. Finally in Section 4, we discuss interpretations of our model as well as make suggestions for future work. 2 METHODS We varied and explored different combinations of parameters, namely: (1) environments {Harvest, Cleanup}, (2) reward network features {prospective, retrospective}, (3) matchmaking {random, assortative}, and (4) reward network evolution {individual, shared, none}. We describe these in the following sections. 2.1 Intertemporal social dilemmas In this paper, we consider Markov games [37] within a MARL set- ting. Specifically we study intertemporal social dilemmas [28, 35], defined as games in which individually selfish actions produce individual benefit on short timescales but have negative impacts on the group over a longer time horizon. This conflict between the two timescales characterizes the intertemporal nature of these games. The tension between individual and group-level rational- ity identifies them as social dilemmas (e.g. the famous Prisoner's Dilemma). We consider two dilemmas, each implemented as a partially ob- servable Markov game on a 2D grid (see Figure 1), with N = 5 players playing at a time. In the Cleanup game, agents tried to col- lect apples (reward +1) that spawned in a field at a rate inversely related to the cleanliness of a geographically separate aquifer. Over time, this aquifer filled up with waste, lowering the respawn rate of apples linearly, until a critical point past which no apples could spawn. Episodes were initialized with no apples present and zero spawning, thus necessitating cleaning. The dilemma occurred be- cause in order for apples to spawn, agents must leave the apple field and clean, which conferred no reward. However if all agents declined to clean (defect), then no rewards would be received by any. In the Harvest game, again agents collected rewarding apples. The apple spawn rate at a particular point on the map depended on the number of nearby apples, falling to zero once there were no apples in a certain radius. There is a dilemma between the short- term individual temptation to harvest all the apples quickly and the consequential rapid depletion of apples, leading to a lower total yield for the group in the long-term. All episodes last 1000 steps, and the total size of the playable area is 25×18 for Cleanup and 38×16 for Harvest. Games are partially observable in that agents can only observe via a 15×15 RGB window, centered on their current location. The action space consists of moving left, right, up, and down, rotating left and right, and the ability to tag each other. This action has a reward cost of 1 to use, and causes the player tagged to lose 50 reward points, thus allowing for the possibility of punishing free-riders [21, 45]. The Cleanup game has an additional action for cleaning waste. 2.2 Modeling social preferences as intrinsic motivations In our model, there are three components to the reward that enter into agents' loss functions (1) total reward, which is used for the policy loss, (2) extrinsic reward, which is used for the extrinsic value function loss and (3) intrinsic reward, which is used for the intrinsic value function loss. The total reward for player i is the sum of the extrinsic reward and an intrinsic reward as follows: (cid:16) (cid:17) , ri(si , ai) = r E i (si , ai) + ui(fi) . (1) i (s, a) is the environment reward obtained The extrinsic reward r E by player i when it takes action ai from state si, sometimes also written with a time index t. The intrinsic reward u(f) is an aggregate social preference across features f and is calculated according to the formula, ui(fiθ) = vTσ WTfi + b (2) where σ is the ReLU activation function, and θ = {W, v, b} are the parameters of a 2-layer neural network with 2 hidden nodes. These parameters are evolved based on fitness (see Section 2.3). The elements of v = (v1, v2) approximately correspond to a linear combination of the coefficients related to advantagenous and dis- advantagenous inequity aversion mentioned in [28], which were found via grid search in this previous work, but are here evolved. The feature vector fi is a player-specific vector quantity that agents can transform into intrinsic reward via their reward network. It's composed of features fij derived from all players 2, so that each 2Note that we use both i and j to index over the players, but i makes reference to the player receiving the intrinsic reward, while j indexes the players sending the features over which the intrinsic reward of player i is defined. Figure 1: Screenshots from (a) the Cleanup game, (b) the Harvest game. The size of the agent-centered observation window is shown in (b). The same size observation was used in all experiments. Figure 2: (a) Agent Aj adjusts policy πj(s, aϕ) using off-policy importance weighted actor-critic (V-Trace) [11] by sampling from a queue with (possibly stale) trajectories recorded from 500 actors acting in parallel arenas. (b) The architecture (shown only for 1 agent) includes a visual encoder (1-layer convolutional neural net with 6 3x3 filters, stride 1, followed by two fully- connected layers with 32 units each), intrinsic and extrinsic value heads (V I and V E), a policy head π, and a long-short term memory (LSTM, with 128 hidden units), which takes last intrinsic and extrinsic rewards (u(f) and r E) and last action as input. The reward network weights are evolved based on total episode return. player has access to the same set of features, with the exception that its own feature is demarcated specially (by always occupying the first element of the vector). The features themselves are a function of recently received or expected future (extrinsic) reward for each agent. In Markov games the rewards received by different players may not be aligned in time. Thus, any model of social preferences should not be overly influenced by the precise temporal alignment of different players' rewards. Intuitively, they ought to depend on comparing temporally averaged reward estimates between players, of pop. policy networks {{ϕ1, h1}, ..., {ϕN , hN }} Algorithm 1 Training Pseudocode - Shared Reward Network Require: P: and hyper-parameters Require: R: pop. of reward networks {θ1, ..., θN }, where θ = {W, v, b} Require: S: procedure to sample from P and R and return 5 players Require: U: procedure to update/evolve weights given a population of Require: F: procedure to calculate and assign fitness given the sampled fitness-scored individuals players 1: while not done do 2: 3: 4: 5: 6: p1:5 = p1, ..., p5 ← S(P, R) env ∼ p(T) 1:5 ← 0 e0 for t := 1 to T do τ t = (r E,t 1:5 , V E,t ▷ Sample 5 players ▷ Sample from distribution p over environments T ▷ Initialize temporally decayed reward ▷ Run for T = 1000 steps 1:5) ← env(p1:5) ▷ Play players in ▷ Calculate feature vectors for players 1:5 , ot 1:5, at environment, collect outputs in a trajectory τ = {τ 1, ..., τ T } for j := 1 to 5 do ▷ Decayed extrinsic reward if retrospective then j ← η e t−1 j else if prospective then j + r E,t fj = e t fj ← V E,t for i := 1 to 5 do j i ← vTσ(cid:0)WTfi + b(cid:1) fi ← reorder(fi, i) ut ▷ Value estimate from extrinsic value head ▷ Calculate intrinsic rewards for players ▷ Calculate intrinsic reward for i := 1 to 5 do ϕi ← RL(ϕi, τ , ui, hi) Fϕk , Fθk ▷ RL training for each ϕ ← F(τ) ▷ Calculate smoothed fitnesses associated with each reward and policy network sampled in this episode for (ϕk, hk) ∈ P do if available_to_evolve(ϕk, hk) then ▷ If burn-in period has passed, update population based on smoothed fitness of individuals 7: 8: 9: 10: 11: 12: 13: 14: 15: 16: 17: 18: 19: 20: 21: 22: 23: (ϕk, hk) ← U(P, Fϕk ) for θk ∈ R do if available_to_evolve(θk) then θk ← U(R, Fθk ) rather than instantaneous values. Therefore, we considered two different ways of temporally aggregating the rewards. The retrospective method derives intrinsic reward from whether an agent judges that other agents have been actually (extrinsically) rewarded in the recent past. The prospective variant derives intrinsic reward from whether other agents are expecting to be (extrinsically) rewarded in the near future.3 For the retrospective variant, fij = et j , where the temporally decayed reward et j for the agents j = 1, . . . , N are updated at each timestep t according to j = η et−1 et j + r E,t j , (3) and η = 0.975. The prospective variant uses the value estimates V E j (see Figure 2b) for fij and has a stop-gradient before the reward network module so that gradients don't flow back into other agents (as in for example DIAL from [15]). 3Our terms prospective and retrospective map onto the terms intentional and conse- quentialist respectively as used by [36, 49]. 2.3 Architecture and Training We used the same training framework as in [29], which performs distributed asynchronous training in multi-agent environments, including population-based training (PBT) [30]. We trained a popu- lation of N = 50 agents4 with policies {πi}, from which we sampled 5 players in order to populate each of 500 arenas (where arena is an instantiation of a single episode of the environment) running in par- allel. Within each arena, an episode of the environment was played with the sampled agents, before resampling new ones. Agents were sampled using one of two matchmaking processes (described in more detail below). Episode trajectories lasted 1000 steps and were written to queues for learning, from which weights were updated using V-Trace (Figure 2a). The set of weights evolved included learning rate, entropy cost weight, and reward network weights θ5. The parameters of the policy network ϕ were inherited in a Lamarckian fashion as in [30]. Furthermore, we allowed agents to observe their last actions ai,t−1, last intrinsic rewards (ui,t−1(fi)), and last extrinsic rewards i,t−1(si , ai)) as input to the LSTM in the agent's neural network. (r E The objective function was identical to that presented in [11] and comprised three components: (1) the value function gradient, (2) policy gradient, and (3) entropy regularization, weighted according to hyperparameters baseline cost and entropy cost (see Figure 2b). Evolution was based on a fitness measure calculated as a moving average of total episode return, which was a sum of apples collected minus penalties due to tagging, smoothed as follows: i = (1 − ν)F n−1 F n i t r E,t + νRn (4) i , where ν = 0.001 and Rn is the total episode return i obtained on episode n by agent i (or reward network i in the case of the shared reward network evolution (see Section 2.5 for details). Training was done via joint optimization of network parameters via SGD and hyperparameters/reward network parameters via evo- lution in the standard PBT setup. Gradient updates were applied for every trajectory up to a maximum length of 100 steps, using a batch size of 32. Optimization was via RMSProp with epsilon=10−5, momentum=0, decay rate=0.99, and an RL discount factor of 0.99. The baseline cost weight (see Mnih et al. [42]) was fixed at 0.25, and the entropy cost was sampled from LogUniform(2 × 10−4,0.01) and evolved throughout training using PBT. The learning rates were all initially set to 4 × 10−4 and then allowed to evolve. PBT uses evolution (specifically genetic algorithms) to search over a space of hyperparameters rather than manually tuning or performing a random search, resulting in an adaptive schedule of hyperparameters and joint optimization with network parameters learned through gradient descent [30]. There was a mutation rate of 0.1 when evolving hyperparameters, using multiplicative perturbations of ±20% for entropy cost and learning rate, and additive perturbation of ±0.1 for reward network parameters. We implemented a burn-in period for evolution of 4 × 106 agent steps, to allow network parameters and hyperparameters = i 4Similar to as in [11], we distinguish between an "agent" which acts in the environment according to some policy, and a "learner" which updates the parameters of a policy. In principle, a single agent's policy may depend on parameters updated by several separate learners. 5We can imagine that the reward weights are simply another set of optimization hyperparameters since they enter into the loss. Figure 3: (a) Agents assigned and evolved with individual reward networks. (b) Assortative matchmaking, which preferentially plays cooperators with other cooperators and defectors with other defectors. (c) A single reward network is sampled from the population and assigned to all players, while 5 policy networks are sampled and assigned to the 5 players individually. After the episode, policy networks evolve according to individual player returns, while reward networks evolve according to aggregate returns over all players. to be used in enough episodes for an accurate assessment of fitness before evolution. 2.4 Random vs. assortative matchmaking Matches were determined according to two methods: (1) random matchmaking and (2) assortative matchmaking. Random matchmak- ing simply selected uniformly at random from the pool of agents to populate the game, while cooperative matchmaking first ranked agents within the pool according to a metric of recent cooperative- ness, and then grouped agents such that players of similar rank played with each other. This ensured that highly cooperative agents played only with other cooperative agents, while defecting agents played only with other defectors. For Cleanup, cooperativeness was calculated based on the amount of steps in the last episode the agent chose to clean. For Harvest, it was calculated based on the difference between the the agent's return and the mean return of all players, so that having less return than average yielded a high cooperativeness ranking. Cooperative metric-based matchmaking was only done with either individual reward networks or no re- ward networks (Figure 3b). We did not use cooperative metric-based matchmaking for our multi-level selection model, since these are theoretically separate approaches. 2.5 Individual vs. shared reward networks Building on previous work that evolved either the intrinsic reward [29] or the entire loss function [27], we considered the reward network weights to be hyperparameters that could be evolved in parallel with the policy parameters (Figure 3a). Distinct from these methods, we separately evolved the reward network within its own population, thereby allowing different modules of the agent to compete only with like components. This allowed for independent exploration of hyperparameters via separate credit assignment of fitness, and thus considerably more of the hyperparameter land- scape could be explored compared with using only a single pool. In addition, reward networks could be randomly assigned to any policy network, and so were forced to generalize to a wide range of policies. In a given episode, 5 separate policy networks were paired with the same reward network, which we term a shared re- ward network. In line with [30], the fitness determining the copying of policy network weights and evolution of optimization-related hyperparameters (entropy cost and learning rate) were based on individual agent return. By contrast, the reward network parame- ters were evolved according to fitness based on total episode return across the group of co-players (Figure 3c). Figure 4: Total episode rewards, aggregated over players. (a), (b): Comparing retrospective (backward-looking) reward evolution with assortative matchmaking and PBT-only baseline in (a) Cleanup and (b) Harvest. (c), (d): Comparing prospective (forward- looking) with retrospective (backward-looking) reward evolution in (c) Cleanup and (d) Harvest. The black dotted line indicates performance from [28]. The shaded region shows standard error of the mean, taken over the population of agents. This contribution is distinct from previous work which evolved intrinsic rewards [e.g. 29] because (1) we evolve over social features rather than a remapping of environmental events, and (2) reward network evolution is motivated by dealing with the inherent tension in ISDs, rather than merely providing a denser reward signal. In this sense it's more akin to evolving a form of communication for social cooperation, rather than learning reward-shaping in a sparse- reward environment. We allow for multiple agents to share the same components, and as we shall see, in a social setting, this winds up being critical. Shared reward networks provide a biologically principled method that mixes group fitness on a long timescale and individual reward on a short timescale. This contrasts with hand-crafted means of aggregation, as in previous work [5, 38]. 3 RESULTS As shown in Figure 4, PBT without using an intrinsic reward net- work performs poorly on both games, where it asymptotes to 0 total episode reward in Cleanup and 400 for Harvest (the number of apples gained if all agents collect as quickly as they can). Figures 4a-b compare random and assortative matchmaking with PBT and reward networks using retrospective social features. When using random matchmaking, individual reward network agents per- form no better than PBT at Cleanup, and only moderately better at Harvest. Hence there is little benefit to adding reward networks over social features if players have separate networks, as these tend to be evolved selfishly. The assortative matchmaking experiments used either no reward network (u(f) = 0) or individual reward net- works. Without a reward network, performance was the same as the PBT baseline. With individual reward networks, performance was very high, indicating that both conditioning the internal rewards on social features and a preference for cooperative agents to play to- gether were key to resolving the dilemma. On the other hand, shared reward network agents perform as well as assortative matchmaking and the handcrafted inequity aversion intrinsic reward from [28], even using random matchmaking. This implies that agents didn't necessarily need to have immediate access to honest signals of other agents' cooperativeness to resolve the dilemma; it was enough to simply have the same intrinsic reward function, evolved accord- ing to collective episode return. Videos comparing performance of the PBT baseline with the retrospective variant of shared reward network evolution can be found at https://youtu.be/medBBLLM4c0 and https://youtu.be/yTjrlH3Ms9U. Figures 4(c) and (d) compare the retrospective and prospective variants of reward network evolution. The prospective variant, although better than PBT when using a shared reward network, generally results in worse performance and more instability. This is likely because the prospective variant depends on agents learning good value estimates before the reward networks become useful, whereas the retrospective variant only depends on environmentally provided reward and thus does not suffer from this issue. Inter- estingly, we observed that the prospective variant does achieve very high performance if gradients are allowed to pass between agents via the value estimates V E (data not shown); however, this j constitutes centralized learning, albeit with decentralized execution (see [15]). Such approaches are promising but less consistent with biologically plausible mechanisms of multi-agent learning which are of interest here and so were not pursued. We next plot various social outcome metrics in order to better capture the complexities of agent behavior (see Figure 5). Equality is calculated as E(1 − G(R)), where G(R) is the Gini coefficient over individual returns. Figure 5b demonstrates that, in Harvest, having the prospective version of reward networks tends to lead to lower equality, while the retrospective variant has very high equality. Equality in Cleanup is more unstable throughout training, since it's not necessarily optimal, but tends to be lower overall than for Harvest, even when performance is high, indicating that equality might be harder to achieve in different games. Tagging measures the average number of times a player fined another player through- out the episode. The middle panel of Figure 5b shows that there is a higher propensity for tagging in Harvest when using either a prospective reward network or an individual reward network, com- pared to the retrospective shared reward network. This explains the performance shown in Figure 4, as being tagged results in a very high negative reward. Tagging in the Cleanup task is overall much lower than in Harvest. Sustainability measures the average time step on which agents received positive reward, averaged over the episode and over agents. We see in the bottom panel of 5b that having no reward network results in players collecting apples extremely quickly in Harvest, compared with much more sustain- able behavior with reward networks. In Cleanup, the sustainability metric is not meaningful and so this was not plotted. Finally, we can directly examine the weights of the final retro- spective shared reward networks which were best at resolving the ISDs. Interestingly, the final weights evolved in the second layer suggest that resolving each game might require a different set of social preferences. In Cleanup, one of the final layer weights v2 evolved to be close to 0, whereas in Harvest, v1 and v2 evolved to be of large magnitude but opposite sign. We can see a similar pattern with the biases b. We interpret this to mean that Cleanup re- quired a less complex reward network: it was enough to simply find other agents' being rewarded as intrinsically rewarding. In Harvest, however, a more complex reward function was perhaps needed in order to ensure that other agents were not over-exploiting the apples. We found that the first layer weights W tended to take on arbitrary (but positive) values. This is because of random match- making: co-players were randomly selected and thus there was little evolutionary pressure to specialize these weights. We demonstrated that the reward network weights evolved dif- ferently for Cleanup versus Harvest, indicating that the two tasks necessitate different forms of social cooperation for optimal per- formance. This highlights the advantage of evolving rather than hand-crafting the weighting between individual reward and group reward, as optimal weightings cannot necessarily be anticipated for all environments. Evolving such weightings thus constitutes a form of meta-learning, wherein an entire learning system, including intrinsic reward functions, is optimized for fast learning [13, 53]. Here we have extended these ideas to the multi-agent domain. Why does evolving intrinsic social preferences promote coop- eration? Firstly, evolution ameliorates the intertemporal choice problem by distilling the long timescale of collective fitness into the short timescale of individual reinforcement learning, thereby improving credit assignment between selfish acts and their tempo- rally displaced negative group outcomes [28]. Secondly, it mitigates the social dilemma itself by allowing evolution to expose social signals that correlate with, for example, an agent's current level of selfishness. Such information powers a range of mechanisms for achieving mutual cooperation like competitive altruism [24], other- regarding preferences [8], and inequity aversion [12]. In accord, laboratory experiments show that humans cooperate more readily when they can communicate [32, 46]. 4 DISCUSSION Real environments don't provide scalar reward signals to learn from. Instead, organisms have developed various internal drives based on either primary or secondary goals [2]. Here we examined intrinsic rewards based on features derived from other agents in the environment, in order to establish whether such social signals could enable the evolution of altruism to solve intertemporal social dilemmas. In accord with evolutionary theory [1, 43], we found that naïvely implementing natural selection via genetic algorithms did not lead to the emergence of cooperation. Furthermore, assortative matchmaking was sufficient to generate cooperative behavior in cases where honest signals were available. Finally, we proposed a new multi-level evolutionary paradigm based on shared reward networks that achieves cooperation in more general situations. The shared reward network evolution model was inspired by multi-level selection; yet it does not correspond to the prototypical case of that theory since its lower level units of evolution (the policy networks) are constantly swapping which higher level unit (reward network) they are paired with. Nevertheless, there are a variety of ways in which we see this form of modularity arise in nature. For example, free-living microorganisms occasionally form multi- cellular structures to solve a higher order adaptive problem, like slime mold forming a spore-producing stalk for dispersal [58], and many prokaryotes can incorporate plasmids (modules) found in their environment or received from other individuals as functional parts of their genome, thereby achieving cooperation in social dilemmas [20, 40]. Alternatively, in humans a reward network may represent a shared "cultural norm", with its fitness based on cultural information accumulated from the groups in which it holds sway. In this way, the spread of norms can occur independently of the success of individual agents [3]. Note that in this work, we have assumed that agents have perfect knowledge of other agents' rewards, while in real-world systems Figure 5: Social outcome metrics for (a) Cleanup and (b) Harvest. Top: equality, middle: total amount of tagging, bottom: sus- tainability. The shaded region shows the standard error of the mean. to, for instance, a smile/frown or other locally observable social signals), identifiability across episodes, or even deception. The approach outlined here opens avenues for investigating alternative evolutionary mechanisms for the emergence of coop- eration, such as kin selection [19] and reciprocity [56]. It would be interesting to see whether these lead to different weights in a reward network, potentially hinting at the evolutionary origins of different social biases. Along these lines, one might consider study- ing an emergent version of the assortative matchmaking model along the lines suggested by [25], adding further generality and power to our setup. Finally, it would be fascinating to determine how an evolutionary approach can be combined with multi-agent communication to produce that most paradoxical of cooperative behaviors: cheap talk. Figure 6: Distribution of layer 2 weights and biases of evolved retrospective shared reward network at 1.5 × 108 training steps for (a) Cleanup, and (b) Harvest. this is not typically the case. This assumption was made in order to disentangle the effects of cultural evolution from the quality of the signals being evolved over. Natural next steps include adding partial observability or noise to this signal (to make it more analogous ACKNOWLEDGMENTS We would like to thank Simon Osindero, Iain Dunning, Andrea Tacchetti, and many DeepMind colleagues for valuable discussions and feedback, as well as code development and support. REFERENCES [1] Robert Axelrod and William D. Hamilton. 1981. The Evolution of Cooperation. Science 211, 4489 (1981), 1390 -- 1396. http://www.jstor.org/stable/1685895 [2] Gianluca Baldassarre and Marco Mirolli. 2013. Intrinsically Motivated Learning in Natural and Artificial Systems. 1 -- 458 pages. [3] Robert Boyd and Peter J. Richerson. 2009. Culture and the evolution of human cooperation. Philosophical Transactions of the Royal Society of London B: Biologi- cal Sciences 364, 1533 (2009), 3281 -- 3288. https://doi.org/10.1098/rstb.2009.0134 arXiv:http://rstb.royalsocietypublishing.org/content/364/1533/3281.full.pdf [4] ZSH Chan, HW Ngan, AB Rad, and TK Ho. 2002. Alleviating'overfitting'via genetically-regularised neural network. Electronics Letters 38, 15 (2002), 1. [5] Yu-Han Chang, Tracey Ho, and Leslie P Kaelbling. 2004. All learning is local: Multi-agent learning in global reward games. In Advances in neural information processing systems. 807 -- 814. [6] Nuttapong Chentanez, Andrew G. Barto, and Satinder P. Singh. 2005. Intrinsically Motivated Reinforcement Learning. In Advances in Neural Information Processing Systems 17, L. K. Saul, Y. Weiss, and L. Bottou (Eds.). MIT Press, 1281 -- 1288. http:// papers.nips.cc/paper/2552-intrinsically-motivated-reinforcement-learning.pdf [7] Edoardo Conti, Vashisht Madhavan, Felipe Petroski Such, Joel Lehman, Kenneth O Stanley, and Jeff Clune. 2017. Improving exploration in evolution strategies for deep reinforcement learning via a population of novelty-seeking agents. arXiv preprint arXiv:1712.06560 (2017). [8] David J Cooper and John H Kagel. 2016. Other-regarding preferences. The handbook of experimental economics 2 (2016), 217. [9] Charles Darwin. 1859. On the Origin of Species by Means of Natural Selection. Murray, London. or the Preservation of Favored Races in the Struggle for Life. [10] Richard Dawkins. 1976. The selfish gene Oxford university press. New York [11] Lasse Espeholt, Hubert Soyer, Remi Munos, Karen Simonyan, Volodymir Mnih, Tom Ward, Yotam Doron, Vlad Firoiu, Tim Harley, Iain Dunning, et al. 2018. IMPALA: Scalable distributed Deep-RL with importance weighted actor-learner architectures. arXiv preprint arXiv:1802.01561 (2018). [12] Ernst Fehr and Klaus M. Schmidt. 1999. A Theory of Fairness, Competition, and Cooperation*. The Quarterly Journal of Economics 114, 3 (1999), 817 -- 868. https://doi.org/10.1162/003355399556151 [13] Chrisantha Thomas Fernando, Jakub Sygnowski, Simon Osindero, Jane Wang, Tom Schaul, Denis Teplyashin, Pablo Sprechmann, Alexander Pritzel, and An- drei A Rusu. 2018. Meta Learning by the Baldwin Effect. arXiv preprint arXiv:1806.07917 (2018). [14] Jeffrey A Fletcher and Michael Doebeli. 2009. A simple and general explanation for the evolution of altruism. Proceedings of the Royal Society of London B: Biological Sciences 276, 1654 (2009), 13 -- 19. [15] Jakob N. Foerster, Yannis M. Assael, Nando de Freitas, and Shimon Whiteson. 2016. Learning to Communicate with Deep Multi-Agent Reinforcement Learning. CoRR abs/1605.06676 (2016). arXiv:1605.06676 http://arxiv.org/abs/1605.06676 [16] Jakob N. Foerster, Richard Y. Chen, Maruan Al-Shedivat, Shimon Whiteson, Pieter Abbeel, and Igor Mordatch. 2017. Learning with Opponent-Learning Awareness. CoRR abs/1709.04326 (2017). arXiv:1709.04326 http://arxiv.org/abs/1709.04326 [17] Andy Gardner and Stuart A West. 2010. Greenbeards. Evolution: International Journal of Organic Evolution 64, 1 (2010), 25 -- 38. [18] Alan Grafen et al. 1990. Do animals really recognize kin?. Animal Behaviour 39, (1976). 1 (1990), 42 -- 54. 2002), 15 -- 21. [19] Ashleigh Griffin and Stuart West. 2002. Kin selection: Fact and fiction. 17 (01 [20] Ashleigh S Griffin, Stuart A West, and Angus Buckling. 2004. Cooperation and competition in pathogenic bacteria. Nature 430, 7003 (2004), 1024. [21] Özgür Gürerk, Bernd Irlenbusch, and Bettina Rockenbach. 2006. The competitive advantage of sanctioning institutions. Science 312, 5770 (2006), 108 -- 111. [22] W.D. Hamilton. 1964. The Genetical Evolution of Social Behaviour. I. Journal of [23] William D Hamilton. 1964. The genetical evolution of social behaviour. II. Journal Theoretical Biology 7, 1 (July 1964), 1 -- 16. of theoretical biology 7, 1 (1964), 17 -- 52. [24] Charlie L Hardy and Mark Van Vugt. 2006. Nice guys finish first: The competitive altruism hypothesis. Personality and Social Psychology Bulletin 32, 10 (2006), 1402 -- 1413. [25] Joseph Henrich. 2004. Cultural group selection, coevolutionary processes and large-scale cooperation. Journal of Economic Behavior & Organization 53, 1 (2004), 85 -- 88. [26] Geoffrey E Hinton and Steven J Nowlan. 1987. How learning can guide evolution. Complex systems 1, 3 (1987), 495 -- 502. [27] Rein Houthooft, Richard Y. Chen, Phillip Isola, Bradly C. Stadie, Filip Wol- ski, Jonathan Ho, and Pieter Abbeel. 2018. Evolved Policy Gradients. CoRR abs/1802.04821 (2018). arXiv:1802.04821 http://arxiv.org/abs/1802.04821 [28] Edward Hughes, Joel Z Leibo, Matthew G Phillips, Karl Tuyls, Edgar A Duéñez- Guzmán, Antonio García Castañeda, Iain Dunning, Tina Zhu, Kevin R McKee, Raphael Koster, et al. 2018. Inequity aversion improves cooperation in intertem- poral social dilemmas. In Advances in neural information processing systems (NIPS). Montreal, Canada. [29] Max Jaderberg, Wojciech M Czarnecki, Iain Dunning, Luke Marris, Guy Lever, Antonio Garcia Castaneda, Charles Beattie, Neil C Rabinowitz, Ari S Morcos, Avraham Ruderman, et al. 2018. Human-level performance in first-person multi- player games with population-based deep reinforcement learning. arXiv preprint arXiv:1807.01281 (2018). [30] Max Jaderberg, Valentin Dalibard, Simon Osindero, Wojciech M Czarnecki, Jeff Donahue, Ali Razavi, Oriol Vinyals, Tim Green, Iain Dunning, Karen Si- monyan, et al. 2017. Population based training of neural networks. arXiv preprint arXiv:1711.09846 (2017). [31] Vincent Jansen and Minus van Baalen. 2006. Altruism through beard chromody- namics. 440 (04 2006), 663 -- 6. [32] Marco A Janssen, Robert Holahan, Allen Lee, and Elinor Ostrom. 2010. Lab experiments for the study of social-ecological systems. Science 328, 5978 (2010), 613 -- 617. [33] Laurent Keller and Kenneth G. Ross. 1998. Selfish genes: A green beard in the red fire ant. 394 (08 1998), 573 -- 575. [34] Max Kleiman-Weiner, Mark K. Ho, Joseph L. Austerweil, Michael L. Littman, and Joshua B. Tenenbaum. 2016. Coordinate to cooperate or compete: Abstract goals and joint intentions in social interaction. In CogSci. [35] Joel Z Leibo, Vinicius Zambaldi, Marc Lanctot, Janusz Marecki, and Thore Graepel. 2017. Multi-agent reinforcement learning in sequential social dilemmas. In Proceedings of the 16th Conference on Autonomous Agents and MultiAgent Systems. International Foundation for Autonomous Agents and Multiagent Systems, 464 -- 473. [36] Adam Lerer and Alexander Peysakhovich. 2017. Maintaining cooperation in complex social dilemmas using deep reinforcement learning. CoRR abs/1707.01068 (2017). arXiv:1707.01068 http://arxiv.org/abs/1707.01068 [37] Michael L Littman. 1994. Markov games as a framework for multi-agent rein- forcement learning. In Machine Learning Proceedings 1994. Elsevier, 157 -- 163. [38] Maja J Mataric. 1994. Learning to behave socially. In Third international conference on simulation of adaptive behavior, Vol. 617. 453 -- 462. [39] John Maynard Smith and Eors Szathmary. 1997. The major transitions in evolution. Oxford University Press. [40] Sorcha E McGinty, Daniel J Rankin, and Sam P Brown. 2011. Horizontal gene transfer and the evolution of bacterial cooperation. Evolution: International Journal of Organic Evolution 65, 1 (2011), 21 -- 32. [41] Geoffrey F Miller, Peter M Todd, and Shailesh U Hegde. 1989. Designing Neural Networks using Genetic Algorithms.. In ICGA, Vol. 89. 379 -- 384. [42] Volodymyr Mnih, Adrià Puigdomènech Badia, Mehdi Mirza, Alex Graves, Timo- thy P. Lillicrap, Tim Harley, David Silver, and Koray Kavukcuoglu. 2016. Asyn- chronous Methods for Deep Reinforcement Learning. CoRR abs/1602.01783 (2016). arXiv:1602.01783 http://arxiv.org/abs/1602.01783 [43] Martin A. Nowak. 2006. Five Rules for the Evolution of Cooperation. Science 314, 5805 (2006), 1560 -- 1563. https://doi.org/10.1126/science.1133755 arXiv:http://science.sciencemag.org/content/314/5805/1560.full.pdf [44] Martin A Nowak and Karl Sigmund. 2005. Evolution of indirect reciprocity. Nature 437, 7063 (2005), 1291. [45] Pamela Oliver. 1980. Rewards and punishments as selective incentives for collec- tive action: theoretical investigations. American journal of sociology 85, 6 (1980), 1356 -- 1375. [46] Elinor Ostrom, James Walker, and Roy Gardner. 1992. Covenants with and without a sword: Self-governance is possible. American political science Review 86, 2 (1992), 404 -- 417. [47] Liviu Panait and Sean Luke. 2005. Cooperative multi-agent learning: The state of the art. Autonomous agents and multi-agent systems 11, 3 (2005), 387 -- 434. [48] Julien Pérolat, Joel Z. Leibo, Vinícius Flores Zambaldi, Charles Beattie, Karl Tuyls, and Thore Graepel. 2017. A multi-agent reinforcement learning model of common- pool resource appropriation. CoRR abs/1707.06600 (2017). arXiv:1707.06600 http://arxiv.org/abs/1707.06600 [49] Alexander Peysakhovich and Adam Lerer. 2018. Consequentialist conditional cooperation in social dilemmas with imperfect information. In International Con- ference on Learning Representations. https://openreview.net/forum?id=BkabRiQpb [50] Mitchell A Potter and Kenneth A De Jong. 2000. Cooperative coevolution: An architecture for evolving coadapted subcomponents. Evolutionary computation 8, 1 (2000), 1 -- 29. [51] Francisco C Santos, Jorge M Pacheco, and Tom Lenaerts. 2006. Cooperation prevails when individuals adjust their social ties. PLoS computational biology 2, 10 (2006), e140. [52] Satinder Singh, Richard L Lewis, and Andrew G Barto. 2009. Where do rewards come from. In Proceedings of the annual conference of the cognitive science society. 2601 -- 2606. [53] Satinder Singh, Richard L Lewis, Andrew G Barto, and Jonathan Sorg. 2010. Intrinsically motivated reinforcement learning: An evolutionary perspective. IEEE Transactions on Autonomous Mental Development 2, 2 (2010), 70 -- 82. [54] J Maynard Smith. 1964. Group selection and kin selection. Nature 201, 4924 (1964), 1145. [55] Kenneth O Stanley and Risto Miikkulainen. 2002. Evolving neural networks through augmenting topologies. Evolutionary computation 10, 2 (2002), 99 -- 127. [56] Robert L Trivers. 1971. The evolution of reciprocal altruism. The Quarterly review of biology 46, 1 (1971), 35 -- 57. [57] Francisco Úbeda and Edgar A Duéñez-Guzmán. 2011. Power and corruption. Evolution: International Journal of Organic Evolution 65, 4 (2011), 1127 -- 1139. [58] Stuart A West, Ashleigh S Griffin, Andy Gardner, and Stephen P Diggle. 2006. Social evolution theory for microorganisms. Nature Reviews Microbiology 4, 8 (2006), 597. [59] Daan Wierstra, Tom Schaul, Jan Peters, and Juergen Schmidhuber. 2008. Natural evolution strategies. In Evolutionary Computation, 2008. CEC 2008.(IEEE World Congress on Computational Intelligence). IEEE Congress on. IEEE, 3381 -- 3387. [60] David Sloan Wilson. 1975. A theory of group selection. Proceedings of the national academy of sciences 72, 1 (1975), 143 -- 146. [61] Chern Han Yong and Risto Miikkulainen. 2001. Cooperative coevolution of multi-agent systems. University of Texas at Austin, Austin, TX (2001).
1506.09032	1	1506	2015-06-30T11:11:58	Organization of Multi-Agent Systems: An Overview	[ "cs.MA" ]	In complex, open, and heterogeneous environments, agents must be able to reorganize towards the most appropriate organizations to adapt unpredictable environment changes within Multi-Agent Systems (MAS). Types of reorganization can be seen from two different levels. The individual agents level (micro-level) in which an agent changes its behaviors and interactions with other agents to adapt its local environment. And the organizational level (macro-level) in which the whole system changes it structure by adding or removing agents. This chapter is dedicated to overview different aspects of what is called MAS Organization including its motivations, paradigms, models, and techniques adopted for statically or dynamically organizing agents in MAS.	cs.MA	cs	International Journal of Intelligent Information Systems 2015; 4(3): 46-57 Published online June 26, 2015 (http://www.sciencepublishinggroup.com/j/ijiis) doi: 10.11648/j.ijiis.20150403.11 ISSN: 2328-7675 (Print); ISSN: 2328-7683 (Online) Organization of Multi-Agent Systems: An Overview Hosny Ahmed Abbas1, , Samir Ibrahim Shaheen2, Mohammed Hussein Amin1 1Department of Electrical Engineering, Assiut University, Assiut, Egypt 2Department of Computer Engineering, Cairo University, Giza, Egypt Email Address: [email protected] (H. A. Abbas), [email protected] (S. I. Shaheen), [email protected] (M. H. Amin) To cite this article: Hosny Ahmed Abbas, Samir Ibrahim Shaheen, Mohammed Hussein Amin. Organization of Multi-Agent Systems: An Overview. International Journal of Intelligent Information Systems. Vol. 4, No. 3, 2015, pp. 46-57. doi: 10.11648/j.ijiis.20150403.11 Abstract: In complex, open, and heterogeneous environments, agents must be able to reorganize towards the most appropriate organizations to adapt unpredictable environment changes within Multi-Agent Systems (MAS). Types of reorganization can be seen from two different levels. The individual agents level (micro-level) in which an agent changes its behaviors and interactions with other agents to adapt its local environment. And the organizational level (macro-level) in which the whole system changes it structure by adding or removing agents. This chapter is dedicated to overview different aspects of what is called MAS Organization including its motivations, paradigms, models, and techniques adopted for statically or dynamically organizing agents in MAS. Keywords: Multi-Agent Systems, Organization, Organizational Models, Dynamic Reorganization, Self-Organization 1. Introduction [3]. Further, the Complexity and highly distribution are the key characteristics of modern real world systems. The complexity of the near future and even present applications can be characterized as a combination of aspects such as great number of components taking part in the applications, knowledge and control have to be distributed, the presence of non-linear processes in the system, the fact that the system is more and more often open, its environment dynamic and the interactions unpredictable increasing complexity, heterogeneity, and openness of modern software systems have reached a point that imposes new demands on their engineering that conventional engineering approaches will stand powerless in front of future systems increase in scale and complexity either vertically (control and information layers) or horizontally (physical distribution). It doesn’t mean that conventional engineering techniques will become obsolete and have to be thrown away. Absolutely, they only need to be integrated with new as, decomposition, autonomy, modularity, and adaptivity can be collectively combined in one system. MAS are considered as a promising engineering (i.e., architectural) style for developing adaptive software systems able to handle the continuous increase in their complexity as a result of their open, heterogeneous, and continuous evolution nature. They model styles where is expected technologies. It engineering concepts such systems consisting of many the system as distributed autonomous agents cooperate together to achieve system goals. The ability of agents to dynamically reorganize to adapt working environment dynamic changes is a key feature provided by MAS. It is obvious that the natural way to model a complex system is in terms of multiple autonomous components that can act and interact in flexible ways in order to achieve their objectives, and also that agents provide a suitable abstraction for modeling subsystems, components and their relationships [22]. Ferber [23] described how agents, as a form of distributed artificial intelligence, are suitable for use in application domains which are widely distributed. MAS are currently considered as the most representatives among artificial systems dealing with complexity and highly distribution [24]. MAS allow the design and implementation of software systems using the same ideas and concepts that are the very founding of human societies and habits. These systems often rely on the delegation of goals and tasks among autonomous software agents, which can interact and collaborate with others to achieve common goals [34]. In other words, an agent falls somewhere between a simple event-triggered program and one with human collaborative abilities [36]. In contrast to initial MAS research, which concerned individual agents’ aspects such as agents’ architectures, agents’ 47 Hosny Ahmed Abbas et al.: Organization of Multi-Agent Systems: An Overview to the environment conditions and mental capabilities, behaviors, etc, the current research trend of MAS is actively interested in the adaptivity, environment, openness and the dynamics of these systems. Also, there is a great attention towards the MAS technique as a way to design self-organized systems. In open environments, agents must be able to adapt towards the most appropriate organizations according their unpredictable changes. Agent organizations are considered as an emergent area of MAS research that relies on the notion of openness and heterogeneity of MAS and imposes new demands on traditional MAS models [44]. MAS that have the ability to dynamically reorganize (regardless of the type of reorganization, self or enforced) will be adaptive enough to survive against their dynamic and continuously changing working environments. Dynamic reorganization can take many forms, for instance, agents can dynamically change their roles, behaviors, locations, acquaintances, or the whole system organization structure can be dynamically changed. following An agent organization can also be defined as a social entity composed of a specific number of members (agents) that accomplish several distinct tasks or functions and that are structured and communication interrelationships in order to achieve the main aim of the organization. Thus, agent organizations assume the existence of global common goals, outside the objectives of any individual agent, and they exist independently of agents [64][65]. some specific topology This chapter is dedicated to provide a comprehensive overview of MAS organization including its motivations, paradigms, and familiar organizational models. The remaining of this chapter is organized as follows: Section 2 explores MAS literature to identify the motivations towards agent organizations. Section 3 presents different approaches and paradigms used to organize agents within multi-agent systems. Section 4 introduces what is called organizational models, which concern the abstractions, languages, approaches and techniques for modeling dynamically reorganized MAS. And Section 5 concludes the article and highlights future work. 2. Motivations to MAS Organization This section is dedicated to identify from MAS literature the suggested motivations to give increasing attention to MAS organization. Basically, a MAS is formed by the collection of autonomous agents situated in a certain environment, respond to their environment dynamic changes, interact with other agents, and persist to achieve their own goals or the global system goals. There are two viewpoints of MAS engineering, the first one is the agent-centered MAS (ACMAS) in which the focus is given to individual agents. With this viewpoint, the designer concerns the local behaviors of agents and also their interactions without concerning the global structure of the system. The global required function of the system is supposed to emerge as a result of the lower level individual agents interactions in a bottom-up way. Picard et al. [13] stated that the agent-centered approach takes the agents as the “engine” for the system organization, and agent organizations implicitly exist as observable emergent phenomena, which states a unified bottom-up and objective global view of the pattern of cooperation between agents. Further, Picard gives the ant colony [15] as an example, where there is no organizational behavior and constraints are explicitly and directly defined inside the ants. The main idea is that the organization is the result of the collective emergent behavior due to how agents act their individual behaviors and interact in a common shared and dynamic environment. The key problems of the ACMAS viewpoint are unpredictability and uncertainty. Because the whole is more than the sum of its parts [14], this approach can lead to undesirable emergent behaviors that may impact system performance, as a result, this approach might be not suitable to design and engineer complex multi-agent systems. The MAS applications engineered by the ACMAS approach are closed for agents that are not able to use the same type of coordination and behavior, and that all global characteristics and requirements are implemented in the individual agents and not outside them [10]. Weyns [11] stated that giving the responsibility of system organization implicitly to individual agents, as in the ACMAS approach, in addition to their functional responsibilities is not adequate because it is a type of dual responsibility, which is very complex to engineer and not suitable for handling real world complexity and other emerged characteristics such as highly distribution, unpredictability, uncertainty, and continuous evolution. The second viewpoint of MAS engineering is what is called organization-centered MAS (OCMAS) in which the structure of the system is given a bigger attention through the explicit abstraction of agent organization. With that approach, the designer designs the entire organization and coordination patterns on the one hand, and the agents’ local behaviors on the other hand. It is considered as a top-down approach because the organization abstraction imposes some rules or norms used by agents to coordinate their local behaviors and interactions with other agents. The OCMAS viewpoint has been promoted by many pioneers in MAS research. For instance, Jennings and Wooldridge [2] stated that MAS contribute to the software engineering (SE) discipline as a way to simplify the design of complex software systems but considering MAS with no real structure isn’t suitable for handling current software systems complexity, and higher order abstractions should be used and some way of structuring the society is typically needed to reduce system complexity, to increase system efficiency, and to more accurately model the problem being tackled. Odell et al. [4] stated that the current practice of MAS design tends to be limited to individual agents and small face-to-face groups of agents that operate as closed systems which is not adequate to model and design of complex adaptive systems. Also Gutknecht and Ferber [66] argued that taking organizational concepts, such as groups, roles, structures, dependencies, etc, as first class citizens, and relating them to the behavior of agents is a key issue for building large scale and complex systems. In another article, Ferber [6] also stated that International Journal of Intelligent Information Systems 2015; 4(3): 46-57 48 representing a MAS as an organization consists of roles enacted by agents arranged (statically or dynamically) to form groups of agents, can handle many drawbacks such as system complexity, uncertainty, and system dynamism. truly heterogeneous Gasser [3] stated that we simply have hardly any real experience building realistically coordinated multi-agent systems that work together and almost no basis for systematic reflection and analysis of that experience. Further, Horling et al. [5] stated that our real world getting more complex and highly distributed and that should be reflected in new software engineering paradigms such as MAS. Therefore, the adoption of higher order abstract concepts like organizations, societies, communities, and groups of agents can reduce systems complexity, increase its efficiency, and improve system scalability. (Coordination, Organization, Institutions and Norms in agent systems) which concerns the OCMAS viewpoint. in the near future, The OCMAS viewpoint is more adequate for engineering complex adaptive multi-agent systems, which are expected to be, the mainstream approach for engineering large-scale and even ultra-large scale application domains especially with the evolving topic of the Internet of Things (IoT) [17], which concerns devices capable to communicate via the Internet and manipulate an enormous amount of data. Examples of such application domains are CPS (Cyber-Physical Systems) [16], Smart Grids [18], global SCADA (Supervisor Control and Data Acquisition) [19], Pervasive Computing [20], Ubiquitous Computing [21], etc. The next section explores the familiar paradigms of MAS organization. together helps in a system should work Establishing an organizational structure that specifies how agents the achievement of effective coordination in MAS [39]. Broek [7] stated that complexity of real world applications needs to be tackled from higher abstraction order such as organizations which can be used to limit the scope of interactions, provide strength in numbers, reduce or manage uncertainty, and formalize high-level goals which no single agent may be aware of. Further, Hübner [8] confirmed that organizations provide a framework for structuring and managing agents’ interactions and serve as a kind of tuning of the agents autonomy level. Furthermore, Burns et al. [12] stated that in organization theory [25][26], it is commonly accepted that different types of organizational structure are suitable for particular environmental conditions and one of the main reasons for creating organizations is to provide stable means for coordination that enable the achievement of global goals. Moreover, Corkill et al. [36] stated that as agent-based systems become more widespread and complex, designed organization will become an important aspect of effective system performance, and the possible situations where organization design will be very important such as, large number of agents, long duration of agent activities, more repetitive activities, more activities require shared resources, more collaborative the activities, more specialized agents, less capable agents, and less slack resources are available. Also, they emphasized that no one organization is right for every situation. they suggested In nutshell, proposing a way for statically or dynamically organizing MAS, has been given great attention by MAS researchers, as a promising approach for handling the challenging issue of engineering complex and large-scale software systems. The adoption of the ACMAS or OCMAS viewpoints mainly depends on the nature of application domain and the degree of system complexity. The developers interested in bottom-up self-organized systems will prefer the ACMAS approach and the developers interested in top-down system reconfiguration will prefer the OCMAS approach. In the MAS literature there are two communities each adopts and concerns one of the two engineering approaches. The first one is SASO (Self-Adaptive and Self-Organizing systems) which concerns the ACMAS viewpoint. And the second one is COIN 3. Paradigms of MAS Organization the organization management and Originally, the organization abstract is inspired from business human organizations, which are constituted of a number of roles, so a key concept in the design of OCMAS is that of roles, which define normative behavioral repertoires for agents [4]. A role is defined as an abstract description of some activity or functionality, for instance in a business human organization we may see a role like Manager who is responsible of the coordination between other organization members (roles). In MAS, agents are supposed to enact roles according to the capabilities of each agent. It is also possible that one agent can enact many roles in the same time. The role enacted by an agent has a direct effect on the agent behavior and interaction with other roles (agents). Odell et al. [27] described two familiar ways for assigning roles to agents, endogenously by emergent self-organization as the system runs, or exogenously by the system designer when the system is constructed or modified. The adoption of human organization theory was the focus of distributed systems in general before multi-agent systems, which systems [30][31][32][33]. themselves distributed are they Modern organizations (real or virtual) are characterized by information flows, and their complex structure, dense incorporation of technology, also information characterized by highly dynamic, constantly changing, organic structure and show hardly identified, not formalized, non-linear behavior [28][29]. These challenges enforce the urgent need to a new way of engineering multi-agent systems. [37] described an agent organization as an entity that is composed of a set of agents, working together to achieve a shared purpose through a division of labor, integrated by decision processes continuously through time. Further, Galbraith pointed out that an organization consists of patterns of behavior and interaction that are relatively stable and change slowly over time. from human organizations, Galbraith Inspired Shehory [1] defined MAS organization as the way in which multiple agents are organized to form a multi-agent system. The relationships and interactions among the agents and 49 Hosny Ahmed Abbas et al.: Organization of Multi-Agent Systems: An Overview specific roles of agents within the organization are the focus of multi-agent organization. The use of organizations provides a new way for describing the structures and the interactions that take place in MAS. Dignum [10] stated that agent organization can be understood from two perspectives: organization as a process and organization as an entity. In other words, organizations can be considered as the process of organizing a set of individual agents, thus in this sense it is used to refer to constraints (structures, norms and patterns) found in a social context that shape the actions and interactions of agents [53]. In other situations, it can be considered as an entity in itself, with its own requirements and objectives and is represented by (but not identical to) a group of agents. In fact, agent organizations demand the integration of both perspectives and rely for a great extent on the notion of openness and heterogeneity of MAS. Figure 1 illustrates how a MAS can be seen from two levels, the individual agents’ level and the organizational level. The organizational level presents a higher order abstraction of the lower agents’ level. Figure 1. Organizational level vs. individual level in MAS. Ferber et al. [6] proposed a set of general principles that should be taken into account when designing MAS with organizational dimension: 1. The organizational level describes the “what” and not the “how”. In other words, the organizational level imposes a structure into the pattern of agents’ activities, but does not describe how agents behave. 2. No agent description and therefore no mental issues at the organizational level. The organizational level should not say anything about the way agents would interpret this level. 3. An organization provides a way for partitioning a system, each partition (or agent group) constitutes a context of interaction for agents. Thus, a group is an organizational unit in which all members are able to interact freely. Ferber principles provide important general guidelines for OCMAS research. They identify precisely the logical relation between agents and their organization regardless of the nature of organization (i.e. a process or an entity). The first principle concerns the autonomy of agents. Agents should be autonomous but they may be guided by some general organizational norms or constraints. Full autonomy is not a preferred agent characteristic in MAS research, we can only find a type of full autonomy with humans because they have perfect rational minds, but agents (software or hardware) designed for specific missions in certain application domains and the concept of safety imposes some constraints on agents’ autonomy, in these situations, a designed organization, where agents give up some degree of self-motivation and autonomy can be an appropriate choice [36]. The second principle concerns the unawareness of agents about the existence of the organizational level, which according to Ferber should be transparent from agents. In other words, agents should be affected indirectly by the change of system organization (i.e., through environment). The third principle concerns system modularity. Organizations provide a way for, statically or dynamically, decomposing the system. Modularity and flexibility of system decomposition enhance system maintainability. teams, congregations, societies, the major organizational paradigms used Horling and Lesser [5] also stated that organizational design employed by an agent system can have a significant, quantitative effect on its performance characteristics, and they surveyed in multi-agent systems. These include hierarchies, holarchies, coalitions, federations, markets, and matrix organizations. Also, they provided a description of each paradigm, and discuss its advantages and disadvantages, further, they provided examples of how each organization paradigm may be instantiated and maintained. Table 1 provides a summary of Horling and Lesser [5] work. The Table contains a number of methods by which MAS could be organized and highlights the key characteristics, benefits, and drawbacks of each organization paradigm. Similar work was provided by Carley and Gasser [35]. The main conclusion of these surveys is that no single organization paradigm is necessarily better than all others in all situations. The selection made by a designer should be dictated by the needs imposed by the system's goals, the resources at hand, and the environment in which the participants will exist. In other words, an organization paradigm that can be described as a fit-to-all paradigm does not exist (at least till now!). A MAS can be statically (in design time) organized using any of the organization paradigms presented in Table 1, not only this but also hybrids of these and others in addition to dynamic changes from one organization style to another are also possible [1] with the price of implementation complexity. The later case is called dynamic reorganization which is currently a very active research area within MAS discipline. The next subsections present in more details the concept of dynamic reorganization and relevant concepts, self-organization and emergence. its captivating 3.1. Dynamic Reorganization Earlier proposed MAS organization mechanisms tackled with organizational aspects at design time, that approach requires some important initial knowledge about the exact purposes and objectives of the system-to-be and every interaction to which it may be confronted in the future have to be known in design time [41]. However, the openness, complexity, and heterogeneity of modern software systems impose new demands and requirements on agent-oriented software engineering (AOSE) [71], which is concerned with International Journal of Intelligent Information Systems 2015; 4(3): 46-57 50 the development of feasible, effective, and adaptive MAS. Building adaptive MAS (AMAS) able to handle openness, complexity, and highly distribution of modern real world applications has recently attracted great attention. Table 1. Analysis of Some of Possible MAS Organization Paradigms (adopted from [5]). Paradigm Key Characteristic Benefits Decomposition Maps to many common domains; handles scale well Holarchy Decomposition with autonomy Exploit autonomy of functional units Coalition Dynamic, goal-directed Exploit strength in number Team Congregation Group level cohesion Long-lived, utility-directed Address larger grained problems; task-centric Facilitates agent discovery Society Open system Public services; well defined conventions Federation Middle-agents Market Competition through pricing Matchmaking, brokering, translation services, facilitates dynamic agent pool Good at allocation, increased utility through centralization, increased fairness through bidding Matrix Multiple managers Resource sharing, multiple influenced agents Compound Concurrent organizations Exploit benefits of several organizational styles Drawbacks Potentially brittle, can lead to bottlenecks or delays Must organize holons, lake of predicable performance Short-term benefits may not outweigh organization construction costs Increased communication Sets may be overly restrictive Potentially complex, agents may require additional society-related capabilities Intermediaries become bottlenecks Potential for collusion, malicious behaviour, allocation decision complexity can be high Potential of conflicts, need for increased agent sophistication Increased sophistication, drawbacks of several organizational styles AMAS designed to be capable to adapt themselves to unforeseen situations in an autonomous manner. They can be realized by enabling the system to dynamically reorganize to adapt its environment changes [42]. Dynamic reorganization is a way to design and develop AMAS. It can be described as the change of MAS structure and behavior as a result of internal (local) or external (supervisory) demand. The external demand can be for example human intervention. The internal demand emerges from the system itself as an autonomous system to adapt environments changes. Generally, dynamic reorganization in MAS takes place as a result of individual agents' interactions. However, in many application domains the environment can stimulate MAS reorganization (e.g., when removing or adding environment resources), the system may reorganize to adapt the change of environment. In other words, reorganization is the answer to change in the environment. Dignum et al. [40] identified two types of MAS dynamic reorganization, emergent Organization in which global behavior cannot be specified in advance, but emerges from the interaction of local behaviors. In other words, agents’ interactions may eventually create dynamic organizations [44]. Thus, emergent organizational behavior is primarily a bottom-up process in which agents look for interaction and local control decisions that have been effective in the past and give similar decisions preference in the future. The ACMAS viewpoint concerns this type of reorganization. The other type of reorganization is called designed organization, which has an explicit the coordination of the agents participating. Designed systems are design created and task-environment to develop an explicit organizational structure, that is then elaborated by the individual agents into appropriate behaviors. Designed organization exhibits predicable and controllable behavior, dynamic change implies the need for highly intelligent and organization information knowledge interaction structure that determines using communicative agents (at least some of them) that can reason about and negotiate change. Designed organization is the main concern of the OCMAS approach. In human organizations, it has been proven that designed organizations perform better than those that emerge naturally. This viewpoint holds for agent organizations as well, that is because the global behavior of emergent organizations cannot be predicted and changes cannot be guided, which makes this type less suitable for situations where coordinated and goal-directed global action is required. Picard et al. [13] added the agents’ awareness /unawareness of the existence of the organization structure as a dimension of the organization modification process and he identified four cases: 1. The agents don’t represent the organization, although the observer can see an emergent organization. In some sense, they are unaware that they are part of an organization. 2. Each agent has an internal and local representation of cooperation patterns which it follows when deciding what to do. This local representation is obtained either by perception, communication or explicit reasoning. 3. The organization exists as a specified and formalized schema, made by a designer but agents don’t know anything about it and even do not reason about it. They simply comply with it as if the organizational constraints were hard-coded inside them. 4. Agents have an explicit the organization which has been defined. The agents are able to reason about it and to use it in order to initiate cooperation with other agents in the system. The agents are able to reason about it and to use it in order to initiate cooperation with other agents in the system. representation of Case 1 and 2 considered as ACMAS and case 3 and 4 considered as OCMAS. The importance of Picard classification of MAS dynamic reorganization is that nearly 51 Hosny Ahmed Abbas et al.: Organization of Multi-Agent Systems: An Overview most of known reorganization methods fit to a specific case or multiple cases. Similar classification proposed by Sichman et al. [44], but he used the concept of observer in the same position as the agent awareness of the organization. Table 2 provides the global picture of possible types of MAS organization with examples. As shown in the table MAS the organization awareness/unawareness of the presence of the organizational level. individual agents about according classified to is Table 2. The global picture of MAS organization. Agents unaware Agents aware ACMAS=Emergent organization Organization is observed. It is implicitly programmed in agents, interactions, and environment. Organization is observed. Coalition mechanisms programmed in the agents. Concerned Community SASO COIN Examples Swarm-based systems [72] Contract-Net Interaction Protocol [73] OCMAS = Designed organization Organization is a design model. It may be hard coded in the agents. Organization is programmed in the agents and/or in specialized middleware services. Concerned Community COIN Examples AOSE methodologies such as: MASE [74], INGENIAS [75]. Picard also, after finishing his valuable study proposed a comprehensive definition of dynamic reorganization as follows: “Reorganization is a process, endogenous or exogenous, concerning systems in which organization is explicitly manipulated through specifications, constraints or other means, in order to ensure an adequate global behavior, when the organization is not adapted. Agents being aware of the organization state and structure, they are capable of manipulating primitives to modify their social environment. This process can be both initiated by an external entity or by agents themselves, by reasoning directly on the organization (roles, organizational specification) and the cooperation patterns (dependencies, commitments, powers).” This definition assumes that the agents are aware of the existence of the organizational level, thus it concerns the OCMAS viewpoint. But, what if agents are unaware of the organization level? According to Picard, in this case the dynamic reorganization process is called self-organization which defined by Picard as follows: “Self-organization is an endogenous and bottom-up process concerning systems in which only local information and representations are manipulated by agents unaware of the organization as a whole, in order to adapt the system to the environmental pressure by modifying the organization, therefore by changing directly the system configuration influences, differentiation), or the environment of the system, by local interactions and propagation, by avoiding predefined model biases.” neighborhoods, (topology, indirectly This definition states that self-organization represents the ACMAS viewpoint. In a self-organized system, agents are unaware of the organization level, the reorganization process is decentralized, implicit, endogenous, and agents are responsible of the system dynamic reorganization, which is often initiated by an environmental change. In a dynamically reorganizing system where agents are aware of the organization level, this process can be decentralized or not, but always explicit and directly performed by entities (designer or COIN Organizational models such as: AGR [58] MOISE+ [67] agents) manipulating organizational primitives. Therefore, the awareness is a key dimension added by Picard to identify self-organized MAS. The next section provides detailed review of the self-organization concept in MAS. 3.2. Self-Organization recently in concepts appeared self-configuration, etc) are Self- properties [43] (i.e., self-organization, self-healing, the most self-adaptation, captivating software engineering. They remind us of Einstein ideas about Time Machine, which was and still a far dream of human to travel through time. Human also dreams to design a system, regardless of its nature (software or hardware), able to do all things by itself. A system has all known self-* properties will be amazing. Actually, this type of systems is imaginary (at least till now!); we can only see this system in science fiction movies. However, it is possible to design systems with one or more of self-* properties for predefined purposes and under certain circumstances. The first use of the term self-organization returns to Ashby [56], in 1947, he stated that a system is said to be self-organized if it changed its own organization rather than being changed by an external entity. Self-organization in software systems received great attention since the last few years. It is an attractive way to handle the dynamic requirements in software in general and MAS in specific. It refers to a process where a system changes its internal organization to adapt to changes in its goals and its working environment without explicit external control. Understanding the mechanisms that can be used to model, assess and engineer self-organizing behavior in MAS is currently an issue of major interest [38]. Picard’s definition of self-organization (see previous section) can be rephrased as follows: Self-organization is a process where some form of overall order or coordination arises out of the local interactions between the components of an initially disordered system. This process is spontaneous: it is not necessarily directed or controlled by any agent or subsystem inside or outside of the system. It is often triggered International Journal of Intelligent Information Systems 2015; 4(3): 46-57 52 by random fluctuations, which are triggered and amplified by positive feedback. The resulting organization is wholly decentralized or distributed over all the components of the system, it is typically very robust and able to survive and self-repair substantial damage or perturbations. The roots of the term self-organization return to the work of Glansdorff and Prigogine [46] through thermodynamics studies. They discovered that open systems decrease their entropy (order comes out of disorder) when an external energy is applied on the system. Matter organizes itself under this external pressure to reach a new state where entropy has decreased. Nature is full of self-organization forms and patterns, for instance social behavior of insects like ants or termites, which formed as a result of indirect communication through environment without the need for any type of direct interaction, this type of interaction is called Stigmergy. Social behavior of humans is also self-organized and gives rise to emergent complex global behaviors. Human beings typically work with local information and through local direct or indirect interactions producing complex societies [45]. Researchers from variety of disciplines who were interested in self-organization in nature found MAS as the adequate engineering style for modeling and simulation of the self-organization phenomena and after a period of time the situation reversed as the MAS researchers, who are concerned with AOSE gave a greet attention to bio-inspired models for developing complex, open, and heterogeneous MAS-based applications. Self-organization and emergence are currently the main focus of AOSE researchers. The adoption of naturally inspired methods and approaches for engineering self-organized MAS is currently a very active research area [47][48]. Mechanisms such as direct interactions [49], Stigmergy [50], reinforcement [51], and agents’ cooperation [52] are widely used to design MAS with self-organization behavior. Another relevant and interesting concept is that of emergence, which can be considered as a process takes place in complex systems (which may or may not be self-organized). Self-organization results from emergence, but there is no guarantee that a self-organized system will always generate emergent phenomena. Understanding how to engineer systems that are capable of presenting self-organized behavior and desirable emergence is currently a very active research area too. The next section introduces briefly the concept of emergence. 3.3. Emergence A lot of confusion exists about the meaning of the two relevant terms emergence and self-organization. One of the sources of the confusion comes from the fact that a combination of both phenomena often occurs in dynamical systems [53]. In MAS domain, self-organization and emergence concepts are recently getting great focus as a way to engineer open, heterogeneous, and complex MAS-based applications such as complex adaptive systems (CAS) [54], which are fluidly changing collections of distributed interacting components that react to both their environments and to one another. The familiar definition of emergence is as a phenomenon where global behavior arises from the interactions between the local parts of the system. This general and vague definition indicates that still there is no consensus of a clear definition for emergence. Also, it indicates the absence of clear understating of its nature. In contrast to the reductionism theory [55], which allows a system to be reduced to the sum of its parts, the emergent global behavior cannot be predicted by observing its parts local behaviors. An accepted operational definition of emergence was proposed by De Wolf and Holvoet [53] as follows: “A system exhibits emergence when there are coherent emergents at the macro-level that dynamically arise from the interactions between the parts at the micro-level. Such emergents are novel with respect to the individual parts of the system.” This definition uses the concept of an ‘emergent’ as a general term to denote the result of the process of emergence, i.e., properties, behavior, structure, patterns, etc. The ‘level’ mentioned refers to certain points of view. The macro-level considers the system as a whole and the micro-level considers the system from the viewpoint of the individual entities that make up the system. The concept of emergence is very complex and it is not fall in the scope of this article, interested readers are invited to explore the emergence relevant references. 3.4. Discussion their Static design of MAS is not adequate for modern real world applications, which characterized by increasing complexity heterogeneity, and openness. Even closed systems in which the number of agents is constant with time should have a type of adaptive dynamic behaviors. All possible behaviors of modern systems cannot be captured at design time and that requires these systems to be adaptive able to adapt changes in their working environments. Dynamic reorganization is currently a familiar way for developing adaptive MAS. As shown in Section 2, the adoption of organizational aspects within MAS is promoted and recommended by pioneers of MAS research. Dynamic reorganization can be described as the change of MAS structure and behavior as a result of internal or external demand. The external demand can be for example human intervention. The internal demand emerges from the system itself as an autonomous system to adapt environments changes. Self-organization is a dynamical and adaptive process where systems acquire and maintain structure themselves, without external control. In self-organizing systems, robustness is used in terms of adaptivity in the presence of perturbations and change. A self-organizing system is expected to cope with that change and to maintain its organization autonomously. Emergence emphasizes the presence of a novel coherent macro-level emergent (property, behavior, structure, etc) as a result of the interactions between micro-level parts. A combination of emergence and self-organization is a promising approach to engineer large-scale multi-agent 53 Hosny Ahmed Abbas et al.: Organization of Multi-Agent Systems: An Overview systems. In most systems that are considered in MAS literature, emergence and self-organization occur together. Research in MAS and CPS communities focuses on such systems. In very complex (multi-agent) systems, i.e. distributed, open, large, situated in a dynamic context, etc., the combination of is recommended. self-organization emergence and When a researcher proposes an approach to dynamically reorganize a multi-agent system to adapt environments’ changes, he actually proposes what the MAS domain consensus agreed to call as an Organizational (or organization) Model. MAS organizational models will play a critical role in the development of future larger and more complex MAS. The main concern of organizational models is to describe the structural and dynamical aspects of organizations [9]. They have proven to be a useful tool for the analysis and design of multi-agent systems. Furthermore, they provide a framework to manage and engineer agent organizations, dynamic reorganization, self-organization, emergence, and autonomy within multi-agent systems. The next section introduces organizational models in some details. 4. Organizational Models software agent-oriented Organizational models have been recently used in agent theory for modeling coordination in open systems and to ensure social order in MAS applications [64]. The adoption of organizational models is currently given great importance within most engineering methodologies. The motivation to this direction is that in open environments, agents must be able to adapt towards the most appropriate organizations according to the environment conditions and their unpredictable changes. As a result, organizational models should guarantee the ability of organizations to dynamically reorganize as a response to dynamic environment changes. Organizational models are responsible of how efficiently and effectively organizations carry out their tasks, they have been recently used in agent theory for modeling coordination in open systems and to ensure social order in multi-agent system applications [7]. From the business management discipline an organizational model, also called as organizational structure, defines an organization through its framework, including lines of authority, communications, duties and resource allocations. A model is driven by the organization’s goals and serves as the context in which processes operate and business is done. The ideal model depends on the nature of the business and the challenges it faces. In turn, the model determines the number of roles needed and their required skill sets. In MAS, the purpose of an organizational model is to enhance the analysis and design of OCMAS, so it's usually integrated with a particular agent-based software engineering methodology. Before exploring some of the familiar organizational models proposed for modeling complex MAS, it is a suitable time to show the difference between them and MAS development methodologies. In general, a methodology is a body of methods employed by a discipline. A method is a procedure for attaining something. A methodology aims to prescribe all the elements necessary for the development of a software system [57]. AOSE community concerns creating development methodologies suitable for the development of agent-oriented or agent-based software. Typically, a development methodology (agent-oriented or not) comprises an ordered set of phases such planning, analysis, design, implementation, deployment. An organizational model is a tool adopted within a development methodology for modeling the system-to-be. Typically, it starts in the analysis phase but can expand through design and implementation phases, or in other cases it can expanded through the whole development life cycle. validation, and The next section explores some of familiar proposed MAS organizational models focusing in their tackled organizational aspects, their advantages, and their disadvantages. 4.1. Familiar Organizational Models There is a lot of MAS organizational models proposed in the literature; each of them tackles MAS organization from a different viewpoint. Some of them adopt the ACMAS viewpoint, others adopt the OCMAS viewpoint, and some adopt a hybrid approach concerns both ACMAS and OCMAS viewpoints. In what follows, three of familiar organizational models are introduced. 4.1.1. AGR and AGRE Ferber et al. [6] proposed a very concise and minimal OCMAS model called AGR, for Agent/Group/Role, also known as the AALAADIN model [58]. The authors of AGR model proposed a set of notations and a methodological framework to help the designer to build MAS using AGR. Further, they presented a set of diagrams (organizational structure, cheeseboard diagram, and organizational sequence diagrams), which may represent the different aspects (static and dynamic) of OCMAS. Their model is based on the dynamic creation of agents groups (agents partitioning) and dynamic forming of hierarchies of groups (Holarchies). They pointed out that their AGR-based model can be integrated with Gaia [59] MAS development methodology to complete the analysis and design phases of MAS development. Figure 2 presents the AGR meta-model. Figure 2. The AGR Meta-Model. International Journal of Intelligent Information Systems 2015; 4(3): 46-57 54 The core concepts on which the AGR model is based are agent, group and role. The agent in AGR is assumed to be an active, communicating entity which plays roles within groups, with no restrictions on its internal architecture. The group is defined as the basic unit of agent aggregation. Each agent is part of one or more of these groups. The role is an abstract representation of an agent function, service or identification within a group. Each agent can handle several roles, and each role handled by an agent remains local to a group. Other important abstract concepts are also shown the AGR meta-model shown in Figure 9.1, they are Group Structure and Organizational Structure. The group structure is an abstract representation of the roles required in this group and their interaction relationships and protocols. The organization structure is the set of group structures expressing the design of a multi-agent organizational scheme. In other paper, Ferber et al. [60] presented an extension of the AGR organizational model, called AGRE (AGR + Environment), which simply geometrical) environments. This extension is based on the concept of a space which can be seen either as a physical area or as a social group. includes physical (or of agents heterogeneous relationships. On The main advantages of the AGR/AGRE models are: supporting architectures, heterogeneous communication languages, and dynamic role-group the disadvantages are multi-role agents which make agents internally complex; roles sharing between Groups can cause overloaded agents, agents ask to join groups which require highly knowledgeable agents, use of mediator agents (brokers) which can be a source of bottlenecks, and very few known real applications. the other hand, 4.1.2. MOISE for multI-agent SystEms); Hannoun et al. [67] proposed MOISE (Model of Organisation for modeling organizational aspects of MAS. Similar to the AGR model, their model is based on three major concepts: roles, organizational links (roles relations), and groups. What distinguishes the MOISE model is that it tries to integrate both viewpoints, ACMAS and OCMAS. By this way MOISE gives the chance to the designer to model totally or partially the social behavior of system agents by specify the possible organizational structures and that will be useful for system verification and validation. On the other hand, for the sake of flexibility, agents should be able to reason about their social behaviors and have a direct influence on system dynamic reorganization. The MOISE model is structured along with three levels: (i) for each agent, definition of the tasks that it is responsible of (individual level), (ii) aggregation of agents in large structures (aggregate level), (iii) global structuring and interconnection of the agents and structures with each other (society level). The organization in MOISE is viewed as a normative set of rules that constrains the agents’ behaviors [67]. The MOISE organizational model was extended by Hübner et al. to MOISE+ [68] to create an organization-centered model for independently specify the structural and functional aspects then link them by a deontic aspect. Another extension to MOISE+ [69] was done to add dynamic reorganization process to adapt environment changes. 4.1.3. MACODO in Weyns et al. [61] presented an organizational model for context-driven dynamic agent organizations. The model defines abstractions that support application developers to describe dynamic reorganization. The organizational model is part of an integrated approach, called MACODO (Middleware Architecture for COntext-driven Dynamic agent Organizations); life-cycle management of dynamic organizations is separated from the agents, organizations are first-class citizens, and their dynamics are governed by laws. Moreover, the authors provided a formal specification to describe and specify the semantics of their organizational model abstractions using Z specification language [62], which is based on set theory and first order predicate calculus. The main concern of MACODO is to directly relate organization dynamics to context changes in the environment. this model, the is We argue that the main drawback of the MACODO organizational model the pure dynamically created organization; we argue that an organization should be tackled from the two perspectives, static and dynamic for the sake of long-term system stability. Dynamically creating and vanishing of organizations without keeping an amount of static behavior can impact system stability and prevents it from reaching an equilibrium state. In other words, there should be an amount of balance between static and dynamic organizational behaviors. The authors applied their model to traffic monitoring application. To our best knowledge, there is no any other real world application designed with MACODO. 4.2. Discussion In MAS literature, there is large number of organizational models proposed by MAS researchers from all over the world to support the organizational aspects within MAS. Some of these models adopt the ACMAS viewpoint, others are concerned with the OCMAS viewpoint, and others adopt a hybrid approach combines both viewpoints. For the sake of this article size, it is not possible to explore all proposed organizational models in details, but interested readers can see [70], which is a handbook of research on MAS organizational models recent organizational model. Bellow, we provide our remarks about these models: contains many of (cid:1) This large number of organizational models indicates that concerning organizational aspects within MAS is currently a very interesting research area. (cid:1) It also emphasize that till now there is no a fit-to-all organizational models that can be used to design MAS-based systems in all application domains. (cid:1) Nearly, each of these models was dedicated to specific real world application domain and it is not applied to other applications. 55 Hosny Ahmed Abbas et al.: Organization of Multi-Agent Systems: An Overview (cid:1) Some of these organizational models tackle with organization structure issues at design time (pure static), and others tackle them at will (pure dynamic). References (cid:1) In some of them the organization abstraction is not dynamic explicit reorganization is given to individual agents in addition to their functional responsibilities. responsibility and the of (cid:1) Most of them considered the intra-organization and did not tackle with inter-organization reorganization. How to model the interaction among organizations? (cid:1) Few organizational models tackled both static and dynamic aspects of organizations and environments. (cid:1) In most of them the individual agent initiates to join a certain organization and this require that the agent has a reasonable knowledge about the services of each organization to select the appropriate one to join. We argue that letting the organization itself to select suitable agents to award it a role is the better approach because organization knowledge is more global than that of an individual agent. Based on these remarks and limitations of most previously proposed MAS organizational models, we proposed a novel organizational model for engineering complex large-scale MAS-based applications. The new MAS organizational model was called NOSHAPE and conceptually presented in [76]. NOSHAPE is supposed to handle all the limitations of other the overlapping relationships among higher order abstraction entities such as organizations of agents, worlds of organizations, and even universes of worlds within MAS to realize and utilize their captivating characteristics. related models. It exploits 5. Conclusion and Future Work the to including reorganize system-to-be MAS organization can be considered as a process to dynamically adapt environment dynamic changes. Or, it can be considered as an entity facilitates the partitioning of the system-to-be. Organizations are a typical way to structure and manage interactions among agents. Establishing an organizational structure that specifies how agents in a system should work together helps the achievement of effective coordination in MAS. This chapter provided a comprehensive overview about MAS organization its motivations, paradigms, models, and other related concepts such as self-organization and emergence. In MAS literature, we found very large number of organizational models proposed to support dynamic reorganization of the MAS, this large number of organizational models indicates that concerning organizational aspects within MAS is currently a very active and interesting research area and that till now there is no a fit-to-all MAS organizational model that can be used for engineering all possible application domains. This conclusion motivates us to propose a novel organizational model for engineering complex and highly distributed large-scale MAS such as modern industrial networks (i.e., SCADA [19]). [1] Shehory, O. Architectural properties of multi-agent systems. Technical Report CMU-RI-TR-98-28, The Robotics Institute, Carnegie Mellon University, Pittsburgh, Pennsylvania 15213, 1998. [2] Jennings, N. R., & Wooldridge, M. Agent-Oriented Software Engineering. in Bradshaw, J. ed. Handbook of Agent Technology, AAAI/MIT Press, 2000. [3] Gasser, L. (2001). Perspectives on Organizations in Multi-Agent Systems. and Applications, Michael Luck et al (pp. 1–16). Berlin: Springer-Verlag. doi:10.1007/3-540-47745-4_1. In Multi-Agent Systems [4] Odell, J. J., Parunak, H. V. D., & Fleischer, M. (2003). The role of roles in designing effective agent organizations. In Software Engineering for Large-Scale Multi-Agent Systems (pp. 27–38). Springer Berlin Heidelberg. doi:10.1007/3-540-35828-5_2. [5] Horling, B., & Lesser, V. (2004). A survey of multi-agent organizational paradigms. The Knowledge Engineering Review, 19(4), 281–316.doi:10.1017/ S0269888905000317. [6] Ferber, J., Gutknecht, O., & Michel, F. (2004). From agents to organizations: an organizational view of multi-agent systems. In Agent-Oriented Software Engineering IV (pp. 214–230). Springer Heidelberg. doi:10.1007/978-3-540-24620-6_15. Berlin [7] Van Den Broek, E. L., Jonker, C. M., Sharpanskykh, A., & Treur, J. (2006). Formal modeling and analysis of organizations. In Coordination, Organizations, Institutions, and Norms in Multi-Agent Systems (pp. 18-34). Springer Berlin Heidelberg. [8] Hübner, J. F., Vercouter, L., & Boissier, O. (2009). Instrumenting multi-agent organisations with artifacts to support reputation processes. In Coordination, Organizations, Institutions and Norms in Agent Systems IV (pp. 96–110). Springer Heidelberg. doi:10.1007/978-3-642-00443-8_7. Berlin [9] Ferber, J., Michel, F., & Báez, J. (2005). AGRE: Integrating for environments with organizations. multi-agent systems (pp. 48-56). Springer Berlin Heidelberg. In Environments [10] Dignum, V. (2009). The role of organization in agent systems. Handbook of Research on Multi-Agent Systems: Semantics and Dynamics of Organizational Models, 1-16. [11] Weyns, D., Haesevoets, R., & Helleboogh, A. (2010). The MACODO organization model for context-driven dynamic agent organizations. ACM Transactions on Autonomous and Adaptive Systems (TAAS), 5(4), 16. [12] Burns, T., & Stalker, G. (1961). The Management of Innovation, Tavistock, London. [13] Picard, G., Hübner, J. F., Boissier, O., & Gleizes, M. P. (2009, June). Reorganisation and self-organisation in multi-agent systems. In 1st International Workshop on Organizational Modeling, ORGMOD (pp. 66-80). [14] Upton, J., Janeka, I., & Ferraro, N. (2014). The whole is more than the sum of its parts: aristotle, metaphysical. Journal of Craniofacial Surgery, 25(1), 59-63. International Journal of Intelligent Information Systems 2015; 4(3): 46-57 56 [15] A. Drogoul, B. Corbara, and S. Lalande. MANTA: New experimental results on the emergence of (artificial) ant societies. In Nigel Gilbert and Rosaria Conte, editors, Artificial Societies: the Computer Simulation of Social Life, pages 119–221. UCL Press, London, 1995. [16] Rajkumar, R. R., Lee, I., Sha, L., & Stankovic, J. (2010, June). Cyber-physical systems: the next computing revolution. In Proceedings of the 47th Design Automation Conference (pp. 731-736). ACM. [17] Mattern, F., & Floerkemeier, C. (2010). From the Internet of Computers to the Internet of Things. In From active data management to event-based systems and more (pp. 242-259). Springer Berlin Heidelberg. [18] James Momoh (2012), Smart Grid: Fundamentals of Design and Analysis, Wiley-IEEE Press; 1 edition (March 20, 2012). [19] Abbas, H. A. (2014). Future SCADA challenges and the promising solution: the agent–based SCADA. International Journal of Critical Infrastructures, 10(3), 307-333. [20] Saha, D., & Mukherjee, A. (2003). Pervasive computing: a paradigm for the 21st century. Computer, 36(3), 25-31. [21] Friedewald, M., & Raabe, O. (2011). Ubiquitous computing: impacts. Telematics and technology An overview of Informatics, 28(2), 55-65. [22] Jennings NR (2001) An agent-based approach for building complex software systems. Communications of the ACM 44 (4):35–41 [23] Ferber J (1999) Multi-Agent Systems: An Introduction to Distributed Artificial Intelligence. Addison-Wesley, Harlow, England. [24] Wooldridge, M.: An Introduction to Multi-Agent Systems. Wiley, New York (2002). [25] Posey, Rollin B. (March 1961). "Modern Organization Theory edited by Mason Haire". Administrative Science Quarterly 5 (4): 609–611. [26] Hertz, D. and R. Livingston. (1950). Contemporary Organizational theory: A review of current concepts and methods. Human Relations, 3 (4), 373-394. [27] Odell, J. J., Parunak, H. V. D., & Fleischer, M. (2003). The role of roles in designing effective agent organizations. In Software Engineering for Large-Scale Multi-Agent Systems (pp. 27-38). Springer Berlin Heidelberg. [28] Van Den Broek, E. L., Jonker, C. M., Sharpanskykh, A., & Treur, J. (2006). Formal modeling and analysis of organizations. In Coordination, Organizations, Institutions, and Norms in Multi-Agent Systems (pp. 18-34). Springer Berlin Heidelberg. [29] Miles, R. E., Snow, C. S., Mathews, J. A., Miles, G., & Coleman, H. J. (1997). Organizing in the knowledge age: Anticipating the cellular form. The Academy of Management Executive, 11(4), 7-20. [30] Lesser, V. R., & Corkill, D. D. (1981). Functionally accurate, and systems. Systems, Man cooperative distributed Cybernetics, IEEE Transactions on,11(1), 81-96. [31] Corkill, D. D. (1980). An organizational approach to planning in distributed problem-solving systems. Technical Report 80-13, Department of Computer and Information Science, University of Massachusetts, Amherst, Massachusetts 01003. [32] Jay Galbraith. Designing Complex Organizations. Addison-Wesley, 1973. [33] Fox, M. S. (1981). An organizational view of distributed systems. Systems, Man and Cybernetics, IEEE Transactions on, 11(1), 70-80. [34] Giovanna Di Marzo Serugendo et al (2011), “Self-organizing Software, From Natural to Artificial Adaptation”, Springer. [35] M. Carley and L. Gasser. Computational organization theory. In G. Weiss, editor, Multiagent Systems: A Modern Approach to Distributed Arti_cial Intelligence, pages 299.330. MIT Press, 1999. [36] Corkill, D. D., & Lander, S. E. (1998). Diversity in agent organizations. Object Magazine, 8(4), 41-47. [37] Jay R. Galbraith. Organization Design. Addison-Wesley, 1977. [38] Giovanna di Marzoserugendo et al, “Self-organization in multi-agent systems”, The Knowledge Engineering Review, Vol. 20:2, 165–189., 2005, Cambridge University Press. [39] K. S. Barber and C. E. Martin, ‘Dynamic reorganization of the 5th in Proceedings of decisionmaking groups’, Autonomous Agents, (2001). [40] Dignum, V., Dignum, F., & Sonenberg, L. (2004, September). Towards dynamic reorganization of agent societies. In Proceedings of Workshop on Coordination in Emergent Agent Societies at ECAI (pp. 22-27). [41] Bernon, C., Camps, V., Gleizes, M. P., & Picard, G. (2005). Engineering adaptive multi-agent systems: The adelfe methodology. Agent-oriented methodologies, 172-202. [42] Guessoum, Z., Briot, J. P., Marin, O., Hamel, A., & Sens, P. (2003). Dynamic and adaptive replication for large-scale reliable multi-agent systems. InSoftware engineering for large-scale multi-agent systems (pp. 182-198). Springer Berlin Heidelberg. [43] Berns, A., & Ghosh, S. (2009, September). Dissecting self-* properties. In Self-Adaptive and Self-Organizing Systems, 2009. SASO'09. Third IEEE International Conference on (pp. 10-19). IEEE. [44] Sichman, J. S., Dignum, V., & Castelfranchi, C. (2005). Agents' organizations: a concise overview. Journal of the Brazilian Computer Society, 11(1), 3-8. [45] Serugendo, G. D. M., Gleizes, M. P., & Karageorgos, A. (2006). Self-Organisation and Emergence in MAS: An Overview. Informatica (Slovenia),30(1), 45-54. [46] P. Glansdorff and I. Prigogine. Thermodynamic study of Structure, Stability and Fluctuations. Wiley, 1971. [47] Mano, J. P., Bourjot, C., Lopardo, G., & Glize, P. (2006). Bio-inspired mechanisms for artificial self-organised systems. Informatica, 30(1), 55-62. [48] Giovanna (2011) Di Marzo Serugendo et al, “Self-organizing Software, From Natural to Artificial Adaptation”, Springer, 2011. [49] Zambonelli, F., Gleizes, M. P., Mamei, M., & Tolksdorf, R. (2004, May). Spray computers: frontiers of self-organization. In Autonomic Computing, 2004. Proceedings. International Conference on (pp. 268-269). IEEE. 57 Hosny Ahmed Abbas et al.: Organization of Multi-Agent Systems: An Overview [50] Karuna, H., Valckenaers, P., Saint-Germain, B., Verstraete, P., Zamfirescu, C. B., & Van Brussel, H. (2005). Emergent forecasting using a stigmergy approach in manufacturing coordination and control. In Engineering Self-Organising Systems (pp. 210-226). Springer Berlin Heidelberg. [51] Weyns, D., Schelfthout, K., Holvoet, T., & Glorieux, O. (2004). Role based model for adaptive Agents. In Fourth Symposium on Adaptive Agents and Multiagent Systems at the AISB'04 Convention. Garcia-Fornes, A., & Espinosa, A. (2007). Supporting agent organizations. In Multi-Agent Systems and Applications V (pp. 236-245). Springer Berlin Heidelberg. [65] Dignum, V., Dignum, F.: A Landscape of Agent Systems for the Real World. Tech. Report Utrecht University (2007) [66] Gutknecht, O., & Ferber, J. (1998). A model for social structures in multi-agent systems (Vol. 98040). Technical Report RR LIRMM. [52] M.P. Gleizes, V. Camps, and P. Glize. A theory of emergent computation based on cooperative self-organisation for adaptive artificial systems. Fourth European Congress of Systems Science. Valencia, 1999. [67] Hannoun, M., Boissier, O., Sichman, J. S., & Sayettat, C. (2000). MOISE: An organizational model for multi-agent systems. In Advances in Artificial Intelligence (pp. 156-165). Springer Berlin Heidelberg. [53] De Wolf, T., & Holvoet, T. (2004). Emergence and self-organisation: a statement of similarities and differences. Engineering Self-Organising Systems, 3464, 1-15. [54] Akgün, A. E., Keskin, H., & Byrne, J. C. (2014). Complex adaptive systems theory and firm product innovativeness. Journal of Engineering and Technology Management, 31, 21-42. [55] Streng, W. (2005). Reductionism versus Holism–Contrasting Approaches.Consilience. Interdisciplinary Communications, 2006, 11-14. [56] Ashby, W.R.: Principles of self-organizing dynamic systems. Journal of General Psychology 37 (1947) 125–128. [68] Hübner, J. F., Sichman, J. S., & Boissier, O. (2002, July). Moise+: towards a structural, functional, and deontic model for mas organization. In Proceedings of the first international joint conference on Autonomous agents and multiagent systems: part 1 (pp. 501-502). ACM. [69] Hübner, J. F., Sichman, J. S., & Boissier, O. (2004). Using the MOISE+ for a Cooperative Framework of MAS Reorganisation. In Advances in artificial intelligence–SBIA 2004 (pp. 506-515). Springer Berlin Heidelberg. [70] Dignum, V. (Ed.). (2009). Handbook of Research on Multi-Agent Systems: Semantics and Dynamics of Organizational Models: Semantics and Dynamics of Organizational Models. IGI Global. [57] Giorgini, P., & Henderson-Sellers, B. (2005). Agent-oriented methodologies: an introduction. Agent-oriented Methodologies, 1-19. [71] Jennings, N. R. (1999). Agent-oriented software engineering. In Multiple Approaches to Intelligent Systems (pp. 4-10). Springer Berlin Heidelberg. [58] Ferber, J. and Gutknecht, O., Aalaadin: a meta-model for the analysis and design of organizations in multi-agent systems. in Third International Conference on Multi-Agent Systems, (Paris, 1998), IEEE, 128-135. [72] Duan, Junhua, Yi-an Zhu, and Shujuan Huang. "Stigmergy agent and swarm-intelligence-based multi-agent system." Intelligent Control and Automation (WCICA), 2012 10th World Congress on. IEEE, 2012. [59] Wooldridge, M., Jennings, N. R., & Kinny, D. (2000). The Gaia methodology for agent-oriented analysis and design. Autonomous Agents and Multi-Agent Systems, 3(3), 285-312 . [73] Smith, R. G. (1980). The contract net protocol: High-level communication and control in a distributed problem solver. IEEE Transactions on computers, (12), 1104-1113. [60] Ferber, J., Michel, F., & Báez, J. (2005). AGRE: Integrating for environments with organizations. multi-agent systems (pp. 48-56). Springer Berlin Heidelberg. In Environments [61] Weyns, D., Haesevoets, R., & Helleboogh, A. (2010). The MACODO organization model for context-driven dynamic agent organizations. ACM Transactions on Autonomous and Adaptive Systems (TAAS), 5(4), 16. [62] Spivey, J. M. (1989). The Z notation (Vol. 1992). New York: Prentice Hall. [63] Coutinho, L. R., Sichman, J. S., & Boissier, O. (2005, October). Modeling organization in mas: A comparison of models. In First Workshop on Software Engineering for Agent-oriented Systems (pp. 1-10). [64] Argente, E., Palanca, J., Aranda, G., Julian, V., Botti, V., [74] S.A. DeLoach. Methodologies and Software Engineering for Agent Systems. The Agent-Oriented Software Engineering Handbook Series : Multiagent Systems, Artificial Societies,and Simulated Organizations, volume 11, chapter The MaSE Methodology. Kluwer Academic Publishing (available via Springer), 2004. [75] Juan Pav´on and Jorge J. G´omez-Sanz. Agent oriented software engineering with ingenias. In Vladim´ır Mar´ık, Jorg P. Muller, and Michal Pechoucek, editors, CEEMAS, volume 2691 of Lecture Notes in Computer Science, pages 394–403. Springer, 2003. [76] Abbas, H. A. (2014). Exploiting the Overlapping of Higher Order Entities within Multi-Agent Systems. International Journal of Agent Technologies and Systems (IJATS), 6(3), 32-57. doi:10.4018/ijats.2014070102.
1702.03614	1	1702	2017-02-13T02:50:55	Multitask diffusion adaptation over networks with common latent representations	[ "cs.MA", "stat.ML" ]	Online learning with streaming data in a distributed and collaborative manner can be useful in a wide range of applications. This topic has been receiving considerable attention in recent years with emphasis on both single-task and multitask scenarios. In single-task adaptation, agents cooperate to track an objective of common interest, while in multitask adaptation agents track multiple objectives simultaneously. Regularization is one useful technique to promote and exploit similarity among tasks in the latter scenario. This work examines an alternative way to model relations among tasks by assuming that they all share a common latent feature representation. As a result, a new multitask learning formulation is presented and algorithms are developed for its solution in a distributed online manner. We present a unified framework to analyze the mean-square-error performance of the adaptive strategies, and conduct simulations to illustrate the theoretical findings and potential applications.	cs.MA	cs	Multitask diffusion adaptation over networks 1 with common latent representations Jie Chen, Member, IEEE, Cédric Richard, Senior Member, IEEE, Ali H. Sayed, Fellow Member, IEEE 7 1 0 2 b e F 3 1 ] A M . s c [ 1 v 4 1 6 3 0 . 2 0 7 1 : v i X r a Abstract Online learning with streaming data in a distributed and collaborative manner can be useful in a wide range of applications. This topic has been receiving considerable attention in recent years with emphasis on both single-task and multitask scenarios. In single-task adaptation, agents cooperate to track an objective of common interest, while in multitask adaptation agents track multiple objectives simultaneously. Regularization is one useful technique to promote and exploit similarity among tasks in the latter scenario. This work examines an alternative way to model relations among tasks by assuming that they all share a common latent feature representation. As a result, a new multitask learning formulation is presented and algorithms are developed for its solution in a distributed online manner. We present a unified framework to analyze the mean-square-error performance of the adaptive strategies, and conduct simulations to illustrate the theoretical findings and potential applications. Multitask learning, distributed optimization, common latent subspace, online adaptation, diffusion strategy, collaborative processing, performance analysis. Index Terms I. INTRODUCTION Multi-agent networks usually consist of a large number of interconnected agents or nodes. Interconnections between the agents allow them to share information and collaborate in order to solve complex tasks collectively. Examples abound in the realm of social, economic and biological networks. Distributed algorithms over such networks offer a valuable alternative to centralized solutions with useful properties such as scalability, robustness, and decentralization. When endowed with adaptation abilities, these algorithms enable agents to continuously learn Copyright (c) 2016 IEEE. Personal use of this material is permitted. However, permission to use this material for any other purposes must be obtained from the IEEE by sending a request to [email protected]. The work of J. Chen was supported in part by the NSFC grant 61671382. The work of C. Richard was supported in part by the Agence Nationale pour la Recherche, France, (ODISSEE project, ANR-13-ASTR-0030). The work of A. H. Sayed was supported in part by NSF grants CIF-1524250 and ECCS-1407712. A short and preliminary version of this work appears in the conference publication [1]. J. Chen is with CIAIC of School of Marine Science and Technology, Northwestern Polytechinical University, Xi'an, 710072, China (e- mail: [email protected]). C. Richard is with the Université de Nice Sophia-Antipolis, CNRS, France (e-mail: [email protected]), in collaboration with Morpheme team (INRIA Sophia-Antipolis). A. H. Sayed is with the department of electrical engineering, University of California, Los Angeles, CA 9005-1594, USA (email: [email protected]). 2 and adapt in an online manner to concept drifts in their data streams [2], [3]. Broadly, distributed strategies for online parameter estimation can be applied to single-task or multi-task scenarios. In the first case, agents cooperate with each other to estimate a single parameter vector of interest, such as tracking a common target. Reaching consensus among the agents is critical for successful inference in these problems. In the multitask case, the agents cooperate to estimate multiple parameter vectors simultaneously, such as tracking a collection of targets moving in formation [4]. Extensive studies have been conducted on adaptive distributed strategies for single-task problems. Existing techniques include incremental [5]–[8], consensus [9]–[11], and diffusion strategies [1], [2], [12]–[17]. Incremental techniques require determining a cyclic path that runs across all nodes, which is generally a challenging (NP-hard) task to perform. Besides, feature makes the incremental strategies sensitive to link failures and problematic for adaptation. Consensus techniques aim to reach an agreement among nodes on the estimate of interest via local information exchanges, but they have been shown [2], [3] to suffer from instability problems when used in the context of adaptive networks due to an inherent asymmetry in the update equations. Diffusion techniques, on the other hand, have been shown to have superior stability and performance ranges [18] than consensus-based implementations. For these reasons, we shall focus on diffusion-type implementations in this paper. Besides single-task scenarios, there are also applications where it is desirable to estimate multiple parameter vectors at the same time, rather than promote consensus among all agents [19]. For example, geosensor networks that monitor dynamic spatial fields, such as temperature or windspeed variations in geographic environments, require node-specific estimation problems that are able to take advantage of the spatial correlation between the measurements of neighboring nodes [20], [21]. A second example is the problem of collaborative target tracking where agents track several objects simultaneously [4], [19]. Motivated by these applications, there have been several variations of distributed strategies to deal with multitask scenarios as well. Existing strategies mostly depend on how the tasks relate to each other and on exploiting some prior information. In a first scenario, nodes are grouped into clusters, and each cluster of nodes is interested in estimating its own parameter vector. Although clusters may generally have distinct though related estimation tasks to perform, the nodes may still be able to capitalize on inductive transfer between clusters to improve their estimation accuracy. Multitask diffusion strategies were developed to perform estimation under these conditions [4], [22]. One useful way to do so is to employ regularization. A couple of other useful works have also addressed variations of this scenario where the only available information is that clusters may exist in the network but nodes do not know which other nodes share the same estimation task [23]–[25]. In [26], the authors use multitask diffusion adaptation with a node clustering strategy to identify a model between the gait information and electroencephalographic signals. In [27], the authors consider the framework in [4] to devise a distributed strategy that allows each node in the network to locally adapt inter-cluster cooperation weights. The authors in [28] promote cooperation between clusters with (cid:96)1-norm co-regularizers. They derive a closed- form expression of the proximal operator, and introduce a strategy that also allows each node to automatically set its inter-cluster cooperation weights. The works in [29], [30] propose alternative node clustering strategies. In a 3 second scenario, it is assumed that there are parameters of global interest to all nodes in the network, a collection of parameters of common interest within sub-groups of nodes, and a set of parameters of local interest at each node. A diffusion strategy was developed to perform estimation under these conditions [31], [32]. Likewise, in the works [33]–[35], distributed algorithms are derived to estimate node-specific parameter vectors that lie in a common latent signal subspace. In another work [36], the diffusion LMS algorithm is extended to deal with structured criteria built upon groups of variables, leading to a flexible framework that can encode various structures in the parameters. An unsupervised strategy to differentially promote or inhibit collaboration between nodes depending on their group is also introduced. Alternatively, in recent years, there has been an increasing interest in modeling relations between tasks by assuming that all tasks share a common feature representation in a latent subspace [37]–[39]. The authors in [38] proposed a non-convex method based on Alternating Structure Optimization (ASO) for identifying the task structure. A convex relaxation of this approach was developed in [40]. In [39], the authors showed the equivalence between ASO, clustered multitask learning [41], [42] and their convex relaxations. The efficiency of such task relationships has been demonstrated in these works for clustering and classification problems. In our preliminary work [1], we introduced this framework within the context of distributed online adaptation over networks. Useful applications can be envisaged. First, consider the case where the common subspace is spanned by certain selected columns of the identity matrix. This means that a subset of the entries of the parameter vector to be estimated are common to all nodes while no further restriction is imposed on the other entries. Another example concerns beamforming for antenna arrays with a generalized side-lobe canceller (GSC). The latent subspace corresponds to the space where interfering signals reside [43]. A third example deals with cooperative spectrum sensing in cognitive radios, where the common latent subspace characterizes common interferers [31]. Drawing on these motivations, this paper deals with distributed learning and adaptation over multitask networks with common latent representation subspaces. Algorithms are designed accordingly, and their performance analyzed. The contributions of this work include the following main aspects: • We formulate a new multitask estimation problem, which assumes that all tasks share a common latent subspace representation in addition to node-specific contributions. Additional constraints can be incorporated if needed. This work contrasts with earlier works [4], [28], where the inductive transfer between learning tasks is promoted by regularizers. It also differs from [31], which considers direct models by stacking local and global variables in an augmented parameter vector. Moreover, the work [38] uses a similar inductive transfer model but the common latent subspace is unknown and embedded into a joint estimation process. Our work is the first one to introduce an online estimation algorithm over networks. Estimating the common latent subspace of interest within this context is a challenging perspective. • We explain how this formulation can be tailored to fit individual application contexts by considering additional model constraints. We illustrate this fact by considering two convex optimization problems and the associated distributed online algorithms. The first algorithm is a generalization in some sense of the diffusion LMS 4 algorithm, which can be retrieved by defining the low-dimensional common latent subspace as the whole parameter space. The second algorithm uses (cid:96)2-norm regularization to account for the multitask nature of the problem. This opens the way to other regularization schemes depending on the application at hand. • We present a unified framework for analyzing the performance of these algorithms. This framework also allows to address the performance analysis of the multitask algorithms in [4], [19], [44], [45] in a generic manner, though these analyses were performed independently of each other in these works. The rest of the paper is organized as follows. Section II introduces the multitask estimation problem considered in this paper. Then, two distributed learning strategies are derived in Section III by imposing different constraints on common and node-specific representation subspaces. Section IV provides a general framework for analyzing distributed algorithms of this form. In Section V, experiments are conducted to illustrate the characteristics of these algorithms. Section VI concludes the paper and connects our work with several other learning strategies. Notation. Normal font x denotes scalars. Boldface small letters x denote vectors. All vectors are column vectors. Boldface capital letters X denote matrices. The asterisk (·)∗ denotes complex conjugation for scalars and complex- conjugate transposition for matrices. The superscript (·)(cid:62) represents transpose of a matrix or a vector, and (cid:107)·(cid:107) is the (cid:96)2-norm of its matrix or vector argument. Re{·} and Im{·} denote the real and imaginary parts of their complex argument, respectively. Matrix trace is denoted by trace(·). The operator col{·} stacks its vector arguments on the top of each other to generate a connected vector. The operator diag{·} formulates a (block) diagonal matrix with its arguments. Identity matrix of size N × N is denoted by I N . Kronecker product is denoted by ⊗, and expectation is denoted by E{·}. We denote by Nk the set of node indices in the neighborhood of node k, including k itself, and Nk its set cardinality. II. MATCHED SUBSPACE ESTIMATION OVER MULTITASK NETWORKS A. Multitask estimation problems over networks Consider a connected network composed of N nodes. The problem is to estimate an L × 1 unknown vector wo k at each node k from collected measurements. At each time n, node k has access to local streaming measurements {dk(n), xk,n}, where dk(n) is a scalar zero-mean reference signal, and xk,n is a 1 × L zero-mean row regression k,nxk,n} > 0. The data at agent k and time n are assumed to be related vector with covariance matrix Rx,k = E{x∗ via the linear model: dk(n) = xk,nwo k + zk(n) (1) k is an unknown complex parameter vector, and zk(n) is a zero-mean i.i.d. noise with variance σ2 where wo z,k = E{zk(n)2}. The noise signal zk(n) is assumed to be independent of any other signal. Let Jk(w) be a differentiable convex cost function at agent k. In this paper, we shall consider the mean-square-error criterion: Jk(w) = E{dk(n) − xk,nw2} (2) It is clear from (1) that each Jk(w) is minimized at wo a more general sense) as a task. Depending on whether the minima of all Jk(w) are achieved at the same wo not, the distributed learning problem can be single-task or multitask oriented [4]. k to estimate (or model in k or k. We refer to each parameter wo With single-task networks, all agents aim at estimating the same parameter vector wo shared by the entire network, that is, 5 (3) for all k ∈ {1, ..., N}. Several popular collaborative strategies, such as diffusion LMS [1], [2], [13], [14], were derived to address this problem by seeking the minimizer of the following aggregate cost function: k = wo wo N(cid:88) J glob(w) = Jk(w) (4) in a distributed manner. Since the individual costs (2) admit the same solution, wo is also the solution of (4). It has been shown that using proper cooperative strategies to solve (4) can improve the estimation performance [2], k=1 [3]. With multitask networks, each agent aims at determining a local parameter vector wo k. It is assumed that some similarities or relations exist among the parameter vectors of neighboring agents so that cooperation can still be meaningful, namely, (5) where the symbol ∼ refers to a similarity relationship in some sense, which can be exploited to enhance performance. Depending on the problem characteristics, this property can be promoted in several ways, e.g., by introducing some k ∼ wo wo (cid:96) if (cid:96) ∈ Nk regularization term, or by assuming a common latent structure. Networks may also be structured into clusters where agents within each cluster estimate the same parameter vector [4], [44]. B. Node-specific subspace constraints Although agents aim to estimate distinct minimizers wo k, exploiting relationships between solutions can make cooperation among agents beneficial. Regularization is one popular technique for introducing prior information about the solution. It can improve estimation accuracy though it may introduce bias [4], [19], [46]. In this paper, we explore an alternative strategy that assumes that the hypothesis spaces partially overlap. Specifically, we assume that each wo k can be expressed in the form: wo k = Θuo + o k (6) where Θuo is common to all nodes with Θ denoting an L × M matrix with known entries and uo an unknown M × 1 parameter vector (common to all nodes), and where o k is an unknown node-specific component. We assume that matrix Θ = [θ1, . . . , θM ] is full-rank with M ≤ L. Overcomplete sets of column vectors {θ1, . . . , θM} may be advantageous in some scenarios but this usually requires to impose further constraints such as sparsity over uo. We shall not discuss this case further in order to focus on the main points of the presentation. Model (6) means 6 that all tasks share the same parameter vector Θuo, which lies in the subspace spanned by columns of Θ. This subspace representation can be useful in several applications. For instance, consider the case where Θ is composed of selected columns of the identity matrix I L. This means that a subset of the entries of wo k are common to all agents while no further assumptions are imposed on the other entries. This situation is a natural generalization of the single-task scenario. Another example concerns beamforming problems with a generalized sidelobe canceller (GSC), where Θ acts as a blocking matrix to cancel signal components that lie in the constraint space [43]. In machine learning, formulation (6) is referred to as the alternating structure optimization (ASO) problem [38], [39]. The subspace Θ is, however, learnt simultaneously via a non-convex optimization procedure. In what follows, we shall assume that Θ is known by each agent. Before proceeding further, we clarify the difference between model (6) addressed here and in our preliminary work [1], and the model studied in [31], [32], [35], [47], [48]. In these last works, the authors consider particular information access models where global and local components are assumed to be related to distinct regressors. The centralized problem can then be formulated by stacking the global and local regressors, and by considering a parameter vector augmented accordingly. In our work, motivated by applications of the latent space model in batch-mode learning, we address the problem where the parameter vectors to be estimated lie in global and local latent subspaces. We do not need to distinguish explicitly between global and local regressors. Instead, as shown in the sequel, some extra conditions are needed so that model (6) is identifiable. Among other possibilities, we shall investigate two strategies where constraints on Θ and o k are imposed. Replacing (6) into (2), the global cost function is expressed as a function of a common parameter u and node- specific perturbations {k}N k=1: J glob(cid:0)u,{k}N k=1 (cid:1) = N(cid:88) k=1 E(cid:8)dk(n) − xk,n(Θu + k)2(cid:9) (7) k by each agent to benefit from the cooperative estimation of u. Problem (7) is still We expect the estimation of wo insufficient for estimating the tasks {wo k}. This is because the decomposition wk = Θ u + k is not unique. Indeed, given any optimum solution {¯u, ¯k}, and any s = Θx, we can generate another optimum solution by considering the shift {¯u − x, ¯k + s}. This ambiguity prevents us from deriving collaboration strategies based on u. From the point of view of convex analysis, the Hessian matrix of (7) is rank deficient and no unique solution exists. III. PROBLEM FORMULATIONS AND SOLUTION ALGORITHMS Problem (7) can be modified to make it well-determined and more meaningful. In this section, among other possibilities, we investigate two strategies that consist of imposing further constraints and derive the corresponding distributed algorithms. These two formulations guarantee the uniqueness of the solution and have clear interpreta- tions. A. Node-specific subspace constraints 7 We restrict the node-specific components {k}N k=1 to lie in the complementary subspace to span(Θ). The problem can be formulated as: min u,{k}N k=1 J glob(cid:0)u,{k}N k=1 (cid:1) subject to k ∈ span(Θ⊥), ∀k = 1, . . . , N (8) where the L − M columns of matrix Θ⊥ span the complementary subspace to span(Θ), that is, Θ∗ Θ⊥ = 0. We write: (9) where ξk is a column vector of size (L − M ). Now, replacing (9) into (8), the optimization problem becomes unconstrained and the objective function is given by: k = Θ⊥ ξk (cid:1) J glob(cid:0)u,{ξk}N N(cid:88) N(cid:88) k=1 = = k=1 k=1 E(cid:8)dk(n) − xk,n(Θu + Θ⊥ξk)2(cid:9) (cid:17) E{dk(n)2} + u Rx,k ∗Θ∗(cid:16) N(cid:88) (cid:110) N(cid:88) (cid:111) − 2 Re k=1 (cid:110) N(cid:88) Θu + N(cid:88) (cid:111) k=1 ∗ kΘ∗ ⊥Rx,kΘ⊥ξk + 2 Re ξ u (cid:110) ∗Θ∗ N(cid:88) (cid:111) Rx,kΘ⊥ξk k=1 − 2 Re ∗ dx,kΘu p where Rx,k = E{x∗ k,nxk,n} is the covariance matrix of xk,n, and pdx,k = E{dk(n)x∗ between the input data xk,n and the reference output data dk(n). ∗ dx,kΘ⊥ξk p k=1 k=1 (10) k,n} is the covariance vector Lemma 1: Problem (8) has a unique solution with respect to u and {k}N k=1 if the perturbations {k}N k=1 lie in a subspace orthogonal to span(Θ). Proof of Lemma 1 is provided in Appendix A. We shall now derive a distributed algorithm to seek the minimizer k, we of (8). Focusing on the terms that depend on u in (10), and setting parameters ξk to their optimum values ξo consider first the global cost function over the variable u: ∗Θ∗ u Rx,kΘu + 2 Re ∗Θ∗ u Rx,kΘ⊥ξo k ∗ dx,kΘu p + gk(ξo k) (cid:111) − 2 Re (cid:110) (cid:111) (cid:17) (cid:110) (cid:16) N(cid:88) N(cid:88) k=1 k=1 Ju,k(u) J glob u (u) = = k in (10). The term(cid:80)N (11) E{dk(n)2} is discarded because where gk(ξo k) collects all the terms depending only on ξo it is constant with respect to the arguments u and {ξk}N (u) has a unique minimizer for all nodes over the network, we can use a single-task adapt-then-combine (ATC) diffusion strategy to estimate uo [13], [15]. We introduce a right-stochastic matrix C with nonnegative entries c(cid:96)k such that: glob k=1. Since J u k=1 c(cid:96)k = 1, and c(cid:96)k = 0 if k /∈ N(cid:96) (12) N(cid:88) k=1 8 With each node k, we associate the local cost over the variable u: Observe that (cid:80)N k=1 J loc glob u,k(u) = J u respect to u, it can be expressed at each node k as follows: glob (u) because matrix C is right-stochastic. Since J u J loc u,k(u) = c(cid:96)kJu,(cid:96)(u) (13) (cid:88) (cid:96)∈Nk (cid:88) (cid:88) (cid:96)(cid:54)=k (cid:96)(cid:54)=k J glob u (u) = J loc u,k(u) + = J loc u,k(u) + J loc u,(cid:96)(u) (cid:107)u − uo(cid:107)2∇2J loc u,(cid:96) (u) is quadratic with (14) u,(cid:96) denotes the Hessian matrix of J loc where ∇2J loc by any positive semi-definite matrix Σ, i.e., (cid:107)u(cid:107)2 characterization of eigenvalues [13, Sec. 3.1], we approximate ∇2J loc (cid:107)u − uo(cid:107)2∇2J loc Minimizing (14) in two successive steps yields: ≈ b(cid:96)k(cid:107)u − uo(cid:107)2. u,(cid:96) u,(cid:96)(u) with respect to u, and (cid:107)u(cid:107)2 Σ is the squared norm of u weighted Σ = u∗Σu. Following an argument based on the Rayleigh-Ritz u,(cid:96) by a multiple of the identity matrix, so that (cid:88) φk,n = uk,n−1 − µ∇J loc uk,n = φk,n + µ u,k(uk,n−1) b(cid:96)k(uo − uk,n−1) (cid:96)(cid:54)=k (15) (16) where µ is a positive step size. Its choice to ensure stability of the algorithm will be elaborated on later in Sec. IV. Now, note the following. First, iteration (16) requires knowledge of uo, which is not available. Each node (cid:96) has a readily available, however, an approximation for uo, which is φk,n. Therefore, we replace uo by φk,n in (16). Second, φk,n at node k is generally a better estimate for uo than uk,n−1 since it is obtained by incorporating information from the neighbors through (15). Therefore, we replace uk,n−1 by φk,n in (16). Then, absorbing coefficients b(cid:96)k into another set of nonnegative coefficients that satisfies: a(cid:96)k = 1, and a(cid:96)k = 0 if (cid:96) /∈ Nk, (17) (cid:96)=1 which means that matrix A with entries a(cid:96)k is left-stochastic, using an instantaneous approximation of the gradient, and limiting the summation in (16) to the neighbors of node (cid:96) (see [13], [15] for more details on a similar derivation in the context of single-task diffusion strategies), we can update uk,n as follows: (cid:2)d(cid:96)(n) − x(cid:96),n(Θuk,n−1) − x(cid:96),n(Θ⊥ξ(cid:96),n−1)(cid:3) c(cid:96)kΘ∗ ∗ (cid:96),n x (cid:88) (cid:96)∈Nk φk,n = uk,n−1 + µ uk,n = a(cid:96)kφ(cid:96),n (cid:88) (cid:96)∈Nk (18) (19) where ξk,n−1 is an estimate for the unknown minimizer ξo k, to be evaluated as explained further ahead in (21). N(cid:88) Focusing on the terms that depend on {ξk}N k=1 in (10), and setting parameter u to its optimum value uo, we consider the global cost function over the variables ξk: (cid:110) ξ ⊥Rx,kΘuo(cid:111) − 2 Re (cid:110) kΘ∗ ∗ ∗ dx,kΘ⊥ξk p (cid:111)(cid:17) (cid:48) k(uo) + g (20) ⊥Rx,kΘ⊥ξk + 2 Re (cid:1) (cid:0){ξk}N (cid:16) N(cid:88) N(cid:88) k=1 ξ k=1 kΘ∗ ∗ Jξ,k(ξk) k=1 glob J ξ = = 9 (21) (25) (26) k(ξo where g(cid:48) k) collects all the terms depending only on uo in (10). Now since the parameters ξk are node-specific, if no further constraints are imposed, they can be updated independently of each other via an LMS-type update: ξk,n = ξk,n−1 + µΘ∗ ∗ ⊥x k,n (cid:2)dk(n) − xk,n(Θuk,n−1+ Θ⊥ξk,n−1)(cid:3) At each time instant n, node k updates its parameters uk,n−1 and ξk,n−1 using (18)–(19) and (21), respectively. The local estimate wk,n is then given by: wk,n = Θuk,n + Θ⊥ξk,n (22) It is interesting to note that we can rewrite the algorithm without using the auxiliary variables uk,n and {ξk,n}N by substituting the relations: k=1, uk,n = (Θ∗Θ)−1Θ∗ ⊥Θ⊥)−1Θ∗ ξk,n = (Θ∗ wk,n ⊥wk,n (23) (24) into (18)–(19) and (21), respectively. Selecting C = I N to avoid exchanging raw data and node-specific components, we can implement the update of wk,n−1 to an intermediate value ψk,n as follows: ψk,n (a) = Θφk,n + Θ⊥ξk,n (b) = Θuk,n−1 + Θ⊥ξk,n−1 + µ (cid:104) = wk,n−1 + µSΘx ∗ k,n (ΘΘ∗ + Θ⊥Θ∗ (cid:2)dk(n) − xk,nwk,n−1 (cid:3) (cid:105) k,n[dk(n) − xk,n(Θuk,n−1 + Θ⊥ξk,n−1)] ∗ ⊥)x with SΘ = ΘΘ∗ + Θ⊥Θ∗ ⊥. For step (a), we use (22) with the intermediate value φk,n of uk,n in (18) and ξk,n. Step (b) follows from their adaptation steps (18) and (21). Now substituting (19) in (22) to aggregate the intermediate estimates of uk,n from the neighbors of node k, we arrive at the combination step: wk,n (22) = Θuk,n + Θ⊥ξk,n (c) = Θ a(cid:96)k(Θ∗Θ)−1Θ∗ ψk,n+Θ⊥(Θ∗ ⊥Θ⊥)−1 Θ∗ ⊥ψk,n = a(cid:96)kP Θψ(cid:96),n + P Θ⊥ψk,n (cid:88) (cid:88) (cid:96)∈Nk (cid:96)∈Nk 10 where P Θ = Θ(Θ∗Θ)−1Θ∗ and P Θ⊥ = I L − P Θ are the projection matrices over subspaces Θ and Θ⊥. For step (c), we use (23)–(24) with the intermediate estimate ψk,n. Finally, we arrive at the ATC strategy summarized in Algorithm 1. The first step in (28) is an adaptation step where node k uses the data realizations {dk(n), xk,n} to update its existing estimate wk,n−1 to an intermediate value ψk,n. All other nodes in the network are performing a similar step. The second step in (29) is an aggregation step. To update its intermediate estimate to wk,n, each node k combines the existing estimates of its neighbors in the common latent subspace Θ to build up a common representation, and refines it with a node-specific value in Θ⊥. In the special case when A = I N , so that no information exchange is performed, the ATC strategy reduces to a non-cooperative solution where each node k runs its own individual descent algorithm. Matrix SΘ in the adaptation step (28) is positive-definite. It arises from the calculation of the gradient of (10) with respect to u and ξk. The algorithm can be simplified by replacing SΘ by I L in (28) without compromising the convergence of the method (as analyzed further ahead in Section IV). We then arrive at the recursion: ψk,n = wk,n−1 + µ x ∗ k,n (cid:2)dk(n) − xk,nwk,n−1 (cid:3) (27) Strictly speaking, observe that SΘ = I L if, and only if, the columns of Θ and Θ⊥ form an orthonormal basis of IRL. Note that the adaptation step (27) is the LMS solution for minimizing the cost in (10) with respect to wk. Before leaving this section, we would like to point out that the algorithm described in [31], which addresses direct models by stacking global and local variables in an augmented parameter vector, may be used to solve problem (8), provided that an appropriate variable change is performed in order to make the latent variables uk,n and ξk,n explicit in wk,n. The resulting algorithm has the same performance as Algorithm 1 defined by (28), (29), but, obviously, they do not have the same form since they do not operate in the same domain. This structural difference has a major consequence for Algorithm 1. As already explained, it can be further tuned by replacing the matrix SΘ in (28) by any positive definite matrix while ensuring convergence of the method. This extra degree of freedom will be taken into account in the analysis of the algorithm, where the only condition on SΘ is to be positive definite. We will also show that setting SΘ to I L, besides simplifying Algorithm 1, can greatly improve its performance. B. Node-specific subspace constraints with norm-bounded projections The second formulation we consider is to relax the constraint that node-specific components {k}N k=1 must lie in span(Θ⊥). We now assume that they are norm-bounded in some sense. The problem is formulated as follows: (cid:1) J glob(cid:0)u,{k}N N(cid:88) k=1 min u,{k}N k=1 subject to (cid:107)P Θk(cid:107)2 ≤ ν1, k=1 N(cid:88) k=1 (cid:107)P Θ⊥k(cid:107)2 ≤ ν2 (30) Algorithm 1: ATC diffusion LMS with node-specific subspace constraints Parameters: Preset - positive step-size µ for all agents; - left-stochastic combination matrix A; - full-rank matrix Θ with columns {θ1, . . . , θM}. Initialization: Set initial weights wk,0 = 0 for all k ∈ {1, ..., N}. Algorithm: At each time instant n ≥ 1, and for each agent k, update wk,n as: (cid:2)dk(n) − xk,nwk,n−1 (cid:3) ∗ ψk,n = wk,n−1 +µSΘ x k,n wk,n = a(cid:96)kP Θψ(cid:96),n + P Θ⊥ψk,n (cid:88) (cid:96)∈Nk 11 (28) (29) Since the objective function and the constraints are convex in (u,{k}N formulated as a regularized optimization problem that consists of minimizing a global cost of the form [49]: k=1), the constrained problem (30) can be J glob(cid:0)u,{k}N k=1 N(cid:88) (cid:1) = E(cid:8)dk(n) − xk,n(Θu + k)2(cid:9) + η1 N(cid:88) (cid:107)P Θk(cid:107)2 + η2 (cid:107)P Θ⊥k(cid:107)2 (31) N(cid:88) k=1 k=1 k=1 where η1 and η2 are positive regularization parameters that are related to the bounds ν1 and ν2. Lemma 2: Problem (30) has a unique solution with respect to u and {k}N Proof of Lemma 2 is provided in Appendix B. Other norms such as the general (cid:96)p,q-norm may be used with k in (30), depending on the application. Some form of regularization on u may also be included. However, using the (cid:96)2-norm with k in (30) enables us to solve the problem with respect to wk, without using the auxiliary variables u and {k}N k=1. Indeed, let us rewrite (31) as follows: k=1. J glob(cid:0)u,{wk}N k=1 (cid:1) = E(cid:8)dk(n) − xk,nwk2(cid:9) + η1 N(cid:88) (cid:107)P Θ(wk − Θu)(cid:107)2 + η2 N(cid:88) The optimality condition relative to u gives: k=1 k=1 Θ∗ P Θ(wo k − Θuo) = 0 N(cid:88) k=1 Node k can apply a steepest-descent iteration to minimize the cost in (35) with respect to {wk}N the gradient vector of (35) we get: ∇J glob = (Rx,kwk − pdx) + η1 (Θ∗Θ)−1Θ∗ wo k k=1 (cid:13)(cid:13)(cid:13)P Θwk − 1 N(cid:88) N (cid:0)P Θwk − 1 N(cid:88) (cid:96)=1 P Θw(cid:96) + η2 (cid:13)(cid:13)(cid:13)2 (cid:1) + η2P Θ⊥wk N(cid:88) k=1 (cid:105)∗ P Θw(cid:96) N (cid:96)=1 N(cid:88) k=1 (cid:104) from which the optimal parameter vector uo can be expressed as: N(cid:88) N(cid:88) E(cid:8)dk(n) − xk,nwk2(cid:9) + η1 uo = 1 N k=1 J glob(cid:0){wk}N k=1 (cid:1) = N(cid:88) k=1 Substituting (34) into (32), and using that P Θ is Hermitian and idempotent (i.e., P Θ = P 2 Θ), yields: (cid:107)P Θ⊥wk(cid:107)2 (32) (33) (34) (cid:107)P Θ⊥wk(cid:107)2 (35) k=1. Computing (36) 12 Starting from an initial condition wk,0, we arrive at the steepest descent iteration: (cid:104) (cid:16) wk,n = wk,n−1 − µ (Rx,kwk,n−1 − pdx) + η2P Θ⊥wk,n−1 − µη1 P Θwk,n−1 − 1 N P Θw(cid:96),n−1 N(cid:88) (cid:96)=1 (cid:17) (cid:105) (37) This iteration indicates that the update term involves adding two correction terms to wk,n−1. Among many other forms, we can implement the update in two successive steps by adding one correction term at a time: (cid:104) (Rx,kwk,n−1 − pdx) + η2P Θ⊥wk,n−1 ψk,n = wk,n−1 − µ (cid:16) (cid:17) P Θwk,n−1 − 1 wk,n = ψk,n − µη1 N P Θw(cid:96),n−1 N(cid:88) (cid:105) (cid:96)=1 Step (38) updates wk,n−1 to an intermediate value ψk,n. We now revise (38)–(39) to achieve a diffusion LMS type algorithm. The intermediate value ψ(cid:96),n at node (cid:96) is generally expected to be a better estimate for wo (cid:96) than w(cid:96),n−1 since it is updated by the first step (38). Therefore, we replace w(cid:96),n−1 by ψ(cid:96),n in the second step (39) as follows to get: wk,n = ψk,n − µη1 = (ψk,n−P Θψk,n)+ (cid:16) P Θψk,n − 1 N N(cid:88) (cid:17) (cid:16) N(cid:88) (1−µη1)P Θψk,n+ P Θψ(cid:96),n (cid:96)=1 (cid:96)=1 (cid:17) ψ(cid:96),n µη1 N (40) N for (cid:96) (cid:54)= k, and akk = Observe that P Θ⊥ψk,n = ψk,n − P Θ ψk,n, and introduce the coefficients a(cid:96)k = µη1 1 − µη1 + µη1 N . We get: wk,n = P Θ⊥ψk,n + a(cid:96)kP Θψ(cid:96),n (41) Considering that each node in the network can only share information with its neighbors, and using instantaneous approximations for Rx,k and pdx, we arrive at: ψk,n = (I L − µη2P Θ⊥)wk,n−1 + µx wk,n = a(cid:96)kP Θψ(cid:96),n + P Θ⊥ψk,n (cid:88) ∗ k,n (cid:96)∈Nk (cid:2)dk(n) − xk,nwk,n−1 (cid:3) with akk = 1 − µη1 + µη1 µk, these coefficients are nonnegative and satisfy (cid:80)N Nk for (cid:96) ∈ Nk and (cid:96) (cid:54)= k. Note that, for sufficiently small step-sizes (cid:96)=1 a(cid:96)k = 1 for all k. We will treat these coefficients as free parameters that can be chosen by the designer according to these conditions (i.e., nonnegative coefficients that add Nk and a(cid:96)k = µη1 up to one on each column of matrix A). We summarize this statement in Algorithm 2. Algorithms 1 and 2 employ the same aggregation step in (29) and (45). Node k combines the intermediate estimates of its neighbors in the common subspace Θ without affecting the local contribution in the complementary subspace Θ⊥. The norm constraint (30) in Θ⊥ leads to a leaky-LMS alike term in the adaptation step (44). (38) (39) (42) (43) N(cid:88) (cid:96)=1 Algorithm 2: ATC diffusion LMS with node-specific subspace constraints (norm-bounded projections) Parameters: Preset - positive step-size µ for all agents; - full-rank matrix Θ with columns {θ1, . . . , θM}. Initialization: Set initial weights wk,0 = 0 for all k = 1, ..., N. Algorithm: For each instant n ≥ 1, and for each agent k, update wk,n−1: ψk,n = (I L − µη2P Θ⊥)wk,n−1 + µx a(cid:96)kP Θψ(cid:96),n + P Θ⊥ψk,n wk,n = (cid:88) (cid:96)∈Nk ∗ k,n (cid:2)dk(n) − xk,nwk,n−1 (cid:3) 13 (44) (45) Let us now examine two special cases of Algorithm 2. First, in the case where Θ = 0, problem (31) reduces to a regularized least-mean squares problem with wk = k. That is, the algorithm reduces to the non-cooperative leaky-LMS algorithm. On the other hand, if Θ = I L, the algorithm reduces to diffusion LMS. Before leaving this section, we briefly discuss the complexity of Algorithms 1 and 2. Both algorithms have the same adapt-then-combine structure as the diffusion LMS except that each node needs to project data on Θ and Θ⊥. This means that each node k only needs to update the L× 1 parameter vectors ψk,n and wk,n at each time instant. Next, each node k needs to transmit wk,n to its Nk − 1 neighbors. A projection performed by a matrix-vector product has a computational complexity of O(L log2 L) [50]. All the other operations performed by each node have a complexity of O(L). IV. PERFORMANCE AND CONVERGENCE ANALYSES In this section, we examine the convergence properties and network performance of the proposed adaptive strategies. We shall first describe a convergence framework for a family of distributed algorithms, where Algorithms 1 and 2 are special cases. Quantities specifically related to Algorithms 1 or 2 will be distinguished by superscripts (1) and (2), respectively. In order to perform the analysis, we collect information from across the network into block vectors and matrices. Let us denote by wn and wo the block weight vector at instant n and the block optimum weight vector, both of size LN × 1, that is wn = col{w1,n, . . . , wN,n} wo = col{wo N} 1, . . . , wo We denote the difference between the optimum wo k and the instantaneous estimate wk,n by: vk,n = wo k − wk,n (46) (47) (48) 14 We collect the weight error vectors vk,n from across all nodes into the block weight error vector: vn = col{v1,n, . . . , vN,n} (49) Assumption 1: (Independent inputs) The regression vectors xk,n arise from a stationary random process that is k,nxk,n} > 0. A direct consequence of temporally stationary, white, and independent over space with Rx,k = E{x∗ this condition is that xk,n is independent of v(cid:96),m for all (cid:96) and m ≤ n. A. Mean weight behavior analysis The estimation error in (28) and (44) can be rewritten as a function of vk,n: dk(n) − xk,nwk,n−1 = zk(n) + xk,nvk,n−1 (50) In what follows, we first show that the weight error update relations for both Algorithms 1 and 2 are of the form: vn = Bnvn−1 − µ gn − r, (51) with Bn an LN × LN time-dependent matrix, gn an LN × 1 zero-mean time-dependent vector, and r a constant LN × 1 vector. Consequently, it will be possible to represent their mean weight behavior in the form of a state- transition equation with a bounded driving term: (52) with B = E{Bn−1}. Let H x,n be the block diagonal matrix of size LN × LN, and pzx,n the vector of length LN × 1, defined as follows: E{vn} = B E{vn−1} − r H x,n (cid:44) diag{x N,nxN,n} ∗ ∗ 1,nx1,n, . . . , x N,n} (cid:44) col{z1(n)x ∗ ∗ 1,n, . . . , zN (n)x pzx,n The expectation of H x,n and pxz,n are given by: H x (cid:44) E{H x,n} = diag{Rx,1, . . . , RN,2} 1) Mean weight behavior of Algorithm 1: Define the intermediate weight error vector (cid:101)ψk,n: pzx (cid:44) E{pzx,n} = 0 (cid:101)ψk,n = wo k − ψk,n (cid:101)ψn = col{(cid:101)ψ1,n, . . . ,(cid:101)ψN,n} (53) (54) (55) (56) (57) (58) (59) and collect these vectors from across all nodes into the block weight error vector: for (cid:101)ψn: Subtracting wo k from both sides of the update relation (28), and using relation (50), leads to the update equation (cid:101)ψn = (I LN − µDSΘH x,n)vn−1 − µDSΘpzx,n where DSΘ = diag{SΘ, . . . , SΘ} is an LN × LN block diagonal matrix with SΘ as diagonal entries. Let A = A⊗ I L. Defining DP Θ and DP Θ⊥ as the LN × LN block diagonal matrices with P Θ and P Θ⊥ as diagonal entries, respectively, equation (29) can be written in vector form as: Combining this equation with (59), the weight error update relation can be written in a single expression: Subtracting wo from both sides of the above expression, we have: vn =(cid:0)A(cid:62) vn =(cid:0)A(cid:62) DP Θ + DP Θ⊥ Now we denote several terms in the weight error expression (62) by: wn =(cid:0)A(cid:62) (cid:1)ψn DP Θ + DP Θ⊥ (cid:1)(cid:101)ψn −(cid:0)A(cid:62) DP Θ + DP Θ⊥ DP Θ + DP Θ⊥ − I LN (cid:1)(cid:2)(I LN − µDSΘH x,n)vn−1 − µDSΘpzx,n (cid:1)(I LN − µDSΘH x,n) n =(cid:0)A(cid:62) (cid:1)DSΘpzx,n n =(cid:0)A(cid:62) DP Θ + DP Θ⊥ DP Θ + DP Θ⊥ g(1) r(1) = (A(cid:62) − I LN )DP Θwo, B(1) (cid:1)wo (cid:3) − (A(cid:62) − I LN )DP Θwo 15 (60) (61) (62) (63) (64) (65) (66) (67) (68) (70) (71) (72) and the associated expected values: =(cid:0)A(cid:62) B(1) (cid:44) E{B(1) n } DP Θ + DP Θ⊥ n } = 0 g(1) (cid:44) E{g(1) (cid:1)(I LN − µDSΘH x) With the above notation, the weight error update relation (62) can be written as: vn = B(1) n vn−1 − µ g(1) n − r(1) Taking the expectation on both sides of (68), and using Assumption 1, we arrive at the mean weight behavior for Algorithm 1: E{vn} = B(1) E{vn−1} − r(1) (69) 2) Mean weight behavior of Algorithm 2: Subtracting wo k from both sides of the update relation (44), and using relation (50), yields: Subtracting wo from both sides of (45), we have: (cid:101)ψn = (I − µη2DP Θ⊥ − µH x,n)vn−1 − µ(pzx,n − η2DP Θ⊥ wo) (cid:1)(cid:101)ψn −(cid:0)A(cid:62) (cid:1)wo vn =(cid:0)A(cid:62) (cid:1)(cid:2)(I LN −µη2DP Θ⊥−µH x,n)vn−1 vn =(cid:0)A(cid:62) − µ(pzx,n − η2DP Θ⊥ wo)(cid:3) − (A(cid:62) − I LN )DP Θ wo DP Θ + DP Θ⊥ − I LN DP Θ + DP Θ⊥ DP Θ +DP Θ⊥ Combining this equation with (70), the weight error update relation can be written in a single expression: 16 where we used the fact that I LN = DP Θ +DP Θ⊥ . Next, we denote several terms in the weight error expression (72) by: n =(cid:0)A(cid:62) n =(cid:0)A(cid:62) B(2) DP Θ+DP Θ⊥ DP Θ + DP Θ⊥ g(2) r(2) = (A(cid:62) − I LN)DP Θwo − µη2 (cid:0)A(cid:62) (cid:1)(I LN −µη2DP Θ⊥−µH x,n) (cid:1)pzx,n (cid:1)DP Θ⊥ wo DP Θ + DP Θ⊥ (73) (74) (75) (76) (77) (78) and the associated expected values: =(cid:0)A(cid:62) B(2) (cid:44) E{B(2) n } DP Θ +DP Θ⊥ n } = 0 g(2) (cid:44) E{g(2) (cid:1)(I LN −µη2DP Θ⊥ −µH x) With the above notation, the weight error update relation (72) can be written as: vn = B(2) n vn−1 − µ g(2) n − r(2) Taking the expectation on both sides of (78), and using Assumption 1, we get the mean weight behavior of Algorithm 2: E{vn} = B(2) E{vn−1} − r(2) (79) 3) Stability in the mean: The mean-weight error recursions (69) and (79) are of the same form as (52). The convergence of such recursive state-transition equations, with a bounded driving term, is determined by the stability of matrix B. Algorithm parameters should be chosen to satisfy the mean stability condition ρ(B) < 1, where ρ(·) denotes spectral radius of its matrix argument. In this case, the bias of the algorithm will be given by: E{vn} = −(I LN − B)−1r lim n→∞ (80) We shall now establish two results that provide ranges for selecting the step size µ to ensure convergence in the mean for each algorithm. Theorem 1: (Stability in the mean for Algorithm 1) Assume data model (1) and Assumption 1 hold. We select a doubly stochastic matrix A. Assume {Θ, Θ⊥} forms an orthonormal basis of IRL. Then, for any initial condition, Algorithm 1 asymptotically converges in the mean if the step-size satisfies: maxk λmax(Rx,k) where λmax(·) denotes the maximum eigenvalue of its matrix argument. 0 < µ < 2 (81) Proof: The convergence of (69) is determined by the stability of matrix B(1). The required mean stability condition is met by selecting µ so that: ρ(cid:0)(A(cid:62) DP Θ + DP Θ⊥ )(I LN − µSΘH x)(cid:1) < 1 (82) 17 (83) (84) (85) (86) (87) (88) (cid:13)(cid:13)2 (cid:13)(cid:13)P Θ⊥xi (cid:13)(cid:13)2 i=1 = j=1 aji = 1 with aji ≥ 0, Jensen's inequality guarantees: (cid:13)(cid:13)2(cid:17) i=1 j=1 i=1 j=1 aji P Θxj (cid:1)x DP Θ + DP Θ⊥ (cid:13)(cid:13)2 = aji P Θxj + P Θ⊥xi (cid:13)(cid:13)(cid:0)A(cid:62) (cid:13)(cid:13)(cid:13)2 +(cid:13)(cid:13)P Θ⊥xi Let x = col{x1, . . . , xN} be any block vector of size LN × 1. We have: Given that A is left stochastic, namely,(cid:80)N (cid:13)(cid:13)(cid:13) N(cid:88) (cid:13)(cid:13)(cid:13)2 +(cid:13)(cid:13)P Θ⊥xi (cid:13)(cid:13)(cid:13) N(cid:88) N(cid:88) (cid:16)(cid:13)(cid:13)(cid:13) N(cid:88) (cid:13)(cid:13)(cid:13)2 N(cid:88) (cid:13)(cid:13)(cid:13)2 ≤ N(cid:88) (cid:13)(cid:13)2 (cid:13)(cid:13)P Θxj (cid:13)(cid:13)2(cid:17) ≤ N(cid:88) N(cid:88) N(cid:88) (cid:13)(cid:13)P Θxj (cid:13)(cid:13)2 + N(cid:88) N(cid:88) (cid:13)(cid:13)P Θ⊥xi (cid:13)(cid:13)2 + (cid:13)(cid:13)P Θxj where for step (a) we use that A is right stochastic, namely,(cid:80)N (cid:13)(cid:13) ≤ 1 Consequently, the quantity in (84) can be upper-bounded as follows: (cid:16)(cid:13)(cid:13)(cid:13) N(cid:88) (cid:13)(cid:13)A(cid:62) = (cid:107)x(cid:107)2 N(cid:88) aji P Θxj aji P Θxj i=1 j=1 aji j=1 i=1 j=1 j=1 i=1 j=1 aji (a) = i=1 aji = 1. We conclude that: (89) We know that the spectral radius of any matrix X satisfies ρ(X) ≤ (cid:107)X(cid:107), for any induced norm. Then we have: DP Θ + DP Θ⊥ ρ(cid:0)(A(cid:62) DP Θ + DP Θ⊥ )(I LN − µSΘH x)(cid:1) ≤(cid:13)(cid:13)A(cid:62) (89)≤ (cid:13)(cid:13)I LN − µSΘH x (cid:13)(cid:13) The mean stability condition is thus met by selecting µ so that:(cid:13)(cid:13)I LN − µSΘH x DP Θ + DP Θ⊥ (cid:13)(cid:13) (cid:13)(cid:13)(cid:13)(cid:13)I LN − µSΘH x (cid:13)(cid:13) < 1. In the case where {Θ, Θ⊥} (90) (91) forms an orthonormal basis of IRL, then SΘ = I L. This leads us to the condition in (81). Theorem 2: (Stability in the mean for Algorithm 2) Assume data model (1) and Assumption 1 hold. We select a doubly stochastic matrix A. Then, for any initial condition, Algorithm 2 asymptotically converges in the mean if the step-size satisfies: 0 < µ < 2 maxk λmax(η2P Θ⊥ + Rx,k) (92) Proof: The convergence of (79) is determined by the stability of matrix B(2). Considering that: DP Θ + DP Θ⊥ )(I LN − µη2 DP Θ⊥ − µH x)(cid:1) ≤(cid:13)(cid:13)I LN − µη2 DP Θ⊥ − µH x ρ(cid:0)(A(cid:62) DP Θ + DP Θ⊥(cid:107) ≤ 1, the mean stability condition is met by selecting µ so that(cid:13)(cid:13)I LN − µη2 DP Θ⊥ − (cid:13)(cid:13) < 1. This leads us to the condition in (92). Furthermore, by Weyl's theorem, we have λmax(η2P Θ⊥+Rx,k) ≤ (cid:13)(cid:13) (93) since (cid:107)A(cid:62) µ H x 18 η2 + λmax(Rx,k) since P Θ⊥ and Rx,k are Hermitian matrices and λmax(P Θ⊥) = 1. This leads to the sufficient condition: 0 < µ < 2 η2 + maxk λmax(Rx,k) (94) B. Mean-square error behavior analysis We now study the mean-square error behavior of Algorithms 1 and 2. To this end, we consider the general update Σ of the weight vector vn relation (52) since both algorithms are of this form. From (51), the squared norm (cid:107)vn(cid:107)2 weighted by any positive semi-definite matrix Σ, i.e., (cid:107)vn(cid:107)2 Σ − 2 Re{r Σ = v∗ ∗ΣBnvn−1} − 2µ Re{g nΣvn, satisfies the following relation: nΣ(Bnvn−1 − r)} ∗ − µ2(cid:107)gn(cid:107)Σ + (cid:107)r(cid:107)2 Σ = (cid:107)vn−1(cid:107)2 B∗ (cid:107)vn(cid:107)2 (95) nΣBn Under the independence assumption, and considering that gn includes the zero-mean noise term zn which is independent of any other signal, taking expectations of both sides of (95) leads to: In the above expression, Σ is any positive semi-definite matrix that the user is free to choose in order to derive different performance metrics, and Σ(cid:48) = E{B n} in the second term on the RHS of (96). For the two presented algorithms, G is respectively given by: (A(cid:62) nΣBn}. Let G be the expected value of E{gng∗ ∗ (97) z,N Rx,N} D DP Θ +DP Θ⊥ )∗ z,1Rx,1, . . . , σ2 ∗ SΘ G(1) = (A(cid:62) G(2) = diag{σ2 DP Θ + DP Θ⊥ )DSΘdiag{σ2 z,1Rx,1, . . . , σ2 z,N Rx,N} E{(cid:107)vn(cid:107)2 Σ} = E{(cid:107)vn−1(cid:107)2 Σ(cid:48)} + µ2 trace{ΣE{gng n}} + (cid:107)r(cid:107)2 ∗ Σ − 2 Re(cid:8)E{r ∗ΣBnvn−1}(cid:9) (96) (98) (99) (100) (101) With G, equation (96) is expressed as: E{(cid:107)vn(cid:107)2 Σ} = E{(cid:107)vn−1(cid:107)2 Σ(cid:48)} + µ2trace{ΣG} + (cid:107)r(cid:107)2 Σ − 2 Re(cid:8)r ∗Σ B E{vn−1}(cid:9) Vectorizing matrices Σ and Σ(cid:48) by σ = vec(Σ) and σ(cid:48) = vec(Σ(cid:48)), it can be verified that: where the (LN )2 × (LN )2 matrix K is given by: (cid:48) = Kσ σ K = E{B n ⊗ B (cid:62) n} ≈ B ∗ (cid:62) ⊗ B ∗ The above approximation can be used provided that the step size is sufficiently small so that the influence of the second-degree term in µ can be neglected [13]. Equation (99) can then be expressed as: σ} = E{(cid:107)vn−1(cid:107)2 where we use the notation (cid:107)·(cid:107)Σ and (cid:107)·(cid:107)σ interchangeably, and E{(cid:107)vn(cid:107)2 (cid:16) sn−1 = vec µ2 G + rr Kσ} + s (cid:62) n−1 σ ∗ − 2 Re(cid:8)B E{vn−1}r ∗(cid:9)(cid:17) (102) (103) Theorem 3: (Mean-square stability) Assume data model (1) and Assumption 1 hold. Assume further that the step size µ is sufficiently small to guarantee the stability in the mean of the algorithms, and to ensure that (102) can be used as a reasonable representation for the evolution of the (weighted) mean-square error. Mean-square stability of 19 cooperative algorithms characterized by (51) requires the step-size µ to be chosen such that it ensures the stability of matrix K (in addition to the mean stability condition ρ(B) < 1). Proof: Iterating (102) starting from n = 0, we find that E{(cid:107)vn(cid:107)2 σ} = (cid:107)v0(cid:107)2 n−iKi−1σ (cid:62) K nσ + (104) with initial condition v0 = wo−w0. Provided that K is stable, the first term on the RHS of (104) converges to zero as n → ∞. We know from (52) that E{vn} is bounded because (52) is a BIBO stable recursion with a bounded driving term r. The second term on the RHS of (104) then converges as n → ∞. We conclude that E{(cid:107)vn(cid:107)2 σ} converges to a bounded value as n → ∞, and the algorithm is mean-square stable. i=1 s Theorem 4: (Transient MSD) Consider a sufficiently small step size µ to ensure mean and mean-square stabilities. E{(cid:107)vn(cid:107)2} of the cooperative algorithms characterized by (51), obtained by setting The MSD learning curve ζn = 1 N Σ = 1 N I LN , evolves according to the following recursion for n ≥ 1: (cid:2)(γn−1 + sn−1)(cid:62)vec(I LN ) − (cid:107)v0(cid:107)2 ζn = ζn−1 + 1 N (cid:62) γn = K γn−1 + (K − I (LN )2)(cid:62) sn−1 (cid:3) (I (LN )2−K)K n−1σ n(cid:88) with initial conditions ζ0 = 1 N (cid:107)v0(cid:107)2 and γ0 = 0. Proof: Comparing (104) at instants n and n − 1, we can relate E{(cid:107)vn(cid:107)2 (cid:62) n(cid:88) σ} = E{(cid:107)vn−1(cid:107)2 σ} − (cid:107)v0(cid:107)2 E{(cid:107)vn(cid:107)2 (I (LN )2−K)K n−1σ + s = E{(cid:107)vn−1(cid:107)2 σ} − (cid:107)v0(cid:107)2 Introducing the notation γn−1 = (I (LN )2−K)K n−1σ +s (cid:104) n(cid:88) n−iKi−1 − n−1(cid:88) (cid:62) s i=1 (cid:62) n−1σ+ s i=1 σ} and E{(cid:107)vn−1(cid:107)2 σ} as follows: n−iKi−1σ − n−1(cid:88) n−1−iKi−1σ (cid:62) n(cid:88) n−1−iKi−1(cid:105)(cid:62) n−iKi−1σ−n−1(cid:88) i=1 i=2 (cid:62) (cid:62) s s s (cid:62) n−1−iKi−1σ we can reformulate the recursive expression (107) as follows: i=2 i=1 E{(cid:107)vn(cid:107)2 σ} = E{(cid:107)vn−1(cid:107)2 σ} − (cid:107)v0(cid:107)2 (I (LN )2−K)K n−1σ + (γn−1 + sn−1)(cid:62) σ γn = K (cid:62) γn−1 + (K − I (LN )2)(cid:62) sn−1 with γ0 = 0. To derive the transient curve for the MSD, replace σ by 1 N vec{I LN}. Corollary 1: (Steady-state MSD) If the step size is chosen sufficiently small to ensure mean and mean-square convergence, then the steady-state MSD, defined as ζ∞ = limn→∞ ζn, is given by: ζ∞ = 1 N ∞ (I (LN )2 − K)−1vec(I LN ) (cid:62) s (111) (105) (106) (107) (108) (109) (110) 20 with s∞ = limn→∞ sn determined by (103), using E{v∞} = limn→∞ E{vn} determined by (80). Proof: From expression (102), we get: E{(cid:107)vn(cid:107)2 (I (LN )2−K)σ} = s (cid:62) ∞σ lim n→∞ Observe that the MSD calculation requires us to choose σ that satisfies: This leads to expression (111). (I (LN )2 − K) σ = 1 N vec(I LN ) V. SIMULATIONS (112) (113) In this section, we report simulation results that illustrate the theoretical results. All agents were initialized with zero parameter vectors wk,0 = 0 for all k. Simulated curves were obtained by averaging over 100 runs as we obtained sufficiently smooth curves to check the consistency with theoretical results. A. Algorithm validation We considered a network consisting of 12 agents with interconnections shown in Fig. 1(a). The parameter vectors to be estimated were of length L = 5. The input data xk,n were generated from circularly-symmetric zero-mean complex Gaussian distributions. White input data were considered first, by setting: Next, correlated input data, characterized by the following covariance matrix, were considered: Rx,k = σ2 x,k I 5 √−1 the imaginary unit. The modeling noises zk,n were i.i.d. zero-mean circularly-symmetric Gaussian with j = z,k were sampled from U(0.8, 1.2) and variables, independent of any other signals. The variances σ2 U(0.18, 0.22), respectively. Their values are shown in Fig. 1(b). We considered two sets of subspace basis vectors. The first set is the standard basis: x,k and σ2 Θ1 = [e1, e2, . . . , eM ], (116) where ei denotes a vector of length N with 1 at the ith entry and 0 otherwise. Its orthogonal complementary subspace is spanned by Θ1,⊥ = [eM +1, . . . , eL]. This setup can be interpreted as a variable selection process for  Rx,k = σ2 x,k × 1 −.4 + .3j 1 −.4 − .3j .2 + .1j −.4 − .3j .1 + .05j .02 − .02j .2 − .1j −.4 + .3j 1 .1 − .05j .2 − .1j −.4 + .3j .2 + .1j −.4 − .3j .1 + .05j .2 + .1j −.4 − .3j 1  .02 + .02j .1 − .05j .2 − .1j −.4 + .3j 1 (114) (115) 21 (a) Network topology. (b) Agent input and noise variances. Fig. 1. Network topology and input-noise variances. information exchange, where the first M entries of the optimal parameter vectors are identical across the network. Parameter M was set to 3. The second set of basis vectors is a complex Vandermonde matrix: (117)  Θ2 = 1 e−jψ1 ··· 1 e−jψ2 ··· e−j(L−1)ψ1 e−j(L−1)ψ2 ··· ··· ··· ··· 1 e−jψM ··· e−j(L−1)ψM  with ψk = 2πd sin θk. Matrix Θ2 can represent the array manifold of a uniform linear array (ULA) with inter- λo element space d, operating at wavelength λo with impinging signal directions of angles θk. Parameter M was set to 3, with θ1 = π 2 . We considered three settings to validate the theoretical results. 6 , θ2 = π 4 , θ3 = π 3 and d = λo In the first setting, we assumed that model (6) matches the observation data. The entries of the coefficient vectors k were sampled from the Gaussian distribution N (0, 1). The step-size parameter µ for Algorithm 1 was uo and ξo successively set to 0.01 and 0.02. A uniform combination matrix A with a(cid:96)k = Nk−1 was used. With Θ1, note that matrix SΘ is equal to I 5. With Θ2, it was successively set to ΘΘ∗ + Θ⊥Θ∗ ⊥ as in (29), and to I 5. The transient behavior and the steady-state MSD were determined theoretically. The results with subspace settings Θ1 and Θ2, for white and correlated input data, are shown in Fig. 2. It can be observed that setting SΘ to I 5 for Θ2 leads to a better convergence behavior. For Algorithm 2, we did not set the parameter η1 explicitly but we used the same combination matrix A as for Algorithm 1. Parameters (µ, η2) were set to (0.02, 0.01) with white input data. With correlated input data, the following combinations (µ, η2) were considered: {(0.01, 0.01); (0.01, 0.02); (0.02, 0.01)}. The results are shown in Fig. 3. The simulation results match the theoretical results, and illustrate the trade-off between the convergence speed and the steady-state MSD. It can also be observed with Algorithm 2 that a small value for η2 is preferable since constraining the norm of node-specific components in the complementary subspace Θ⊥ introduces a bias that can degrade the performance. As leaky-LMS, this kind of regularization can improve the stability of the algorithm for some particular problems and practical applications, at the cost of an extra estimation bias. We then considered another scenario in order to illustrate the interest of the extra degree of freedom provided 156211374891210<2x0.80.850.90.9511.051.11.151.2<2z0.180.1850.190.1950.20.2050.210.2150.2212345678910111222 (a) Using Θ1. (b) Using Θ2. Fig. 2. Learning curves and model validation of Algorithm 1 with different settings. (a) Using Θ1. (b) Using Θ2. Fig. 3. Learning curves and model validation of Algorithm 2 with different settings. (a) Using Θ1. (b) Using Θ2. Fig. 4. Learning curves and model validation of the algorithms using ξk with small variances. by η2 in Algorithm 2. Experimental setups were left unchanged with correlated inputs except for the entries of k, which were sampled from Gaussian distribution N (0, 0.01). We successively set η2 to 0, 0.1 and 1 in order to ξo progressively constrain the variance of ξk. Note that with η2 = 0, Algorithm 2 reduces to Algorithm 1. The results with Θ1 and Θ2 are provided in Fig. 4. The result with non-cooperative LMS is also provided as a reference. In the second setting, we assumed that the node-specific components o k in (6) do not strictly lie in the comple- mentary subspace Θ⊥. To evaluate the robustness of our algorithms and the power of the analytical models, we Iterationn05001000150020002500300035004000MSD(dB)-30-25-20-15-10-50510Simulated transient MSDTheoretical transient MSDTheoretical steady-state MSD7=0:02,whiteinput7=0:01,whiteinput7=0:01,correlatedinput7=0:02,correlatedinputIterationn05001000150020002500300035004000MSD(dB)-30-25-20-15-10-50510Simulated transient MSDTheoretical transient MSDTheoretical steady-state MSD7=0:01,whiteinput7=0:02,correlatedinput7=0:02,whiteinput7=0:01,correlatedinput7=0:01,correlatedinput,S#7=0:01,whiteinput,S#Iterationn05001000150020002500300035004000MSD(dB)-30-25-20-15-10-50510Simulated transient MSDTheoretical transient MSDTheoretical steady-state MSD7=0:02,22=0:01,whiteinput7=0:02,22=0:01,correlatedinput7=0:01,22=0:02,correlatedinput7=0:01,22=0:01,correlatedinputIterationn05001000150020002500300035004000MSD(dB)-25-20-15-10-50510Simulated transient MSDTheoretical transient MSDTheoretical steady-state MSD7=0:01,22=0:01,correlatedinput7=0:02,22=0:01,correlatedinput7=0:02,22=0:01,whiteinput7=0:01,22=0:02,correlatedinputIterationn05001000150020002500300035004000MSD(dB)-30-25-20-15-10-50Simulated transient MSDTheoretical transient MSDTheoretical steady-state MSDnon-cooperative22=0(coincidentwithAlg1)22=0:122=1Iterationn05001000150020002500300035004000MSD(dB)-35-30-25-20-15-10-50510Simulated transient MSDTheoretical transient MSDTheoretical steady-state MSDnon-cooperative22=0(coincidentwithAlg1)22=0:122=123 Fig. 5. Learning curves of the algorithms using ξk with small variances. set: o k = Θνo k + Θ⊥ξo k (118) where νo k) lie in span(Θ) but differ from one node to another. The entries of uo and νo k are zero-mean circular Gaussian variables. This setting refers to a non-ideal situation because components k were sampled from Θ(uo + νo Gaussian distributions N (0, 1) and N (0, 0.01), respectively. The step-size µ was set to 0.01 for Algorithms 1 and 2. Parameter η2 in Algorithm 2 was set to 0.01. Subspace Θ1 and white input signals were considered to test the model. The transient behavior and the steady-state MSD were determined theoretically. The simulation results provided in Fig. 5 match the theoretical results, and illustrate that cooperation among nodes can still be beneficial when optimal solutions in the subspace Θ are different but close to each other. This is another illustration of the conclusion reached in [19] for single-task diffusion LMS operating in multitask environments. In the third setting, we exploited the leaky property of Algorithm 2 to promote its use in real applications. It is well known that the (non-cooperative) leaky LMS algorithm introduces an estimation bias compared to the (non- cooperative) LMS, but improves its robustness when applied to practical applications [51]. In particular, it avoids the so-called weight-drift problem of the LMS algorithm [52]. To highlight this phenomenon in the context of diffusion adaptation, we assumed that, say, the last tap/channel of node #1 was failing to work and was providing consistent null-valued readings, i.e., [xn,1]5 = 0 for all n. We also assumed that, e.g., finite-precision effect was corrupting the combination step (29), or (45), with an additive non-zero mean disturbance qk. The poor conditioning of regressors associated with a non-zero mean disturbance is known to possibly lead to a weight-drift problem. We considered the same experimental setup as in the first experiment with the standard basis Θ1. We picked each entry of the random vectors qk according to the Gaussian distribution N (10−4, 10−8). We set η2 to 0.1. All the vectors wk were initialized to 0. Fig. 6 shows the behavior of the weight vector at node #1 for (a) Algorithm 1 with SΘ = I 5, and (b) Algorithm 2. We can observe the drift of the 5th entry of w1 with Algorithm 1. Algorithm 2 alleviates this effect. Iterationn050010001500MSD-30-25-20-15-10-50510Theoreticalsteady-stateMSDTheoreticaltransientMSDSimulatedtransientMSDnon-cooperativeAlgorithm2,22=0:01Algorithm124 (a) Weight behavior of Algorithm 1. (b) Weight behavior of Algorithm 2. Fig. 6. Weight behavior of Algorithms 1 and 2 with singular inputs and non-zero mean disturbance. B. Target localization We now consider a target localization problem. Cooperative localization with a diffusion strategy was already addressed in the case of a single target [13], and of multiple nearby targets [4]. We focus here on the case where targets lie in a manifold. To make the presentation clearer, we assumed that the targets were collinear in IR3. Their locations were estimated by the network with 100 nodes shown in Fig. 7(a). Each node randomly selected a target to localize. Let R be a member of the rotation group SO(3) defined by the matrix R = Rx(θx) Ry(θy) Rz(θz), where Rx(θx), Ry(θy) and Rz(θz) are rotation matrices that rotate vectors by an angle of θx,y,z around x, y and z axis, respectively. The coordinate vector wo q of each target q was generated as follows: wo q = R1,2u + qr3 (119) where R1,2 is the matrix composed of the first and second columns of R, and r3 corresponds to the third column of R. As illustrated in Fig. 7(b), this model means that all targets lie on a common line defined by point R1,2u and direction vector r3. Parameter q characterizes the location of each target q on this line. We considered the problem of estimating u (common to all targets) and the parameters q for seven targets. We set the angles and the parameter vectors in (119) as follows: π 6 , θx = v = [1 2](cid:62) θy = π 3 , θz = π 4 1 = 0, 2 = 1, 3 = 3, 4 = 4, 5 = 7, 6 = 7.5, 7 = 9 The distance between each agent k and target q can be expressed in the inner product form: rkq = xkq(wo q − pk) (120) (121) (122) (123) where pk is the location of agent k, and xkq is the unit-norm row vector pointing from pk to wo that agents were aware of their location pk. Let dkq = rkq + xkq pk, that is, dkq = xkq wo q. We assumed q. The problem was to Realpartofw1-0.6-0.4-0.200.20.40.60.811.2Imaginarypartofw1-0.5-0.4-0.3-0.2-0.100.10.20.30.40.50.60.70.80.9151342Realpartofw1-0.6-0.4-0.200.20.40.60.8Imaginarypartofw1-0.4-0.3-0.2-0.100.10.20.30.40.50.61354225 (a) Network topology. (b) Collinear targets. Fig. 7. Network topology and locations of targets. (a) Network MSD. (b) Estimation results (non-coop. LMS). (c) Estimation results (Algorithm 1). Fig. 8. Estimated network MSD and estimation results for a single realization. q from noisy streaming measurements {dkq(n), xkq,n} collected by each agent k, and governed by the estimate wo linear model [13]: dkq(n) = xkq,nwo q + zkq(n) with kq diag{αk1(n), αk2(n)} xkq,n = [1 − βk(n)] xkq + x ⊥ (124) with zkq(n) a zero-mean temporally and spatially i.i.d. Gaussian noise of variance σ2 z. As shown in (124), the measured direction vector xkq,n was assumed to be a noisy realization of the unit-norm vector pointing from pk to q, with x⊥ kq a unit-norm orthogonal contribution to xkq. Random variables αk1(n), αk2(n), βk(n) and zk(n) were wo zero-mean Gaussian with standard deviation σα1 = σα2 = 0.1, σβ = 0.001 and σz = 0.3, respectively. We ran the (non-cooperative) LMS algorithm at each node, and Algorithm 1, with Θ = R1,2 and Θ⊥ = r3. The step-size µ was set to 0.1. A uniform combination matrix A with a(cid:96)k = Nk−1 was used for Algorithm 1, where Nk denotes the cardinality of the neighborhood of node k. Figure 8(a) compares the MSD of these strategies. Figures 8(b) and 8(c) show one realization of the target locations estimated with the (noncooperative) LMS algorithm and Algorithm 1. This experiment illustrates the advantage of cooperative strategies over the non-cooperative one. 55.25.45.65.86−3−2.8−2.6−2.4−2.2−200.20.40.60.81xyzR1,2ur3xyzIterationn01000200030004000500060007000MSD(dB)-25-20-15-10-5051015Non-cooperativeLMSCooperativealgorithm−10−50−0.500.511.522.5012345x y zTrueEstimated−10−50−0.500.511.522.5012345x y zTrueEsitamted26 VI. CONCLUSION AND PERSPECTIVES In this paper, we formulated an online multitask adaptation problem that assumes that all tasks share a common latent feature representation, locally refined by node-specific contributions. This model can be extended into inter- esting directions by imposing new constraints, depending on applications. Based on this principle, we derived two cooperative algorithms and analyzed their performance. Although this work considers that common representation subspaces are known a priori, it paves the way towards more general frameworks. APPENDIX A PROOF OF LEMMA 1 The uniqueness of the solution of (8) follows from the strict convexity of (10), which is ensured by the positive definiteness of its Hessian matrix. For the quadratic cost (10), the Hessian matrix with respect to the vector of stacked variables col{u, ξ1, . . . , ξN} is block diagonal [3, App. B], with blocks given by the following matrix X and its transpose:  X 0 Θ∗(cid:0)(cid:80)N 0 X k=1 Rx,k ⊥Rx,1Θ Θ∗ ...  (cid:1)Θ Θ∗Rx,1Θ⊥ . . . Θ∗Rx,N Θ⊥ (cid:62) Θ∗ ⊥Rx,1Θ⊥ 0 ... Θ∗ ⊥Rx,N Θ 0 Θ∗ ⊥Rx,N Θ⊥ ∇2J glob = with X =   (125) where Θ and Θ⊥ have full column rank. The positive definiteness of (125) can be checked by verifying the positive definiteness of each term Θ∗ Rx,kΘ and of the Schur complement relative to the block diagonal corner of X, namely, [53] Schur(X) = Rx,kΘ−Θ∗ Rx,kΘ⊥(Θ∗ ⊥Rx,kΘ⊥)−1Θ∗ (126) ⊥Rx,kΘ(cid:3) N(cid:88) (cid:2)Θ∗ k=1 where each inverse (Θ∗ is positive definite since it is the Schur complement of the block Θ∗ ⊥Rx,kΘ⊥)−1 exists since Θ⊥ has full column rank. Each term inside the summation (126) Rx,k Θ in the positive definite matrix:  Θ∗ Rx,kΘ Θ∗ Θ∗ ⊥Rx,kΘ Θ∗ Rx,kΘ⊥ ⊥Rx,kΘ⊥  = [Θ Θ⊥]∗ Rx,k[Θ Θ⊥] > 0 (127) This guarantees the positive definiteness of (126). It follows that the cost in (10) is strictly convex and has a unique minimizer. APPENDIX B PROOF OF LEMMA 2 Without loss of generality, assume that η1 > η2. Otherwise, replace (129) by: E(cid:8)dk(n) − xk,n(Θu + k)2(cid:9) + η1 N(cid:88) k=1 (cid:107)k(cid:107)2 + (η2 − η1) Recalling that P Θ⊥ = I L − P Θ, the objective function (31) can be written as follows: k=1 J glob(cid:0)u,{k}N J glob(cid:0)u,{k}N N(cid:88) (cid:1) = (cid:1) = (η1 − η2) k=1 k=1 E(cid:8)dk(n) − xk,n(Θu + k)2(cid:9) + η2 N(cid:88) k=1 (cid:107)P Θk(cid:107)2 + N(cid:88) (cid:124) k=1 (cid:123)(cid:122) J glob 1 N(cid:88) k=1 27 (128) (129) (cid:107)P Θ⊥k(cid:107)2 N(cid:88) (cid:107)k(cid:107)2 k=1 (cid:125) The uniqueness of the minimizer of (31) follows from its strict convexity. For the quadratic cost in (129), the 1 with respect to the vector of stacked variables col{u, 1, . . . , N} is again block diagonal, with Hessian of J its blocks determined by the matrix Y below and its transpose: glob ∇2J glob 1 = 0 (cid:62) 0 Y  Y Θ∗(cid:0)(cid:80)N  (cid:1)Θ with Y =  k=1 Rx,k Rx,1Θ ... Rx,N Θ Θ∗Rx,1 Rx,1 + η2 I 0 . . . ... Θ∗Rx,N  Rx,k(Rx,k + η2I)−1Rx,kΘ(cid:3) Rx,N + η2 I 0 (130) (131) (132) (133) (134) The positive definiteness of (130) can be checked by verifying the positive definiteness of each term Rx,k + η2 I and of the Schur complement relative to the right block diagonal corner in (130), namely, [53] N(cid:88) (cid:2)Θ∗ Schur(Y ) = Rx,kΘ−Θ∗ Since they are positive definite, each covariance matrix Rx,k can be decomposed as follows: k=1 Rx,k = U k diag{λk,1, . . . , λk,L} U ∗ k where the λk,i are the eigenvalues of Rx,k, which are real and positive, and U k is the corresponding matrix of eigenvectors. Since U k is an orthonormal matrix, each term in the summation (132) can be written as: Θ∗ =Θ∗ Rx,kΘ − Θ∗ Udiag Rx,k(Rx,k + η2I)−1Rx,kΘ , . . . , λk,L − λ2 k,L k,1 λk,1 − λ2 λk,1+η2 λk,L+η2 (cid:110) (cid:111) ∗Θ > 0 U block diagonal matrix diag{Rx,1+η2I, . . . , Rx,N +η2I} is positive definite. Finally, since (η1−η2)(cid:80)N Since Θ has full column rank, the above matrix and the Schur complement (132) are positive definite. In addition, the k=1 (cid:107)P Θk(cid:107)2 in (129) is convex, problem (129) is strictly convex and problem (30) has a unique solution. 28 REFERENCES [1] 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 on Machine Learn. for Signal Process. (MLSP), Reims, France, Sept. 2014, pp. 1–6. [2] A. H. Sayed, "Adaptive networks," Proceedings of the IEEE, vol. 102, no. 4, pp. 460–497, 2014. [3] A. H. Sayed, "Adaptation, learning, and optimization over networks," in Foundations and Trends in Machine Learning, vol. 7, pp. 311–801. NOW Publishers, Boston-Delft, Jul. 2014. [4] 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. [5] D. P. Bertsekas, "A new class of incremental gradient methods for least squares problems," SIAM J. Optimiz., vol. 7, no. 4, pp. 913–926, Nov. 1997. [6] M. G. Rabbat and R. D. Nowak, "Quantized incremental algorithms for distributed optimization," IEEE J. of Sel. Topics Areas Commun., vol. 23, no. 4, pp. 798–808, Apr. 2005. [7] D. Blatt, A. O. Hero, and H. Gauchman, "A convergent incremental gradient method with constant step size," SIAM J. Optimiz., vol. 18, no. 1, pp. 29–51, Feb. 2007. [8] C. G. Lopes and A. H. Sayed, "Incremental adaptive strategies over distributed networks," IEEE Trans. Signal Process., vol. 55, no. 8, pp. 4064–4077, Aug. 2007. [9] A. Nedic and A. Ozdaglar, "Distributed subgradient methods for multi-agent optimization," IEEE Trans. Autom. Control, vol. 54, no. 1, pp. 48–61, Jan. 2009. [10] S. Kar and J. M. F. Moura, "Distributed consensus algorithms in sensor networks: Link failures and channel noise," IEEE Trans. Signal Process., vol. 57, no. 1, pp. 355–369, Jan. 2009. [11] K. Srivastava and A. Nedic, "Distributed asynchronous constrained stochastic optimization," IEEE J. Sel. Topics Signal Process., vol. 5, no. 4, pp. 772–790, Aug. 2011. [12] A. H. Sayed, S.-Y Tu, J. Chen, X. Zhao, and Z. Towfic, "Diffusion strategies for adaptation and learning over networks," IEEE Sig. Process. Mag., vol. 30, no. 3, pp. 155–171, May 2013. [13] A. H. Sayed, "Diffusion adaptation over networks," in Academic Press Library in Signal Processing, R. Chellapa and S. Theodoridis, Eds., pp. 323–454. Academic Press, Elsevier, 2014. [14] 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. [15] F. S. Cattivelli and A. H. Sayed, "Diffusion LMS strategies for distributed estimation," IEEE Trans. Signal Process., vol. 58, no. 3, pp. 1035–1048, Mar. 2010. [16] J. Chen and A. H. Sayed, "Diffusion adaptation strategies for distributed optimization and learning over networks," IEEE Trans. Signal Process., vol. 60, no. 8, pp. 4289–4305, Aug. 2012. [17] 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. [18] S.-Y. Tu and A. H. Sayed, "Diffusion strategies outperform consensus strategies for distributed estimation over adaptive networks," IEEE Trans. Signal Process., vol. 60, no. 12, pp. 6217–6234, Dec. 2012. [19] 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. [20] 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. [21] Y. Hirata, D. P. Mandic, H. Suzuki, and K. Aihara, "Wind direction modelling using multiple observation points," Philosophical Transactions of the Royal Society A, vol. 366, pp. 591–607, 2008. [22] R. Nassif, C. Richard, A. Ferrari, and A. H. Sayed, "Multitask diffusion LMS with sparsity-based regularization," in Proc. IEEE Int. Conf. Acoust., Speech, Signal Process. (ICASSP), Brisbane, Australia, 2015, pp. 3516–3520. 29 [23] 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. [24] J. Chen, C. Richard, and A. H. Sayed, "Adaptive clustering for multitask diffusion networks," in Proc. European Signal Process. Conf. (EUSIPCO), Nice, France, Sept. 2015, pp. 200–204. [25] R. Nassif, C. Richard, J. Chen, A. Ferrari, and A. H. Sayed, "Diffusion LMS over multitask networks with noisy links," in Proc. IEEE Int. Conf. Acoust., Speech, Signal Process. (ICASSP), Shanghai, China, 2016. [26] S. Monajemi, K. Eftaxias, S. Sanei, and Ong S.-H., "An informed multitask diffusion adaptation approach to study tremor in Parkinson's disease," IEEE J. Sel. Topics Signal Process., vol. 10, no. 7, pp. 1306–1314, Oct. 2016. [27] Y. Wang, W. P. Tay, and W. Hu, "Multitask diffusion LMS with optimized inter-cluster cooperation," in Proc. IEEE Stat. Signal Process. Workshop (SSP), Palma de Mallorca, Spain, 2016, pp. 1–5. [28] N. Roula, 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. [29] S. Monajemi, S. Sanei, Ong S.-H., and A. H. Sayed, "Adaptive regularized diffusion adaptation over multitask networks," in Proc. IEEE Int. Workshop on Machine Learn. for Signal Process. (MLSP), Boston, USA, Sept. 2015, pp. 1–5. [30] S. Khawatmi, A. M. Zoubir, and A. H. Sayed, "Decentralized clustering over adaptive networks," in Proc. European Signal Process. Conf. (EUSIPCO), Nice, France, 2015, pp. 2696–2700. [31] J. Plata-Chaves, N. Bogdanovi´c, and K. Berberidis, "Distributed diffusion-based LMS for node-specific adaptive parameter estimation," IEEE Trans. Signal Process., vol. 63, no. 13, pp. 3448–3460, Jul. 2015. [32] J. Plata-Chaves, H. H. Bahari, M. Moonen, and A. Bertrand, "Unsupervised diffusion-based LMS for node-specific parameter estimatin over wireless sensor networks," in Proc. IEEE Int. Conf. Acoust., Speech, Signal Process. (ICASSP), Shanghai, China, Mar. 2016, pp. 4159–4163. [33] 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. [34] A. Bertrand and M. Moonen, "Distributed adaptive estimation of node-specific signals in wireless sensor networks with a tree topology," IEEE Trans. Signal Process., vol. 59, no. 5, pp. 2196–2210, May 2011. [35] N. Bogdanovi´c, J. Plata-Chaves, and K. Berberidis, "Distributed incremental-based LMS for node-specific adaptive parameter estimation," IEEE Trans. Signal Process., vol. 62, no. 20, pp. 5382–5397, Oct. 2014. [36] J. Chen, S. K. Ting, C. Richard, and A. H. Sayed, "Group diffusion LMS," in Proc. IEEE Int. Conf. Acoust., Speech, Signal Process. (ICASSP), Shanghai, China, March 2016. [37] J. Baxter, "A model of inductive bias learning," J. Artif. Intell. Res., vol. 12, no. 1, pp. 149–198, Feb. 2000. [38] R. K. Ando and T. Zhang, "A framework for learning predictive structures from multiple tasks and unlabeled data," J. Mach. Learn. Res., vol. 6, pp. 1817–1853, 2005. [39] J. Zhou, J. Chen, and J. Ye, "Clustered multi-task learning via alternating structure optimization," in Proc. Adv. Neural Inf. Process. Syst., Granada, Spain, Dec. 2011, pp. 702–710. [40] J. Chen, L. Tang, J. Liu, and J. Ye, "A convex formulation for learning shared structures from muliple tasks," in Proc. Ann. Int. Conf. Machine Learning (ICML), Montreal, Canada, Jun. 2009, pp. 137–144. [41] L. Jacob, F. Bach, and J.-P. Vert, "Clustered multitask learning: A convex formulation," Adv. Neural Inf. Process. Syst., vol. 21, pp. 745–752, 2009. [42] G. Obozinski, B. Taskar, and M. Jordan, "Joint covariate selection and joint subspace selection for multiple classification problems," Stat. Comput., vol. 20, no. 2, pp. 231–252, 2010. [43] L. J. Griffiths and C. W. Jim, "An alternative approach to linear constrained adaptive beamforming," IEEE Trans. Antenn. Propag., vol. 30, no. 1, pp. 27–34, Jan. 1982. [44] X. Zhao and A. H. Sayed, "Clustering via diffusion adaptation over networks," in Proc. Int. Workshop Cognitive Inf. Process. (CIP), Parador de Baiona, Spain, May 2012, pp. 1–6. 30 [45] R. Nassif, C. Richard, A. Ferrari, and A. H. Sayed, "Performance analysis of multitask diffusion adaptation over asynchronous networks," in Proc. Asilomar Conf. Signals, Systems and Computers (ASILOMAR), Pacific Grove, CA, USA, Nov. 2014, pp. 788–792. [46] J. Chen and C. Richard, "Performance analysis of diffusion LMS in multitask networks," in Proc. IEEE Int. Workshop on Compt. Adv. in Multi-Sensor Adaptive Process. (CAMSAP), Saint Martin, France, Dec. 2013, pp. 137–140. [47] J. Plata-Chaves, N. Bogdanovi´c, and K. Berberidis, "Distributed incremental-based RLS for node-specific parameter estimation over adaptive networks," in Proc. European Signal Process. Conf. (EUSIPCO), Marrakech, Morocco, Sept. 2013, pp. 1–5. [48] N. Bogdanovi´c, J. Plata-Chaves, and K. Berberidis, "Distributed diffusion-based LMS for node-specific parameter estimation over adaptive networks," in Proc. IEEE Int. Conf. Acoust., Speech, Signal Process. (ICASSP), Florence, Italy, May 2014, pp. 7223–7227. [49] S. Boyd and L. Vandenberghe, Convex Optimization, Cambridge University Press, 2004. [50] I. Gohberg and V. Olshevsky, "Fast algorithms with preprocessing for matrix-vector multiplication problems," J. Complexity, vol. 10, no. 4, pp. 411–427, 1994. [51] A. H. Sayed, Fundamentals of Adaptive Filtering, J. Wiley & Sons, Hoboken, NJ, 2003. [52] V. H. Nascimento and A. H. Sayed, "Unbiased and stable leakage-based adaptive filters," IEEE Trans. Signal Process., vol. 47, no. 12, pp. 3261–3276, 1999. [53] F. Zhang, The Schur Complement and its Applications, Springer, 2005.
1803.11457	1	1803	2018-03-30T13:47:36	Co-evolution and morphogenetic systems	[ "cs.MA", "cs.CY" ]	The emerging field of morphogenetic engineering proposes to design complex heterogeneous system focused on the paradigm of emergence. Necessarily at the interface of disciplines, its concepts can be defined through multiple viewpoints. This contribution aims at linking a co-evolutionary perspective on such systems with morphogenesis, and therein at bringing a novel conceptual approach to the bottom-up design of complex systems which allows to fully consider co-evolutive processes. We first situate systems of interest at the interface between biological and social systems, and introduce a multidisciplinary perspective on co-evolution. Building on Holland's signals and boundaries theory of complex adaptive systems, we finally suggest that morphogenetic systems are equivalent to combinations of co-evolutionary niches. This introduces an entry to morphogenetic engineering focused on co-evolution between components of a system. Applications can be found in a broad range of subjects, which we illustrate with the example of planning in territorial systems, suggesting an extended scope for the relevance of morphogenetic engineering concepts.	cs.MA	cs	Co-evolution and morphogenetic systems Juste Raimbault1,2 1UPS CNRS 3611 ISC-PIF, Paris, France 2UMR CNRS 8504 G´eographie-cit´es, Paris, France [email protected] Abstract The emerging field of morphogenetic engineering proposes to design complex heterogeneous system focused on the paradigm of emergence. Necessarily at the interface of disciplines, its concepts can be defined through multiple viewpoints. This contribution aims at linking a co- evolutionary perspective on such systems with morphogenesis, and therein at bringing a novel conceptual approach to the bottom-up design of com- plex systems which allows to fully consider co-evolutive processes. We first situate systems of interest at the interface between biological and social systems, and introduce a multidisciplinary perspective on co-evolution. Building on Holland's signals and boundaries theory of complex adaptive systems, we finally suggest that morphogenetic systems are equivalent to combinations of co-evolutionary niches. This introduces an entry to mor- phogenetic engineering focused on co-evolution between components of a system. Applications can be found in a broad range of subjects, which we illustrate with the example of planning in territorial systems, sug- gesting an extended scope for the relevance of morphogenetic engineering concepts. 1 Introduction The standard opposition between bottom-up self-organized systems and top- down architected has been challenged by the field of morphogenetic engineering (ME), which focuses on the design of systems from the bottom-up: tuning the behavior of agents and processes at the micro level allows in this paradigm to obtain a desired functionality at the macro level, without any centralization (Doursat et al., 2013). Doursat (2011) identifies two streams of research em- bracing this approach, namely complex systems controlled by Information and Communication Technologies (ICT) and ICT inspired by complex systems, typ- ically in the spirit of bio-inspired designs. Swarm chemistry (Sayama, 2009) is a good illustration how simple rules produce complex emergent patterns in swarm of particles, and how these can be interactively evolved (Sayama et al., 2009). The type of systems to which ME applies stays however limited, as it for example does not grasp social or territorial systems. The theoretical and practical implications of an extension in scope of ME would be significant, as it 1 would for example converge with bottom-up approach to planning for territorial systems (Batty, 2007). We propose here such a extension in scope, in an indirect way through the conceptual exercise of a new entry on morphogenetic systems from the point of view of co-evolution. The aim of this paper is thus twofold: (i) develop the concept of co-evolution and show how it connects to morphogenesis and ME; (ii) show how this link between co-evolution and morphogenesis opens new potentialities for the application of ME concepts. The rest of the paper is organized as follows: we first develop the parallel between biological and social systems, what will suggest the exploration of the concept of co-evolution, on which a multidiscplinary view is then constructed. We recall then the concept of morphogenesis and introduce its connexion with co-evolution. This allows us to conclude with an illustration of a possible scope extension of ME to territorial systems. 2 A parallel between biological and social sys- tems 2.1 Epistemology of life The parallel between social and biological systems is not rare, altough it some- times relies on analogies as for example in West's Scaling theory which applies similar growth equations starting from scaling laws, with however inverse con- clusions concerning the relation between size and pace of life (Bettencourt et al., 2007). Scaling relations do not hold when we try to apply them to a single ant, and they must be applied to the whole ant colony which is then the organism studied. When adding the property of cognition, we confirm that it is the rel- evant level, since the colony shows advanced cognitive properties, such as the resolution of spatial optimization problems, or the quick answer to an external perturbation. Human social organizations, cities, could be seen as organisms ? Banos (2013) extends the metaphor of the urban anthill but recalls that the parallel stops quickly. We can however show that some concepts from the epistemology of biology can be useful to understand social systems. Monod (1970) develops epistemological principles for the study of life. Living organisms answer to three essential properties that differentiate them from other systems: (i) the teleonomy , i.e. the property that these are "objects with a project", project that is reflected in their structure and the structure of artifacts they produce; (ii) the importance of morphogenetic processes in their constitu- tion; (iii) the property of the invariant reproduction of information defining their structure. Monod furthermore introduces research directions towards a theory of cultural evolution. Teleonomy is crucial in social structures, since any organization aims at satisfying a set of objectives, even if in general it will not succeed and the objectives will co-evolve with the organization. This notion of multi-objective optimization is typical of complex socio-technical systems. 2 2.2 Morphogenesis and co-evolution We postulate that the concept of morphogenesis is an essential tool to under- stand social systems, with a definition very similar to the one used in biology. We sum it up as the existence of relatively autonomous processes guiding the growth of the system and implying causal circular relations between form and function, that witness an emergent architecture in the sense of Doursat et al. (2012). For social systems, isolating the system is more difficult and the notion of boundary will be less struct than for a biological system. Finally, the reproduction of information is at the core of cultural evolution, through the transmission of culture and memetics, the difference being that the ratio of scales between the frequency of transmission and mutation and cross- over processes or other non-memetic processes of cultural production is relatively low, whereas is many orders of magnitude in biology. An example shows that the parallel is not always absurd : Gabora and Steel (2017) proposes an auto- catalytic network model for cognition, that would explain the apparition of cultural evolution through processes that are analogous to the ones that occurred at the apparition of life. But even if processes are at the origin analogous, the nature of evolution is then quite different, as show van der Leeuw et al. (2009), Darwinian criteria for evolution being not sufficient to explain the evolution of our organized societies. 2.3 Levels of emergence One point that also must retain our attention is the greater difficulty to define levels of emergence for social systems: Roth (2009) underlines the risk to fall into ontological dead-ends if levels were badly defined. He argues that more generally we must go past the single dichotomy micro-macro that is used as a caricature of the concepts of weak emergence, and that ontologies must often be multi-level and imply multiple intermediate levels. This echoes to the question of the exis- tence of strong emergence in social structures, that in sociological terms corre- sponds to the idea of the existence of "collective beings" (Angeletti and Berlan, 2015). Morin (1980) indeed distinguishes living systems of the second type (multi-cellular) and of the third type (social structures), but precises that the subjects of the latest are necessarily unachieved. This thematic entry on social and biological systems suggests therefore a cru- cial conceptual role of morphogenesis and co-evolution in understanding these systems, and therefore a possible link between morphogenetic systems and co- evolution. We will now develop this concept in more details. 3 A multidisciplinary approach to co-evolution Co-evolution is indeed used in several disciplines, but does not have a unified definition. We propose to take advantage of this disparity of views and con- struct a multidisciplinary approach. We sequentially review several disciplines, 3 starting from biology with a progressive shift to social systems. 3.1 Biology The concept of co-evolution in biology is an extension of the well-known con- cept of evolution, that can be tracked back to Darwin. Durham (1991) recalls the components and systemic structures that are necessary for the presence of evolution, that are: 1. Processes of transmission, implying transmission units and transmission mechanisms. 2. Processes of transformation, that necessitates sources of variation. 3. Isolation of sub-systems such that the effects of previous processes are observable in differentiations. Therefore, a population submitted to constraints (often conceptually syn- thesized as a fitness) that condition the transmission of the genetic heritage of individuals (transmission), and to random genetic mutations (transformation), will indeed be in evolution in the spatial territories it populates (isolation), and by extension the species to which it can be associated. Co-evolution is then defined as an evolutionary change in a characteristic of individuals of a population, in response to a change in a second population, which in turn responds by evolution to the change in the first, as synthesized by Janzen (1980). He furthermore highlights the subtlety of the concept and warns against its unjustified uses: the presence of a congruence between two characteristics that seem adapted one to the other does not necessarily imply a co-evolution, since one species could have adapted alone to one characteristic already present in the other. This rough presentation partly hides the real complexity of ecosystems: pop- ulations are embedded in trophic networks and environments, and co-evolutionary interactions would imply communities of populations from diverse species, as presented by Strauss et al. (2005) under the appellation of diffuse co-evolution. Similarly, spatio-temporal dynamics are crucial in the realization of these pro- cesses: Dybdahl and Lively (1996) study for example the influence of the spatial distribution on patterns of co-evolution for a snail and its parasite, and show that a higher speed of genetic diffusion in space for the parasite drive the co- evolutionary dynamics. The essential concepts to retain from the biological point of view are thus: (i) existence of evolution processes, in particular transmission and transformation; (ii) in circular schemas between populations in the case of co-evolution; and (iii) in a complex territorial frame (spatio-temporal and environmental in the sense of the rest of the ecosystem). 4 3.2 Cultural evolution This development on co-evolution was brought by the parallel between biological and social systems. The evolution of culture is theorized within a proper field, and witnesses many co-evolutive dynamics. Mesoudi (2017) recalls the state of knowledge on the subject and future issues, such as the relation with the cumulative nature of culture, the influence of demography in evolution processes, or the construction of phylogenetic methods allowing to reconstruct branches of past evolutionary trees. To give an example, Carrignon et al. (2015) introduces a conceptual frame for the co-evolution of culture and commerce in the case of ancient societies for which there are archeological data, and proposes its implementation with a multi-agent model which dynamics are partly validated by the study of stylized facts produced by the model. The co-evolution is here indeed taken in the sense of a mutual adaptation of socio-spatial structures, at comparable time scales, in this more general frame of cultural evolution. Cultural evolution would even be indissociable from genetic evolution, since Durham (1991) postulates and illustrates a strong link between the two, that would themselves be in co-evolution. Bull et al. (2000) explore a stylized model including two types of replicant populations (genes and memes) and shows the existence of phase transitions for the results of the genetic evolution process when the interaction with the cultural replicant is strong. 3.3 Sociology The concept was used in sociology and related disciplines such as organisation studies, following the parallel done before the same way as cultural evolution. In the field of the study of organisations, Volberda and Lewin (2003) develop a conceptual frame of inter-organisational co-evolution in relation with internal management processes, but deplore the absence of empirical studies aiming at quantifying this co-evolution. In the context of production systems manage- ment, Tolio et al. (2010) conceptualize an intelligent production chain where product, process and the production system must be in co-evolution. 3.4 Economic geography In economic geography, the concept of co-evolution has also largely been used. The idea of evolutionary entities in economy comes in opposition to the neo- classical current which remains a majority, but finds a more and more relevant echo (Nelson and Winter, 2009). Schamp (2010) proceeds to an epistemological analysis of the use of co-evolution, and opposes the view of a neo-schumpeterian approach to economy which considers the emergence of populations that evolve from micro-economic rules (what would correspond to a direct and relatively isolationist reading of biological evolution) to a systemic approach that would consider the economy as an evolutive system in a global perspective (what would correspond to diffuse co-evolution that we previously developed), to propose a 5 precise characterization that would correspond to the first case, assuming co- evolving institutions. The most important for our purpose is that he underlines the crucial aspect of the choice of populations and of considered entities, of the geographical area, and highlights the importance of the existence of causal circular relations. Diverse examples of application can be given. Colletis (2010) introduces a framework for the co-evolution of territories and technology (questioning for ex- ample the role of proximity on innovations), that reveals again the importance of the institutional aspect. The framework proposed by Ter Wal and Boschma (2011) couples the evolutionary approach to companies, the literature on indus- tries and innovation in clusters, and the approach through complex networks of connexions between the latest in the territorial system. In environmental economics, Kallis (2007) show that "broad" approaches (that can consider most of co-dynamics as co-evolutive) are opposed to stricter approaches (in the spirit of the definition given by Schamp (2010)). 3.5 Geography In geography, the stream of research which is the closest empirically and the- oretically to notions of co-evolution corresponds to the evolutive urban theory. The concept of co-evolution was present since the foundations of the theory (Pumain, 1997): the system of cities as a complex adaptive system is composed of subsystems that are interdependent in a complex way, often with circular causalities. The corrsponding models include this vision in an implicit way, but co-evolution is not explicitly highlighted nor precisely defined, in terms that would be quantifiable or structurally identifiable. Paulus (2004) gives empiri- cal proofs of mechanisms of co-evolution through the study of the evolution of economic profiles of French cities. The interpretation used by Schmitt (2014) is based on an entry by the evolutive urban theory, and fundamentally consists in a reading of systems of cities as highly interdependent entities. 3.6 Physics Finally, we can mention that the term of co-evolution is also used in physics. Its use for physical systems may induce some debates, depending if we sup- pose or not that the transmission assumes a transmission of information 1. In the case of a purely physical ontological transmission (physical beings), then a large part of physical systems are evolutive. Hopkins et al. (2008) develop a cosmological frame for the co-evolution of cosmic heterogenous objects which presence and dynamics are difficultly explained by more classical theories (some types of galaxies, quasars, supermassive black holes). Antonioni and Cardillo 1Information is defined within the shanonian theory as an occurence probability for a chain of characters. Morin (1976) shows that the concept of information must indeed be thought conjointly to a given context of the generation of a self-organizing negentropic system, i.e. realizing local decreases in entropy in particular thanks to this information. This type of system is necessarily alive. We consider here this complex approach to information. 6 (2017) study the co-evolution between synchronization and cooperation proper- ties within a Kuramoto oscillators network. 3.7 A synthetic definition of co-evolution Most of these approaches fit in the theory of complex adaptive systems de- veloped by Holland (2012): it takes any system as an imbrication of systems of boundaries, that filter signals or objects. Within a given limit, the corre- sponding subsystem is relatively autonomous from the outside, and is called an ecological niche, in a direct correspondence with highly connected communities within trophic or ecological networks. This way, interdependent entities within a niche are said to be co-evolving. We will come back on that approach below, when establishing the link with morphogenesis. We retain from this multidisciplinary view of co-evolution the following im- portant points: 1. The presence of evolution processes is primary, and their definition is al- most always based on the existence of transmission and transformation processes. 2. Co-evolution assumes entities or systems, belonging to distinct classes, which evolutive dynamics are coupled in a circular causal way. Approaches can differ depending on the assumptions of populations of these entities, singular objects, or components of a global system then in mutual inter- dependency without a direct circularity. 3. The delineation of systems and subsystems, both in the ontological space (definition of studied objects), but also in space and time, and their distri- bution in these spaces, is fundamental for the existence of co-evolutionary dynamics, and it seems in a large number of cases, of their empirical char- acterization. We propose therefore a synthetic entry for the specific case of transporta- tion networks and territories, echoing these points. It verifies the three following specifications. First of all, evolutive processes correspond to transformations of components of the system at the different scales: for the example of territorial systems transformation of cities on the long time, of their networks, trans- mission between cities of socio-economic characteristics carries by microscopic agents but also cultural transmission, reproduction and transformation of agents themselves (firms, households, operators)2. These evolutive processes may imply a co-evolution. Within a territorial system, can simultaneously co-evolve: (i) given entities (a given infrastructure and given characteristics of a given territory for example, i.e. individuals), when their mutual influence will be circularly causal (at the corresponding scale); (ii) 2This list is based on assumptions of the evolutive urban theory developed by Pumain (2008). It can not be exhaustive, since what would be the "ADN of a city" remains an open question. 7 populations of entities, what will be translated for example as such type of in- frastructure and given territorial components co-evolve at a statistical level in a given geographical region; (iii) all the components of a system when there exists strong global interdependencies at a large scale. Our approach is thus funda- mentally multi-scale and articulates different significations at different scales. Finally, the constraint of an isolation implies, in relation with the previous point, that co-evolution and the articulation of significations will have a meaning if there exists spatio-temporal isolations of subsystems in which different co- evolutions operate. 4 Morphogenesis Morphogenesis is at the very heart of concepts in ME. We have furthermore seen that it naturally arises when comparing biological and social systems. We will precise now its meaning and unveil its link with co-evolution. 4.1 Definition of morphogenesis The morphogenesis of a system consists in evolution rules that produce the emer- gence of its successives states and its form. It corresponds to self-organization, with the additional property that an emergent architecture exists (Doursat et al., 2012), in the sense of causal circular relations between the form and the function. Morphogenesis can be tracked back to first contributions in biology such as (Thompson, 1942), but the first interdisciplinary perspective can be with few doubts attributed to Turing (1952) which introduced a reaction-diffusion model to explain the growth of the embryo. Progresses towards the understanding of embryo morphogenesis (in particular the isolation of particular processes pro- ducing the differentiation of cells from an unique cell) have been made only re- cently with the use of complexity approaches in integrative biology (Delile et al., 2016). To show how morphogenesis applies to social systems, we can consider the ex- ample of urban systems. The idea of urban morphogenesis, i.e. of self-consistent mechanisms that would produce the urban form, is used in the field of architec- ture and urban design (as for example the generative grammar of "Pattern Lan- guage" of Alexander (1977)), in relation with theories of urban form (Moudon, 1997). This idea can be pushed to small spatial extents such as the one of the building (Whitehand et al., 1999) but it is in our case more relevant to consider it at a mesoscopic scale, in terms of land-use changes within an intermediate scale of territorial systems, with similar ontologies as the urban morphogenesis modeling literature. For example Bonin et al. (2012) describes a model of ur- ban morphogenesis with qualitative differentiation, whereas Makse et al. (1998) give a model of urban growth based on a mono-centric population distribution perturbed with correlated noises. Raimbault et al. (2014) show with a model of co-evolution between transportation networks and housing how the interplay 8 between the urban form and functions plays a determining role in the emergent behavior of the urban system. The concept of morphogenesis is important in our approach in link with modularity and scale. Modularity of a complex system consists in its decom- position into relatively independent sub-modules, and the modular decompo- sition of a system can be seen as a way to disentangle non-intrinsic correla- tions (Kolchinsky et al., 2015) (what can be thought in analogy to a block di- agonalisation of a first order dynamical system). In the context of large-scale cyber-physical systems design and control, similar issues naturally raise and specific techniques are needed to scale up simple control methods (Wang et al., 2017). The isolation of a subsystem yields a corresponding characteristic scale. Isolating possible morphogenesis processes implies a controlled extraction (con- trolled boundary conditions e.g.) of the considered system, corresponding to a modularity level and thus a scale. When local processes are not enough to explain the evolution of a system (with reasonable variations of initial condi- tions), a change in scale is necessary, caused by an underlying phase transition in modularity. Finally, it is important to remark that a subsystem for which morphogenesis makes sense, which boundaries are well defined and which processes allow it to maintain itself as a network of processes, is close to an auto-poietic system in the extended sense of Bourgine and Stewart (2004)3. These systems regulate then their boundary conditions, what underlines the importance of boundaries that we will finally develop. 4.2 Linking morphogenesis and co-evolution We can now introduce the link between morphogenesis and co-evolution. It is build on Holland's approach to complex adaptive systems (CAS) with his theory of CAS as agents which fundamental property is to process signals through to their boundaries (Holland, 2012). In this theory, complex adaptive systems form aggregates at diverse hierarchical levels, which correspond to different level of self-organization, and boundaries are vertically and horizontally entangled in a complex way. That approach introduces the notion of niche as a relatively inde- pendent subsystem in which resources circulate (the same way as communities in a network): numerous illustrations such as economical niches or ecological niches can be given. Agents within a niche are then said to be co-evolving. This notion of co-evolution coincides with the definition we gave, in partic- ular at the level of a population of entities. The decomposition of a system in niches relates then to the modular decomposition that came in the morpho- genesis approach. We postulate therefore the following fundamental conceptual link: the presence of morphogenetic processes in a system is equivalent to a de- composition in niches in which its components are co-evolving. In other words, 3Which are however not cognitive, making these morphogenetic systems not alive in the sense of auto-poietic and cognitive. Given the difficulty to define the delineation of cities for example, we will leave open the issue of the existence of auto-poietic territorial systems, and will consider in the following a less restrictive point of view on boundaries. 9 the properties of a morphogenetic system are essentially determined by the co- evolutive processes between its elements. This proposition does not fundamentally changes the way to consider mor- phogenetic engineering, but gives a complementary entry, with strong basis in biology and other disciplines. We believe that this epistemological loop demon- strates the relevance of ME to study and design such kind of systems, since it exhibits a correspondence between the core concept of morphogenesis and an approach that is essentially multidisciplinary. This legitimates an extent of the scope of ME, as we will finally illustrate for the case of planning of territorial systems. 5 An example of application: planning of terri- torial systems as morphogenetic engineering Territorial and urban planning correspond to a well established tradition of top- down approach to the design of systems. It has been shown by Hall (1982) that this approach is far from being optimal both in terms of correspondence to the expected evolution of the corresponding territorial system, but also in terms of performance indicators. On the other hand, Jacobs (1961) early suggested that bottom-up processes and neighborhood effects were crucial for the well-being of a city, close to seeing it as a morphogenetic living organism. Planning practices remain however highly architectured and hierarchical. Moreover, the evolutive urban theory (Pumain, 1997) suggests that terri- torial systems are mostly composed of co-evolutive entities. In particular, the interactions and strong interdependencies between cities would be significant drivers of their growth processes (Bretagnolle and Pumain, 2010), through the hierarchical diffusion of innovation which cycles are furthermore a potential explanation for urban scaling laws (Pumain et al., 2006). This type of system integrates co-evolution niches both as system components (the set of co-evolving agents) but also as the spatial support of the system (the regional territories that are relatively modular one to another). The importance of co-evolutive processes in territorial systems and the con- ceptual link we established above suggest the relevance of applying morphogene- sis to territorial systems: considering the appropriated co-evolutive niches, what would be equivalent to the appropriated morphogenetic subsystems, a problem of territorial planning can be mapped to a problem of ME. It would allow to answer questions such as how to obtain a desired territorial functionality from the bottom-up, by local actions on the behavior of agents or local interventions on the urban environment, what is in fact the exact logic of ME applied to the planning of territorial systems. We claim that this transfer could not be made directly, as the notion of morphogenesis of territorial systems is fuzzy and ontologies, scales and boundaries of the considered system would be difficult to consider without the conceptual entry of co-evolution. There naturally remain unanswered questions on the practical implemen- 10 tation of such a bottom-up ME approach to territorial planning, since similar planning approaches applying complex systems paradigms such as fractal cities by Frankhauser et al. (2018) still keep a high level of top-down control. These questions should be the object of future work that can be investigated for ex- ample through modeling. This illustration shows how this conceptual extension of ME by linking mor- phogenesis and co-evolution allows the application of ME to new types of sys- tems, in particular territorial systems that are a given case of socio-technical systems. We believe broader applications should be found, since an action on the theoretical domain of knowledge (here a conceptual development) must im- pact the empirical domain (here the fields of application) because of their close relation (Raimbault, 2017). 6 Conclusion In conclusion, a parallel between biological and social systems done in the spirit of extending the scope of ME has lead us to a multidisciplinary approach to co-evolution. We then linked it with morphogenesis, suggesting an entry to morphogenetic systems from the point of view of co-evolution. The application to the planning of territorial systems as morphogenetic design shows the poten- tialities of this conceptual extension. Our contribution is significant since it is to the best of our knowledge the first time that (i) an approach to co-evolution embracing a broad spectrum of disciplines is introduced; (ii) an explicit link between morphogenesis and co-evolution is proposed. References Alexander, C. (1977). A pattern language: towns, buildings, construction. Oxford university press. Angeletti, T. and Berlan, A. (2015). Les etres collectifs en question. Trac´es. Revue de Sciences humaines, (29):7 -- 22. Antonioni, A. and Cardillo, A. (2017). Coevolution of synchronization and cooperation in costly networked interactions. Physical Review Letters, 118(23):238301. Banos, A. (2013). Pour des pratiques de mod´elisation et de simulation lib´er´ees en g´eographies et shs. HDR. Universit´e Paris, 1. Batty, M. (2007). Cities and complexity: understanding cities with cellular automata, agent- based models, and fractals. The MIT press. Bettencourt, L. M., Lobo, J., Helbing, D., Kuhnert, C., and West, G. B. (2007). Growth, innovation, scaling, and the pace of life in cities. Proceedings of the national academy of sciences, 104(17):7301 -- 7306. Bonin, O., Hubert, J.-P., et al. (2012). Mod`ele de morphog´en`ese urbaine: simulation d'espaces qualitativement diff´erenci´es dans le cadre du mod`ele de l'´economie urbaine. In 49`e colloque de l'ASRDLF. Bourgine, P. and Stewart, J. (2004). Autopoiesis and cognition. Artificial life, 10(3):327 -- 345. 11 Bretagnolle, A. and Pumain, D. (2010). Simulating urban networks through multiscalar space-time dynamics: Europe and the united states, 17th-20th centuries. Urban Studies, 47(13):2819 -- 2839. Bull, L., Holland, O., and Blackmore, S. (2000). On meme -- gene coevolution. Artificial life, 6(3):227 -- 235. Carrignon, S., Montanier, J.-M., and Rubio-Campillo, X. (2015). Modelling the co-evolution In Winter Simulation Conference (WSC), 2015, of trade and culture in past societies. pages 3949 -- 3960. IEEE. Colletis, G. (2010). Co-´evolution des territoires et de la technologie: une perspective institu- tionnaliste. Revue d' ´Economie R´egionale & Urbaine, (2):235 -- 249. Delile, J., Doursat, R., and Peyri´eras, N. (2016). Mod´elisation multi-agent de l'embryogen`ese animale. Mod´elisations, simulations, syst`emes complexes, pages 581 -- 624. Doursat, R. (2011). The myriads of alife: Importing complex systems and self-organization into engineering. In Artificial Life (ALIFE), 2011 IEEE Symposium on, pages 1 -- 8. IEEE. Doursat, R., Sayama, H., and Michel, O. (2012). Morphogenetic engineering: toward pro- grammable complex systems. Springer. Doursat, R., Sayama, H., and Michel, O. (2013). A review of morphogenetic engineering. Natural Computing, 12(4):517 -- 535. Durham, W. H. (1991). Coevolution: Genes, culture, and human diversity. Stanford Univer- sity Press. Dybdahl, M. F. and Lively, C. M. (1996). The geography of coevolution: comparative popu- lation structures for a snail and its trematode parasite. Evolution, 50(6):2264 -- 2275. Frankhauser, P., Tannier, C., Vuidel, G., and Houot, H. (2018). An integrated multifractal modelling to urban and regional planning. Computers, Environment and Urban Systems, 67:132 -- 146. Gabora, L. and Steel, M. (2017). Autocatalytic networks in cognition and the origin of culture. arXiv preprint arXiv:1703.05917. Hall, P. (1982). Great Planning Disasters: With a New Introduction, volume 1. Univ of California Press. Holland, J. H. (2012). Signals and boundaries: Building blocks for complex adaptive systems. MIT Press. Hopkins, P. F., Hernquist, L., Cox, T. J., and Keres, D. (2008). A cosmological framework for the co-evolution of quasars, supermassive black holes, and elliptical galaxies. i. galaxy mergers and quasar activity. The Astrophysical Journal Supplement Series, 175(2):356. Jacobs, J. (1961). The death and life of great american cities. New-York, NY: Vintage. Janzen, D. H. (1980). When is it coevolution. Evolution, 34(3):611 -- 612. Kallis, G. (2007). When is it coevolution? Ecological Economics, 62(1):1 -- 6. Kolchinsky, A., Gates, A. J., and Rocha, L. M. (2015). Modularity and the spread of pertur- bations in complex dynamical systems. Physical Review E, 92(6):060801. Makse, H. A., Andrade, J. S., Batty, M., Havlin, S., Stanley, H. E., et al. (1998). Modeling urban growth patterns with correlated percolation. Physical Review E, 58(6):7054. 12 Mesoudi, A. (2017). Pursuing darwin's curious parallel: Prospects for a science of cultural evolution. Proceedings of the National Academy of Sciences, 114(30):7853 -- 7860. Monod, J. (1970). Le hasard et la N´ecessit´e. Points. Morin, E. (1976). La M´ethode, tome 1. la nature de la nature. Le Seuil. Morin, E. (1980). La M´ethode, tome 2. La Vie de la Vie. Le Seuil. Moudon, A. V. (1997). Urban morphology as an emerging interdisciplinary field. Urban morphology, 1(1):3 -- 10. Nelson, R. R. and Winter, S. G. (2009). An evolutionary theory of economic change. Harvard University Press. Paulus, F. (2004). Co´evolution dans les syst`emes de villes: croissance et sp´ecialisation des aires urbaines fran¸caises de 1950 `a 2000. PhD thesis, Universit´e Panth´eon-Sorbonne-Paris I. Pumain, D. (1997). Pour une th´eorie ´evolutive des villes. Espace g´eographique, 26(2):119 -- 134. Pumain, D. (2008). The socio-spatial dynamics of systems of cities and innovation processes: a multi-level model. The Dynamics of Complex Urban Systems, pages 373 -- 389. Pumain, D., Paulus, F., Vacchiani-Marcuzzo, C., and Lobo, J. (2006). An evolutionary theory for interpreting urban scaling laws. Cybergeo: European Journal of Geography. Raimbault, J. (2017). An applied knowledge framework to study complex systems. In Complex Systems Design & Management, pages 31 -- 45. Raimbault, J., Banos, A., and Doursat, R. (2014). A hybrid network/grid model of urban In 4th International Conference on Complex Systems morphogenesis and optimization. and Applications, pages 51 -- 60. Roth, C. (2009). Reconstruction failures: Questioning level design. In Epistemological Aspects of Computer Simulation in the Social Sciences, pages 89 -- 98. Springer. Sayama, H. (2009). Swarm chemistry. Artificial life, 15(1):105 -- 114. Sayama, H., Dionne, S., Laramee, C., and Wilson, D. S. (2009). Enhancing the architecture of interactive evolutionary design for exploring heterogeneous particle swarm dynamics: An in-class experiment. In Artificial Life, 2009. ALife'09. IEEE Symposium on, pages 85 -- 91. IEEE. Schamp, E. W. (2010). On the notion of co-evolution in economic geography. The handbook of evolutionary economic geography, page 432. Schmitt, C. (2014). Mod´elisation de la dynamique des syst`emes de peuplement: de SimpopLo- cal `a SimpopNet. PhD thesis, Paris 1. Strauss, S. Y., Sahli, H., and Conner, J. K. (2005). Toward a more trait-centered approach to diffuse (co) evolution. New Phytologist, 165(1):81 -- 90. Ter Wal, A. L. and Boschma, R. (2011). Co-evolution of firms, industries and networks in space. Regional studies, 45(7):919 -- 933. Thompson, D. W. (1942). On growth and form. Cambridge University Press. Tolio, T., Ceglarek, D., ElMaraghy, H., Fischer, A., Hu, S., Laperri`ere, L., Newman, S. T., and V´ancza, J. (2010). Species -- co-evolution of products, processes and production systems. CIRP Annals-Manufacturing Technology, 59(2):672 -- 693. 13 Turing, A. M. (1952). The chemical basis of morphogenesis. Philosophical Transactions of the Royal Society of London, 237:1 -- 37. van der Leeuw, S., Lane, D., and Read, D. (2009). The Long-Term Evolution of Social Organization, pages 85 -- 116. Springer Netherlands, Dordrecht. Volberda, H. W. and Lewin, A. Y. (2003). Co-evolutionary dynamics within and between firms: From evolution to co-evolution. Journal of management studies, 40(8):2111 -- 2136. Wang, Y.-S., Matni, N., and Doyle, J. C. (2017). Separable and Localized System Level Synthesis for Large-Scale Systems. arXiv preprint arXiv:1701.05880. Whitehand, J., Morton, N., and Carr, C. (1999). Urban morphogenesis at the microscale: how houses change. Environment and Planning B: Planning and Design, 26(4):503 -- 515. 14
1502.07601	2	1502	2015-03-03T19:15:42	Data Driven Validation Framework for Multi-agent Activity-based Models	[ "cs.MA" ]	Activity-based models, as a specific instance of agent-based models, deal with agents that structure their activity in terms of (daily) activity schedules. An activity schedule consists of a sequence of activity instances, each with its assigned start time, duration and location, together with transport modes used for travel between subsequent activity locations. A critical step in the development of simulation models is validation. Despite the growing importance of activity-based models in modelling transport and mobility, there has been so far no work focusing specifically on statistical validation of such models. In this paper, we propose a six-step Validation Framework for Activity-based Models (VALFRAM) that allows exploiting historical real-world data to assess the validity of activity-based models. The framework compares temporal and spatial properties and the structure of activity schedules against real-world travel diaries and origin-destination matrices. We confirm the usefulness of the framework on three real-world activity-based transport models.	cs.MA	cs	Data Driven Validation Framework for Multi-agent Activity-based Models Jan Drchal, Michal Certick´y, and Michal Jakob Faculty of Electrical Engineering Czech Technical University in Prague [email protected] {certicky,jakob}@agents.fel.cvut.cz Abstract. Activity-based models, as a specific instance of agent-based models, deal with agents that structure their activity in terms of (daily) activity schedules. An activity schedule consists of a sequence of activ- ity instances, each with its assigned start time, duration and location, together with transport modes used for travel between subsequent ac- tivity locations. A critical step in the development of simulation models is validation. Despite the growing importance of activity-based models in modelling transport and mobility, there has been so far no work fo- cusing specifically on statistical validation of such models. In this paper, we propose a six-step Validation Framework for Activity-based Models (VALFRAM) that allows exploiting historical real-world data to assess the validity of activity-based models. The framework compares tempo- ral and spatial properties and the structure of activity schedules against real-world travel diaries and origin-destination matrices. We confirm the usefulness of the framework on three real-world activity-based transport models. 1 Introduction Transport and mobility have recently become a prominent application area for multi-agent systems and agent-based modelling [Chen and Cheng, 2010]. Models of transport systems offer an objective common ground for discussing policies and compromises [de Dios Ort´uzar and Willumsen, 2011], help to understand the underlying behaviour of these systems and aid in the actual decision making and transport planning. Large-scale, complex transport systems, set in various socio-demographic contexts and land-use configurations, are often modelled by simulating the be- haviour and interactions of millions of autonomous, self-interested agents. Agent- based modelling paradigm generally provides a high level of detail and allows representing non-linear patterns and phenomena beyond traditional analytical approaches [Bonabeau, 2002]. Specific subclass of agent-based models, called activity-based models, address particularly the need for realistic representation of travel demand and transport-related behaviour. Unlike traditional trip-based models, activity-based models view travel demand as a consequence of agent's 2 needs to pursue various activities distributed in space and understanding of travel decisions is secondary to a fundamental understanding of activity be- haviour [Jones et al., 1990]. Gradual methodological shift towards such a behaviourally-oriented model- ling paradigm is evident. An early work on the topic is represented by the CARLA model, developed as part of the first comprehensive assessment of behaviourally-oriented approach at Oxford [Jones et al., 1983]. Later work is represented by the SCHEDULER model -- a cognitive architecture producing activity schedules from long- and short-term calendars and perceptual rules [Garling et al., 1994], TRANSIMS -- an integrated system of travel forecasting models, including activity scheduler [Smith et al., 1995], or ALBATROSS -- the first model of complete activity scheduling process automatically estimated from data [Arentze and Timmermans, 2000]. In order to produce dependable and useful results, the model needs to be valid1 enough. In fact, validity is often considered the most important property of models [Klugl, 2009]. The process of quantifying the model validity by deter- mining whether the model is an accurate representation of the studied system is called validation and the validation process needs to be done thoroughly and throughout all phases of model development [Law, 2009]. Despite the growing adoption of activity-based models and the generally acknowledged importance of model validation, a validation process for activity- based models in particular has not yet been standardized by a detailed method- ological framework. Validation techniques and guidelines are addressed in most modelling textbooks [Balci, 1994,Law, 2007] and have even been instantiated in the form of a validation process for general agent-based models [Klugl, 2009]; however, such techniques are still too general to provide concrete, practical methodology for the key validation step: statistical validation against real-world data. In this paper, we address this gap and propose a validation framework en- titled VALFRAM (Validation Framework for Activity-based Models), designed specifically for statistically quantifying the validity of activity-based transport models. The framework relies on the real-world transport behaviour data and quantifies the model validity in terms of clearly defined validation metrics. We illustrate and demonstrate the framework on several activity-based transport models of a real-world region populated by approximately 1 million citizens. 2 Preliminaries 2.1 Activity-based Models Activity-based models [Ben-Akivai et al., 1996] are multiagent models in which the agents plan and execute so-called activity schedules -- finite sequences of ac- tivity instances interconnected by trips. Each activity instance needs to have a 1 Valid model is a model of sufficient accuracy (precision). We use these terms inter- changeably in the following text. 3 specific type (e.g. work, school or shop), location, desired start time and dura- tion. Trips between activity instances are specified by their main transport mode (e.g. car or public transport). 2.2 Validation Methods Validation methods in general are usually divided into two types: -- Face validation subsumes all methods that rely on natural human intelli- gence such as expert assessments of model visualizations. Face validation shows that model's behaviour and outcomes are reasonable and plausible within the frame of the theoretic basis and implicit knowledge of system experts or stake-holders. Face validation is in general incapable of produc- ing quantitative, comparable numeric results. Its basis in implicit expert knowledge and human intelligence also makes it difficult to standardize face validation in a formal methodological framework. In this paper, we therefore focus on statistical validation. -- Statistical validation (sometimes called empirical) employs statistical mea- sures and tests to compare key properties of the model with the data gathered from the modelled system (usually the original real-world system). From a higher-level perspective, VALFRAM can be viewed as an activity- based model-focused implementation of the statistical validation step of a more comprehensive validation procedure for generic agent-based models, introduced in [Klugl, 2009], as depicted in Figure 1. Besides the face and statistical valida- tion, this procedure features other complementary steps such as calibration and sensitivity analysis. Fig. 1: Higher-level validation procedure for agent-based models in general, in- troduced in [Klugl, 2009]. VALFRAM implements the statistical validation step specifically for activity-based models. Being set in the context of activity-based modelling, the VALFRAM frame- work is concerned with the specific properties of activity schedules generated by the agents within the model. These properties are compared to historical real-world data in order to compute a set of numeric similarity metrics. In case of activity-based models, implemented by VALFRAMConcept, Implementation, VerificationRun-able ModelFace ValidationSensitivity AnalysisStatistical ValidationNot validOver-simplificationUnsuccessfulNot plausibleNot successfulPlausible ModelMinimalModelCalibrationPlausibility CheckCalibratedModelDeployableModel4 3 VALFRAM description In this section a detailed description of VALFRAM is given. We cover validation data, validation objectives and finally measures defined by VALFRAM. 3.1 Data A requirement for statistical validation of any model is data capturing the re- levant aspects of the behaviour of modelled system, against which the model is validated. To validate an activity-based model, the VALFRAM framework requires two distinct data sets gathered in the modelled system: 1. Travel Diaries: Travel diaries are usually obtained by long-term surveys (tak- ing up to several days), during which participants log all their trips. The resulting data sets contain anonymized information about every participant (usually demographic attributes such as age, gender, etc.), and a collec- tion of all their trips with the following properties: time and date, duration, transport mode(s) and purpose (the activity type at the destination). More detailed travel diaries also contain the locations of the origin and the desti- nation of each trip. 2. Origin-Destination Matrix (O-D Matrix): The most basic O-D matrices (sometimes called trip tables) are simple two-dimensional square matrices displaying the number of trips travelled between every combination of origin and destination locations during a specified time period (e.g. one day or one hour). The origin and destination locations are usually predefined, mutually exclusive zones covering the area of interest and their size determines the level of detail of the matrix. In real-world systems, O-D matrices may be obtained by roadside monitoring, household surveys or derived from mobile phone networks [Caceres et al., 2007]. 3.2 VALFRAM Validation Objectives The VALFRAM validation framework is concerned with a couple of specific properties of activity schedules produced by modelled agents. These particular properties need to correlate with the modelled system in order for the model to accurately reproduce the system's transport-related behaviour. At the same time, these properties can actually be validated based on available data sets -- travel diaries and O-D matrices. In particular, we are interested in: A. Activities and their: 1. temporal properties (start times and durations), 2. spatial properties (distribution of activity locations in space), 3. structure of activity sequences (typical arrangement of successive activity types). B. Trips and their: 1. temporal properties (transport mode choice in different times of day; durations of trips), 2. spatial properties (distribution of trip's origin-destination pairs in space), 3. structure of transport mode choice (typical mode for each destination activity type). 5 3.3 VALFRAM Validation Metrics To validate these properties of interest, we need to perform six validation steps (A1, A2, A3, B1, B2, B3), as depicted in Table 1 and detailed in the rest of this section. In each validation step, we compute specific numeric metrics (statis- tics). For all metrics, higher values of these statistics indicate a larger difference between the model and validation set, i.e., lower accuracy. A.Activities B.Trips 1.Time 2.Space 3.Structure distribution Task Compare the distributions of start times and dura- tions for each activity type using Kolmogorov-Smirnov (KS) statistic. Compare of each activity type in 2D space using RMSE. Plot heat maps for additional feedback. i) Compare activity counts within activity schedules using χ2 statistics. ii) Com- pare distributions of activ- ity schedule subsequences as n-grams profiles using χ2 statistics. Data set Travel Diaries Task Compare the distribution of selected modes by time of day and the distribution of travel times by mode using χ2 and KS statistics. Data set Travel Diaries Space-aware Travel Diaries Compute the distance be- tween generated and real- world O-D matrix using RMSE. Origin- Destination Matrix Travel Diaries Compare the distribution of selected transport mode for each type of target activity type using χ2 statistics. Travel Diaries Table 1: Six validation steps of VALFRAM framework and corresponding vali- dation data sets needed for each of them. A1. Activities in Time: The comparison of activity distributions in time is realized by means of a well-established Kolmogorov-Smirnov two-sample statis- tic [Hollander et al., 2013]. VALFRAM applies the method to start time distri- butions p(startact. type) as well as to duration distributions p(durationact. type). The statistic is defined as the maximum deviation between the empirical cumulative distribution functions FM and FV which are based on the model and validation data distributions: dKS = supx FM (x)− FV (x). The values lie in the interval [0, 1]. Figure 2a shows an example application of the Kolmogorov-Smirnov statistic comparing two different models to validation data. A2. Activities in Space: The comparison of activity distributions in space is performed separately for every activity type. Unlike in the previous step, the distributions are two-dimensional (latitude, longitude or projected coordinates). The process consists of the following steps. First, bivariate empirical cumulative distribution functions (ECDFs) FM and FV are constructed using coordinate data for both model and validation data, respectively. Second, FM and FV are 6 (a) work activity start time (b) Modeled area, sleep activity Fig. 2: Start time distributions for work activity shown for validation data and two different models (a) including Kolmogorov-Smirnov statistics. Modelled area including sleep activity spatial PDF visualized as a heat map (b). regularly sampled getting matrices EM and EV both having m rows and n columns. Third, Root Mean Squared Error (RMSE) of the two matrices is com- ij ≤ 1, puted using decdf = the measure decdf is again limited to the [0, 1] interval. ij ≤ 1 and EV /(mn). As EM ij − EV (cid:0)EM (cid:80)n (cid:1)2 ij i=1 j=1 (cid:113)(cid:80)m Figure 2b shows the spatial probability distribution function (PDF) of sleep type activities on the validation set visualized as a heat map. The probability distribution was approximated from data using Gaussian kernels. Similar heat maps might be helpful when developing a model as they can show where problems or imprecisions are. A3. Structure of Activities: In the previous steps, we examined the activity distributions in time and space. In this step, we consider the activity composition of the entire activity schedules. We propose a measure which compares distribu- tions of activity counts in activity schedules as well as a measure comparing the distribution of possible activity type sequences. Activity Count: The comparison of activity counts in activity schedules is based on a well-known Pearson's chi-square test [Sokal and Rohlf, 1994]. The procedure is performed separately for each activity type. First, frequencies f M and f V for the count i are collected for both model and validation data. Va- i lidation data frequencies f V i and in turn validation frequencies sV scaled to match the sum of model's frequen- i i chi-square statistic is computed as i are then used to get count proportions pV and sV i f M i sV i cies ((cid:80) χ2 =(cid:80) i i = (cid:80) (cid:0)f M i − sV i (cid:1)2 i ). Using f M i /sV i . Activity Sequences: We also compare activity sequence distributions. The method is based on the well-established text mining techniques [Cavnar et al., 1994] [Manning, 1999]. Particularly, we compare n-gram profiles using chi-square statis- tic. N-gram is a continuous subsequence of the original sequence having a length exactly n. Consider an example activity schedule consisting of the following activ- ValidationModelAModelB051015200start(h)p(startwork)dKSA=0.56dKSB=0.227 ity sequence: (cid:104)none, sleep, work, leisure, sleep, none(cid:105)2. The set of all 2-grams (bigrams) is then: {(cid:104)none, sleep(cid:105),(cid:104)sleep, work(cid:105),(cid:104)work, leisure(cid:105),(cid:104)leisure, sleep(cid:105), (cid:104)sleep, none(cid:105)}. We create an n-gram profile by counting frequencies of all n-grams in a range n ∈ {1, 2,··· , k} for all activity schedules. All the N n-grams are then sorted by their counts in a decreasing order so that the counts are fi ≥ fj for any two n-grams i and j where 1 ≤ i < j ≤ N (for a tie fi = fj one should sort in the lexicographical order). We only work with a proportion P of n-grams having the highest count in the profile. More precisely, we take only the first M i=1 fi is true. n-grams, where M is the highest value for which(cid:80)M i=1 fi ≤ P(cid:80)N In order to compare n-gram profiles of model and validation data, we employ chi-square statistic matching both profiles by the corresponding n-grams (only n-grams found in both profiles are considered). B1. Trips in Time: The validation of trips in time consists of two sub-steps: a comparison of mode distributions for a given time of day and a comparison of travel time distributions for selected modes. Modes by Time of Day: The comparison of mode distributions for a given time of day, i.e., p(modetime range), is based on exactly the same approach which we used to compare activity counts (validation step A3): the χ2 statistic is computed for mode frequencies of trips starting in a selected time interval. We suggest computing χ2 statistic for twenty four one-hour intervals per day, although other partitionings are possible. Travel Times per Mode: Travel time distributions for modes p(travel timemode) are validated in the same way as activities in time (see validation step A1) using Kolmogorov-Smirnov statistic dKS. B2. Trips in Space: In order to validate trip distributions in space, we pro- pose a symmetrical dissimilarity measure based on O-D matrix comparison. The algorithm is realized in three consecutive steps. First, O-D matrices are rear- ranged to use a common set of origins and destinations. Second, both matrices are scaled to make trip counts comparable. Third, RMSE for all elements which have non zero trip count in either of the matrices is computed. The algorithm starts with two O-D matrices: model matrix M and validation matrix V . Each element Mij (or Vij) represents a count of trips between origin i and destination j. The positional information (i.e., latitude/longitude or other type of coordinates) is denoted mi, mj ∈ CM for model and similarly vi, vj ∈ CV for validation data where CM and CV are sets of all possible coordinates (e.g., all traffic network nodes). Note that in most practical cases CM (cid:54)= CV . As an example we can have precise GPS coordinates generated by the model, however, only approximate or aggregated trip locations from validation travel diaries. As we have to work with the same locations in order to compare the O-D matrices, we need to select a common set of coordinates C. In practice, this would be typically the validation data location set (C = CV ) while all locations from CM must be projected to 2 Note, that none activities are added to the beginning and end of the activity schedule in order to preserve information about initial/terminal activity. 8 it by replacing each mi by its closest counterpart in C. This might eventually lead to resizing of the O-D matrix M as more origins/destinations might get aggregated into a single row/column. In many cases the total number of trips in M and V can be vastly different. The second step of the algorithm scales both M and V to a total element sum ij and V (cid:48) of one: M(cid:48) now represents a relative traffic volume between origin i and destination j. . Each element of both M(cid:48) Mij(cid:80) (cid:80) and V (cid:48) ij = j Mij ij i Finally, we compute the O-D matrix distance using the following equation: i j Vij ij = Vij(cid:80) (cid:80) (cid:118)(cid:117)(cid:117)(cid:116) (cid:80) (cid:0)M(cid:48) (cid:80) (cid:12)(cid:12)(cid:8)(i, j) : M(cid:48) j i ij − V (cid:48) ij > 0 ∨ V (cid:48) (cid:1)2 ij > 0(cid:9)(cid:12)(cid:12) . ij dOD = (1) ij, V (cid:48) Note that the equation is RMSE computed over all origin-destination pairs which appear either in M(cid:48) ij or in both. We have decided to ignore the elements which are zero in both matrices as these might represent trips which might not be possible at all (i.e., not connected by the transport network). Possible values ij ≤ 1). of dOD lie in interval [0, 1] (the upper bound is given by M(cid:48) B3. Mode for Target Activity Type: The validation of the mode choice for target activity type is again based on χ2 statistic. Here, we collect counts per each mode for each target activity of choice. ij ≤ 1 and V (cid:48) 4 VALFRAM Evaluation In general, we expect a statistical validation framework to meet three key con- ditions: 1. The framework quantifies the precision of the validated models in a way which allows comparing model's accuracy in replicating different aspects of the beahviour of the modelled system. 2. Data required for validation are available. 3. Validation results produced by the framework correlate with the expectations based on expert insight and face validation. VALFRAM meets conditions 1 and 2 for activity-based models by explic- itly expressing the spatial, temporal and structural properties of activities and trips, using only travel diaries and O-D matrices. To evaluate it with respect to condition 3, we have built three different activity-based models, formulated hypotheses about them based on our expert insight and used VALFRAM to validate both of them. 4.1 Evaluation Models The first model, denoted MA (model A), is a rule-based model inspired by ALBA- TROSS3 [Arentze and Timmermans, 2000]. The second model, denoted MB, is a 3 Although we call MA the rule-based model, it estimates activity count, durations and occasionally start times using linear-regression models based on data. All other activity schedule properties are based on rules constructed using expert knowledge. 9 fully data-driven model based on Recurrent Neural Networks (RNNs). More spe- cifically, the model employs fully-connected Long-Short Term Memory (LSTM) units [Hochreiter and Schmidhuber, 1997] and several sets of softmax output units. Given the training dataset based on travel diaries, the model is trained to repetitively take current activity type and its end time as input in order to pro- duce a trip (including trip duration and main mode) and the following activity (defined by type and duration). As MB is currently unable to generate spatial component of the schedules (e.g., activity locations), VALFRAM steps A2 and B2 are evaluated on a predecessor of MA denoted M(cid:48) Auses a less sophisticated approach to select activity locations. A (model A(cid:48)). M(cid:48) All MA, MB and M(cid:48) A models were used to generate a sample of 100,000 ac- tivity schedules. Our validation set V contained approximately 1,800 schedules. Such a disproportion is typical in reality, since obtaining real-world data tends to be more costly than obtaining synthetic data from model. All the data used in this study cover a single workday. An overview of the modelled area is depicted in Figure 2b. In the following text we present five hypotheses based on our insight of mo- dels. Note that all VALFRAM steps A1 through B3 are performed in order to evaluate them. 4.2 Test Hypotheses Hypothesis 1: The rule-based model MA uses very simple linear classifier for decisions on activity start times, so it will likely perform worse than the RNN- based model in their assignment. On the other hand, the activity scheduler in MA performs schedule optimization, during which it adapts activity durations according to rules psychologically plausible. This should produce more realistic behaviour than the purely data-driven RNN model4. Step A1 of VALFRAM confirms the hypothesis. Figure 3a depicts the distri- butions p(startwork) for validation data V and models MA and MB. The values KS > dB dA KS indicate the higher precision of the RNN model, with the most sig- nificant difference in the case of work and school activities. On the other hand, Figure 3b shows that MA outperforms MB in terms of activity durations. Hypothesis 2: Activity sequences of real-world system tend to be harder to replicate using simple rule-based models than robust data-driven approaches. Results of the step A3 (activity counts) for all the activity types are shown in Table 2. The data-driven model MB outperforms MA with the exception of the leisure activity (which we later found to be insufficiently covered by the RNN training data). Note that both MA and MB give the same χ2 value for the sleep activity which is caused by the fact that both models generate daily schedules having strictly two sleep activities in the current setup. For the step A3 (activity sequences) we got the following results for both models using the proportion P = 0.9 and k = 11 (same as the longest sequence in data): 4 At least given the limited size of the RNN training dataset. 10 (a) start time (b) duration Fig. 3: An example of activity in time comparison. The values of dKS are shown for both models MA and MB. MB outperforms MA on start times while the situation is the opposite for durations. χ2 ≈ 8.4 × 105 for MA and χ2 ≈ 2.6 × 105 for MB showing superiority of the RNN model. Model MA MB sleep 21468.1 21468.1 work 2889.3 255.7 school leisure 1750.3 542.2 293.8 4625.7 shop 974.2 773.8 Table 2: Activity counts for selected activities (χ2 statistic). Model MB outper- forms model MA with the exception of the leisure activity type. Hypothesis 3: While rule-based model optimizes the whole daily activity plans, RNN-based model works sequentially and schedules new activity based only on the previous ones. Therefore, it will be less precise towards the end of the day. By a further analysis of step A3 (activity sequences), which involved the comparison of a set of n-grams having highest frequency difference, we have, indeed, found that the RNN model tends to be less precise towards the end of the generated activity sequence resulting in schedules not ended by the sleep activity in a number of cases. Moreover, Figure 4 shows a comparison of mode by time of day selection χ2 values (step B1) for MA and MB showing that although MB is initially more precise it eventually degrades and the rule-based model MA prevails. Hypothesis 4: Unlike the rule-based model, the RNN model has no access to trip-planning data (i.e., transport network, timetables) which will decrease its performance in selecting trip modes. KS = 0.37 < dB For the step B1 (travel times per mode) we got dA KS = 0.31 for car and dA KS = 0.43 for public transport modes. Results of the step B3 are summarized in Table 3 also supporting the superiority of MA in modelling mode selection. Hypothesis 5: Model M(cid:48) activity location selection. A will be inferior to MA as it uses an oversimplified KS = 0.22 < dB For the step A2 this is clearly demonstrated in Figure 5 by dA ecdf for the leisure and shop activities (only activity types affected by the algorithm ecdf < dA(cid:48) sleepworkschoolleisureshopModelAModelB00.51dKSsleepworkschoolleisureshop00.51dKS11 Fig. 4: Modes by the time of day. The figure shows a comparison of χ2 values for car and public transport modes for one hour intervals between 3:00 and 23:00. Model MA MB sleep 562 2875.2 work 1371.7 3437.9 school leisure 1120 12817.3 475.1 7286.2 shop 5 2507.3 Table 3: Transport mode selection for target activity type (χ2 statistic). Model MA outperforms model MB in four out of five activity types. OD = 3.7 × 10−4 < dA(cid:48) selecting activity locations). For the step B2 we get dA OD = 4.8 × 10−4 which again supports the hypothesised improvement of A over A(cid:48). Fig. 5: Activities in space: comparison of Model A to Model A(cid:48). M(cid:48) to MA for flexible activities (dA ecdf ) based on 18 × 31 ECDF matrices. ecdf < dA(cid:48) A is inferior 5 Conclusion We have introduced a detailed methodological framework for data-driven statis- tical validation of multiagent activity-based transport models. The VALFRAM framework compares activity-based models against real-world travel diaries and origin -- destination matrices data. The framework produces several validation metrics quantifying the temporal, spatial and structural validity of activity sched- ules generated by the model. These metrics can be used to assess the accuracy of the model, guide model development or compare the model accuracy to other models. We have applied VALFRAM to assess and compare the validity of three activity-based transport models of a real-world region comprising around 1 mil- lion inhabitants. In the test application, the framework correctly identified strong χ234567891011121314151617181920212223ModelAModelB01000200030004000intervalstartsleepworkschoolleisureshopModelAModelA'0.000.020.040.060.08decdf12 and weak aspects of each model, which confirmed the viability and usefulness of the framework. References Arentze and Timmermans, 2000. Arentze, T. and Timmermans, H. (2000). Albatross: a learning based transportation oriented simulation system. Eirass Eindhoven. Balci, 1994. Balci, O. (1994). Validation, verification, and testing techniques through- out the life cycle of a simulation study. Annals of operations research, 53(1):121 -- 173. Ben-Akivai et al., 1996. Ben-Akivai, M., Bowman, J. L., and Gopinath, D. (1996). Travel demand model system for the information era. Transportation, 23(3):241 -- 266. Bonabeau, 2002. Bonabeau, E. (2002). Agent-based modeling: Methods and tech- niques for simulating human systems. Proceedings of the National Academy of Sci- ences, 99(suppl 3):7280 -- 7287. Caceres et al., 2007. Caceres, N., Wideberg, J., and Benitez, F. (2007). Deriving origin destination data from a mobile phone network. IET, 1(1):15 -- 26. Cavnar et al., 1994. Cavnar, W. B., Trenkle, J. M., et al. (1994). N-gram-based text categorization. Ann Arbor MI, 48113(2):161 -- 175. Chen and Cheng, 2010. Chen, B. and Cheng, H. H. (2010). A review of the applica- tions of agent technology in traffic and transportation systems. IEEE Transactions on Intelligent Transportation Systems, 11(2):485 -- 497. de Dios Ort´uzar and Willumsen, 2011. de Dios Ort´uzar, J. and Willumsen, L. (2011). Modelling Transport. Wiley. Garling et al., 1994. Garling, T., Kwan, M.-p., and Golledge, R. G. (1994). Computational-process modelling of household activity scheduling. Transportation Research Part B: Methodological, 28(5):355 -- 364. Hochreiter and Schmidhuber, 1997. Hochreiter, S. and Schmidhuber, J. (1997). Long short-term memory. Neural computation, 9(8):1735 -- 1780. Hollander et al., 2013. Hollander, M., Wolfe, D. A., and Chicken, E. (2013). Nonpara- metric Statistical Methods. Wiley, 3rd edition. Jones et al., 1983. Jones, P. M., Dix, M. C., Clarke, M. I., and Heggie, I. G. (1983). Understanding travel behaviour. Number Monograph. Jones et al., 1990. Jones, P. M., Koppelman, F., and Orfeuil, J.-P. (1990). Activ- ity analysis: State-of-the-art and future directions. Developments in dynamic and activity-based approaches to travel analysis, pages 34 -- 55. Klugl, 2009. Klugl, F. (2009). Agent-based simulation engineering. PhD thesis, Habil- itation Thesis, University of Wurzburg. Law, 2007. Law, A. M. (2007). Simulation modeling and analysis, 4th edition. McGraw-Hill New York. Law, 2009. Law, A. M. (2009). How to build valid and credible simulation models. In Simulation Conference (WSC), Proceedings of the 2009 Winter, pages 24 -- 33. IEEE. Manning, 1999. Manning, C. D. (1999). Foundations of Statistical Natural Language Processing. MIT press. Smith et al., 1995. Smith, L., Beckman, R., Anson, D., Nagel, K., and Williams, M. E. (1995). Transims: Transportation analysis and simulation system. In Fifth National Conference on Transportation Planning Methods Applications-Volume II: A Com- pendium of Papers Based on a Conference Held in Seattle, Washington. Sokal and Rohlf, 1994. Sokal, R. R. and Rohlf, F. J. (1994). Biometry: The Principles and Practices of Statistics in Biological Research. W. H. Freeman, 3rd edition.
0808.3569	3	0808	2008-09-01T20:34:05	Offloading Cognition onto Cognitive Technology	[ "cs.MA", "cs.CL" ]	"Cognizing" (e.g., thinking, understanding, and knowing) is a mental state. Systems without mental states, such as cognitive technology, can sometimes contribute to human cognition, but that does not make them cognizers. Cognizers can offload some of their cognitive functions onto cognitive technology, thereby extending their performance capacity beyond the limits of their own brain power. Language itself is a form of cognitive technology that allows cognizers to offload some of their cognitive functions onto the brains of other cognizers. Language also extends cognizers' individual and joint performance powers, distributing the load through interactive and collaborative cognition. Reading, writing, print, telecommunications and computing further extend cognizers' capacities. And now the web, with its network of cognizers, digital databases and software agents, all accessible anytime, anywhere, has become our 'Cognitive Commons,' in which distributed cognizers and cognitive technology can interoperate globally with a speed, scope and degree of interactivity inconceivable through local individual cognition alone. And as with language, the cognitive tool par excellence, such technological changes are not merely instrumental and quantitative: they can have profound effects on how we think and encode information, on how we communicate with one another, on our mental states, and on our very nature.	cs.MA	cs	Offloading Cognition onto Cognitive Technology / Page 1 To Appear in: Itiel E. Dror & Stevan Harnad (Eds) Cognition Distributed: How Cognitive Technology Extends Our Minds. Amsterdam: John Benjamins. Offloading Cognition onto Cognitive Technology Itiel E. Dror School of Psychology University of Southampton Southampton , United Kingdom SO171BJ Stevan Harnad Chaire de recherché de Canada Institut des sciences cognitives Université du Québec à Montréal Montréal, Québec, Canada H3C3P8 & School of Electronics and Computer Science University of Southampton Southampton , United Kingdom SO171BJ ABSTRACT: "Cognizing" (e.g., thinking, understanding, and knowing) is a mental state. Systems without mental states, such as cognitive technology, can sometimes contribute to human cognition, but that does not make them cognizers. Cognizers can offload some of their cognitive functions onto their thereby extending technology, cognitive performance capacity beyond the limits of their own brain power. Language itself is a form of cognitive technology that allows cognizers to offload some of their cognitive functions onto the brains of other cognizers. Language also extends cognizers' individual and joint performance powers, distributing the load through interactive and collaborative cognition. Reading, writing, print, telecommunications and computing further extend cognizers' capacities. And now the web, with its network of cognizers, digital databases and software agents, all accessible anytime, anywhere, has become our “Cognitive Commons,” in which distributed cognizers and cognitive technology can interoperate globally with a speed, scope and degree of interactivity inconceivable through local individual cognition alone. And as with language, the cognitive tool par excellence, such technological changes are not merely instrumental and quantitative: they can have profound effects on how we Offloading Cognition onto Cognitive Technology / Page 2 think and encode information, on how we communicate with one another, on our mental states, and on our very nature. Introductory Overview: With the development and wide use of cognitive technologies (Dror, 2007; Dascal & Dror, 2005), questions arise as to their effects on their human users and society, as well as on their own scope and limits: Can cognitive technologies (i) increase cognitive capacities and thus enhance human efficiency? (ii) affect how people and society go about achieving their goals? (iii) highlight and transform how we view ourselves and our goals? (iv) modify how we cognize and thus change our mental states and nature? (v) give rise to new forms of cognition (such as distributed cognition) and mental states that are either distributed across or even embodied in cognitive technology? These issues are examined as follows: (1) The notion of an "extended mind" -- with mental states (i.e., felt states) “distributed” beyond the narrow bounds of the individual brain – is not only as improbable as the notion that the US government can have a distributed migraine headache, but arbitrary. (2) "Cognition" -- if it is simply defined as the ability to do the kinds of things that cognizers like us can do, plus the underlying functional mechanisms for doing them -- can be arbitrarily defined to be as wide or as narrow as we like . (3) Vagueness about the nature, locus and scope of cognizing leads to a dissociation of “cognitive states” from mental states. However, their co- occurrence had been our only basis for distinguishing cognitive performance capacity from other capacities and functionality (animate or inanimate, narrow or wide). (4) If cognitive states are indeed not mental states, it follows that "cognitive technology" is not just something used by cognizers, but a functional part of the cognitive states themselves, because the boundary between user and tool disappears, and cognitive states become merely instances of functional states in general. (5) We then do not need the terms "cognitive" and "distributed cognition" at all, and can just talk about relatively complex and wide or narrow functional states, leaving it a coincidence and mystery (at least at this stage) that every single case of what we used to call “cognitive” also happened to be mental. (6) A way to resolve this is to accept that only mental states are cognitive states, that cognition is only narrow, and that the only place it is "distributed" is within a single cognizer's brain. Offloading Cognition onto Cognitive Technology / Page 3 (7) The only kind of “technology” that might really turn out to be intrinsically cognitive, rather than just being a tool used by cognizers, would be a robot that could pass the Turing Test (TT) -- because such a TT-scale robot would almost certainly have mental states, and hence it would be a cognizer in its own right. (8) Whatever distributed activity was going on within the functional mechanism generating such a TT robot’s performance capacity would then indeed be a case of distributed cognition (exactly as the distributed activity within our own brains is distributed cognition) – even if not all the components of its generating mechanism were located inside the robot’s head. (9) The “cognitive technology” used by such a TT robot, however, would still not be part of its distributed cognitive (hence mental) state, just as it is not a part of ours. (10) Nor would a group of such TT robots, interacting and collaborating, be a case of distributed cognition; it would merely be a case of collaborative cognition among individual (narrow) TT-robot cognizers, just as it is in the case of a group of collaborating human cognizers. (11) Cognitive technology does, however, extend the scope and power of cognition, exactly as sensory and motor technology extends the scope and power of the bodily senses and movement. (12) Just as we can see further with telescopes, move faster with cars, and do more with laser microsurgery than we can do with just our unaided hands and heads, so we can think faster and further, and do more, with language, books, calculators, computers, the web, algorithms, software agents, plus whatever is in the heads of other cognizers. (13) Both sensorimotor technology and cognitive technology extend our bodies’ and brains’ performance capacities as well as giving us the feeling of being able to do more than just our bodies and brains alone can do. (14) Sensorimotor and cognitive technology can thus generate a perceptual change, rather like virtual reality (VR), making us feel a difference in our body image and causal power (perhaps not unlike what the physical metamorphosis from caterpillar to butterfly might feel like, as one sensed one’s newfound somatic capacity to fly). (15) This change in perceived body image is indeed a change in mental state; but although its distal inputs and outputs certainly extend wider than the body (as all sensory inputs and all motor outputs do), the functional mechanism of that altered mental state is still just proximal -- skin and in – exactly as when it is induced by VR technology. Offloading Cognition onto Cognitive Technology / Page 4 (16) Hence, although sensorimotor and cognitive technology can undeniably extend our bodies’ sensorimotor and cognitive performance powers in the outside world, only their sensorimotor input and output contact points with our bodies are part of our cognitive (= mental) state, not the parts that extend beyond. (17) Perhaps it could be otherwise too, as in the case of a hypothetical TT- robot whose generating mechanism is indeed partly located outside its body: Maybe parts of our brain could be removed and still functionally integrated with the rest wirelessly, through telemetry or some other action at a distance: But that would just be a widened, spatially distributed body. (18) The resultant distributed cognitive state would still not be the same thing as considering a telescope, car, library or calculator as parts of a distributed cognitive state (for either a human or a TT robot): Those would still just be parts of the sensorimotor I/O to and from the cognizer’s body. (19) We are not aware of the generating mechanism underlying our cognitive capacity, however, only of its outcome: Hence retrieving a word from memory or retrieving a word via a Google search feels much the same to us. (20) Does the fact that cognizing is a conscious mental state, yet we are unconscious of its underlying functional mechanism, mean that the underlying functional mechanism could include Google, Wikipedia, software agents and other human cognizers’ heads after all? That question is left open for the reader. (21) The worldwide web, a distributed network of cognizers, digital databases and sofware agents, has become our “Cognitive Commons,” in which cognizers and cognitive technology can share cognizing anytime and anywhere, and interact globally with a speed, scope and degree of interactivity that yield distributed cognizing with performance powers inconceivable within the scope of individual cognition. (22) Such changes go beyond mere quantitative increase in efficiency and performance power. As we increase our use and reliance on cognitive technologies, they effect and modify how we cognize, how we do things and what we do. Just as motor technology extended our physical ability and modified our physical life, cognitive technology extends our cognitive ability and modifies our mental life. Part I: What Distributed Cognition Is Not. Meaning: Narrow and Wide. Philosophers, in wrestling with the problem of meaning (“Is meaning in the head or is it in the world?”) have sometimes resorted to Offloading Cognition onto Cognitive Technology / Page 5 saying that there are two kinds of meaning, “narrow” and “wide,” the former located between the ears and the latter distributed across the entire universe -- both the Newtonian universe of distant stars and the Platonic universe of the eternal truths of logic and mathematics. The wide meaning of “apple,” for example, includes not only whatever it is that I may have in mind when I think of or say “apple,” but also what apples really are, out there in the world.1 That, however, is all metaphysics, and concerns the existence and “reality” of some elusive entity called “meaning.” The mission of cognitive science is more modest: Humans and other organisms have certain functional capacities, including metabolism, reproduction, and locomotion. It is clear that each of these capacities is “narrow,” even though it sometimes involves a local interaction between the organism and part of the world around it (be that other organisms or the inanimate world). Performance Capacity. Movement itself, inasmuch as it includes the movements of parts of the organism, and not just the whole of the organism, covers everything that we are able to do; and that, in turn, extends naturally to all of our cognitive capacities – what we are able to think, deduce, understand, etc. – encompassing also the internal mechanisms that generate those capacities. So far, that makes all of cognition narrow: skin and in. It is not that we do not (as in metabolism and reproduction) interact with objects (and skins) outside our own skin. Although usually it is not particularly illuminating to speak of eating and digestion as a dyadic function, “distributed” between predator and prey. Reproductive function is for the most part decidedly dyadic and to that extent distributed more widely than a single organism.2 Distributed Perception? Gibson (1966), too, has stressed that even something as seemingly passive as seeing is in fact interactive, with the locomotory organism perceiving things in terms of their sensorimotor “affordances” – what our dynamic bodies are able to do with external things. Don’t look for the purely sensory property that all “chairs” share: their real invariant is that they afford “sittability upon” – a property that cannot even be defined without reference to the shape and motor capacity of our bodies as well as the shape of things in the external world. Does it follow from this that the perceptual state of perceiving something is a distributed state that includes the perceiving organism as well as the external object or event or action that is being perceived? And – to extend this question further – is the cognitive state of thinking or knowing about something a distributed state, 1 This is even more evident when it comes to what is meant by “superstrings” or “prime number.” 2 In the case of sexual reproduction, ab ovo, and in the case of asexual reproduction, a posteriori, so to speak. Indeed, there is perhaps a lesson to be learned about cognitive function from the two forms of reproductive function, since both are “productive” of something beyond the narrow borders of the particular organism in question. Offloading Cognition onto Cognitive Technology / Page 6 consisting of the cognizing organism plus the external object or event or action (or property or state) that is being cognized (Clark & Chalmers 1998; Wilson 2004)? Let us defer reply until we consider a few more cases, noting only that this question about whether perception/cognition is just (i) internal and local or (ii) internal/external and distributed is similar to the question of whether meaning is narrow or wide. Physical States: Narrow and Wide: A trivial answer would be that every physical state is “distributed” in that nothing is ever causally isolated from everything else. So in singling out (“individuating”) any physical “state” we are individuating an arbitrary subset of the total state of the universe: This chair is not causally isolated from the ground it rests upon, nor the ground from the rest of the planet, spinning about the solar system, etc. By that token, all states are wide – as wide as the world, including oneself, sitting on the chair. But the fact that there is no such thing as an absolutely isolated local entity or state is not what we mean when we ask whether cognitive states are narrow or wide. Otherwise, the state of a toaster, toasting bread, is wide too, and includes not only the toaster and the bread, but also the events transpiring on faraway Alpha Centauri. But, leaving aside the physics and metaphysics of wide causality and action-at-a- distance: what about just the toaster and the bread? Does the “state” of a toaster, toasting bread, include the bread, being toasted? It seems obvious that this distinction, too, is arbitrary, hence trivial: We can include the toaster in a distributed hybrid state and call that a state of the toaster, or a state of toasting. Or we can say that the toaster does what it does, and the bread gets done to it whatever is done to it, but we will consider their states as distinct, acting upon one another (more the toaster acting on the bread than vice versa, unless the toast catches fire) but not a joint, distributed state worth speaking of as such, in useful discussions of either toasters or bread, and their respective functional states and properties. Autonomous Systems. But although there are no states or systems that are completely isolated causally, there are surely “things” – like chairs and bread and toasters -- that are sufficiently isolated to be called autonomous things. Some of these autonomous things will be (again, only relatively) static, like chairs and bread, and some relatively dynamic, like a toaster – if plugged in and functioning. Some of these autonomous things may also be parts of other, wider autonomous things. Toasters have functional parts that can do what they do on their own, in isolation from the toaster. A toaster, in turn, may be part of a more elaborate device that toasts as well as butters, fills and wraps your sandwich; or simply a component in a modular commercial kitchen. So in this approximate way, bracketing the issue of wide causality, we arrive at the notion of autonomous systems, like toasters, composed sometimes of components Offloading Cognition onto Cognitive Technology / Page 7 that are themselves autonomous systems. Let us call those subcomponents autonomous modules, and note that any autonomous system could in principle also be an autonomous module in one (or many) wider autonomous systems. But is a toaster really autonomous? Don’t we have to build it, plug it in to the electrical system, and then put in the bread, and set the level, etc.? Are the toaster and bread and ourselves just part of a still wider distributed system, the one with the real autonomy, while the toaster and the bread are merely “slave” systems, with no autonomy of their own? We cannot avoid extending our relentless questions to asking what we really mean by “autonomy”: Is anything really autonomous, apart from the universe itself, or God almighty? This is again the question of causal isolation, and maybe we can again finesse it by settling for commonsense approximations: A system is autonomous if it can do what it does “on its own.” It’s just that systems differ in what they can do on their own. A toaster is an autonomous system that can only toast bread -- and that, only if a person plugs it in, puts in the bread, presses the switch. A person is an autonomous system that can (among other things) plug in a toaster, put in bread, and press the switch. And so it goes. Both autonomy and functional capacity look modular, and superordinate autonomous systems may include the distributed modularity of many component autonomous systems. We can easily get lost in this mereological maze, so let us avoid the lure of “general system theory” and just note that, yes, there are quasi-autonomous things and quasi-autonomous states, and those things and states may themselves be distributed parts of other, wider quasi-autonomous things and states. That’s all indisputable even before we get to the question of cognizing and distributed cognition. But before we broach that question, we alas have to ask yet another basic question: What is cognition? Cognitive and Vegetative Function. To a first approximation, we have already answered this: cognition is whatever gives cognitive systems the capacity to do what they can do. It is the causal substrate of performance capacity. Cognitive systems (“cognizers”) include ourselves and perhaps other animals (and possibly also extraterrestrial creatures, if they exist). Do they include anything more? Are living systems the only cognizers? Are cognizers necessarily local, or can they be distributed? And is the capacity underlying everything that we cognizers can do cognitive, or only the capacity underlying some of what we can do? One question at a time. Let us first agree that not everything a human being can do is cognitive. Breathing, for example, except in some special cases, is not cognitive; neither is balance, again, except in some special cases. What are the special cases? They are when we breathe or adjust our balance consciously. Otherwise, breathing and balance are unconscious and automatic – we might call them “vegetative” rather than cognitive functions. Offloading Cognition onto Cognitive Technology / Page 8 Consciousness. But surely consciousness itself cannot be the mark of cognition either, because although when we take conscious control of our breathing or our balance that is undoubtedly cognitive, we are not really conscious of how we control breathing or balance. If we suddenly feel we are suffocating or falling over, we “command” our lungs to breathe and our limbs to right themselves, but we are hardly conscious of how our commands are implemented. It is physiologists who must discover how we manage to do those things. And, by the same token, if we do something that we are more accustomed to calling cognitive, such as perceiving a chair, understanding the meaning of a word, or remembering the product of seven times nine, all of which we cognize consciously, we are nevertheless unconscious of how we manage to perceive a chair as a chair,3 how we understand the meaning of, say “cognitive,” or even how we retrieve (from wherever we “stored” it decades ago) the product of seven and nine. Fear not, dear reader, we have not been forced into the clutches of the metaphysical problem of Free Will here. We simply need to make the observation that what makes some of our capacities cognitive rather than vegetative ones is that we are conscious while we are executing them, and it feels like we are causing them to be executed – not necessarily that we are conscious of how they get executed (Libet 1985). But that’s not enough. We have an initial approximate criterion for what performance capacities count as cognitive: It is the ones we execute consciously, which just means that we normally have to be conscious while we are executing them (This criterion is actually flawed, but we will fix it later.) Is There Cognizing Without Consciousness? Now the second question: Are there any other cognitive systems besides ourselves and animals? We have already noted that not all of our performance capacities are cognitive: the cognitive ones are the ones we execute consciously (although we are not conscious of how they are executed by our brains). The question of whether systems other than animals like us cognize is hence related to the question of whether or not there can be cognizing without consciousness: It concerns which organisms are conscious, and whether nonconscious -- perhaps even nonliving -- systems, can be cognizers too. The “Other-Minds” Problem. Let us quickly agree (with Hume and Descartes) that there is absolutely no way for one to know for sure whether anyone (or anything) but oneself is conscious. (This is called the “other minds” problem, and it is insoluble.) Hence we already have a problem here, if consciousness is the mark of the cognitive. We can’t know for sure who or what is or isn’t conscious. But do things get even worse? Doomed to be left agnostic about whether anyone or anything else is conscious, are we even more agnostic, then, about whether nonconscious systems can cognize? 3 i.e., how to detect its Gibsonian “affordances.” Offloading Cognition onto Cognitive Technology / Page 9 What Is Alive? Here there may be useful lessons to be drawn from the problem of life: Very similar questions have been raised about what it is for a system to be alive. We used to think there had to be a “vital force.” Now we know better; life is just the state of certain dynamical systems, having certain structural and functional properties, including molecular ones. The properties of living systems are all objective and observable, so once it has been ascertained that those properties are indeed present, there is no vitalist homologue of the “other minds” problem to trouble us, about whether or not the system is really in a biotic state, i.e., “really alive.” The observable, objective properties of living systems exhaust all there is to being alive (other, perhaps, than the “other minds” problem itself, for those who hold that all living systems must be conscious!). Biotic States: Narrow and Wide? The same question of distributedness – “narrow” versus “wide” life – arises also with living things: We all know the case of the amoeba, which is an individual, autonomous, one-celled organism, definitely alive in its own right. But when individual amoebae find themselves together in a certain chemical gradient, they coalesce and become a further, superordinate, fungus-like organism called a slime-mold. This is “distributed” life, in the sense that it is wider than any of the individual amoebae (who nevertheless remain alive too), and encompasses the entire slime mold, which is then an autonomous, superordinate, living organism. Something like the slime mold was also probably the origin of all multicellular organisms, all the way up to ourselves: We are instances of “wide” life, distributed over all our individual living (though only minimally autonomous) cells. Distributed Life. So far, so good. But, can a group of organisms working and functioning as one, be an individual organism? What about a sports team or an army unit? Or a colony of ants or bees? Even more controversially, some have gone on to argue that an entire biological species may also be an individual organism -- a wider, superordinate organism, distributed over all its members, much the way the slime mold is a superordinate organism distributed over its individual modules, the amoebae (Hull 1976). And it can get even wider, some arguing that Earth itself is a superordinate organism, “Gaia,” distributed over the entire biosphere (Lovelock 2000). Perhaps some exobiologists will want to argue that if there is life elsewhere in the universe, then all instances of biotic systems are distributed subcomponents of yet another individual mega-organism. We will not settle the question of “distributed life” here one way or the other, except to note that (apart from the relatively coherent multi-cellular organisms “supervening” on individual living cells) the criteria for individuating wider and wider forms of life begin to look just as arbitrary as the extension of physical states (on the grounds that no physical sub-state is totally isolated causally) to the size of the entire universe. Nor is it clear any more what, if anything, is at stake when we Offloading Cognition onto Cognitive Technology / Page 10 can call many distributed things one superordinate thing at will, mixing and matching according to taste. We should try to avoid such a state of affairs with distributed cognition. The questions of distributed life and of distributed cognition, however, are not independent, because (to a first approximation) it is living organisms that cognize (those of them that do), and it is likewise living organisms that are conscious (those of them that are). Spatial and Causal Disjointness. Up to the level of continuous multi-cellular organisms, we can agree about what is and is not a living organism. Consensus and coherence collapse only when we move to the level of the species or Gaia, as both the spatial distance and the causal interactions among the component organisms get distributed more and more widely and loosely: Fungi are the biggest uncontestable organisms on the planet. Some of them can grow underground to a size of over 2,000 acres and live more than 2,000 years. Their individual fruiting bodies, the mushrooms (that we mistake as being the whole organism), though remote from one another in space, are all continuously connected. What about a coral colony, or, better, an ant colony? Is it such a stretch from the spatially continuous and tightly coupled causal interactions of the amoebae that constitute a slime mold to the only somewhat more spatially disjoint and less tightly coupled causal interactions of the ants that constitute a colony? Within multi- cellular organisms there is action at a distance (for example, via chemical gradients) as well as coherent but distributed activity (as in a neural network). And we all know that “spatial continuity” breaks down at microscopic scales. Fortunately, in individuating organisms there are other criteria besides spatial and temporal continuity. There is DNA, which can help resolve (up to cloning) whether or not two bits are (or were) indeed parts of the same organism. But genetic relatedness is only relative, which is what allows some to argue that species are individuals and that Gaia is a mega-organism. Distributed Mental States? We can avoid having to wrestle with the metaphysical problem of individual identity in making our bets as to whether something is a case of individual life or just multiple life, interacting. Siamese twins offer a clue: Why are we ready to contemplate the possibility that Gaia, or an entire species, or an ant colony, might be one single, widely distributed, physically disjoint organism, yet we are not ready to consider that Siamese twins, no matter how tightly fused they are physically, are one single organism? The example illustrates how tightly interconnected and fused the questions of distributed life and distributed cognition really are (at least in our minds): The reason we would never dream of saying that Siamese twins are one single distributed organism is that they have two different minds. And distinctness (or identity) of minds trumps all of our other intuitions and inclinations, insofar as Offloading Cognition onto Cognitive Technology / Page 11 individuating either organisms or cognizers is concerned. Suppose Siamese twins could share every last body part yet could still have two distinct minds: not as in the ambiguous case of multiple personality disorder, where the “minds” (if they are really different minds at all) come and go serially, like masks, but where they are always jointly present, and you can communicate with them, and they with one another, simultaneously, exactly as in the case of ordinary Siamese twins (except that both cannot speak at exactly the same time). We would still have profound difficulty seeing them as one and the same “organism” – or perhaps we should say that the notion of an individual “organism” would simply lose its meaning for us in such a situation: They would be one “biotic system,” in some technical sense, but two distinct “cognizers.”4 Is this “animism” (which was probably always latent in the “vitalism” that has since been discredited by molecular and evolutionary biology), just in our minds? Should we be accepting objective, system-based functional inventories of what does or does not count as a distinct cognizer, as we do with what does or does not count as being alive? Or does our subjective sense have some privileged say in the matter? Mind-Reading. The very same mentalistic intuition that underlies how many cognizers we perceive within a single organism can cut the other way too: The reason most of us are not ready to see an ant colony, a species, a corporation, a government, a sports team, an army unit or Gaia as either an individual organism or a cognizer is that we do not perceive any of them as having a mind. We can, with a little effort, see a tree or a fungus, a coral, an amoeba or a slime mold as a living organism, like us, especially if biologists tell us it is so; but we would have great difficulty seeing any of them as cognizers – unless we are ready to see them as being conscious (having a mind). We do have natural “mind-reading” abilities and inclinations (Whiten 1991), as along with fertile imaginations. When we are children, and our animism is at its ”widest,” we are ready to see a tree as watching or even waving at us, or to believe that it hurts the tree when we kick it. Past a certain age, children also have a great deal of difficulty not believing that it hurts a dog, or another child, if they kick it.5 We perceive other minds because we can (sometimes) detect the Gibsonian “affordances” (perhaps via our “mirror neurons”) of being in a mental state: We know what it’s like to have a mind, because we each have one. The rest is our “mirror neurons,” detecting when another mind is in a mental state like our own, because it is doing something like what we would be doing in that mental state. In other words, we mind-read through a combination of having a mind and perceiving its bodily performance correlates in others (Gallese & Goldman 1998). 4 In contrast a baby born with extra limbs will always be considered as a single organism, regardless of how many extra limbs it has, as long as it has just one cognizing mind. 5 Although without moral training, that is not necessarily enough to prevent the child from kicking it! Offloading Cognition onto Cognitive Technology / Page 12 The "other minds" problem does not go away; our 'mind reading' is not based on flawless deductive reasoning. The logic that similar consequences must have similar causes (I have a mind and know its bodily outcome, therefore such bodily outcomes in others must be caused by a mind) gives raise to false positives and false negatives. Hence we are fallible mind-readers. Seeing the tree as having a mind is probably a false positive. Might seeing Gaia as not having a mind be a false negative? Perhaps. But let us be clear about exactly what we would be getting right or wrong, when we made a correct “hit” versus a false positive or negative: Living and Cognizing. If Gaia, or a species (e.g., the earth’s elephant population as a whole) did have a mind, that would mean, roughly speaking, that it was the kind of thing that was capable of having a headache (not necessarily having a head, just having a headache), say, a migraine. The migraine is just a stand-in, here, for our intuitions about what it is to have a mind at all. To have a mind is to be in a mental state, and a mental state is simply a felt state: To have mind is to feel something – to feel anything at all (e.g., a migraine).6 And make no mistake about it: you must have a mind – i.e. you must be in a mental state, you must be able to feel -- in order to have any inkling at all of what it means to have a mind! A toaster will not get that from a dictionary definition. Outside minds there is nothing but mindless (feelingless, insentient) functionality. The Migraine Test. The migraine is merely our stand-in for the capacity to feel anything at all -- in other words, for being conscious. We all know what it feels like to have a headache. All feelings are pretty much like that, mutatis mutandis, from what it feels like to perceive a chair, to what it feels like to understand the meaning of a word or to remember the product of seven and nine. Note that what is essential for having a mind is not having the performance capacity itself – being able to detect the presence of the chair, being able to define or reply correctly to the word, being able to retrieve “sixty-three ” -- nor is it essential to have an understanding of the underlying causal mechanism of that performance capacity (knowing how we manage to do it). The essential thing for having a mind is being able to feel what it is like to have and execute the capacity – or to feel anything at all (e.g., a migraine). This is the consciousness that accompanies cognizing (though without necessarily any consciousness of how the cognizing actually works). Suppose it was somehow true that Gaia (or the entire elephant species, or an ant colony) was indeed a superordinate living organism, distributed across everything in the earth’s biosphere (or across all elephants, or all the ants in a colony). And suppose the reason we wrongly thought Gaia was not an organism was that we couldn’t imagine such a distributed system as being capable of having a migraine (or any other mental state). We could of course have been wrong about that too: Maybe 6 Having a mind, being in a mental state, being conscious, being in a conscious state, feeling, being in a feeling state, feeling anything at all -- all of these are synonymous. Offloading Cognition onto Cognitive Technology / Page 13 Gaia could have a migraine. (Because of the other-minds problem, there is no way to be sure one way or the other.) But suppose we were right that Gaia has no mind yet wrong that Gaia is not a living organism. In that case, our mind-reading mirror-neurons would have been right – they detected no mind. But they would nevertheless have steered us into a false negative, because Gaia, though mindless, is nevertheless alive. In contrast, the child’s mirror-neurons commit a false positive on the migraine test, wrongly inferring that a tree does have a mental state, though it does not, but the child is nevertheless right (though for the wrong reasons) that the tree is alive. In both cases, being an organism was conflated, animistically, with having a mind. This is an error; living and feeling are not necessarily the same thing. There can be living organisms that have no mental states and there can be nonliving systems that do have mental states. Cognitive States and Mental States. Can the same distinction be made, however, if we apply the same mind-reading criterion to being a cognizer, rather than to being a living organism? We (or rather, our mind-reading mirror neurons) insisted, in the case of the Siamese twins with only one body, that even if Biology were to tell us that they were one single organism, they would still be two distinct cognizers, if they had two distinct minds: They would not have one, shared mind, even though they did have one, shared body. And if they had a migraine, it would be two migraines, even if it was implemented in one and the same head -- just as when something is a “headache” for the US Congress, it is at most N distinct headaches in the heads of N distinct members of congress, with no further superordinate entity feeling an N+1st headache (or feeling anything at all). There is no such thing as a distributed migraine – or, rather, a migraine cannot be distributed more widely than one head. And as migraines go, so goes cognizing too -- and with it cognition: Cognition cannot be distributed more widely than a head -- not if a cognitive state is a mental state. Cognitive Technology: Tools R Us? Does this settle the question of distributed cognition, or does it beg it? The case for distributed cognition is based mostly on cognitive technology: the argument is that even something as simple as an external piece of paper with a phone number on it is a piece of cognitive technology -- a peripheral device on which data are stored. If the phone number were encoded inside one’s brain, as a memory, there would be no dispute at all about its being part of the (internally) distributed cognitive state of, say, knowing or finding that phone number. Why, then, would we no longer consider that same datum as part of that distributed cognitive state just because its locus happened to be outside the cognizer’s body? Moreover, once we realize that our cognitive states include data that are located on an external piece of paper, then it becomes apparent that they include far more than that -- widely distributed things, ranging from everything located in our libraries and on the Web, to every auxiliary device, process or datum that may enter into any cognizing or its outcome, extending also to everything located in the narrow heads Offloading Cognition onto Cognitive Technology / Page 14 of all other individual cognizers (Hollan et al. 2000). Whether we want to include in a cognitive state everything that can potentially enter into anyone’s cognizing or only what actually enters into someone’s cognizing, either way, on this extended view, cognition is looking exceedingly wide. The Extended Mind. This wide view of distributed cognition is also called the “extended mind hypothesis” (Clark & Chalmers 1998; Wilson 2004). It has some affinities with wide theories of meaning, in which apples themselves, or the truths about apples, are taken to be part of the distributed meaning of “apple,” which is thereby extended beyond what may be going on within the narrow confines of the head of any individual, local cognizer. The extended mind is also reminiscent of the notion of “wide” toaster-states (in which the bread is part of a distributed state of the toaster), thereby also inheriting the apparent arbitrariness of such state extensions. We must accordingly ask ourselves why we would want to contemplate such arbitrary extensions of what it is to have or to be a mind, hence to be a cognizer and to cognize? Why would it even cross our minds? The answer is again the (insoluble) other-minds problem: Since there is no way of knowing for sure whether any cognizer other than oneself has a mind, there is even less way of knowing whether or not there can be cognizing without a mind, or even of knowing what the actual geographic boundaries of a mind are. We know, objectively, what cognition does. Doing is performance capacity. Cognitive science is also beginning to make some progress in explaining the functional mechanism generating that performance capacity (i.e., how our brains do it). We also know that so far the only systems that our adult “mind-reading” capacities have tentatively identified as being cognizers have been living organisms like ourselves. Our confidence that we have detected an “other mind” increases the more the candidate resembles ourselves, both in its appearance and in its performance capacities. That is in fact the (narrow) meaning of “cognition”: the kinds of things that I and other living organisms can do, using our minds. But there are other candidates that also seem to be able to do some of the things that living organisms like us can do, and not just the vegetative things, but the things we have identified as cognitive, when performed by us. Computers and robots are not only doing more and more of what only living organisms like us used to be able to do, but even the functional mechanisms that biology and cognitive science are proposing to explain how organisms do it often turn out to draw on the same functional mechanisms that explain how computers and robots do it. (Indeed, the functional explanation often comes from the fact that we have programmed computers and designed robots to do what we do, and in so doing, we have also provided a potential explanation of how our own brains do it). So, if it walks and quacks like a duck, and even its internal mechanism is like that of a duck, it’s only natural to assume it’s some kind of duck too. Offloading Cognition onto Cognitive Technology / Page 15 The Turing Test. This is the rationale and the methodology behind the Turing Test (TT; Turing 1950; Epstein et al. 2008): If we can design a system that can do everything that we can do -- and do it so well that even our mind-reading mirror- neurons perceive it as having a mind -- then we have no more (or less) reason to doubt that it has a mind that we have for doubting that other human beings have minds (again because of the other-minds problem). In addition, the TT-passing candidate – which has to be a robot, because a computer alone cannot have all of our sensorimotor capacities – will provide us with at least one explanation of the functional mechanism underlying our own cognitive capacity. Hence the question of narrow vs. wide cognition is also a question about what is and is not part of the functional mechanism of a TT-passing robot: What needs to be “inside” such a robot in order to pass the TT? Note that this is not the question of whether we need to pack all its functions inside its head, the way ours are packed inside our heads. It is conceivable that the mechanism of the TT robot could be more widely distributed: some of it inside and some of it outside its body, integrated wirelessly, perhaps, from some central location. The states consisting of the joint activity of the robot-internal and the robot-external components of the mechanism that give the robot the capacity to pass the TT would be indisputably distributed cognitive states. But those hypothetically distributed robot states (if they are possible at all) do not settle the question we are inquiring about here. Nor would they settle it even if it were somehow possible to breed people with parts of their brains physically located outside their bodies and their joint activity integrated through wireless telemetry or some such. Such a hypothetical distributed robot (or person) could even have a distributed migraine. But what we would really have then would be a robot (or person) with an extended (or distributed) body. The constituents of its mental states would all still be (distributed) within that one distributed body. Our brains, after all, are still parts of our bodies, even if they could be removed, all or in part, temporarily or permanently, the way our hearts have been, and somehow kept functionally integrated with our bodies wirelessly. This is all cog-sci-fi. But the point of the example is to show that this sort of hypothetical cognitive state – distributed across multiple parts of a robot’s functional mechanism (or even multiple parts of an organism’s brain) that happen to be widely separated in space but coordinated wirelessly (Dennett 1981) – does not address the question of whether or not cognitive technology is part of our cognitive state too. An affirmative answer to the question of whether, if the parts of my brain that control the left and right sides of my body could be moved out of my brain and two miles apart, while still being able to remotely coordinate my walking, does not address the question of whether cars or calculators are a part of my mind. 7 7 Please note that if cognitive technology is only a tool and not part of our mind and a cognizer itself, this does not imply that it does not have profound effects on how we cognize --more on this later. Offloading Cognition onto Cognitive Technology / Page 16 Software Agents. It is not just today’s (sub-TT) robots that appear to be doing some of the things we cognizers do: Software agents seem to be doing it too, including communities of distributed software agents, interacting among themselves, trawling the net, executing local as well as distributed computations on local and distributed data, and displaying -- not just individually but also collectively -- performance capacities that, in living organisms like ourselves, we would have taken to be the result of cognizing (Dror 2007). These autonomous devices – both hardware and software -- are, of course, like toasters. But they are “cognitive” toasters, in that they operate not on bread but on informational inputs, to generate, as output, performance that we would have called cognitive if we had been the ones doing it. Particularly in the case of the joint activity of distributed software agents, autonomously crawling the web, it is obvious why the question of whether a datum is internal or external becomes arbitrary. The datum may be the input to one agent or the output of another, and the distributed cognitive system consists of all the agents and their inputs and outputs together anyway. It makes little sense, nor is it of much use, to try to say which is the bread and which is the toaster in such cases. But is what these systems are doing (whether they are local pieces of hardware or distributed digital data and the software agents programmed to process them) cognizing, or just something that is similar to what ordinarily requires cognizing to do? The question seems to be as undecidable as whether or not Gaia is really a living organism. Part II: What Distributed Cognition Is. Wide-Body Beings. In Part I we argued that inasmuch as cognition is mentation (i.e., insofar as cognizing is thinking), there can only be distributed cognitive states where there can be distributed mental states. Within the head there are narrowly distributed cognitive states, since neural states are presumably not all local and punctate. If the mechanism that generates mental states and bodily performance capacity (normally the brain) could be more widely distributed in space (beyond the head), and still be integrated somehow so as to generate coordinated mental states and bodily function, then that too would be widely distributed cognition, whether in a hypothetical TT-scale robot or a hypothetically re-engineered organism, but that would also be a widely distributed body. Distributed cognition would still not be wider than the body. Can there be distributed cognition beyond the bounds of the body and the brain? In particular, can external cognitive technology serve as a functional part of our cognitive states, rather than just serving as input to and output from them? Mental States Are Conscious States. Let us consider brain states, rather than just mental or cognitive states. We have agreed that not everything our bodies do is Offloading Cognition onto Cognitive Technology / Page 17 cognitive. Some of it, like breathing, balance, or temperature control is vegetative. So, too, are the brain states that implement those vegetative functions. We have also agreed that although cognizing is conscious, we are not conscious of how cognizing is implemented. When we recognize a chair, or understand a word, or retrieve the product of seven and nine from our memory, the outcome, a conscious experience, is delivered to us on a platter. We are not conscious of how we recognized the chair, understood the word, or retrieved “63”. Hence the brain states that implement those cognitive functions are not conscious either. Are unconscious brain states mental? Are Unconscious Brain States Mental? The natural answer would seem to be: no. Unconscious states are unconscious states. The states of a toaster are unconscious and certainly not mental. Until further notice, “conscious states” is synonymous with “mental states.” The brain states implementing vegetative function are not mental either: Presumably a person in a chronic vegetative state is as unconscious as a toaster (although, because of the other-minds problem, we can never be sure about either the toaster or the comatose person). The only reason we want to call the brain states that occur while we are conscious mental states is that they occur while we are in a conscious state, and they physically implement that conscious state. (We are on the fuzzy boundary of the mind/body problem here.) But just as vegetative states such as the regulation of breathing, which occur unconsciously while we are conscious, are nevertheless not themselves mental, nor part of our mental state, why would we want to call the unconscious state that “delivers” our conscious mental state mental? When you say to yourself “what is seven times nine?” and then “sixty-three” pops up, you are certainly conscious of thinking “sixty-three.” So that’s definitely mental; and so is the brain state that corresponds to your thinking “sixty-three.” But what about the brain state that actually found and delivered the “sixty-three”? You are certainly not conscious of that, although you were just as conscious while your brain was finding and delivering “sixty-three” as while you were breathing, though you don’t feel either of those states. Neural vs Google Storage and Retrieval. Let us make the retrieval interval longer then, just to make the problem more vivid: You are trying to remember the name of a poet. You know he wrote “Tell me not in mournful numbers, life is but an endless dream” and his name is on the tip of your tongue, but you just can’t retrieve it. You go to sleep, and next morning “Henry Wadsworth Longfellow” immediately pops up. You were not even awake during the brain state that retrieved it. So what difference does it make if you recall it through an unconscious retrieval state in your brain, or by Googling it (again relying on a state in some remote computer and database of which you are not conscious)? Are they not both based on an unconscious, nonmental state, in the first case narrow and neural, inside your brain, in the second case wide and computational, distributed between your brain and a computer hundreds or thousands of miles away? Offloading Cognition onto Cognitive Technology / Page 18 Distributed Databases. And what about a modern child, who has never bothered to memorize the multiplication tables, as you did, because a computer is always at hand? The only way he ever retrieves 7 x 9 is to key it in, and read off the product . He blindly consults his computer when you blindly consult your memory: What’s the difference? Never mind computers: the poet’s name could be read out of a static book that indexes poems’ first lines. Or you could just ask somebody who knows to tell you who wrote those lines. What difference does it make if the database in which the datum is stored, outside your awareness, is in your brain, or on the shelf of a library, or in someone else’s brain? Offloading Brainwork. The beginning of cognitive technology was surely language, which allowed cognizers to “offload” a lot of brainwork onto other brains that could do it for you, and deliver you the results, and vice-versa (Cangelosi & Harnad 2001; Dascal 2004). Are our own neural states, plus Google states, plus book states, plus the neural states in the heads of other cognizers all parts of distributed cognitive states – and if so, whose cognitive states? I am presumably the cognizer of my narrow cognitive states, but who is the cognizer of the wide ones? Or are cognitive states just sui generis, rather than belonging to anyone in particular? Neural firings in brains, plus keystrokes on computers, bits coursing across fibre optic cables, remote disk activity, print in a library book, neural states in other people’s brains – all just parts of wide, distributed, disembodied cognitive states, taking place here, there, and everywhere: cognizing, with no cognizer?8 At the very least, we need to pinpoint the cognizer of the distributed cognitive state. Let us say it is the user of the cognitive technology, and that what we are asking is whether the technology outside the body is part of or merely I/O to/from a narrow cognitive state inside his brain? Sensorimotor Technology and Augmented Reality. Let us start by considering a kindred kind of technology, perhaps not quite cognitive, only sensorimotor, with the corresponding states being sensorimotor states rather than fully cognitive ones: Sensorimotor technology probably began in our species’ prehistory with tools and weapons, which extended our performance capacities dramatically. Let us consider a relatively recent tool: If you look at a star through a telescope, is that a distributed sensorimotor state, consisting of your brain and retina plus the telescope (and perhaps also the star), in which your visual capacity is augmented by the telescope’s power of refraction? Or is it just input to your narrow, skin-and-in sensorimotor state – input augmented by the telescope? 8 (rather like a distributed life, with no organism living it; or a distributed migraine, with no one experiencing it)? Isn’t cognition with no cognizer cognizing it like a feeling with no feeler feeling it?) Offloading Cognition onto Cognitive Technology / Page 19 If you are driving a car, is that an extended sensorimotor state, in which your body is moving at speeds in excess of what it can manage alone, narrowly? The wider, distributed sensorimotor state might include the car and its locomotor capacity. Or is it just output from your narrow, skin-and-in sensorimotor state (in this case a slow movement of your foot on the pedal) – output augmented by the horse-power of our external vehicle? Another example would be operating a crane, and the extended power to reach and manipulate objects that are too far, big and heavy to be manipulated in your narrower sensorimotor state. Is this a widely augmented I, or just I/O to narrow old me? But before we dismiss too quickly the notion of a wider sensorimotor state, note that some of us have literally experienced a change in our felt body image when driving a large car: Our sense of our own width, pulling through a narrow squeeze, extends to the width of our car, not just our narrow body. This change in body image is not unlike the effect induced by distorting prisms, Virtual Reality, or even surgery, prosthetic limbs, and neurological re-adaptation. Tadpoles mutating into frogs and caterpillars mutating into butterflies might be undergoing similar sensorimotor changes in their body images and powers because of real changes in their (narrow) bodies: Are technology-extended bodies all that different? The Advent of Language. The effects of cognitive technology can be similar to those of sensorimotor technology. Language evolved neurologically for speech and its interactive tempo. We can accelerate recorded speech technologically beyond the rate we can speak it, yet still understand it. Beyond a certain speed, speech becomes gibberish – yet we can read and understand written language at far faster speeds (probably because hearing is a more serial medium of processing and vision is more parallel). It is virtually certain that there was no specific neural adaptation for reading, which was a technological invention of less than 10,000 years ago. In contrast, the language areas of our brain were shaped genetically several hundred thousand years ago, altering our neural hardware and radically extending our cognitive powers. If spoken language widened our cognitive powers biologically, didn’t reading and writing widen them technologically in much the same way? Language As Distributed Cognition? Is language itself a form of distributed cognition? How does the knowledge in other people’s heads, conveyed to us auditorily, differ from the knowledge in books, conveyed to us visually? Both allow us to access information without needing to gather it the hard way, through our own direct, time-consuming, risky and uncertain sensorimotor experience. Writing and speaking also allow us to offload our knowledge and memory outside our own narrow bodies, rather than having to store it all internally. Individual cognizers write books, but Wikipedia, for example, seems to be growing spontaneously Offloading Cognition onto Cognitive Technology / Page 20 according to an independent, collective agenda of its own, more like the joint activity of a colony of ants. Computers, distributed digital databases and automated algorithms have augmented both the speed and the computing power of our brains, and that newfound speed and power is capable of inducing changes in our mental self-image not unlike the ones that sensorimotor technology can induce in our body image: If being deprived of one’s spectacles or one’s automobile feels rather like the loss of eyes or limbs, being deprived of one’s computer or cell-phone feels like the loss of one’s intrinsic cognitive and communicative capacity. Interactive Cognition. Human discourse is certainly interactive cognition, indeed collaborative cognition, and the speed and distance at which we could speak, and understand our interlocutors, set biological limits on the rate and scope of that collaborative cognition, hundreds of thousands of years ago. The speed of verbal thought probably co-evolved with language and is probably of the same order of magnitude as the speed of oral speech. Although reading speed is much faster than listening speed, writing (and typing) speed is not as fast as speaking (and the turn- around time of letter delivery is certainly slower than conversational speed). So in real-time interactions, at the speed of thought, we still prefer to talk rather than write. It is only recently that cognitive technology (in this case, email and texting) has accelerated the potential speed of written interactions in almost real-time to something closer to the speed of thought. Web-based threaded discussion lists, and especially their quote/commentary capability, not only accelerate this interaction still further – allowing individual cognizers to interact with the text itself in real time. They also increase the scope of this almost-real-time interaction among distributed minds and distributed texts; and global posting and immediate accessibility potentially make the collaboration almost instantaneous (Harnad 2004). If the human brain was biologically optimized for interactive cognition at speaking speed, and writing technology slowed down that interactive cycle (in exchange for the other benefits of the transmission and archiving of a written record) for thousands of years, then digital online technology has now once again accelerated the interaction to the speed of thought, increasing its power and productivity by orders of magnitude, and distributing it globally and instantaneously. It is this newfound interactivity (not passive radio, television or film) that is at last truly turning Gaia into McLuhan’s (1962) “global village.” Cognizers and Tools. So where does this leave the question of distributed cognition? It is still cognizers who cognize -- the tool-users, not the tools.9 Yet there 9 Did some of the ambiguity arise from the fact that we fell into the habit (perhaps because of funding agency contingencies) of overusing (for funding purposes) a rather vague and equivocal noun and adjective – “cognition” and “cognitive” – instead of a less impressive verb and gerund – “cognize” and “cognizing” – to ask these questions that are basically about thinking and knowing? Would we have Offloading Cognition onto Cognitive Technology / Page 21 is no doubt that cognitive technology has radically widened the scope of human cognizing10. Could “cognitive technology” be the brain’s way of off-loading some of its otherwise far greater encoding and processing burden? If so, then the worldwide web, a distributed network of cognizers, digital databases and software agents, has effectively become our “Cognitive Commons,” in which distributed cognizers and cognitive technology can interoperate globally with a speed, scope and degree of interactivity that generate cognitive performance powers that would be inconceivable within the scope of individual local cognition alone. Cognitive Technology and the Human Mind. Is cognitive technology limited to increasing the cognitive performance capacity of its users? No. We have argued that cognitive tools are not themselves cognizers, nor do they have -- or serve as distributed substrates of -- mental states. But their effects go well beyond making human cognition more efficient and productive. Just as noncognitive technology (cars, planes, machinery) transformed our somatic lives, so the offloading of brain function onto cognitive technology is now transforming our cerebral lives. Physical technology altered the frequency, intensity, and manner of our muscle use, altering our muscular development (even introducing new ‘technological diseases’, such as carpal tunnel syndrome). Cognitive technology will do likewise, but instead of affecting our muscles it will affect our brain development, organization and capacities. Changing how we think, learn, and communicate, our cognitive tools are reshaping our minds. been ready to say that a library was doing “distributed knowing,” or that it was part of a “thinking state” distributed across brains and book-shelves? Or that “collaborative cognizing” was any more “distributed” than collaborative thinking or knowing (or worrying)? 10 Epigenetics is perhaps a biological precedent for this (Waddington 1942): Not every trait of an organism needs to be genetically encoded in its (“narrow”) DNA. If there are stable environmental influences that can be relied upon to “canalize” the expression of genes without having to be written into the blueprint, that takes a needless load off the narrow code, and even allows it to be more flexible toward wider environmental contingencies. (Perhaps the neural counterpart of Eprigenetics should be called “Epinoetics.”) Offloading Cognition onto Cognitive Technology / Page 22 References Cangelosi, A. and Harnad, S. (2001) The Adaptive Advantage of Symbolic Theft Over Sensorimotor Toil: Grounding Language in Perceptual Categories. Evolution of Communication 4: 117-142. Clark, A. & Chalmers, DJ. (1998) The Extended Mind. Analysis 58(1): 7-19 Dascal, M. 2004. Language as a cognitive technology. In B. Gorayska & J.L. Mey (eds.), Cognition and Technology: Co-existence, Convergence, and Evolution (pp. 37-62). Amsterdam: John Benjamins. Dascal, M. & Dror, I. E. (2005). The impact of cognitive technologies: Towards a pragmatic approach. Pragmatics & Cognition, 13 (3), 451-457. Dennett, D. (1981) Where am I? In D. Dennett, Brainstorms: Philosophical Essays on Mind and Psychology (pp. 310-323). Cambridge, MA: MIT Press. Dror, I. E. (2007). Land mines and gold mines in cognitive technologies. In I. E. Dror (ed.), Cognitive Technologies and the Pragmatics of Cognition (pp1-7). Amsterdam: John Benjamins. Epstein, R, Roberts, G & Beber, G. (2008) (Eds) Parsing the Turing Test: Philosophical and Methodological Issues in the Quest for the Thinking Computer Springer Gallese, V & Goldman, A. (1998) Mirror neurons and the simulation theory of mind- reading. Trends in Cognitive Sciences 2(12): 493-501 Gibson, JJ. (1966) The Senses Considered as Perceptual Systems. Greenwood Publishing Group Harnad, S. (2004) Back to the Oral Tradition Through Skywriting at the Speed of Thought. In: Salaun, J-M & Vendendorpe, C. (Eds.). Le défi de la publication sur le web: hyperlectures, cybertextes et méta-éditions. Presses de l'enssib. Harnad, S. and Dror, I. (2006) Distributed Cognition: Cognizing, Autonomy and the Turing Test. Pragmatics & Cognition, 14 (2), 209-123. Hollan, J, Hutchins, E & Kirsh D (2000) Distributed cognition: toward a new foundation for human-computer interaction research. ACM Transactions on Computer-Human Interaction 7(2): 174-1962 Hull, DL (1976) Are Species Really Individuals? Systematic Zoology 25 (2): 174-191 Offloading Cognition onto Cognitive Technology / Page 23 Libet, B. 1985. Unconscious cerebral initiative and the role of conscious will in voluntary action. Behavioral and Brain Sciences 8: 529-566. Lovelock, James (2000). Gaia: A New Look at Life on Earth. Oxford University Press McLuhan, M. (1962) The Gutenberg Galaxy: The Making of Typographic Man. University of Toronto Press Turing, A.M. (1950) Computing Machinery and Intelligence. Mind 49 433-460 Waddington, CH. (1942), The epigenotype. Endeavour 1, 18–20. Whiten, A. (Ed.) (1991). Natural theories of mind: Evolution, development, and simulation of everyday mindreading. Oxford: Blackwell Wilson, R. A. (2004) Boundaries of the Mind: The Individual in the Fragile Sciences - Cognition. Cambridge University Press
1004.4454	1	1004	2010-04-26T09:43:28	Crowd simulation influenced by agent's socio-psychological state	[ "cs.MA" ]	The aim our work is to create virtual humans as intelligent entities, which includes approximate the maximum as possible the virtual agent animation to the natural human behavior. In order to accomplish this task, our agent must be capable to interact with the environment, interacting with objects and other agents. The virtual agent needs to act as real person, so he should be capable to extract semantic information from the geometric model of the world where he is inserted, based on his own perception, and he realizes his own decision. The movement of the individuals is representing by the combination of two approaches of movement which are, the social force model and the based-rule model. These movements are influenced by a set of socio-psychological rules to give a more realistic result.	cs.MA	cs	JOURNAL OF COMPUTING, VOLUME 2, ISSUE 4, APRIL 2010, ISSN 2151-9617 HTTPS://SITES.GOOGLE.COM/SITE/JOURNALOFCOMPUTING/ Crowd simulation influenced by agent’s socio- psychological state F. Cher if, and R. Ch ighoub 48 Abstract — The aim our work is to create virtual humans as intelligent entities, which includes approximate the maximum as possible the virtual agent animation to the natural human behavior. In order to accomplish this task, our agent must be capable to interact with the environment, interacting with objects and other agents. The virtual agent needs to act as real person, so he should be capable to extract semantic information from the geometric model of the world where he is inserted, based on his own perception, and he realizes his own decision. The movement of the individuals is representing by the combination of two approaches of movement which are, the social force model and the based-rule model. These movements are influenced by a set of socio-psychological rules to give a more realistic result. Index Terms—Intelligent Agent, virtual crowd, cognitive map, social force, based-rule model ——————————  —————————— V 1 INTRODUCTION IRTUAL human simulations are becoming each time more popular, and many systems are availa- ble targeting several domains, as autonomous agents, humans factors analysis, training, education, vir- tual prototype simulation-based design, and entertain- ment simulations with autonomous virtual humans, or actors, may use different techniques for the behavioral programming. Most common approaches are based on scripts and hierarchical finite state machines, but many other techniques exist, as the parallel transitions network. Some constraints arise when we deal with crowds of vir- tual actors different from the modeling of virtual individ- uals; crowds are ubiquitous feature of everyday life. Hu- man crowds are ubiquitous in the real word, making their simulation a necessity for realistic interactive environ- ments, physically correct crowd models also have appli- cations outside of computer graphics in psychology, transportation research, and architecture. People have long assembled collectively to observe, to celebrate, or to protest at various happenings. The collec- tive assemblages or gatherings called crowds have been the object of scientific inquiry since the end of 19th century. With computers it becomes possible not only to observe human crowds in the real world, but also to simulate vari- ous phenomena from the domain of collective behavior in the virtual environments. Collective behaviors have been studied and modeled with very different purposes. Besides single work concerned with generic crowd simulation, most approaches were application specific, focusing on different aspects of the crowd behavior. As a consequence they employ different modeling techniques ranging from those that do not distinguish individuals such as flow and network models, to those that represent each individual as being controlled by rules based on physical laws or beha- vioral models. Applications include animation production systems used in entertainment industry, crowd behavior models used in training of military personnel or policemen, crowd motion simulations to support architectural design both for everyday use and for emergency evacuation con- ditions, simulations of physical aspects of crowd dynamics and finally sociological and behavioral simulations. Within the framework of our work, we propose a mi- croscopic model of simulation of a virtual crowd with high density in a dynamic and more complex environment. In order to navigate in a complex environment, we need to have an efficient abstract representation of the virtual envi- ronment [4] where the agents can rapidly perform way- finding. For this reason, we use two different approaches for represent our space, which are cell graphs and portal graph. These two approaches of abstract representation can also be used to store some pre-computed information about the environment that will speed up the naviga- tion and also be helpful to achieve fast perception for local motion computation. This virtual environment is similar to our real world, filled with an important number of intelligent entities (e.g. virtual agents, autonomous agents, intelligent objects). The aim of our work is to create virtual humans as intelligent entities in these space, which includes approximate the maximum as possible the virtual agent animation to the natural human beha- vior. In order to accomplish this task, our agent must be capable to interact with the environment, interacting with objects and other agents. The virtual agent needs to act as real person, so he should be capable to extract se- mantic information from the geometric model of the world where he is inserted, based on his own perception, and he realizes his own decision. The movement of the individuals is representing by the combination of two approaches of movement which are, the force social model and the rule- based model. These movements are influenced by a set of socio-psychological rules to give a more realistic result. 2 RELATED WORK ————————————————  Chighoub Rabiaa, Computer science department, LESIA, Biska University, BP 145, Biskra 07000.  Cherif Foudil, Computer science department, LESIA, Biska University, BP 145, Biskra 07000. JOURNAL OF COMPUTING, VOLUME 2, ISSUE 4, APRIL 2010, ISSN 2151-9617 HTTPS://SITES.GOOGLE.COM/SITE/JOURNALOFCOMPUTING/ Computers power increase recently allowed to populate interactive virtual worlds [9] with numerous inhabitants [8]. Crowds are now common in movies and more and more in video games, especially in real-time strategy games. Simulation of real-time virtual crowds is still a difficult challenge given that available computation-time is mainly dedicated to rendering; a need for fast simulation techniques exists. The topic of modeling of pedestrian streams is not new and has been done for many years. The previous focus was on modeling pedestrian streams in urban environments. In these earlier studies, the aim was to determine the dimension of the parameters of walk- ways. Later the scope was extended to the field of emer- gency. In the 90’s the simulation of pedestrian streams was integrated in the simulation of intermodal transport facili- ties. Nowadays, there exist quite a three different ap- proaches for modeling pedestrian streams. 1. Macroscopic simulation: Models of this category are field-based simulation models that only deal with densities and flux. The approach describes pedestrian flows with differential equations, and is based on the idea that the movement of pede- strians can be handled analogous to fluids and gases. 2. Microscopic simulations: these models consider the movement of pedestrians at the individual lev- el [1]. They treat pedestrians as individuals and model the individual’s behavior with the expecta- tion that dynamic crowd behaviors will emerge from the interactions between individuals and the environment. 3. Mesoscopic simulation: Models of this approach are a mixture of macroscopic and microscopic models, simulating not individual entities but groups of similar entities, The idea of grouping individuals was transformed to mesoscopic pede- strian flow model. That means we do not model a single pedestrian but we use groups of pedestrians and every group has its own rules of behavior. The flow of a single pedestrian is integrated into a flow of pedestrian groups. This simplification is valid because one of the main interesting results from the simulation model is not the state of a single person but the number of persons in a particular area at time t. Figure 1 shows the classification of the mesoscopic approach in relation to the other approaches. In our work, we interest with microscopic approach, many microscopic simulations already exist, and one way to group them is to divide them into cellular automata models, behavioral force models, and rule-based models. The key difference between these techniques is the me- thod in which the entity is represented, how the entity is controlled, and how it interacts with other entities. The social forces model [5, 6] calculates forces acting on agents to determine movement, with excessive forces leading to agent injuries. The model considers the effect that each agent has upon all the other agents, almost as if the model were a simulation of an n-body problem in astrophysics. Physical forces (e.g. friction encountered 49 Fig. 1. Relations of three approaches of modeling pedestrian when brushing past another person, or elastic force due to body compressions) are modeled, as are social forces (desire to change direction to avoid another). One problem with Helbing’s model is that of computa- tional complexity. Simulation update is O(n2) due to the calculation of the effect that each agent (and obstacle) has on all the other agents. This may limit the model’s ability to simulate many agents. Braun et al [2] expand Helbing’s social force model. In their model, each pede- strian is assigned a ‘family’ identifier and an ‘altruism’ level. These acts as forces in the model to tend to make some of the pedestrians form groups with others. Reynolds [15] described a distributed behavior model for simulating flocks of birds formed by actors en- dowed with perception skills. In fact, the birds (or ‘boids’) maintain proper position and orientation within the flock by balancing their desire to avoid collisions with neighbors, to match the velocity of neighbors and to move towards the center of the flock. Reynolds’s work shows realistic animation of groups by applying simple local rules within the flock structure. Following this ap- proach, Terzopoulos et al. [20] presented the simulation of virtual fish groups, where each fish is endowed with perception (artificial vision), locomotion control (based on a mass-spring system used to propel the fish in the water) and with behaviors (based on a group of parameters, such as hunger degree and predators fear). 3 ARCHITECTURE OVERVIEW We have developed a realistic simulator of high-density crowds of autonomous virtual agents composing of an important number of pedestrians autonomous in recon- structed large environments, demonstrating realistic hu- man activity. In this context, a microscopic architecture has been proposed for handling high-density crowds of auto- nomous agents moving in a natural manner in dynamically changing virtual environments filled with obstacles. JOURNAL OF COMPUTING, VOLUME 2, ISSUE 4, APRIL 2010, ISSN 2151-9617 HTTPS://SITES.GOOGLE.COM/SITE/JOURNALOFCOMPUTING/ Model of environment 50 A g e n t ’ s S o c i o - P s y c h o l o g i c a l s t a t e Agent n Agent 1 Group 1 Group L Virtual Interaction Emergent bihavior Agent2 Agent K C o l l e c v t i i v o e r b e h a - I n d i v i d v u i o a r l B e h a - Agent Agent2 Agent m Agent 1 Fig. 2. Our system architecture In this architecture (proposal), a virtual agent can be de- fined as an autonomous entity in a virtual environment. It should not only look like, but also behave as a living organ- ism in a synthetic 3D world, and be able to interact with the world and its inhabitants. The virtual agent must re- ceive its environment through sensors or perception, it receives from the environment the list of entities (agents, obstacles…) that are in its field of view, so he should be capable to extract semantic information from the geometric model of the world where he is inserted, and based on his own perception, he decides its decision and acts as real people. The complexity of pedestrian behavior comes from the presence of collective behavioral patterns (as clustering, lanes and queues) evolving from the interactions among a large number of individuals. This empirical evidence leads to consider two different approaches: pedestrians as a flow and pedestrians as a set of individuals or agents. 3.1 Model of environment In any location, and particularly in a city, virtual humans need to be aware of their environment in order for them to avoid any collision, know where their goals are, and how to reach them. Many different approaches have been de- veloped to tag an environment with information. Our sys- tem can handle two approaches of environment represen- tation, grid and topological graph, by combining both pa- radigms – grid-based and topological—the approach pre- sented here gains the best of both worlds: accura- cy/consistency and efficiency. 1. Grid-based approach, represent environments by evenly spaced grids [18, 17]. Each grid cell may, for example, indicate the presence of an obstacle in the corresponding region of the environment. This approach must be fine enough to capture every important detail of the world [7, 19]. 2. Topological approaches represent environments by graphs. Nodes in such graphs correspond to distinct situations, places, or landmarks (such as doorways) [19]. They are connected by arcs if there are direct paths between them. [18, 17]. Conse- quently, they permit fast planning. The process of representation of the environment needs: 1. Our system receives as an input an arbitrary build- ing model in 3 dimensions and creates grid map. To build the grid graph, by cutting out the envi- ronment a unit of cells, then; one encapsulates all information (static) necessary to the process of na- vigation. 2. Once we have the grid decomposition, we start an iterative conquering process starting from the top left corner cell that is empty. We assign a positive number to this cell that will represent the room ID in the cell and portal graph, and then this ID is propagated using a breadth-first traversal. The propagation of the cell ID continues until the en- tire room is bounded by cells having either 0 (wall) or -1 (door) [10]. Fig. 3. Model of environment Association of information Extraction of a structure in 2d Informed Topological map Informed Grid-map Topological map Environment in 3D Grid-map 3. Once all the cells have been identified, we need to generate the cell and portal graph by joining the rooms through the doors. This is car- ried out by traversing the grid representation from left to right, top to bottom, looking for doors. When a door is found, a portal is created that will join the two cells appearing at both sides of the door [11]. In the last step, points of attraction will be asso- ciated at each door and at each painting. In order to perform global navigation, we also need to store information about paths within the building from each cell to each of the exits in the building. 4. JOURNAL OF COMPUTING, VOLUME 2, ISSUE 4, APRIL 2010, ISSN 2151-9617 HTTPS://SITES.GOOGLE.COM/SITE/JOURNALOFCOMPUTING/ 51 Model of representation of environnement Each cell will contain one or more alternative paths to each exit [11]. 3.2 Model of agent Our simulation consists of high-density crowd of auto- nomous virtual human agents existing in dynamic, com- plex, virtual 3d environment. In order to behave in be- lievable way these agents must act in accordance with their surrounding environment, be able to react to its changes, to the other agents and also to the actions of real humans interacting with the virtual world. Figure 4 gives overview of the agent model, this model of the agent follow the Sense-Decide-Act (SDA) cycle, where it can be broken up into subsystems, each one associated in turn with more specific routines. From the perception sub- module, the individual perceive the environment accord- ing to its position, personality and knowledge, and he obtains the environment information on a semicircular front region of the agent, it detects the positions, orienta- tions and speeds of other agents, and obstacles. Moreo- ver, the agent cans predicted future positions changes in speed and orientation of each character and of objects in the VE. Fig.4. Model of agent Based on information perceived, the internal state and the socio-psychological traits of the agent (current beha- vior, panic, role, impatience…), the Decision sub-model calculates the orientation, the velocity and next position of the agent, finally the Motion sub-model execute and realize this decision. 3.2.1 Perception Humans can perceive a field of view (FOV) from 120° to 180°, the latter being the most common we can simulate human perception by having the virtual agents be only aware of those objects falling within a specified angle from their direction of movement (assuming the head is oriented in the same direction). This is calculated from the dot Position of obstacles, and position of agents, orientation of agents, distance between goals, others agents, obstacles and points of attraction. A g e n t ’ s S o c i o - P s y c h o l o g i c a l s t a t e Perception Motion New position, velocity and orientation Decision product between the direction of movement vector and the vector joining the current position with each object in the room. Since the dot product gives us the cosi that value is bigger than 0 it means the object falls within the agent’s FOV [21]. In figure 5, the agent perceives its environment; and determines which object exists in its zone of vision, wall, the obstacle and another agent J. 3.2.2 Motion The main motivation for this model is the fact that differ- ent people can react in different ways depending on their individual characteristics and on their psychological traits. Pedestrian movement shows the characteristics of a pedestrian compared to cars or other vehicles such as a bicycle, a pedestrian has more flexibility to move in two dimensions continue space, a pedestrian also has more flexibility to stop and go and a pedestrian also has the smallest velocity and the highest fluctuation in accelera- tion and velocity compared to vehicles. In our system, which consists in simulating a dense human crowd, the local movements of agents are inspired by the combina- tion of two approaches which are the approach based- rules and the approach of social force. Reynolds [15, 16] described a distributed behavioral model for simulating flocks of birds formed by actors endowed with perception skills with many purposes. In fact, the birds (or boids) maintain proper position and orientation within the flock by balancing their desire to avoid collisions with neigh- bors, to match the velocity of neighbors and to move toward the center of the flock. Reynolds work shows realistic animation of groups by applying simple local rules within the flock structure. This model gives results realize for a crowd with low density. To mitigate the insufficiencies of the approach containing rules, we com- bined this approach with the approach of social force inspired of the model of Helbing [5]. It is based on phys- ics and the socio-psychological forces in order to describe the human behavior of crowd in situations of panic, this model can simulate a crowd with high density but does not give realistic results. The displacement of the agents is modeled by a whole of social forces, and then Nuria [12, 13] hears this model by the incorporation of psycho- logical and geometrical rules. Therefore the module of movement of our pedestrian consists in calculating the new position and the new for each agent by using a certain force of attraction and re- pulsion, these forces make it possible the agent to show a broad variety of behaviors in an individual or collective way. 1. Advance Force: in the normal situation and in the absence of other pedestrians, the movement of a pedestrian should be directed from the current lo- cation P(t) toward the destination point E(t), there should be a “advance force” that directs the pede- strian to move, the advance force makes the pede- strian path almost in a straight line. So the agent perceives the environment and if its zone of vision does not contain obstacles or pedestrians, then it advances easily towards its goal by using its stable speed. Agent i Agent j Wall Agen t H JOURNAL OF COMPUTING, VOLUME 2, ISSUE 4, APRIL 2010, ISSN 2151-9617 HTTPS://SITES.GOOGLE.COM/SITE/JOURNALOFCOMPUTING/ Fig.5. Process of perception 2. Acceleration Force: this behavior is influenced by the existence of panic, in this case if the zone of vi- sion of our agent is empty, contains neither ob- stacles nor other virtual humans, the agent aug- ments its speed, and is directed towards the door by using a maximum speed. V Obstacle Vi Agen Vi Agent Point of attraction Point of attraction (goal or door) Fig.6.The Goal-attractive Force 3. The Occupant-attractive Force: because of the complex relationship among people, the cluster phenomena are observed, otherwise, people in danger tend to move towards the area with a large population, when the “follow” phenomena are of- ten observed, which means occupants tend to fol- low the movement of majority. In the model, the occupant-attractive force is considered between some specified occupants when they are sufficient- ly far from each other (the distance between any two of them is no less than r1, and occupants may choose the same moving direction as most people have been taken around him (within a distance r2). 4. The Goal-attractive Force: In normal situation, pe- destrians are sometimes attracted by window dis- plays, sights, special performances (street artists), or unusual events at specified places (in certain area), otherwise, in the process of evacuation, oc- cupants are attracted by the doors (in our model, the doors are marked by points of attractions). Both situations can be modeled by (often tempo- rally decaying) attractive forces, in a similar way like effects repulsive, but with an opposite sign and a longer range of the interactions, it is general- ly assigned according to exits' quite selected in the 52 building. The magnitude of attraction force is in- fluenced by the type of situation (panic or normal), and the distance between the agent and the door (point of attraction), if this distance is short, the force is large. V Vi Agent i Agent j Agent j Personal S Fig.7.The Occupant-Repulsive Force 5. The Occupant-Repulsive Force: This repelling power works to avoid the collision between the pedestrians. To avoid the collision, it is supposed that each pedestrian has a ray of influence which represents its conscience of safety. The force is produced when the rays of influences of pede- strians overlap 6. The Obstacle-Repulsive Force: The response of col- lision, which is ensured by the repulsive forces, against the obstacles and the walls applies when the personal space of an agent covers with their zones. The covering of wall occurs when the dis- tance between the center of the agent and the cen- ter of the wall is smaller than the personal space of the agent. Repulsive Force Agent i Vi Vj 4 RESULTS AND PERFORMANCES We have chosen the museum like an example of applica- tion of our system, the museum is regarded as a closed environment that its structure topology is represented by a unit of room of different size, connected between them by doors and passageways. Each room is generally to contain obstacles distributed on its surface, and it is deco- rated by a unit with the images suspended on the walls and by trimmings positioned in various localizations, in this environment our virtual agents, the normal situation, can move and see a number preset of the images and the trimmings which decorate the museum, but if the situa- JOURNAL OF COMPUTING, VOLUME 2, ISSUE 4, APRIL 2010, ISSN 2151-9617 HTTPS://SITES.GOOGLE.COM/SITE/JOURNALOFCOMPUTING/ tion changes, i.e. an alarm announces the event of panic, all the human virtual ones try to find the best way to leave the environment. Generally the pedestrian move- ments are influenced by a set of psychological and socio- logical rules which play a role of balancing of rationality in the behavior and the decision-making that we use to illustrate a broad more realistic variety of behaviors. Sev- eral situations have been studied to present the influence of socio-psychological rules in the realism of the beha- viors human; we can resume the following situations: 1. The behavior of avoidance of collision and the avoidance of obstacles, these two behaviors are af- fected by the type of situation (normal and of pan- ic), the personal space of the agent (broad, aver- age, and narrow), and finally the level of patience of agent (patient and impatient). 2. Attractions forces, the agent are guided by three gravitational attractions which are Attraction force towards a point of attraction representing an un- der-goal (painting), Attraction force towards an exit (door), Attraction force towards another indi- vidual. This force is influenced by the distance be- tween the individual and the goal (painting, doors or another agent), the distance which must respect between the agent and its goal, the situation (nor- mal or of panic), knowledge of pedestrian (com- plete or partial), and the role of pedestrian (guide, or follower). L’espace personnel The agent changes its direction to avoid the collision The agent advances in its new direction, the change is rather fine since personal space is narrow (a) The avoidance of normal collision in case, such as personal space is narrow, and the impatient agent. 53 Personnel Space Advance in the new direction Change of direction to avoid the collision (b) The avoidance of collision, such as personal space is average, and the impatient agent, the collision is treated with a broad distance (2 meter). The patient agent changes its direction at a short distance, and it advances in its new direction (c) The avoidance of collision, such as personal space is narrow, and the patient agent, the collision is treated with a short distance (1 meter). Fig. 8. The behavior of avoidance of collision influenced by socio-psychological state A patient agent detects the collision, but it advances does not avoid it Personnel Space JOURNAL OF COMPUTING, VOLUME 2, ISSUE 4, APRIL 2010, ISSN 2151-9617 HTTPS://SITES.GOOGLE.COM/SITE/JOURNALOFCOMPUTING/ 54 5 CONCLUSION We have proposed a microscopic model of simulation of a crowd with high density of human vrtuels in a dynamic environment. We started with the description of an envi- ronment of simulation, not only from point of view topo- graphic, but also from the semantic aspect. For this, we used the two approaches of representation of the envi- ronment knowing, the graph of grid and the topological graph. The originality of our work is the introduction of a number of socio-psychological rules influencing the be- havior of crowd. The integration of these rules in our model and the combination of the approach based rules and the approach of social force have conducted to realis- tic simulations. Our system gives acceptable results, but optimizations and a whole of prospect remain to be supplemented like: 1. The addition of a cognitive layer in the human be- haviors, the training in its process of navigation and the reasoning. 2. To integrate a responsible module to produce movements realistic, like walk, to run it, etc.... 3. One can also introduce the concept of group and the behaviors collective, the agents that are very close can build groups and these groups are stu- died like only one entity, to minimize the compu- ting time. REFERENCES [1] V.J. Blue and J.L. Adler, “Cellular automata microsimulation of bi- directional pedestrian flows”, Transportation Research Record, Journal of the Transportation Research Board, pp. 135—14, 2000. [2] A. Braun, S.R. Musse, L.P.L de Oliveira, and B.E.J. Bodmann, “Model- ing individual behaviors in crowd simulation.”, CASA2003, 16th Inter- national Conference on Computer Animation and Social Agents., Rutgers University, New Brunswick, New Jersey, USA, pp. 143–148. Mai 2003. [3] S. Bandi and D. Thalmann, “Space Discretization for Efficient Human Navigation”, Computer Graphics Forum, 17(3), pp. 195-206, 1998. [4] N. Farenc., R. Boulic and D. Thalman, “An informed environment dedicated to the simulation of virtual humans in urban contex”, In Pro- ceedings of EUROGRAPHICS1999, Annual Conference Series, pp 309- 318, 1999. [5] D. Helbing, I. Farkas and T. Vicsek, “Simulating dynamical features of escape panic”. Nature, 407(1), pp 487–490, 2000. [6] D. Helbing, I. Farkas, P. Molnar and T. Vicsek, “Simulating of Pede- strian Crowds in Normal and Evacuation Situations”, M.Schreckenberg, S.D. Sharma (ed.) Pedestrian and Evacuation Dy- namics. Springer Verlag Berlin and Heidelberg, pp 21-58, 2001. J.J. Kuffner, “Goal-directed navigation for animated characters using real-time path planning and control”. Lecture Notes in Computer Science, 1537, pp 171–179, 1998. [8] F. Lamarche, and S. Donikian, “Crowd of Virtual Humans: a New Approach for Real Time Navigation in Complex and Structured Envi- ronments”, Computer Graphics Forum 23(3): pp 509-518, 2004. [7] [9] S.R. Musse and D. Thalmann, “Hierarchical Model for Real Time Simulation of Virtual Human Crowds”, IEEE Transactions on Visuali- sation and Computer Graphics, vol. 7, n.2, pp 152-164. April-June, 2001. [10] S. Paris,, “ Caractérisation des niveaux de services et modélisation des circulations de personnes dans les lieux d’échanges” , PhD disserta- tion, Dept. of, Rennes I Univ., 10 octobre 2007. [11] N. Pelechano, “Modeling Realistic High Density Autonomous Agent Crowd Movement: Social Forces, Communication, Roles And Psycho- logical Influences”, PhD dissertation, Dept. of, Pennsylvania Univ., 2006. [12] N. Pelechano, K. O'Brien, B. Silverman, and N. Badler, “Crowd simula- tion incorporating agent psychological models, roles and communica- tion”. First Int. Workshop on Crowd Simulation, pp 21-30, 2005. [13] N. Pelechano, K. O'Brien, B. Silverman, and N. Badler, “Controlling Individual Agents in High-Density Crowd Simulation”. University of Pennsylvania, Eurographics/ACM SIGGRAPH Symposium on com- puter Animation, 2006. [14] S. Paris, J. Pettré, and S. Donikian, “Pedestrian Reactive Navigation for Crowd Simulation: a Predictive Approach”, EUROGRAPHICS 2007 / D. Cohen-Or and P. Slavík, 2007. [15] C.W. Reynolds, “Flocks, herds and schools: a distributed behavioural model”, SIGGRAPH’87, Vol. 21(4) of Computer Graphics, ACM Press, Anaheim (USA), pp 25-34 1987. [16] C.W. Reynolds, “Steering Behaviors For Autonomous Characters”, GDC 99. A. Yu (Ed.) Miller Freeman, San Fransisco, Pp 763-782 1999. [17] W. Shao, “Animating Autonomous Pedestrians”, PhD dissertation, Dept. of, New York Univ. 2006. [18] W. Shao and D. Terzopoulos, “Environmental modeling for autonom- ous virtual pedestrians”. 2005 SAE Symposium on Digital Human Modeling for Design and Engineering, 2005. [19] S. Thrun and A. Bücken, “Integrating grid-based and topological maps for mobile robot navigation”. Proc. of the AAAI Thirteenth National Conference on Artificial Intelligence, pp 944–951. AAAI Press / MIT Press, 1996. [20] X. Tu and D. Terzopoulos, “Artificial Fishes: physics, locomotion, perception, behavior”, SIGGRAPH 94 Conference Proceedings, pp 43- 50, Orlando, FL, USA, 1994. [21] M. Villamil, S. Musse, and L. Luna, “A model for Generating and Animating Groups of Virtual Agents”, IVA, Kloster-Irsee, Allmend, 2003. Dr. Cherif Foudil. is currently working as an Associate Professor of computer science at Computer Science Department, Biskra Uni- versity, Algeria. Dr. Cherif holds PhD degree in computer science. The topic of his doctoral dissertation was Behavioral Animation: simulation of a crowd of virtual humans. He also possesses B. Sc. (engineer) in computer science from Constantine University 1985, M. Sc. in computer science from Bristol University, UK 1989. He is currently the head of “animation and artificial life” team in LESIA laboratory. His current research interest is in artificial intelligence, artificial life, crowd simulation, behavioural animation Chighoub Rabiaa. A Phd candidate in computer science, her cur- rent research interest is crowd simulation. © 2010 Journal of Computing http://sites.google.com/site/journalofcomputing/
0708.3220	1	0708	2007-08-23T17:53:54	Optimal strategies in the average consensus problem	[ "cs.MA", "cs.NI", "math.OC" ]	We prove that for a set of communicating agents to compute the average of their initial positions (average consensus problem), the optimal topology of communication is given by a de Bruijn's graph. Consensus is then reached in a finitely many steps. A more general family of strategies, constructed by block Kronecker products, is investigated and compared to Cayley strategies.	cs.MA	cs	Optimal strategies in the average consensus problem Jean-Charles Delvenne Ruggero Carli Sandro Zampieri 7 0 0 2 g u A 3 2 ] A M . s c [ 1 v 0 2 2 3 . 8 0 7 0 : v i X r a Abstract- We prove that for a set of communicating agents to compute the average of their initial positions (average consensus problem), the optimal topology of communication is given by a de Bruijn's graph. Consensus is then reached in a finitely many steps. A more general family of strategies, constructed by block Kronecker products, is investigated and compared to Cayley strategies. I. INTRODUCTION Coordination algorithms for multiple autonomous vehicles and decentralized estimation techniques for handling data coming from distributed sensor networks have attracted large attention in recent years. This is mainly motivated by that both coordinated control and distributed estimation have applications in many areas, such as coordinated flocking of mobile vehicles [26], [27], cooperative control of unmanned air and underwater vehicles [4], [3], multi-vehicle tracking with limited sensor information [19], monitoring very large scale areas with fine resolution and collaborative estimation over wireless sensor networks [24]. Typically, both in coordinated control and in distributed estimation the agents need to communicate data in order to execute a task. In particular they may need to agree on the value of certain coordination state variables. One expects that, in order to achieve coordination, the variables shared by the agents, converge to a common value, asymptotically. The problem of designing controllers that lead to such asymptotic coordination is called coordinated consensus, see for exam- ple [12], [20], [15] and references therein. Generalisation to high order consensus [22] and nonholonomic agents [18], [11], [28] have also been explored. One of the simplest consensus problems that has been mostly studied consists in starting from systems described by an integrator and in finding a feedback control yielding consensus, namely driving all the states to the same value [20]. The informa- tion exchange is modeled by a directed graph describing in which pair of agents the data transmission is allowed. The situation mostly treated in the literature is when each agent has the possibility of communicate its state to the other agents that are positioned inside a neighborhood [26], [15] and the communication network is time-varying [27], [15]. Robustness to communication link failure [8] and the effects of time delays [20] has been considered recently. Randomly time-varying networks have also been analyzed in [14]. Moreover, a first analysis involving quantized data J.-C. Delvenne is with the Institute for Mathematical Sciences, 53 Prince's Gate, London SW7 2PG, United and Information Engineering, Italy Imperial College, Kingdom, [email protected]. R. Carli S. Zampieri are with the Department of Universit`a 6/a, {carlirugzampi}@dei.unipd.it. di Padova, Via Gradenigo 35131 Padova, transmission has been proposed in [7], [16]. In this paper we consider the consensus problem from a different perspective. We are interested to characterize the relationship between the amount of information exchanged by the agents and the achievable control performance. More precisely we assume that N agents are given initial positions in the euclidean space, and move in discrete-time in order to reach the average of their initial positions. This problem is also called average coordinated consensus. Every agent asks several agents their position before taking a decision to modify its own position. We impose that, in order to limit costs of communication, every agent communicates with only ν agents (including itself), where ν < N . This means that in the graph describing the communications between agents, the max in-degree is at most ν. In this paper, we exhibit a family of strategies for solving this problem based on de Bruijn's graphs and we prove that according to a suitable criteria this is the best that one can do. Precisely we compute its performances according two criteria: rate of convergence to the average of their initial positions and an LQR criterion. We find that a deadbeat strategy is optimal according to the rate of convergence, and nearly optimal according to the LQR criterion. Finally, we compare it with an another strategy having limited communication and exhibiting symmetries: the Cayley strategies [6], [5]. It should be noted however that our strategy is limited to the case where the number of agents is an exact power of ν. Whether it is possible to build a linear time-invariant deadbeat strategy for any number of agents (for a given ν) remains an open problem. The paper is organized as follows. In Section II we provide some basic notions of graph theory and some notational conventions. In Section III we formally define the average consensus problem. In Section IV we introduce the block Kronecker strategy. In Section V we show that the block Kronecker strategy is the quickest possible strategy and we compare it with the Cayley strategy. In section VI we evaluate the performance of the block Kronecker strategy according to suitable quadratic criteria. Finally we gather our conclusions in Section VII. II. PRELIMINARIES ON GRAPH THEORY Before defining the problem we want to solve, we sum- marize some notions on graph theory that will be useful throughout the rest of the paper. Let G = (V,E) be a directed graph where V = (1, . . . , N ) is the set of vertices and E ⊂ V × V is the set of arcs or edges. If (i, j) ∈ E we say that the arc (i, j) is outgoing from i and incoming in j. The adjacency matrix A is a {0, 1}-valued square matrix indexed by the elements in V defined by letting Aij = 1 if and only if (i, j) ∈ E. Define the in-degree of a vertex j as Pi Aij and the out-degree of a vertex i as Pj Aij . In our setup we admit the presence of self-loops. A graph is called in-regular (out-regular) of degree k if each vertex has in-degree (out-degree) equal to k. A path in G consists of a sequence of vertices i1i2 . . . . . . ir such that (iℓ, iℓ+1) ∈ E for every ℓ = 1, . . . , r − 1; i1 (resp. ir) is said to be the initial (resp. terminal) vertex of the path. A cycle is a path in which the initial and the terminal vertices coincide. A vertex i is said to be connected to a vertex j if there exists a path with initial vertex i and terminal vertex j. A directed graph is said to be connected if, given any pair of vertices i and j, either i is connected to j or j is connected to i. A directed graph is said to be strongly connected if, given any pair of vertices i and j, i is connected to j. Finally some notational conventions. Let A any matrix belonging to RN ×N . With Tr A we denote the trace of A, i.e. the sum of the diagonal entries. We say that A is nonnegative, denoted A ≥ 0, or positive, denoted A > 0, if the entries of A are respectively nonnegative or positive. III. PROBLEM FORMULATION We suppose that the positions of all N agents are listed into one vector of dimension N . If the agents move, say, in R3, it seems that we would need a 3N -dimensional vector. However we will suppose that the positions are scalar, as every linear strategy on scalar positions, if applied separately on every component of the position, trivially extends to strategies for higher dimensions. More precisely the problem of our interest can be formal- ized in the following way. Consider N > 1 identical systems whose dynamics are described by the following discrete time state equations x+ i = xi + ui i = 1, . . . , N where xi ∈ R is the state of the i-th system, x+ represents the updated state and ui ∈ R is the control input. More compactly we can write i (1) where x, u ∈ RN . The goal is to design a feedback control law x+ = x + u u = Kx, K ∈ RN ×N yielding the average consensus, namely a control such that all the xi's become asymptotically equal to the average of the initial condition. More precisely, our objective is to obtain K such that, for any initial condition x(0) ∈ RN , the closed loop system yields x+ = (I + K)x, lim t→∞ x(t) = α1 where 1 := [1, . . . , 1]T and α = 1 N 1T x(0). (2) (3) Writing x(t) as a linear combination of the eigenvectors of I + K, it is almost immediate to see that the average consensus problem is solved if and only if the following three conditions hold: (A) Every row and every column of I + K sums to one. Hence it has eigenvalue 1 with 1 as left and right eigenvector. (B) The eigenvalue 1 of I + K has algebraic multiplicity one (namely it is a simple root of the characteristic polynomial of I + K). (C) All the other eigenvalues are strictly inside the unit circle. For nonnegative matrices, namely for matrices having all the components nonnegative, condition (A) is called double stochasticity, condition (B) is ergodicity and condition (C) is a consequence of double stochasticity. We do not require our matrices to be nonnegative, even though it will appear that the optimal matrices are. Observe now that the fact that the element in position i, j of the matrix I + K is different from zero, means that the system i needs to know exactly the state of the system j in order to compute its feedback action. This implies that the j-th agent must communicate his state xj to i-th agent. In this context a good description of the communication effort required by a specific feedback K is given by the directed graph GI+K with set of vertices {1, . . . , N} in which there is an arc from j to i whenever in the feedback matrix K the element (I + K)ij 6= 0. The graph GK is said to be the communication graph associated with K. Conversely, given any directed graph G with set of vertices {1, . . . , N}, a feedback K is said to be compatible with G if GI+K is a subgraph of G (we will use the notation GI+K ⊆ G). In the sequel, we will impose the following constraint on the communication graph: the max in-degree of the nodes is ν. This models the fact that communication lines are costly to establish or operate, and every agent has the right to talk to a limited number of other agents. Note that for compatibility with usual conventions we consider that ν counts all arcs entering a node, including self-loops (which could be considered as 'free communication' in most technological situations). Without this constraint, the problem becomes trivial: choose the complete graph, and the consensus is reached in one step. We therefore add the following constraint on I + K: (D) Every row of I + K contains at most ν non-zero elements. From this point of view we would like to obtain a matrix I + K satisfying (A),(B),(C),(D) and minimizing a suitable performance index. The simplest control performance in- dex is the exponential rate of convergence to the average consensus. When we are dealing with average consensus controllers it is meaningful to consider the displacement from the average of the initial condition ∆(t) := x(t) −(cid:18) 1 N 1T x(0)(cid:19) 1 . It is immediate to check that, ∆(t) = x(t) −(cid:0) 1 (since the average position 1 N t) and that it satisfies the closed loop equation N 1T x(t) is the same at all times 1T x(t)(cid:1) 1 ∆+ = (I + K)∆ . (4) Notice moreover that the initial conditions ∆(0) are such that 1T ∆(0) = 0 . (5) Hence the asymptotic behavior of our consensus problem can equivalently be studied by looking at the evolution (4) on the hyperplane characterized by the condition (5). The speed of convergence toward the average of the initial condition can be defined as follows. Let P any matrix satisfying conditions (A),(B),(C). Define ρ(P ) = (cid:26) 1 maxλ∈σ(P )\{1} λ if dim ker(P − I) > 1 if dim ker(P − I) = 1 , which is called the essential spectral radius of P . As the dominant eigenvalues of P t is one and the others are smaller in magnitude than ρ(P )t, the essential spectral radius says how quickly P t converges to the rank-one matrix 1/N 11T , where N is the dimension of P . In this context the index ρ(I + K) seems quite appropriate for analyzing how per- formance is related to the communication effort associated with a graph. The smaller the essential spectral radius, the quicker the system will converge to the average of the initial condition. However in control theory, strategies that converge in finite time or very quickly are sometimes dismissed on the ground that they lead to large values of update values u(t) = x(t + 1) − x(t), that can be physically impossible or very costly to implement. Hence a strategy is often required to optimize an LQR cost, taking into account both the quickness convergence and the norm of updates values. Therefore another suitable measure of performance could be the following quantity: J = E(Xt≥0 x(t) − x(∞)2 + γu(t)2), (6) where x(t) is the vector of positions at time t, x(∞) = lim∞ x(t) is the vector whose every entry is the average of initial positions, u(t) = x(t + 1) − x(t) is the update vector at time t, the initial positions are supposed to be uncorrelated random variables with unit variance, E denotes the expectation, x2 = xT x is the euclidean norm and γ is a nonnegative real. We will prove that the optimal topology of communication (in the meaning of speed of convergence) is given by a de Brujin's graph. We will call the control strategies based on such graph block Kronecker strategies, as explained in the next section. For these strategies we will evaluated (6) and we will compare them to another family of strategies based on a regular communication graph having the same degree ν: the Cayley strategies [6], [5]. IV. BLOCK KRONECKER STRATEGIES In this section, we define block Kronecker strategies. Let A be a n × n matrix satisfying (A),(B),(C),(D) and k be a nonnegative integer. Note that if A is full then n ≤ ν (since the number of non-zero elements cannot exceed ν). Then we build an nk × nk matrix M in the following way. Let A =   a0 a1 ... an−1   be a row-partition of the matrix A, where ai ∈ R1×n. Then M is the matrix M = For example, if .   (7) ... Ink−1 ⊗ a0 Ink−1 ⊗ a1   Ink−1 ⊗ an−1 A = (cid:18)α β β α(cid:19) (with α + β = 1) and k = 3, then M =   α β α β α β α β β α β α β α β α   This is a kind of block Kronecker product. A general theory of block Kronecker product is built in [17]. We only need a more restricted definition, detailed below. The new matrix M is a matrix of larger dimension than A and satisfying conditions (A),(B),(C),(D): (A) and (D) follow from the definition, while (B) and (C) are proved below. Hence it can play the role of the matrix I + K in Section III.We start by some reminders on Kronecker product, define the block Kronecker product and explore the properties of the latter. A. Kronecker product We recall that the Kronecker product A⊗B of the matrices A and B is the matrix [aijB]i,j, whose dimensions are the product of dimensions of A and B. Some useful properties of the Kronecker product are the following: • AB ⊗ CD = (A ⊗ C)(B ⊗ D); • Tr A ⊗ B = Tr ATr B; • the eigenvalues of A ⊗ B are all possible products of • the eigenvectors of A ⊗ B are all possible Kronecker products of an eigenvector of A with an eigenvector of B. an eigenvalue of A with an eigenvalue of B; The Kronecker product is sometimes called tensor product. Let us see why. For instance consider the matrices B, C, D of sizes mB × nB, mC × nC, mD × nD. The Kronecker product has size mBmC mD × nBnCnD, and an arbitrary element of B⊗ C⊗ D can be denoted (B⊗ C⊗ D)abc,def = BadCbeDcf , where the index written as abc denotes the number c + bmD + amCmD and the index def is the number f + enD + dnCnD; we suppose that the indices start form zero: a = 0, . . . , mB − 1, etc. If B, C, D happen to be square matrices of size n, this notation coincides with the usual notation in base n of an index running from 0 to n3 − 1. This notation of the Kronecker product is very close to the tensor product used in algebra and differential geometry. The only difference is that B ⊗ C ⊗ D, viewed as a tensor product, is considered as a 6-dimensional array with a, b, c, d, e, f as separate indices, instead of a matrix (i.e., a 2-dimensional array). All this immediately extends to more than three matrices. B. Block Kronecker product Let us now consider the following variant of Kronecker product, that we call block Kronecker product. Consider for instance two matrices B (of size n3 × n3) and C (of size n2×n2). The block Kronecker product of B and C is defined as follows: its element of index abcde, ghijk is the element Bcde,ghiCab,jk (notice the shift of the first indices by two places). We will denote this matrix by B⊙C. This definition applies to any two square matrices whose dimensions are powers of n. In general, we can write (B ⊙ C)p,q = (B ⊗ C)σt(p),q, where σ operates a cyclic permutation by one place to the left on the digits of p in base n, and C is of size nt. The matrix M defined by Equation (7) can be expressed as M = (I ⊗ ··· ⊗ I) ⊙ A (where the n × n identity matrix I is repeated k − 1 times). If we write the index of M in base n, then Mi1...ik,j1...jk = Ii2,j1 Ii3,j2 . . . Iik−1,jk Ai1,j0. This form is particularly useful to compute the behavior of M from the properties of the block Kronecker product, which we now explore. As a first property, we can easily see that (B ⊙ C)T = CT ⊙ BT . We can also prove the following lemma. (8) Lemma 4.1: For any matrices A, B, C, D, E, F for which all the products below are meaningful, we have ((A ⊗ B) ⊙ C)((D ⊗ E) ⊙ F ) = BD ⊙ (CE ⊗ AF ). (9) Proof: We write, using Einstein's convention (indices repeated twice in an expression are implicitly summed over), [((A ⊗ B) ⊙ C)((D ⊗ E) ⊙ F )]u,w = = ((A ⊗ B) ⊙ C)u,v((D ⊗ E) ⊙ F )v,w = Au2,v1 Bu3,v2 Cu1,v3Dv2,w1Ev3,w2Fv1,w3, where u, v, and w, interpreted as sequences of digits in base n, have been partitioned into u1u2u3, v1v2v3, and w1w2w3 in an appropriate way. This is possible if B and D have same size, as well as C and E, and A and F . Then the expression above can be regrouped as (BD)u3,w1(CE)u1,w2(AF )u2,w3 = = (BD ⊙ (CE ⊗ AF ))u,w, which ends the proof. In particular, if B = D = 1 we have (A ⊙ C)(E ⊙ F ) = (CE ⊗ AF ). If we choose C = E = 1 instead, we have (10) (11) (A ⊗ B)(D ⊙ F ) = BD ⊙ AF. The following proposition provides an interesting charac- terization of the powers of any order of the matrix M . Proposition 4.1: For A a square matrix, M defined by s k−s {z {z Equation (7), and any integers r ≥ 0 and 0 ≤ s < k, ⊙ (Ar+1 ⊗ ··· ⊗ Ar+1) , } M rk+s = (Ar ⊗ ··· ⊗ Ar) } where the exponents in the right-hand side sum to rk + s. Proof: We prove the claim by induction on rk + s. It is true by definition for rk + s = 1. The induction step is easily proved by applying Equation (9). Indeed, [(Ar ⊗ (Ar ⊗···⊗ Ar))⊙(Ar+1⊗···⊗Ar+1)][((I⊗···⊗I)⊗(I⊗···⊗I))⊙A)] can be written as (Ar ⊗ ··· ⊗ Ar) ⊙ (Ar+1 ⊗ ··· ⊗ Ar+1 ⊗ (ArA)). The argument is correct also for limit cases s = 0 and s = k − 1. In particular we have the following. Corollary 4.1: For A a square matrix and M defined by Equation (7), M k = A ⊗ ··· ⊗ A. Moreover, if A satisfies (A),(B),(C) the essential spectral radius of M is the kth root of the essential spectral radius of A. Proof: The first part is a particular case of Proposition 4.1. From the properties of Kronecker product, we know the spectrum of M k is composed of all possible products of k eigenvalues of A. Hence the largest eigenvalue in absolute value, different from 1, of the matrix M k results to be 1k−1λ, where λ denotes the largest eigenvalue in absolute value, different from 1, of the matrix A. This also proves also that conditions (B) and (C) are verified for M when they are for A. If we take A = 1/n11T , (12) of size n, then M k is the matrix 1/nk11T of size nk with all identical elements. Thus we have a strategy converging exactly in k steps. We comment further on this example in the next section. Another property of M that will prove useful is stated in the next proposition. Proposition 4.2: For A a square matrix, M defined by Equation (7), and any integers r ≥ 0 and 0 ≤ s < k, M T rk+s Ar M rk+s = AT r k−s ⊗ Ar ⊗ ··· ⊗ AT r {z ⊗ AT r+1 Ar+1 ⊗ ··· ⊗ AT r+1 } {z s Ar+1 , } where the sums of exponents is rk + s. D. Design decentralisation Proof: From Proposition 4.1, we know that M rk+s = (Ar ⊗···⊗ Ar)⊙ (Ar+1 ⊗···⊗ Ar+1). Hence, by Equation (8), M T rk+s ) ⊙ (Ar ⊗ ··· ⊗ Ar). These two expressions are multiplied using Equation (10). ⊗ ··· ⊗ AT r+1 = (AT r+1 Now we would like to compute Tr M T t M t+1. This will be useful later when we will evaluate the performance of the block Kronecker strategy. We first need the following lemma. Lemma 4.2: Let B0, B1, . . . , Bk−1 be k square matrices of same dimensions. If l ≤ k is relatively prime to k, then Tr (B0 ⊗ B1 ⊗ ··· ⊗ Bl−1) ⊙ (Bl ⊗ ··· ⊗ Bk−1) = Tr B0BlB2lB3l ··· B(k−1)l, where the indices are understood modulo k. Proof: If we use Einstein's convention (repeated indices are summed over), we can write Tr (B0 ⊗ B1 ⊗ ··· ⊗ Bl−1) ⊙ (Bl ⊗ . . . ⊗ Bk−1) = [(B0 ⊗ B1 ⊗ ··· ⊗ Bl−1) ⊙ (Bl ⊗ ··· ⊗ Bk−1)]p,p = (B0)pk−l,p0(B1)pk−l+1,p1 ··· (Bl−1)pk−1,pl−1 (Bl)p0,pl ··· (Bk)pk−l−1,k = (B0)pk−l,p0(Bl)p0,pl(B2l)pl,p2l (B3l)p2l,pl ··· (B(k−1)l)p(k−2)l ,p(k−1)l = Tr B0BlB2l . . . B(k−1)l, where p = p0p1 . . . pk−1. Proposition 4.3: For A and M as defined above, if A is normal (i.e., AT A = AAT ) then Tr M T t M t+1 = Tr AT t At+1. AT r+1 Proof: We know that M T t Ar+1, if t = rk + s for some 0 ≤ s < k. Hence M t+1 = (AT r M t = AT r Ar ⊗ . . . ⊗ M T t Ar⊗···⊗AT r+1 is equal which by Equation (11) Ar+1) ⊙ AT r AT r+1 the same trace as AT r Ar+1AT r and A commute, this is also the trace of AT t Ar+1)((I⊗···⊗I)⊙A), to (AT r Ar ⊗ ··· ⊗ Ar+1. By Lemma 4.2, this matrix has Ar+1. As AT At+1. Ar . . . AT r+1 C. De Bruijn's graph The communication graph of M is (a subgraph of) a de Bruijn graph, which has nk vertices and arcs from any i to ni, ni + 1, ni + 2, . . . and ni + k − 1 (all modulo nk). In particular, if A is given by Equation (12), then M is the adjacency matrix of a de Bruijn graph, normalized so as for every row to sum to one. This graph was introduced by de Bruijn [10] in 1946 and has been considered for efficient distribution of information in different context such as in parallel computing [23] and peer-to-peer networks [13]. This paper can be seen as an extension of this idea to consensus problems. The process itself of convergence to consensus is de- centralised, in the sense that every agent acts on its own. However the communication strategy (who talks to whom?) must be designed once for all beforehand. This can be done in centralised way, where a new external agent dispatch to every other agent their own strategy. This can also be done in a decentralised way, where every agent is attributed a number i between 0 and N − 1 and then finds the agents of number νi, νi + 1,. . . , νi + ν− 1. Achieving this in the most effective way is a problem of its own, and is not treated in this paper. V. THE QUICKEST POSSIBLE STRATEGY We have seen that starting from A with all identical entries, we get arbitrarily large matrices M converging in finite time k. If we write N = nk the dimension of M , this convergence time is log N/ log n = log N/ log ν, where ν is the maximal in-degree of the graph of communication for M . We can see that no strategy, whether linear or not, whether time- invariant or not, can converge more rapidly. Indeed, to reach the average of the initial conditions, every agent must have information about all other agents, but it can only know ν other positions in one step of time, ν2 in two steps of time, etc. Hence the propagation of information needs around log N/ log ν steps to connect all agents. This reasoning is made rigorous in the following proposition. Proposition 5.1: Let M ∈ RN ×N such that M ≥ 0. Let ν be defined as above. Then M k > 0 implies νk ≥ N. Proof: The fact M k > 0 implies that for any pair of nodes (i, j) there exists in the graph GM a path connecting i to j of length k. Hence there are at least N 2 paths of length k. Let now Pi denote the number of paths having length i. The previous consideration implies that Pk ≥ N 2. On the other hand it is easy to see that P1 ≤ νN and in general that Pi ≤ νiN from which we get that Pk ≤ νkN . Hence νkN ≥ N 2 from which it results that νk ≥ N . The above proposition considers only the time-invariant case. An identical result can be found for the time-varying case, showing that there is no difference, in terms of speed of convergence toward the meeting point, between the time- invariant and the time-varying cases. This can be seen an a posteriori justification of our interest in the class of the time-invariant strategies. A linear time-invariant strategy converges in finite time iff its essential spectral radius is 0. For a strategy converging in infinite time, the essential spectral radius is a good measure of the convergence to the average of the initial conditions, as already mentioned. A. Comparison between block Kronecker strategy and Cay- ley strategy In this subsection we propose a comparison of the block Kronecker strategy with another strategy whose underlying communication graph has limited max in-degree and exhibits strong symmetries: the Cayley strategy. First we recall the concept of Cayley graph defined on Abelian groups [2], [1]. Let G be any finite Abelian group (internal operation will always be denoted +) of order G = N , and let S be a subset of G containing zero. The Cayley graph G(G, S) is the directed graph with vertex set G and arc set E = {(g, h) : h − g ∈ S} . Notice that a Cayley graph is always in-regular, namely the in-degree of each vertex is equal to S. Notice also that strong connectivity can be checked algebraically. Indeed, it can be seen that a Cayley graph G(G, S) is strongly connected if and only if the set S generates the group G, which means that any element in G can be expressed as a finite sum of (not necessarily distinct) elements in S. If S is such that −S = S we say that S is inverse-closed. In this case the graph obtained is undirected. A notion of Cayley structure can also be introduced for matrices. Let G be any finite Abelian group of order G = N . A matrix P ∈ RG×G is said to be a Cayley matrix over the group G if Pi,j = Pi+h,j+h ∀ i, j, h ∈ G . It is clear that for a Cayley matrix P there exists a π : G → R such that Pi,j = π(i − j). The function π is called the generator of the Cayley matrix P . Notice how, if P is a Cayley matrix generated by π, then GP is a Cayley graph with S = {h ∈ G : π(h) 6= 0}. Moreover, it is easy to see that for any Cayley matrix P we have that P 1 = 1 if and only if 1T P = 1T . This implies that a Cayley stochastic matrix is automatically doubly stochastic. In this case the function π associated with the matrix P is a probability distribution on the group G. Among the multiple possible choices of the probability distribution π, one is particularly simple, namely π(g) = 1/S for every g ∈ S. Example 1: Let us consider the group ZN of integers modulo N and the Cayley graph G(ZN , S) where S = in this case S is inverse-closed. {−1, 0, 1}. Notice that Consider the uniform probability distribution π(0) = π(1) = π(−1) = 1/3 The corresponding Cayley stochastic matrix is given by P =   1/3 1/3 1/3 1/3 1/3 0 0 0 ··· ··· 1/3 1/3 1/3 ··· ... ··· 0 ··· ... 0 ... 0 0 ... 1/3 0 1/3 0 0 0 0 ... ... 1/3 1/3   . (13) Notice that in this case we have two symmetries. The first is that the graph is undirected and the second that the graph is circulant. These symmetries can be seen in the structure of the transition matrix P that, indeed, turns out to be both symmetric and circulant [9]. Example 2: Let us now consider the group ZN × ZN and the Cayley graph G(ZN × ZN , S) where S = {(1, 0); (0, 0); (0, 1)}. Again consider the uniform probabil- ity distribution π((0, 0)) = π((1, 0)) = π((0, 1)) = 1/3 The corresponding Cayley stochastic matrix is given by the following block circulant matrix belonging to RN 2×N 2 0 0 0 0 0 0 0 0 0 ... ... ... 0 0 P1   (14) , P2 = 1 3 I. (15)  P = 0 P1 P2 0 ... 0 0 0 P1 P2 ... ... 0 0 P1 P2 0 0 ... P2 ··· ··· ··· ··· ··· where P1, P2 ∈ RN ×N are such that  P1 = 1/3 1/3   0 ... 1/3 0 ··· 1/3 1/3 ··· ... ··· 0 ··· ... 0 0 0 ... 0 0 0 ... 1/3   This example can be generalized to the more general case of the discrete d-dimensional tori Zd N , extensively studied in the literature regarding the peer-to-peer networks [21], [25]. Now we recall an interesting result regarding the essential spectral radius of the Cayley stochastic matrices. Assume that P ∈ RN ×N is a Cayley stochastic matrix generated by a suitable π and assume that S = ν, where S is as previously defined. Moreover assume that 0 ∈ S. Notice that this last fact implies that Pii > 0, ∀ i : 1 ≤ i ≤ N. Then it follows that ρ ≥ 1 − C/N 2/(ν−1), where C > 0 is a constant independent of S and N the number of agents. This result was proved in [6]. On the other side, the block Kronecker strategy con- structed from any matrix A has essential spectral radius λ1/k, where λ is the essential spectral radius of A, as stated in Corollary 4.1. If 0 < λ < 1, then λ1/k behaves like 1 − µ/k for large k and some µ. Recall that k is log N/ log n. Hence this is better than abelian Cayley strategies. In conclusion, block Kronecker strategies have a better essential spectral radius, hence a quicker convergence speed, than Cayley strategies. For the particular choice given by Equation (12), we converge in finite time, and this time is the smallest possible over all linear strategies with the same constrained degree. B. Simulation result As an illustration, we present a simulative comparison between the Cayley strategy and the block Kronecker strat- egy. The network considered consists of N = 81 agents. The matrix P for the Cayley strategy is the matrix (13), whereas the matrix M for the block Kronecker strategy is built starting from A =   1/3 1/3 1/3 1/3 1/3 1/3 1/3 1/3 1/3   with n = 3 and k = 4. The initial conditions has been chosen randomly inside the interval [−50, 50]. In both cases the in-degree is 3. Notice that, as depicted in Figure 1, the block Kronecker strategy reaches the average of the initial conditions in a finite number of steps whereas, the Cayley strategy, after the same numbers of steps, is still far from converging toward the meeting point. We get a lower bound on J1 by summing only the first k terms: Block Kronecker Strategy Cayley Strategy 50 40 30 20 10 0 −10 −20 −30 −40 −50 1 50 40 30 20 10 0 −10 −20 −30 −40 −50 1 2 3 4 5 6 7 8 9 10 2 3 4 5 6 7 8 9 10 Fig. 1. The block Kronecker strategy (left) converges in finite time, while the Cayley strategy (right) has a relatively slow convergence VI. LQR COST In this section we want to evaluate the performance of the block Kronecker strategy according to the quadratic cost J = J1 + γJ2, where J1 = EPt≥0(x(t) − x(∞))T (x(t) − x(∞)) accounts for the quickness of convergence, J2 = EPt≥0(x(t+1)−x(t))T (x(t+1)−x(t)) limits the norm of the updates, and γ weights the respective importance of those two factors. Precisely we evaluate J for any block Kronecker strategy derived from a normal matrix A. Remember that the initial state x(0) is supposed to be characterized by a identity covariance matrix. We start with a lemma which provides an upper and a lower bound for J1. Lemma 6.1: If A is a normal n × n matrix satisfying conditions (A),(B),(C), and ρ is the essential spectral radius of A, then the J1 cost of the corresponding block Kronecker strategy of size nk satisfies: JL ≤ J ≤ JU , 1−Tr (AT A)/n − k and JU = JL + k Proof: Classical arguments lead to write: where JL = N 1−(Tr (AT A)/n)k 1−ρ2 (Tr AT A − 1). J1 = EXt≥0 = Xt≥0 = Xt≥0 = Xt≥0 = Xt≥0 = Xt≥0 (x(t) − x(∞))T (x(t) − x(∞)) E (x(t) − x(∞))T (x(t) − x(∞)) E Tr (x(t) − x(∞))T (x(t) − x(∞)) E Tr (x(t) − x(∞))(x(t) − x(∞))T Tr (M t − E)E(x(0)x(0)T )(M t − E)T Tr (M t − E)(M t − E)T with E = 1/nk11T . Now, Tr (M t − E)(M t − E)T = Tr M tM tT − Tr E = Tr M tM tT − 1. When M is derived by block Kronecker product from a normal matrix A, this is equal to (Tr (AT A)r)k−s(Tr (AT A)r+1)s if t = rk + s, according to Proposition 4.2. k−1 J1 = Xr≥0 Xs=0 Xs=0 ≥ k−1 = k−1 ((Tr (AT A)r)k−s(Tr (AT A)r+1)s − 1) Xs=0 (Tr (AT A)0)k−s(Tr (AT A)1)s − k (16) nk−s(Tr AT A)s − k, The last summation is a geometric series that can be evalu- ated, leading to the bound J1 ≥ N 1 − (Tr AT A/n)k 1 − Tr AT A/n − k This proves the left inequality in the claim. For the right inequality, we find an upper bound on the terms neglected in the lower bound (16). As normal matrices can be diagonalized by a unitary transformation, the eigenvalues of AT A, which we denote 1, λ1, λ2, . . . , λn−1, are precisely the squared module of the eigenvalues of A. In i. particular, ρ2 = λ1, and the trace of (AT A)t is 1 +P λt The terms neglected in the lower bound (16) are Xr≥1 k−1 Xs=0 ((1 +Xi λr i )k−s(1 +Xi λr+1 i )s. For every r, we bound every of the k terms by the highest (for which s = 0). Hence the neglected terms are bounded from above by: Xr≥1 ((1 +Xi λr i )k − 1) = kXr≥1 P (λr 1, . . . , λr n−1), where P is a polynomial in the variables λ1, . . . , , λn−1 with no independent term: all monomials have degree at least one. Now we can sum all corresponding monomials this is a geometric series of progres- for r = 1, 2, . . .: k) is at most sion at most λ1. Hence Pr≥1 P (λr 1 Hence J1 differs from our lower bound by at most (Tr AT A − 1). P (λ1, . . . , λk) = 1 1, . . . , λr 1−λ1 1−λ1 k 1 (Tr AT A − 1). 1−λ1 Thus J1 = N 1−(Tr (AT A)/n)k 1−Tr (AT A)/n + O(log N ). Now we Lemma 6.2: Under the assumptions of Lemma 6.1, if ρi estimate J2. denote the eigenvalues of A different from one, 2J1 − N −Xi 1 1 − ρi2 ≤ J2 ≤ 2J1 − N. Proof: First we notice, adapting the first steps of the proof of the preceding proposition, that J2 = P Tr (M − I)T (M T )tM t(M − I), with I the identity. This involves terms of the form (M T )t+1M t. More precisely, J2 = Xt≥0 Tr (M T t+1 M t− M t+1 − M T t+1 M t+1 + M T t M t) M t − 1) − N− (Tr M T t M t+1 − 1) = 2Xt≥0 − M T t (Tr M T t − 2Xt≥0 1 1−λi At. The first term of the last member is precisely 2J1, the last At+1−1). From Cauchy-Schwartz inequality applied to Frobenius At+1 ≤ pTr AT t+1At+1Tr AT tAt ≤ term is, thanks to Proposition 4.3, 2Pt≥0(Tr AT t norm, Tr AT t Tr AT t Hence Pt≥0(Tr AT t Pi λi = ρi2. Hence, J = N(cid:16)(1 + 2γ) 1 − (Tr (AT A)/n)k 1 − Tr (AT A)/n − γ + O(log N/N )(cid:17). i = , where, as argued in the proof of Lemma 6.1, At+1 − 1) ≤ Pt≥0Pi λt Since the trace of AT A is the sum of squares of ele- ments of A, we see that the coefficient of N (neglecting the O(log N/N )) term) is optimized by the matrix A = 1/n11T , whatever the value of γ is. In this case, the lower bound obtained on J1 is exact, since only k terms are non- zero. The optimal cost is then 1 − 1/n − γ + O(log N/N )(cid:17), 1 − 1/N J = N(cid:16)(1 + 2γ) with ν = n. Hence there is here no trade-off between J1 and J2 among the family of block Kronecker strategies, in contrast with the general LQR theory. Note that the optimal control strategy for unconstrained degree (every agent knows every position) is easily solved by a scalar algebraic Riccati equation, leading to the optimal cost J = N (1+√1 + 4γ)/2. If γ is small and n is large, then the optimal finite-time block Kronecker approaches the un- bounded degree optimal solution with a cost approximately equal to (1 + γ)N . VII. CONCLUSIONS We have introduced a family of strategies for a consensus problem, whose graph of communication is de Bruijn's graph. We have shown that they can converge in finite logarithmic time, which is optimal. We have evaluated the LQR cost of these strategies, proving their quasi-optimality if the cost of update is small and the degree of the graph not too low. This work can be extended in several directions, including: • designing strategies valid for any N , not only exact powers of n; • tackling the continuous-time case, where no deadbeat strategy can exist; • estimating the LQR cost for Cayley strategies; • finding strategies that minimize the LQR cost for any cost γ of the update; VIII. ACKNOWLEDGEMENTS This work was partly developed during the stay of one of the authors (J.-C. D.) at Department of Information Engineering of Universit`a di Padova. Stimulating discussions with J. Hendrickx are gratefully acknowledged. REFERENCES [1] N. Alon and Y. Roichman. Random Cayley graphs and expanders. Random Structures and Algorithms, 5:271–284, 1994. [2] L. Babai. Spectra of Cayley graphs. Journal of Combinatorial Theory, Series B, 27:180–189, 1979. [3] R. W. Beard, J. Lawton, and F. Y. Hadaegh. A coordination architec- IEEE Transaction on Control ture for spacecraft formation control. Systems Technology, 9:777–790, 2001. [4] P. Bhatta and N. E. Leonard. Stabilization and coordination of underwater gliders. In 41st IEEE Conference on Decision and Control, 2002. [5] R. Carli, F. Fagnani, A. Speranzon, and S.Zampieri. Communication In American Control constraints in coordinated consensus problem. Conference (ACC '06), June 2006. [6] R. Carli, F. Fagnani, A. Speranzon, and S. Zampieri. Communication constraints in the average consensus problem. Accepted for publication to Automatica. [7] R. Carli, F. Fagnani, and S. Zampieri. On the state agreement with quantized information. In Proceedings of the 17th International Sym- posium on Mathematical Theory of Networks and Systems (MTNS), Kyoto, Japan, pages 1500–1508, July 2006. [8] J. Cortes, S. Martinez, and F. Bullo. Robust rendezvous for mobile autonomous agents via proximity graphs in arbitrary dimensions. IEEE Trans. Automat. Control. [9] P. J. Davis. Circulant matrices. A Wiley-Interscience Publication, Pure and Applied Mathematics. John Wiley & Sons, New York-Chichester- Brisbane, 1979. [10] N. G. de Bruijn. A combinatorial problem. Koninklijke Nederlandse Akademie v. Wetenschappen, 49:758–764, 1946. [11] D. V. Dimarogonas and K. J. Kyriakopoulos. On the rendezvous IEEE Transactions on problem for multiple nonholonomic agents. automatic control, 52(5):916–922, 2007. [12] J. A. Fax and R. M. Murray. Information flow and cooperative control IEEE Trans. Automat. Control, 49(9):1465– of vehicle formations. 1476, 2004. [13] P. Fraigniaud and P. Gauron. D2b: A de Bruijn based content- addressable network. Theor. Comput. Sci., 355(1):65–79, 2006. [14] Y. Hatano and M. Meshabi. Agreement of random networks. In IEEE Conference on Decision and Control, 2004. [15] A. Jadbabaie, J. Lin, and A. S. Morse. Coordination of groups of mobile autonomous agents using nearest neighbor rules. IEEE Trans. Automat. Control, 48(6):988–1001, 2003. [16] A. Kashyap, T. Basar, and R. Srikant. Consensus with quantized information updates. In Proceedings of CDC Conference, san Diego, 2006. [17] R. H. Koning, H. Neudecker, and T. Wansbeek. Block kronecker products and the vecb operator. Linear algebra and its applications, 149:165–184, 1991. [18] Z. Lin, B. Francis, and M. Maggiore. Necessary and sufficient graph- ical conditions for formation control of unicycles. IEEE Transactions on automatic control, 50(1):121–127, 2005. [19] M. Mazo, A. Speranzon, K. H. Johansson, and X. Hu. Multi-robot tracking of a moving object using directional sensors. In Proceedings of the International Conference on Robotics and Automation (ICRA), 2004. [20] R. Olfati-Saber and R. M. Murray. Consensus problems in networks of agents with switching topology and time-delays. IEEE Trans. Automat. Control, 49(9):1520–1533, 2004. [21] S. Ratnasamy, P. Francis, M. Handley, R. M. Karp, and S. Shenker. A scalable content-addressable network. In SIGCOMM., pages 161–172, 2001. [22] W. Ren, K. L. Moore, and Y. Chen. Necessary and sufficient graphical conditions for formation control of unicycles. ASME Journal of Dynamic Systems, Measurement, and Control, to appear, 2007. [23] M. R. Samatham and D. K. Pradhan. The de Bruijn multiprocessor network: A versatile parallel processing and sorting network for vlsi. IEEE Trans. Comput., 38(4):567–581, 1989. [24] A. Speranzon, C. Fischione, and K. Johansson. Distributed and collaborative estimation over wireless sensor networks. In Proceedings of the IEEE Conference on Decision and Control (CDC'06), pages 1025–1030, December 2006. [25] I. Stoica, R. Morris, D. R. Karger, M. F. Kaashoek, and H. Balakr- ishnan. Chord: A scalable peer-to-peer lookup service for internet applications. In SIGCOMM., pages 161–172, 2001. [26] H. G. Tanner, A. Jadbabaie, and G. J. Pappas. Stable flocking of mobile agents, part i: fixed topology. In IEEE Conference on decision and control, 2003. [27] H. G. Tanner, A. Jadbabaie, and G. J. Pappas. Stable flocking of In IEEE Conference on mobile agents, part ii: dynamic topology. decision and control, 2003. [28] H. G. Tanner, S. G. Loizou, and K. J. Kyriakopoulos. Nonholonomic IEEE navigation and control of cooperating mobile manipulators. Transactions on robotics and automation, 19(1):53–64, 2003.
1908.10404	2	1908	2019-09-03T03:43:47	The Effectiveness of Managed Lane Strategies for the Near-term Deployment of Cooperative Adaptive Cruise Control	[ "cs.MA", "eess.SP" ]	Traffic simulation is a cost-effective way to test the deployment of Cooperative Adaptive Cruise Control (CACC) vehicles in a large-scale transportation network. By using a previously developed microscopic simulation testbed, this paper examines the impacts of four managed lane strategies for the near-term deployment of CACC vehicles under mixed traffic conditions. Network-wide performance measures are investigated from the perspectives of mobility, safety, equity, and environmental impacts. In addition, the platoon formation performance of CACC vehicles is evaluated with platoon-orientated measures, such as the percentage of platooned CACC vehicles, average platoon depth, and vehicle-hour-platooned that is proposed in this paper under the imperfect DSRC communication environment. Moreover, managed lane score matrices are developed to incorporate heterogeneous categories of performance measures, aiming to provide a more comprehensive picture for stakeholders. The results show that mixing CACC traffic along with non-CACC traffic across all travel lanes is an acceptable option when the market penetration (MP) is lower than 30% for roadways where a managed lane is absent. Providing CACC with priority access to an existing managed lane, if available, is also a good strategy for improving the overall traffic performance when the MP is lower than 40%. When the MP reaches above 40%, a dedicated lane for CACC vehicles is recommended, as it provides greater opportunity for CACC vehicles to form platoons. The facilitation of homogeneous CACC traffic flow could make further improvements possible in the future.	cs.MA	cs	The Effectiveness of Managed Lane Strategies for the Near-term Deployment of Cooperative Adaptive Cruise Control Zijia Zhonga,∗, Joyoung Leeb aDepartment of Civil and Environmental Engineering, University of Delaware bJohn A. Reif, Jr. Department of Civil and Environmental Engineering, New Jersey Institute of Technology 9 1 0 2 p e S 3 ] A M . s c [ 2 v 4 0 4 0 1 . 8 0 9 1 : v i X r a Abstract Traffic simulation is a cost-effective way to test the deployment of Cooperative Adaptive Cruise Control (CACC) vehicles in a large-scale transportation network. By using a previously developed microscopic simulation testbed, this paper examines the impacts of four managed lane strategies for the near-term deployment of CACC vehicles under mixed traffic conditions. Network-wide performance measures are investigated from the perspectives of mobility, safety, equity, and envi- ronmental impacts. In addition, the platoon formation performance of CACC vehicles is evaluated with platoon-orientated measures, such as the percentage of platooned CACC vehicles, average platoon depth, and vehicle-hour-platooned that is proposed in this paper under the imperfect DSRC communication environment. Moreover, managed lane score matrices are developed to incorpo- rate heterogeneous categories of performance measures, aiming to provide a more comprehensive picture for stakeholders. The results show that mixing CACC traffic along with non-CACC traffic across all travel lanes is an acceptable option when the market penetration (MP) is lower than 30% for roadways where a managed lane is absent. Providing CACC with priority access to an existing managed lane, if available, is also a good strategy for improving the overall traffic performance when the MP is lower than 40%. When the MP reaches above 40%, a dedicated lane for CACC vehicles is recommended, as it provides greater opportunity for CACC vehicles to form platoons. The facilitation of homogeneous CACC traffic flow could make further improvements possible in the future. Keywords: Managed Lane, Cooperative Adaptive Cruise Control, Microscopic Traffic Simulation, Platoon Formation, Mixed Traffic Conditions, Imperfect DSRC Communication 1. Introduction Connected and Automated Vehicle (CAV) technology will revolutionize the way vehicles are operated. Among its wide-ranging applications, Cooperative Adaptive Cruise Control (CACC) is identified as one of the thrust applications by governments, industries, and academia all around the world in improving mobility, environment, and more importantly safety. The field experiments of ∗Corresponding author: Zijia Zhong Email addresses: [email protected] (Zijia Zhong), [email protected] (Joyoung Lee) Preprint submitted to Elsevier September 4, 2019 CACC systems have been accelerated as the technology matures in recent years. The highlights of the deployments, to name a few, include the United State Department of Transportation (USDOT)- approved CAV test sites (Xiang et al., 2009), the California PATH's Automated Highway System (AHS) program (Bu et al., 2010; Milan´es et al., 2014; Rajamani and Shladover, 2001), the En- ergy ITS Project (Tsugawa et al., 2011), SARTRE (van Willigen et al., 2013), KONVOI (Wille et al., 2007), and the European Truck Challenging 2016 (Eckhardt, 2016). Despite policy en- couragement from government agencies and the progress made in academic research, large-scale field deployment, however, are still considered premature at the current stage, in terms of safety, technological, and budgetary concerns. Simulation is one of the best approaches to bridge the gap between prototype testing CACC technology and the large-scale real-world deployment by provid- ing a virtual environment. Simulation can also improve the quality of the testing by expanding the testing scope to a wide spectrum of scenarios without incurring a significant amount of additional cost. Managed lanes (MLs), as defined by the Federal Highway Administration (FHWA), are high- way lanes that are set aside and operated under a variety of fixed and/or real-time strategies in response to local goals and objectives (Kim et al., 2015), such as traffic operation improvement, emission reduction, safety enhancement etc. As solutions for efficient lane management, ML strategies have been implemented nationwide with various forms (e.g., high-occupancy vehicle (HOV) lane, high-occupancy toll (HOT) lane, exclusive bus lane (XBL)). Pairing ML strategies with CACC (designated as CACC-ML) becomes a logical step in promoting CACC, because of CACC's distinct operational characteristics in comparison to human-driven vehicles (HVs). Fur- thermore, the provision of CACC-ML can provide road users with stronger incentives either to purchase CACC vehicles that meet the minimum functional requirements or to factor in the option of retrofitting certified aftermarket CACC systems to a reasonable extent. In this paper, the effectiveness of deploying CACC-ML strategies is examined with the goal of deriving policy recommendations for the initial introduction of CACC vehicles into mixed traf- fic. The organization of the remainder of the paper is as follows. Relevant research regarding deploying CACC vehicles in mixed traffic is reviewed in Section 2, followed by the microscopic simulation framework in Section 3. The simulation results are presented and discussed in Section 4. Conclusions and future work are discussed in Section 5. 2. Literature Review Increasing practices of MLs have been observed over the years and case studies show that the desired travel patterns or behaviors can be incentivized by ML strategies. For instance, the California Clean Air Vehicle Decal program (Shewmake and Jarvis, 2014) was designed to pro- mote user adoption of energy-efficient and low-emission hybrid vehicles by allowing the owners to drive on HOV lanes without even meeting the occupancy requirement. It is believed that the adoption of CACC vehicles can be facilitated with policy alike. The CACC-ML provision is en- visioned to have three phases. First, the adoption of CACC vehicles is incentivized by allowing the use of the ML free of charge. At this stage, the following headway of CACC vehicles on the ML may be comparable to that has been observed from HVs for safety reason in mixed traffic. Then a shorter following headway for CACC vehicles could be implemented to further increase 2 the carrying capacity of the ML when the demand as well as the familiarity of road users to CACC increases. Lastly, when the CACC vehicles reach a critical level of MP, the transition to CACC- only lane could be eventually made. At this stage, high-performance driving enabled by CACC can be achieved because of a homogeneous CACC traffic on the ML (Zhong, 2018). The benefits of CACC for enhancing the performance of a freeway system have been reported. The capacity of 4,250 vph per lane was observed by Van Arem et al. (2006) on a 6-km segment with uniformly distributed ramps under full market penetration. Shladover et al. (2012) concluded the capacity of 3,600 vph per lane at 90% MP in the absence of on-ramp demand for a one- lane freeway. Arnaout and Arnaout (2014) evaluated CACC under moderate, saturated, and over- saturated demand on a hypothetical 4-lane highway under different market penetrations. They found that 9,400 vehicles (out of 10,000 vehicles) could be served in an hour when the CACC reached 40% MP. Lee et al. (2014) found the mobility benefits of CACC began to show at 30% MP. Songchitruksa et al. (2016) assessed the improvements of CACC on the 26-mile I-30 freeway in Dallas, TX and observed the highest throughput being 4,400 vph at 50% MP. Insofar, studies on MLs for CAVs are relatively scarce in the literature. Traditional lane man- agement experience may not be directly transferrable to CACC deployment because of the the difference in the underlying operational principles. There are two major approaches for study- ing CACC-ML: analytical modeling and computer simulation. Hussain et al. (2016) proposed an analytical ML model for determining the optimal lanes to be allocated for CAV traffic under single-lane and ML environment. Three types of CAV following headway (conservative, neu- tral, and aggressive) were assigned. The maximization of the overall system throughput could be computed under the assumption that both CAVs and HVs are randomly distributed in a freeway facility. From a macroscopic perspective, Qom et al. (2016) used simulation-based dynamic traffic assignment to study the diversion of CACC vehicles onto a ML with toll incentives from dynamic pricing. Using Vissim, Zhang et al. (2018) investigated the operational capacity of CACC-ML with dedicated CACC on/off ramps. It was discovered that the ramp density should be placed no less than 5 km (3 miles) apart to ensure a reasonable level of throughput. Liu et al. (2018) found that deploying the ML and vehicle awareness device (VAD)-equipped vehicles were proven to be helpful under low and medium market penetrations. The increase in the pipeline capacity (under the ideal condition with no lane changes, uniform desired speed, etc.) was estimated between 8% to 23%. On a subsequent test on an 18-km segment of the SR-99, they observed great improve- ment on traffic flow pattern when the MP was at 40% or higher on the corridor-level speed heat map. A CACC vehicle is able to operate in a safe manner with a much smaller following headway due to the V2V communication with its predecessor(s) (Nowakowski et al., 2011; Shladover et al., 2018; Tsugawa et al., 2016). Clustering CACC vehicles, therefore, becomes a crucial task in operation. Note that CACC "clustering" and "platooning" are used interchangeably. According to Shladover et al. (2015), three types of clustering strategies could be expected in operation: 1) ad hoc clustering, 2) local coordination, and 3) global coordination. Ad hoc clustering assumes that the CACC vehicles arrive in random sequence and the CACC vehicles do not actively seek clustering opportunities. Therefore, the probability of driving around other CACC vehicles is highly correlated to MP. The ad hoc clustering has been adopted in most of the previous studies thus far because of its simplicity on implementation. 3 Local coordination, commonly known as active platooning, could be employed to help with clustering. Under this concept, CACC vehicles actively identify and approach an existing CACC cluster or other free-agent CACC vehicles to form a new cluster. The local coordination has been discussed by Liang et al. (2013), D´avila (2012), and Lee et al. (2014). One of the main chal- lenges to local coordination is the determination of the relative position of a vehicle with sufficient accuracy and reliability. Solutions have been proposed to aid the localization of CACC vehi- cles (Shladover et al., 2015), for instance, 1) infrastructure lane identification by radio frequency identification (RFID), 2) vehicle-based lane identification by on-broad sensors (e.g., GPS, inertial measurement unit, and camera), 3) vehicle-based confirmation of the predecessor by using visual or infrared-camera-visible marking, and 4) visual confirmation by the drivers. Lastly, global coor- dination uses advance planning to coordinate vehicles traveling with the same origin-destination pair even before the CACC vehicles entering the highway (Larson et al., 2013). The most likely application for global coordination would be long-haul trucking or lengthy commute trip on con- gested highways. Equity issues is another crucial aspect for MLs, and it has to be properly addressed, due to its vital role in gaining public support. The perception of exclusivity for different road users could be a sensitive matter. The Victoria Transport Policy Institute (2018) conducted extensive research regarding the equity and social justice for the impact of a ML in CAV context. It identified that a dedicated lane for CACC could impact various social classes (Litman, 2017), since not too many people are able to afford CACC vehicles. The issue becomes more pronounced when speed limit privilege (e.g., higher speed limit) is assigned to CACC lane. Hence, the impacts of a CACC lane among users and non-users need to be quantified for further assessment. More mobility options could be provided (e.g., access to transit) in order to gain the benefits of CACC-ML for different user groups. Learned from existing MLs, additional considerations include: 1) cost incentives to make CACC accessible for all via tax deduction, subsidies, etc.; 2) possible shared mobility for CACC vehicles; and 3) availability of alternative routes for other road users. The CACC-ML strategy could provide technical accommodation, economic incentive, and mo- bility improvement for users who decide whether to purchase CACC-equipped vehicles. CACC- ML could be instrumental in the near-term deployment of CACC. The effectiveness of CACC-ML strategies can be evaluated systematically in a cost-efficient way with microscopic simulation. 3. Methodology 3.1. Simulation Framework A segment of a major commuter corridor, the Interstate Highway 66 (I-66) located in Fairfax County, Virginia, is used to evaluate the potential operation impacts for deploying CACC vehicles in the near term. The 8-km (5-mile) stretch of the roadway (shown in Fig. 1) is located outside of the Washington D.C. Beltway (I-495) with four lanes in each direction and two interchanges. Currently, an HOV lane is implemented in the leftmost lane in each direction. No physical barrier is in place between the General Purpose Lanes (GPLs) and the HOV lane. The I-66 network has been used in previous studies by Li et al. (2019), Transportation Research Board (2018), Zhong (2018), and Zhong and Lee (2018). The calibration was conducted with three independent data sources: INRIX travel time, remote traffic microwave sensor data (RTMS) flow data, and video 4 camera data to ensure a realistic representation of the real-world traffic conditions (Leidos, 2016; Li et al., 2019). The first two were the primary sources of fine-tuning the simulation network. An overview of the traffic condition for each origin is shown in Table 2. The demand is defined in a 15-min interval. Table 1: Independent Data Source for Calibration Data Source RTMS data INRIX data Video camera footage (cid:88) (cid:88) (cid:88) point segment point Volume Speed Occupancy Travel Time Measurement Type (cid:88) (cid:88) Fig. 1: Interstate highway 66 (I-66) segment Table 2: Network Traffic Demand Zone ID Description Peak/Avg. GP Demand, vph Peak/Avg. HOV Demand, vph Z1 Z2 Z3 Z4 I-66 East or I-495 Exit 62 Nutley St. Exit 60 SR 123 I-66 West 4197/3998 712/382 1411/1146 - 1885/1796 335/180 1411/1146 - The simulation framework is developed based on PTV Vissim (PTV AG, 2018b), Vissim Com- ponent Object Model (COM) interface (PTV AG, 2015), and the Vissim DriverModel.DLL appli- cation programming interface (PTV AG, 2017) (VEDM thereafter). The overall simulation frame- work is shown in Fig. 2. Vissim is a multimodal microscopic traffic simulation software, where each entity (e.g., car, train, or pedestrian) is simulated individually. Such capability is one of the most crucial elements for simulation of CACC vehicles. The VEDM is programmed to replace the default Wiedemann car-following model (Wiedemann, 1991) with the custom CACC behavioral model. The CACC vehicles are equipped with automated longitudinal control that is based on the Enhanced Intelligent Driver Model (E-IDM) (Kesting et al., 2010). The behavior model of the 5 Fig. 2: Vissim-based microscopic simulation testbed EIDM is expressed in Eq. 1, 2, and 3. )δ − ( s∗( x, xlead) )] s0 x = (1 − c) xIDM + c[ xCAH + b · tanh( xIDM− xCAH √ s∗( x, xlead) = s0 + xT + b )] x( x − xlead) if x = xIDM ≥ xCAH otherwise xdes a[1 − ( x  x2·min( xlead, x) (1) (2) (3) xCAH = lead−2x·min( xlead, x) x2 min( xlead, x) − ( x− xlead)2Θ( x− xlead) 2x 2 ab xlead( x − xlead) ≤ −2x min( xlead, x) otherwise where, a is the maximum acceleration; b is the desired deceleration; c is the coolness factor; δ is the free acceleration exponent; x is the current speed of the subject vehicle; xdes is the desired speed, xlead is the speed of the lead vehicle; s0 is the minimal distance; x is the acceleration of the subject vehicle; xlead is the acceleration of the lead vehicle; xIDM is the acceleration calculated by the original IDM model (Treiber et al., 2000). The minimal distance can be calculated by Eq.2 where T is the desired time gap; and xCAH is the acceleration calculated by the constant-acceleration heuristic (CAH) component as shown in Eq.3, where Θ is the Heaviside step function. The local coordination of CACC vehicles is incorporated into the VEDM as well, where a free-agent CACC vehicle (that is not in a platoon) detects its surrounding traffic, trying to join an existing platoon or form a platoon with other free-agent CACC vehicles. Thus far, three types of platoon formation have been used: front-, mid-, and rear- joining (Amoozadeh et al., 2015; Lee et al., 2014). In this study, only rear-joint is allowed for jointing a platoon in order to control the variables of the evaluation. The decision-making procedure for a free agent CACC vehicle is exhibited in Fig. 3. 6 Fig. 3: CACC clustering algorithm The Minimizing Overall Braking Induced by Lane Change model (MOBIL) (Kesting et al., 2007) as expressed in Eq.4 is adopted as the mechanism to prevent lane changing of a free-agent CACC vehicle that is potentially disruptive to the surrounding traffic. x − x + p( xn − xn + xo − xo) > ∆ xth (4) where p is the politeness factor; x is the new estimated acceleration for the subject vehicle; x is the current acceleration of the subject vehicle; xn is estimated acceleration of the new trailing vehicle; xn is the previous acceleration of the new trailing vehicle; xo is the new estimated acceleration for the old trailing vehicle; xo is the previous acceleration of the old trailing vehicle. Note that the estimated acceleration is calculated by the E-IDM model. When a potential platooning opportunity is identified via V2V communication, the CACC system estimates the impacts on the immediate vehicles based on MOBIL should the lane change 7 be initiated. A lane change is only executed when the expected negative impact is below the pre-determined threshold, ∆ xth. In the event of multiple clustering alternatives, the one with the smallest value (less impact) in the left hand side of the Eq.4 is chosen, as illustrated in Fig. 4. Fig. 4: MOBIL-based lane change decision The parameters for the E-IDM models are shown in table 3. Note that two types of headway are implemented in the longitudinal control. The Tintra represents the intra-platoon headway, which is typically smaller owing to the communication within a CACC platoon. The larger headway, Tinter, is used when a CACC vehicle follows a non-CACC vehicle. The minimum number of CACC vehicles for a platoon is set as three and the maximum number of CACC vehicles for a platoon is set as seven. Table 3: CACC Vehicle Control Parameters Parameter Tintra Tinter 1.2 s value 1 s a s0 1m 2m/s2 b c 2m/s2 0.99 θ 4 xdes 105 km/h p 0.9 ∆ xth 1m/s2 While the imperfect wireless communication is not the main focus in this paper, it is worth mentioning that an analytical probability module that was derived from a packet-level network simulator (i.e. ns-2 (Chen et al., 2007)) by Killat and Hartenstein (2009) was also incorporated into the VEDM. One-hop DSRC broadcast reception probability can be calculated during the sim- ulation based on transmission power, transmission distance, and communication channel loading. The outcomes of the communication have an influence on the CACC platoon operation. The de- tails regarding the implementation of this module in VEDM is available in Zhong and Lee (2018). The behaviors of a CACC vehicle in a platoon is dictated by the successful V2V communication. Here the CACC system is assumed with SAE Level 3 vehicle automation that requires a recep- tive fallback-ready driver(Society of Automotive Engineers International, 2014). Fig. 5 illustrates the system fallback in the event of automated driving system (ADS) failure, such as loss of DSRC packets. It is executed at each simulation time step on an individual vehicle basis, and it determines the transition of the control authority between the ADS and the human driver. 8 Fig. 5: CACC system fallback loop 3.2. Managed Lane Scenarios Eligibility, accessibility, and pricing are the three primary factors in designing a ML (Federal Highway Administration, 2008). In this study, only the first two factors are considered. A three- phase ML deployment for CACC is envisioned as covered in the literature review. Subsequently, four ML strategies, shown in table 4, are formulated for testing. • Base case (BASE): This scenario serves as the base condition of the I-66 segment for this study. As stated previously, an HOV lane is implemented in the leftmost lane of the roadway and the network has been calibrated. • Unmanaged lane (UML): In this strategy, the HOV lane was revoked, and current HOVs are treated as GP vehicles. CACC vehicles are not given priority use of any lane, and they operate along with general purpose (GP) vehicles/ HVs. • Mixed managed lane (MML): This strategy aims to utilize the current lane use configura- tion of the network and provide priority usage of HOV lane for CACC vehicles. It investi- gates the effectiveness of introducing CACC vehicle to an existing ML facility. 9 • Dedicated CACC lane (DL): The exclusive access to the leftmost lane for CACC vehicles is studied. A homogeneous CACC traffic is believed to be beneficial for the CACC operation. The merging impact of CACC vehicles toward the leftmost lane can be studied as well. Like in UML, the HOVs are treated as GP vehicles in this strategy. • Dedicated CACC lane with access control (DLA): This strategy is essentially a dedicated CACC lane mentioned above but with access control where CACC vehicles are only able to move in or out of the ML at designated locations. Therefore, the weaving activities are aggregated at certain locations of the segment. It is formulated to insulate the CACC vehicle platoons form the potential impacts of the weaving activities. Table 4: Scenario settings for two cases. XXXXXXXXXXXX Strategy Lane BASE UML MML DL DLA 4th(leftmost) H G+C C+H C C 3rd G 2nd G 1st G G+C G+C G+C G G G G G G G G G Access Control N N N N Y 1H-HOV, G-GP, C-CACC Since MLs are commonly designed to be left-lane concurrent with at-grade entry/exit, the weaving which requires crossing multiple GPLs could have a negative impact on mobility and safety of the overall traffic. Therefore, the weaving length of the incoming CACC vehicle to the leftmost managed lane has to be considered when it comes to a ML design. California Department of Transportation (2018) suggested the minimal distance of 800 ft (243 m) per lane; whereas Washington State Department of Transportation (2006) proposed the minimal distance of 500 ft (152m) between an access zone and immediate on- or off- ramp. The access control level is a trade-off among accessibility, efficiency, and safety. The optimal access control level was studied by Saad et al. (2018) and the weaving length of 1,000 ft (333m) per lane change was recommended. Cai et al. (2018) investigated the safety impact of weaving lengths for lane change maneuver using a driving simulator. The 1,000 ft (333 m) optimal length for each lane change is based on the three evaluative measures: speed standard deviation, minimum time to collision, and potential collision events. According to the aforementioned recommendations, 1,000-ft (333-m) weaving length per lane for the location of the ML entry is adopted in this study. 4. Simulation Result The effectiveness of CACC-ML is discussed from the perspectives of overall network per- formance, safety, environmental impact, equity, and platooning performance. To factor in the variability of simulation, five replications are run for each combination of strategy and MP. 10 4.1. Network Performance The network-wide improvement of CACC under the four ML strategies are evaluated. Being developed by Caltrans PeMS (Chen et al., 2001), the Q value is the ratio of vehicle mile traveled (VMT) and vehicle hour traveled (VHT). The VHT represents the input of a freeway system, whereas the VMT gauges the output of a freeway system. It can be also considered as the speed on an aggregated level. Hence, the higher the value of Q, the better performance or more productive of a highway system. As shown in Fig. 6, introducing CACC vehicles to the network helps to improve the network efficiency even when MP is below 15%. UML achieves the highest efficiency among all strategies across the MP range in terms of Q values. For low MP range (less than 30%), the implementation of DL or DLA seems to be premature, because it decreases the efficiency of the network drastically. Fig. 6: Network Q value Fig. 7: Planning time index 11 The Planning Time Index (PTI) is defined as the ratio of the 95-percentile peak period travel time to the free-flow travel time (Federal Highway Administration, 2006). It is used for evaluating the travel time reliability of the ML strategies. As shown in Fig. 7, the original PTI for traversing the network is 1.50, which indicate that a driver needs to allocate 1.5 times of the free-flow travel time in order to ensure a 95% on-time arrival. The PTIs for UML and MML exhibit a decreasing trend as more CACC vehicles are introduced into the network. At 40% MP, the lowest PTIs for UML and MML are observed as 1.14 and 1.15, respectively. Travelers in both DL and DLA experience higher PTIs than the BASE when MP is below 30%. With a decreasing trend, at 50% MP, the DL and DLA reach the same level of the PTIs as observed for UML and MML at 40% MP. 4.2. Safety The standard deviation of speed, as a Safety Surrogate Assessment Model (SSAM) (Lee et al., 2014; Li et al., 2017), has been used to assess the safety performance (Gettman and Head, 2003). Generally, a lower value in the standard deviation of speed represents a more stable traffic flow (Songchitruksa et al., 2016). In any of the ML strategy, the standard deviation of speed exhibits a decreasing trend as the MP goes up as shown in Fig. 8. The improvement in terms of the SSAM is observed in UML and MML even at low MP range. The DL and DLA start with higher values of standard deviation than that of the BASE. They both resumes to the baseline level at 30% MP. Between DL and DLA, the former display a marginally lower standard deviation at each MP level that is below 50%. Additionally, the speed standard deviation of UML is consistently lower than that of MML throughout all levels of MP. Fig. 8: Standard deviation of speed 4.3. Environmental Impact The VT-Micro model (Rakha et al., 2004) is implemented in the Vissim EmissionModel.DLL (PTV AG, 2018a) to collect second-by-second fuel consumption for each vehicle. The entire network is treated as a system and the average fuel consumption for each vehicle per second is 12 shown in Fig. 9. All the testing scenarios yield a lower fuel consumption with the exception of 20% and 25% MP for both DL and DLA. The higher the MP, the greater the reduction in fuel consumption is observed. Fig. 9: Average fuel consumption 4.4. Equity Equity refers to fairness when it comes to distributing impacts. Horizontal transportation eq- uity, in the context of CACC-ML, is the impact distribution among CACC users and non-CACC users. Traffic speed and roadway level-of-service can be used as indicators to measure impacts that affect equity (Litman, 2002). When the performance on GPLs is degraded, negative equity is distributed to GP users. To measure the equity distribution among all four ML strategies, the segment travel time for traversing the entire segment on the mainline is compared. For each Violin plot shown in Fig. 10, the widths on either side are scaled by the number of observations. The quartiles of the distributions are also shown in the shade of the violin plot. Under UML, a narrower distribution of travel time is observed even at 5% MP, which means both CACC or GP vehicles do not experience negative equity. Across the rest of the MP levels, the distributions of GP travel time remain nearly unchanged but with fewer high values observed in the BASE. With the increase of CACC vehicles, it is also observed that the median values of the distributions remain the same in UML. For MML where HOVs are grouped into GP vehicles, improvement only shows when MP reaches 10% and the trend is maintained until MP reaches 50%. For DL and DLA, the travel times get worsen when MP is below 30% and the improvement only shows above 35% MP. Among all the scenarios, the CACC travel time distributions have lower variances and medians than GP vehicles that are operated in the prevailing traffic condition. 13 Fig. 10: Travel time comparison based on vehicle type 4.5. Platooning Performance In this section, a set of rarely-examined MOEs is introduced. They are related to CACC clus- tering and made possible by the high customizability of VEDM. One of the major benefits of CACC is the compact vehicle platoons with short intra-platoon following distance enabled by the V2V communication. As such, the number of vehicles within platoons is positively correlated to the improvement in throughput. The average percentage of CACC vehicles is shown in Fig. 11. With the CACC demand increases, the percentage of the number of platooned CACC vehicles de- creases in both UML and MML, which seems counterintuitive. In fact, this indicates the increased difficulties for additional CACC vehicles to form platoons. When the slope of a curve is positive as shown in DL, the platooning opportunity increases as the presence of CACC vehicles increases. If a curve linearly increases at a 45-degree slope, it is an indication that all the additional CACC vehicles get the same likelihood of forming platoons. For a negative slope of the curve, the cor- responding ML strategy is not able to offer clustering opportunity at the same level as MP grows. Between 20% and 50% MP, DLA outperforms other strategies and it is only marginally surpassed 14 by DL at 55 % MP. Fig. 11: Percentage of CACC vehicle in platoons During the simulation run, each CACC vehicle registers its sequence in a platoon. Due to the decentralized execution for each CACC vehicle in VEDM, the mean platoon depth (Amoozadeh et al., 2015) is calculated. The mean platoon depth represents the mean platoon position reported by each CACC vehicle and it is a good indicator of the average size of platoons in the network. As shown in Fig. 12, there is a major distinction between the two types of ML strategies: the type with dedicated CACC lane (i.e. DL and DLA) and the type without. The smallest and greatest gaps observed between these two types of ML strategies are 0.16 at 25% MP and 0.25 at 35%, respectively. Fig. 12: Average platoon position The formation and dissolution of CACC platoons occur dynamically under mixed traffic con- ditions due to various reasons (e.g. vehicles exiting a freeway, non-CACC vehicles cutting in, or 15 communication disruption). At certain timeframe, there is a percentage of CACC vehicles that are in platoons, utilizing the intra-platoon, short following distance. On the other hand, there is the duration of the CACC vehicles in platoons. The dynamic changing of the two aforementioned measures during operation makes it difficult to spatially and temporally measure the effectiveness of forming CACC platoons. To account for this phenomenon, the vehicle hour platooned (VHP) is proposed in this study to quantify the temporospatial platooning performance for the entire net- work. The VHP is the accumulation of the number of CACC vehicles that are in platoons within an hour. In Fig. 13, the convex shape in UML and MML curves indicate that a non-linear growth of VHP as the MP increases. Compared to UML and MML, DL and DLA yield consistently higher values of VHP. The increasing trends in DL and DLA exhibit a linear pattern. Fig. 13: Vehicle hour platooned A score matrix system that factors in all the aforementioned performance measures is de- veloped. Table 5 summarizes the score assignment rule. The first four heterogeneous types of performance measures (i.e. mobility, safety, equity, and environmental impact) are determined based on the comparison with the BASE, whereas the CACC platooning score is determined by the comparisons among the four tested strategies at each level of MP. Table 5: Vehicle Control Parameters Parameter Traditional performance measure (mobility, safety, equity, and environmental impact) Evaluation Score improvement: 1 neutral: 0 degradation:-1 CACC platoon formation ranked among 4 strategies 1st: 4, 2nd: 3, 3rd: 2, 4th: 1 Fig. 14(a) and (b) show the evaluation scores for the network performance and platooning performance, respectively. Since the calculation of the two score categories is different, normal- 16 ization is performed before summation to obtain the scores in Fig. 14(c) that represent the overall evaluation scores between the critical MP range. When MP is below 25%, both UML and MML strategies are suitable options for CACC deployment, depending on whether an existing ML strat- egy is already in place. When the MR reaches above 30%, both MML and DLA strategies are viable options. Either DL or DLA is able to create homogeneous CACC traffic; however, DLA is able to provide better operational accommodation for CACC. The recommendations are summa- rized in Fig. 15. (a) traffic performance score (b) platooning performance score (c) normalized sum of score Fig. 14: Managed lane evaluation score Fig. 15: Managed lane policy recommendations 5. Conclusions This study focuses on assessing the effectiveness of four managed lane (ML) strategies for the near-term deployment of CACC under mixed traffic conditions. An integrated microscopic simulation testbed is employed to conduct the large-scale traffic simulation on the calibrated I-66 segment in Fairfax County, Virginia. Note that the types of the ML strategies tested in this study are not intended to be exhaustive. Additionally, the control parameters for CACC vehicles could vary when it comes to implementation. The desired time gaps used in this study are relatively con- servative (i.e., 1 s and 1.2 s), and it would become even shorter as the MP grows. The powertrain parameters used in this study could also be different in the future as lighter vehicles (without the 17 heavy safety frame) become viable due to drastically fewer crashes. Nevertheless, the simulations reveal several important aspects of deploying CACC in mixed traffic, most of which have rarely been discussed and yet the stakeholders have to take into account. Besides the traditional network- wide performance measures, a set of platoon performance measures are put forward to evaluate the platoon formation of CACC vehicles. Additionally, the evaluation score matrices considering various aspects of CACC deployment were proposed to demonstrate the overall suitability for each ML strategy. For locations without existing managed lane policy, mixing CACC traffic across all travel lanes is an acceptable option when the market penetration rate is below 30%. Mixed managed lane (e.g., HOV + CACC, HOT + CACC) is a versatile option for providing priority lane usage for CACC vehicles, especially in the locations where ML strategies have already been implemented. It is premature to implement CACC lane when market penetration (MP) is below 30%, due to the degradation of the traffic condition on the general purpose lanes. When the mid-range MP (30% to 55%) is reached, CACC lane begins to show its advantages for facilitating CACC clustering and forming homogeneous CACC traffic that is ultimately necessary for the further harness of the benefits of CACC (e.g., higher speed limit and short headway intra-platoon car following). Further exploration of the operation of CACC-ML is very much needed, including the ways to optimize the operational efficiency for the managed lane in conjunction with other CAV-based systems, such as speed harmonization, automated merging assistance, one-way vehicle communi- cation devices, etc. In addition, the clustering strategy for CACC-ML in mixed traffic conditions is an under-explored area at the current stage. Such operation should be performed at a tactical driving level beyond the basic operational level as defined by Society of Automotive Engineers In- ternational (2014). Moreover, the interactions between CAVs and HVs is another area that deserve more attentions. Thus far, the majority of the studies have been focusing on CAVs. The potential impacts on HVs from CAVs (e.g., platoon blockage, induced weaving) should not be overlooked, especially when active clustering strategy is employed. 18 Appendix A. Definitions for the abbreviations used Abbreviation V2V DOT CACC CAV DSRC MP HOV HOT XBL GP HV AHS ADS SARTRE PATH ML RFID GPS COM VEDM E-IDM CAH MOBIL UML MML DL DAL VMT VHT PTI SSAM MOE VHP DDT ODD Definition vehicle-to-vehicle department of transportation cooperative adaptive cruise control connected and automated vehicle dedicated short-range communication market penetration high occupancy vehicle high-occupancy toll exclusive bus lane general purpose vehicle human-driven vehicle automated highway system automated driving system Safe Road Trains for the Environment project Partners for Advanced Transportation Technology managed lane radio frequency identification global positioning system component object model Vissim external driver model enhanced intelligent driver model constant-acceleration heuristic minimizing overall braking induced by lane change unmanaged lane mixed managed lane dedicated lane dedicated lane with access control vehicle mile traveled vehicle hour traveled planning time index safety surrogate assessment model measure of effectiveness vehicle hour platooned dynamic driving task operational design domain 19 References Amoozadeh, M., Deng, H., Chuah, C.-N., Zhang, H. M., Ghosal, D., apr 2015. Platoon management with cooperative adaptive cruise control enabled by VANET. Vehicular Communications 2 (2), 110 -- 123. URL https://www.sciencedirect.com/science/article/pii/S2214209615000145 Arnaout, G., Arnaout, J.-P., 2014. Exploring the effects of cooperative adaptive cruise control on highway traffic flow using microscopic traffic simulation. Transportation Planning and Technology. Bu, F., Tan, H.-S., Huang, J., 2010. Design and field testing of a cooperative adaptive cruise control system. In: American Control Conference (ACC), 2010. IEEE, pp. 4616 -- 4621. Cai, Q., Saad, M., Abdel-Aty, M., Yuan, J., Lee, J., 2018. Safety impact of weaving distance on freeway facili- ties with managed lanes using both microscopic traffic and driving simulations. Transportation Research Record, 0361198118780884. California Department of Transportation, 2018. High-occupancy vehicle guidelines. Tech. rep. http://www.dot.ca.gov/trafficops/tm/docs/HOV_Guidelines-English-Edition-Jan2018. URL pdf Chen, C., Petty, K., Skabardonis, A., Varaiya, P., Jia, Z., 2001. Freeway performance measurement system: mining loop detector data. Transportation Research Record: Journal of the Transportation Research Board (1748), 96 -- 102. Chen, Q., Schmidt-Eisenlohr, F., Jiang, D., Torrent-Moreno, M., Delgrossi, L., Hartenstein, H., 2007. Overhaul of ieee 802.11 modeling and simulation in ns-2. In: Proceedings of the 10th ACM Symposium on Modeling, analysis, and simulation of wireless and mobile systems. ACM, pp. 159 -- 168. D´avila, A., 2012. Sartre report on infrastructure and environment. Tech. rep., Technical Report for European Commis- sion under the Framework 7 Programme Project 233683 Deliverable 5.2). Cambridge, UK: Ricardo UK Limited, 2012. http://www. sartre-project. eu/en/publications/Documents/SARTRE 5 002 PU. pdf. Eckhardt, J., 2016. European truck platooning challenge 2016. The Hague, Delta3. Federal Highway Administration, 2006. Making it there on time, all the time. Federal Highway Administration, 2008. Managed Lanes: A Primer. Tech. rep., Washington DC, United States. Gettman, D., Head, L., 2003. Surrogate safety measures from traffic simulation models. Transportation Research Record: Journal of the Transportation Research Board (1840), 104 -- 115. Hussain, O., Ghiasi, A., Li, X., 2016. Freeway lane management approach in mixed traffic environment with connected autonomous vehicles. arXiv preprint arXiv:1609.02946. Kesting, A., Treiber, M., Helbing, D., 2007. General lane-changing model mobil for car-following models. Trans- portation Research Record 1999 (1), 86 -- 94. Kesting, A., Treiber, M., Helbing, D., 2010. Enhanced intelligent driver model to access the impact of driving strate- gies on traffic capacity. Philosophical Transactions of the Royal Society of London A: Mathematical, Physical and Engineering Sciences 368 (1928), 4585 -- 4605. Killat, M., Hartenstein, H., 2009. An empirical model for probability of packet reception in vehicular ad hoc networks. EURASIP Journal on Wireless Communications and Networking 2009, 4. Kim, J., Park, S., Na, S., Lee, S., 2015. Sustainable highway operations: High occupancy lane and movable lane. International Journal of Transportation 3 (3), 1 -- 10. Larson, J., Kammer, C., Liang, K.-Y., Johansson, K. H., 2013. Coordinated route optimization for heavy-duty vehicle platoons. In: Intelligent Transportation Systems-(ITSC), 2013 16th International IEEE Conference on. IEEE, pp. 1196 -- 1202. Lee, J., Bared, J., Park, B., 2014. Mobility impacts of cooperative adaptive cruise control (cacc) under mixed traffic conditions. In: 93rd Annual Meeting of the Transportation Research Board, Washington, DC. Leidos, 2016. Simulation of evolutionary introduction of cooperative adaptive cruise control equipped vehicles into traffic. Tech. rep., Saxton Transportation Operations Laboratory. Li, T., Ma, J., Hale, D. K., 2019. High-occupancy vehicle lanes on the right: an alternative design for congestion reduction at freeway merge, diverge, and weaving areas. Transportation Letters 0 (0), 1 -- 13. URL https://doi.org/10.1080/19427867.2019.1584347 Li, Y., Wang, H., Wang, W., Xing, L., Liu, S., Wei, X., 2017. Evaluation of the impacts of cooperative adaptive cruise control on reducing rear-end collision risks on freeways. Accident Analysis & Prevention 98, 87 -- 95. 20 Liang, K.-Y., Mårtensson, J., Johansson, K. H., 2013. When is it fuel efficient for a heavy duty vehicle to catch up with a platoon? In: 7th IFAC Symposium on Advances in Automotive Control, Tokyo, Japan, September 4-7, 2013. Litman, T., 2002. Evaluating transportation equity. World Transport Policy & Practice 8 (2), 50 -- 65. Litman, T., 2017. Autonomous vehicle implementation predictions. Victoria Transport Policy Institute Victoria, Canada. Liu, H., Kan, X. D., Shladover, S. E., Lu, X.-Y., Ferlis, R. E., 2018. Modeling impacts of cooperative adaptive cruise control on mixed traffic flow in multi-lane freeway facilities. Transportation Research Part C: Emerging Technologies 95, 261 -- 279. URL http://www.sciencedirect.com/science/article/pii/S0968090X18310313 Milan´es, V., Shladover, S. E., Spring, J., Nowakowski, C., Kawazoe, H., Nakamura, M., 2014. Cooperative adaptive cruise control in real traffic situations. IEEE Trans. Intelligent Transportation Systems 15 (1), 296 -- 305. Nowakowski, C., Shladover, S. E., Cody, D., 2011. Cooperative Adaptive Cruise Control : Testing Drivers' Choices of Following Distances (January). PTV AG, 2015. Introduction to the com api. PTV AG. D-76131 Karlsruhe, Germany. PTV AG, 2017. PTV Vissim Drivermodel DLL Interface Documentation. PTV Group, Karlsruhe, Germany. PTV AG, 2018a. Emission Modelling. http://vision-traffic.ptvgroup.com/en-us/products/ptv-vissim/use-cases/ PTV AG, 2018b. Ptv vissim. Retrieved from PTV Group: http://vision-traffic. ptvgroup. com/enus/products/ptv- URL emissions-modelling/ vissim/use-cases/junction-geometry. Qom, S. F., Xiao, Y., Hadi, M., 2016. Evaluation of cooperative adaptive cruise control (cacc) vehicles on managed lanes utilizing macroscopic and mesoscopic simulation. In: Transportation Research Board 95th Annual Meeting. No. 16-6384. Rajamani, R., Shladover, S. E., 2001. An experimental comparative study of autonomous and co-operative vehicle- follower control systems. Transportation Research Part C: Emerging Technologies 9 (1), 15 -- 31. Rakha, H., Ahn, K., Trani, A., 2004. Development of vt-micro model for estimating hot stabilized light duty vehicle and truck emissions. Transportation Research Part D: Transport and Environment 9 (1), 49 -- 74. Saad, M., Abdel-Aty, M., Lee, J., Wang, L., 2018. Access design safety analysis for managed lanes including acces- sibility level and weaving length. Tech. rep. Shewmake, S., Jarvis, L., 2014. Hybrid cars and hov lanes. Transportation Research Part A: Policy and Practice 67, 304 -- 319. Shladover, S. E., Nowakowski, C., Lu, X.-Y., 2018. Using Cooperative Adaptive Cruise Control ( CACC ) to Form High-Performance Vehicle Streams Definitions , Literature Review and Operational Concept Alternatives. Tech. rep., University of California Berkeley: California Partners for Advanced Transportation Technology. URL https://escholarship.org/uc/item/3w6920wz Shladover, S. E., Nowakowski, C., Lu, X.-Y., Ferlis, R., 2015. Cooperative adaptive cruise control: Definitions and operating concepts. Transportation Research Record: Journal of the Transportation Research Board (2489), 145 -- 152. Shladover, S. E., Su, D., Lu, X.-Y., 2012. Impacts of cooperative adaptive cruise control on freeway traffic flow. Transportation Research Record 2324 (1), 63 -- 70. Society of Automotive Engineers International, 2014. Surface Vehicle Recommended Practice. Taxonomy and Defi- nitions for Terms Related to Driving Automation Systems for On-Road Motor Vehicles 4970, 724 -- 776. Songchitruksa, P., Bibeka, A., Lin, L. I., Zhang, Y., et al., 2016. Incorporating driver behaviors into connected and automated vehicle simulation. Tech. rep., Center for Advancing Transportation Leadership and Safety (ATLAS Center). The Victoria Transport Policy Institute, 2018. The victoria transport policy institute. URL http://www.vtpi.org/ Transportation Research Board, 2018. Dedicating Lanes for Priority or Exclusive Use by Connected and Automated Vehicles. The National Academies Press, Washington, DC. URL https://www.nap.edu/catalog/25366 Treiber, M., Hennecke, A., Helbing, D., 2000. Congested Traffic States in Empirical Observations and Microscopic 21 http://link.aps.org/doi/10.1103/PhysRevE.62.1805http://www.theo2.physik. Tsugawa, S., Jeschke, S., Shladover, S. E., 2016. A review of truck platooning projects for energy savings. IEEE Simulations. Physical Review E 62 (2), 1805. URL uni-stuttgart.de/treiber/ Trans. Intelligent Vehicles 1 (1), 68 -- 77. Tsugawa, S., Kato, S., Aoki, K., 2011. An automated truck platoon for energy saving. In: Intelligent Robots and Systems (IROS), 2011 IEEE/RSJ International Conference on. IEEE, pp. 4109 -- 4114. Van Arem, B., Van Driel, C. J., Visser, R., 2006. The impact of cooperative adaptive cruise control on traffic-flow characteristics. IEEE Transactions on intelligent transportation systems 7 (4), 429 -- 436. van Willigen, W., Haasdijk, E., Kester, L., 2013. A multi-objective approach to evolving platooning strategies in intelligent transportation systems. In: Proceedings of the 15th annual conference on Genetic and evolutionary computation. ACM, pp. 1397 -- 1404. Washington State Department of Transportation, 2006. Hot lane buffer and mid-point access design review. Tech. rep. URL https://www.wsdot.wa.gov/research/reports/fullreports/651.1.pdf Wiedemann, R., 1991. Modelling of rti-elements on multi-lane roads. In: Drive Conference (1991: Brussels, Bel- Wille, M., Rowenstrunk, M., Debus, G., 2007. Konvoi: Electronically coupled truck convoys. Human Factors for gium). Vol. 2. assistance and automation, 243. Xiang, W., Gozalvez, J., Niu, Z., Altintas, O., Ekici, E., 2009. Wireless access in vehicular environments. Zhang, X., Ma, J., Smith, B., Liu, J., 2018. Operational performance evaluation of the managed lane strategy for early deployment of cooperative adaptive cruise control. Tech. rep. Zhong, Z., 2018. Assessing the effectiveness of managed lane strategies for the rapid deployment of cooperative adaptive cruise control technology. Ph.D. thesis, Newark, NJ. URL pdf http://archives.njit.edu/vol01/etd/2010s/2018/njit-etd2018-033/njit-etd2018-033. Zhong, Z., Lee, J., 2018. Simulation Framework for Cooperative Adaptive Cruise Control with Empirical DSRC Module (forthcoming). The 44th Annual Conference of the IEEE Industrial Electronics Society. URL http://arxiv.org/abs/1810.06510 22
1209.5331	1	1209	2012-09-24T16:53:41	Reliability of swarming algorithms for mobile sensor network applications	[ "cs.MA" ]	There are many well-studied swarming algorithms which are often suited to very specific purposes. As mobile sensor networks become increasingly complex, and are comprised of more and more agents, it makes sense to consider swarming algorithms for movement control. We introduce a natural way to measure the reliability of various swarming algorithms so a balance can be struck between algorithmic complexity and sampling accuracy.	cs.MA	cs	Reliability of swarming algorithms for mobile sensor network applications Poster Summary for FFT 2012 Steven Senger University of Delaware [email protected] 1 Abstract There are many well-studied swarming algorithms (for example, [2], [4], [5], [6], [7], [8], and [9]) which are often suited to very specific purposes. As mobile sensor networks become increasingly complex, and are comprised of more and more agents, it makes sense to consider swarming algorithms for movement control. We introduce a natural way to measure the reliability of various swarming algorithms so a balance can be struck between algorithmic complexity and sampling accuracy. The main idea is to utilize relatively well-developed tools (see [1]) from areas like frame theory to provide theoretical guidance for algorithm selec- tion. 1 2 Frames, frame potential, and energy Frames were first introduced in [3]. A frame is a set of vectors which, like a basis of a vector space, span a certain space of interest, but with perhaps (many) more vectors than necessary to span the space. This theory is imme- diately relevant for transmitting complex information through noisy or lossy media. More precisely, let Hd be a d-dimensional Hilbert space, then the set of vectors F = {fj}n j=1 ⊂ Hd, is a frame if there exist constants A and B, called frame bounds, such that for any y ∈ Hd, the following holds: Akyk2 ≤ X hy, fji2 ≤ Bkyk2, j where h·, ·i denotes the inner product associated to the Hilbert space. The more control one has on the frame bounds, the more "well-behaved" a given frame is. One useful measure of how well a given frame behaves is the so-called frame potential, introduced in [1]. The lower the frame potential, the better the frame is suited to many tasks. Retaining the notation above, the frame potential of a given frame is defined to be F P (F ) = n n X X i=1 j=1 hfi, fji2. The frame potential is related to a very physically motivated quantity used in many areas of mathematics called energy. Again, using the same notation as above, the energy of a collection of vectors is defined to be E(F ) = 1 (cid:0)n 2(cid:1) n n X X k i=1 j>i 1 fi − fjk2 . 2 3 Mobile sensor networks As the number of agents in mobile sensor networks increases, it will be nec- essary to consider asymptotic analysis of the algorithms which control them. It stands to reason that tools from frame theory could be usefully employed to study mobile sensor networks by viewing them as frames with dynamic vectors. Here is an example of such a result. An excellent example would be a swarm of drones flying just above a forest fire, regularly measuring quantities like temperature and humidity. This swarm of sensors is moving coherently in a some direction. Each agent of the swarm is taking samples of some quantity at regular intervals along the way. The agents should be well-distributed, so that the the region under consideration is sampled with minimum bias. Also, whatever quantities are being sampled will be assumed to not change too much, on average, over some small, fixed distance. Notice that the complexity of a given algorithm can strongly influence the possible velocity for such applications. Theorem 1. Let X (t) = {xj(t)}n j=1 ⊂ Rd be the positions of agents in a swarm at time t. Let vj(t) denote the velocity of the agent at position xj at time t, and let vavg = 1 n n X j=1 vj. Suppose further that the desired resolution in space of the sampling is such that each agent has a minimum separation of δ > 0 from each other agent, and that vj(t) − vavg < cδ, for some small constant c, independent of n. Fi- nally, let ρ(t) denote the ratio of the area which is covered by δ-balls centered at the agents to the area of the smallest cube containing the swarm at time t. Then for all times t such that E(X (t)) = 1 (cid:0)n 2(cid:1) n n X X i=1 j>i 1 kxi(t) − xj(t)k2 ≤ C, for some value C, independent of n, ρ(t) ≥ C ′, for some other value C ′, which is also independent of n. 3 4 Conclusion Energy can be determined by simulation before any swarm is given a navi- gation algorithm, so that various parameter sets can be tested cheaply. If a given algorithm or parameter set has energy which depends on the number of agents, there could be an uneven distribution of sensors. However, the theo- rem above guarantees that the distribution of sensors will be asymptotically robust. References [1] Benedetto and Fickus, Finite normalized tight frames, Adv. Comput. Math. 18, pp. 357 -- 385 (2003). [2] Cui, Hardin, Ragade, Elmaghraby, A Swarm-based Fuzzy Logic Control Mobile Sen- sor Network for Hazardous Contaminants Localization, 2004 IEEE International Conference on Mobile Ad-hoc and Sensor Systems (2004). [3] Duffin and Schaeffer, A class of nonharmonic Fourier series, Trans. Amer. Math. Soc. 72 (1952) pp. 341 -- 366. [4] Kadrovach and Lamont, A Particle Swarm Model for Swarm-based Networked Sensor Systems, Proceedings of the 2002 ACM symposium on Applied computing, New York, NY, (2002). [5] Leonard, Paley, Lekien, Sepulchre, Fratantoni, and Davis, Collective Motion, Sensor Networks, and Ocean Sampling, Proceedings of the IEEE, Vol. 95, No. 1, January 2007. [6] Miller, Kolpas, Juchem Neto, and Rossi, A continuum three-zone model for swarms, Bull Math Biol. 2012 Mar; 74(3):536 -- 61. Epub 2011 Jul 29. [7] Smith, Chao, Li, Caron, Jones, and Sukhatme, Planning and Implementing Trajecto- ries for Autonomous Underwater Vehicles to Track Evolving Ocean Processes Based on Predictions from a Regional Ocean Model, International Journal of Robotics Re- search, 29(12) 1475 -- 1497, (2010). [8] Tillet, Rao, and Sahin, Cluster-Head Identification in Ad Hoc Sensor Networks us- ing Particle Swarm Optimization, 2002 IEEE International Conference on Personal Wireless Communications (2002). [9] Wang, Wang, Ma, An Improved Co-evolutionary Particle Swarm Optimization for Wireless Sensor Networks with Dynamic Deployment, Sensors 2007, 7, pp. 354 -- 370. 4
1212.0023	1	1212	2012-11-30T22:00:56	Emergence of Self-Organized Amoeboid Movement in a Multi-Agent Approximation of Physarum polycephalum	[ "cs.MA", "q-bio.CB" ]	The giant single-celled slime mould Physarum polycephalum exhibits complex morphological adaptation and amoeboid movement as it forages for food and may be seen as a minimal example of complex robotic behaviour. Swarm computation has previously been used to explore how spatiotemporal complexity can emerge from, and be distributed within, simple component parts and their interactions. Using a particle based swarm approach we explore the question of how to generate collective amoeboid movement from simple non-oscillatory component parts in a model of P. polycephalum. The model collective behaves as a cohesive and deformable virtual material, approximating the local coupling within the plasmodium matrix. The collective generates de-novo and complex oscillatory patterns from simple local interactions. The origin of this motor behaviour is distributed within the collective rendering is morphologically adaptive, amenable to external influence, and robust to simulated environmental insult. We show how to gain external influence over the collective movement by simulated chemo-attraction (pulling towards nutrient stimuli) and simulated light irradiation hazards (pushing from stimuli). The amorphous and distributed properties of the collective are demonstrated by cleaving it into two independent entities and fusing two separate entities to form a single device, thus enabling it to traverse narrow, separate or tortuous paths. We conclude by summarising the contribution of the model to swarm based robotics and soft-bodied modular robotics and discuss the future potential of such material approaches to the field.	cs.MA	cs	Emergence of Self-Organized Amoeboid Movement in a Multi-Agent Approximation of Physarum polycephalum Jeff Jones and Andrew Adamatzky Centre for Unconventional Computing, University of the West of England, Coldharbour Lane, Bristol BS16 1QY, UK. [email protected], [email protected] Abstract: The giant single-celled slime mould Physarum polycephalum exhibits complex morphological adaptation and amoeboid movement as it forages for food and may be seen as a minimal example of complex robotic behaviour. Swarm computation has previously been used to explore how spatio- temporal complexity can emerge from, and be distributed within, simple component parts and their interactions. Using a particle based swarm approach we explore the question of how to generate collective amoeboid movement from simple non-oscillatory component parts in a model of P. polycephalum. The model collective behaves as a cohesive and deformable virtual material, approximating the local coupling within the plasmodium matrix. The collective generates de-novo and complex oscillatory patterns from simple local interactions. The origin of this motor behaviour is distributed within the collective rendering is morphologically adaptive, amenable to external influence, and robust to simulated environmental insult. We show how to gain external influence over the collective movement by simulated chemo-attraction (pulling towards nutrient stimuli) and simulated light irradiation hazards (pushing from stimuli). The amorphous and distributed properties of the collective are demonstrated by cleaving it into two independent entities and fusing two separate entities to form a single device, thus enabling it to traverse narrow, separate or tortuous paths. We conclude by summarising the contribution of the model to swarm based robotics and soft-bodied modular robotics and discuss the future potential of such material approaches to the field. Keywords: amoeboid movement, swarm intelligence, distributed computation, collective behaviour, Physarum polycephalum Introduction: Classical robotics approaches typically connect separate sensory, control, and locomotion systems, using complex parts with little redundancy. Nature inspired robotics takes inspiration from aggregate populations, collective transport, segmentation of component parts, and soft-bodied motion in living systems. Swarm approaches to nature inspired computing seek to elucidate the sensory mechanisms and individual interactions which generate the complexity patterning and movement seen at very different scales in natural systems including car traffic dynamics (Helbing, 2001), human walking patterns (Helbing et al., 1998), (Helbing et al., 2001), flocking and schooling (Reynolds, 1987), collective insect movement (Buhl et al., 2006), (Kube and Bonabeau, 2000), and bacterial patterning (Matsushita et al., 1999), (Ben-Jacob, 2003). In all these examples there is a population of entities in space, coupled by sensory information. The different sensory coupling mechanisms between individuals in these apparently disparate systems has also been abstracted at a minimal level of self-propelled particles to find common mechanisms (Vicsek et al., 1995). Swarm approaches typically approximate collective movement at the population level. Distributed generation and control of movement also occurs at the individual level (Kennedy et al., 2001). Specific examples include flagellated movement of bacteria (Zhang et al., 2009), eukaryotes (Dreyfus et al., 2005), ciliated transport (Suh et al., 2000), peristaltic propulsion (Trimmer et al., 2006), (Saga and Nakamura, 2004), and amoeboid movement (Yokoi et al., 2003), (Umedachi et al., 2009). Many organisms make use of segmentation, employing identical or similar sub-units to form their body plan and generate structure and movement. Such modularity is attractive to computer science and robotics but presents challenges in terms of how to configure, communicate between, and control the devices. Population based approaches have also proved useful in tackling some of these issues. Internal propagation of simulated hormone signals was used to control gait patterns in simulated caterpillar-like modular robots (Salemi et al., 2001). Reconfiguration stimulated by crystalline mechanical transformation within a population of automata was demonstrated in (Rus and Vona, 2000) and future possibilities (and difficulties) in communicating between entities at larger scales with massive populations were discussed in (Goldstein et al., 2009). The exploitation of external environmental attractants (“seeds” and “scents”) was used to effect the reconfiguration of a simulated modular robotic system composed of simple finite state machines for the creation of specific manipulation structures (Bojinov et al., 2002), and pre-desired shapes (Stoy, 2006). Progress in biologically inspired robotics has also been made by considering the use of even simpler structures which straddle the boundary of non-living physical materials and living organisms including those acting as biological fibres and membranes (Zhang, 2003), lipid self assembly in terms of networks (Lobovkina et al., 2008), pseudopodium-like membrane extension (Lobovkina et al., 2009) and even those exhibiting simple chemotaxis responses (Lagzi et al.). Some engineering and biological insights have already been gained by studying the structure and function of what might be termed ‘semi-biological’ materials and the complex behaviour seen in such minimal examples raises questions about the lower bounds necessary for the emergence of apparently intelligent behaviour. An ideal hypothetical candidate for a biological machine would be an organism which is capable of the complex sensory integration, movement and adaptation of a living organism, yet which is also composed of a relatively simple material that is amenable to simple understanding and control of its properties. We suggest that the myxomycete organism, the true slime mould Physarum polycephalum, is a suitable candidate organism which meets both criteria; i.e. it is a complex organism, but which is composed of relatively simple materials. A giant single-celled organism, P. polycephalum is an attractive biological candidate medium for emergent motive force because the basic physical mechanism during the plasmodium stage of its life cycle is a self-organized system of oscillatory contractile activity which is used in the pumping and distribution of nutrients within its internal transport network. The organism is remarkable in that the control of the oscillatory behaviour is distributed throughout the almost homogeneous medium and is highly redundant, having no critical or unique components. The plasmodium is amorphous in shape and ranges from the microscopic scale to up to many square metres in size. It is a single cell syncytium formed by repeated nuclear division, comprised of a sponge-like actomyosin complex co-occurring in two physical phases. The gel phase is a dense matrix subject to spontaneous contraction and relaxation, under the influence of changing concentrations of intracellular chemicals. The protoplasmic sol phase is transported through the plasmodium by the force generated by the oscillatory contractions within the gel matrix. Protoplasmic flux, and thus the behaviour of the organism, is affected by changes in pressure, temperature, space availability, chemoattractant stimuli and illumination (Carlile, 1970; Durham and Ridgway, 1976; Kishimoto, 1958; Nakagaki et al., 1996; Nakagaki et al., 2000b; Takamatsu et al., 2000; Ueda et al., 1975). The P. polycephalum plasmodium can thus be regarded as a complex functional material capable of both sensory and motor behaviour. Indeed P. polycephalum has been described as a membrane bound reaction-diffusion system in reference to both the complex interactions within the plasmodium and the rich computational potential afforded by its material properties (Adamatzky et al., 2008). The study of the computational potential of the P. polycephalum plasmodium was initiated by Nakagaki et al. (Nakagaki et al., 2000a) who found that the plasmodium could solve simple maze puzzles. This research has been extended and the plasmodium has demonstrated its performance in, for example, path planning and plane division problems (Shirakawa and Gunji, 2010), (Shirakawa et al., 2009), spanning trees and proximity graphs (Adamatzky, 2007),(Adamatzky, 2008), simple memory effects (Saigusa et al., 2008), the implementation of individual logic gates (Tsuda et al., 2004) and P. polycephalum inspired models of simple adding circuits (Jones and Adamatzky, 2010). From a robotics perspective it was shown that by its adaptation to changing conditions within its environment, the plasmodium may be considered as a prototype micro-mechanical manipulation system, capable of simple and programmable robotic actions including the manipulation (pushing and pulling) of small scale objects (Adamatzky and Jones, 2008), transport and mixing of substances (Adamatzky, 2010c) and as a guidance mechanism in a biological—mechanical hybrid approach where the response of the plasmodium to light irradiation was used to provide feedback control to a robotic system (Tsuda et al., 2007). A P. polycephalum inspired approach to amoeboid robotics was demonstrated by Umedachi et. al. (Umedachi et al., 2009) in which an external ring of coupled oscillators, each connected to passive and tuneable springs was coupled to a fluid filled inner bladder. The compression of the peripheral springs mimicked the gel contractile phase and the flux of sol within the plasmodium was approximated by the coupled transmission of water pressure to inactive (softer) springs, thus deflecting the peripheral shape of the robot. The resulting movement exhibited flexible behaviour and amoeboid movement. This paper is also motivated by the complex sensory and foraging behaviour of P. polycephalum, however it takes a different approach from that of Umedachi et al. Instead of trying to build an amoeboid robot with pre-existing oscillatory components, we investigate how oscillatory behaviour may emerge from the local interactions between simple component parts to generate self-organized amoeboid movement. The method uses a swarm based, or multi-agent, population exploiting self-organization to behave as a collective virtual material. We demonstrate the emergence of the oscillatory phenomena within the material, its utilisation for collective, amorphous and controllable amoeboid movement, and mechanisms for external control of the material by simulated chemoattraction and repulsion by light hazards. We conclude by discussing some novel properties of the approach within the context of existing research into swarm robotics and modular robotics. Methods: To investigate the use of emergent oscillatory phenomena for collective movement we employ a simple extension to the particle approximation of P. polycephalum network adaptation in (Jones, 2010b) which was shown to generate dynamical emergent transport networks. In this approach a plasmodium is composed of a population of mobile particles with very simple behaviours, residing within a 2D diffusive environment. A discrete 2D lattice (where the features of the environment arena are mapped to greyscale values in a 2D image) stores particle positions and also the concentration of a local factor which we refer to generically as chemoattractant. The ‘chemoattractant’ factor actually represents the hypothetical flux of sol within the plasmodium. Free particle movement represents the sol phase of the plasmodium. Particle positions represent the fixed gel structure (i.e. global pattern) of the plasmodium. Particles act independently and iteration of the particle population is performed randomly to avoid introducing any artifacts from sequential ordering. Particle behaviour is divided into two distinct stages, the sensory stage and the motor stage. In the sensory stage, the particles sample their local environment using three forward biased sensors whose angle from the forward position (the sensor angle parameter, SA), and distance (sensor offset, SO) may be parametrically adjusted (Figure 1a). The offset sensors represent the overlapping and intertwining filaments within the transport networks and plasmodium, generating local coupling of sensory inputs and movement to form networks of particles (Figure 1c) and sheets of particles (Figure 1d). The SO distance is measured in pixels and a minimum distance of 3 pixels is required for strong local coupling to occur. The coupling effect increases as SO increases. [Sensory stage] - Sample chemoattractant map values - if (F > FL) && (F > FR) - Continue facing same direction - Else if (F < FL) && (F < FR) Rotate by RA towards larger of FL and FR - Else if (FL < FR) Rotate right by RA - Else if (FR < FL) Rotate left by RA - Else Continue facing same direction Figure 1: Agent particle morphology, algorithm, and aggregation types. a) Schematic illustration of agent particle showing central position and offset sensors. b) Pseudocode of agent sensory algorithm to align with strongest source from the forward biased sensors. c) Schematic illustration of agent particles in transport network configuration d) Schematic illustration of agent particles within simulated plasmodial sheet. During the sensory stage each particle changes its orientation to rotate (via the parameter rotation angle, RA) towards the strongest local source of chemoattractant (Figure 1b). After the sensory stage, each particle executes the motor stage and attempts to move forwards in its current orientation (an angle from 0—360) by a single pixel. Each lattice site may only store a single particle and—critically—particles deposit chemoattractant into the lattice only in the event of a successful forwards movement (Figure 2a). If the next chosen site is already occupied by another particle the default (non-oscillatory) behaviour is to abandon the move, remain in the current position, and select a new random direction (Figure 2b). Diffusion of the collective chemoattractant signal is achieved via a simple 3x3 mean filter kernel with a damping parameter (set to 0.07) to limit the diffusion distance of the chemoattractant. The low level particle interactions result in complex pattern formation. The population spontaneously forms dynamic transport networks showing complex evolution and quasi-physical emergent properties, including closure of network lacunae, apparent surface tension effects and network minimisation. An exploration of the possible patterning parameterisation was presented in (Jones, 2010a). Although the particle model is able to reproduce many of the network based behaviours seen in the P. polycephalum plasmodium such as spontaneous network formation, shuttle streaming and network minimisation, the default motor behaviour does not exhibit oscillatory phenomena and inertial surging movement, as seen in the organism. This is because the default action when a particle is blocked (i.e. when the chosen site is already occupied) is to randomly select a new orientation, resulting in very fluid network evolution resembling the relaxation evolution of soap films and the lipid nanotube networks seen in (Lobovkina et al., 2008). The oscillatory phenomena seen in the plasmodium are thought to be linked to the spontaneous assembly/disassembly of actomyosin and cytoskeletal filament structures within the plasmodium which generate contractile forces on the protoplasm within the plasmodium. The resulting shifts between gel and sol phases prevent (gel phase) and promote (sol phase) cytoplasmic streaming within the plasmodium. To mimic this behaviour in the particle model requires only a simple change to the motor stage. Instead of randomly selecting a new direction if a move forward is blocked, the particle increments separate internal co-ordinates until the nearest cell directly in front of the particle is free. When a cell becomes free, the particle occupies this new cell and deposits chemoattractant into the lattice (Figure 2c). The effect of this behaviour is to remove the fluidity of the default movement of the population. The result is a surging, inertial pattern of movement dependent on population density (the population density specifies the initial amount of free movement within the population). The strength of the inertial effect can be damped by a parameter (pID) which determines the probability of a particle resetting its internal position coordinates, lower values providing stronger inertial movement. When this simple change in motor behaviour is initiated surging movements are seen and oscillatory domains of chemoattractant flux spontaneously appear within the virtual plasmodium showing characteristic behaviours: temporary blockages of particles (gel phase) collapse into sudden localised movement (solation) and vice versa. The oscillatory domains themselves undergo complex evolution including competition, phase changes and entrainment. We utilise these dynamics below to investigate the possibility of generating useful patterns of regular oscillations which may be coupled to provide motive force. Figure 2: Particle motor behaviour in non-oscillatory and oscillatory conditions a) A particle moves forward to occupy a vacant cell and deposits chemoattractant at the new location. b) A non-oscillatory particle will select a new random location if the preferred cell is occupied. c) If the preferred cell is occupied, the oscillatory particle will maintain the direction until a cell in front is free. Model Setup The emergence of oscillatory behaviour in the model corresponds to differences in distribution of protoplasm within the plasmodium and subsequent changes in thickness of the plasmodium. In the P. polycephalum plasmodium, changes in thickness of the plasmodium membrane are used to provide impetus (pumping of material through the vein network, or bulk movement of the plasmodium). There is known to be a reciprocal relationship between the spontaneous contraction of the plasmodium and the subsequent transport of protoplasm away from that region (Takamatsu et al., 2000). Thus the region undergoing contraction becomes thinner (allowing more light transmission when illuminated) as protoplasm is being transported away from the contraction zone. Protoplasmic tubes connected to the contraction zone become thicker as more protoplasm is transported from the zone (thus allowing less light transmission). In the computational model the transport of particles represents the free flux of protoplasm within the material and a local increase in flux (particle movement) is indicated in the supplementary video recordings by a temporary increase in greyscale brightness at this location. A decrease in the bulk movement of particles represents local congestion (lack of transport) and is indicated by a decrease in greyscale brightness (since deposition of chemoattractant factor only occurs in the event of successful forward movement). Supplementary video recordings showing the initiation and entrainment of oscillatory phenomena, and recordings from specific figures in the paper text can be found at: http://uncomp.uwe.ac.uk/jeff/collectivemovement.htm. For visual clarity in the static images, the greyscale images are inverted (dark areas indicate greater flux). The particle population environment is a 2D lattice, represented by a digitised image configured to represent the habitat of the experimental plasmodium. The habitat is composed of ‘wall’ regions where particle occupancy and movement cannot occur, ‘vacant’ regions where occupancy and movement are possible and, where relevant, ‘stimulus’ regions which provide attraction stimuli, or repulsion stimuli, to the particle population. In the experiments there is an initial period where oscillatory behaviour is not initially activated and this results in self-organized regularly spaced domains within the collective. When oscillatory motor behaviour is induced these regular domains collapse and small domains of disorganized oscillators begin to emerge. The oscillation domains are composed of local areas of flux (brighter) and obstruction (darker). Over time these domains compete and coalesce, causing entrainment of the population into regular oscillation patterns, influenced by both the particle sensory parameters and also by the shape of the experimental arenas. Emergence of Amoeboid Movement P. polycephalum utilises internal protoplasmic transport to migrate towards nutrient sources and away from hazardous sources (Durham and Ridgway, 1976), and adapts its gross body plan to changing environments (Nakagaki et al., 2004; Nakagaki et al., 2007). Small plasmodia can shift the entire plasmodium away from unfavourable conditions such as bacterial or fungal contamination. The plasmodium is also notable for its ability to survive physical damage; fragments of plasmodium excised can survive independently and individual plasmodia may be fused to form a single organism. To utilise a vehicular analogy, P. polycephalum not only represents the internal mechanicals (motive force mechanism, transmission coupling), but also the moving vehicle itself, and is a vehicle which can survive the removal of parts, the introduction of new foreign parts and the repair of damaged parts. We set out to explore the behaviour of the particle collective to assess its behaviour when compared to a fragment of P. polycephalum plasmodium. When oscillatory motor behaviour is not used the particle collective condenses into a uniform circular shape as the initial transport network condenses (Figure 3a). The non-oscillatory blob shows regular vacancy domains (dark areas) and the fluid particle motion afforded by the non-oscillatory motor condition ensures that the blob is cohesive and takes a minimal shape. The non-oscillatory blob is also resilient to external perturbation. When excited by an externally applied source of chemoattractant (Figure 3a, mouse position in fourth image), the deformation of the collective induced by the stimulus as it is attracted to the stimuli is repaired when the stimulus is removed, the collective returning to its minimal shape. When oscillatory motor behaviour is initialised the regular domains collapse as the particle motion becomes less fluid (Figure 3b) and oscillations travel through the collective. Because the small collective is not constrained by any externally applied pattern the oscillations distort the shape of the collective. When the pID parameter is further reduced to 0.01 there is even greater restriction on the fluidity of individual particle movement and the oscillations become stronger and distort the collective’s boundary significantly. The large shift of a mass of particles causes the collective to move across its environment. The cohesion of the collective is maintained but other SA/RA parameter settings, combined with lower sensor interaction (SO) distance, can result in the fragmentation of the collective (see supplementary material for examples of oscillation patterns using different SA/RA combinations). a) b) c) Figure 3: Initiation of oscillatory behaviour and amoeboid movement in simulated plasmodium Images show flux within simulated plasmodium over time. Images anchored to fixed location. a) Condensation of blob material in non-oscillatory behaviour shows vacancy domains and resilience to deformation. Collective of 9380 particles, SA90, RA45, SO15 b) Initial non-oscillatory collective with regular vacancy domains (left) and onset of oscillatory behaviour at 1950, 2039, 2057, 2086 and 2122 scheduler steps. pID set to 0.05. All other parameters as in (a) c) Reduction of pID to 0.01 results in stronger oscillations and amoeboid movement at 4648, 4684. 4720, 4768 and 4836 scheduler steps. All other parameters as in (a) Persistent Movement in a Small Blob Fragment The amoeboid movement seen in Figure 3 occurs because the oscillation waves distort the boundary of the collective whilst it is still able to maintain a cohesive whole. Because the population maintains its cohesion, any distortion of the boundary on one side must result in a shift in population distribution from the opposite side (since the collective is non-compressible and occupies a fixed area). The diameter of the collective (a function of the number of pixel sized particles comprising it) must be large enough to for oscillations to emerge and to confine an oscillation pattern within it. When the collective is comprised of only a relatively small number of particles, the distortion of the boundary forms an approximately semicircular domed shape and the small number of particles ensures that the collective cannot maintain a fully circular shape. However the persistence of forward movement generated by the oscillatory motor behaviour at low pID values causes the dome shape itself to be maintained over time, and the small blob fragment is able to move forwards (Figure 4a). Movement of the fragment is relatively smooth and different from the pulsatile motion observed in larger collectives. Particles move towards the front of the domed profile (Figure 4b, dark pixels) and then, over time, move to the side. Particles at the sides of the fragment ultimately fall behind only to re-enter the dome at the centre. Higher pID values result in more frequent changes of direction of the fragment as the dome shaped front profile cannot be maintained. If the population size is increased, the single sided dome shape cannot be maintained and the resultant motion becomes pulsatile and chaotic. The movement of small blob fragments in the particle collective mimics amoeboid motion observed in P. polycephalum on a non-nutrient substrate (Figure 4c) which has been shown to be equivalent to the propagation of wave fragments in sub-excitable media (Adamatzky et al., 2008; Adamatzky, 2010b). Both artificial and real blobs exhibit reflection (reversion of direction) when encountering the boundary of their environment and can both be directionally guided by the placement of attractants and repellents (Adamatzky, 2010a). a) Figure 4: Persistent forward movement of blob fragments in small simulated plasmodia b) c) Images show flux within simulated plasmodium fragments over time. Images anchored to fixed location. a) Non-oscillatory condition, initiation of oscillatory behaviour and (final 3 images) self perpetuating transport of blob fragment. Chemoattractant flux concentration is greatest at the front of the dome shape b) Enlargement showing particle composition of the moving blob fragment. The persistent shape is maintained despite repeated turnover of component parts. Population size 900 particles, SA90, RA45, SO9, pID 0.001 c) Independently moving small plasmodial fragment of Physarum polycephalum External Influence of Collective Movement – Attraction and Repulsion Movement of the P. polycephalum plasmodium is strongly affected by local environmental conditions. Attractant sources (such as increasing temperature gradients and chemoattractant nutrients) cause the plasmodium to move and grow towards the attractants whereas repulsive sources (salts, dry regions) cause the plasmodium to try to avoid such regions (Durham and Ridgway, 1976). The plasmodium is able to integrate many separate localised inputs to compute its response to the environment. One method in which this is achieved is by the modulation of local oscillation patterns in response to attractants or hazards – attractants tend to increase localised oscillation strength and hazards decrease oscillation strength. We set out to see if a localised response to external influences could be used to govern the collective movement of the particle population. Attractant sources were previously used as a method to confine the collective to a region by pinning it down. By externally presenting an attractant source (effectively a simulated nutrient source) near to a cohesive blob of virtual material (circle in Figure 5a) a concentration gradient emerged from the source (Figure 5b). When the diffusion gradient reached the sensors of the closest particles at the front of the collective it provoked local movement towards the source. The cohesion within the collective resulted in a pseudopodium-like extension of the border region which extended towards the source (Figure 5c), ultimately engulfing it. Travelling waves spontaneously emerged within the collective which were directed at the source, causing the collective to shift its position towards the source (Figure 5d). Consumption of the source was simulated by simply decrementing the value projected to the diffusion field when the source was covered by a particle. When the source was consumed by the population, the collective regained its previous approximately circular shape. Figure 5: Control of amoeboid movement by attraction b) a) c) d) Images show flux within simulated plasmodium over time. Images anchored to fixed location. a) A persistent source of chemoattractant is placed into the diffusion field which provokes the extension of the collective. The collective engulfs the source by moving via travelling waves towards the source. When the source is exhausted the collective re-adopts its original approximately circular shape. b) Schematic illustration of attraction of the collective by the nutrient source (c) Migration of leading particles towards the source (d) Emergence of travelling waves pulling the collective to engulf the source. To approximate the repulsion of the collective to hazardous sources such as the simulated response to irradiation by visible light we added a condition to the sensory stage of the algorithm to the effect that if any particles of the collective were in a region exposed to ‘light’ (a defined area within the arena), those particles would have their sensitivity to chemoattractant diminished whilst they remained in this region (this achieved by multiplying the sampled sensor values with a weighting factor less than 1, lower values generating a stronger response to irradiation). The effect of exposing regions of the collective to simulated light damage was that the collective immediately started to move away from the irradiated region (Figure 6a). Specifically, oscillation waves moved from the irradiated region towards the unexposed regions. The shift of particles from the irradiated region eventually moved the collective away from the stimulus. The a) cause of movement away from the light can be found at the interface between irradiated and unexposed areas. Before irradiation (Figure 6b) all regions of the collective are equally attractive to the particles (subject to fluctuations caused by discrepancies in particle movement and intrinsic oscillations within the collective). There is a strong coupling between the particles in the collective caused by the offset sensor distance. Some of the particles at the interface of the irradiated region will receive input from the unexposed region and will be attracted to that area because the chemoattractant concentration in unexposed areas is perceived as greater due to the damping in irradiated regions (Figure 6c). The movement of particles near the interface towards unexposed regions causes both new vacant spaces (Figure 6d) and also an increase in chemoattractant concentration (because only mobile particles deposit chemoattractant). This results in a further increased attraction to the interface region, until eventually the entire collective has migrated from the irradiated region. Figure 6: Avoidance of simulated light irradiation by particle collective b) c) d) Images show flux within simulated plasmodium over time. Images anchored to fixed location. a) Area within dashed box is stimulated with simulated light irradiation. Particle collective oscillates sending travelling waves towards the region which is unexposed, moving the collective away from the irradiated region. b) Schematic illustration of condition before irradiation – all particles sense equal concentration of chemoattractant c) Irradiated areas (top) are perceived as weaker concentration d) Particles at the irradiation interface are more attracted to unexposed areas. Migration across interface causes chemoattractant deposition, causing further attraction to region and vacant space. Morphological Adaptation of the Collective The previous results demonstrated that the collective changes its shape during self-oscillatory behaviour and also in response to simulated attractants and hazards. The collective retains its typically circular shape due to the cohesion of the individual particles making up the population. When the morphology of the collective is disturbed by its movement towards, or away from, externally applied stimuli it can reform the original shape when the stimulus is removed. An adaptive morphology is a very desirable property in robotic devices since it imbues the robot with great flexibility of size and movement, enabling it to traverse environments which traditionally are difficult to navigate (for example narrow spaces, gratings etc.). This feature is only possible because the properties of the movement and guidance of the collective are distributed throughout the collective and not located in fixed sized and inflexible units as is the case with conventional robotic systems. This is also the case with the P. polycephalum plasmodium which adapts its shape and growth patterns in response to its environment. One of the most remarkable properties of the plasmodium is the ability to survive external damage beyond simple attraction and repulsion. A piece of plasmodium excised from the growing tip can survive, and indeed continue to move and grow as an independent entity. Furthermore two independent plasmodia can fuse to form a single plasmodium when placed in close proximity. These phenomena are not only desirable from a robotics perspective in terms of resilience and damage repair, but offer new and as yet little explored opportunities in robotic movement and control. We set out to find if the particle collective could also replicate these highly desirable features as seen in the real plasmodium. We took a large single oscillating collective (5000 particles) and applied a narrow band of hazardous simulated light irradiation through its centre (Figure 7a, dashed box represents irradiated area). Particles immediately began to surge away from the irradiated region on both sides and the collective narrowed in diameter and became further pinched in shape until the collective was cleaved into two independently controllable blobs. The cleavage mechanism can be applied in different ways. For example both resulting blobs can be of equal size and have similar oscillatory properties. Alternately it is possible to cleave the collective in such a way to have one large stationary blob and one smaller mobile blob (recall that a blob which is small enough will be able to move spontaneously in a persistent direction). It is possible to guide each blob independently using either a pulling type mechanism (externally applied attractants) or a pushing type mechanism (simulated irradiation). In Figure 7b we guide the lower right blob (arrowed) towards the larger blob by pushing it from its opposite side with simulated irradiated light. As the blobs become closer (specifically, to a separation distance which is sufficient for the border particles in each blob to sense the chemoattractant flux in the other blob) the closest border regions of each blob surge towards each other and a single larger collective is formed by the fusion. Cleavage and guidance of the particle collective by simulated hazards reproduces the control of P. polycephalum plasmodial migration (Figure 7c) by repellents (Adamatzky, 2010a). a) b) Figure 7: Controlling blob morphology by splitting and fusion. c) Images show flux within simulated plasmodium over time. Images anchored to fixed location. a) ‘Blob’ of aggregate particles is split by applying simulated light irradiation (dashed box), disrupting particle flux in that region. The single blob separates into two smaller blobs, each capable of individual oscillation and external control. b) Fusion of two independent particle aggregates. The blob on the lower right is guided diagonally upwards in the direction of the arrow towards the larger blob. The two independent blobs fuse forming a single aggregate. c) Cleavage of plasmodium of P. polycephalum. Fragment of plasmodium initialised at triangular region. Crystal of potassium chloride placed at circular region. Diffusion of potassium chloride cleaves the plasmodium into two independent fragments. As an example of the robotic flexibility endowed by the adaptive morphology of the collective, and the guidance mechanisms enabled by its external control, we show in Figure 8 three examples of how a blob can be guided externally (in this case a repulsive pushing by simulated irradiation) in order to traverse narrow, separated or tortuous paths where the path may be much narrower than the diameter of the collective itself. In Figure 8a, the blob has a diameter of approximately 52 pixels and is guided through a channel of width 30 pixels by automatically reconfiguring its morphology. In Figure 8b, the blob is split into two parts by irradiating a rectangular region at the entrance of the channel. The two separated blobs may be guided separately by simulated radiation and are re-fused near the exit of the channels. Finally, in Figure 8c, the blob automatically separates its structure in response to the different obstacles and re-forms its shape by mutual attraction and cohesion of the particles. No fine control of the individual components, nor complex pre-determinism of path choice, is necessary; the collective is guided by simple avoidance of the simulated irradiation (the irradiation location is not shown, but follows the previous examples of simply exposing the rearward part of the collective to push it forwards). When clear of the obstacles, the original circular shape is reformed. If the blob is returned backwards through the obstacle path again the path chosen can be somewhat different to the original path with the same result. a) b) Figure 8: Exploiting the adaptive morphology of the collective. c) Aggregate blob of simulated plasmodium is pushed by the application of external simulated light irradiation (illuminated at the left side of the aggregate, not shown). a) Large aggregate moves through a channel narrower than the aggregate diameter b) Aggregate is cleaved by hazard irradiation (boxed) and guided separately before being refused into original size c) Aggregate is pushed through tortuous narrow grating and reforms original aggregate shape. Conclusion and Discussion: We have examined the problem of generating collective amoeboid movement in a multi-agent approximation of the plasmodium of true slime mould Physarum polycephalum using self-organization between simple non-oscillatory component parts. P. polycephalum is attractive because it satisfies many physical and computational properties which are desirable in robotics applications (self-oscillatory, simple components, distributed sensory and motor control, integration of multiple sensory stimuli, amorphous and adaptive shape, amenable to external influence, resilience to damage, self repair) and can be regarded as a living example of a so-called smart material because of the way in which it combines robotic and control functions in a distributed manner throughout its constituent material. How can such complex behaviour emerge from the interactions between simple component parts? Answering this question may provide insights into the development of non-living smart materials which would have the advantages demonstrated by the P. polycephalum plasmodium, but without some of its limitations, such as slow speed, fragility and unpredictability. Because of its simple components and structure, any search for a ‘secret source’ of the plasmodium’s complexity would be fruitless. Instead we set about answering the question by posing in reverse: rather than try to find out how the organism produces such complex behaviour from simple parts, is it possible to artificially generate similarly complex behaviour from simple parts and interactions? We utilised a previous particle model of emergent transport networks where a collective of identical particles with identical behaviour was used to construct and minimise synthetic and emergent transport networks. We modified its motor algorithm in a very simple way so that, instead of smooth network flow, we obtained a more resistant flow of particles. The addition of transient resistance to particle flux was sufficient to generate complex oscillatory behaviour strongly similar to that observed in the plasmodium itself. The condensation of the collective in its virtual environment resulted in cohesive blob-like sheets of virtual material held together by cohesion which exhibited spontaneous internal oscillations. The oscillations were composed of regular spatial domains of mobile particles and temporarily restricted particles. The wave-like propagation of the oscillatory domains resulted in a collective amoeboid movement as the perimeter of the particle collective was deformed by the internal oscillations. External control of the blob sheets was achieved by stimulating the collective with simulated attractants (pulling the collective towards the source) and repellents (‘pushing’ the collective away from simulated light irradiation). We were also able to reproduce the resilience of the P. polycephalum plasmodium to damage by cleaving the collective into two separate and independent blobs and fusing two independent blobs to form a single functional blob. Finally the adaptive morphology of the collective was demonstrated by guiding the collective through an obstacle field narrower than the diameter of the collective itself. The flexible modularity and morphology seen in this approach builds upon previous work on the external influence and control of robotic devices (Bojinov et al., 2002; Salemi et al., 2001) by enabling separate and functional modules of arbitrary size to be created and re-formed from a single collective. The results demonstrate how very simple and local low-level interactions within a simple material can generate complex and emergent behaviour which appear to transcend the capabilities of the simple matter of which they are composed. As noted by previous swarm approaches, there need not be anything magical or special about the properties of these materials; the complexity emerges merely from their interactions (Buhl et al., 2006; Reynolds, 1987; Vicsek et al., 1995). This complex behaviour is harnessed effortlessly by organisms such as P. polycephalum as part of a parsimonious survival strategy, enabling their persistence in unpredictable, changeable and hazardous environmental conditions. By understanding the generative mechanisms underlying the complex behaviour it may be possible to incorporate these features within real physical materials for small scale robotic devices. By utilising the robotic substrate material itself for distributed computation, transport and movement it may be possible to reduce the total number of component parts and also reduce the number of different types of components, thus further simplifying the production of the devices. Acknowledgements: The work was partially supported by the Leverhulme Trust research grant F/00577/1 “Mould intelligence: biological amorphous robots”. References: Adamatzky, A., 2007. Physarum machines: encapsulating reaction–diffusion to compute spanning tree. Naturwissenschaften. 94, 975-980. Adamatzky, A., 2008. Developing proximity graphs by Physarum Polycephalum: Does the plasmodium follow Toussaint hierarchy? Parallel Processing Letters. 19, 105-127. Adamatzky, A., De Lacy Costello, B., Shirakawa, T., 2008. Universal computation with limited resources: Belousov-Zhabotinsky and Physarum Computers. International Journal of Bifurcation and Chaos. 18, 2373-2389. Adamatzky, A., Jones, J., 2008. Towards Physarum robots: computing and manipulating on water surface. Journal of Bionic Engineering. 5, 348-357. Adamatzky, A., 2010a. Routing Physarum with repellents. The European Physical Journal E: Soft Matter and Biological Physics. 31, 403-410. Adamatzky, A., 2010b. If BZ medium did spanning trees these would be the same trees as Physarum built. Physics Letters A. 373, 952-956. Adamatzky, A., 2010c. Manipulating substances with Physarum polycephalum. Materials Science and Engineering: C. 30, 1211-1220. Ben-Jacob, E., 2003. Bacterial self-organization: co-enhancement of complexification and adaptability in a dynamic environment. Philosophical Transactions: Mathematical, Physical and Engineering Sciences. 361, 1283-1312. Bojinov, H., Casal, A., Hogg, T., 2002. Multiagent control of self-reconfigurable robots. Artificial Intelligence. 142, 99-120. Buhl, J., Sumpter, D.J.T., Couzin, I.D., Hale, J.J., Despland, E., Miller, E.R., Simpson, S.J., 2006. From disorder to order in marching locusts. Science. 312, 1402. Carlile, M.J., 1970. Nutrition and chemotaxis in the myxomycete Physarum polycephalum: the effect of carbohydrates on the plasmodium. Journal of General Microbiology. 63, 221. Dreyfus, R., Baudry, J., Roper, M.L., Fermigier, M., Stone, H.A., Bibette, J., 2005. Microscopic artificial swimmers. NATURE-LONDON-. 7060, 862. Durham, A.C., Ridgway, E.B., 1976. Control of chemotaxis in Physarum polycephalum. Journal of Cell Biology. 69, 218-223. Goldstein, S.C., Mowry, T.C., Campbell, J.D., Ashley-Rollman, M.P., De Rosa, M., Funiak, S., Hoburg, J.F., Karagozler, M.E., Kirby, B., Lee, P., 2009. Beyond audio and video: Using claytronics to enable pario. AI Magazine. 30, 29. Helbing, D., MolnÃ¡r, P., Schweitzer, F., 1998. Computer simulations of pedestrian dynamics and trail formation. Arxiv preprint cond-mat/9805074. Helbing, D., 2001. Traffic and related self-driven many-particle systems. Reviews of Modern Physics. 73, 1067-1141. Helbing, D., Molnar, P., Farkas, I.J., Bolay, K., 2001. Self-organizing pedestrian movement. Environment and planning B. 28, 361-384. Jones, J., 2010a. Characteristics of Pattern Formation and Evolution in Approximations of Physarum Transport Networks. Artificial Life. 16, 127-153. Jones, J., 2010b. The Emergence and Dynamical Evolution of Complex Transport Networks from Simple Low-Level Behaviours. International journal of Unconventional Computing. 6, 125-144. Jones, J., Adamatzky, A., 2010. Towards Physarum Adders. BioSystems. 101, 51-58. Kennedy, B., Melhuish, C., Adamatzky, A. Eds.), 2001. Biologically Inspired Robots. SPIE Press. Kishimoto, U., 1958. Rhythmicity in the protoplasmic streaming of a slime mould, Physarum polycephalum. The Journal of General Physiology. 41, 1223-1244. Kube, C.R., Bonabeau, E., 2000. Cooperative transport by ants and robots. Robotics and Autonomous Systems. 30, 85-101. Lagzi, I., Soh, S., Wesson, P.J., Browne, K.P., Grzybowski, B.A., Maze solving by chemotactic droplets. Journal of the American Chemical Society. 132, 1198-1199. Lobovkina, T., Dommersnes, P.G., Tiourine, S., Joanny, J.F., Orwar, O., 2008. Shape optimization in lipid nanotube networks. The European Physical Journal E-Soft Matter. 26, 295-300. Lobovkina, T., Gözen, I., Erkan, Y., Olofsson, J., Weber, S.G., Orwar, O., 2009. Protrusive growth and periodic contractile motion in surface-adhered vesicles induced by Ca2+-gradients. Soft Matter. 6, 268-272. Matsushita, M., Wakita, J., Itoh, H., Watanabe, K., Arai, T., Matsuyama, T., 1999. Formation of colony patterns by a bacterial cell population. Physica A: Statistical Mechanics and its Applications. 274. Nakagaki, T., Umemura, S., Kakiuchi, Y., Ueda, T., 1996. Action spectrum for sporulation and photoavoidance in the plasmodium of Physarum polycephalum, as modified differentially by temperature and starvation. Photochemistry and photobiology. 64, 859-862. Nakagaki, T., Yamada, H., Toth, A., 2000a. Maze-solving by an Amoeboid Organism. Nature. 407. Nakagaki, T., Yamada, H., Ueda, T., 2000b. Interaction between cell shape and contraction pattern in the Physarum plasmodium. Biophysical Chemistry. 84, 195-204. Nakagaki, T., Kobayashi, R., Nishiura, Y., Ueda, T., 2004. Obtaining multiple separate food sources: behavioural intelligence in the Physarum plasmodium. Proceedings of the Royal Society B: Biological Sciences. 271, 2305-2310. Nakagaki, T., Saigusa, T., Tero, A., Kobayashi, R., 2007. Effects of amount of food on path selection in the transport network of an amoeboid organism. Vol., ed.^eds. World Scientific Pub Co Inc, pp. 94. Reynolds, C., 1987. Flocks, herds and schools: A distributed behavioral model. ACM SIGGRAPH Computer Graphics. 21, 25-34. Rus, D., Vona, M., 2000. A physical implementation of the self-reconfiguring crystalline robot. Vol. 2, ed.^eds. IEEE, pp. 1726-1733 vol. 2. Saga, N., Nakamura, T., 2004. Development of a peristaltic crawling robot using magnetic fluid on the basis of the locomotion mechanism of the earthworm. Smart materials and structures. 13, 566. Saigusa, T., Tero, A., Nakagaki, T., Kuramoto, Y., 2008. Amoebae anticipate periodic events. Physical Review Letters. 100, 18101. Salemi, B., Shen, W.M., Will, P., 2001. Hormone-controlled metamorphic robots. Vol. 4, ed.^eds. IEEE, pp. 4194-4199. Shirakawa, T., Adamatzky, A., Gunji, Y., Miyake, Y., 2009. On simultaneous construction of Voronoi diagram and Delaunay triangulation by Physarum polycephalum. International Journal of Bifurcation and Chaos. 19, 3109-3117. Shirakawa, T., Gunji, Y., 2010. Computation of Voronoi diagram and collision-free path using the plasmodium of Physarum polycephalum. International journal of Unconventional Computing. 6, 79-89. Stoy, K., 2006. Using cellular automata and gradients to control self-reconfiguration. Robotics and Autonomous Systems. 54, 135-141. Suh, J.W., Darling, R.B., Bohringer, K.F., Donald, B.R., Baltes, H., Kovacs, G.T., 2000. Fully programmable MEMS ciliary actuator arrays for micromanipulation tasks. Vol. 2, ed.^eds. IEEE, pp. 1101-1108 vol. 2. Takamatsu, A., Fujii, T., Endo, I., 2000. Control of interaction strength in a network of the true slime mold by a microfabricated structure. BioSystems. 55, 33-38. Trimmer, B.A., Takesian, A.E., Sweet, B.M., Rogers, C.B., Hake, D.C., Rogers, D.J., 2006. Caterpillar locomotion: A new model for soft-bodied climbing and burrowing robots. Vol. 1, ed.^eds. Citeseer, pp. 1–10. Tsuda, S., Aono, M., Gunji, Y., 2004. Robust and emergent Physarum logical-computing. BioSystems. 73, 45-55. Tsuda, S., Zauner, K., Gunji, Y., 2007. Robot control with biological cells. BioSystems. 87, 215-223. Ueda, T., Terayama, K., Kurihara, K., Kobatake, Y., 1975. Threshold phenomena in chemoreception and taxis in slime mold Physarum polycephalum. The Journal of General Physiology. 65, 223. Umedachi, T., Kitamura, T., Takeda, K., Nakagaki, T., Kobayashi, R., Ishiguro, A., 2009. A Modular Robot Driven by Protoplasmic Streaming. Distributed Autonomous Robotic Systems 8. 193-202. Vicsek, T.s., CzirÃ³k, A.s., Ben-Jacob, E., Cohen, I., Shochet, O., 1995. Novel Type of Phase Transition in a System of Self-Driven Particles. Physical Review Letters. 75, 1226-1229. Yokoi, H., Nagai, T., Ishida, T., Fujii, M., Iida, T., 2003. Amoeba-like robots in the perspective of control architecture and morphology/materials. Morpho-Functional Machines: The New Species, Springer-Verlag, Tokyo. 99-129. Zhang, L., Abbott, J.J., Dong, L., Kratochvil, B.E., Bell, D., Nelson, B.J., 2009. Artificial bacterial flagella: Fabrication and magnetic control. Applied Physics Letters. 94, 064107. Zhang, S., 2003. Fabrication of novel biomaterials through molecular self-assembly. Nature biotechnology. 21, 1171-1178.
1801.07579	1	1801	2018-01-20T01:27:45	A Work Zone Simulation Model for Travel Time Prediction in a Connected Vehicle Environment	[ "cs.MA", "cs.LG" ]	A work zone bottleneck in a roadway network can cause traffic delays, emissions and safety issues. Accurate measurement and prediction of work zone travel time can help travelers make better routing decisions and therefore mitigate its impact. Historically, data used for travel time analyses comes from fixed loop detectors, which are expensive to install and maintain. With connected vehicle technology, such as Vehicle-to-Infrastructure, portable roadside unit (RSU) can be located in and around a work zone segment to communicate with the vehicles and collect traffic data. A PARAMICS simulation model for a prototypical freeway work zone in a connected vehicle environment was built to test this idea using traffic demand data from NY State Route 104. For the simulation, twelve RSUs were placed along the work zone segment and sixteen variables were extracted from the simulation results to explore travel time estimation and prediction. For the travel time analysis, four types of models were constructed, including linear regression, multivariate adaptive regression splines (MARS), stepwise regression and elastic net. The results show that the modeling approaches under consideration have similar performance in terms of the Root of Mean Square Error (RMSE), which provides an opportunity for model selection based on additional factors including the number and locations of the RSUs according to the significant variables identified in the various models. Among the four approaches, the stepwise regression model only needs variables from two RSUs: one placed sufficiently upstream of the work zone and one at the end of the work zone.	cs.MA	cs	A Work Zone Simulation Model for Travel Time Prediction in a Connected Vehicle Environment Xuejin Wen* Ohio State University * [email protected] Abstract A work zone bottleneck in a roadway network can cause traffic delays, emissions and safety issues. Accurate measurement and prediction of work zone travel time can help travelers make better routing decisions and therefore mitigate its impact. Historically, data used for travel time analyses comes from fixed loop detectors, which are expensive to install and maintain. With connected vehicle technology, such as Vehicle-to-Infrastructure, portable roadside unit (RSU) can be located in and around a work zone segment to communicate with the vehicles and collect traffic data. A PARAMICS simulation model for a prototypical freeway work zone in a connected vehicle environment was built to test this idea using traffic demand data from NY State Route 104. For the simulation, twelve RSUs were placed along the work zone segment and sixteen variables were extracted from the simulation results to explore travel time estimation and prediction. For the travel time analysis, four types of models were constructed, including linear regression, multivariate adaptive regression splines (MARS), stepwise regression and elastic net. The results show that the modeling approaches under consideration have similar performance in terms of the Root of Mean Square Error (RMSE), which provides an opportunity for model selection based on additional factors including the number and locations of the RSUs according to the significant variables identified in the various models. Among the four approaches, the stepwise regression model only needs variables from two RSUs: one placed sufficiently upstream of the work zone and one at the end of the work zone. KEYWORDS: Connected Vehicle, Road Side Unit, Work Zone Travel Time, PARAMICS 1 Introduction The ability to provide accurate traffic state predictions in real-time has long been identified as a critical need for Intelligent Transportation Systems (ITS), especially under abnormal conditions such as traffic accidents (1) and inclement weather (2). Work zone (WZ) is a main issue that may lead to abnormal traffic congestions. It typically requires one or two lanes of the freeway segment closed, resulting in temporary capacity reduction. Based on the report of Federal Highway Authority (FHWA), just in 2014 nearly 24 percent of non-recurring freeway delay, equivalent to about 888 million hours, is typically attributed to work zones (3). Work zone may also lead to safety issues. In 2015 New York State DOT (NYSDOT) reported 646 crashes in capital construction work zones on New York roads and bridges, resulting in six motorist fatalities and 147 injuries to motorists, contractor employees and NYSDOT staffs (4). A few studies have focused on providing current or predicted travel times through WZs in the freeway network. For example, the variable message signs (VMS) can display the travel time and distance to the end of the work zone to motorists (5). Lee and Kim (2006) designed and implemented an Automated Work Zone Information System (AWIS) to provide road users with real-time delay information during peak hours so that the traffic can be diverted to detour routes (6). Ghosh-Dastidar and Adeli (2006) built a neural network model using simulated data to predict the delay and queue length at freeway work zones (7). Bae et al. (2017) also applied a neural network model to predict the traffic volume and percentages of trucks at the upstream of the work zone, then they applied a BPR function to calculate the work zone travel time based on the ratio of volume and capacity (8). The estimation and prediction of travel times requires measurement of traffic states. Traditionally the traffic detectors are installed on freeways to record traffic characteristics such as flow, speed and occupancy, which are used as the data source for travel time prediction (9). For example, California Freeway Performance Measurement System (PeMS) has installed 43,426 detectors on freeways with a total distance of 30,599.4 miles by 2017. However, based on the latest statistics, 37.1% of PeMS detectors have been out of function (11). The installation and maintenance costs of these huge amounts of fixed detectors are therefore very high. On the other hand, if the deployment is not dense, there may not be enough detectors around the work zone segment to collect the necessary information. Connected vehicle is an emerging technology recently. It allows each vehicle to be fully connected to the other vehicles and to the infrastructures like traffic light and Variable Message Sign (VMS). The U.S. Department of Transportation has made the development of vehicle-to-vehicle (V2V) and vehicle-to-infrastructure (V2I) communication a major research priority to improve the safety and efficiency of the transportation system. Connected vehicles can store their vehicle activities such as speed, position, turn signal, brake status and so on with a frequency in On-board Units (OBUs). At the same time, there are roadside units (RSUs) installed at intersection, interchanges and other locations to provide communication interface to the vehicles. When a connected vehicle passes by the coverage area of a RSU, the vehicle activities data in OBU is transmitted to RSU and the OBU buffer will be cleared (9). In this study, we propose to utilize portable roadside unit (RSU) for freeway work zone travel time prediction in a V2I environment. Comparing with the fixed detectors, the locations and numbers of the reusable and portable RSUs can be adjusted based on work zone types, traffic demand and the connected vehicle penetration rates. Simulation is an important approach for this kind of experiments with advantages like safe, convenient and almost no cost (10). Here a real-world work zone simulation model in connected vehicle environment is built using PARAMICS®, microsimulation software by QuadstoneParamics. The work zone travel time prediction models are also built to identify candidate locations of RSUs and verify the connected vehicle techniques can improve the travel time prediction accuracy. The rest of the paper is organized as follows. The next section will introduce the PARAMICS work zone simulation model for NY State Route 104. Following that, the simulation results are processed and relevant variables are extracted. The work zone travel time prediction models are then built based on these variables. Conclusions and recommendations are given at last. Work Zone Simulation Model The configuration of the work zone simulation model is observed from NY State Route 104. As shown in Figure 1, it is a piece of two-lane freeway with no ramps and the traffic are from left to the right. The length of the 2 one-lane-closed segment is 2 miles. A warning arrow panel and a VMS sign are located at 0.1 miles and 0.7 miles from the work zone beginning separately. Start of Queue Figure 1. Layout of WZ from NY-104 After that, the simulation model was built using PARAMICS (12). Twelve portable RSUs were introduced in this model (Figure 1). RSU 1 is at the end of the work zone, the rest RSUs 2-12 are distributed every 0.2 miles from the beginning of work zone to 2 miles at the upstream of work zone. Including a dense set of RSUs in the simulation enables us to explore the impact of RSU location on travel time prediction. As will be discussed, through the development of work zone travel time prediction models, the significant variables as well as the RSUs that provide those information can be identified, and therefore the number and locations of the RSUs also become obvious. The hourly traffic volume data in 2014 for this freeway segment are acquired from NY Department of Transportation (13). As shown in Figure 2, for each month, we took the average by hour and got the typical traffic volumes. Then for the 24-hour traffic volumes of each month, we conducted 10 runs of simulations with different random seeds so that stochastic behaviour is considered. In this study, the market penetration rate of connected vehicles was set to 100%. The PARAMICS API was implemented in C++ to simulate the V2I applications and collect the data. 3000 2500 2000 1500 1000 500 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 Jan July Feb Aug March Sep April Oct May Nov June Dec Figure 2. Hourly Traffic Volumes Averaged by Month 3 Simulation Data Processing In this connected vehicle simulation model, the RSU has the function to record the ID and timestamp when a vehicle passes by it. The vehicle will also transmit its previous trajectory data to the RSU. Finally, we extract the following variables from the simulation data. Note that all these variables are aggregated by five minutes. a) 𝑇𝑇𝑖,1, 𝑖 = 2,3, … , 12. Based on the vehicle ID and time stamp, we calculate the travel time from each of the upstream RSU to the one at the end of the work zone. For example, "𝑇𝑇2,1" is the travel time from RSU 2 to RSU 1. b) UpstreamFlow. The traffic volume at the upstream RSU 12. c) DownstreamFlow. The traffic volume at the downstream RSU 1. d) WorkZoneEndAcc. The acceleration rate at the end of work zone, which is based on the vehicle's trajectory collected by RSU 1. e) Qlength. The length of the queue at the upstream of the work zone. In this study, a queue is defined to satisfy three criteria: the speed is less than 10 miles/hour, the inter-vehicle distance is less than 0.01 miles and the number of vehicles is at least 5. The queue information can be extracted from the trajectory data stored in RSU 1. f) StartQueue. The location of the start of the queue as shown in Figure 1. g) 𝑇𝑇𝑤𝑧. The work zone travel time. Note that for most of the previous studies, they measured the work zone travel time from a fixed location at the upstream to the end of the work zone (5) (7) (8). Although this simplifies the problem, however, it is not accurate because the length and location of the upstream queues are changing. The approach we measure 𝑇𝑇𝑤𝑧 is as follows. If there was a queue, it is the time from the leftmost vehicle in the queue to the end of work zone. If there was no queue, it is the travel time through the exact work zone. With the five-minute aggregation, for 24-hour traffic volumes of each month, we can get 2,880 observations of variables a) - g) by running the simulation model 10 times. In total, 28,800 records from the first 10 months (January to October) are used for the training of the travel time prediction model, and the rest 5,760 records from November to December are used for the testing of the model performance. Work Zone Travel Time Prediction and RSU Location Let 𝑇𝑇𝑤𝑧(𝑡) denote the travel time through a work zone at time step t and Lahead be the "look ahead" time. In this study, Lahead is set at 5 time steps, which corresponds to predicting travel time through the work zone 25 minutes into the future (data is aggregated every 5 minutes, so 5 time steps equates to 25 minutes). To predict the future travel time, we use an autoregressive with exogenous input model structure as follows: 𝑇𝑇𝑤𝑧(𝑡 + 𝐿𝑎ℎ𝑒𝑎𝑑) = 𝐴(𝑞)𝑇𝑇𝑤𝑧(𝑡) + ∑ 𝐵𝑖(𝑞)𝑢𝑖(𝑡) ; 𝑛𝑢 𝑖=1 (1) where 𝐴(𝑞) = ∑ 𝑛𝑎 𝑗=1 𝑎𝑗𝑞−𝑗 , 𝐵𝑖(𝑞) = ∑ 𝑛𝑏𝑖 𝑗=1 𝑏𝑗𝑞−𝑗 ; 𝑢𝑖 refers to one of 16 variables (11 travel time variables 𝑻𝑻𝒊,𝟏, 𝑖 = 2,3, … , 12 , UpstreamFlow, DownstreamFlow; WorkZoneEndAcc, Qlength and StartQueue), 𝑛𝑢 = 16; 𝐴(𝑞) and 𝐵𝑖(𝑞) are the backshift operators, which operates on an element of a time series to produce the previous element; 𝑛𝑎 and 𝑛𝑏𝑖 are the backshift step lengths; 4 The coefficients 𝑎𝑗 and 𝑏𝑗 are derived using the algorithm that predicts travel times in the work zone. In this study, we applied the following four regression techniques: simple linear regression model; multivariate adaptive regression splines (MARS), with up to 2nd order nonlinear interactions in the input variables (14); a stepwise model with risk inflation criterion (15); and an elastic net model (16). All the experiments are conducted using statistical software R. Table 1 summarizes the significant variables selected by the four models. The root mean squared error (RMSE) between the real and predicted work zone travel times is also calculated for each model. Table 1. Significant Variables and RMSEs of the Four Models (* means that variable is significant for that model) Models Linear Regression MARS Stepwise Elastic Net 16 Variables 𝑇𝑇2,1 𝑇𝑇3,1 𝑇𝑇4,1 𝑇𝑇5,1 𝑇𝑇6,1 𝑇𝑇7,1 𝑇𝑇8,1 𝑇𝑇9,1 𝑇𝑇10,1 𝑇𝑇11,1 𝑇𝑇12,1 UpstreamFlow DownstreamFlow WorkZoneEndAcc Qlength StartQueue * * * * * * * * * * * * * * * * * * * * * * * RMSE (seconds) 31.46 28.37 34.64 30.82 As can be seen in Table 1, first for the variable selection, the linear regression model selects four significant variables, which are collected by three RSUs 1, 10 and 12; the MARS model identifies six significant variables from RSU 1, 8, 10, 12; the Elastic Net model is the most complicate one with 11 significant variables from all the other RSUs except 2 and 3; and the Stepwise model is the simplest one with only two significant variables from RSU 1 and 12. One interesting observation is that the variables from the RSU 10 and RSU 12 are the most significant ones, e.g. 𝑇𝑇10,1 is selected three times and 𝑇𝑇12,1 are included by all the four models. This tells that the location and length of the queues caused by the work zone are probably around RSU 10 and RSU 12. Second, for the travel time prediction performance, MARS model performs the best with a RMSE of 28.37 seconds. It's a little surprising that the Elastic Net model has a little worse RMSE of 30.82 seconds with the requirement of so many variables to build the model. The linear regression model ranks the third, its RMSE is 31.46 seconds. Finally the RMSE of the simplest Stepwise model with two variables is around 34.64 seconds. Third, we can also make the portable RSU location plans based on Table 1. The performance of Elastic Net model is not the best and it needs 10 RSUs, which is unrealistic. Therefore this plan can be discarded. For the other three models, there is a trade-off between the algorithm accuracy and the RSU resources. The MARS model has the smallest RMSE but it needs variables from 4 RSUs. Its performance is followed by the linear regression model which requires 3 RSUs. The RMSE of the stepwise model is only about 6 and 3 seconds higher than MARS and linear regression model, respectively. And it just requires two RSUs: RSU 12 and RSU 1, which, from a practical viewpoint might represent a reasonable trade-off between algorithm performance and sensing requirements. We are interested to further check the performance of the Stepwise Model. Figure 3 shows the real work 5 zone travel time (red) and the predicted one (blue) for the testing set using this Stepwise model. Figure 4 shows the histogram of the error between the real work zone travel time and the predictions. Figure 3. Real Travel Times and Stepwise Model Predictions for the Testing Dataset Figure 4. Histogram of Travel Time Difference between Real Values and Predictions Figure 3 shows that in general the Stepwise model can capture the trend of work zone travel times very well. However, it looks like that the predictions from Stepwise model (blue line) are lower than the real work zone travel times (red line) when the ground truth values are high. Figure 4 shows the distribution of the difference between the real travel times and the predictions. We can see that the frequency with the absolute difference less than 10 seconds is very high. Most of the time, the difference is within [-20, 20] interval. This once again shows the Stepwise model has a reasonable performance. Conclusion This paper focuses on travel time prediction for freeway work zones in a connected vehicle environment. A work zone segment motivated by NY State 104 was simulated using PARAMICS. The simulation model considered 6 a Vehicle-to-Infrastructure setting by using portable RSUs to communicate with connected vehicles and collect traffic data. Compared with loop detectors, the portable RSUs can save installation and maintenance cost, and their number and locations can be adjusted. To determine candidate RSU locations, twelve RSUs were put along the work zone segment in the simulation model and sixteen traffic variables were extracted from the V2I data. For travel time prediction and variable selection, four different modelling approaches were investigated, including linear regression, MARS, stepwise and elastic net. The results showed that the models have similar RSME performance, ranging from 28 seconds to 35 seconds. Through the comparison of the identified significant variables, a suggested number and locations of the portable RSUs were also determined. We found that the stepwise model provides a practical sensing plan in that only two RSUs are required: one sufficiently upstream from the work zone and one at the end of work zone. For our future study, more experiments are necessary. For example, we can explore more scenarios with different connected vehicle penetration rates and work zone types. We can also adjust the RSU location plans dynamically based on various traffic demands by peak-hour and non-peak-hour, weekday and weekend, inclement and clear weather and so on. References 1. Lin, L., Wang, Q., & Sadek, A. W. (2016). A combined M5P tree and hazard-based duration model for predicting urban freeway traffic accident durations. Accident Analysis & Prevention, 91, 114-126. 2. Lin, L., Ni, M., He, Q., Gao, J., & Sadek, A. W. (2015). Modeling the impacts of inclement weather on freeway traffic speed: exploratory study with social media data. Transportation Research Record: Journal of the Transportation Research Board, (2482), 82-89. 3. Federal Highway Administration. Facts and Statistics – Work Zone Mobility, http://www.ops.fhwa.dot.gov/wz/resources/facts_stats/mobility.htm, accessed on Jan 27, 2017. 4. New York State DOT. Work Zone Awareness. https://www.dot.ny.gov/news/press-releases/2016/2016-04-12, accessed on Jan 27, 2017. 5. Zwahlen, H., & Russ, A. (2002). Evaluation of the accuracy of a real-time travel time prediction system in a freeway construction work zone. Transportation Research Record: Journal of the Transportation Research Board, (1803), 87-93. 6. Lee, E. B., & Kim, C. (2006). Automated work zone information system on urban freeway rehabilitation: California implementation. Transportation Research Record: Journal of the Transportation Research Board, (1948), 77-85. 7. Ghosh-Dastidar, S., & Adeli, H. (2006). Neural network-wavelet microsimulation model for delay and queue length estimation at freeway work zones. Journal of Transportation Engineering, 132(4), 331-341. 8. Bae, J., Choi, K., and Oh, J.H. (2017). Multi-Contextual Machine-Learning Approach to Modeling Traffic Impact of Urban Highway Work Zones. Transportation Research Record: Journal of the Transportation Research Board, Transportation Research Board of the National Academies, accepted for publication by the Artificial Intelligence Committee of TRB, in press. 9. Kianfar, J., & Edara, P. (2013). Placement of roadside equipment in connected vehicle environment for travel time estimation. Transportation Research Record: Journal of the Transportation Research Board, (2381), 20- 27. 10. Lin, L., Wang, Q., & Sadek, A. W. (2014). Border crossing delay prediction using transient multi-server queueing models. Transportation Research Part A: Policy and Practice, 64, 65-91. 11. PeMs - Caltrans Performance Measurement. http://pems.dot.ca.gov/?dnode=State&content=detector_health&tab=dh_summary, accessed on Jan 27, 2017. 12. Quadstone PARAMICS. http://www.paramics-online.com/, accessed on Feb 3, 2016. 13. New York Department Transportation. of https://www.dot.ny.gov/divisions/engineering/technical- services/highway-data-services/hdsb, accessed on June 01, 2016. 14. Friedman, J. H. (1991). Multivariate Adaptive Regression Splines. The Annals of Statistics. 19: 1 15. Efroymson,M. A. (1960). Multiple regression analysis, Mathematical Methods for Digital Computers, Ralston A. and Wilf,H. S., (eds.), Wiley, New York. 16. Zou, Hui; Hastie, Trevor (2005). Regularization and Variable Selection via the Elastic Net. Journal of the Royal Statistical Society, Series B: 301–320. 7
1806.02578	1	1806	2018-06-07T09:27:57	Investigating Spatiotemporal Dynamics and Synchrony of Influenza Epidemics in Australia: An Agent-Based Modelling Approach	[ "cs.MA", "physics.soc-ph", "q-bio.PE" ]	In this paper we present ACEMod, an agent-based modelling framework for studying influenza epidemics in Australia. The simulator is designed to analyse the spatiotemporal spread of contagion and influenza spatial synchrony across the nation. The individual-based epidemiological model accounts for mobility (worker and student commuting) patterns and human interactions derived from the 2006 Australian census and other national data sources. The high-precision simulation comprises 19.8 million stochastically generated software agents and traces the dynamics of influenza viral infection and transmission at several scales. Using this approach, we are able to synthesise epidemics in Australia with varying outbreak locations and severity. For each scenario, we investigate the spatiotemporal profiles of these epidemics, both qualitatively and quantitatively, via incidence curves, prevalence choropleths, and epidemic synchrony. This analysis exemplifies the nature of influenza pandemics within Australia and facilitates future planning of effective intervention, mitigation and crisis management strategies.	cs.MA	cs	Investigating Spatiotemporal Dynamics and Synchrony of Influenza Epidemics in Australia: An Agent-Based Modelling Approach Oliver M. Cliff1, Nathan Harding1, Mahendra Piraveenan1, E. Yagmur Erten1,2, Manoj Gambhir3, and Mikhail Prokopenko1,4 1 Centre for Complex Systems, Faculty of Engineering and IT, University of Sydney, Sydney, NSW 2006, Australia 2 Department of Evolutionary Biology and Environmental Studies, University of Zurich, Winterthurerstrasse 190, 8057 4 Marie Bashir Institute for Infectious Diseases and Biosecurity, University of Sydney, Westmead, NSW 2145, Australia Zurich, Switzerland 3 IBM Research, Melbourne, Australia Abstract In this paper we present AceMod, an agent-based modelling framework for studying influenza epidemics in Australia. The simulator is designed to analyse the spatiotemporal spread of contagion and influenza spatial synchrony across the nation. The individual-based epidemiological model accounts for mobility (worker and student commuting) patterns and human interactions derived from the 2006 Australian census and other national data sources. The high-precision simulation comprises 19.8 million stochastically generated software agents and traces the dynamics of influenza viral infection and transmission at several scales. Using this approach, we are able to synthesise epidemics in Australia with varying outbreak locations and severity. For each scenario, we investigate the spatiotemporal profiles of these epidemics, both qualitatively and quantitatively, via incidence curves, prevalence choropleths, and epidemic synchrony. This analysis exemplifies the nature of influenza pandemics within Australia and facilitates future planning of effective intervention, mitigation and crisis management strategies. Keywords: computational epidemiology, agent-based modelling, discrete-time simulation, influenza, epidemics, synchrony, demographics 1. Introduction Early recognition and curtailing of outbreaks of infectious diseases is crucial to policy making and pub- lic health. Without efficient intervention, international travel can result in the spread of highly infectious pathogens around the globe within weeks of the initiation of the outbreak, as evidenced by the 2009 H1N1 (swine flu) pandemic. Moreover, infectious diseases are large burdens on the economy. For instance, a 2007 study estimated that the annual cost of seasonal influenza to the health system in Australia was AU$828-884 million [1]. Effective strategies for mitigating outbreaks should include a suitable combination of immunisation, vaccination, and palliative care [2, 3, 4] and planning these strategies requires reliable fore- casting through simulation of various "what if?" scenarios. Without efficient intervention, highly infectious pathogens, such as swine flu, result in international travel carrying the virus around the globe within weeks to months of the initiation of the outbreak, causing a worldwide public health emergency [5, 6, 7, 8]. Intergration of large-scale datasets and explicit simulations of entire populations down to the scale of single individuals has considerably improved the accuracy of epidemiological models [9]. Furthermore, an effective intervention requires not only an accurate modelling and surveillance of the disease dynamics, but also implementation of the underlying contact networks and their topologies. This requires a detailed theoretical and practical understanding of the interplay between epidemic processes, mobility patterns (es- pecially over long distances), and population data. By running multiple computer simulations while varying the sources of infection, we can then ascertain the average social and health impact, as well as "zoom in" on specific pathways and patterns of epidemics. Preprint submitted to Simulation Modelling Practice and Theory June 8, 2018 While there are currently many useful models and tools to assist in the analysis and prediction of con- tagion processes, none are adequate in terms of precise and integrated tracking, predicting and mitigating epidemics at the national-level for Australia. It remains a challenge to generate and simulate a realistic and dynamic contact network, accounting for both mobility and human interactions representing the de- mographics of Australia through an individual-based epidemiological model. Moreover, as evidenced by Cauchemez et al. [10], modeling the transmission of respiratory diseases must specifically account for struc- turing of schools, grades, and classes due to the "back-and-forth waves of transmission between the school, the community, and the household." These challenges are particularly important for investigating epidemics in Australia, where the population is concentrated mainly around urban areas. We have addressed this need by developing AceMod, the Australian Census-based Epidemic Model that employs a discrete-time and stochastic agent-based model to investigate complex outbreak scenarios at var- ious spatiotemporal levels. The simulator comprises 19.8 million software agents where each agent contains a set of attributes of an anonymous individual. The agents are generated such that the distributions at multiple scales concur with key demographic statistics of the 2006 Australian census data. The next layer of the model includes mobility patterns (with respect to work, study and other activities) in order to charac- terise potential interactions between spatially distributed agents. The final layer describes local transmission dynamics by combining agent health characteristics, such as susceptibility and immunity to diseases, with a natural history model for influenza. The simulation runs in 12 hour cycles ("day" and "night") over the course of an epidemic such that agents interact in different social mixing groups depending on the cycle. Given a surrogate population generated in this way, we run multiple instances of each scenario, varying the disease infectiousness and outbreak locations to give insight into the spread of influenza pandemics around Australia. The novel aspects of AceMod include (i) the spatial fidelity of the stochastically generated population, calibrated to Australian Census and the Australian Bureau of Statistics (ABS) datasets that utilise multi- scale distributions over school data and hierarchical mixing groups; (ii) its refined models for the transmission and natural history of the simulated infection, based on the latest available epidemiological studies; and (iii) the focus on measures of spatiotemporal complexity of influenza epidemics. We evaluate the spread of influenza in space and time through Australia via incidence curves, prevalence choropleths, and epidemic synchrony. Incidence curves characterise the disease by plotting the number of newly ill individuals (i.e., the incidence) at each time step. Here, ill refers to infected agents that are showing symptoms. This illustrates the severity of the disease over time and can be used to compare different pathogens or mitigation strategies. Then, in order to qualitatively study the evolving spatial distribution of the epidemic over time, we record the percentage of ill individuals (i.e., the prevalence) in each community at a given time step and give snapshots of this distribution through choropleths taken at key times of the epidemic. Finally, we examine spatial hierarchies in disease prevalence by analysing the variance in the timing of the epidemic peaks in each community (i.e., the synchrony). The spatiotemporal synchrony of disease spread between communities correlates to the size of the communities, suggesting the disease transmission is more associated with social connectivity rather than geographic distance. 2. Related work Well-established stochastic models of epidemics often use SIR and SEIR differential equations for the population dynamics of susceptible (S), exposed (E), infectious (I), and removed/recovered (R) individu- als [11]. While these models are suitable for analysing the general behaviour of an epidemic on larger scales, with a focus on global variables, they do not allow us to make accurate predictions at a finer resolution. In many scenarios it is beneficial to trace the dynamics in a more fine-grained way especially during the initial or final stages of an outbreak, when person-to-person transmission processes dominate [12]. As a result, stochastic agent-based discrete-time simulation models were developed to capture how the uncertainty in disease diffusion varies in different social groups and affects the overall analysis and predictions of epidemics [12, 5]. These models have been used to assess vaccination and antiviral prophylaxis strategies on a local level, highlighting importance of detailed modelling of contact patterns [13, 14, 12]. These models 2 further allow us to investigate strategies at large-scale regional and national levels for containing an emerging pandemic influenza strain at its source [5, 6]. Early efforts include an agent-based simulation by Elveback et al., where they modeled a small artificial population to represent a community and defined multiple contexts in which individuals interact, in order to explore the disease spread. Later works [5, 12] leveraged increased computational power to further this methodology, considering populations on a larger scale based on exact census data. A suite of scalable agent- based modelling software simulators was described in a series of papers[15, 16, 17, 18, 19]. Simdemics was designed as a general purpose simulation environment supporting 300 million agents, aimed at modelling pandemic planning and response to a range of crisis scenarios, such as H1N1/H5N1 pandemics and bio- terrorism events [18]. A number of simulators were built upon Simdemics framework. For example, a high- performance computing modelling environment, EpiSims, designed as a distributed discrete event simulator, modelled disease transmission via sub-groups of people in a given location. The location objects were distributed across CPUs and people objects moved among locations, using various census, land-use and population-mobility data. It has been successfully applied to modelling of smallpox [15] and pandemic influenza plan [20]. The EpiSimdemics algorithm [16, 19] was based on contagion diffusion across a person- location graph, a bipartite graph where the nodes represented people and locations, while the edges indicated a person's presence at a location. A person-location graph of the entire population of the United States included 300 million nodes (people) and 1.5 billion edges, and ran on the 352,000 core NCSA BlueWaters system, so that 120 days of an epidemic were simulated in 12 seconds. While EpiSimdemics built the graph implicitly, the next extension, Epifast [17], used an explicitly given network to simulate SEIR compartmental model in a distributed memory system with a master-slave computation, and run 10 times faster than EpiSimdemics for equivalent graphs. In order to inform policy at a national and state level, it is imperative for an epidemic model to have a high degree of spatial fidelity, which is the primary motivation and focus of our work. Due to the high concentration of the Australian population around urban areas, modelling Australian demographics provides a unique challenge. For instance, GLEAM [21] is a global-scale simulation based on a gridded population of the world. In order to capture global disease dynamics, GLEAM discretises the world into a grid of 25×25 km cells. While this method effectively captures global disease dynamics, it does not provide a resolution fine enough for modelling disease transmission at the individual level in Australia. A more recent simulator, FluTE [22], facilitates modelling epidemic scenarios in the United States but is not immediately suitable to address the Australian population, mobility, and school structure. Agent-based simulators rely on detailed models of the transmission and the natural history of the pathogen through a host individual [5]. Recent work by Cauchemez et. al. [10] emphasizes the importance of high precision disease parameters when high fidelity census data is being used. The authors recognised "a lack of detailed data on transmission outside the household setting" and quantified the parameters of disease spread within the school and home environment by observing an outbreak of H1N1(2009) in an elementary school in Pennsylvania. Longini et al. [5] provided a detailed description of the disease natural history for influenza, specifying the progression of a disease from pathological onset to its eventual resolution within a host. There is significant research into estimating the proportion of individuals who become symptomatic (and hence more infectious) [23, 24, 25, 26], informing our study. 3. Simulation model Our simulation model comprises two basic elements: (i) a surrogate Australian population based on real- world data for demographics and worker flow; and (ii) a stochastic agent-based model for disease spread. The demographics induce numerous statistical areas (social mixing groups) within which individuals associate and provide context for susceptible agents to contract an infectious disease. The model runs in cycles of two 12-hour periods ("day" and "night") and an agent is associated with a set of daytime and nighttime groups depending on the cycle. Given the census data, we are able to generate mobility networks via the reported worker flow between these mutually exclusive sets. The framework of developing contact networks from census data and integrating a transmission model is extensively used in literature to stochastically model influenza pandemics [14, 5, 12, 22, 21]. 3 Figure 1: Spatial distribution of the population of 19.8 million Australian residents according to 2006 Australian census data. The states and territories of Australia are (in order of population size): New South Wales (NSW), Victoria (VIC), Queensland (QLD), Western Australia (WA), South Australia (SA), Tasmania (TAS), Australian Capital Territory (ACT), Northern Territory (NT). Insets: distributions for SLAs based in and around Sydney (red) and Brisbane (green) on the same scale. 4 3.1. Demographics and mobility We utilise the 2006 Australian census data to distribute the model population of 19.8 million individuals into 1422 statistical local areas (SLAs). Figure 1 shows the spatial distribution of the Australian population among the SLAs. These areas are further subdivided into census collection districts (CDs) of approximately 225 dwellings each in metropolitan settings. The census provides population (e.g., age, sex, employment status) and housing (e.g., household size and composition) statistics at this geographic level. This data is then used to stochastically generate all households and individuals within each CD. Each individual in the population is a member of several mixing groups derived from census data and prior work [5, 12, 22]. The contact (and transmission) probability of any disease is, in general, decreasing with the group size, and numerous empirical studies report on these statistics for different contexts [5, 12, 10, 22]. During the daytime, the mixing groups comprise work groups for adults, and classrooms, grades, and schools for students. During the nighttime, individuals interact in households, household clusters, communities, and neighbourhoods. We first stochastically generate the population based on housing statistics and then subsequently assign social mixing groups by geographic locality. Specifically, the census contains the frequency distribution of household sizes in each CD. In addition, households are composed of either lone, group, or family types, and the distribution of these compositions are known for each CD. We further subdivide the family type into the following categories: single (S), couples with children (CWC), couples without children (CWOC), or single parent families (SPF). This yields a conditional distribution for family compositions, given household size, in each CD. We then generate agents for the adults and children living in every residence by sampling a family type conditioned on the household size. Households have the highest pairwise contact probability and four consecutive houses are assigned to a larger social group denoted a cluster. The communities (CDs) and neighbourhoods (SLAs) within which these residents live then provide context for unspecified and occasional person-to-person contact (e.g., supermarkets, theatres, etc.). In this study, the population is stochastically generated once, and used in all simulations. It is possible to run simulations over multiple stochastically generated populations; however, this did not produce significantly different results for the key variables investigated in this work. In the daytime, most working-age adults interact with other individuals in their immediate workplace, whereas students associate through local schools, grades, and classes. To establish working communities, the Australian census data subdivides SLAs into destination zones (DZNs). Although these geographic areas do not concord with CDs, the worker flow from CDs to DZNs can be used to accurately capture the short- to medium-distance population mobility important for disease spread [5]. Adults from each CD are randomly assigned a DZN based on this data, and the workers in each DZN are split into working groups of around 20 adults to represent the colleagues with whom an individual is likely to work with on a daily basis. The commute distance distribution of workers given by this approach is shown in the appendix in Fig. 8(a) and is distributed similarly to reports by the Australian Bureau of Infrastructure, Transport and Regional Economics (BITRE) [27]. Finally, accurate contact networks in schools, grades, and classes are paramount to the demographics given the close and frequent contact of students with one another [10]. Unfortunately, the census does not contain information about school attendance, nor spatial distribution of students. However, the Australian Bureau of Statistics (ABS) provides state-wide frequency distribution tables specifying the number of schools with (non-uniform) intervals of the school attendance. Due to the concentration of the Australian population around urban areas, we employ a biased sampling technique to place these schools within DZNs that have more students nearby, i.e., SLAs that are more populated are biased to have more schools (see Appendix B.2.1). Using this approach, the average commute distance of students is illustrated in Fig. 8(b) and is distributed according to reports by the BITRE [27]. 3.2. Transmission model At an individual level, we are concerned with the transition probability of a susceptible agent i becoming infected during a given time period (step) n. Let the random variable Xi(n) denote the state of the individual 5 (a) Natural history of the disease. (b) Simulated generation time. Figure 2: Natural history of the disease and corresponding simulated generation time. The disease dynamics are modelled as having a linear increase followed by a linear decrease, as illustrated in 2(a). In the figure, the area under the curve is shaded according to the proportion of people at least that infectious after disease onset (darker representing a higher proportion). If an agent becomes symptomatic, their infectiousness doubles (dashed blue line) from that day onward. Moreover, 67% of agents become symptomatic; of these agents, 30% start showing symptoms on day 1, 50% on day 2, and the remaining 20% on day 3. We obtain empirical generation times from simulations resulting from this model, shown in 2(b) for a number of R0 values. The confidence intervals range from 3.35 to 3.39 days (also shown on Fig. 2(a)), depending on R0 and, in general, the generation time has a slight downward trend as a function of disease severity. at time n. That is, we are interested in the infection transition probability: pi(n) = Pr{Xi(n) = latent Xi(n − 1) = susceptible} , (1) where latent is the state denoting exposure to the disease to that agent (see Sec. 3.3). Section 3.1 defined a number of mixing groups Gi(n) that each agent i interacts within where each group g ∈ Gi(n) is associated with a static set of agents Ag. The infection probability is thus computed as a product of all of the possible infectious contacts during that step: (2) pi(n) = 1 − (cid:89) g∈Gi(n) j∈Ag\i (cid:34) (cid:89) (cid:0)1 − pg j→i(n)(cid:1)(cid:35) , where pg g. Let nj denote the time agent j becomes infected, then j→i(n) denotes the instantaneous probability of transmission from agent j to agent i in contact group j→i(n) = κ f (n − nj j, i) qg pg j→i. (3) The scaling factor κ is a free parameter that allows variation in the contagiousness of the simulated epidemic, while keeping other parameters constant. Changing this parameter thus results in different values of the reproductive ratio R0; we investigate the precise dependence between these two variables in Sec. 4.1. The function f : N → [0, 1] characterises infectivity of case j over time; importantly, f (n − nj j, i) = 0 when n < nj. In this paper we assume infectivity linearly decreases over time. Finally, qg j→i is the probability of transmission from agent j to i at peak infectivity, which we discuss below. Each susceptible individual can become infected at the end of each step, which is determined by Bernoulli trial with transition probability pi(n). Note that the probabilities of transmission qg j→i are very low in practice, and so the instantaneous probabilities of transmission pg In this work, we derive qg j→i(n), scaled by κ, remain significantly below 1. j→i from two sources: (i) the recent study of H1N1 (2009) by Cauchemez et al. [10], in setting transmission probabilities for a number of relevant mixing groups (households and schools), as well as (ii) a recent study using contact probabilities for mixing groups [22]. The latter approach included mixing groups that were not considered by [10] (household clusters, neighbourhoods and working groups). Furthermore, as pointed out by Chao et al. [22], these contact probabilities are in agreement with other 6 00.511.52R03.343.353.363.373.383.39Generationtimeagent-based models [20, 28]. Thus, the transmission between two agents qg is described by either a transmission probability βg j→i or a contact probability cg j→i. j→i within each contact group g The transmission probabilities contained in Table C.4 are obtained from a study of H1N1 (2009) by Cauchemez et al. [10] which determined the transition rates of infection βg j→i between an individual j and a susceptible individual i, using reverse jump Monte Carlo sampling methods, while fitting of the transition rates varied slightly depending on the mixing context j. We transform the rate of infection βg j→i into the daily probability of transmission from an infected individual j to a susceptible individual i as follows [10] j→i = 1 − exp(cid:0)βg qg j→i (cid:1) . (4) In addition, when the transition rates are not available in the study Cauchemez et al. [10] we utilise the contact probabilities given in Table C.3, reported by [22]. These probabilities represent the likelihood, within each daily period, of having a contact of sufficient duration and closeness for a transmission of the infection to be possible between the two individuals i and j in this social setting. Each contact probability cg j→i for a mixing group g is multiplied by a factor ρ scaling to the probability of transmission: qg j→i = ρ cg j→i. (5) The factor ρ allows us to calibrate the transmission probabilities set by Eq. (5) within mixing groups. This is achieved by matching the attack rates of the corresponding mixing groups to the values of the context- dependent rates reported in [10]. This ensures consistency between Eq. (4) and Eq. (5). 3.3. Natural history of disease The natural history model describes the course of a disease from onset to resolution [5]. Each state has associated properties (e.g., infectious or non-infectious) depending on the disease being modelled. In this paper, we define a few states for influenza: susceptible, latent, infectious symptomatic, infec- tious asymptomatic, and recovered. Similar models have been established in earlier work based on observations of influenza dynamics [5]. The model, illustrated in Fig. 2(a), includes states that account for nuances in influenza that make the pathogen highly transmissible. The latent period captures the time from the exposure to the acquisition of infectiousness, contributing to the spread of the disease. From the latent state, 67% of individuals develop symptoms (i.e., transition to symptomatic) [23], leaving the remainder as asymptomatic agents that are half as infectious [5]. More recent studies of the proportion of symptomatic cases suggest a wider range of estimates – within 25–67% [24, 26] – but, in general, this depends on the study design and definitions of infection and symptomatic illness (with estimates for asymptomatic cases ranging between 0–100%) [25]. Hence, we adopted the established estimates of [23] that have been used in various simulations [12, 29, 22]. An individual that will eventually show signs of illness will typically do so a number of days after contracting the disease. During the interim, these agents are considered as infectious as agents in the asymptomatic state. Specifically, out of the individuals that will become ill, 30% will start showing symptoms one day after disease onset, 50% after two days, and the remaining 20% on day three. These dynamics result in the generation times (i.e., the times between the infection of a primary case and one of its secondary cases) between 3.35 and 3.39 days, depending on the R0, as illustrated in Fig. 2(b). These results are in line with the literature on generation times in a number of empirical studies [6] and established simulations [5, 12, 22]. In this framework, the prevalence at a given time is the number of ill individuals in the symptomatic state. This is then used to compute the illness attack rate, R0, and many other quantities of interest. 3.4. Scenario description In order to sufficiently model the domestic spread of the disease, we assume that the Australian popu- lation is exposed to the strain once it is a global pandemic. The geographic isolation of Australia suggests that people arriving from all other countries are equally likely to be already carrying the disease. Hence, we follow the approach of [12] by introducing the disease at international airports, where the number infected is proportional to the air traffic inflow, obtained from BITRE [30]. 7 Airport code SYD MEL BNE PER OOL ADL CNS DRW TSV State NSW VIC QLD WA QLD SA QLD NT QLD City Sydney Melbourne Brisbane Perth Gold Coast Adelaide Cairns Darwin Townsville Passengers 40884 25859 14250 11449 3022 2214 1874 597 105 Figure 3: Daily incoming passengers per Australian international airport obtained from BITRE [30] along with a map detailing the airport locations. The AceMod simulation itself is dynamically seeded such that new individuals are randomly infected at every time step n instead of only at the beginning. There are 10 international airports in Australia and the number of "seeds" surrounding each airport is proportional to the average daily number of incoming passengers at that airport (see Tab. 3). Note that the simulation is a closed system such that only existing and susceptible agents can be infected. Moreover, only individuals living in SLAs with geographic centroids inside a given (seeding) radius of each airport are considered for infection. The number of agents infected within this region are binomially distributed B(P, N ) with P and N chosen such that, on average, 0.04% of incoming arrivals infect exactly one agent (in agreement with the previous studies of [12]). The epidemic characteristics are relatively insensitive to variations in the initial seeding, as has been shown before by [12] and further confirmed by our study. However, the outbreak scenario does affect community synchrony, which we explore in Sec. 4.2. 3.5. Calibrating the model Correct implementation of the transmission model requires calibrating a number of free parameters to known empirical results. That is, by rescaling the contact probabilities and transmission probabilities, we are able to match previously reported attack rates in schools, households, and other social mixing groups. The first problem with calibrating the model is that contact groups all vary in size. The contact prob- abilities, reported in Tab. C.3, are determined for communities of fixed size following the study of [12]. However, depending on the size of the community, these probabilities should be rescaled such that in larger communities people are less likely to be in contact with a lower proportion of the population. We do so by linearly rescaling these probabilities based on the number of agents in each stochastically generated group. The second issue we address is calibrating the likelihood that a susceptible individual will contract a disease, given contact with another infectious individual, i.e., tuning ρ from Eq. (5). This is done by matching the known proportion of transmissions within each different social group. That is, the fractions of influenza transmissions within households, school, and other settings (work, communities, etc.) have been estimated as 30–40%, 20%, and 40-50%, respectively [29, 10]. Using these calibration procedures, the illness attack rates within different mixing groups in our simulation were matched to these studies, validating the simulations. 4. Simulation results We illustrate the use of the model by studying influenza pandemics in Australia for a variety of simulated pathogens. We quantify the transmissibility of a strain by the reproductive ratio R0 and study the epidemic curves of the disease for a number of different R0 values. In addition to analysing the (national) epidemic curve, we study the spatial distribution of the pandemic by prevalence proportion choropleths and disease synchrony. 8 Figure 4: Reproductive ratio R0, attack rate and transmission scaling factor κ for simulated influenza epidemics in Australia. 4.1. Reproductive ratio and the epidemic curve In the study of disease transmission, R0 is defined by "the expected number of secondary cases produced by a typical infected individual during its entire period of infectiousness in a completely susceptible popu- lation" [31]. It is thus a key value that coarsely describes the rate of growth of an infection and whether it will spread in the population. We illustrate both analytically and empirically that we can approximate R0 by a linear function of κ for a biologically plausible range of values. A similar result was empirically shown by Chao et al. [22], who noted that R0 is linear in κ (from Eq. (3)). Using Eq. (3) and taking "typical" to mean "in expectation", we obtain the R0 from Eq. (2): R0 = E Xi(N ) = symptomatic XJ (0) = symptomatic (cid:17) (cid:17)(cid:17) (cid:16)(cid:88) (cid:16) (cid:16)(cid:89) (cid:89) (cid:88) (cid:88) E n i = E = pI (n) XJ (0) = symptomatic (cid:16) 1 − (cid:89) (cid:0)1 − pg j→i(n)(cid:1)(cid:17) g∈Gj (n) , (6) j i n g f (n−nj i, j)qg outer products in Eq. (6), we note that the dominant terms are the first-order components(cid:80) κ (cid:80) where Gi(n) = Gj(n) by the commutative property of the mixing groups. By expanding the inner and j→i(n) = j→i. Thus, we expect an approximately linear relationship of R0 for reasonable values of κ. In order to validate our conjecture, we compute the sample population R0 by performing local simulations for a given set of parameters. Within each simulation, one agent is seeded at random, the number of cases directly caused by this agent is counted, then R0 is computed as the average of this number. Using this approach, we evaluate R0 for a number of κ values, as shown in Fig. 4, where the result is a linear and homoscedastic function; the simulation results thus support our analytical reasoning. Figure D.10 in the appendix illustrates the frequency distribution of secondary cases for κ = {1.0, 2.0, 3.0, 4.0}. g pg For a given R0, we measure the severity of the epidemic via the prevalence, incidence, and infection attack rate of a disease. The prevalence and incidence are defined by the number of current and newly ill agents at each discrete time step. Both quantities are commonly used to characterise the time dependent global state of the population with respect to the epidemic [12, 22, 21]. Tracing the incidence proportion (normalised by population size) over time produces an epidemic curve. In addition, the infection attack rate captures the proportion of the population that became ill over the duration of the epidemic [32] and is quantified by cumulating the incidence proportion. The attack rate is thus an effective summary measure for the time independent disease spread. Figure 5 illustrates the global epidemic state for simulated pathogens. As expected, a simulated strain with a higher R0 is associated with higher attack rate and an earlier incidence peak. Table 1 gives corre- sponding key features of the incidence curve for the range of R0 values used. The reported illness attack rates match expected empirical and simulated results [12, 29, 22], validating our model. 9 00.511.52R001234Transmissionscalingfactor50.20.4AttackRateFigure 5: The epidemic curve illustrates the trajectory of the epidemic by tracing the incidence over time for each R0 value. This illustrates the behaviour of the simulated influenza epidemics in Australia with R0 = {1.0, 1.25, 1.5, 1.75, 2.0}. We performed 10 trial runs for each R0 value, then plot the mean (solid lines) and standard deviation (shaded area). Moreover, Tab. 2 collates the national and the community average attack rates for the different age groups. We compute the national rate as the cumulative number of ill individuals in the country, giving a weighted average of the attack rate depending on the SLA size. The community based measure is an unweighted average of each SLA attack rate, obtained by taking the mean of the attack rate in that com- munity. The table illustrates the significant impact on student-aged individuals, having approximately four times higher incidence than the other groups. Further, the community and national attack rates converge for all age groups for higher R0, indicating the attack rate saturates uniformly for all SLAs. This is further evidenced by the synchrony increasing with higher R0 in Tab. 1. 4.2. Spatial spread of the disease We demonstrate the spatiotemporal spread of the simulated pandemics qualitatively and quantitatively through prevalence proportion choropleths and epidemic synchrony. Both tools are employed at the com- munity (SLA) level in order to identify important hubs and turning points of the developing epidemic. In Fig. 6 we present the spatial distribution for a realisation of the simulated influenza pandemics for R0 = 1.5 and R0 = 2. The snapshots are taken from key days that were identified from the epidemic curve in Fig. 5: onset inflexion point1 (day 30); peak of R0 = 2.0 (day 50); peak of R0 = 1.5 (day 62); and intersection point (day 88). Due to the seeding of the pathogen at major airports and the contact network, the epidemic tends to spread primarily around the coast before it can be traced inland, where often entire SLAs are left unaffected in central Australia. Moreover, the incidence typically peaks first in dense urban communities, followed by sporadic outbreaks in the less populous areas. This further motivates quantifying this behaviour by analysing the synchrony of community-based epidemic curves. Epidemic synchrony is characterised by the phase coherence of incidence curves in specified geographic locations [33]. In this simulation, however, each SLA follows a relatively smooth transition, allowing us to simplify the definition. We denote asynchrony of a set of communities by the variance of their incidence peaks (in days) and, consequently, synchrony is the reciprocal of this quantity. Figure 7 gives results on the synchrony of incidence peaks in AceMod simulations. It has been established that more populous regions of the United States exhibit more synchronised epidemics, i.e., that there exists a "hierarchical spatial spread" 1The inflexion point is approximated to be between the two days. 10 (a) R0 = 1.5 (b) R0 = 2.0 Figure 6: Prevalence proportion choropleths illustrating the spatial distribution of simulated influenza pandemics in Australia for R0 = 1.5 and R0 = 2. For the colour scale, the minimum prevalence (green) is 5 × 10−3 and maximum prevalence (red) is 8 × 10−2. We plot the distribution for days 30, 50, 62, and 88, with time going down the page. The pandemic spreads from these urban hubs inwards toward central Australia. Both simulations are sample realisations comprising the same demographics (contact) and mobility networks, as well as identical seeding at the same rate at major international airports around Australia (see Tab. 3). The epidemic peaks at larger cities at similar times, whereas less populous areas are less likely to synchronise. 11 (a) (b) (c) Figure 7: Synchronisation of epidemic curves between the different communities (SLAs) in Australia illustrating spatial hier- archies. Figure 7(a) investigates how synchronisation is related to community (SLA) size, i.e., communities of similar size are grouped together and their peak incidence times are compared. Synchrony is computed for 10 partitions (bins), with around 142 SLAs per interval. Figure 7(b) traces the pairwise synchrony between all pairs of SLAs against the pairwise distance, averaged over 10 partitions (bins). Figure 7(c) illustrates the effect of different outbreak scenarios on the synchrony of each community. This is achieved by seeding the pathogen at a varying number of SLAs, whilst varying the proportion of those SLAs initially infected. To the right of the figure we illustrate example attack rates (cumulative number of ill individuals) for given datapoints in the contour plot. The bold black lines illustrate the attack rate for the most populous SLAs in each state, whereas grey lines are all other SLAs. We observe an increased synchrony if the infection starts in a larger number of SLAs and with a higher proportion of infected individuals within these SLAs. 12 Table 1: Characteristics of the incidence curves, averaged over 10 trials. Basic reproductive number R0 Rate of spread: 1K ill∗ 10K ill∗ 100K ill∗ 1M ill∗ Peak of epidemic∗ Daily number of new cases at peak activity 1.0 29 44 60 80 84 1.25 22 33 44 58 63 1.5 21 30 39 50 59 1.75 18 25 33 42 47 2.0 17 24 31 39 47 85.3 K 140 K 189 K 257 K 328 K Number of days with > 100K ill 44 55 52 51 48 Cumulative number of ill individuals 3.4 M 5.0 M 6.4 M 7.7 M 8.8 M Synchrony of community epidemic peaks (10−3) M – Million; K – Thousand. ∗ Days after initial introduction. 1.38 2.67 4.33 6.39 7.70 Table 2: Attack rates, per 10K individuals, at the community and national level for different age demographics, averaged over 10 trials. The attack rate is the proportion of the population which became ill. The community attack rate is calculated as the average of the attack rates of each SLA. Age Group R0 = 1.0 R0 = 1.25 R0 = 1.5 R0 = 1.75 R0 = 2.0 0-4 5-18 19-34 35-64 65+ Overall 0-4 5-18 19-34 35-64 65+ Overall 1164 4140 967 975 939 1552 1280 4520 1042 1082 1003 1734 1898 5080 1698 1710 1665 2320 2020 5400 1774 1824 1724 2500 2600 5560 2500 2500 2450 3050 2730 5880 2590 2630 2520 3230 3270 5860 3250 3250 3210 3710 3370 6140 3320 3360 3260 3870 3870 6070 3970 3950 3920 4320 3960 6310 4010 4060 3940 4450 Cumulative number of community illnesses per 10K∗ Cumulative number of national illnesses per 10K∗ ∗Compared to the number of agents in that age group (e.g., per 10K 19-34 year olds). of influenza [33]. We investigate this hypothesis for Australian demographics (that are most prominently characterised by concentration of the population around urban and coastal regions). In Fig. 7(a) we group communities of similar size together and their synchrony is computed. Using the synchrony measure, we study the phenomena of both hierarchical and wave-like spread. While the hierarchical spread is captured by the synchrony across SLAs of similar sizes, the wave-like spread is associated with pairwise geographical distance. That is, to investigate the hierarchical hypothesis, we partition all SLAs into groups according to their population size and calculate the synchrony across all peak times within each group. This is shown in Fig. 7(a), where we see an increase in synchrony as the SLA size grows. In contrast, in order to verify the wave-like hypothesis, we compute pairwise synchrony between all pairs of SLAs, and trace it against the pairwise distance. Fig. 7(b) shows the average pairwise synchrony for ten partitions (bins) of the pairwise distance. Crucially, the average pairwise synchrony does not grow with the increases in the pairwise distance. These results support the hypothesis of a hierarchical spatial spread, suggesting the regional movement of disease depends on the worker/student flow moreso than the geographic distance. 13 We also note that the synchrony with respect to SLA size, shown in Fig. 7(a), is particularly pronounced for two specific SLA sizes. This two-phase characteristic is also evident in incidence plots (cf. Fig. 4.1), suggesting a secondary phase of an epidemic. The precise dependency between such phases and a multi- modal character of the synchrony profiles is a subject of future research. Finally, we study the effect of outbreak scenarios on epidemic synchrony by varying the number of SLAs infected at the beginning of the simulation, as well as the proportion of individuals seeded within these SLAs. The number of SLAs initially infected ranges from 1 to 1422, and the proportion of each SLAs infected is logarithmically spaced between 10−5 and 1. We then compare the synchronisation between community incidence peaks when we vary these two parameters, as represented with the contour plot shown in Fig. 7(c). 5. Discussion and future work In this paper we present the AceMod simulator as an approach for modelling epidemics in Australia using census data. Specifically, we used the 2006 Australian Census data to construct a contact network and develop a stochastic agent-based model for disease spread. The detailed data on population demographics and mobility allowed us to generate specific dynamics of influenza viral infection and transmission. More- over, the transmission and disease natural history models were calibrated to empirical data on influenza transmission, allowing us to validate the model. The agent-based model facilitated fine-grained analysis of the spatiotemporal dynamics of influenza spread both qualitatively and quantitatively through incidence curves, prevalence choropleths, and epidemic synchrony. This study has highlighted several interesting dependencies in large-scale epidemics, calibrated for Aus- tralian demographics and connectivity. Firstly, we showed that the reproductive ratio R0 can be approxi- mated by a linear function of a simple factor that scales the probability of disease transmission. Secondly, the epidemic curves, produced in our simulations, over a typical range of reproductive ratios, concur with the general ranges of the community and national attack rates, and show, in addition, the presence of sec- ondary phases following the primary peaks. The spatiotemporal spread of the simulated pandemics was also illustrated using prevalence proportion choropleths, identifying important hubs and turning points of an epidemic. Most importantly, we verified the hypothesis that the modelled disease spreads in accordance with spatial hierarchies as opposed to a wave-like fashion. This has been achieved by measuring synchrony as a function of both population size and pairwise distance between SLAs. As the number of initially infected SLAs and the proportion of each infected SLAs are increased, the synchrony between communities increases, suggesting that spatial hierarchies disappear depending on how widespread the pandemic already is when it reaches Australia. There are several important assumptions behind the presented results: fidelity of the Australian census datasets, including mobility patterns; knowledge of the transmission rates in specific mixing groups; and understanding of the natural history of the disease. With census datasets improving in their overall quality and coverage, there is an increasing confidence in data-driven simulation studies. In parallel, empirical epidemiological studies continue to narrow down the uncertainty in key modelling parameters. One of the emerging trends is thus an integrated real-time tracking of epidemic spreads with rapid simulation of "what-if" scenarios. It is known that infectious disease dynamics are heavily influenced by social network characteristics. It is then crucial to better represent and characterise the various Australia social networks in order to both ameliorate our predictions and suggest appropriate mitigation strategies. For this, we first aim to leverage the 2011 and 2016 census data to generate surrogate populations. With this, we could also study how the changes in the social network over time affect our predictions under the same disease dynamics. This could in turn help us in tuning our model to account for the predicted future changes in the social network structure. Subsequently, our goal is to study the network characteristics to better understand the relationship between epidemic dynamics and different network properties. For instance, the critical value of the repro- ductive ratio R0 (and its relationship to attack rate) is influenced by the network topology, and various graph-theoretic techniques can be employed to reveal these dependencies [34, 35, 36, 37, 38]. 14 Directed information-theoretic measures, such as transfer entropy [39], are particularly suited to this task in real systems in either the time-averaged or local forms [40]. These quantities have been used previously to infer interaction networks in multi-agent systems [41] and are particularly suited to processes with attracting states [42, 43]. This would allow us to study contact networks in terms of both structural and functional connectivity [44, 45, 46, 47] and their dominant motifs [48, 49, 50]. We can also exploit the predictive power of transfer entropy in identifying critical regimes of network dynamics and epidemic processes in particular [51, 52]. Precursors of phase transitions, identifiable by information dynamics [53, 54, 55] promise to strongly impact on the quality of epidemic forecasting. Other future avenues of inquiry are aimed at developing a more versatile computational epidemiological framework by considering (i) more complicated outbreak scenarios, e.g., bio-terrorism attacks with targeted infections; (ii) different scales of modelling (fast and slow dynamics of disease diffusion); (iii) more diverse interaction types varying across demographic groups, professions, territories, seasons, and other contexts; and (v) compound intervention strategies, ranging from traditional vaccination to high-precision quarantine decisions. 6. Acknowledgments The Authors are grateful to Tim Germann, Joseph Lizier, Cameron Zachreson, Kristopher Fair, David Newth, Philippa Pattison, Stephen Leeder, Tania Sorrell, Alessandro Montalto, and especially Peter Wang, for numerous helpful discussions of various intricacies involved in agent-based modelling of influenza. The Authors were supported through the Australian Research Council grant DP160102742. Appendix A. Australian Census Data We are using data from the 2006 Australian Census obtained from the Australian Bureau of Statistics (ABS). All data contained in the this section can be obtained publicly, with the exception of the work travel data. It should be noted that some of the data needs to be processed using the TableBuilder. The census data has a hierarchical structure based on geography, described below. Appendix A.1. Hierarchical Structure of ABS data For the 2006 census, the ABS divided Australia using the Australian standard geographical classifica- tion (ASGC). The levels present in the ASGC include: census collection district (CD), local government area (LGA), postal area (POA), remoteness classification (RA), statistical division (SD), statistical local area (SLA), state suburb (SSC), statistical sub division (SSD), urban centre or locality (UCL), section of state (SOS), state electoral division (SED), commonwealth electoral division (CED), state or territory (STE). Some of these structures, such as RA and UCL, exist to express geographic concepts specific only to certain regions, such as remoteness and urbanity. Therefore these structures may not cover the entirety of Australia. In this work we use the following ASGC structures: CD, SLA, STE. In addition we also use the destination zone (DZN) classification. Each structure used in this work forms a complete partition of Australia. In addition, the census data in each CD is collected for each household, thus the household (HH) is also a de facto structure of the census data. State (STE). The STE is the highest level structure used by the ABS. This comprises the six states, two territories and one offshore territory. The offshore territory comprises any area which is not part of mainland Australia or Tasmania. Statistical local area (SLA). The SLA is the second highest level structure used in this work, however this is only the fourth highest structure of the ASGC. In the 2006 census there are 1422 SLA's. The SLA has on average 14,064 people. The SLA's aggregate to states, that is no SLA is contained within two states. 15 Census collection district (CD). The CD is the structure with the highest hierarchical level used exclusively for the 'home regions' (see below). Data from a CD is collected by a single ABS representative, hence the name. In total there are 38,200 CDs which have an average population of 523 people. The CD's aggregate to SLA's, that is no CD is contained within two SLA's. Destination zone (DZN). The working group is the structure with the highest hierarchical level used exclu- sively for the 'work regions' (see below). The DZN describes the location that individuals go to work. In the 2006 census there are 9,800 DZN's. The DZNs aggregate to SLAs, i.e., no DZN is contained within two SLAs. In addition, in our simulations we have defined and used some lower level structures which are relevant to the underlying infection dynamics, as described in the following section. These are the household cluster-HC and the working group (WG). Appendix A.2. Australian Bureau of Statistics (ABS) data structures The ABS data is presented in tables. There are a number of tables which can be directly accessed from the ABS database. In addition, ABS provides the TableBuilder tool which we have used to build certain It should be noted that the tables do not provide additional tables according to our simulation needs. any data which could be used to identify individuals or households. Instead they are aggregated at the aforementioned structural levels (CD, SLA, STE). For example, a table built at the CD level could provide the number of individuals within the CD between certain age brackets, the number of households with a given number of members or the number of households who have a particular type of internet connection. The lowest structural level for which a table can be retrieved or built is the CD level. In this work, we retrieved or built the following tables for schools, family, households, age, sex, airport, travel to work.2 School distribution. This was retrieved at the structural level of states (STE). This is the highest resolution at which the ABS offers data about schools. In this table, each row represents a state. Each column represents a range of the number of students attending the school. The states (rows) are NSW, VIC, QLD, SA, WA, TAS, NT, ACT. The ranges (columns) are 0-35, 36-100, 101-200, 201-300, 301-400, 401-600, 601-800, 801-1000, 1000+. Age distribution. This was obtained at the structural level of the CD. In this table each row represents a CD. Each column gives the number of individuals of a specific age group and gender. The columns contain the number of males/females in the given CD who are between 0-4, 5-18, 19-39, 40-65, 65+. Family distribution. This table is constructed from three distinct tables from the ABS. The first table is defined at the level of CD and gives the number of households of a given size. The second table is also defined on the level of CD describes the number of each of four distinct types of family as defined by the ABS-Couples with children (CWC), couples without children (CWOC), single parent families (SPF) and other (OTH). The third table is defined on the national level gives the proportion of households which are family units or non family units. Non family units constitute households of either a single person, or people sharing accommodation without a familial relationship. Household Composition.dat is defined at the structural level of the CD. In this table each row represents a CD. Each column corresponds to a particular combination of household attributes derived from the first, second and third tables. Therefore each cell corresponds to the proportion of households with a particular combination of attributes in a given CD. Worker flow. Each row corresponds to a (CD, DZ) pair which represents a worker flow from a CD (source) to a DZ (target). The first column gives the CD code of the worker flow. The second column gives the DZ code of the flow and the third column gives the number of individuals living in the CD and working in the DZ. 2The travel to work data is not publicly available and was purchased from the ABS. 16 Daily incoming passengers. This table lists all the international airports in Australia. Each row in this table corresponds to a particular international airport and the number of passengers entering that airport from overseas per month on average. This data is taken from The Bureau of Infrastructure, Transport and Regional Economics (BITRE) [30]. Appendix B. Simulated population structure Following [12, 5] in constructing our national simulation model, we use a discrete-time stochastic simu- lation model of disease spread within a structured SLA. The model population is stochastically generated to match CD attributes for gender, age structure, household size, employment status, and location of em- ployment among others. AceMod was implemented in C++11 and run on a High Performance Computing (HPC) service (Artemis at the University of Sydney), comprising 4264 cores of computing capacity, consisting of 136 standard compute nodes, two high memory nodes (512 GB), three very high memory nodes (6 TB), and five GPU compute nodes, under a "Fair Share" resource allocation model (e.g., no more than 288 cores). A typical run of 180 days of an epidemic is simulated on an HPC node in 42 minutes, so that each scenario covering 10 runs varying over infection "seeds" is complete and integrated within seven hours. However, the code is not optimised for performance and includes numerous debugging statistics; thus much lower run-times could be achieved with minimal effort. Each person in the population belongs to one of five age groups: preschool-age children (0–4 years), school-age children (5–18 years), young adults (19–29 years), adults (30–64 years), and older adults (64+ years). The individuals within the model are arranged into one of 4 family structures: singles; couples without children; couples with children; and single parent families. In general, households consist of a number of persons, matching the Australian census data, and are grouped randomly into clusters of four households each. Every person also belongs to a set of close and casual contact (also referred to as "mixing") groups, ranging from their household and household cluster (highest contact probabilities) to workplaces, collection districts, and to the SLA (with the lowest contact probabilities, representing occasional "long distance" interactions in supermarkets, shopping centres, theatres, libraries, mass gatherings, etc.). Depending on their age, school-age children attend primary school or high-school. In each case, children are arranged into 25 person groups referred to as 'classes'. According to the Australian Census data 98.7% of children 5-18 years old attend school, so we allow the remaining 1.3% to mix in the household, household cluster, and community during daytime. Working adults (restricted to those who are 19–64 years old) belong to a working group of a varying size determined by census data. We assume that workers make a contact of sufficient duration and/or closeness to transmit the disease with a subset of the entire workforce at that location. Appendix B.1. Assigning nighttime mixing groups In order to populate the nighttime (home) regions, our simulation iterates through CDs from the census data. For each CD, a cumulative density function (CDF) is constructed from the family distributions as described above. For a particular CD, we then create households one by one. Each household has attributes assigned to it probabilistically based on the CDF, including the number of people living in the household and the type of family living in that household. This process continues until the population of the CD is reached. After the household attributes are chosen, individuals are assigned their own attributes, e.g., a single parent family of size 3 requires one adult and two children. Therefore one person will be assigned as an adult, and two will be assigned as children. The age and gender of each individual is then randomly assigned using data from the age distribution. Appendix B.2. Assigning daytime mixing groups After home regions have been populated, the worker flow information is used to generate mobility data. This data gives the number of workers travelling between a CD and a DZN. This number of workers are then 17 (a) Workers (b) Students Figure B.8: Commute distance distributions. selected randomly from the CD and assigned as workers in the given DZN. Therefore for each (CD, DZN) pair, we randomly select the given number of unassigned, working aged individuals from the corresponding CD and assign them to the corresponding DZN. Figure B.8 shows the distribution of commute distances for working individuals, as well as school students (described in the next subsection), generated by AceMod. Appendix B.2.1. Populating schools Our simulation generates schools and assigns a maximum capacity to these schools according to the dis- tribution of school sizes available from ABS. The schools are then spatially assigned based on the availability of school staff in the relevant DZNs. The number of school staff in each DZN will determine the number of students able to go to school in that DZN based on a fixed staff to student ratio. Each school is randomly assigned to a DZN which has a sufficient number of potential students. The number of potential students in a DZN is then reduced by the size of the school. This continues until all potential schools have been allocated to a DZ. At this point, each DZN has a number of schools with unfilled student and staff places. In order to assign students to schools, we assume that students would go to a school if they live within the catchment zone of the school. This process mimics the process that students are likely to attend a schools within a short distance first, and if there are no places available they would search further afield. Similarly, schools are more likely to accept students from a greater distance if it is evident that there are no other closer options. If no school could be found within the maximum radius, it is assumed that the student does not attend a school. The catchment area of a school is defined by randomly selecting a radius from a distribution, which is a folded normal distribution converted from a normal distribution with mean and variance of 10km and 4km. If the student lives within the catchment zone of many schools, the algorithm will pick one school with probability proportional to the remaining available student places in the school. Once a student place is chosen, the total available student places in the given school are reduced by one. We do not distinguish between public and private schools for the allocation purposes of students. If a student does not live within a catchment zone of any school, the algorithm searches an area with double the original radius and again selects a school with probability proportional to the available places. If there is still no school within the doubled radius, the radius is doubled again. This process is repeated until the algorithm finds a school or reaches the maximum searching radius of 100 km. Once the students are all assigned to school places, we use the student to staff ratio (2:17) (extracted from historical ABS data) and the enrolment numbers of the schools to calculate the number of staff working in each school. These staff members are then selected randomly from the ensemble of people who work in that particular DZN. The remaining people who work in the DZN are assigned to non-school working groups, as follows. The number of remaining workers is divided by the maximum size of a working group (20) and the 18 Table C.3: Daily contact probabilities cg j→i for different contact groups g, reported by [22]. Mixing group g Infected individual j Susceptible individual i Contact probability cg j→i Household cluster Child (<19) Child (<19) Adult (>18) Adult (>18) Working Group Adult (19-64) Neighbourhood Community Any Any Any Any Any Any Any Any Child (<19) Adult (>18) Child (<19) Adult Adult (19-64) Child (0-4) Child (5-18) Adult (19-64) Adult (65+) Child (0-4) Child (5-18) Adult (19-64) Adult (65+) 0.08 0.035 0.025 0.04 0.05 0.0000435 0.0001305 0.000348 0.000696 0.0000109 0.0000326 0.000087 0.000174 Table C.4: Daily transmission probabilities qg by [10]. j→i for different contact groups g, obtained by Eq. (4) where βg j→i are reported Contact Group g Infected Individual j Susceptible Individual i Transmission Probability qg j→i Household size 2 Household size 3 Household size 4 Household size 5 Household size 6 Any Any Any Any Any Any Any Any Any Any School Grade Class Child (<19) Child (<19) Child (<19) Child (<19) Adult (>18) Child (<19) Adult (>18) Child (<19) Adult (>18) Child (<19) Adult (>18) Child (<19) Adult (>18) Child (<19) Child (<19) Child (<19) 0.0933 0.0393 0.0586 0.0244 0.0417 0.0173 0.0321 0.0133 0.0259 0.0107 0.000292 0.00158 0.035 ceiling of this number is the number of non school working groups within the particular DZN. Workers are then randomly assigned to these working groups such that none will exceed a size of 20. The results of this assignment algorithm are summarised in the following figures: • Figure B.9 shows school size distributions; • Figure B.8 show commute to school (gamma) distribution (mean distance to school). Appendix C. Transmission and contact probabilities The transmission probabilities (for Eq. (4)) and contact probabilities (for Eq. (3)) are given in Tab. C.3 and Tab. C.4, respectively. Appendix D. Supplementary results Figure D.10 illustrates the number of secondary cases from a typical outbreak. Figure D.11 shows the prevalence choropleth for the Greater Sydney Area. 19 (a) NSW (b) VIC (c) QLD (d) WA (e) NT (f) SA (g) TAS (h) ACT Figure B.9: School size distributions. 20 02004006000200400600Frequency050010001500200025003000School size (students)-100010002004000200400600Frequency050010001500200025003000School size (students)-200020001002003000100200300Frequency050010001500200025003000School size (students)-200020001002000100200300Frequency050010001500200025003000School size (students)-100010002040600204060Frequency050010001500200025003000School size (students)-20020050100150050100150Frequency050010001500200025003000School size (students)-500500204060050100Frequency050010001500200025003000School size (students)-1001001020300102030Frequency050010001500200025003000School size (students)-20020(a) κ = 1.0 (b) κ = 2.0 (c) κ = 3.0 (d) κ = 4.0 Figure D.10: Number of secondary cases for simulations with κ = {1.0, 2.0, 3.0, 4.0}. 21 24681012141618Number of secondary cases100101102103104Frequency24681012141618Number of secondary cases100101102103104Frequency24681012141618Number of secondary cases101102103104Frequency24681012141618Number of secondary cases101102103104FrequencyFigure D.11: Prevalence proportion choropleths of Sydney for R0 = 1.5 and R0 = 2. We plot the distribution for days 30, 50, 62, and 88, with time going down the page. Both simulations are realisations comprising the same demographics (contact) and mobility networks, as well as identical seeding at the same rate at major international airports around Australia (see Tab. 3). 22 References [1] A. T. Newall, P. A. Scuffham, B. Hodgkinson, Economic report into the cost of influenza to the Australian health system, Tech. rep. (2007). [2] T. M. Lanzieri, S. R. Bialek, I. R. Ortega-Sanchez, M. Gambhir, Modeling the potential impact of vaccination on the epidemiology of congenital cytomegalovirus infection, Vaccine 32 (30) (2014) 3780–3786. [3] B. A. Lopman, V. E. Pitzer, R. Sarkar, B. Gladstone, M. Patel, J. Glasser, M. Gambhir, C. Atchison, B. T. Grenfell, W. J. Edmunds, et al., Understanding reduced rotavirus vaccine efficacy in low socio-economic settings, PloS one 7 (8) (2012) e41720. [4] K. Simmons, M. Gambhir, J. Leon, B. Lopman, Duration of immunity to norovirus gastroenteritis, Emerging infectious diseases 19 (8) (2013) 1260. [5] I. M. Longini, A. Nizam, S. Xu, K. Ungchusak, W. Hanshaoworakul, D. A. Cummings, M. E. Halloran, Containing pandemic influenza at the source, Science (New York, N.Y.) 309 (5737) (2005) 1083–1087. [6] N. M. Ferguson, D. A. T. Cummings, S. Cauchemez, C. Fraser, S. Riley, A. Meeyai, S. Iamsirithaworn, D. S. Burke, Strategies for containing an emerging influenza pandemic in Southeast Asia, Nature 437 (8) (2005) 209–214. [7] E. O. Nsoesie, R. J. Beckman, M. V. Marathe, Sensitivity analysis of an individual-based model for simulation of influenza epidemics, PLOS ONE 7 (10) (2012) 0045414. [8] E. O. Nsoesie, J. S. Brownstein, N. Ramakrishnan, M. V. Marathe, A systematic review of studies on forecasting the dynamics of influenza outbreaks, Influenza and other respiratory viruses 8 (3) (2014) 309–316. [9] R. Pastor-Satorras, C. Castellano, P. Van Mieghem, A. Vespignani, Epidemic processes in complex networks, arXiv:1408.2701 (2014). [10] S. Cauchemez, A. Bhattarai, T. L. Marchbanks, R. P. Fagan, S. Ostroff, N. M. Ferguson, D. Swerdlow, S. V. Sodha, M. E. Moll, F. J. Angulo, et al., Role of social networks in shaping disease transmission during a community outbreak of 2009 H1N1 pandemic influenza, Proceedings of the National Academy of Sciences 108 (7) (2011) 2825–2830. [11] M. J. Keeling, P. Rohani, Modeling infectious diseases in humans and animals, Princeton University Press, 2011. [12] T. C. Germann, K. Kadau, I. M. Longini, C. A. Macken, Mitigation strategies for pandemic influenza in the United States, Proceedings of the National Academy of Sciences 103 (15) (2006) 5935–5940. [13] M. E. Halloran, I. M. Longini, A. Nizam, Y. Yang, Containing bioterrorist smallpox, Science 298 (5597) (2002) 1428–1432. [14] I. M. Longini, M. E. Halloran, A. Nizam, Y. Yang, Containing pandemic influenza with antiviral agents, American Journal of Epidemiology 159 (7) (2004) 623–633. [15] S. Eubank, H. Guclu, V. A. Kumar, M. V. Marathe, A. Srinivasan, Z. Toroczkai, N. Wang, Modelling disease outbreaks in realistic urban social networks, Nature 429 (6988) (2004) 180. [16] C. L. Barrett, K. R. Bisset, S. G. Eubank, X. Feng, M. V. Marathe, Episimdemics: an efficient algorithm for simulating the spread of infectious disease over large realistic social networks, in: High Performance Computing, Networking, Storage and Analysis, 2008. SC 2008. International Conference for, IEEE, 2008, pp. 1–12. [17] K. R. Bisset, J. Chen, X. Feng, V. Kumar, M. V. Marathe, Epifast: a fast algorithm for large scale realistic epidemic simulations on distributed memory systems, in: Proceedings of the 23rd international conference on Supercomputing, ACM, 2009, pp. 430–439. [18] C. Barrett, K. Bisset, J. Leidig, A. Marathe, M. V. Marathe, An integrated modeling environment to study the co-evolution of networks, individual behavior and epidemics, AI Magazine 31 (1) (2010) 75–87. [19] K. R. Bisset, A. M. Aji, E. Bohm, L. V. Kale, T. Kamal, M. V. Marathe, J.-S. Yeom, Simulating the spread of infectious disease over large realistic social networks using charm++, in: Parallel and Distributed Processing Symposium Workshops & PhD Forum (IPDPSW), 2012 IEEE 26th International, IEEE, 2012, pp. 507–518. [20] M. E. Halloran, N. M. Ferguson, S. Eubank, I. M. Longini, D. A. Cummings, B. Lewis, S. Xu, C. Fraser, A. Vullikanti, T. C. Germann, et al., Modeling targeted layered containment of an influenza pandemic in the United States, Proceedings of the National Academy of Sciences 105 (12) (2008) 4639–4644. [21] D. Balcan, B. Gon¸calves, H. Hu, J. J. Ramasco, V. Colizza, A. Vespignani, Modeling the spatial spread of infectious diseases: The global epidemic and mobility computational model, Journal of computational science 1 (3) (2010) 132–145. [22] D. L. Chao, M. E. Halloran, V. J. Obenchain, I. M. Longini Jr, FluTE, a publicly available stochastic influenza epidemic simulation model, PLoS computational biology 6 (1) (2010) e1000656. [23] F. Carrat, E. Vergu, N. M. Ferguson, M. Lemaitre, S. Cauchemez, S. Leach, A.-J. Valleron, Time lines of infection and disease in human influenza: a review of volunteer challenge studies, American journal of epidemiology 167 (7) (2008) 775–785. [24] A. C. Hayward, E. B. Fragaszy, A. Bermingham, L. Wang, A. Copas, W. J. Edmunds, N. Ferguson, N. Goonetilleke, G. Harvey, J. Kovar, et al., Comparative community burden and severity of seasonal and pandemic influenza: results of the flu watch cohort study, The Lancet Respiratory Medicine 2 (6) (2014) 445–454. [25] N. H. Leung, C. Xu, D. K. Ip, B. J. Cowling, The fraction of influenza virus infections that are asymptomatic: a systematic review and meta-analysis, Epidemiology (Cambridge, Mass.) 26 (6) (2015) 862. [26] L. Feng, P. Yang, T. Zhang, J. Yang, C. Fu, Y. Qin, Y. Zhang, C. Ma, Z. Liu, Q. Wang, et al., Technical guidelines for the application of seasonal influenza vaccine in China (2014–2015), Human vaccines & immunotherapeutics 11 (8) (2015) 2077–2101. [27] Bureau of Infrastructure, Transport and Regional Economics (BITRE), Australia's commuting distance: cities and regions, BITRE (2015). [28] J. Mossong, N. Hens, M. Jit, P. Beutels, K. Auranen, R. Mikolajczyk, M. Massari, S. Salmaso, G. S. Tomba, J. Wallinga, et al., Social contacts and mixing patterns relevant to the spread of infectious diseases, PLoS medicine 5 (3) (2008) e74. 23 [29] Y. Yang, J. D. Sugimoto, M. E. Halloran, N. E. Basta, D. L. Chao, L. Matrajt, G. Potter, E. Kenah, I. M. Longini, The transmissibility and control of pandemic influenza a (H1N1) virus, Science 326 (5953) (2009) 729–733. [30] Bureau of Infrastructure, Transport and Regional Economics (BITRE), Airport traffic data, BITRE (2017). URL https://bitre.gov.au/publications/ongoing/airport_traffic_data.aspx [31] O. Diekmann, J. A. P. Heesterbeek, J. A. Metz, On the definition and the computation of the basic reproduction ratio r0 in models for infectious diseases in heterogeneous populations, Journal of mathematical biology 28 (4) (1990) 365–382. [32] M. Porta, A dictionary of epidemiology, Oxford University Press, 2014. [33] C. Viboud, O. N. Bjørnstad, D. L. Smith, L. Simonsen, M. A. Miller, B. T. Grenfell, Synchrony, waves, and spatial hierarchies in the spread of influenza, Science 312 (5772) (2006) 447–451. [34] M. Piraveenan, M. Prokopenko, A. Y. Zomaya, Assortativeness and information in scale-free networks, European Physical Journal B 67 (2009) 291–300. [35] M. Kitsak, L. Gallos, S. Havlin, F. Liljeros, L. Muchnik, H. Stanley, H. Makse, Identification of influential spreaders in complex networks, Nature Physics 6 (11) (2010) 888–893. [36] M. Piraveenan, M. Prokopenko, A. Y. Zomaya, Assortative mixing in directed biological networks, IEEE Transactions on computational biology and bioinformatics 9 (2012) 66–78. [37] G. Daraganova, P. Pattison, J. Koskinen, B. Mitchell, A. Bill, M. Watts, S. Baum, Networks and geography: Modelling community network structures as the outcome of both spatial and network processes, Social networks 34 (1) (2012) 6–17. [38] M. Piraveenan, M. Prokopenko, L. Hossain, Percolation Centrality: Quantifying Graph-Theoretic Impact of Nodes during Percolation in Networks, PLoS ONE 8 (1) (2013) e53095+. [39] T. Schreiber, Measuring information transfer, Physical Review Letters 85 (2000) 461. [40] J. T. Lizier, M. Prokopenko, A. Y. Zomaya, Local information transfer as a spatiotemporal filter for complex systems, Physical Review E 77 (2) (2008) 026110. [41] O. M. Cliff, J. T. Lizier, X. R. Wang, P. Wang, O. Obst, M. Prokopenko, Quantifying long-range interactions and coherent structure in multi-agent dynamics, Art. Life. [42] O. M. Cliff, M. Prokopenko, R. Fitch, An information criterion for inferring coupling of distributed dynamical systems, Frontiers in Robotics and AI 3 (2016) 71. [43] O. M. Cliff, M. Prokopenko, R. Fitch, Minimising the Kullback–Leibler divergence for model selection in distributed nonlinear systems, Entropy 20 (2) (2018) 51. [44] K. J. Friston, Functional and effective connectivity in neuroimaging: A synthesis, Human Brain Mapping 2 (1994) 56–78. [45] C. J. Honey, R. Kotter, M. Breakspear, O. Sporns, Network structure of cerebral cortex shapes functional connectivity on multiple time scales, Proceedings of the National Academy of Sciences 104 (24) (2007) 10240–10245. [46] L. M. A. Bettencourt, G. J. Stephens, M. I. Ham, G. W. Gross, Functional structure of cortical neuronal networks grown in vitro, Physical Review E 75 (2) (2007) 021915. [47] J. T. Lizier, M. Piraveenan, D. Pradhana, M. Prokopenko, L. S. Yaeger, Functional and structural topologies in evolved neural networks, in: Advances in Artificial Life: Tenth European Conference on Artificial Life (ECAL '09), Vol. 5777-5778 of LNCS/LNAI, Springer, 2009. [48] R. Milo, S. Shen-Orr, S. Itzkovitz, N. Kashtan, D. Chklovskii, U. Alon, Network motifs: simple building blocks of complex networks., Science 298 (5594) (2002) 824–827. [49] O. Sporns, R. Kotter, Motifs in brain networks, PLoS Biology 2 (11) (2004) e369. [50] M. Piraveenan, M. Prokopenko, A. Y. Zomaya, Local assortativeness in scale-free networks, Europhysics Letters 84 (2) 28002. [51] J. T. Lizier, S. Pritam, M. Prokopenko, Information dynamics in small-world Boolean networks, Artificial Life 17 (4) (2011) 293–314. [52] E. Y. Erten, J. T. Lizier, M. Piraveenan, M. Prokopenko, Criticality and information dynamics in epidemiological models, Entropy 19 (5). [53] X. R. Wang, J. T. Lizier, M. Prokopenko, Fisher information at the edge of chaos in random Boolean networks, Artificial Life 17 (4) (2011) 315–329. [54] M. Prokopenko, J. T. Lizier, O. Obst, X. R. Wang, Relating Fisher information to order parameters, Physical Review E 84 (4) (2011) 041116. [55] R. E. Spinney, M. Prokopenko, J. T. Lizier, Transfer entropy in continuous time, with applications to jump and neural spiking processes, Physical Review E 95 (3) (2017) 032319. 24
1905.11828	2	1905	2019-06-04T08:32:31	AsymDPOP: Complete Inference for Asymmetric Distributed Constraint Optimization Problems	[ "cs.MA" ]	Asymmetric distributed constraint optimization problems (ADCOPs) are an emerging model for coordinating agents with personal preferences. However, the existing inference-based complete algorithms which use local eliminations cannot be applied to ADCOPs, as the parent agents are required to transfer their private functions to their children. Rather than disclosing private functions explicitly to facilitate local eliminations, we solve the problem by enforcing delayed eliminations and propose AsymDPOP, the first inference-based complete algorithm for ADCOPs. To solve the severe scalability problems incurred by delayed eliminations, we propose to reduce the memory consumption by propagating a set of smaller utility tables instead of a joint utility table, and to reduce the computation efforts by sequential optimizations instead of joint optimizations. The empirical evaluation indicates that AsymDPOP significantly outperforms the state-of-the-arts, as well as the vanilla DPOP with PEAV formulation.	cs.MA	cs	AsymDPOP: Complete Inference for Asymmetric Distributed Constraint Optimization Problems Yanchen Deng , Ziyu Chen∗ , Dingding Chen , Wenxin Zhang and Xingqiong Jiang [email protected], {chenziyu,dingding}@cqu.edu.cn, [email protected], [email protected] College of Computer Science, Chongqing University 9 1 0 2 n u J 4 ] A M . s c [ 2 v 8 2 8 1 1 . 5 0 9 1 : v i X r a Abstract Asymmetric distributed constraint optimization problems (ADCOPs) are an emerging model for co- ordinating agents with personal preferences. How- ever, the existing inference-based complete algo- rithms which use local eliminations cannot be ap- plied to ADCOPs, as the parent agents are required to transfer their private functions to their children. Rather than disclosing private functions explicitly to facilitate local eliminations, we solve the prob- lem by enforcing delayed eliminations and propose AsymDPOP, the first inference-based complete al- gorithm for ADCOPs. To solve the severe scala- bility problems incurred by delayed eliminations, we propose to reduce the memory consumption by propagating a set of smaller utility tables instead of a joint utility table, and to reduce the computa- tion efforts by sequential optimizations instead of joint optimizations. The empirical evaluation in- dicates that AsymDPOP significantly outperforms the state-of-the-art, as well as the vanilla DPOP with PEAV formulation. 1 Introduction Distributed constraint optimization problems (DCOPs) [Modi et al., 2005; Fioretto et al., 2018] are a fundamental frame- work in multi-agent systems where agents coordinate their decisions to optimize a global objective. DCOPs have been adopted to model many real world problems including ra- dio frequency allocation [Monteiro et al., 2012], smart grid [Fioretto et al., 2017] and distributed scheduling [Mah- eswaran et al., 2004b; Li et al., 2016]. Most of complete algorithms for DCOPs employ either distributed search [Hirayama and Yokoo, 1997; Modi et al., 2005; Gershman et al., 2009; Yeoh et al., 2010] or inference [Petcu and Faltings, 2005b; Vinyals et al., 2011] to optimally solve DCOPs. However, since DCOPs are NP-hard, complete algorithms cannot scale up due to exponential overheads. incomplete algorithms [Maheswaran et al., 2004a; Thus, Zhang et al., 2005; Okamoto et al., 2016; Rogers et al., 2011; Zivan and Peled, 2012; Chen et al., 2018; Ottens et al., 2017; ∗Corresponding Author Fioretto et al., 2016] are proposed to trade optimality for smaller computational efforts. Unfortunately, DCOPs fail to capture the ubiquitous asym- metric structure [Burke et al., 2007; Maheswaran et al., 2004b; Ramchurn et al., 2011] since each constrained agent shares the same payoffs. PEAV [Maheswaran et al., 2004b] attempts to capture the asymmetric costs by introducing mir- ror variables and the consistency is enforced by hard con- straints. However, PEAV suffers from scalability problems since the number of variables significantly increases. More- over, many classical DCOP algorithms perform poorly when applied to the formulation due to the presence of hard con- straints [Grinshpoun et al., 2013]. On the other side, AD- COPs [Grinshpoun et al., 2013] are another framework that captures asymmetry by explicitly defining the exact payoff for each participant of a constraint without introducing any variables, which has been intensively investigated in recent years. Solving ADCOPs involves evaluating and aggregating the payoff for each constrained agent, which is challenging in asymmetric settings due to a privacy concern. SyncABB and ATWB [Grinshpoun et al., 2013] are asymmetric adaption of SyncBB [Hirayama and Yokoo, 1997] and AFB [Gershman et al., 2009], using an one-phase strategy to aggregate the in- dividual costs. That is, the algorithms systematically check each side of a constraint before reaching a complete assign- ment. Besides, AsymPT-FB [Litov and Meisels, 2017] is the first tree-based algorithm for ADCOPs, which uses forward bounding to compute lower bounds and back bounding to achieve one-phase check. Recently, PT-ISABB [Deng et al., 2019] was proposed to improve the tree-based search by im- plementing a non-local elimination version of ADPOP [Petcu and Faltings, 2005a] to provide much tighter lower bounds. However, since it relies on an exhaustive search to guaran- tee the optimality, the algorithm still suffers from exponential communication overheads. On the other hand, although com- plete inference algorithms (e.g., DPOP [Petcu and Faltings, 2005b]) only require a linear number of messages to solve DCOPs, they cannot be directly applied to ADCOPs with- out PEAV due to their requirement for complete knowledge of each constraint to facilitate variable elimination. Accord- ingly, the parents have to transfer their private cost functions to their children, which leaks at least a half of privacy. In this paper, we adapt DPOP for solving ADCOPs for the first time by deferring the eliminations of variables. Specif- ically, we contribute to the state-of-the-art in the following aspects. • We propose AsymDPOP, the first complete inference- based algorithm to solve ADCOPs, by generalizing non- local elimination [Deng et al., 2019]. That is, instead of eliminating variables at their parents, we postpone the eliminations until their highest neighbors in the pseudo tree. In other words, an agent in our algorithm may be responsible for eliminating several variables. • We theoretically analyze the complexity of our algo- rithm where the space complexity of an agent is not only exponential in the number of its separators but also the number of its non-eliminated descendants. • We scale up our algorithm by introducing a table-set propagation scheme to reduce the memory consumption and a mini-batch scheme to reduce the number of op- erations when performing eliminations. Our empirical evaluation indicates that our proposed algorithm signif- icantly outperforms the state-of-the-art, as well as the vanilla DPOP with PEAV formulation. 2 Backgrounds In this section we introduce the preliminaries including DCOPs, ADCOPs, pseudo tree, DPOP and non-local elim- ination. 2.1 Distributed Constraint Optimization Problems A distributed constraint optimization problem [Modi et al., 2005] is defined by a tuple (cid:104)A, X, D, F(cid:105) where • A = {a1, . . . , an} is the set of agents • X = {x1, . . . , xm} is the set of variables • D = {D1, . . . , Dm} is the set of domains. Variable xi takes values from Di • F = {f1, . . . , fq} is the set of constraint functions. Each function fi : Di1 × ··· × Dik → R≥0 specifies the cost assigned to each combination of xi1, . . . , xik. For the sake of simplicity, we assume that each agent con- trols a variable (and thus the term "agent" and "variable" could be used interchangeably) and all constraint functions are binary (i.e., fij : Di × Dj → R≥0). A solution to a DCOP is an assignment to all the variables such that the total cost is minimized. That is, (cid:88) fij∈F X∗ = arg min di∈Di,dj∈Dj fij(xi = di, xj = dj) 2.2 Asymmetric Distributed Constraint Optimization Problems While DCOPs assume an equal payoff for each participant of each constraint, asymmetric distributed constraint optimiza- tion problems (ADCOPs) [Grinshpoun et al., 2013] explicitly define the exact payoff for each constrained agent. In other words, a constraint function fi : Di1×···×Dik → Rk≥0 in an ADCOP specifies a cost vector for each possible combination Figure 1: An ADCOP with two variables and a constraint Figure 2: A pseudo tree of involved variables. And the goal is to find a solution which minimizes the aggregated cost. An ADCOP can be visualized by a constraint graph where the vertexes denote variables and the edges denote constraints. Fig. 1 presents an ADCOP with two variables and a constraint. Besides, for the constraint be- tween xi and xj, we denote the private function for xi and xj as fij and fji, respectively. 2.3 Pseudo Tree A pseudo tree [Freuder and Quinn, 1985] is an ordered ar- rangement to a constraint graph in which different branches are independent. A pseudo tree can be generated by a depth- first traverse to a constraint graph, categorizing constraints into tree edges and pseudo edges (i.e., non-tree edges). The neighbors of an agent ai are therefore categorized into its parent P (ai), pseudo parents P P (ai), children C(ai) and pseudo children P C(ai) according to their positions in the pseudo tree and the types of edges they connect through. We also denote its parent and pseudo parents as AP (ai) = P (ai) ∪ P P (ai), and its descendants as Desc(ai). Besides, we denote its separators, i.e., the set of ancestors which are constrained with ai and its descendants, as Sep(ai) [Petcu and Faltings, 2006]. Finally, we denote ai's interface de- scendants, the set of descendants which are constrained with Sep(ai), as ID(ai). Fig. 2 presents a pseudo tree in which the dotted edge is a pseudo edge. 2.4 DPOP and Non-local Elimination DPOP [Petcu and Faltings, 2005b] is an inference-based complete algorithm for DCOPs based on bucket elimination [Dechter, 1999]. Given a pseudo tree, it performs a bottom- up utility propagation phase to eliminate variables and a value propagation phase to assign the optimal assignment for each variable. More specifically, in the utility propagation phase, an agent ai eliminates its variables from the joint utility table by computing the optimal utility for each possible assignment to Sep(ai) after receiving the utility tables from its children, and sends the projected utility table to its parent. In the value propagation phase, ai computes the optimal assignments for its variables by considering the assignments received from its transfer their private functions to their children, which would lead to an unacceptable privacy loss. Inspired by non-local elimination, we consider an alterna- tive privacy-protecting approach to aggregate constraint func- tions. That is, instead of deferring to its parent, we postpone the elimination of a variable to its highest (pseudo) parent. In this way, all the functions involving the variables have been aggregated and the variable can be eliminated from the util- ity table optimally. Note that the utility table is a summation of the utility tables from children and local constraints. As a result, although the utility table which contains the variable's private functions is propagated to its ancestors without elim- ination, the ancestors can hardly infer the exact payoffs in these private functions. We refer the bottom-up utility propa- gation phase as generalized non-local elimination (GNLE). Before diving into the details of GNLE, let's first consider the following definitions. Definition 1 (dims). The dims(·) is a function which returns the set of dimensions of a utility table. Definition 2 (Slice). Let S be a set of key-value pairs. S[K] is a slice of S over K such that S[K] = {(k = v) ∈ Sk ∈ K} Definition 3 (Join [Vinyals et al., 2011]). Let U, U(cid:48) be two utility tables and DU = ×xi∈dims(U )Di, DU(cid:48) = ×xi∈dims(U(cid:48))Di be their joint domain spaces. U ⊗ U(cid:48) is the join of U and U(cid:48) over DU⊗U(cid:48) = ×xi∈dims(U )∪dims(U(cid:48))Di such that (U⊗U(cid:48))(V ) = U (V[dims(U )])+U(cid:48)(V[dims(U(cid:48))]),∀V ∈ DU⊗U(cid:48) Algorithm 1 presents the sketch of GNLE. The algorithm begins with leaf agents sending their utility tables to their par- ents via UTIL messages (line 2-3). When an agent ai receives a UTIL message from a child ac, it joins its private functions w.r.t. its (pseudo) children in branch ac (line 5-6), and elimi- nates all the belonging variables whose highest (pseudo) par- ents are ai from the utility table (line 7). Here, EV (ai, ac) is given by EV (ai, ac) = P C(ai) ∩ Desc(ac) ∪ {ac}\ID(ai) Then ai joins the eliminated utility table with its running util- ity table utili. It is worth mentioning that computing the set of elimination variables EV (ai, ac) does not require agents to exchange their relative positions in a pseudo tree. Specif- ically, each variable is associated with a counter which is initially set to the number of its parent and pseudo parents. When its (pseudo) parent receives the UTIL message con- taining it, the counter decreases. And the variable is added to the set of elimination variables as soon as its counter equals zero. After receiving all the UTIL messages from its children, ai propagates the utility table utili to its parent if it is not the root (line 12). Otherwise, the value propagation phase starts (line 10). 3.2 Value Propagation Phase In contrary to the one in vanilla DPOP which determines the optimal assignment locally for each variable, the value as- signment phase in AsymDPOP should be specialized to ac- commodate the non-local elimination. Specifically, since a parent, and propagates the joint assignment to its children. Although DPOP only requires a linear number of messages to solve a DCOP, its memory consumption is exponential in the induced width. Thus, several tradeoffs including ADPOP [Petcu and Faltings, 2005a], MB-DPOP [Petcu and Faltings, 2007] and ODPOP [Petcu and Faltings, 2006] have been pro- posed to improve its scalability. However, DPOP cannot be directly applied to asymmetric settings as it requires the total knowledge of each constraint to perform optimal elimination locally. PT-ISABB [Deng et al., 2019] applies (A)DPOP into solving ADCOPs by performing variable elimination only to a subset of constraints to build look-up tables for lower bounds, and uses a tree-based search to guarantee the optimality. The algorithm further reinforces the bounds by a non-local elimination scheme. That is, in- stead of performing elimination locally, the elimination of a variable is postponed to its parent to include the private func- tion enforced in the parent's side and increase the integrity of the utility table. 3 Asymmetric DPOP The existing complete algorithms for ADCOPs use complete search to exhaust the search space, which makes them un- suitable for large scale applications. In fact, as shown in our experimental results, these search-based algorithms can only solve the problems with the agent number less than 20. Hence, in this section, we propose a complete, privacy- protecting, and scalable inference-based algorithm for AD- COPs built upon generalized non-local elimination, called AsymDPOP. An execution example can be found in the ap- pendix (https://arxiv.org/abs/1905.11828). 3.1 Utility Propagation Phase In DPOP, a variable could be eliminated locally without loss of completeness after receiving all the utility messages from its children since all functions that involve the variable have been aggregated. However, the conclusion does not hold for ADCOPs. Taking Fig. 2 as an example, x4 cannot be elim- inated locally since the private functions f14 and f24 are not given. Thus, the local elimination to x4 w.r.t. f41 and f42 would lead to overestimate bias and offer no guarantee on the completeness. A naıve solution would be that x1 and x2 Algorithm1:GNLEforaiWhenInitialization:1utili←⊗aj∈AP(ai)fij2ifaiisaleafthen3sendUTIL(utili)toP(ai)WhenreceivedUTIL(utilc)fromac∈C(ai):4utilci←utilc5foreachaj∈(PC(ai)∩Desc(ac))∪{ac}do6utilci←utilci⊗fij7utili←utili⊗minEV(ai,ac)utilci8ifaihasreceivedallUTILfromC(ai)then9ifaiistherootthen10startValuePropagationphase11else12sendUTIL(utili)toP(ai)4 Tradeoffs As shown in Section 3.3, AsymDPOP suffers serious scal- ability problems in both memory and computation. In this section, we propose two tradeoffs which make AsymDPOP a practical algorithm. 4.1 Table-Set Propagation Scheme: A Tradeoff between Memory and Privacy The utility propagation phase of AsymDPOP could be prob- lematic due to the unacceptable memory consumption when the pseudo tree is poor. Consider the pseudo tree shown in Fig. 3. Since every agent is constrained with the root agent, according to the GNLE all the variables can only be elim- inated at the root agent, which incurs a memory consump- tion of O(dn) due to the join operation in each agent. Here, d = maxi Di. Besides, a large utility table would also incur unacceptable computation overheads due to the join opera- tions and the elimination operation (line 5-7). We notice that utility tables are divisible before performing eliminations. Thus, instead of propagating a joint and high- dimension utility table to a parent agent, we propagate a set of small utility tables. In other words, we wish to reduce the unnecessary join operations (i.e., line 1 and line 7 in the case of EV (ai, ac) = ∅) which could cause a dimension increase during the utility propagation phase. On the other side, com- pletely discarding join operations would incur privacy leak- ages. For example, if an chooses to propagate both fn,n−1 and fn1 without performing the join operation to an−1, then an−1 would know the private functions of an directly. Thus, we propose to compromise the memory consumption and the privacy by a parameter kp controlling the the maximal num- ber of dimensions of each local utility table. We refer the tradeoff as a table-set propagation scheme (TSPS). Specifically, when an agent sends a UTIL message to its parent, it first joins its private functions w.r.t. its parents with the received utility tables whose dimensions contain the di- mensions of these private functions. Notice that the first step does not incur a dimension increase, and can reduce the num- ber of utility tables. Finally, it propagates the set of utility tables to its parent. n−1} to an−2. Consider the pseudo tree shown in Fig. 3 again. Assume that kp = 3 and then agent an would propagate the utility set utiln = {fn,n−1 ⊗ fn1} to an−1. Since there is no elim- ination in an−1, it is unnecessary to perform the join oper- ation in line 7. Thus, an−1 would propagate the utility set utiln−1 = {fn−1,n−2 ⊗ fn−1,1, utiln It can be concluded that TSPS in the example only requires O(nd3) space, which is much smaller than the one required by GNLE. Formally, we have the following theorem. Theorem 2. The size of each utility table of an agent ai in TSPS is no greater than max(min(AP (ai),kp), max d Proof. According to Theorem 1, the dimension of each util- ity table from child ac ∈ C(ai) is a subset of Sep(ac) ∪ ID(ac) ∪ {ac}. Since TSPS omits the join operation in line 7, the maximal size of received utility tables of ai is (ID(ai)∩(ID(ac)∪{ac})+Sep(ac)) (ID(ai)∩(ID(ac)∪{ac})+Sep(ac)) ac∈C(ai) max ac∈C(ai) d Figure 3: A chain-like pseudo tree variable is eliminated at its highest (pseudo) parent, the par- ent is responsible for selecting the optimal assignment for that variable. Thus, the value messages in our algorithm would contain not only the assignments for ancestors, but also as- signments for descendants. Algorithm 2 presents the sketch of value propagation phase. The phase is initiated by the root agent selecting the opti- mal assignment (line 14). Given the determined assignments either from its parent (line 16) or computed locally (line 15), agent ai selects the optimal assignments for the eliminated variables in each branch ac ∈ C(ai) by a joint optimization over them (line 17-20), and propagates the assignments to- gether with the determined assignments to ac (line 21-22). The algorithm terminates when each leaf agent receives a VALUE message. 3.3 Complexity Analysis Theorem 1. The size of a UTIL message produced by an agent is exponential in the number of its separator and its interface descendants. Proof. We prove the theorem by showing a UTIL message produced by agent ai contains the dimensions of Sep(ai) and ID(ai) ∪ {ai}. The UTIL message must contain the dimen- sions of ID(ai) ∪ {ai} since ai is not the highest (pseudo) parent of aj,∀aj ∈ ID(ai) ∪ {ai}. On the other hand, ac- cording to the definition to ID(ai), the UTIL message can- not contain the dimensions of Desc(ai)\ID(ai) since for each aj ∈ Desc(ai)\ID(ai) it must exist an agent ak ∈ Desc(ai) ∪ {ai} such that ak is the highest (pseudo) parent of aj and thus the variable is eliminated at ak (line 7). Finally, the UTIL message contains Sep(ai) according to [Petcu and Faltings, 2005b] (Theorem 1). Thus, the size of the UTIL message is exponential in Sep(ai) + ID(ai) + 1 and the theorem is concluded. Algorithm2:ValuePropagationforaiWhenInitialization:13ifaiistherootthen14v∗i←argminxiutili15PropagateValue({(xi=v∗i)})WhenreceivedVALUE(Assign)fromP(ai):16PropagateValue(Assign)FunctionPropagateValue(Assign):17foreachac∈C(ai)do18Assignci←Assign[Sep(ac)∪{ac}∪ID(ac)]19ifEV(ai,ac)(cid:54)=∅then20V∗←argminEV(ai,ac)utilci(Assign[dims(utilci)])21Assignci←Assignci∪{(xj=V∗[xj])∀xj∈EV(ai,ac)}22sendVALUE(Assignci)toac Figure 4: A set of utility tables Since the local utility is partitioned into utility tables accord- ing to kp, the maximal size of local utility tables of ai is dmin(AP (ai),kp) Thus the theorem holds. 4.2 Mini-batch Elimination Scheme: A Tradeoff between Time and Space TSPS could factorize a big utility table to a set of smaller util- ity tables, which allows us to reduce the computational efforts when performing eliminations by a decrease-and-conquer strategy. Taking Fig. 3 as an example, to perform the elimi- nation, a1 in GNLE has no choice but to optimize a big utility table over variables x2, . . . xn (line 7), which requires O(dn) operations. Instead, combining with TSPS (kp = 2) we could exploit the the structure of each small utility table by arrang- ing the min operators among them to reduce computational complexity. That is, instead of performing min x2,...xn u1,2,...,n we perform xn (f12 +f21 +···+min (f1n +fn1 +fn,n−1 +fn−1,n)··· ) min x2 which can be solved recursively from xn to x2 and the overall complexity is O(nd3). In other words, we reduce the com- putational complexity by exploiting the independence among utility tables to avoid unnecessary traverses. However, completely distributing the min operators into every variable would incur high memory consumption as a min operator could implicitly join utility tables to a big and indivisible table. Although the problem can be alleviated by carefully arranging the min operators, it is not easy to find the optimal sequence of eliminations in practice. Consider the utility tables shown as a factor graph in Fig. 4 where square nodes represent utility tables and circle nodes represent vari- ables. And the red circles represent the variables to be elim- inated. Obviously, no matter how to arrange the elimination sequence, a 3-ary utility table must appear when eliminating x1 or x2. Instead, if we jointly optimize both x1 and x2, the maximal number of dimensions are 2. We thus overcome the problem by introducing a parame- ter ke which specifies the minimal number of variables op- timized in a min operator (i.e., the size of a batch), and re- fer the tradeoff as a mini-batch elimination scheme (MBES). Specifically, when performing an elimination operation, we first divide elimination variables into several groups whose variables share at least a common utility table. For each vari- able group, we divide the variables into batches whose sizes are at least ke if it is possible. For each batch, we perform (a) network load (b) NCLOs Figure 5: Performance comparison under different agent numbers optimization to the functions that are related to the batch and replace these functions with the results. The process termi- nates when all the variable groups are exhausted. Note that dividing variables into disjoint variable groups in the first step is crucial since optimizing independent variables jointly is equivalent to optimizing them individually. Taking Fig. 4 for example, if we set ke = 2 and let x2, x5 be a batch, a 4-ary utility table over x1, x3, x4 and x6 still appear even if x2 and x5 is jointly optimized. 5 Experimental Results We empirically evaluate the performances of AsymDPOP and state-of-the-art complete algorithms for ADCOPs including PT-ISABB, AsymPT-FB, SyncABB, ATWB and the vanilla DPOP with PEAV formulation (PEAV DPOP) in terms of the number of basic operations, network load and privacy leak- age. For inference-based algorithms, we consider the max- imal number of dimensions during the utility propagation phase as an additional metric. Specifically, we use NCLOs [Netzer et al., 2012] to measure the hardware-independent runtime in which the logic operations in inference-based al- gorithms are accesses to the utility tables while the ones in 812162024# of Agents100101102103104105Network Load (KB)812162024# of Agents102104106108NCLOs(cid:1)(cid:2)(cid:3)(cid:4)(cid:5)(cid:6)(cid:7)(cid:8)(cid:1)(cid:4)(cid:4)(cid:1)(cid:9)(cid:6)(cid:10)(cid:11)(cid:2)(cid:12)(cid:13)(cid:4)(cid:11)(cid:2)(cid:12)(cid:14)(cid:5)(cid:1)(cid:4)(cid:4)(cid:11)(cid:15)(cid:1)(cid:16)(cid:17)(cid:11)(cid:18)(cid:11)(cid:1)(cid:9)(cid:6)(cid:10)(cid:17)(cid:11)(cid:18)(cid:11)(cid:19)(cid:1)(cid:2)(cid:20)(cid:3)(cid:1)(cid:4)(cid:1)(cid:5)(cid:20)(cid:1)(cid:21)(cid:1)(cid:9)(cid:6)(cid:10)(cid:17)(cid:11)(cid:18)(cid:11)(cid:19)(cid:1)(cid:2)(cid:20)(cid:22)(cid:4)(cid:1)(cid:5)(cid:20)(cid:1)(cid:21)(cid:1)(cid:9)(cid:6)(cid:10)(cid:17)(cid:11)(cid:18)(cid:11)(cid:19)(cid:1)(cid:2)(cid:20)(cid:23)(cid:4)(cid:1)(cid:5)(cid:20)(cid:1)(cid:21)(cid:1)(cid:9)(cid:6)(cid:10)(cid:17)(cid:11)(cid:18)(cid:11)(cid:19)(cid:1)(cid:2)(cid:20)(cid:22)(cid:4)(cid:1)(cid:5)(cid:20)(cid:24)(cid:21)(cid:1)(cid:9)(cid:6)(cid:10)(cid:17)(cid:11)(cid:18)(cid:11)(cid:19)(cid:1)(cid:2)(cid:20)(cid:23)(cid:4)(cid:1)(cid:5)(cid:20)(cid:24)(cid:21)(a) network load (a) maximum dimensions (b) NCLOs (b) NCLOs Figure 6: Performance comparison under different densities Figure 7: Performance comparison of different batch sizes search-based algorithms are constraint checks. For the net- work load, we measure the size of information during an ex- ecution. Finally, we use entropy [Brito et al., 2009] to quan- tify the privacy loss [Litov and Meisels, 2017; Grinshpoun et al., 2013]. For each experiment, we generate 50 random instances and report the medians as the results. In our first experiment, we consider the ADCOPs with the domain size of 3, the density of 0.25 and the agent number varying from 8 to 24. Fig. 5 presents the experimental results under different agent numbers. It can be concluded that com- pared to the search-based solvers, our AsymDPOP algorithms exhibit great superiorities in terms of network load. That is due to the fact that search-based algorithms explicitly exhaust the search space by message-passing, which is quite expen- sive especially when the agent number is large. In contrast, our proposed algorithms incur few communication overheads since they follow an inference protocol and only require a linear number of messages. On the other hand, although PEAV DPOP also uses the inference protocol, it still suffers from a severe scalability problem and can only solve the prob- lems with the agent number less than 12. The phenomenon is due to the mirror variables introduced by PEAV formulation, which significantly increases the complexity. More specifi- cally, a UTIL message in PEAV DPOP contains the dimen- sions of mirror variables, which significantly increases the memory consumption. Fig. 6 presents the performance comparison under differ- ent densities. Specifically, we consider the ADCOPs with the agent number of 8, the domain size of 8 and the den- sity varying from 0.25 to 1. It can be concluded from Fig. 6(a) that AsymDPOP with TSPS (i.e., kp < w∗ where w∗ is the induced width) incurs significantly less communica- tion overheads when solving dense problems, which demon- strates the merit of avoiding unnecessary join operations. Be- sides, it is interesting to find that compared to the one of AsymDPOP without TSPS (i.e., kp = w∗), the network load of AsymDPOP(kp < w∗,·) increases much slowly. That is due to the fact that as the density increases, eliminations are more likely to happen at the top of a pseudo tree. On the other hand, since unnecessary join operations are avoided in TSPS, eliminations are the major source of the dimension in- crease. As a result, agents propagate low dimension utility tables in most of the time. For NCLOs, search-based algo- rithms like AsymPT-FB and PT-ISABB outperform AsymD- POP algorithms when solving dense problems due to their effective pruning. 0.250.40.550.70.851Density101102103104105Network Load (KB)0.250.40.550.70.851Density103104105106107108NCLOs(cid:1)(cid:2)(cid:3)(cid:4)(cid:5)(cid:6)(cid:7)(cid:8)(cid:1)(cid:4)(cid:4)(cid:1)(cid:9)(cid:6)(cid:10)(cid:11)(cid:2)(cid:12)(cid:13)(cid:4)(cid:11)(cid:2)(cid:12)(cid:14)(cid:5)(cid:1)(cid:4)(cid:4)(cid:11)(cid:15)(cid:1)(cid:16)(cid:17)(cid:11)(cid:18)(cid:11)(cid:1)(cid:9)(cid:6)(cid:10)(cid:17)(cid:11)(cid:18)(cid:11)(cid:19)(cid:1)(cid:2)(cid:20)(cid:3)(cid:1)(cid:4)(cid:1)(cid:5)(cid:20)(cid:1)(cid:21)(cid:1)(cid:9)(cid:6)(cid:10)(cid:17)(cid:11)(cid:18)(cid:11)(cid:19)(cid:1)(cid:2)(cid:20)(cid:22)(cid:4)(cid:1)(cid:5)(cid:20)(cid:1)(cid:21)(cid:1)(cid:9)(cid:6)(cid:10)(cid:17)(cid:11)(cid:18)(cid:11)(cid:19)(cid:1)(cid:2)(cid:20)(cid:23)(cid:4)(cid:1)(cid:5)(cid:20)(cid:1)(cid:21)(cid:1)(cid:9)(cid:6)(cid:10)(cid:17)(cid:11)(cid:18)(cid:11)(cid:19)(cid:1)(cid:2)(cid:20)(cid:22)(cid:4)(cid:1)(cid:5)(cid:20)(cid:24)(cid:21)(cid:1)(cid:9)(cid:6)(cid:10)(cid:17)(cid:11)(cid:18)(cid:11)(cid:19)(cid:1)(cid:2)(cid:20)(cid:23)(cid:4)(cid:1)(cid:5)(cid:20)(cid:24)(cid:21)812162024# of Agents051015Maximum Dimension(cid:1)(cid:2)(cid:3)(cid:4)(cid:5)(cid:1)(cid:6)(cid:1)(cid:3)(cid:7)(cid:8)(cid:9)(cid:5)(cid:1)(cid:6)(cid:1)(cid:10)(cid:1)(cid:2)(cid:11)(cid:12)(cid:3)(cid:1)(cid:4)(cid:11)(cid:13)(cid:14)(cid:3)(cid:7)(cid:8)(cid:9)(cid:5)(cid:1)(cid:6)(cid:1)(cid:10)(cid:1)(cid:2)(cid:11)(cid:12)(cid:3)(cid:1)(cid:4)(cid:11)(cid:12)(cid:14)(cid:3)(cid:7)(cid:8)(cid:9)(cid:5)(cid:1)(cid:6)(cid:1)(cid:10)(cid:1)(cid:2)(cid:11)(cid:12)(cid:3)(cid:1)(cid:4)(cid:11)(cid:1)(cid:14)(cid:3)(cid:7)(cid:8)(cid:9)(cid:5)(cid:1)(cid:6)(cid:1)(cid:10)(cid:1)(cid:2)(cid:11)(cid:15)(cid:3)(cid:1)(cid:4)(cid:11)(cid:13)(cid:14)(cid:3)(cid:7)(cid:8)(cid:9)(cid:5)(cid:1)(cid:6)(cid:1)(cid:10)(cid:1)(cid:2)(cid:11)(cid:15)(cid:3)(cid:1)(cid:4)(cid:11)(cid:12)(cid:14)(cid:3)(cid:7)(cid:8)(cid:9)(cid:5)(cid:1)(cid:6)(cid:1)(cid:10)(cid:1)(cid:2)(cid:11)(cid:15)(cid:3)(cid:1)(cid:4)(cid:11)(cid:1)(cid:14)(cid:3)(cid:7)(cid:8)(cid:9)(cid:5)(cid:1)(cid:6)(cid:1)(cid:10)(cid:1)(cid:2)(cid:11)(cid:5)(cid:1)(cid:3)(cid:1)(cid:4)(cid:11)(cid:1)(cid:14)812162024# of Agents102104106108NCLOs(cid:1)(cid:2)(cid:3)(cid:4)(cid:5)(cid:1)(cid:6)(cid:1)(cid:3)(cid:7)(cid:8)(cid:9)(cid:5)(cid:1)(cid:6)(cid:1)(cid:10)(cid:1)(cid:2)(cid:11)(cid:12)(cid:3)(cid:1)(cid:4)(cid:11)(cid:13)(cid:14)(cid:3)(cid:7)(cid:8)(cid:9)(cid:5)(cid:1)(cid:6)(cid:1)(cid:10)(cid:1)(cid:2)(cid:11)(cid:12)(cid:3)(cid:1)(cid:4)(cid:11)(cid:12)(cid:14)(cid:3)(cid:7)(cid:8)(cid:9)(cid:5)(cid:1)(cid:6)(cid:1)(cid:10)(cid:1)(cid:2)(cid:11)(cid:12)(cid:3)(cid:1)(cid:4)(cid:11)(cid:1)(cid:14)(cid:3)(cid:7)(cid:8)(cid:9)(cid:5)(cid:1)(cid:6)(cid:1)(cid:10)(cid:1)(cid:2)(cid:11)(cid:15)(cid:3)(cid:1)(cid:4)(cid:11)(cid:13)(cid:14)(cid:3)(cid:7)(cid:8)(cid:9)(cid:5)(cid:1)(cid:6)(cid:1)(cid:10)(cid:1)(cid:2)(cid:11)(cid:15)(cid:3)(cid:1)(cid:4)(cid:11)(cid:12)(cid:14)(cid:3)(cid:7)(cid:8)(cid:9)(cid:5)(cid:1)(cid:6)(cid:1)(cid:10)(cid:1)(cid:2)(cid:11)(cid:15)(cid:3)(cid:1)(cid:4)(cid:11)(cid:1)(cid:14)(cid:3)(cid:7)(cid:8)(cid:9)(cid:5)(cid:1)(cid:6)(cid:1)(cid:10)(cid:1)(cid:2)(cid:11)(cid:5)(cid:1)(cid:3)(cid:1)(cid:4)(cid:11)(cid:1)(cid:14)three ingredients: generalized non-local elimination which facilitates the aggregation of utility in an asymmetric environ- ment, table-set propagation scheme which reduces the mem- ory consumption and mini-batch elimination scheme which reduces the operations in the utility propagation phase. We theoretically show its complexity and our empirical evalua- tion demonstrates its superiorities over the state-of-the-art, as well as the vanilla DPOP with PEAV formulation. Acknowledgements We would like to thank the anonymous reviewers for their valuable comments and helpful suggestions. This work is supported by the Chongqing Research Pro- gram of Basic Research and Frontier Technology un- der Grant No.:cstc2017jcyjAX0030, Fundamental Research Funds the Central Universities under Grant No.: 2018CDXYJSJ0026, National Natural Science Foundation of China under Grant No.: 51608070 and the Graduate Research and Innovation Foundation of Chongqing, China under Grant No.: CYS17023 for References [Brito et al., 2009] Ismel Brito, Amnon Meisels, Pedro Meseguer, and Roie Zivan. Distributed constraint sat- isfaction with partially known constraints. Constraints, 14(2):199 -- 234, 2009. [Burke et al., 2007] David A Burke, Kenneth N Brown, Mustafa Dogru, and Ben Lowe. Supply chain coordina- In The tion through distributed constraint optimization. 9th International Workshop on DCR, 2007. [Chen et al., 2018] Ziyu Chen, Yanchen Deng, Tengfei Wu, and Zhongshi He. A class of iterative refined max-sum al- gorithms via non-consecutive value propagation strategies. Autonomous Agents and Multi-Agent Systems, 32(6):822 -- 860, 2018. [Dechter, 1999] Rina Dechter. Bucket elimination: A uni- fying framework for reasoning. Artificial Intelligence, 113(1-2):41 -- 85, 1999. [Deng et al., 2019] Yanchen Deng, Ziyu Chen, Dingding Chen, Xingqiong Jiang, and Qiang Li. PT-ISABB: A hybrid tree-based complete algorithm to solve asymmet- ric distributed constraint optimization problems. In Pro- ceedings of 18th International Conference on Autonomous Agents and Multiagent Systems, pages 1506 -- 1514, 2019. [Fioretto et al., 2016] Ferdinando Fioretto, William Yeoh, and Enrico Pontelli. A dynamic programming-based MCMC framework for solving DCOPs with GPUs. In Proceedings of 22nd International Conference on Princi- ples and Practice of Constraint Programming, pages 813 -- 831, 2016. [Fioretto et al., 2017] Ferdinando Fioretto, William Yeoh, Enrico Pontelli, Ye Ma, and Satishkumar J Ranade. A dis- tributed constraint optimization (DCOP) approach to the In Proceed- economic dispatch with demand response. ings of the 16th International Conference on Autonomous Agents and MultiAgent Systems, pages 999 -- 1007, 2017. Figure 8: Privacy loss under different tightness To demonstrate the effects of different batch sizes, we benchmark MBES with different ke when combining with TSPS with different kp on the configuration used in the first experiment. Specifically, we measure the maximal number of dimensions generated in the utility propagation phase, in- cluding the intermediate results of MBES. Fig. 7 presents the experimental results. It can be seen that AsymDPOP with a small batch size significantly reduces NCLOs but produces larger intermediate results, which indicates the necessity of the tradeoff. Besides, the performance of MBES significantly degenerates when combining with TSPS with a larger kp. That is because the utility tables contain more dimensions in the scenario, and a utility table would be traversed more frequently when performing eliminations. In the last experiment, we consider the privacy loss of dif- ferent algorithms when solving asymmetric MaxDCSPs with different tightness. In particular, we consider the asymmetric MaxDCSPs with 10 agents, the domain size of 10, the den- sity of 0.4 and the tightness varying from 0.1 to 0.8. Fig. 8 presents the results. It can be concluded from the figure that as the tightness grows, the search-based algorithms leak more privacy while the inference-based algorithms leaks less privacy. That is due to the fact that search-based algorithms rely on a direct disclosure mechanism to aggregate the private costs. Thus, search-based algorithms would leak more pri- vacy when solving the problems with high tightness as they need to traverse more proportions of the search space. In con- trast, inference-based algorithms accumulate utility through the pseudo tree, and an agent ai could infer the private costs of its (pseudo) child ac when the utility table involving both xi and xc is a binary table which is not a result of eliminations or contains zero entries. Thus, AsymDPOP(kp = 2,·) leaks almost a half of privacy. On the other hand, since the num- ber of prohibit combinations grows as the tightness increases, AsymDPOP(kp ≥ 3,·) incurs much lower privacy loss when solving high tightness problems. 6 Conclusion In this paper we present AsymDPOP, the first complete in- ference algorithm for ADCOPs. The algorithm incorporates 0.10.20.30.40.50.60.70.8Tightness00.10.20.30.40.50.60.7Privacy(cid:1)(cid:2)(cid:3)(cid:4)(cid:5)(cid:6)(cid:7)(cid:8)(cid:1)(cid:4)(cid:4)(cid:1)(cid:9)(cid:6)(cid:10)(cid:11)(cid:2)(cid:12)(cid:13)(cid:4)(cid:11)(cid:2)(cid:12)(cid:14)(cid:5)(cid:1)(cid:4)(cid:4)(cid:1)(cid:9)(cid:6)(cid:10)(cid:15)(cid:11)(cid:16)(cid:11)(cid:17)(cid:1)(cid:2)(cid:18)(cid:3)(cid:1)(cid:4)(cid:1)(cid:5)(cid:18)(cid:1)(cid:19)(cid:1)(cid:9)(cid:6)(cid:10)(cid:15)(cid:11)(cid:16)(cid:11)(cid:17)(cid:1)(cid:2)(cid:18)(cid:20)(cid:4)(cid:1)(cid:5)(cid:18)(cid:1)(cid:19)(cid:1)(cid:9)(cid:6)(cid:10)(cid:15)(cid:11)(cid:16)(cid:11)(cid:17)(cid:1)(cid:2)(cid:18)(cid:21)(cid:4)(cid:1)(cid:5)(cid:18)(cid:1)(cid:19)[Fioretto et al., 2018] Ferdinando Fioretto, Enrico Pontelli, and William Yeoh. Distributed constraint optimization problems and applications: A survey. Journal of Artificial Intelligence Research, 61:623 -- 698, 2018. [Freuder and Quinn, 1985] Eugene and Michael J. Quinn. Taking advantage of stable sets of variables in constraint satisfaction problems. In Proceedings of the 9th International Joint Conference on Artificial Intelligence, pages 1076 -- 1078, 1985. Freuder C. [Gershman et al., 2009] Amir Gershman, Amnon Meisels, and Roie Zivan. Asynchronous forward bounding for dis- tributed COPs. Journal of Artificial Intelligence Research, 34:61 -- 88, 2009. [Grinshpoun et al., 2013] Tal Grinshpoun, Alon Grubshtein, Roie Zivan, Arnon Netzer, and Amnon Meisels. Asym- metric distributed constraint optimization problems. Jour- nal of Artificial Intelligence Research, 47:613 -- 647, 2013. and Makoto Yokoo. Distributed partial constraint satisfaction problem. In International Conference on Principles and Practice of Constraint Programming, pages 222 -- 236, 1997. [Hirayama and Yokoo, 1997] Katsutoshi Hirayama [Li et al., 2016] Shijie Li, Rudy R. Negenborn, and Gabriel Lodewijks. Distributed constraint optimization for ad- dressing vessel rotation planning problems. Engineering Applications of Artificial Intelligence, 48:159 -- 172, 2016. [Litov and Meisels, 2017] Omer Litov and Amnon Meisels. Forward bounding on pseudo-trees for DCOPs and AD- COPs. Artificial Intelligence, 252:83 -- 99, 2017. [Maheswaran et al., 2004a] Rajiv Maheswaran, Distributed Jonathan P Pearce, and Milind Tambe. algorithms for DCOP: A Graphical-Game-Based ap- In Proceedings of ISCA PDCS, pages 432 -- 439, proach. 2004. T [Maheswaran et al., 2004b] Rajiv T. Maheswaran, Milind Tambe, Emma Bowring, Jonathan P. Pearce, and Pradeep Varakantham. Taking DCOP to the real world: Efficient complete solutions for distributed multi-event scheduling. In Proceedings of the 3rd International Conference on Au- tonomous Agents and Multiagent Systems, pages 310 -- 317, 2004. [Modi et al., 2005] Pragnesh Jay Modi, Wei-Min Shen, Milind Tambe, and Makoto Yokoo. Adopt: asynchronous distributed constraint optimization with quality guaran- tees. Artificial Intelligence, 161(1-2):149 -- 180, 2005. [Monteiro et al., 2012] Tania L Monteiro, Guy Pujolle, Marcelo E Pellenz, Manoel C Penna, and Richard Demo Souza. A multi-agent approach to optimal channel assign- ment in WLANs. In Wireless Communications and Net- working Conference, pages 2637 -- 2642, 2012. [Netzer et al., 2012] Arnon Netzer, Alon Grubshtein, and Amnon Meisels. Concurrent forward bounding for dis- tributed constraint optimization problems. Artificial Intel- ligence, 193:186 -- 216, 2012. [Okamoto et al., 2016] Steven Okamoto, Roie Zivan, and Aviv Nahon. Distributed breakout: Beyond satisfaction. In Proceedings of the 25th International Joint Conference on Artificial Intelligence, pages 447 -- 453, 2016. [Ottens et al., 2017] Brammert Ottens, Christos Dimi- trakakis, and Boi Faltings. DUCT: An upper confidence bound approach to distributed constraint optimization problems. ACM Transactions on Intelligent Systems and Technology, 8(5):69, 2017. [Petcu and Faltings, 2005a] Adrian Petcu and Boi Faltings. In Interna- Approximations in distributed optimization. tional conference on principles and practice of constraint programming, pages 802 -- 806, 2005. [Petcu and Faltings, 2005b] Adrian Petcu and Boi Faltings. A scalable method for multiagent constraint optimization. In Proceedings of the 19th International Joint Conference on Artificial Intelligence, pages 266 -- 271, 2005. [Petcu and Faltings, 2006] Adrian Petcu and Boi Faltings. ODPOP: An algorithm for open/distributed constraint op- timization. In Proceedings of the 21st national conference on Artificial intelligence, pages 703 -- 708, 2006. [Petcu and Faltings, 2007] Adrian Petcu and Boi Faltings. MB-DPOP: A new memory-bounded algorithm for dis- In Proceedings of the 20th Inter- tributed optimization. national Joint Conference on Artificial Intelligence, pages 1452 -- 1457, 2007. [Ramchurn et al., 2011] Sarvapali D Ramchurn, Perukrish- nen Vytelingum, Alex Rogers, and Nick Jennings. Agent- based control for decentralised demand side management in the smart grid. In The 10th International Conference on Autonomous Agents and Multiagent Systems, pages 5 -- 12, 2011. [Rogers et al., 2011] Alex Rogers, Alessandro Farinelli, Ruben Stranders, and Nicholas R Jennings. Bounded ap- proximate decentralised coordination via the max-sum al- gorithm. Artificial Intelligence, 175(2):730 -- 759, 2011. [Vinyals et al., 2011] Meritxell Vinyals, Juan A Rodriguez- Aguilar, and Jes´us Cerquides. Constructing a unifying theory of dynamic programming DCOP algorithms via the generalized distributive law. Autonomous Agents and Multi-Agent Systems, 22(3):439 -- 464, 2011. [Yeoh et al., 2010] William Yeoh, Ariel Felner, and Sven Koenig. BnB-ADOPT: An asynchronous branch-and- bound DCOP algorithm. Journal of Artificial Intelligence Research, 38:85 -- 133, 2010. [Zhang et al., 2005] Weixiong Zhang, Guandong Wang, Zhao Xing, and Lars Wittenburg. Distributed stochastic search and distributed breakout: properties, comparison and applications to constraint optimization problems in sensor networks. Artificial Intelligence, 161(1-2):55 -- 87, 2005. [Zivan and Peled, 2012] Roie Zivan and Hilla Peled. Max/min-sum distributed constraint optimization through value propagation on an alternating DAG. In Proceedings of the 11th International Conference on Autonomous Agents and Multiagent Systems, pages 265 -- 272, 2012. A AsymDPOP A.1 An Example for AsymDPOP Fig.2 presents a pseudo tree. For better understanding, we take the agent a2 to explain the concepts in a pseudo tree. Since a1 is the only ancestor constrained with a2 via a tree edge, we have P (a2) = {a1}, P P (a2) = ∅ and Sep(a2) = {a1}. Similarly, since a3 and a4 are the descendants con- strained with a2 via tree edge, we have C(a2) = {a3, a4}, P C(a2) = ∅, Desc(a2) = {a3, a4}. Particularly, since a4 ∈ Desc(a2) is constrained with a1 ∈ Sep(a2), we have ID(a2) = {a4}. AsymDPOP begins with the leaf agents (a3 and a4) send- ing their utility tables (util3 and util4) to their parent a2, where util3 = f32 and util4 = f41 ⊗ f42. When a2 receives the UTIL message from its child a3 (assume a3's UTIL message has arrived earlier), it stores the received utility table (i.e., util3 2 = util3 = f32), and joins its private function f23 to update the table util3 2 (i.e., 2 = f32⊗f23). According to the definition of EV (ai, ac) util3 that all the belonging variables' highest (pseudo) parent is ai in branch ac. Here, EV (a2, a3) is given by EV (a2, a3) = P C(a2) ∩ Desc(a3) ∪ {a3}\ID(a2) = ∅ ∩ ∅ ∪ {a3}\{a4} = {a3} Thus, a2 eliminates variable x3 from the utility table util3 2. After that, a2 joins the eliminated result to the running utility table util2 (util2 has been initialized to f21). Similarly, upon receipt of the UTIL message from a4, a2 saves the received 2 = util4 = f41 ⊗ f42), then updates utility table (i.e., util4 the utility table util4 2 by joining the private function f24 (i.e., 2 = f41 ⊗ f42 ⊗ f24). Since EV (a2, a4) = ∅, a2 joins util4 2 to the running utility table util2 directly without any util4 elimination. Since a2 has received all the UTIL messages from its children, it propagates the utility table util2 to its parent a1. Here, we have util2 = f21 ⊗ (f41 ⊗ f42 ⊗ f24) ⊗ (min f23 ⊗ f32) x3 When a1 receives the UTIL message from a2, it saves the received utility table (i.e., util2 1 = util2) and joins its private 1 ⊗ functions f12 and f14 to update util2 (f12 ⊗ f14)). After eliminating x2 and x4 (EV (a1, a2) = {a2, a4}), a1 joins the eliminated result into its running utility table util1 (util1 has been initialized to null). Here, we have util1 = min x2,x4 f12⊗f14⊗(f21⊗(f41⊗f42⊗f24)⊗(min 1 (i.e., util2 1 = util2 x3 f23⊗f32)) Since a1 is the root agent and receives all the UTIL messages from its children, it selects the optimal assignment v∗ 1 for it- self and the optimal assignments v∗ 4 for the eliminated variables a2 and a4. That is, 2 and v∗ v∗ 1 = argmin (v∗ 2, v∗ x1 4) = argmin x2,x4 util1, 1(x1 = v∗ 1) util2 Then a1 propagates a VALUE message including the optimal assignment to its child a2, where the optimal assignment is {(x1 = v∗ 1), (x2 = v∗ 2), (x3 = v∗ 3)}. Figure 9: The sketch of AsymDPOP with TSPS and MBES (mes- sage passing) When receiving the VALUE message from a1, a2 assigns 2 for itself. Since EV (a2, a3) = {a3} and the value v∗ EV (a2, a4) = ∅, a2 selects the optimal assignment for a3 by performing optimization over util3 2 with the determined assignment of x2, that is v∗ 3 = argmin (util3 2(x2 = v∗ 2)) x3 3)} and {(x1 = v∗ Then a2 propagates the optimal assignment {(x2 = 4)} to v∗ 2), (x3 = v∗ a3 and a4, respectively. Once a3 and a4 receives the VALUE messages, they assign for themselves. Since all leaf agents have received VALUE messages, the algorithm terminates. 1), (x2 = v∗ 2), (x4 = v∗ B AsymDPOP with TSPS and MBES B.1 Pseudo Code for AsymDPOP with TSPS and MBES Fig.11 and Fig.12 present the sketch of AsymDPOP with TSPS and MBES which consists of two phases: utility set propagation phase and value propagation phase. Different from the utility propagation phase in AsymDPOP, the utility set propagation phase applies the TSPS to trade off between memory and privacy through propagating multiple small util- ity tables, and utilizes the MBES to trade off between time and space by sequential optimizations. The utility set propagation phase begins with leaf agents sending their utility tables to their parents via UTILSET mes- sages (line 2-3). Note that, the utility tables are obtained from the Function PartitionF. That is, leaf agents partition their lo- cal utilities into multiple smaller utility tables according to kp (line 32-41). If there is any residue, ai adds them to the utility tables (line 42-43). When receiving the UTILSET message from its child i ← ac, ai stores the received utility tables (i.e., utilSetc Algorithm3:AsymDPOPwithTSPSandMBESforai(messagepassing)WhenInitialization:1utilSeti←∅2ifaiisaleafthen3sendUTILSET(PartitionF())toP(ai)WhenreceivedUTILSET(utilSetc)fromac∈C(ai):4utilSetci←utilSetc5foreachaj∈(PC(ai)∩Desc(ac))∪{ac}do6if∃u∈utilSetci,s.t.dims(fij)⊂dims(u)then7u←u⊗fij8utilSeti←utilSeti∪Elimination(utilSetci,EV(ai,ac))9ifaihasreceivedallUTILSETfromC(ai)then10ifaiistherootthen11v∗i←argminxi(⊗u∈utilSetiu)12PropagateValue({(xi=v∗i)})13else14utilSeti←utilSeti∪PartitionF()15while∃u,u(cid:48)∈utilSeti,s.t.dims(u(cid:48))⊂dims(u)do16u←u⊗u(cid:48),utilSeti←utilSeti\{u(cid:48)}17sendUTILSET(utilSeti)toP(ai)WhenreceivedVALUE(Assign)fromP(ai):18PropagateValue(Assign)Figure 11: A chain-like pseudo tree utilSetc), then joins its private functions w.r.t. its (pseudo) children in branch ac into the utility tables utilSetc i (line 4-7). Note that the join operation of its private functions w.r.t. its children does not increase the number of dimen- sions and should be applied accordingly to the related util- ity tables. Then the Function Elimination are used to im- plement sequential eliminations on the utility tables utilSetc i In the Function Elimination, with MBES (line 8, 26-30). variables in EV (ai, ac) are firstly divided into several groups (GEV ) such that the variables in each group share at least one common utility table (line 26). And then, for the variables in each group, ai partitions them into several sets (or batches) with the Function PartitionEV according to ke (line 27, 49- 58). After obtaining the eliminated variable sets (EV Set), ai traverses the sets to optimize the utility tables (utilSet) (line 28-30). In detail, for each set, MBES performs optimization to the utility tables that are related to the eliminated variable in the set and replaces these utility tables with the optimized utility table. When receiving all the UTILSET messages from its chil- dren, ai adds its local utility tables computed by the Function FunctionF into the optimized utility tables, then sends these utility tables to its parent if it is not the root agent (line 14-17). Otherwise, the value propagation phase starts (line 11-12). The value propagation phase is roughly the same as the one in AsymDPOP. The root agent ai selects the optimal as- signments for itself and the eliminated variables (EV (ai, ac)) belong to each branch ac ∈ C(ai). And since the utility set propagation uses the TSPS, there may have more than one utility table. So before selecting the optimal assignments for branch ac (line 23), ai joins all the tables in utilSetc i with the assignments received from its parent (line 18) or computed locally (line 11-12) . After that, ai sends a VALUE message to its child ac (line 25). The algorithm terminates when each leaf agent receives a VALUE message. B.2 An Example of AsymDPOP with TSPS and MBES Figure 10: The sketch of AsymDPOP with TSPS and MBES (auxil- iary functions) For the convenience of further explanation, we denote the joint utility table uijk as a function of three variables xi, xj and xk. And the index of the function is sorted in alphabetical order. We take Fig.13 as an example to demonstrate the Algo- rithm 3. And suppose that kp = 3 and ke = 1. The algo- rithm begins with the leaf agent a5 by sending a UTILSET message with a utility table u245 to its parent a4. Here, u245 is computed by Function PartitionF. Specifically, a5 groups its private functions related with AP (a5) into a function set {f52, f54}, since the dimension of its local utility table is Algorithm3:AsymDPOPwithTSPSandMBESforai(auxiliaryfunctions)FunctionPropagateValue(Assign):19foreachac∈C(ai)do20Assignci←Assign[Sep(ac)∪{ac}∪ID(ac)]21ifEV(ai,ac)(cid:54)=∅then22u←⊗u(cid:48)∈utilSetciu(cid:48)(Assign[dims(u(cid:48))])23V∗←argminEV(ai,ac)(u)24Assignci←Assignci∪{(xj=V∗[xj])∀xj∈EV(ai,ac)}25sendVALUE(Assignci)toacFunctionElimination(utilSet,EV):26U←utilSet,GEV←GroupEV(EV,utilSet)27EVSet←∪EV(cid:48)∈GEVPartitionEV(EV(cid:48))28foreachEV(cid:48)∈EVSetdo29U(cid:48)←{u∀u∈U,dims(u)∩EV(cid:48)(cid:54)=∅}30U←(U\U(cid:48))∪{minEV(cid:48)⊗u∈U(cid:48)f}31returnUFunctionPartitionF():32U←∅,u←null33orderAP(ai)accordingtotheirlevels34foreachaj∈AP(ai)do35if∃u(cid:48)∈utilSeti,s.t.dims(fij)⊂dims(cid:0)u(cid:48)(cid:1)then36u(cid:48)←u(cid:48)⊗fij37else38ifdims(u)≥kpthen39U←U∪{u},u←fij40else41u←u⊗fij42ifu(cid:54)=∅then43U←U∪{u}44returnUFunctionGroupEV(EV,utilSet):45GEV←{dims(u)∩EV∀u∈utilSet}46while∃EV(cid:48),EV(cid:48)(cid:48)∈GEV,s.t.EV(cid:48)∩EV(cid:48)(cid:48)(cid:54)=∅do47EV(cid:48)←EV(cid:48)∪EV(cid:48)(cid:48),GEV←GEV\EV(cid:48)(cid:48)48returnGEVFunctionPartitionEV(EV):49EVSet←∅,EV(cid:48)←∅50foreachev∈EVdo51ifEV(cid:48)≥kethen52EVSet←EVSet∪{EV(cid:48)},EV(cid:48)←{ev}53else54EV(cid:48)←EV(cid:48)∪{ev}55ifEV(cid:48)(cid:54)=∅then56randomselectanelementEV(cid:48)(cid:48)∈EVSet57EV(cid:48)(cid:48)←EV(cid:48)(cid:48)∪EV(cid:48)58returnEVSet 5)} to a3. Upon receipt of the VALUE message from a2, v∗ a3 assigns for itself and sends the assignments received from a2 to its child a4. And a4 performs just like a3. The algo- rithm terminates when the leaf agent a5 receives the VALUE message and assigns for itself. C Experiment Results C.1 The Experimental Configuration The Experiment with Different Agent Numbers • Problem type: Random ADCOPs • Agent numbers: [8, 24] • Density: 0.25 • Domain size: 3 The Experiment with Different Densities • Problem type: Random ADCOPs • Agent numbers: 8 • Density: [0.25, 1.0] • Domain size: 8 The Experiment with Different Domain size • Problem type: Random ADCOPs • Agent numbers: 8 • Density: 0.4 • Domain size: [4, 14] The Experiment with Different Tightness • Problem type: ADCSPs • Agent numbers: 10 • Density: 0.4 • Domain size: 10 • Tightness: [0.1, 0.8] C.2 The Experiment Results three which satisfies kp = 3. Thus, a5 obtains a 3-ary utility table u245 by joining the functions in the function set (i.e., u245 = f54 ⊗ f52). When a4 receives the UTILSET message from its child a5, it saves the utility table (u245) and updates the utility table by joining its private functions f45 and f42 (i.e., u245 = u245 ⊗ (f45 ⊗ f42)). Since a4 is not the highest (pseudo) parent of its any descendants (EV (a4, a5) = ∅), it does not need to perform elimination. Further more, since a4 has received all the UTILSET messages from its children, it sends the utility tables {u245, u34} to its parent a3. Here, the utility table u34 is computed by the Function PartitionF through dealing with the residual function f43. Upon receipt of the UTILSET message from a4, similarly, a3 saves the utility tables and updates the utility table u34 by joining its private functions f34 (i.e., u34 = u34 ⊗ f34). Since a3 is not the highest (pseudo) parent of a4 or any other de- scendants, there is no elimination at a3. Because the residual functions f32 and f31 could not be joined into the utility tables {u245, u34} without increase the number of dimensions, a3 deals them with Function PartitionF and gets a utility table u123. After that, it sends the utility tables {u123, u245, u34} to its parent a2. Once a2 receives the UTILSET message from a3, it also saves the utility tables firstly, and joins the relative private functions into the corresponding utility tables (i.e., u245 = u245 ⊗ (f25 ⊗ f24), u123 = u123 ⊗ (f23 ⊗ f21)). Since that is a2 is the highest (pseudo) parent of a4 and a5, EV (a2, a3) = {a4, a5}, the eliminations are preformed by the Function Elimination. In Function Elimination, a2 firstly partitions the eliminated variables x4 and x5 into a group, as they share a common utility table f542. And since ke = 1, a2 divides these variables into two batches and elim- inates them from the utility tables {u123, u245, u34} one by one. After eliminating x5 (assume x5 is eliminated first), a2 gets the utility tables {u123, u24, u34}. Then a2 obtains the new utility tables {u123, u2, u3} by eliminating x4 (i.e., (u34)). After that, a2 updates u2 = min x4 the utility table u321 by a joint operation of the utility tables {u2, u3} (i.e., u123 = u123 ⊗ u2 ⊗ u3), and sends the updated utility table {u123} to its parent a1. (u24), u3 = min x4 When a1 receives the UTILSET message from a2, it saves the utility table, and then joins its private functions f12 and f13 into the utility table u123 (i.e., u123 = u123 ⊗ f12 ⊗ f13). Since EV (a1, a2) = {a2, a3} and the variables x2 and x3 are both relative to the utility table u321, they are partitioned into a single group. But since ke = 1, a1 still eliminates the variables x2 and x3 from the utility table {u123}, respectively, and obtains the eliminated utility table {u1}. Since it is the root agent and has received all the UTILSET messages from its children, a1 chooses the optimal value v∗ 1 for itself based on the utility table {u1}. After that, the value propagation phase starts. a1 selects the 2 and v∗ optimal values v∗ 3 for a2 and a3, then propagates the assignment {(x1 = v∗ 3)} to its child a2. When receiving the VALUE message from a1, a2 selects the optimal assignments for a4 and a5, and sends the assign- ment {(x1 = v∗ 4), (x5 = 2), (x3 = v∗ 1), (x2 = v∗ 1), (x2 = v∗ 2), (x3 = v∗ 3), (x4 = v∗ (a) Network Load (b) NCLOs Figure 12: Performance comparison under different agent numbers (a) Network Load (b) NCLOs Figure 13: Performance comparison under different densities 812162024# of Agents100101102103104105Network Load (KB)812162024# of Agents102104106108NCLOs(cid:1)(cid:2)(cid:3)(cid:4)(cid:5)(cid:6)(cid:7)(cid:8)(cid:1)(cid:4)(cid:4)(cid:1)(cid:9)(cid:6)(cid:10)(cid:11)(cid:2)(cid:12)(cid:13)(cid:4)(cid:11)(cid:2)(cid:12)(cid:14)(cid:5)(cid:1)(cid:4)(cid:4)(cid:11)(cid:15)(cid:1)(cid:16)(cid:17)(cid:11)(cid:18)(cid:11)(cid:1)(cid:9)(cid:6)(cid:10)(cid:17)(cid:11)(cid:18)(cid:11)(cid:19)(cid:1)(cid:2)(cid:20)(cid:3)(cid:1)(cid:4)(cid:1)(cid:5)(cid:20)(cid:1)(cid:21)(cid:1)(cid:9)(cid:6)(cid:10)(cid:17)(cid:11)(cid:18)(cid:11)(cid:19)(cid:1)(cid:2)(cid:20)(cid:22)(cid:4)(cid:1)(cid:5)(cid:20)(cid:1)(cid:21)(cid:1)(cid:9)(cid:6)(cid:10)(cid:17)(cid:11)(cid:18)(cid:11)(cid:19)(cid:1)(cid:2)(cid:20)(cid:23)(cid:4)(cid:1)(cid:5)(cid:20)(cid:1)(cid:21)(cid:1)(cid:9)(cid:6)(cid:10)(cid:17)(cid:11)(cid:18)(cid:11)(cid:19)(cid:1)(cid:2)(cid:20)(cid:22)(cid:4)(cid:1)(cid:5)(cid:20)(cid:24)(cid:21)(cid:1)(cid:9)(cid:6)(cid:10)(cid:17)(cid:11)(cid:18)(cid:11)(cid:19)(cid:1)(cid:2)(cid:20)(cid:23)(cid:4)(cid:1)(cid:5)(cid:20)(cid:24)(cid:21)0.250.40.550.70.851Density101102103104105Network Load (KB)0.250.40.550.70.851Density103104105106107108NCLOs(cid:1)(cid:2)(cid:3)(cid:4)(cid:5)(cid:6)(cid:7)(cid:8)(cid:1)(cid:4)(cid:4)(cid:1)(cid:9)(cid:6)(cid:10)(cid:11)(cid:2)(cid:12)(cid:13)(cid:4)(cid:11)(cid:2)(cid:12)(cid:14)(cid:5)(cid:1)(cid:4)(cid:4)(cid:11)(cid:15)(cid:1)(cid:16)(cid:17)(cid:11)(cid:18)(cid:11)(cid:1)(cid:9)(cid:6)(cid:10)(cid:17)(cid:11)(cid:18)(cid:11)(cid:19)(cid:1)(cid:2)(cid:20)(cid:3)(cid:1)(cid:4)(cid:1)(cid:5)(cid:20)(cid:1)(cid:21)(cid:1)(cid:9)(cid:6)(cid:10)(cid:17)(cid:11)(cid:18)(cid:11)(cid:19)(cid:1)(cid:2)(cid:20)(cid:22)(cid:4)(cid:1)(cid:5)(cid:20)(cid:1)(cid:21)(cid:1)(cid:9)(cid:6)(cid:10)(cid:17)(cid:11)(cid:18)(cid:11)(cid:19)(cid:1)(cid:2)(cid:20)(cid:23)(cid:4)(cid:1)(cid:5)(cid:20)(cid:1)(cid:21)(cid:1)(cid:9)(cid:6)(cid:10)(cid:17)(cid:11)(cid:18)(cid:11)(cid:19)(cid:1)(cid:2)(cid:20)(cid:22)(cid:4)(cid:1)(cid:5)(cid:20)(cid:24)(cid:21)(cid:1)(cid:9)(cid:6)(cid:10)(cid:17)(cid:11)(cid:18)(cid:11)(cid:19)(cid:1)(cid:2)(cid:20)(cid:23)(cid:4)(cid:1)(cid:5)(cid:20)(cid:24)(cid:21)(a) Network Load (b) NCLOs Figure 14: Performance comparison under different domain sizes (a) Maximum Dimension (b) NCLOs Figure 15: Performance comparison under different batch sizes 468101214Domain Size100101102103104105106Network Load (KB)468101214Domain Size103104105106107108NCLOsATWBSyncABBAsymPT-FBPT-ISABBPEAVDPOPAsymDPOP(kp=w$;ke=1)AsymDPOP(kp=2;ke=1)AsymDPOP(kp=3;ke=1)AsymDPOP(kp=2;ke=1)AsymDPOP(kp=3;ke=1)812162024# of Agents051015Maximum Dimension(cid:1)(cid:2)(cid:3)(cid:4)(cid:5)(cid:1)(cid:6)(cid:1)(cid:3)(cid:7)(cid:8)(cid:9)(cid:5)(cid:1)(cid:6)(cid:1)(cid:10)(cid:1)(cid:2)(cid:11)(cid:12)(cid:3)(cid:1)(cid:4)(cid:11)(cid:13)(cid:14)(cid:3)(cid:7)(cid:8)(cid:9)(cid:5)(cid:1)(cid:6)(cid:1)(cid:10)(cid:1)(cid:2)(cid:11)(cid:12)(cid:3)(cid:1)(cid:4)(cid:11)(cid:12)(cid:14)(cid:3)(cid:7)(cid:8)(cid:9)(cid:5)(cid:1)(cid:6)(cid:1)(cid:10)(cid:1)(cid:2)(cid:11)(cid:12)(cid:3)(cid:1)(cid:4)(cid:11)(cid:1)(cid:14)(cid:3)(cid:7)(cid:8)(cid:9)(cid:5)(cid:1)(cid:6)(cid:1)(cid:10)(cid:1)(cid:2)(cid:11)(cid:15)(cid:3)(cid:1)(cid:4)(cid:11)(cid:13)(cid:14)(cid:3)(cid:7)(cid:8)(cid:9)(cid:5)(cid:1)(cid:6)(cid:1)(cid:10)(cid:1)(cid:2)(cid:11)(cid:15)(cid:3)(cid:1)(cid:4)(cid:11)(cid:12)(cid:14)(cid:3)(cid:7)(cid:8)(cid:9)(cid:5)(cid:1)(cid:6)(cid:1)(cid:10)(cid:1)(cid:2)(cid:11)(cid:15)(cid:3)(cid:1)(cid:4)(cid:11)(cid:1)(cid:14)(cid:3)(cid:7)(cid:8)(cid:9)(cid:5)(cid:1)(cid:6)(cid:1)(cid:10)(cid:1)(cid:2)(cid:11)(cid:5)(cid:1)(cid:3)(cid:1)(cid:4)(cid:11)(cid:1)(cid:14)812162024# of Agents102104106108NCLOs(cid:1)(cid:2)(cid:3)(cid:4)(cid:5)(cid:1)(cid:6)(cid:1)(cid:3)(cid:7)(cid:8)(cid:9)(cid:5)(cid:1)(cid:6)(cid:1)(cid:10)(cid:1)(cid:2)(cid:11)(cid:12)(cid:3)(cid:1)(cid:4)(cid:11)(cid:13)(cid:14)(cid:3)(cid:7)(cid:8)(cid:9)(cid:5)(cid:1)(cid:6)(cid:1)(cid:10)(cid:1)(cid:2)(cid:11)(cid:12)(cid:3)(cid:1)(cid:4)(cid:11)(cid:12)(cid:14)(cid:3)(cid:7)(cid:8)(cid:9)(cid:5)(cid:1)(cid:6)(cid:1)(cid:10)(cid:1)(cid:2)(cid:11)(cid:12)(cid:3)(cid:1)(cid:4)(cid:11)(cid:1)(cid:14)(cid:3)(cid:7)(cid:8)(cid:9)(cid:5)(cid:1)(cid:6)(cid:1)(cid:10)(cid:1)(cid:2)(cid:11)(cid:15)(cid:3)(cid:1)(cid:4)(cid:11)(cid:13)(cid:14)(cid:3)(cid:7)(cid:8)(cid:9)(cid:5)(cid:1)(cid:6)(cid:1)(cid:10)(cid:1)(cid:2)(cid:11)(cid:15)(cid:3)(cid:1)(cid:4)(cid:11)(cid:12)(cid:14)(cid:3)(cid:7)(cid:8)(cid:9)(cid:5)(cid:1)(cid:6)(cid:1)(cid:10)(cid:1)(cid:2)(cid:11)(cid:15)(cid:3)(cid:1)(cid:4)(cid:11)(cid:1)(cid:14)(cid:3)(cid:7)(cid:8)(cid:9)(cid:5)(cid:1)(cid:6)(cid:1)(cid:10)(cid:1)(cid:2)(cid:11)(cid:5)(cid:1)(cid:3)(cid:1)(cid:4)(cid:11)(cid:1)(cid:14)(a) Network Load (b) NCLOs (c) Privacy Figure 16: Performance comparison under different tightness 0.10.20.30.40.50.60.70.8Tightness101102103104105Network Load (KB)ATWBSyncABBAsymPT-FBPT-ISABBAsymDPOP(kp=w$;ke=1)AsymDPOP(kp=2;ke=1)AsymDPOP(kp=3;ke=1)0.10.20.30.40.50.60.70.8Tightness102103104105106107108109NCLOsATWBSyncABBAsymPT-FBPT-ISABBAsymDPOP(kp=w$;ke=1)AsymDPOP(kp=2;ke=1)AsymDPOP(kp=3;ke=1)0.10.20.30.40.50.60.70.8Tightness00.10.20.30.40.50.60.7Privacy(cid:1)(cid:2)(cid:3)(cid:4)(cid:5)(cid:6)(cid:7)(cid:8)(cid:1)(cid:4)(cid:4)(cid:1)(cid:9)(cid:6)(cid:10)(cid:11)(cid:2)(cid:12)(cid:13)(cid:4)(cid:11)(cid:2)(cid:12)(cid:14)(cid:5)(cid:1)(cid:4)(cid:4)(cid:1)(cid:9)(cid:6)(cid:10)(cid:15)(cid:11)(cid:16)(cid:11)(cid:17)(cid:1)(cid:2)(cid:18)(cid:3)(cid:1)(cid:4)(cid:1)(cid:5)(cid:18)(cid:1)(cid:19)(cid:1)(cid:9)(cid:6)(cid:10)(cid:15)(cid:11)(cid:16)(cid:11)(cid:17)(cid:1)(cid:2)(cid:18)(cid:20)(cid:4)(cid:1)(cid:5)(cid:18)(cid:1)(cid:19)(cid:1)(cid:9)(cid:6)(cid:10)(cid:15)(cid:11)(cid:16)(cid:11)(cid:17)(cid:1)(cid:2)(cid:18)(cid:21)(cid:4)(cid:1)(cid:5)(cid:18)(cid:1)(cid:19)
1903.06979	1	1903	2019-03-16T19:59:41	A Principal-Agent Model of Systems Engineering Processes with Application to Satellite Design	[ "cs.MA" ]	We present a principal-agent model of a one-shot, shallow, systems engineering process. The process is one-shot in the sense that decisions are made during one time step and that they are final. The term shallow refers to a one-layer hierarchy of the process. Specifically, we assume that the systems engineer has already decomposed the problem in subsystems, and that each subsystem is assigned to a different subsystem engineer. Each subsystem engineer works independently to maximize their own expected payoff. The goal of the systems engineer is to maximize the system-level payoff by incentivizing the subsystem engineers. We restrict our attention to requirement-based system-level payoffs, i.e., the systems engineer makes a profit only if all the design requirements are met. We illustrate the model using the design of an Earth-orbiting satellite system where the systems engineer determines the optimum incentive structures and requirements for two subsystems: the propulsion subsystem and the power subsystem. The model enables the analysis of a systems engineer's decisions about optimal passed-down requirements and incentives for sub-system engineers under different levels of task difficulty and associated costs. Sample results, for the case of risk-neutral systems and subsystems engineers, show that it is not always in the best interest of the systems engineer to pass down the true requirements. As expected, the model predicts that for small to moderate task uncertainties the optimal requirements are higher than the true ones, effectively eliminating the probability of failure for the systems engineer. In contrast, the model predicts that for large task uncertainties the optimal requirements should be smaller than the true ones in order to lure the subsystem engineers into participation.	cs.MA	cs	A Principal-Agent Model of Systems Engineering Processes with Application to Satellite Design Salar Safarkhani†, Vikranth Reddy Kattakuri†, Ilias Bilionis†∗, Jitesh Panchal† †School of Mechanical Engineering, Purdue University, West Lafayette, Indiana 47907-2088 ∗Corresponding author, Email: [email protected] 9 1 0 2 r a M 6 1 ] A M . s c [ 1 v 9 7 9 6 0 . 3 0 9 1 : v i X r a Abstract -- We present a principal-agent model of a one-shot, shallow, systems engineering process. The process is "one-shot" in the sense that decisions are made during one time step and that they are final. The term "shallow" refers to a one-layer hierarchy of the process. Specifically, we assume that the systems engineer has already decomposed the problem in subsystems, and that each subsystem is assigned to a different subsystem engineer. Each subsystem engineer works independently to maximize their own expected payoff. The goal of the systems engineer is to maximize the system-level payoff by incentivizing the subsystem engineers. We restrict our attention to requirement-based system- level payoffs, i.e., the systems engineer makes a profit only if all the design requirements are met. We illustrate the model using the design of an Earth-orbiting satellite system where the systems engineer determines the optimum incentive structures and requirements for two subsystems: the propulsion subsystem and the power subsystem. The model enables the analysis of a systems engineer's decisions about optimal passed-down require- ments and incentives for sub-system engineers under different levels of task difficulty and associated costs. Sample results, for the case of risk-neutral systems and subsystems engineers, show that it is not always in the best interest of the systems engineer to pass down the true requirements. As expected, the model predicts that for small to moderate task uncertainties the optimal requirements are higher than the true ones, effectively eliminating the probability of failure for the systems engineer. In contrast, the model predicts that for large task uncertainties the optimal requirements should be smaller than the true ones in order to lure the subsystem engineers into participation. I. INTRODUCTION The ubiquitous problem of schedule and cost over-runs during the development of large-scale complex systems is well documented within systems engineering literature [1]. Various remedies have been proposed to address these un- sustainable trends, including, better methods and tools for managing complexity, better incentive mechanisms, and tran- sition from document-based systems engineering to model- based systems engineering (MBSE). The research community has related these trends to the fundamental way in which systems engineering processes are carried out. Requirements engineering, which is one of the foundational processes within systems engineering, has been identified as a key source of the inefficiency. Collopy, for example, argues that the use of requirements in systems engineering is an ineffective way of coordination between systems engineer and subsystems engineers [2]. Therefore, there is a need within systems en- gineering to model and analyze the requirements engineering process. There have been few efforts in addressing this need. By modeling systems engineering processes as multi-disciplinary design optimization problems, Collopy et al. show that us- ing requirements within systems engineering processes cre- ates design trade conflicts among different subsystems, re- sulting in dead losses within the system [2]. Collopy and Hollingsworth [1] espouse the use of value driven design (VDD) as a better alternative to requirements engineering, wherein, objectives for extensive attributes are passed down instead of requirements. Their model is applicable for settings where the incentives of the subsystem engineers are well aligned with the objectives of the systems engineers. This assumption may be valid when both the systems engineers and sub-systems engineers are within the same organization. If, on the other hand, the subsystems engineers are independent decision makers with private information and driven by their own objectives, their model is inappropriate, and superiority of VDD is not clear. To model realistic systems engineering processes, there is a need to model interactive decisions of self-interested actors using game theory. Vermillion and Malak [3] take initial steps in that direction by modeling the interactions between a sys- tems engineer and subsystems engineers using principal-agent models. They adapt the generalized principal-agent model to model the situation of a systems engineer delegating work to subsystem engineers as a one-shot game. Their adaption is primarily focused on incorporating behavioral aspects such as deviations from expected utility maximization within the general principal-agent model. While incorporating behavioral aspects within principal- agent models is an important step forward, we believe that there is still a lack of models that account for unique aspects of systems engineering, specifically, the information available to systems engineers, the state of technology, the uncertainty in the ability to achieve specific outcomes, and the level of difficulty of the tasks. To address this gap, we develop a principal-agent model [4] of a simple systems engineering process in which decisions are made once and all involved individuals have their own private interests. Note that our model is an oversimplification of real systems engineering processes which are iterative and in which information and outcomes flow back and forth between the systems engineer and subsystem engineers until a final decision is made. Our model should be considered as a first step towards modeling full fledged systems engineering processes. Using our frame- work, we study the optimal mechanisms within the class of requirement-based incentives. We illustrate the model using a satellite design case study with two subsystems: power and propulsion. Specifically, we show how historical data can be used to infer the parameters of our process model to this case study. The paper is organized as follows. We start Sec. II by describing our systems engineering process model in general terms and we cast the selection of subsystem incentives as a mechanism design problem. In Sec. II.A, we justify some assumptions for the nature of the subsystem engineers which simplify the model. In Sec. II.B, we focus the discussion on the class of requirement-based incentives for all agents. In Sec. II.C and II.D, we non-dimensionalize the equations and we describe how the parameters can be inferred from readily available historical data. In Sec. III, we apply the model to the design of a spacecraft taking into account two subsystems (power and propulsion). Finally, in Sec. IV, we present our numerical study and in Sec. V our conclusions. II. MODELING A SINGLE-SHOT, SHALLOW SYSTEMS ENGINEERING PROCESS We consider a model of a single-shot (evolves in one time-step), shallow (considers one-layer interactions between a systems engineer and multiple subsystem engineers) systems engineering process. The systems engineer (SE) has already decomposed the problem in N subsystems. The SE assigns the design of each subsystem to a subsystem engineer (sSE). Each sSE designs independently to maximize their own expected payoff, and returns the design outcome back to the SE. The goal of the SE is to incentivize the sSEs to produce subsystem designs that maximize the expected system-level payoff by choosing appropriate contracts. We start by formulating the problem of optimal subsystem contract design in its full generality. Then, we make simplifying assumptions about the form of the sSEs' quality and utility functions, and, we study the optimality requirement-based contracts. Let i = 1, . . . , N be a label indexing the sSEs. The i-th sSE chooses a normalized effort level ei ∈ [0, 1]. This measures the percentage of maximum effort that the sSE can allocate to this specific project within a predefined time framework, e.g., in a fiscal year. Second, the units of the effort depend on the nature of the sSE. If the sSE is an individual that works for the same organization as the SE, then the effort ei could be measured in terms of the percentage of that the individual dedicates to the project. Alternatively, if the sSE is an external contractor, e.g., another company, then effort could be measured in terms of the percentage of the available yearly resources that the contractor dedicates to this particular project. We denote the cost of effort to the sSE as ci(ei). In economic terms, ci(ei) is an opportunity cost, i.e., the monetary gain the sSE could receive, but forfeits, to participate in this particular project. Let (Ω,F, P) be a probability space associated with random states of nature. We model the design qualities that the i-th sSE can produce as a stochastic process qi(ei, ω). That is, the 2 quality function qi(ei, ω) gives the design quality that the i- th sSE can achieve by choosing an effort level ei if the state of nature is ω ∈ Ω. The quality function is normalized so that zero corresponds to the quality of the current state-of- the-art. Note that we have deliberately chosen to ignore the dependence of qi(ei, ω) on any private information. In other words, we assume that the form of the quality function is common knowledge. Each of the sSEs enters a contract with the SE. The contracts describe transfer functions, ti(qi), which specify the transport of monetary funds from the SE to the sSE contingent on the quality of the design that the sSE produces. Therefore, the payoff to the sSE is: πi (ei, ω) = ti (qi (ei, ω)) − ci (ei) . (1) We assume that the sSE selects an optimal effort level ex- ante, i.e., before they observe the future state of nature ω. If we further assume that the sSE is risk-neutral, then rationality implies that they should select their effort level by maximizing their expected payoff: ei ∗ (ti (·)) = arg max ei∈[0,1] Eω [πi (ei, ω)] . (2) Note the dependence on the transfer function. To ensure that the sSEs are willing to participate in this project, their optimal expected payoff must be positive. Oth- erwise, the sSEs have no incentive to be part of the project as their expected monetary benefit is smaller than their oppor- tunity cost. Therefore, the SE must choose transfer functions that enforce the participation constraints: Eω [πi (ei ∗ (ti (·)) , ω)] ≥ 0, (3) for i = 1, . . . , N. ties: The SE obtains from each sSE the following design quali- ∗ (ti (·) , ω) = qi (ei qi (4) If V (q1, . . . , qN ) is the net present value of any cash flows that result from a system with subsystem qualities q1, . . . , qN , then the total payoff to the SE is: ∗ (ti (·)) , ω) . π (t (·) , ω) = V (q1 ∗ (t1 (·) , ω) ,··· , qN ∗ (tN (·) , ω)) − N(cid:88) i=1 ∗ (ti (·) , ω)) , ti (qi (5) where we defined t (·) = (t1 (·) ,··· , tN (·)). Assuming that the SE makes ex-ante, risk-neutral decisions, they must select transfer functions that solve: t∗ (·) = arg max Eω [π (t (·) , ω)] , (6) t(·) then the SE does not subject to the N participation constraints defined in Eq. (3). Of course, if the optimal expected payoff Eω[π(t∗(·), ω)] is negative, in the first place. In what follows, we study this mechanism design problem by making specific assumptions for the form of the SE value function, the sSEs' quality functions and costs, and the form of the possible transfer functions. initiate the project A. Assumptions about the subsystem engineers Our final assumption is that the sSE's cost grows linearly with effort as: ci (ei) = ciei, (9) where ci > 0. This assumption is reasonable for almost all types of sSEs. In particular, when sSE is an individual engineer, ci could be the average industry salary per unit effort. For a contractor, ci could be the expected payoff per unit effort of the next most profitable project that they could be engaging in. B. Optimal requirement-based incentives We assume that the SE has N requirements, r1, . . . , rN , one to be satisfied by each subsystem. These requirements arise from the business case of the project. Mathematically, the system design is successful if qi > ri for all i = 1, . . . , N. If the value of a successful system design is V0, then the value function of the SE is: V (q1, . . . , qN ) = V0 H (qi − ri) , (10) N(cid:89) i=1 where H (·) is the Heaviside function: 1, if x ≥ 0, 0, otherwise. H(x) = (cid:40) The random field qi(ei, ω) captures the common state of knowledge about what is technologically possible in the design quality of subsystem i. Using the Karhunen-Lo`eve expansion [5], the random field qi(ei, ω) can be written as: qi(ei, ω) = q0 i (ei) + λikξik(ω)φik(ei), (7) ∞(cid:88) (cid:112) k=1 where q0 i (ei) is the mean of the random field, λik, φik(ei) are the eigenvalues and eigenvectors of its covariance function, respectively, and the random variables ξik are zero mean, unit variance, and uncorrelated. Assuming stationarity of the process, these quantities can, in principle, be estimated sta- tistically from historical data of marginal investments versus increases in product quality. As a first approximation, we trun- cate the series at k = 1 keeping only the largest eigenvalue: qi(ei, ω) ≈ q0 i (ei) + λi1ξi1(ω)φi1(ei). (cid:112) √ Furthermore, we approximate the zero-mean and unit-variance random variable ξi1 as a standard normal random variable ξ (the standard normal is the maximum entropy distribution with zero-mean and unit variance). We also assume that the first eigenvector is approximately constant, φi1(ei) ≈ const, and we introduce the new variable σi = λiφi1(ei). Without loss of generality, we may take that σi > 0. Finally, we take the first order Taylor expansion of q0 i ), recalling that we scale the quality so that zero corresponds to the current state-of-the-art, which can be delivered without any effort. This is reasonable since we are considering a one-shot systems engineering process which, necessarily, takes place in a limited amount of time. For larger timescales, we expect qi(ei, ω) to be curved: Concave for mature technologies, and convex followed by concave for emerging technologies. Furthermore, we take ai > 0 since more effort can only lead to increased design quality. To summarize, we model the quality as: i (ei) = aiei+O(e2 qi (ei, ω) = aiei + σiξ (ω) , (8) where ai, σi > 0, and ξ ∼ N (0, 1). The parameter ai depends on the skills of the sSE as well as on the maturity of the underlying technology. That is, a skillful sSE produces a higher increase in quality from the state-of-the-art than a less skilled sSE. Therefore, keeping the maturity of the technology fixed, ai expected to grow as the skills of the sSE are improved. Similarly, keeping the skills of the sSE fixed, ai decreases as a function of the maturity of the underlying technology. The more mature the underlying technology is, the more difficult it becomes to obtain a given increase in quality. The parameter σi behaves in exactly the opposite way. A skillful sSE produces design qualities that vary less, therefore σi decreases as skill improves. On the other hand, we expect that subsystem designs that depend on mature technologies are more predictable, therefore σi decreases as technological maturity increases. 3 We restrict our attention to the study of requirement-based transfer functions: ti (qi; ψi) = ψi1 + ψi2 H (qi − ψi3) . (11) The first parameter, ψi1, specifies the amount that is going to be paid simply for agreeing to participate in the project, guaranteeing a design quality at least the same as the cur- rent state-of-the-art. The second parameter, ψi2, specifies the amount to be paid if the subsystem engineer meets specific requirements. The third parameter, ψi3, specifies the require- ment that the subsystem engineer has to meet. Note that the optimal passed-down requirement is, in general, different than the real requirement, ri. We start with the optimal decision of the i-th sSE given a fixed contract ti (qi; ψi). The payoff is: πi(ei, ω) = ψi1 + ψi2 H (aiei + σiξi(ω) − ψi3) − ciei. To take the expectation over ω, we use the following result: Eω [H (λ + σξ(ω))] = Φ (12) where Φ(·) is the cumulative distribution function (CDF) of a standard normal random variable. Taking the expectation over ω, we get: σ , (cid:18) λ (cid:19) Eω [πi(ei, ω)] = ψi1 − ciei + ψi2Φ So, the optimal effort level is: i (ti (·; ψi)) = arg max e∗ ei∈[0,1] (cid:26) ψi1 − ciei + ψi2Φ (cid:19) (cid:18) aiei − ψi3 (cid:18) aiei − ψi3 σi . σi (cid:19)(cid:27) . (13) We solve this problem using Brent's method [6] as imple- mented in SciPy [7]. The SE problem consists of maximizing: (cid:18) aie∗ i (ti (·; ψi)) − ri σi (cid:19) Eω [π (t (·, ψ) , ω)] =V0 Φ i=1 N(cid:89) − N(cid:88) − N(cid:88) i=1 i=1 ψi1 ψi2Φ (cid:18) aie∗ (cid:19) i (ti (·; ψi)) − ψi3 σi − cie∗ i (ti (·; ψi)) ≥ 0, (15) (cid:19) , (14) (16) subject to the participation constraints: (cid:18) aie∗ ψi1 + ψi2Φ and the bounds: i (ti (·; ψi)) − ψi3 σi ψik ≥ 0. for k = 1, 2, 3 and i = 1, . . . , N. In practice, the optimal solution is always within the following bounds: 0 ≤ ψi1, ψi2 ≤ 2ci, and ri − 3σi < ψi3 < ri + 3σi. (17) We solve this problem using the constrained optimization by linear approximation (COBYLA) algorithm [8] as imple- mented in [7]. C. Non-dimensionalization of the equations Without loss of generality, we can pick all the subsys- tem requirements to be ri = 1. This can be achieved by appropriately scaling all ai's and all σi's. In other words, the quality function of each subsystem will be measured in units of the corresponding business-imposed requirement. Then, the inverse of the coefficient ai, can be interpreted as the effort that needed to meet the subsystem requirement for sSE if there is no uncertainty. We will consider two levels of ai's corresponding to different levels of subsystem design difficulty: (i) (hard) ai = 1.5, (ii) (easy) ai = 2. Having scaled the quality function in this way, the variance parameter σi can also be interpreted as the amount of uncertainty in the quality of the final design as a percentage of the requirement. We will consider three levels of uncertainty: (i) (low) σi = 0.05, (ii) (moderate) σi = 0.1, and (iii) (high) σi = 0.2. Finally, also without loss of generality, we may set V0 = 1. This can be achieved by appropriately scaling the transfer functions and the opportunity costs of all sSEs. It amounts to measuring all monetary quantities, in terms of the SE's maximum value. For the opportunity costs of the sSEs, we will consider two levels: (i) (low) ci = 0.01, (ii) (high) ci = 0.05. D. Extracting subsystem parameters from historical data To study decision making in the context of a real ap- plication, one needs to extract all parameters, ai, ci, σi, . . . , from historical data. To this end, let Qi denote the quality of the i-th subsystem in physical units, and Ii the cumulative investment per firm on this technology. Historical data, say Si(cid:88) Si(cid:88) s=1 s=1 Di = {(Iis, Qis)}Si s=1, of these quantities are readily available for many technologies. We model the relationship between Qi and Ii as: Qi = Qi0 + Ai(Ii − Ii0) + Σiξi(ω), (18) where Qi0 and Ii0 is the current state of these variables, ξi(ω) ∼ N (0, 1), and Ai and Σi are parameters to be estimated from the all available data Di. We use a maximum likelihood estimator [9] for Ai and Σi. This is equivalent to least squares estimate for Ai: Ai = arg min Ai [Qi0 + Ai(Iis − Ii0) − Qis]2 , (19) and to setting Σi equal to the mean residual square error: Σi = 1 Si [Qi0 + A∗ i (Iis − Ii0) − Qis]2 . (20) Now, let Qir be the required quality for subsystem i in physical units. The scaled quality of a subsystem qi, can be defined as: qi = Qi − Qi0 Qir − Qi0 , (21) with this definition, we get qi = 0 for the state-of-the-art, and qi = 1 for the requirement. Substituting Eq. (18) in Eq. (21) while making use of the maximum likelihood estimates for Ai and Σi, we get qi = Ai Qir − Qi0 (Ii − Ii0) + Σi Qir − Qi0 ξi(ω). (22) From this equation, we see that the uncertain parameter of our previous discussions, can be obtained from σi = Σi Qir − Qi0 . (23) Finally, let Ti and Ci represent the time for which the ith sSE is to be hired and the cost of the engineer per unit time, respectively. Ti is just the duration of the systems engineering process we consider. The opportunity cost Ci can be read inferred from the balance sheets of publicly traded firms related to the technology. We can now define the effort variable ei as: ei = Ii − Ii0 TiCi . (24) (25) From this and Eq. (22), we see that ai is given by: ai = TiCi Ai Qir − Qi0 . III. ILLUSTRATIVE EXAMPLE: SPACECRAFT DESIGN PROBLEM A. Spacecraft systems design During the initial proposal phase of satellite development for scientific applications, the principal investigator puts forward an estimate of how the goals of the project will be achieved through engineering means. These project goals concern the 4 the heart overarching science objectives of the mission which corre- spond to the instrument design at the satellite. However, in the process of successfully launching a scientific instrument into space, all the necessary components to power the instruments, actuate the spacecraft, and transmit data must also be included in the satellite payload. These mission requirements are translated to specific functional requirements for each subsystem of the spacecraft. Typically, a spacecraft consists of seven main sub- systems [10], namely, electrical power subsystem (EPS), propulsion, attitude determination and control (ADC), on- board processing, telemetry, tracking and command (TT&C), structures and thermal subsystems. For the proposed study, we will focus our attention on two subsystems (N = 2): EPS and propulsion. Simplifying the analysis, we assume that the design of these subsystems will be assigned to two (subsystem) engineers in a one-shot fashion. Note that, the systems engineering process of the spacecraft is an iterative process and the information and results are exchanged and flow back and forth between the SE and sSEs in each iteration. Our model is a very crude approximation of reality. The goal of the SE is to optimally incentivize the sSEs to produce subsystem designs that meet the mission's requirements. B. Electrical power subsystem The EPS is designed and configured to perform several key functions, the primary being a continuous and reliable source of peak and average electrical power for the life of the mission. It consists of a power source, energy storage, power conversion/distribution and power regulations and con- trol equipment. Typically, for earth orbiting satellites, one employs solar photo-voltaic arrays as a primary energy source and batteries as secondary power storage units. Silicon solar cells are the most commonly used photo-voltaic cells for space applications because of their low cost and high availability. In this study, the design quality of interest for the EPS sSE is chosen to be the solar cell efficiency, i.e., Q1 is the efficiency of Si-based solar photo-voltaic cells expressed as a percentage value, and I1 is the average cumulative investment in solar cell research per firm. To estimate the relationship between Q1 and I1, we first considered the global trends of commercial Si module efficiencies and cumulative investments in solar-cell research over 2001 -- 2008. Over this period, a total of hundred solar cell research and development companies received venture capital (VC) funding [11]. Based on this information, we assume that, on average, thirteen companies are participating in solar cell R&D every year. The variation of commercial crystalline-Si module effi- ciency (Q1 (%) ) with time is taken from [12]. The trend of cumulative global VC investment per firm (I1 (millions USD)) in crystalline Si-cell technology over time is obtained from [11]. In 2008, the state-of-the-art was Q10 = 19% and the cumulative investment per firm was I10 = 102.4 million USD. A maximum likelihood fit of the parameters in Eq. (18) results in a regression coefficient of A1 = 0.035% per million USD, and standard deviation Σ1 = 0.15%. We can visualize the data and the maximum likelihood fit in Fig. 1. Fig. 1: Spacecraft case study (EPS subsystem): Historical data (2001 -- 2008) of solar cell efficiencies vs cumulative investment per firm. The solid line and the shaded area correspond to the maximum likelihood fit of a linear regression model and the corresponding 95% prediction intervals, respectively. The cost parameter C1 is estimated by considering the total pay towards employees salaries over the number of employees in R&D jobs in global PV industry. According to the statistics [12], around 2,320 employees were involved in such jobs by the end of 2008. The value of C1 is the median salary of a solar cell development engineer which is approximately equal to 100,000 (0.1 million) USD based on data from [13]. Substituting the values of Q10, C1 and A1 in Eq. 25 yields the following relationship between the parameters Q1r, T1 and a1: a1 = 0.0035T1 Q1r − 19 (26) The variation of required quality (efficiency) for sSE-1 (Q1r (%)) with respect to the time for which one person from sSE firm is to be hired (T1 years) is shown in Fig. 2, for two different levels of EPS design difficulty. C. Propulsion subsystem Space propulsion systems essentially provide thrust to lift the launch vehicle along with its payload from the launch pad and place the payload into low-Earth orbit. They assist payload transfer between lower and higher orbits or into trajectories based on the type of mission. Finally, they provide thrust for attitude control and orbit corrections [10]. Based on mission profiles, performance requirements for propulsion systems include thrust, total impulse, and duty cycle specifications for which, the specific impulse and propellant density are the key parameters. Chemical combustion systems, which are the most common systems for space applications can be divided into three basic categories: liquid. solid and hybrid. The terminology refers 5 Fig. 2: Spacecraft case study (EPS subsystem): Variation of Q1r (percentage efficiency) w.r.t T1 (time in years) to the physical state of the stored propellants. In this study, the design quality of interest (Q2) for the propulsion sSE is chosen to be the delivered specific impulse (Isp, measured in seconds) of solid propellants. Specific impulse is defined as the ratio of thrust to weight flow rate of the propellant and is a measure of energy content of the propellants [10]. It signifies the energy to thrust conversion efficiency. I2 is the cumulative investment on chemical propulsion research and technology by NASA over a period of ten fiscal years from 1979 -- 1988. Trends in delivered specific impulse (Q2 (sec.)) and in- vestments by NASA (I2 (millions USD)) in chemical propul- sion technology with time are obtained from [14] and [15], respectively. The state-of-the-art solid propellant technology corresponds to a Q20 value of 252 sec. and I20 value of 149.1 million USD. In this case, the maximum likelihood fit of the parameters results in a regression coefficient of A2 = 0.0133 sec. per million USD, and standard deviation Σ2 = 0.12 sec. The corresponding data and the maximum likelihood fit are illustrated in Fig. 3 The value of C2 in this case is approximately same as that of C1, i.e., C2 = 100,000 USD according to the data obtained from [16]. Substituting the values of Q20, C2 and A2 in Eq. 25 yields the following relationship between the parameters Q2r, T2 and a2: (27) a2 = 0.0013T2 Q2r − 252 Fig. 4 shows the variation of required specific impulse for sSE- 2 (Q2r (sec.)) with time for which one propulsion engineer has to be hired (T2 years) for two different levels of subsystem design difficulty. IV. RESULTS We present numerical examples of one-shot SE processes with two identical sSEs (N = 2). We investigate 4 different cases consisting of all possible combinations of two difficulty levels (easy (ai = 2) vs hard (ai = 1.5)) and sSE opportunity 6 Fig. 3: Spacecraft case study (propulsion subsystem): His- torical data (1979 -- 1988) of specific impulse of solid mono- propellants vs cumulative investment per firm. The solid line and the shaded area correspond to the maximum likelihood fit of a linear regression model and the corresponding 95% prediction intervals, respectively. Fig. 4: Spacecraft case study (propulsion subsystem): Variation of Q2r (Specific impulse) w.r.t T2 (time in years) cost levels (low (ci = 0.01) vs high (ci = 0.05)). For each case, we consider three scenarios with different uncertainty levels: σ = 0.05, 0.1, and 0.2. For each scenario, we obtain the ∗ by solving the mechanism design problem optimal contract ψ of Sec. II within the class of requirement-based incentives. Finally, we study the sensitivity of the SE's expected payoff on the passed-down requirement by changing the values of ψ13 in the range [0, 2]. All numerical results can be found in Fig. 5 and its captions. Specifically, Figs. 5a, 5b, 5c, and 5d include the results for a hard-task -- low-cost-sSE, hard-task -- high-cost-sSEs, easy-task -- low-cost-sSE, and easy-task -- high- cost-sSE, respectively. The optimal contracts of each scenario can be read from the captions of the associated subfigures. i1 and ψ∗ We start by commenting on the properties of the payment amounts ψ∗ i2. We will refer to ψ∗ i1 as the partic- ipation payment, i.e., the fixed payment made to the sSE independently of the design outcome, and to ψ∗ i2 as the bonus payment, i.e., the payment made to the sSE if the passed-down requirements are met in our one-shot systems engineering model. First, we observe that across all the scenarios the participation payment increases as uncertainty grows. This makes sense, since the sSE is expected to ask for a higher certain gain to accept a higher risk task. Counter-intuitively, the bonus payment is independent of the uncertainty level. Second, for a fixed opportunity costs the participation payment decreases with increasing task difficulty while the bonus pay- ment behaves in the opposite way (it increases with increasing task difficulty). This means that for harder tasks the SE hedges themselves by shifting some of the participation payment to the bonus. Finally, as the opportunity cost increases, both the participation and bonus payments increase. Let us focus on the optimal passed-down requirements (ψ∗ i3). Across all scenarios, the optimal passed-down require- ment increases as the uncertainty becomes larger. This is an example of the SE attempting to increase the probability of the actual requirement being met as uncertainty increases. More- over, we observe that the optimal passed-down requirement is independent of the opportunity cost. This is due to the fact that the SE value function is requirement-based. Furthermore, we observe that for easy tasks, captions of Figs. 5c and 5d, the optimal passed-down requirement is always greater than the actual requirement ri = 1. This result can be intuitively understood as follows. Since the task is easy, the sSE will definitely reach the actual requirement with sufficient effort, even in the presence of significant uncertainties. By setting the requirement threshold higher than the actual one, the SE forces the sSE to use more effort, effectively increasing the probability of success to almost certainty. In contrast, for hard tasks, Figs 5a and 5b, the optimal passed-down requirements are lower than the actual system requirement. Since the task is hard, there is a high probability that the sSE may not be able to achieve a very high passed-down requirement. To lure the sSEs to participate, the SE needs to lower the passed-down requirements below the actual threshold. At a first glance, this may look like the SE will not be able to gain a positive expected payoff. However, due to uncertainty in the final outcome, there is a significant probability that the sSE will produce better than the requested result. Naturally, we observe that the expected SE payoff is in- creasing as the opportunity costs go down and that it decreases as the task difficulty increases. For easy tasks, the expected SE payoff decreases as the uncertainty increases, Figs. 5c and 5d. On the contrary, for hard tasks the expected SE payoff increases as a function of the uncertainty, Figs. 5a and 5b. As we saw in the previous paragraph, this makes sense for a hard task that, an increase in uncertainty makes it more likely that the actual requirement is met. V. CONCLUSION We presented a principle agent model of a one-shot, shallow systems engineering process. We assumed that all agents are risk neutral and, thus, they maximize their expected payoff. We modeled the quality function of subsystem engineers as a linear function of effort plus some Gaussian noise. Using a spacecraft design case study, we demonstrated how the parameters of our model can be estimated from historical data. Finally, we posed the optimal mechanism design problem within the class of requirement-based incentives. Our one-shot model of systems engineering process chal- lenges the intuitive belief that one should ask for higher requirements as the design task becomes more difficult. Our model predicted that for a hard task, the optimal passed-down requirement should be less than the actual requirement. The reason is that in this way the sSE is lured into participation while the SE may still meet the requirement because the design outcome may actually be better than anticipated. Our result does not mean that this common belief is wrong. After all, it is a very simple model, capturing only one iteration of systems engineering process. This result may change if the quality function is not linear or if the design quality noise becomes skewed. Even if the modeling choices were spot on, at the present stage, it is too simple to be truly descriptive. There may be mechanisms beyond the ones included in our model that incentivize the SE to ask for a requirement higher than what our model predicts? One reason is that they may underestimate the difficulty of the task. Another reason is that they may want to hedge themselves against dishonest behavior of sSEs, e.g., the sSE may put a design on their back pocket for later use. Third, the systems engineering process may be taking place iteratively and asking for a higher requirement may be an effective way to probe what is possible. From the perspective of the sSE, why would they accept to participate in a hard task with an exceedingly high requirement? Of course, they may also underestimate the difficulty of the task. Alternatively, they may be offered a participation payment that is high enough so that they do not care that it is impossible to meet the passed- down requirement. Finally, they may believe that they will be able to renegotiate the contract in the future, especially if the SE has a history of doing so. All these intricacies, and many more, are not captured by our model. They are topic of on going research towards a theoretical foundation of systems engineering design that accounts for human behavior. ACKNOWLEDGMENT This material is based upon work supported by the National Science Foundation under Grant No. 1728165. REFERENCES [1] P. D. Collopy and P. M. Hollingsworth, "Value-Driven Design," Journal of Aircraft, vol. 48, no. 3, pp. 749 -- 759, 2011. [Online]. Available: https://doi.org/10.2514/1.C000311 7 (a) Hard (ai = 1.5) - low cost (ci = 0.01). σ = 0.05: ψ1 = (0.000, 0.007, ψ13), ψ2 = (0.000, 0.007, 0.945); σ = 0.10: ψ1 = (0.001, 0.007, ψ13), ψ2 = (0.001, 0.007, 0.953); σ = 0.20: ψ1 = (0.002, 0.007, ψ13), ψ2 = (0.002, 0.007, 0.968). (b) Hard (ai = 1.5) - high cost (ci = 0.05). σ = 0.05: ψ1 = (0.004, 0.033, ψ13), ψ2 = (0.004, 0.033, 0.945); σ = 0.10: ψ1 = (0.007, 0.033, ψ13), ψ2 = (0.007, 0.033, 0.953); σ = 0.20: ψ1 = (0.012, 0.033, ψ13), ψ2 = (0.012, 0.033, 0.968). (c) Easy (ai = 2.0) - low cost (ci = 0.01). σ = 0.05: ψ1 = (0.001, 0.005, ψ13), ψ2 = (0.001, 0.005, 1.087); σ = 0.10: ψ1 = (0.002, 0.005, ψ13), ψ2 = (0.002, 0.005, 1.205); σ = 0.20: ψ1 = (0.003, 0.005, ψ13), ψ2 = (0.003, 0.005, 1.221). (d) Easy (ai = 2.0) - high cost (ci = 0.05). σ = 0.05: ψ1 = (0.005, 0.025, ψ13), ψ2 = (0.005, 0.025, 1.087); σ = 0.10: ψ1 = (0.009, 0.025, ψ13), ψ2 = (0.009, 0.025, 1.142); σ = 0.20: ψ1 = (0.014, 0.025, ψ13), ψ2 = (0.014, 0.025, 1.221). Fig. 5: The SE's expected payoff vs. passed-down requirement (ψ13) for several types of sSE [2] P. Collopy, "Adverse Impact of Extensive Attribute Requirements on the Design of Complex Systems," in 7th AIAA ATIO Conf, 2nd CEIAT Int'l Conf on Innov and Integr in Aero Sciences,17th LTA Systems Tech Conf; followed by 2nd TEOS Forum. American Institute of Aeronautics and Astronautics, dOI: 10.2514/6.2007-7820. [Online]. Available: https://arc.aiaa.org/doi/abs/10.2514/6.2007-7820 [3] S. D. Vermillion and R. J. Malak, "Using a Principal-Agent Model to Investigate Delegation in Systems Engineering," no. 57052, p. V01BT02A046, 2015. [Online]. Available: http://dx.doi.org/10.1115/ DETC2015-47778 [4] J. Meluso and J. Austin-Breneman, "Gaming the System: An Agent- Based Model of Estimation Strategies and Their Effects on System Performance," p. V02AT03A050, Aug. 2017. [Online]. Available: http://dx.doi.org/10.1115/DETC2017-68202 [5] R. Ghanem, Stochastic finite elements: a spectral approach. New York: Springer-Verlag, 1991, open Library ID: OL1865197M. [6] R. P. Brent, Algorithms for Minimization Without Derivatives. Mineola, N.Y: Dover Publications, Apr. 2013. [7] E. Jones, T. Oliphant, P. Peterson, and others, SciPy: Open source scientific tools for Python, 2001. [Online]. Available: http: //www.scipy.org/ [8] S. Lucidi and M. Sciandrone, "On the Global Convergence of for Unconstrained Optimization," SIAM Derivative-Free Methods Journal on Optimization, vol. 13, no. 1, pp. 97 -- 116, Jan. 2002. [Online]. Available: http://epubs.siam.org/doi/abs/10.1137/S1052623497330392 [9] M. E. Tipping and C. M. Bishop, "Probabilistic Principal Component Analysis," Journal of the Royal Statistical Society: Series B (Statistical Methodology), vol. 61, no. 3, pp. 611 -- 622, Jan. 1999. [Online]. Available: http://onlinelibrary.wiley.com/doi/10.1111/1467-9868.00196/ abstract [10] J. R. Wertz, D. F. Everett, and J. J. Puschell, Space mission engineering: the new SMAD. Hawthorne, CA: Microcosm Press : Sold and distributed worldwide by Microcosm Astronautics Books, 2011, oCLC: 747731146. [11] "A Historical Analysis (2000-2007) Energy Technologies Fi- nance." [Online]. Available: https://financere.nrel.gov/finance/content/ historical-analysis-investment-solar-energy-technologies-2000-2007 of Renewable Energy Investment Project Solar in [12] S. Price, R. Margolis, G. Barbose, J. Bartlett, K. Cory, T. Cou- ture, J. DeCesaro, P. Denholm, E. Drury, M. Frickel, C. Hemmeline, 8 T. Mendelsohn, S. Ong, A. Pak, L. Poole, C. Peterman, P. Schwabe, A. Soni, B. Speer, R. Wiser, J. Zuboy, and T. James, "2008 solar technologies market report." [13] "Solar Engineer Salaries." [Online]. Available: https://www. environmentalscience.org/career/solar-engineer [14] "Solid." [Online]. Available: http://www.astronautix.com/s/solid.html [15] "NASA Historical Data Books." [Online]. Available: https://history. nasa.gov/SP-4012/vol6/cover6.html [16] "Propulsion Engineer line]. space-exploration-technologies--dragon-propulsion-engineer--hawthorne, -ca Available: Salaries." [On- https://www.paysa.com/salaries/ 9
1109.6838	1	1109	2011-09-30T14:42:14	Distributed Air Traffic Control : A Human Safety Perspective	[ "cs.MA", "cs.AI" ]	The issues in air traffic control have so far been addressed with the intent to improve resource utilization and achieve an optimized solution with respect to fuel comsumption of aircrafts, efficient usage of the available airspace with minimal congestion related losses under various dynamic constraints. So the focus has almost always been more on smarter management of traffic to increase profits while human safety, though achieved in the process, we believe, has remained less seriously attended. This has become all the more important given that we have overburdened and overstressed air traffic controllers managing hundreds of airports and thousands of aircrafts per day. We propose a multiagent system based distributed approach to handle air traffic ensuring complete human (passenger) safety without removing any humans (ground controllers) from the loop thereby also retaining the earlier advantages in the new solution. The detailed design of the agent system, which will be easily interfacable with the existing environment, is described. Based on our initial findings from simulations, we strongly believe the system to be capable of handling the nuances involved, to be extendable and customizable at any later point in time.	cs.MA	cs	Distributed Air Traffic Control : A Human Safety Perspective 1 1 0 2 p e S 0 3 ] A M . s c [ 1 v 8 3 8 6 . 9 0 1 1 : v i X r a Sarvesh Nikumbh Centre for Modeling and Joeprakash Nathaman [email protected] Rahul Vartak [email protected] Simulation University of Pune Pune 411007, India [email protected] ABSTRACT The issues in air traffic control have so far been addressed with the intent to improve resource utilization and achieve an optimized solution with respect to fuel comsumption of aircrafts, efficient usage of the available airspace with min- imal congestion related losses under various dynamic con- straints. So the focus has almost always been more on smarter management of traffic to increase profits while hu- man safety, though achieved in the process, we believe, has remained less seriously attended. This has become all the more important given that we have overburdened and over- stressed air traffic controllers managing hundreds of airports and thousands of aircrafts per day. We propose a multiagent system based distributed approach to handle air traffic ensuring complete human (passenger) safety without removing any humans (ground controllers) from the loop thereby also retaining the earlier advantages in the new solution. The detailed design of the agent sys- tem, which will be easily interfacable with the existing en- vironment, is described. Based on our initial findings from simulations, we strongly believe the system to be capable of handling the nuances involved, to be extendable and cus- tomizable at any later point in time. Keywords Air traffic control, multiagent system, holding pattern gener- ation, landing sequence, decentralization, automated traffic flow management. INTRODUCTION 1. Air traffic is increasing rapidly in many developing and de- veloped countries. The busiest airports are swamped every- day and incidents involving overworked and overstressed air traffic controllers (ATCs) is increasing. In the current sys- tem, ATCs have some awareness of things related to major airport closures, but they have to make multiple decisions about how to move and re-route traffic just based on their experience and little help to guide them. That's better than nothing, when there are about 16000 ATCs at more than 5000 public airports in the USA itself and more than 13000 worldwide [6]. The end result is decisions that are not al- ways optimal for the larger, national needs. The Controllers are on the job for more than 16 to 17 hours on the trot. The resulting fatigue and stress invariably transfers on their de- cision making, which may result in fatal or penalizing errors. As a matter of fact, there have been many such incidences in the past couple of years. It is this aspect of transferred risk of safety of passengers due to overburdened controllers or humans is what we try to address in this work while the usual optimization of resources is inherently achieved. 2. THE MULTIAGENT SYSTEM In this work we propose a multiagent system based dis- tributed approach for such modern air traffic controlling. Modeled as an inherently fault tolerant system, it handles efficient holding pattern and landing sequence generation. It provides a solution out of the classical centralized rout- ing strategies which are often slow to respond to developing weather conditions or airport operations, and which allow minor local delays to cascade into large regional congestions. We restrict our approach to operate only in the vicinity of airports. In this work, we attempt to address the issues of : • What constitutes an agent in this system? • What actions can the agents take to impact air traffic? • Its interface with the environment. Focussing on the airport vicinities lets us proceed with the obvious choice of an aircraft being an agent though in lit- erature there are instances of fixes -- which indicate ground locations throughout the airspace -- as choice of agents. This choice of aircrafts as agents helps us achieve the aim of being able to tackle issues arising in-flight, on-board that trigger last minute decision changes [4]. 2.1 The Design Each aircraft is responsible with various tasks according to the phase of flight it is currently in. The various phases of flight are viz. arrival into airport vicinity (henceforth called the airspace), at entry gate, on path, holding pattern, to metering fix, at metering fix, final descent, on runway and backtrack. Their service requirement is then dynami- cally determined by arising circumstances. On arrival into airspace, every aircraft shares a copy of queue Qi that has a buffer of size Bi to accommodate other aircrafts (whose number is limited for this airspace). We have holding pat- terns formed which are stacks of aircrafts hovering in the airspace at different altitudes maintaining the queue. These aircrafts in the stack form sets of agent systems. All of those at nearly the same altitude (or groupable altitudes) form a system 'A' which is lead by the first amongst them in the queue to land. Similarly there will be systems 'B', 'C' and so on at higher altitudes lead by the corresponding leaders. The leaders in all the stages also form a system amongst themselves together with the aircrafts which have just en- tered into flight on takeoff. This helps them be updated with the direct information of the aircraft's flight just upon enter- ing the airspace. The non-leaders in each stage would have this information relayed to themselves through the leaders. Once the leader at the lowest altitude takes to landing and leaves the system it passes on the leadership to the next in line. Each aircraft here, on entering into any agent system, is updated with the corresponding current status by a provi- sional agent -- the run time Directory Facilitator (DF). It is responsible for providing every agent in the system with the required information about other agents currently alive and their co-ordinates, state and phase description. The ATC isn't thrown out-of-the-loop and is infact our next agent in the system. It picks up major properties as mentioned in [3]. It overhears and overviews completely the decisions made in the entire process. ATC is the main passive supervision service provider requiring very little amount of service for itself. It shares the queue Qi with buffer of size Bi to track all the aircrafts and their movements. It is also responsible for handling communication with other running processes such as the Surface Movement Radars (SMR), TRACONs (Terminal Radar Approach CONtrol). It does so with the help of the other provisional agent of the system -- InPro- cess interfacing agent. It is responsible for detecting agents in the environment and continually updates the DF agent with this piece of information. It reads input from other external processes or entities on the traffic control system such as Flight Management System (FMS), User Request Evaluation Tool (URET) and many others. The other im- portant agent of the system is the TRACON. It is the second control facility located within the vicinity of a large airport typically controling aircraft within a 30-50 nautical mile ra- dius of the airport between the surface and 18,000 feet. The actual airspace boundaries and altitudes assigned to a termi- nal control are based on factors such as traffic flows, neigh- bouring airports and terrain, and vary widely from airport to airport. It is entrusted with ATCs' responsibilities in their absence. They also notify the ATC regarding any air- craft entering the concerned airspace. They thus form the fall-back agents offering a stronger fault-tolerance. We can also handle communication with 'fixes', the agents for the greater airspace facilitating seamless interface between other agent systems elsewhere in the environment. This way we are guaranteed to gain from the advantages of both -- fixes and aircrafts -- as agents. Each agent can now effect a re- route or a ground delay or cause a change in the sequence by prompt decision making. It prevents linguistic/mental com- munication errors triggered by humans in the loop. These 'well-aware'aircrafts would thus be in a position to tackle issues of emergency re-routes or re-sequencing without any fatal delay. 3. SIMULATIONS We based our simulations on artificial data with typical con- ditions viz. free-flights maintained as per the path-corridor correlation [5] and the data-exchange between agents follows the co-operative data exchange described in [1]. The combi- nations of dynamic causes for route or flight-plan changes in the participating agents were chosen by Monte Carlo meth- ods [2]. The resulting variations observed in the pre-fixed flight paths were succesfully handled under various scenar- ios. The simulations were conducted for Mumbai airport which has two runways (landing and takeoff) crossing each other. We achieved an average of about 38 flights per hour with average successful handling of 86.67 % of all flights with variations in their flight-paths, without human intervention. The results are averaged over 25 runs. 4. FUTURE WORK There have been issues observed previously in the Indian airspace caused due to some weak links in the administra- tion. For instance, a well-known public carrier, in 2008, was to land at Hyderabad airport but was allegedly unaware of the new Shamshabad airport inaugurated only a few days before, leading the confused pilots to fly the aircraft to Delhi and then Mumbai (reported in the News [4] and also docu- mented on Wikipedia). Such incidences can be avoided given that a secure information-exchange protocol can ensure re- laying of important notifications across aircrafts. Also, once the system gets into operation, over time it could learn by it- self the various traffic patterns observed over a single day for an airspace and bring about an improvement in the overall performance thus optimizing usage of invaluable resources. Also we could have more well defined reward mechanisms [6] implemented for the agents improving their performance as time progresses. 5. CONCLUSION A fundamental element of this research has been to investi- gate the air traffic operational environments and the effect of and on all the humans in the loop. The need for automat- ing certain tasks and still leaving the monitoring part to the ground air traffic controller has been made visible. The mul- tiagent system and its internal details are discussed. The simulations are performed on the basis of work presented previously in literature. We do not expect the system to yield readily acceptable results in any given scenario. It needs to be rigorously tested to maintain its performance graph. Despite these limitations, we believe that our propo- sition and its demonstration clearly and quite precisely ex- plain the handling of all the nuances in the system and are customizable with relative ease of intuition. We are confi- dent that it can be extended to a variety of similar real-world problems. 6. REFERENCES [1] Aircraft charter world : http://www.aircraft-charter-world.com/. January/February 2009. [2] W. Stirling. J.C. Hill, J.K. Archibald and R. Frost. A multi-agent system architecture for distributed air traffic control. USA. American Institute of Aeronautics and Astronautics. [3] F. Legras and C. Tessier. Lotto: Group formation by overhearing in large teams. LNAI, 12(2922):254 -- 270, 2004. [4] R. Niles. KLM: It knows all about shamshabad : http://www.avweb.com/avwebflash/news/. January/February 2009. [5] Baloch Fariha, Ragothaman Vijay and Pendse. Ravi. Unassisted aircraft landing via co-operative data exchange. In Proceedings of the 3rd Annual GRASP Symposium. Wichita State University, 2007. [6] Kagan. Tumer and Adrian. Agogino. Improving air traffic management with a learning multiagent system. IEEE Intelligent Systems, 24(1):18 -- 22, January/February 2009.
0805.1785	1	0805	2008-05-13T06:32:12	Distributed Self Management for Distributed Security Systems	[ "cs.MA", "cs.AI" ]	Distributed system as e.g. artificial immune systems, complex adaptive systems, or multi-agent systems are widely used in Computer Science, e.g. for network security, optimisations, or simulations. In these systems, small entities move through the network and perform certain tasks. At some time, the entities move to another place and require therefore information where to move is most profitable. Common used systems do not provide any information or use a centralised approach where a center delegates the entities. This article discusses whether small information about the neighbours enhances the performance of the overall system or not. Therefore, two information-protocols are introduced and analysed. In addition, the protocols are implemented and tested using the artificial immune system SANA that protects a network against intrusions.	cs.MA	cs	Distributed Self Management for Distributed Security Systems Michael Hilker University of Luxembourg Faculty of Science, Technology, and Communications 6, Rue Richard Coudenhove-Kalergi, L-1359 Luxembourg [email protected] Abstract. Distributed system as e.g. artificial immune systems, com- plex adaptive systems, or multi-agent systems are widely used in Com- puter Science, e.g. for network security, optimisations, or simulations. In these systems, small entities move through the network and perform cer- tain tasks. At some time, the entities move to another place and require therefore information where to move is most profitable. Common used systems do not provide any information or use a centralised approach where a center delegates the entities. This article discusses whether small information about the neighbours enhances the performance of the over- all system or not. Therefore, two information-protocols are introduced and analysed. In addition, the protocols are implemented and tested us- ing the artificial immune system SANA that protects a network against intrusions. 1 Introduction Distributed systems are widely used in computer science in order to solve prob- lems distributed on several nodes. Examples are simulation or optimisation prob- lems, which run on several nodes in parallel or the network security where a network is secured against intrusion like e.g. viruses, worms, and trojans. These distributed systems are e.g. multi-agent systems [1], complex adaptive systems [2], and artificial immune systems [3,4,5]. These systems use small entities - e.g. agents or artificial cells - that move through the network, perform certain tasks for network security, and work mostly autonomously. These entities receive local information about the situation in the network and contain a limited knowledge about the overall goal - here secure a network against intrusions. An example is that an entity knows how to identify one certain intrusion and receives the status of the node where the entity resides [6]. The performance of the overall systems originates from the performance of all the small entities, which leads to waste of computational power and decreased performance of the overall system when the entities are not properly distributed over all nodes. A key problem in the distributed systems is that the entities work au- tonomously and have to decide in some timestep whether they move to another node and if yes to which node - this problem is part of the self-management of a distributed system. This problem is significant because the distributed security system should secure the whole network and all nodes should have at least a cer- tain number of entities. This article discusses whether exchanging a small piece of information about the nearby situation increases the performance of the secu- rity system or not - local information about a node with some additional global information about the nearby region of the network. Therefore, two information protocols describing how and which information is exchanged are introduced. Thereafter, the two approaches are implemented in an artificial immune system that uses autonomous artificial cells in order to secure a network against intru- sions. Lastly, some discussions about the test results and the analysis conclude the paper. The domain of the application is network security. It describes the research in order to secure a network against intrusions. Intrusions are e.g. viruses, worms, and trojans. There are several approaches in order to secure a network against intrusions, e.g. antivirus software, packet filters, and intrusion detection systems [7]. Most of these common used security systems lack from the local installation because then they only secure one network node and require lots of compu- tational power to check the whole network traffic. Furthermore, the security solutions for a network consist of several systems, e.g. antivirus software in each node, packet filters in each router, and intrusion detection systems at the internet gateway and email server. However, these systems are not connected and do not work together. In order to protect a network against the upcoming intrusions, a more sophisticated approach is required where an example are the distributed systems. Thus, a distributed approach is chosen in the artificial immune system SANA [4] where artificial cells perform the required tasks and build up a dis- tributed intrusion detection system. These cells - the entities - move through the network and require the information where to roam next. The next section intro- duces the existing approaches for sharing movement information in distributed systems. 2 Current situation Currently, there are two different approaches in order to manage the movement of the entities in a distributed system: 1. The entities do not work fully autonomously. A centralised system, which knows the situation of the whole network, delegates the entities - mostly agents - to their destinations. Thus, the system is again dependent on a centralised system, which is also a single point of failure and is a contradiction to the idea of a distributed system. 2. When the entities work fully autonomously, in most systems they receive no information when and where to move. The movement is normally modelled using some probabilistic events with a distribution over all neighbour nodes. Both systems have significant disadvantages. The centralised approach re- quires lots of bandwidth because the center needs information from all over the 2 network and the center is a single point of failure. The distributed approach reduces the bandwidth significantly but it is possible that the entities move to inappropriate nodes where e.g., some nodes have too many entities and other nodes suffer from the low number of entities. Consequently, this article discusses how small pieces of information can help the entities to do the right decision. 3 Information Management and Protocols For investigating research in the information management, the following situa- tion in the network is assumed: An artificial immune system [4] with lots of artificial cells - the entities - pro- tects a network against intrusions [6]. Therefore, a type of artificial cell knows how to detect and remove exactly one intrusion where there are several instances of this type of artificial cells, cp. the immune cells in the human immune sys- tem [8]. The whole system does not contain any centralised system and the artificial cells decide autonomously when and where to move. The artificial im- mune system presents each packet to all artificial cells in this node and the cells decide whether the packet contains an intrusion or not. These cells should be distributed over all nodes so that each node contains approximately the same number of cells. In addition, the artificial immune system contains artificial cells that move through the network and check the nodes whether they are infected by an intrusion or not; e.g. in looking for backdoors or significant files and pro- cesses on the node. These artificial cells require the information where to go because the checks should be done uniformly in time distributed over all nodes; this means that each node is checked regularly but not too often. This prob- lem is the same as the general problem of the entities moving to another node. The artificial cells are the entities and the infrastructure provides the required resources for enabling the information management. For the first problem - the movement of artificial cells that check packets - the notification of insufficient security is introduced and for the second problem - the movement of artificial cells that check nodes - the trails of entities is intro- duced. In both approaches, the required resources must be limited. In the first approach, only one network packet is at most required for each connection in the network. The second approach requires only some storage in the network nodes that depends on the number of different types of artificial cells. Furthermore, the artificial cells receive some small pieces of information and, henceforth, do not increase that much in size and complexity. 3.1 Notification of insufficient Security In this approach, the problem of entities that check packets is discussed [9]. In general, this approach can be used for all entities, which perform regularly cer- tain tasks and which are required in lots of nodes of the network. The approach is motivated by the cytokines and hormones of the humen immune system reg- ulating the immune response and its immune cells motivate this approach [8]. 3 Fig. 1. Distribution of Notification: The node marked with N has an insufficient security level and notifies this to its neighbours. The receiving time of this in- formation is indicated using the colour of the nodes where the more black the earlier the information arrives. The nodes without a colour do not receive the information because they are too far away. For detecting that a node n has an insufficient number of entities, it has a pa- rameter that represents the minimum number of required security min secnode. Each entity returns on demand how many security it provides: secentity. The node n asks every time step all entities residing in this node how many security sece; a weighted each entity provides and sums this up: sectime = sum for a more sophisticated calculation can be also used. If the security in this time step is below the required security min secn > sectime, the node n starts a notification process in order to allure nearby entities. e∈entity node (cid:80) This notification works as follows: 1. Using local communication, the node informs the entities that are working in this node about the insufficient security level. These entities are then more affine to stay than to move. However, the entities are not forced to stay, they still decide on their own. 4 2. The node furthermore sends a network packet to all neighbours that indicate the low security level and the packet also describes the missing level. 3. The nodes receiving this packet forwards it to all entities working in these nodes. These entities then get more affine to move to this node than to stay or to move to another node. The change of the affinity is dependent on the level of lacking security that is stored in the packet. 4. With this workflow, each node maybe receives each time step several packets describing insufficient security levels of several nodes. Therefore, the node will forward only the packet with the highest missing security level. 5. In order to receive a global notification, each node forwards these notifica- tion packets to their neighbours. The value describing the missing value is therefore decreased by a decreasing function fd and if the value is still above a threshold, the packet is forwarded to all neighbours with the decreased value. 6. The notification about insufficient security in a node is not stored. The distribution of the information that a node has insufficient security level is visualised in figure 1. The entities - here artificial cells - react to this information as follows. Each entity has a function fm that describes the probability that it moves to an- other node. Another function fw describes where the entity moves next. These functions can have several parameters as the time the entity did not move, the network load, and lacking security level. Therefore, each time step a probabilistic event is done that succeeds with probability fm and if this event succeeds a sec- ond probabilistic event using fw describes the node where the entity will move. When the entity receives a notification that a node has an insufficient value, the function fm will be increased in order to increase the affinity to move and the second function fw is so changed that the entity will always move towards the node with insufficient security. The function fm is visualised in figure 2. This approach is implemented in the artificial immune system SANA that is developed at the University of Luxembourg for the purpose of network security (section 6). Each node with low security level sends the notification each time step and the entities react to it. In the simulations, the functions are set as follows: – min secnode is set to 20 where there are 60 different types of artificial cells in the artificial immune system – secentity is the number of known intrusions, mostly 1 – the decreasing function fd is set to old value − 1 – if a node has an insufficient security level the artificial cells in this node will not leave it – the movement function fm is shown in figure 2 – the where to move function fw is only used if there is not any notification, otherwise the entities move towards this node 5 Fig. 2. Response function fm of an entity: fm(0) describe the normal affinity to move; fm(x) describes the moving probability when lacking security x is received; maximum limits response to prevent overreaction. 3.2 Trails of Entities This approach provides information to entities performing regular checks in nodes as identifying infected nodes. In general, the approach can be used for entities performing a tasks that must be done regularly but not too often in a node. Therefore, an approach is used that is motivated by artificial ant colonies [10]. Each node n contains storage for each outgoing connection (v, w). In this storage, for each type of entity e is a value stored vv,w,e that describes when the last check took place. When an entity does a check in the node and moves to the next node, it increases this value vv,w,e and every time step the node decreases this value vv,w,e until it reaches zero. The entities in this node use this information in order to evaluate where to move next. Assume the following situation: an entity is in node n and finishes the check. The next step is to move to another node and the entity receives information when an entity of the same type moves to the neighbours, i.e. the values vv,w,e for each (v, w) in n for type e. Then, it uses this information for a probabilistic event where the P r(move to neighbour ni) depends on the value that describes when the node ni was lastly checked. In addition, the value is not a counting of time steps. For the increasing of the value is a function fi and for the decrease a function fd used. Then it is possible to when the last check did place at the neighbouring node as well as how often in the past the node was checked by this type of artificial cell. An example of fi is discussed in figure 3. This approach is also implemented in the artificial immune system where the storage uses float-values. The increasing function fi is set to c1+c2·exp(old value) 1000. The decreasing with c1 and c2 as a constant value, e.g. c1 = 10 and c2 = 1 6 function fd is set to old value − c3 where c3 is a constant value, e.g. 2. Thus, the value increases exponentially and decreases linearly. The probabilistic event sets the probability for each node dependent on the required security in the neighbour nodes. 4 Testing Results As already explained, the two techniques are implemented in an artificial im- mune system. The simulations are done using a network simulator implementing a packet-switching network. The network is a company-like network with about 500 nodes with different server, workstations, and services. An Internet gateway connects the network to the Internet and different intrusions are implemented, which are mostly viruses, worms, and trojans arriving over the network or di- rectly by internal attacks. The results are discussed in the next two subsections: 4.1 Notification of insufficient Security The notification of insufficient security works more than acceptable. When the parameters and functions are chosen properly, the system adapts quickly so that nodes with insufficient security receive artificial cells quickly. Furthermore, the system does not overreact which means that the artificial cells react too heavily and all cells move immediately to a node with insufficient security. Then, the alternating status is prevented: this means that all cells oscillate between two nodes because one of these two nodes has always an insufficient security level. The required resources are limited. For each connection is exactly one net- work packet for each time step used. The required computational power is also limited because the processing of the information is quite fast. The performance increase of the artificial immune system is observed. The performance with no- tification of insufficient security levels is significant higher than in the version without enabled notification. For example in one scenario, about 80% of the intrusions are identifies without the usage of the notification system; this low value occurs because some nodes have too much and some nodes too few artificial cells. When the notification system is enabled, the system identifies about 98% of the intrusions because all nodes have a guaranteed security value. Compared to the centralised approach with an entity management server, the notification of insufficient security requires about 80% less bandwidth and the performance increases by 15% due to the quicker adaption to weak points in the network. The reason for this is that the system quickly solves security problems and the attackers cannot use nodes with insufficient security in order to install intrusions. Additionally, the system distributes the artificial cells in a better way. If there is a node with too many artificial cells that requires too much computational power, there are also nodes with insufficient security and the artificial cells will solve this problem using the information. Furthermore, it is possible to adapt the system to the current situation using changes of the parameters. In addition, important nodes are secured with more entities, which is configured through changes in the notification system. 7 Fig. 3. Trails: an entity in the left node wants to move, reads therefore the trails, and receives the values indicating the trail. Thereafter, it makes a probabilistic event where it chooses uniformly a number from [1, 289]. If the number is in [1, 100] it moves to the node with trail value 1. Else, if the number is in [101, 197], it moves to the node with trail value 5. Else, if the number is in [198, 288], it moves to the node with trail value 10. Else, if the number is 289, it moves to the node with trail value 100. This probabilistic event reflects the trail values of the neighbouring nodes. 4.2 Trails of Entities The trails of entities work well. The system provides information to the entities where their tasks are required and the entities move there. The entities also save computational power because between two checks is a certain time period. In our simulations, without the trails only about 70% of the nodes are checked and some nodes are very often checked about every second time-step. With the trails systems, all nodes are regularly checked and too often checks are successfully prevented where the required computational power is reduced by 60%. Unfortunately, when the network is fragmented in several parts and these parts are connected by a few connections, the trails of entities do not work properly. The reason for this is that many entities must use these connections, they receive a high value, and the probability to move over these connections is reduced. Fortunately, this problem can be solved in using other functions in these nodes: one solution is to use a decreasing function with a significant higher gradient or to not use the trails in this node, which results to a probability of #neighbours for all neighbours. 1 5 Analysis In the analysis of the two techniques, the focus lays on the question how the selected information helps the entities and what other information might be 8 interesting. In addition, the required resources are analysed and a balancing be- tween these two parameters are done. Definition: Local information comes from the network node and is not sent over the network. Global information describes the situation in another node and is sent over the network. Question: Are the information locally or globally? In the notification of insufficient security, the information is local and global because the entities receive the information from the neighbouring nodes directly and using forwarding from nodes that are nearby. However, using feasible pa- rameters, the entities do not receive information about insufficient security in nodes that are far away. The reason for this is that if they receive information about all nodes with insufficient security, the required resources are equal to the centralised approaches. Using the parameters, it is possible to trade off the proportion between the required bandwidth and the distance where information are forwarded. In the trails of entities, the information is locally. There is no exchange be- tween the nodes about this information and the entities receive only the infor- mation of a node. Question: Does information that is more global increase the performance? In the notification of insufficient security, using the parameters it is possible to choose between required bandwidth and global information. However, when the system is so configured that the cells receive information about all nodes with insufficient security, the entities receive too many information and will travel too often through the network. Furthermore, it is uninteresting that an entity moves through the whole network only to fix a problem at a far-away node. Then it is more interesting to solve the problem in the area around the node where only localised information is required. The tests showed that a local setting with a distribution of some hops results in the best performance and limits the bandwidth greatly. In addition, the distribution area of the notification depends on the missing security value: the bigger the missing security level, the greater is the distribution area, which is a more adaptive approach. In the trails of entities, some global information would increase the perfor- mance. Especially in the situation explained in the last section that connections with high load will not be traversed by entities, a view to other nodes would make the decision of an entity where to move more sophisticated. Then, the en- tities would receive a more global view about the nearby network. However, then the system must invest more bandwidth and the selection which information to store in which node is quite difficult. Question: Which other information are interesting to use for these two problems? For the protocol of notification of insufficient security, it would be interesting to know how many artificial cells are on the way towards this node. Furthermore 9 it would be interesting to receive information about more than one node so that the system does not only adapt to the node which has the highest value of lacking security value. However, both would implicate lots of bandwidth and high storage space for saving the information. Furthermore, the workflow in the entities would be more complex. For the protocol of trails of entities, it would be interesting to receive infor- mation about a part of a network and not only from the neighbours of this node. However, this would again increase the required bandwidth and storage space. Question: Which other problems occur in the self-management of a distributed security system? In the notification of insufficient security, the following questions are also of interest: – Which node has too many entities. Too many entities is defined so that the resources - mainly computational power - is not sufficient with the required resources from all entities. It is highly interesting that these entities move to the nodes with insufficient security level in order to balance the system. – Is it possible to design moving chains where several entities move from node to node and after one time step all nodes have a sufficient security level? If yes, which information is required so that all entities can initiate this moving? For the trails of entities, the following questions are also interesting: – How to use these trails so that not too many checks are performed on a node? If too many checks are performed simultaneous, the node has to invest too much computational power. – Is it maybe possible to use these trails in order to find nodes that are con- trolled by an intrusion in absorbing cells? The problem in distributed security systems is that an attacker occupies a node and has access to the security systems. Then, the attacker can remove all incoming entities and after some time the system will lack of too few entities. The system can analyse the trail values in the nodes that are nearby the infected node and maybe there are association between an occupied node and trail values. These questions will be the challenges for the future research in distributed self management of distributed security systems. 6 Applications Currently, the information management is used in the artificial immune system SANA, which protects a network against intrusions [9,11]. Therefore, SANA fa- cilitates lots of small artificial cells performing autonomously the tasks required for network security - the organic computing approach. Common used secu- rity components like antivirus softwares, firewalls, packet filters, and intrusion detection systems complement the system to benefit from their performance. 10 The artificial cell communication enables the collaboration between the different artificial cells where the infrastructure supplies the cells with additional informa- tion. The artificial lymph nodes implement a meeting and response point for a small part of the network and the central nativity and training stations generate and release continuously novel artificial cells including the newest security tech- niques. With the specialised artificial cells is it possible to implement nearly all approaches of network security. The in this article introduced self management organises the cells so that the system is always properly secured and there is no weak point in the network. The information management is more sophisticated implemented where the security components and artificial cells exchange con- tinuously status information in order to adapt the overall system and internal thresholds as well as parameters - implementation of the danger theory [12] and the tunable activation threshold [13] of the immune cells. For the response when an intrusion is found, the components develop a defense strategy and perform this or quarantine the node with informing the administrator for further tasks. Maintenance - updates, extensions, and regular self-checks - are performed by the system using specialised cells where the maintenance is automated and not proper working as well as outdated components are reported to the administra- tor. However, the administrator is able to access and configure all nodes and components of the system through an administration GUI. Infected nodes are identified through regular checks as well as through distributed analysis of data gathered by various security components, i.e. the distributed analysis of network packets through intrusion detection systems and artificial cells [6]. Due tos the usage of common used components, the performance is at least as good as of common used security systems. Due to the dynamic artificial cells is the system hard to predict and to attack, contains no single point of failure, and adapts to the current situation in the network. In addition, a real distributed intru- sion detection system is implemented in all nodes with simplified administration workflows [9]. Furthermore, the notification of low security can be used in nearly all dis- tributed systems where the number of entities can be measured and thereafter modeled using the parameters and functions. The trails can be used in all scenar- ios where a type of entity should visit all nodes regularly but not too frequently. These systems are most distributed systems as the multi-agent systems, the mas- sively distributed systems, complex adaptive systems, and the artificial immune systems. 7 Conclusion This article discusses two information protocols for distributed systems and analyses the performance. The two protocols increase the performance of a dis- tributed system enormously because the entities receive information where their tasks are required. In addition, the required bandwidth is limited because a cen- tralised system is not used. The next steps are to introduce more information protocols and evaluate theoretically how the performance would increase when 11 more bandwidth is invested. Furthermore, the enhancing of the distributed se- curity systems that base on the artificial immune system as archetype is a next step. Acknowledgments The PhD-Project SANA - Security Analysis in iNternet trAffic - is financially supported by the University of Luxembourg. I like to thank Christoph Schommer for helpful discussions and the Ministre Luxembourgeois de l'education et de la recherche for additional financial support. References 1. Bellifemine, F., Poggi, A., Rimassa, G.: JADE - a FIPA-compliant agent frame- work. In: Proceedings of the Practical Applications of Intelligent Agents. (1999) 2. Vargas, P.A., de Castro, L., von Zuben, F.: Artificial immune systems as complex adaptive systems. In: Proceedings of the 1st International Conference on Artificial Immune Systems (ICARIS). Volume 1. (2002) 115–123 3. Aickelin, U., Greensmith, J., Twycross, J.: Immune system approaches to intrusion detection - a review. Proceedings of the 3rd International Conference on Artificial Immune Systems (ICARIS 2004) (2004) 4. Hilker, M., Schommer, C.: Sana - security analysis in internet traffic through artificial immune systems. Proceedings of the Trustworthy Software Workshop, Saarbruecken, Germany (2006) 5. Hofmeyr, S.A., Forrest, S.: Immunity by design: An artificial immune system. Proceedings of the Genetic and Evolutionary Computation Conference 2 (1999) 1289–1296 6. Hilker, M., Schommer, C.: Agnosco - identification of infected nodes with artificial ant colonies. Proceedings of RASC (2006) 7. Roesch, M.: Snort - lightweight intrusion detection for networks. LISA 13 (1999) 229–238 8. Janeway, C.A., Travers, P., Walport, M., Shlomchik, M.: Immunobiology: the Immune System in Health and Disease. Sixth edn. Garland Publishing (2004) 9. Hilker, M., Schommer, C.: Description of bad-signatures for network intrusion detection. AISW-NetSec 2006 during ACSW 2006, CRPIT 54 (2006) 10. Dorigo, M., Maniezzo, V., Colorni, A.: The Ant System: Optimization by a colony of cooperating agents. IEEE Transactions on Systems, Man, and Cybernetics Part B: Cybernetics 26(1) (1996) 29–41 11. Hilker, M., Schommer, C.: Sana - network protection through artificial immunity. In: Proceedings of the 2nd International Workshop on Theory of Computer Viruses (TCV 2007). (2007) 12. Aickelin, U., Bentley, P., Cayzer, S., Kim, J., McLeod, J.: Danger theory: The link between ais and ids? In: Proceedings of the Second International Conference on Artificial Immune Systems (ICARIS 2003). LNCS 2787, pp. 147-155, Edinburgh, UK (2003) 13. Grossman, Z., Singer, A.: Tuning of activation thresholds explains flexibility in the selection and development of t cells in the thymus. Proceedings of the National Academy of Sciences 93 (1996) 14747–14752 12
1811.01683	1	1811	2018-11-05T13:53:10	Supply Chain Management analysis: a simulation approach of the Value Chain Operations Reference model (VCOR)	[ "cs.MA" ]	The impact of globalization and worldwide competition has forced firms to modify their strategies towards a real time operation with respect to customer's requirements. This behaviour, together with the communication possibilities offered by the actual Information and Communication Technologies, allows the top management to move towards the concept of extended enterprise in which a collaborative link is established among suppliers, commercial partners and customers. When the information flows involve each actor of the chain, from suppliers to the final distribution centers, the extended enterprise becomes a virtual firm, that can be defined as a set of stand-alone operational units that acts to reconfigure themselves as a value chain in order to adapt to the business opportunities given by the market. The present work is intended to verify through a simulation approach the quantitative advantages that can be obtained by the introduction of the Value Chain concept into the Supply Chain Management (SCM). The paper, after a description of the two most known (SCM) methods - SCOR and VCOR - makes a comparison between them by the customer's point of view. In the second part of the work a simulation model has been developed to verify the advantage that the VCOR is able to obtain, validating it on an industrial case study.	cs.MA	cs	Supply Chain Management analysis: a simulation approach of the Value Chain Operations Reference model (VCOR) Yacine Ouzrout * LIESP-Lyon2 Laboratory, Université of Lyon 2, 160 Bd de l'université 69676 Bron Cedex, France, E-mail: [email protected] Phone: +33.478.774.484 *Corresponding author Matteo M. Savino Department of Engineering, University of Sannio, Piazza Roma, 21, 82100, Benevento, Italy, E-mail: [email protected] Abdelaziz Bouras LIESP-Lyon2 Laboratory, Université of Lyon 2, 160 Bd de l'université 69676 Bron Cedex, France E-mail: [email protected] Carlo Di Domenico Hermes Reply S.p.a, Corso Francia, 110, Torino, Italy. E-mail: [email protected] 1. Introduction Scientific literature related on Supply Chain Management (SCM) is rich of publications, but the reality is that there is a lag between practice and theory (Balan, et. al, 2006; Simchi-Levi et al, 2000; Neubert et al, 2004). Mass media and Internet have speeded up the diffusion of new products; at the same time, technical innovation and market severe competition promote rapid obsolescence of existing products and technologies. When a company succeeds in developing a new product category, other competitors may soon emerge. The market originator must endure not only the substantial risk of whether the market would materialize or not, but also the difficulty of recovering major costs, such as research and development and advertisements. Increasingly, the supply chain becomes the mechanism for coping with these problems because it is often inefficient for any single company to produce a whole product (Abdullah et al., 2004; Feller et al, 2006). Hence, modern business is essentially the competition of one supply chain (SC) with another. SC dynamics is the interaction processes of the participants from different departments and companies. A positive aspect of supply chain dynamics is effective collaboration, which may lead to higher performance. A negative aspect is independent decision making, which may create various delays and aggravate forecasting errors (Simatupang and Sridhran, 2002; Abu- Suleiman et al. 2005). This research stresses the interest on the possible quantitative advantages given by the introduction of the Value Chain concept into the SCM through a simulation approach. Discrete event simulation, continuous time-differential equations, discrete time difference models and operational research techniques are some of the commonly used quantitative modelling techniques to evaluate and design supply chains (Lee et al., 2002; Terzi and Cavalieri, 2004). The correct identification of key variables and their interactions, together with determining how the information can be better managed enables the utility of the entire supply chain. In this work we use discrete event formalism to model and study the supply chains. Discrete-event simulation is chosen for its capability to represent physical and information flows along with their respective delays, in an information feedback control type of setting. Our main research interest is to clarify the critical factors for minimizing the negative effects of supply chain dynamics and to gain insight on how to effectively manage them. To achieve these objectives we developed a simulation model to implement the VCOR and to verify the possible advantages that it is able to obtain. 2. The Value Chain management theory A Supply Chain can be defined as a system network that provides raw materials, transforms them into intermediate commodities and/or in finished goods and distributes them to the customers through a delivery system (Christopher et al, 2002). The aim is to produce and distribute the right quantities, to the right locations, at the right time, while reducing costs and maintaining a high level of service. SCM is concerned with smoothness, economically driven operations and maximising value for the end customer through quality delivery. The limitations are however due to the fact that SCM as a concept does not extend far enough to capture customer's future needs and how these get addressed, and furthermore, it does not encompass the post- delivery, post-evaluation and relationship building aspects (Al-Mudimigh et al, 2004). Another important theory can be defined as strategic in the context of SCM, the concept of value chain management. The Value Chain was described and popularized by Michael Porter in his best-seller, Competitive Advantage: Creating and Sustaining Superior Performance (Porter, 1985). Porter defined Value as the amount that buyers are willing to pay for what a firm provides, and he conceived the Value Chain as the combination of generic activities operating within a firm, activities that work together to provide value to customers. Porter linked up the value chains between firms to form what he called a Value System. However, in the present era of greater outsourcing and collaboration, the linkage between multiple firms' value creating processes has more commonly become called as Value Chain. As this name implies, the primary focus in value chains is on the benefits that accrue to customers, the interdependent processes that generate value, and the resulting demand and funds flows that are created. Feller, Shunk, and Callarman (Feller at al. 2006; Jan Olhager et al, 2006) exposed some important considerations about the value concept. First is that value is a subjective experience that is dependent on context. The same product or service has not the same value in different parts of the world or in different situations. Second, value occurs when needs are met through the provision of products, resources, or services. Finally, value is an experience and it flows from the customer. Clemmer (1990) affirms that customer value is layered and has been described by three concentric rings. In the center ring is the product value which is the technical value derived from providing a source of supply. A second ring of service value is provided by the services that surround the product such as personal care and warranty service. The third ring has been called the new service/quality battleground, and was made popular by business thinkers such as Peters and Waterman ("In Search of Excellence"). This third level of value is achieved by providing enhanced service, to "make your customer successful" rather than just satisfied. The huge importance of focusing on the customer has forced the integration of the optimization techniques of the Supply Chain Management, Customer Relationship Management (CRM) and Product Lifecycle Management (PLM), as shown in figure 1. Figure 1 - Integration of PLM, SCM, and CRM from www.process-wizard.com The value chain theory focuses on the management of the product lifecycle (PLM) which is a strategic business approach that helps enterprises to achieve its business goals of reducing costs, improving quality and shortening time to market, contemporarily innovating its products, services, and business operation (CIMdata, 2002). By increasing an enterprise's flexibility to respond quickly to changing market pressures and competitors, the value chain helps companies to: deliver more innovative products and services; § § Reduce costs; § § Shorten time to market while achieving the targeted Return On Improve quality; Investment (ROI); § Establish more collaborative, and improved relationships with their customers and suppliers (Garetti et al, 2005). The value chain management includes also the Customer Relationship Management (CRM) which consists on the creation, the development, the palimony and the optimization of long period relationships more profitable among consumers and firm. The success of CRM is based on the understanding of the consumers' needs and desires, and it fulfils setting such desires to the center of the business, integrating them with the firm's strategy, the people, the technology and the business process (Munari, 2004). 3. Supply Chain and Value Chain Modeling The first general framework for SCM, the Supply Chain Reference Model (SCOR) was developed by the Supply Chain Council (SCOR, 2005); the model is rather general, defining the supply chain standard processes and establishing standard terminology in quite broad terms. SCOR spans customer and market interactions and the physical material transactions. This model can also help manufacturers to carry-out benchmarking against other well-established companies, for that the model proposes some best practices and Key Performance Indicators (KPI). The SCOR-model has been developed to describe the business activities associated with all phases of satisfying a customer's demand; it consists of a plan, source, make and deliver process elements (level 1 of the model) which revolve around the entire supply chain. The main assumption of the model is that by integrating the process elements along the supply chain, companies should become more competitive. But the support functions such as administration, R & D, and customer services are not included (Bolstorff and Rosenbaum, 2003). the European enterprises have realized Nowadays manufacturing in Europe is going on a deep change. The increase of the productivity has concentrated the attention on the approach for the achievement of competitive advantage through efficiency actions and process optimization. Besides, that a complementary kind of initiative is necessary for the product design and engineering processes. It is indispensable, in fact, to keep a better know-how in order to remain competitive and to be able to offer new and advanced products to the market. The introduction of concepts like PLM has become essential to acquire new clients and new market segments, and to adopt concepts of Value Chain Management all over the network. In late 2003 and early 2004 a series of meetings culminated in the development of the first iteration of the Value Chain Operations Reference (VCOR) model. Participants in these meetings came from a global pool of business process knowledge experts many of whom worked for large end-user, consulting or software companies, domain specific not-for-profit organizations, or academia; they created the Value Chain Group: VCG (Value Chain Group, 2005). The VCOR model is able to achieve some benefits to firms summarized as follows: • A standard based approach to define essential collaborations between trading partners. • Agreement on product life cycle objectives and how to achieve them. • Reusable process templates based on best practices. • Integration of existing and new information management systems. • A fast response to changes while maintaining and extending value chain performance. 3.1. The VCOR Framework to their The VCOR model was developed from the perspective of being a value chain framework with the development of Seven Performance Attributes linking the three domains of Product Development, SCM and Customer Chain across the supply networks together. The structure of VCOR model supports and enables companies three critical domains: Product Developments, Supply Network Integration and Customer Success, using one reference model to support the vision of an integrated value chain (Heinzel, 2005). To achieve this goal VCOR uses a "process based, common language" of syntax and semantics while, at the same time, create a base for the successful Service Oriented Architecture Game Plan. The main objective of this model is to increase the performance of the total chain and support the current evolution; for that, it proposes four different modeling layers: integrate Strategic Level: The Top Level of the model includes all the high level processes in Value Chains and is represented through the Process Categories Plan -- Govern -- Execute. The Level is defined to be the Strategic Level of the Model, meaning that this is where high level decisions are made regarding how to gain a competitive advantage for the Value Chain in scope. The VCOR Strategic Level has three Macro-processes: Plan: it balances the current strategic objectives with current asset status and produces decisions on activities to move the organization toward the goals; Govern: decision based process which identifies and enables a value chain by establishing the rules, policies and procedures to control the implementation of Plan and Execute processes within the Value Chain; Execute: it transforms the customer requirements to production processes. The Execute Processes operate within the limits of the Management criteria and to the parameters defined by the Plan Processes. Tactical Level: The second level of the model contains "abstract" processes decomposed from the Strategic Level to implement and fulfill the strategic goals set in the top level of the model hierarchy a set of tactics needs to be developed and realized. The Tactical Level can be described being instituted for "Horizontal Value Chain Process Re-Engineering". The VCOR model processes decomposes from Strategic to the Tactical Level with Plan and Govern, keeping their respective naming in the first part of the process notations on this level as they influence each of the Execution Processes. The figure 2 illustrates the decomposition of the strategic level into component processes. The Market Process has deep impact on the entire chain and for this reason is extended over the diagram. The Plan and Govern processes have the same name (e.g. Plan Research, Govern Develop and so on). To describe these different processes, VCOR defines three main groups: Market, Research & Develop, Acquire, Build & Fulfill, and Sell & Support. Figure 2- VCOR Layers Diagram from EMEA-VCG Operational Level: The third level of the model represents specific processes in the value chain related to actual activities being executed clarifies. On this level focus is usually vertical business process improvements or business process re-engineering as usually it is named. In a value chain perspective this is the level where fine-tuning occurs. 3.2. VCOR vs. SCOR: Metrics and Performances In order to measure the performance of the chain, the Supply Chain Council has introduced five metrics (see table 1) and KPI (Key Performance Indicators) in the first test supply chain reliability, responsiveness, flexibility, costs and efficiency in managing assets (working and fixed capital). For the remaining processes (level two) and elements (level three) there are specific KPI to test the performance of each part of the SC configuration. Moreover, for each process, some best practices are defined to simplify the analysis of the chain. These are reliable points of reference, given by the experience of enterprises leader in their own field, to follow in order to improve the performance without trying unpredictable strategies that could be dangerous. A company cannot be best in all metrics, so it should wisely target its strength and differentiates itself in the market, while ensuring that it stays competitive in the others (Bolstorff and Rosenbaum, 2003; Hausman, 2002) Actually, the most part of companies do not choose to improve all indicators but they focus on some of these, building their strength points. According to the definition of Value Chain Group, a metric is "a quantifiable variable that reflects a specific state of business performance during process execution within a strategic value chain context". In VCOR a metric is characterized by different features: Metric Name, Metric Definition, Priority Dimension, Metric Class & Sub-Class, Formula, Input Requirements, Dimension, Calculation Rules etc. The table 1 describes the definition of the seven dimensions defined in the VCOR model and compares them to the SCOR's model definition. level of the SCOR model to Table 1 SCOR strategic performance metrics vs. VCOR priority dimension SCOR Definition: The performance of the SC in delivering: the correct Delivery/ Reliability product, to the correct place, at the correct time, in the correct condition, packaging and quantity, with the correct documentation, to the correct customer. Metrics: Delivery performance; Fill rates; Perfect order fulfillment. VCOR Description: The ability to deliver the correct product to the correct market and customer on time. Metrics: Deliver Performance; Request Date; Product Release Variance; Forecast Accuracy. SCOR Definition: The velocity at which a SC provides products to customer. Metrics: Order fulfillment lead times. VCOR Description: The cycle time taken to deliver a product or service Responsiveness/ to the customer. Velocity Metrics: Order Fulfillment Lead Time; Product Development Lead Time. SCOR Definition: The agility of a supply chain in responding to Flexibility/ marketplace changes to gain or maintain competitive advantage. Adaptability Metrics: SC response time; Production flexibility. VCOR Description: The capability in responding to market change to gain Costs or maintain competitive advantage. Metrics: Delivery Adaptability; Value Ch. Agility; Ideation Yield. SCOR Definition: The costs associated with operating the SC. Metrics: Cost of goods sold; Total SCM cost; Value-added employee productivity Warranty/return processing costs. VCOR Description: The cost associated with operating a value chain. Metrics: Cost of Quality, Design Cost Ratio; Logistic Cost Ratio; Manufacturing, Sales & Marketing Cost Ratio. SCOR Definition: The effectiveness of an organization in managing assets to Effectiveness / Asset support demand satisfaction (Fixed and working capital). Metrics: Cash-to-cash cycle time; Inventory supply; Asset turns. VCOR Description: The effectiveness of an organization in managing assets of the Value Ch. to support market and customer satisfaction. Metrics: Asset Turnover; Cash Conversion Cycle; Design Realization; Inventory Supply. Metrics: - SCOR Definition: - VCOR Description: The ability to strategically leverage internal & external sources of ideas and introduce them to market through multiple paths. Metrics: Product Innovation Index; R&D Profit Contribution. SCOR Definition: - VCOR Description: The capability to develop positive collaborative Metrics: - Innovation Customer customer relationships. Metrics: Customer Growth Rate &Retention Rate, Market Share. 3.3. VCOR vs. SCOR: Synthesis The two models have been created in order to give a reference point to those companies that applied the principles of SCM and Value Chain Management. The SCOR is an affirmed and reliable model and has preceded the VCOR by few years. For this reason they have some analogies inherent in the structure and differences in their scope. Both models tried to be standards, improved by Supply Chain Council and Value Chain Group respectively, to simplify their correct realization in real cases and facilitate the communication between trading partners. Moreover, they are based on high-level generic process categories to fit better for all types of firms. But the feature that perhaps is the reason of their success is the capability to calculate the level of performance for every single process of the chain in a very simple way and, at the same time, to introduce best practices to facilitate the management of productive processes. It is crucial to found strategic choices on objective data that can be compared with the metrics filed during years of activity or with performance indicators of other companies if possible. Differences between the SCOR and VCOR are that VCOR is an enterprise model using a framework and taxonomy that features governance and the decision making processes at its highest levels with interconnectivity to all Enterprise processes. The interaction of process elements in the supply chain domain are mostly "transactional" as opposed to interaction of Enterprise business process elements involving higher order information processing in the decision making process. A question has arisen: does a company that successfully implements the SCOR model really needs to change framework facing other efforts? The answer is not trivial, but it is possible to state that the paybacks gained with the VCOR can benefit an enterprise enough to precede its own competitors, extending its market. Furthermore, the extension from SCOR to VCOR is not very intricate because VCOR's basic design uses the framework and methodology of a unified General Systems model that can be applied to most organizational type (Feller et al, 2006). Expansion occurs by reaching out into existing vertical domains of SCM, PLM and CRM to integrate their respective business process elements into a unified Value Chain model similar to SCOR. In this sense the VCOR can be considered the natural extension of the SCOR model. In the next sections we will describe a simulation model realized to allow companies to evaluate in advance the possible advantages of the Value Chain implementation. This model includes some of the metrics and KPI presented in the first part of the paper. 4. The simulation of the Value Chain processes In this section we front the problem of the development of a simulation architecture for the implementation of a VCOR model. Nowadays in literature there are few examples of global simulated Supply Chain because it is no easy to model the entire chain and implement it by specific and dedicated software (Lee et al, 2002; Kleijnen, 2005); moreover the number sensitively increases if we consider the examples that apply the SCOR (Boucher et al, 2003, Jain and Leong 2005, Herrmann et al, 2003). This section presents first the advantages of the simulation in comparison to the mathematical models and finally the realization of a generic architecture to simulate the VCOR model. Many researchers tried to solve the production- distribution issues in the form of analytical problems. The objective in this type of method is to minimize the overall production and distribution costs in multi- facility, multi-product and multi-period problems. The algorithms are mainly based on heuristics or on network flow; but it is obvious that the complexity of the problem grows in function of the number of elements in the chain: number of products, resources, constraints... if we also introduce the uncertainty the modelling and the problem solving became prohibitive. In this context of dynamic, stochastic and complex systems it is difficult to analytically model the problem, but the limits of the mathematical analysing can be solved by computer simulation. The benefits in using simulation in Supply Chains can be summarized as follow (Kelton et al, 2004): • Capacity to capture data for analysis: users may model unexpected • • • • events in certain areas and understand the impact of these to the SC It can decrease drastically the risks inherent to changes in planning: users may test several alternatives before making the planning change. Investigating the impact of: innovations within the SC; eliminating an existing infrastructure or adding a new one within the SC; strategic operational changes, such as process, location and use of new facilities, the fusion or the separations of some components of two SC; manufacturing products inside the company; creating new suppliers or subcontracting some processes; Investigating relations between suppliers and other components of the Supply Chains to rationalize the number and size of order lots, using as a basis the total of costs, quality, flexibility and responsibilities; Investigating the opportunities to decrease the varieties of product components and standardize them throughout the Supply Chains. These main reasons have lead us to choose a simulation approach to analyse the SC dynamic and behaviour and to define a simulation model to implement the VCOR processes. The next paragraphs describe the generic supply chain simulation model that we propose; from the configuration to the description of the different blocks in ARENA. 4.1. Model Implementation The objectives of this section are to describe the VCOR model simulation architecture and to illustrate the implementation of its different elements. Since we want to track not only the material flow but also the information and cost flow in order to obtain metrics and performance indicators, we propose to store all the simulation data in a database which leads us to save the parameters and input data for the initialization and the simulation execution. To better explain the application of VCOR to our chain we started from the level 1 of the model, i.e. the Top Level, following a top-down approach. Once defined the Macro-Processes involved in each element of the chain, we choose its configuration and later we depict the level 3 elements implicated in the process. To describe the model we will use the following example (figure 3). Figure 3 - Supply Chain Level 2 Configuration We started from the central element of the chain: only the Firm has a production process that transforms raw materials in finished goods (3 products). For this reason the Firm is the only participant that has a Make process in its configuration. The other members, like Retailer and Suppliers, have only package processes that can be simply included in the Deliver process. For some products the production follows the rule of the Make-to-Stock, whilst for other products a Make-to-Order production is assigned. These considerations are summarized in the figure 3: the processes P1, P2, P3 and P4 represent the development and the establishment of activities over specified time periods that represent a projected appropriation of supply chain resources (materials, production, and delivery) to meet supply chain requirements. One of the most important aims of the VCOR model is to describe and manage information, costs and material flows. In order to achieve this goal the simulation tool is based on the orders management. When any member needs a certain quantity of products it becomes a client and sends an order signal to its predecessor (i.e. the provider) in the chain. An order is characterized by its identifier, quantity of product, product type, timestamps and status. When a client makes an order, it updates the Demand database of its provider. Every time there is a change in the status of the order, the information is updated in the database. This mechanism allows us to exactly know the real position of commodities, monitoring both the material and the information flow. To build our simulation model we used the discrete event oriented ARENA software; the necessary SCOR/VCOR Level 3 elements have been realized in ARENA blocks and then gathered and organized in sub-models in order to set up the different processes. In the following paragraphs we describe, as example, the implementation of different actors of the chain in the model: a. Retailer: the retailer is composed by a Source and a Deliver process with two different configurations: Make-to-Stock and Make-to-Order. An Arena block is implemented with a rule which involves that each request made by the consumer becomes an order of the Retailer to the Firm. In the figure 4 some elements of the Source process are shown. The element called Schedule Product Deliveries S2.1 has the role to check periodically the three inventory levels stored into the database. If the effective level is under a "s" value (chosen by the managers), this module sends an order to the Firm, writing information as the "order id", "client id", "quantity", "product type", etc. in the Firm's Demand database. Figure 4 - Retailer VCOR Source process With the application of VCOR model, the Retailer configuration defines a new process: the Support process which has the objective to solve the Customers problems when a delivered lot is defective because of the transport or a bad installation. The figure 5 illustrates this process. Figure 5 - Retailer VCOR Support process The Manage Incident U2 has the role to register in the Support database a defective order, the related quantity of products returned, the customer identifier and the timestamps. The Resolve Problem U3 resolves the problem writing a new order in the Retailer Demand base with a quantity equal to the number of damaged products. Once the order is delivered the Monitor Experience U7 registers the operation and modifies a variable in the Customer Behaviour model that increments the vote. The Educate Customer U5 simulates a Call Center, it has the important task to decrement the percentage of defective products decrementing the value of a variable in a Customer ARENA module. b. The firm: The Firm has the role to transform raw materials coming from Suppliers into products for the Retailer. Its configuration includes a Source, a Deliver and a Make process. The following figure describes the ARENA representation of the "A5- Receive Order" process Figure 6 -- Firm's Receive Order process The Firm has also four processes derived by PLM and CRM of the Value Chain Management. The Market, Research and Develop processes belong to PLM, while the Sell and the Support to the CRM. With these new processes a company tries to analyze the market and takes consequently strategic decisions, like the restyling of a product or the acquisition of new technologies. Since it is very difficult to reproduce a complex market analysis or a restyling of a product, the ARENA blocks simulate these events in terms of required time and associated costs. The Analyze Market M1 (figure 7) periodically checks the Customer satisfaction level in the Retailer Demand database and, if the vote goes down under a specified threshold, it activates the Architect Solution M4 module. This block finds the product type that has the least number of sales and decides to adapt it. In order to change a product, the Firm needs to modify its production line with the introduction of new technology. Once acquired the new equipment with R3 and R7 modules, Introduce Technology R8 changes the production process in the VCOR Build Product B3 that corresponds to the M2.3 element. Figure 7 - Firm Market, Research, Sell and Develop processes The Develop process materially changes the Bill of Material of the product and it is responsible to launch the product. In fact, the D8 sets the variable "Innovation Factor" in the Customer Behaviour model and, as consequence, the customer's vote increases considerably. The Sell process tries to identify the clients in the market in order to develop relationships and proposals. The Qualify Target S2 classifies the Clients on the basis of their priority and determines which of them can be supplied by the Firm. It calculates the total number of products required by the Clients and finalizes the contracts (S8) that must not exceed the fixed percentage of the dedicated production capacity. 4.2. The Customer Behavior model In the simulation model, the Customers are able to express an opinion, in the form of a vote, which represents the satisfaction level for each single order delivered. This is necessary because it is at the base of the Value Chain philosophy in which the main goal is to give to the market what it needs, obtaining a feedback for the improvements introduced on the product and/or for the introduction of new products. For this reason the VCOR simulation needs a consumer behaviour model. Some models used questionnaires (Chen and Hughes, 2004), some others applied complex neural networks (Wen-Bao, 2007). In our simulation approach we have adopted the Adaptive Learning Model (Hopkins 2006), which consists in a mathematical model which is able to simulate the satisfaction level of the consumer and to pass the data to the corresponding simulation blocks. According to the Adaptive Learning Model two customers generally express their level of satisfaction through a vote for each type of product. This vote decrements its value in exponential way which is coherent if we consider that a product has a life cycle and that at the end it becomes obsolete. The consumer satisfaction can increase if the support service, introduced in the VCOR example, is able to resolve the clients' problems. The factor that can drastically increment the customer satisfaction level is the introduction of a new product as result of a market analysis. This is possible in the mathematical model thanks to an "Innovation Factor" that is enabled when an old product is modified to meet the demand of the market. Our model differs from the Hopkins's one for the following two main aspects: § § the introduction of a forgetting factor α in place of the coefficient δ, in order to model the behaviour of the customer to "forget" the improvement on a product with the time; the input function ui(k), which in the original model expresses the input of the customer, has been modified taking into account the following variables: - The product cost variation Δp: expressed in percentage, according to the market rules. The delay d: expressed in percentage with respect to the delivery lead time of the product. The Quality parameter q: expressed as a ration of conform products over the total product purchased (KPI). - - The adopted consumer behaviour model can be now expressed with Hopkins notation, taking into account our modifications: -=+ ( ) a1 ) 1 × ( ) kx i ×+ a ( )ku i (1) ( kx i Where: × I n s kf )( bx ×+×+×+D×+×+ hfd xi(k) : the vote at instant k of the i customer. α: forgetting factor (0<α<1) ui(k) the user input function In our model ui(k) has the following expression ( ) ku = i Where: s indicates if the last request of support was correctly satisfied or not. ξ takes into account how important is the quality of service. In is a Boolean parameter which indicates if the product delivered is a new one (1) or not (0). It changes its status to one when a new product is introduced in the market and to zero after a first order for this product is delivered. The purpose of the parameter is to increment the customer's vote. (2) ( )k q d p x j ¹ i = yi(k) is the output of the system (i.e. the vote): f(k) is a specific function depending on the vote given at k time. It is based on the specific industrial context and depends essentially on the customer satisfaction detection system adopted by the firm in its Quality Management System. In our application case a measurement system will be adopted with respect to the firm's customer satisfaction requirements. i ( )kx ( ) ky i 5. The case study Our value chain simulation model has been tested on a supply chain related to an agro-food main firm of the southern Italy producing fruit and tomato juices. This firm is involved in a supply chain and uses an Enterprise Resource Planning system (ERP) to integrate the information systems of its different sites. In this case study, we propose different scenarios related to three types of products (figure 8) which represent the 30% of the entire production of the firm and have a market share up to 15%. Figure 8 -The products chosen for the simulation test Here the SCOR we be also used to model and analyze the different processes of the chain. The firm receives fruits, vegetables and chemicals for product preservation from three different suppliers and provides their treatment, packaging and distribution to the retailers. The firm has a set of 10 big retailers that distribute the products to the final customers (supermarkets and small stores). In this study, we have selected only one Retailer with the highest number of customers (7 big supermarkets) and, among the customers, the two customers with the highest value of products sold (2 supermarkets that have a sale rate up to 20%). The supplier has one warehouse, while the firm and the retailer have two warehouses. The described supply chain can be structured as shown in figure 9, where the material flows can be seen too. Figure 9 - Supply Chain structure The Supplier 1 is specialized in the production of only one component (named Raw 1), the Supplier 3 for the Raw 3 component, while the Supplier 2 can supply the other two types of components (Raw 1 and Raw 2) The entire configuration of raw materials distribution necessary for the production of finished goods is given in the following table: Table 2 Production of raw materials by suppliers Production of raw materials Supplier 1 Supplier 2 Supplier 3 Raw 1 Raw 2 Raw 3 X X X X Based on a first analysis we obtained the data related to the demand of the three described products. To do this we have monitored the demand of the two customers during a period of one year, divided in observation periods of one month; the data retrieved are shown in tables 3 for the customer 1 and 2, in which Mi are the observation months. The reported values are in number of boxes for each product. We can see that for the product 1 the demand can be assumed as constant, while it is variable in a seasonal behavior for products 2 and 3. Table 3 Product demand for the Customer # 1 and Customer # 2 Dem. 1 Prod. 1 Prod. 2 Prod. 3 Dem. 2 Prod. 1 Prod. 2 Prod. 3 - - M1 M2 M3 M4 M5 M6 M7 M8 M9 M10 M11 M12 250 260 245 247 255 257 250 251 253 255 250 241 550 659 580 650 770 850 890 790 700 650 590 500 - M1 M2 M3 M4 M5 M6 M7 M8 M9 M10 M11 M12 300 310 312 295 311 320 301 305 313 300 295 297 - 70 165 140 145 250 355 397 410 380 371 280 210 - - - - - - - - - - - - - - - - - - - - . 5.1. SCOR/VCOR model Simulation The validation of the simulation model has been conducted with an experimental campaign, before with SCOR model and then with the extension of the VCOR ARENA modules. This section presents the most important results of the experiments conducted and some comparisons, with the real data of the firm. The unit time of the run is expressed in hours and the run length is two days. The most important parameters of the participant are shown in the table 4, where the numbers in square parenthesis refer respectively to product 1, 2 and 3. Table 4 Participant Retailer Firm Parameters Value used in the simulation Parameter Value Products Inventory Level [1, 2] [0, 0] Rescheduling Frequency Deliver 2 hours 2.5 hours Rescheduling Frequency Source Finished Goods Inventory (FGI) [500, 500, 300] Raw Inventory Level (Kg) [200, 200, 200] Rescheduling Frequency Deliver 2.5 hours Rescheduling Frequency Source Rescheduling Frequency Make Max Daily Production Capacity 3 hours 3 hours 185 Suppliers FGI Inventory Level Rescheduling Frequency Deliver 4 hours Rescheduling Frequency Source 4 hours 500 Figure 10 illustrates the Retailer Delivery Time obtained by the run of the simulations. It is possible to note the peak at the 11th order (18 hours) is caused by the lead time for the replenishment of inventory; the mean time is 5.32 hours. Figure 10 - Retailer Delivery Time The following figure represents how the simulator has calculated the delivery time for the Firm. There are few orders delivered because at the end of the simulation some orders were in Finished Goods Inventory (FGI), "Open" or "In Transit" status. The mean time is rather low, due to the FGI initial stocks. Figure 11 - Firm Delivery Time (mean time: 2.97 hours) We have compared the data obtained with some historical data of the firm, obtaining a satisfactory congruence with the results of the simulation, which difference with real data has never been less than 12% in the worst case. As a further experimental validation of the model, we have noticed that during the entire simulation, the supplier number 2 has never sent an order to its supplier because of the Inventory Level calculated by the simulator which was enough to satisfy the Firm's demand. In fact, a survey conducted on the warehouse holding costs of the supplier number 2 has shown that they are up to the 87% of the total supply chain costs. 5.2. VCOR specifications and Customer Satisfaction Scenario: Simulation Results The VCOR specification has been added through the Market and the Research processes. In this case we have made a previous survey related to the customer satisfaction level with the previous model for the product 1 (the one with the highest numbers of sells). The data have been acquired through the questionnaires of customer satisfaction distributed to the customers during the period in which the orders have been set. Figure 12 shows the results obtained, in which we can see a progressive decreasing of the vote related to the product with the increasing of the number of orders. Figure 12 - Customer Satisfaction trend with SCOR The extension to the VCOR has been made through a particularization of the Customer behavior model presented in section 4.2; in our model the function (2) related to the vote at instant k has a particular expression for the f(k) relation, in order to accomplish the specification of the votes found on the Customer Satisfaction modules used by the firm in its Quality Management System. In this application case the relation ui(k) has the following expression: ( ) ku i (3) ×+×+×+D×+×+ hfd bxa ] ( ) kx i [ ) a1569 ( -× - ( )k in which the notations are the same of equation (2) with the following = × I n x j ¹ i s × p d q additional considerations: • The function f(k) has its values included into the range [0-10], according to the customer satisfaction system of the firm • The pD variation has to be expressed as a percentage of the previous price and the weight b has to be positive • The delay d has to be expressed in percentage with respect to the mean lead time of the product. • The Quality level q has to be expressed in percentage with respect to the mean quality level. The simulation model has been run for the subsequent of the 19 orders. The figure 13 shows the customer satisfaction level in the simulation with VCOR model from the order number 20 to the order number 39, including the Support process in the Retailer and the whole PLM part in the Firm. The simulation has generated a high elevation of the trend after the 20th order due to the introduction in the market of the renewed product (a new taste and nutritional features have been added due to surveys conducted to the customers with and additional survey data base introduced by the CRM module). 10 9 8 7 6 5 4 3 2 1 0 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 Customer Satisf action Tr end Figure 13 -- Customer Satisfaction trends after VCOR simulation As a further validation of the proposed simulation model, a post-sales survey has been conducted on the product for which the simulation has been run. In Figure 14 we can see the real data, coming from the above mentioned Customer Satisfaction questionnaires. We can see that the values and the related trend reflect the simulation analysis with a maximum error of not more than 20% in the worst case. 10 9 8 7 6 5 4 3 2 1 0 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 Cust. Satisf . Tr end Figure 14 -- The results of the survey on customer satisfaction 5.3. Delivery Times Scenario: Simulation Results Another interesting result from the Value Chain simulation is related to the delivery times of the Retailer. In the histogram of figure 15 we report the simulation result of the entire period (from the first to the 37th order). With respect to the SCOR case we have noticed an increasing of the mean delivery time (7.84 hours); this increasing has been investigated inside the firm, finding the main two factors: • With the introduction of the Sell process, the Firm decreases its finished goods inventory level to satisfy the new Customers demand, slowing down the Retailer's orders. • The Support process of the Retailer increases the mean to 2.3 hours; it is due to the support time. Figure 15 - Retailer Delivery Time The Firm's behavior is more complex and for this reason two simulations have been necessary. The first one analyzes the Firm without the value chain part, in which the mean time is 5.36 hours (figure 16), and the second one investigates the whole performance. Figure 16 - Firm Delivery time With the introduction of the Sell process in the Firm, we can notice that in the previous case the firm has satisfied about fifteen orders and now, with the application of CRM part of VCOR, more than thirty. The high increment is mainly addressed to the acquisition of new customers that guarantees fixed financial entrances, to which Retailer's demand must be added. Moreover, the production increases from 296 to almost 500 finished products, stressing the productive cycle to higher level. To sum up, there are some benefits deriving from the adoption of the VCOR model. To better understand the improvements obtained by this choice, some indicators have been calculated to test the performance of the chain: - Stock Rotation Indicator: it shows the sale rapidity of the commodities. It is calculated as: SRI = Sales Profit Meanvalueof the stock in which a big value indicates the capability of the company to use our products. - Stock Mean Time: it indicates the mean time of undelivered materials = or goods. It is calculated as: SMI where Period is the simulation timeline, expressed in hours. Period SRI influence the effective profit. It is calculated as: SPI Sales Profit Costs - = Sales Profit - Sales Profitability Indicator: it gives an idea on how the costs In table 6 the enhancements can be seen; we can notice a huge improvement in all of the Firm's indicators, especially for the SMI for the finished goods and the raw materials inventory. Table 6 Retailer's Performance Indicator Indicator SRI for FGI SMI for FGI SRI for Raw Materials SMI for Raw Materials SPI VCOR 22.4 2.13 hours 1.2 39.8 hours 13% SCOR 12.34 3.88 hours 0.69 68.8 hours 12% This last indicator underlines that with the VCOR a lot of commodity is used (sold or used for the production) in the 45% of the time required with the SCOR simulation. We can also see that the SPI remains almost unchanged. However, if we consider that in the second simulation there are also the Support costs, the fact that the SPI is constant means that the Support costs do not affect much the total costs of the Retailer. 6. Conclusions This paper has discussed the study of the two most known models used to implement the concepts of Supply Chain Management and Value Chain Management, SCOR and VCOR, through a simulation approach. Starting from an analysis of their standard architectures, using a bottom-up approach, and their performance indicators and metrics, a simulation framework has been developed under ARENA. Once implemented the SCOR model the tool was extended to VCOR. The simulation models have been applied into an agro-food firm in which the main experimental results given by the simulation runs have been compared with the real data of the firm. This has been done to assess the model proposed and to verify the difference between the simulated data and the real ones. As confirmed by the results of these simulations, the adoption of VCOR model needs of bigger financial and organizational efforts, but these efforts can be fully repaid by the benefits in terms of quality of service, market extension, competitiveness, flexibility, "quick response", innovation and other features essential for the firms to survive in the global market. Finally, the results of some simulations have been presented, including a comparison of the two models in terms of the main Key Performance Indicators that are crucial for the supply chain analysed. A future work could complete the implementation of the templates to all the processes of VCOR model, extending the flexibility and the reusability of the realized tool. 7. References Abu-Suleiman, A. Boardman, B. and Priest, J.W. (2005) "A framework for an integrated Supply Chain Performance Management System" International Journal of Production Research 43(15) 3287 -- 3296(10). Al-Mudimigh, A.S., Zairi, M. and Ahmed., A.M. (2004) "Extending the concept of supply chain: The effective management of value chains" International Journal of Production Economics, Vol. 87, Issue 3, pp. 309-320. Balan, S. Prem Vrat, Pradeep Kumar (2006) "Assessing the challenges and opportunities of global supply chain management" International Journal of Value Chain Management (IJVCM), Vol.1 n°2, pp.105 -- 116. Bolstorff, P. and Rosenbaum, R. (2003) "Supply Chain Excellence: A Handbook for Dramatic: Improvement Using the SCOR Model" Eds. New York: AMACOM. Boucher, X. Kuhn, H., and Janke, J. (2003) "Integrated Modeling and Simulation of Inter-Organizational Business Processes". MOSIM'03-Toulouse (France). Chen, C.K. Hughes J. (2004) "Using Ordinal Regression Model to Analyze Student Satisfaction Questionnaires" Association for Institutional Research Applications. Vol. 1, May, 26. Christopher, M. and Lee H.L. (2002) Supply Chain Confidence. The Key to Effective Supply Chains Through Improved Visibility and Reliability. White Paper des Stanford Global Supply Chain Management Forum. http://www.stanford.edu. Clemmer, J. (1990) "The Three Rings of Perceived Value." The Canadian Manager. Jun ;15(2):12-15. Feller, A. Shunk, D. and Callarman, T. (2006) "Value Chains Versus Supply Chains". BPT Trends, www.bpttrends.com Garetti, M., Terzi, S., Bertacci, N., Brianza, M. (2005) "Organisational change and knowledge management in PLM Implementation", International Journal of Product Lifecycle Managemement, , Vol. 1, No. 1, pp. 43-51. Hausman, Warren H., (2002) "Supply Chain Performance Metrics, Stanford Global", Supply Chain Management Forum, http://www.stanford.edu. Heinzel, H. (2005) "The value chain operations reference model as a framework to integrate and manage cross-enterprise business processes", PLM'05: international conference on product life cycle management, Eds.Inderscience. Herrmann, J.W., Lin, E., and Pundoor, G. (2003) "Supply Chain Simulation modeling using the Supply Chain Operations Reference Model". Proceedings of DETC'03 Design Engineering Technical Conferences and Computers and Information in Engineering Conference , Chicago, Illinois USA, September 2-6, pp.1-9. Hopkins, E. (2006) "Adaptive Learning Models of Consumer Behavior" Eds. Mimeo. Jain, S. and Leong, S. (2005) "Stress testing a Supply Chain using simulation". Proceedings of the Winter Simulation Conf. ISBN: 07803951902, pp. 8-005. Kelton, W.D., Sadowski, R.P., and Sturrock, D.T.(2004) "Simulation with ARENA" Eds. McGraw-Hill Higher education, ISBN 0072919817. Kleijnen, J.P.C. (2005) "Supply chain simulation tools and techniques: a survey" International Journal of Simulation & Process Modelling, Volume 1, Nos. 1/2. Munari, L. (2004) "Il Customer Relationship Management: un'innovazione di processo." in L'innovazione Finanziaria-Osservatorio Newfin Neubert, G., Ouzrout, Y. & Bouras, A. (2004) Integration and Information Systems within the Supply Chain: A Review. International Journal of Technology Management (IJTM), ISSN0267-5730, Indersciences Publishers, pp. 259-273. Olhager, J. Selldin, E. and Wikner, J. (2006) "Decoupling the value chain" International Journal of Value Chain Management (IJVCM), Vol.1 n°1, pp.19 -- 32. Porter, M., (1985) Competitive Advantage, Creating and Sustaining Superior Performance, The Free Press, New York. SCOR Supply Chain Council "Supply Chain Operations Reference Model SCOR" www.supply-chain.org Simatupang, T.M., and Sridhran R. (2002) "The Collaborative Supply Chain", International Journal of Logistics Management. Simchi-Levi, D., Kaminsky, P., and Simchi-Levi, E. (2000). Designing and managing the supply chain. McGraw-Hill Higher Education. Value Chain Group (2005) "VCOR Quick Reference V1R4" www.value-chain.org, Terzi, S. and Cavalieri, S. (2004). "Simulation in the supply chain context: a survey". Computers in Industry n°53, pp. 3 -- 16. Wen-Bao, L. (2007) "The exploration of customer satisfaction model from a comprehensive perspective" Int. Journal in Expert Systems with Applications Vol. 33, Issue 1, pp. 110-121
1711.09564	2	1711	2017-11-28T04:36:52	AUPCR Maximizing Matchings : Towards a Pragmatic Notion of Optimality for One-Sided Preference Matchings	[ "cs.MA", "cs.DS" ]	We consider the problem of computing a matching in a bipartite graph in the presence of one-sided preferences. There are several well studied notions of optimality which include pareto optimality, rank maximality, fairness and popularity. In this paper, we conduct an in-depth experimental study comparing different notions of optimality based on a variety of metrics like cardinality, number of rank-1 edges, popularity, to name a few. Observing certain shortcomings in the standard notions of optimality, we propose an algorithm which maximizes an alternative metric called the Area under Profile Curve ratio (AUPCR). To the best of our knowledge, the AUPCR metric was used earlier but there is no known algorithm to compute an AUPCR maximizing matching. Finally, we illustrate the superiority of the AUPCR-maximizing matching by comparing its performance against other optimal matchings on synthetic instances modeling real-world data.	cs.MA	cs	AUPCR Maximizing Matchings : Towards a Pragmatic Notion of Optimality for One-Sided Preference Matchings Girish Raguvir J∗, Rahul Ramesh∗, Sachin Sridhar∗, Vignesh Manoharan∗ Department of Computer Science and Engineering Indian Institute of Technology Madras, India 7 1 0 2 v o N 8 2 ] A M . s c [ 2 v 4 6 5 9 0 . 1 1 7 1 : v i X r a Abstract We consider the problem of computing a matching in a bi- partite graph in the presence of one-sided preferences. There are several well studied notions of optimality which include pareto optimality, rank maximality, fairness and popularity. In this paper, we conduct an in-depth experimental study com- paring different notions of optimality based on a variety of metrics like cardinality, number of rank-1 edges, popularity, to name a few. Observing certain shortcomings in the stan- dard notions of optimality, we propose an algorithm which maximizes an alternative metric called the Area under Pro- file Curve ratio (AUPCR). To the best of our knowledge, the AUPCR metric was used earlier but there is no known al- gorithm to compute an AUPCR maximizing matching. Fi- nally, we illustrate the superiority of the AUPCR-maximizing matching by comparing its performance against other optimal matchings on synthetic instances modeling real-world data. Introduction The problem of assigning elements of one set to elements of another set is motivated by important real-world scenarios like assigning students to universities, applicants to jobs and so on. In many of these applications, members of one or both the sets rank each other in an order of preference. The goal is to compute an assignment that is "optimal" with respect to the preferences. In this paper we focus on the one-sided preference list model where members of one set rank a subset of elements in the other set in a linear order (that is, preferences are assumed to be strict). Several notions of optimality like pareto-optimality, rank-maximality, fairness and popularity have been considered in literature (We give formal defini- tions of each of these notions later). For each of the above mentioned notions of optimality, there are efficient algo- rithms studied in the literature to compute the specified op- timal matching. (Abraham et al. 2004) describe an algo- rithm that computes a maximum cardinality pareto-optimal matching. (Abraham et al. 2007) present an algorithm to compute a popular matching while (Irving et al. 2004; Huang et al. 2016) propose algorithms that optimize the head/tail of the matching profile (rank-maximal and fair re- spectively). Maximizing one metric could however result in poor performance on other yardsticks of measure. When ∗All authors contributed equally comparing two matchings, it is difficult to measure the qual- ity of the two matchings using a single scalar value. They can be compared using a variety of metrics like cardinal- ity, number of matched Rank-1 edges or cardinality, none of which can serve as a sole indicator of optimality. Profile based matchings, like Rank-maximal or Fair matchings which optimize for the head or the tail of the pro- file can turn out to be biased under certain circumstances. An alternative is to consider the Area under Profile Curve Ratio metric introduced in (Diebold and Bichler 2017). This met- ric aims to maximize a measure, that is a weighted sum of matched edges, with the weight proportional to its position in the preference list In this work, we first present a comprehensive experimen- tal study of the well-studied notions of optimality and com- pare them using different measures of matching quality. We then describe the AUPCR metric, and propose algorithms to compute an AUPCR maximizing matching, and a maximum cardinality AUPCR maximizing matching. Finally, we empirically evaluate different matching algo- rithms on synthetic graphs generated from generator models specified by (Michail 2011) using various metrics. The gen- erated graphs fall into two categories, one having uniformly random preference lists and the other having highly corre- lated preference lists. Our analysis is inspired by the analysis of (Michail 2011), and we additionally consider a ranking system in which the matchings are ranked based on multi- ple metrics. These rankings are consequently aggregated to obtain a single rank, which we use as a coarse indicator of matching quality. The AUPCR maximizing matching is experimentally shown to have good performance across evaluated metrics on the considered data-sets, and we believe this matching is well suited for practical applications. Preliminaries Consider a set A of applicants and a set P of posts. Every applicant a has preference list over a subset of the posts in P. This list is a linear order (strict list) and is called the pref- erence list of a over P. The problem is readily represented as a bipartite graph with vertices V = A ∪ P and an edge (a, p) is present if p is acceptable to a. Preferences of ap- plicants are encoded by assigning ranks to edges. Each edge (a, p) has a rank i if a considers p as its i-th most preferred post. A matching M ⊆ E is a collection edges such that no two edges share an endpoint. Let A∪P = n and E = m. We now define formally the different notions of optimality. Maximum Cardinality Pareto Optimal Matching A matching M is said to be Pareto-optimal if there is no other matching M(cid:48) such that some applicant is better off in M(cid:48) while no applicant is worse off in M(cid:48) than in M(an ap- plicant is worse of in M if it is matched to an less preferred vertex compared to M(cid:48)) . Maximum cardinality Pareto op- √ timal matchings(POM) can be computed in O(m n) time using the algorithm given by (Abraham et al. 2004). Rank Maximal Matching The notion of rank-maximal matchings was first introduced by Irving under the name of greedy matchings (Irving 2003). A rank-maximal matching is a matching in which the num- ber of rank one edges is maximized, subject to which the number of rank two edges is maximized and so on. An- other way of defining rank-maximal matchings is through their signatures. Given that r is the largest rank given to a choice across all preference lists, we define the signature of a matching M as an r-tuple (x1, x2,··· , xr, xr+1) where, for 1 ≤ i ≤ r, xi represents the number of applicants matched to one of their i-th preferences (xr+1 denotes the number of unmatched applicants). Let (x1, x2,··· , xr, xr+1) and (y1, y2,··· , yr, yr+1) denote the signatures of M and M(cid:48) respectively. We say that M (cid:31) M(cid:48) w.r.t. rank-maximality if there exists an index k such that 1 ≤ k ≤ r and for 1 ≤ i ≤ k, xi = yi and xk > yk. A matching M is rank- maximal if M has the best signature w.r.t. rank-maximality. Through the rest of the paper, we denote this matching as RMM. For the purposes of our experimental evaluation, we implement a simple combinatorial algorithm (Irving et al. 2004) to compute a rank maximal matching. The running time of the algorithm is O(min(C n, n + C) · m). √ Popular Matching To define popularity, we translate preferences of applicants over posts to preferences of applicants over matchings. An applicant a prefers matching M to M(cid:48) if either a is matched in M and unmatched in M(cid:48), or a is matched in both M and M(cid:48) but has better rank in M than in M(cid:48). A matching M is more popular than M(cid:48) if the number of applicants who pre- fer M to M(cid:48) is more than the number of those who prefer M(cid:48) to M. A matching M is popular if there is no matching that is more popular than M. A linear time algorithm to compute a maximum cardinality popular matching for strict prefer- ences is given in (Abraham et al. 2007). The more popular than relation is not transitive, and hence it is possible that a popular matching does not exist. When a popular matching does not exist, one can attempt to obtain the least unpopular matching. We consider the unpopularity factor given in (Mc- Cutchen 2008). An algorithm given in (Huang et al. 2011) finds a popular matching if it exists. Through the rest of the paper, we denote this matching as POPM. Fair Matching Fair matchings can be considered as complementary to rank- maximal matchings. A fair matching is always a maximum cardinality matching, subject to this, it matches the least number of applicants their last preferred post, subject to this, least number of applicants to their second last pre- ferred post and so on. Fair matchings can be conveniently defined using signatures. Let (x1, x2,··· , xr, xr+1) and (y1, y2,··· , yr, yr+1) denote the signatures of two match- ings M and M(cid:48) respectively. We say that M (cid:31) M(cid:48) w.r.t. fairness if there exists a index k, such that 1 ≤ k ≤ r + 1 and for k < i ≤ r + 1, xi = yi, and xk < yk. A matching is fair if it is of maximum cardinality, and subject to that it has the best signature according to the above defined criteria. Recently (Huang et al. 2016) gave a combinatorial algorithm to compute fair matchings. Through the rest of the paper, we denote this matching as FM. AUPCR Maximizing Matching Fair and Rank-maximal matchings are profile based match- ings that are geared towards minimizing the tail or maxi- mizing the head of the profile. However, optimizing for the peripheral portions of a profile may not be necessarily rep- resentative of a good matching in many practical settings. This encouraged us to look into a metric called Area Under Profile Curve Ratio (AUPCR) which, in a sense, seemed to capture the entire signature of a matching. Formulation of AUPCR The Area Under Profile Curve Ratio (AUPCR), introduced under the context of matchings by (Diebold and Bichler 2017) is a measure of second order stochastic dominance of the profile. It is a useful metric that can be used to compare multiple signatures and is very similar in nature to the highly popular Area Under Curve of Receiver Operating Character- istic (Hanley and McNeil 1982). For a matching M of a bipartite graph G(A ∪ P, E) with ni(M ) representing the number of applicants matched to their i'th preference, AUPCR(M) is defined as the ratio of Area Under Profile Curve (AUPC) and Total Area (TA) where TA(M ) = AP (ai, pj) ∈ M : rank(ai, pj) ≤ r (1) AUPC(M ) = P(cid:88) P(cid:88) r=1 = (P − i + 1)ni(M ) Giving us, i=1 AUPCR(M ) = (cid:80)P i=1(P − i + 1)ni(M ) AP (2) (3) One can visualize this quantity by considering Figure 1. For an instance with A = 8, P = 6 and signature of matching M given by (4, 0, 2, 1, 1, 0), the area under the shaded region corresponds to AUPC(M) (= 4+4+6+7+8+8 = 37) while the area of bounding rectangle corresponds to TA(M)(= 8× 6 = 48). With these computed, AUPCR(M) is essentially the ratio of the two and is given by 37 48 ≈ 0.771. Figure 1: AUPCR - Visualization A matching that maximizes this measure can be vaguely seen as a "softer" version of the rank maximal matching: it does not give up matching low ranked edges entirely in order to match a large number of high ranked edges. Based on this we consider two problems: • AUPCR Maximizing Matching - the problem of finding a matching which maximizes the AUPCR metric. We de- note such a matching as AMM. • Max Cardinality AMM - the problem of finding a match- ing with the maximum cardinality among all matchings with maximum AUPCR. We denote such a matching as MC-AMM. In this paper, we formulate algorithms to address the above defined problems and show that the Max Cardinal- ity AUPCR maximizing matching performs favorably on a variety of other standard metrics typically used to compare matchings in practical settings. weight of a matching as w(M ) =(cid:80) Algorithm - AUPCR Maximizing Matching The problem of finding an AUPCR maximizing match- ing can be reduced to the problem of finding a maximum weighted perfect matching. Given a bipartite graph G(A ∪ P, E) and a weight we for each edge e ∈ E, we define the e∈E we. Then, the max- imum weighted perfect matching problem is find a matching M which matches all vertices in A (M is a perfect matching) and maximizes w(M ). Given an bipartite graph G(A ∪ P, E) with edges repre- senting preferences of A, we construct G(cid:48)(A(cid:48) ∪ P(cid:48), E(cid:48)) as follows: 1. A(cid:48) = A1 ∪P and P(cid:48) = P1 ∪ A2 where A1, A2 are copies of A and P1,P2 are copies of P. 2. For each edge e ∈ E of rank i, add edge between corre- sponding vertices of A1 and P1 with weight P − i + 1. Similarly, add an edge between A2 and P2 with the same weight. 3. Add edges with weight 0 from vertices in A1 to their copies in A2. Add similar edges between P1 and P2. We refer to these edges as identity edges. Proof of Correctness Claim. If M is a max weighted perfect matching in G(cid:48), then M restricted to A1 ∪ P1 is a AMM in G. Proof. Let M1 be the matching obtained by restricting M to A1∪P1 and M2 obtained by restricting M to A2∪P2. Since M is a perfect matching, all vertices of G(cid:48) must be matched. So, if a vertex in A1 ∪ P1 is not matched in M1, it must be matched to its copy in A2 ∪ P2 via the identity edge. This means that its copy is also unmatched in M2. So, M1 and M2 match the same set of vertices. Since the identity edges have 0 weight, w(M ) = w(M1) + w(M2) Since M1 and M2 match the same set of vertices, one can copy the edges matched in M1 to A2 ∪ P2. This means that w(M1) = w(M2) and w(M ) = 2w(M1). Maximizing w(M ) is equivalent to maximizing w(M1). (cid:88) e∈M1 we = P − re + 1 (cid:88) P(cid:88) e∈M1 We also have w(M1) = = ni(P − i + 1) = APAUPCR(M1) i=1 Hence, if M is a maximum weight perfect matching in where re is the rank of edge e and ni is the number of edges of M1 with rank i.This means that maximizing w(M1) max- imizes AUPCR(M1). G(cid:48), M1 is a max AUPCR matching in G. Algorithm - Max Cardinality AMM The problem of finding a Max Cardinality AMM can also be reduced to an instance of max weighted perfect match- ing. The reduction is the same as the max AUPCR case, but we add a negative weight of − 1 A+P to the identity edges going from A1 to A2. Proof of Correctness Claim. If M is a max weighted perfect matching in G(cid:48), then M restricted to A1 ∪ P1 is a Max Cardinality AMM in G. Proof. As before, we can prove that w(M1) = w(M2). However, w(M ) = w(M1) + w(M2) + w(I) where I is the set of identity edges from A1 to A2 in M. If M1 leaves kA vertices in A unmatched and kP vertices in P unmatched, then M2 also leaves the same vertices unmatched. So, we have 2(kA + kP ) identity edges in I and hence w(I) = −2 kA + kP A + P Since kA + kP < A + P, we have −2 < w(I) ≤ 0 and 2(w(M1) − 1) < w(M ) ≤ 2w(M1) Let M(cid:48) be a Max AUPCR matching extended to G(cid:48) and M(cid:48) be its restriction to A1 ∪ P1. Since w(M(cid:48)) ≤ w(M ) 1 2(w(M(cid:48) 1) − 1) < 2w(M1) 1) − w(M1) < 1 ⇒w(M(cid:48) ⇒APAUPCR(M(cid:48) ⇒AUPCR(M(cid:48) 1) − APAUPCR(M1) < 1 1) − AUPCR(M1) < 1 AP From the definition of AUPCR, we can see that if two match- ings have different AUPCR, then the difference is ≥ 1AP. So, M(cid:48) 1 and M1 have the same AUPCR, which means that M1 is an AUPCR maximizing matching in G. The cardinality of M1 is M1 = A − kA = P − kP. Writing w(M ) in terms of M1, w(M ) = 2w(M1) − A + P − 2M1 A + P All AUPCR maximizing matchings will have the same w(M1), which means that maximizing w(M ) maximizes M1. So, M1 is a maximum cardinality AUPCR maximiz- ing matching in G. √ The time complexity of the algorithm to find maximum weighted matching presented is O(m n log n) (Duan and Su 2012). Since both our algorithms construct a graph with 2n vertices and m + n2 edges and find a max weighted matching, the time complexity would be O(n2√ n log n). Algorithm Running Time √ O(m √ POM (Abraham et al. 2004) n) O(min(C √ RMM (Irving et al. 2004) n, n + C)m) O(Cm √ FM (Huang et al. 2016) n log n) O(m POPM (Huang et al. 2011) AMM, MC-AMM (Our work) O(n2√ n) n log n) Table 1: C is the maximum rank of any edge in the matching, n = A + P and m = E Experiments ing. Evaluation Metrics The matchings obtained from each algorithm are evaluated with respect to the following metrics. • Cardinality: The number of edges present in the match- • Unpopularity measure: The unpopularity measure u(M, M(cid:48)) measures how far away a matching M is from a popular(least unpopular) matching M(cid:48). Let p(M1, M2) be the number of applicants that prefer M1 over M2. Then u(M, M(cid:48)) for matching M is defined as the ratio of p(M(cid:48), M ) − p(M, M(cid:48)) to the total number of applicants. • Rank 1: The number of matched rank 1 edges • AUPCR: The AUPCR metric is second order stochastic dominance of the profile as defined in Equation 3. • Ranks less than half the preference list size (RHPL): This counts the number of applicants who have been matched to a post with a rank better than or equal to half the length of their preference list. • Average rank: For a matching M, this is the average rank of all matched edges. Although this is similar to the AUPCR metric, the average rank is computed only over the matched edges while AUPCR accounts for unmatched edges. • Worst rank: For a matching M, this is the highest (worst) • Time: The time taken to find the matching. This is imple- mentation dependent, and the algorithms used have been mentioned earlier along with their time complexities. rank among all matched edges in M. Instances For our experiments, we consider two structured instance generators, namely: Highly Correlated and Uniform Ran- dom. These generators are similar in nature to (Michail 2011), but we consider only instances with strict prefer- ence lists. Though all the algorithms described above, ex- cept Maximum Cardinality Pareto Optimal, can also handle instances with ties, we went with this choice to have a set of instances upon which all the algorithms could be compared and analyzed. If one thinks about it, this choice is not too restrictive as in practical scenarios preference lists are often strict and devoid of ties. Uniform Random (UNI) Similar to HC, UNI instances are also parameterized by a density d with 0 ≤ d ≤ 1. Every applicant has a preference list size of l = (cid:98)np · d(cid:99). These preference lists are chosen uniformly at random from the set of permutations of l posts. Let an applicant a's adjacency list be (p1, p2, ..., pl). Then p1 is ranked 1 by a, p2 is ranked 2, and so on. Unlike HC, preference list length is identical across all applicants. Highly Correlated (HC) These instances are generated based on a global preference ordering(say P ) for the set of posts; one that all the applicants agree upon. A HC instance is parameterized by a density d with 0 ≤ d ≤ 1. For every vertex pair (u, v) with u ∈ A, v ∈ P, an edge (u, v) is added with a probability d. Once the graph has been constructed, the applicants rank the posts as per the global preference list: the best post, as per P, an applicant is connected to is assigned rank 1, and so on. Experiment Setup The number of applicants are equal to the number of posts in any graph and is varied from 50 to 900 in steps of 50. Or- thogonally, the density parameter d for HC and UNI is varied from 0.02 to 0.20 in steps 0.02. The reason for this choice of range is that real world datasets are not very dense in nature . Each instance is averaged over 50 random seeds. There ex- ists one more level of averaging across different density(d) values to get one value for each metric for each problem size(number of applicants). The variant of Max-AUPCR, that does not not enforce maximum cardinality is used. Surprisingly, this still yields (a) AUPCR (b) Avg Rank (c) Popularity (d) Rank 1 Figure 2: Uniform Random (a) AUPCR (b) Avg Rank (c) Popularity (d) Rank 1 Figure 3: Highly Correlated a max-cardinality matching without exception. For POPM, in cases where popular matchings don't exist, the least un- popular matching is utilized. The code was executed using the Amazon web services(AWS) based EC2 service on a t2.micro instance(1 GB Ram, 1 CPU, Intel Xeon processor). Experimental Results Comparing matchings based on rank means POM RMM POPM FM AMM 1.00 Card. Unpop. 5.00 Rank 1 3.80 AUPCR 3.81 RHPL 4.28 Avg Rank 4.98 Worst Rank 4.66 2.94 2.00 1.00 4.85 2.89 2.58 3.27 2.00 1.00 1.00 3.28 2.58 3.81 3.39 1.00 1.00 3.99 3.01 3.18 2.00 1.99 1.00 1.26 1.14 2.34 1.22 1.00 1.94 Rank Mean 3.93 2.79 2.44 2.11 1.62 Table 2: Rank means of algorithms on metrics for UNI For this analysis, we consider a set of the evaluation met- rics which we believe characterizes preference matchings in general. For a given metric and graph instance, we rank the algorithms in terms of performance with the best one getting a rank of 1 and worst one getting a rank of 5. We then aver- age this rank across all instances and this value corresponds POM RMM POPM FM AMM Card. 1.00 Unpop. 5.00 Rank 1 4.93 AUPCR 3.25 RHPL 4.09 Avg Rank 5.00 Worst Rank 3.66 2.89 3.07 1.00 4.84 1.96 2.18 4.73 2.11 1.00 1.59 3.92 3.02 3.53 3.42 1.00 1.00 3.94 2.00 2.99 3.96 2.00 1.00 4.75 1.11 3.20 1.09 1.01 1.99 Rank Mean 3.84 2.95 2.66 2.69 1.73 Table 3: Rank means of algorithms on metrics for HC to an entry in Table 2. The rank mean is computed by tak- ing the average of the entries along the column. This value is intended to serve as a measure of overall performance. As seen from the table, each chosen metric has a sub- set of the algorithms performing best. It is however impor- tant to note that AMM performs competitively in almost all metrics. This observation is also qualitatively supported from the fact that the rank mean attained by AMM is lowest among all algorithms for both UNI and HC instances. This empirically shows that AMM is able to achieve a much de- sired balance, making it a very compelling choice for many practical preference matching problems. Comparing the Matchings on different metrics Some interesting observations for some metrics are as fol- lows : • Cardinality : As expected, POM and FM have the largest cardinality since they compute maximum cardinal- ity matchings. However, it was observed that AMM with- out exception returned a maximum cardinality matching. While this may not universally true(as proved in conse- quent section) this is a useful property in practice • RHPL : The RHPL is one metric that no matching in par- ticular optimizes for. It is peculiar to note that AUPCR maximizes this metric indicating that it is indeed a more general notion of optimality. • Rank 1 : It was observed that both popular and rank max- imal matchings have similar if not same number of rank 1 edges. While the head of the signature is maximized, it is observed that both these matchings display poor perfor- mances on metrics that account for the entirety or the tail of the signature. • Time : Dictated by the computational time complexities of the respective algorithms, the times were vastly differ- ent for FM and AMM compared to the other three match- ings. In graphs with 900 vertices(in each partition), the FM took 512.45 seconds,AMM executed in 204.78 sec- onds while POP and PM were executed in less than 5 sec- onds. Is an AMM always a maximum cardinality matching? An AMM need not always be a maximum cardinality match- ing. Consider the instance with A = {a1, a2, a3, a4}, P = {b1, b2, b3, b4} and the preferences given by a1 : (b1, 1) a2 : (b1, 1), (b2, 2) a3 : (b2, 1), (b1, 2), (b3, 3) a4 : (b3, 1), (b1, 2), (b4, 3) As shown in Figure 4 and Figure 5, for this instance, AMM has a cardinality of 3 while a maximum cardinality matching has a cardinality of 4. Figure 4: AUPCR Maximizing Matching Understanding AMM The strongly positive empirical performance of AMM, in various metrics of importance as shown above, leads us to ask some interesting questions. Is an AMM Pareto optimal? Yes, AMM is a Pareto optimal matching. Theorem. AUPCR maximizing matching is Pareto optimal. Proof. Assume to the contrary that an AUPCR maximizing matching M is not Pareto optimal. This means there exists a matching M(cid:48) where every applicant in M(cid:48) is at least as well off as in M and at least one applicant in M(cid:48) is better off than M. Consider a vertex v ∈ A. Let rM (v) be the rank of the post that v is matched to (rM (v) = P + 1 if v is unmatched), and rM(cid:48)(v) be defined analogously. AU P CR(M(cid:48)) − AU P CR(M ) = ((P − rM(cid:48)(v) + 1) − (P − rM (v) + 1)) (rM(cid:48)(v) − rM (v)) (cid:88) (cid:88) v∈A = v∈A > 0 The last inequality follows from the fact that every term of the summation is non negative and at least one term is posi- tive by our assumption that M is not Pareto optimal. Since AUPCR(M(cid:48)) - AUPCR(M) > 0, M is not an AUCPR maximizing matching, a contradiction, and so M must be Pareto optimal. Figure 5: Maximum Cardinality Matching Do all AMMs have the same cardinality? All AMMs need not have the same cardinality. Consider the instance with A = {a1, ...a6} and P = {b1, ...b6} and the preferences given by a1 : (b6, 1), (b3, 2), (b1, 3) a2 : (b2, 1), (b3, 2), (b1, 3) a3 : (b4, 1), (b5, 2), (b2, 3) a4 : (b1, 1), (b4, 2), (b6, 3) a5 : (b5, 1), (b2, 2), (b1, 3) a6 : (b4, 1), (b2, 2), (b5, 3) As shown in Figure 6 and Figure 7, both are AUPCR max- imizing matchings, with an AUPCR of 0.833, but they have different cardinalities. This example also shows that multi- ple AMMs can exist for a given instance. Figure 6: An AMM with M = 5 Figure 8: An AMM with matching with signature (3, 3, 0, 0, 1) Figure 7: An AMM with M = 6 Is an AMM always more "rank maximal" than a FM? An AMM matching need not be more rank-maximal than the fair matching. Consider the instance with A = {a1, a2, a3, a4, a5, a6, a7}, P = {b1, b2, b3, b4, b5, b6, b7} and the preferences given by a1 : (b1, 1) a2 : (b2, 1) a3 : (b3, 1), (b4, 2), a4 : (b1, 1), (b5, 2), (b4, 3) a5 : (b1, 1), (b6, 2), (b2, 3), (b5, 4) a6 : (b1, 1), (b2, 2), (b7, 3), (b6, 4), (b3, 5) a7 : (b7, 1) An AMM matching for the above graph is as show in Figure 8 and it's signature is given by (3, 3, 0, 0, 1). One can easily see that the FM shown in Figure 9 is more rank-maximal with a signature (4, 0, 1, 2, 0). Figure 9: A Fair matching with signature (4, 0, 1, 2, 0) Conclusion In this work, we introduce the notion of an AUPCR maxi- mizing matching. We describe two variants with one max- imizing the AUPCR, and the other maximizing the cardi- nality subject to maximizing the AUPCR. We empirically evaluate our algorithm on standard synthetically generated datasets and highlight that AUPCR maximizing matching achieves this much needed middle-ground with respect to the different notions of optimality. The overall performance of the AUPCR matching is superior in comparison to other matchings when all metrics are cumulatively used for com- parison. Extending the AUPCR matching and finding algo- rithms with reduced time complexity is left as future work. J.; References J.; Cechl´arov´a, K.; [Abraham et al. 2004] Abraham, D. Manlove, D. F.; and Mehlhorn, K. 2004. Pareto optimality in house allocation problems. In International Symposium on Algorithms and Computation, 3–15. Springer. [Abraham et al. 2007] Abraham, D. Irving, R. W.; Kavitha, T.; and Mehlhorn, K. 2007. Popular matchings. SIAM Journal on Computing 37(4):1030–1045. [Diebold and Bichler 2017] Diebold, F., and Bichler, M. 2017. Matching with indifferences: A comparison of algo- rithms in the context of course allocation. European Journal of Operational Research 260(1):268–282. [Duan and Su 2012] Duan, R., and Su, H.-H. 2012. A scaling algorithm for maximum weight matching in bipartite graphs. In Proceedings of the twenty-third annual ACM-SIAM sym- posium on Discrete Algorithms, 1413–1424. SIAM. [Hanley and McNeil 1982] Hanley, J. A., and McNeil, B. J. 1982. The meaning and use of the area under a receiver operating characteristic (roc) curve. Radiology 143(1):29– 36. [Huang et al. 2011] Huang, C.-C.; Kavitha, T.; Michail, D.; and Nasre, M. 2011. Bounded unpopularity matchings. Al- gorithmica 61(3):738–757. [Huang et al. 2016] Huang, C.-C.; Kavitha, T.; Mehlhorn, K.; and Michail, D. 2016. Fair matchings and related prob- lems. Algorithmica 74(3):1184–1203. [Irving et al. 2004] Irving, R. W.; Kavitha, T.; Mehlhorn, K.; Michail, D.; and Paluch, K. 2004. Rank-maximal match- In Proceedings of the fifteenth annual ACM-SIAM ings. symposium on Discrete algorithms, 68–75. Society for In- dustrial and Applied Mathematics. [Irving 2003] Irving, R. W. 2003. Greedy matchings. Tech- nical R eport Tr-2003-136, University of G lasgow. [McCutchen 2008] McCutchen, R. The least- unpopularity-factor and least-unpopularity-margin criteria for matching problems with one-sided preferences. LATIN 2008: Theoretical Informatics 593–604. [Michail 2011] Michail, D. 2011. An experimental compari- son of single-sided preference matching algorithms. Journal of Experimental Algorithmics (JEA) 16:1–4. 2008.
1912.05362	1	1912	2019-12-11T14:43:22	Jason-RS, a Collaboration between Agents and an IoT Platform	[ "cs.MA", "cs.CY" ]	In this article we start from the observation that REST services are the most used as tools of interoperability and orchestration in the Internet of Things (IoT). But REST does not make it possible to inject artificial intelligence into connected objects, ie it cannot allow autonomy and decision-making by the objects themselves. To define an intelligence to a connected object, one can use a Beleive Desire Intention agent (BDI an intelligent agent that adopts human behavior) such as Jason Agentspeak. But Jason AgentSpeak does not guarantee orchestration or choreography between connected objects. There are platforms for service orchestration and choreography in IoT, still the interconnection with artificial intelligence needs to be built. In this article, we propose a new approach called Jason-RS. It is a result of pairing Jason BDI agent with the web service technologies to exploit the agent capacity as a service, Jason-RS turn in Java SE and it does not need any middleware. The architecture that we propose allows to create the link between Artificial Intelligence and Services choreography to reduce human intervention in the service choreography. In order to validate the proposed approach, we have interconnected the Iot BeC 3 platform and the REST agent (Jason-RS). The decision-making faculty offered by Jason-RS is derived from the information sent by the objects according to the different methods of REST (GET, POST, PUT, and DELETE) that Jason-RS offers. As a result, the objects feed the inter-agent collaborations and decision-making inside the agent. Finally, we show that Jason-RS allows the Web of Objects to power complex systems such as an artificial intelligence responsible for processing data. This performance is promising.	cs.MA	cs	Jason-RS, a Collaboration between Agents and an IoT Platform Hantanirina Felixie RAFALIMANANA1,Jean Luc RAZAFINDRAMINTSA1, Syl- vain CHERRIER², Thomas MAHATODY1, Laurent GEORGE2, Victor MANANTSOA1 1 Laboratory for Mathematical and Computer Applied to the Development Systems, University 2 Université Paris-Est, Institut Gaspard Monge (LIGM),77454 Marne-la-Vallée Cedex 2 of Fianarantsoa, Madagascar [email protected] 1 [email protected] [email protected] [email protected] [email protected] [email protected] Abstract. In this article we start from the observation that REST services are the most used as tools of interoperability and orchestration in the Internet of Things (IoT). But REST does not make it possible to inject artificial intelligence into connected objects, ie it cannot allow autonomy and decision-making by the ob- jects themselves. To define an intelligence to a connected object, one can use a Beleive Desire Intention agent (BDI an intelligent agent that adopts human be- havior) such as Jason Agentspeak. But Jason AgentSpeak does not guarantee or- chestration or choreography between connected objects. There are platforms for service orchestration and choreography in IoT, still the interconnection with ar- tificial intelligence needs to be built. In this article, we propose a new approach called Jason-RS. It is a result of pairing Jason BDI agent with the web service technologies to exploit the agent capacity as a service, Jason-RS turn in Java SE and it does not need any middleware. The architecture that we propose allows to create the link between Artificial Intelligence and Services choreography to re- duce human intervention in the service choreography. In order to validate the proposed approach, we have interconnected the Iot BeC3 platform and the REST agent (Jason-RS). The decision-making faculty offered by Jason-RS is derived from the information sent by the objects according to the different methods of REST (GET, POST, PUT, and DELETE) that Jason-RS offers. As a result, the objects feed the inter-agent collaborations and decision-making inside the agent. Finally, we show that Jason-RS allows the Web of Objects to power complex systems such as an artificial intelligence responsible for processing data. This performance is promising. Keywords: BDI Agent, BeC3, Internet of Thing, Jason. 2 1 Introduction The tendency towards the Web of Object (WoO) does not cease to increase from time to time. Given the number of objects to facilitate the daily life, the number of possible interactions and the applications that one can imagine become more powerful and rich. But what are the ways to make these application in a generic way? The object web orientation is considered as a great resource not only for managing objects [1] but also for questions of flexibility of exchanges. In the Web of Object (WoO) one is rather directed towards the use of Web Service [1]. We note that the most used Web Service type is REST1 [2] [1] [3]. REST (Repre- sentational State Transfer) or RESTful is an architecture style for building applications (Web, Intranet, Web Service). It is a set of conventions, rules and good practices to respect and not a technology in its own right. Not only REST is stateless [4] (meaning that the service consumer does not have to store information about the way he is using the service); it is design to reduce the coupling between software pieces. In the Iot, REST may allow reconfiguration of the object, it reuse in a different application for a multiplatform application, REST service structure is easy to describe. REST allows machine to machine communication and is simplier to do than SOAP. The communi- cation is done through the HTTP protocol. Nevertheless the possibility of using the resource via the web given by the REST, it still lacks intelligent data processing. Our contribution aim to add a complex data processing of the object web. We then opt for the reuse of our search fruit in [5] which we call Jason-RS or Jason-Rest for more pre- cision. Jason-RS gives the REST Web Service ability to the BDI agent Service [5]. Jason is a multi-agent platform based on the Agentspeak-L language [6] [7]. It is a platform for BDI agent, that is to say goal-oriented software agent, it is a smart agent. We have to look at the choice of the BDI agent because it is close to human behavior. [8] [9]. So the connexion of the latter with the REST Web Service allows him to publish his ca- pacity as a Web Service. The processing of a complex action can be done inside Jason (prediction, composition of services, constraint logic programming) and could be pub- lished as a Web Service. Jason-RS can serve as an intelligent treatment center. We focus in the collaboration between agent and Iot platform cause connected object had no intelligence, designed differently, so with this collaboration we can open in the intelligence of Iot with a sensor less, little human intervention. In our approach, we collect the information from the connected object and by Jason-RS we delegate the intelligent agent works. All change in the object behavior is sent to the agent BDI (as a perception) in order to provide a decision or to change the agent comportment. Through this perception that the agent can decide and delegate a task to agent or service chore- ography to object. 1 https://www.supinfo.com/articles/single/5642-qu-est-ce-qu-une-api-rest-restful 3 The rest of this article is organized as follows: Section 2 will discuss the related work of our work. Section 3 describes in general the architecture of our contribution. In sec- tion 4 we will evaluate our work through an architecture. In section 5 we discuss a discussion and open challenge. We will conclude this article with a conclusion (section 6) 2 Related work In this article we propose a new form of Web Service that the WoO can use. The use of REST in WoO is not an innovative work as researchers have already done research for the implementation of the latter to allow connected objects to trade [10] [11]. The In- ternet of Things is a system of physical objects that can be discovered, monitored, con- trolled or interact with electronic devices that communicate on various network inter- faces and can potentially be connected to the wider Internet. The webs of object is a concept say recent in computing that offers a future with which object of the daily life can integrate into the web [12]. The Web of Object is considered as a subset of the Internet of Things. WoO applies to standards and software infrastructures such as REST, HTTP and URIs to create applications and services that combine and interact with a variety of network devices. The key point of the WoO is that its implementation does not involve the reinvention of the means of communication given that they use the existing standards. It also gives an easy access to a wide popu- lation of programmer that used to handle such an architecture. It facilitates the integra- tion of the IoT in the internet. It is also a solution that can open the connection of the Industry 4.0 [13]. View the exponential increase of the connected object, to reduce the cost, the re- quired performance and to facilitate communication in the object internet, the W3C recently launched the working group on the Web of Objects. They aim to fight the fragmentation of IoT. They developed the initial standards for the Web of Objects [14]. To facilitate service choreography for Iot Cherrier describes in his thesis a new ap- proach based on a platform to standardize the interoperability of heterogeneous objects [15]. Castellani and Al focused on the implementation of both CoAP and EXI technology for communication between objects by adopting the REST approach [16]. They imple- mented the CoAP protocol to enable communication between distributed connected objects. Then they also implemented the EXI library for compressing XML data into binary. In the paper [2], they discuss how to create an event SOA. They found a new ap- proach for coordinating IoT's services with the resource service. They named EDSOA the result of their research. It is an asynchronous SOA and responsive it can be consid- ered as an event based service. We have inspired by this approach for the technical aspect of Jason-RS. The Web of Object is used in almost any IoT field even in the field of health. On what [17] proposed a new instance of cloud server and set up data collection services 4 on the heterogeneous devices they use. This aims to improve online health service by properly implementing the architectural web object principle for remote control. Based on WoO, [4] describes the object-based web architecture that adopts the prin- ciple of Resource Oriented Architecture (ROA), hence the basis of RESTful. They im- plemented a smart gateway for smart meters. This allows the connected object manage- ment servers to keep track of the status and movements of the objects.[2] the HTTP protocol is used to facilitate access to the service. The answer emitted is in JSON for- mat. The architecture proposed by [4] allows developers to use web programming lan- guages such as (HTML, Python,...). All of the research cited above are interesting from the point of view of exposure and communication capacity in the Internet of Things. Each with its strong point in communication and interconnection in the world of IoT. They all almost used the REST Web Service. But for us, a simple service is not enough not only to bathe the connected objects in the artificial intelligence but also to treat in intelligent way the data captured at the level of the objects. This treatment helps automatic decision-making thanks to the intelligent BDI agent that has the same reasoning mode as humans. We focus our research on the intelligent Web Service, that is, a Web Service associated with artificial intelligence in order to reduce human intervention to the orchestration of services in connected objects and decision-making. We have exploited here our previous research work, Jason-RS. An intelligent Web Service with BDI agents that perform the resolu- tion and automation of certain tasks thanks to their ability to cooperate and exchange using the AgenSpeak (L) agent language. These agents may then exhibit functionality as a Web Rest Service. Jason-RS has the methods used in REST (GET, PUT, POST, and DELETE) to ensure the dialog with heterogeneous clients. 3 Proposed Architecture 3.1 General architecture In this section we discuss the general architecture of our approach. As we are in the world of the IoT there is a mean of communication: the network. There are many types of networks today for the connectivity of objects such as wifi, Bluetooth for common devices, LoraWAN for IoT devices using long distance communication, and ZigBee, 6LowPan, EnOcean, Zwave for short range communication devices, etc. As connected objects do not have enough memory space and very limited energy to feed themselves permanently so it is important to outsource most of its function. Although the design of its objects is different (ie the objects are heterogeneous), a standardization platform is important to allow inter-object choreography. Taking into account all these parameters we have our control center server constituted by the Jason-RS server and the IoT server illustrated by Fig. 1. Generally the IoT server is a centralized platform of (Cloud, big data) whether it is an IoT server that administers the objects. Jason plays a complex problem solver role as well as the complex system such as optimization ... The autonomous and goal-oriented capability through the BDI agent that it implements allows it to perform a task reasonably. This capability is now ex- ploitable through a distributed component as a service [5] through Jason-RS. Then the 5 objects can use it not only to send the sensor data but also to initiate a decision-making process at the agent level. Communication between the IoT and the web is done through the HTTP protocol. With Jason-RS we can use all the existing methods of the REST and most importantly there is an interaction with the agent. As a result we can send data to the central control servers via the POST method. We can also retrieve the data or decision made by the GET method. JASON-RS recognizes directly the task that will send back to the agent. At the agent level as soon as the service is called by the objects or other platform, the agent reacts and changes state in relation to the perception it receives from it envi- ronment. Fig. 1. General architecture of the JASON-RS implementation in the Web of Object We have already studied in our article [5] the publication of Jason's BDI agent ca- pability as a Web Service. We have proposed an approach allows deploying and run- ning Web Services (REST and SOAP) pairs with a Jason BDI agent in java SE envi- ronment. We didn't use a modern Web-App server or application server, and we didn't developed a different middleware cause it requires a lot of time. Deploying an agent inside a server is a tedious task. Then in our strategy, we reused the existing Java fram- works called Non-Blocking Input Output. A small illustration for brightening the archi- tecture is shown in Fig.2. 6 Fig. 2. General architecture of the JASON-RS micro-service [5] Jason-RS offers an opportunity to exploit the world-class logic, which is embedded in Jason. It can associate them with other elements or heterogeneous application. When we can manipulate first-order logic, we can imagine a kind of possibility of complex problem solving as well as the common problem in artificial intelligence. What is in- teresting here is that the reasoning mode of Jason agent is near the human reasoning mode to achieve a goal. Jason-RS aims to outsource the capacity present in Jason as a service. This initial Jason-RS that we have exploited with the collaboration with the Iot plat- form to intelligently process the data received from the connected object. By exploiting the horn clause and the first-order logic included in Jason we will use this ability to decide the action of objects. Then we take into account the behavior of the objects as perception for the agents. To validate our approach we tested the interoperability of Jason-RS and heterogene- ous objects choreographed with BeC3. 3.2 Application of the architecture with a python object and BeC3 Scenario. To illustrate and motivate our approach, we introduce a concrete application using the following scenario. We take the example of the disposal of household waste in our paper [5]. Waste disposal agents have waste disposal costs that vary according to time constraints, people, distances traveled and evacuation loads. A decision-maker who gives these opinions to optimize the evacuation of waste thanks to the information 7 that the evacuator agents gives him. Information of each of its agents could be changed thanks to the perception emitted by different connected objects. The Fig. 3 illustrates the application of this scenario. We used BeC3 for the service Mashup. BeC3 is a com- poser who allows a user to choreograph heterogeneous objects. In our experimentation we have created sensor objects choreographed in BeC3 with other objects. This sensor sends the data captured via Jason-RS in POST for the agents to change the perception of the Jason agent (receiver agent). A decision-maker analyzes the present data and the cost of performing the tasks for each of his agents in order to decide which one is du- rable, and which one can perform a task at lower cost. The sensor is controlled by the Python version of service compatible with BeC3. When the service is launched, it can send a POST request to BeC3 to identify itself and then another POST request to set up its feature handled by the device. Any mashup designed by a user can then be assured by a communication the service on each object. The de- cision made by the agent could be consumed via a GET request to the Jason-RS. Fig. 3. Application interaction between object choreography and Jason-RS 4 Result and discussion We tested this approach in two different objects: a python object, a mobile phone. We used Jason-RS to communicate the BeC3 platform with choreographed objects and BDI agents who take the place of a decision maker instead of a human being (such as it is originally designed in BeC3. 8 There is a demo script (in Python) APIDemo.py in Pylite project (resources dir). We have used xmpp2 server for login, its specificity is the presence. So as we describe in the section 3 we must log in BeC3 in order to create a virtual programmable object: POST /login with content { "username": {xmpp login}, "password": {xmpp password}, "service": {xmpp service (for example im.bec3.com)} } After login, create a BeC3 object with the following JSON format: POST /feature with content { "name": "tot2", "path" : "truc/bidul21", "type": "switch", "details": "ooo", "widget": "none", "mqtt": false } Available types are accelerometer, button, buzzer, gauge, gps, These object can be delete, update. DELETE /feature/{id} PUT /feature/{id} { "data": {data (dependent of object type} } One object is associate with one BDI agent. To connect BeC3 an URL is available from Jason-RS. To change the agent BDI perception a POST request is send by this url format: POST jason_rs_based_url/object_agent/ { "data":value_from_python_object } To receive a decision from BDI agent, Jason-RS provide an URL that we can con- sume: GET jason_rs_based_url/agent_decider/decision 2 https://xmpp.org/ The following sequence diagram presented in Fig.4 shows the exchange between these two platforms 9 Fig. 4. Sequence diagram for communication between object a Jason agent by Jason-RS We analyzed the performance of our approach via a browser. Table 1. The calculation time of the exchanges between the objects and Jason-RS Device Python Object Smart Phone Method REST GET POST, GET Task Duration (millisecond) Consumation WS Sending data 200-450 428-1000 We tested our approach on both devices. During our experiment the return is almost <= 1000 ms. Our test area are done through the WIFI network so the discussion field is open for other types of networks available for the Internet of Things. Agent-based internet of things is an interested field of research now [18], in the face of these works, we focus in the technical side of things, and we compare in table 2 what differentiates our approach compared to the state of the art. Table 2. Comparison with a similar research domain work Work in field Ja- son agent and Agent flexibility in Horn Logic Taking in count internet Heterogeneity of IoT IoT [19] [20] ? ? yes ? ? yes 10 [21] Jason-RS and IoT Platform ? yes yes yes ? yes So in our approach we taking in count the heterogeneity of IoT by using a platform BeC3. To permit an automatic decision making we exploit the horn logic provided in Jason. This collaboration permit a flexibility in interacting between agent and hetero- geneous connected object. 5 Conclusion This article describes our contribution on service improvement in the world of Web of Object. We have integrated our previous Jason-RS research work into the sensor device connected to the web. Object of the IoT do not support the execution of artificial intelligence within themselve because there resources are very limited. Our approach is towards the decentralization of this AI treatment with Jason-RS. Not only is Jason- RS a Web Service but it has an artificial intelligence BDI engine running. So the com- plex data processing in the connected web object is provided by the Jason. The com- munication is done through HTTP protocol as it is more flexible and commonly adopted in the world of web. We tested our approach with a Mashup service platform for heter- ogeneous connected objects called BeC3.Our approach therefore allows the object con- nected to the web to open up to data processing with artificial intelligence as well as complex systems itself. As perspective we are considering the automatic injection of service inside the connected object to reduce the human task of defining the service choreography. References 1. Laine, M. (2012). Restful Web Services for the internet of things. Online]. Saatavilla: http://media.tkk.fi/webservices/personnel/markku_laine/restful_web_services_for _the_in- ternet_of_things.pdf. 2. Gupta, P., Mokal, T. P., Shah, D. D., & Satyanarayana, K. V. V. (2018). Event-Driven SOA- Based IoT Architecture. In International Conference on Intelligent Computing and Applica- tions (pp. 247-258). Springer, Singapore. 3. Wilde, E. (2007). Putting things to REST. 4. Guinard, D., Trifa, V., & Wilde, E. (2010, November). A resource oriented architecture for the Web of Objects. In IoT (pp. 1-8). 5. Rafalimanana, H. F., Razafindramintsa, J. L., Ratovondrahona, A. J., Mahatody, T., & Manantsoa, V. (2018, December). Publish a Jason Agent BDI Capacity as Web Service REST and SOAP. In International conference on the Sciences of Electronics, Technologies of Information and Telecommunications (pp. 163-171). Springer, Cham. 6. A. S. Rao, « AgentSpeak(L): BDI agents speak out in a logical computable language », in Agents Breaking Away, vol. 1038, W. Van de Velde et J. W. Perram, Éd. Berlin, Heidelberg: Springer Berlin Heidelberg, 1996, p. 42-55. 11 7. A. Ricci, R. H. Bordini, et J. F. Hübner, « AgentSpeak(ER): An Extension of AgentSpeak(L) improving Encapsulation and Reasoning about Goals », p. 3, 2018. 8. Norling, E. (2003, July). Capturing the quake player: using a bdi agent to model human behaviour. In Proceedings of the second international joint conference on Autonomous agents and multiagent systems (pp. 1080-1081). ACM. 9. Kamphorst, B., van Wissen, A., & Dignum, V. (2009, November). Incorporating BDI agents into human-agent decision making research. In International Workshop on Engineering So- cieties in the Agents World (pp. 84-97). Springer, Berlin, Heidelberg. 10. Ciortea, A., Boissier, O., & Ricci, A. (2017, October). Beyond physical mashups: Autono- mous systems for the Web of Objects. In Proceedings of the Eighth International Workshop on the Web of Objects (pp. 16-20). ACM. 11. Kovatsch, M., Hassan, Y. N., & Mayer, S. (2015, October). Practical semantics for the In- ternet of Things: Physical states, device mashups, and open questions. In 2015 5th Interna- tional Conference on the Internet of Things (IOT) (pp. 54-61). IEEE. 12. D. Guinard, « A Web of Objects application architecture: Integrating the real-world into the web », PhD Thesis, ETH Zurich, 2011. 13. Kim, J. H. (2017). A review of cyber-physical system research relevant to the emerging IT trends: industry 4.0, IoT, big data, and cloud computing. Journal of Industrial Integration and Management, 2(03), 1750011. 14. « W3C Web of Objects at W3C ».https://www.w3.org/WoO/. 15. Cherrier, S., Salhi, I., Ghamri-Doudane, Y. M., Lohier, S., & Valembois, P. (2014). BeC 3: behaviour crowd centric composition for IoT applications. Mobile Networks and Applica- tions, 19(1), 18-32. 16. Castellani, A. P., Gheda, M., Bui, N., Rossi, M., & Zorzi, M. (2011, June). Web Services for the Internet of Things through CoAP and EXI. In 2011 IEEE International Conference on Communications Workshops (ICC) (pp. 1-6). IEEE. 17. L. Pescosolido, R. Berta, L. Scalise, G. M. Revel, A. De Gloria, et G. Orlandi, « An IoT- inspired cloud-based Web Service architecture for e-Health applications », in 2016 IEEE International Smart Cities Conference (ISC2), Trento, Italy, 2016, p. 1-4. 18. Savaglio, C.; Ganzha, M.; Paprzycki, M.; Bădică, C.; Ivanović, M.; Fortino, G. Agent-based Internet of Things: State-of-the-art and research challenges. Future Gener. Comput. Syst. 2020, 102, 1038 -- 1053. 19. Ciortea, A., Boissier, O., Zimmermann, A., & Florea, A. M. (2016, November). Responsive decentralized composition of service mashups for the internet of things. In Proceedings of the 6th International Conference on the Internet of Things (pp. 53-61). ACM. 20. Singh, M. P., & Chopra, A. K. (2017, June). The internet of things and multiagent systems: Decentralized intelligence in distributed computing. In 2017 IEEE 37th International Con- ference on Distributed Computing Systems (ICDCS) (pp. 1738-1747). IEEE. 21. Novais, P., & Konomi, S. (2016, October). From raw data to agent perceptions for simula- tion, verification, and monitoring. In Intelligent Environments 2016: Workshop Proceedings of the 12th International Conference on Intelligent Environments (Vol. 21, p. 66). IOS Press.
1207.2235	1	1207	2012-07-10T06:47:11	NAAS: Negotiation Automation Architecture with Buyer's Behavior Pattern Prediction Component	[ "cs.MA" ]	In this era of "Services" everywhere, with the explosive growth of E-Commerce and B2B transactions, there is a pressing need for the development of intelligent negotiation systems which consists of feasible architecture, a reliable framework and flexible multi agent based protocols developed in specialized negotiation languages with complete semantics and support for message passing between the buyers and sellers. This is possible using web services on the internet. The key issue is negotiation and its automation. In this paper we review the classical negotiation methods and some of the existing architectures and frameworks. We are proposing here a new combinatory framework and architecture, NAAS. The key feature in this framework is a component for prediction or probabilistic behavior pattern recognition of a buyer, along with the other classical approaches of negotiation frameworks and architectures. Negotiation is practically very complex activity to automate without human intervention so in the future we also intend to develop a new protocol which will facilitate automation of all the types of negotiation strategies like bargaining, bidding, auctions, under our NAAS framework.	cs.MA	cs	NAAS: Negotiation Automation Architecture with Buyer’s Behavior Pattern Prediction Component Debajyoti Mukhopadhyay1, Sheetal Vij2, Suyog Tasare2 Department of Information Technnology1 Department of Computer Engineering2 Maharashtra Institute of Technology Pune 411038, India {debajyoti.mukhopadhyay, sheetal.sh, suyog59}@gmail.com Abstract: In this era of “Services” everywhere, with the explosive growth of E- Commerce and B2B transactions, there is a pressing need for the development of intelligent negotiation systems which consists of feasible architecture, a reliable framework and flexible multi agent based protocols developed in specialized negotiation languages with complete semantics and support for message passing between the buyers and sellers. This is possible using web services on the internet. The key issue is negotiation and its automation. In this paper we review the classical negotiation methods and some of the existing architectures and frameworks. We are proposing here a new combinatory framework and architecture, NAAS. The key feature in this framework is a component for prediction or probabilistic behavior pattern recognition of a buyer, along with the other classical approaches of negotiation frameworks and architectures. Negotiation is practically very complex activity to automate without human intervention so in the future we also intend to develop a new protocol which will facilitate automation of all the types of negotiation strategies like bargaining, bidding, auctions, under our NAAS framework. Keywords: NAAS, Negotiation architecture, Negotiation automation, Negotiation framework, Protocols, Agents, Web Services 1. Introduction to Negotiation Negotiation is the process between two or multiple entities where everybody comes to some useful consensus or agreement as a result. Like it or not everybody is a negotiator in some ways without even knowing it. We did it as kids for trading toys, cards and still we do it for the raise in salary, purchasing things in our personal lives. Many people try to avoid this blatant negotiation procedure consciously because they don’t like it but end up in either negotiating or losing in the bargain. Now the question is how to automate this process using the latest technologies and advancements in computing. Here are some key concepts behind the theory of negotiation [6, 7, and 8]. 1.1 Negotiation Types 1. Distributive or fixed pie negotiation. It involves people who have never had a previous interactive relationship neither they are likely to do so in near future. Everybody gets a fixed pie. 2. Integrative negotiation or everybody wins something or win-win scenario. This means to join several parts into whole. This needs some cooperation and higher degree of trust from every entity in negotiation. Ideal ly it is difficult to achieve and most difficult to automate because trust and forming relationships on that is human characteristic and difficult to implement in machines and computing [9]. 1.2 Negotiation Tactics The level of detail in the best negotiator actually understands the human mind and how to use this in different voice tones and expressions for the best possible negotiation outcome for all the parties, ideally. A few common tactics that are used in negotiation are outright refusal, conditioning, calling bluffs [10]. 1.3 Negotiation protocol types It is the set of rules that govern the interaction between participants. Depending upon the types the negotiation can be categorized as bidding, auction, bargaining. [11] In this paper we are proposing a new architecture and a framework for automating the negotiation process. The key feature in this framework is a component for prediction or probabilistic behavior pattern recognition of a buyer, along with the other classical approaches of negotiation frameworks and architectures. 2. Related Work Substantial research work has been done on various negotiation protocols, languages and set of parameters, frameworks on each of the negotiation types like auctioning, bidding, bargaining, simple request and response methods etc. We reviewed the following frameworks and architectures to come up with certain conclusions and assessed if any provision is given in those for observing the buyer’s behavior. Hudert, Guido and Ludwig [1] have defined a framework for augmenting WS- Agreement by Open grid Forum (OGF) standard which actually defines an XML based structural definition of Service Level Agreements [2]. This framework is proposed only in relation with WS-Agreement protocol where parties interested in negotiating an agreement first run the negotiation Meta protocol to establish which negotiation protocol is used. [1] Subsequently, the protocol is executed to determine the resulting, negotiated WS-Agreement document. Finally, winner is determined and acceptance, rejection is performed again according to the WS-Agreement protocol standard (Fig. 1). Fig. 1. Agreement creation process Tung Bui and Gachet [3] found out that web services (using UDDI, WSDL, SOAP, and XML) can be used as a market broker, to help in discovering the supply/demand, arbitrate the pricing, find the most appropriate service for a given request, to modify the request and services and generate the contract. They have given the basic architecture as shown in Fig. 2. Fig. 2. Electronic market for web services In this framework the authors have used BPEL4WS[4,5] and WSCI and proposed following support roles as information, Communication, Negotiation, Bargaining, Execution of transactions , all of them with steps in pre -transaction, transaction and post transaction phases in electronic market. Authors have given detailed diagrams for the seven web services like topology of web services for negotiation and bargaining, service discovery, Adaptation and pricing, Service ranking, service bargaining, best price adaptation, contract composition, here the client is seen as a negotiation manager. This architecture has its limitations like assumption of trust and there is no competitive strategy to distort the cooperative spirit of e -market. There is no fixed pricing, brokerage and addressing of services, no technical requirements as security and logging mechanisms discussed. Bin Wu and Chaozhen Guo [12] present a new web service negotiation mechanism and new web service composition coordinated negot iation architecture to solve the problems which mainly occur in such architectures of web service composition based on agent. The problems like unreasonable use of time and data link resources in the condition of multi-negotiation concurrency which leads to inefficiency of negotiation, lack of effective processing when confronting negotiation failure. Here the authors are applying asynchronous communication theory to the process of Web Service Negotiation and extend effective processing in case of negotiation failure. There are disadvantages as overhead of time spent in processing failure may be more than the benefit it brings, if the bad status of network (Fig. 3 and Fig. 4). Fig. 3. Original web service composition coordinated negotiation architecture Fig. 4. Improved web service composition coordinated negotiation architecture (IWSCCNA) It divides NA into SPNA and SRNA. Here a pseudo algorithm of process handling negotiation failure is given. So the research focuses only on elaborating architecture (Improved Vs OWSCCNA) for proper processing in case of negotiation failure in web services. Jin and Segev [13] have proposed a framework for negotiation processes that provide a consistent model for supporting a comprehensive range of negotiations in dyn amic next generation e-business environment. It has five components negotiation requirements, negotiation structure, negotiation process, negotiation protocol and strategy. The authors have described each of their framework components in stepwise details and claim that this model is the most flexible and practical where protocol and strategy are separated in the design, as shown in Fig. 5 Fig. 5. Negotiation Framework Chhetri, Lin, Goh, jun, Jian, Kowalczyk [14] have proposed an agent based coordinated negotiation architecture to ensure collective functionality, end to end QoS and coordination of complex services and they describe how this architecture can be used in different application domains and also how the negotiation system on the service provider’s side can be implemented both as an agent based negotiation as well as a web service based negotiation system. This work is based on ASAPM, adaptive agreement and process management, aims at developing intelligent agent based techniques and tools to facilitate the adaptive service management and process management. The overall architecture of ASAPM includes following four components (Fig. 6). Fig. 6. ASAPM Coordinated negotiation architecture is proposed on the above as shown in Fig. 7. Fig. 7. Coordinated Negotiation Architecture This two level approach enables the reuse of this architecture in any application domain. This paper does not contain the decision making strategies both at the negotiation level and at the coordination level. All these architectures are mostly implemented in FIPA compliant JADE Agent Framework and WS2JADE toolkit which enables the integration of JADE agents and web services. Stanley, Huang, Yihua, Haifei Li, Wang, Liu, Lee, Lam [15] have presented the design and implementation of a replicable, internet based negotiation server for conducting bargaining type negotiations between enterprises involved in e -commerce and e-business where enterprises can be buyers and sellers of product/services or participants of a complex supply chain involved in purchasing, planning and scheduling. The use of negotiation servers to conduct automated negotiation has been demonstrated by the authors. A content specification language for information registration, a negotiation protocol and its primitive operations, an automated negotiation process, a cost benefit decision model and the architecture of an implemented system have been described in this. This is based on object oriented, active database technology in contrast to the existing systems which are based mostly on distributed agent technology. Their negotiation server is analogous to web server which provide following negotiation services as a registration service, a negotiation proposal processing service, an event and rule management service, a cost-benefit analysis service. The system architecture consists of following components, external to WAN components as shown in Fig. 8. Fig. 8. System Architecture Cao, Chi, Liu [17] have also described an automated negotiation archite cture based on software multi agents using SOA (Service oriented architecture) and web services technology. Here is a reference to Negotiation as a Service where the authors observed that in businesses, negotiation process is better given as a service, not as software and more suitably the entities can be seen in a role of a service provider. Authors correlate it with SaaS (software as a service) where software providers deploy application software on their own servers and customers search, access software services via Internet, consume services as per demand and pay the software providers based on time and number of services consumed. So the users buy software service as an alternative to permanent license and SaaS subverts the traditional life cycle of software i.e. design- development-testing-installation. User need not concern about the upgrade and maintenance of software. Here the authors have given a six layered architecture for the automated negotiation as in Fig. 9. Fig. 9. Architecture for Automated Negotiation Though this prototype has been developed using SOA and UDDI, WSDL, XML technologies , it could not be a standard for practical implementation of negotiation automation due to many pervasive issues and conflicts in standards and overhead due to parsing of SOAP messages and XML documents on every exchange. Precisely in all the above architectures, even if they can be converted into a prototype of automated negotiation, they cannot be a fixed standard for the practical implementation of the same. Also there is no consideration of understanding and predicting the buyer’s behavior in all the above architectures, which is if added it can be a very sophisticated and futuristic approach of negotiation automation. 3. NAAS Automation Architecture We are proposing an automated negotiation system in the form of web service calling it as NAAS (Negotiation as a Service). As we saw in survey, using SOA (Service oriented architecture) we can make the system running in any place as a service node that is integrated with third party e-commerce platforms so the system can play role of negotiation service provider in the real business environment. The benefits are , we can obtain stable visiting quantity, maintenance and upgrade of system can be completed on the server independently, saving human and material resources, automated negotiation system can make use of the existing basic facilities provided by e-commerce platform i.e. security, authentication, transaction management etc. ,saving costs of development. With NAAS we are trying to overcome the issues we saw in the previous architectural styles. Specifically we are trying to add a module which can be developed on a strong algorithmic base i.e. using Association rules or Markov models to predict and then further analyzing the buyer’s behavior for making the process of negotiation automation complete. The main reason for buyer’s behavior prediction module in the basic design is to make the negotiation system intelligent, sophisticated and futuristic so that a better standard for negotiation automated can be created and existing designs can be enhanced. NAAS Architecture Components: 1.Service registry (databases types, directory) 2.Negotiation support system (Negotiation support system provider, Negotiation service requestor) 3.Protocol on internet module, MTS and service discovery by using web services on UDDI, HTTP, SOAP 4.Advertisement publishing from provider/seller 5.Negotiation service requestor/ buyer 6 .Buyer’s behavior pattern prediction (proposed in auction, in bargaining, in bidding for further research by the authors) 7.Business logic module and agent management module (external to NAAS) 8.Strategies, decision modules (external to NAAS). Working: The seller will publish its information about product and the prices etc. On the service registry via web services and web container to which a negotiation support system will interact on some negotiation protocol or all the existing protocols. A MTS is on the internet to transfer the requests from buyers a s well as product information from the sellers. A database can be maintained for all these service related queries and information with different ontology, if this architecture is considered in details. In our buyer’s behavior prediction component, some key features like age, gender, culture, type of product, feedback can be taken as input to the system for the decision making and predicting the buyer’s choices, behavior according to the region, country and states. This will help us in tapping the particular market and applying further strategies in negotiation system according to the buyer’s behavioral aspects for all the types in negotiation like bidding, bargaining and auctions for the desired product on sale (Fig. 10). Fig. 10. NAAS Architecture 4. Discussion on buyer’s behavior prediction component Due to addition of the buyer’s behavior pattern prediction, this architecture can be a flexible, reliable because if in advance of any negotiation process in a multiple buyers and sellers scenario on web based automated negotiation system, we are able to analyze how many potential buyers can be in this negotiation process and what kind of negotiation behavior the particular set of buyers go into. We are proposing a buyer’s behavior prediction component as shown in Fig. 11. Fig. 11. Buyer’s behavior prediction component First, data processor processes the input and then forwards the required values to the prediction logic. Our proposed prediction logic computes the offer values according to the feature which is submitted by the buyer. According to the present proposed prediction logic, we can generate such five types of offer values according to the particular feature (e.g. , age, gender, culture, type of product, buyer’s demand or feedback) For e.g. Feature (F1) = Age then we can divide this into 3 categories like Age[10 - 30] , [30 - 50 ], [50 – 70], we can allocate a weighted value for each of these age groups which will be divided by count (C) of buyer(s) to calculate the percentage. This percentage can be used by seller(s) to negotiate on the Offer Value (OV) of that particular feature (F1...F5). Please see Fig. 12. Fig. 12. Feature Categories This leads to a very successful implementation of various negotiation strategies (auction, bidding, bargaining) etc. There is no existing architecture on negotiation which facilitates the intelligent buyer’s behavior prediction in automated negotiation. We are working on the survey of appropriate existing algorithms for the prediction. 5. Conclusion and Future Work The overall survey of existing architectures leads us to come up with shortfalls and requirements in negotiation automation. NAAS can provide the advantages and further applications related to this. The authors of this manuscript are trying to as sess the work done in this area and come up with some conclusions about how to construct as well as deploy the buyer’s behavior pred iction component for the e-negotiation system. We intend to find out an appropriate algorithmic base to implement this scheme. We also intend to create a new negotiation protocol for our NAAS framework in the next phase of work. References 4. BPEL4WS, 1. Sebastian Hudert, Heiko Ludwig, Guido Wirtz: “A Negotiation Protocol Framework for WS- agreement”; IEEE; 2006 2. Sebastian Hudert, Torsten Eymann, Heiko Ludwig, Guido Wirtz: “A negotiation Protocol Description Language or Automated service Level Agreement Negotiations”; IEEE; 2007 3. Tung Bui, Alexandre Gachet: “Web Services for Negotiation and Bargaining in Electronic Markets: Design Requirements and Implementation Framework”; IEEE; 2005 Introduction to http://www.research.ibm.com/convsupport/papers/BPEL%20&%20Conversations.htm 5. Jin Baek Kim, Arie Segev, Ajit Patankar, Min Gi Cho: “Web Services and BPEL4WS for Dynamic eBusiness Negotiation Processes”; IEEE; 2004 6. Raymond Y.K. Lau: “Towards a web services and intelligent agents-based negotiation system for B2B eCommerce”; IEEE; 2007 7. Yonglei Yao, Li Ma: “Automated Negotiation for Web Services”; IEEE; 2010 8. Negotiation Types, http://www.negotiations.com/articles/negotiation-types/ 9. NegotiationTactics, http://www.negotiationtactics.net/ 10. Negotiation Protocols, http://www.w3.org/2001/03/WSWS-popa/paper19 11. Bin Wu, Chaozen Guo: “The Reasearch and Improvement on the Coordinated - Negotiation Architecture of Web Service Composition based on Agent in the condition of Multi -Negotiation Concurrency”; IEEE; 2010 12. Jin Baek Kim : “a Framework for Dynamic eBosiness Negotiation Processes”; IEEE; 2003 13. Mohan Chhetri, Jian Lin, SukKeong Goh, Jian Zhang, Ryszard Kowalczyk, Jun Yan : “A Coordinated Architecture for the agent –based Service Level Agreement Negotiation of Web Composition”; IEEE; 2006 14. Stanley Su, Chunbo Huang, Joachim Hammer, Yihua Huang, Haifei Li, Liu Wang, Youzhong Liu, Chanyote, Lee, Lam: “An internet based negotiation server for e-commerce”; 2001 Springer Journal VLDB : 72-90; 2001 15. Mu-kun Cao (School of Management, Xiamen University, China), Robert Chi, Ying Liu : “Developing Multi Agent automated negotiation service based on service oriented architec ture”; Service Science Journal, Vol.1. ; 2009
1612.00480	1	1612	2016-12-01T21:38:03	A Scalable and Adaptable Multiple-Place Foraging Algorithm for Ant-Inspired Robot Swarms	[ "cs.MA", "cs.RO" ]	Individual robots are not effective at exploring large unmapped areas. An alternate approach is to use a swarm of simple robots that work together, rather than a single highly capable robot. The central-place foraging algorithm (CPFA) is effective for coordinating robot swarm search and collection tasks. Robots start at a centrally placed location (nest), explore potential targets in the area without global localization or central control, and return the targets to the nest. The scalability of the CPFA is limited because large numbers of robots produce more inter-robot collisions and large search areas result in substantial travel costs. We address these problems with the multiple-place foraging algorithm (MPFA), which uses multiple nests distributed throughout the search area. Robots start from a randomly assigned home nest but return to the closest nest with found targets. We simulate the foraging behavior of robot swarms in the robot simulator ARGoS and employ a genetic algorithm to discover different optimized foraging strategies as swarm sizes and the number of targets are scaled up. In our experiments, the MPFA always produces higher foraging rates, fewer collisions, and lower travel and search time compared to the CPFA for the partially clustered targets distribution. The main contribution of this paper is that we systematically quantify the advantages of the MPFA (reduced travel time and collisions), the potential disadvantages (less communication among robots), and the ability of a genetic algorithm to tune MPFA parameters to mitigate search inefficiency due to less communication.	cs.MA	cs	A Scalable and Adaptable Multiple-Place Foraging Algorithm for Ant-Inspired Robot Swarms* Qi Lu1§, Melanie E. Moses1,2, and Joshua P. Hecker1 6 1 0 2 c e D 1 ] A M . s c [ 1 v 0 8 4 0 0 . 2 1 6 1 : v i X r a Abstract- Individual robots are not effective at exploring large unmapped areas. An alternate approach is to use a swarm of simple robots that work together, rather than a single highly capable robot. The central-place foraging algorithm (CPFA) is effective for coordinating robot swarm search and collection tasks. Robots start at a centrally placed location (nest), explore potential targets in the area without global localization or central control, and return the targets to the nest. The scalability of the CPFA is limited because large numbers of robots produce more inter-robot collisions and large search areas result in substantial travel costs. We address these problems with the multiple-place foraging algorithm (MPFA), which uses multiple nests distributed throughout the search area. Robots start from a randomly assigned home nest but return to the closest nest with found targets. We simulate the foraging behavior of robot swarms in the robot simulator ARGoS and employ a genetic algorithm to discover different optimized foraging strategies as swarm sizes and the number of targets are scaled up. In our experiments, the MPFA always produces higher foraging rates, fewer collisions, and lower travel and search time compared to the CPFA for the partially clustered targets distribution. The main contribution of this paper is that we systematically quantify the advantages of the MPFA (reduced travel time and collisions), the potential disadvantages (less communication among robots), and the ability of a genetic algorithm to tune MPFA parameters to mitigate search inefficiency due to less communication. I. INTRODUCTION A large number of simple individual robots working together has the potential to be useful for tasks which a tra- ditional single expensive, specialized and complicated robot is not able to handle, such as searching in large unmapped areas [1], distributed contaminant cleanup, and rescue [2]. Robot swarms can also be involved in sophisticated problem solving, including cooperative transportation, de-mining, and space exploration [3]–[6]. We focus on developing a scalable, decentralized search- and-collection algorithm based on ant-like foraging [7]–[9]. The swarm can adapt to changes in swarm size and the number of targets through real-time response to conditions without external or off-line intervention. Each robot in the swarm makes real-time in-situ decisions on whether to communicate using pheromones, forego communication but individually return to search a location, or return to the collection zone to gather additional information from other *This work was supported by a James S. McDonnell Foundation Complex Systems Scholar Award, and NASA MUREP #NNX15AM14A funding for the Swarmathon. 1The Department of Computer Science, University of New Mexico, Albuquerque, NM 87106, USA. 2External Faculty, Santa Fe Institute, Santa Fe, NM 87501, USA. §Correspondence: [email protected] robots. The robot behaviors are modeled after those of a particular genus of desert seed harvester ants that [10], [11] are restricted to foraging in short-time windows during which not all available targets can be collected; So they are designed to collect as many targets as possible, but not for optimal complete collection [12]. Here, we present the multiple-place foraging algorithm (MPFA) with multiple nests that robots depart from and return to. The robots make on-line decisions to switch to new collection zones based on proximity to their last-found target. The MPFA was presented in our recent work [13] and it shown that distributing 2, 4, or 8 nests in the MPFA produce higher foraging rates and lower average travel time compared to the central-place foraging algorithm (CPFA) developed by Hecker and Moses [14], [15]. Here we compare the scalability and adaptation of the MPFA to the CPFA when increases the number of robots and the number of targets. In the MPFA we deploy 4 nests uniformly in the same size of search arena. A set of real-valued parameters specifying the individual robot controllers is evolved by a genetic algorithm (GA) to optimize the foraging strategy in the multi-physics robot simulator Autonomous Robots Go Swarming (ARGoS) [16]. Every robot in the swarm uses the same controller. We evolve foraging strategies for different swarm sizes (4, 8, 16, 32 and 64) and number of targets (128, 256, 512, 1024 and 2048). We observe the average foraging rate, collision time, travel and search time change as swarm size and the number of targets increase. The remainder of this paper is organized as follows. Section II introduces related work. The design of the MPFA and the description of evolution are provided in Section III and Section IV. The configuration of the MPFA in ARGoS and the experimental results are in Section V and Section VI. Section VII discusses the conclusions. II. RELATED WORK Central-place foraging is commonly studied in swarm robotics [17], [18]. Hecker and Moses utilized and formal- ized the behaviors from Flanagan and Letendre's ant field studies [10], [11], [19] to create the CPFA. The algorithm is well designed and applied to real physical robots, which are designed on the iAnt robots platform [20]. The error- tolerance, flexibility, and scalability were evaluated on both simulated and real robot swarms [21]. However, the simu- lated robots were not physics-based and collisions between robots were not considered. The studies on task allocation by Hsieh et al [22]–[24] showed that a bio-inspired approach to the deployment of a homogeneous swarm of robots to multiple sites. The robots autonomously redistribute themselves among the candidate sites to ensure task completion by optimized stochastic con- trol policies. It models the swarm as a hybrid system where agents switch between maximum transfer rates and constant transition rates. In our method, the robots are distributed and initialized to multiple nests evenly. Then, they transit to other nests autonomously based on the distribution of remainder targets and the evolved search strategy. The search strategy is evolved by GA automatically and all the robots use the same strategy. There are few studies on multiple-place foraging in bio- logical systems. The polydomous colonies of Argentine ants are comprised of multiple nests spanning hundreds of square meters [25], [26]. A study by Chapman et al [27] showed that a community of spider monkeys can be considered as multiple central place foragers (MCPF). They select a sleeping site close to current feeding areas, and the MCPF strategy entails the lowest travel costs. A study by Tindo et al [28] showed that wasps living in multiple nests have greater survival rate and increased productivity. However, multiple-place foraging has not been systematically com- pared to central placed foraging in robotic swarms which we do here. III. THE DESIGN OF THE MPFA In the MPFA, robots are evenly distributed around nests. They start from a nest but return to the closest nest to their position after finding a target or giving up the search. The use of multiple collection points is the fundamental difference between the CPFA and the MPFA; all other components of the two foraging algorithms are kept deliberately identical in order to test for the effect of multiple nests on swarm foraging efficiency. The behavior of an individual robot in an MPFA foraging round is shown in Fig. 1. Each robot transitions through a series of states as it forages for targets. This differs from the CPFA [21] in how the robots return to nests which are in steps 4 and 5. In the MPFA, robots initially disperse from the nests closest to them, followed by random selected travel paths (step 1). An uninformed correlated random walk is used to search targets when robots stop to follow the paths (step 2) [29]. Robots navigate home to nests closest to them when they retrieve targets or give up search (step 4 and 5) [30]. Robots that find targets will detect the local target density before return to nests (step 3) [31]. Robots that are more likely to return to previously found sites using site fidelity or pheromone recruitment (step 6), then they search the sites thoroughly with informed walk (step 7). In our design, the robots search globally just as in the CPFA – they can travel in the entire arena. As in the CPFA, pheromone trails are simulated using a list of pheromone-like waypoints to identify target-rich areas. When a robot returns to a nest, it will probabilistically select a waypoint from the nest's list and travel to that location. It shares information Fig. 1. The flow chart of an individual robot's behavior in MPFA during an experiment. (pheromone waypoints) locally at its current nest (see Fig. 2). In contrast to the CPFA, pheromone waypoints are globally available to all robots. Since robots always return to the closest nest with a found target, the sensed information relevant to a given target neighborhood is always associated with the nest closest to the position of the identified neighborhood. Thus, if a robot follows a pheromone waypoint from a nest, then the distance from the nest to the destination of the pheromone must be the shortest distance to the target neighborhood identified by the pheromone. Fig. 2. The placement of nests and targets in ARGoS. 1024 targets (black points) and 16 robots (larger blue points) are placed in a 15× 15m2 arena, 4 nests (red circles) are distributed uniformly in the arena. The targets are arranged in a partially clustered distribution. Colored lines indicate pheromone trails with different strength. A small area is magnified to show a robot, colored pheromone waypoints, a large cluster of targets, and a single target. IV. THE GA EVOLUTION We implement the CPFA and MPFA on a foraging task for different experiments in ARGoS. Furthermore, we use a GA to identify MPFA parameters that maximize foraging strategy. We implement our GA using GAlib [32] following parameters described by Hecker and Moses [21]. The set Travel on the PathSearch withInformed WalkStartReach SearchLocationSense LocalResource DensitySearch with Uninformed WalkRandomselect pathSite fidelityPheromonesGive up search Find resourceReturn to TheClosest NestSwitch to searchRetrievingreourceSwitch tosearch1235467of seven MPFA parameters is identical to the set of CPFA parameters. The movement, sensing, and communication of each single robot are evolved and evaluated. The parameters are described in the following, • Probability of switching to search: The robot has the probability of switching from travel to uninformed random search. The probability is initialized from a uniform random distribution, U(0, 1). • Probability of returning to nest: The robot has the probability of giving up search and returning to nest. It is initialized from a uniform random distribution, U(0, 1). • Uninformed search variation: If the robot searches using a correlated uninformed random walk, the succes- sive turning angles θt is defined by θt = N (θt−1, σ), where θt−1 is the turning angle in the current step, and σ is the standard deviation or uninformed search variation, which determines the turning angle of the next step. σ is initialized from a uniform random distribution, U(0, π). • Rate of informed search decay: If the robot searches using an informed correlated random walk, the standard deviation of the successive turning angles σ decays as a function of time t, σ = ω + (2π − ω)e−λidt, where λid is the rate of informed search decay. λid is initialized from an exponential decay function exp(5). • Rate of laying pheromone and rate of site fidelity: The information decisions are governed by parameteri- zation of a Poisson cumulative distribution function as defined by POIS(k, λ), where k is the likelihood of detecting at least k additional resources, and λ is the rate of laying pheromone or the rate of site fidelity. It is initialized from a uniformed random distribution, U(0, 20). The robot returns to a previous location if POIS(k, λ) > U(0, 1). If k is large, the robot is likely to return to the same location using information on its next foraging trip. • Rate of pheromone decay: The pheromone decays exponentially over time t as defined by e−λpdt, where λpd is the rate of pheromone decay. It is initialized from an exponential decay function exp(10). We repeat the evolutionary process 10 times for the CPFA as well as for the MPFA, in order to generate 10 independently evolved foraging strategies for each experimental configura- tion. In summary, using a swarm size of 40 robots, we evaluate each swarm 8 times on different random placements of targets in the partially clustered distribution to determine their fitness. We use a 50% uniform crossover rate and a 5% Gaussian mutation rate with a standard deviation of 0.02. We use elitism to keep the individual with the highest fitness. We altered the termination criteria of the GA in order to hasten parameter convergence and ran the GA for a maximum of 100 generations. The GA terminates based on three criteria: the number of generations, the convergence of fitness, and the diversity of swarm sizes, which are introduced in GAlib [32]. The GA will stop if the fitness is convergent and the diversity of the population is low. Otherwise, it will stop after 100 generations. Our code is available on GitHub1. In our GA, 89% of the evolution terminates on the con- vergence of fitness and the diversity of swarm sizes. Across 10 independent evolutionary runs, all evolved parameter sets were nearly equally fit: The standard deviation in fitness was at most 5% of the mean fitness value. The fitness of the best parameter set, evaluated on 100 target placements, is shown in Fig. 3 and Fig. 4. V. EXPERIMENTAL CONFIGURATION IN ARGOS Table I shows the experimental configuration in ARGoS. To test scalability, the number of targets is always 1024, and the number of robots is scaled to be 4, 8, 16, 32 or 64. We set different foraging time windows for each swarm, depending on the swarm size. The selected times allow the evolved swarms to collect approximately half of the targets. The foraging time of robots are the same across all experiments: by multiplying the number of robots by the foraging time, we have 480 robot-minutes (or 8 robot-hours) in our experiments (see Table I). To test adaptation, the number of robots is always 32. The number of targets is 128, 256, 512, 1024 or 2048. The foraging time is set independently for each experiment so that approximately 40% of the targets are collected by the best evolved strategy. All experiments are replicated 100 times. The locations of targets and robots are initialized randomly in the 100 replicates. TABLE I EXPERIMENTAL CONFIGURATION IN ARGOS Scalability Adaptation Time (minutes) Robots Targets Robots Targets Time (minutes) 4, 8, 16, 32 or 64 1024 120, 60, 30, 15 and 7.5 32 128, 256, 512, 1024 or 2048 5, 8, 10, 12 and 30 The targets are placed in a partially clustered distribution. This distribution has various sizes of square clusters. The targets are placed either in a large cluster, a medium cluster or individual targets in a uniform random distribution (see Fig. 2). Both algorithms are tested in a simulated arena size of 15× 15m2. The CPFA has one center nest and the MPFA has 4 uniformly and evenly distributed nests. VI. RESULTS We compare the efficiency of the CPFA and the 4 nest MPFA on foraging rate, collision time, and travel and search time when the swarm sizes and the number of targets are scaled up. We identify statistical differences using a t-test, and we identify whether performance varies systematically by calculating a log-linear regression in which the perfor- mance are compared to the log2 of the swarm sizes or the number of targets. 1https://github.com/BCLab-UNM/iAnt-ARGoS/tree/ lukey_development A. Foraging Efficiency B. Collision Efficiency The total foraging rate of each swarm is the sum of the total collected targets in the swarm. We measure the average foraging rate, which is the number of targets per robot collected in every minute. Fig. 3 shows the average foraging rate as the swarm size increases. The average foraging efficiency of the MPFA exceeds that of the CPFA in all cases, by up to 66% in the case of 64 robots. In our simulation, if the distance between two robots is less than 0.25m, each robot will detect a collision. Each robot senses the location of the other and turns left or right in order to avoid a collision, moving approximately 8 cm before resuming traveling. The collision time is the time required to avoid a collision. The total collision time of each swarm is the sum of the total collision time for all robots in the swarm. We measure the average collision time, which is the collision time per robot in collecting a target. The "per robot, per target" collision makes the comparison fairly. For "per robot", it is obvious that a larger swarm results in more collisions, but the rate of increase is not obvious. It is easier to analyze the trend of collision rates on each robot rather than on the swarm when the swarm sizes are different. For "per target", calculating 'per target' definitely dilute the comparison. However, the collision is higher if more targets are collected. We should consider the collision rate based on the foraging rate (In our results, the MPFA always has higher foraging rate). The average collision time as swarm size increases is shown in Fig. 5. Fig. 3. The average efficiency (targets collected per robot, per minute) for the CPFA (p = 0.08) and MPFA (p = 0.04) decrease as the swarm size increases. The p value is from the average of collected targets and the log2 of the swarm size. Results are for 100 replicates. The percentage of improvement is labelled. Fig. 4 shows the average foraging rate as the number of targets increases. The average foraging efficiency of the MPFA exceeds that of the CPFA in all cases, by up to 50% in the case of 2048 targets. Fig. 5. The average efficiency (collision time per robot, per target) for the CPFA (p = 0.06) and MPFA (p = 0.10) increase as the swarm size increases. The collision time for the MPFA is less than the collision time for the CPFA. We also see that the collision time for the MPFA is reduced as the number of targets increases (see Fig. 6). C. Travel and Search Efficiency Foraging time is composed of two distinct activities. When a robot departs from its nest, it travels to a location where it starts to search for targets. Once at the destination, the robot engages in a localized search. Once a target is discovered, the robot takes approximately the same travel time back to the nest. Some robots take approximately the same travel time back to the location of the discovered target if they are recruited by pheromones, but their search time is reduced. We measure the average travel time and search time spent to collect one target by a robot. The average travel time for Fig. 4. The average efficiency (targets collected per robot, per minute) for the CPFA (p = 0.04) and MPFA (p = 0.001) decrease as the number of targets increases. The p value is from the average of collected targets and the log2 of the number of targets. The efficiency is always higher for the MPFA. Fig. 6. The average efficiency (collision time per robot, per target) for the CPFA and MPFA as the number of targets increases (p = 0.03). Fig. 8. The average travel time (per robot, per target) for the CPFA (p = 0.001) and MPFA (p = 0.03) as the number of targets increases. the MPFA (see Fig. 7) is less than the CPFA for all swarm sizes. Fig. 7. The average travel time (per robot, per target) for the CPFA and MPFA decrease as the swarm size increases (p = 0.04). Fig. 9. The average search time (per robot, per target) for the CPFA and MPFA as the swarm size increases (p = 0.03). The travel time for the MPFA (see Fig. 8) is also less than the CPFA as the number of targets increases. Fig. 9 shows that the average search time decreases as the number of robots increases. The search time for the MPFA is less than the CPFA. The search time for the CPFA decreases faster than the MPFA. The improvement is up to 34% in the first case for 4 robots and it is down to 19% in the last case for 64 robots. The search time decreases as the number of targets increases (see Fig. 10). The search time for the MPFA decreases faster than the CPFA. The improvement goes up to 31% in the last case. VII. DISCUSSION This paper explores how swarm size and the number of targets affect foraging rates, collision time, travel and search time. Not surprisingly, increasing the swarm sizes or the number of targets decreases the average foraging rate (see Fig. 10. The average search time (per robot, per target) for the CPFA and MPFA as the number of targets increases (p = 0.05). Fig. 3 and Fig. 4), but decreases slower for the MPFA. This implies that the MPFA is more efficient in larger swarms or in an environment with more targets. The average collision time for the MPFA is much less than the CPFA as the swarm size or the number of targets increases (see Fig. 5 and Fig. 6). The collision time for the CPFA increases faster as the number of targets increases (see Fig. 6). We hypothesize that the more targets there are, the harder robots will spread out in the CPFA. It demonstrates that the MPFA has the advantage of avoiding collisions in large swarm size or in an environment with a large number of targets. The increase of swarm sizes makes the average travel time for the MPFA decrease faster than the CPFA (see Fig. 7). This shows that the MPFA has the advantage of reducing travel time as the swarm size increases. We hypothesize that the evolved probability of returning to nest increases faster as the swarm size increases. The more robots there are, the more likely robots will minimize time traveling in the MPFA. On the other hand, this makes the search time has smaller difference (see Fig. 9). It is obvious that the addition of more nests makes the travel time less for the MPFA. However, the information (pheromone waypoints) is distributed to multiple nests. In contrast to the CPFA, pheromone waypoints are globally available to all robots. So, there are tradeoffs among com- munication (and therefore search time) and travel time and congestion. In addition, the MPFA may get the benefit from all resources are not eventually be moved to one nest. However, we can consider a "high-speed" delivery (multiple targets can be moved in one round) in the future. This may not make the results too much different. The search time for the MPFA decreases faster with increasing numbers of targets (see Fig. 10). This shows that the MPFA has the advantage of reducing search time in an environment with large number of targets. We hypothesize that the evolved probability of laying pheromone increases and it is higher for the MPFA, or the rate of pheromone decay decreases and it is lower for the MPFA as the number of targets increases. The more targets there are, the more likely pheromone will be laid, or slower pheromone decay. These discoveries reveal that the MPFA improves foraging rates when the swarm size or the number of targets are scaled up. This is not only because of the simple intuitive reduction in travel time, but also because of the significant improvement in avoiding collisions. Overall, the MPFA has better performance as the swarm size or the number of targets increases. In the future work, we will discover the trends on the evolved seven parameters and confirm the above hypothesis for the random, partially clustered and clustered resource dis- tributions. In addition, we will consider the cost of deploying multiply nests and evolve the optimized number of nests for different resource distributions. ACKNOWLEDGMENT We gratefully acknowledge members of the Moses Biolog- ical Computation Lab for their assistance with the multiple- place foraging swarm robotics project. Thanks to Antonio Griego for developing the CPFA algorithm in ARGoS. Thanks to Matthew Fricke for discussing the issues in the experiments. REFERENCES [1] D. P. Stormont, Autonomous rescue robot swarms for first responders, In: Proceedings of the 2005 IEEE International Conference on Com- putational Intelligence for Homeland Security and Personal Safety, IEEE; March, 2005. pp. 151-157. [2] G. Kantor, S. Singh, R. Peterson,D. Rus, A. Das, V. Kumar, et al. Distributed Search and Rescue with Robot and Sensor Teams, In: The 5th international conference on field and service robotics, vol. 24. Germany: Springer; 2006. pp. 529-538. [3] R. A. Brooks and A. M. Flynn, Fast Cheap and Out of Control: A Robot Invasion of the Solar System, Journal of the British Interplan- etary Society, 1989, vol. 42, pp. 478-485. [4] G. A. Landis, Robots and humans: synergy in planetary exploration, Acta Astronautica, 2004, vol. 55, no. 12, pp. 985-990. [5] W. Fink and J. M. Dohm, M. A. Tarbell, T. M. Hare and V. R. Baker, Next-generation robotic planetary reconnaissance missions: A paradigm shift, Planetary and Space Science, vol. 53, no. 14, pp.1419- 1426, 2005. [6] Karl A. Stolleis, Joshua P. Hecker, Melanie E. Moses, "The Ant and the Trap: Evolution of Ant-Inspired Obstacle Avoidance in a Multi- Agent System," Proceedings of Earth & Space 2016 Engineering for Extreme Environments, ASCE, Reston, VA, USA. [7] D. M. Gordon and A. W. Kulig, Founding, Foraging, and Fighting: Colony Size and the Spatial Distribution of Harvester Ant Nests. Ecology, vol. 77, no. 8, pp. 2393-2409, Ecological Society of America, 1996. [8] A. F. T. Winfield, Foraging Robots, in Encyclopedia of Complexity and Systems Science, pp. 3682-3700, Springer New York, 2009. [9] W. Liu and A. F. T. Winfield, Modeling and Optimization of Adaptive Foraging in Swarm Robotic Systems, Int. J. Rob. Res. vol. 29, no. 14, pp. 1743-1760, Dec. 2010. [10] T. P. Flanagan, K. Letendre, W. Burnside, G. M. Fricke, and M. E. Moses, How ants turn information into food. Artificial Life (ALIFE), 2011 IEEE Symposium, pp. 178-185, April 2011. [11] T. P. Flanagan, K. Letendre, W. R. Burnside, G. M. Fricke, and M. E. Moses, Quantifying the Effect of Colony Size and Food Distribution on Harvester Ant Foraging. PLoS ONE, vol. 7, no. 7, July 2012. [12] J. P. Hecker, J. C. Carmichael, and M. E. Moses, Exploiting clusters for complete resource collection in biologically-inspired robot swarms. IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), September, 2015. [13] Q. Lu, J. P. Hecker and M. E. Moses, The MPFA: A multiple-place foraging algorithm for biologically-inspired robot swarms, IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), Deajeon, October, 2016 (In review). [14] J. P. Hecker, K. Letendre, K. Stolleis, D. Washington, and M. E. Moses, Formica Ex Machina: Ant Swarm Foraging from Physical to Virtual and Back Again, Proceedings of the 8th International Conference on Swarm Intelligence, ANTS'12, pp. 252-259, Berlin, Heidelberg, Springer-Verlag, 2012. [15] J. P. Hecker, and M. E. Moses, An Evolutionary Approach for Robust Adaptation of Robot Behavior to Sensor Error, Proceedings of the 15th Annual Conference Companion on Genetic and Evolutionary Computation, GECCO '13 Companion, ACM, no. 8, pp. 1437-1444, New York, NY, USA, 2013. [16] C. Pinciroli, V. Trianni, R. O'Grady, G. Pini, A. Brutschy, M. Bram- billa, N. Mathews, E. Ferrante, G. D. Caro, F. Ducatelle, M. Birattari, L. M. Gambardella and M. Dorigo, ARGoS: a Modular, Parallel, Multi-Engine Simulator for Multi-Robot Systems, Swarm Intelligence. vol. 6, no. 4, pp. 271-295, Springer. Berlin, Germany, 2012. [17] E. S¸ahin, S. Girgin, L. Bayindir and A. Turgut, Swarm Robotics, Swarm Intelligence. Natural Computing Series, Springer Berlin Hei- delberg, pp. 87-100, 2008. [18] M. Brambilla, E. Ferrante, M. Birattari and M. Dorigo, Swarm robotics: a review from the swarm engineering perspective, Swarm Intelligence. vol. 7, no. 1, pp. 1-41, 2013. [19] K. Letendre, and M. E. Moses, Synergy in Ant Foraging Strategies: Memory and Communication Alone and in Combination. Proceedings of the 15th Annual Conference on Genetic and Evolutionary Compu- tation, GECCO '13, ACM, no. 8, pp. 41-48, New York, NY, USA, 2013. [20] M. E. Moses, J. P. Hecker and K. Stolleis, The iAnt project. http://iant.cs.unm.edu, 2014. [23] [21] J. P. Hecker and M. E. Moses, Beyond pheromones: evolving error- tolerant, flexible, and scalable ant-inspired robot swarms, Swarm Intelligence, Springer US, vol. 9, pp. 43-70, 2015. [22] M. A. Hsieh, ´A. Hal´asz, S. Berman and V. Kumar, Biologically inspired redistribution of a swarm of robots among multiple sites, Swarm Intelligence, vol. 2, no. 2, pp. 121-141, 2008. ´A. Hal´asz, M. A. Hsieh, S. Berman and V. Kumar, Dynamic redistribu- tion of a swarm of robots among multiple sites, IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pp. 2320-2325, 2007. [24] S. Berman, ´A. Hal´asz, M. A. Hsieh and V. Kumar, Navigation-based optimization of stochastic strategies for allocating a robot swarm among multiple sites, Proceeding of the 47th IEEE Conference on Decision and Control, Cancun, Mexico, Dec. 9-11, 2008. [25] T. P. Flanagan, N. M. Pinter-Wollman, M. E. Moses and D. M. Gordon, Fast and flexible: Argentine ants recruit from nearby trails, PLOS ONE, vol. 8, no. 8, September, 2013. [26] Benefits of Dispersed Central-Place Foraging: An Individual-Based Model of a Polydomous Ant Colony The American Naturalist, 2009, vol. 173, no. 6, pp. 772-778. [27] C. Chapman, L. Chapman and R. McLaughlin, Multiple central place travel consequences of using many foraging by spider monkeys: sleeping sites, Oecologia. vol. 79, no. 4, pp. 506-511, 1989. [28] M. Tindo, M. Kenne and A. Dejean, Advantages of multiple foundress colonies in Belonogaster juncea juncea L.: greater survival and in- creased productivity, Ecological Entomology, vol. 33, no. 2, pp. 293- 297, 2008. [29] J. H. Fewell, Directional fidelity as a foraging constraint in the western harvester ant, Pogonomyrmex occidentalis, Oecologia, vol. 82, pp. 4551, 1990. [30] T. O. Crist and J. A. MacMahon, Individual foraging components of harvester ants: Movement patterns and seed patch fidelity, Insectes Sociaux, vol. 38, no. 4, pp. 379396, 1991. [31] B. H´olldobler, Recruitment behavior, home range orientation and territoriality in harvester ants, Pogonomyrmex, Behavioral Ecology and Sociobiology, vol. 1, no. 1, pp. 344, 1976. [32] M. Wall, GAlib: A C++ library of genetic algorithm components, Mechanical Engineering Department, Massachusetts Institute of Tech- nology, vol. 87, pp. 54, 1996.
1210.7600	1	1210	2012-10-29T09:19:56	Study on the Availability Prediction of the Reconfigurable Networked Software System	[ "cs.MA", "cs.SE" ]	This paper describes multi-agent based availability prediction approach for the reconfigurable networked software system.	cs.MA	cs	Study on the Availability Prediction of the Reconfigurable Networked Software System Ling-zhong Meng * , Min-yan Lu, Jing Lai, Xiao-jie Xu School of Reliability and System Engineering, Beihang University, Beijing, 100191, China This paper describes a multi-agent based availability prediction approach for the reconfigurable networked software system based on the complex system’s model and simulation. Firstly, the definition of the features and availability of reconfigurable networked software system is introduced. Secondly, on the basis of the multi-agent based networked software system modeling method, an availability prediction method of the multi-agent based reconfigurable networked software system is proposed after the analysis of the reconfigurable behavior model. Finally, in order to verify the validity of this proposed method, an actual project is taken to predict its reliability. And the simulation result indicates that the availability of reconfigurable networked software system is significantly improved. Keywords: multi-agent system, reconfigurable, networked software system, availability prediction. Received 01 June 2011; Accepted 31 August 2011 1. INTRODUCTION Networked Software System (NSS) is a kind of Internet-based online complex software system of which topology structure and activity can be evolutionary dynamically [1]. At present, the network-based software system applications and services have become an important pillar of national economy’s sustainable development, social life and national safety, and with the rapid development of the network service, NSS is also widely used in various fields, such as on-line bank, e-government, ERP software, military C4I software and so on. For some stand-alone software, in order to meet the needs of users the software needs to be manual repaired, or can recover by rebooting within a certain time once some parts fail. However, for the Networked Software System used in the critical areas, it will cause serious consequences if the software failure cannot be repaired in time. Reconfigurable software system refers to system which can be reconfigured in structure when its modules fail, according to different design principles by utilizing the dynamic adjustment of the structure in order to complete the customers’ requirements. It can improve the availability of NSS to apply reconfigurable technology * Ling-zhong Meng, email: [email protected] to the network software system, so that the software system failures can be restored in a relatively short period of time. DESEREC is an integrated project of the sixth framework program of the European Union under the "Information Society Technologies" priority, strategic objective "Towards a global dependability and security framework"[2]. One of the research content is “Enable Services Dynamic Configuration” which is to establish a secure, trusted network and service platform. NSS is constituted by a number of subsystem, the elements which compose of the subsystem call each other to complete the required tasks, while the software emerged from the overall system reliability, safety and so on. The traditional methods include the availability researches which based on the features and services[3,4], the evolution analysis of component-based model,, system reliability modeling and assessment methods based on Bayesian Networks [5,6,7] and component-based software reliability prediction, metrics [8] and analysis approach [9,10]. However, these traditional methods cannot be applied in the NSS reliability prediction due to its large scale, complex functional interaction, the complexity of the composition of the components and the reliability of its components will change with outside influence. The American scholar Santa Fe Institute has proposed the concept of the Agent to meet the complex behaviors of the complex system 1 on the macro level. The system is seen as composed of by several autonomous agents. The interactions between the agents are the source of the macro mode of the system. Through the establishment of the agent model, these relationships can better be understood and explained. Agent technology is also applied in the field of software engineering. Agent-oriented software engineering is considered an important means of supporting the development of complex software systems [11]. On this basis, the Mobile Agent-based Mobile Agent mobile system’s reliability assessment study [12, 13, 14] and security research [15] have been developed by applying the agent to different software design and mobile agent’s network services. In this article, using the multi-agent modeling and simulation methods, the reconfigurable NSS is regarded as a complex system to research its reliability and availability [16]. The structure of the paper is as follows. Section 2 reviews a brief summary of the basic concepts of agent and the main ideas of software availability. Section 3 is to study the multi-agent based reliability modeling and simulation method based the analysis of the reconfigurable NSS features. Section 4 shows experimental studies for verifying the proposed simulation with the combination of the actual reconfigurable NSS. Section 5 contains the conclusions. 2. CONCEPTS AND METHODS The dependability is proposed firstly by French scholar Laprie as a collective term in the 1985. In his view, dependability is a nature of the computer system, making the services provided by the computer system has been considered to be trustworthy. It provides a conceptual framework and a clear pattern of development for building dependability computer systems [17]. Laprie made it to be clear that dependability of software contain availability, reliability, safety, confidentiality, integrity, and maintainability in 2004[13]. Software Reliability is the probability of failure-free software operation for a specified time in a specified environment. Agents are the actors that are able to presence to meet the design goals in a particular environment with autonomy, flexibility operations [19]. Agent's structure includes the various abstract components of itself, the status and role of each abstract component in the running agent and the interaction of these abstract components. It currently includes three categories: response structure, thinking structure and hybrid structure [20]. is a Multi-agent Systems (Multi-Agent System, MAS) composition of two or more of the Agents, each Agent is an independent behavior entity, and encapsulates the state and behavior, and interacts with each other through the communication among them. MAS architecture can be divided into centralized, distributed and hybrid. The basic idea of MAS is to consider many subsystems as the composition of multiple autonomous agents, and the interaction between the agents is the origin of the macroscopic properties of the system. Through the establishment of multi-agent model, these systems can be better understood and analyzed. The essential characteristic of multi-agent simulation method is to establish a conceptual model of the actual system from the multi-agent perspective. The process is as follows: 1) Analyze the actual system; 2) establish a conceptual model; 3) build the simulation model; 4) analyze of simulation results; 5) conclusion. 3. MULTI-AGENT BASED RECONFIGURABLE NSS AVAILABILITY STUDY Firstly the concept of the characteristics of reconfigurable NSS and the availability of the system is analyzed, then the multi-agent reconfigurable NSS modeling method is studied which includes the modeling methods of the component agent and connector agent, multi-agent coordination and interaction modeling approaches and reconfigurable modeling method, the multi-agent based reconfigurable NSS availability simulation method is proposed. The process of the multi-agent based reconfigurable NSS modeling and simulation method is as illustrated in Fig.1. Analyze Actual System Modeling Identify Component and connection Identify Node and Edge Modeling Various Agent Modeling Node and Edge Modeling Agent Interaction Modeling Node Interaction Build Reconfigurable Mechanism Modeling Network Topology Evolution Simulation Build Multi-Agent Simulation System Obtain Simulation Conclusions Verify Run Analyze Simulation Results Predict NSS Availability Fig. 1. Modeling and simulation framework based on MAS 3.1 Study of the reconfigurable NSS’s Availability The availability of reconfigurable NSS can be service-oriented, rather than the entire system-oriented. A reconfigurable NSS is able to ensure all services can be completed on schedule when its subsystem’s component fail by the reconfiguration method, then the allocation strategy can be determined successful. For ease of presentation, the concepts used in the text are defined as follows: Definition 2.1 Software Service Software service refers to a set of software function provided by the NSS for the customers. Definition 2.2 Software Business The software business refers to the consumers’ use of software services, including those aggregations of services involved in the software business, the services’ usage mode and performance and quality requirements. The reconfigurable NSS is a typical complex system, composed by a variety of software, and it has a very close relationship with the external environment and customers, its features are as flows: 1) Systemic. Reconfigurable NSS are closely related to the external environment, and reconfigurable NSS itself is an organic whole with a lot of interaction among subsystems. In general reconfigurable NSS is a multi-objective, multi-discipline and developmental system; 2) Evolvability. Under certain conditions, the reconfigurable NSS can be reconfigured with the occurrence of software failures to make the structure change, while the composition of software systems is constantly changing over time; 3) Random. Many random factors affect reconfigurable NSS, such as the spread of the virus, customers’ impact, software defects, therefore reconfigurable NSS is a kind of random system; 4) Continuity. Reconfigurable NSS can provide continuous software business, management and maintenance of the database is online, generally will not stop. The availability of the NSS measures in operational availability (A 0 ), refers to the total time of the software business provided by NSS, including operation time (OT), standby time (ST), the time of 2 corrective maintenance (TCM), the time of preventive maintenance (TPM), the measurement model is as shown in (1). (1) And MUT refers to the average working time, namely OT + ST, while the MDT is the average down time, namely TCM + TPM. 3.2. Multi-agent based reconfigurable NSS Modeling the composition of software architecture, to Referring reconfigurable NSS are described with hierarchical models, object-oriented model and event-driven model. Hierarchical model is the basic structure of the entire software system simulation, object-oriented model aim to simulate the specific components in each layer, event-driven model used to simulate the interaction and communication among the components. 1. The modeling research with the component agent and connector agent Based on software architecture analysis methods, software system can be considered as a combination of many components and connectors. In this mode, component is abstracted as a node, while connector is abstracted as a line. Therefore, the nodes and lines constitute the topology network. Multiple subsystems use connections to connect with each other, constituting the software topology network. Components and connections are abstracted as different structural types of agent, therefore nodes and lines of the topology network get more intelligent to promote the multi-agent availability prediction. Component agent receives the required information from the network environment, make integration or information the information judgment according to the internal state, and then match the information in the Knowledge Base to formulate a planning, and finally generate a series of actions depending on the target, and thus have an impact on the external network environment. Therefore the component agent utilize the cognitive agent to accomplish the model, its formal description is as the following six-point groups. Com_Agent :: = < Comi, Com_set, Com_KB, Com_PS, Com_GS, Com_AS > Comi refers to the serial number of the component agent, Com_set refers to the set of states of the components of Agent, including component reliability, status and other information; Com_KB refers to the knowledge base of component Agent; the Com_PS refers to the component planning a collection of the Agent; Com_GS refers to the target set of component Agent; Com_AS refers to the action collection of component Agent. Connector agents receives the required information from the network environment, then makes the information integration, and react with the internal rule library, and finally process information back to the network environment. Therefore connector agent utilize the reactive agent to accomplish the model, its formal description is as the following four-point groups. Con_Agent :: = < Coni, Con_set, Con_PS, Con_AS > Coni refers to the connector agent's serial number; Con_set refers to the state collection of the connector agent, including the connection reliability, status and other information; the Con_PS refers to the connector agent’s planning collection; Con_AS refers to the connector agent's action set. 2. Research on multi-agent based coordination and interaction modeling NSS consists of multiple subsystems. Each subsystem is composed by several component agent and connector agent. Subsystems connect with each other through different connection agent to transfer the data and control. A number of software services will be called in order to complete the user's software business, while each service involve several component agent and connector agent, namely, the simulation of each user's business use requires multiple component agent and connector agent. Therefore, the modeling and control of multi-agent coordination and interaction become very important. Based on this, a model which is suitable for NSS’s multi-agent coordination and interaction is proposed here, shown in Fig. 2. Software Business Environment Information Bus_Agent Service Rules Service Strategy Ser_Agent Service Learning Action Output Fig. 2. Multi-agent coordination and interaction model The proposed Bus_Agent and Ser_Agent is used to run a multi-agent coordination and interaction model, Bus_Agent refers to the agent which manages multiple services’ use, the Ser_Agent set up the users’ service of NSS, their formal description is as follows: Bus_Agent :: =< Busi, Bus_set, Bus_ser_set, Bus_state> Busi refers to the serial number of the Bus_Agent; Bus_set refers to the composition of the Bus_Agent collection, including the business’ profiles, the transition probability of branches; Bus_ser_set refers to the set of services of business management; Bus_state refers to the status of Bus_Agent. Ser_Agent:: =< Seri, Ser_S, Ser_P, Ser_Per, Ser_Trans > Seri refers to the serial number of the Ser_Agent; Ser_S refers to the set of states; Ser_P refers to the set of perception; Ser_Per presents the perceptual function of E → Ser_P, the state-mapping of the environment refers to the perceptual input, E is the set of the environment; Ser_Trans refers to the decision-makingfunction of Ser_P * Ser_S → Ser_S, changing with according to the perceptual input and the current state to match the actual service. 3. Study on reconfigurable modeling However, in the case of software failures occur, not all of the component agents and connector agents can be reconfigured to meet the user’s requirement of the software business. During the design, the critical software business generally will be designed reconfigurable. The formal modeling of multi-agent reconfigurable process is as follows: Re_Model :: = < Re_seti, <Re_act_set>, Re_PS, Re_Cor > Re_seti refers to the serial number of the multiple agent set which participates in reconfigurable process; <Re_act_set> refers to the collection of each agent's action, in a reconfigurable process, each agent’s effective actions constitute the set of actions; Re _PS refers to the service rule library involved in reconfiguration; Re_Cor refers to the service strategies, including coordination, collaboration and consultation between the multiple agent. Re_act_set:: =< Re_act_seti, <Re_act _com>, <Re_act _con>> Re_act _com::= <Fcomi, Rcomi_set > Re_act _con:: = <Fconi, Rconi_set > The Re_act_seti refers to the serial number of the reconfigurable action; <Re_act_com> refers to the set of reconfigurable component agent; <Re_act_con> refers to the set of reconfigurable connector agent, Fcomi means the serial number of the failed component agent; Rcomi_set refers to a group of component agent which can complete the Fcomi function of the failed component agent through reconfigurable actions; Fconi means the serial number of the failed connector agent; Rconi_set refers to a group of connector agent which can complete the Fconi function of the failed connector agent through reconfigurable actions. 3 3.3. Study on the availability simulation of multi-agent based reconfigurable NSS There are some other functional agents in the availability simulation of the reconfigurable NSS except for component agent, connector agent, business agent (Bus_Agent), service agent (Ser_ Agent) or reconfigurable agent. These agents are used to implement specific functions to ensure the smooth of the simulation operations and the collection of simulation data. These agents include Schedule Agent (Sch_Agent), Monitor Agent (Mon_Agent), Corresponce Agent (Cor_Agent), Data Agent (Dat_Agent) and GUI Agent (Gui_Agent), etc. Schedule Agent is to set the simulation time, decide when the simulation system starts and terminates the simulation when the conditions are met simulation; Monitor Agent collected the data during the simulation in order to achieve real-time display; Corresponce Agent used to provide multiple means of communication among the agents; Data Agent can handle various data during the simulation process to describe the system’s macro performance; GUI Agent is used to receive input data and send output data. The simulation architecture is shown in Fig. 3. Dat_Agent Mon_Agent Re_Model Ser_Agent Sch_Agent Bus_Agent Com_Agent Con_Agent Cor_Agent Fig. 3. Reconfigurable NSS simulation architecture Various types of agent's properties and behavior models are described as follows: Sch_Agent=<Schi, Sch_clock, Sch_Cori>, Schi is for the sequence number of the schedule agent; Sch_clock is for the simulation clock; Sch_Cori is for the connect-out Corresponce agent. Mon_Agent=<Moni, Mon_type, Mon_ Cori >, Moni is for the sequence number of the monitor agent; Mon_type is for the type of the monitor agent; Mon_ Cori is for the connect-out corresponce agent. Cor_Agent=<Cori, Cor_type, in_agents, out_agents>, Cori is for the sequence number of the Corresponce Agent; Cor_type is for the type of the Corresponce Agent, includes monitor agent corresponce, component agent corresponce, connect Agent corresponce; in_agents is for the connect-in agent; out_agents is for the connect-out agent. Dat_Agent=<Dati, Data_type>, Dati is for the sequence number of data agent; Data_type is for the type of data agent, Gui Agent=<Guii, Gui_type>, Guii is for the sequence number of the GUI agent; Gui_type is for the type of the GUI Agent. In the simulation platform, a variety of agents’ behavior is scheduled to coordinate by schedule agent in order to ensure the normal operations of the simulation system. The simulation clock utilizes the equal step method to promote service agent, connector agent, component agent, and ensure the time synchronization of each agent and update the status at the same time. The statistical time of different kinds of time in the availability metrics model is arrived through simulation clock, thus the prediction of the reconfigurable NSS availability will be obtained. 4. APPLICATION An information processing system is selected to study its availability simulation here. The system is responsible for receiving all kinds of data to meet a variety of requirements of management and application. Meanwhile, the system output different kinds of special information for other systems in order to meet the diverse requirements of the users. Thus, this demand the systems must have very high availability. This information processing system is taken as an application case to study the availability prediction based on the multi-agent modeling and simulation method though designing it as a reconfigurable system. As stated above, this proposed application is study on the simulation platform ‘Netlogo’. Netlogo is a modeling and simulation platform, which is programmable and able to simulate the evolution of complex object behavior. It’s especially suitable for the complex systems’ modeling and simulation which is constantly changing with the interaction time. In order to facilitate the system simulation, the following principles are set for the implementation of simulation. 1) Both component agent and connector agent have only two states, failure mode and normal running state; 2) The failure of the component agent and connector agent are independent with each other; 3) In order to accelerate the operation of the software system, the reliability of component agent and connector agent are both set to 0.9999; 4) Reconfigured behavior can be instantaneous with key business-related component failure occurs; 5) The using probability of each software business is the same. On this basis, multi-agent method is utilized to make the availability model of the reconfigurable NSS. The simulation system has been run to get the final state which is shown in Fig. 4. Fig. 4. Availability of simulation results The nodes represent the component agent, amount to 306. Independent node refers to this component agent can provide some of the software business alone. Directed lines represent the connector Agent, amount to 459. Black nodes refers to the agent cannot be reconfigured to returned to normal once failure occurs, and these agents are the agent composed of non-critical software business. The dark gray nodes refers to the agent can be reconfigured to returned to normal once failure occurs, and these agents are the agent composed of critical software business. The simulation system’s availability curve is showed in Fig. 5 after running for some time. The horizontal axis represents the running time; the vertical axis represents the availability of the system, the value of availability metric interval is set to [50%, 100%] in order to display expediently. The black curve represents the availability variation tendency of the non-reconfigurable system; the gray curve represents the availability variation‘s tendency of the 4 14. C. Chowdhury, S. Neogy. Reliability Estimate of Mobile Agent System for Qos Manet Applications. Proceedings Annual Reliability and Maintainability Symposium, (2011) Jan 24-27; Florida, USA 15. E. M. Holloway, G. B. Lamont, and G. L. Peterson. Network Security Using Self Organized Multi Agent Swarms. IEEE Symposium on Computational Intelligence in Cyber Security, (2009) March 30 -April 2; Nashville,USA 16. L. Meng, M. Lu, and J. Lai. The Multi-Agent Based Reliability Simulation Approach for Networked Software System. International Journal of Advancements in Computing Technology. 3, 10 (2011) 17. J. C. Laprie. Dependable Computing and Fault Tolerance: Concepts and Terminology, IEEE the 15th International symposium on fault-tolerant computing, (1985) June 15-18; Wisconsin, USA 18. A. Avizienis, J.C. Laprie. Randell B. et al. Basic Concept and Taxonomy of Depenable and Secure Computing. Dependable and Secure Computing. 1, 1 (2004) 19. M. Wooldridge, N. R. Jennings. Agent Theories, Architectures, and Langeuages: A Survey.Workshop on Agent Theories Architecture and languages, (1995)August 19-20; Montreal, Canada 20. M. Wooldridge, N. R. Jennings. Intelligent Agent: Theory and Practice. Knowledge Engineering Review. 10, 2 (1995) reconfigurable system. Fig. 5. Comparison to reconfigurable NSS and non-reconfigurable NSS According to the analysis of the above results, the availability of the two systems is similar in the early running period of NSS. However, with the advance of simulation time, the reconfigurable system’s availability is significantly higher than the availability of non-configuration system, and the high part is gradually increasing. This indicates that reconfigurable behavior has a significant role in the availability of NSS. Therefore, in some critical system, using the reconfigurable method can improve this kind systems’ availability. 5. CONCLUSIONS According to the multi-agent technology, this paper analyzes and solves the availability prediction of reconfigurable NSS from the perspective of complex systems. Multi-agent based modeling and simulation method is used to verify that the availability of reconfigurable system have obvious improvement than non- reconfigurable System with running time increasing. In the future work, we can do further research from the following aspects: To Coordinate the number and costs of the reconfigurable component agent and connector agent with development costs; and based on works above, to research the selection criteria of the reconfigurable component agent and connector agent. References and Notes Reconfigurability, Information Systems. 1. Z. Feng, K. He, and Y. Ma. Towards Individualized Requirements Specification Evolution for Networked Software Based On Aspect. Making Globally Distributed Software Development a Success Story, (2008) May 10-11; Leipzig, Germany 2. Dependability and Security by Enhanced http://www.deserec.eu/index.html 3. Caban D. Dependability of Large Reconfigurable DepCos-RELCOMEX, (2009) June 30 -July 2; Brunow, Poland 4. Y. O. Boon, Y. Ch. Huah, Y. N Cheah. Dynamic Service Placement and Redundancy to Ensure Service Availability During Resource Failures. Information Technology (ITSim), (2010) Jun 15-17; Kuala Lumpur, Malaysia 5. E. L. Droguett, M. Moura, and C. M. Jacinto. A Semi-Markov Model with Bayesian Belief Network Based Human Error Probability for Availability Assessment of Downhole Optical Monitoring Systems. Simulation Modeling Practice and Theory. 16, 10 (2008) 6. C. W. Kang, M. W. Golay. A Bayesian Belief Network-Based Advisory System for Operational Availability Focused Diagnosis of Complex Nuclear Power Systems. Expert Systems with Applications. 17, 1 (1999) 7. M. Neil, D. Marquez. Availability Modeling of Repairable Systems Using Bayesian Networks. Engineering Applications of Artificial Intelligence, (2010) 8. M. Palviainen, A. Evesti, and E. Ovaska. The Reliability Estimation, Prediction and Measuring of Component-Based Software. The Journal of Systems and Software. 84, 6 (2011) 9.C. Hsu, C. Huang. An Adaptive Reliability Analysis Using Path Testing for Complex Component-Based Software Systems. IEEE Transactions on Reliability. 60, 1 (2011) 10. H. Hu. Reliability Analysis for Component-Based Software System in Open Distributed Environments. International Journal of Computer Science and Network Security. 7, 5 (2007) 11. N. R. Jenning. An Agent-Based Approach for Building Complex Software Systems. Communication of ACM. 44, 4 (2001) 12. M. Daoud, Q. H. Mahmoud. A Fuzzy Approach to Reliability Estimation of Mobile Agent-Based Systems. IEEE International Conference on Systems, Man and Cybernetics, (2007) Oct 7-10; Montreal, Canada 13. C. Chowdhury, S. Neogy. Estimating Reliability of Mobile Agent System for Mobile Ad Hoc Networks. Third International Conference on Dependability, (2010) July 18 - 25; Venice, Italy 5
1010.0034	1	1010	2010-09-30T23:03:07	Spectral Control of Mobile Robot Networks	[ "cs.MA", "eess.SY", "math.OC" ]	The eigenvalue spectrum of the adjacency matrix of a network is closely related to the behavior of many dynamical processes run over the network. In the field of robotics, this spectrum has important implications in many problems that require some form of distributed coordination within a team of robots. In this paper, we propose a continuous-time control scheme that modifies the structure of a position-dependent network of mobile robots so that it achieves a desired set of adjacency eigenvalues. For this, we employ a novel abstraction of the eigenvalue spectrum by means of the adjacency matrix spectral moments. Since the eigenvalue spectrum is uniquely determined by its spectral moments, this abstraction provides a way to indirectly control the eigenvalues of the network. Our construction is based on artificial potentials that capture the distance of the network's spectral moments to their desired values. Minimization of these potentials is via a gradient descent closed-loop system that, under certain convexity assumptions, ensures convergence of the network topology to one with the desired set of moments and, therefore, eigenvalues. We illustrate our approach in nontrivial computer simulations.	cs.MA	cs	Spectral Control of Mobile Robot Networks Michael M. Zavlanos, Victor M. Preciado and Ali Jadbabaie 0 1 0 2 p e S 0 3 ] A M . s c [ 1 v 4 3 0 0 . 0 1 0 1 : v i X r a Abstract -- The eigenvalue spectrum of the adjacency matrix of a network is closely related to the behavior of many dynami- cal processes run over the network. In the field of robotics, this spectrum has important implications in many problems that require some form of distributed coordination within a team of robots. In this paper, we propose a continuous-time control scheme that modifies the structure of a position-dependent network of mobile robots so that it achieves a desired set of adjacency eigenvalues. For this, we employ a novel abstraction of the eigenvalue spectrum by means of the adjacency matrix spectral moments. Since the eigenvalue spectrum is uniquely determined by its spectral moments, this abstraction provides a way to indirectly control the eigenvalues of the network. Our construction is based on artificial potentials that capture the distance of the network's spectral moments to their desired values. Minimization of these potentials is via a gradient descent closed-loop system that, under certain convexity assumptions, ensures convergence of the network topology to one with the desired set of moments and, therefore, eigenvalues. We illustrate our approach in nontrivial computer simulations. I. INTRODUCTION A wide variety of coordinated tasks performed by teams of mobile robots critically rely on the topology of the un- derlying communication network and its spectral properties. Examples include, cooperative manipulation [1] -- [3], surveil- lance and coverage [4] -- [6], distributed averaging [7] -- [9], for- mation control [10] -- [12], flocking [13, 14], and multi-robot placement [15] -- [17], that all require some form of network connectivity, structure and, oftentimes, spectral properties. In this paper we address the problem of controlling a network of mobile robots to a topology with a desired eigenvalue spectrum. This effort is a first step towards the design of controllers that allow robots to perform their assigned tasks, while optimizing coordination within the team. The eigenvalue spectra of a network provide valuable information regarding the behavior of many dynamical pro- cesses running within the network [18]. For example, the eigenvalue spectra of the Laplacian and adjacency matrices of a graph affects the mixing speed of Markov chains [19], the stability of synchronization of a network of nonlinear oscillators [20, 21], the spreading of a virus in a network [22, 23], as well as the dynamical behavior of many decen- tralized network algorithms [24]. Similarly, the second small- est eigenvalue of the Laplacian matrix (also called spectral gap) is broadly considered a critical parameter that influences the stability and robustness properties of dynamical systems Michael M. Zavlanos is with the Dept. of Mechanical Engineer- ing, Stevens Institute of Technology, Hoboken, NJ 07030, USA, [email protected]. Victor. M. Preciado and Ali Jadbabaie are with the Dept. of Electrical and Systems Engi- neering, University of Pennsylvania, Philadelphia, PA 19104, USA, {preciado,jadbabai}@seas.upenn.edu. that are implemented over information networks [25, 26]. Optimization of the spectral gap has been studied both in a centralized [27] -- [29] and decentralized context [30]. In this paper, we propose a novel framework to control the structure of a network of mobile robots to achieve a desired eigenvalue spectrum. In particular, we focus on the spectrum of the weighted adjacency matrix of the network, with weights that are (decreasing) functions of the inter- robot distances. This construction is relevant, for example, in modeling the signal strength in wireless communication networks. Although our framework performs well with dif- ferent distance metrics, in this paper, we focus on the (cid:96)1 norm (Manhattan distance) primarily for analytical reasons, since its composition with convex functions preserves their convexity. Additionally, this metric has potential applications in indoor navigation where the presence of obstacles forces the signals to propagate in grid-like environments. We employ a novel abstraction of the eigenvalue spectrum in terms of the associated spectral moments, and define artificial potentials that capture the distance between the network's spectral moments and their desired values. These potentials are minimized via a gradient descent algorithm, for which we show convergence to the globally optimal mo- ments. Since the eigenvalue spectrum is uniquely determined by the associated spectral moments, our approach provides a way to indirectly control a network's eigenvalues. This work is related to [31], which addresses a similar problem for static robots in discrete environments. This formulation, however, is more appropriate for robotics applications, where communication depends continuously on the robot motion. The rest of this paper is organized as follows. In Section II- A, we introduce some graph-theoretical notation and useful results. In Section II-B, we formulate the control problem under consideration. We introduce an artificial potential and derive the associated motion controllers in Section III-A and discuss convergence of our approach in Section III-B. Finally, in Section IV, we illustrate our approach with several computer simulations. II. PRELIMINARIES & PROBLEM DEFINITION A. Notation and Preliminaries In this section we introduce some nomenclature and results needed in our exposition. Let G = (V , E , W ) denote a weighted undirected graph, with V = [n] being a set of n nodes, E ⊆ V × V a set of e undirected edges, and W ∈ Re a set of weights associated to the edges. If {i, j} ∈ E we call nodes i and j adjacent (or neighbors), which we denote by i ∼ j. In this paper, we consider graphs without self-loops, i.e., {i,i} (cid:54)∈ E for all i ∈ V . We denote by ai j = a ji ∈ R+ the + weight associated with edge {i, j} ∈ E , and assume that ai j = 0 for {i, j} (cid:54)∈ E . We define a walk w of length k from node v0 to vk to be an ordered sequence of nodes (v0,v1, ...,vk) such that vi ∼ vi+1 for i = 0,1, ...,k−1. We define the weight aw of a walk w = (v0,v1, ...,vk) as aw = ∏k−1 A weighted graph can be algebraically represented via its weighted adjacency matrix, defined as the n× n symmetric matrix AG = [ai j], where ai j is the weight of edge {i, j}. In this paper we are particularly interested in the spectral properties of AG . Since AG is a real symmetric matrix, it has a set of real eigenvalues λ1 ≥ λ2 ≥ ... ≥ λn. The powers of the adjacency matrix AG can be related to walks in G : i=0 avi,vi+1. Lemma 1: Let G be a weighted undirected graph with no G can be written in terms self-loops. The (i, j)-th entry of Ak of walks in G as follows: (cid:105) (cid:104) Ak G = ∑ w∈W (k) i, j aw, i j where W (k) i, j node i to j in the complete graph Kn 1. is the set of all closed walks of length k from Proof: The above result comes directly from an alge- braic expansion of(cid:2)Ak (cid:105) (cid:104) G (cid:3) Ak G v0,vk v0,vk n ∑ v1=1 = (cid:104) (cid:105) . First, notice that av0,v1 Ak−1 G . v1,vk (cid:104) Using the above rule in a simple recursion, we can expand the entries of the decreasing powers of AG , to obtain av0,v1av1,v2...avk−1,vk Ak G = (cid:105) v0,vk 1≤v1,...,vk−1≤n ∑ k−1 ∏ i=0 = ∑ w∈W (k) i, j avi,vi+1, (cid:104) Ak−1 (cid:105) which is the statement of our lemma. The following result will be useful in our derivations: Lemma 2: Let G be a weighted undirected graph with no self-loops. Then, we have that G ) ∂tr(Ak ∂ ai j . G i j = 2k (1) Proof: First, notice that for any two matrices S = [si j] and B = [bi j], we have that tr (SB) = ∑i, j si jb ji. Consider S and B to be symmetric matrices. Then, we can write S = U + L, where U and L are upper and lower triangular matrices, respectively, with L = U T . Let i < j (the same holds for j < i), hence ∂ ∂ si j tr(SB) = = = (cid:34) ∂ ∂ si j ∂ ∑ ∂ si j ∂ ∂ ui j ∑ i< j i< j [tr(UB) + tr(LB)] ui jb ji +∑ i> j li jb ji ui jb ji +∑ ui jbi j j>i (cid:35) = b ji + bi j = 2bi j. 1The complete graph Kn, is the undirected graph with n nodes in which every pair of distinct vertices is connected by a unique edge. Then, we have that G ) ∂tr(Ak ∂ ai j = = k ∂ si j which completes the proof. ∂tr(SG Ak−1 G ) ∂ si j ∂tr(Ak−1 G SG ) ∂ si j + ∂tr(SG Ak−1 G ) ∂tr(AG SG Ak−2 G ) ∂ si j + ... + (cid:104) Ak−1 G (cid:105) i j , = 2k B. Problem Definition 1 . . . xT n (cid:2)xT Consider a group of n mobile robots and define by xi(t) ∈ Rd the position of robot i at time t ≥ 0. Let x = (cid:3)T ∈ Rdn denote the stacked column vector of x = u =(cid:2)uT all robot positions, so that xir is the r-th coordinate of the i-th robot position. We assume that we can control the position of the robots by the simple kinematic law (2) where ui (t) ∈ Rd is the control input applied to robot i, and uir (t) is the r-th coordinate of ui (t). (cid:3)T ∈ Rdn, For a given set of robot positions, {xi}n i=1, we define the weighted adjacency matrix of the network of robots as A(x) = [ai j(x)] with uT n . . . 1 ai j(x) (cid:44) e−c(cid:107)xi−x j(cid:107)z, (3) where c > 0 is a constant and z ∈ {1,2} denotes the (cid:96)1 or (cid:96)2 norm. Notice that A(x) is a symmetric matrix with position- dependent real eigenvalues, {λi (x)}n i=1. We define, further, the k-th spectral moment of A (x) by (cid:104) (cid:105) Ak (x) , (4) mk (x) (cid:44) 1 n n ∑ i=1 λ k i (x) = 1 n tr for 1 ≤ k ≤ n, where the last equality follows from diag- onalizing Ak (x). For any finite network with n nodes, its eigenvalue spectrum is uniquely defined by the sequence of n moments (m1,m2, ...,mn). Therefore, we can simultane- ously control the whole set of eigenvalues of a network by controlling the first n spectral moments of A (x). This gives rise to the following problem that we aim to address: {m(cid:63) k}n k=1 denote a desired set of spectral moments. Design control laws ui for all robots i = 1, . . . ,n so that the adjacency matrix A(x) of the position-dependent robot network has spectral moments that satisfy mk(x) → m(cid:63) Problem 3 (Control of spectral moments): Let k for all k = 1, . . . ,n. Controlling the eigenvalue spectrum of the adjacency matrix of a network is particularly important, since it is related to the behavior of interesting network dynamical properties [18]. In Section III we propose a gradient descent algorithm to address Problem 3 and discuss its convergence properties. In Section IV we illustrate our approach in numerical simulations. III. CONTROL OF SPECTRAL MOMENTS A. Controller Design Assume a given sequence of desired spectral moments {m(cid:63) k}n k=1 and define the cost function f (x) (cid:44) n ∑ k=1 1 4k (mk (x)− m(cid:63) k)2 , (5) where mk (x) is defined in (4). Define, further, the gradient descent control law u (cid:44) −∇x f (x). (6) i j (x) (cid:44)sgn(xir − x jr). Hence, we can write (10) using a s(r) Hadamard product as follows (11) ∂ ai j (x) ∂ xir = −c [A (x)◦ Sr (x)]i j . Ak−1(cid:105) (cid:104) Substituting (9) and (11) in (8), we have ∑ (cid:104) (A◦ Sr)Ak−1(cid:105) (mk (x)− m(cid:63) k) (mk (x)− m(cid:63) k) = − c n = − c n ∂ f (x) ∂ xir ∑ ∑ k=1 j=1 ji n n n k=1 , ii [A◦ Sr]i j n uir = (cid:18)1 with no self-loops has(cid:0)n ∑ c n k=1 n 2 Then, we can show the following result: Lemma 4: Let the adjacency matrix A = A (x) be defined as in (3) and assume that z = 1 ((cid:96)1 norm). Then, an explicit expression for the entries of u is given by (cid:19)(cid:104) (A◦ Sr)Ak−1(cid:105) , ii trAk − m(cid:63) k where Sr (x) = [si j] with si j (cid:44)sgn(xir − x jr). Proof: The adjacency matrix of an undirected graph (cid:1) independent entries (for example, the upper triangular entries). Each one of these entries are a function of the vector of positions x. Hence, applying the chain rule, we have the following expansion for the partial derivative of the cost function with respect to the entries xir: ∂ f (x) ∂ xir = ∑ 1≤p<q≤n ∂ f ∂ apq ∂ apq ∂ xir . (7) Furthermore, a particular entry apq depends solely on the position xp and xq; therefore, we have that ∂ apq = 0 if both ∂ xir p and q are different than i. Thus, for a fixed i, only the following summands in (7) survive ∂ api ∂ xir ∂ ai j ∂ xir ∂ f ∂ api ∂ f ∂ ai j ∂ f (x) ∂ xir ∂ aiq ∂ xir ∂ f ∂ aiq p=1 n ∑ ∑ ∑ (8) q=i = + = n , i j=1 where we have used that ai j = a ji and aii = 0. We now analyze each one of the partial derivatives in (8). First, from (5) and (4), we have that ∂ f ∂ ai j = = n ∑ k=1 1 n 1 2k (mk (x)− m(cid:63) (mk (x)− m(cid:63) ∑ k) n k) (cid:104) k=1 (cid:17)(cid:19) (cid:16) (cid:18)1 (cid:105) ∂ ∂ ai j n Ak−1 (x) tr Ak (x) , (9) i j where we have used Lemma 2 in the last equality. Second, from (3) we have that ∂ ai j (x) ∂ xir = −ce−c(cid:107)xi−x j(cid:107)1 ∂ ∂ xir = −cai jsgn (xir − x jr) , (cid:12)(cid:12)xir − x jr (cid:12)(cid:12) (the fact for xir (cid:54)= x jr is not differen- tiable at xir = x jr does not affect our analysis). Let us define the antisymmetric matrix Sr (x) = with that s(r) i j (x) (cid:105) (10) (cid:12)(cid:12) (cid:12)(cid:12)xir − x jr (cid:104) where both matrices A and Sr depend on x. Then, from (4), we obtain the statement of our lemma. An efficient relaxation of the spectral control Problem 3 results from controlling a truncated sequence of spectral moments (m1, ...,ms), for s < n. In this case, we can define a cost function s 1 k=1 ∑ k)2 , fs(x) = (12) and an associated control law u = −∇x fs(x). An explicit expression for u can be obtained by following the steps in the proof of Lemma 4, which result in 4k (mk (x)− m(cid:63) (cid:19)(cid:104) (A◦ Sr)Ak−1(cid:105) (cid:18)1 . (13) trAk − m(cid:63) k s u(s) ir = c n ∑ k=1 n ii Although controlling a truncated sequence of moments is not mathematically equivalent to controlling the whole eigenvalue spectrum of the adjacency, we observe a very good overall matching between the eigenvalues obtained from the relaxed problem and the desired eigenvalue spec- trum, especially for the eigenvalues of largest magnitude, which are usually the most relevant in dynamical problems (Section IV). B. Convergence Analysis To simplify convergence analysis of the closed-loop sys- its dynamics to the open set F = k, ∀ 1 ≤ k ≤ s}.2 We can ensure that x ∈ F tem (2), we restrict {x mk (x) > m(cid:63) for all time by adding the barrier potential bs(x) (cid:44) s ∑ k=1 εk 4k (cid:0)mk (x)− m(cid:63) 1 k (cid:1)2 , (14) to (5), for sufficiently small constants ε1, . . . ,εs > 0 (assum- ing that the initial state of the system is already in F ). The convergence properties of the resulting closed loop system x = −∇x( fs(x) +bs(x)) are discussed in the following result. k}s Theorem 5: Let {m(cid:63) k=1 denote a desired set of adjacency spectral moments and assume that x(0) ∈ F . Then, for sufficiently small ε1 . . . ,εs > 0, the closed loop system x = −∇x ( fs(x) + bs(x)) ensures that the moments {mk( x)}s approximate arbitrarily well the desired set {m(cid:63) k}s k=1. (15) k=1, where x = limt→∞ x(t), 2It is shown in Theorem 5 that this construction does not restrict convergence of the state variables to the desired equilibria. Proof: The time derivative of fs(x) + bs(x) is given by d dt ( fs(x) + bs(x)) = ∇x( fs(x) + bs(x)) x = −(cid:107)∇x( fs(x) + bs(x))(cid:107)2 2 ≤ 0 (16) and so the closed loop system is stable and x will converge to a minimum of fs(x) + bs(x). Since bs(x) → ∞ whenever x → ∂ F , where ∂ F = {x mk (x) = m(cid:63) k, ∀ 1 ≤ k ≤ s} = {x fs(x) = 0} denotes the boundary of the set F , equation (16) also implies that the set F is an invariant of motion for the system under consideration. Let P0 = {x xir(0) ≤ x jr(0) ⇒ xir ≤ x jr, ∀ i, j,r} denote the polytope defined by the relative positions of the robots with respect to their initial configuration x(0), so that x(0) ∈ P0 ∩ F (see Fig. 13). In what follows we show that as ε1, . . . ,εs → 0, the state variable x asymptotically reaches a value in the set P0 ∩ ∂ F , where fs(x) = 0. To see this, observe first that lim ε1,...,εs→0 ( fs(x) + bs(x)) = fs(x) for all x ∈ P0 ∩ F , i.e., for small enough ε1, . . . ,εs > 0, the potential fs(x) +bs(x) can be approximated by fs(x) in P0∩ F . Moreover, note that fs(x) is convex in the set P0 ∩ F . Convexity of fs(x) follows from the fact that −cxir − x jr is affine in P0 for any c > 0 and, therefore, −∑cxir − x jr is also affine for any number of terms in the summation. This implies that e−∑cxir−x jr is convex in P0 as a composition of a convex and an affine function and, therefore, trAk(x)− m(cid:63) k = 1 n = 1 n ∑∏e−cxir−x jr − m(cid:63) n ∑e−∑cxir−x jr − m(cid:63) 1 k k k is also convex as a sum of convex functions. Clearly, ntrAk(x)−m(cid:63) k is nonnegative for any x ∈ P0∩ F . Since any 1 power greater than one of a nonnegative and convex function is also convex, every one of the terms(cid:0) 1 (cid:1)2, for ntrAk(x)− m(cid:63) lim ε1,...,εs k = 1, . . . ,s, is convex, which implies that fs(x) is also convex in P0 ∩ F . Taking the limit of (16) we have that dt ( fs(x) + bs(x)) = fs(x) ≤ 0 in P0 ∩ F . Therefore, convexity of fs(x) along with the condition fs(x) ≤ 0 implies that x will converge to a global minimum of fs(x) in P0 ∩ F (recall that F is an invariant of motion for the system under consideration). Since { fs(x)} = fs(x)x∈P0∩∂ F = 0, inf d x∈P0∩(F∪∂ F ) we conclude that the system will converge to a network with spectral moments {mk(x)}s k=1 that are almost equal to the desired {m(cid:63) k}s k=1. The quality of the approximation depends on how small the constants ε1, . . . ,εs > 0 are. What remains is to show that ∂ F ⊂ P0. For this, assume that ∂ F (cid:54)⊂ P0. Then, if the desired set of moments {m(cid:63) k}s 3The polytope P0 essentially defines an ordering of the state variables. k=1 Fig. 1. Plot of the function e−xir−x jr. Note the boundary xir = x jr of two adjacent convex polytopes P. i (x(cid:63)) = m(cid:63) is realizable, there exists another polytope P (cid:63) (cid:54)= P0 such that ∂ F ⊂ P (cid:63). Equivalently, there exists a configuration x(cid:63) ∈ P (cid:63) such that A(x(cid:63)) has eigenvalues {λi(x(cid:63))}n i=1 with ∑n i=1 λ k k for all k = 1, . . . ,s. The result follows from the observation that P0 can be obtained from P (cid:63) by changing the relative positions of the robots in Rd. Mathematically, this means that there exists a permutation matrix Π ∈ Rnd×nd, i.e., an orthogonal matrix with 0 or 1 entries, such that that if P (cid:63) = {x Ex ≥ 0} with E ∈ Rn2d×nd, then P0 = {x E(Πx) ≥ 0}. Therefore, if x(cid:63) ∈ P (cid:63), then ΠT x(cid:63) ∈ P0. Moreover, A(ΠT x(cid:63))⊗ Id = Π(A(x(cid:63))⊗ Id)ΠT , i.e., a permutation of the robots' coordinates results in a permutation of the entries of (A(x)⊗ Id), where ⊗ indicates the Kronecker product between matrices. Since Π is orthog- onal, A(ΠT x(cid:63))⊗Id and A(x(cid:63))⊗Id have the same eigenvalues. Therefore, there exists a configuration x = ΠT x(cid:63) ∈ P0 such that A( x) has eigenvalues {λi( x)}n i=1 λ k i ( x) = m(cid:63) k for all k = 1, . . . ,s.4 This implies that ∂ F ⊂ P0. In the above discussion, we have used the fact that if i=1 are the eigenvalues of Y ∈ Rn×n and {zi}m {yi}n i=1 are the eigenvalues of Z ∈ Rm×m, then the eigenvalues of Y ⊗ Z are {y1z1, . . . ,y1zm, . . . ,ynz1, . . . ,ynzm}. This implies that the eigenvalues of A(x) ⊗ Id are essentially the eigenvalues of A(x), each one with multiplicity d. i=1 with ∑n Remark 6 (Barrier functions): The barrier functions bs(x) are necessary in the proof of Theorem 5. If not there, the quantities mk(x)− m(cid:63) k are not guaranteed to be positive and, fs(x) is not necessarily convex in therefore, the potential P0. This causes technical difficulties in ensuring a global minimum of fs(x). Also, the analysis in Theorem 5 assumes that ε1, . . . ,εs → 0. This essentially restricts the influence of the barrier potential bs(x) to a small neighborhood of the set ∂ F , so that the potential fs(x) remains unaffected out- 4Equivalently, this means that the two graphs are isomorphic. −1−0.500.51−1−0.500.5100.10.20.30.40.50.60.70.80.91Fig. 2. Hexagonal formation of mobile robots. side this neighborhood. In practice, the closer the constants ε1, . . . ,εs are to zero, the better the approximation of the desired spectral moments. This approximation can be made arbitrarily good. Remark 7 (Distance metric): In the preceding analysis, we have employed the (cid:96)1 norm (z = 1) as a distance metric to define the entries of the adjacency matrix (see (3)). This choice is mainly due to technical reasons, since it ensures that trAk(x) and, therefore, fs(x), are convex functions. From a practical point of view, the (cid:96)1 metric has potential appli- cations in indoor navigation where the presence of obstacles forces the signals to propagate in grid-like environments. Nevertheless, our numerical simulations indicate that the control law herein proposed also performs well for z = 2 ((cid:96)2 norm). We leave a rigorous proof of this case for future work. IV. NUMERICAL SIMULATIONS In this section we present examples of spectral design of mobile robot networks and discuss performance of our proposed approach. Example 8 (Hexagonal Formation): Consider the prob- lem of controlling the structure of a network of mobile robots to match the eigenvalue spectrum of the hexagonal network on 7 nodes shown in Fig. 2. The target eigenvalues of the weighted adjacency matrix of this formation are i }7 {λ (cid:63) and the corresponding spectral moments are i=1 ={−0.51, −0.47, −0.40, −0.40, 0.05, 0.05, 1.70} , k}7 {m(cid:63) k=1 = {0, 0.53, 0.64, 1.22, 2.02, 3.47, 5.90} . The initial configuration of the mobile robot network is that of a random geometric graph on 7 nodes uniformly distributed in the square [0,1]× [0,1]. We applied the pro- posed control law (15) and studied the evolution of the spectral moments of the network's adjacency matrix. Fig. 3 shows the evolution of the second, third and fourth moment.5 5A similar behavior is observed for the higher order moments. Note that the first spectral moment is always equal to zero for simple graphs. Fig. 3. Evolution of second (blue), third (green) and fourth (red) spectral moments towards the moments of the hexagonal formation. As expected, they all converge to the desired values. The asymptotic values of all spectral moments of the network are {mk}7 k=1 = {0, 0.53, 0.65, 1.23, 2.04, 3.50, 5.95} , (cid:8)m(cid:63) (cid:9)7 k which are very close to the desired sequence of moments k=1, and slightly on the larger side, as predicted by Theorem 5. The eigenvalues of the weighted adjacency matrix of the final configuration are i=1 ={−0.52, −0.48, −0.42, −0.40, 0.02, 0.10, 1.70} , {λi}7 which are also very close to the desired values. Notice that our approach fits better those eigenvalues further away from the origin, since they are more heavily weighted in the expression of spectral moments, than those close to zero. An alternative approach to overcome this limitation would be to modify our cost function (5) to assign more weight to eigenvalues of small magnitude. {mk}4 As discussed in Section III-A, we also consider a relax- ation of Problem 3, that involves a truncated sequence of the first four moments of the network. In this case, the cost function is given by (12) and the closed loop system is defined in (15), for s = 4. The asymptotic values of the first four spectral moments of the network are k=1 = {0, 0.55, 0.65, 1.26} and eigenvalues of the final configuration are {λi}7 i=1 ={−0.55, −0.50, −0.42, −0.31, −0.09, 0.18, 1.71} , which are remarkably close to the set of desired eigenvalues, especially those far from zero. The main advantages of using a relaxation of Problem 3 are (i) that it reduces the computational cost of the controller, since it only requires the terms trAk and Ak−1, and (ii) that it improves the stability of the numerical behavior of the gradient descent algorithm. Example 9 (Random Geometric Network): In this exam- ple we consider the problem of matching the eigenvalue spectrum of a particular realization of a random geometric Fig. 4. Random geometric graph with 10 nodes. The width of each edge is proportional to its weight. Fig. 5. Evolution of second (blue), third (green) and fourth (red) spectral moments towards the moments of the random geometric graph. graph on 10 nodes that are uniformly distributed in [0,1]× [0,1]. This target realization is illustrated in Fig. 4, where the thickness of each edge is proportional to its weight. The eigenvalues of the weighted adjacency matrix of this network are {λ (cid:63) i=1 = {−0.89, −0.85, −0.84, −0.79, −0.77, −0.68, i }10 −0.61, 0.02, 0.27, 5.16} , and the corresponding sequence of spectral moments is k}k>0 = {0, 3.11, 13.45, 71.60, 368.36, 1905, . . .} . {m(cid:63) We applied the control law (15) using the cost function (5) and starting with a different realization of the random geometric graph on 10 nodes in the square [0,1] × [0,1]. In this case, matching the whole set of spectral moments directly provides us with a very good approximation of the eigenvalues spectrum. As before, we also considered the relaxation of Problem 3 involving only a truncated sequence consisting of the first 4 moments. The evolution of the second, third and fourth moments of the network is shown in Fig. 5. The asymptotic values of the spectral moments of the network are {mk}k>0 = {0, 3.13, 13.46, 71.63, . . .} , which are remarkably close to the desired ones and, as in the previous example, slightly on the larger side. Similarly, the final set of eigenvalues is {λ (cid:63) i=1 = {−0.94, −0.87, −0.85, −0.83, −0.74, −0.59, i }10 −0.55, −0.27, 0.50, 5.16} , which is a very good fit to the given eigenvalues, especially for those far from zero. V. CONCLUSIONS AND FUTURE RESEARCH In this paper, we proposed a novel control framework to modify the structure of a mobile robot network in order to control the eigenvalue spectrum of its adjacency matrix. We introduced a novel abstraction of the eigenvalue spectrum by means of the spectral moments and derived explicit gradient descent motion controllers for the robots to obtain a network with the desired set of moments. Since the eigenvalue spectrum is uniquely determined by the associated spectral moments, our approach provides a way of controlling the eigenvalues of mobile networks. Convergence to the desired moments was always guaranteed due to convexity of the proposed cost functions. Efficiency of our approach was illustrated in nontrivial computer simulations. The adjacency matrix eigenvalue spectrum is relevant to the performance of many distributed coordination algorithms run over a network. Therefore, our approach is particularly useful in providing network structures that are optimal with respect to networked coordination objectives. Future work involves theoretical guarantees for the Eu- clidean distance metric (z = 2 in (3)), as well as extension of our results to the spectrum of Laplacian matrix of the network. Moreover, the relaxation of Problem 3 to a trun- cated sequence of moments does not guarantee (mathemati- cally) a good fit of the complete distribution of eigenvalues. Therefore, a natural question is to characterize the set of graphs most of whose spectral information is contained in a relatively small set of low-order moments. REFERENCES [1] K. M. Lynch and M. T. Mason, "Stable pushing: Mechanics, controlla- bility and planning," International Journal of Robotics Research, vol. 15, pp. 533-556, 1996. [2] M. J. Mataric, M. Nilsson, and K. Simsarian, "Cooperative multi- robot box-pushing," IEEE/RSJ International Conference on Intelligent Robots and Systems, pp. 556-561, 1995. [3] Z. Wang and V. Kumar, "Object closure and manipulation by multiple cooperating mobile robots," International Symposium on Distributed Autonomous Robotic Systems, pp. 394-399, 2002. [4] J. Cortes, S. Martinez, T. Karatas, and F. Bullo, "Coverage control for mobile sensing networks," IEEE Transactions on Robotics and Automation, vol. 20, pp. 243-255, 2004. [5] L. E. Parker, "Distributed algorithms for multi-robot observation of multiple moving targets," Autonomous Robots, vol. 12, pp. 231-255, 2002. [6] S. Poduri and G. S. Sukhatme, "Constrained coverage for mobile sensor networks," IEEE International Conference on Robotics and Automation, pp. 165-172, 2004. [7] A. Jadbabaie, J. Lin, and A. S. Morse, "Coordination of groups of mobile autonomous robots using nearest neighbor rules," IEEE Transactions on Automatic Control, vol. 48, pp. 988-1001, 2003. [8] R. Olfati-Saber, J. A. Fax, and R. M. Murray, "Consensus and cooperation in networked multi-robot systems," Proceedings of the IEEE, vol. 95, pp. 215-33, 2007. [9] W. Ren and R. Beard, "Consensus seeking in multi-robot systems un- der dynamically changing interaction topologies," IEEE Transactions on Automatic Control, vol. 50, pp. 655-661, 2005. [10] J. P. Desai, J. P. Ostrowski, and V. Kumar, "Modeling and control of formations of nonholonomic mobile robots," IEEE Transactions on Robotics and Automation, vol. 17, pp. 905-908, 2001. [11] J. Lawton, B. Young, and R. Beard, "A decentralized approach to elementary formation maneuvers," IEEE International Conference on Robotics and Automation, pp. 2728-2733, 2005. [12] M. Mesbahi and F. Y. Hadaegh, "Formation flying control of mul- tiple spacecraft via graphs, matrix inequalities and switching," AIAA Journal of Guidance, Control and Dynamics, vol. 24, pp. 369-377, 2001. [13] R. Olfati-Saber, "Flocking for multi-robot dynamic systems: Algo- rithms and theory," IEEE Transactions on Automatic Control, vol. 51, pp. 401-420, 2006. [14] H. G. Tanner, A. Jadbabaie, and G. J. Pappas, "Flocking in fixed and switching networks," IEEE Transactions on Automatic Control, vol. 52, pp. 863-868, 2007. [15] S. L. Smith and F. Bullo, "Target assignment for robotic networks: Asymptotic performance under limited communication," American Control Conference, pp. 1155-1160, 2007. [16] M. M. Zavlanos and G. J. Pappas, "Dynamic assignment in distributed motion planning with local coordination," IEEE Transactions on Robotics, vol. 24, pp. 232-242, 2008. [17] S. Kloder and S. Hutchinson, "Path planning for permutation-invariant multi-robot formations," IEEE Transactions on Robotics, vol. 22, pp. 650-665, 2006. [18] V.M. Preciado, Spectral Analysis for Stochastic Models of Large-Scale Complex Dynamical Networks, Ph.D. dissertation, Dept. Elect. Eng. Comput. Sci., MIT, Cambridge, MA, 2008. [19] D.J. Aldous, "Some Inequalities for Reversible Markov Chains," J. London Math. Soc., vol. 25, pp. 564-576, 1982. [20] L.M. Pecora and T.L. Carroll, "Master Stability Functions for Synchro- nized Coupled Systems," Physics Review Letters, vol. 80, pp. 2109- 2112, 1998. [21] V.M. Preciado, and G.C. Verghese, "Synchronization in Generalized Erdos-R´enyi Networks of Nonlinear Oscillators," IEEE Conference on Decision and Control, pp. 4628-4633, 2005. [22] M. Draief, A. Ganesh, and L. Massouli´e, "Thresholds for Virus Spread on Networks," Annals of Applied Probability, vol. 18, pp. 359-378, 2008. [23] V.M. Preciado, and A. Jadbabaie, "Spectral Analysis of Virus Spread- ing in Random Geometric Networks," IEEE Conference on Decision and Control , 2009. [24] N.A. Lynch, Distributed Algorithms, Morgan Kaufmann Publishers, 1997. [25] A. Fax, and R. M. Murray, "Information Flow and Cooperative Control of Vehicle Formations," IEEE Transactions on Automatic Control, vol. 49, pp. 1465-1476, 2004. [26] R. Olfati-Saber, and R. M. Murray, "Consensus Problems in Networks of robots with Switching Topology and Time-Delays," IEEE Transac- tions on Automatic Control, vol. 49, pp. 1520-1533, 2004. [27] A. Ghosh, and S. Boyd, "Growing Well-Connected Graphs," Proc. of the 45th IEEE Conference on Decision and Control, pp. 6605-6611, 2006. [28] R. Grone, R. Merris, and V.S. Sunder, "The Laplacian Spectrum of a Graph," SIAM Journal Matrix Analysis and Applications, vol. 11, pp. 218-238, 1990. [29] Y. Kim, and M. Mesbahi, "On Maximizing the Second-Smallest Eigenvalue of a State Dependent Graph Laplacian," IEEE Transactions on Automatic Control, vol. 51, pp. 116-120, 2006. [30] M.C. DeGennaro, and A. Jadbabaie, "Decentralized Control of Con- nectivity for Multi-robot Systems," IEEE Conference on Decision and Control, pp. 3628-3633, 2006. [31] V. M. Preciado, M. M. Zavlanos, A. Jadbabaie and G. J. Pappas, "Distributed control of the Laplacian spectral moments of a network," American Control Conference, pp. 4462-4467, 2010.
1505.04142	1	1505	2015-05-15T18:03:53	An informational study of the evolution of codes and of emerging concepts in populations of agents	[ "cs.MA" ]	We consider the problem of the evolution of a code within a structured population of agents. The agents try to maximise their information about their environment by acquiring information from the outputs of other agents in the population. A naive use of information-theoretic methods would assume that every agent knows how to "interpret" the information offered by other agents. However, this assumes that one "knows" which other agents one observes, and thus which code they use. In our model, however, we wish to preclude that: it is not clear which other agents an agent is observing, and the resulting usable information is therefore influenced by the universality of the code used and by which agents an agent is "listening" to. We further investigate whether an agent who does not directly perceive the environment can distinguish states by observing other agents' outputs. For this purpose, we consider a population of different types of agents "talking" about different concepts, and try to extract new ones by considering their outputs only.	cs.MA	cs	An informational study of the evolution of codes and of emerging concepts in populations of agents Andr´es C. Burgos1 Daniel Polani2 1,2Adaptive Systems Research Group, University of Hertfordshire, Hatfield, UK 1Email address: [email protected], Tel: +44 1707 28 4490 Abstract We consider the problem of the evolution of a code within a structured population of agents. The agents try to maximise their information about their environment by acquiring information from the outputs of other agents in the population. A naive use of information- theoretic methods would assume that every agent knows how to "interpret" the information offered by other agents. However, this assumes that one "knows" which other agents one observes, and thus which code they use. In our model, however, we wish to preclude that: it is not clear which other agents an agent is observing, and the resulting usable information is therefore influenced by the universality of the code used and by which agents an agent is "listening" to. We further investigate whether an agent who does not directly perceive the environment can distinguish states by observing other agents' outputs. For this purpose, we consider a population of different types of agents "talking" about different concepts, and try to extract new ones by considering their outputs only. Keywords: information theory, code evolution, semantics, emerging concepts 1 1 Introduction If we consider organisms capable of processing information, then we can argue that they must be able to internally assign meaning to the symbols they perceive in a code-based manner [10]. For instance, bacteria perceives chemical molecules in their environment and interprets them in order to better estimate environmental conditions and (stochastically) decide their phenotype [24, 1, 23, 27]. Plants detect airborne signals released by other plants, being able to interpret them as attacks of pathogens or herbivores [13, 29]. Therefore, a correspondence between environmental conditions and chemical molecules must be established. It is in this way that Barbieri characterises codes, and he proposes three fundamental characteristics for them: they connect two independent worlds; they add meaning to information; and they are community rules [2]. Codes connect two independent worlds by establishing a correspondence or mapping between them. These worlds are independent and thus there are no material constraints for establishing arbitrary mappings. The meaning of information comes exclusively from the mapping: symbols by themselves are meaningless. Finally, the third property requires that the correspondence between the two worlds constitutes an integrated system. For instance, human languages establish a correspondence between words and objects [2]; in bacteria it is between chemical molecules and environmental and social conditions [35, 36]. Words (or chemical molecules) by themselves do not have any meaning, and each individual of a population can define, arbitrarily to some extent, their own set with its mapping. However, populations of individuals sharing the same code are ubiquitous in nature. How is it that codes come to be shared by many individuals when their constitution involve arbitrary choices for each individual? This question is what we are investigating in the present paper. For this work, we assume a simple scenario where organisms live in a fluctuating environment. If they can perfectly predict the future environmental conditions, they can prepare themselves by adopting a proper phenotype, and, therefore, survive. However, when uncertainty about the environment remains, organisms will follow a bet-hedging strategy [31, 28], where they try to maximise their long-term growth rate by adopting the phenotype that matches the environment in proportions based on the information they have about it. For example, seeds of annual plants germinate stochastically in different periods in relation to the probability of rainfalls, and their chances of survival are maximised when they match this probability [6]. The relation between information and long-term growth rate can be expressed elegantly in information theoretic terms, where an increase in the environmental information of an organism 2 is translated into an increase in its long-term growth rate [30, 17, 18, 8, 26]. Such models achieve the maximisation of the long-term growth rate by maximising an organism's information about the environment. If we assume this behaviour in organisms, then those obtaining additional environmental information (other than that from their sensors, which we assume it does not completely eliminate environmental uncertainty) from other individuals will have an advantage over those that do not, since they would be able to better predict the future conditions. However, for individuals to be able to communicate with each other, they must be able to translate symbols into environmental conditions, where the output of these symbols results from an individual's code. We consider the code of an individual as a stochastic mapping from its sensors states to a set of outputs. For this study, we consider outputs (or messages) of individuals (or agents) as conventional signs. In semiotics, the science of all processes in which signs are originated, stored, communicated, and being effective [10], two types of signs are traditionally recognised: conventional signs and natural signs [7]. In conventional signs there is no physical constraint on the possible mappings, they are established by conventions. Although in physical systems there can be limitations to the possible mappings that can be implemented, in this work we assume complete freedom of choice. On the other hand, in natural signs, there is always a physical link between the signifier and signified, such as smoke as a sign of fire, odours as signs of food, etc. [3]. In this work, we are not interested in the particular detailed mechanisms by which an agent implements its code, nor how the agent decodes the outputs of other agents. Instead, we focus on the theoretical limits on the amount of environmental information an agent can possibly acquire resulting from different scenarios of population structure and codes distribution. The natural framework to analyse such quantities is information theory [30]. However, it does not take semantic aspects into account, it only deals with frequencies of symbols instead of what they symbolise. Codes, on the other hand, add meaning to information, which makes the integration of sciences such as semiotics with information theory non-trivial [9, 4]. In the following section, we present an information-theoretic model which incorporates the necessity of conventions by dropping from the model the usual implicit assumption of knowing the identity of the communicating units. 2 Model To introduce the model in a progressive manner, let us first consider three agents, θ1, θ2 and θ3. Each of these agents depend on the same environmental conditions for survival, which are 3 modelled by a random variable µ. Agents acquire information about the environment through their sensors, which are modelled by random variables Yθ1, Yθ2 and Yθ3 , all three conditioned on µ, for agents θ1, θ2 and θ3, respectively. We assume each agent acquires the same amount and aspects of environmental information from µ, i.e. p(Yθ1µ) = p(Yθ2µ) = p(Yθ3µ). Let us further assume that the information each agent acquires about the environment does not eliminate its uncertainty, i.e. H(µYθi) > 0 for 1 ≤ i ≤ 3. The code of an agent is a stochastic mapping from its sensor states into a set of outputs, and is represented by the conditional probabilities p(Xθ1Yθ1), p(Xθ2Yθ2) and p(Xθ3Yθ3) for agents θ1, θ2 and θ3, respectively (see Fig. 1). µ Yθ1 Yθ2 Yθ3 Xθ1 Xθ2 Xθ3 Figure 1: Bayesian network representing the relantionship between the sensor and output variables of three agents. Let us assume that agent θ1 perceives only the outputs of agents θ2 and θ3. One possible way of computing the information about the environment agent θ1 has is to consider the mutual information between µ and the joint distribution of the sensor of θ1 and the outputs of θ2 and θ3: I(µ; Yθ1, Xθ2, Xθ3). However, by writing down this quantity, we are implicitly assuming that agent θ1 "knows" which output corresponds to θ2 and which output corresponds to θ3. Therefore, in this consideration, an agent can theoretically do the translations of the outputs according to some internal model of other agents and infer the mentioned amount of information about its environment. 2.1 Indistinguishable sources of messages For this study, on the contrary, we consider an agent observing other agents' messages, but under the assumption that the originator of a message cannot be identified. In this way, the total amount of information an agent can infer from the outputs of other agents will depend on to which extent it either can identify who the other agents are or can rely on them using a coding scheme that does not depend too much on their particular identity. For instance, if agents θ2 and θ3 both agree on the output for each of the environmental conditions, then agent θ1 should be able to infer 4 more environmental information than if they disagree on the output for each of the environmental conditions, given that agent θ1 does not know which of the agents it is observing. To model this idea, let us assume a random variable Θ(cid:48) denoting the selected agent. This agent depends on the same environmental conditions for survival as θ1, which are modelled, as above, by a random variable µ. Agents acquire information about the environment through their sensors, which are modelled by a random variable YΘ(cid:48) conditioned on the index variable denoting the agent under consideration, Θ(cid:48), and µ. The amount of acquired sensory information of a specific agent θ(cid:48) about µ is given by I(µ; Yθ(cid:48)). As above, the code of an agent is a stochastic mapping from its sensor states into a set of messages, and is represented by the conditional probability p(Xθ(cid:48)Yθ(cid:48)) for an agent θ(cid:48) (see Fig. 2). µ Yθ1 Xθ1 YΘ(cid:48) XΘ(cid:48) Θ(cid:48) Figure 2: Bayesian network representing the relationships as described above (see text). However, now we want to model the fact that we do not know which agent is observed. In the case with maximum uncertainty, Θ is uniformly distributed, and then this parametrisation of the codes considers the outputs of all agents in Θ(cid:48) altogether, such that if we are not observing Θ(cid:48), we cannot identify whose agent's output we are observing. In Eq. 3 and Eq. 4 we show two examples of codes for agents θ2 and θ3, while their sensor states are define by the Eq. 2 (Eq. 1 defines the sensors states of agent θ1). We compute how much information about the environment there is when the outputs of both agents (θ2 and θ3) are considered together by agent θ1. (cid:18) y1 1 − (cid:19) y2 1 − P r(Yθ1µ) := µ1 µ2 (1) P r(YΘ(cid:48)µ, Θ(cid:48)) :=  y1 1 − 1 −  (2) y2 1 − 1 − θ2, µ1 θ2, µ2 θ3, µ1 θ3, µ2 If we assume p(θ2) = p(θ3) = 1/2, and p(µ1) = p(µ2) = 1/2 and = 0.01, then if we consider the codes shown in Eq. 3, we have that I(µ; Yθ1 , XΘ(cid:48)) = 0.97872 bits, where Θ(cid:48) consists of agents θ2 and θ3. However, had θ2 and θ3 "opposite" codes as shown in Eq. 4, then I(µ; Yθ1 , XΘ(cid:48)) = 0.9192 5 P r(XΘ(cid:48)YΘ(cid:48), Θ(cid:48)) :=  x1 x2 0 1 1 0 0 1 0 1  (3) θ2, y1 θ2, y2 θ3, y1 θ3, y2 P r(XΘ(cid:48)YΘ(cid:48), Θ(cid:48)) :=  x1 x2 0 1 1 0 1 0 1 0  (4) θ2, y1 θ2, y2 θ3, y1 θ3, y2 bits, which is exactly I(µ; Yθ1), that is, I(µ; XΘ(cid:48)Yθ1) = 0 bits (agent θ1 cannot acquire any side information from the outputs of agents θ2 and θ3). We should note here that the sensor states y1 and y2 of agents θ2 and θ3 in the conditional probability shown in Eq. 1 and 2 refer almost deterministically to the same environmental condition, and therefore the loss of side information is thus entirely due to the incompatible codes. The conditional probabilities of sensor states given the environmental conditions further defined throughout the paper are also assumed to be almost deterministically. 2.2 Environmental information of a population The model shown in Fig. 2 considers the environmental information of agent θ1, ignoring its own output Xθ1. Nevertheless, agents ignoring their outputs is contrary to our assumption over the incapability of agents to identify the sources of the outputs. On the other hand, we are assuming a specific type of communication, one which could be classified as persistent within the different classifications of stigmergy ([37, 33, 22], see [14] for a summary). To incorporate this option in the model shown in Fig. 2, we could consider the state space of Θ(cid:48) as the set {θ1, θ2, θ3}. Then, to express not only the environmental information of agent θ1, but the average environmental information of the whole population, we can parametrise the agent by a random variable Θ (defined over the same state space, representing the same set of agents as Θ(cid:48)), such that p(YΘµ, Θ) = p(YΘ(cid:48)µ, Θ(cid:48)) (i.e., YΘ(cid:48) is i.i.d. to YΘ, and vice versa). µ YΘ YΘ(cid:48) Θ XΘ(cid:48) Θ(cid:48) Figure 3: Bayesian network representing the sensor variables of a set of agents indexed by the random variable Θ, and the sensor and output variables of a copy of the set of agents indexed by Θ named Θ(cid:48). 6 In this way, the average environmental information of a population of the agents selected by Θ is given by I(µ; YΘ, XΘ(cid:48)) (see Fig. 3). This measure can be consider as the objective function to maximise in our model. However, we would be making two important assumptions: first, this objective function assumes agents have access to the environmental conditions µ, which they indirectly do but only through their sensors; and second, every agent would perceive the output of every other agent, including itself. In this work, instead, we propose that agents follow a behaviour such that it maximises the similarity of their outputs (via their codes) with those of which the agent perceives. A consequence of this behaviour is that the average information about µ is also maximised. In addition, we will introduce a potentially flexible "population structure", so that we can specify which agents interact with which. 2.3 Code similarity First, we introduce a copy of the codes of the agents, such that when we instantiate the variables XΘ and XΘ(cid:48), the probabilities are the same. The structure of the population is then given by p(Θ, Θ(cid:48)) = p(Θ)p(Θ(cid:48)). However, the conditional independence of Θ and Θ(cid:48) restricts significantly the diversity of the structures that can be represented. Θ perceive the outputs of all the agents selected by Θ(cid:48) and vice versa. In order to model a general interaction structure between agents, we consider p(Θ, Θ(cid:48)) not independent, as shown in In such cases, the agents selected by the Bayesian network in Fig. 4, where we introduce a helper variable Ξ. This allows different agents selected by Θ to perceive outputs from exclusive agents selected by Θ(cid:48). µ YΘ YΘ(cid:48) Θ XΘ XΘ(cid:48) Θ(cid:48) Ξ Figure 4: Bayesian network representing the relantionship of the variables in the model of code evolution. YΘ(cid:48) is an i.i.d copy of YΘ and XΘ(cid:48) is an i.i.d. copy of XΘ. Θ(cid:48) covers the same set of agents as Θ, but its probability distribution is not necessary the same. 7 We define the objective function as I(XΘ; XΘ(cid:48)), that is the average code similarity of a pop- ulation of agents according to the population structure p(Θ, Θ(cid:48)). For instance, if the interaction probability of two agents is zero, then the similarity of the codes of these two agents is irrelevant for the objective function. On the other hand, they interact with probability bigger than zero (p(θ, θ(cid:48)) > 0, for some agents θ and θ(cid:48)), then how similar their codes are will influence I(XΘ; XΘ(cid:48)). If we consider our system as a process in time, then at each time-step two agents are chosen according to p(Θ, Θ(cid:48)). Agent Θ reads the output of agent Θ(cid:48) (generated via its code, which is i.i.d over time), and let us assume that it stores the pair (YΘ, XΘ(cid:48)), i.e. its current sensor state together with the perceived output. If this is repeated a large number of times, then the total amount of environmental information that can be inferred from the collected statistics by the population is bounded by I(µ; YΘ, XΘ(cid:48)). This is the theoretical limit to which we refer in the introduction, and for this study we are not interested in how the inference is computed. However, we implicitly assume that agents decode the perceived outputs according to their codes. 2.4 Distance between two codes In order to visualise the evolution of codes, we define the distance between the codes of two agents θi and θj as the square root of the Jensen-Shannon divergence [40, 19] between them. This measure has the property that 0 ≤ JSD(θi, θj) ≤ 1 when log2 is used, and the square root yields a metric. Let us note that this distance requires the sensor states Y to be named identically (for the corresponding states of µ) among agents in order to be meaningful. As we stated above, this is (closely) the case in all our experiments. This requirement over the sensor states discards the possibility of using other measures such as mutual information. dist(θi, θj) = = (cid:113) JSD(cid:0)p(XθiYθi)p(XθjYθj )(cid:1) (cid:114) 1 D(cid:0)p(XθiYθi)p(XθkYθk )(cid:1) + (cid:0)p(XθiYθi) + p(XθjYθj )(cid:1). 2 where p(XθkYθk ) = 1 2 D(cid:0)p(XθjYθj )p(XθkYθk )(cid:1) 1 2 (5) 8 3 Methods To illustrate the behaviour of our model, we consider four different scenarios, which are described in Sec. 4. The common parameters for the first two experiments are the following: the population consists of 25 agents; the amount and quality of the acquired sensory information is the same for every agent, that is p(YΘµ) = p(YΘ(cid:48)µ). For the third scenario, the only difference is that we con- sider only 15 agents, since the dimensions to consider with a flexible structure grows quadratically with the number of agents. The optimisation algorithm used in the following experiments is CMA-ES (Covariance Matrix Adaptation Evolution Strategy), which is a stochastic derivative-free method for non-linear op- timisation problems [12]. We utilised the implementation provided by the Shark library v3.0.0 [15] with its default parameters, which implements the CMA-ES algorithm described in [11]. The evolutionary algorithm used for optimisation does not intend to represent the actual evolution of the codes. Instead, we are interested in the solutions of this optimisation process, which are representative of the possible outcomes of evolution. To visualise the evolution of the codes of the agents, we use the method of multidimensional scaling provided by R version 2.14.1 (2011-12-22). This method takes as input the distance matrix between codes, and plots them in a two-dimensional space preserving the distances as well as possible. To visualise, not only the distances between the resulting codes, but also how they relate to the distances between initial codes, we provide a distance matrix of both initial and resulting codes. The initial codes are randomly set by the evolutionary algorithm. 4 Results In this section, we analyse the outcome of the four different scenarios where code similarity is maximised. While the outcomes are particular for one simulation, they are illustrative of the richness that the model is able to capture, which is described for each scenario. The outcomes are typical solutions, and we cannot perform statistics over simulations since the many solutions are qualitatively different. However, the outcome of each scenario is presented together with a description of alternative outcomes, giving indicators of achievement of local/global optimum. 9 4.1 Well-mixed population In the first scenario, each agent θi perceives the output of every other possible agent θj with the same probability, that is p(θi, θj) = 1/252 for every i, j ∈ [1, 25]. The maximum average code similarity is bounded by I(YΘ; YΘ(cid:48)) = 1.71908 bits, which is achieved under two conditions: first, every code must be a one-to-one mapping; second, the code must be universal. This is indeed the outcome of the performed optimisation, as we show in Fig. 5: the optimised codes (blue points) converged into a universal code (the distance between any of them is zero). Each red (diamond) point correspond to an initial code. Figure 5: 2-dimensional plot of code distance: red points are codes at the beginning of the optimisation process; blue points are codes at the end of the optimisation process (where the distance between every pair of codes is zero). The resulting code adopted by the population is a one-to-one mapping between sensor states and outputs, and any of the 24 possible one-to-one mappings is a global maximum (there are 4 sensor states and 4 possible outputs). However, it is still interesting to briefly analyse the possible paths towards a universal and optimal code. In Fig. 6, we show the distribution of the adopted codes by the agents of the population in an iteration of the optimisation process where the average code similarity is I(XΘ; XΘ(cid:48)) = 1.18276 bits. Here, the most popular code is the suboptimal code shown in Fig. 6 (a). This results from the particular initialised codes, driving the agents temporarily towards a suboptimal code. However, once any of the many-to-one codes becomes (nearly) universally adopted, then any code's deviation improving the code similarity will eventually drive the convention towards optimality. The fact that it does not need simultaneous changes in the code increases the likeliness of improving the code similarity. 10 −0.6−0.4−0.20.00.20.40.6−0.4−0.20.00.20.4dimension 1dimension 2llllllllllllllllllllllllllinitial codesfinal codesFigure 6: Representation of the codes p(xy) by a heat-map using inverse grayscale. For each evolved code, we output the number of agents adopting it. This code distribution was achieved with 25 agents in a well-mixed population. 4.2 Spatially-structured population In another set-up, we assume the agents are structured in a 5 × 5 grid, where p(θ, θ(cid:48)) = 1/105 if θ and θ(cid:48) are neighbours or when θ = θ(cid:48) (see Fig. 8 for a representation of the structure). After randomly initialising the codes, the performed optimisation plateaued on an average code similarity of I(XΘ; XΘ(cid:48)) = 1.13536 bits. As in the former scenario, here the optimal solution is also a universal code with a one-to-one mapping. However, in this case, the result is not a universal code, as can be appreciated in Fig. 7. Spatially structured populations are sensitive to the initial codes and how codes are updated. a a c e e a a c e e b b d f f b b b g h b b b b d Figure 7: 2-dimensional plot of code distance: Figure 8: Representation of the spatial structure points in diamond shape represent codes at the be- utilised for the experiment. Agents are assumed to ginning of the optimisation process; rounded points be distributed in a grid: an edge from one agent to represent codes at the end of the optimisation pro- another means that one agent perceives the output cess. The points are coloured in order to be able to of the other. Agents are labelled (see Fig. 9) and relate this plot with the figure beside it. coloured according to their adopted code. The resulting code distribution among the population is shown in Fig. 9, with 8 different codes 11 y1y2y3y4x1x2x3x4x1x2x3x4x1x2x3x4x1x2x3x4x1x2x3x4(a)20(b)1(c)2(d)1(e)1−0.6−0.4−0.20.00.20.40.6−0.4−0.20.00.20.4dimension 1dimension 2llllllllllllllllllllllllllinitial codesfinal codesin the population. Where well-mixed populations evolved the use of common codes, agreement on codes only occurred among neighbours in spatially structured populations. As a consequence, many local conventions are established within neighbourhoods, and, once this situation is reached, the improvement of the total code similarity requires simultaneous changes to the agent's codes. For instance, the code shown in Fig. 9 (e) could increase the average similarity of the population if p(x2y1) = 1, as it is in the rest of the codes. However, for this to happen (in this particular case), at least two agents need to change their code simultaneously (otherwise the average similarity decreases), which makes the deviation from the resulting code distribution unlikely. Figure 9: Representation of the codes p(xy) by a heat-map using inverse grey scale. For each evolved code, we output the number of agents adopting it. This code distribution was achieved with 25 agents in a grid structure. 4.3 Flexible population structure For the third scenario, we let the structure co-evolve with the codes without any constraint (the probability distribution of the interaction between agents, p(Ξ), is optimised together with the codes). In this case, the resulting average code similarity is nearly optimal, but the code is not necessarily universal. This is because, when the structure is not fixed, agents form roughly disconnected clusters of related codes. In this process, the interaction probability of agents with unrelated codes will vanish. However, once the clusters are formed, if it is not a single isolated agent (such that no other agent perceives its output), then codes of agents are universal within each cluster. This is exemplified by the code distribution and population structure we obtained (see Fig. 10). Here, we have two clusters with universal codes, one optimal (in red) and the other suboptimal (in yellow). Agents with dissimilar codes from every other agent they interact with will become isolated in the optimisation process, as the example shows for two agents (light and dark blue). To summarise, the optimal code similarity equals I(YΘ; YΘ(cid:48)), and is achieved, for instance, when all agents adopt the same one-to-one mapping. Nevertheless, the interaction probability allows agents to form disconnected clusters of related codes, where several one-to-one mappings could 12 y1y2y3y4x1x2x3x4x1x2x3x4x1x2x3x4x1x2x3x4x1x2x3x4x1x2x3x4x1x2x3x4x1x2x3x4(a)4(b)9(c)2(d)2(e)4(f)2(g)1(h)1result while still achieving optimality. Theoretically, we could have as many one-to-one mappings as the minimum between the amount of agents and the total one-to-one mapping combinations (24 in this case). Figure 10: Each node in the graph corresponds to the code of an agent. There is a weighted edge between agent θi and θj if p(θi, θj) > 0 (which is the weight). We omit weights of edges in the graph since they all are roughly of similar value. The temperature colours on top of the nodes indicate the amount of environmental information they would contribute to any agent perceiving only that agents output. 4.4 Emerging concepts in a well-mixed heterogeneous population So far, we have only considered populations of agents that acquired the same aspects of information from µ (i.e., p(Yθiµ) = p(Yθjµ) for any pair of agents θi, θj). The assumption was that the information that was relevant for the survival of the agents was the same among the agents of the population, and this was represented by µ. Now, we consider a more general scenario, where different types of agents acquire different aspects from the environmental conditions µ. We investigate whether it is possible for an agent that does not directly perceive the environment at 13 all (we call this type of agent "blind") to predict conditions based solely on the outputs of other agents. We consider a well-mixed population, such that different types of agents are forced to talk to each other. Considerations with a flexible population structure are not interesting for our purposes, since in these cases, each type of agent forms a cluster disconnected from clusters of other types. This was confirmed by simulations which are not shown here. Let us illustrate the idea with a relatively simple scenario: we consider five types of agents (we denote the i-th type φi), where each type can only distinguish whether the current state of the environment belongs to its coloured region or not. The environment consists of 9 states, and the probability of each state is uniformly distributed. We illustrate this environment by a 3×3 grid, as shown in Fig. 11, although the square does not denote the physical structure of the environment. Then, the outputs of each type of agent will be related to the regions they capture. For instance, for agents of type φ2 with the same deterministic code, if P r(µ ∈ {1, 2, 4, 5}Xθ = x) equals one (for all θ of type φ2), then x will signify that this agent is currently in the region coloured in red in Fig. 11. We say that a population of agents has a joint concept of the environment if by considering its representation of the environmental information they capture, we can obtain information about the environment, i.e. we require that I(µ; XΘ) > 0. For instance, the symbol x in the example above, assuming that it is only utilised by agents of the same type, can be understood as representing the concept "top-left" of the grid. 1 2 3 4 5 6 7 8 9 states of µ type φ1 type φ2 type φ3 type φ4 type φ5 Figure 11: Representation of the conditional probabilities p(Yθµ) for an agent θ of each type. These are defined such that each type of agent can only distinguish between the coloured region and the white region. For instance, the sensor of type φ2 is defined as P r(Y = y1µ) = 1 if µ ∈ {1, 2, 4, 5}, and zero otherwise, and P r(Y = y2µ) = 1 if µ /∈ {1, 2, 4, 5}, and zero otherwise. For type φ1, P r(Y = y1µ) = 0.5 and P r(Y = y2µ) = 0.5 (Y = 2 for all types of agents). The amount of environmental information that an agent θ of type φ1 (a blind agent) captures is I(µ; Yθ) = 0 bits, while all agents θ of the other types capture I(µ; Yθ) = 0.991076 bits (note that the total entropy in µ to be resolved is H(µ) = 3.16993 bits). Throughout this study, we considered that agents predict the environment by considering their perceptions together with the outputs of other agents. The blind agent, instead, since it is not able to capture any direct cue 14 from µ, we consider capable of perceiving the outputs of both of the agents selected by Θ and Θ(cid:48). With this relaxed consideration, we say a blind agent has a concept of the environment if I(µ; XΘ, XΘ(cid:48)) > 0, i.e. we consider the maximum amount of information an agent can possibly infer from the joint outputs XΘ and XΘ(cid:48). Let us recall that the structure of the population is well-mixed, and thus the distribution of outputs of all agents is considered, including the blind ones, which are not able to express (via their outputs) any particular concept by themselves (for a blind agent θ, I(µ; Xθ) ≤ I(µ; Yθ) = 0, i.e. I(µ; XΘ) vanishes). Therefore, whether a blind agent has some concept of the environment will depend, first, on the universality of the codes of each type of agent (agents representing the same information with different symbols may create ambiguities). Second, on the cardinality of X) utilised by the population. A small alphabet will force agents to the alphabet of X (i.e. represent different concepts of the environment with the same symbols, while a large alphabet is likely to result in exclusive representations of concepts for each type of agent. Taking this into account, we ask, is it possible for a blind agent to identify concepts of the environment? If so, how are these concepts related to the concepts of the individual agents (other than the blind ones)? Is the size of the available alphabet related to the quality of the concepts? To study these questions, we performed different experiments varying the size of the alphabet X, where the rest of the parameters remained the same. In these experiments, we optimised the similarity of codes for a population composed of 20 agents, with 4 agents of each of the five types. In Table 1 we show that the cardinality of the alphabet of X affects the limit of the amount of information a blind agent can possibly infer about the environment. Now, if we measure the uncertainty of the environment for a blind agent for each combination of outputs XΘ and XΘ(cid:48), we find that for some of them, it is zero. For instance, with X = 7, we found that when P r(µ = 5XΘ = 1, XΘ(cid:48) = 2) = 1.0 (see Fig. 12, where only combinations with XΘ ≤ XΘ(cid:48) are shown). These distributions are also valid when swapping the values of XΘ and XΘ(cid:48), since in the well-mixed population the structure is symmetric. Looking at the example of the conditional probability in Fig. 12, we can find many other concepts, although none of them -- apart from the one already discussed -- can uniquely identify a state of the environment. For instance, we have that P r(µXΘ = 3, XΘ(cid:48) = 6) = 0.33 when µ ∈ {3, 5, 7}, which is a concept for being on a particular diagonal of the environment. In Fig. 13 we show the resulting codes (which are universal for each type, including the blind one) for this particular experiment. Here, the types φ2 (red) and φ5 (purple) utilise the same 15 X 2 3 4 5 6 7 8 9 I(µ; XΘ, XΘ(cid:48)) 0.34621 0.56555 0.71620 0.95467 1.08139 1.18362 1.30919 1.30919 Table 1: Results of experiments where the size of the alpha- bet of a population varies. The maximum amount of environ- mental information that a blind agent can infer is achieved with X = 8 and remains equal for bigger alphabets. As the Figure 12: Conditional probabil- ity p(µXΘ, XΘ(cid:48) ) in inverse grey-scale. Each row represents a combination of values of XΘ and XΘ(cid:48) , and each column size of the alphabet decreases, this information also decreases. represents a state of µ. symbols to represent different environmental conditions. By using a small size of the alphabet for X, we force ambiguities in the population, but these will be chosen (by evolution) such that they are minimal. In this way, we maximise the amount of information we can infer from the outputs (although this can be a local optimum). For instance, the outputs of the blind agents (type φ1) for all the experiments never overlapped that of other types (unless we use X = 2, where there is no choice). In other words, blind agents always choose one symbol so that they minimise the amount of utilised symbols from the whole population. Figure 13: Representation of codes p(XΘYΘ, Θ) by a heat-map using inverse grayscale for the experiment with X = 7. For each node, the rows represent a sensor state y, while the columns represent an output state x. The colours on top of the nodes are used to distinguish the type of agent to whom the code belongs, and colours are related to those shown in Fig. 11. In all the performed experiments, we found that for values of X ≥ 6, the blind agent can perfectly predict the environmental state µ = 5 for at least one combination of outputs XΘ and XΘ(cid:48). Interestingly, this new concept, which in this particular experiment can be called the "centre" 16 µ1µ2µ3µ4µ5µ6µ7µ8µ9XΘ=1,XΘ(cid:48)=1XΘ=1,XΘ(cid:48)=2XΘ=1,XΘ(cid:48)=3XΘ=1,XΘ(cid:48)=4XΘ=1,XΘ(cid:48)=5XΘ=1,XΘ(cid:48)=6XΘ=1,XΘ(cid:48)=7XΘ=2,XΘ(cid:48)=2XΘ=2,XΘ(cid:48)=3XΘ=2,XΘ(cid:48)=4XΘ=2,XΘ(cid:48)=5XΘ=2,XΘ(cid:48)=6XΘ=2,XΘ(cid:48)=7XΘ=3,XΘ(cid:48)=3XΘ=3,XΘ(cid:48)=4XΘ=3,XΘ(cid:48)=5XΘ=3,XΘ(cid:48)=6XΘ=3,XΘ(cid:48)=7XΘ=4,XΘ(cid:48)=4XΘ=4,XΘ(cid:48)=5XΘ=4,XΘ(cid:48)=6XΘ=4,XΘ(cid:48)=7XΘ=5,XΘ(cid:48)=5XΘ=5,XΘ(cid:48)=6XΘ=5,XΘ(cid:48)=7XΘ=6,XΘ(cid:48)=6XΘ=6,XΘ(cid:48)=7XΘ=7,XΘ(cid:48)=7of the world or environment, cannot be obtained by looking to individual concepts only. 5 Discussion We considered four different scenarios of code evolution: in the first one, all agents perceived the outputs of all other agents, including itself. We argued that two main stages of evolution can be recognised: in the first stage, a universal code is established, which can be optimal or not. If it is not optimal, then a second stage will achieve optimality. The same result was obtained in [34], in a model of the evolution of the genetic code (represented as a probabilistic mapping between codons and amino acids), although universality and optimality were simultaneously achieved. In the mentioned work, which developed further the ideas of [38, 39], the authors argue that the universality of the genetic code is a consequence of early communal evolution, mediated by horizontal gene transfer (HGT) between primitive cells. In this evolutionary process, they argue, larger communities will have access (through the exchange of genetic material) to more innovations, leading to faster evolution than smaller ones. Then, "it is not better genetic codes that give an advantage but more common ones" [34]. Although their model does not explicitly show this property, it is captured in our model. We show that a more common, but not optimal code is widely adopted within a population (see Fig. 6). However, in our model, a code imposes itself as universal not because it provides access to more innovations (in our model there is no "code exchange", only the outputs are shared), but because the population structure forces the adoption of the most popular code. After this stage, further changes in the code of the agents eventually lead to optimality. In another related work, [21] explored the origins of language in a scenario consisting of artificial agents with a coupled perception and production of speech sounds. Although this work is focused on plausible mechanisms for the origin of language, it assumes the same similarity principle as we do (hearing a vocalisation increases the probability of producing similar vocalisations), arriving to the same outcome (a universal language, or code). Other works have considered similar principles in the evolution of languages: for instance, the naming game [32] and the imitation game [5]. However, these models assume some common conventions in order to evolve new ones. In this study, our main assumption was that the population of agents depended on common environmental conditions. Our second scenario, where the structure of the population is a grid, showed how establishing local conventions in early stages of evolution constrains the outcome of the code distribution, 17 since to reconcile different conventions, several simultaneous changes are needed. On the other hand, in our third scenario, where we let the structure of the population change simultaneously with the codes themselves, such situations are avoided by "disconnecting" clusters with dissimilar conventions. This property enhances evolution, and can potentially lead to the adoption of several different conventions within an increasingly fragmenting, or "speciating" population. Our last scenario assumed perceptual constraints on the environmental information of each agent, an we looked at emerging concepts within a well-mixed population. This scenario was studied in [20], where, as well as in our study, new conceptualisations of the world emerged as a result of considering together the concepts of every agent. In both studies, the new concept was not representable individually by any agent. Differently from the mentioned study, the new concepts obtained in our study were the result of a simple similarity maximisation principle, while in the work of [20], concepts were obtained through the modelling of an explicit fitness function. The evolution of conventional codes could be interpreted, in the widest sense, as a form of cultural evolution. For instance, considering the definition of culture given by [25]: "Culture is information capable of affecting individuals' behavior that they acquire from other members of their species through teaching, imitation, and other forms of social transmission.", it could be argued that a form of cultural information is present in organisms, such as bacteria or plants. Although there is a dependence among the different dimensions on which information is transmitted in organisms (if we assume the dimensions to be, for instance, genetic, epigenetic, behavioural and symbol-based, as proposed by [16]), our model assumes freedom of choice in one dimension, without direct influence on the others. Finally, communication between individuals of a population opens up the possibility of "signal cheaters", which could be either individuals that do not produce signals themselves but still perceive those of the others, or individuals who exploit other individual's learned responses to symbols to their advantage. However, our model does not allow such behaviour, since the code producing the outputs functions, implicitly, as the interpreter of the perceived signals. 6 Conclusion In the proposed model, we introduced a key assumption which allowed us to evolve, for some structures, universal and optimal codes. This assumption states that an agent cannot distinguish the sources of the outputs it perceives from other agents. Following from this, a universal code will necessary introduce semantics by relating symbols to environmental conditions (via the internal 18 states of the agent). Our model proposes an information-theoretic way of measuring the similarity within a population of codes. In this work, we proposed, as an evolutionary principle, that agents try to maximise their side information about the environment indirectly by maximising their mutual code similarity. This be- haviour produces several interesting outcomes in the code distribution of a structured population. Depending on the population structure, it captures the evolution of a universal and optimal code (well-mixed population structure), while also the evolution of different codes organised in clusters (in a freely evolving structure), which allows the establishment of optimal as well as suboptimal conventions. Finally, we considered a well-mixed heterogeneous population with perceptual constraints on the agents about the environment, and showed how, just by looking at the outputs of agents, it is possible to extract concepts that relate to the environment, concepts that none of the agents of the population could individually represent. References [1] G´abor Bal´azsi, Alexander van Oudenaarden, and James J Collins. Cellular decision making and biological noise: from microbes to mammals. Cell, 144(6):910 -- 25, March 2011. [2] Marcello Barbieri. The organic codes-An introduction to semantic biology. Genetics and Molecular Biology, 2003. [3] Marcello Barbieri. Biosemiotics: a new understanding of life. Die Naturwissenschaften, 95(7):577 -- 99, July 2008. [4] G´erard Battail. Applying Semiotics and Information Theory to Biology: A Critical Compar- ison. Biosemiotics, 2(3):303 -- 320, October 2009. [5] Bart De Boer. Self-organization in vowel systems. Journal of Phonetics, 2000. [6] Dan Cohen. Optimizing reproduction in a randomly varying environment. Journal of Theo- retical Biology, 12(1):119 -- 129, September 1966. [7] J Deely. On 'semiotics' as naming the doctrine of signs. Semiotica, 2006. [8] Matina C. Donaldson-Matasci, Carl T Bergstrom, and Michael Lachmann. The fitness value of information. Oikos, 119(2):219 -- 230, 2010. 19 [9] Donald Favareau. The evolutionary history of biosemiotics. Introduction to biosemiotics, pages 1 -- 67, 2007. [10] Dennis Gorlich, Stefan Artmann, and Peter Dittrich. Cells as semantic systems. Biochimica et biophysica acta, 1810(10):914 -- 23, October 2011. [11] Nikolaus Hansen and Stefan Kern. Evaluating the CMA evolution strategy on multimodal test functions. Parallel Problem Solving from Nature-PPSN VIII, 2004. [12] Nikolaus Hansen and A Ostermeier. Completely derandomized self-adaptation in evolution strategies. Evolutionary computation, 9(2):159 -- 95, January 2001. [13] Martin Heil and Richard Karban. Explaining evolution of plant communication by airborne signals. Trends in ecology & evolution, 25(3):137 -- 44, March 2010. [14] Francis Heylighen. Stigmergy as a generic mechanism for coordination: definition, varieties and aspects. Technical report, Working paper, 2011. [15] Christian Igel, Verena Heidrich-meisner, and Tobias Glasmachers. Shark. Journal of Machine Learning Research, 9:993 -- 996, 2008. [16] Eva Jablonka, Marion J Lamb, and Anna Zeligowski. Evolution in four dimensions: Genetic, epigenetic, behavioral, and symbolic variation in the history of life, volume 5. MIT press Cambridge, MA, 2005. [17] J Kelly. A new interpretation of information rate. IEEE Transactions on Information Theory, 2(3):185 -- 189, 1956. [18] Edo Kussell and Stanislas Leibler. Phenotypic diversity, population growth, and information in fluctuating environments. Science (New York, N.Y.), 309(5743):2075 -- 8, September 2005. [19] J. Lin. Divergence measures based on the Shannon entropy. IEEE Transactions on Informa- tion Theory, 37(1):145 -- 151, 1991. [20] Marco Moller and Daniel Polani. Common concepts in agent groups, symmetries, and con- formity in a simple environment. Artificial Life XI, 2008. [21] PY Oudeyer. The self-organization of speech sounds. Journal of Theoretical Biology, 2005. [22] H Van Dyke Parunak. A survey of environments and mechanisms for human-human stigmergy. In Environments for Multi-Agent Systems II, pages 163 -- 186. Springer, 2006. 20 [23] Theodore J Perkins and Peter S Swain. Strategies for cellular decision-making. Molecular systems biology, 5(326):326, January 2009. [24] Thomas G Platt and Clay Fuqua. What's in a name? The semantics of quorum sensing. Trends in microbiology, 18(9):383 -- 7, September 2010. [25] Peter J Richerson and Robert Boyd. Not by genes alone: How culture transformed human evolution. University of Chicago Press, 2005. [26] Olivier Rivoire and Stanislas Leibler. The Value of Information for Populations in Varying Environments. Journal of Statistical Physics, 142(6):1124 -- 1166, March 2011. [27] Martin Schuster, D Joseph Sexton, Stephen P Diggle, and E Peter Greenberg. Acyl- homoserine lactone quorum sensing: from evolution to application. Annual review of mi- crobiology, 67:43 -- 63, January 2013. [28] J Seger and H. Jane Brockmann. What is bet-hedging? Oxford surveys in evolutionary biology, 1987. [29] Jyoti Shah. Plants under attack: systemic signals in defence. Current opinion in plant biology, 12(4):459 -- 64, August 2009. [30] C.E. Shannon. A mathematical theory of communication. Bell Systems Technical Journal, 27:379 -- 423, 1948. [31] Montgomery Slatkin. Hedging one's evolutionary bets. Nature, 250:704 -- 705, 1974. [32] L Steels. A self-organizing spatial vocabulary. Artificial life, 1995. [33] Guy Theraulaz and Eric Bonabeau. A brief history of stigmergy. Artificial life, 5(2):97 -- 116, January 1999. [34] Kalin Vetsigian, Carl R Woese, and Nigel Goldenfeld. Collective evolution and the genetic code. Proceedings of the National Academy of Sciences of the United States of America, 103(28):10696 -- 10701, 2006. [35] Christopher M Waters and Bonnie L Bassler. Quorum sensing: cell-to-cell communication in bacteria. Annual review of cell and developmental biology, 21:319 -- 46, January 2005. [36] Stuart a West, Ashleigh S Griffin, Andy Gardner, and Stephen P Diggle. Social evolution theory for microorganisms. Nature reviews. Microbiology, 4(8):597 -- 607, August 2006. 21 [37] Edward O Wilson. Sociobiology. Harvard University Press, 2000. [38] Carl R Woese. On the evolution of cells. Proceedings of the National Academy of Sciences of the United States of America, 99(13):8742 -- 7, June 2002. [39] Carl R Woese. A new biology for a new century. Microbiology and Molecular Biology Reviews, 68(2):173 -- 186, 2004. [40] a K Wong and M You. Entropy and distance of random graphs with application to struc- tural pattern recognition. IEEE transactions on pattern analysis and machine intelligence, 7(5):599 -- 609, May 1985. 22
1909.03475	1	1909	2019-09-08T14:18:31	An Architectural Style for Self-Adaptive Multi-Agent Systems	[ "cs.MA", "cs.SE" ]	Modern distributed software systems often operate in dynamic environments in which operation conditions change continuously and subsystems may come and go at will, e.g. intelligent traffic management and multi-robot systems. To manage these dynamics, these systems have to self-adapt their structures and behaviors dynamically. While we have witnessed significant progress over the past decade in the manner in which such systems are designed, persistent challenges remain. In particular, dealing with distribution and decentralized control remains one of the major challenges in self-adaptive systems. This report presents an architecture style that supports software architects with designing architectures for a family of decentralized self-adaptive systems. The architecture style structures the software in a number of interacting autonomous entities (agents) that cooperatively realize the system tasks. Multi-agent systems derived from the architectural style realize flexibility (agents adapt their behavior and interactions to variable operating conditions) and openness (agents cope autonomously with other agents that enter and leave the system). The architectural style consists of five related patterns that distill domain-specific architectural knowledge derived from extensive experiences with developing various multi-agent systems. The architectural patterns are specified using pi-ADL, a formal architectural description language supporting specification of dynamic architectures. This specification provides architects with a rigorous description of the architecture elements of the patterns, their interactions and behavior. We illustrate how we have applied the architectural style with excerpts of two cases from our practice: an experimental system for anticipatory traffic routing and an industrial logistic system for automated transportation in warehouse environments.	cs.MA	cs	An Architectural Style for Self-Adaptive Multi-Agent Systems DANNY WEYNS KU Leuven, Belgium and Linnaeus University, Sweden FLAVIO OQUENDO IRISA - Universit´e de Bretagne-Sud, France Modern distributed software systems often operate in dynamic environments in which oper- ation conditions change continuously and subsystems may come and go at will, e.g. intelligent traffic management and multi-robot systems. To manage these dynamics, these systems have to self-adapt their structures and behaviors dynamically. While we have witnessed significant progress over the past decade in the manner in which such systems are designed, persistent chal- lenges remain. In particular, dealing with distribution and decentralized control remains one of the major challenges in self-adaptive systems. This report presents an architecture style that sup- ports software architects with designing architectures for a family of decentralized self-adaptive systems. The architecture style structures the software in a number of interacting autonomous entities (agents) that cooperatively realize the system tasks. Multi-agent systems derived from the architectural style realize flexibility (agents adapt their behavior and interactions to variable oper- ating conditions) and openness (agents cope autonomously with other agents that enter and leave the system). The architectural style consists of five related patterns that distill domain-specific architectural knowledge derived from extensive experiences with developing various multi-agent systems. The architectural patterns are specified using π-ADL, a formal architectural description language supporting specification of dynamic architectures. This specification provides architects with a rigorous description of the architecture elements of the patterns, their interactions and be- havior. We illustrate how we have applied the architectural style with excerpts of two cases from our practice: an experimental system for anticipatory traffic routing and an industrial logistic system for automated transportation in warehouse environments. Categories and Subject Descriptors: D.2.11 [Software Engineering]: Software Architectures -- Domain-specific architectures; I.2.11 [Artificial Intelligence]: Distributed Artificial Intelligence -- Multi-agent systems General Terms: Design Additional Key Words and Phrases: Architectural style, patterns, self-adaptation, decentralized control, flexibility, openness; Authors' addresses: D. Weyns, Department of Computer Science, KU Leuven, 3000 Leuven Belgium and Lin- naeus University, 351 95, Sweden; [email protected]. F. Oquendo, IRISA Research Institute - Univer- sit´e de Bretagne-Sud, France, ENSIBS, BP 573, F-56017 Vannes Cedex, France; [email protected]. Permission to make digital/hard copy of all or part of this material without fee for personal or classroom use provided that the copies are not made or distributed for profit or commercial advantage, the ACM copyright/server notice, the title of the publication, and its date appear, and notice is given that copying is by permission of the ACM, Inc. To copy otherwise, to republish, to post on servers, or to redistribute to lists requires prior specific permission and/or a fee. c(cid:13) 2012 ACM 0000-0000/2001/0000-0001 $5.00 ACM Transactions Journal Format, Vol. 2, No. 3, 09 2001, Pages . · 2 D. Weyns and F. Oquendo 1. INTRODUCTION Due to the increasing integration and dynamicity of software systems, providing and main- taining the required qualities of software constitutes a huge engineering challenge. Mod- ern distributed software systems, such as mobile multimedia applications, intelligent traffic management, and multi-robot systems usually comprise different subsystems that collabo- rate with each other as peers. Such systems typically have to operate in dynamic environ- ments in which operation conditions change continuously and subsystems may come and go at will. To manage these dynamics and maintain the required functionality and quali- ties, these systems have to self-adapt their structures and behaviors dynamically [Cheng et al. 2009; Weyns 2019]. Several researchers have argued that software architecture provides the right level of abstraction and generality to deal with the challenges of self- adaptability [Oreizy et al. 1999; Garlan et al. 2004; Kramer and Magee 2007; Weyns et al. 2012]. While we have witnessed significant progress over the past decade in the manner in which such systems are designed, persistent challenges remain [Dobson et al. 2006; Hueb- scher and McCann 2008; Salehie and Tahvildari 2009; Cheng et al. 2009; de Lemos et al. 2012; de Lemos et al. 2017]. In particular, dealing with distribution and decentralized control remains one of the major challenges in self-adaptive systems [Kramer and Magee 2007; Brun et al. 2009; Weyns et al. 2013]. This report presents an architecture style that supports software architects with design- ing a family of distributed self-adaptive systems in which control is decentralized. The architecture style distills the knowledge and expertise acquired from eight years of experi- ence with designing and developing multi-agent systems. Multi-agent systems are a class of decentralized self-adaptive systems, comprising of autonomous agents that inhabit some complex dynamic environment, sense and act autonomously in this environment, and by doing so realize a set of goals or tasks for which they are designed [Maes 1995; Jennings et al. 1998; Sterling and Taveter 2009; Weyns and Georgeff 2010]. The architecture style structures the software in a number of such interacting autonomous agents that coopera- tively realize the system tasks with two key quality properties: flexibility, i.e. agents adapt their behavior and interactions to variable operating conditions, and openness, i.e., agents cope autonomously with other agents that enter and leave the system. The architecture style consists of five related architectural patterns that provide generic solution schemes for recurring design problems. These architecture patterns make the ap- plication of the architecture style more operational. To document the architectural patterns we use a template inspired by the approach for documenting architecture styles presented in [Clements et al. 2010]. The elements of the architecture patterns and their relation- ships are specified in π-ADL [Oquendo 2004], a formally founded architecture description language. This specification provides a rigorous description of the architecture elements, their structures, and behaviors. The specification is type checked using the π-ADL.NET tool [Qayyum and Oquendo 2008]. The work presented in this report builds upon previous work [Weyns 2009] in which we have introduced a set of patterns for multi-agent systems. This report extends that work in multiple ways, including a refinement and elaborated discussion of the patterns of the architecture style, the application of the patterns to two applications including a system for anticipatory vehicle routing that was recently developed, and a systematic and complete specification of the structural and behavioral aspects of the architectural style in π-ADL. In addition, this report provides methodological guidelines for how to use the architecture ACM Transactions Journal Format, Vol. 2, No. 3, 09 2001. An Architectural Style for Self-Adaptive Multi-Agent Systems · 3 style in practice, and a classification of related work on patterns for multi-agent systems. Overview The report is structured as follows. In Section 2 we introduce situated multi-agent systems, the family of decentralized self-adaptive systems that are supported by the architecture style. We discuss the background of the architecture style and introduce two applications that we have built with the style and use for demonstration in the remainder of the paper. Section 3 gives a high-level overview of architectural style, explains the template we use to document patterns, and introduces the necessary basics of π-ADL. Section 4 is the heart of the report, providing a detailed description of the five patterns of the architectural style. Section 5 provides methodological guidelines for using the architecture style. Section 6 discusses related work, and Section 7 draws conclusions and looks at future work. For space reasons, we have omitted some parts of the π-ADL specification. For a complete π-ADL specification, we refer to the online Appendix A. 2. BACKGROUND This section introduces situated multi-agent systems, the family of self-adaptive systems supported by the architecture style. We give a brief overview of the applications that un- derly the architecture style and we introduce two applications from our practice that we use for illustration in the remainder of the report. Finally, we summarize the properties that determine the target domain of the architecture style. 2.1 Situated Multi-Agent Systems A multi-agent system is in essence a system that is structured as a set of autonomous agents. [Wooldridge and Jennings 1995] define an agent as "a computer system, situated in some environment, that is capable of flexible autonomous action in order to meet its design objectives." A multi-agent system is one that consists of a number of agents, which interact with one-another. [Durfee and Lesser 1989] define a multi-agent system as "a loosely coupled network of problem solvers (agents) that interact to solve problems that are beyond the individual capabilities or knowledge of each problem solver." Characteristics of multi- agent systems are as follows: (1) each agent has incomplete information or capabilities for solving the problem and, thus, has a limited viewpoint; (2) there is no system global control; (3) data is distributed; and (4) computation is asynchronous. Multi-agent systems are known for quality attributes such as adaptability, openness, robustness, and scalability, making them particularly interesting to deal with the demanding challenges of complex distributed software applications. Situated multi-agent systems are one family of multi-agent systems. The focus of sit- uated multi-agent systems is on modularization of agent behavior, efficient decision mak- ing, and indirect coordination [Maes 1990; Ferber 1999; Bandini et al. 2002; Mamei and Zambonelli 2004; Parunak et al. 2005; Weyns 2010]. This contrasts with delibera- tive approaches of multi-agent systems that emphasize knowledge representation, ratio- nality, planning, and direct communication [Rao and Georgeff 1995; Sycara et al. 1996; Wooldridge 2000; Pokahr et al. 2005; Bordini et al. 2007]. A particularly interesting as- pect of situated multi-agent systems is indirect coordination [Weyns et al. 2005; Weyns et al. 2007; Weyns et al. 2015]. Indirect coordination reduces the coupling between agents and provides a means for managing complexity by separation of concerns. Classical ex- amples of indirect coordination are computational fields [Mataric 1994; Mamei and Zam- ACM Transactions Journal Format, Vol. 2, No. 3, 09 2001. · 4 D. Weyns and F. Oquendo In bonelli 2004] and digital pheromones [Bonabeau et al. 1999; Parunak et al. 2005]. the approach of computational fields, agents coordinate their behavior by following the shape of virtual force fields that are spread through a shared coordination infrastructure. Environment dynamics and movements of the agents induce changes in the surface of the fields, realizing a feedback cycle that influences the agents' movement. In the ap- proach of digital pheromones, situated agents coordinate their behavior through digital markers (pheromones) in a shared coordination medium, similar to social ants. Agents can drop pheromones to form paths that can guide the agents to locations of interest. Digital pheromones evaporate over time. This mechanism ensures that paths to outdated informa- tion will fade away. These forms of indirect coordination are often constructed on top of basic mechanisms of shared memory based on tuple spaces [Carriero and Gelernter 1989; Schelfthout et al. 2005; Murphy et al. 2006; Mamei and Zambonelli 2009]. 2.2 Underlying Expertise of the Architecture Style The architecture style embodies architectural knowledge gained from the design and devel- opment of various multi-agent system applications. We extensively used the Packet -- World, a simple robotic application, as a study case for investigation and experimentation [Weyns et al. 2005]. We derived expertise from the design and development of a distributed peer-to- peer file sharing system [Schelfthout and Holvoet 2004; Weyns et al. 2004b]. Our focus in this application was on coordination mechanisms inspired by principles of social ants. We have applied multi-agent systems in several experimental robotic applications focussing on the roles agents play to set up collaborations [Steegmans et al. 2004; Weyns et al. 2004a]. We have employed a decentralized multi-agent system architecture in an industrial trans- portation system for controlling automatic guided vehicles [Weyns et al. 2005; Weyns et al. 2008]. A particular interest in this industrial application was on advanced coordina- tion mechanisms that enable situated agents to flexibly adapt their behavior to changes in the stream of tasks that enter the system and other environmental dynamics. We have been using agents in the domain of intelligent transportation systems, for example for traffic jam monitoring [Weyns et al. 2010], supply chains [Haesevoets et al. 2014], and anticipatory vehicle routing [Claes et al. 2011]. The focus here is on situated agents that exploit a co- ordination medium to dynamically organize themselves based on the changing context in which they are situated. Recently, we have also been studying how we can apply similar ideas in the management of dynamic supply chains [Haesevoets 2012]. We now introduce two cases that we use to demonstrate how we have applied the archi- tectural style and realized the patterns in our work. We start by introducing an experimental system for anticipatory vehicle routing. Then we introduce an industrial logistic system for automated transportation. 2.2.1 Anticipatory Vehicle Routing System. The first application is an experimental system to avoid traffic congestion based on anticipatory vehicle routing [Weyns et al. 2007; Claes et al. 2011]. Fig. 1 shows a high-level component model of the system. This application is situated in the domain of intelligent transportation systems, a world- wide initiative to exploit information and communication technology to improve traffic.1 The large number of vehicles today, the limited capacity of the road networks, and un- predictable dynamics of the traffic context (such as a fluctuating amount of traffic, chang- 1http://www.itsworldcongress.com/ ACM Transactions Journal Format, Vol. 2, No. 3, 09 2001. An Architectural Style for Self-Adaptive Multi-Agent Systems · 5 Fig. 1. Component model of the simulated anticipatory vehicle routing system ing behavior of drivers, traffic jams, and road accidents) make traffic routing a particu- larly challenging problem. Existing technology such as satellite navigation devices exploit broadcast traffic information to trigger vehicles to reroute and bypass traffic jams. Al- though such mechanisms allow substantial performance gain already, they only react upon traffic jams after they occurred. In our research, we study anticipatory vehicle routing that aims to encompass this by using forecast of traffic density. The approach presented here assumes that the road infrastructure is equipped with elec- tronic devices which provide some computation power and are connected through a net- work. Vehicles are equipped with smart devices which can sense that vehicle's local traffic conditions and which can communicate with other vehicles and electronic devices of the road infrastructure. On each vehicle device, a vehicle agent is deployed (see Figure 1) that has knowledge about that vehicle's start location and destination, and its current state such as position and speed. The vehicle agent provides information to the driver to reach its destination with minimal congestion overhead. On each infrastructure device, an infrastructure agent is deployed that has knowledge about the traffic conditions of a part of the road network and maintains bookings for future traffic. Each vehicle agent explores the relevant paths in the environment, and based on the evaluation of the possible alternatives, it chooses one path and records it as an intention. Therefore, the vehicle agent informs the corresponding infrastructure agents of the time the vehicle intends to occupy the corresponding road elements, i.e. the agents need to book these reservations. To explore paths and book reservations vehicle agents use so called ant agents. Ant agents are smart messages that travel on behalf of a vehicle agent along a path in the environment visiting a sequence of infrastructure agents [Holvoet et al. 2009]. A vehicle virtual environment reflects the real traffic environment. This virtual environ- ment is a distributed software entity that is deployed on each device in the traffic system. Vehicle agents and infrastructure agents use the virtual environment to enable the ant agents to explore feasible paths in the environment and make bookings for the selected paths. ACM Transactions Journal Format, Vol. 2, No. 3, 09 2001. (cid:87)(cid:105)(cid:114)(cid:101)(cid:108)(cid:101)(cid:115)(cid:115)(cid:69)(cid:116)(cid:104)(cid:101)(cid:114)(cid:110)(cid:101)(cid:116)(cid:86)(cid:101)(cid:104)(cid:105)(cid:99)(cid:108)(cid:101)(cid:86)(cid:105)(cid:114)(cid:116)(cid:117)(cid:97)(cid:108)(cid:69)(cid:110)(cid:118)(cid:105)(cid:114)(cid:111)(cid:110)(cid:109)(cid:101)(cid:110)(cid:116)(cid:86)(cid:101)(cid:104)(cid:105)(cid:99)(cid:108)(cid:101)(cid:65)(cid:103)(cid:101)(cid:110)(cid:116)(cid:73)(cid:110)(cid:102)(cid:114)(cid:97)(cid:115)(cid:116)(cid:114)(cid:117)(cid:99)(cid:116)(cid:117)(cid:114)(cid:101)(cid:86)(cid:105)(cid:114)(cid:116)(cid:117)(cid:97)(cid:108)(cid:69)(cid:110)(cid:118)(cid:105)(cid:114)(cid:111)(cid:110)(cid:109)(cid:101)(cid:110)(cid:116)(cid:73)(cid:110)(cid:102)(cid:114)(cid:97)(cid:115)(cid:116)(cid:114)(cid:117)(cid:99)(cid:116)(cid:117)(cid:114)(cid:101)(cid:65)(cid:103)(cid:101)(cid:110)(cid:116)(cid:73)(cid:110)(cid:102)(cid:114)(cid:97)(cid:115)(cid:116)(cid:114)(cid:117)(cid:99)(cid:116)(cid:117)(cid:114)(cid:101)(cid:86)(cid:105)(cid:114)(cid:116)(cid:117)(cid:97)(cid:108)(cid:69)(cid:110)(cid:118)(cid:105)(cid:114)(cid:111)(cid:110)(cid:109)(cid:101)(cid:110)(cid:116)(cid:73)(cid:110)(cid:102)(cid:114)(cid:97)(cid:115)(cid:116)(cid:114)(cid:117)(cid:99)(cid:116)(cid:117)(cid:114)(cid:101)(cid:65)(cid:103)(cid:101)(cid:110)(cid:116)(cid:83)(cid:105)(cid:109)(cid:117)(cid:108)(cid:97)(cid:116)(cid:101)(cid:100)(cid:84)(cid:114)(cid:97)(cid:102)(cid:102)(cid:105)(cid:99)(cid:69)(cid:110)(cid:118)(cid:105)(cid:114)(cid:111)(cid:110)(cid:109)(cid:101)(cid:110)(cid:116)(cid:83)(cid:105)(cid:109)(cid:117)(cid:108)(cid:97)(cid:116)(cid:101)(cid:100)(cid:67)(cid:111)(cid:109)(cid:109)(cid:117)(cid:110)(cid:105)(cid:99)(cid:97)(cid:116)(cid:105)(cid:111)(cid:110)(cid:73)(cid:110)(cid:102)(cid:114)(cid:97)(cid:115)(cid:116)(cid:114)(cid:117)(cid:99)(cid:116)(cid:117)(cid:114)(cid:101)(cid:67)(cid:111)(cid:109)(cid:112)(cid:111)(cid:110)(cid:101)(cid:110)(cid:116)(cid:75)(cid:69)(cid:89)(cid:85)(cid:115)(cid:101)(cid:115)· 6 D. Weyns and F. Oquendo Flexibility and openness are key requirements in this domain as vehicles face contin- uously changing operating conditions, such as changing traffic conditions, new vehicles that enter and leave the system, etc. Therefore, vehicle agents may revise there intentions and book alternative routes when needed. To that end, the vehicle agents send new ant agents at regular times to explore changes in the environment and refresh bookings. This mechanism makes the system robust to dynamics in the traffic environment. Our experiments are done in simulation of a real-world traffic environment, which in- clude a model of the physical traffic environment and the communication network. The simulated environment maintains the state of elements of the physical system, provides the means for monitoring the state of elements in the system, determines the outcome of the actuator actions of the agents, and it collects and delivers messages. 2.2.2 Automated Transportation System. The second application is an industrial trans- portation system that was developed between 2004 and 2007 in a collaboration between researchers of DistriNet Labs and a team of engineers and developers of Egemin2, a leading company that provides full life cycle support for automated transportation systems [Weyns et al. 2005; Weyns et al. 2008; Weyns 2010]. An automated transportation system consists of a number of automatic guided vehicles (AGVs) that need to work together to transport loads in an industrial environment. Trans- port tasks are generated by client systems, typically an enterprise resource planning (ERP) system. The main functionalities that an AGV transportation system has to fulfill are as- signing incoming transportation tasks to appropriate AGVs, routing the AGVs through the warehouse efficiently while avoiding collisions and deadlocks, and recharging the AGVs' batteries. An AGV transportation system has to deal with dynamic and changing operating con- ditions. The stream of transportation tasks that enters the system is typically irregular and unpredictable, AGVs can leave and re-enter the system for maintenance, production ma- chines may have variable waiting times, etc. All kinds of disturbances can occur, supply of goods can be delayed, certain areas in the warehouse may temporarily be closed for main- tenance services, loads can block paths, AGVs can fail, etc. Despite these challenging operating conditions, the system is expected to operate efficiently and robustly. Traditionally, the AGV systems deployed by Egemin are directly controlled by a central server. The server plans the schedule for the system as a whole, dispatches commands to the AGVs, and continually polls their status. System operators monitor the system and intervene whenever needed. This results in reliable solutions. The central point of control also enables easy diagnosis of errors. However, a shift in user requirements challenges the centralized architecture. Customers increasingly request for systems that are able to adapt their behavior with changing circumstances autonomously, i.e. without intervention of an operator. For example, the system should adapt itself when the stream of transportation tasks changes dynamically, or when disturbances occur. Such self-adaptability requires flexibility and openness of the solution. Flexibility enables the system to deal with dynamic operating conditions autonomously, and openness enables it to deal with AGVs entering and leaving the system. Applying a situated multi-agent system opens perspectives to improve flexibility and openness of the system: the AGVs can adapt themselves to the current situation in their 2http://www.egemin.com/ ACM Transactions Journal Format, Vol. 2, No. 3, 09 2001. An Architectural Style for Self-Adaptive Multi-Agent Systems · 7 Fig. 2. High-level deployment model of a transportation system vicinity, order assignment is dynamic, the system can deal autonomously with AGVs leav- ing and reentering the system, etc. Fig. 2 shows a high-level deployment model of the agent-based system. Each vehicle is controlled by an AGV agent, and each transportation task in the system is represented by a transport agent that is deployed at the transport base. To realize the system requirements, AGV agents and transport agents have to coordinate, e.g. for task assignment, for collision avoidance, etc. Therefore, the agents exploit a local virtual en- vironment. The states of neighboring local virtual environments are synchronized using middleware services that run on top of the .NET platform. 2.3 Target Domain of the Architecture Style We summarize the key characteristics and requirements shared by the systems that underly the architecture style: -- The software systems are subject to dynamic and changing operating conditions, such as dynamically changing workloads and variations in availability of resources and services. -- Activity in the systems is inherently localized, i.e. global control or access to resources is difficult to achieve or even infeasible. -- Important stakeholder requirements are (1) flexibility: the software systems have to adapt to variable operating conditions autonomously; and (2) openness: the software systems have to cope autonomously with parts that come and go during execution. In addition, we consider the following constraints in terms of the domain: -- The speed of changes in the environment are orders of magnitude lower than the speed of communication and the execution of the control software. -- Quasi continual communication access to the distributed software system is available, e.g., via a LAN. -- We consider distributed software systems within a single ownership domain. ACM Transactions Journal Format, Vol. 2, No. 3, 09 2001. (cid:65)(cid:71)(cid:86)(cid:77)(cid:105)(cid:100)(cid:100)(cid:108)(cid:101)(cid:119)(cid:97)(cid:114)(cid:101)(cid:83)(cid:101)(cid:114)(cid:118)(cid:105)(cid:99)(cid:101)(cid:115)(cid:77)(cid:105)(cid:100)(cid:100)(cid:108)(cid:101)(cid:119)(cid:97)(cid:114)(cid:101)(cid:83)(cid:101)(cid:114)(cid:118)(cid:105)(cid:99)(cid:101)(cid:115)(cid:84)(cid:114)(cid:97)(cid:110)(cid:115)(cid:112)(cid:111)(cid:114)(cid:116)(cid:66)(cid:97)(cid:115)(cid:101)(cid:65)(cid:71)(cid:86)(cid:86)(cid:105)(cid:114)(cid:116)(cid:117)(cid:97)(cid:108)(cid:69)(cid:110)(cid:118)(cid:105)(cid:114)(cid:111)(cid:110)(cid:109)(cid:101)(cid:110)(cid:116)(cid:84)(cid:66)(cid:86)(cid:105)(cid:114)(cid:116)(cid:117)(cid:97)(cid:108)(cid:69)(cid:110)(cid:118)(cid:105)(cid:114)(cid:111)(cid:110)(cid:109)(cid:101)(cid:110)(cid:116)(cid:46)(cid:78)(cid:69)(cid:84)(cid:65)(cid:71)(cid:86)(cid:86)(cid:105)(cid:114)(cid:116)(cid:117)(cid:97)(cid:108)(cid:69)(cid:110)(cid:118)(cid:105)(cid:114)(cid:111)(cid:110)(cid:109)(cid:101)(cid:110)(cid:116)(cid:65)(cid:71)(cid:86)(cid:77)(cid:105)(cid:100)(cid:100)(cid:108)(cid:101)(cid:119)(cid:97)(cid:114)(cid:101)(cid:83)(cid:101)(cid:114)(cid:118)(cid:105)(cid:99)(cid:101)(cid:115)(cid:46)(cid:78)(cid:69)(cid:84)(cid:46)(cid:78)(cid:69)(cid:84)(cid:65)(cid:71)(cid:86)(cid:65)(cid:103)(cid:101)(cid:110)(cid:116)(cid:65)(cid:71)(cid:86)(cid:65)(cid:103)(cid:101)(cid:110)(cid:116)(cid:84)(cid:114)(cid:97)(cid:110)(cid:115)(cid:112)(cid:111)(cid:114)(cid:116)(cid:65)(cid:103)(cid:101)(cid:110)(cid:116)(cid:84)(cid:114)(cid:97)(cid:110)(cid:115)(cid:112)(cid:111)(cid:114)(cid:116)(cid:65)(cid:103)(cid:101)(cid:110)(cid:116)(cid:67)(cid:108)(cid:105)(cid:101)(cid:110)(cid:116)(cid:115)(cid:75)(cid:69)(cid:89)(cid:67)(cid:111)(cid:109)(cid:112)(cid:111)(cid:110)(cid:101)(cid:110)(cid:116)(cid:80)(cid:114)(cid:111)(cid:99)(cid:101)(cid:115)(cid:115)(cid:105)(cid:110)(cid:103)(cid:110)(cid:111)(cid:100)(cid:101)(cid:115)(cid:115)(cid:105)(cid:116)(cid:117)(cid:97)(cid:116)(cid:101)(cid:100)(cid:105)(cid:110)(cid:119)(cid:97)(cid:114)(cid:101)(cid:104)(cid:111)(cid:117)(cid:115)(cid:101)(cid:87)(cid:105)(cid:114)(cid:101)(cid:108)(cid:101)(cid:115)(cid:115)(cid:69)(cid:116)(cid:104)(cid:101)(cid:114)(cid:110)(cid:101)(cid:116)· 8 D. Weyns and F. Oquendo These characteristics and constraints determine the target domain (family of systems) that are supported by the architecture style for multi-agent systems. 3. OVERVIEW OF THE ARCHITECTURE STYLE In this section, we give a general introduction of the architectural style. We start with a brief overview of the patterns that constitute the architecture style. Then, we outline the pattern template that we use for the detailed description of the patterns. We conclude the section with a brief overview of the basic language primitives of π-ADL. 3.1 Patterns of the Architecture Style Fig. 3 shows an overview of the different patterns of the architecture style with the rela- tionships between the patterns. Fig. 3. Overview of the patterns of the architectural style The basic patterns of the architecture style are Situated Agent and Virtual Environment. A situated agent is an autonomous problem solving entity that encapsulates its state and controls its behavior. The responsibility of an agent is to realize the application specific tasks it is assigned. Situated agents are cooperative entities that are able to flexibly adapt their behavior and interactions with changing conditions. Agents are situated in a virtual environment that maintains a virtualization of the relevant parts of the world and serves as a coordination medium for the agents, i.e. the virtual environment mediates both the interactions among agents and the access to resources. Besides the basic patterns, there are three additional patterns that refine the situated agent pattern: Selective Perception, Protocol-based Communication, and Behavior-Based Action Selection. A situated agent can use selective perception to sense the environment and up- date its knowledge about the world. Situated agents can use protocol-based communication to exchange messages according to prescribed communication protocols, i.e. well-defined ACM Transactions Journal Format, Vol. 2, No. 3, 09 2001. An Architectural Style for Self-Adaptive Multi-Agent Systems · 9 sequences of messages. Finally, a situated agent can use behavior-based action selection to decide about the actions it wants to perform in the environment. Selective perception, protocol-based communication, and behavior-based action selection are supported by the virtual environment. The patterns have several variability points. For example, for the design of agents that only communicate indirectly via the virtual environment, the protocol-based communica- tion pattern can be omitted. We elaborate on variability in Section 5. 3.2 Pattern Template To document the patterns of the architecture style we use the following template: (1) The name of the pattern. (2) A primary presentation that shows the elements and their relationships in the pattern. We use component and connector models to describe the pattern's units of execution. (3) A description of the architectural elements and their responsibilities. We illustrate the description with examples from the two cases introduced in Section 2. The architec- tural elements are specified in π-ADL. Some fragments are omitted for brevity. For a complete formal specification, we refer to the online Appendix A. (4) A rationale that motivates the design of the pattern. This pattern template follows the approach for documenting architectural styles presented in [Clements et al. 2010]. We discuss related patterns from the field of multi-agent systems in Section 6. 3.3 π-ADL in a Nutshell The patterns of the architectural style are specified in π-ADL, a formal architectural de- scription language. It is a theoretically well-founded language based on the higher-order typed π-calculus, supported by a specification and verification toolset providing services for automated checking of architectural properties. nent is an architectural abstraction that delivers some system functionalities. Formal specification improves the understanding of the specification, enables reasoning about the patterns, and provides a firm foundation for employing the patterns in practice. The patterns are type checked using the π-ADL.NET tool [Qayyum and Oquendo 2008]. In this section, we briefly explain the main language constructs of π-ADL. We limit the explanation to the subset of language constructs that we use to describe the patterns. • abstraction: the basic architectural element in π-ADL. Abstraction is a unit of execution. • component: the basic computational element that we use in the description. A compo- • connection: a basic interaction point. A connection provides a communication channel between architectural elements that allows value passing. Values are sent and received through connections via the out and in prefixes respectively. Value passing via a con- nection is synchronous. • port: an interface between a component and its environment. Ports are described in • unifies: expresses the linking of two connections. Unifies connects connections together. • relays: expresses the binding between external ports and ports of components. If con- nections have the same names in different ports, identifying ports is enough to express unifications. terms of connections. Protocols may be enforced by ports and among ports. ACM Transactions Journal Format, Vol. 2, No. 3, 09 2001. · D. Weyns and F. Oquendo 10 • type: a value type which can be declared either a basic type or a constructed type. A type is declared within the scope of a behavior. • Integer, Boolean, and String: regular basic types. • Any: an unspecified data type, represents any arbitrary combination of data types. We use any as an abstract data type that has to be made concrete when the patterns are applied to a concrete domain. • view: a collection of named elements, possibly of different types. • sequence: an indexed collection of elements of the same type. • set: an abstract collection of elements of a particular type. • location: an abstract storage of elements of a particular type. • behavior: an expression of the process of a component in terms of actions and interac- tions with its environment through connections. • function: a part of functionality within a behavior which performs a specific task. A function may take input of different types and produce output of a particular type. • compose: expresses the capability of a behavior to parallel compose the capabilities of multiple behaviors. The composed behaviors can proceed independently and can interact via attached connections. • choose: expresses the capability of a behavior to choose the capability of one of set of behaviors. If all the sub-behaviors are blocking on an input, the first one to resume execution will continue while the others will terminate. • unobservable: expresses the capability to enact an action invisibly, i.e. internally, silently. Non-observable behavior is internal to an abstraction. 4. ARCHITECTURAL STYLE We now present the patterns of the architectural style using the pattern template. We start with the two basic patterns: Virtual Environment and Situated Agent. Then we explain the three additional patterns that refine a situated agent: Selective Perception, Protocol-Based Communication, and Behavior-Based Action Selection. 4.1 Virtual Environment 4.1.1 Primary Presentation. The primary presentation of the virtual environment pattern is shown in Figure 4. The virtual environment provides a domain-specific infrastructure that supports the ac- tivities of situated agents (perception, communication, performing actions). In particular, it provides a virtualization of the external environment, possibly augmented with support for coordination.3 For example, the virtual environment may represent the physical context as a map and this map may be equipped with facilities for agents to add marks on particu- lar locations, supporting indirect coordination among the agents. The virtual environment 3Environment is an overloaded term in software engineering. The term is used for development environment, execution environment, physical environment, etc. In this report, we refer with environment (or external environ- ment) to the part of the external world (which may comprise both physical or software elements) with which the system interacts, and in which the effects of the system will be observed and evaluated [Jackson 1997]. ACM Transactions Journal Format, Vol. 2, No. 3, 09 2001. An Architectural Style for Self-Adaptive Multi-Agent Systems · 11 Fig. 4. Primary presentation of virtual environment requires middleware services for observing the external environment, transmitting mes- sages, and executing actions in the environment. At the top level, the virtual environment is defined as follows: component VirtualEnvironment is abstraction () { type Focus is view[focusName : String, focusParams : Any]. type Sense is view[agentId : String, focus : Focus]. type Representation is Any. type Action is view[agentId : String, actionName : String, actionParams : Any]. type Operation is Any. type Message is view[id : Integer, sender : String, receiver : String, protocol : String, performative : String, content : Any]. type Observation is Any. type Observed is Any. type StateItem is view[name : String, val : Any]. type StateItems is set(StateItem). type SynchronizationUpdate is Any. type Transmission is Any. port perceive is { connection sense is in(Sense). connection sensed is out(Representation). } port act is { connection action is in(Action) }. port communicate is { connection sendMsg is in(Message). port operate is { connection operation is out(Operation) }. connection receiveMsg is out(Message) }. ACM Transactions Journal Format, Vol. 2, No. 3, 09 2001. · 12 D. Weyns and F. Oquendo port synchronize is { connection syncOut is out(SynchronizationUpdate). connection syncIn is in(SynchronizationUpdate) }. port observe is { connection observation is out(Observation). connection observed is in(Observed) }. port transmit is { connection transmission is out(Transmission). connection deliver is in(Transmission) }. //composition behavior is compose { actionService is ActionService() perceptionservice is PerceptionService() communicationservice is CommunicationService() state is State() synchronization is Synchronization() { sense relays perceptionService::sense perceptionService::readWriteState unifies state::readWriteState perceptionService::sensed relays sensed and and and and } where and and ... } } } The π-ADL specification first defines the different types and ports. Some types include the Any data type. These types have to be made concrete when the pattern is applied to a concrete domain. Then, the composition of the various component instances is defined (for brevity, only the composition of the perception service is showed). The components are further refined and explained in the following subsection. 4.1.2 Architectural Elements. The State repository has a central role in the virtual environment. The repository con- tains data that is shared between the components of the virtual environment. The π-ADL specification of state is defined as follows: component State is abstraction() { type StateItem is view[name : String, val : Any]. type StateItems is set(StateItem). port readWriteState is { connection stateTemplate is in(StateItems). connection readState is out(StateItems). connection writeState is in(StateItems) }. assuming { protocol is (via stateTemplate receive StateItems. true.via readState send StateItems) or (via writeState receive StateItems)* }. stateRepository is location(set(StateItems)). behavior is { write is function( state : StateItems, repository : set(StateItems) ) : set(StateItems) { unobservable }. read is function( template : StateItems, repository : set(StateItems) ) : ACM Transactions Journal Format, Vol. 2, No. 3, 09 2001. An Architectural Style for Self-Adaptive Multi-Agent Systems · 13 StateItems { unobservable }. choose { //read state via stateTemplate receive state_template : StateItems. via readState send read(state_template, stateRepository). behavior(). //write state via writeState receive state_items : StateItems. stateRepository := write(state_items, stateRepository). behavior(). or } } } State items are abstract data that represent elements of interest maintained by the virtual environment. State items are stored in the state repository. The state repository provides a port that allows clients to read state based on a state template, and write state. The behavior specifies how reading and writing is executed. Reading requires a template of state items to select the state items of interest. The internal behavior of read and write are invisible and have to be made concrete for a particular domain. State stored in the state repository typically includes an abstract representation of ex- ternal resources and additional state that is used for coordination purposes. An example of state related to external resources in the vehicle routing system is a map of the traffic environment. The physical road network is mapped onto a graph representation of discrete road elements, i.e. road segments and crossroads, that are connected via edges. The road network is concretely realized as follows: public class Edge<RoadElement> { private int id; private RoadElement elem; ... } public class RoadNetwork<RoadElement> { private ArrayList<Edge<RoadElement>> edges; public RoadNetwork(){ edges = new ArrayList<Edge<RoadElement>>(); } ... } Edge has an id and a road element of the type RoadElement, corresponding to the name and value of a state item. The road network models a state repository that comprises a list of edges. Road network implements regular accessors and mutators to read and write state. For our case study, we used OpenStreepMap to generate the graph of the city of Leuven, which consists of over 1600 road elements as shown in Figure 5. In the vehicle routing system, additional state maintained by the state repository are the current position of the vehicle on the map, and road segments that are marked indicating that they are temporarily blocked. Similarly in the AGV transportation system, the state related to external resources in- ACM Transactions Journal Format, Vol. 2, No. 3, 09 2001. · 14 D. Weyns and F. Oquendo (a) Rendering of the OpenStreetMap data (b) Graph model based on OpenStreetMap data Fig. 5. Example of a graph model used in the vehicle routing system. The figure on the left shows a rendering of the OpenStreetMap data of the city center of Leuven. The figure on the right shows the graph with over 1600 roads. cludes a map of the warehouse that is modeled as a graph of segments. A segment repre- sents an atomic part of the warehouse infrastructure along which an AGV can be instructed to drive (typically 3 to 5 meter). Examples of additional state are virtual marks situated on the map. As an example, consider Fig. 6 that shows a fusion view4 of the AGV local virtual environments of three AGVs. Fig. 6. Path projections in the virtual environment. 4The view is generated on a remote machine that collects the state of the virtual environment of the AGVs via a wireless network and fuses the information into one image. ACM Transactions Journal Format, Vol. 2, No. 3, 09 2001. (cid:1)(cid:2)(cid:3)(cid:4)(cid:5)(cid:6)(cid:5)(cid:5)(cid:7)(cid:8)(cid:5)(cid:4)(cid:9)(cid:10)(cid:11)(cid:9)(cid:5)(cid:12)(cid:11)(cid:9)(cid:5)(cid:12)(cid:6)(cid:5)(cid:5)(cid:7)(cid:8)(cid:5)(cid:4)(cid:9)(cid:10)(cid:6)(cid:5)(cid:5)(cid:7)(cid:8)(cid:5)(cid:4)(cid:9)(cid:10)(cid:6)(cid:5)(cid:5)(cid:7)(cid:8)(cid:5)(cid:4)(cid:9)(cid:10)(cid:13)(cid:5)(cid:14)(cid:5)(cid:12)(cid:4)(cid:5)(cid:5)(cid:15)(cid:5)(cid:16)(cid:16)(cid:5)(cid:4)(cid:17)(cid:18)(cid:9)(cid:1)(cid:2)(cid:3)(cid:2)(cid:4)(cid:5)(cid:2)(cid:2)(cid:6)(cid:2)(cid:7)(cid:3)(cid:2)(cid:8)(cid:13)(cid:9)(cid:19)(cid:20)(cid:21)(cid:5)(cid:12)(cid:20)(cid:22)(cid:2)(cid:6)(cid:5)(cid:8)(cid:4)(cid:23)(cid:23)(cid:8)(cid:24)(cid:5)(cid:25)(cid:8)(cid:5)(cid:5)(cid:26)(cid:16)(cid:5)(cid:16)(cid:4)(cid:5)(cid:16)(cid:16)(cid:27)(cid:12)(cid:28)(cid:26)(cid:23)(cid:23)(cid:12)(cid:26)(cid:16)(cid:16)(cid:8)(cid:5)(cid:6)(cid:5)(cid:4)(cid:4)(cid:9)(cid:10)(cid:29)(cid:5)(cid:4)(cid:4)(cid:5)(cid:6)(cid:5)(cid:8)(cid:16)(cid:16)(cid:26)(cid:12)(cid:23)(cid:23)(cid:26)(cid:26)(cid:23)(cid:23)(cid:12)(cid:26)(cid:16)(cid:26)(cid:5)(cid:2)(cid:4)(cid:11)(cid:26)(cid:23)(cid:23)(cid:12)(cid:26)(cid:16)(cid:24)(cid:12)(cid:5)(cid:7)(cid:5)(cid:24)(cid:4)(cid:23)(cid:30)(cid:31)(cid:12)(cid:5)(cid:32)(cid:2)(cid:6)(cid:33)(cid:5)(cid:12)(cid:5)(cid:8)(cid:7)(cid:5)(cid:12)(cid:24)(cid:34)(cid:35)(cid:23)(cid:35)(cid:2)(cid:23)(cid:4)(cid:23)(cid:23)(cid:8)(cid:8)(cid:23)(cid:23)(cid:4)(cid:5)(cid:6)(cid:5)(cid:28)(cid:20)(cid:12)(cid:5)(cid:21)(cid:20)(cid:19)(cid:9)(cid:13)(cid:36)(cid:2)(cid:4)(cid:32)(cid:5)(cid:8)(cid:33)(cid:9)(cid:24)(cid:5)(cid:26)(cid:23)(cid:23)(cid:12)(cid:26)(cid:16)(cid:24)(cid:12)(cid:5)(cid:7)(cid:5)(cid:24)(cid:4)(cid:23)(cid:30)(cid:37)(cid:28)(cid:15)(cid:20)(cid:24)(cid:8)(cid:2)(cid:6)(cid:12)(cid:23)(cid:31)(cid:28)(cid:2)(cid:5)(cid:16)(cid:26)(cid:16)(cid:26)(cid:12)(cid:23)(cid:23)(cid:26)(cid:26)(cid:12)(cid:23)(cid:23)(cid:26)(cid:16)(cid:5)(cid:24)(cid:8)(cid:23)(cid:18)(cid:26)(cid:8)(cid:5)(cid:12)(cid:5)(cid:13)(cid:18)(cid:5)(cid:27)(cid:14)(cid:5)(cid:8)(cid:24)(cid:5)(cid:25)(cid:8)(cid:5)(cid:5)(cid:26)(cid:16)(cid:5)(cid:16)(cid:12)(cid:27)(cid:27)(cid:14)(cid:12)(cid:5)(cid:21)(cid:38)(cid:27)(cid:32)(cid:5)(cid:20)(cid:39)(cid:2)(cid:3)(cid:14)(cid:5)(cid:4)(cid:16)(cid:5)(cid:20)(cid:25)(cid:5)(cid:24)(cid:31)(cid:9)(cid:40)(cid:27)(cid:4)(cid:2)(cid:5)(cid:12)(cid:5)(cid:8)(cid:4)(cid:23)(cid:23)(cid:8)(cid:19)(cid:5)(cid:5)(cid:12)(cid:32)(cid:4)(cid:5)(cid:40)(cid:23)(cid:15)(cid:24)(cid:5)(cid:25)(cid:4)(cid:23)(cid:23)(cid:32)(cid:8)(cid:5)(cid:40)(cid:5)(cid:2)(cid:1)(cid:8)(cid:23)(cid:23)(cid:4)(cid:26)(cid:16)(cid:4)(cid:22)(cid:20)(cid:12)(cid:5)(cid:21)(cid:30)(cid:9)(cid:16)(cid:25)(cid:2)(cid:8)(cid:20)(cid:41)(cid:41)(cid:41)(cid:20)(cid:4)(cid:23)(cid:23)(cid:8)(cid:42)(cid:5)(cid:12)(cid:32)(cid:2)(cid:8)(cid:23)(cid:8)(cid:32)(cid:20)(cid:11)(cid:5)(cid:12)(cid:7)(cid:2)(cid:5)(cid:16)(cid:26)(cid:4)(cid:23)(cid:23)(cid:8)(cid:9)(cid:10)(cid:11)(cid:12)(cid:26)(cid:23)(cid:23)(cid:12)(cid:26)(cid:16)(cid:8)(cid:5)(cid:24)(cid:5)(cid:25)(cid:5)(cid:8)(cid:2)(cid:5)(cid:4)(cid:6)(cid:5)(cid:5)(cid:25)(cid:21)(cid:31)(cid:12)(cid:2)(cid:9)(cid:12)(cid:2)(cid:3)(cid:4)(cid:23)(cid:23)(cid:8)(cid:38)(cid:27)(cid:32)(cid:5)(cid:20)(cid:39)(cid:2)(cid:3)(cid:14)(cid:5)(cid:4)(cid:16)(cid:5)(cid:25)(cid:5)(cid:24)(cid:15)(cid:5)(cid:2)(cid:7)(cid:5)(cid:12)(cid:24)(cid:16)(cid:26)(cid:12)(cid:23)(cid:23)(cid:26)(cid:43)(cid:9)(cid:16)(cid:20)(cid:1)(cid:5)(cid:20)(cid:13)(cid:23)(cid:5)(cid:16)(cid:4)(cid:23)(cid:23)(cid:8)(cid:13)(cid:5)(cid:8)(cid:12)(cid:2)(cid:20)(cid:30)(cid:5)(cid:5)(cid:12)(cid:23)(cid:5)(cid:12)(cid:26)(cid:4)(cid:23)(cid:23)(cid:8)(cid:19)(cid:5)(cid:5)(cid:12)(cid:32)(cid:4)(cid:5)(cid:40)(cid:23)(cid:15)(cid:24)(cid:5)(cid:25)(cid:8)(cid:5)(cid:5)(cid:26)(cid:16)(cid:5)(cid:16)(cid:12)(cid:27)(cid:27)(cid:14)(cid:12)(cid:5)(cid:21)(cid:43)(cid:5)(cid:19)(cid:20)(cid:11)(cid:23)(cid:8)(cid:20)(cid:13)(cid:9)(cid:9)(cid:19)(cid:4)(cid:23)(cid:23)(cid:8)(cid:8)(cid:23)(cid:23)(cid:4)(cid:8)(cid:5)(cid:8)(cid:3)(cid:2)(cid:26)(cid:16)(cid:5)(cid:4)(cid:5)(cid:37)(cid:43)(cid:9)(cid:9)(cid:16)(cid:20)(cid:42)(cid:4)(cid:9)(cid:12)(cid:44)(cid:27)(cid:2)(cid:8)(cid:4)(cid:23)(cid:23)(cid:8)(cid:26)(cid:23)(cid:23)(cid:12)(cid:26)(cid:16)(cid:26)(cid:5)(cid:2)(cid:4)(cid:11)(cid:19)(cid:5)(cid:5)(cid:12)(cid:32)(cid:4)(cid:5)(cid:40)(cid:23)(cid:15)(cid:15)(cid:23)(cid:40)(cid:5)(cid:4)(cid:32)(cid:12)(cid:5)(cid:5)(cid:19)(cid:9)(cid:10)(cid:11)(cid:12)(cid:22)(cid:24)(cid:5)(cid:8)(cid:33)(cid:9)(cid:14)(cid:5)(cid:8)(cid:25)(cid:5)(cid:24)(cid:29)(cid:26)(cid:5)(cid:5)(cid:8)(cid:24)(cid:12)(cid:9)(cid:5)(cid:14)(cid:5)(cid:8)(cid:4)(cid:23)(cid:23)(cid:8)(cid:26)(cid:23)(cid:23)(cid:12)(cid:26)(cid:16)(cid:8)(cid:5)(cid:4)(cid:9)(cid:45)(cid:8)(cid:5)(cid:12)(cid:5)(cid:46)(cid:43)(cid:41)(cid:37)(cid:28)(cid:15)(cid:20)(cid:24)(cid:8)(cid:2)(cid:6)(cid:12)(cid:23)(cid:31)(cid:8)(cid:23)(cid:23)(cid:4)(cid:8)(cid:3)(cid:2)(cid:25)(cid:5)(cid:32)(cid:27)(cid:9)(cid:28)(cid:20)(cid:24)(cid:8)(cid:2)(cid:8)(cid:9)(cid:15)(cid:8)(cid:23)(cid:23)(cid:4)(cid:8)(cid:5)(cid:8)(cid:3)(cid:2)(cid:26)(cid:16)(cid:5)(cid:4)(cid:5)(cid:37)(cid:15)(cid:9)(cid:8)(cid:2)(cid:8)(cid:24)(cid:20)(cid:28)(cid:9)(cid:27)(cid:32)(cid:5)(cid:25)(cid:2)(cid:3)(cid:8)(cid:4)(cid:23)(cid:23)(cid:8)(cid:30)(cid:23)(cid:4)(cid:24)(cid:5)(cid:7)(cid:5)(cid:12)(cid:24)(cid:16)(cid:26)(cid:12)(cid:23)(cid:23)(cid:26)(cid:21)(cid:5)(cid:12)(cid:14)(cid:27)(cid:27)(cid:12)(cid:16)(cid:5)(cid:16)(cid:26)(cid:5)(cid:5)(cid:8)(cid:25)(cid:5)(cid:24)(cid:8)(cid:23)(cid:23)(cid:4)(cid:8)(cid:5)(cid:8)(cid:3)(cid:2)(cid:26)(cid:16)(cid:5)(cid:4)(cid:5)(cid:37)(cid:8)(cid:23)(cid:23)(cid:4)(cid:8)(cid:5)(cid:8)(cid:3)(cid:2)(cid:26)(cid:16)(cid:5)(cid:4)(cid:5)(cid:37)(cid:29)(cid:26)(cid:5)(cid:5)(cid:8)(cid:24)(cid:12)(cid:9)(cid:5)(cid:14)(cid:5)(cid:8)(cid:4)(cid:23)(cid:23)(cid:8)(cid:10)(cid:9)(cid:8)(cid:16)(cid:5)(cid:2)(cid:24)(cid:8)(cid:5)(cid:27)(cid:12)(cid:20)(cid:11)(cid:23)(cid:8)(cid:20)(cid:36)(cid:23)(cid:5)(cid:47)(cid:5)(cid:8)(cid:7)(cid:5)(cid:12)(cid:24)(cid:33)(cid:4)(cid:23)(cid:23)(cid:8)(cid:26)(cid:23)(cid:23)(cid:12)(cid:26)(cid:16)(cid:5)(cid:12)(cid:5)(cid:13)(cid:15)(cid:5)(cid:12)(cid:16)(cid:5)(cid:4)(cid:23)(cid:12)(cid:5)(cid:8)(cid:25)(cid:5)(cid:24)(cid:13)(cid:5)(cid:12)(cid:5)(cid:16)(cid:26)(cid:12)(cid:23)(cid:23)(cid:26)(cid:4)(cid:5)(cid:24)(cid:8)(cid:2)(cid:16)(cid:32)(cid:2)(cid:5)(cid:33)(cid:35)(cid:16)(cid:8)(cid:5)(cid:32)(cid:27)(cid:18)(cid:26)(cid:23)(cid:23)(cid:12)(cid:26)(cid:16)(cid:5)(cid:12)(cid:5)(cid:13)(cid:26)(cid:23)(cid:23)(cid:12)(cid:26)(cid:16)(cid:26)(cid:16)(cid:5)(cid:2)(cid:28)(cid:26)(cid:16)(cid:5)(cid:14)(cid:47)(cid:45)(cid:5)(cid:48)(cid:20)(cid:32)(cid:12)(cid:23)(cid:27)(cid:32)(cid:22)(cid:30)(cid:12)(cid:9)(cid:5)(cid:8)(cid:14)(cid:5)(cid:4)(cid:32)(cid:16)(cid:26)(cid:12)(cid:23)(cid:23)(cid:26)(cid:4)(cid:5)(cid:24)(cid:8)(cid:2)(cid:16)(cid:32)(cid:2)(cid:5)(cid:33)(cid:35)(cid:16)(cid:8)(cid:5)(cid:32)(cid:27)(cid:18)(cid:13)(cid:14)(cid:10)(cid:15)(cid:23)(cid:7)(cid:9)(cid:27)(cid:26)(cid:5)(cid:12)(cid:45)(cid:23)(cid:8)(cid:16)(cid:16)(cid:26)(cid:12)(cid:23)(cid:23)(cid:26)(cid:30)(cid:12)(cid:9)(cid:5)(cid:19)(cid:16)(cid:26)(cid:12)(cid:23)(cid:23)(cid:26)(cid:26)(cid:23)(cid:23)(cid:12)(cid:26)(cid:16)(cid:5)(cid:4)(cid:4)(cid:9)(cid:13)(cid:32)(cid:12)(cid:23)(cid:23)(cid:24)(cid:8)(cid:3)(cid:2)(cid:36)(cid:20)(cid:8)(cid:5)(cid:21)(cid:21)(cid:5)(cid:12)(cid:14)(cid:27)(cid:27)(cid:12)(cid:16)(cid:5)(cid:14)(cid:5)(cid:16)(cid:26)(cid:32)(cid:12)(cid:23)(cid:23)(cid:24)(cid:8)(cid:3)(cid:2)(cid:36)(cid:20)(cid:8)(cid:5)(cid:21)(cid:26)(cid:16)(cid:5)(cid:14)(cid:5)(cid:16)(cid:12)(cid:27)(cid:27)(cid:14)(cid:12)(cid:5)(cid:21)(cid:26)(cid:23)(cid:23)(cid:12)(cid:26)(cid:16)(cid:5)(cid:16)(cid:12)(cid:27)(cid:27)(cid:14)(cid:12)(cid:5)(cid:21)(cid:21)(cid:9)(cid:4)(cid:33)(cid:27)(cid:2)(cid:46)(cid:5)(cid:8)(cid:16)(cid:26)(cid:12)(cid:23)(cid:23)(cid:26)(cid:26)(cid:23)(cid:23)(cid:12)(cid:26)(cid:16)(cid:5)(cid:16)(cid:4)(cid:5)(cid:16)(cid:16)(cid:27)(cid:12)(cid:28)(cid:28)(cid:23)(cid:8)(cid:6)(cid:16)(cid:26)(cid:12)(cid:23)(cid:23)(cid:26)(cid:10)(cid:5)(cid:35)(cid:33)(cid:5)(cid:4)(cid:16)(cid:5)(cid:14)(cid:5)(cid:16)(cid:26)(cid:21)(cid:5)(cid:8)(cid:20)(cid:36)(cid:2)(cid:3)(cid:8)(cid:24)(cid:23)(cid:23)(cid:12)(cid:32)(cid:26)(cid:23)(cid:23)(cid:12)(cid:26)(cid:16)(cid:16)(cid:12)(cid:5)(cid:40)(cid:8)(cid:3)(cid:2)(cid:36)(cid:8)(cid:23)(cid:23)(cid:4)(cid:8)(cid:5)(cid:24)(cid:9)(cid:26)(cid:12)(cid:5)(cid:13)(cid:17)(cid:16)(cid:49)(cid:8)(cid:23)(cid:23)(cid:4)(cid:12)(cid:5)(cid:2)(cid:35)(cid:12)(cid:5)(cid:10)(cid:20)(cid:4)(cid:23)(cid:23)(cid:8)(cid:2)(cid:32)(cid:12)(cid:23)(cid:15)(cid:43)(cid:27)(cid:4)(cid:5)(cid:16)(cid:20)(cid:11)(cid:23)(cid:8)(cid:32)(cid:5)(cid:8)(cid:7)(cid:5)(cid:45)(cid:40)(cid:26)(cid:4)(cid:23)(cid:23)(cid:8)(cid:39)(cid:9)(cid:9)(cid:12)(cid:45)(cid:23)(cid:8)(cid:8)(cid:5)(cid:8)(cid:16)(cid:26)(cid:12)(cid:23)(cid:23)(cid:26)(cid:26)(cid:23)(cid:23)(cid:12)(cid:26)(cid:16)(cid:5)(cid:4)(cid:4)(cid:9)(cid:13)(cid:8)(cid:23)(cid:23)(cid:4)(cid:5)(cid:32)(cid:8)(cid:9)(cid:48)(cid:20)(cid:5)(cid:32)(cid:27)(cid:38)(cid:15)(cid:14)(cid:16)(cid:32)(cid:4)(cid:5)(cid:14)(cid:4)(cid:5)(cid:5)(cid:12)(cid:23)(cid:15)(cid:10)(cid:5)(cid:35)(cid:33)(cid:5)(cid:4)(cid:16)(cid:5)(cid:14)(cid:5)(cid:16)(cid:26)(cid:15)(cid:14)(cid:16)(cid:39)(cid:2)(cid:5)(cid:27)(cid:25)(cid:5)(cid:20)(cid:10)(cid:5)(cid:35)(cid:33)(cid:5)(cid:4)(cid:16)(cid:5)(cid:16)(cid:26)(cid:5)(cid:5)(cid:8)(cid:25)(cid:5)(cid:24)(cid:26)(cid:23)(cid:23)(cid:12)(cid:26)(cid:16)(cid:26)(cid:16)(cid:5)(cid:5)(cid:30)(cid:20)(cid:5)(cid:24)(cid:2)(cid:4)(cid:2)(cid:5)(cid:13)(cid:21)(cid:5)(cid:12)(cid:14)(cid:27)(cid:27)(cid:12)(cid:16)(cid:5)(cid:14)(cid:5)(cid:16)(cid:26)(cid:11)(cid:9)(cid:5)(cid:12)(cid:14)(cid:2)(cid:23)(cid:32)(cid:27)(cid:35)(cid:26)(cid:26)(cid:16)(cid:5)(cid:14)(cid:5)(cid:16)(cid:12)(cid:27)(cid:27)(cid:14)(cid:12)(cid:5)(cid:21)(cid:24)(cid:5)(cid:25)(cid:8)(cid:5)(cid:5)(cid:26)(cid:16)(cid:5)(cid:16)(cid:4)(cid:5)(cid:33)(cid:35)(cid:5)(cid:10)(cid:26)(cid:23)(cid:23)(cid:12)(cid:26)(cid:16)(cid:8)(cid:5)(cid:12)(cid:5)(cid:33)(cid:6)(cid:2)(cid:32)(cid:5)(cid:12)(cid:31)(cid:20)(cid:5)(cid:16)(cid:12)(cid:5)(cid:41)(cid:19)(cid:9)(cid:33)(cid:16)(cid:27)(cid:45)(cid:16)(cid:23)(cid:12)(cid:30)(cid:26)(cid:23)(cid:23)(cid:12)(cid:26)(cid:16)(cid:45)(cid:5)(cid:9)(cid:4)(cid:7)(cid:16)(cid:32)(cid:27)(cid:9)(cid:30)(cid:26)(cid:16)(cid:5)(cid:14)(cid:5)(cid:16)(cid:12)(cid:27)(cid:27)(cid:14)(cid:12)(cid:5)(cid:21)(cid:28)(cid:12)(cid:9)(cid:27)(cid:25)(cid:5)(cid:12)(cid:16)(cid:16)(cid:26)(cid:12)(cid:23)(cid:23)(cid:26)(cid:8)(cid:2)(cid:5)(cid:4)(cid:40)(cid:16)(cid:7)(cid:9)(cid:35)(cid:23)(cid:43)(cid:17)(cid:26)(cid:8)(cid:2)(cid:29)(cid:26)(cid:23)(cid:23)(cid:12)(cid:26)(cid:16)(cid:5)(cid:16)(cid:4)(cid:5)(cid:16)(cid:16)(cid:27)(cid:12)(cid:28)(cid:48)(cid:2)(cid:32)(cid:32)(cid:5)(cid:12)(cid:16)(cid:26)(cid:12)(cid:23)(cid:23)(cid:26)(cid:26)(cid:23)(cid:23)(cid:12)(cid:26)(cid:16)(cid:16)(cid:2)(cid:27)(cid:12)(cid:15)(cid:21)(cid:5)(cid:12)(cid:14)(cid:27)(cid:27)(cid:12)(cid:16)(cid:5)(cid:14)(cid:5)(cid:16)(cid:26)(cid:12)(cid:5)(cid:9)(cid:14)(cid:8)(cid:5)(cid:8)(cid:3)(cid:2)(cid:35)(cid:27)(cid:40)(cid:23)(cid:15)(cid:15)(cid:23)(cid:40)(cid:27)(cid:35)(cid:2)(cid:3)(cid:8)(cid:5)(cid:8)(cid:14)(cid:9)(cid:5)(cid:12)(cid:43)(cid:23)(cid:8)(cid:16)(cid:5)(cid:8)(cid:2)(cid:27)(cid:16)(cid:16)(cid:26)(cid:12)(cid:23)(cid:23)(cid:26)(cid:26)(cid:23)(cid:23)(cid:12)(cid:26)(cid:16)(cid:24)(cid:12)(cid:5)(cid:7)(cid:8)(cid:5)(cid:6)(cid:2)(cid:22)(cid:26)(cid:23)(cid:23)(cid:12)(cid:26)(cid:16)(cid:24)(cid:9)(cid:26)(cid:12)(cid:5)(cid:13)(cid:26)(cid:16)(cid:5)(cid:14)(cid:5)(cid:16)(cid:12)(cid:27)(cid:27)(cid:14)(cid:12)(cid:5)(cid:21)(cid:26)(cid:23)(cid:23)(cid:12)(cid:26)(cid:16)(cid:12)(cid:5)(cid:6)(cid:8)(cid:9)(cid:1)(cid:15)(cid:23)(cid:12)(cid:5)(cid:5)(cid:4)(cid:14)(cid:5)(cid:4)(cid:32)(cid:48)(cid:5)(cid:32)(cid:2)(cid:8)(cid:24)(cid:5)(cid:8)(cid:16)(cid:26)(cid:12)(cid:23)(cid:23)(cid:26)(cid:49)(cid:16)(cid:17)(cid:13)(cid:5)(cid:12)(cid:26)(cid:9)(cid:24)(cid:5)(cid:8)(cid:25)(cid:2)(cid:3)(cid:8)(cid:24)(cid:23)(cid:23)(cid:12)(cid:32)(cid:48)(cid:9)(cid:5)(cid:16)(cid:5)(cid:4)(cid:7)(cid:5)(cid:12)(cid:24)(cid:32)(cid:23)(cid:4)(cid:15)(cid:23)(cid:40)(cid:27)(cid:35)(cid:2)(cid:3)(cid:8)(cid:5)(cid:8)(cid:14)(cid:9)(cid:5)(cid:12)(cid:48)(cid:2)(cid:32)(cid:32)(cid:5)(cid:12)(cid:16)(cid:26)(cid:12)(cid:23)(cid:23)(cid:26)(cid:15)(cid:5)(cid:2)(cid:46)(cid:5)(cid:12)(cid:16)(cid:7)(cid:5)(cid:12)(cid:24)(cid:11)(cid:12)(cid:2)(cid:5)(cid:16)(cid:5)(cid:8)(cid:33)(cid:9)(cid:19)(cid:38)(cid:27)(cid:32)(cid:5)(cid:20)(cid:48)(cid:9)(cid:8)(cid:32)(cid:5)(cid:4)(cid:23)(cid:23)(cid:8)(cid:26)(cid:23)(cid:23)(cid:12)(cid:26)(cid:16)(cid:8)(cid:5)(cid:8)(cid:5)(cid:2)(cid:26)(cid:8)(cid:5)(cid:26)(cid:2)(cid:8)(cid:5)(cid:31)(cid:26)(cid:23)(cid:23)(cid:12)(cid:26)(cid:16)(cid:12)(cid:5)(cid:32)(cid:32)(cid:2)(cid:48)(cid:48)(cid:5)(cid:45)(cid:2)(cid:20)(cid:11)(cid:23)(cid:8)(cid:32)(cid:5)(cid:12)(cid:14)(cid:23)(cid:5)(cid:12)(cid:5)(cid:8)(cid:4)(cid:23)(cid:23)(cid:8)(cid:42)(cid:9)(cid:8)(cid:26)(cid:5)(cid:2)(cid:8)(cid:16)(cid:26)(cid:12)(cid:23)(cid:23)(cid:26)(cid:26)(cid:23)(cid:23)(cid:12)(cid:26)(cid:16)(cid:45)(cid:9)(cid:9)(cid:7)(cid:5)(cid:12)(cid:5)(cid:31)(cid:42)(cid:9)(cid:8)(cid:26)(cid:5)(cid:2)(cid:8)(cid:16)(cid:26)(cid:12)(cid:23)(cid:23)(cid:26)(cid:8)(cid:23)(cid:23)(cid:7)(cid:5)(cid:16)(cid:12)(cid:5)(cid:14)(cid:23)(cid:36)(cid:21)(cid:5)(cid:16)(cid:16)(cid:5)(cid:8)(cid:16)(cid:26)(cid:12)(cid:23)(cid:23)(cid:26)(cid:50)(cid:20)(cid:42)(cid:9)(cid:8)(cid:26)(cid:5)(cid:2)(cid:8)(cid:16)(cid:26)(cid:12)(cid:23)(cid:23)(cid:26)(cid:8)(cid:23)(cid:23)(cid:4)(cid:8)(cid:5)(cid:4)(cid:9)(cid:45)(cid:4)(cid:9)(cid:11)(cid:28)(cid:12)(cid:27)(cid:16)(cid:16)(cid:5)(cid:4)(cid:16)(cid:5)(cid:16)(cid:26)(cid:12)(cid:23)(cid:23)(cid:26)(cid:48)(cid:5)(cid:32)(cid:2)(cid:8)(cid:24)(cid:5)(cid:8)(cid:16)(cid:26)(cid:12)(cid:23)(cid:23)(cid:26)(cid:21)(cid:5)(cid:12)(cid:14)(cid:27)(cid:27)(cid:12)(cid:16)(cid:5)(cid:14)(cid:5)(cid:16)(cid:26)(cid:15)(cid:5)(cid:2)(cid:46)(cid:5)(cid:12)(cid:16)(cid:7)(cid:5)(cid:12)(cid:24)(cid:28)(cid:12)(cid:9)(cid:27)(cid:25)(cid:5)(cid:12)(cid:16)(cid:16)(cid:26)(cid:12)(cid:23)(cid:23)(cid:26)(cid:31)(cid:2)(cid:5)(cid:26)(cid:5)(cid:12)(cid:20)(cid:37)(cid:9)(cid:27)(cid:26)(cid:5)(cid:12)(cid:5)(cid:5)(cid:4)(cid:16)(cid:26)(cid:12)(cid:23)(cid:23)(cid:26)(cid:48)(cid:2)(cid:51)(cid:14)(cid:9)(cid:8)(cid:32)(cid:5)(cid:12)(cid:5)(cid:8)(cid:32)(cid:2)(cid:3)(cid:6)(cid:18)(cid:5)(cid:9)(cid:40)(cid:9)(cid:4)(cid:32)(cid:20)(cid:1)(cid:5)(cid:35)(cid:9)(cid:27)(cid:51)(cid:4)(cid:23)(cid:23)(cid:8)(cid:48)(cid:5)(cid:32)(cid:2)(cid:8)(cid:24)(cid:5)(cid:8)(cid:33)(cid:9)(cid:19)(cid:26)(cid:23)(cid:23)(cid:12)(cid:26)(cid:16)(cid:5)(cid:16)(cid:4)(cid:5)(cid:33)(cid:35)(cid:5)(cid:10)(cid:26)(cid:23)(cid:23)(cid:12)(cid:26)(cid:16)(cid:16)(cid:27)(cid:2)(cid:8)(cid:5)(cid:16)(cid:8)(cid:23)(cid:43)(cid:8)(cid:23)(cid:23)(cid:4)(cid:26)(cid:33)(cid:35)(cid:2)(cid:46)(cid:5)(cid:24)(cid:12)(cid:5)(cid:11)(cid:13)(cid:23)(cid:8)(cid:32)(cid:7)(cid:9)(cid:9)(cid:24)(cid:33)(cid:9)(cid:19)(cid:31)(cid:9)(cid:45)(cid:40)(cid:16)(cid:26)(cid:12)(cid:23)(cid:23)(cid:26)(cid:4)(cid:5)(cid:24)(cid:8)(cid:2)(cid:16)(cid:32)(cid:2)(cid:5)(cid:33)(cid:35)(cid:16)(cid:8)(cid:5)(cid:32)(cid:27)(cid:18)(cid:18)(cid:27)(cid:32)(cid:5)(cid:8)(cid:16)(cid:35)(cid:33)(cid:5)(cid:2)(cid:32)(cid:16)(cid:2)(cid:8)(cid:24)(cid:5)(cid:4)(cid:1)(cid:12)(cid:5)(cid:5)(cid:19)(cid:16)(cid:26)(cid:12)(cid:23)(cid:23)(cid:26)(cid:10)(cid:2)(cid:8)(cid:32)(cid:5)(cid:12)(cid:7)(cid:12)(cid:9)(cid:5)(cid:32)(cid:5)(cid:12)(cid:16)(cid:16)(cid:26)(cid:12)(cid:23)(cid:23)(cid:26)(cid:38)(cid:8)(cid:46)(cid:5)(cid:17)(cid:18)(cid:2)(cid:5)(cid:14)(cid:5)(cid:17)(cid:11)(cid:12)(cid:9)(cid:27)(cid:25)(cid:16)(cid:26)(cid:12)(cid:23)(cid:23)(cid:26)(cid:24)(cid:5)(cid:25)(cid:8)(cid:5)(cid:5)(cid:26)(cid:16)(cid:5)(cid:16)(cid:45)(cid:23)(cid:23)(cid:39)(cid:11)(cid:23)(cid:4)(cid:4)(cid:5)(cid:2)(cid:16)(cid:26)(cid:12)(cid:23)(cid:23)(cid:26)(cid:10)(cid:5)(cid:35)(cid:33)(cid:5)(cid:4)(cid:16)(cid:5)(cid:16)(cid:26)(cid:12)(cid:23)(cid:23)(cid:26)(cid:26)(cid:23)(cid:23)(cid:12)(cid:26)(cid:16)(cid:26)(cid:33)(cid:35)(cid:12)(cid:27)(cid:28)(cid:26)(cid:23)(cid:23)(cid:12)(cid:26)(cid:16)(cid:8)(cid:5)(cid:9)(cid:12)(cid:30)(cid:2)(cid:5)(cid:18)(cid:2)(cid:5)(cid:18)(cid:34)(cid:45)(cid:5)(cid:12)(cid:2)(cid:6)(cid:23)(cid:4)(cid:23)(cid:23)(cid:8)(cid:13)(cid:9)(cid:14)(cid:5)(cid:8)(cid:2)(cid:5)(cid:12)(cid:16)(cid:32)(cid:12)(cid:5)(cid:5)(cid:19)(cid:13)(cid:23)(cid:4)(cid:14)(cid:5)(cid:16)(cid:26)(cid:12)(cid:23)(cid:23)(cid:26)(cid:2)(cid:5)(cid:18)(cid:29)(cid:40)(cid:9)(cid:9)(cid:12)(cid:25)(cid:5)(cid:24)(cid:16)(cid:26)(cid:12)(cid:23)(cid:23)(cid:26)(cid:10)(cid:2)(cid:32)(cid:32)(cid:5)(cid:8)(cid:16)(cid:26)(cid:12)(cid:23)(cid:23)(cid:26)(cid:8)(cid:23)(cid:23)(cid:4)(cid:16)(cid:26)(cid:27)(cid:9)(cid:28)(cid:20)(cid:6)(cid:12)(cid:2)(cid:1)(cid:22)(cid:24)(cid:5)(cid:4)(cid:23)(cid:8)(cid:26)(cid:2)(cid:5)(cid:12)(cid:4)(cid:23)(cid:23)(cid:8)(cid:18)(cid:5)(cid:2)(cid:8)(cid:23)(cid:23)(cid:4)(cid:4)(cid:2)(cid:12)(cid:40)(cid:23)(cid:8)(cid:5)(cid:2)(cid:26)(cid:26)(cid:33)(cid:35)(cid:34)(cid:34)(cid:4)(cid:7)(cid:5)(cid:12)(cid:26)(cid:20)(cid:36)(cid:9)(cid:27)(cid:26)(cid:5)(cid:12)(cid:16)(cid:16)(cid:26)(cid:12)(cid:23)(cid:23)(cid:26)(cid:26)(cid:23)(cid:23)(cid:12)(cid:26)(cid:16)(cid:16)(cid:3)(cid:2)(cid:12)(cid:23)(cid:31)(cid:26)(cid:23)(cid:23)(cid:12)(cid:26)(cid:16)(cid:8)(cid:5)(cid:14)(cid:9)(cid:28)(cid:29)(cid:26)(cid:23)(cid:32)(cid:16)(cid:14)(cid:5)(cid:16)(cid:26)(cid:26)(cid:23)(cid:23)(cid:12)(cid:26)(cid:16)(cid:8)(cid:5)(cid:40)(cid:23)(cid:33)(cid:35)(cid:29)(cid:26)(cid:23)(cid:23)(cid:12)(cid:26)(cid:16)(cid:8)(cid:5)(cid:40)(cid:23)(cid:33)(cid:35)(cid:29)(cid:37)(cid:12)(cid:23)(cid:5)(cid:8)(cid:5)(cid:8)(cid:32)(cid:9)(cid:8)(cid:35)(cid:6)(cid:2)(cid:5)(cid:18)(cid:29)(cid:35)(cid:33)(cid:23)(cid:40)(cid:5)(cid:8)(cid:16)(cid:26)(cid:12)(cid:23)(cid:23)(cid:26)(cid:36)(cid:9)(cid:4)(cid:14)(cid:5)(cid:8)(cid:40)(cid:9)(cid:9)(cid:12)(cid:26)(cid:24)(cid:23)(cid:8)(cid:24)(cid:26)(cid:23)(cid:23)(cid:12)(cid:26)(cid:16)(cid:5)(cid:16)(cid:4)(cid:5)(cid:33)(cid:35)(cid:5)(cid:10)(cid:48)(cid:23)(cid:45)(cid:7)(cid:5)(cid:12)(cid:24)(cid:8)(cid:23)(cid:23)(cid:4)(cid:5)(cid:4)(cid:4)(cid:5)(cid:40)(cid:23)(cid:15)(cid:10)(cid:5)(cid:35)(cid:33)(cid:5)(cid:4)(cid:16)(cid:5)(cid:16)(cid:26)(cid:12)(cid:23)(cid:23)(cid:26)(cid:13)(cid:14)(cid:10)(cid:22)(cid:12)(cid:23)(cid:16)(cid:45)(cid:27)(cid:16)(cid:20)(cid:48)(cid:27)(cid:5)(cid:4)(cid:5)(cid:8)(cid:16)(cid:14)(cid:5)(cid:16)(cid:26)(cid:26)(cid:23)(cid:23)(cid:12)(cid:26)(cid:16)(cid:2)(cid:5)(cid:45)(cid:19)(cid:4)(cid:23)(cid:23)(cid:25)(cid:21)(cid:28)(cid:12)(cid:27)(cid:16)(cid:16)(cid:5)(cid:4)(cid:16)(cid:5)(cid:16)(cid:26)(cid:12)(cid:23)(cid:23)(cid:26)(cid:26)(cid:16)(cid:5)(cid:14)(cid:5)(cid:16)(cid:45)(cid:23)(cid:23)(cid:39)(cid:39)(cid:23)(cid:23)(cid:45)(cid:16)(cid:5)(cid:16)(cid:26)(cid:12)(cid:23)(cid:23)(cid:26)(cid:28)(cid:27)(cid:12)(cid:35)(cid:33)(cid:26)(cid:16)(cid:26)(cid:12)(cid:23)(cid:23)(cid:26)(cid:15)(cid:4)(cid:5)(cid:2)(cid:8)(cid:20)(cid:28)(cid:5)(cid:24)(cid:2)(cid:3)(cid:8)(cid:33)(cid:9)(cid:19)(cid:31)(cid:23)(cid:8)(cid:9)(cid:12)(cid:23)(cid:45)(cid:23)(cid:4)(cid:23)(cid:23)(cid:8)(cid:19)(cid:9)(cid:33)(cid:16)(cid:8)(cid:5)(cid:12)(cid:9)(cid:33)(cid:35)(cid:29)(cid:8)(cid:23)(cid:23)(cid:4)(cid:6)(cid:35)(cid:3)(cid:2)(cid:48)(cid:20)(cid:5)(cid:1)(cid:20)(cid:16)(cid:23)(cid:4)(cid:16)(cid:2)(cid:8)(cid:23)(cid:26)(cid:29)(cid:26)(cid:23)(cid:23)(cid:12)(cid:26)(cid:16)(cid:5)(cid:16)(cid:4)(cid:5)(cid:33)(cid:35)(cid:5)(cid:10)(cid:13)(cid:5)(cid:8)(cid:32)(cid:12)(cid:2)(cid:6)(cid:20)(cid:37)(cid:9)(cid:8)(cid:16)(cid:35)(cid:2)(cid:5)(cid:8)(cid:35)(cid:5)(cid:16)(cid:26)(cid:12)(cid:23)(cid:23)(cid:26)(cid:29)(cid:4)(cid:27)(cid:2)(cid:16)(cid:16)(cid:26)(cid:12)(cid:23)(cid:23)(cid:26)(cid:26)(cid:23)(cid:23)(cid:12)(cid:26)(cid:16)(cid:16)(cid:2)(cid:27)(cid:4)(cid:29)(cid:31)(cid:23)(cid:12)(cid:6)(cid:16)(cid:26)(cid:12)(cid:23)(cid:23)(cid:26)(cid:28)(cid:12)(cid:9)(cid:5)(cid:6)(cid:16)(cid:26)(cid:12)(cid:23)(cid:23)(cid:26)(cid:32)(cid:23)(cid:40)(cid:16)(cid:27)(cid:2)(cid:12)(cid:5)(cid:26)(cid:27)(cid:48)(cid:1)(cid:5)(cid:20)(cid:28)(cid:5)(cid:12)(cid:2)(cid:9)(cid:26)(cid:16)(cid:26)(cid:12)(cid:23)(cid:23)(cid:26)(cid:29)(cid:35)(cid:33)(cid:9)(cid:9)(cid:8)(cid:46)(cid:2)(cid:35)(cid:33)(cid:26)(cid:4)(cid:23)(cid:23)(cid:8)(cid:45)(cid:9)(cid:6)(cid:26)(cid:12)(cid:23)(cid:23)(cid:11)(cid:11)(cid:4)(cid:23)(cid:45)(cid:2)(cid:8)(cid:24)(cid:5)(cid:8)(cid:16)(cid:26)(cid:12)(cid:23)(cid:23)(cid:26)(cid:11)(cid:23)(cid:23)(cid:12)(cid:26)(cid:6)(cid:9)(cid:45)(cid:18)(cid:27)(cid:32)(cid:5)(cid:8)(cid:16)(cid:35)(cid:33)(cid:5)(cid:2)(cid:32)(cid:16)(cid:2)(cid:8)(cid:24)(cid:5)(cid:4)(cid:4)(cid:5)(cid:24)(cid:8)(cid:2)(cid:16)(cid:32)(cid:2)(cid:5)(cid:33)(cid:35)(cid:16)(cid:8)(cid:5)(cid:32)(cid:27)(cid:18)(cid:37)(cid:9)(cid:8)(cid:16)(cid:26)(cid:23)(cid:8)(cid:26)(cid:2)(cid:8)(cid:20)(cid:10)(cid:5)(cid:27)(cid:8)(cid:2)(cid:5)(cid:12)(cid:16)(cid:26)(cid:12)(cid:23)(cid:23)(cid:26)(cid:11)(cid:23)(cid:23)(cid:12)(cid:26)(cid:16)(cid:26)(cid:12)(cid:23)(cid:23)(cid:26)(cid:15)(cid:12)(cid:9)(cid:45)(cid:45)(cid:5)(cid:16)(cid:26)(cid:12)(cid:23)(cid:23)(cid:26)(cid:48)(cid:5)(cid:8)(cid:52)(cid:20)(cid:11)(cid:5)(cid:12)(cid:7)(cid:5)(cid:5)(cid:35)(cid:6)(cid:16)(cid:26)(cid:12)(cid:23)(cid:23)(cid:26)(cid:26)(cid:23)(cid:23)(cid:12)(cid:26)(cid:16)(cid:12)(cid:5)(cid:26)(cid:16)(cid:9)(cid:35)(cid:5)(cid:1)(cid:20)(cid:4)(cid:23)(cid:26)(cid:2)(cid:11)(cid:26)(cid:23)(cid:23)(cid:12)(cid:26)(cid:16)(cid:26)(cid:12)(cid:23)(cid:23)(cid:11)(cid:8)(cid:2)(cid:5)(cid:4)(cid:40)(cid:4)(cid:9)(cid:9)(cid:33)(cid:35)(cid:16)(cid:5)(cid:24)(cid:9)(cid:13)(cid:8)(cid:23)(cid:23)(cid:4)(cid:8)(cid:5)(cid:3)(cid:2)(cid:23)(cid:26)(cid:16)(cid:38)(cid:20)(cid:8)(cid:23)(cid:11)(cid:20)(cid:4)(cid:27)(cid:23)(cid:31)(cid:34)(cid:16)(cid:16)(cid:26)(cid:12)(cid:23)(cid:23)(cid:26)(cid:31)(cid:23)(cid:12)(cid:6)(cid:32)(cid:23)(cid:4)(cid:4)(cid:23)(cid:23)(cid:8)(cid:31)(cid:23)(cid:12)(cid:6)(cid:16)(cid:26)(cid:12)(cid:23)(cid:23)(cid:26)(cid:31)(cid:12)(cid:5)(cid:4)(cid:23)(cid:26)(cid:5)(cid:8)(cid:16)(cid:26)(cid:12)(cid:23)(cid:23)(cid:26)(cid:22)(cid:8)(cid:24)(cid:5)(cid:4)(cid:16)(cid:20)(cid:31)(cid:4)(cid:5)(cid:2)(cid:8)(cid:8)(cid:23)(cid:23)(cid:4)(cid:23)(cid:45)(cid:23)(cid:12)(cid:9)(cid:8)(cid:23)(cid:31)(cid:22)(cid:2)(cid:6)(cid:16)(cid:26)(cid:12)(cid:23)(cid:23)(cid:26)(cid:1)(cid:2)(cid:5)(cid:16)(cid:26)(cid:16)(cid:5)(cid:16)(cid:26)(cid:12)(cid:23)(cid:23)(cid:26)(cid:10)(cid:23)(cid:12)(cid:24)(cid:27)(cid:5)(cid:12)(cid:2)(cid:26)(cid:5)(cid:20)(cid:18)(cid:5)(cid:19)(cid:53)(cid:14)(cid:12)(cid:5)(cid:4)(cid:23)(cid:23)(cid:8)(cid:43)(cid:54)(cid:20)(cid:31)(cid:54)(cid:20)(cid:10)(cid:2)(cid:8)(cid:35)(cid:6)(cid:5)(cid:4)(cid:5)(cid:12)(cid:16)(cid:16)(cid:26)(cid:12)(cid:23)(cid:23)(cid:26)(cid:31)(cid:23)(cid:27)(cid:4)(cid:20)(cid:18)(cid:5)(cid:7)(cid:12)(cid:27)(cid:8)(cid:16)(cid:26)(cid:12)(cid:23)(cid:23)(cid:26)(cid:26)(cid:47)(cid:27)(cid:12)(cid:4)(cid:9)(cid:37)(cid:11)(cid:23)(cid:23)(cid:12)(cid:26)(cid:16)(cid:26)(cid:12)(cid:23)(cid:23)(cid:26)(cid:26)(cid:23)(cid:23)(cid:12)(cid:26)(cid:16)(cid:4)(cid:9)(cid:9)(cid:33)(cid:35)(cid:16)(cid:3)(cid:2)(cid:48)(cid:8)(cid:23)(cid:23)(cid:4)(cid:24)(cid:12)(cid:5)(cid:7)(cid:8)(cid:2)(cid:27)(cid:1)(cid:21)(cid:2)(cid:5)(cid:8)(cid:16)(cid:5)(cid:16)(cid:26)(cid:12)(cid:23)(cid:23)(cid:26)(cid:26)(cid:23)(cid:23)(cid:12)(cid:26)(cid:16)(cid:9)(cid:47)(cid:23)(cid:28)(cid:20)(cid:5)(cid:1)(cid:36)(cid:5)(cid:4)(cid:32)(cid:23)(cid:32)(cid:2)(cid:24)(cid:33)(cid:5)(cid:2)(cid:32)(cid:16)(cid:16)(cid:26)(cid:12)(cid:23)(cid:23)(cid:26)(cid:29)(cid:26)(cid:17)(cid:10)(cid:23)(cid:23)(cid:12)(cid:26)(cid:5)(cid:8)(cid:16)(cid:26)(cid:12)(cid:23)(cid:23)(cid:26)(cid:28)(cid:9)(cid:8)(cid:32)(cid:24)(cid:5)(cid:8)(cid:9)(cid:26)(cid:5)(cid:8)(cid:4)(cid:23)(cid:23)(cid:8)(cid:26)(cid:16)(cid:5)(cid:14)(cid:16)(cid:32)(cid:23)(cid:26)(cid:29)(cid:26)(cid:23)(cid:23)(cid:12)(cid:26)(cid:16)(cid:6)(cid:12)(cid:23)(cid:31)(cid:26)(cid:23)(cid:23)(cid:12)(cid:26)(cid:16)(cid:16)(cid:27)(cid:2)(cid:4)(cid:23)(cid:16)(cid:5)(cid:11)(cid:20)(cid:16)(cid:23)(cid:5)(cid:12)(cid:32)(cid:8)(cid:34)(cid:30)(cid:9)(cid:5)(cid:32)(cid:5)(cid:20)(cid:13)(cid:23)(cid:23)(cid:12)(cid:32)(cid:4)(cid:23)(cid:23)(cid:8)(cid:26)(cid:23)(cid:23)(cid:12)(cid:26)(cid:16)(cid:16)(cid:26)(cid:8)(cid:2)(cid:18)(cid:20)(cid:6)(cid:2)(cid:12)(cid:5)(cid:32)(cid:5)(cid:12)(cid:42)(cid:26)(cid:23)(cid:23)(cid:12)(cid:26)(cid:16)(cid:8)(cid:5)(cid:24)(cid:8)(cid:2)(cid:45)(cid:23)(cid:4)(cid:11)(cid:26)(cid:23)(cid:23)(cid:12)(cid:26)(cid:16)(cid:8)(cid:5)(cid:26)(cid:12)(cid:23)(cid:23)(cid:10)(cid:17)(cid:26)(cid:29)(cid:42)(cid:12)(cid:5)(cid:32)(cid:5)(cid:12)(cid:2)(cid:6)(cid:20)(cid:18)(cid:2)(cid:8)(cid:26)(cid:16)(cid:16)(cid:26)(cid:12)(cid:23)(cid:23)(cid:26)(cid:15)(cid:9)(cid:4)(cid:9)(cid:8)(cid:5)(cid:4)(cid:20)(cid:28)(cid:5)(cid:24)(cid:23)(cid:27)(cid:4)(cid:26)(cid:4)(cid:23)(cid:23)(cid:8)(cid:55)(cid:5)(cid:24)(cid:5)(cid:4)(cid:23)(cid:23)(cid:8)(cid:8)(cid:23)(cid:23)(cid:4)(cid:8)(cid:9)(cid:12)(cid:28)(cid:8)(cid:23)(cid:23)(cid:4)(cid:26)(cid:4)(cid:27)(cid:23)(cid:24)(cid:5)(cid:28)(cid:20)(cid:4)(cid:5)(cid:8)(cid:9)(cid:4)(cid:9)(cid:15)(cid:18)(cid:5)(cid:9)(cid:20)(cid:1)(cid:23)(cid:12)(cid:26)(cid:5)(cid:4)(cid:23)(cid:23)(cid:8)(cid:29)(cid:26)(cid:12)(cid:2)(cid:3)(cid:32)(cid:5)(cid:12)(cid:16)(cid:16)(cid:26)(cid:12)(cid:23)(cid:23)(cid:26)(cid:13)(cid:9)(cid:14)(cid:5)(cid:8)(cid:2)(cid:5)(cid:12)(cid:16)(cid:32)(cid:12)(cid:5)(cid:5)(cid:19)(cid:26)(cid:23)(cid:23)(cid:12)(cid:26)(cid:16)(cid:8)(cid:5)(cid:14)(cid:23)(cid:48)(cid:26)(cid:23)(cid:23)(cid:12)(cid:26)(cid:16)(cid:16)(cid:23)(cid:26)(cid:29)(cid:20)(cid:8)(cid:23)(cid:43)(cid:8)(cid:23)(cid:23)(cid:4)(cid:19)(cid:9)(cid:33)(cid:26)(cid:2)(cid:36)(cid:10)(cid:23)(cid:26)(cid:23)(cid:32)(cid:2)(cid:4)(cid:23)(cid:23)(cid:8)(cid:26)(cid:23)(cid:23)(cid:12)(cid:26)(cid:16)(cid:8)(cid:5)(cid:6)(cid:23)(cid:31)(cid:11)(cid:2)(cid:8)(cid:6)(cid:5)(cid:8)(cid:4)(cid:23)(cid:23)(cid:8)(cid:55)(cid:9)(cid:27)(cid:26)(cid:16)(cid:26)(cid:12)(cid:23)(cid:23)(cid:26)(cid:26)(cid:23)(cid:23)(cid:12)(cid:26)(cid:16)(cid:26)(cid:16)(cid:45)(cid:9)(cid:6)(cid:8)(cid:41)(cid:20)(cid:5)(cid:32)(cid:3)(cid:2)(cid:4)(cid:28)(cid:21)(cid:2)(cid:5)(cid:8)(cid:16)(cid:5)(cid:16)(cid:26)(cid:12)(cid:23)(cid:23)(cid:26)(cid:26)(cid:23)(cid:23)(cid:12)(cid:26)(cid:16)(cid:16)(cid:12)(cid:5)(cid:32)(cid:5)(cid:48)(cid:24)(cid:5)(cid:25)(cid:8)(cid:5)(cid:5)(cid:26)(cid:16)(cid:5)(cid:16)(cid:26)(cid:9)(cid:33)(cid:35)(cid:16)(cid:12)(cid:23)(cid:34)(cid:1)(cid:2)(cid:3)(cid:4)(cid:5)(cid:32)(cid:12)(cid:5)(cid:5)(cid:19)(cid:18)(cid:5)(cid:40)(cid:5)(cid:4)(cid:16)(cid:26)(cid:12)(cid:23)(cid:23)(cid:26)(cid:43)(cid:54)(cid:28)(cid:54)(cid:20)(cid:11)(cid:23)(cid:8)(cid:20)(cid:10)(cid:9)(cid:8)(cid:16)(cid:16)(cid:26)(cid:12)(cid:23)(cid:23)(cid:26)(cid:1)(cid:5)(cid:6)(cid:5)(cid:8)(cid:16)(cid:26)(cid:12)(cid:23)(cid:23)(cid:26)(cid:26)(cid:23)(cid:23)(cid:12)(cid:26)(cid:16)(cid:5)(cid:8)(cid:8)(cid:9)(cid:56)(cid:8)(cid:23)(cid:7)(cid:23)(cid:12)(cid:28)(cid:15)(cid:9)(cid:8)(cid:2)(cid:8)(cid:24)(cid:20)(cid:18)(cid:5)(cid:9)(cid:40)(cid:9)(cid:4)(cid:32)(cid:20)(cid:41)(cid:17)(cid:16)(cid:26)(cid:12)(cid:23)(cid:23)(cid:26)(cid:26)(cid:23)(cid:23)(cid:12)(cid:26)(cid:16)(cid:23)(cid:24)(cid:5)(cid:48)(cid:20)(cid:2)(cid:12)(cid:8)(cid:5)(cid:13)(cid:39)(cid:5)(cid:12)(cid:14)(cid:2)(cid:57)(cid:12)(cid:16)(cid:16)(cid:26)(cid:12)(cid:23)(cid:23)(cid:26)(cid:29)(cid:26)(cid:12)(cid:2)(cid:3)(cid:32)(cid:5)(cid:12)(cid:16)(cid:16)(cid:26)(cid:12)(cid:23)(cid:23)(cid:26)(cid:18)(cid:23)(cid:8)(cid:24)(cid:5)(cid:8)(cid:32)(cid:23)(cid:23)(cid:4)(cid:4)(cid:23)(cid:23)(cid:8)(cid:26)(cid:23)(cid:23)(cid:12)(cid:26)(cid:16)(cid:8)(cid:5)(cid:32)(cid:12)(cid:23)(cid:23)(cid:24)(cid:9)(cid:28)(cid:21)(cid:2)(cid:5)(cid:8)(cid:16)(cid:5)(cid:16)(cid:26)(cid:12)(cid:23)(cid:23)(cid:26)(cid:26)(cid:23)(cid:23)(cid:12)(cid:26)(cid:16)(cid:16)(cid:45)(cid:5)(cid:4)(cid:4)(cid:2)(cid:36)(cid:26)(cid:23)(cid:23)(cid:12)(cid:26)(cid:16)(cid:12)(cid:5)(cid:26)(cid:16)(cid:5)(cid:5)(cid:45)(cid:5)(cid:24)(cid:12)(cid:27)(cid:28)(cid:8)(cid:23)(cid:23)(cid:4)(cid:5)(cid:32)(cid:12)(cid:9)(cid:9)(cid:11)(cid:48)(cid:23)(cid:14)(cid:5)(cid:8)(cid:16)(cid:26)(cid:12)(cid:23)(cid:23)(cid:26)(cid:43)(cid:54)(cid:28)(cid:54)(cid:20)(cid:11)(cid:23)(cid:8)(cid:20)(cid:10)(cid:9)(cid:8)(cid:16)(cid:16)(cid:26)(cid:12)(cid:23)(cid:23)(cid:26)(cid:1)(cid:2)(cid:3)(cid:4)(cid:5)(cid:32)(cid:12)(cid:5)(cid:5)(cid:19)(cid:18)(cid:5)(cid:40)(cid:5)(cid:4)(cid:16)(cid:26)(cid:12)(cid:23)(cid:23)(cid:26)(cid:8)(cid:23)(cid:23)(cid:7)(cid:5)(cid:16)(cid:6)(cid:23)(cid:23)(cid:8)(cid:5)(cid:32)(cid:4)(cid:5)(cid:30)(cid:13)(cid:14)(cid:10)(cid:26)(cid:23)(cid:23)(cid:12)(cid:26)(cid:16)(cid:26)(cid:16)(cid:45)(cid:9)(cid:6)(cid:8)(cid:41)(cid:20)(cid:5)(cid:32)(cid:3)(cid:2)(cid:4)(cid:28)(cid:26)(cid:23)(cid:23)(cid:12)(cid:26)(cid:16)(cid:23)(cid:2)(cid:16)(cid:5)(cid:12)(cid:5)(cid:33)(cid:21)(cid:20)(cid:23)(cid:2)(cid:12)(cid:23)(cid:10)(cid:1)(cid:2)(cid:5)(cid:16)(cid:26)(cid:16)(cid:5)(cid:14)(cid:5)(cid:16)(cid:26)(cid:26)(cid:23)(cid:23)(cid:12)(cid:26)(cid:16)(cid:8)(cid:5)(cid:32)(cid:4)(cid:2)(cid:30)(cid:26)(cid:16)(cid:5)(cid:14)(cid:5)(cid:16)(cid:26)(cid:16)(cid:5)(cid:2)(cid:1)(cid:11)(cid:27)(cid:27)(cid:12)(cid:6)(cid:12)(cid:27)(cid:2)(cid:16)(cid:5)(cid:8)(cid:4)(cid:23)(cid:23)(cid:8)(cid:5)(cid:26)(cid:2)(cid:16)(cid:16)(cid:40)(cid:2)(cid:4)(cid:2)(cid:33)(cid:31)(cid:30)(cid:5)(cid:4)(cid:32)(cid:5)(cid:8)(cid:23)(cid:23)(cid:6)(cid:16)(cid:5)(cid:14)(cid:5)(cid:16)(cid:26)(cid:18)(cid:2)(cid:3)(cid:8)(cid:20)(cid:58)(cid:59)(cid:1)(cid:5)(cid:6)(cid:5)(cid:8)(cid:16)(cid:26)(cid:12)(cid:23)(cid:23)(cid:26)(cid:1)(cid:2)(cid:3)(cid:4)(cid:5)(cid:32)(cid:12)(cid:5)(cid:5)(cid:19)(cid:43)(cid:27)(cid:16)(cid:26)(cid:27)(cid:16)(cid:20)(cid:18)(cid:2)(cid:40)(cid:16)(cid:2)(cid:27)(cid:16)(cid:16)(cid:26)(cid:12)(cid:23)(cid:23)(cid:26)(cid:8)(cid:2)(cid:5)(cid:4)(cid:40)(cid:16)(cid:40)(cid:2)(cid:4)(cid:2)(cid:33)(cid:31)(cid:21)(cid:2)(cid:5)(cid:8)(cid:16)(cid:5)(cid:16)(cid:26)(cid:12)(cid:23)(cid:23)(cid:26)(cid:21)(cid:2)(cid:14)(cid:9)(cid:4)(cid:2)(cid:16)(cid:26)(cid:12)(cid:23)(cid:23)(cid:26)(cid:19)(cid:5)(cid:5)(cid:12)(cid:32)(cid:3)(cid:2)(cid:32)(cid:7)(cid:34)(cid:43)(cid:9)(cid:46)(cid:5)(cid:19)(cid:20)(cid:41)(cid:41)(cid:17)(cid:16)(cid:26)(cid:12)(cid:23)(cid:23)(cid:26)(cid:15)(cid:5)(cid:12)(cid:6)(cid:33)(cid:9)(cid:19)(cid:32)(cid:12)(cid:5)(cid:5)(cid:19)(cid:8)(cid:23)(cid:23)(cid:7)(cid:5)(cid:16)(cid:6)(cid:23)(cid:23)(cid:8)(cid:5)(cid:32)(cid:4)(cid:5)(cid:30)(cid:5)(cid:26)(cid:2)(cid:16)(cid:16)(cid:40)(cid:2)(cid:4)(cid:2)(cid:33)(cid:31)(cid:31)(cid:33)(cid:2)(cid:4)(cid:2)(cid:40)(cid:16)(cid:16)(cid:2)(cid:26)(cid:5)(cid:26)(cid:23)(cid:23)(cid:12)(cid:26)(cid:16)(cid:3)(cid:2)(cid:32)(cid:7)(cid:34)(cid:1)(cid:2)(cid:5)(cid:16)(cid:26)(cid:16)(cid:5)(cid:16)(cid:26)(cid:12)(cid:23)(cid:23)(cid:26)(cid:26)(cid:23)(cid:23)(cid:12)(cid:26)(cid:16)(cid:4)(cid:5)(cid:7)(cid:9)(cid:39)(cid:20)(cid:32)(cid:4)(cid:27)(cid:9)(cid:8)(cid:12)(cid:34)(cid:1)(cid:23)(cid:24)(cid:9)(cid:7)(cid:5)(cid:12)(cid:26)(cid:16)(cid:26)(cid:12)(cid:23)(cid:23)(cid:26)(cid:38)(cid:9)(cid:16)(cid:26)(cid:5)(cid:12)(cid:26)(cid:27)(cid:8)(cid:8)(cid:5)(cid:4)(cid:43)(cid:27)(cid:16)(cid:26)(cid:27)(cid:16)(cid:20)(cid:18)(cid:2)(cid:40)(cid:16)(cid:2)(cid:27)(cid:16)(cid:16)(cid:26)(cid:12)(cid:23)(cid:23)(cid:26)(cid:26)(cid:23)(cid:23)(cid:12)(cid:26)(cid:16)(cid:8)(cid:2)(cid:5)(cid:4)(cid:31)(cid:26)(cid:16)(cid:5)(cid:14)(cid:5)(cid:16)(cid:26)(cid:16)(cid:5)(cid:2)(cid:1)(cid:1)(cid:2)(cid:5)(cid:16)(cid:26)(cid:16)(cid:5)(cid:40)(cid:9)(cid:9)(cid:12)(cid:26)(cid:15)(cid:10)(cid:21)(cid:2)(cid:5)(cid:8)(cid:16)(cid:5)(cid:14)(cid:5)(cid:16)(cid:26)(cid:21)(cid:2)(cid:5)(cid:8)(cid:16)(cid:5)(cid:14)(cid:5)(cid:16)(cid:26)(cid:28)(cid:2)(cid:5)(cid:12)(cid:7)(cid:5)(cid:5)(cid:6)(cid:16)(cid:26)(cid:12)(cid:23)(cid:23)(cid:26)(cid:13)(cid:14)(cid:10)(cid:21)(cid:2)(cid:14)(cid:9)(cid:4)(cid:2)(cid:16)(cid:26)(cid:12)(cid:23)(cid:23)(cid:26)(cid:18)(cid:2)(cid:3)(cid:8)(cid:20)(cid:58)(cid:59)(cid:26)(cid:23)(cid:23)(cid:12)(cid:26)(cid:16)(cid:6)(cid:5)(cid:5)(cid:7)(cid:12)(cid:5)(cid:2)(cid:28)(cid:21)(cid:2)(cid:5)(cid:8)(cid:16)(cid:5)(cid:14)(cid:5)(cid:16)(cid:26)(cid:28)(cid:12)(cid:27)(cid:24)(cid:7)(cid:5)(cid:12)(cid:24)(cid:40)(cid:23)(cid:32)(cid:19)(cid:5)(cid:5)(cid:12)(cid:32)(cid:3)(cid:2)(cid:32)(cid:7)(cid:34)(cid:31)(cid:12)(cid:9)(cid:14)(cid:2)(cid:8)(cid:35)(cid:2)(cid:5)(cid:33)(cid:27)(cid:2)(cid:16)(cid:26)(cid:23)(cid:23)(cid:12)(cid:26)(cid:16)(cid:8)(cid:5)(cid:12)(cid:9)(cid:15)(cid:8)(cid:23)(cid:23)(cid:4)(cid:2)(cid:5)(cid:45)(cid:8)(cid:5)(cid:22)(cid:15)(cid:14)(cid:10)(cid:23)(cid:12)(cid:26)(cid:5)(cid:4)(cid:23)(cid:12)(cid:5)(cid:8)(cid:4)(cid:23)(cid:23)(cid:8)(cid:22)(cid:54)(cid:29)(cid:9)(cid:4)(cid:14)(cid:23)(cid:47)(cid:16)(cid:26)(cid:12)(cid:23)(cid:23)(cid:26)(cid:26)(cid:23)(cid:23)(cid:12)(cid:26)(cid:16)(cid:3)(cid:2)(cid:32)(cid:7)(cid:34)(cid:26)(cid:23)(cid:23)(cid:12)(cid:26)(cid:16)(cid:2)(cid:4)(cid:9)(cid:14)(cid:2)(cid:21)(cid:8)(cid:23)(cid:23)(cid:4)(cid:2)(cid:5)(cid:45)(cid:8)(cid:5)(cid:22)(cid:13)(cid:5)(cid:12)(cid:19)(cid:16)(cid:26)(cid:4)(cid:23)(cid:23)(cid:8)(cid:60)(cid:20)(cid:18)(cid:29)(cid:13)(cid:48)(cid:9)(cid:8)(cid:32)(cid:5)(cid:16)(cid:26)(cid:12)(cid:23)(cid:23)(cid:26)(cid:18)(cid:2)(cid:3)(cid:8)(cid:20)(cid:59)(cid:61)(cid:10)(cid:23)(cid:12)(cid:26)(cid:5)(cid:4)(cid:23)(cid:12)(cid:5)(cid:8)(cid:4)(cid:23)(cid:23)(cid:8)(cid:26)(cid:23)(cid:23)(cid:12)(cid:26)(cid:16)(cid:16)(cid:8)(cid:23)(cid:45)(cid:5)(cid:16)(cid:9)(cid:9)(cid:37)(cid:20)(cid:45)(cid:5)(cid:4)(cid:4)(cid:2)(cid:36)(cid:8)(cid:23)(cid:23)(cid:4)(cid:2)(cid:5)(cid:45)(cid:8)(cid:5)(cid:22)(cid:18)(cid:5)(cid:8)(cid:26)(cid:5)(cid:32)(cid:12)(cid:5)(cid:5)(cid:19)(cid:8)(cid:23)(cid:23)(cid:4)(cid:2)(cid:5)(cid:45)(cid:8)(cid:5)(cid:22)(cid:26)(cid:23)(cid:23)(cid:12)(cid:26)(cid:16)(cid:16)(cid:8)(cid:5)(cid:6)(cid:5)(cid:4)(cid:4)(cid:9)(cid:39)(cid:20)(cid:12)(cid:5)(cid:26)(cid:5)(cid:2)(cid:31)(cid:26)(cid:23)(cid:23)(cid:12)(cid:26)(cid:16)(cid:40)(cid:2)(cid:36)(cid:34)(cid:35)(cid:33)(cid:26)(cid:27)(cid:12)(cid:5)(cid:8)(cid:16)(cid:26)(cid:12)(cid:23)(cid:23)(cid:26)(cid:8)(cid:23)(cid:23)(cid:4)(cid:5)(cid:32)(cid:2)(cid:5)(cid:25)(cid:5)(cid:5)(cid:11)(cid:8)(cid:23)(cid:23)(cid:4)(cid:16)(cid:4)(cid:5)(cid:14)(cid:27)(cid:5)(cid:12)(cid:26)(cid:29)(cid:20)(cid:8)(cid:3)(cid:2)(cid:26)(cid:29)(cid:30)(cid:5)(cid:45)(cid:5)(cid:5)(cid:8)(cid:26)(cid:5)(cid:16)(cid:26)(cid:12)(cid:23)(cid:23)(cid:26)(cid:29)(cid:40)(cid:23)(cid:23)(cid:12)(cid:16)(cid:26)(cid:12)(cid:23)(cid:23)(cid:26)(cid:60)(cid:20)(cid:18)(cid:29)(cid:13)(cid:15)(cid:9)(cid:8)(cid:2)(cid:8)(cid:24)(cid:20)(cid:34)(cid:4)(cid:7)(cid:5)(cid:12)(cid:26)(cid:4)(cid:23)(cid:23)(cid:8)(cid:18)(cid:2)(cid:3)(cid:8)(cid:20)(cid:59)(cid:61)(cid:36)(cid:5)(cid:12)(cid:6)(cid:33)(cid:27)(cid:2)(cid:46)(cid:5)(cid:8)(cid:16)(cid:26)(cid:12)(cid:23)(cid:23)(cid:26)(cid:31)(cid:4)(cid:23)(cid:26)(cid:26)(cid:5)(cid:17)(cid:18)(cid:9)(cid:16)(cid:26)(cid:12)(cid:23)(cid:23)(cid:26)(cid:15)(cid:10)(cid:21)(cid:2)(cid:5)(cid:8)(cid:16)(cid:5)(cid:16)(cid:26)(cid:5)(cid:5)(cid:8)(cid:25)(cid:5)(cid:24)(cid:26)(cid:23)(cid:23)(cid:12)(cid:26)(cid:16)(cid:24)(cid:27)(cid:12)(cid:28)(cid:26)(cid:23)(cid:23)(cid:12)(cid:26)(cid:16)(cid:5)(cid:16)(cid:9)(cid:18)(cid:17)(cid:6)(cid:5)(cid:5)(cid:7)(cid:12)(cid:9)(cid:15)(cid:8)(cid:23)(cid:23)(cid:4)(cid:32)(cid:2)(cid:12)(cid:26)(cid:16)(cid:34)(cid:20)(cid:8)(cid:2)(cid:24)(cid:8)(cid:2)(cid:8)(cid:9)(cid:15)(cid:43)(cid:9)(cid:46)(cid:5)(cid:19)(cid:20)(cid:36)(cid:23)(cid:27)(cid:26)(cid:5)(cid:12)(cid:16)(cid:16)(cid:26)(cid:12)(cid:23)(cid:23)(cid:26)(cid:26)(cid:23)(cid:23)(cid:12)(cid:26)(cid:16)(cid:16)(cid:8)(cid:23)(cid:45)(cid:12)(cid:27)(cid:33)(cid:35)(cid:29)(cid:20)(cid:4)(cid:5)(cid:12)(cid:23)(cid:15)(cid:26)(cid:23)(cid:23)(cid:12)(cid:26)(cid:16)(cid:6)(cid:12)(cid:5)(cid:15)(cid:43)(cid:9)(cid:46)(cid:5)(cid:19)(cid:20)(cid:31)(cid:2)(cid:5)(cid:12)(cid:12)(cid:5)(cid:16)(cid:26)(cid:12)(cid:23)(cid:23)(cid:26)(cid:26)(cid:23)(cid:23)(cid:12)(cid:26)(cid:16)(cid:6)(cid:12)(cid:5)(cid:15)(cid:11)(cid:12)(cid:5)(cid:32)(cid:5)(cid:16)(cid:26)(cid:12)(cid:23)(cid:23)(cid:26)(cid:26)(cid:23)(cid:23)(cid:12)(cid:26)(cid:16)(cid:8)(cid:5)(cid:32)(cid:45)(cid:5)(cid:5)(cid:28)(cid:18)(cid:2)(cid:8)(cid:32)(cid:5)(cid:8)(cid:16)(cid:5)(cid:16)(cid:26)(cid:12)(cid:23)(cid:23)(cid:26)(cid:8)(cid:23)(cid:23)(cid:4)(cid:26)(cid:2)(cid:9)(cid:8)(cid:5)(cid:28)(cid:20)(cid:12)(cid:5)(cid:26)(cid:5)(cid:31)(cid:26)(cid:23)(cid:23)(cid:12)(cid:26)(cid:16)(cid:5)(cid:16)(cid:32)(cid:12)(cid:23)(cid:23)(cid:24)(cid:5)(cid:9)(cid:13)(cid:26)(cid:23)(cid:23)(cid:12)(cid:26)(cid:16)(cid:26)(cid:12)(cid:9)(cid:9)(cid:40)(cid:8)(cid:5)(cid:12)(cid:5)(cid:46)(cid:43)(cid:41)(cid:26)(cid:23)(cid:23)(cid:12)(cid:26)(cid:16)(cid:5)(cid:32)(cid:9)(cid:48)(cid:20)(cid:8)(cid:23)(cid:11)(cid:34)(cid:4)(cid:7)(cid:12)(cid:5)(cid:35)(cid:33)(cid:26)(cid:20)(cid:48)(cid:9)(cid:32)(cid:5)(cid:8)(cid:7)(cid:23)(cid:35)(cid:33)(cid:16)(cid:26)(cid:12)(cid:23)(cid:23)(cid:26)(cid:29)(cid:40)(cid:23)(cid:23)(cid:8)(cid:16)(cid:5)(cid:20)(cid:15)(cid:12)(cid:9)(cid:9)(cid:8)(cid:4)(cid:23)(cid:23)(cid:8)(cid:26)(cid:23)(cid:23)(cid:12)(cid:26)(cid:16)(cid:16)(cid:8)(cid:23)(cid:45)(cid:5)(cid:16)(cid:9)(cid:9)(cid:37)(cid:20)(cid:45)(cid:5)(cid:4)(cid:4)(cid:2)(cid:36)(cid:22)(cid:4)(cid:19)(cid:20)(cid:39)(cid:9)(cid:14)(cid:5)(cid:45)(cid:7)(cid:5)(cid:12)(cid:4)(cid:23)(cid:23)(cid:8)(cid:8)(cid:23)(cid:23)(cid:4)(cid:27)(cid:23)(cid:6)(cid:23)(cid:12)(cid:15)(cid:26)(cid:23)(cid:23)(cid:12)(cid:26)(cid:16)(cid:26)(cid:5)(cid:2)(cid:4)(cid:11)(cid:8)(cid:23)(cid:23)(cid:4)(cid:26)(cid:33)(cid:35)(cid:12)(cid:27)(cid:28)(cid:20)(cid:5)(cid:16)(cid:8)(cid:23)(cid:23)(cid:40)(cid:29)(cid:26)(cid:23)(cid:23)(cid:12)(cid:26)(cid:16)(cid:26)(cid:16)(cid:5)(cid:5)(cid:19)(cid:25)(cid:27)(cid:5)(cid:22)(cid:43)(cid:23)(cid:8)(cid:20)(cid:11)(cid:23)(cid:8)(cid:32)(cid:5)(cid:14)(cid:5)(cid:4)(cid:32)(cid:5)(cid:4)(cid:23)(cid:23)(cid:8)(cid:38)(cid:27)(cid:32)(cid:5)(cid:7)(cid:23)(cid:23)(cid:8)(cid:19)(cid:9)(cid:33)(cid:26)(cid:33)(cid:35)(cid:23)(cid:31)(cid:26)(cid:23)(cid:23)(cid:12)(cid:26)(cid:16)(cid:26)(cid:16)(cid:45)(cid:9)(cid:6)(cid:5)(cid:9)(cid:21)(cid:26)(cid:23)(cid:23)(cid:12)(cid:26)(cid:16)(cid:5)(cid:16)(cid:32)(cid:2)(cid:45)(cid:29)(cid:1)(cid:2)(cid:5)(cid:16)(cid:26)(cid:16)(cid:5)(cid:16)(cid:26)(cid:5)(cid:5)(cid:8)(cid:25)(cid:5)(cid:24)(cid:13)(cid:9)(cid:5)(cid:24)(cid:23)(cid:23)(cid:12)(cid:32)(cid:16)(cid:5)(cid:16)(cid:26)(cid:12)(cid:23)(cid:23)(cid:26)(cid:8)(cid:23)(cid:23)(cid:4)(cid:16)(cid:40)(cid:2)(cid:4)(cid:2)(cid:33)(cid:31)(cid:48)(cid:2)(cid:35)(cid:33)(cid:23)(cid:12)(cid:32)(cid:20)(cid:11)(cid:23)(cid:4)(cid:14)(cid:5)(cid:6)(cid:5)(cid:8)(cid:16)(cid:26)(cid:12)(cid:23)(cid:23)(cid:26)(cid:34)(cid:27)(cid:24)(cid:27)(cid:16)(cid:26)(cid:20)(cid:11)(cid:5)(cid:12)(cid:45)(cid:5)(cid:47)(cid:4)(cid:5)(cid:8)(cid:4)(cid:23)(cid:23)(cid:8)(cid:24)(cid:5)(cid:25)(cid:16)(cid:4)(cid:5)(cid:14)(cid:2)(cid:27)(cid:1)(cid:8)(cid:23)(cid:23)(cid:4)(cid:32)(cid:8)(cid:23)(cid:26)(cid:16)(cid:12)(cid:5)(cid:5)(cid:36)(cid:19)(cid:9)(cid:9)(cid:33)(cid:35)(cid:16)(cid:5)(cid:25)(cid:12)(cid:23)(cid:21)(cid:26)(cid:23)(cid:23)(cid:12)(cid:26)(cid:16)(cid:16)(cid:24)(cid:8)(cid:2)(cid:8)(cid:8)(cid:2)(cid:25)(cid:12)(cid:5)(cid:14)(cid:38)(cid:18)(cid:5)(cid:9)(cid:40)(cid:9)(cid:4)(cid:32)(cid:20)(cid:28)(cid:5)(cid:9)(cid:16)(cid:2)(cid:5)(cid:12)(cid:4)(cid:23)(cid:23)(cid:8)(cid:38)(cid:27)(cid:32)(cid:5)(cid:7)(cid:23)(cid:23)(cid:8)(cid:22)(cid:4)(cid:19)(cid:20)(cid:39)(cid:9)(cid:14)(cid:5)(cid:45)(cid:7)(cid:5)(cid:12)(cid:4)(cid:23)(cid:23)(cid:8)(cid:8)(cid:23)(cid:23)(cid:4)(cid:8)(cid:9)(cid:9)(cid:12)(cid:15)(cid:20)(cid:5)(cid:16)(cid:8)(cid:23)(cid:23)(cid:40)(cid:29)(cid:55)(cid:2)(cid:5)(cid:6)(cid:5)(cid:4)(cid:2)(cid:8)(cid:24)(cid:5)(cid:8)(cid:16)(cid:26)(cid:12)(cid:23)(cid:23)(cid:26)(cid:26)(cid:23)(cid:23)(cid:12)(cid:26)(cid:16)(cid:4)(cid:9)(cid:9)(cid:33)(cid:35)(cid:29)(cid:37)(cid:12)(cid:2)(cid:16)(cid:26)(cid:2)(cid:23)(cid:8)(cid:4)(cid:23)(cid:23)(cid:8)(cid:28)(cid:5)(cid:5)(cid:45)(cid:32)(cid:5)(cid:8)(cid:40)(cid:23)(cid:32)(cid:1)(cid:27)(cid:2)(cid:14)(cid:5)(cid:4)(cid:16)(cid:25)(cid:5)(cid:24)(cid:32)(cid:4)(cid:5)(cid:14)(cid:5)(cid:8)(cid:2)(cid:27)(cid:12)(cid:28)(cid:19)(cid:5)(cid:5)(cid:12)(cid:32)(cid:8)(cid:5)(cid:3)(cid:2)(cid:28)(cid:60)(cid:20)(cid:18)(cid:29)(cid:13)(cid:22)(cid:16)(cid:40)(cid:4)(cid:23)(cid:8)(cid:23)(cid:32)(cid:5)(cid:4)(cid:23)(cid:23)(cid:8)(cid:34)(cid:4)(cid:7)(cid:5)(cid:12)(cid:26)(cid:20)(cid:1)(cid:5)(cid:3)(cid:9)(cid:8)(cid:24)(cid:33)(cid:5)(cid:16)(cid:26)(cid:12)(cid:23)(cid:23)(cid:26)(cid:18)(cid:2)(cid:3)(cid:8)(cid:20)(cid:59)(cid:61)(cid:18)(cid:5)(cid:9)(cid:40)(cid:9)(cid:4)(cid:32)(cid:20)(cid:28)(cid:5)(cid:9)(cid:16)(cid:2)(cid:5)(cid:12)(cid:4)(cid:23)(cid:23)(cid:8)(cid:15)(cid:9)(cid:12)(cid:26)(cid:12)(cid:2)(cid:3)(cid:6)(cid:16)(cid:5)(cid:16)(cid:26)(cid:12)(cid:23)(cid:23)(cid:26)(cid:8)(cid:23)(cid:23)(cid:4)(cid:12)(cid:5)(cid:14)(cid:3)(cid:2)(cid:11)(cid:18)(cid:9)(cid:4)(cid:23)(cid:8)(cid:32)(cid:5)(cid:8)(cid:8)(cid:23)(cid:23)(cid:7)(cid:5)(cid:16)(cid:12)(cid:5)(cid:14)(cid:23)(cid:36)(cid:15)(cid:9)(cid:8)(cid:2)(cid:8)(cid:24)(cid:20)(cid:28)(cid:9)(cid:27)(cid:32)(cid:5)(cid:25)(cid:2)(cid:3)(cid:8)(cid:4)(cid:23)(cid:23)(cid:8)(cid:8)(cid:23)(cid:23)(cid:4)(cid:8)(cid:3)(cid:2)(cid:25)(cid:5)(cid:32)(cid:27)(cid:9)(cid:28)(cid:20)(cid:24)(cid:8)(cid:2)(cid:8)(cid:9)(cid:15)(cid:26)(cid:16)(cid:5)(cid:14)(cid:5)(cid:16)(cid:8)(cid:5)(cid:2)(cid:21)(cid:26)(cid:16)(cid:5)(cid:14)(cid:5)(cid:16)(cid:6)(cid:23)(cid:23)(cid:8)(cid:5)(cid:32)(cid:4)(cid:5)(cid:30)(cid:62)(cid:63)(cid:64)(cid:55)(cid:20)(cid:30)(cid:23)(cid:16)(cid:26)(cid:33)(cid:27)(cid:2)(cid:16)(cid:7)(cid:5)(cid:12)(cid:24)(cid:62)(cid:63)(cid:62)(cid:65)(cid:36)(cid:2)(cid:8)(cid:32)(cid:5)(cid:6)(cid:2)(cid:8)(cid:32)(cid:13)(cid:5)(cid:2)(cid:4)(cid:2)(cid:24)(cid:17)(cid:1)(cid:12)(cid:2)(cid:5)(cid:14)(cid:27)(cid:4)(cid:32)(cid:2)(cid:24)(cid:33)(cid:5)(cid:2)(cid:32)(cid:16)(cid:35)(cid:9)(cid:4)(cid:4)(cid:5)(cid:24)(cid:5)(cid:17)(cid:18)(cid:19)(cid:18)(cid:6)(cid:2)(cid:7)(cid:3)(cid:2)(cid:8)(cid:26)(cid:9)(cid:27)(cid:12)(cid:2)(cid:16)(cid:26)(cid:2)(cid:8)(cid:19)(cid:9)(cid:12)(cid:45)(cid:23)(cid:26)(cid:2)(cid:9)(cid:8)(cid:10)(cid:2)(cid:8)(cid:2)(cid:5)(cid:45)(cid:5)(cid:8)(cid:2)(cid:8)(cid:16)(cid:26)(cid:2)(cid:26)(cid:27)(cid:27)(cid:26)(cid:15)(cid:9)(cid:8)(cid:2)(cid:8)(cid:6)(cid:4)(cid:2)(cid:3)(cid:6)(cid:34)(cid:26)(cid:33)(cid:5)(cid:8)(cid:5)(cid:27)(cid:45)(cid:48)(cid:5)(cid:32)(cid:2)(cid:8)(cid:24)(cid:5)(cid:8)(cid:33)(cid:9)(cid:19)(cid:1)(cid:2)(cid:3)(cid:4)(cid:5)(cid:1)(cid:2)(cid:3)(cid:4)(cid:5)(cid:1)(cid:2)(cid:3)(cid:4)(cid:5)(cid:1)(cid:2)(cid:3)(cid:4)(cid:5)(cid:1)(cid:2)(cid:3)(cid:4)(cid:5)(cid:1)(cid:2)(cid:3)(cid:4)(cid:5)(cid:1)(cid:2)(cid:3)(cid:4)(cid:5)(cid:1)(cid:2)(cid:3)(cid:4)(cid:5)(cid:1)(cid:2)(cid:3)(cid:4)(cid:5)(cid:11)(cid:23)(cid:23)(cid:12)(cid:26)(cid:6)(cid:9)(cid:45)(cid:1)(cid:2)(cid:3)(cid:4)(cid:5)(cid:1)(cid:2)(cid:3)(cid:4)(cid:5)(cid:1)(cid:2)(cid:3)(cid:4)(cid:5)(cid:9)(cid:4)(cid:20)(cid:21)(cid:22)(cid:7)(cid:21)(cid:23)(cid:2)(cid:24)(cid:25)(cid:7)(cid:26)(cid:1)(cid:2)(cid:3)(cid:4)(cid:5)(cid:30)(cid:23)(cid:16)(cid:26)(cid:33)(cid:27)(cid:2)(cid:16)(cid:7)(cid:5)(cid:12)(cid:24)(cid:31)(cid:12)(cid:9)(cid:14)(cid:2)(cid:8)(cid:35)(cid:2)(cid:5)(cid:32)(cid:9)(cid:45)(cid:5)(cid:2)(cid:8)(cid:15)(cid:5)(cid:16)(cid:16)(cid:5)(cid:4)(cid:17)(cid:18)(cid:9)(cid:29)(cid:2)(cid:8)(cid:26)(cid:17)(cid:1)(cid:9)(cid:8)(cid:23)(cid:26)(cid:27)(cid:16)(cid:40)(cid:23)(cid:12)(cid:6)(cid:41)(cid:10)(cid:22)(cid:37)An Architectural Style for Self-Adaptive Multi-Agent Systems · 15 To avoid collisions, AGV agents have to lock the segments they want to drive. Therefore, the virtual environment uses path projections that are realized as follows: public class PathProjection { private int id; private int priority; private Hull hull; private ArrayList projection; private Status status; ... public void lock() public void clear(ArrayList segments) ... { ... } } { ... } A path projection has an identifier, a priority that is used to give precedence to AGVs with important transportation tasks, a hull that represents the physical area an AGV occu- pies on the floor, a projection that consists of an array of segments of the intended path, and finally a status which is requested or locked. When an AGV agent projects the path projec- tion in the AGV local virtual environment, the path projection will have status requested. The local virtual environment then coordinates with other local virtual environment and marks the path projection locked when the conditions for the AGV are safe to move on. This coordination is realized by the synchronization component discussed below. When the AGV drives forward, the path projection is gradually cleared. The Perception Service provides the functionality for agents to sense the environment. The π-ADL specification of the perception service is defined as follows: component PerceptionService is abstraction () { type Focus is view[focusName : String, focusParams : Any]. type Sense is view[agentId : String, focus : Focus]. type Representation is Any. type Observation is Any. type Observed is Any. ... port perceive is { connection sense is in(Sense). port observe is { connection observation is out(Observation). connection sensed is out(Representation). } connection observed is in(Observed) }. ... perceptionTypes is location(set(StateItems)). behavior is { analyzePerception is function ( sense : Sense ) : StateItems { unobservable }. virtualPerception is function ( items : StateItems, types : set(StateItems) : Boolean { unobservable }. generateRepresentation is function( sense : Sense, state : StateItems) : Representation { unobservable }. ... ACM Transactions Journal Format, Vol. 2, No. 3, 09 2001. · 16 D. Weyns and F. Oquendo via sense receive sense_request : Sense. sense_items := analyzePerception( sense_request ). if virtualPerception( sense_items, perceptionTypes ) then { //perceive local state repository unobservable. via stateTemplate send sense_items. via readState receive perceived_state. via sensed send generateRepresentation( sense_request, perceived_state ). behavior(). } else { //perceive external environment unobservable. via observation send generateObservation( sense_items ). unobservable. via observed receive external_observed. observation_state := analyzeObserved( external_observed ). via sensed send generateRepresentation( sense_request, observation_state ). behavior(). } } } The perception service supports selective perception. Selective perception enables an agent to direct its perception at the relevant aspects according to its current task. This facilitates better situation awareness. To direct its perception an agent selects a focus. A focus has a set of domains-specific parameters that determine aspects such as element types of interest, operating range, etc. Sensing results in a representation. A representation is a domain-specific data structure that represents the sensed elements in a form that can be interpreted by the agents. When an agent invokes a perception request, the perception service analyzes it and de- termines whether it concerns a virtual perception of the local state or a perception of the external environment. In case of a virtual perception, a representation is generated based on state derived from the repository. In case of an external perception, an observation is generated, and a representation is generated from the observed state. An observation is a domain-specific command that allows observing state of the external environment. Examples of foci in the AGV transportation system are a focus to observe the positions of nearby vehicles and a focus to observe the status of an operating space. An example in the vehicle routing system is a focus to sense a path on the road infrastructure: public PathFocus(String name, int maxDist, Node from, Node to) { ... } pfocus = new PathFocus("paths", 1500, n15, n125); This focus selects the traffic routes with a maximum distance of 1500 m from a given node (n15) to another node (n125). The nodes refer to the elements of the road elements ACM Transactions Journal Format, Vol. 2, No. 3, 09 2001. An Architectural Style for Self-Adaptive Multi-Agent Systems · 17 of the road network. Nodes are an abstraction of road elements used by the agents. The perception service maps nodes to road elements of the road network maintained by the state repository of the virtual environment. Note that the result of a perception with this focus may dynamically change when certain road areas are temporary not accessible as reflected in the map. Such map updates can be maintained by the synchronization components (discussed below). The representation that results from sensing a path on the road infrastructure is an object of graph paths that is realized as follows: public class GraphPath { private ArrayList<Node> path; ... } public class GraphPaths { private ArrayList<GraphPath> paths; ... } Graph paths contains a list of paths, each path consisting of a list of nodes. Agents can interpret graph paths to support their decision making. Both for the AGV transportation systems and the vehicle routing system, we only use virtual perceptions. Synchronization of the state of the virtual environment with the ex- ternal environment is realized by the synchronization component (discussed below). An example of a virtual perception in the AGV transportation systems is a perception request to check the status of a path projection (see Fig. 6). An example in the vehicle routing system is a perception request to sense traffic roads with a path focus: public class PathSense( PathFocus pf ) { ... } psense = nes PathSense(pfocus); virtualEnv.sense(id, psense); Action Service is responsible to deal with agents' actions. The π-ADL specification of the action service is defined as follows: component ActionService is abstraction () { type Action is view[agentId : String, actionName : String, actionParams : Any]. type Operation is Any. ... port act is { connection action is in(Action) }. port operate is { connection operation is out(Operation) }. ... actionTypes is location(set(StateItems)). behavior is { analyzeAction is function ( act : Action ) : StateItems { unobservable }. virtualAction is function( state : StateItems, ACM Transactions Journal Format, Vol. 2, No. 3, 09 2001. · 18 D. Weyns and F. Oquendo types :location(set(StateItems)) : Boolean { unobservable }. updateState is function( state: StateItems ) : StateItems { unobservable }. generateOperation is function( state : StateItems ) : Operation { unobservable }. action_items : StateItems. via act receive action_request : Action. action_items := analyzeAction( action ). if virtualAction( action_items, actionTypes ) then { //apply action to local state unobservable. via writeState send updateState( action_items ). behavior(). } else { //apply external action unobservable. via operate send generateOperation( action_items ). behavior(). } } } Situated agents perform actions to realize their tasks. When an agent invokes an action, the action service analyzes the action based on the action types to determine whether the action is concerned with modifying the state of the virtual environment (virtual action) or modifying the state of external resources (external action). In the former case, the local action is executed on the state repository of the virtual environment. In the latter case, an operation is generated and executed. An operation is an action that can be executed in the external environment. An example of a virtual action in the AGV transportation system is an agent that projects a requested path on the map in the AGV local virtual environment (see Fig. 6). public Project(String name, int prio, Hull hull, Path path) { ... } ... project = new Project("project", 0, hull, path); virtualEnv.invoke(id, project); The project action has a name, the priority of the transport currently performed by the requesting AGV, the hull of the AGV, and the path for the projection (i.e., a list of nodes that correspond to segments). The excerpt illustrates how a concrete action is instantiated and then executed on the virtual environment by an agent with identifier id. The virtual environment of in the vehicle routing system does not provide virtual actions. An example of an external action in the vehicle routing system is an vehicle agent that advises the driver about a preferable route. An example in the AGV transportation system is an AGV agent that commands the vehicle to pick a load: public Pick(String name, Node node) { ... } ACM Transactions Journal Format, Vol. 2, No. 3, 09 2001. An Architectural Style for Self-Adaptive Multi-Agent Systems · 19 ... pick = new Pick("pick", n45); virtualEnv.invoke(id, pick); The pick action consists of a name and the node where the load is located. Node is an abstraction of a position on the map of the environment, maintained by the agents. This action is then then converted to an operation for Egemin's AGV steering system E'nsor R(cid:13) (Egemin Navigation System On Robot). For an AGV with a single fork lift, the pick action only requires the segment where the load is located. The physical execution of the action, such as staying on track, turning, and the manipulation of the load is handled by E'nsor. Actually, the control commands provided by the E'nsor interface are coded in a low-level digital format. The virtual environment therefore uses the underlying E'pia R(cid:13) middleware (Egemin Platform for Integrated Automation) to invoke external action, which translates the operations to Ensor control commands. Similarly, to inform a driver about a route in the vehicle routing system, the action is translated to interface commands for a dashboard driver. To support such external actions, the action service can provide supporting functionality to translate action requests to low- level operations. Alternatively, supporting middleware services are available that support action translation. The Communication Service is responsible for managing the communicative interactions among agents. The service receives messages, provides the necessary infrastructure to buffer messages, and deliver messages to the appropriate agents. Communication service is defined as follows: Component CommunicationService is abstraction () { type Message is view[id : Integer, sender : String, receiver : String, protocol : String, performative : String, content : Any]. type Transmission is Any. ... port communicate is { connection sendMsg is in(Message). port transmit is { connection transmission is out(Transmission). connection deliver is in(Transmission) }. connection receiveMsg is out(Message) }. ... addresses is location(set(StateItems)). behavior is { analyzeMessage is function ( msg : Message ) : ItemStates { unobservable }. localCommunication is function( msg_items: StateItems, addrs : location(set(StateItems) ) : Boolean { unobservable }. generateTransmission is function( msg_items : StateItems, addrs : location(set(StateItems) ) : Transmission { unobservable }. analyzeTransmission is function( trans : Transmission, addrs : location(set(StateItems) ) : StateItems { unobservable }. generateMessage is function( trans_items : StateItems, addrs : location(set(StateItems) ) : ACM Transactions Journal Format, Vol. 2, No. 3, 09 2001. · 20 D. Weyns and F. Oquendo Message { unobservable }. ... choose { //send message unobservable. via sendMsg receive message_in. message_items := analyzeMessage(message_in). if localCommunication(message_items, addresses) then //deliver message locally via receiveMsg send message_in. behavior(). else //deliver message remotely transmission_out = generateTransmission(message_state, addresses). via transmission send transmission_out. behavior(). or //deliver message unobservable; via deliver receive transmission_in. transmission_data := analyzeTransmission(transmission_in, addresses). message_out := generateMessage(transmission_data, addresses). via receiveMsg send message_out. behavior(). } } An agent communication message consists of message identifier, the names of the sender and receiver, the message performative (inform, request, propose, etc.), the name of the protocol, and the content of the message. The content is described in a content language which is based on a shared ontology. The ontology defines a vocabulary of words that enables agents to refer unambiguously to concepts and relationships between concepts in the domain when exchanging messages. Such message descriptions enable a designer to express the communicative interactions independently of the underlying communication technology. When the message service receives a message, it analyzes the message and checks whether it concerns a message for a local or a remote addressee. In the former case, it forwards the message to the right agent, in the latter case a transmission is generated and the message is transmitted to the remote agent. To actually transmit the messages, the communication service makes use of a distributed communication system provided by un- derlying middleware. The communication service can provide supporting functions for exchanging messages in a distributed setting, such as white and yellow page services, or alternatively, the service may depend on the underlying middleware for such functions. When the communication service receives a transmission from a remote host, it analy- ses it and generates a message from the received data in the agent message format. The message is then forwarded to the agent. In the AGV transportation systems, we developed two approaches for task assignment, one based on computational fields (we explain this approach below) and another approach in which agents exchange messages [Weyns et al. 2008]. In the message-based approach, transport agents and AGV agents exchange messages with one another to agree on the ACM Transactions Journal Format, Vol. 2, No. 3, 09 2001. An Architectural Style for Self-Adaptive Multi-Agent Systems · 21 assignment of transportation tasks. Here is an example of a message for a call for proposals sent by a transport agent: public Message(int id, String sender, String receiver, String protocol, String performative, Content c) { ... } ... transport = new Content("regular", "5", "n60"); cfp = new Message(77, "ta14", "20m", "DynCNET", "cfp", transport); virtualEnv.send(id, cfp); The identifier of this message is 77. The message is sent by transport agent ta14. Re- ceivers are all AGV agents within a scope of 20 m from the location of the transport. The protocol is DynCNET (dynamic Contract Net [Weyns et al. 2008]) and the performative of the message is cfp (call for proposals). We elaborate on the DynCNET protocol when we discuss the protocol-based communication pattern. The content of the message con- tains 3 elements: regular, i.e., the type of the transport; 5, i.e., the priority of the transport; and n60, i.e., the location of the transport. The receivers use the performative to parse the content of the message correctly. An example of a message in the vehicle routing system is a vehicle agent that sends a reservation ant (i.e., a smart message) to book a particular path along the traffic network. public Message(int id, String protocol, ... String sender, String receiver, String performative, Content c) { } ... booking = new Content(...); ant = new Message(44, "va174", "ia75", "PropagateIntention", "booking", booking); virtualEnv.send(ant); This message with identifier 44 is sent by vehicle agent va147. The first receiver is infrastructure agent ia75. The name of the protocol is propagate intention. The performa- tive says that the message is a booking. The booking contains a sequence of entries that correspond with road segments along the path of the booking, including the names of the infrastructure agents along the path, and timing data. When an infrastructure agent receives the message, it parses the content to extract the data to book the local reservation. Then the agent acknowledges its reservation by marking the entry of its booking in the content of the message. Finally, the agent updates the sender with its id and the receiver with the id of the next infrastructure agent in the list, and forwards the messages. Once the infrastructure agent at the final destination has made its booking it sends the booking message back to the vehicle agent to notify the agent about the booking. Synchronization is responsible for synchronizing state of the virtual environment with external entities, i.e., virtual environments on neighboring nodes as well as external re- sources. The π-ADL specification of synchronization is defined as follows: component Synchronization is abstraction () { type SynchronizationUpdate is Any. ... ACM Transactions Journal Format, Vol. 2, No. 3, 09 2001. · 22 D. Weyns and F. Oquendo port synchronize is { connection syncOut is out(SynchronizationUpdate). connection syncIn is in(SynchronizationUpdate) }. ... synchronizationItems is location( set(StateItems) ). behavior is { selectItems is function( item : location(set(StateItems)) : StateItems { unobservable }. localSync is function( state : StateItems, location( set(StateItems) )) : Boolean { unobservable }. ... generateSyncUpdate is function( state : StateItems ) : SynchronizationUpdate { unobservable }. decodeSyncUpdate is function( update : SynchronizationUpdate) : StateItems { unobservable }. choose { //locally initiated synchronization unobservable. via stateTemplate send selectItems(synchronizationItems). via readState receive state_items : StateItems. if localSync(state_items, synchronizationItems) //local state update via writeState send updateState(state_items). else //external synchronization unobservable. via syncOut send generateSyncUpdate(state_items). behavior(). or //remotely initiated synchronization unobservable. via syncIn receive sync_update : SynchronizationUpdate. via writeState send decodeSyncUpdate(sync_update). behavior(). } } } A synchronization update defines a synchronization interaction with the external envi- ronment. This abstract type has to be instantiated for the domain at hand. A synchroniza- tion can be initiated either locally or remotely. In the former case, a distinction can be made between a synchronization that updates the local state of the virtual environment, and one that triggers an external update. In the latter case, the received synchronization update is decoded and the resulting data is used to update the local state repository. An example of synchronization in the AGV transportation system is the coordination required to resolve conflicts between path projections requested by AGV agents. To re- alize this coordination, the synchronization components of the AGV virtual environments rely on the ObjectPlaces middleware [Schelfthout 2006]. We developed ObjectPlaces to encapsulate the tedious management tasks associated with distribution in mobile applica- tions, and applied it to the AGV transportation system. ObjectPlaces enables to set up an ACM Transactions Journal Format, Vol. 2, No. 3, 09 2001. An Architectural Style for Self-Adaptive Multi-Agent Systems · 23 automatically up-to-date collection of data objects on a set of nodes in the network, and maintain these objects in the face of dynamically changing conditions. To resolve con- flicts between path projections requested by different AGV agents, ObjectPlaces uses a mutual exclusion protocol. For a detailed discussion of path projection coordination with ObjectPlaces, we refer the interested reader to [Schelfthout et al. 2006]. We illustrate synchronization in the AGV transportation with another example. As ex- plained above, besides the message-based approach, we also developed an approach for transport assignment that is based computational fields. These fields are emitted by trans- port agents. The computational fields are distributed data structures spread in through the virtual environment, which attract AGV agents. Each AGV agent of an idle AGV com- bines the sensed fields and follows the gradient of the combined field, guiding its AGV to a load [Weyns et al. 2008]. Fig. 7 shows a snapshot from the simulation of an industrial AGV transportation system in which a number of transport agents emit a field in the virtual environment. The range of the fields is variable and depends on the priority of the tasks. In this case, the synchronization components of the local virtual environments are responsible for maintaining the fields. Fig. 7. Transport fields emitted in the virtual environment This field-based approach employs both types of locally initiated synchronizations. In particular, if a task is not assigned after a predefined amount of time, its priority is in- ACM Transactions Journal Format, Vol. 2, No. 3, 09 2001. Figure6.TransportsemitfieldsinthevirtualenvironmenttoattractidleAGVsCollisionavoidance.AGVagentsmarkthepaththeyaregoingtodriveintheirlocalvirtualenvironmentusinghulls.ThehullofanAGVisthephysicalareatheAGVoccupies.AseriesofhullsdescribethephysicalareaanAGVoccupiesalongacertainpath.Iftheareaisnotmarkedbyotherhulls(theAGV'sownhullsdonotintersectwithothers),theAGVcanmovealongandactuallydriveoverthereservedpath.Incaseofaconflict,theinvolvedlocalvirtualenvironmentsusetheprioritiesofthetransportedloadsandthevehiclestode-terminewhichAGVcanmoveon.AGVagentsmonitorthelocalvirtualenvironmentandonlyinstructtheAGVtomoveonwhentheyareallowed.Afterwards,theAGVagentsre-movethemarkingsinthelocalvirtualenvironment.TheleftpartsofFig.8andFig.9showtwoAGVscoor-dinatetoavoidacollision(weexplaintherightpartofthefiguresbelow).ThesnapshotsshowafusionviewofthetwolocalvirtualenvironmentsthatiscollectedviaremoteaccessofthetwoAGVs.InFig.8twoAGVsarecoordinating.BothAGVsareprojectinghullsinthevirtualenvironment.Atthispointthereisanoverlapofhulls:theAGVontop(boxmarkedwithX)isreadytomovetotheleft,thesmallerAGVatthebottomismanoeuvringtoturnright.TheAGVatthebottomhasalreadyreservedthetrajectoryoccupiedbyitshull,andthusithasprioritytotheAGVatthetopthatmustwait.InFig.9,theAGVatthebottomhaspassedthecurve.ThereisnolongeranoverlapofhullsandtheAGVatthetophasstarteddrivingtotheleft.4.2.3ObjectPlaces:MiddlewareforMobileApplicationsThemobilityoftheAGVsimposehighlydynamicopera-tionconditionsandinherentdistributionofresources.Atyp-icalapproachinmainstreamsoftwareengineeringistosup-portdistributionwithasuitablemiddleware.Wehavede-velopedamiddlewareformobileapplicationscalledOb-jectPlaces(Schelfthout2006).MobileapplicationssuchasanAGVtransportationsystemarecharacterizedby(1)theirneedtotakeintoaccounttheirphysicalenvironment(usu-allycalledcontext)explicitly;and(2)theirneedtodealwithdynamicsandunexpectedeventsoriginatingfromtheircon-· 24 D. Weyns and F. Oquendo creased. Such update is triggered locally by the synchronization component and updates the local state of the field in the virtual environment. On the other hand, a change in priority of a field triggers the synchronization component to spread of field in the virtual environ- ment to neighboring nodes. This is an example of a locally initiated synchronization that triggers an external update. public class FieldData { private int id; private int priority; private Node source; ... } public class TaskField { private Task task; private FieldData fieldData; ... } ... //transport agent creates a field and then performs an emit action fdata = new FieldData(120, 3, n60); field = new TaskField(t60, fdata); emit = new Emit('emit', field); virtualEnv.invoke(ta8, emit); //virtual environment distributes the field objectplace.put(field, callback); ... The excerpt shows how transport agent ta8 located at node n60 creates a field with id 120 and priority 3 for transportation task t60, and then emits the field in the virtual environment. Subsequently, the virtual environment distributes the field to other nodes using the put operation provided by the underlying ObjectPlaces middleware (the put operation is an example of a domain-specific realization of a synchronization update. For emitting a field, only the nodes within a certain distance from the transport agent that emits the field will receive the field. This distance is determined based on the priority of the field (in the field data). When an AGV enters or leaves the range determined by this distance, the virtual environment of the AGV will be notified by the middleware and the field will be added or removed respectively. When the execution of the task is not started after a certain time, the priority of the task will be increased and the virtual environment will update the field in the objectplace. This is turn will trigger the update to the other nodes via the middleware. An example of an externally initiated synchronization in the traffic routing application is the periodic update of the position of a car on the virtual road network. To that end, the synchronization component will use the digital data derived from the position sensor of the car to update the position of the car in the state repository of the virtual environment. 4.1.3 Design Rationale. Three primary principles that underlay the design of the virtual environment are separa- tion of concerns, synchronization processes, the shared data style. Action service, perception service, and communication service provide operations cor- responding to the various ways situated agents can access the virtual environment. By ACM Transactions Journal Format, Vol. 2, No. 3, 09 2001. An Architectural Style for Self-Adaptive Multi-Agent Systems · 25 separating the various concerns, the decomposition of the virtual environment yields a flexible modularization that can be tailored to a broad family of application domains. For example, for applications in which agents interact via marks in the virtual environment but do not communicate via message exchange, the communication service can be omitted. The synchronization component contributes to the flexibility and openness of the multi- agent system. Synchronization can encapsulate processes that maintain a representation of the state of external resources in the state repository, which reflect dynamics in the environment. Agents can perceive these dynamics, which allow them to flexibly adapt their behavior when operating conditions change. Furthermore, synchronization processes can keep track of agents that come and go and reflect this in the state repository of the virtual environment. This indirection reduces the coupling between agents enhancing the openness of the multi-agent system. The shared data style results in low coupling between elements that have clear and dis- junctive responsibilities. This makes the elements better understandable in isolation. Fur- thermore, it improves reuse (elements are not dependent on too many other elements) and modifiability (changes in one element do not affect other elements or the changes have only a local effect). 4.2 Situated Agent 4.2.1 Primary Presentation. The primary presentation of the situated agent pattern is shown in Fig. 8. Fig. 8. Primary presentation of situated agent A situated agent is an autonomous entity that is situated in a virtual environment. A situated agent encapsulates its state and controls its behavior, aiming to realize the tasks it is assigned. Situated agents are able to adapt their behavior according to changing operating ACM Transactions Journal Format, Vol. 2, No. 3, 09 2001. · 26 D. Weyns and F. Oquendo conditions. A situated agent is a cooperative entity. The realization of the overall tasks results from interaction among agents. At the top level, the π-ADL specification of situated agent is defined as follows: component SituatedAgent is abstraction () { //type definitions type Focus is view[focusName : String, focusParams : Any]. type Sense is view[agentId : String, focus : Focus]. type Representation is Any. type Action is view[agentId : String, actionName : String, actionParams : Any]. type Message is view[id : Integer, sender : String, receiver : protocol : String, performative : String, content : Any]. port perceive is { connection sense is out(Sense). connection sensed is in(Representation) }. port act is { connection action is out(Action) }. port communicate is { connection send is out(Message). connection receive is in(Message) }. //composition behavior is compose { perception is Perception() communication is Communication() actionSelection is ActionSelection() currentKnowledge is CurrentKnowledge() and and and } where { perception::sense relays sense perception::request unifies communication::request perception::request unifies actionSelection::request sensed relays perception::sensed communication::send relays send receive relays communication::receive actionSelection::action relays action currentKnowlege::readWriteKnowledge unifies and and and and and and and communication::readWriteKnowledge and currentKnowlege::readWriteKnowledge unifies actionSelection::readWriteKnowledge and currentKnowlege::readWriteKnowledge unifies perception::readWriteKnowledge } } The specification starts with defining types and ports of a situated agent. Then, the com- position of the constituent components of a situated agent is defined. We introduce these components in the following subsection. An elaborated discussion of perception, action selection and communication follows in the next sections that discuss refined patterns for each of these components. 4.2.2 Architectural Elements. The Current Knowledge repository contains data that is shared among the components of a situated agent. The π-ADL specification of current knowledge is defined as follows: component CurrentKnowledge is abstraction() ACM Transactions Journal Format, Vol. 2, No. 3, 09 2001. An Architectural Style for Self-Adaptive Multi-Agent Systems · 27 { type KnowledgeItem is view[name : String, val : Any]. type Knowledge is set(KnowledgeItem). port readWriteKnowledge is { connection knowledgeTemplate is in(Knowledge). connection readKnowlege is out(Knowledge). connection writeKnowledge is in(Knowledge) }. assuming { protocol is (via knowledgeTemplate receive Knowledge. true.via readState send Knowledge.) or (via writeState receive Knowledge.)* }. currentKnowledge is location(set(Knowledge)). behavior is { write is function( knowledge : Knowledge, repository : set(Knowledge) ) : location(set(Knowledge)) { unobservable. }. read is function( template : Knowledge, repository : set(Knowledge) ) : Knowledge { unobservable }. choose { //read knowledge via knowledgeTemplate receive knowledge_template : Knowledge. unobservable. via readKnowledge send read(knowledge_template, currentKnowledge). behavior(). or //write knowledge via writeKnowledge receive knowledge_items : Knowledge. unobservable. currentKnowledge := write(knowledge_items, currentKnowledge). behavior(). } } } Knowledge items are abstract elements that represent something of interest to the agent. Such items can refer to elements of the external environment, elements in the virtual envi- ronment, or internal elements. An agent's current knowledge is maintained in a repository that provides a shared port that allows components to read and write knowledge. The pro- tocol of the read write port enforces the atomic completion of either reading or writing the repository. Reading knowledge requires a template and after processing the requested knowledge is returned. Reading and writing are further specified in the behavioral part of the definition. Examples of knowledge in the vehicle routing system are a path along the road network represented as a graph of nodes (see the definition of graph path in section 4.5.2 where we discuss the perception service), the current position of the car as a node of the graph, the status of the traffic in the monitored area, bookings for road segments, etc. An example of knowledge in the AGV transportation systems is operating space that an AGV agent uses to keep track of its path projections in the virtual environment. ACM Transactions Journal Format, Vol. 2, No. 3, 09 2001. · 28 D. Weyns and F. Oquendo public class OperatingSpace { private Path requestedPath; private Path lockedPath; ... public void locked(Path p) { ... } public void release(Path p) { ... } ... } An operating space keeps track of a requested path and a locked path. The requested path has been projected in the virtual environment, but is not locked yet (at least the AGV agent has not observed this). The locked path is reserved for the AGV to move safely. Operation space provides a lock operation to add a projected path to the locked path once it is locked, and an operation to release the path behind the AGV when it moves. Perception is responsible for collecting data for the agent. This data concerns either state of the virtual environment or state of the environment. The π-ADL specification of per- ception is defined as follows: component Perception is abstraction() { type Focus is view[focusName : String, focusParams : Any]. type Filter is view[filterName : String, filterParams : Any]. type PerceptionRequest is view[id : String, focus : Focus, filter : Filter]. type Sense is view[agentId : String, focus : Focus]. type Representation is Any. ... port request is { connection requestPerception is in(PerceptionRequest). port perceive is { connection sense is out(Sense). connection notifyRequest is out(String)}. connection sensed is in(Representation) }. ... behavior is { generateSense is function( p_request : PerceptionRequest) : Sense { unobservable }. via requestPerception receive request : PerceptionRequest. via sense send generateSense(request). unobservable. via sensed receive representation : Representation. unobservable. via writeKnowledge send update : Knowledge. via notifyRequest send request::id. behavior(). } } The perception component enables an agent to sense the virtual environment for infor- mation. To support selective perception, a perception request includes a selected focus and filter. We explained the concept of a focus in section 4.5.2. A filter allows an agent to select only those elements of a representation that match a particular selection criteria. ACM Transactions Journal Format, Vol. 2, No. 3, 09 2001. An Architectural Style for Self-Adaptive Multi-Agent Systems · 29 Perception requests can be triggered by the communication component and the action selection component. When perception receives a request, it generates a sense that is used to sense the virtual environment, resulting in a representation. Perception processes the representation, updates the agent's current knowledge, and then notifies the perception requester. An example of a filter in the AGV transportation system is a filter to select the nearest transportation task that resulted from sensing the tasks within a certain distance. An exam- ple in the vehicle routing system is a filter to select the shortest path from a set of sensed paths: public ShortestPath(Sting name, GraphPaths paths) { ... } pfilter = new ShortestPath('shortestPath', sensedPaths); This filter selects the shortest path from a given set of sensed paths. The representation of graph paths is discussed in section 4.5.2. We elaborate on the perception when we discuss the selective perception pattern. Behavior-Based Action Selection is responsible for realizing the agent's objectives by invoking actions in the virtual environment. Actions either intend to manipulate the state of the virtual environment or manipulate external resources. The π-ADL specification of action selection is defined as follows: component ActionSelection is abstraction() { type Action is view[agentId : String, actionName : String, actionParams : Any]. ... port act is { connection action is out(Action) }. behavior is { unobservable. via action send action : Action. behavior(). } } After the definition of types and ports, the abstract behavior of action selection is defined. During internal processing, action selection can decide to select an action and subsequently, the selected action is invoked in the virtual environment. The following excerpt illustrates how an action is realized in the vehicle routing system: public InstructDriver(String name, GraphPath path) { ... } ... intention = new InstructDriver('intention', path); virtualEnv.invoke(id, intention); The action advises the driver with a preferable path in the environment. For examples of actions in the AGV transportation system, we refer to the discussion of the action ser- vice in section 4.5.2, where we discuss several examples. We further elaborate on decision making when we discuss the behavior-based action selection pattern. ACM Transactions Journal Format, Vol. 2, No. 3, 09 2001. · 30 D. Weyns and F. Oquendo Communication is responsible for communicative interactions with other agents. Mes- sage exchange enables agents to share information and set up collaborations. The π-ADL specification of communication is defined as follows: component Communication is abstraction () { type Message is view[id : Integer, sender : String, receiver : String, protocol : String, performative : String, content : Any]. ... port communicate is { connection sendMsg is out(Message). connection receiveMsg is in(Message) }. behavior is { choose { //receive message via receiveMsg receive message : Message. unobservable. behavior(). or //send message unobservable. via sendMsg send message : Message. unobservable. behavior(). or //silence unobservable. behavior(). } } } After type and port declarations, the behavior of communication is defined. The com- munication component either receives a message that is internally processed, or it sends a message after internal processing, or its behavior is silent (e.g., a conversation is termi- nated without further notice). We have presented concrete realizations of messages for both the vehicle routing system and the AGV transportation system in section 4.5.2, where we discuss the communication service. We elaborate on communication and give a concrete example of a protocol when we explain the protocol-based communication pattern. 4.2.3 Design Rationale. The primary principles that underlay the design of situated agent are decentralized control and locality, perception on demand, and parallelism of action selection and communica- tion. In a situated multi-agent system, control is decentralized and divided among the agents. Situated agents manage the dynamic and changing operating conditions locally and au- tonomously, enabling them to flexibly adapt their behavior when needed. However, decen- tralized control (the system tasks are realized by agents that make local decisions) typically requires more communication. The performance of the system may be affected by the com- munication links between agents, which could impose a major bottleneck. Furthermore, agents' decision making is based on local information, which may result in suboptimal ACM Transactions Journal Format, Vol. 2, No. 3, 09 2001. An Architectural Style for Self-Adaptive Multi-Agent Systems · 31 system behavior. These tradeoffs and should be kept in mind throughout the design and development of a situated multi-agent system. Selective perception enables an agent to focus its attention to the relevant aspects in the environment according to its current tasks. When selecting actions and communicating messages with other agents, decision making and communication typically request per- ceptions to update the agents knowledge. By selecting an appropriate focus and filter, the agent directs its attention to the current aspects of its interest, and adapts it attention when the operating conditions change. The overall behavior of a situated agent is the result of the coordination of action se- lection and communication. Action selection is responsible for selecting suitable actions to perform tasks. Communication is responsible for the communicative interactions with other agents. However, the two components coordinate to complete the agent's tasks more efficiently. For example, agents can send each other messages with requests for information that enable them to act more purposefully, or support each other in a collaboration. En- suring that both action selection and communication behave in a coordinated way requires a careful design. On the other hand, the separation of functionality for coordination (via communication) from the functionality to perform actions to complete tasks has several advantages. Besides improving reusability, it allows both functions to act in parallel and at a different pace. In many applications, sending messages and executing actions happen at different tempo. A typical example domain is robotics, but it applies to any application in which the time required for performing actions in the environment differs significantly from the time to communicate messages. Separation of communication from performing actions enables agents to reconsider the coordination of their behavior while they perform actions, improving adaptability and efficiency. 4.3 Selective Perception 4.3.1 Primary Presentation. The primary presentation of the selective perception pattern is shown in Fig. 9. Selective perception is responsible for collecting runtime information to update the agent's current knowledge. Perception requests can be initiated by the communication component and the action selection component. The π-ADL specification of communica- tion is defined as follows: component Perception is abstraction() { type Focus is view[focusName : String, focusParams : Any]. type Filter is view[filterName : String, filterParams : Any]. ... port perceive is { connection sense is out(Sense). connection sensed is in(Representation) }. port request is { connection requestPerception is in(PerceptionRequest) connection notifyRequest is out(String) }. port readWriteKnowledge is { connection knowledgeTemplate is out(Knowledge). connection readKnowledge is in(Knowledge). connection writeKnowledge is out(Knowledge) }. behavior is compose { sensing is Sensing(). ACM Transactions Journal Format, Vol. 2, No. 3, 09 2001. · 32 D. Weyns and F. Oquendo Fig. 9. Primary presentation of selective perception and interpreting is Interpreting(). and descriptions is Descriptions(). and filtering is Filtering(). } where { requestPerception relays sensing::applySensing and sensing::sense relays sense and sensed relays sensing::sensed and sensing::interpret unifies interpreting::interpret and interpreting::readDescriptions unifies descriptions::readDescriptions and interpreting::filter unifies filtering::filter and requestPerception relays filtering::requestPerception and filtering::readWriteKnowledge relays readWriteKnowledge and filtering::notifyRequest relays notifyRequest } } After type and port declarations, the composition of the constituent components of per- ception is defined. We now discuss these components further in detail. 4.3.2 Architectural Elements. The perception components comprises one data repository: descriptions, and three com- ponents: sensing, interpreting, and filtering. To explain the various elements, we follow the logical thread of successive activities that take place from the moment a requester takes the initiative to sense the virtual environment until the agent's current knowledge is updated and the requester is notified. Sensing is defined as follows: component Sensing is abstraction() { ACM Transactions Journal Format, Vol. 2, No. 3, 09 2001. An Architectural Style for Self-Adaptive Multi-Agent Systems · 33 ... type PerceptionRequest is view[id : String, focus : Focus, filter : Filter]. type Sense is view[agentId : String, focus : Focus]. type Representation is Any. ... port perceive is { connection sense is out(Sense). port apply is { connection applySensing is in(PerceptionRequest) }. port interpret is { connection percept is out(Representation) }. connection sensed is in(Representation) }. behavior is { generateSense is function(request : PerceptionRequest) : Sense { unobservable. }. via applySensing receive request : PerceptionRequest. via sense send generateSense(request). unobservable. via sensed receive representation : Representation. unobservable. via percept send representation. behavior(). } } Sensing is initiated by a perception request with a focus. Sensing applies sense requests in the virtual environment and sends the resulting representation to the interpreting com- ponent. We have explained and illustrated the concepts of focus and representation in the discussion of the virtual environment pattern, see section 4.5.2. Descriptions is defined as follows: component Descriptions is abstraction() { type Representation is Any. type Description is Any. port readDescriptions is { connection representation is in(Representation). connection description is out(set(Description)) } assuming { protocol is { (via representation receive Any.true. via description send Any.) } }. descriptions is location(set(Description)). behavior is { select is function(rep : Representation, des : set(Description) ) : set(Description) { unobservable. }. unobservable. via representation receive repres : Representation. via description send select(repres, descriptions). behavior(). } } Description maintains a repository of descriptions to interpret representations. A de- scription specifies a particular pattern of the data contained in a representation. An exam- ACM Transactions Journal Format, Vol. 2, No. 3, 09 2001. · 34 D. Weyns and F. Oquendo ple in the vehicle routing system is a description used by infrastructure agents to map a representation of the traffic state to a level of traffic congestion in a monitored range: public class TrafficObservation { private double density; private double intensity; private double averageSpeed; ... } public class CongestionLevel { private TrafficObservation observation: ... public CongestionLevel(TrafficObservation o) { ... } } Traffic observation is a representation of the current traffic in a monitored range. This representation comprises three traffic variables in the monitored range: the current density (i.e., the number of vehicles per length unit), intensity (i.e., the number of vehicles per time unit) and average speed. Given a traffic representation, the congestion level description allows to determine the current the traffic state in the monitored range. Interpreting is defined as follows: component Interpreting is abstraction() { type Representation is Any. type Description is Any. type Knowledge is set(KnowledgeItems). ... port interpret is { connection interpretIn is in(Representation) }. port readDescriptions is { connection representation is out(Representation). connection description is in(set(Description)) }. port filter is { connection interpretOut is out(Knowledge) }. behavior is { interpreting is function(r : Representation, ds : set(Description)) : Knowledge { unobservable. }. via interpretIn receive repres : Representation. via representation send repres. unobservable. via description receive descriptions : Description. via interpretOut send interpreting(repres, descriptions). behavior(). } } The interpreting component uses a set of descriptions to extract knowledge from a rep- resentation. This knowledge represent elements sensed in the virtual or the external envi- ronment that can be used to update the current knowledge of the agent. An example in the traffic monitoring system is the traffic state derived from a traffic observation using a congestion level description: public class TrafficState { ACM Transactions Journal Format, Vol. 2, No. 3, 09 2001. An Architectural Style for Self-Adaptive Multi-Agent Systems · 35 private GraphPath path; private Status status; ... } The path represents the part of the road network that is monitored by an infrastructure agent. Status is the actual status of the traffic on the path, ranging from free flow to jammed. The infrastructure agent employs the traffic state to provide predictive traffic intensity in- formation which is used to improve bookings requested by vehicle agents. For details on this prediction mechanism, we refer the interested reader to [Claes et al. 2011]. Filtering is defined as follows: component Filtering is abstraction() { ... type Filter is view[filterName : String, filterParams : Any]. type PerceptionRequest is view[id : String, focus : Focus, filter : Filter]. port filter is { connection filterIn is in(Knowledge) }. port apply is { connection applyFiltering is in(PerceptionRequest). connection notifyRequest is out(String) }. port readWriteKnowledge is { connection knowledgeTemplate is out(Knowledge). connection readKnowledge is in(Knowledge). connection writeKnowledge is out(Knowledge) }. behavior is { filtering is function(percept : Knowledge, filter : Filter) : Knowledge { unobservable. }. unobservable. via applyFiltering receive request : PerceptionRequest. unobservable. via filterIn receive percept : Knowledge. via writeKnowledge send filtering(percept, request::filter). via notifyRequest send request::id. behavior(). } } Filtering filters a percept using a selected filter. A filter imposes conditions on a percept that determine whether the data elements of the percept can pass the filter or not. The filtering component uses the filtered percept to update the agent's current knowledge, and then notifies the requester. We discussed an example of a filter in the vehicle routing system that selects the shortest path from a given set of sensed paths. Here is an example in the AGV transportation system: public ParkLocation(Sting name, Node pos, List<Node> locations) { ... } tfilter = new ParkLocation('nearestParking', myPos, sensedLocs); This filter selects the nearest park location for the given AGV position from a given set of observed park locations. ACM Transactions Journal Format, Vol. 2, No. 3, 09 2001. · 36 D. Weyns and F. Oquendo 4.4 Behavior-Based Action Selection 4.4.1 Primary Presentation. The primary presentation of the behavior-based action selection pattern is shown in Fig. 10. Fig. 10. Primary presentation of behavior-based action selection To select actions, a situated agent employs a behavior-based action selection mechanism. The main advantages of behavior-based action selection are efficiency and flexibility to deal with dynamism in the environment. The π-ADL specification of action selection is defined as follows: component ActionSelection is abstraction() { ... type Action is view[agentId : String, actionName : String, actionParams : Any]. port act is { connection invoke is out(Action) }. port request is { connection requestPerception is out(PerceptionRequest). connection notifyRequest is in(String) }. port readWriteKnowledge is { ... }. behavior is compose { decisionMaking is DecisionMaking(). and roles is Roles(). and commitments is SituatedCommitments(). } where { decisionMaking::updateReadRoles unifies roles::updateReadRoles ACM Transactions Journal Format, Vol. 2, No. 3, 09 2001. An Architectural Style for Self-Adaptive Multi-Agent Systems · 37 and decisionMaking::readWriteCommitments unifies commitments::readWriteCommitments decisionMaking::act relays act decisionMaking::knowledgeTemplate relays knowledgeTemplate and and ... } } Central to behavior-based action selection is the decision making component that em- ploys roles and situated commitments to select actions. A role represents a coherent part of functionality of a situated agent. Roles are relatively simple behavioral modules that tightly couple sensing to action. A situated commitment defines an engagement of an agent in a collaboration that takes into account the context of the collaboration. Decision making se- lects the next action of the agent based on the agent's roles and the current commitments. We discuss roles and situated commitments now in detail. 4.4.2 Architectural Elements. The action selection component comprises three components: roles, situated commit- ments, and behavior. Roles is defined as follows: component Roles is abstraction() { type Role is view[name : String, behaviors : Any, actions : set(Action)]. port updateReadRoles is { connection updateRoles is in(Knowledge). connection readRoles is out(set(Role)). } roles is set(Role). behavior is { update is function(k : Knowledge, roles : set(Role)) : set(Role) { unobservable }. choose { //update roles via updateRoles receive knowledge : Knowledge. unobservable. roles := update(knowledge, roles). behavior(). or //read roles unobservable. via readRoles send(roles). behavior(). } } The roles component comprises a set of roles. A role has a name, a set of behaviors, and a set of actions. The roles can be updated with new knowledge, and the set of roles can be ACM Transactions Journal Format, Vol. 2, No. 3, 09 2001. · 38 D. Weyns and F. Oquendo read. The main roles of the AGV agent in the AGV transportation system are "Working," "Charging" and "Parking," which respectively denote the roles to perform a transportation task, recharge the AGV's battery, and park the vehicle when it is idle. For the concrete realization of the roles we used an extension of free-flow trees [Tyrrell 1993; Bryson 2001; Steegmans et al. 2004]. Fig. 11 shows the design of the tree for the working role of an AGV agent. Fig. 11. Design of the free-flow tree for the working role of AGV agents A free-flow tree is composed of a hierarchy of nodes representing the role behavior with leave nodes representing actions. To select an action, activity is injected at the top node of the tree. While the activity flows along the nodes additional activity may be injected based on stimuli sensed by the agent (ovals in Fig. 11). When the activity arrives at the leave nodes the action with the highest activity is the preferable action for the role in the given context. The main behaviors of the working role are moving, performing a job (pick or drop a load), and finishing job (i.e., the actual manipulation of the load during a pick or drop). The behaviors are affected by stiumuli related to the actual position of the AGV to the pick or drop location, and the reservation of hulls to avoid collisions. Skip move is selected when no new move action is required (e.g, when the AGV waits on a crossroad), and skip finished is selected during the manipulation of a load. Commitments is defined as follows: component Commitments is abstraction() { type SituatedCommitment is view[name : String, context : Any, condition : Boolean, effect : Any]. port updateReadCommitments is { connection updateCommitments is in(Knowledge). connection readCommitments is out(set(Commitment)). ACM Transactions Journal Format, Vol. 2, No. 3, 09 2001. An Architectural Style for Self-Adaptive Multi-Agent Systems · 39 } commitments is set(SituatedCommitment). behavior is { update is function(k : Knowledge, cs : set(SituatedCommitment)) : set(SituatedCommitment) { unobservable }. choose { //update commitments via updateCommitments receive knowledge : Knowledge. unobservable. roles := update(knowledge, commitments). behavior(). or //read commitments unobservable. via readCommitments send(commitments). behavior(). } } The commitments component comprises a set of situated commitments. A situated com- mitment is defined by a name, a context, an activation condition, and an effect. Com- mitments can be updated with new knowledge, and the set of commitments can be read. Agents agree on mutual commitments in a collaboration via direct communication (see the Communication pattern below). Once the agents have agreed on a collaboration, the mu- tual situated commitments will affect the selection of actions in favor of the agents roles in the collaboration. An example of a situated commitment in the vehicle routing system is "current inten- tion" that represents the booking of a path a vehicle agent at the infrastructure agents. This commitment will guide the driver along the selected path to the destination. The main commitments in the AGV transportation system are "Work" and "Charge." Work is a mu- tual commitment between an AGV agent and a transport agent to perform a transportation task. Charge is commitment of an AGV agent to charge its battery enabling it to perform future transportation tasks. Such commitment contributes implicitly to the realization of the system tasks. Fig. 12 shows the design of the work commitment for AGV agents. A situated commitment in the AGV transportation system is represented by a connec- tor between a set of source roles (in) and one target role (out). When activated, addi- tional activity is injected to the target role, proportional to the activity levels of the source roles [Steegmans et al. 2004; Weyns et al. 2008]. The work commitment connects the parking role (single source role) with the working role (target role) and is activated when the AGV has accepted a job (these elements define the context and condition of the situated commitment). When the work commitment is activated it injects an additional amount of activity in the working role (i.e., the activity of the top node of the parking role is added to the top node of the working role). This additional activity will stimulate the working role to start performing the committed task. Decision Making is defined as follows: ACM Transactions Journal Format, Vol. 2, No. 3, 09 2001. · 40 D. Weyns and F. Oquendo Fig. 12. Design of the work commitment of AGV agents component DecisionMaking is abstraction() { type Role is view[name : String, behaviors : Any, actions : set(Action)]. type SituatedCommitment is view[name : String, context : Any, condition Boolean, effect : Any]. type Action is view[agentId : String, actionName : String, actionParams : Any]. ... port updateReadRoles is { connection updateRoles is in(Knowledge). connection readRoles is out(set(Roles)). } port updateReadCommitments is { connection updateCommitments is in(Knowledge). connection readCommitments is out(set(SituatedCommitment)). } port act is { connection selectAction is out(Action).}. ... externalActions is location(set(Action)). selectedAction is Action. behavior is { selectAction is function (roles : set(Role), commitments : set(SituatedCommitment) : Action { unobservable. }. externalAction is function (action : Action, actions : set(Action)) : Boolean { unobservable. }. unobservable. via updateRoles send roleKnowledge : Knowledge. via updateCommitments send commitmentKnowledge : Knowledge. via readRoles receive roles : set(Roles). via readCommitments receive commitments : set(SituatedCommitment). selectedAction := selectAction(roles, commitments). if (externalAction(selectedAction, externalActions)) then ACM Transactions Journal Format, Vol. 2, No. 3, 09 2001. An Architectural Style for Self-Adaptive Multi-Agent Systems · 41 via selectAction send selectedAction. behavior(). else unobservable. behavior(). } } The decision making component has a repository with external actions that define effec- tive actions that can be invoked in the virtual environment. Actions that are not included in this set are internal to the agent and may be used to update the agent's current knowledge. Decision making starts with updating the state of the roles and situated commitments based on the agent's current knowledge. Next, the roles and situated commitments are read and an action is selected. Subsequently, the component checks whether the selected action is an external action, and if so the action it is invoked. We illustrate the realization of the action selection function for the two example applications. We start with an excerpt of the decision making of an AGV agent that integrates the roles and situated commitments illustrated in Fig. 11 and Fig. 12. To manage the complexity, we designed behavior-based action selection of the AGV agent in two subsequent steps. In the first step, a free-flow tree is used to select an action at a high-level of abstraction. Examples of such actions are move, pick, and park. In the next step, the selected action is further refined taking into account collision avoidance and deadlock avoidance. For example, for a move action, the next segment is added if the corresponding path projection has been locked. We illustrate here the fist step, for additional information, we refer the interested reader to [Weyns and Holvoet 2007; Weyns 2010]. Fig. 13 shows a snapshot of the main part of the free-flow tree of an AGV agent in action. Fig. 13. Runtime snapshot of the free-flow tree of an AGV agent. The part of the tree that deals with battery charging is omitted. ACM Transactions Journal Format, Vol. 2, No. 3, 09 2001. · 42 D. Weyns and F. Oquendo The action selection function of the AGV agent is executed every second. To select an action, activity is injected via the root node of the tree. The root node feeds its activity to the working node and the parking node. These nodes feed the activity down through the hierarchy until the activity arrives at the action nodes, i.e. the leaf nodes of the tree. The level of activity that flows from one node to another may be multiplied by a weight factor that is associated with the edge. For example, the connection between the root node and the top node of the parking role has a weight factor of 2. The colors of the nodes in Fig. 13 give a visual representation on the level of activities of the different parts of the behavior (the exact values are indicated next to each node). While the activity flows through the tree, some of the nodes receive additional activity from stimuli, e.g., the nodes just beneath the working role (R2 and R3) receive additional activity from the nearJob stimulus. When activated, the situated commitment work, that connects the parking role with the working role, injects an additional amount of activity in the working role (equal to the activity in the parking role). In the situation shown in Fig. 13, the situated commitment is active and as a result the activity level of the working role is 3, i.e., the sum of the activity 1 from the root node and 2 from the parking role via the situated commitment. When all action nodes have collected their activity the node with the highest activity level is selected for execution. In the example, the activity collected by the pick node is dominant (value 33) so the agent will select this action. The concrete node where to perform the action will be determined in the second step of action selection (see also the definition of the pick action in the discussion of the action service in section 4.5.2). We conclude the discussion of the behavior-based action selection pattern with an ex- cerpt of the design that illustrates the decision making of the vehicle agents in the antici- patory vehicle routing application. Figure 14 shows the subsequent actions selected by the vehicle agent. 1: loop 2: 3: 4: 5: 6: 7: 8: 9: 10: 11: 12: 13: 14: end loop // Update the feasible paths paths ← getAlternatives(currentLocation, destination) // Select the best route selectedP ath ← choose(paths) // Decide whether to revise intention if reviseIntention(currentIntention, selectedP ath) then currentIntention ← selectedP ath end if // Trigger propagation of the current intention propagateIntention(currentIntention) // Instruct the driver instructDriver(currentIntention) Fig. 14. Decision making of a vehicle agent in the anticipatory vehicle routing system The behavior of the vehicle agent is contained in a single role, with the current intention (selected path) as a single type of situated commitment. On line 3, the vehicle agent updates its knowledge with feasible paths explored by exploration ants. On line 5, the agent selects the shortest path based on the set of feasible paths. On line 7, the vehicle agent decides on whether to deviate to this new path. Whatever the agent decides, on line 11, the vehicle agent triggers sending out an intention ant to the infrastructure agents along the path of its ACM Transactions Journal Format, Vol. 2, No. 3, 09 2001. An Architectural Style for Self-Adaptive Multi-Agent Systems · 43 current intention. Finally, the driver is instructed about the current intention. The interested reader finds additional information in [Weyns et al. 2007; Claes et al. 2011]. 4.4.3 Design Rationale. The primary principles that underlay the design of action selection are behavior-based decision making, role and situated commitment as first class concepts. Behavior based decision making enables an agent to act efficiently. A situated agent se- lect its actions based on the locally perceived state of the environment and limited internal state. The reactive nature of behavior based decision making enables an agent to respond efficiently when the conditions in the system and the environment change. Nevertheless, locality and reactivity of decision making may affect that overall efficiency due to a lack of global and long term view on the system. Behavior based decision making is therefore suitable for interaction driven problem domains, rather than data driven domains. This fits the target domain of the architecture style. The design of a behavior-based decision making mechanism may be a complex task for non-trivial agents. To manage this complexity, we introduced roles and situated commit- ments for situated agents. Roles and situated commitments provide design abstractions that allow to structure the agent behavior. Furthermore, roles and situated commitments pro- vide the means for agents to set up collaborations in a particular context. A role represents a coherent part of functionality that can serve as a building block for social organization of an agent system. A situated commitment is defined based on the context in which agents are situated, which typically refers to their roles. Explicitly naming roles and situated com- mitments enables agents to set up collaborations, reflected in mutual commitments. Such collaboration can be set up by exchanging messages (see the Protocol-Based Communi- cation pattern below). Once a collaboration is established, the mutual commitments will affect the selection of actions in favor of the agents' roles in the collaboration. The so- cial behavior is typically driven by the context in which agents are situated (e.g., in field based transport assignment, agents are driven by the fields that are emitted in the virtual environment by the tasks). The latter results in high degree of openness (when a task or AGV enters the system, the agents take into account the new field, and when a task or AGV leaves the system the field simply disappears). 4.5 Protocol-Based Communication 4.5.1 Primary Presentation. The primary presentation of the protocol-based communication pattern is shown in Fig. 15. Protocol-based communication enables situated agents to share information and set up collaborations. The communication component processes incoming messages, and pro- duces outgoing messages according to well-defined communication protocols. A commu- nication protocol specifies a set of possible sequences of messages. The π-ADL specifica- tion of the communication component is defined as follows: component Communication is abstraction() { type Message is view[id : Integer, protocol_ : String, sender : String, protocol : String, performative : String, content : Any]. ... ACM Transactions Journal Format, Vol. 2, No. 3, 09 2001. · 44 D. Weyns and F. Oquendo Fig. 15. Protocol-based communication pattern port communicate is { connection send is out(Message). connection receive is in(Message) }. ... behavior is compose { messageReceiving is MessageReceiving(). inbox is Inbox(). messageDecoding is MessageDecoding(). communicating is Communicating(). conversatations is Conversations(). messageEncoding is MessageEncoding(). outbox is Outbox(). messageSending is MessageSending(). and and and and and and and } where { receive relays messageReceiving::receive messageReceiving::addIn unifies inbox::addIn messageDecoding::removeIn unifies inbox::removeIn messageDecoding::decode unifies communicating::decode communicating::knowledgeTemplate relays knowledgeTemplate readKnowledge relays communicating::readKnowledge communicating::writeKnowledge relays writeKnowledge communicating::requestPerception relays requestPerception and and and and and and and ACM Transactions Journal Format, Vol. 2, No. 3, 09 2001. An Architectural Style for Self-Adaptive Multi-Agent Systems · 45 and communicating::readUpdateConversation unifies conversations::readUpdateConversation communicating::encode unifies messageEncoding::encode messageEncoding::addOut unifies outbox::addOut messageSending::removeOut unifies outbox::removeOut messageSending::send relays send and and and and } } Message receiving accepts messages from the communication service and adds them to the inbox message buffer. Message decoding selects messages from the inbox and decodes the messages. The communicating component interprets the decoded messages and up- dates the agent knowledge and conversations. Communicating also initiates and continues conversations when the necessary conditions hold (based on the agent's current knowledge and the ongoing conversations). Ongoing conversations are stored in the conversations repository. A conversation is a communicative interaction following a well-defined com- munication protocol that defines of a series of protocol steps. Message encoding encodes newly composed messages and add them to the outbox message buffer. Finally, message sending selects pending messages from the outbox buffer and passes it to the communica- tion service of the virtual environment. 4.5.2 Architectural Elements. We limit the discussion to the central components: message decoding, conversations, and communicating. Message encoding is basically the inverse of message decoding. The other components, message receiving, inbox, outbox, and message sending, have standard functions. The online Appendix A provides a rigorous definition of these components. Message Decoding is defined as follows: component MessageDecoding is abstraction() { ... type Message is view[id : Integer, sender : String, receiver : String, protocol : String, performative : String, content : Any]. type MessageData is view[id : Integer, sender : String, receiver : String, protocol : String, performative : String, content : sequence[Knowledge]]. port removeInbox is { connection removeIn is in(Message) }. port decodeMessage is { connection decode is out(MessageData) }. language : Any. ontology : Any. behavior is { decode is function( msg : Message, ontology : Any, language : Any ) : MessageData { unobservable. }. unobservable. via removeIn receive message : Message. via decodeMsg send decode(message, language, ontology). behavior(). } } ACM Transactions Journal Format, Vol. 2, No. 3, 09 2001. · 46 D. Weyns and F. Oquendo Message data is data extracted from a message in a form that can be interpret by the communicating component. Message decoding selects a received message from the inbox buffer and decodes the message according to the given content language and ontology. The content language defines the format of the messages, i.e. the subsequent fields the message is composed of, and the ontology defines a shared vocabulary of words that agents use to represent domain concepts and relationships between the concepts. As an example, in the AGV transportation system, we defined a content language for the messages of the DynCNET protocol that allows tasks agents and AGV agents to agree on task assignment. DynCNET offers an alternative to field-based task assignment in the AGV transportation system, which we illustrated in section 4.5.2. We elaborate on the protocol below. The content language defines for each performative a set of strings that represent fields of the corresponding messages. The ontology defines the domain of values for each field. For example, the domain of the task priority field defines 5 priority levels represented as integers. During message decoding, the content of the message is parsed based on the performative of the message using the content language and the ontology. Conversations is defined as follows: component Conversations is abstraction() { ... type MessageData is view[id : Integer, sender : String, receiver : String, protocol : String, performative : String, content : sequence[Knowledge]]. type Conversation is sequence[MessageData]. port readUpdateConversation is { connection selectConversation is in(Integer). connection readConversation is out(Converstation). connection updateConversation is in(MessageData). connection addConversation is in(MessageData). connection terminateConversation is in(Integer) }. type conversations is location(set(Conversation)). behavior is { add is function(data : MessageData, convs : set(Conversation) ) : set(Conversation) { unobservable. }. write is function( id : Integer, convs : set(Conversation) ) : set(Conversation) { unobservable. }. read is function( conv : Conversation, conv : Conversation ) : set(Conversation) { unobservable. }. terminate is function(id : Integer, convs : set(Conversation) ) : set(Conversation) { unobservable. }. choose { //add conversation via addConversation receive data : MessageData. conversations := add(data, conversations). behavior(). or ACM Transactions Journal Format, Vol. 2, No. 3, 09 2001. An Architectural Style for Self-Adaptive Multi-Agent Systems · 47 //read conversation via selectConversation receive id : Integer. via readConversation send readConversation(id, conversations). behavior(). //write conversation via updateConversation receive conversations := write(data, conversations). behavior(). data : MessageData. //delete conversation via terminateConversation receive id : Integer. conversations := terminate(id, conversations). behavior(). or or or } } } A conversation is a sequence of message data. The conversation component maintains a set of ongoing conversations. Conversations offers four functions that allows the com- municating component to manage the conversations repository. Add includes a new con- versation to the set of conversations, read returns a conversation based on a given id, write updates an ongoing conversation, and delete removes the conversation with a given id from the set of conversations. Communicating, the central component of the communication pattern, is defined as fol- lows: component Communicating is abstraction () { type MessageData is view[id : Integer, sender : String, receiver : String, protocol : String, performative : String, content : sequence[Knowledge]]. type Conversation is sequence[MessageData]. type Protocol is view[name : String, steps : Any]. ... port readUpdateConversation is { ... }. port decodeMessage is { connection decode is in(MessageData) }. port encodeMessage is { connection encode is out(MessageData) }. ... protocols : set(Protocol). conversationIds : set(Integer). behavior is { extractKnowledge is function (data : MessageData) : Knowledge { unobservable. }. addCommunication is function (data : MessageData) : Knowledge { unobservable. }. updateCommunication is function (data : MessageData) : Knowledge { unobservable. }. closeCommunication is function (c : Conversation) : Knowledge { unobservable. }. ongoingConversation is function (id : Integer, ids : set(Integer)) : Boolen { unobservable. }. ACM Transactions Journal Format, Vol. 2, No. 3, 09 2001. · 48 D. Weyns and F. Oquendo continueConversation is function (k : Knowledge, ids : set(Integer)) : Boolen { unobservable. }. terminateConversation is function (ps : set(Protocol), c: Conversation, k : Knowledge) : Boolean { unobservable. }. initiateConversation is function (ps : set(Protocol), k : Knowledge) : MessageData { unobservable. }. encodeMessage is function (ps : set(Protocol), c : Conversation, k : Knowledge) : MessageData { unobservable. }. getId is function (k : Knowledge, ids : set(Integer)) : Integer { unobservable. }. addId is function (id : Integer, ids : set(Integer)) : set(Integer) { unobservable. }. msg_knowledge : Knowledge. enc_msg_data : Messagedata. conv_id : Integer. choose { //handle incoming message via decode receive dec_msg_data : MessageData. msg_knowledge := extractKnowledge(dec_msg_data). if(ongoingConversation(dec_msg_data::id, conversationIds)) then { via updateConversation send dec_msg_data. via writeKnowledge send updateCommuncation(dec_msg_data). } else { via addConversation send dec_msg_data. via writeKnowledge send addCommuncation(dec_msg_data). } behavior(). } or { //handle outgoing message unobservable. via readKnowledge receive knowledge : Knowledge. if(continueConversation(knowledge, conversationIds)) then { via selectConversation send getId(knowledge, conversationIds). via readConversation receive conversation : Conversation. enc_msg_data := encodeMessage(protocols, conversation, knowledge). via encode send enc_msg_data. via updateConversation send enc_msg_data. via writeKnowledge send updateCommunication(enc_msg_data). } else { enc_msg_data := initiateConversation(protocols, knowledge). via addConversation send enc_msg_data. conversationIds := addId(id, conversationIds). via writeKnowledge send addCommunication(enc_msg_data). } behavior(). } ACM Transactions Journal Format, Vol. 2, No. 3, 09 2001. An Architectural Style for Self-Adaptive Multi-Agent Systems · 49 or { unobservable. via readKnowledge receive knowledge : Knowledge. conv_id := getId(knowledge, conversationIds). via selectConversation send conv_id. via readConversation receive conversation : Conversation. if(terminiateConversation(protocols, conversation, knowledge)) then { //terminate conversation via terminateConversation send conv_id. conversationIds := deteleId(conv_id, conversationIds). via writeKnowledge send closeCommunication(conversation). } else { //silence unobservable. } behavior(). } } } Central to message-based communication are protocols. A protocol is characterized by a name and a set of protocol steps. To manage the communicative interactions, the com- municating component maintains a set of conversation identifiers. The behavior of the communicating component consists of three parts: handling incoming messages, handling outgoing messages, and silent behavior. An incoming message can refer to an ongoing conversation or initiate a new conversation. An outgoing message can be the continuation of an ongoing conversation or the initiation of a new conversation, depending on the con- ditions as reflected in the agent's current knowledge. Silent behavior can either terminate a conversation when the necessary conditions hold, or have no effects. We illustrate protocol-based communication with DynCNET, a protocol for transport as- signment that we developed for the AGV transportation system. DynCNET is an extension of the well-know contract net protocol Contract NET [Smith 1980], with "Dyn" referring to support for dynamic task assignment. The default protocol consists of four steps: (1) the initiator sends a call for proposals; (2) the participants respond with proposals; (3) the initiator notifies the provisional winner; and finally, (4) the selected participant informs the initiator that the task is started. These four steps are basically the same as in the standard Contract NET protocol.The flexibility of DynCNET is based on the provisional agreement between initiator and participant, and the possible revision of the assignment of the task between the third and fourth step of the protocol. These state diagrams in Fig. 16 and Fig. 17 show the behavior of the AGV agent and the transport agent respectively. When a task enters the system and it is ready to be executed (task-ready), the corresponding transport agent enters the Active state in which it remains until the task is completed (task -- completed) (see Figure 17). As soon as an AGV agent is ready -- to -- work it enters the Working state in which it remains until the task is executed (ready) (see Figure 16). Let us first explain adaptation from the viewpoint of an AGV agent. When an AGV agent is ready to execute a task, it enters the Voting state. As long as the agent has not received a provisional accept, it answers cfp's with proposals. When the AGV ACM Transactions Journal Format, Vol. 2, No. 3, 09 2001. · 50 D. Weyns and F. Oquendo Fig. 16. DynCNET protocol for an AGV agent. In the shaded state, the agent can switch the provisional agree- ment. The format of a state transition is event [guard] / actions. agent receives a provisional -- accept message, it enters the Intentional state. As soon as the AGV starts the task (task -- started), it sends a bound message to the transport agent to inform the latter that the execution of the task is started. The AGV agent is then committed to execute the task.5 However, if a new opportunity occurs before the task is started, i.e. the AGV agent receives a provisional -- accept which is a better offer, the agent changes to the Switch Initiator state. The agent then retracts from the earlier provisional task assignment (send(retract)), and switches to the more suitable task (SwitchTask()) entering again the Intentional state. AGV agents use the function TaskInOutScope() to determine whether new tasks enter and leave their area of interest (see Figure 16). This function is realized by perception component. Let us know look at adaptation from the transport agent viewpoint. As long as the trans- port agent has not selected an AGV to execute the task (! haveWinner), it sends period- ically (Timer()) cfp's to the AGV agents in scope. Based on the received proposals from the AGV agents, it selects a winner, sends a provisional -- accept message, and enters the Assigned state. As soon as the transport agent receives a bound message from the selected AGV agent, it enters the state Executing in which the task is effec- tively started. However, if a new opportunity occurs before the task is started, i.e. the trans- port agent receives a proposal from an AGV agent which is better than the current provi- sionally accepted proposal, the transport agent changes to the Switch Participant 5The transport agent's state changes from Assigned to Executing when it receives the bound message from the AGV agent (see Figure 17). ACM Transactions Journal Format, Vol. 2, No. 3, 09 2001. An Architectural Style for Self-Adaptive Multi-Agent Systems · 51 Fig. 17. DynCNET protocol for a transport agent. agreement. The format of a state transition is event [guard] / actions. In the shaded state, the agent can switch the provisional state. In this state the transport agent sends an abort message to the provisionally as- signed AGV agent, and subsequently sends a provisional -- accept message to switch to the more suitable AGV agent (newWinner). The function AGVInOutScope() no- tifies the transport agent when AGVs enter and leave its area of interest. This function is realized by the agent's perception component. 4.5.3 Design Rationale. The primary principles that underlay the design of protocol-based communication are coordination through message exchange and structured interactions through protocols. Coordination through message exchange complements indirect coordination via marks in the virtual environment such as field-based coordination. Direct communication allows situated agents to exchange information and set up explicit collaborations. By exchanging messages, agents can agree to take up roles in a collaboration, reflected in mutual com- mitments. Such commitments steer the behavior of the agents towards the realization of shared objectives. Collaborations based on mutual commitments require a careful design as it creates explicit dependencies between agents. Communication defined in terms of protocols structures the communication, which fits the reactive nature of situated agents. In each step of a communicative interaction, condi- tions determine the agent's behavior in the conversation. Conditions not only depend on the status of the ongoing conversations and the content of received messages, but also on the actual conditions in the environment reflected in the agent's current knowledge. This ACM Transactions Journal Format, Vol. 2, No. 3, 09 2001. · 52 D. Weyns and F. Oquendo contributes to the flexibility of the agent's behavior. 5. METHODOLOGICAL GUIDELINES We explained the key characteristics and requirements of the systems targeted by the ar- chitectural style in section 2.3. From our experience, we derive a set of guidelines to apply the style for systems within the target scope. Our objective is to offer the reader some guidance with respect to important aspects of the design of a situated multi-agent systems when using the architectural style. What agents to select? A key design decision of a situated multi-agent system is to se- lect the type and number of agents in a system. There is no standard rule to select agents. However, as a rule of thumb, active entities that perform tasks in an autonomous way are primary candidates to become agents. In both the systems we used in this report, we have selected two types of agents: one type that performs the actual tasks for the users (AGV agents that perform transportation tasks and vehicle agents that guide drivers), and another type that supports the first type if agent (task agents and infrastructure agents). In purely technical systems, agents are typically selected for processes that perform tasks, and re- sources that encapsulate some activities. In socio-technical systems, agents often represent social actors in the system (in addition to processes and resources). For example, in [Hae- sevoets et al. 2010], we selected different types of situated agents in a supply chain that represent the different actors in a supply chain. Responsibilities of agents and virtual environment. Another key decision in the design of a situated multi-agent system is assigning responsibilities to agents and the virtual envi- ronment. In particular, for decisions regarding coordination, the designer often has a choice to either assign the responsibility to agents are assign it the virtual environment. The key underlying principle here is complexity management through separation of concerns. An excellent example that illustrates this aspect was the choice between the protocol-based and field-based approach to assign tasks in the AGV transportation systems. Using a protocol puts the burden of coordination on the agents. The task of the communication service is simply routing messages. Using the field-based approach puts the burden of spreading and maintaining the fields on the virtual environment. In this option, the agents can observe fields that guide them to a task. Finding a good balance that takes into account the different coordination tasks is not trivial. Variant selection. The patterns of the architectural style have several built-in variability mechanisms. Situated agents can or cannot make use of protocol-based communication. In the latter case, the communication component of situated agent and the communication service of the virtual environment can be omitted. The perception service may or may not require direct access to the external environment. In the latter case, the synchronization component can be responsible to synchronize the state repository of the virtual environ- ment with corresponding elements in the environment. We applied this approach in the AGV transportation system. The state of the virtual environment can be coded in the same format as the knowledge used by the agents. For such a setting, the interpreting and de- scriptions components of agents can be omitted. Other variants are filters of the perception component, commitments of behavior-based action selection, and synchronization of the virtual environment. Supporting middleware. As any modern distributed software system, a situated multi- ACM Transactions Journal Format, Vol. 2, No. 3, 09 2001. An Architectural Style for Self-Adaptive Multi-Agent Systems · 53 agent system requires common middleware service such as basic support for distribution, support for persistency, transactions, security, etc. Such services can be provided by com- mon middleware that provides a basic design and deployment platform. Often, additional domain-specific services are required, such as communication services or dedicated ser- vices to interact with the external environment. The integration with this underlying mid- dleware and supporting services is a crucial aspect in the design of a situated multi-agent system. 6. RELATED WORK Over the last decade, many patterns for multi-agent systems have been documented. To obtain a good overview of the state of the art on these patterns, we have performed a systematic literature review. The review covered the main publication venues of the field since 1998 resulting in 206 patterns documented in in 39 articles.6 The list of articles with the patterns is attached in the online Appendix B. We limit the discussion here to a classification of the patterns that we derived from the study, and we explain how the patterns of the architectural style presented in this report fit in this classification. The study material and additional results are available via http://homepage.lnu.se/staff/daweaa/SLR- MASPatterns.htm The objective of classifying design patterns for multi-agent systems is to provide an intellectual graspable overview of the huge space of existing patterns. The classification offers engineers a general picture of the pattern space of multi-agent systems, and helps those who are not familiar with the domain to get an easy jump-start to understand the pat- tern space. Several researchers have proposed classifications of design patterns for MAS, but these classifications are either bound to a specific catalog of patterns, or to an devel- opment methodology [Kendall et al. 1998; Aridor and Lange 1998; Tahara et al. 1999; Schelfthout et al. 2002; Oluyomi 2006; Chella et al. 2010]. The classification presented in this report covers the full space of patterns for multi-agent systems as document at the time of writing. The classification resulted from the analysis of several types of data that were collected during the review, including pattern categories (source of inspiration), pattern associations (explicitly documented relations between patterns), and short description of the patterns. From the analysis of the data, we derived four dimensions of patterns for multi-agent sys- tems: inspiration, abstraction, focus, and granularity. Fig. 18 shows a graphical overview of the dimensions, illustrated with example patterns. Metaphors and analogies help in understanding complex systems. The inspiration di- mension has four categories that provide intuitive domains from which patterns are de- rived. Examples of nature-inspired patterns are gradient fields [De Wolf et al. 2007] that is inspired by the fields in nature, and delegate ant MAS [Holvoet et al. 2009] that is in- spired by behavior of social insects. Examples of society-inspired patterns are emergent society [Kendall et al. 1998] and holonic society [Chella et al. 2010] that get their inspira- tion from the way societies emerge and structure themselves. Examples of human-inspired patterns are receptionist [Silva et al. 2000] and gossip [Fernandez-Marquez et al. 2011]. 6Publications from SpringerLink, ACM Digital Library, IEEE Explore, Science Direct, CiteseerX, Google Scholar were searched automatically. Publications of The Knowledge Engineering Review (KER), Transactions on Autonomous and Adaptive Systems (TAAS) and the Journal of Autonomous Agents and Multi-Agent Systems (JAAMAS) were searched manually. ACM Transactions Journal Format, Vol. 2, No. 3, 09 2001. · 54 D. Weyns and F. Oquendo Fig. 18. Classification of patterns for multi-agent systems with example patterns. Finally, locker [Aridor and Lange 1998] and blackboard [Deugo et al. 1999] are example of patterns that get there inspiration from artifacts in our environment. The abstraction dimension classifies patterns either as conceptual or concrete. Both these main categories are further refined in subcategories that refer to stages in the soft- ware life cycle where the patterns can be used. Actors [Mouratidis and Weiss 2006] and pyramid [Kolp et al. 2006] are examples of patterns that are useful in early phases in the life cycle, while broker [Hayden et al. 1999] and synchronizer [Schelfthout et al. 2002] can be applied in detailed design and the implementation phase. The focus dimension has two categories: structural and behavioral. Structural patterns are useful to deal with the decomposition of a system, while behavior patterns are use- ful to deal with interaction aspects. Examples of structural patterns are reflective black- ACM Transactions Journal Format, Vol. 2, No. 3, 09 2001. An Architectural Style for Self-Adaptive Multi-Agent Systems · 55 board [Silva et al. 2003] and joint venture [Kolp et al. 2006], the former focusing a particu- lar coordination structure for an agent system, the latter focusing on the way a community of agents is organized. Examples of behavioral patterns are market maker [Deugo et al. 1999], negotiating agents [Deugo et al. 1999], and conversation [Kendall et al. 1998]. These patterns provide different approaches to support interactions among agents. Finally, the granularity dimension refers to the scope of the patterns, i.e., the system or parts of the system. Hierarchical contracting [Kolp et al. 2006] and emergent soci- ety [Kendall et al. 1998] are examples of patterns that apply to a multi-agent system as a whole. Messenger [Aridor and Lange 1998] and virtual environment [Weyns 2009] are patterns that apply to parts of a multi-agent system. Intention [Kendall et al. 1998] and InteRRaP [Lind 2002] are patterns that support the design of individual agents. We conclude with classifying the patterns presented in this report, see Table I. Table I. Classification applied to the patterns of the architectural style Pattern Virtual environment Situated agent Inspiration nature-inspired nature-inspired Selective perception artifact-inspired Behavior based action selection society-inspired Protocol based communication society-inspired Abstraction conceptual (design) conceptual (design) conceptual (design) conceptual (design) conceptual (design) Focus structural (architectural) structural (architectural) behavioral Granularity subsystem level agent level agent level behavioral agent level behavioral (interactions) agent level The patterns of the architectural style take inspiration from nature (environment, sens- ing, synchronization,e tc.), artifacts (focus, filter), and society (protocols, roles, situated commitments, etc.). The five patterns are useful during architectural design. The primary patterns (virtual environment and situated agent) are structural, while the refined patterns are of behavioral nature. 7. CONCLUSIONS AND FUTURE WORK To further mature the domain of software architecture, [Shaw and Clements 2006] argue for the creation of reference materials that give engineers access to the field's systematic knowledge. This report presents an architectural style for multi-agent systems that doc- uments well-proven design expertise for a family of self-adaptive software systems. The structures and behaviors of the pattern elements are specified in π-ADL, a formal ADL supporting automated verification of architectural properties. These specifications provide rigorously defined templates to guide architects when applying the patterns in practice. The templates are under-specified providing the necessary flexibility for applying the patterns for a wide range of systems. We have illustrated how we have realized the various patterns of the style with excerpts from an experimental as well as an industrial case study. Whereas the presented architectural style consolidates design knowledge, expertise from building new systems brings new know-how that may be useful to extend or evolve the ACM Transactions Journal Format, Vol. 2, No. 3, 09 2001. · 56 D. Weyns and F. Oquendo style. In recent work, we have built a number of applications in which the pattern real- izations of the presented architecture style have been extended with additional features. In one extension, we have encapsulated the life-cycle management of agent organizations as a reusable service, which is modeled as an intermediate component between agent and virtual environment [Weyns et al. 2010]. In another extension, we have added self-healing components to a situated agent system that deals with particular types of faults [Vromant et al. 2011]. These self-healing components monitor agents and the virtual environment de- ployed on different nodes and make adaptations when faults occur. Both these extensions may eventually lead to two new patterns that extend the architectural style. To conclude, we hope that the systematic knowledge documented in this report will contribute to a better understanding of multi-agent system architectures and their value for designing complex self-adaptive systems. Online Appendix https://people.cs.kuleuven.be/danny.weyns/Appendix-1-2.pdf Acknowledgement The research presented in this report consolidates multiple collaborative research efforts spread over eight years. We acknowledge the valuable contributions of our colleagues in these efforts, including Kurt Schelfhout, Alexander Helleboogh, Nelis Boucke, Ro- brecht Haesevoets, Rutger Claes, Elke Steegmans, Bartosz Michalik, Tom Holvoet, Us- man Iftikhar, Jan Wielemans, Tom Lefever, Rudy Vanhoutte, Wim Van Betsbrugge, Jan Vercammen, Zawar Qayyum, and Qin Xiong. REFERENCES ARIDOR, Y. AND LANGE, D. B. 1998. Agent design patterns: elements of agent application design. In Proceed- ings of the second international conference on Autonomous agents. ACM, New York, NY, USA, 108 -- 115. BANDINI, S., MANZONI, S., AND SIMONE, C. 2002. Dealing with space in multi -- agent systems: a model for situated mas. In AAMAS '02: Proceedings of the first international joint conference on Autonomous agents and multiagent systems. ACM, New York, NY, USA, 1183 -- 1190. BONABEAU, E., DORIGO, M., AND THERAULAZ, G. 1999. Swarm intelligence: from natural to artificial systems. Oxford University Press, Inc., New York, NY, USA. BORDINI, R., HUBER, J., AND OOLDRIDGE, M. 2007. Programming Multi-Agent Systems in AgentSpeak. Wiley. BRUN, Y., MARZO SERUGENDO, G., GACEK, C., GIESE, H., KIENLE, H., LITOIU, M., M ULLER, H., PEZZ `E, M., AND SHAW, M. 2009. Software engineering for self-adaptive systems. Springer-Verlag, Berlin, Heidel- berg, Chapter Engineering Self-Adaptive Systems through Feedback Loops, 48 -- 70. BRYSON, J. 2001. Intelligence by Design, Principles of Modularity and Coordination for Engineering Complex Adaptive Agents. PhD Dissertation, MIT, USA. CARRIERO, N. AND GELERNTER, D. 1989. Linda in context. Communications of the ACM 32, 444 -- 458. CHELLA, A., COSSENTINO, M., GAGLIO, S., SABATUCCI, L., AND SEIDITA, V. 2010. Agent-oriented soft- ware patterns for rapid and affordable robot programming. J. Syst. Softw. 83, 557 -- 573. CHENG, B. H., LEMOS, R., GIESE, H., INVERARDI, P., AND MAGEE, J., Eds. 2009. Software Engineering for Self-Adaptive Systems. Springer-Verlag, Berlin, Heidelberg. CHENG, B. H., LEMOS, R., GIESE, H., INVERARDI, P., MAGEE, J., ANDERSSON, J., BECKER, B., BEN- COMO, N., BRUN, Y., CUKIC, B., MARZO SERUGENDO, G., DUSTDAR, S., FINKELSTEIN, A., GACEK, C., GEIHS, K., GRASSI, V., KARSAI, G., KIENLE, H. M., KRAMER, J., LITOIU, M., MALEK, S., MIRAN- DOLA, R., M ULLER, H. A., PARK, S., SHAW, M., TICHY, M., TIVOLI, M., WEYNS, D., AND WHITTLE, J. 2009. Software engineering for self-adaptive systems: A research roadmap. 1 -- 26. ACM Transactions Journal Format, Vol. 2, No. 3, 09 2001. An Architectural Style for Self-Adaptive Multi-Agent Systems · 57 CLAES, R., HOLVOET, T., AND WEYNS, D. 2011. A decentralized approach for anticipatory vehicle routing using delegate multi-agent systems. IEEE Transactions on Intelligent Transportation Systems 12, 2. CLEMENTS, P., BACHMANN, F., BASS, L., GARLAN, D., IVERS, J., LITTLE, R., NORD, R., AND STAFFORD, J. 2010. Documenting Software Architectures: Views and Beyond; 2nd Edition. Addison Wesley. DE LEMOS, R. ET AL. 2012. Software engineering for self-adaptive systems ii. Springer-Verlag, Berlin, Heidel- berg, Chapter Software Engineering for Self-Adaptive Systems: A Second Roadmap. DE LEMOS ET AL., R. 2017. Software engineering for self-adaptive systems: Research challenges in the provi- sion of assurances. In Software Engineering for Self-Adaptive Systems III: Assurances. Springer, 3 -- 30. DE WOLF, T. ET AL. 2007. Design patterns for decentralised coordination in self-organising emergent systems. In Engineering Self-Organising Systems. Lecture Notes in Computer Science, vol. 4335. Springer, Berlin, Heidelberg, Germany, 28 -- 49. DEUGO, D., OPPACHER, F., KUESTER, J., AND OTTE, I. V. 1999. Patterns as a means for intelligent software engineering. In Proceedings of the 1999 Conference on Artificial Intelligence. 605 -- 611. DOBSON, S., DENAZIS, S., FERN ´ANDEZ, A., GAITI, D., GELENBE, E., MASSACCI, F., NIXON, P., SAFFRE, F., SCHMIDT, N., AND ZAMBONELLI, F. 2006. A survey of autonomic communications. ACM TAAS 1, 223 -- 259. DURFEE, E. AND LESSER, V. 1989. Negotiating Task Decomposition and Allocation Using Partial Global Planning. Distributed Artificial Intelligence 2, 229 -- 244. FERBER, J. 1999. Multi-Agent Systems: An Introduction to Artificial Intelligence. ISBN 0-201-36048-9, Addison-Wesley. FERNANDEZ-MARQUEZ, J., ARCOS, J., DI MARZO SERUGENDO, G., AND CASADEI, M. 2011. Descrip- In Engineering of Autonomic and tion and composition of bio-inspired design patterns: The gossip case. Autonomous Systems (EASe), 2011 8th IEEE International Conference and Workshops on. 87 -- 96. GARLAN, D., CHENG, S., HUANG, A., SCHMERL, B., AND STEENKISTE, P. 2004. Rainbow: Architecture- Based Self-Adaptation with Reusable Infrastructure. IEEE Computer 37, 10. HAESEVOETS, R. 2012. Macodo: Architecture-Centric Support for Dynamic Service Collaborations. Ph.D Dissertation. Supervisors: Danny Weyns and T. Holvoet, Katholieke Universiteit Leuven, Belgium. HAESEVOETS, R., WEYNS, D., AND HOLVOET, T. 2014. Architecture-centric support for adaptive service collaborations. ACM Trans. Softw. Eng. Methodol. 23, 1 (Feb.), 2:1 -- 2:40. HAESEVOETS, R., WEYNS, D., TORRES, M. H. C., HELLEBOOGH, A., HOLVOET, T., AND JOOSEN, W. 2010. A middleware model in alloy for supply chain-wide agent interactions. In Agent-Oriented Software Engineering XI. Lecture Notes in Computer Science. Springer. HAYDEN, S. C., CARRICK, C., AND YANG, Q. 1999. Architectural design patterns for multiagent coordination. In Proceedings of the 3rd international conference on autonomous agents. HOLVOET, T., WEYNS, D., AND VALCKENAERS, P. 2009. Patterns of delegate mas. In Proceedings of the 2009 Third IEEE International Conference on Self-Adaptive and Self-Organizing Systems. SASO '09. IEEE Computer Society, Washington, DC, USA, 1 -- 9. HUEBSCHER, M. C. AND MCCANN, J. A. 2008. A survey of autonomic computing: degrees, models, and applications. ACM Computer Surveys 40, 7:1 -- 7:28. JACKSON, M. 1997. The meaning of requirements. Ann. Softw. Eng. 3, 5 -- 21. JENNINGS, N., SYCARA, K., AND WOOLDRIDGE, M. 1998. A Roadmap of Agent Research and Development. Autonomous Agents and Multiagent Systems 1. KENDALL, E. A., MURALI KRISHNA, P. V., PATHAK, C. V., AND SURESH, C. B. 1998. Patterns of intelligent In Proceedings of the second international conference on Autonomous agents. ACM, and mobile agents. 92 -- 99. KOLP, M., GIORGINI, P., AND MYLOPOULOS, J. 2006. Multi-agent architectures as organizational structures. Autonomous Agents and Multi-Agent Systems 13, 3 -- 25. KRAMER, J. AND MAGEE, J. 2007. Self-managed systems: an architectural challenge. In FOSE '07: Future of Software Engineering. IEEE Computer Society. LIND, J. 2002. Patterns in agent-oriented software engineering. In Proceedings of the 3rd international con- ference on Agent-oriented software engineering III. AOSE'02. Springer-Verlag, Berlin, Heidelberg, Germany, 47 -- 58. MAES, P. 1990. Situated Agents can have Goals. Designing Autonomous Agents, MIT Press. ACM Transactions Journal Format, Vol. 2, No. 3, 09 2001. · 58 D. Weyns and F. Oquendo MAES, P. 1995. Artificial Life Meets Entertainment: Life like Autonomous Agents. Communications of the ACM 38, 11, 108 -- 114. MAMEI, M. AND ZAMBONELLI, F. 2004. Co-Fields: A Physically Inspired Approach to Distributed Motion Coordination. IEEE Pervasive Computing 3, 2, 52 -- 61. MAMEI, M. AND ZAMBONELLI, F. 2009. Programming pervasive and mobile computing applications: the tota approach. ACM Transaction on Software Engineering and Methodology 18, 4. MATARIC, M. J. 1994. Leaning to behave socially. In SAB94: Proceedings of the third international conference on Simulation of adaptive behavior : from animals to animats 3. MIT Press, Cambridge, MA, USA, 453 -- 462. MOURATIDIS, H. AND WEISS, M. 2006. Patterns for modelling agent systems with tropos. In Software En- gineering for Multi-Agent Systems IV, A. Garcia, R. Choren, C. Lucena, P. Giorgini, T. Holvoet, and A. Ro- manovsky, Eds. Lecture Notes in Computer Science, vol. 3914. Springer-Verlag, 207 -- 223. MURPHY, A. L., PICCO, G. P., AND ROMAN, G.-C. 2006. Lime: A coordination model and middleware supporting mobility of hosts and agents. ACM Trans. Softw. Eng. Methodol. 15, 3 (July), 279 -- 328. OLUYOMI, A. 2006. Phd. Ph.D. thesis, The University of Melbourne. OQUENDO, F. 2004. Pi-ADL: an Architecture Description Language based on the higher-order typed pi-calculus for specifying dynamic and mobile software architectures. SIGSOFT Softw. Eng. Notes 29, 3, 1 -- 14. OREIZY, P., GORLICK, M., TAYLOR, R., HEIMHIGNER, D., JOHNSON, G., MEDVIDOVIC, N., QUILICI, A., ROSENBLUM, D., AND WOLF, A. 1999. An architecture-based approach to self-adaptive software. Intelligent Systems and Their Applications 14, 54 -- 62. PARUNAK, H. V. D., BRUECKNER, S., AND SAUTER, J. 2005. Digital Pheromones for Coordination of Un- manned Vehicles. In Environments for Multiagent Systems, E4MAS. Lecture Notes in Computer Science, Vol. 3374. Springer. PARUNAK, H. V. D., BRUECKNER, S., SAUTER, J., AND MATTHEWS, R. 2005. Global Convergence of Local In 4th Joint Conference on Autonomous Agents and Multiagent Systems. Utrecht, The Agent Behaviors. Netherlands. POKAHR, A., BRAUBACH, L., AND LAMERSDORF, W. 2005. A BDI architecture for goal deliberation. In 4th International Joint Conference on Autonomous Agents and Multiagent Systems. ACM, New York, NY, USA, 1295 -- 1296. QAYYUM, Z. AND OQUENDO, F. 2008. The pi-adl.net project: an inclusive approach to adl compiler design. W. Trans. on Comp. 7, 5, 414 -- 423. RAO, A. AND GEORGEFF, M. 1995. BDI Agents: From Theory to Practice. In 1st International Conference on Multiagent Systems. The MIT Press. SALEHIE, M. AND TAHVILDARI, L. 2009. Self-adaptive software: Landscape and research challenges. ACM Transactions on Autonomous and Adaptive Systems 4. SCHELFTHOUT, K. 2006. Supporting Coordination in Mobile Networks: A Middleware Approach. PhD Dis- sertation: Katholieke Universiteit Leuven, Belgium. SCHELFTHOUT, K., CONINX, T., HELLEBOOGH, A., HOLVOET, T., STEEGMANS, E., AND WEYNS, D. 2002. Agent Implementation Patterns. In OOPSLA Workshop on Agent-oriented Methodologies. Seattle, WA USA. SCHELFTHOUT, K. AND HOLVOET, T. 2004. A pheromone-based coordination mechanism applied in peer-to- peer. Lecture Notes in Computer Science, Vol. 2872. Springer. SCHELFTHOUT, K., WEYNS, D., AND HOLVOET, T. 2005. Middleware for Protocol-based Coordination in In 3rd International Workshop on Middleware for Pervasive and Ad-hoc Computing. Dynamic Networks. ACM Press, Grenoble, France. SCHELFTHOUT, K., WEYNS, D., AND HOLVOET, T. 2006. Middleware that enables protocol-based coordina- tion applied in automated guided vehicle control. Distributed Systems on Line 7. SHAW, M. AND CLEMENTS, P. 2006. The golden age of software architecture. IEEE Software 23, 2, 31 -- 39. SILVA, I. C., DA SILVA, A. R., AND MEIRA, N. 2000. A set of agent patterns for a more expressive approach. In Proceedings of the EuroPLOP 2000. 331 -- 346. SILVA, O., GARCIA, A., AND LUCENA, C. 2003. The reflective blackboard pattern: Architecting large multi- agent systems. In Software Engineering for Large-Scale Multi-Agent Systems, A. Garcia, C. Lucena, F. Zam- bonelli, A. Omicini, and J. Castro, Eds. Lecture Notes in Computer Science, vol. 2603. Springer Berlin, Hei- delberg, 73 -- 93. ACM Transactions Journal Format, Vol. 2, No. 3, 09 2001. An Architectural Style for Self-Adaptive Multi-Agent Systems · 59 SMITH, R. 1980. The Contract Net Protocol: High Level Communication and Control in a Distributed Problem Solver. In IEEE Transactions on Computers, C-29(12). STEEGMANS, E., WEYNS, D., HOLVOET, T., AND BERBERS, Y. 2004. A Design Process for Adaptive Behavior of Situated Agents. In Agent-Oriented Software Engineering V. Lecture Notes in Computer Science, Vol. 3382. Springer. STERLING, L. AND TAVETER, K. 2009. The Art of Agent-Oriented Modeling. MIT Press. SYCARA, K., PANNU, A., WILLAMSON, M., ZENG, D., AND DECKER, K. 1996. Distributed intelligent agents. IEEE Expert 11, 6 (dec), 36 -- 46. TAHARA, Y., OHSUGA, A., AND HONIDEN, S. 1999. Agent system development method based on agent patterns. In Proceedings of the 21st International Conference on Software Engineering. ICSE '99. ACM, New York, NY, USA, 356 -- 367. TYRRELL, T. 1993. Computational Mechanisms for Action Selection. PhD Dissertation, University of Edin- burgh. VROMANT, P., WEYNS, D., MALEK, S., AND ANDERSSON, J. 2011. On interacting control loops in self- adaptive systems. In Proceedings of the 6th International Symposium on Software Engineering for Adaptive and Self-Managing Systems. SEAMS '11. ACM, New York, NY, USA, 202 -- 207. WEYNS, D. 2009. A Pattern Language for Multiagent Systems. In Joint Working International Conference on Software Architecture and European Conference on Software Architecture (WICSA/ECSA). Cambridge, UK. WEYNS, D. 2010. Architecture-Based Design of Multi-Agent Systems. ISBN 978-3-642-01063-7, Springer. WEYNS, D. 2019. Software Engineering of Self-Adaptive Systems. Chapter in Handbook of Software Engi- neering, Springer. WEYNS, D., BOUCK ´E, N., AND HOLVOET, T. 2008. A field-based versus a protocol-based approach for adap- tive task assignment. Autonomous Agents and Multi-Agent Systems 17, 2, 288 -- 319. WEYNS, D. AND GEORGEFF, M. 2010. Self-adaptation using multiagent systems. IEEE Softw. 27, 1, 86 -- 91. WEYNS, D., HAESEVOETS, R., HELLEBOOGH, A., HOLVOET, T., AND JOOSEN, W. 2010. The macodo mid- dleware for context-driven dynamic agent organzations. ACM Transaction on Autonomous and Adaptive Sys- tems 5, 1, 3.1 -- 3.29. WEYNS, D., HELLEBOOGH, A., AND HOLVOET, T. 2005. The Packet-World: a Test Bed for Investigating Situ- ated Multi-Agent Systems. In Software agent-based applications, platforms, and development kits. Whitestein Series in Software Agent Technology. WEYNS, D. AND HOLVOET, T. 2007. Architectural design of a situated multiagent system for controlling automatic guided vehicles. International Journal of Agent-Oriented Software Engineering 1, 4. WEYNS, D., HOLVOET, T., AND HELLEBOOGH, A. 2007. Anticipatory vehicle routing using delegate multi- agent systems. In Intelligent Transportation Systems Conference, 2007. ITSC 2007. IEEE. IEEE, 87 -- 93. WEYNS, D., HOLVOET, T., SCHELFTHOUT, K., AND WIELEMANS, J. 2008. Decentralized control of automatic guided vehicles: Applying multi-agent systems in practice. In Development Track OOPSLA, ACM SIGPLAN International Conference on Object-Oriented Programming, Systems, Languages, and Applications. WEYNS, D., MALEK, S., AND ANDERSSON, J. 2012. Forms: Unifying reference model for formal specification of distributed self-adaptive systems. ACM TAAS 7, 1, 8:1 -- 8:61. WEYNS, D., MICHEL, F., ET AL. 2015. Agent environments for multi-agent systems -- a research roadmap. In Revised Selected and Invited Papers of the 4th International Workshop on Agent Environments for Multi-Agent Systems IV - Volume 9068. Springer-Verlag New York, Inc., New York, NY, USA, 3 -- 21. WEYNS, D., OMICINI, A., AND ODELL, J. 2007. Environment as a First-Class Abstraction in Multiagent Systems. Autonomous Agents and Multi-Agent Systems 14, 1. WEYNS, D., PARUNAK, H. V. D., MICHEL, F., HOLVOET, T., AND FERBER, J. 2005. Environments for Multiagent Systems, State-of-the-art and Research Challenges. D. Weyns, H. V. D. Parunak, and F. Michel, Eds. Lecture Notes in Computer Science, Vol. 3374. Springer Verlag. WEYNS, D., SCHELFTHOUT, K., HOLVOET, T., AND LEFEVER, T. 2005. Decentralized control of E'GV In 4th Joint Conference on Autonomous Agents and Multiagent Systems, Industry transportation systems. Track. ACM Press, New York, NY, USA, Utrecht, The Netherlands. WEYNS, D., SCHMERL, B., GRASSI, V., MALEK, S., MIRANDOLA, R., PREHOFER, C., WUTTKE, J., AN- DERSSON, J., GIESE, H., AND GOSCHKA, K. 2013. Software engineering for self-adaptive systems ii, ACM Transactions Journal Format, Vol. 2, No. 3, 09 2001. · 60 D. Weyns and F. Oquendo lncs vol. 7475. Springer-Verlag, Berlin, Heidelberg, Chapter On Patterns for Decentralized Control in Self- Adaptive Systems. WEYNS, D., STEEGMANS, E., AND HOLVOET, T. 2004a. Integrating Free-Flow Architectures with Role Mod- els Based on Statecharts. In Software Engineering for Multi-Agent Systems III. Lecture Notes in Computer Science, Vol. 3390. Springer. WEYNS, D., STEEGMANS, E., AND HOLVOET, T. 2004b. Towards Active Perception in Situated Multi-Agent Systems. Applied Artificial Intelligence 18, 9-10, 867 -- 883. WOOLDRIDGE, M. 2000. Reasoning about Rational Agents. MIT Press. WOOLDRIDGE, M. AND JENNINGS, N. 1995. Intelligent agents: theory and practice. The Knowledge Engi- neering Review 10, 2. ACM Transactions Journal Format, Vol. 2, No. 3, 09 2001.
1905.01360	1	1905	2019-04-20T18:30:58	Skynet: A Top Deep RL Agent in the Inaugural Pommerman Team Competition	[ "cs.MA", "cs.AI", "cs.LG" ]	The Pommerman Team Environment is a recently proposed benchmark which involves a multi-agent domain with challenges such as partial observability, decentralized execution (without communication), and very sparse and delayed rewards. The inaugural Pommerman Team Competition held at NeurIPS 2018 hosted 25 participants who submitted a team of 2 agents. Our submission nn_team_skynet955_skynet955 won 2nd place of the "learning agents'' category. Our team is composed of 2 neural networks trained with state of the art deep reinforcement learning algorithms and makes use of concepts like reward shaping, curriculum learning, and an automatic reasoning module for action pruning. Here, we describe these elements and additionally we present a collection of open-sourced agents that can be used for training and testing in the Pommerman environment. Code available at: https://github.com/BorealisAI/pommerman-baseline	cs.MA	cs	Skynet: A Top Deep RL Agent in the Inaugural Pommerman Team Competition Chao Gao∗, Pablo Hernandez-Leal, Bilal Kartal, and Matthew E. Taylor Borealis AI {pablo.hernandez, bilal.kartal, matthew.taylor}@borealisai.com Edmonton, Alberta, Canada Abstract The Pommerman Team Environment is a recently proposed benchmark which involves a multi-agent domain with challenges such as partial observability, decentralized execution (without communication), and very sparse and delayed rewards. The inaugural Pommerman Team Competition held at NeurIPS 2018 hosted 25 participants who submitted a team of 2 agents. Our submission nn_team_skynet955_skynet955 won 2nd place of the "learning agents" category. Our team is composed of 2 neural networks trained with state of the art deep reinforcement learning algorithms and makes use of concepts like reward shaping, curriculum learning, and an automatic reasoning module for action pruning. Here, we describe these elements and additionally we present a collection of open-sourced agents that can be used for training and testing in the Pommerman environment. Code available at: https://github.com/BorealisAI/ pommerman-baseline Keywords: Deep Reinforcement Learning; Multi-Agent Deep Reinforcement Learning; Pommerman Acknowledgements 9 1 0 2 r p A 0 2 ] A M . s c [ 1 v 0 6 3 1 0 . 5 0 9 1 : v i X r a ∗[email protected] Department of Computing Science, University of Alberta. Work performed as intern at Borealis AI 1 Introduction The Pommerman environment for benchmarking Multi-Agent Learning is based on the classic console game Bomberman. The team competition involves 4 bomber agents initially placed at the four corners of an 11×11 board. Each two diagonal agents form a team. At every step, each agent issues an action simultaneously from 6 discrete candidate moves: moving left, right, up, down, placing a bomb, or stop. The bomb action is legal as long as the agent's ammo is greater than 0, and any illegal action is superseded with stop by the environment. Each cell on the board can either be a passage, a rigid wall, or wood. Only passage is passable for the agent. Wood would be destroyed if it is covered by an exploding bomb's flame while a rigid location is unaffected. Also, when wood is destroyed a powerup might appear (according to some probability) at the same location. There are 3 types of powerups: ExtraBomb, EnableKick, and ExtraBlast for which the agent respectively, increases its ammo by 1, acquires the ability to kick bombs, and rises its bombs' blast strength by 1. The game starts with a procedurally generated random map and each agent initially has an ammo of 1 and blast strength of 2. Whenever an agent places a bomb, it explodes after 10 time steps, producing flames that have a lifetime of 2 time steps and a radius of the bomb placing agent's blast strength. The game ties when both teams have at least one agent alive after 800 steps. The team environment is also partially observable, meaning each agent can only see the local board with a radius of 4 cells, see Figure 1(a). Pommerman is a challenging benchmark for multi-agent learning and model-free reinforcement learning, due to the following characteristics: Sparse and deceptive rewards: the former refers to the fact that the only non-zero reward is obtained at the end of an episode. The latter refers to the fact that quite often a winning reward is due to the opponents' involuntary suicide, which makes reinforcing an agent's action based on such a reward deceptive. Note that suicide happens frequently during learning since an agent has to place bombs to explode wood to move around on the board, while due to terrain constraints, in some cases, performing non-suicidal bomb placement requires complicated, long-term, and accurate planing. Delayed action effects: the only way to make a change to the environment (e.g., bomb wood or kill an agent) is by means of bomb placement, but the effect of such an action is only observed when the bomb's timer decreases to 0; more complications are added when a placed bomb is kicked to another position by some other agent. Imperfect information: an agent can only see its nearby areas. This makes the effect of certain actions, such as bomb kicking, unpredictable. Indeed, even detecting which agent placed the exploding bomb is intractable in general because of the hidden information of the board. Uninformative multiagent credit assignment: In the team environment, the same episodic reward is given to two members of the team. It may not be clear how to assign credit to individual agents. For example, consider an episode where an agent eliminates an opponent but then commits suicide, and its teammate eliminates the remaining opponent. Under this scenario, both team members get a positive reward from the environment, but this could reinforce the suicidal behaviour of the first agent. Similarly, one agent could eliminate both opponents whereas its teammate just camps; both agents would get positive rewards, reinforcing a lazy agent [7]. 2 Skynet955 nn_team_skynet955_skynet955 is a team composed of two identical neural networks, where skynet955 is trained after equipping the neural net agent with an "ActionFilter" module for 955 iterations. The philosophy is to instill prior knowledge to the agent by telling the agent what not to do and let the agent discover what to do by trial-and-error. The benefit is twofold: 1) the learning problem is simplified since suicidal actions are removed and bomb placing becomes safe; and 2) superficial skills such as not moving into flames and evading bombs in simple cases are handled. Below we describe the main components of our team: the ActionFilter and the reinforcement learning aspect. ActionFilter We designed the filter to speed up learning so that agents can focus on higher level strategies. The filter thus serves as a safety check to provide more efficient exploration. The ActionFilter is implemented by rules shown in Table 1. Note that for the "avoiding suicide" rules, skynet955 implemented a simple version of them (e.g., a moving bomb was simply treated as static); a full implementation would arguably make the agent stronger. It is worth mentioning that the above ActionFilter is extremely fast. In our experiments, together with neural net evaluation, each action takes ≈ 3 ms on a GTX 1060 GPU on average, while the time limit in the competition is 100 ms per move. Also, we note that another natural approach for "bomb placement" pruning is conducting a lookahead search using "avoding suicide" rules; this is perhaps better than the crude rules described above. We think this ActionFilter will be useful for learning agents within the Pommerman domain, since it can constrain the action space during learning without significantly affecting the overall team strategy. Ideally, one can even start model-free learning with the filter, and once an agent acquires some basic 1 (a) (b) Figure 1: (a) An example team environment Pommerman game. On the right is shown the partial observation (a limited area around) for each agent. (b) Winning percentage of the learning team against a team of Static opponents and a team of Smart random opponents. Table 1: ActionFilter rules Avoiding Suicide Bomb Placement Not going to positions that are flames on the next step. Not going to doomed positions, i.e., positions where if the agent were to go there the agent would have no way to escape. For any bomb, doomed positions can be computed by referring to its blast strength, blast range, and life, together with the local terrain. Not place bombs when teammate is close, i.e., when their Manhattan distance is less than their combined blast strength. Not place bombs when the agent's position is covered by the blast of any previously placed bomb. skills, it can be unplugged from the neural network so that strategy bias by filter is completely removed. For this reason, we have open-sourced an implementation of the ActionFilter.1 Reinforcement Learning As depicted in Figure 2, the architecture contains 4 convolution layers, followed by policy and value heads. The input contains 14 features planes, each of shape 11×11, similar to [9]. It then convolves using 4 layers of convolution, each has 64 3×3 kernels; the result thus has shape 11×11×64. Then, each head convolves using 2 1×1 kernels. Finally, the output is squashed into action probability distribution and value estimation, respectively. Such a two-head architecture is a natural choice for Action-Critic algorithms, as it is generally believed that forcing policy and value learning to use shared weights could reduce over-fitting [10]. Instead of using an LSTM to track the history observations, we use a "retrospective board" to remember the most recent value at each cell of the board, i.e., for cells outside of an agent's view, in the "retrospective board" its value is filled with what was present when the agent saw that cell the last time. The input feature has 14 planes in total, where the first 10 are extracted from the agent's current observation, the rest 4 are from "retrospective board." We initially performed experiments with an LSTM to track all previous observations; however, due to computational overhead on top of the already prolonged training in Pommerman domain, we decided to replace it with the "retrospective board" idea, which yielded similar performance but was significantly faster. 1 https://github.com/BorealisAI/pommerman-baseline Figure 2: Architecture used for the skynet955 agents 2 05001,0001,500020406080100IterationWiningPercentageAgainstStaticTeam05001,0001,5002,00005101520IterationWiningPercentageAgainstSmartRandomTeamTable 2: Reward Shaping for skynet955 agents Going to a cell not in a 121-length FIFO queue gets 0.001. Picking up kick gets 0.02. Picking up ammo gets 0.01. Picking up blast strength gets 0.01. At the end of a game, dead agent in the winning team gets 0.5. For draw games, all agents receive 0.0. On one enemy's death gets 0.5. On a teammate's death gets −0.5. (cid:34) (cid:88) The neural net is trained by PPO [10], minimizing the following objective: o(θ;D) = (st,at,Rt)∈D −clip( πθ(atst) θ (atst) πold , 1−, 1+)A(st, at)+ ·max α 2 (cid:104) (vθ(st)−Rt)2, (vold θ (st)+clip(vθ(st)−vold θ (st),−, )−Rt)2(cid:105)(cid:35) , (1) where θ is the neural net, D is sampled by πold θ , and is a tuning parameter. Refer to OpenAI baseline for details [10]. Curriculum Learning Training is conducted by letting two identical neural net players compete against a set of curricu- lum [1] opponents: (i) Static opponent teams, where opponents do not move or place bombs. Competing against a team of Static opponents teaches our agents to get closer to opponents, place a bomb, and move away to a safe zone. The trained neural net is then used against the second opponent in the curriculum. (ii) SmartRandomNoBomb: players that do not place bombs. Smart random means it has the ActionFilter as described earlier and the action taken is random (except that bomb placing is disallowed). The reason we let SmartRandomNoBomb not place bombs is that the neural net can focus on learning true "killing" skills, not a skill that solely relies on the opponent's strategy flaw (e.g., the provided baseline SimpleAgent has a significant flaw where the competitor can diagonally block and make SimpleAgent be killed by its own bomb). This avoids the "false positive" reward signal caused by opponent's involuntary suicide. Competing against a team of SmartRandomNoBomb helps our agents to learn better battling skills such as using the topological map to corner the opponents and pursuing opponents. Reward shaping: To cope with the sparse reward problem, a dense reward function is added during the learning, see Table 2. It should be noted the above hand designed reward function is still noisy in the sense that an agent's contribution was not clearly separated. Results: Figure 1(b) shows the learning curves against Static and SmartRandomNoBomb teams. In our training, each iteration contains 120 games, produced in parallel by 12 actor workers. The curves show that, against Static agents, the neural net achieved wining percentage around 70%, while against SmartRandomNoBomb, it never reached 20%. We note that because the opponents do not place bombs, the rest of the games are almost all draws. The learning seems to be slow, in part because playing against SmartRandomNoBomb, a large number of games were ended with draws, which gives reward signal 0 in our training. In the competition, our team was composed of two identical neural net models, at each step, for each of our agent, each action typically costs one to several milliseconds, while the time limit is 100ms per move. The submitted agent skynet955 was the neural net model at iteration 955 obtained in training against SmartRandomNoBomb team, as by the time of submission only around 1000 iterations were finished. Throughout our training and testing, only the V0 environment (which has no wall collapsing) was used. By contrast, in the competition, the V1 (which has wall collapsing, meaning at certain time step, the boarder passages will suddenly change to rigid walls, and any agent that happens to be in any of the corresponding cells dies) was used. An open-sourced collection of agents for Pommerman Against SimpleAgent, it is not difficult to train a non-placing- bomb neural net agent that wins by diagonally blocking and forcing SimpleAgent to get stuck on its self-placed bomb.2 This strategy flaw of SimpleAgent stems from its hand-crafted heuristic strategy for enemy engagement and bomb placement. Learning agents exploit this flaw and the learned policy does not generalize against other opponents types. Therefore, we propose and open source different agents to be used for training, see Table 3. StaticAgent is an extremely simple agent that always executes the stop action, the advantage of using this agents is that rewards are noise-free. SmartRandomNoBomb agent is a very challenging opponent. It moves randomly among the filtered actions, therefore learning against it provides better generalization. It does not place bombs, and thus never commits suicide. In contrast to SimpleAgent and CautiousAgent, it does not use Dijkstra, and it takes less time to act. CautiousAgent is based on a modification of SimpleAgent, the main idea is to only let this agent place a bomb when it is certain to kill an opponent. However, with this adaptation, the "weak" opponent SimpleAgent is instantly turned into a "strong" player. Killing this agent requires quite advanced skills. When learning against this opponent, a large number 2https://youtu.be/3yUhI46Xx8o 3 Our open-sourced agents: Description of different agents. Table 3: StaticAgent, SmartRandom, SmartRandomNoBomb and CautiousAgent, will be helpful to baseline against and to train against. The bottom 3 agents are relatively stronger than the provided baseline, less prone to exploitation due to higher level of stochasticity, and fast decision makers to be used during training. SimpleAgent (Provided by Pommer- man) StaticAgent A heuristic agent that uses Dijkstra and rules for navigation, tune-up collection, and simple attacks. A boring agent that always executes the stop action. Helpful for learning agents as rewards are noise-free. Takes random actions from the filtered action space. Stochastic but careful actions render it a competitive opponent for RL agents. Similar to SmartRandom but force the agent not place bomb at all. SmartRandom (Random agent + Ac- tionFilter) SmartRandomNoBomb (Random agent + ActionFilter + No Bombing) CautiousAgent pleAgent) Skynet955 (Neural Network agent) (Modified Sim- The modification enables the agent to place a bomb if and only if it guarantees a kill. Agent that is trained with PPO using reward shaping, ActionFilter, and opponent curriculum learning. of games ended with draws, or the challenger is killed. For example, in our tests, our Skynet agent can easily achieve 70% wining percentage against SimpleAgent team, but only around 10% against CautiousTeam, if not excluding draws. 3 Conclusions and Future work In spite of the good performance of our skynet agents, there are still many potential avenues for future research. For example, recent innovations, such as curiosity [2, 8], centralized learning with decentralized execution [5], or difference rewards [4] have shown to be able to produce good results in related domains, it remains to experimentally verify how effective they would be in the challenging domain of Pommerman. One challenge of multi-agent learning in partial observable environments is the credit assignment for each individual agent's behavior [3, 6]. Whenever there is a success or failure, it is important to correctly identify who is to reward or blame. Unfortunately, the original environment provided by Pommerman does not provide any such information. However, if centralized training [5] can be used by revising the environment, information might be helpful in devising more accurate reward function, for example: identification of bombs' owners and bombs' kickers; in general for any relevant event occurred (wood destruction, enemy's death, etc.) identifying which agent is to reward or blame could be based on who is responsible for the corresponding exploding bomb. Lastly, an ensemble consists of multiple neural net models may also improve the playing performance in the competition setting. References [1] Y. Bengio, J. Louradour, R. Collobert, and J. Weston. Curriculum learning. In Proceedings of the 26th annual international conference on machine learning, pages 41 -- 48. ACM, 2009. [2] Y. Burda, H. Edwards, A. Storkey, and O. Klimov. Exploration by random network distillation. arXiv preprint arXiv:1810.12894, 2018. [3] S. Devlin, L. Yliniemi, D. Kudenko, and K. Tumer. Potential-based difference rewards for multiagent reinforcement learning. In AAMAS, pages 165 -- 172, 2014. [4] J. N. Foerster, G. Farquhar, T. Afouras, N. Nardelli, and S. Whiteson. Counterfactual Multi-Agent Policy Gradients. In 32nd AAAI Conference on Artificial Intelligence, 2017. [5] J. N. Foerster, N. Nardelli, G. Farquhar, T. Afouras, P. H. S. Torr, P. Kohli, and S. Whiteson. Stabilising Experience Replay for Deep Multi-Agent Reinforcement Learning. In International Conference on Machine Learning, 2017. [6] P. Hernandez-Leal, B. Kartal, and M. E. Taylor. Is multiagent deep reinforcement learning the answer or the question? A brief survey. arXiv preprint arXiv:1810.05587, 2018. [7] L. Panait and S. Luke. Cooperative Multi-Agent Learning: The State of the Art. Autonomous Agents and Multi-Agent Systems, 11(3), Nov. 2005. [8] D. Pathak, P. Agrawal, A. A. Efros, and T. Darrell. Curiosity-driven exploration by self-supervised prediction. In International Conference on Machine Learning (ICML), volume 2017, 2017. [9] C. Resnick, W. Eldridge, D. Ha, D. Britz, J. Foerster, J. Togelius, K. Cho, and J. Bruna. Pommerman: A multi-agent playground. arXiv preprint arXiv:1809.07124, 2018. [10] J. Schulman, F. Wolski, P. Dhariwal, A. Radford, and O. Klimov. Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347, 2017. 4
1001.4122	1	1001	2010-01-23T02:31:26	Distributed Control of the Laplacian Spectral Moments of a Network	[ "cs.MA", "cs.CE" ]	It is well-known that the eigenvalue spectrum of the Laplacian matrix of a network contains valuable information about the network structure and the behavior of many dynamical processes run on it. In this paper, we propose a fully decentralized algorithm that iteratively modifies the structure of a network of agents in order to control the moments of the Laplacian eigenvalue spectrum. Although the individual agents have knowledge of their local network structure only (i.e., myopic information), they are collectively able to aggregate this local information and decide on what links are most beneficial to be added or removed at each time step. Our approach relies on gossip algorithms to distributively compute the spectral moments of the Laplacian matrix, as well as ensure network connectivity in the presence of link deletions. We illustrate our approach in nontrivial computer simulations and show that a good final approximation of the spectral moments of the target Laplacian matrix is achieved for many cases of interest.	cs.MA	cs	Distributed Control of the Laplacian Spectral Moments of a Network Victor M. Preciado, Michael M. Zavlanos, Ali Jadbabaie, and George J. Pappas 1 0 1 0 2 n a J 3 2 ] A M . s c [ 1 v 2 2 1 4 . 1 0 0 1 : v i X r a Abstract -- It is well-known that the eigenvalue spectrum of the Laplacian matrix of a network contains valuable infor- mation about the network structure and the behavior of many dynamical processes run on it. In this paper, we propose a fully decentralized algorithm that iteratively modifies the structure of a network of agents in order to control the moments of the Laplacian eigenvalue spectrum. Although the individual agents have knowledge of their local network structure only (i.e., myopic information), they are collectively able to aggregate this local information and decide on what links are most beneficial to be added or removed at each time step. Our approach relies on gossip algorithms to distributively compute the spectral moments of the Laplacian matrix, as well as ensure network connectivity in the presence of link deletions. We illustrate our approach in nontrivial computer simulations and show that a good final approximation of the spectral moments of the target Laplacian matrix is achieved for many cases of interest. I. INTRODUCTION A wide variety of distributed systems composed by au- tonomous agents are able to display a remarkable level of self-organization despite the absence of a centralized coordinator [1, 2]. For example, the structure of many "self-engineered" networks, such as social and economic networks, emerges from local interactions between agents aiming to optimize their local utilities [3]. Motivated by the implications of a network's Laplacian spectrum on its structure (i.e., number of connected components) and behavior of dynamical processes implemented on it (i.e., speed of convergence of distributed consensus algorithms), we propose a distributed model of graph evolution in which autonomous agents can modify their local neighborhood in order to control a set of moments of the network Laplacian spectrum. The eigenvalue spectra of a network provide valuable information regarding the behavior of many dynamical pro- cesses running within the network [4]. For example, the eigenvalue spectrum of the Laplacian matrix of a graph affects the mixing speed of Markov chains [5], the stability of synchronization of a network of nonlinear oscillators [6, 7], the spreading of a virus in a network [8, 9], as well as the dynamical behavior of many decentralized network algorithms [10]. Similarly, the second smallest eigenvalue of the Laplacian matrix (also called spectral gap) is broadly considered a critical parameter that influences the stability This work is partially supported by the ONR MURI HUNT and AFOR MURI Complex Networks programs Victor M. and George J. Pappas Preciado, Michael M. Jad- babaie, Labora- tory, Uni- versity USA {preciado,zavlanos,jadbabai,pappasg}@seas.upenn.edu are with GRASP Science, 19104, Applied PA Pennsylvania, School of Engineering and Zavlanos, Ali of Philadelphia, and robustness properties of dynamical systems that are im- plemented over information networks [11, 12]. Optimization of the spectral gap has been studied by several authors both in a centralized [13] -- [15] and decentralized context [16]. In contrast, our approach focuses on controlling the moments of the Laplacian eigenvalue spectrum. In this way, we can influence the behavior of certain dynamical processes run within the network. As we show, the benefit of controlling the spectral moments, and especially the lower order ones, lies in lower computational cost and elegant distributed implementation. A major challenge in our approach is to efficiently control the spectral moments of a network in a fully distributed fash- ion while maintaining network connectivity in the presence of link deletions. Our work is related to [17], where a fully distributed algorithm is proposed to compute the full set of eigenvalues and eigenvectors of a matrix representing the network topology. However, our approach is computationally cheaper since computation of the spectral moments does not require a complete eigenvalue decomposition, but can be performed distributively by averaging local network infor- mation, such as node degrees. On the other hand, control of the network structure to the desired set of spectral moments is based on greedy actions (link additions and deletions) that are the result of distributed agreement protocols between the agents. We show that our distributed topology control algorithm is stable and converges to a network with spectral moments "close" to the desired. The performance of our algorithm is illustrated in computer simulations. The rest of this paper is organized as follows. In Section II, we formulate the problem under consideration and review some terminology. In Section III, we derive closed-form expressions for the first four moments of the Laplacian spectrum in terms of graph properties that can be measured locally. Based on these expressions, we introduce a dis- tributed algorithm to compute these moments. In Section IV, we propose an efficient distributed algorithm to control of the spectral moments of a network. Finally, in Section V, we illustrate our approach with several computer simulations. II. PRELIMINARIES & PROBLEM DEFINITION Let G = (V, E) denote a graph on n nodes, where V = {v1, . . . , vn} denotes the set of nodes and E ⊆ V × V is the set of edges. If (vi, vj) ∈ E whenever (vj, vi) ∈ E we say that G is undirected and call nodes vi and vj adjacent (or neighbors), which we denote by vi ∼ vj. The set of all nodes adjacent to node v constitutes the neighborhood of node v, defined by N v = {w ∈ V : (v, w) ∈ E}, and the number of those neighbors is called the degree of node v, denoted by deg v. In this paper, we consider finite simple graphs, meaning that two nodes are connected by at most one edge and self-loops are not allowed. We define a walk from v0 to vk of length k to be an ordered sequence of nodes (v0, v1, ..., vk) such that vi ∼ vi+1 for i = 0, 1, ..., k − 1. We say that a graph G is connected if there exists a walk between every pair of nodes. If v0 = vk, then the walk is closed. A closed walk with no repeated nodes (with the exception of the first and last nodes) is called a cycle. Triangles and quadrangles are cycles of length three and four, respectively. Let d (v, w) denote the distance between two nodes v and w, i.e., the minimum length of a walk from v to w. We say that v and w are k-th order neighbors if d (v, w) = k and define the k-th order neighborhood of a node v as the set of nodes within k = {w ∈ V : d (v, w) ≤ k}. a distance k from v, i.e., N v A k-th order neighborhood, induces a subgraph Gv k ⊆ G k defined as the set of with node-set N v edges in E that connect two nodes in N v k . We say that a graphical property PG is locally measurable within a radius k around a node v if PG is exclusively a function of the k ). For example, both neighborhood subgraph, i.e., PG = f (Gv the degree and the number of triangles touching a node are locally measurable within a radius 1. Also, the number of quadrangles touching a node is locally measurable within a radius 2. k and edge-set E v Graphs can be algebraically represented via the adjacency and Laplacian matrices. The adjacency matrix of an undi- rected graph G, denoted by AG = [aij ], is an n×n symmetric matrix defined entry-wise as aij = 1 if nodes vi and vj are adjacent and aij = 0 otherwise.1 The powers of the adjacency matrix is related to walks in a graph. In particular we have the following results [18]: Lemma 2.1: The number of closed walks of length α joining node vi to itself is given by the i-th diagonal entry of the matrix Aα G. G)ii = deg vi, (A3 2Ti and (cid:0)A4 G(cid:1)ii = 2Qi + (deg vi)2 +Pj∈Ni Corollary 2.2: Let G be a simple graph. Denote by Ti and Qi the number of triangles and quadrangles touching node vi, respectively. Then (AG)ii = 0, (A2 G)ii = (deg vj − 1). Arranging the node degrees on a diagonal matrix yields the degree matrix DG = diag (deg vi). Then, the Laplacian matrix LG of a graph G can be defined by LG = DG − AG. Let λ1 ≤ λ2 ≤ ... ≤ λn be the eigenvalues of LG, where 1 is the vector of all ones. One can prove that LG is positive semidefinite and λ1 = 0. Furthermore, G is connected if and only if λ2 > 0, or equivalently, if ker LG = span{1} [18]. As a result, we have the following well-known result: Theorem 2.3 ( [19]): Consider a fixed undirected graph G on n nodes and let θi(t) ∈ R denote the state variable of node i. Let θ(t) = [θi(t)] ∈ Rn be the vector of all states and assume θ(t) = −LGθ(t). Then the network G is connected if and only if, lim t→∞ θ(t) = 1 n n Xi=1 θi(0)1 ∈ span{1}. (1) 1For simple graphs with no self-loops, aii = 0 for all i. 2 for all initial conditions θ(0) ∈ Rn. Theorem 2.3 says that the graph G is connected if and only if all nodes eventually reach a consensus on their state values θi(t), for all initial conditions. Therefore, connectivity of a network G can be verified almost surely by comparing the asymptotic state values (1) of all agents, for any random initialization. Note that a similar result can be obtained by application of a finite-time maximum (or minimum) consensus [20]. A. Problem Definition Consider a discrete-time sequence of graphs {G(s)}s≥1 where s ∈ {1, 2, . . . } is the discrete time index. We denote by {λi(s)}s≥1 the set of Laplacian eigenvalues of G(s). We define the k-th spectral moment of the Laplacian matrix of G(s) at time s as: mk (s) , 1 n n Xi=1 λk i (s). Similarly, the k-th centralized spectral moment of the Lapla- cian can written as: ¯mk(s) , 1 n (λi(s) − m1 (s))k n Xi=1 (cid:18)k r(cid:19) (−1)k−r mr(s)mk−r Xr=0 k 1 = (s). (2) The first four centralized spectral moments of the Laplacian corresponds to the mean, variance, skewness and kurtosis of the eigenvalue spectrum and they play a central role in this paper. Define further the error function: CME (G(s)) = 4 Xk=0h( ¯mk (s))1/k − ( ¯m∗ k)1/ki2 , (3) where ¯m∗ k denotes a given set of desired centralized mo- ments. Since the k-th moment is the k-th power-sum of the Laplacian eigenvalues, we include the exponents 1/k in the above error function with the purpose of assigning the same dimension to the summands in (3). Then, the problem addressed in this paper is: Problem 1: Given an initially connected graph G(0), de- sign a distributed algorithm that iteratively adds or deletes links from G(s), so that the connectivity of G(s) is main- tained for all time s and the error function CME(G(s)) is locally minimized for large enough s. In what follows, we first propose a distributed algorithm to efficiently compute and update CME(G(s)) without any ex- plicit eigendecomposition (Section III). Then, in Section IV, we propose a greedy algorithm where the most beneficial edge addition/deletion is determined based on a distributed agreement over all possible actions that satisfy network connectivity (Theorem 2.3). In this framework, the time variable s increases by one whenever an action is taken (i.e., an addition or deletion of a link). For simplicity, we assume that actions are taken one at a time, although this assumption can be relaxed to accommodate more complex action schemes. III. DISTRIBUTED COMPUTATION OF SPECTRAL MOMENTS In what follows, we assume that the agents in the network have very limited knowledge of the network topology. In particular, we assume that every agent v only knows the topology of the second-order neighborhood subgraph around it, Gv 2 . (This is the case, for example, for many online social networks, where each individual can retrieve a list of his friends' friends.) Then, computing the first four Laplacian spectral moments relies on counting the presence of certain subgraphs, such a triangles and quadrangles, in every agent's neighborhood and averaging these quantities via distributed average consensus. In particular, since the matrix trace opera- tor is conserved under diagonalization (in general, under any similarity transformation) the first three spectral moments of the Laplacian matrix of a graph can be written as mk (LG) = 1 N tr Lk G = 1 n tr (DG − AG)k = 1 n k Xp=0 (cid:18)k p(cid:19)(−1)p tr (Ap GDk−p G ), (4) for k ≤ 3, where we have used the fact that the trace is preserved under cyclic permutations (i.e., tr ADD=tr DAD= tr DDA). We cannot use Newton's binomial expansion for the forth moment; nevertheless, we may still obtain the following closed form solution: m4 (LG) = 1 = 1 n tr (DG − AG)4 G(cid:1) − 4tr(cid:0)D3 GAG(cid:1) + 4tr(cid:0)D2 n (cid:2)tr(cid:0)D4 +2tr (DGAGDGAG) − 4tr(cid:0)DGA3 GA2 G(cid:1) G(cid:1) + tr(cid:0)A4 G(cid:1)(cid:3) . (5) Expanding the traces that appear in (4) and (5) we get GAp tr(cid:0)Dq G(cid:1) = n Xi=1 (deg vi)q(cid:0)Ap G(cid:1)ii and tr (DGAGDGAG) = n n Xi=1 Xj=1 (deg viaij) (deg vjaji) = n n Xi=1 Xj=1 (deg vi deg vj) aij = n Xi=1 deg vi Xj∈Ni deg vj, which substituted back in equations (4) and (5) give the following expression for k ≤ 3 mk (LG) = 1 n n k Xi=1 Xr=0 (cid:18)k r(cid:19)(−1)r (deg vi)k−r (cid:0)Ar G(cid:1)ii . (6) For k = 4, we can also simplify the Laplacian spectral moment, which now becomes m4 (LG) = 1 n n Xi=1 h(deg vi)4 − 4 (deg vi)3 (AG)ii + +4 (deg vi)2 (AG)2 deg vj− 3 Substituting the expressions for (AG)r ii form Lemma 2.1 and Corollary 2.2 in equations (6) and (7) we obtain the first four spectral moments of the Laplacian matrix LG as a function of the second-order neighborhood subgraphs only m1 (LG) = m2 (LG) = m3 (LG) = m4 (LG) = 1 n 1 n 1 n 1 n n (8a) (8b) n deg vi, Xi=1 Xi=1 h(deg vi)2 + deg vii , Xi=1 h(deg vi)3 + 3 (deg vi)2 − 2Tii , Xi=1 h (deg vi)4 + 4 (deg vi)3 + (deg vi)2 − n n (8c) (8d) − deg vi + (2 deg vi + 1) Xj∈Ni deg vj − 8Ti deg vi + 2Qi  . Note that the expressions for the spectral moments in equations (8) are all averages of locally measurable quan- tities (within a 2-hop neighborhood), namely, node degrees, triangles and quadrangles touching the node. Hence, we can apply consensus and use the result of Theorem 2.3 to obtain the first four moments in a distributed way. IV. DISTRIBUTED CONTROL OF SPECTRAL MOMENTS A. Possible Local Actions The possible actions (or control variables) we consider are local link additions and local link deletions. A link addition is local if it connects a node with another node within its second-order neighborhood. Since agents in the network only know their local neighborhood, a fully distributed algorithm must limit edge additions to be local. (One could extend the algorithm to allow connections between nodes being further than two hops away, but this option would require much more computation and communication.) Let N i 1(s) and N i d(s) , (cid:8)(i, j) ∈ E(s) j ∈ N i 2(s) denote the sets of neighbors and two-hop neighbors of node i at time s ≥ 0, respectively. Since any of the two nodes adjacent to a link can take an action to delete that link, we need to decide which of the two nodes has the authority to delete the link. To avoid ambigu- ities, we define the set of edges that node i has authority to remove as: E i 1(s), i > j(cid:9). Similarly, to disambiguate between nodes adding a (still non-existing) link between them, we define the set of potential edges that node i can create as: E i a(s) , 1(s), i > j(cid:9)2. In other words, a(s), or delete a node i can either add a link (i, j) ∈ E i link (i, j) ∈ E i d(s). Note that link deletions may violate network connectivity. In this case, those link deletions should be excluded from the set of allowable actions. In the next two sections we address the cases of link deletions and link additions separately. (cid:8)(i, j) ∈ E(s) j ∈ N i 2(s)\N i −4 (deg vi) (AG)3 ii + 2 deg vi Xj∈Ni iii . ii + (AG)4 (7) 2Note that since the indices of all nodes in the network are distinct, this definition results in a unique assignment of links to nodes. 1) Link Deletions: Network connectivity is typically in- ferred from the number of trivial eigenvalues of the Laplacian matrix. However, such approaches are not applicable in our framework, since we assume no global knowledge of the network topology G(s), but only knowledge of local neighborhoods. Instead, we employ finite-time-maximum consensus which is a distributed algorithm and converges to equal values on nodes belonging to the same connected component of a graph (Theorem 2.3). Therefore, if deletion of a link violates connectivity, both nodes adjacent to that link will lie in different connected components and will have different consensus values. In what follows, we extend this idea to simultaneously checking connectivity for all possible edge deletions in the graph using a single high-dimensional consensus algorithm. Consider node j that has authority to remove any of the links in the set E j d(s). Each one of these links needs to be checked with respect to connectivity and each connectivity verification relies on a scalar consensus update, according to Theorem 2.3. Therefore, checking all links in E j d(s) requires E j d(s) consensus updates.3 We associate with every link in E j d(s) a consensus variable, and stacking all these variables in a vector we obtain the state vector xjj (s) ∈ RE j d(s). Running a distributed consensus over the network, requires participation of all other nodes i 6= j. This is possible by defining the state variables xij (s) ∈ RE j d(s). All vectors xij (s) are initialized randomly and are updated by node i according to the following maximum consensus: Case I: If (i, j) 6∈ E i not neighbors, then d(s), i.e., if nodes i and j are d(s) ∪ E j xij(s) := max k∈N i 1 (s) {xij (s), xkj (s)} , (9) with the maximum taken elementwise on the vectors, Case II: If (i, j) ∈ E j and node j has authority over link (i, j), then d(s), i.e., if nodes i and j are neighbors [xij (s)](i,j) := and k∈N i max 1 (s)\{j}(cid:8)[xij (s)](i,j), [xkj (s)](i,j)(cid:9) , (10) [xij (s)](l,j) := max k∈N i 1 (s)(cid:8)[xij (s)](l,j), [xkj (s)](l,j)(cid:9) , (11) for l 6= i, where [xkj (s)](l,j) ∈ R denotes the entry of xkj (s) corresponding to the link (l, j), Case III: If (i, j) ∈ E i and node i has authority over link (i, j), then d(s), i.e., if nodes i and j are neighbors k∈N i max 1 (s)\{j}(cid:8)[xii(s)](i,j), [xki(s)](i,j)(cid:9) . [xii(s)](i,j) := (12) Once consensus (9) -- (12) has converged, node i compares the entries [xii(s)](i,j) and [xji(s)](i,j) for all (i, j) ∈ d(s). Since, violation of connectivity due to deletion of E i the link (i, j) would result in nodes i and j being in dif- ferent connected components of the network, [xii(s)](i,j) = 3We define by X the cardinality of the set X. 4 d for all j ∈ V ; Algorithm 1 Connectivity verification Require: xij ∈ RE i Require: Tij = [0 . . . 1i . . . 0]T , ∀ j ∈ V; 1: if ∃ j ∈ V such that min{Tij} = 0 then 2: 3: 4: else if min{Tij} = 1 for all j ∈ V then 5: 6: end if Update xij by (9) -- (12); Tij := maxk∈N 1 σ by (13); {Tij, Tkj}; Update E i i [xji(s)](i,j) implies that the reduced network is still con- nected. Hence, we can define a set E i σ(s) , (cid:8)(i, j) ∈ E i d(s)[xii(s)](i,j) = [xji(s)](i,j)(cid:9) , (13) containing the safe links adjacent to node i that if deleted, connectivity is maintained. 2) Connectivity Verification: The connectivity verification of link deletions, discussed in Section IV-A.1, is illustrated in Alg. 1. Convergence of the finite-time consensus (9) -- (12) is captured by a vector of tokens Tij(s) ∈ {0, 1}n, initialized as Tij(s) := [0 . . . 1i . . . 0]T for all j ∈ V and indicating that node i has initialized the consensus variables for link deletions for which node j is responsible. In particular, when all tokens of all nodes have been collected (line 4, Alg. 1), then consensus has converged and the set of safe σ(s) can be computed (line 5, Alg. 1). Note link deletions E i that node i does not need to know the neighbor sets of its non-neighbors in the network, neither their size. Instead, the vectors xij (s) are initialized both in values and dimension as soon as vectors xkj (s) are received from a neighbor 1(s). Clearly, xjj (s) are initialized first and then k ∈ N i propagated in the network via maximum consensus until the information they contain reaches node i. B. Most Beneficial Local Action As discussed in Problem 1, the objective of this work is to minimize the error function CME(G(s)). For this we propose a greedy algorithm, which for every time s selects the action that maximizes the quantity CME(G(s))−CME(G(s+1)), if such an action exists, and terminates if no such action exists. By construction, such an algorithm converges to a network that locally minimizes CME(G(s)), while in Section V, we show that it performs well in practice too. In what follows we first compute the effect of a link addition or deletion on the four first spectral moments. Although distributed consensus could be used to compute the new moments resulting from each possible action, as in Section III, this would clearly be computationally very expensive. Instead, we can achieve this goal locally and with minor computational overhead, based on the observation that the addition or removal of an edge (i, j) only influences the degrees of nodes i and j, as well as and the triangles and quadrangles touching their neighboring nodes. Hence, agents i and j can communicate to compute the net increment in the spectral moments due to the addition or deletion of edge (i, j). In particular, we get the following expressions for the increments in the first three moments ∆m ±(i,j) 1 = ± 2 n , ∆m ±(i,j) 2 = ∆m ±(i,j) 3 = 2 n 1 n [1 ± (di + dj + 1)] , (cid:2)(3 ± 6) (di + dj) ± 3(d2 i + d2 j ) + (6 ± 2) ∓ 6Tij(cid:3) where the notation ± (i, j) indicates a link addition (+) or deletion (−) and the dependence on time s has been dropped for simplicity. (Similarly, one can obtain a complicated closed-form expression for ∆m±(i,j) , which we omit due to space limitations.) Then, agent's i copy of the k-th spectral moment mi k(s) becomes 4 m±(i,j) k (s) = mi k(s) + ∆m±(i,j) k (s), and the associated centralized moment ¯m±(i,j) (s) can be computed using (2). Then, for all possible actions discussed in Section IV-A, agent i computes the error function k CME±(i,j)(s) = 4 Xk=0 (cid:20)(cid:16) ¯m±(i,j) k (s)(cid:17)1/k − ( ¯m∗ k)1/k(cid:21)2 . Then, the local most beneficial action to the target centralized moments, namely, the action that results in the maximum decrease in the error function CME(·), can be defined by4 νi(s) , max(cid:26)argmin where CMEi(s) denotes agent i's copy of CME(s), and the largest decrease in the error associated with action νi(s) becomes: (cid:0)CME±(i,j)(s) − CMEi(s)(cid:1)(cid:27) , j CMEi(s) = CME±(i,νi(s))(s) if minj(CME±(i,j)(s) − CME(s)) ≤ 0 and CMEi(s) = D, otherwise. Note that CMEi(s) is nontrivially defined only if the exists a link adjacent to node i that if added or deleted decreases the error function CME(·). Otherwise, a large value D > 0 is assigned to CMEi(s) to indicate that this action is not beneficial to the final objective. We can include all information of a best local action in the vector vi(s) , (cid:2)i νi(s) CMEi(s) ¯m±(i,νi(s))(s)(cid:3)T ∈ R7 containing the local best action (i, νi(s)), the associated distance to the desired moments CMEi(s), and the vector of centralized moments ¯m±(i,νi(s))(s) due to this action. In the following section we discuss how to compare all local actions vi(s) for all nodes i ∈ V to obtain the one that decreases the distance to the desired moments the most. 4The max in the expression bellow indicates that in case of ties in the min, the highest index node wins. 5 Algorithm 2 Globally most beneficial action Require: vi , [i νi CMEi ¯m±(i,νi)]T ; Require: Ti := [0 . . . 1i . . . 0]T ; 1: if min{Ti} = 0 then 2: 3: 4: else if min{Ti} = 1 and [vi]3 < D then 5: 6: else if min{Ti} = 1 and [vi]3 = D then 7: 8: end if vi := vj, with j = max{argmink∈N i Ti := maxj∈N i {Ti, Tj}; 1 1 Update Ni, ¯mi and CMEi according to (14) -- (17); No beneficial action. Algorithm has converged; {[vi]3, [vk]3}; C. From Local to Global Action To obtain the overall most beneficial action, all local actions need to be propagated in the network and compared against each other. For this we require minimal storage and communication as well as no propagation of any information regarding the network topology. As in Section IV-A, we achieve this goal using a finite time minimum consensus algorithm. In particular, the desired local actions vi(s) are propagated in the network, along with vectors of tokens Ti(s) ∈ {0, 1}n, initialized as Ti(s) := [0 . . . 1i . . . 0]T , indicating that node i has transmitted its desired action. During every iteration of the algorithm, every node i communicates with its neighbors and updates its vector of tokens Ti(s) (line 3, Alg. 2), as well as its desired action vi(s) with the action vj(s) correspond- ing to the node j that contains the smallest distance to the target moments [vj (s)]3, i.e., j ∈ argmink∈N1i(s){[vi(s)]3, [vk(s)]3}. In case of ties in the distances to the targets [vj(s)]3, i.e., if 1 (s){[vi(s)]3, [vk(s)]3} contains more than one argmink∈N i nodes, then the node j with the largest label is selected (line 2, Alg. 2). Note that line 2 of Alg. 2 is essentially a minimum consensus update on the entries [vi(s)]3 and will converge to a common outcome for all nodes when they have all been compared to each other, which is captured by the condition minn j=1 Tij(s) = 1 (lines 4 and 6, Alg. 2). When the consensus has converged, if there exists a node whose desired action decreases the distance to the target moments, i.e., if [vi(s)]3 < D (line 4, Alg. 2), then Alg. 2 terminates with a greedy action and node i updates its set of neighbors N i 1(s) and vector of centralized moments ¯mi(s) (line 5, Alg. 2). If the optimal action is a link addition, i.e., if [vi(s)]2 6∈ N i 1(s) ∪ {[vi(s)]2} . (14) On the other hand, if the optimal action is a link deletion, i.e., if [vi(s)]2 ∈ N i 1(s), then 1(s + 1) := N i N i 1(s), then 1(s + 1) := N i N i 1(s)\ {[vi(s)]2} . (15) In all cases, the centralized moments and error function are updated by ¯mi(s + 1) := [vi(s)]4:7 (16) Reset 3 Reset 1 Run 1 Run 2 Run 3 Reset 2 Fig. 1. Synchronization: Assume node i is in run 1. Necessary for node i to transition to run 2 is that all other nodes are also in run 1, since otherwise node i will be missing tokens from the nodes that are not in run 1 yet (currently in run 3) and Alg. 2 will not be able to converge. Once node i transitions to run 2, it initializes all variables for that run with the latest values from run 1, while it maintains the variables of run 1 for nodes that are still in run 1 and it clears all variables of run 3 since, no node is in this run any more. and CMEi(s + 1) := [vi(s)]3, (17) respectively, where [vi(s)]4:7 = [[vi(s)]4 . . . [vi(s)]7]T . Fi- nally, if all local desired actions increase the distance to the target moments, i.e., if [vi(s)]3 = D (line 6, Alg. 2), then no action is taken and the algorithm terminates with a network topology with almost the desired spectral properties. This is because no action exists that can further decrease the distance to the target moments. D. Synchronization Communication time delays, packet losses, and the asym- metric network structure, may result in runs of the algorithm starting asynchronously, outdated information being used for future decisions, and consequently, nodes reaching different decisions for the same run. In the absence of a common global clock, the desired synchronization is ideally event triggered, where by a triggering event we understand the time instant that a message Msg[i] has been received by 1. We achieve such a any of node i's neighbors j ∈ N i synchronization by labeling every algorithm run in the set {1, 2, 3} and requiring that all information exchange takes place among neighbors that are in equally labeled runs [21]. Essentially, "fast" nodes wait for their "slower" peers and, hence, all nodes are always synchronized in the sequence {1, 2, 3, 1, 2, 3, . . . } (Fig. 1). V. NUMERICAL SIMULATIONS In this section we illustrate our algorithm with several numerical examples. 1 = 1.8, ¯m∗ 2 = 7.56, ¯m∗ Example 1 (Star networks): Consider a star network on 10 nodes. The first four central moments of the associated Laplacian matrix are: ¯m∗ 3 = 54.14, and ¯m∗ 4 = 453.49. Our objective is to control the topology of a randomly initialized network on 10 nodes so that it eventually has the same set of moments as the given star network. We observe in Fig. 2 that our algorithm decreases the error function (blue line) to zero in finite time. Similar performances are observed when we repeat this procedure for star networks of any size. Furthermore, although we are con- trolling the first four moments solely, the resulting network structures are exactly the star topologies whose moments we were trying to approximate. The perfect reconstruction observed in this case is due to the fact that a star graph is a 6 Fig. 2. Convergence of the error function CME(·) for the star graph (blue plot) and the two-stars graph (red plot). The subgraph in the upper right corner shows the behavior of the error function in a neighborhood of zero. Observe that our algorithm can match the first four moments of the star network with zero error in finite time, but can not exactly match the moments of a the two-stars graph, although the final error is very small. Fig. 3. by our algorithm (b). Structures of the two-stars network (a) and the network returned extreme case in which the graph topology is uniquely defined by their eigenvalue spectrum. Moreover, if each agent in a star network has access to its second-order neighborhood, it has access to the complete star topology. Example 2 (Two-stars network): Although our approach works very well for star networks, the case of two-stars networks points out one of its weaknesses, namely, its lim- itation in modeling network communities. In this example, we consider two star graphs on 10 nodes each, and connect their two central hubs with a link. The resulting graph is the two-stars graph shown in Fig. 3(a). As before, we initialize our algorithm with a random graph on 20 nodes and try to approximate the first four central moments of the two- stars graph. In Fig. 2, we observe that the error function (red line) quickly reaches a neighborhood of zero but does not reach zero exactly. Therefore, although our algorithm tries to generate the two hubs in the two-star network, its local nature will not allow it recover the highly-structured two- stars graph. Instead, it returns the final network shown in Fig. 3(b). Nevertheless, the eigenvalue spectra between the desired two-star network and the network in Fig. 3(b) are still very similar, as shown in Fig. 4. Example 3 (Chain vs. ring networks): The objective of 7 Fig. 4. Empirical cumulative distribution functions for the eigenvalues of the two-stars graph (blue) and the graph returned by our algorithm (red). this example is to illustrate how two structurally very similar (but topologically different) target graphs, such as a chain and a ring, may affect the performance of our algorithm. In particular, if we run our algorithm to control the first four centralized moments of an initially random graph towards the moments of a chain graph, we observe how the error function converges exactly to zero in finite time. Furthermore, the final result of our algorithm is an exact reconstruction of the chain graph. Nevertheless, when transforming the target graph from a chain graph into a ring graph (by adding a single link), an exact reconstruction is very difficult. In Fig. 5, we illustrate some graphs returned by our algorithm for different initial conditions when we control the set of moments toward the moments of a ring network on 20 nodes. Observe that, although the algorithm tends to create long cycles and the majority of nodes have degree two, it fails to recreate the exact structure of the ring graph due to the local nature of the algorithm (as in Example 2). However, although the structure of the resulting networks is not exactly the desired ring graph, their spectral properties are remarkably close to those of a ring. In Fig. 6, we illustrate the empirical cumulative distribution functions of the eigenvalues of the ring graph (blue plot), versus the four empirical cumulative distribution functions corresponding to the graphs in Fig. 5. Example 4 (Small-Worlds): In our final example, we use our algorithm to control the moments of a randomly gener- ated graph to those of a small-world network. We consider small-world graphs as defined in [22], namely, we take a ring of n nodes, and connect each node in the ring with all the nodes in its 3-hop neighborhood. Then, we randomly rewire a fraction of the resulting edges with probability p as proposed by Watts and Strogatz [22]. Our objective is to approximate the first four centralized moments of a random instance of a small world graph with n = 40 nodes and link probability p = 1/n. We observed a fast convergence of the error to a neighborhood of zero, i.e., CME(27 iterations) = .0009, which suggests (although does not guarantee) a good approximation between the spectra of the target small-world graph and the graph returned by our algorithm. We repeated Fig. 5. Networks returned by our algorithm when trying to match the first four central moments of a ring on 20 nodes. Fig. 6. Empirical cumulative distribution of eigenvalues for the ring graph with 20 nodes (blue plot) and the graphs in Fig. 5(a) (red), Fig. 5(b) (green), Fig. 5(c) (magenta) and Fig. 5(d) (cyan). this process for a link probability p = 4/n and similar results were obtained. We should note, however, that although the spectral properties between the target small-world graphs and the graphs returned by our algorithm are remarkably similar (Fig. 7), we expect the structures to be quite different as in Example 3 (although this difference is not easy to observe from a direct visualization of both graph structures). VI. CONCLUSIONS AND FUTURE RESEARCH In this paper, we have described a fully decentralized al- gorithm that iteratively modifies the structure of a network of agents with the objective of controlling the spectral moments of the Laplacian matrix of the network. Although we assume that each agent has access to local information regarding the graph structure, we show that the group is able to collectively aggregate their local information to take a global optimal decision. This decision corresponds to the most beneficial 8 [10] N.A. Lynch, Distributed Algorithms, Morgan Kaufmann Publishers, 1997. [11] A. Fax, and R. M. Murray, "Information Flow and Cooperative Control of Vehicle Formations," IEEE Transactions on Automatic Control, vol. 49, pp. 1465-1476, 2004. [12] R. Olfati-Saber, and R. M. Murray, "Consensus Problems in Networks of Agents with Switching Topology and Time-Delays," IEEE Trans- actions on Automatic Control, vol. 49, pp. 1520-1533, 2004. [13] A. Ghosh, and S. Boyd, "Growing Well-Connected Graphs," Proc. of the 45th IEEE Conference on Decision and Control, pp. 6605-6611, 2006. [14] R. Grone, R. Merris, and V.S. Sunder, "The Laplacian Spectrum of a Graph," SIAM Journal Matrix Analysis and Applications, vol. 11, pp. 218-238, 1990. [15] Y. Kim, and M. Mesbahi, "On Maximizing the Second-Smallest Eigenvalue of a State Dependent Graph Laplacian," IEEE Transactions on Automatic Control, vol. 51, pp. 116-120, 2006. [16] M.C. DeGennaro, and A. Jadbabaie, "Decentralized Control of Con- nectivity for Multi-Agent Systems," Proc. of the 45th IEEE Conference on Decision and Control, San Diego, CA, Dec. 2006, pp. 3628-3633. [17] D. Kempe, and F. McSherry, "A Decentralized Algorithm for Spectral Analysis," Journal of Computer and System Science, vol. 74, pp. 70- 83, 2008. [18] N. Biggs, Algebraic Graph Theory, Cambridge University Press, 2nd Edition, 1993. [19] A. Jadbabaie, J. Lin, and A. S. Morse, "Coordination of Groups of Mobile Autonomous Agents Using Nearest Neighbor Rules," IEEE Transactions on Automatic Control, vol. 48(6), pp. 988-1001, 2003. [20] J. Cortes, "Distributed algorithms for reaching consensus on general functions," Automatica, vol. 44(3), pp. 726-737, 2008. [21] M. M. Zavlanos, and G. J. Pappas, "Distributed Connectivity Control of Mobile Networks," IEEE Transactions on Robotics, vol. 24(6), pp. 1416-1428, 2008. [22] D.J. Watts, and S. Strogatz, "Collective Dynamics of Small World Networks," Nature, vol. 393, pp. 440-42, 1998. Fig. 7. world graphs in the example for p = 1/n (left) and p = 4/n (right). Empirical cumulative distrubution of eigenvalues for the small- link addition/deletion in order to minimize an error function that involves the first four Laplacian spectral moments of the network. The aggregation of the local information is achieved via gossip algorithms, which are also used to ensure network connectivity throughout the evolution of the network. Future work involves identifying sets of spectral moments that are reachable by our control algorithm. (We say that a sequence of spectral moments is reachable if there exists a graph whose moments match the sequence of moments.) Fur- thermore, we observed that fitting a set of low-order moments does not guarantee a good fit of the complete distribution of eigenvalues. In fact, there are important spectral parameters, such as the algebraic connectivity, that are not captured by a small set of spectral moments. Nevertheless, we observed in numerical simulations that fitting the first four moments of the eigenvalue spectrum often achieves a good reconstruction of the complete spectrum. Hence, a natural question is to describe the set of graphs most of whose spectral information is contained in a relatively small set of low-order moments. REFERENCES [1] N. Wiener, The Mathematics of Self-Organising Systems. Recent Developments in Information and Decision Processes, Macmillan, 1962. [2] H. Haken, Synergetics: An Introduction, 3rd Edition, Springer-Verlag, 1983. [3] M.O. Jackson, Social and Economic Networks, Princeton University Press, 2008. [4] V.M. Preciado, Spectral Analysis for Stochastic Models of Large-Scale Complex Dynamical Networks, Ph.D. dissertation, Dept. Elect. Eng. Comput. Sci., MIT, Cambridge, MA, 2008. [5] D.J. Aldous, Some Inequalities for Reversible Markov Chains, J. London Math. Soc. (2)25, pp. 564-576, 1982. [6] L.M. Pecora, and T.L. Carroll, "Master Stability Functions for Syn- chronized Coupled Systems," Physics Review Letters, vol. 80(10), pp. 2109-2112, 1998. [7] V.M. Preciado, and G.C. Verghese, "Synchronization in Generalized Erdos-R´enyi Networks of Nonlinear Oscillators," Proc. of the 44th IEEE Conference on Decision and Control, pp. 4628-4633, 2005. [8] M. Draief, A. Ganesh, and L. Massouli´e, "Thresholds for Virus Spread on Networks," Annals of Applied Probability, vol. 18, pp. 359-378, 2008. [9] V.M. Preciado, and A. Jadbabaie, "Spectral Analysis of Virus Spread- ing in Random Geometric Networks," Proc. of the 48th IEEE Confer- ence on Decision and Control, 2009. (to appear).
1910.10109	1	1910	2019-10-22T16:53:28	Distributed interference cancellation in multi-agent scenarios	[ "cs.MA", "cs.LG", "cs.RO", "eess.SY", "eess.SP", "eess.SY" ]	This paper considers the problem of detecting impaired and noisy nodes over network. In a distributed algorithm, lots of processing units are incorporating and communicating with each other to reach a global goal. Due to each one's state in the shared environment, they can help the other nodes or mislead them (due to noise or a deliberate attempt). Previous works mainly focused on proper locating agents and weight assignment based on initial environment state to minimize malfunctioning of noisy nodes. We propose an algorithm to be able to adapt sharing weights according to behavior of the agents. Applying the introduced algorithm to a multi-agent RL scenario and the well-known diffusion LMS demonstrates its capability and generality.	cs.MA	cs	DISTRIBUTED INTERFERENCE CANCELLATION IN MULTI-AGENT SCENARIOS Mahdi Shamsi, Alireza Moslemi Haghighi, Farokh Marvasti Multimedia and Signal processing Lab (MSL), Advanced Communications Research Institute (ACRI), EE Department of Sharif University of Technology, Tehran, I.R.Iran 9 1 0 2 t c O 2 2 ] A M . s c [ 1 v 9 0 1 0 1 . 0 1 9 1 : v i X r a ABSTRACT This paper considers the problem of detecting impaired and noisy nodes over network. In a distributed algorithm, lots of processing units are incorporating and communicating with each other to reach a global goal. Due to each one's state in the shared environment, they can help the other nodes or mislead them (due to noise or a deliberate attempt). Previous works mainly focused on proper locating agents and weight assignment based on initial environment state to minimize malfunctioning of noisy nodes. We propose an algorithm to be able to adapt sharing weights according to behavior of the agents. Applying the introduced algorithm to a multi-agent RL scenario and the well-known diffusion LMS demonstrates its capability and generality. Index Terms -- Distributed Algorithm, Noise Reduction, Distributed Estimation, Multi-agent Reinforcement Learning, Noise Propagation. havior, and hence performance of the nodes, can be changed. This could be under the influence of different factors, like noise increasing or an intervening procedure such as attacking and hacking. This article is organized as follows. In Section 2, we study on noise propagation in three typical scenarios over a graph. We begin with a discrete probabilistic non-linear state model and then simplify it to a linear sharing framework. Then we propose an algorithm for malfunctioning detection in a dis- tributed scenario. Section 3 is dedicated to distributed es- timation; the Adapt Then Combine (ATC) scenario is stud- ied as a prominent approach for diffusion Least Mean Square (LMS)[13, 14]. Capability of the proposed method is illus- trated by applying to the ATC with a high noisy node in differ- ent situations [of dense and sparse connection graph]. In Sec- tion 4, a manipulated RL environment is introduced to be as a multi-agent approach for studying an intervened scenario.At last, Section 5 concludes the paper. 1. INTRODUCTION 2. NOISE PROPAGATION In this correspondence, the problem of error propagation in a distributed algorithm is investigated. There is a variety of examples in which distributed algorithms are used such as distributed estimation[1, 2], sensor networks[3], distributed control systems[4, 5], swarm intelligence algorithms[6, 7], multi-agent reinforcement learning scenarios[8, 9] and signal processing on graph[10, 11, 12] just to name but a few. Con- sidering the general viewpoint of graph over networks; each algorithm in its core consists of nodes(agents) as the com- puting units and links as the communication path. It should be noted that distributed algorithms can provide us with ro- bustness, reliability and faster convergence in exchange for a larger number of low-cost processing units. A simplifying assumption, for developing and analyzing distributed algo- rithms, is considering the shared environment to be akin to all of the graph nodes[15]. Assuming a stationary behavior of an environment, several algorithms have been proposed for optimizing, locating and placing a node in the initial state. There are also certain ways of analyzing the environment for assigning weights to each node, Since there could be short- comings due to relying on each node's observed environment. However, in a non-stationary scenario, the environmental be- In this section, we introduce error propagation as a general issue in distributed algorithms. For the purpose of generality, some simplifications are applied to this procedure. From the control systems modeling perspective, all the distributed con- tents and functionalities are stated by overall system's state equations. Every node can be assumed an agent or an esti- mator. Hence the connecting links are interpreted as corre- sponding connections and sharing procedures. In this point of view, we are not only interested in system's steady state. The transient behavior of nodes are also a major concern. In this system, it is assumed that there are N nodes which are in corporation with each other to reach a global goal. In a general case, all of A, B, C and D are nonlinear operators. A general state equations are shown as (cid:126)x(n) = A((cid:126)x(n)) + B((cid:126)u(n)); (cid:126)y(n) = C((cid:126)x(n)) + D((cid:126)u(n)). By assuming a graph as a distributed platform, functional- ity of each can be analyzed in terms of signal flow graph paradigm. This assumption leads us to interpret each operator more clearly. Hence, with minor alterations, the operator A becomes the well-known adjacency matrix of the main graph. For the purpose of simplicity, these operators are assumed to be linear. Furthermore, it is sufficient to study information sharing procedures, which are commonly linear like in recent distributed estimation methods[1]. (cid:126)x(n) = A.(cid:126)x(n) + B.(cid:126)u(n); (cid:126)ai = (cid:126)ai mean((cid:126)ai) where (cid:126)a and (cid:126)ai are the ithrows of A and A (Adjacency ma- trix). Another simplification that we should mention is that we are using a symmetric directional graph. It means, each pair of connected nodes can communicate with each other. However, the effect of the shared value of each node in the destination node is determined by the number of incoming links to the latter one. Generally, in combination step of shar- ing procedures of distributed algorithms, all incoming esti- mated values are averaged. Thus, sum of elements in each row of A should be normalized to 1. As a desirable feature of distributed structure, at least in many cases, isomorphism of the main graph is considered in the following way; each two nodes can be connected indepen- dent of the others and all nodes have the same level of impor- tance. There are trends focusing on optimizing significance of nodes based on their processing power or environmental turbulence[15, 16]. These approaches mainly concentrate on proper locating of the agents[17]; in other words, optimizing the placement procedure. Hence, by assuming an arbitrary node labeled as i, all other nodes have the same expected data flow from i. Due to the isomorphism of the graph, with malfunctioning of an arbitrary node (which could be due to the high level of noise or hacking attacks), all the other nodes are supposed to re- ceive the same level of wrong (misleading) data on average. It is clear that, these methods consider a stationary condition. Hence, after proceeding through certain number of steps of the main function of the system, it seems reasonable to think of the present environmental state. For a better comprehension, and without loss of general- ity, we only consider the case of a single impaired agent. It can be easily extended to the larger number of damaged nodes because of the simplifying linearity assumption. It means we can study two arbitrary nodes in the graph and analyze the procedure of exchanging data. In order to predict the incom- ing data from the node i to the arbitrary destination node j, we should know expected number of paths between them. It can be done through three different cases. By defining p (cid:44) M−1 N−2 and s (cid:44) M N−1 where N and M are respectively number of all nodes and average number of neighbors of an arbitrary node, we have 3 cases as shown in the following table. In the 2nd case, the paths are supposed to be determined at the destina- tion node. It is claimed that, the third case can be interpreted as the impulse response of the overall system between two arbitrary nodes of the pre-described graph. Hence, one node can intro- duce a pandemic interference over the graph. It must be men- <i> K>1 case i 1 2 3 general no-loops no self-loops = expected #paths of length K η P (N − 2, k − 1) × s × pk−1 (cid:80)K−1 (N − 2)K−1 × sK−1 × p (cid:96)=0 (N − 2)K−(cid:96)−1 × sK−(cid:96)−1 × p Table 1: Expected #paths of length K between two arbitrary nodes [= η K ], in a graph with N nodes and, on average, M neighbors for each one; for all 3 cases: η = s. <i> <i> 1 tioned that, even in simple distributed methods, each node is continuously compensating the incoming distortion based on its predefined functionality. However we are insisting on their limited capability of compensation. In order to prevent noise propagation in a graph, we introduce a simple method of weighting and successor decision making on node elimina- tion. In each step, each node receive some data from its neigh- bors (including itself). Using a simple procedure of weighting it will then use all of the incoming estimations thoroughly; in which, a node attempts to reach a predefined goal via opti- mizing a cost function (the goal can also be non-stationary). Although different ordering of these two steps results in vari- ous algorithms, it would not harm the main body of our pro- posed algorithm. Additionally, we can assume a malfunction- ing node in some general graph. In the combination step, each node can determine the distance between its own estimation and its neighbors' shared estimations. Based on these dis- tances, it can assign a proper weight to each one. Hence it can eliminate detected malfunctioning nodes if it seems neces- sary. It can also put some sharing restrictions on the detected impaired nodes to isolate them from following the righteous agents. It can start dissembling and after being trusted, do its job (just as an example of smart simple hacking). The pro- posed algorithm is shown as in Alg.1. Based on the main tar- for t = 1, . . . , T do for i ∈ {1, . . . , N} do Adaptation: xi(t) = g(xi(t − 1), θi(t)) Combination: xi(t) = h(xj(t)j ∈ Ni) end end Algorithm 1: The Algorithm of Malfunctioning Detection: g(.) and h(.) are respectively adaption and combination op- erators; T and N are the number of iterations and the total number of nodes; xi(t), xi(t), θi(t) and Ni are the estimate of the ith node, its temporary version, its insight from the sharing environment at time t and the set of its Neighbors. get algorithm, the distance calculating function, the weighting method, and the eliminating strategy can be properly defined. 3. DISTRIBUTED ESTIMATION In this section, we illustrate the capability of our proposed al- gorithm for detecting noisy estimator in a distributed estima- tion method. We assume well-known ATC strategy[18]. The main algorithm, with the respect to Alg.1, can be summarized as follows[18] (cid:40) di(t) ← aT i,txopt + νi(t) ei(t) ← di(t) − aT i,txi(t − 1) xi(t) = xi(t − 1) − µi,tei(t)ai,t g(xi(t − 1), θi(t)) = (cid:40) h(xj(t)j ∈ Ni) = αi,j(t) ← ( ∀j ∈ Ni : ci,j(t) ← xi(t) =(cid:80) (cid:107)xi(t)−xj (t)(cid:107)2 )e ζ exp(αi,j (t)) (cid:80) cl,iexp(αi,k(t)) k∈Ni ci,lxl(t) l∈Ni where for the ith node at time instance t. It is an implementa- ai,t νi(t) µi,t ζ measurement vector measurement noise adaptation rate normalization coefficient e determining decision type description variance: σ2 i step-size related to σi uniform: e = 0 hard: e → ∞ soft: e ∈ (0,∞) tion of diffusion LMS algorithm which has given an outstand- ing performance[18]. It was mentioned that in each iteration, we can have an adaptation step and a combination step. How- ever, we should be cautious, since with low value of adapta- tion rate, possibly all of the estimators in each neighborhood have near estimations. Hence, simply feeding the distance calculator with their values can mislead the total procedure. In order to simulate the proposed algorithm, we assume a graph having N nodes with s as an existence probability of a link between each pair of nodes (ρ). The target signal has a length of L and is generated with normal distribution (with given sparsity of nonzero components: s). In fact, we just extend the adaptation time (Ta). In order to resolve this is- sue, each node starts its own simple adaptation step multiple times. After an agreed adaptation time slot, sharing method is run. As shown in Fig.1, the heavily-noisy estimator is dis- lodged from the combination step. In the both cases of dense and sparse graph, improvement is clear. Since the noisy one still has access to the intact estimations, it can follow the es- timation's trend, as the fluctuations can be seen in the Mean Square Deviation (MSD) curves. By tuning the weighting pa- rameters (ζ and m), it can be prohibited. 4. MULTI-AGENT RL In reinforcement learning scenario we will rely on some other mathematical foundations to be helped with, which in the fol- lowing paragraphs it will be explored in more details. In this (a) ρ = 0.1, σnoise = 0.04, N = 50, ζ = 0.015, e = 2. (b) Inact nodes connected to an im- paired node with uniform weights, ρ = 1, σnoise = 0.04, N = 10, e = 0. (c) ρ = 0.2, σnoise = 0.4, N = 100, ζ = 0.05, e = 8. (d) ρ = 1, σnoise = 0.04, N = 10, ζ = 0.015, e = 8. Fig. 1: Distributed Estimation Performance (ATC scenario), MSD vs #iteration; L = 100, s = 0.5, µ = 0.001, Ta = 10, σimpaired = 10m × σnoise, #simulation = 1000. case, exploring the search space while exploiting on the best strategy which has been found so far, will give the so-called agent the ability to adapt to its environment [19]. We will consider the classic RL framework in which an agent is in- teracting with a finite state Markovian environment on a dis- crete time scale [20]. Our final goal will be maximizing the expected of the utility function Q over all possible policies, considering different R(s, a) rewards for each pair of state- actions [21]. Bellman's equation describes the following in terms of both, the expected reward and expected outcome of the random transition[20]. The notation of π and γ allude respectively to the certain policy and the discount factor[19]. Based on the optimality equation(1), several algorithms have been proposed[20], which in this setting we are going to focus on one of the well-known methods, Q-Learning [19, 20], as shown by 2 = E[R(s, a)] + γEP [ max ∗ (cid:48) Q (s Qnew(st, at) ← (1 − α)Q(st, at) . . . (cid:48) (cid:48) t+1, a + α(R(st, at) + γ max ∗ {a(cid:48)∈A} Q {a(cid:48)∈A} Q(S )) (cid:48) )] , a (1) (2) In the case of finite state-action pairs we use the tool of Q- table for our convenience. Q-table is a simple representation of the Q-values in each pair of state-action. Since, in the be- ginning of the process our initialized Q-table acts randomly, our total regret [22] diverges linearly from its optimal state. Thus, to overcome this issue, another modification is to use an -decay exploration rate [22]. In this case, our initial ex- ploration rate will be reduced through the process of learning. = min + (max − min) × e−#step×DecayRate In a multi-agent setting, we rely on multiple agents for com- pleting a given task. All of the above definitions could be extended for larger number of agents. There is just an extra point that here agents will interact with one another. Con- sidering the case of fully-connected network, all agents will participate for filling the Q-table. Passing the Q-values, will give them an access to their local neighborhoods. As a result, we expect to converge faster than before. In a simple case, each agent would fill its own Q-table with averaging on all observations they have been given. Finally, after a large num- ber of episodes we may expect to see that their Q-tables have converged to a single Q-table, which could be used for all of them afterwards. 4.1. Problem and Model definition: Given the uncertainties of the real-world problems we should always be prepared for some obstacles. For instance, in the case of multi-agent learning, one could mention the broken state of an agent as a critical situation. Since they are all transferring information and using those information for their actions, a major turbulence could be misleading. So, we can think of some adjustments applied to the prior methods. Here we are going to detect those broken agents and rooting them out from our decision-making, by a sort of swarm-intelligence algorithm. In this way of analysis, agents themselves will decide on that which agents are broken. Consider an arbi- trary agent i and its neighborhood as Ni, in the case of fully- connected network. Then for each j ∈ Ni, in addition to what their given observation is, we would also get their fakeness in- dices as well. Finally, agent i will decide based on its given information that which agent(s) is (are) malfunctioning. Here, we just consider the simple case of one broken agent. How- ever, the results could be generalized to more complex situa- tions. Furthermore, we would also assume that our algorithm will converge which, results in finding the correct target. (cid:40) f 0 i = fj (cid:44) Fakeness labels from nearby agents k = arg max{j∈Ni}(Qj − Qi) ; Qj − Qi > 0 N one ; o.w Then we do the majority vote. So, first for each agent in the system s, we count its appearance as a broken (fake) agent. Finally, we decide whether our maximum appearance should be reported, based on comparing its votes against the total number of votes. ∀n ∈ S; ni = 1(n=f 0 i ) + 1{n=fj} (cid:88) j∈Ni (cid:40) fi = k = arg max{n∈S}(ni) ; ki > N one ; o.w (cid:80) j∈Ni 2 1 Again, due to the randomness being involved in the beginning of the learning, we do not adjust our weights right after a sin- gle step. Instead, we wait for 10 steps as an example and then decide based on those 10 steps that which agents were behav- ing badly. This would help us to avoid random decisions and to reduce some weights unexpectedly. In order to average our observations we would also normalize them by reducing the weight of the broken agent linearly by the factor of 0 < λ < 1 It should be kept in mind that we consider the agent itself as one of its neighbors as well. g(Qm−1 i , θi(t)) = r ← R(st, at) + γ max{a(cid:48)∈A} Qm−1 i (st, at) = (1 − α)Qm−1 Qm i i (st, at) + αr (st+1, a(cid:48)) (cid:40) (cid:40) h(Qm i (st, at)j ∈ Ni) = fi ← [f akeness voting] 1 ; if: fi ∈ Ni ci,fi (m) ← 1−λ(cid:80) −1+(cid:80) j∈Ni (cid:80) i (st, at) =(cid:80) j∈Ni j∈Ni l∈Ni ci,j (m) ← 1+ Qm λ 1 1 ; ∀fi ∈ Ni \ fi j (st, at) ci,l(m) Qm For comparing our results, we have used the OpenAI gym package in python 3.7.2. The SnowFlake environment, with 8 × 8 grid is our main basis. We have also manipulated the environment for putting different agents in different starting points. As it could be seen from the table below, our im- paired diagnosis algorithm outperforms the classical approach in which we ignore the impaired agents. It should be noted that here we have the simple case of one broken agent, which is detected correctly. We also consider a task to be done when- ever there is an agent which has made it to the target point. Parameter Number of episodes Maximum number of steps Learning Rate Discount Factor Minimum Exploration Rate Maximum Exploration Rate Exploration Decay Rate Number of Agents λ success rate (with detection) success rate (without detection) case 1 1e6 1000 0.8 0.97 0.001 1 0.001 3 0.7 %84 %53 case 2 1e5 10000 0.8 0.97 0.001 1 0.001 3 0.9 %78 %59 case 3 1e6 1000 0.7 0.97 0.001 1 0.001 3 0.7 %71 %55 5. CONCLUSION By emphasizing on the role of non-stationarity in an interact- ing network, we introduced a novel approach for analyzing it. The main focus of our analysis was also on the problem of error diagnosis and canceling certain interferences caused by damaged agents in network. We also applied our proposed algorithm to two different frameworks with the same issue of an impaired node. In both cases of distributed estimation and multi-agent RL, by detecting and reducing the effect of the impaired node, we surpassed our prior performance. 6. REFERENCES [1] Federico S Cattivelli and Ali H Sayed, "Diffusion lms strategies for distributed estimation," IEEE Transactions on Signal Processing, vol. 58, no. 3, pp. 1035 -- 1048, 2009. [2] Hadi Zayyani, Mehdi Korki, and Farrokh Marvasti, "A distributed 1-bit compressed sensing algorithm robust to impulsive noise," IEEE Communications Letters, vol. 20, no. 6, pp. 1132 -- 1135, 2016. [3] Sijia Liu, Sundeep Prabhakar Chepuri, Makan Fardad, Engin Mas¸azade, Geert Leus, and Pramod K Varshney, "Sensor selection for estimation with correlated mea- surement noise," IEEE Transactions on Signal Process- ing, vol. 64, no. 13, pp. 3509 -- 3522, 2016. [4] Xiaohua Ge, Fuwen Yang, and Qing-Long Han, "Dis- tributed networked control systems: A brief overview," Information Sciences, vol. 380, pp. 117 -- 131, 2017. [5] Dan Zhang, Sing Kiong Nguang, and Li Yu, "Dis- tributed control of large-scale networked control sys- tems with communication constraints and topology switching," IEEE Transactions on Systems, Man, and Cybernetics: Systems, vol. 47, no. 7, pp. 1746 -- 1757, 2017. [6] Aboul Ella Hassanien and Eid Emary, Swarm intelli- gence: principles, advances, and applications, CRC Press, 2018. Proceedings of the IEEE, vol. 106, no. 5, pp. 808 -- 828, 2018. [12] Nathanael Perraudin and Pierre Vandergheynst, "Sta- tionary signal processing on graphs," IEEE Transactions on Signal Processing, vol. 65, no. 13, pp. 3462 -- 3477, 2017. [13] Jie Chen, C´edric Richard, and Ali H Sayed, "Diffusion lms over multitask networks," IEEE Transactions on Signal Processing, vol. 63, no. 11, pp. 2733 -- 2748, 2015. [14] Reza Abdolee and Benoit Champagne, "Diffusion lms strategies in sensor networks with noisy input data," IEEE/ACM Transactions on Networking (TON), vol. 24, no. 1, pp. 3 -- 14, 2016. [15] Elaheh Mohammadi, Amin Gohari, and Farokh Mar- vasti, "A square root sampling law for signal recovery," IEEE Signal Processing Letters, vol. 26, no. 4, pp. 562 -- 566, 2019. [16] Flavio Zabini, Alex Calisti, Davide Dardari, and Andrea Conti, "Random sampling via sensor networks: Estima- tion accuracy vs. energy consumption," in 2016 24th European Signal Processing Conference (EUSIPCO). IEEE, 2016, pp. 130 -- 134. [17] Firdaus E Udwadia, "Methodology for optimum sensor locations for parameter identification in dynamic sys- tems," Journal of Engineering Mechanics, vol. 120, no. 2, pp. 368 -- 390, 1994. [7] Christian Blum and Gross Roderich, "Swarm intelli- gence in optimization and robotics," in Springer hand- book of computational intelligence, pp. 1291 -- 1309. Springer, 2015. [18] Sheng-Yuan Tu and Ali H Sayed, "Diffusion strategies outperform consensus strategies for distributed estima- tion over adaptive networks," IEEE Transactions on Sig- nal Processing, vol. 60, no. 12, pp. 6217 -- 6234, 2012. [8] Jakob N Foerster, Gregory Farquhar, Triantafyllos Afouras, Nantas Nardelli, and Shimon Whiteson, "Counterfactual multi-agent policy gradients," in Thirty-Second AAAI Conference on Artificial Intelli- gence, 2018. [9] Ryan Lowe, Yi Wu, Aviv Tamar, Jean Harb, Ope- nAI Pieter Abbeel, and Igor Mordatch, "Multi-agent actor-critic for mixed cooperative-competitive environ- ments," in Advances in Neural Information Processing Systems, 2017, pp. 6379 -- 6390. [10] Siheng Chen, Rohan Varma, Aliaksei Sandryhaila, and Jelena Kovacevi´c, "Discrete signal processing on graphs: Sampling theory," IEEE transactions on signal processing, vol. 63, no. 24, pp. 6510 -- 6523, 2015. [11] Antonio Ortega, Pascal Frossard, Jelena Kovacevi´c, Jos´e MF Moura, and Pierre Vandergheynst, "Graph sig- nal processing: Overview, challenges, and applications," [19] Richard S Sutton and Andrew G Barto, Reinforcement learning: An introduction, MIT press, 2018. [20] Marc G Bellemare, Will Dabney, and R´emi Munos, "A distributional perspective on reinforcement learning," in Proceedings of the 34th International Conference on Machine Learning-Volume 70. JMLR. org, 2017, pp. 449 -- 458. [21] Susanne Still and Doina Precup, "An information- theoretic approach to curiosity-driven reinforcement learning," Theory in Biosciences, vol. 131, no. 3, pp. 139 -- 148, 2012. [22] Ian Osband, Benjamin Van Roy, Daniel Russo, and Zheng Wen, "Deep exploration via randomized value functions," arXiv preprint, arXiv:1703.07608, 2017.
1310.3875	1	1310	2013-10-14T22:50:53	Cucker-Smale flocking with alternating leaders	[ "cs.MA" ]	We study the emergent flocking behavior in a group of Cucker-Smale flocking agents under rooted leadership with alternating leaders. It is well known that the network topology regulates the emergent behaviors of flocks. All existing results on the Cucker-Smale model with leader-follower topologies assume a fixed leader during temporal evolution process. The rooted leadership is the most general topology taking a leadership. Motivated by collective behaviors observed in the flocks of birds, swarming fishes and potential engineering applications, we consider the rooted leadership with alternating leaders; that is, at each time slice there is a leader but it can be switched among the agents from time to time. We will provide several sufficient conditions leading to the asymptotic flocking among the Cucker-Smale agents under rooted leadership with alternating leaders.	cs.MA	cs	Cucker-Smale flocking with alternating leaders Zhuchun Li and Seung-Yeal Ha 1 Abstract We study the emergent flocking behavior in a group of Cucker-Smale flocking agents under rooted leadership with alternating leaders. It is well known that the network topology regulates the emergent behaviors of flocks. All existing results on the Cucker-Smale model with leader-follower topologies assume a fixed leader during temporal evolution process. The rooted leadership is the most general topology taking a leadership. Motivated by collective behaviors observed in the flocks of birds, swarming fishes and potential engineering applications, we consider the rooted leadership with alternating leaders; that is, at each time slice there is a leader but it can be switched among the agents from time to time. We will provide several sufficient conditions leading to the asymptotic flocking among the Cucker-Smale agents under rooted leadership with alternating leaders. Cucker-Smale agent, alternating leaders, rooted leadership, flocking Index Terms I. INTRODUCTION The purpose of this paper is to study the emergent flocking phenomenon to the generalized Cucker- Smale (C-S) model with alternating leaders. Roughly speaking, the terminology "flocking" represents the phenomena that autonomous agents, using only limited environmental information, organize into an ordered motion, e.g., flocking of birds, herds of cattle, etc. These collective motions have gained increasing interest from the research communities in biology, ecology, sociology and engineering [1], [13], [20], [23], [27], [29], [30] due to their various applications in sensor networks, formation of robots and spacecrafts, financial markets and opinion formation in social networks. In [11], [12], Cucker and Smale proposed a nonlinear second-order model to study the emergent behavior of flocks. Let xi, vi ∈ Rd be the position and velocity of the i-th C-S agent, and ψij ≥ 0 be interaction weight between j and i-th agents. Then, the discrete-time C-S model reads as xi(t + 1) = xi(t) + hvi(t), i = 1, 2, . . . , N, N vi(t + 1) = vi(t) + h ψij(x(t)) [vj(t) − vi(t)] , (1) Xj=1 1 where h is a time-step. For (1), the "asymptotic flocking" means ψij(x(t)) = (1 + xi(t) − xj(t)2)β , β ≥ 0, sup t∈N xi(t) − xj(t) < ∞, lim t→∞ (vi(t) − vj(t)) = 0, ∀ i 6= j. The study of flocking behavior of multi-agent system based on mathematical models dates back to [20], [30], even before Cucker and Smale. However, the significance of C-S model lies on the solvability of the model and phase-transition like behavior from the unconditional flocking to conditional flocking, as 2. the decay exponent β increases from zero to some number larger than 1 The particular choice of weight function ψij in (1) is a crucial ingredient which makes this model attractive. We note that in the original C-S model (1), the agents are interacting under the all-to-all Z. Li is with the Department of Mathematics, Harbin Institute of Technology, Harbin 150001, China (e-mail: [email protected]). S.-Y. Ha is with the Department of Mathematical Sciences, Seoul National University, Seoul 151-747, Korea (e-mail: [email protected]). This work was carried out when Z. Li visited Seoul National University; he appreciated the support from PARC-SNU. 2 distance depending couplings ψij = ψji > 0 for all i 6= j. Later, Cucker-Smale's results were extended in several directions, e.g., stochastic noise effects [3], [10], [16], collision avoidance [2], [8], steering toward preferred directions [9], bonding forces [26], and mean-field limit [6], [18], [19]. In particular, an unexpected application was proposed by Perea, G´omez and Elosegui [27] who suggested to use the C-S flocking mechanism [11] in the formation of spacecrafts for the Darwin space mission. Recently, the C-S flocking mechanism was also applied to the modeling of emergent cultural classes in sociology and the stochastic volatility in financial markets [1], [15], [21]. In this paper, we consider the Cucker-Smale flocking under a switching of leadership topology with alternating leaders. It is well known that the interaction topology is an important component to understand the dynamics of multi-particle systems and vice versa. Biological complex systems are ubiquitous in our nature and indeed take various interaction topologies. The first work in relation with the C-S model other than all-to-all topology is due to J. Shen, who introduced the C-S model under hierarchical leadership [28]. A more general topology with leadership including hierarchical one was introduced by Li and Xue in [25], namely, the rooted leadership. Unfortunately, the analysis given in [25] cannot be applied to the continuous-time C-S model in a general setting. The continuous-time C-S model with a rooted leadership was studied in the framework of fast-slow dynamical systems in [17] for some restricted situation. Recently, a topology with joint rooted leadership was also considered in [24], in which a "joint" connectivity is imposed only along some time interval, instead of every time slices. Note that in previous works [25], [24] involving interaction topologies with leadership, the leader agent is assumed to be fixed in temporal evolution of flocks. This is not realistic. We can often observe that the leaders in migrating flocks can be changed during their migration. For example, as the large flocks of birds make a long journey from continent to continent, the leader birds located in the front of the flock endure larger resistance from the neighboring airs, e.g., wind. So leaders have to spend more energy than other followers. To save the energy of leader birds, leaders change alternatively. Of course, we can also find alternating leaders in our human social systems, for example, the periodic election of political leaders. Motivated by these situations, we study the asymptotic flocking behavior of the C-S model with alternating leaders. For the flocking analysis of the C-S model, most existing studies assume all-to-all and symmetric couplings so that the conservation of momentum is guaranteed, which is crucial in the energy estimates [11], [12], [18]. In contrast, when the interaction topology is not symmetric, there is no general systematic approach for flocking estimate. The induction method is applied to hierarchical leadership [7], [28], which relies on the triangularity of the adjacency matrix. Another useful tool, the self-bounding argument developed by Cucker-Smale in [11], [12], can be applied to different topologies; however, it requires a flocking estimate that relies on the topologies. For all-to-all coupling, the estimate is made on the matrix 2-norm through the spectrum of symmetric graph Laplacian. For rooted leadership, the authors in [25], [24] employed the (sp) matrices [31], [32] to study the infinity norm of a reduced Laplacian. Note that for all these special cases, the asymptotic velocity for flocking is a priori known, either the mean value of the initial state or just that of the leader. Thus, we can study the dynamics of newly defined variables, i.e., the fluctuations around the average velocity, or the states relative to the fixed leader, which can be bounded, to study the flocking behavior. However, in the case of alternating leaders, we do not have the accurate information on the asymptotic velocity of the flock and the dynamics of referenced variables (see (5)) cannot be given by nonnegative matrices as in [25], [24]. To overcome this difficulty, we consider the combined dynamics of the original system and the reference system. We employ the estimates in [4], [5] for the first-order consensus problem to find a priori estimate for the original system. From this, we can estimate the evolution of referenced velocity to support the self-bounding argument. The rest of this paper is organized as follows. In Section 2, we describe our model and present a consensus estimate that is useful in this work. In Section 3, we provide the flocking estimates for the discrete-time C-S model under rooted leadership with alternating leaders. In Section 4, we present some numerical simulations. Finally, Section 5 is devoted to the summary of this paper. Notation: Given x ∈ RN, we use the notations x∞ and x to denote the infinity norm (maximum norm) and 2-norm (Euclidean norm) of the vector respectively. For a N × N matrix A ∈ RN×N, we use 3 to denote the infinity norm, that is, the maximum absolute row sum of A, and for two N × N kAk∞ matrices A = (aij) and B = (bij), we use A◦ B to denote the element-wise product, i.e., A◦ B = (aijbij). II. PRELIMINARIES In this section we introduce the C-S flocking model under rooted leadership with alternating leaders. A useful estimate for the "flocking" matrix will be presented as well. A. Flocks with alternating leaders In this subsection, we present a brief description of the C-S flocking model under rooted leadership with alternating leaders. Consider a group of agents {1, 2, . . . , N}. For the description of interaction topology, we use graph the- ory [14] as follows. A digraph G = (V,E) (without self-loops) representing N particles with interactions, is defined by V := {1, 2, . . . , N}, E ⊆ V × V \ {(i, i) : i ∈ V}. We say (j, i) ∈ E if and only if j is a neighbor of i, i.e., j influences i. As an information flow chart, we may write j → i if and only if (j, i) ∈ E. A directed path from j to i (of length n + 1) comprises a sequence of distinct arcs of the form j → k1 → k2 → · · · → kn → i. On the other hand, once the directed neighbor graph G = (V,E) is chosen, the associated adjacency matrix, denoted by χ(G) = (χij), is given by χij =(cid:26) 1, if (j, i) ∈ E, if (j, i) /∈ E. Then, the C-S model on the digraph graph G is given by (1) with the second equation replaced by 0, vi(t + 1) = vi(t) + h χijψij(x(t)) [vj(t) − vi(t)] . N Xj=1 Xj=1 1 Thus, there is an interaction from j to i with strength ψij(x(t)) as long as it exists. In order to take the interaction weight ψij(x) into account, we refer to the matrix χ ◦ Ψx := (cid:0)χijψij(x(t))(cid:1) as the weighted adjacency matrix of the C-S system on the digraph G. Below, we use the symbol I to denote a finite set indexing all admissible digraphs Gp = (Vp,Ep) and let σ : N → I be a switching signal. At each time point t ∈ N, the system is registered on an admissible graph Gσ(t), and thus has a weighted adjacency matrix given by χσ(t) ◦ Ψx. In this setting, we write the system as the C-S system undergoing switching of the neighbor graphs with a switching signal σ: xi(t + 1) = xi(t) + hvi(t), i = 0, 1, . . . , N, N vi(t + 1) = vi(t) + h χσ(t) ij ψij(x(t)) [vj(t) − vi(t)] , (2) ψij(x(t)) = (1 + xi(t) − xj(t)2)β . We now introduce the definition of C-S model under rooted leadership with alternating leaders. For this we first present the rooted leadership [25]. Definition 2.1: 1) The system (2) is under rooted leadership at time t, if for the digraph Gσ(t), there exists a unique vertex, say rt ∈ V, such that the vertex rt does not have an incoming path from others, but any other vertex in V has a directed path from rt. The vertex rt represents the leader in the flock. 2) The system (2) is under rooted leadership with alternating leaders, if the system (2) is under rooted leadership at each time slice, but the leader rt is not fixed for all time. Note that the leader can be changed from time to time; thus, the asymptotic velocity is not a priori known, even if the flocking can be achieved. 4 B. Consensus estimates In this subsection, we present a convergence estimate given in [4] for the first-order consensus problem. Given a sequence of stochastic matrices (also known as Markov matrices [22]) F1, F2,· · · ∈ RN×N, for consensus we expect that the product Ft · · · F2F1 converges to a rank one matrix, i.e., has the same row vectors. For a single stochastic matrix F , under some connectivity condition of its associated graph, the matrix iteration F k converges to the rank one matrix 1π with π being the left-eigenvector of F , i.e., πF = π. To deal with the case of time-dependent state transition matrices, we introduce some notations following [4]. Let F be a stochastic matrix and we denote by ⌊F⌋ the row vector whose jth element is the smallest element of the jth column of S. Let [F ] = F − 1⌊F⌋, then we have k[F ]k∞ = 1 − ⌊F⌋1, where 1 = (1, 1, . . . , 1)T. In some sense, [F ] measures how much the matrix F is different with a rank one matrix. It is obvious that a product of stochastic matrices must be a stochastic matrix. For an infinite sequence of stochastic matrices F1, F2, . . . , the limit ⌊· · · Ft · · · F2F1⌋ := lim t→∞⌊Ft · · · F2F1⌋ always exists [4], even if the product Ft · · · F2F1 itself does not have a limit. In order to form a consensus, we expect the product Ft . . . F2F1 to converge to a rank one stochastic matrix, i.e., a matrix of the form 1c. If this is true, then the limit must be 1⌊· · · Ft · · · F2F1⌋. In the following, we will say that the matrix product Ft . . . F2F1 converges to 1⌊· · · Ft · · · F2F1⌋ exponentially fast at a rate no slower than λ if there exist nonnegative constants b and λ < 1 such that (3) We write A ≥ B if A − B is a nonnegative matrix. The following result gives a sufficient condition to the exponential convergence. (cid:13)(cid:13)Ft · · · F2F1 − 1⌊· · · Ft · · · F2F1⌋(cid:13)(cid:13)∞ ≤ bλt, t ≥ 1. Proposition 2.1: [4] (1) For any pair of stochastic matrices F1 and F2, we have (2) Let b and λ < 1 be nonnegative constants. Suppose that F1, F2, . . . is an infinity sequence of stochastic matrices with [F2F1] ≤ [F2][F1]; k[Ft · · · F2F1]k∞ ≤ bλt, t ≥ 0. Then, the matrix product Ft . . . F2F1 converges to 1⌊· · · Ft · · · F2F1⌋ exponentially fast at a rate no slower than λ. Therefore, if each of the matrices in the sequence F1, F2, . . . satisfies k[Fi]k∞ ≤ λ, then Ft . . . F2F1 does converge to 1⌊· · · Ft · · · F2F1⌋ exponentially. Since k[F ]k∞ = 1 − ⌊F⌋1, k[F ]k∞ < 1 if and only if the matrix ⌊F⌋ has at least one positive element, i.e., the matrix F has at least one nonzero column. For a stochastic matrix F = (Fij) ∈ RN×N, we define the associated digraph as G = (V,E) with V = {1, 2, . . . , N} and E = {(j, i) : Fij > 0}. We call a graph G is a strongly rooted graph if there exists some vertex j such that (j, i) ∈ E for all i 6= j. For such a j, we say that it is a strong root of the graph and the graph is strongly rooted at j. In what follows, we also say G is a rooted graph if there exists a vertex which has a path to any other agent; such a vertex is called the root of the graph. Then we arrive at the following result. Lemma 2.1: [4] The digraph associated to a stochastic matrix F is strongly rooted if and only if k[F ]k∞ < 1. We next introduce the composition of digraphs. By the composition of digraphs Gp with Gq, denoted by Gq ◦ Gp, we mean the digraph with the vertex set V and arc set defined in such a way so that (i, j) is an arc of the composition just in case there is a vertex k such that (i, k) is an arc of Gp and (k, j) is an arc of Gq. Denote their associated flocking matrices by Fp and Fq respectively. Then, we see that the flocking matrix associated to the digraph Gq ◦ Gp is exactly the matrix product FqFp. Proposition 2.2: [4] Let Gσ(1),Gσ(2), . . . , be a sequence of rooted graph, then for any t1 ∈ N, the graph Gσ(t1+(N−1)2) ◦ · · · ◦ Gσ(t1+2) ◦ Gσ(t1+1) is a strongly rooted graph. Based on this result, we can obtain a strongly rooted graph from the composition of rooted leadership with alternating leaders. In this section, we introduce the C-S flocking matrix and a reference system to study the C-S model with alternating leaders. III. FLOCKING ANALYSIS A. A reference system and the flocking matrix 5 We consider a group of particles {1, 2, . . . , N} whose dynamics is governed by (2). Let x = (x1, x2, . . . , xN )⊤ and v = (v1, v2, . . . , vN )⊤ ∈ (R3)N be the position and velocity vector of the flock, respectively. In order to simplify the notations, for a given solution {(x(t), v(t))} to system (2) under a switching signal σ, we write N ψij(t) := ψij(x(t)), di(t) := χσ(t) ij ψij(t). (4) Xj=1,j6=i In order to use a self-bounding argument, we introduce a reference system for the N-flocks. We use the last agent as the reference and set x := (x1, . . . , xN−1)⊤ = (x1 − xN , . . . , xN−1 − xN )⊤, v := (v1, . . . , vN−1)⊤ = (v1 − vN , . . . , vN−1 − vN )⊤. (5) It is obvious that the asymptotic flocking behavior is equivalent to the boundedness of x together with the zero convergence of v. If we set xN = 0 and vN = 0, then xi − xj2 = xi − xj2 ≤ 2(xi2 + xj2) ≤ 2x2, This means that the C-S communication weights satisfy 1 ≤ i, j ≤ N. 1 ψij(t) = (1 + xi(t) − xj(t)2)β ≥ Note that the dynamics of x(t) and v(t) are governed by 1 (1 + 2x(t)2)β . x(t + 1) = x(t) + hv(t), where the matrix Pσ(t) ∈ R(N−1)×(N−1) is given by v(t + 1) = Pσ(t)v(t), (Pσ(t))ij =( 1 − hdi − hχσ(t) hχσ(t) N i ψN i(t), ij ψij(t) − hχσ(t) N j ψN j(t), i = j, i 6= j, (6) (7) where i, j = 1, 2, . . . , N − 1. If the flock has a fixed leader, say N, then χσ(t) N j ≡ 0 for all j 6= N, and thus the matrix Pσ(t) is a nonnegative matrix provided a sufficiently small h. However, if the leader agent changes from time to time, we cannot expect the matrix Pσ(t) to be always a nonnegative matrix. Therefore, the approach in [25], [24] cannot be applied in this case. To carry out a flocking estimate, we will not use the explicit difference equation of v(t), instead we derive a direct estimate of v(t) through v(t). This is the key idea in this study apart from the previous works. To estimate v(t), we derive a compact form from the system (2) as follows: x(t + 1) = x(t) + hv(t), v(t + 1) = (Id − hLσ(t))v(t) =: Fσ(t)v(t), (8) where Lσ(t) ∈ RN×N is the weighted Laplacian of digraph Gσ(t), that is, · · · −χσ(t) 1N ψ1N (t) · · · −χσ(t) 2N ψ2N (t) ... · · · dN (t) · · · d1(t) 12 ψ12(t) −χσ(t) 21 ψ21(t) d2(t) · · · · · · −χσ(t) N 1 ψN 1(t) −χσ(t) N 2 ψN 2(t) Lσ(t) =  −χσ(t)   The matrix Fσ(t) := Id − hLσ(t) is called the (C-S) flocking matrix at time t associated to the neighbor graph Gσ(t). If we choose a small h > 0 such that all diagonal entries of Fσ(t) are nonnegative, then Fσ(t) is a stochastic matrix, i.e., a nonnegative matrix with each row sum being 1. 6 . B. Basic estimates In this subsection we present an estimate of v(t) through v(t). We first concentrate on the flocking matrix Fσ(t) that determines the dynamics of v(t). To do this, we will employ the estimates in [4] (see Subsection II-B). Note that for flocking under rooted leadership (see Definition 2.1), the neighbor graph is a rooted graph with the leader agent acting as the root. Thus, Proposition 2.2 and Lemma 2.1 imply that the product of (N − 1)2 flocking matrices must satisfy (cid:13)(cid:13)[Fσ(t1+(N−1)2−1) · · · Fσ(t1+1)Fσ(t1)](cid:13)(cid:13)∞ < 1. Inspired by this fact, we will present an estimate for the product of flocking matrices which can be applied to the analysis of the second-order C-S model. We first estimate the convergence of Fσ(t) · · · Fσ(1)Fσ(0) by Proposition 2.1. Proposition 3.1: Suppose that {(x(t), v(t))}t∈N is a solution of C-S model (2) with alternating leaders. (9) Assume and Then, we have x(t) ≤ B, t ∈ N, h < 1 N + 1 . (N −1)2(cid:3) , (cid:13)(cid:13)Fσ(t) · · · Fσ(1)Fσ(0) − 1⌊· · · Fσ(t) · · · Fσ(1)Fσ(0)⌋(cid:13)(cid:13)∞ ≤(cid:16)1 − (hR)(N−1)2(cid:17)(cid:2) t+1 (1+2B2)β . 1 where R = Proof: According to (6) and (9), for any t ≥ 0, we have 1 (1 + 2B2)β ≤ ψij(t) ≤ 1, 1 ≤ i, j ≤ N. Moreover, by the assumption (10), we have (Fσ(t))ii = 1 − hdi(t) ≥ 1 N + 1 > h (1 + 2B2)β , 1 ≤ i ≤ N. From (12) and (13) we see that under the assumptions (9) and (10), Fσ(t) is a stochastic matrix with nonzero entries not less than hR := (1+2B2)β . Consequently, all the nonzero elements of the matrix product Fσ(t) · · · Fσ(1)Fσ(0) must be not less than (hR)t+1. Note that if the system (2) is under rooted leadership at time t, then the neighbor graph Gσ(t) is a rooted graph with the leader acting as the root. We recall Proposition 2.2 to see that the composition of neighbor graph along a time interval of length h (10) (11) (12) (13) (N − 1)2 must be a strongly rooted graph. By Lemma 2.1, this means that any (N − 1)2-product of flocking matrices satisfies 7 or equivalently, (cid:13)(cid:13)[Fσ(t1+(N−1)2−1) · · · Fσ(t1+1)Fσ(t1)](cid:13)(cid:13)∞ 1 − ⌊Fσ(t1+(N−1)2−1) · · · Fσ(t1+1)Fσ(t1)⌋1 < 1. < 1, Thus, the matrix Fσ(t1+(N−1)2−1) · · · Fσ(t1+1)Fσ(t1) has at least one nonzero column. Because all of the nonzero elements of Fσ(t1+(N−1)2−1) · · · Fσ(t1+1)Fσ(t1) are not less than (hR)(N−1)2, we find that i.e., This implies that for all t ∈ N, 1 − ⌊Fσ(t1+(N−1)2−1) · · · Fσ(t1+1)Fσ(t1)⌋1 ≤ 1 − (hR)(N−1)2 (cid:13)(cid:13)[Fσ(t1+(N−1)2−1) · · · Fσ(t1+1)Fσ(t1)](cid:13)(cid:13)∞ ≤ 1 − (hR)(N−1)2 kFσ(t) · · · Fσ(1)Fσ(0)k∞ ≤(cid:16)1 − (hR)(N−1)2(cid:17)(cid:2) t+1 (N −1)2(cid:3) . . , (14) (15) We now combine (15) and Proposition 2.1 to obtain (11). Next, we use Proposition 3.1 to estimate the evolution of v(t). Proposition 3.2: Suppose {(x(t), v(t))}t∈N is a solution of C-S model (2) with alternating leaders. If (9) and (10) hold, then for the referenced velocity v(t) we have Proof: For the simplicity of notation, we denote the asymptotic velocity alignment state for the v(t)∞ ≤ 2(cid:16)1 − (hR)(N−1)2(cid:17)(cid:2) t+1 (N −1)2(cid:3) v(0)∞. (16) solution {(x(t), v(t))} as Then, it follows from (11) that we have v∞ = 1⌊· · · Fσ(t) · · · Fσ(1)Fσ(0)⌋v(0). v(t) − v∞∞ =(cid:12)(cid:12)Fσ(t) · · · Fσ(1)Fσ(0)v(0) − 1⌊· · · Fσ(t) · · · Fσ(1)Fσ(0)⌋v(0)(cid:12)(cid:12)∞ (N −1)2(cid:3) v(0)∞. Due to the definition of referenced variables (x(t), v(t)), we easily find ≤(cid:16)1 − (hR)(N−1)2(cid:17)(cid:2) t+1 v(t)∞ ≤(cid:12)(cid:12)(cid:0)v(t) − v∞(cid:1) −(cid:0)vN (t) − v∞(cid:1)(cid:12)(cid:12)∞ (N −1)2(cid:3) v(0)∞. Here, we use vN (t) itself to denote the N-duplication of vN (t) ∈ R3. ≤ 2(cid:16)1 − (hR)(N−1)2(cid:17)(cid:2) t+1 C. Flocking behavior 8 To carry out the flocking analysis, we use the spectral norm (or 2-norm) of vectors and matrices. However, the previous estimates for v(t) is given by the infinity norm. Due to the equivalence of norms in finite-dimensional space, there exists a constant λ ≥ 1, such that for all v ∈ (R3)N−1, (17) where · denotes the 2-norm of vector. Before we present the main result, we quote an elementary lemma from [11] without the proof. Consider the following algebraic equation: v∞ ≤ v ≤ λv∞, F (z) := zr − c1zs − c2 = 0. (18) Lemma 3.1: [11] Suppose that the coefficients and exponents in F satisfy c1, c2 > 0 and r > s > 0. Then the equation (18) has a unique positive zero z∗ satisfying z∗ ≤ max{(2c1) 1 r−s , (2c2) 1 r}, and F (z) ≤ 0 for 0 ≤ z ≤ z∗. We are now ready to present our main result on the Cucker-Smale flocking with alternating leaders. Theorem 3.1: Let the discrete-time Cucker-Smale model (2) be under rooted leadership with alternating leaders. Assume that the time step h fulfills (10), and one of the following three hypothesis holds: 1) s < 1; 2) s = 1 and 3) s > 1, and Here, v(0) < h(N−1)2−1 2√2(N − 1)2λ ; 1 s−1 (cid:18) 1 a(cid:19) 1 s−1 "(cid:18)1 s(cid:19) −(cid:18) 1 s(cid:19) s s−1# − b > 1 s−1 8λ2v(0)2s N 2 1 s−1 + a 4√2λv(0) N . a = 2√2λh−(N−1)2+1(N − 1)2v(0), b = 1 + √2x(0), s = 2β(N − 1)2. Then the system (2) or (8) has a time-asymptotic flocking: 1) there exists a constant B0 > 0 such that x(t) ≤ B0, ∀ t ∈ N; 2) v(t) exponentially converges to zero as t → ∞. Proof: Fix a discrete-time mark T ∈ N and define Moreover, there exists x∞ ∈ (R3)N−1 such that x(t) −→ x∞ as t → ∞. x∗ = max 0≤t≤T x(t), T∗ ∈ argmax0≤t≤Tx(t). (19) (20) (21) Then by Proposition 3.2, the estimate (16) holds for newly defined R := (1 + 2x∗2)−β, as long as we restrict t within [0, T ]. That is, By (17) we find v(t)∞ ≤ 2(cid:16)1 − (hR)(N−1)2(cid:17)(cid:2) t+1 (N −1)2(cid:3) v(0)∞, t ∈ [0, T ]. v(t) ≤ λv(t)∞ ≤ 2λ(cid:16)1 − (hR)(N−1)2(cid:17)(cid:2) t+1 ≤ 2λ(cid:16)1 − (hR)(N−1)2(cid:17)(cid:2) t+1 (N −1)2(cid:3) v(0)∞ (N −1)2(cid:3) v(0), t ∈ [0, T ]. 9 By the dynamics of referenced position x(t), i.e., (7), for t ∈ [0, T ], we have t−1 t−1 Pτ =0v(τ ) 2λ(cid:0)1 − (hR)(N−1)2(cid:1)(cid:2) τ +1 Pτ =0 x(t) ≤ x(0) + h ≤ x(0) + h ≤ x(0) + 2hλv(0) ≤ x(0) + 2hλv(0)(N − 1)2 ∞ ≤ x(0) + 2hλv(0)(N − 1)2(hR)−(N−1)2 = x(0) + (N −1)2(cid:3)v(0) (N −1)2(cid:3) Pτ =0(cid:0)1 − (hR)(N−1)2(cid:1)(cid:2) τ +1 Pτ =0(cid:16)1 − (hR)(N−1)2(cid:17)τ √2 2 a(1 + 2x∗2)β(N−1)2. ∞ In particular, for t = T∗, we have x∗ ≤ x(0) + √2 2 a(1 + 2x∗2)β(N−1)2 . Let Z := (1 + 2x∗2) 1 2 , then the above relation and (20) give Z ≤ 1 + √2x∗ ≤ aZ 2β(N−1)2 In order to apply Lemma 3.1 we define F (z) as follows: + b. (22) F (z) = z − azs − b, 1−s , 2b}, and then x∗ ≤ (cid:0) U0 1 (1) Assume s < 1. The relation (22) says that F (Z) ≤ 0. By appealing to Lemma 3.1 we have Z ≤ U0 2 . Note that U0 depends only on a and b, which with U0 ≤ max{(2a) are independent of the pre-assigned time T . Therefore, the bound x(T ) ≤ x(T∗) ≤ B0 :=(cid:0) U0 2 is uniform for all T = 0, 1, 2, . . . . That is, x(t) ≤ B0 for all t ∈ N. that v(t) exponentially converges to 0. Finally, for all t2 > t1, we have 0)−β and use Proposition 3.2 for any time t ∈ N to find Now, we choose B := B0, and R := (1 + 2B2 2−1 2 2−1 2 (cid:1) (cid:1) 1 1 s = 2β(N − 1)2. x(t2) − x(t1) ≤ t2−1 t2−1 t2−1 Pτ =t1 x(τ + 1) − x(τ ) ≤ h ≤ 2hλv(0) Pτ =t1 v(τ ) Pτ =t1(cid:16)1 − (hR)(N−1)2(cid:17)(cid:2) τ +1 (N −1)2(cid:3) Pτ =t1(cid:16)1 − (hR)(N−1)2(cid:17)(cid:2) τ +1 (N −1)2(cid:3) ∞ ≤ 2hλv(0) ≤ 2h1−(N−1)2R−(N−1)2λ(N − 1)2v(0)(cid:16)1 − (hR)(N−1)2(cid:17)[ Note that the right-hand side tends to zero as t1 → +∞ and is independent of t2, by Cauchy principle we deduce that there exists some x∞ ∈ (R3)N−1 such that x(t) → x∞ as t → +∞. (2) Assume s = 1. Then (22) becomes (N −1)2 ] t1 . which implies that x∗ ≤ Z ≤ aZ + b, √2 2 (cid:16)(cid:0) 1 − a(cid:1)2 − 1(cid:17)1/2 b . The right-hand side is positive by our hypothesis and thus gives a uniform bound for x(t). For the remaining parts we proceed as in case (1). zℓ z∗ zu Fig. 1. Shape of F . (3) Assume that s > 1. The derivative F ′(z) = 1 − sazs−1 has a unique zero at z∗ = ( 1 sa) 1 s s−1 − a(cid:0) 1 sa(cid:1) by our hypothesis (19). Since F (0) = −b < 0 and F (z) → −∞ as z → +∞ we see that the shape of F is as in Fig.1. For t ∈ N, let Z(t) = (1 + 2x(t∗)2) is defined as in (21), i.e., t∗ ∈ argmax0≤τ≤tx(τ ). For t = 0 we must have t∗ = 0 and F (z∗) = (cid:0) 1 sa(cid:1) s−1 − b > 0, 1 2 , where t∗ 2 ≤ 1 + √2x(0) = b ≤ ( 1 sa) 1 1 s−1 = z∗. Z(0) = (1 + 2x(0)2) This implies that Z(0) ≤ zℓ. Assume that there exists t ∈ N such that Z(t) ≥ zu and let t0 be the first such t. Then t0∗ = t0 ≥ 1 and for all t < t0, 1 (1 + 2x(t)2) x(t) ≤(cid:0) z∗ x(t0 − 1)2 ≤ zℓ 2 ≤ zℓ ≤ z∗, 2 (cid:1) 2−1 2 ≤ z∗ 2−1 2 . 1 2−1 2 . that is, In particular, On the other hand, Z(t0) ≥ zu gives Thus we have 10 1 s−1 and 2−1 2 . 2−1 x(t0)2 ≥ zu 2 ≥ z∗ 2 − zℓ z∗ 2 2 x(t0)2 − x(t0 − 1)2 ≥ 1 2 ≥ (z∗ − zℓ)z∗. (23) By the intermediate value theorem, there is a ξ ∈ [zℓ, z∗] such that F (z∗) = F ′(ξ)(z∗ − zℓ). Note that F ′(ξ) = 1 − saξs−1 ∈ [0, 1], therefore, z∗ − zℓ ≥ F (z∗). We combine (23) to obtain However, we have x(t0)2 − x(t0 − 1)2 ≥ 1 2 z∗F (z∗). (24) x(t0) − x(t0 − 1) ≤ x(t0) − x(t0 − 1) = hv(t0 − 1) ≤ 2hλv(0)∞ ≤ 2hλv(0). Therefore, 11 x(t0)2 − x(t0 − 1)2 = (x(t0) − x(t0 − 1))2 + 2 (x(t0) − x(t0 − 1))x(t0 − 1) We combine this inequality with (24) to obtain ≤ 4h2λ2v(0)2 + 4hλv(0)x(t0 − 1) ≤ 4h2λ2v(0)2 + 4hλv(0)(cid:16) z∗ 2 (cid:17) . 2−1 1 2 thus, we have 1 2 ≤ 8h2λ2v(0)2 + 4√2hλv(0)z∗, 2−1 2 (cid:17) z∗F (z∗) ≤ 8h2λ2v(0)2 + 8hλv(0)(cid:16) z∗ s−1 "(cid:18) 1 =(cid:18) 1 s(cid:19) a(cid:19) ≤ 8h2λ2v(0)2 (sa) a ≤ 8λ2v(0)2s F (z∗) N 2 s−1 1 1 1 s−1 s s−1# − b s(cid:19) −(cid:18)1 s−1 + 4√2hλv(0) 4√2λv(0) , 1 1 s−1 + N where (10) is used. This contradicts to our hypothesis (19). So we conclude that, for all T ∈ N, Z(t) ≤ zℓ 2 , which is a uniform bound for x(t). Again we can proceed as in case (1) to and then x(t) ≤ (cid:0) z∗ complete the proof. 2 (cid:1) 2−1 1 IV. SIMULATIONS In this section, we present some numerical simulations to show the flocking behavior in a C-S model with alternating leaders. For these simulations, we choose a flock consisting of three agents, labeled by V = {1, 2, 3}, and we take the parameters h = 0.2, and β = 1 4 . We consider the switching in three interaction topologies described by graphs G1 =(cid:0)V,{(1, 2), (1, 3)}(cid:1), G2 =(cid:0)V,{(2, 1), (2, 3)}(cid:1), G3 =(cid:0)V,{(3, 1), (3, 2)}(cid:1). This means that in Gi, the agent i acts as the leader and there are information flows from i to the other two agents. We choose the switching signal σ(t) as σ(t) = (t, mod 3) + 1, t is the discrete time mark. In other words, the sequence of neighbor graphs is given by G1 → G2 → G3 → G1 → G2 → · · · , (25) and at each time step, there is a switching in the neighbor graphs; that is, the dwelling time for each active graph is Td = h = 0.2. For the initial state, we choose the initial position x(0) ∈ (R3)3 with nine coordinates randomly distributed in an interval of length 10; while the coordinates of initial velocities were randomly chosen from an interval of unit length. In Figure 2 we show the evolution of relative positions xi = xi − x3 and the evolution of the norm of total relative velocity (v12 + v22) 2 . These simulations show the asymptotic (exponentially fast) flocking behavior of the C-S model with alternating leaders. 1 In Figure 3 (a), we also exhibit the evolution of velocities v1, v2, and v3 under the switching signal σ. Here, we use different colors to denote different coordinates, and use different markers to describe different agents. Figure 3 (a) indicates that each coordinate of their velocities asymptotically attains an alignment, i.e., they exhibit a velocity consensus asymptotically. To compare the relaxation process of the flocking state under different switching topology, we did some simulations with different switching signal. We chose the similar signal with a sequence (25) but with a different dwelling time Td for each neighbor graph. Precisely, we take Td = 1 seconds, Td = 3 seconds and Td = 7 seconds for the simulations exhibited in Figure 3 (b), (c) and (d), respectively. We observe that in any case the velocities attain a near alignment after t0 = 9 and keep like this. This means that the flocking state is robust about the alternating leaders. Moreover, we observe that the switching signal before the alignment influence the asymptotic value a lot. Since the leader in the first active interaction topology is agent 1, if the dwelling time is longer, the asymptotic velocity is more close to the initial value of the agent 1. 12 x k i 8 6 4 2 0 −2 0 101 100 ¡v 12 +v 22¢ 1 2 10−1 10−2 10−3 0 3 6 3 6 time 9 12 15 (a) The evolution of each coordinate of relative posi- tions xk 3 , i = 1, 2, k = 1, 2, 3. i = xk i − xk (b) The evolution of (v12 +v22) vi = vi − v3, i = 1, 2. 9 12 15 time 2 for the relative velocities 1 Fig. 2. The emergence of flocking in a 3-agent C-S model with alternating leaders V. CONCLUSION We studied the Cucker-Smale flocking under the rooted leadership with alternating leaders. This dy- namically changing interaction topology is motivated by the ubiquitous phenomena in our nature such as the alternating leaders in migratory birds on the long journey, the changing political leaders in human societies, etc. Our study showed that the flocking behavior can occur for such a dynamically changed leadership structure under some sufficient conditions on the initial configurations depending on the decay rate of communications and the size of flocking. REFERENCES [1] S. Ahn, H.-O. Bae, S.-Y. Ha, Y. Kim and H. Lim, Application of flocking mechanism to the modeling of stochastic volatility, Math. Models Methods Appl. Sci. 23 (2013) 1603-1628. [2] S. Ahn, H. Choi, S.-Y. Ha and H. Lee, On collision-avoiding initial configurations to Cucker-Smale type flocking models, Comm. Math. Sci. 10 (2012) 625-643. [3] S. Ahn and S.-Y. Ha, Stochastic flocking dynamics of the Cucker-Smale model with multiplicative white noises, J. Math. Phys. 51 (2010) 103301. [4] M. Cao, A. S. Morse and B. D. O. Anderson, Reaching a consensus in a dynamically changing environment: A graphic approach, SIAM J. Control Optim. 47 (2008) 575-600. [5] M. Cao, A. S. Morse and B. D. O. Anderson, Reaching a consensus in a dynamically changing environment: Vonvergence rates, meansurement delays, and asynchronous events, SIAM J. Control Optim. 47 (2008) 601-623. [6] J. A. Carrillo, M. Fornasier, J. Rosado and G. Toscani, Asymptotic flocking dynamics for the kinetic Cucker-Smale model, SIAM J. Math. Anal. 42 (2010) 218-236. [7] F. Cucker and J.-G. Dong: On the critical exponent for flocks under hierarchical leadership, Math. Models Methods Appl. Sci. 19 (2009) 1391-1404. [8] F. Cucker and J.-G. Dong, Avoiding collisions in flocks, IEEE Trans. Automatic Control 55 (2010) 1238-1243. [9] F. Cucker and C. Huepe, Flocking with informed agents, MathS in Action 1 (2008) 1-25. [10] F. Cucker and E. Mordecki, Flocking in noisy environments, J. Math. Pures Appl. 89 (2008) 278-296. [11] F. Cucker and S. Smale, Emergent behavior in flocks, IEEE Trans. Automat. Control 52 (2007) 852-862. 13 0.4 0.2 k v i 0 −0.2 −0.4 0 0.4 0.2 0 k v i −0.2 −0.4 0 0.4 0.2 k v i 0 −0.2 −0.4 0 0.4 0.2 k v i 0 −0.2 −0.4 0 3 6 time 9 12 15 (a) Td = h = 0.2 3 6 time 9 12 15 3 6 time 9 12 15 (b) Td = 1 3 6 time 9 12 15 (c) Td = 3 (d) Td = 7 Fig. 3. coordinates, respectively. The lines marked by star, circle and dot indicate the agent 1, 2 and 3, respectively. The velocity alignment in a 3-agent C-S model with alternating leaders. The lines in blue, red and black indicate x, y and z [12] F. Cucker and S. Smale, On the mathematics of emergence, Japan. J. Math. 2, (2007) 197-227. [13] F. Cucker, S. Smale and D. Zhou, Modeling language evolution, Found. Comput. Math. 4 (2004) 315 -- 343. [14] R. Diestel, Graph Theory, Graduate Texts in Mathematics. New York, U.S.A.: Springer-Verlag, 1997. [15] S.-Y. Ha, K. Kim and K. Lee, A mathematical model for multi-name Credit based on the community flocking, To appear in Quantitative Finance. [16] S.-Y. Ha, K. Lee and D. Levy, Emergence of time-asymptotic flocking in a stochastic Cucker-Smale system, Commun. Math. Sci. 7 (2009) 453-469. [17] S.-Y. Ha, Z. Li, M. Slemrod and X. Xue, Flocking behavior of the Cucker-smale model under rooted leadership in a large coupling limit, To appear in Quart. Appl. Math. [18] S.-Y. Ha and J.-G. Liu, A simple proof of Cucker-Smale flocking dynamics and mean field limit, Commun. Math. Sci. 7 (2009) 297-325. [19] S.-Y. Ha and E. Tadmor, From particle to kinetic and hydrodynamic description of flocking, Kinetic and Related Models 1 (2008) 415-435. [20] A. Jadbabaie, J. Lin and A. Morse, Coordination of groups of mobile autonomous agents using nearest neighbor rules, IEEE Trans. Autom. Control 48 (2003) 988-1001. [21] J. Kang, S.-Y. Ha, E. Jeong and K.-K. Kang, How do cultural classes emerge from assimilation and distinction? An extension of the Cucker-Smale flocking Model, To appear in J. Math. Sociology. [22] G. Latouche and V. Ramaswami, Introduction to Matrix Analytic Methods in Stochastic Modeling. PH Distributions, ASA SIAM, 1999. [23] N. E. Leonard, D. A. Paley, F. Lekien, R. Sepulchre, D. M. Fratantoni and R. E. Davis, Collective motion, sensor networks and ocean sampling, Proc. IEEE 95 (2007) 48-74. [24] Z. Li, S.-Y. Ha and X. Xue, Emergent phenomena in an ensemble of Cucker-Smale particles under joint rooted leadership, To appear in Math. Models Methods Appl. Sci. [25] Z. Li and X. Xue, Cucker-Smale flocking under rooted leadership with fixed and switching topologies, SIAM J. Appl. Math. 70 (2010) 3156-3174. [26] J. Park, H. Kim and S.-Y. Ha, Cucker-Smale flocking with inter-particle bonding forces, IEEE Tran. Automatic Control 55 (2010) 2617-2623. [27] L. Perea, G. G´omez and P. Elosegui, Extension of the Cucker-Smale control law to space flight formation, J. Guidance, Control and Dynamics 32 (2009) 526-536. [28] J. Shen, Cucker-Smale flocking under hierarchical leadership, SIAM J. Appl. Math. 68 (2007) 694-719. [29] J. Toner and Y. Tu, Flocks, herds, and Schools: A quantitative theory of flocking, Physical Review E. 58 (1998) 4828-4858. [30] T. Vicsek, A. Czir´ok, E. Ben-Jacob, I. Cohen and O. Schochet, Novel type of phase transition in a system of self-driven particles, Phys. Rev. Lett. 75 (1995) 1226-1229. [31] X. Xue and L. Guo, A kind of nonnegative matrices and its application on the stability of discrete dynamical systems, J. Math. Anal. Appl. 331 (2007) 1113-1121. [32] X. Xue and Z. Li, Asymptotic stability analysis of a kind of switched positive linear discrete systems, IEEE Trans. Autom. Control 55 (2010) 2198-2203. 14
1812.04750	2	1812	2019-02-22T03:41:18	An Exchange Mechanism to Coordinate Flexibility in Residential Energy Cooperatives	[ "cs.MA" ]	Energy cooperatives (ECs) such as residential and industrial microgrids have the potential to mitigate increasing fluctuations in renewable electricity generation, but only if their joint response is coordinated. However, the coordination and control of independently operated flexible resources (e.g., storage, demand response) imposes critical challenges arising from the heterogeneity of the resources, conflict of interests, and impact on the grid. Correspondingly, overcoming these challenges with a general and fair yet efficient exchange mechanism that coordinates these distributed resources will accommodate renewable fluctuations on a local level, thereby supporting the energy transition. In this paper, we introduce such an exchange mechanism. It incorporates a payment structure that encourages prosumers to participate in the exchange by increasing their utility above baseline alternatives. The allocation from the proposed mechanism increases the system efficiency (utilitarian social welfare) and distributes profits more fairly (measured by Nash social welfare) than individual flexibility activation. A case study analyzing the mechanism performance and resulting payments in numerical experiments over real demand and generation profiles of the Pecan Street dataset elucidates the efficacy to promote cooperation between co-located flexibilities in residential cooperatives through local exchange.	cs.MA	cs	An Exchange Mechanism to Coordinate Flexibility in Residential Energy Cooperatives Shantanu Chakraborty The University of Melbourne, Melbourne, Australia. [email protected] Centrum Wiskunde & Informatica, Amsterdam, The Netherlands. Pablo Hernandez-Leal and Michael Kaisers {Pablo.Hernandez,Michael.Kaisers}@cwi.nl 9 1 0 2 b e F 2 2 ] A M . s c [ 2 v 0 5 7 4 0 . 2 1 8 1 : v i X r a Abstract -- Energy cooperatives (ECs) such as residential and industrial microgrids have the potential to mitigate increasing fluctuations in renewable electricity generation, but only if their joint response is coordinated. However, the coordination and control of independently operated flexible resources (e.g., storage, demand response) imposes critical challenges arising from the heterogeneity of the resources, conflict of interests, and impact on the grid. Correspondingly, overcoming these chal- lenges with a general and fair yet efficient exchange mechanism that coordinates these distributed resources will accommodate renewable fluctuations on a local level, thereby supporting the energy transition. In this paper, we introduce such an exchange mechanism. It incorporates a payment structure that encourages prosumers to participate in the exchange by increasing their utility above baseline alternatives. The allocation from the pro- posed mechanism increases the system efficiency (utilitarian social welfare) and distributes profits more fairly (measured by Nash social welfare) than individual flexibility activation. A case study analyzing the mechanism performance and resulting payments in numerical experiments over real demand and generation profiles of the Pecan Street dataset elucidates the efficacy to promote cooperation between co-located flexibilities in residential cooperatives through local exchange. Index Terms -- Exchange Mechanism, Energy Cooperative, Multi-agent System, Flexibility, Market Design. I. INTRODUCTION Energy Cooperatives (ECs) of prosumers are gaining con- siderable traction due to their potentials for efficient energy management, reduced grid dependency, and increased local consumption of distributively generated renewable energy. Establishing a local market can facilitate the management of distributed renewable generation that needs to consume as locally as possible [1]. A generic prosumer marketplace can be envisaged as a multiagent system that hosts heterogeneous services with a large number of participant prosumers which take actions au- tonomously [2]. This autonomy and diversity form a complex system where efficient mechanisms such as local markets are crucial [3]. The necessity of establishing a local prosumer marketplace (in the sense of the scope and proximity of the served area) that fulfills the requirements of decentralized production and consumption is therefore of a paramount importance [4]. Flexibility, on the other hand, can be enabled by facilitating Demand Response [5], which is an important element in the prosumer-centric energy market. While previous works have highlighted the need for demand response (DR) markets they have not considered small-scale prosumers as market participants in a local market. One example considers stakeholders such as DSO, TSO (as buyers), aggregators and customers (as sellers) [6]. However, market interactions down to the level of prosumers are not considered. Another work presented an effort-based DR service where the DR partic- ipants are benefited against the time of their stalling effort and showed that by doing so, the system achieves socially efficient allocations [7]. However, the implications of such an innovative design on the real electricity market scenarios are not entertained. Coordination of flexible resources/devices while aggregating their flexibility usually requires a detailed understanding of the underlying characteristics [5] of those devices which may be hard to obtain in many scenarios and thus hinders the scalability of the system. A considerable amount of research is focused on the management of heterogeneous devices (e.g. PV generator, storages, gas micro-turbine) through a centralized controller [8], while hierarchical controls of microgrid through several levels of control architecture [9], [10] show promising applicability in this domain. Additionally, in order to opti- mally manage and control flexibility in demand side, previous works deploy aggregators [11], [12] that essentially combine the available flexibilities and trade them on the customers' behalf. However, this is not a straightforward process and may have negative impacts such as increasing the imbalance of Balancing Responsible Parties (BRPs) and leading to en- gaging in further coordination between the participants [13]. Moreover, the underlying loss emerging from the balancing activity incurred due to energy-only trading are socialized, which means prosumers are priced equally regardless of their marginal contributions to the overall loss. Hence, it is critically important to improve the fairness of the balancing cost attributions. Consequently, the need for a general and fair yet effective coordination and control scheme for flexible resources through local market design becomes crucial. • Proposing a real-time exchange mechanism that promotes efficient allocations in ECs, coordinating heterogeneous and individually-controlled flexible resources. • Proposing a payment structure -- integrated with the proposed mechanism -- that incentivizes the prosumers to participate in the exchange mechanism, rewards them according to their contributions to system loss reduction, In light of these related works and the narrative aforemen- tioned, this paper contributes to the state of the art by: and by doing so, residential community. increases the social welfare of the The rest of the paper is organized as follows: Section II describes the prosumer marketplace and outlines the oper- ation necessary for the exchange mechanism. The proposed control and exchanged mechanism is detailed in Section III. Section IV elaborates the prosumers' utility and the fairness of the mechanism. A simulation case study is presented in Section V. Finally, Section VI concludes the paper with a glimpse of possible follow-up research. II. PROSUMER MARKETPLACE AND OPERATION OUTLINE In this section, we describe a real-time exchange mechanism that operates on a local prosumer marketplace with associated market offers that are necessary for the market formation. Prosumers in a cooperative are represented by software agents, therefore, we will use the terms agent and prosumer inter- changeably. While we discuss it in the context of storage flexibility from batteries, it is applicable to the flexibility of demand response as well. In this section, we describe a real-time market with market offers that are necessary for residential ECs. The market while operating locally, facilitates trading energy and flexibility (of consumption, commonly known as Demand Response) among the prosumers. The prosumers of the cooperative are represented by software agents. Therefore, we will use the terms agent and prosumer interchangeably. The prosumers are assumed to be equipped with PV and Battery (as flexible resources). To this end, we formalize the market offers and associated commodities. A. Market offers A market offer o = (c, p, f ) is defined as a tuple comprising a commodity c ∈ C with a price tag p and a flag stating whether the offer is a bid (f = 1) or an ask (f = 0). The offer can be further extended by adding a quantity n, differentiating between limit orders and fill-or-kill orders l, and an expiry timestamp texp. B. Commodity classes A commodity, in general term, is a marketable item (typ- ically, a good or a service) produced to satisfy demand or need [14]. In the context of the electric power market, a commodity comprises power time series, giving rise to several interesting classes as services. • Energy: C = (q, t), where q and t are equal-length vectors of time and quantity respectively, defining a (possibly continuous) time series with ∀i : qi ≥ 0, with the convention that the last quantity is zero. • Demand Response: C = (q, t), where q and t are vectors of time and quantity respectively, defining a (possibly continuous) time series, possibly imposing the zero-sum i qi = 0 to make this class of commodities complementary to energy This constraint may simplify discrimination from energy commodities but may be counter-productive as it avoids demand response losses to constraint (cid:80) be included in the commodity. However, since losses can- not be standardized, demand response is rather amenable to negotiation than auction-based commodity trading. In this article, however, we limit the contribution to devising exchange mechanisms of Energy commodity while keeping the Demand Response activated by indirectly controlling the local flexibility. The market protocol proceeds in rounds where every agent submits a bid (or ask), allocations are computed by matching bids and asks, establishing the market equilibrium. In this paper, however, we have assumed all bids (and asks) are cleared by a local exchange market. The exchange market clears the offers in near real-time and provides the allocations back to the agents. The operation outline is detailed in the following section. C. Operation outline The high-level operational outline of the proposed mecha- nism is shown in Figure 1. In brief, the mechanism periodically accepts offers from prosumers, computes the allocated ex- change quantities by matching offers, and responds back to the prosumers with those allocations. The additional payments are calculated retrospectively after a predefined operation cycle the of T discrete time periods is completed. Specifically, algorithm proceeds as follows: • Step 1: For each time t ∈ T , repeat Step 1.1 to Step 1.5. -- Step 1.1: Collect the market offers from all prous- mers, segregating bids and asks based on the embed- ded f flag. -- Step 1.2: Tabulate the quantities with round-trip effi- ciencies of each flexible devices of sellers (detailed in Section III). -- Step 1.3: Determine the allocations, using the bids and efficiencies of the devices after applying the proposed exchange mechanism. -- Step 1.4: Respond back to the prosumers with the al- locations, where the prosumers realize the exchange and deploy batteries. -- Step 1.5: Transact the prosumer-specific residual net energy with the Grid. • Step 2: Determine the payments of prosumers as detailed in Section IV. The form of payment includes the transac- tions with the Grid, realized energy exchange with peer prosumers, and allocated flexibility activation. Next, we introduce the baseline cases of No flexibility and Individual control, and proceed to provide a brief description of the proposed exchange mechanism (i.e. Exchange and control), which is elaborated in the subsequent section. • No flexibility: Considers the simplest case where there is no exchange and the flexibility is deactivated (i.e., batteries are not considered). The net-demand is always resolved by interacting with the main Grid. • Individual control: The agents individually control their batteries by charging with excess PV and by discharging to serve their own load. Agents are not allowed to exchange energy. Fig. 1. High-level operational outline of the coordinating exchange mechanism: exchange and control (Section III) followed by payment computation (Section IV). • Exchange and control: Agents are allowed to exchange energy. We assume a centralized exchange is responsible for computing allocations from participants' offers (see Section III). An offer comprises the net demand and the round-trip efficiency of their flexibility (e.g., battery efficiency). By doing so, our proposed mechanism is able to improve allocative efficiency by resolving flexibility needs with the most efficient means. III. EXCHANGE AND CONTROL MECHANISM The optimal allocation for pooled resources is computed by solving a centralized Linear Programming problem. Let, exb(s) ∈ R+ be the energy traded from seller s ∈ S to buyer b ∈ B; where B and S are the set of all buyers and all sellers (both including the Grid {G}), respectively. Aligning with the commodity as Energy (defined in Section II-B), at any discrete time interval t, the buyers and sellers approach to the real-time exchange market by providing a quantity q as Eb(t) and Es(t), respectively. Therefore, for a buyer b the bid structure looks like c ← (q = Eb(t), t) and the market offer is o = (c, f = 0).1 Similarly, for a seller s, the bid structure and the market offer are represented by c ← (q = Es(t), t) and o = (c, f = 1), respectively. In addition to submitting the net demand to the centralized exchanger, the mechanism assumes that sellers submit the round-trip efficiency of their locally controlled flexible devices. Therefore, the bids are restructured as c ← (q =< Es(t), ηs >, t) where, ηs be the round-trip efficiency of the battery (or other flexible resources) of seller s ∈ S − G. The mechanism makes use of these submitted round-trip efficiencies to compute the allocation. The objective is to maximize the local energy exchange and thereby minimizing the Grid exchange. In order to do so, for a particular seller s, the exchanger s = ∞ when s := reformats ηs to ηb {G} or b := {G}. The transformation can be written as, ∀s ∈ S 1The price component p is omitted since the mechanism assumes clearing s, ∀b ∈ B and assigns ηb all bids and asks. (cid:26)ηs, ∞, ηb s = if b ∈ B − {G} otherwise. (1) After receiving the quantities as bids from prosumers, the exchange decides on the allocation that maximizes energy exchange locally (equivalently, minimizes energy transaction with the Grid2). The following objective function serves the purpose of minimization of the Grid exchange with improved allocative efficiencies s · exb(s). ηb (2) (cid:88) (cid:88) b∈B s∈S min ex The above objective function is subjected to the following balancing equations for buyers and sellers, respectively. Eb = Es = CM (b, s) · exb(s) ∀b ∈ B − {G} CM (b, s) · exb(s) ∀s ∈ S − {G}, (3) (cid:88) (cid:88) s∈S b∈B where CM is a "boolean" connectivity matrix representing possible trading restrictions due to the electrical network topology of the cooperatives. These constraints ensure that the demand (and the supply) is met entirely through the exchange with physically connected prosumers. The optimization problem (Eq. 2) with constraints (Eq.3) essentially forms to an assignment problem. In this paper, we assume that the network is not restrictive, and it is therefore modeled by a fully-connected topology i.e. CM (b, s) = 1,∀b ∈ B and ∀s ∈ S. Note that this mechanism is device- agnostic, is not considering the (locally-controlled) storages' internal specification in computing allocations. i.e., it The aforementioned optimization problem with associated constraints is solved by the Linear Programming technique. 2The mechanism does not actually allocate any Grid exchange with prosumers. Rather, it provides the optimal allocation among the prosumers. Having said that, such a network topology is virtually pos- sible by adapting advanced routing technology, e.g. Digital- Grid [15] that provides every prosumer the platform to ex- change power with every other prosumers in the cooperative facilitated by the installed Digital Grid Routers in suitable nodes. IV. PROSUMERS' UTILITY AND FAIRNESS OF THE MECHANISM In this section, we describe the utility of prosumers by applying the coordination mechanism and go on showing how the mechanism increases the utility of prosumers, improving the social welfare [16] and thereby resolving the conflict of interest of the prosumers. For ease of description, we number the mechanisms as No flexibility (0), Individual control (1), and Exchange and control (2), respectively. Let N be the number of prosumers and T be the current operation cycle. Further, let the demand and (PV) generation of agent i at time t be di(t) and pvi(t), respectively. The net demand at period t after applying mechanism m and deploying the local battery of agent i ∈ N is defined as: i,−i(t) + pbi(t), i (t) = di(t) − pvi(t) − Exm (4) Dm where Exm i,−i(t) is the total energy agent i received from other agents −i at time t utilizing mechanism m. Therefore, Exi essentially is the agent specific decision allocated by the exchange described in Section III. We suppose that surplus generated energy is not permitted (or at least not remunerated) to be fed into the grid, i.e. power flow from Grid G to agent i, Dm i (t) is positive, with excess energy spilled: i (t)]+, DmG,i(t) = [Dm (5) where [x]+ = max{x, 0}. Since mechanism 0 and mechanism i,−i(t) = 0 for all i ∈ N. 1 do not impose any exchange, Ex0,1 The dispatched battery power of agent i at t is pbi(t), where pbi(t) < 0 when the battery is discharged. When an agent switches from No flexibility mechanism, the system encounters losses due to the round-trip efficiency of activated batteries. The overall loss contributed by i after switching to mechanism m ∈ {1, 2} is given by (cid:0)D0 i (t)(cid:1) − θm i i (t) − Dm (Exi,−i(t) − pbi(t)) − θm i , (6) where θm is an offset power dispatched from/to the battery to i bring the final state of the charge to the initial one. Finally, total loss of the system using mechanism m is calculated as N(cid:88) i = − N(cid:88) Lm (cid:32) T(cid:88) Lm = pbi(t) + θm i . (7) (cid:33) In order to instill fairness in the mechanism, we adopt a measure, namely difference evaluation function, which has been widely employed in multiagent learning [17]. In this context, the difference evaluation function of an agent i defines the marginal contribution of i in reaching a joint system state (i.e. loss in the system), L utilizing mechanism m. Formally, (8) Lm i = Lm − Lm−i. We define a component function that rewards/penalizes an agent based on its marginal contribution to system loss (i.e. Lm i ). It can be shown from Eq. 7, that the component function links with the agent's individual battery activation and there- fore, provides a way to measure the effect of battery activation and consequent profit sharing with other agents. Formally, for agent i using mechanism m = 2,3 the component function goes Rm i = Lm i − wm i Lm, (9) where wm bution of i using the following softmax function i constitutes the weights proportional to the contri- (10) where (cid:80) wm i = i = 1, which makes (cid:80) , i wm i = 0. It is an important criterion to fairness argument as the total rewards must be shared among all participated agents. i Rm Finally, the utility of an agent i by applying mechanism m is defined as the negative cost it incurs from the (total) payment to the Grid, payment from/to the exchange (with other agents), and the reward defined in Eq. 9. For a constant price p, the utility function takes the following form i(cid:80) eLm j eLm i (cid:34) T(cid:88) T(cid:88) ui(m) = −p DmG,i(t) + α Exm i,−i(t) − (1 − α)Rm i , t t (11) where α = [0, 1) represents the price component (as a fraction of p) an agent pays for exchanged energy. The complementary price component, i.e. (1 − α)) scales the Reward and thereby controls the profit sharing between agents according to their flexibility activation. Inherently, the utilities of No flexibility and Individual control only contain the Grid payment com- ponent. Note, that the price p is set out by the Grid and is in our setting identical for all prosumers in the EC for a particular operation cycle T . The acceptability of the proposed mechanism depends on whether the proposed coordination mechanism (m = 2) satisfies individual rationality of agents, i.e. ui(m = 2) ≥ ui(m = k), (12) where k ∈ {0, 1}. The mechanism searches for the lowest α that satisfies the following constraint (cid:88) (cid:0)DkG,i(t) − DmG,i(t)(cid:1) + Rm i < α (cid:32)(cid:88) (cid:33) Exm i,−i(t) + Rm i , (13) 3Mechanisms 0 and 1 do not engage in exchange and hence do not require component function. (cid:35) T(cid:88) T(cid:88) t t Lm i = = i=1 i=1 t t t for m = 2, k ∈ {0, 1} and i = 1, . . . , N. Typical values of α stay within 0.75 ≤ α < 1.0. In order to validate the efficiency of the mechanisms in improving the social welfare, we define the following utilitarian social welfare function N(cid:88) i N(cid:89) swm = ui(m). (14) Moreover, we quantify the relative fairness of a mechanism based on the Nash social welfare criterion, an established concept of fairness [16], applied to the improvement above a baseline: fmk = (ui(m) − ui(k)) , (15) i where m (cid:54)= k. Utilizing Eq. (4 - 12), it can be shown that m = 2 improves both social welfare functions, i.e. Eq. 14 and Eq. 15, and hence balances efficiency and fairness. The next section supports this claim empirically. V. A CASE STUDY OF LOCAL ENERGY EXCHANGE In this section, we present experimental results of a case study applying the proposed mechanism. The experiments are conducted by taking samples of 10, 20 and 30 prosumers (so called "scalability") over 150 households from real residential data [18]. We present experiment results considering the loss reduction and the payments and social welfare after applying various mechanisms. We use 10 and 20 prosumers cases for Fig. 2. System loss reduction over Individual control with increasing diver- sities in terms of number of prosumers (10, 20) and round-trip efficiencies. loss reduction experiment, whereas 30 prosumers case is used for payment and social welfare experiment. The prosumers are assumed to be equipped with a battery4. In order to experimentally investigate the effect of "diversity" on system performance, the round-trip efficiencies of the flexible re- sources are varied by sampling from normal distributions. The PV generation patterns of a set of randomly chosen prosumers are shifted uniformly with a span of [0,6] hours in order 4A residential Lithium-ion battery of capacity 6.8kWh with rated discharg- ing and charging power of 3kW and 1.3kW, respectively. The SOC is allowed to operate within 10% to 90% of the capacity. The degradation rate of the battery is fixed as 0.1% of the capacity. to form a local prosumer marketplace, simulating diversified installation angles (e.g., east/west facing solar panels). The simulation period covers 7 days, during which the exchange is performed over each quarter-hourly interval. Fig. 3. Example battery power dispatch (of one prosumer) by applying Individual control and Exchange and control mechanism plotted over a day. The exchange mechanism yields a "shifting" in battery scheduling. A. Experiment: Loss Reduction Here, we analyze the performance of the proposed mecha- nism with respect to the loss reduction, as described in Eq. 7. The experiments are performed over a number of trials, where each trial contains an independently chosen subset of the prosumers - the dataset of prosumers contains 150 prosumers. The improvement in loss reduction using the Exchange and control mechanism over the Individual control mechanism (i.e. Lm=2 − Lm=1) for different number of prosumers with increased diversity -- measured by the variance in the dis- tribution of round-trip efficiency of batteries -- is depicted in Figure 2. The positive loss reduction indicates that the Exchange and control mechanism influences the flexibility activation indirectly and thereby results in a system-wise higher efficiency. As the scale of the EC grows, so does the dominance of Exchange and control mechanism over the Individual control. As for the experiment with "diversity", the Exchange and control mechanism takes the round-trip efficiencies into consideration and reduces the loss further with greater diversity. An agent-specific residual battery power dispatch patterns, resulted from the Individual control and the Exchange and control mechanisms for 7 days (plotted in scatter on a 24-hour window) are shown in Figure 3. The storage pattern, emerging from the Exchange and control mechanism, shows a "shifting" tendency from that of the Individual control mechanism, which is resulted from the agent-specific realized allocation. The experiments confirm that the proposed mechanism efficiently coordinates the local flexible resources while resulting in a higher loss reduction when compared to Individual control. B. Experiment: Payment and social welfare Lastly, we turn the analysis toward the payment strategy of the exchange mechanism, and how the strategy makes the mechanism preferable for all individual agents. Figure 4 case study over real demand/generation data for various scales of EC with infused diversifications from underlying flexible devices and a follow-up sensitivity analysis to achieve loss reduction and improvement in social welfare has validated the efficiency and applicability of the proposed exchange mechanism in a residential EC setting. ACKNOWLEDGMENT This research has received funding through the ERA-Net Smart Grids Plus project Grid-Friends, with support from the European Unions Horizon 2020 research and innovation programme, and is a part of Energy Transition Hub research at the University of Melbourne, Australia. REFERENCES [1] C. Rosen and R. Madlener, "Regulatory options for local reserve energy markets: Implications for prosumers, utilities, and other stakeholders," The Energy Journal, vol. 37, 2016. [2] J. Rodr´ıguez-Molina, M. Mart´ınez-N´unez, J.-F. Mart´ınez, and W. P´erez- Aguiar, "Business models in the smart grid: challenges, opportunities and proposals for prosumer profitability," Energies, vol. 7, no. 9, 2014. [3] F. Lezama, J. Soares, P. Hernandez-Leal, M. Kaisers, T. Pinto, and Z. Vale, "Local energy markets: Paving the path towards fully transactive energy systems." IEEE Transactions on Power Systems, 2018. [4] F. Hvelplund, "Renewable energy and the need for local energy markets," Energy, vol. 31, no. 13, pp. 2293 -- 2302, 2006. [5] M. Alizadeh, A. Scaglione, A. Goldsmith, and G. Kesidis, "Capturing aggregate flexibility in demand response," in 53rd IEEE Conference on Decision and Control, Dec 2014, pp. 6439 -- 6445. [6] D. T. Nguyen, M. Negnevitsky, and M. de Groot, "Pool-based demand response exchange; concept and modeling," IEEE Transactions on Power Systems, vol. 26, no. 3, pp. 1677 -- 1685, Aug 2011. [7] F. N. Claessen, M. Kaisers, and H. L. Poutre, "Efficient balancing by effort-based activation of demand response services," in 2016 IEEE Innovative Smart Grid Technologies Conference Europe, 2016, pp. 1 -- 6. [8] H. Kanchev, D. Lu, F. Colas, V. Lazarov, and B. Francois, "Energy management and operational planning of a microgrid with a pv-based active generator for smart grid applications," IEEE Transactions on Industrial Electronics, vol. 58, no. 10, pp. 4583 -- 4592, Oct 2011. [9] J. M. Guerrero, J. C. Vasquez, J. Matas, L. G. de Vicuna, and M. Castilla, "Hierarchical control of droop-controlled ac and dc microgridsa general approach toward standardization," IEEE Transactions on Industrial Electronics, vol. 58, no. 1, pp. 158 -- 172, Jan 2011. [10] S. Chakraborty and T. Okabe, "Optimal demand side management by distributed and secured energy commitment framework," IET Genera- tion, Transmission & Distribution, vol. 10, pp. 3610 -- 3621(11), Novem- ber 2016. [11] A. Agnetis, G. Dellino, G. D. Pascale, G. Innocenti, M. Pranzo, and A. Vicino, "Optimization models for consumer flexibility aggregation in smart grids: The address approach," in IEEE First International Workshop on Smart Grid Modeling and Simulation, 2011, pp. 96 -- 101. [12] S. Ø. Ottesen, A. Tomasgard, and S.-E. Fleten, "Prosumer bidding and scheduling in electricity markets," Energy, vol. 94, pp. 828 -- 843, 2016. en- elec- https://ec.europa.eu/energy/en/studies/ ergy tricity study-effective-integration-distributed-energy-resources-providing-flexibility-electricity, 2015, accessed: 2017-10-04. distributed to of flexibility for system," [13] "Study integration providing resources the [14] A. O'Sullivan and S. Sheffrin, Economics: Principles in Action. Pren- on the effective tice Hall, 2003. [15] R. Abe, H. Taoka, and D. McQuilkin, "Digital grid: Communicative electrical grids of the future," IEEE Transactions on Smart Grid, vol. 2, no. 2, pp. 399 -- 410, June 2011. [16] H. Moulin, Fair division and collective welfare. MIT press, 2004. [17] M. Colby, T. Duchow-Pressley, J. J. Chung, and K. Tumer, "Local approximation of difference evaluation functions," in Proceedings of the 2016 International Conference on Autonomous Agents and Multiagent Systems, 2016, pp. 521 -- 529. [18] "Pecan street inc. dataport," http://www.pecanstreet.org/, 2017, accessed: 2017-03-07. Fig. 4. Prosumer utility improvements of Individual control (horizontal) and Exchange and control (vertical) over No flexibility baseline, with marginal cumulative distributions. Prosumers are above the dashed equal improvements line, hence the new Exchange and control mechanism dominates Individual control, while also vastly improving relative fairness f20 ≈ 8.89 · 1010 (cid:29) f10 ≈ 3.24 · 103. presents the relative increase in utility (i.e. decrease in cost) for each prosumer of an EC with 30 prosumers, comparing the improvements of Individual control and Exchange and control over No flexibility. As seen in the figure, (ui(2) − ui(0)) dominates over (ui(1) − ui(0)) by placing itself over the dashed line and by exhibiting a smoother cumulative distri- bution. Therefore, it implies that the social welfare functions defined in Eq. 14 and Eq. 15 are maximized by Exchange and control mechanism. The Nash social welfare function essentially balances the efficiency and the fairness. Therefore, it is evident that the Exchange and control mechanism is efficient and fair with the proposed payment structure. VI. CONCLUSION In this paper, our first contribution is a market mechanism -- operating on a near real-time prosumer marketplace -- that ensures efficient allocation of flexible energy resources in residential Energy Cooperatives (ECs). Our Exchange and control mechanism is device-agnostic as it does not require detailed modeling of the locally and independently controlled distributed flexible devices (e.g. batteries). The mechanism's payment function rewards participating prosumers based on their marginal contributions to the joint system states (i.e. loss reduction), which encourages the prosumers to participate in the exchange market while also resolving the conflict of interest between prosumers. This results in an improved utilitarian social welfare and improved fairness w.r.t. Nash social welfare. Due to the bidding and exchange execution in a near real-time fashion, the proposed mechanism inherently reduces the necessity of predicting demand and generation that are required for bidding in the day-ahead market. A detailed 0246810ui(1)−ui(0)0246810ui(2)−ui(0)Number of Prosumers: 30
1907.09198	1	1907	2019-07-22T09:31:09	Today Me, Tomorrow Thee: Efficient Resource Allocation in Competitive Settings using Karma Games	[ "cs.MA", "cs.AI", "cs.GT" ]	We present a new type of coordination mechanism among multiple agents for the allocation of a finite resource, such as the allocation of time slots for passing an intersection. We consider the setting where we associate one counter to each agent, which we call karma value, and where there is an established mechanism to decide resource allocation based on agents exchanging karma. The idea is that agents might be inclined to pass on using resources today, in exchange for karma, which will make it easier for them to claim the resource use in the future. To understand whether such a system might work robustly, we only design the protocol and not the agents' policies. We take a game-theoretic perspective and compute policies corresponding to Nash equilibria for the game. We find, surprisingly, that the Nash equilibria for a society of self-interested agents are very close in social welfare to a centralized cooperative solution. These results suggest that many resource allocation problems can have a simple, elegant, and robust solution, assuming the availability of a karma accounting mechanism.	cs.MA	cs	To appear in the Proceeding of the 22nd IEEE Intelligent Transportation Systems Conference, ITSC 2019. Today Me, Tomorrow Thee: Efficient Resource Allocation in Competitive Settings using Karma Games Andrea Censi, Saverio Bolognani, Julian G. Zilly, Shima Sadat Mousavi, Emilio Frazzoli 9 1 0 2 l u J 2 2 ] A M . s c [ 1 v 8 9 1 9 0 . 7 0 9 1 : v i X r a Abstract -- We present a new type of coordination mechanism among multiple agents for the allocation of a finite resource, such as the allocation of time slots for passing an intersection. We consider the setting where we associate one counter to each agent, which we call karma value, and where there is an established mechanism to decide resource allocation based on agents exchanging karma. The idea is that agents might be inclined to pass on using resources today, in exchange for karma, which will make it easier for them to claim the resource use in the future. To understand whether such a system might work robustly, we only design the protocol and not the agents' policies. We take a game-theoretic perspective and compute policies corresponding to Nash equilibria for the game. We find, surprisingly, that the Nash equilibria for a society of self-interested agents are very close in social welfare to a centralized cooperative solution. These results suggest that many resource allocation problems can have a simple, elegant, and robust solution, assuming the availability of a karma accounting mechanism. I. INTRODUCTION The very survival and success of a society with shared resources depends on the rules and protocols agents use to interact with each other. In designing the rules of these societies, there is always a trade-off concerning centralization, efficiency, robustness, and resiliency. A centralized system for resource allocation needs more infrastructure and is less robust and resilient, yet it is the most efficient. A distributed system is more resilient and privacy-preserving. In intelligent transportation systems, we can distinguish the "macro" level of the fleet, and the "micro" level of the vehicles. At the macro level, much research has shown how it is possible to obtain a substantial improvement in the efficiency of a transportation network [1], [2] by optimizing resource use through cooperative approaches; that is, one takes the perspective of a single agent which is able to control centrally a fleet of vehicles. At the micro level there are similar resource allocation problems. Because of the advent of self-driving cars to be used in autonomous mobility on demand networks, the 'micro' coordination problems become interesting, as we study how the codes, customs, and conventions of human drivers can be generalized to a scenario with both artificial and human agents. The authors are with the Institute for Dynamic Systems and Control (IDSC) and the Automatic Control Laboratory (IfA) at ETH Zurich, 8092 Zurich, Switzerland. E-mail: {acensi, bsaverio, jzilly, mousavis, efrazzoli}@ethz.ch c(cid:13)2019 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. Fig. 1. We propose an innovative approach to the problem of resource allocation in a competitive setting based on the notion of "karma", an accounting system that summarizes the agent's actions in the past. The karma system allows agents to accept to give in at a particular interaction, while receiving a karma compensation. This allows an overall more efficient use of resources. Agents interact by meeting in pairs, e.g. {i, j}, and bid on the resource by sending messages {mi, mj} specifying how much karma they are willing to bid in that particular interaction. The agent with the larger bid wins and gets access to the limited resource which in this case is access to an intersection leading to no delay δ in travel time for the winner. The prototypical problem is intersection management. Deciding which car may pass first is a resource allocation problem, in which the resource is the use of the space inside the intersection in a given time interval. Similar resource allocation problems happen also in maneuvers outside of intersections, as drivers compete for the use of space, although the outcome is not as simple as a discrete decision as in intersection management. These interactions happen between independent agents, with competitive goals, and typically are not repeated, as it is rare to encounter the same vehicle again. Therefore, there is little incentive to give in at one interaction; at face value, this appear to be a non-repeated game. Typical human drivers do not act like self-interested agents. Humans have ways to communicate urgency and politely negotiate maneuvers while they drive. Ultimately this is due to the altruism and pro-sociality bias that evolved in our species [3]; the bias makes the single individual intrinsically happy to accommodate somebody who seems to be in a hurry. Our species thrived because individuals are not completely self-interested. When we lived in tribes, deviant antisocial behavior was easily spotted and repressed; now that our social groups are counted in the billions, a set of rules (laws) and corresponding incentives (punishments) help in aligning the individual and societal interests in the handling of common resources [4]. When driving, some of our behaviors derive from these incentives (we do not speed because we are afraid of tickets), but many polite behaviors are due to our visceral intrinsic motivation rather than extrinsic rewards/punishment. How can we ensure that a population of artificial agents, such as self-driving cars, can attain the same efficiency of a pro-social species like humans? In this paper, we consider the problem of resource allocation in a setting that we call Karma Game. The idea is that considerable gains can be realized if an agent is inclined to give in at one interaction, if it is compensated with "karma". Thus, we introduce karma as a way to account for an agent's past actions. (This concept is closer to how "karma" is used in video game mechanics, rather than to how it is understood in Indian religions.) We define a karma protocol with which agents can negotiate the use of resources. The protocol describes the exchange of bidding messages and how karma is updated based on the outcome of the interaction. The protocol does not need a third party, and the primitives needed to implement karma accounting and the interaction are those provided by many blockchains, such as Ethereum [5]. Having fixed the protocol, we study how a population of self-interested agents will use it, by computing the Nash equilibria for the resulting Karma game. We then compare the Nash equilibria of the distributed system with the baseline of the optimal centralized policies. We observe that the efficiency of the system is remarkably similar. The social welfare is thus closely aligned with the self-interest of the agents, assuming the agents have reasonable discount factors. An agent that does not care about the future and lives for the present will also create an inefficient society. II. RELATED WORK a) Intersection control: Traditional intersection control strategies have been substantially based on utilizing control devices such as traffic lights, in which an offline optimization based on historical data can be used to provide a control signal [6], [7]. The main drawback of this control strategy is that it cannot adapt to changes in request patterns and environment. Improving upon classical control strategies, communication-based schemes [8] are based on a competitive scenario, in which different vehicles aim at minimizing their own selfish cost. It is assumed that the urgency ui(t) is a piece of private information of each vehicle i, and is therefore not accessible to other vehicles. This kind of scenario is typically tackled via auctions, which can be designed in order to induce selfish agents to disclose their true urgency [9] -- [16]. For example, in [9], the earliest time-slot in an intersection is auctioned off by an intersection manager among all vehicles at the front of each lane. In [10], having an infinite budget, any agent in a lane can participate in a second-price auction to enhance the winning chance of the agent at the front. In [11], a mechanism based on a first-price auction is proposed for the management of intersections. Two scenarios for single intersection and a network of intersections are considered in [12], and a policy based on a combinatorial auction for assigning the reservations of time-space slots is presented. However, finding the winner of a combinatorial auction is NP-hard [17]. Finally, to schedule the intersection usage, [16] proposes a variant of the Vickrey-Clarke-Groove mechanism in which an intersection unit charges each agent at the front of any lane with a time-token based on its impact on others. Almost all works in the literature which proposed an auction-based approach for the intersection control are static, one-time decision problems. However since the urgencies and the agent's private information change over time, a sequence of decisions needs to be made, resulting in a dynamic resource allocation and a dynamic bidding process [18]. Thus, the utility function of each agent along with the social welfare are defined based on the discounted utility over time. We assume that in every interaction, vehicles are allowed to communicate a scalar message mi(t). The karma value of each agent is a public state ki(t). Both the outcome of the interaction (who goes first) and the update of the public state ki are determined based on a set of rules which are known and verifiable to all agents (as they only depend on public information: the states ki, kj and the messages mi, mj). b) Karma-like concepts: A "karma" system was in- troduced in [19] in the context of file-sharing to prevent "freeloading" in peer-to-peer networks. In this framework, karma represents the standing of each agent in the system, that increases when contributing and decreases when consuming a resource, and thereby incentivizes agents to contribute resources [20]. In this and similar systems, the "value" of the karma is fixed -- in our approach, the agents are free to assign a value to karma according to their goals and current state. We note that our approach departs from the auction-based schemes in the mentioned papers in that to maintain the fair- ness properties between wealthier drivers and those without many funds, it does not require any monetary transactions, and therefore does not require to attach an objective value to the cost incurred by the vehicles. We will discuss later how this sheds light on the true nature of this coordination problem. Any vehicle is assigned an initial karma level. In light of the budget-balance property of our mechanism, the total amount of karma remains constant over the whole transportation network. Also, unlike the assumption in [10], every agent is assumed to have a limited total karma at any time period, which neither is negative, nor exceeds a maximum value. c) Population games: This competitive scenario can be modeled as a repeated game (interactions) between randomly selected agents in a large population. For the analysis of the resulting game, we adopt the approach that is typically used in the study of population games [21], which has its underpinnings in the following abstractions: 1) populations are continuous rather than discrete; the payoffs to a given strategy therefore depend on society's aggregate behavior in a continuous fashion; 2) the aggregate behavior in a population game is described by a "social state", which specifies the empirical distribution of strategy choices (or types) in the population; for simplicity, this social state is generally finite-dimensional. The specific application that we are considering has however some peculiarities, compared to standard population games: for example, each agent's type is also determined by an exogenous time-varying signal (their urgency). Moreover, there is no natural revision protocol or adaptation, and therefore no evolution of the agents. We therefore prefer to present the resulting game in a self- contained and specialized form, without explicitly tapping into that literature for definitions or results. Notice that the game we are formulating is more general than the specific traffic interaction problem, although clearly inspired by that setup. III. RESOURCE ALLOCATION IN A "DRIVE-BY" SCENARIO In this section we introduce a deliberately simple model for vehicle-to-vehicle interaction at intersections. We strove to simplify the model to its core features, in order to isolate the essential phenomena in this problem. We understand the problem of vehicle-to-vehicle interaction at intersections as an example of a "drive-by" scenario, in which: 1) There is a large number of agents in the systems. 2) Agents interact with a random schedule. 3) Each agent interacts many times with other agents over its lifetime. 4) The value of a resource to an agent varies in time according to an exogenous factor. For vehicle-to-vehicle (V2V) interactions at intersection: 1) There is a large number of cars on the road. 2) Cars meet randomly at intersections. 3) Each car encounters many intersections over its lifetime. 4) The value of time saved to a car varies in time according to its urgency on that day. A. Formalization More formally, consider a population N of N vehicles. Each vehicle i ∈ N has an associated urgency process ui(t). The urgency ui at time t indicates the marginal value that agent i gives to a unitary delay in its trip. It is an exogenous process that is not affected by the behavior of the vehicles. The vehicles interact at intersections. Each interaction at time t involves only a pair of vehicles I(t) = {i, j} ⊂ N . Every time two vehicles interact, one of the two vehicles is necessarily delayed by a unitary delay, while the other vehicle does not incur any delay. We therefore have two possible outcomes o(t), that is o(t) ∈ O := {i, j}. Agent i (and, in a completely symmetric way, agent j) incurs a cost c : O × U → R+ that is a function of the outcome and of its own urgency, and is defined as (cid:40) ci(o, ui) = B. Assumptions ui, 0, if o = i; otherwise. (1) We propose the following assumptions about the model. Assumption 1 (Randomness of encounters). The sequence I(t) is random and identically distributed at all times t over the set {I ⊂ N ,I = 2}, and each vehicle has the same probability of belonging to I(t) at a given t. Assumption 2. The urgency processes ui(t) are identical for all vehicles i ∈ N . The urgency at each time t is independent and identically distributed, and takes values in U := {0, U}. We defer the discussion on how to relax these assumptions to Section VIII. For the most part, these assumptions are introduced for technical convenience, as they yield a simpler analysis, a computational advantage (see also Section VI), and a more immediate interpretation of the results. C. Performance measures The focus of this paper is on policies that allow to decide o(t) optimally, where the notion of optimality is to be defined hereafter. We define two measures of social cost for the entire popu- lation, which are associated to two different interpretations. The first measure simply quantifies the expected aggregate cost for the entire system at each interaction: W1 := E(cid:2)(cid:80) (cid:96)∈N c(cid:96)(o(t), u(cid:96)(t))(cid:3). The second measure quantifies the expected rate at which the variance (across agents) of the accumulated cost grows: W2 = lim t→∞ E [var a(t + 1) − var a(t)] (cid:1)2 where (cid:80) (cid:0)a(cid:96) − 1 (cid:80) defined element-wise as a(cid:96)(t) = (cid:80)t var a := 1 N (cid:96)∈N and a denotes the vector of accumulated costs of the agents, τ =0 c(cid:96)(o(t), u(cid:96)(t)). In these expressions, E [·] represents the expectation with respect to both the stochastic urgency processes and the interaction selection process (which are independent processes). k∈N ak N D. Centralized policies In this section, we derive the optimal centralized policies for the simplified intersection management problem that we presented, under the notions of social optimality that we described. These optimal centralized policies will constitute a baseline for the analysis of the policies that emerge in a distributed competitive setting. In a centralized setting, we are allowed to adopt causal policies of the kind o(t) = Π (cid:16)I(t),{u(τ )}t (cid:17) τ =0 ,{o(τ )}t−1 τ =0 , where by u(t), we indicate the past urgencies of all agents. Under Assumptions 1 and 2, the optimal policies for the two social costs W1 and W2 can be computed explicitly. Proposition 1. The social costs W1 and W2 are minimized, respectively, by the policies 1(t) ∈ arg min o∗ (cid:96)∈I(t) u(cid:96)(t) and 2(t) ∈ arg min o∗ (cid:96)∈I(t) a(cid:96)(t − 1) + u(cid:96)(t). (2) (3) If the arg min operation does not return a singleton, then any of the two choices is optimal. Here and thereafter, we assume that arg min ties are resolved via fair coin flipping. We also define a third centralized policy, which prioritizes the minimization of W1 (therefore obtaining the same value for W1 as o∗ 1) and, in case of ties between the urgencies ui and uj (where I = {i, j}), aims at minimizing the unfairness defined by W2: 1,2(t) ∈ arg min o∗ (cid:96)∈I(t) ui(t) = uj(t) ⇒ o∗ 1,2(t) ∈ arg min (cid:96)∈I(t) and u(cid:96)(t) a(cid:96)(t − 1) + u(cid:96)(t). (4) IV. RESOURCE ALLOCATION USING KARMA GAMES In this section, we formulate a mechanism for resource allocation based on the notion of karma. We only design the mechanism and not the agents' policy, which is going to be found automatically through optimization. A. Informal definition of karma interaction mechanism We assume that there is an integer counter ki(t) (karma) associated to each agent bounded by kmax. The agents exchange one message at each interaction. Each agent i can produce a message mi which contains a value not to exceed its current karma: 0 ≤ mi(t) ≤ ki(t). We give this message the semantics of how much karma the agent sees fit to bid on the current interaction. The agent that provides the highest message is allowed to use the resource (go first at the intersection) and must pay the other agent up to the karma value that it has bet. The karma transferred is reduced if the transfer would make the other agent overflow kmax. Suppose that agent i wins betting mi. Then the karma transferred is min(mi, kmax − kj). Remark 1. In this paper we do not delve into the technical implementation of such a scheme, but we would like to remark that it is possible to implement such a scheme, in a completely distributed way, without an arbiter to preside at each interac- tion, by using some of the cryptographic primitives associated to blockchain technology. The counters are implemented using public addresses. Non-refutable messages are implemented using cryptographic commitments. The resolution and the outcome can be easily implemented using the primitives of, for example, Ethereum's Solidity language. B. Formal definition of Karma Game We formalize the discussion so far by defining Karma Games in a way that is slightly more general. Definition 1 (Karma Game in Tabular Format). A Karma Game G is a tuple G = (cid:104)K,M,O,U, p, c, α, γ, φ(cid:105), where: • K is a set of possible public states (karma) of an agent; • M is a set of possible messages of an agent; • O is a set of possible outcomes of an interaction; • U is a set of possible exogenous states of an agent and • c : O × U → R+ is the instantaneous cost for each agent, which depends on the outcome of the interaction and on the exogenous state of the agent; p is a probability distribution on U; • 0 ≤ α < 1 is a discount factor; • γ : K×M×K×M → P(O) is the interaction outcome • φ : K×M×K×M×O → P(K) is the state transition function, as a probability distribution on O; function. The interpretation is as follows. Suppose an agent of karma ki(t) meets an agent of karma kj(t), and they exchange messages mi(t) and mj(t). The function γ gives a distribution on the possible outcome o(t) ∈ O given by o(t) ∼ γ(ki(t), mi(t), kj(t), mj(t)). As for the consequences, φ is the map that specifies the probability distribution of the next value of ki and kj: ki(t + 1) ∼ φ(ki(t), mi(t), kj(t), mj(t), o(t)). The cost for each agent is given by the following series, where time is to be interpreted as ranging over the instants in which the agent participated in an interaction: C = E(cid:104)(cid:80) (cid:105) t=0 αtc(o(t), u(t)) . (5) C. Vehicle interaction as a Karma Game We now put in the form of Definition 1, the model we described so far. K = M is the set of integers up to kmax: K = {0, 1, 2, . . . , kmax}. There are two possible outcomes of each interaction, as explained in Section III: O = {i, j}. For the outcome distribution γ(ki, mi, kj, mj), we have 0, 1, 0.5, if mi > mj, if mi < mj, if mi = mj, P(o = i) = where we defined mi = min(mi, ki). For the state transition function φ(ki, mi, kj, mj, o), we have that with probability 1 ki(t + 1) = ki − min( mi, kmax − kj), min(ki + mj, kmax), if o(t) = i, if o(t) = j. (cid:40) These rules guarantee that • the total amount of karma is conserved. • karma is bounded above by kmax and below by 0. The cost function c is the one already defined in (1). V. ACTING RATIONALLY IN A KARMA GAME We now turn attention to what is the rational behavior of an agent in a Karma Game. An agent's behavior is completely defined by its policy. Definition 2 (Agent policy). In a Karma Game, the agent's policy π is a probability distribution over the possible messages, which varies as a function of the agent's current urgency ui(t) and current karma ki(t): mi(t) ∼ π(ui(t), ki(t)). As an agent, we need to decide what message to send for each combination of urgency ui ∈ U and karma ki ∈ K. In game theory jargon, we speak of a set A of different agent "types"; in this case, A (cid:39) U × K. The traditional notion of "agent type" does not fully capture our setting; because following an interaction, the type of an agent changes as they gain/lose karma. Moreover, the urgency is an exogenous variable that nobody can predict. Still, we use "agent type" in the following. Under our assumptions, it is easy to compute the optimal policy for an agent if the urgency is zero. In that case, the optimal action for the agent is to send a message mi(t) = 0. That is because the agent is indifferent to losing or winning the interaction regarding the cost; and, regarding the karma, the agent prefers to lose the interaction hoping to gain some karma. If an agent has a nonzero urgency, how much karma should she bid today? This does not have an easy answer, except in special cases. For example, if the discount factor α is zero -- the agent does not care about the future, then the optimal policy is to send the maximum message mi(t) = ki(t). In all other cases, we need to characterize and compute Nash equilibria for this game. Figure 2(a) shows a representation of such an optimal policy obtained as a Nash equilibrium. A. Characterization of Nash equilibria for a Karma Game To characterize the equilibrium of the game, we must consider, in addition to the policy, a series of other related quantities. These are: • D ∈ P(K) is the stationary distribution of karma values. • T : K → P(K) is a transition function for the karma levels; Figure 2(e) shows a representation of such a transition function. • c : K → R+ is the expected cost of one interaction, as Figure 2(c) shows a typical stationary distribution. a function of the agent's karma. The transition function T immediately descends from the composition of φ with the policy π and the outcome distribution γ, assuming the karma distribution D for the other agents and p for all agents' urgencies. To express c, it is convenient to define the function ρ(ui, ki, mi) which gives the expected utility of choosing message mi for an agent of type ui, ki. ρ(ui, ki, mi) = Dkj puj· (cid:88) mj∈M (cid:34) (cid:88) (cid:88) uj∈U o∈O (cid:88) kj∈K (cid:88) k(cid:48)∈K πmj (uj, kj) γo(ki, mi, kj, mj)· c(o, ui) + α φk(cid:48)(ki, mi, kj, mj, o)θ(k(cid:48)) (cid:35) . (6) Figure 2(b) shows a representation of a typical ρ. Based on this definition, the expected cost of an interaction is c(ki) = pui πmi(ui, ki)ρ(ui, ki, mi). (7) (cid:88) (cid:88) ui∈U mi∈M We can now define the notion of Nash equilibrium for a Karma Game. Definition 3 (Nash equilibrium for a Karma Game). A policy π is a Nash equilibrium for the Karma Game G if there exist D, T, c that satisfy three properties: P1: Stationarity: D is the equilibrium distribution for the transition map T : D =(cid:80) τ∈K Dτ T (τ ). P2: Bellman: There exists a function θ : K → R+, representing the expected total cost for an agent as a function of the present value of the karma, that satisfies the Bellman- like equation θ(k) = c(k) + α Tτ (k)θ(τ ) (8) (cid:88) τ∈K for the expected interaction cost c defined in (7) and the discount factor α. P3: Rationality: The policy π must yield the best expected outcome: C(π) ≤ C(π(cid:48)) ∀π(cid:48), where C was defined in (5) and can be expressed as (cid:88) ki∈K C = Dkiθ(ki). The next section will be devoted to the numerical compu- tation of a Nash equilibrium for the Karma Game of interest and to the interpretation of the resulting policies and outcome. VI. COMPUTING NASH EQUILIBRIA OF KARMA GAMES In general, Nash equilibria can be computed by iterative algorithms. Starting with an initial policy, one computes the other unknown (stationary distribution, karma utility); then one re-computes the optimal policy. If the recomputed policy is different from the initial one, the delta is a profitable perturbation of the policy. Based on the perturbation, one can make a small update of the policy, and repeat the process until convergence. If this process converges to a distribution, then by definition, we have found a Nash equilibrium as defined above. However, there is in general no guarantee that the iterative process converges. A. Fixed point computation We show here how to rearrange the equations to put them in the form of a fixed point. Suppose we have a current guess of the policy π, the stationary distribution D, and the utility θ. Step 1: Compute the policy π from the previous policy, the stationary distribution D, and the expected utility θ. The policy is computed using (9) based on the values of ρ obtained from (6). Step 2: Compute the transitions T from the policy π and the stationary distribution D. Given the policy and the stationary distribution, we can compute the transitions of the system. For each type ui, ki, we know the distribution of the types it will encounter, and we know their policy. Thus, we can compute the outcomes, and the consequences of the outcomes in terms of what will be the next value of ki. (a) (b) (c) (d) (e) Fig. 2. For a Karma Game, the definition of a Nash equilibrium (Definition 3) involves a policy, from which one can derive the other quantities. (a): Optimal policy as mapping from current karma level to likelihood of sending a given message m; (b) The expected utility of sending each message m as a function of karma levels. (c) Stationary karma distribution across all agents; (d) The expected utility of possessing a certain karma level. (e) The transitions show the likelihood of transitioning from a certain karma level to another level. (a) Iteration #74 (b) Iteration #305 (c) Iteration #705 (d) Iteration #954 Fig. 3. Evolution of policy as the temperature decreases in the simulated annealing procedure. The policy becomes progressively more rational, until we find a Nash equilibrium. (a) (b) (c) (d) Fig. 4. We investigate the effect of discounting future rewards. Displayed are discounting values α in the set {0, 0.3, 0.7, 1.0}. As time discounting α increases from 0 to 1, the future is valued higher and higher and less karma is bid for the same karma levels. rewards. As T → 0, the policy tends to the deterministic policy, where we select the maximum of ρui,ki(·): π(ui, ki, m) T→0−→ if m maximizes ρui,ki (m), 1, 0, otherwise. (cid:40) In the simulations, we gradually decrease the temperature of the system in a series of "eras" (Figure 3). C. Equilibria parametrization in α The parameter α introduced as a cost discounting factor in (5) determines how much importance an agent assigns to future costs. In the limit α (cid:28) 1, the agent is only occupied with minimizing instantaneous costs. When α approaches 1, future costs are deemed almost as important as present costs. To determine the influence this factor has on agent policies, we ran experiments with different α values ranging from 0 to 1 in 0.05 increments. As an overview of the effect, we provide Figure 4 which depicts the gradual changes in policy as α is increased. Similarly we offer Figure 5 as an overview of the effect of time discounting on the best message to send given a karma level. One caveat that we have is that the Nash equilibria are not well defined when α = 1 as some of the series in the formalization do not converge. Still, we also include the results of the algorithm for α = 1. Similarly, we believe that for α → 1 there are numerical instabilities, and in fact we find that there are much larger oscillations. Rather than tuning the optimization parameters for each α, we keep the same parameters, and we still picture the results for α = 0.9 and α = 0.95, without fully believing they are Nash equilibria for the game. VII. POLICY COMPARISON In this section, we are interested in gaining an empirical understanding of different solutions to the proposed distributed interaction problem. a) Evaluation protocol: All simulations of interactions follow the same general procedure. As described in Section III, agents randomly meet in pairs and bid karma if they are urgent in order to pass first in an intersection. All experiments were conducted with 200 agents and a total of 1000 time periods. On each day, there are an average of 0.1 interactions per agent. Agents are urgent with magnitude 3 with probability 0.5 and not urgent (magnitude 0) again with probability 0.5. Each agent has an initial karma level uniformly randomly chosen between 0 and 12. Agents can, through interactions, attain a minimum karma level of 0 and a maximum karma level of 12. In the following, we compare various policies as well as the underlying parameters influencing the agents' policies. We consider two performance metrics which are finite-sample proxies for W1 and W2, respectively: • "Inefficiency": This is the average cost per interaction attained by the agent at the end of the simulation period. Note that this is not the α-discounted factor that each agent is trying to minimize; rather, this is the social welfare -- which roughly corresponds to the case α = 1. Fig. 5. Expected message value given a karma level for mixed policies for different α discounting factors of future costs. Strategies with small discounting factor spend almost all available karma on a message whereas strategies with a large discounting factor save karma for the future. Step 3: Compute the stationary distribution D from the transitions T . This is a standard step - given a transition matrix, compute the equilibrium distribution. It can be done by iteration or by solving an eigenvector problem. Step 4: Compute the expected utility θ from D and π. We can compute the expected utility using (8). The expected daily cost c(k) is computed by setting α = 0 in (6). B. Momentum and simulated annealing We found two simple devices that make the convergence robust, in the sense that the policy converges to the same so- lution no matter the initial conditions of the policy, stationary distribution, and karma utility. 1) Momentum: In Section VI-A, we have defined a way to update the policy π that we can abstract as a function Ψ such that: πnew t = Ψ(πt, Dt, θt). Define the "momentum" τ as a scalar 0 < τ ≤ 1. Then we update the policy as πt+1 = τ πnew t + (1 − τ )πt. For the set of simulations described below the optimization parameters were constant, but we did find in general that for different values of the model properties, the optimization parameters had to be optimized. 2) Simulated annealing: Let T > 0 be a temperature parameter. Rather than looking for a pure strategy, we set π(ui, ki, mi) ∝ exp(−ρui,ki (mi)/T ). (9) For large values of T , agents choose a random action. As T decreases, the agents choose more often actions with good 024681012Karma024681012Expected value of messageNash equilibria (mixed strategies) α = 0.00 α = 0.05 α = 0.10 α = 0.15 α = 0.20 α = 0.25 α = 0.30 α = 0.35 α = 0.40 α = 0.45 α = 0.50 α = 0.55 α = 0.60 α = 0.65 α = 0.70 α = 0.75 α = 0.80 α = 0.85 α = 0.90 α = 0.95 α = 1.00The baselines provide a reference frame to interpret the results for the karma-based policies. We find many interesting nuggets. For example, bid1-always is very inefficient, as inefficient as baseline-random, but it is less unfair. This is because the karma accounting keeps track of previous times when the agent lost, thereby slightly reducing the unfairness even if the policy is trivial. Next consider the performance of bid1-if-urgent. This corresponds to a mechanism in which the agents use the karma message to reveal their urgency. Notwithstanding the fact that this is not an equilibrium for the game (this can be easily verified by noting that this is not a fixed point of the procedure discussed above), what we found surprising is that the efficiency is not as good as some of the Nash equilibria that we find. Next we consider the performance of the Nash equilibrium as a function of α. The sequence draws a hook in the inefficiency/unfairness space. The continuity of this curve also is good evidence that the procedure converged well (as noted before, for α ≥ 0.9 the convergence is not assured). We find the surprising result that for α ≥ 0.4, the Nash equilibria are better in efficiency than the bid1-if-urgent strategy. The reason is that the agents should bid more or less if their karma levels allow -- bidding only 1 is not the best strategy (neither for the agents nor society). For α < 0.4, the agents do worse. The α = 0 "there is no tomorrow" strategy (bid everything if urgent) is particularly bad for society, though not as bad as random: karma still allows some reparations to be made. We observe that for α > 0.4, the karma strategies beat the centralized-urgency strategy in unfairness. There is a minimum unfairness observed for α = 0.8 -- we are not sure how this relates to the parameters of the problem. In these experiments, for α = 0.85, the performance is closest to the centralized-urgency strategy in both inefficiency and unfairness, in fact surprisingly close. In conclusion, we obtain the surprising result that, for agents that are reasonably future-conscious, Nash equilibrium strategies beat heuristic solutions in both efficiency and fairness, and their performance is extremely close to the centralized solutions. VIII. CONCLUSIONS We have demonstrated how the efficient use of a shared infrastructure can emerge from simple coordination protocols among competitive agents, without the need of any monetary transaction or complex decision infrastructures, in sharp contrast to most of the literature. The enabler is the notion of karma: a public state that links the decision of the same agent at different times (as long as each agent reasonably values its own future cost). A solid understanding of the mechanisms that are necessary and sufficient for fair sharing of an infrastructure has the potential to guide the design of scalable solutions in many applications, and in particular for autonomous mobility. Fig. 6. Overview of efficiency and unfairness of random, centralized and karma-based strategies. Random solutions fare the worst in both domains whereas centralized solutions with access to all information are optimal to their respective objectives. Karma-based solutions describe a trend of better efficiency and fairness with increasing discount factor α up to a limit. • "Unfairness": This is the standard deviations of the costs at the end of the simulation period. b) Policies: In addition to the Nash equilibria found for sweeping α between 0 and 1, we consider these other policies, as they are useful reference points: • baseline-random: The winner is decided randomly. • bid1-always: The agents always bid 1. • bid1-if-urgent: The agents bid 1 if the urgency is nonzero, and zero otherwise. • centralized-cost: The policy (3). • centralized-urgency: The policy (2). • centralized-urgency-then-cost: The policy (4). c) Results: The overall results are shown in Figure 6. baseline-random (top right) obtains the worst results, as one might expect. centralized-urgency-then-cost (bottom left) obtains the best results for both fairness and efficiency, as expected. centralized-cost does well in terms of unfairness, as it tries to reduce the spread of the costs, but it is very inefficient. centralized-urgency obtains minimum inefficiency (as pre- dicted), but it does not do anything to reduce the spread of the costs, leading to a relatively high unfairness. 0.40.50.60.7Inefficiency (mean of cost)0.100.120.140.160.180.200.22Unfairness (std-dev of cost)baseline-randombid1-alwaysbid1-if-urgentcentralized-costcentralized-urgencycentralized-urgency-then-costbaseline-randombid1-alwaysbid1-if-urgentcentralized-costcentralized-urgencycentralized-urgency-then-costα = 0.00α = 0.05α = 0.10α = 0.15α = 0.20α = 0.25α = 0.30α = 0.35α = 0.40α = 0.45α = 0.50α = 0.55α = 0.60α = 0.65α = 0.70α = 0.75α = 0.80α = 0.85α = 0.90α = 0.95α = 1.00[14] J. Raphael, E. I. Sklar, and S. Maskell, "An intersection-centric auction- based traffic signal control framework," in Agent-Based Modeling of Sustainable Behaviors, 2017, pp. 121 -- 142. [15] I. K. Isukapati and S. F. Smith, "Accommodating high value-of-time drivers in market-driven traffic signal control," in IEEE Intell. Vehicles Symp., 2017, pp. 1280 -- 1286. [16] M. O. Sayin, C.-W. Lin, S. Shiraishi, J. Shen, and T. Bas¸ar, "Information- driven autonomous intersection control via incentive compatible mechanisms," IEEE Trans. Intell. Trans. Sys., no. 99, pp. 1 -- 13, 2018. [17] N. Nisan, T. Roughgarden, E. Tardos, and V. V. Vazirani, Algorithmic game theory. Cambridge University Press, 2007. [18] D. Bergemann and M. Said, "Dynamic auctions," Wiley Encycl. Operations Research and Management Science, 2010. [19] V. Vishnumurthy, S. Chandrakumar, and E. G. Sirer, "Karma: A secure economic framework for peer-to-peer resource sharing," in Workshop on Economics of Peer-to-peer Systems, vol. 35, no. 6, 2003. [20] F. D. Garcia and J.-H. Hoepman, "Off-line karma: A decentralized currency for static peer-to-peer and grid networks," in 5th International Networking Conference (INC05), 2004, pp. 325 -- 332. [21] W. H. Sandholm, Population Games and Evolutionary Dynamics, ser. Economic Learning and Social Evolution. MIT Press, 2010. REFERENCES [1] S. Samaranayake, K. Spieser, H. Guntha, and E. Frazzoli, "Ridepooling with trip-chaining in a shared-vehicle mobility-on-demand system," in 20th IEEE Intell. Transp. Syst. Conf., 2017, pp. 1 -- 7. [2] C. Ruch, S. Horl, and E. Frazzoli, "Amodeus, a simulation-based testbed for autonomous mobility-on-demand systems," in 21st IEEE Intell. Transp. Syst. Conf., 2018, pp. 3639 -- 3644. [3] R. Kurzban, M. N. Burton-Chellew, and S. A. West, "The evolution of altruism in humans," Annual Review of Psychology, vol. 66, no. 1, pp. 575 -- 599, 2015, pMID: 25061670. [4] E. Ostrom, Governing the commons. Cambridge University Press, 2015. [5] G. Wood et al., "Ethereum: A secure decentralised generalised transaction ledger," Ethereum project yellow paper, 2014. [6] D. I. Robertson, "Transyt: a traffic network study tool," Tech. rep., Rep. LR 253, 1969. [7] A. G. Sims and K. W. Dobinson, "The Sydney Coordinated Adaptive Traffic (SCAT) system philosophy and benefits," IEEE Trans. vehicular tech., vol. 29, no. 2, pp. 130 -- 137, 1980. [8] L. Chen and C. Englund, "Cooperative intersection management: A survey," IEEE Trans. Intell. Trans. Sys., vol. 17, no. 2, 2016. [9] H. Schepperle, K. Bohm, and S. Forster, "Towards valuation-aware agent-based traffic control," in Int. Joint Conf. Auton. Agents, 2007. [10] D. Carlino, S. D. Boyles, and P. Stone, "Auction-based autonomous intersection management," in 16th IEEE Intell. Transp. Syst. Conf., 2013, pp. 529 -- 534. [11] M. W. Levin and S. D. Boyles, "Intersection auctions and reservation- based control in dynamic traffic assignment," Transp. Res. Rec.: J. Transp. Res. Board, no. 2497, pp. 35 -- 44, 2015. [12] M. Vasirani and S. Ossowski, "A market-inspired approach for inter- section management in urban road traffic networks," J. Artif. Intell. Res., vol. 43, pp. 621 -- 659, 2012. [13] M. Mashayekhi and G. List, "A multi-agent auction-based approach for modeling of signalized intersections," in Workshop Synergies Between Multiagent Syst., Mach. Learn. Complex Syst., 2015.
1901.10564	2	1901	2019-02-04T21:46:45	Directed Formation Control of n Planar Agents with Distance and Area Constraints	[ "cs.MA", "eess.SY" ]	In this paper, we take a first step towards generalizing a recently proposed method for dealing with the problem of convergence to incorrect equilibrium points of distance-based formation controllers. Specifically, we introduce a distance and area-based scheme for the formation control of $n$-agent systems in two dimensions using directed graphs and the single-integrator model. We show that under certain conditions on the edge lengths of the triangulated desired formation, the control ensures almost-global convergence to the correct formation.	cs.MA	cs	Directed Formation Control of n Planar Agents with Distance and Area Constraints Tairan Liu, Marcio de Queiroz, Pengpeng Zhang and Milad Khaledyan Abstract -- In this paper, we take a first step towards general- izing a recently proposed method for dealing with the problem of convergence to incorrect equilibrium points of distance-based formation controllers. Specifically, we introduce a distance and area-based scheme for the formation control of n-agent systems in two dimensions using directed graphs and the single- integrator model. We show that under certain conditions on the edge lengths of the triangulated desired formation, the control ensures almost-global convergence to the correct formation. I. INTRODUCTION Formation control is an important problem in multi-agent coordination and cooperation where the objective is for agents to form a prescribed geometric shape in space. This requirement is intrinsic to tasks such as area coverage, perimeter protection, and co-transportation of large objects. One of two methods are typically used in formation control: i) regulate the relative position of certain agent pairs to prescribed values [1], [2], or ii) regulate a set of inter- agent distances (magnitude of the relative position vector) to prescribed values [3], [4]. The first method requires the agents to have a common global coordinate frame or that their local coordinate frames be aligned which may not be feasible in practice. On the other hand, the feedback variables in the second method can be calculated in each agent's local coordinate frame, which do not have to be aligned with a global coordinate frame or with each other. As a result, the desired formation is, at best, only acquired up to translation and rotation; i.e., the agents can converge to any formation that is isomorphic to the desired one. An important consideration in the distance-based method is how to prevent agents from converging to a formation that is equivalent but noncongruent to the desired one (see Section II-A for the formal definitions of equivalency and congruency). Such formations are undesirable because they do not have the same shape or orientation as the prescribed formation, although they satisfy the set of distance con- straints. In other words, the distance constraints do not uniquely define the relative positions of the agents and lead to positional ambiguities [5]. Rigid graph theory provides a partial solution to this problem by requiring the formation graph to be rigid [6], [7]. Specifically, imposing a minimum number of distances to be controlled reduces the undesirable Tairan Liu, Marcio de Queiroz, and Pengpeng Zhang are with the Department of Mechanical and Industrial Engineering, Louisiana State Uni- versity, Baton Rouge, LA 70803, USA (Email: [email protected]; [email protected]; [email protected]). Milad Khaledyan is with the Department of Electrical and Computer Engineering, Uni- versity of New Mexico, Albuquerque, NM 87131, USA (Email: [email protected]). "equilibrium points" to formations that are flipped/reflected versions of the desired one [5]. Then, the determining factor whether convergence is to a congruent formation or a flipped formation is the initial condition of the rigid formation. That is, rigidity distance-based formation controllers only have local stability properties. A few approaches have been recently proposed to address the aforementioned issues with distance-based controllers. In [8], a combination of inter-agent distance and angular con- straints was used to reduce the likelihood of convergence to noncongruent formations in two dimensions (2D). Although the region of attraction of the desired equilibrium can be somewhat enlarged by a proper choice of control gains, the stability of the control proposed in [8] is still local in nature. An extension of this work to 3D appeared in [9] by using area and volume constraints. The control method avoids flipped formations but introduces other undesired equilibrium points due to the multiple local minima of the proposed potential function. A related approach was introduced in [10] for the single-integrator agent model where the signed area of a triangle was employed as a controlled variable to prevent flipped formations. That is, the sign of the area enclosed by the formation along with the inter-agent distances were used to uniquely define the correct formation up to translation and rotation. The formation control law in [10] was based on the gradient of a potential function that incorporates distance error and signed area error terms and on the use of undirected graphs (i.e., bidirectional sensing and control). Convergence analyses were conducted for special cases of 3- and 4-agent planar formations. The purpose of this paper is to explore the approach intro- duced in [10] further. Specifically, we aim to generalize the approach to systems of n agents while introducing explicit, sufficient conditions for convergence to the 2D desired for- mation. The key to our solution is triangulating the directed formation graph to facilitate the use of interconnected system theory. The use of a directed graph has the added benefit of leading to a unidirectional formation controller. Under our solution, mild conditions are imposed on the edge lengths of the interconnected triangles, and the overall formation graph is required to be a Leader-First-Follower (LFF) type of minimally persistent directed graph [11]. We show that our gradient-type control law ensures convergence to the desired formation as long as the leader and first follower are not initially collocated. That is, no restrictions are placed on the initial conditions of the ordinary followers. The closed-loop system is proven to have an almost-global asymptotic equi- librium point corresponding to the desired formation. Thus, the collinear invariant set and flipped formation problems are hereby solved by the proposed control scheme. We note that our result is not a straightforward extension of [10] since the interconnected triangles form coupled nonlinear subsystems, which complicates the stability analysis of the overall system. II. BACKGROUND MATERIAL A. Undirected Graphs An undirected graph G is represented by a pair (V, E), where V = {1, 2, ..., n} is the set of vertices and E = {(i, j) i, j ∈ V, i 6= j} ⊂ V × V is a set of undirected edges. The total number of edges in E is denoted by a ∈ {1, ..., n(n − 1)/2}. The set of neighbors of vertex i ∈ V is represented by (1) Ni(E) = {j ∈ V (i, j) ∈ E}. If p = [p1, ..., pn] ∈ R2n where pi ∈ R2 is the coordinate of the ith vertex, then a framework F is defined as the pair (G, p). 1 2 R(p) = The edge function γ : R2n → Ra is defined as γ(p) = [...,pi − pj2, ...], (i, j) ∈ E (2) such that its mth component, pi − pj, relates to the mth edge of E connecting the ith and jth vertices. The rigidity matrix R : R2n → Ra×2n is given by ∂γ(p) ∂p (3) where we have that rank[R(p)] ≤ 2n − 3 [6]. Frameworks (G, p) and (G, p) are equivalent if γ(p) = γ(p), and are congruent if pi − pj = pi − pj, ∀i, j ∈ V [12]. w − z = T (w) − T (z), ∀w, z ∈ R2. An isometry of R2 is a map T : R2 → R2 satisfying [7] (4) This map includes rotation and translation of the vector w−z. Two frameworks are isomorphic if they are correlated via an isometry. It is obvious that (2) is invariant under isomorphic motions of the framework. A framework F = (G, p) is rigid in R2 if all of its motions satisfy pi(t) = T (pi), ∀i ∈ V and ∀t ∈ [0, 1]; i.e., the family of frameworks F (t) is isomorphic [6], [7]. Some related notions of rigidity are the following. A generic framework (G, p) is infinitesimally rigid if and only if rank[R(p)] = 2n − 3 [7]. A rigid framework is said to be minimally rigid if and only if a = 2n − 3 [5]. If the infinitesimally rigid frameworks (G, p) and (G, p) are equivalent but not congruent, then they are referred to as ambiguous [5] since the edge function cannot uniquely define the framework. Common types of ambiguities are shown in Figure 1. Note that reflected frameworks are an extreme form of flip ambiguity where more than one vertex is flipped. In fact, reflections are the only form of flip ambiguity that can occur in a triangular framework. B. Directed Graphs A directed graph G is a pair (V, Ed) where the edge set Ed is directed in the sense that if (i, j) ∈ Ed then i is the source vertex of the edge and j is the sink vertex. For i ∈ V , the out-degree of i (denoted by out(i)) is the number of edges in Ed whose source is vertex i and sinks are in V − {i}. Fig. 1. Types of ambiguous frameworks. For directed graphs, the notion of rigidity defined in Sec- tion II-A is not enough to maintain the formation structure (see [13] for an example), and two additional concepts are needed. The first one is the notion of a constraint consistent graph. As explained in [13], the intuitive meaning of constraint consistency is that every agent is able to satisfy all its distance constraints when all the others are trying to do the same. A sufficient condition for a directed graph (cid:0)V, Ed(cid:1) in R2 to be constraint consistent is that out(i) ≤ 2 for all i ∈ V (see Lemma 5 of [14]). The second concept is graph persistency, which has the meaning that, provided all agents are trying to satisfy their distance constraints, the structure of the agent formation is preserved [13]. A directed graph is persistent if and only if it is constraint consistent and its underlying undirected graph is infinitesimally rigid (see Theorem 3 of [14]). A persistent graph in R2 is said to be minimally persistent if no single edge can be removed without losing persistence. A necessary condition for a persistent graph in R2 to be minimally persistent is out(i) ≤ 2 for all i ∈ V , while a sufficient condition is minimal rigidity [14]. Starting from two vertices with an edge, a minimally persistent (resp., rigid) graph can be constructed by the Henneberg insertion of type I [15], i.e., iteratively adding a vertex with two outgoing (resp., undirected) edges. Henceforth, we refer to a graph constructed in this manner as a Henneberg graph. C. Signed Area The signed area of a triangular framework, S : R6 → R, is defined as [10] (p3 − p1)⊺ J (p3 − p2) (5) (6) det(cid:20) 1 p1 1 p2 S(p) = 1 2 where 1 1 2 p3(cid:21) = −1 0 (cid:21) . 0 1 J =(cid:20) This quantity is positive (resp., negative) if the vertices are ordered counterclockwise (resp., clockwise). Further, (5) is zero if any two vertices are collocated or the three vertices are collinear. A Henneberg framework can be divided into triangular sub-frameworks. Therefore, the signed area of a Henneberg framework with n vertices and directed edge set Ed, χ : R2n → Rn−2, is defined as det(cid:20) 1 χ(p) =(cid:20)..., ∀(k, i), (k, j) ∈ Ed − {(2, 1)} pk(cid:21) , ...(cid:21) , 1 pj 1 2 (7) pi 1 such that its mth component is related to the signed area of the mth triangle constructed with vertices i, j, and k. For example, the signed area of the framework in Figure 2a is given by the equation has no real solution if Λ > 0 and P > 0, or Λ > 0 and D > 0. Corollary 1: Consider the equation ax4 + bx3 + cx2 + dx + e = 0 (10) (p3 − p1)⊺ J (p3 − p2) , χ(p) =h 1 (p4 − p2)⊺ J (p4 − p3) , 1 2 1 2 2 (p5 − p3)⊺ J (p5 − p4)i, (8) where the three elements of χ(p) are positive, negative, and positive, respectively. For the framework in Figure 2b, it would be (p3 − p1)⊺ J (p3 − p2) , χ(p) =h 1 (p4 − p1)⊺ J (p4 − p2) , 1 2 1 2 2 (p5 − p3)⊺ J (p5 − p4)i. (9) where 1 2 1δ2 δ4 1γ 3 + δ2 2 − δ4 δ2 1 + δ2 a = − 2δ2 1(γ − 2)2 b =δ1(γ 2 − 4)q2δ2 1(cid:18) 3 c = − 2 1 1γ − 3δ2 1 − 2δ2 4 3 − δ4 1δ2 1 + 2δ2 1 + 2δ2 2 − δ4 δ2 1γ 3(2δ2 δ1γ 2(2δ2 d = 1δ2 1δ2 1 + 2δ2 2δ2 1δ2 2 + δ2 2 + 2δ2 3 − δ4 3(cid:19) γ 2 − 4δ2 1(cid:0)δ2 2 − 2δ2 3)× 2δ2 2 + 2δ2 3 3 3 − δ4 3(cid:1) γ 2 + δ2 q2δ2 e = − γ is a positive constant, and δ1, δ2, δ3 are the lengths of the edges of a triangle. If δ2 6= δ3 and 3 − δ4 2 + 2δ2 3 − δ4 3), 3 − δ4 2 − δ4 2δ2 1δ2 1 8 (11) (cid:12)(cid:12)(cid:12)(cid:12) δ2 3 − δ2 δ2 2 1 < 2√2, (cid:12)(cid:12)(cid:12)(cid:12) then there exists a γ > 0 such that (10) has no real solution for γ > max{γ, 2}. Proof: See Appendix VI-B. Fig. 2. Signed area examples. We introduce next is an extension of the concept of congruency that includes the signed area. Definition 1: Henneberg frameworks F = (G, p) and F = (G, p) where G = (V, E) are said to be strongly congruent if they are congruent and χ(p) = χ(p). We represent the set of all frameworks that are strongly congruent to F by SCgt(F ). It is obvious that frameworks that are congruent but not strongly congruent are reflected frameworks. Note that if F is a reflected version of F , then χ(p) = −χ(p). In summary, the signed area function will be used to rule out the the occurrence of framework ambiguities, especially reflections. Lemma 1: Henneberg frameworks F = (G, p) and F = (G, p) are strongly congruent if and only if they are equiv- alent and χ(p) = χ(p). Proof: See Appendix VI-A. D. Quartic Polynomials Lemma 2: [16], [17], [18], [19] For any quartic polyno- mial equation ax4 + bx3 + cx2 + dx + e = 0 where a 6= 0, Λ =256a3e3 − 192a2bde2 − 128a2c2e2 + 144a2cd2e − 27a2d4 + 144ab2ce2 − 6ab2d2e − 80abc2de + 18abcd3 + 16ac4e − 4ac3d2 − 27b4e2 + 18b3cde − 4b3d3 − 4b2c3e + b2c2d2, P =8ac − 3b2, D =64a3e − 16a2c2 + 16ab2c − 16a2bd − 3b4, Remark 1: The geometric meaning of condition (11) is discussed in Remark 3. Although the existence of the lower bound γ in Corollary 1 is guaranteed, a closed-form expres- sion for γ does not exist in general. However, γ can be easily determined by numerical means once δ1, δ2, and δ3 are selected. E. Stability Results Lemma 3: [20] Consider the system x = f (x, u) where x is the state, u is the control input, and f (x, u) is locally Lipschitz in (x, u) in some neighborhood of (x = 0, u = 0). Then, the system is locally input-to-state stable if and only if the unforced system x = f (x, 0) has a locally asymptotically stable equilibrium point at the origin. Lemma 4: [21] Consider the interconnected system Σ1: Σ2: x = f (x, y) y = g(y). (12) If subsystem Σ1 with input y is locally input-to-state stable and y = 0 is a locally asymptotically stable equilibrium point of subsystem Σ2, then [x, y] = 0 is a locally asymptotically stable equilibrium point of the interconnected system. III. PROBLEM STATEMENT pi = ui, i = 1, ..., N the kinematic equation Consider a system of N ≥ 2 mobile agents governed by (13) where pi ∈ R2 is the position of the ith agent relative to an Earth-fixed coordinate frame, and ui ∈ R2 is the velocity- level control input. The desired formation of agents is modeled by the di- rected framework F ∗ = (G∗, p∗) where G∗ = (V ∗, E ∗), i ∈ R2 denotes the dim(E ∗) = a, p∗ = [p∗ desired position of the ith agent. The fixed desired distance separating the ith and jth agents is defined as i, j ∈ V ∗ dji = p∗ i > 0, N ], and p∗ j − p∗ 1, ..., p∗ (14) We assume F ∗ is constructed to satisfy the following con- ditions: 3. Condition 1 out(1) = 0, out(2) = 1, and out(i) = 2,∀i ≥ Condition 2 If there is an edge between agents i and j, the direction must be i ← j if i < j. The above conditions imply that F ∗ should be a LFF- type minimally persistent formation [11], where agent 1 is the leader, agent 2 is the first follower, and agents i for i ≥ 3 are called ordinary followers. The actual formation of agents is modeled by F (t) = (G∗, p(t)) where p = [p1, ..., pN ]. Note that F and F ∗ share the same directed graph, which remains unchanged for all time. The physical meaning of (j, i) ∈ E ∗ in the actual formation is that agent j can measure its relative position to agent i, pj − pi, but not vice versa. The control objective of this paper is to ensure we propose the following control law u1 = 0 u2 = −α2z21 p21 uk = −αk (zki pki + zkj pkj ) − βk SijkJ ⊺ (pki − pkj) (22a) (22b) (22c) for 2 < k ≤ N, i < j < k, and (k, i), (k, j) ∈ E ∗. The control law is only a function of pki, pkj , dki, dkj and dji for i, j ∈ Nk(E ∗). Thus, the control law is distributed since it only requires the kth agent to measure its relative position to neighboring agents in the directed graph. The following theorem states our main result. Theorem 1: Let the initial conditions of the formation F (t) = (G∗, p(t)) be such that p1(0) 6= p2(0), and let F ∗ satisfy (15) (16a) (16b) which is equivalent to saying F (t) → SCgt(F ∗) as t → ∞, pj(t) − pi(t) → dji as t → ∞, i, j ∈ V ∗ and χ(p(t)) → χ(p∗) as t → ∞. The control objective will be quantified by two types of error variables. If the relative position of two agents is defined as pji = pj − pi, the distance error is given by [4] zji = pji2 − d2 ji. (17) The stacked vector of all distances errors is defined as z = [z21, ..., zji, ...],∀(j, i) ∈ E ∗. The area error is defined as [10] (18) ijk = S(p∗) k]. The stacked vector of all area errors ijk, (k, i), (k, j) ∈ E ∗ where Sijk = S(p) with p = [pi, pj, pk] and S ∗ with p∗ = [p∗ is given by S = [ S123, ..., Sijk, ...], ∀(k, i), (k, j) ∈ E ∗. inter-agent distances, a useful formula for calculating S ∗ is given by [22] Sijk = Sijk − S ∗ i , p∗ Since F ∗ is typically specified in terms of the desired j , p∗ ijk S ∗ ijk = ±qdijk (dijk − dji) (dijk − dki) (dijk − dkj ) (19) where dji + dki + dkj dijk = 2 . (20) Note that if the order of agents i, j, k is counterclockwise (resp., clockwise), then (19) takes the positive (resp., nega- tive) sign. IV. CONTROL LAW FORMULATION The control law will be dictated by the choice of potential function associated with the error variables (17) and (18). To this end, we consider the Lyapunov function candidate [10] Vk =  α2 4 αk z 2 21, 4 (cid:16)z 2 ki + z 2 kj(cid:17) + βk S 2 ijk, where αk and βk are positive constants, i < j < k, and (k, i) , (k, j) ∈ E ∗. Based on (21) and its time derivative, if k = 2 if 2 < k ≤ N (21) < 2√2, (23) ki − d2 d2 d2 ji kj (cid:12)(cid:12)(cid:12)(cid:12)(cid:12) (cid:12)(cid:12)(cid:12)(cid:12)(cid:12) for all i, j, k such that 2 < k ≤ N , i < j < k, and (k, i), (k, j) ∈ E ∗. Then, the control (22) with if dki = dkj ji/4 , (24) if dki 6= dkj where γ is determined from Corollary 1, rendershz, Si = 0 asymptotically stable and ensures F (t) → SCgt(F ∗) as t → ∞. kj − d2 d2 d2 ji γ, βk αk >  Proof: The open-loop dynamics for (17) and (18) zji = 2p⊺ ji(uj − ui) (25) (26) are given by and · Sijk = 1 2 [(uk − ui)⊺J pkj + p⊺ kiJ(uk − uj)] , Vk = where (13) was used. Therefore, the time derivative of (21) becomes +zkj p⊺ if k = 2 α2z21 p⊺ 21(u2 − u1), αk(cid:2)zki p⊺ ki(uk − ui) kj(uk − uj)(cid:3) +βk Sijk(cid:2)(uk − ui)⊺J pkj kiJ(uk − uj)(cid:3),   21(cid:1) z21 = −2α2z21p212 = −2α2z21(cid:0)z21 + d2 if 2 < k ≤ N +p⊺ (28) where (17) was used. The solution to the above nonlinear ODE is Step 1: Consider the subsystem composed of agents 1 and 2 only. Substituting (22a) and (22b) into (25) yields (27) d2 21z21(0) d2 21 exp(2d2 it 21α2t) (z21(0) + d2 21) − z21(0) is clear 21,∞(cid:1) where 21 corresponds to agents 1 and 2 being collocated. 21, we can show from (29) that z21(t) > −d2 that z21 ∈ (cid:2)−d2 21 . (29) z21(t) = From (17), z21 = −d2 If z21(0) > −d2 ∀t > 0 as follows: 21 z21(0) d2 21 exp(2d2 z21(t) > −d2 ⇔ ⇔ z21(0) > z21(0) − d2 ⇔ d2 21 exp(2d2 ⇔ z21(0) > −d2 21. 21α2t) (z21(0) + d2 21 exp(2d2 21α2t)(cid:0)z21(0) + d2 > −1 21) − z21(0) 21(cid:1) 21α2t)(cid:0)z21(0) + d2 21(cid:1) > 0 (30) Now, after substituting (22a) and (22b) into (27), we obtain V2 = −α2 2z 2 21p212 ≤ 0. (31) 21 implies p21(t) > 0, we can see that V2 is negative definite 21 (or Since z21(t) > −d2 V2 = 0 only at z21 = 0. Therefore, and z21 = 0 is asymptotically stable for z21(0) > −d2 equivalently, p1(0) 6= p2(0)). Step 2: Consider that a third agent is added to the previous subsystem as shown in Figure 3. We can view this new system as the interconnected system ξ3 = f3(ξ3, Ξ2) Ξ2 = g3(Ξ2) (32b) where ξ3 := [z31, z32, S123] is the state of the error dynamics of agent 3 and Ξ2 = ξ2 := z21. From (35), we get (32a) Fig. 3. Three-agent system. (33) We seek to establish the input-to-state stability of (32a) with respect to input Ξ2 via Lemma 3. When Ξ2 = 0, we can see from (22b) that u2 = 0. Therefore, from (27) with k = 3 under the condition that Ξ2 = 0, we have that Substituting (22c) with k = 3 in (33) gives V3 =hα3 (z31 p31 + z32 p32)⊺ + β3 S123 (p31 − p32)⊺ Ji u3. V3 = −(cid:13)(cid:13)(cid:13)α3 (z31 p31 + z32 p32) + β3 S123J ⊺ (p31 − p32)(cid:13)(cid:13)(cid:13) . (34) If ξ3 = 0 is the only value at which V3 = 0, then (33) is negative definite and (32a) is input-to-state stable. It then follows that the origin of (32), i.e., [Ξ2, ξ3] = 0, is asymptotically stable according to Lemma 4. To this end, note that translational and rotational motions of the triangle will not change the value of V3 since it is a function of the relative position of agents and the triangle area [23], [24]. Thus, without the loss of generality, let p1 = [−d21/2, 0], p2 = [d21/2, 0], and p3 = [x, y] for simplicity. Then, V3 = 0 2 is equivalent to (cid:18)2x2 + 2y2 + (cid:18)2x2 + 2y2 + 21 21 and d2 2 − d2 d21 + 31(cid:19) x 32 − d2 2 (cid:0)2d21x − d2 31(cid:19) y d2 2 − d2 32 − d2 α3 (cid:18) d2 β3 y − d21S ∗ 32(cid:1) / (2d21) + 2 21 31 + d2 32(cid:1) = 0 (35) 123(cid:19) = 0. One solution to (35) is x =(cid:0)d2 31 − d2 and y = 2S ∗ (36) which corresponds to ξ3 = 0. We will show next that β3/α3 can be selected such that this is the only solution to (35). This proof will be conducted for two distinct cases: an isosceles triangle and the non-isosceles case. 123/d21, (Case 2a) Consider that the triangle is such that d32 = d31. (cid:18)2x2 + 2y2 + (cid:18)2x2 + 2y2 + 21 3d2 2 − 2d2 1 2 21 − 2d2 d2 d2 β3 2α3 + and 32(cid:19) x = 0 32(cid:19) y 21(cid:18)y − 1 2q4d2 32 − d2 21(cid:19) = 0. (37) The first equation of (37) implies x = 0 or x2 + y2 = 32 − 3 d2 4 d2 21. Substituting x = 0 into the second equation of (37) yields (cid:18)y − 1 21(cid:19)× 2q4d2 32 − d2 (cid:18)8y2 + 4q4d2 32 − d2 21y + 2β3 α3 d2 21(cid:19) = 0. d2 21 is less than 0. That is, inequality (39) will lead to x = 0 and y = 21y + 2β3 α3 It is easy to show that when , 21 21 > 4 d2 β3 α3 32 − 1 d2 d2 the discriminant of 8y2 + 4p4d2 32 − d2 2p4d2 3 4 into the second equation of (37) gives Now, substituting x2 + y2 = d2 32 − d2 d2 1 21 21 being the only solution to (38). 32 − α3q4d2 β3 32 − d2 21. α3 − 2(cid:19) = After squaring (41) and using (40) again to eliminate y, (38) (39) (40) (41) 2y(cid:18) β3 α3 − 2(cid:19)2 α3(cid:19)2 21(cid:18) β3 = we obtain 2x2(cid:18) β3 −"d2 21(cid:1) β3 α3 +(cid:0)8d2 32 − 6d2 {z φ(β3/α3) + 6d2 . 32# 21 − 8d2 } (42) Since the left hand side of (42) is nonnegative for any β3/α3, this equation has no solution if φ(β3/α3) > 0. This means that φ(β3/α3) should have no real roots, or β3/α3 should be chosen to be greater (resp., smaller) than the largest (resp., smallest) root. The discriminant of φ(β3/α3) is given by 3 d2 If 4 32 < d2 21 < 4d2 and φ(β3/α3) > 0. 21(cid:1)(cid:0)4d2 21(cid:1) . 32 − d2 32 − 3d2 (43) 32, then ∆ < 0 for any β3/α3 > 0 ∆ = 4(cid:0)4d2 32, then ∆ ≥ 0 and φ(β3/α3) has real roots. Since the smallest root is less than zero, the only option for ensuring φ(β3/α3) > 0 is to choose β3/α3 greater than the largest root, i.e., 21 ≤ 4 If d2 3 d2 > 3d2 21 − 4d2 32 − 3d2 β3 d2 α3 21 ≥ 4d2 Note that the case where d2 32 +p(4d2 21 21)(4d2 32 − d2 21) . (44) 32 is not possible since it contradicts the fact that d32 + d31 = 2d32 > d21. Combining the three cases, we see that (41) will have no solution if (44) holds. Finally, it is not difficult to show that (39) is a sufficient condition for (44). Therefore, for the isosceles triangle, the condition for ξ3 = 0 to be the only value where V3 = 0 is given by (39). (46a) 21d2 d4 21d2 2S ∗ 123 where 32 + 2d2 32d2 21 + 2d2 31 − d4 the triangle is not (Case 2b) Consider that isosceles (d32 6= d31). After substituting the first equation of (35) into the second one, eliminating x, and factoring the resulting polynomial of y, we obtain (cid:18)y − d21 (cid:19)(cid:0)c4y4 + c3y3 + c2y2 + c1y + c0(cid:1) = 0 (45) 21(cid:18) β3 α3 − 2(cid:19)2 c4 = − 2d2 − 4#× c3 =d21"(cid:18) β3 α3(cid:19)2 q2d2 32 − d4 α3(cid:19)3 21(cid:18) β3 21(cid:18) 3 1 c2 = − 2 31(cid:1) β3 21(cid:0)d2 − 4d2 α3 d21(cid:18) β3 α3(cid:19)2(cid:18)2d2 1 β3 α3 − 3d2 4 q2d2 32 − d4 21d2 21(cid:18) β3 α3(cid:19)3 c0 = − (cid:0)2d2 32 − d4 (46e) Note that the quartic polynomial in (45) is similar to (10). 31 − d4 α3(cid:19)2 31(cid:19)(cid:18) β3 31(cid:19)× 31 − d4 31(cid:1) . 31 − d4 32 − 2d2 32d2 1 8 21d2 21 − 2d2 31 − d4 31 − d4 d2 21 + d2 21 + 2d2 32 + 2d2 31 (46b) 31 (46d) 32 + 2d2 21 + 2d2 32 + d2 32 + d2 + d2 21d2 32d2 21d2 c1 = 21 d2 × (46c) 2 Thus, by Corollary 1, if (47) and β3/α3 > max{γ, 2} (see proof of Corollary 1 for detail 21 d2 31 − d2 d2 32 < 2√2 (cid:12)(cid:12)(cid:12)(cid:12) (cid:12)(cid:12)(cid:12)(cid:12) of γ), the quartic polynomial has no real solution, and y = 2S ∗ 123/d21 is the only solution to (45). Step k: The process of adding a vertex k with two outgoing edges to any two distinct vertices i and j of the previous graph can be followed one step at a time, resulting at each step in the interconnected system ξk = fk(ξk, Ξk−1) Ξk−1 = gk(Ξk−1) (48a) (48b) where ξk := [zki, zkj, Sijk], (k, i), (k, j) ∈ E ∗ is the state of error dynamics of the kth agent and Ξk−1 := [ξ2, ..., ξk−1]. Note that the asymptotic stability of Ξk−1 = 0 for (48b) was already established in Step k − 1. Therefore, we only need to check the input-to-state stability of (48a) with respect to input Ξk−1. To this end, when Ξk−1 = 0, (27) becomes Vk =hαk (zki pki + zkj pkj)⊺ + βk Sijk (pki − pkj )⊺ Ji uk. Vk = −(cid:13)(cid:13)(cid:13)αk (zki pki + zkj pkj ) + βk SijkJ ⊺ (pki − pkj)(cid:13)(cid:13)(cid:13) Now, substituting (22c) into (49) gives Similar to Step 2, we can show that if the gain ratio βk/αk is selected according to (24) and the edges of triangle ∆ijk satisfy (23), then (50) is negative definite. As a result, (48a) is input-to-state stable and [Ξk−1, ξk] = 0 in (48) is asymptotically stable by Lemma 4. . (50) (49) 2 Repeating this process until k = N leads to the conclusion that [ξ2, ..., ξN ] = 0 is asymptotically stable, which implies z(t) → 0 and χ(p(t)) → χ(p∗) as t → ∞. Given that F ∗ and F (t) have the same edge set and F ∗ is minimally persistent by design, then we have that F (t) → SCgt(F ∗) as t → ∞ from Lemma 1. Remark 2: Theorem 1 only requires that the leader and first follower not be collocated at t = 0. If agents 1 and 2 were initialized at the same position, then u1 = u2 = 0 and they would remain at this position forever. In other words, the condition p1 = p2 is an invariant set. As for the ordinary followers, (22) guarantees formation acquisition regardless of their initial conditions. For example, if agents 2, 3, 4, and 5 in Figure 2 are all initially collocated, then u4 = u5 = 0 at t = 0 which means agents 4 and 5 will not move at first. However, u3 6= 0, so agent 3 will move. This results in u4 6= 0, causing agent 4 to move, and finally u5 becomes nonzero, so agent 5 moves. Remark 3: Condition (23) on the desired formation has the following geometric interpretation. Consider the three i = [−dji/2, 0], vertices in Figure 4 where, for simplicity, p∗ p∗ j = [dji/2, 0], and p∗ k = [x, y]. 1 1Translation and rotation of these vertices as a rigid body will not affect the following analysis since it is only dependent on their distances. kj (cid:12)(cid:12)(cid:12)(cid:12)(cid:12) Given that d2 ki − d2 d2 ji < 2√2 (cid:12)(cid:12)(cid:12)(cid:12)(cid:12) ⇐⇒ (cid:12)(cid:12)(cid:12)(x + dji/2)2 + y2 − (x − dji/2)2 − y2(cid:12)(cid:12)(cid:12) < 2√2 ⇐⇒ 2xdji ⇐⇒ x < √2dji, d2 ji d2 ji < 2√2 (51) any point p∗ k inside the shaded region in Figure 4 satisfies (23). It is important to point out that (23) is sufficient but not necessary for stability. For example, consider a triangular formation with d21 = 1, d31 = 2.1, and d32 = 3, which does not satisfy (23). If however β3/α3 is selected in the range (10.42, 13.55), the stability result of Theorem 1 will hold. In fact, the gain ratio βk/αk and (24) impose a lower bound on the relative weight of the distance error and area error in the potential function (21) in order to guarantee stability. Fig. 4. Geometeric interpretation of (23). Remark 4: Mathematically, the role of the area-based term βk S 2 ijk is to guarantee the existence of a unique minimum for the potential function (21) in the Euclidean plane, and thus avoid the system from converging to an unde- sirable local minima. To illustrate this, consider a triangular formation where p1 = [−1, 0], p2 = [1, 0], p3 = [x, y], and d21 = d31 = d32 = 2, and let W = 1 32) be the potential function with only the distance error terms of (21) with k = 3. In Figure 5, we plot ln(W + 1) and ln(V3 + 1) versus p3 to have a better view of their minima.2 We can clearly see that W (p3) has two minima, corresponding to the desired position for agent 3 and its reflected position, whereas V3(p3) has a unique minimum. 31 + z 2 4 (z 2 V. CONCLUSIONS This paper presented a 2D formation control scheme that uses distance and signed area information to guarantee convergence to the desired formation shape. The asymptotic convergence result is valid under mild conditions on the 2Since functions ln(V + 1) and V are positively correlated, this variable change does not affect the function extrema. Fig. 5. Potential function V3(q3) and corresponding counter plot. a) Potential function W (q3) and corresponding counter plot; b) edge lengths of the triangulated-like framework and when the leader agent and the first follower are not collocated at time zero. The scheme is applicable to systems with any number of agents governed by the single-integrator model. VI. APPENDIX A. Proof of Lemma 1 (Proof of ⇒) If F and F are strongly congruent, then pi − pj = pi − pj, ∀i, j ∈ V and χ(p) = χ(p) by definition. Therefore, since E ⊂ V ×V , we know pi−pj = pi − pj, ∀(i, j) ∈ E, i.e., F and F are equivalent. (Proof of ⇐) If dim(V ) = 3, then framework equivalency and congruency are equivalent, so the conditions for strong congruency are trivially satisfied. If a vertex is added such that dim(V ) = 4, the result- ing framework would have two additional edges and one additional triangle. Consider without loss of generality the framework in Figure 6(a), where the area of the quadrilateral is given by SQ := S123 − S234. Since χ(p) = χ(p), we know that SQ(p) = SQ(p), so it follows from the general quadrilateral area formula [22] that 2 2 1 1 = − kp3 − p1k2 − kp4 − p2k2(cid:1)2i 1 4h4 kp3 − p2k2 kp4 − p1k2 −(cid:0)kp2 − p1k2 + kp4 − p3k2 4h4 kp3 − p2k2 kp4 − p1k2 −(cid:0)kp2 − p1k2 + kp4 − p3k2 Since F and F are equivalent, kp2 − p1k = kp2 − p1k, kp3 − p1k = kp3 − p1k, kp3 − p2k = kp3 − p2k, kp4 − p2k = kp4 − p2k, and kp4 − p3k = kp4 − p3k. Therefore, we have from (52) that kp4 − p1k = kp4 − p1k, so F and F are strongly congruent for dim(V ) = 4. Since the quadrilateral signed area formula in (52) applies to both convex and concave quadrilaterals, a similar analysis exists for all other cases, some of which are shown in Figure 6. − kp3 − p1k2 − kp4 − p2k2(cid:1)2i 1 (52) . REFERENCES [1] R. Olfati-Saber and R. M. Murray, "Consensus problems in networks of agents with switching topology and time-delays," IEEE Transac- tions on Automatic Control, vol. 49, no. 9, pp. 1520 -- 1533, 2004. [2] W. Ren and R. W. Beard, Distributed consensus in multi-vehicle cooperative control. Springer-Verlag London, 2008. [3] F. Dorfler and B. Francis, "Geometric analysis of the formation problem for autonomous robots," IEEE Transactions on Automatic Control, vol. 55, no. 10, pp. 2379 -- 2384, 2010. [4] L. Krick, M. E. Broucke, and B. A. Francis, "Stabilisation of infinitesi- mally rigid formations of multi-robot networks," International Journal of Control, vol. 82, no. 3, pp. 423 -- 439, 2009. [5] B. D. O. Anderson, C. Yu, B. Fidan, and J. M. Hendrickx, "Rigid graph control architectures for autonomous formations," IEEE Control Systems, vol. 28, no. 6, 2008. [6] L. Asimow and B. Roth, "The rigidity of graphs, II," Journal of Mathematical Analysis and Applications, vol. 68, no. 1, pp. 171 -- 190, 1979. [7] I. Izmestiev, "Infinitesimal rigidity of frameworks and surfaces," Lectures on Infinitesimal Rigidity, Kyushu University, Japan, 2009. [8] E. D. Ferreira-Vazquez, E. G. Hernandez-Martinez, J. J. Flores- Godoy, G. Fernandez-Anaya, and P. Paniagua-Contro, "Distance-based formation control using angular information between robots," Journal of Intelligent & Robotic Systems, vol. 83, no. 3-4, pp. 543 -- 560, 2016. [9] E. D. Ferreira-Vazquez, J. J. Flores-Godoy, E. G. Hernandez-Martinez, and G. Fernandez-Anaya, "Adaptive control of distance-based spatial formations with planar and volume restrictions," in Control Applica- tions (CCA), 2016 IEEE Conference on. IEEE, 2016, pp. 905 -- 910. [10] B. D. O. Anderson, Z. Sun, T. Sugie, S. Azuma, and K. Sakurama, "Formation shape control with distance and area constraints," IFAC Journal of Systems and Control, vol. 1, pp. 2 -- 12, 2017. [11] T. H. Summers, C. Yu, S. Dasgupta, and B. D. O. Anderson, "Control of minimally persistent leader-remote-follower and coleader forma- tions in the plane," IEEE Transactions on Automatic Control, vol. 56, no. 12, pp. 2778 -- 2792, 2011. [12] B. Jackson, "Notes on the rigidity of graphs," in Levico Conference Notes, vol. 4, 2007. [13] J. M. Hendrickx, B. D. O. Anderson, J. Delvenne, and V. D. Blon- del, "Directed graphs for the analysis of rigidity and persistence in autonomous agent systems," International Journal of Robust and Nonlinear Control, vol. 17, no. 10-11, pp. 960 -- 981, 2007. [14] C. Yu, J. M. Hendrickx, B. Fidan, B. D. O. Anderson, and V. D. Blondel, "Three and higher dimensional autonomous formations: Rigidity, persistence and structural persistence," Automatica, vol. 43, no. 3, pp. 387 -- 402, 2007. [15] S. Bereg, "Certifying and constructing minimally rigid graphs in the plane," in Proceedings of the Twenty-first Annual Symposium on Computational Geometry. ACM, 2005, pp. 73 -- 80. [16] I. N. Stewart, Galois theory, 4th ed. CRC Press, 2015. [17] L. E. Dickson, Elementary theory of equations. J. Wiley & Sons, Inc., 1914. [18] E. L. Rees, "Graphical discussion of the roots of a quartic equation," The American Mathematical Monthly, vol. 29, no. 2, pp. 51 -- 55, 1922. [19] D. Lazard, "Quantifier elimination: optimal solution for two classical examples," Journal of Symbolic Computation, vol. 5, no. 1-2, pp. 261 -- 266, 1988. [20] H. K. Khalil, Nonlinear control. Pearson Education, Harlow, UK, 2015. [21] H. J. Marquez, Nonlinear control systems: analysis and design. Wiley, Hoboken, NJ., 2003. [22] D. Zwillinger, CRC standard mathematical tables and formulae, 31st ed. CRC Press, 2002. [23] K. Sakurama, "Distributed control of networked multi-agent systems for formation with freedom of special euclidean group," in Decision and Control (CDC), 2016 IEEE 55th Conference on. IEEE, 2016, pp. 928 -- 932. [24] Z. Sun, B. D. O. Anderson, M. Deghat, and H. Ahn, "Rigid forma- tion control of double-integrator systems," International Journal of Control, vol. 90, no. 7, pp. 1403 -- 1419, 2017. Fig. 6. Convex and concave quadrilaterals with five edges. As more vertices are added, each additional vertex will create a quadrilateral, so above process can be repeated to show that F and F are strongly congruent for dim(V ) = n. B. Proof of Corollary 1 Based on Lemma 2, (10) has no real solution if the following quantities are positive: δ12 2 − δ2 16 2 + δ2 1 (cid:0)δ2 2 − δ2 3(cid:1) +(cid:0)δ2 Λ =γ 6 (γ − 2)2h−1 3(cid:1)2(cid:17) × 3(cid:1)2(cid:16)8δ4 1 −(cid:0)δ2 3(cid:1)2(cid:17) γ 7 + f1(δ1, δ2, δ3, γ)i (cid:16)δ4 1(cid:0)δ2 2 − δ2 1 − 2δ2 P = (γ − 2)2(cid:2)8δ6 1γ 3 + f2(δ1, δ2, δ3, γ)(cid:3) (53b) where f1(·) and f2(·) are polynomials in γ of, at most, degree 6 and 2, respectively. Given a polynomial p(γ) = anγn +Pn−1 i=0 aiγi where n ≥ 3 is an odd integer, we can see from Figure 7 that p(γ) = 0 has at least one real root. Denote the largest or the unique real root by γ ∗. Then, (53a) p(γ) > 0 ∀γ > γ ∗, p(γ) < 0 ∀γ > γ ∗, if an > 0 if an < 0. (54) Fig. 7. Odd degree polynomial. Consider that γ 6= 2 in (53a) and (53b). It follows from the conditions δ2 6= δ3 and (cid:12)(cid:12)(cid:0)δ2 1 − 2δ2 δ4 2 − δ2 3) + (δ2 2 + δ2 1(δ2 3 − δ2 3)2 = 1(cid:12)(cid:12) < 2√2 that 2(cid:1) /δ2 1 where γ ∗ (δ1 +δ2 +δ3)(δ1−δ2−δ3)(δ1 +δ2−δ3)(δ1−δ2 +δ3) < 0. Therefore, the coefficient of γ 7 in (53a) is positive, and Λ > 0 for γ > γ ∗ 1 is the lower bound from (54). Likewise, the coefficient of γ 3 in (53b) is positive, and P > 0 for γ > γ ∗ 2 is some lower bound. Thus, the overall sufficient condition for Λ > 0 and P > 0 is given by γ > max{γ, 2} where γ := max{γ ∗ 2 where γ ∗ 2}. 1 , γ ∗
1506.08133	1	1506	2015-06-25T15:32:13	A Study on the Effect of Exit Widths and Crowd Sizes in the Formation of Arch in Clogged Crowds	[ "cs.MA", "physics.soc-ph" ]	The arching phenomenon is an emergent pattern formed by a $c$-sized crowd of intelligent, goal-oriented, autonomous, heterogeneous individuals moving towards a $w$-wide exit along a long $W$-wide corridor, where $W>w$. We collected empirical data from microsimulations to identify the combination effects of~$c$ and~$w$ to the time~$T$ of the onset of and the size~$S$ of the formation of the arch. The arch takes on the form of the perimeter of a half ellipse halved along the minor axis. We measured the~$S$ with respect to the lengths of the major~$M$ and minor~$m$ axes of the ellipse, respectively. The mathematical description of the formation of this phenomenon will be an important information in the design of walkways to control and easily direct the flow of large crowds, especially during panic egress conditions.	cs.MA	cs	A Study on the Effect of Exit Widths and Crowd Sizes in the Formation of Arch in Clogged Crowds Francisco Enrique Vicente G. Castro and Jaderick P. Pabico Institute of Computer Science University of the Philippines Los Baños College 4031, Laguna {fevgcastro,jppabico}@uplb.edu.ph ABSTRACT The arching phenomenon is an emergent pattern formed by a c-sized crowd of intelligent, goal-oriented, autonomous, het- erogeneous individuals moving towards a w-wide exit along a long W -wide corridor, where W > w. We collected empir- ical data from microsimulations to identify the combination effects of c and w to the time T of the onset of and the size S of the formation of the arch. The arch takes on the form of the perimeter of a half ellipse halved along the minor axis. We measured the S with respect to the lengths of the major M and minor m axes of the ellipse, respectively. The mathematical description of the formation of this phe- nomenon will be an important information in the design of walkways to control and easily direct the flow of large crowds, especially during panic egress conditions. INTRODUCTION 1. A crowd is a large group of humans, usually numbering by the hundreds to tens of thousands, who are in the same phys- ical environment, such as a hallway, sharing a common pur- pose, such as exiting through a sole door along the hallway. Although the individuals in the crowd are sharing a common purpose, they may act differently in a crowd than when they are by themselves [17]. The general movement of a crowd is affected by a number of different individuals who are moving independently from each other, but sharing a common envi- ronment, and aiming to reach their respective destinations. The understanding of crowd movement is important to plan- ning and improving shared public places, not only to effec- tively and efficiently facilitate the comfortable movement of individuals, but also to guarantee the individual safety, espe- cially under conditions of danger when quick and orderly evacuation of a mass of individuals is desired. The patterns of behavior of individuals in a crowd is affected directly by the interactions and influences between the indi- viduals and their environment. Because data from real evac- uation is hard to measure and conducting experiments on humans has ethical questions, the characterization of the dynamic aspects of evacuation processes have been confined by experts (engineers, sociologists, and scientists alike) only through simulation and modeling. Because of this, several computer models have been proposed and utilized which aided in the understanding of crowd behaviors under normal, controlled egress, and panic situations. Of course, these models have been separately evaluated either by comparing the existing descriptive real-world data usually compiled by social scientists with the descriptive behavior of the model outputs, or by comparing the quantitative and qualitative outputs of the model with that of another model. One among the many models of crowd dynamics is the social force model (SFM) [17] which describes the motion of an individual as affected by both physical and psychological forces. SFM has been a popular model among researchers and planners because it was successful in describing phe- nomena observed in the real world, such as the “faster- is-slower” in escape panic and the “arching” phenomenon observed as a side-effect to “clogging” on exit ways. How- ever, when using SFM for crowds under low crowd density conditions, empirical results show that the simulated indi- viduals do not behave as expected. In other words, the simulated individuals act irrationally rather than rational ones, as we expect of humans under normal conditions. For example, the simulated individuals go directly and repeat- edly towards an object that blocks an exit way, instead of finding a way to go around the object and exit the hallway. This happens because the desired direction of the simulated individual is always towards the exit points, even in the pres- ence of obstacles between its current position and the exit points. The SFM has been used extensively by many researchers [15– 19] because of its capability to simulate the often observed human crowd phenomena such as flocking and bidirectional lane formation, among others. However, the simulated individuals under SFM do not exhibit some of the common human behaviors observed in a large crowd. These behaviors are: 1. Imitation. According to Festinger [8], there is a drive in a person towards its own self evaluation. This drive forces the person to belong to groups so that he can be associated with others. Thus, in a crowd of moving people, the individuals tend to move into groups whose members they believe would have the same opinion as theirs. 2. Contagion. Using the imitation behavior, people will also tend to “adopt” the behavior of others in the same group. Thus, if a group member senses that there is a discrepancy in a group with respect to behavior, opin- ions or abilities, that same member will reduce the discrepancy by adopting the most common behavior, opinion, or ability [8]. Similar to what Ngoho and Pabico [29] did in the past, Musse et al. [28] proposed a model based on the trajecto- ries of motion captured automatically from video sequences. While Ngoho and Pabico [29] captured the real-time trajec- tory of the velocity field (see for example Figure 1), Musse et al. [28] used an offline method to cluster similar trajecto- ries to extrapolate the velocity field. Because of these, researchers [10] recently proposed the social comparison theory (SCT) that algorithmically mimics the above-mentioned behaviors and was introduced as an alternative improvement to SFM. The key idea of SCT is that humans who lack the objective means to evaluate their current state, such as when in confusion, would com- pare themselves to others that they perceive are similar to themselves (i.e., both imitation and contagion behaviors). In our previous work [5], we used the SCT to design an arti- ficial individual whose behavior resembles that of a human when walking (or traveling) within a crowd. We included in the same model a “field of desire” to allow the simulated individual to change its desired direction of movement volun- tarily, based on its perception of the behavior and location of other individuals and obstacles along the hallway. The “field of desire” is a 100◦ field of vision facing the forward motion of the simulated individual, which the individual uses to obtain spatial information concerning its surround- ings and to choose the desired forward direction towards the exit goal. Our crowd of simulated individuals were able to exhibit the arching, clogging, and bursty exit phenomena usually observed in real crowd situations. The arch exhib- ited at the edge of the clogged crowd resembles that of a half ellipse, with the major axis parallel with the direction of the crowd flow, while the minor axis is coincident with the axis of the exit width. We present in this paper the result of the study we conducted to understand the interactive effects of crowd size c and exit width w in the formation of arch in a clogged crowd moving towards one exit door along and in the middle of a hallway. We will show by empirical means that the onset of arching has a negative correlation with the increasing w, while its slope tends to grow steeper as the c is increased. We will also show that the profile of the arch resembles a half ellipse whose major axis runs along the length of the hallway, while its minor axis is parallel to the width of the exit door. The major axis grows smaller as the exit width is increased, but grows bigger with the increase of the crowd size. The minor axis, on the other hand, grows bigger either as the crowd size is increased, or as the exit width is increased (of course limited only by the width of the hallway). 2. REVIEW OF LITERATURE The approaches used by researchers for crowd simulation maybe classified into two categories namely, macroscopic and microscopic. The macroscopic models are models that describe the dynamics of observable quantities of a crowd motion, such as density and velocity. Examples of such models are the ones used by Hughes [21], Huang et al. [20], and Treuille et al. [35] wherein they captured the density and velocity of the crowd using partial differential equations, the usual mathematical descriptions of the dynamics of fluid. Borrowing some ideas in gas-kinematic models, Helbing and Molnar [16] introduced the SFM to simulate pedes- trian flows, wherein a self-driven particle interacts through social rules and regulations, moves in its desired speed and direction, and attempts not to collide with obstacles, other particles, and surrounding barriers. Thus, in order to reach its destination faster, pedestrians take detours even if the taken route is crowded [18]. However, the choice of detour is dependent on the recent memory of what the traffic was like the last time they took the route, the figure of which was found by Ganem [11] to be polygonal in nature. In agreement with the SFM, Weidmann [38] observed that, as long as it is not necessary to go faster, such as going down a ramp, a pedestrian prefers to walk with his or her desired speed, which corresponds to the most comfortable walking speed. However, Weidmann [38] further observed that pedestrians keep a certain distance from other pedes- trians and borders. The distance between the pedestrians decreases as the density of the crowd increases. The pedes- trians themselves cause delays and obstructions. Arns [2] observed that the motion of the crowd is similar to the motion of gases and fluid, while [18] suggested that it is similar to granular flow as well. Helbing et al. [18], in his extension of the SFM, showed that many aspects of traffic flow can be reflected by self- driven many-particle systems. In this system, he identified the various factors that govern the dynamics of the particles such as the specification of the desired velocities and direc- tions of motion, the geometry of the boundary profiles, the heterogeneity among the particles, and the degree of fluc- tuations. One such observable pattern is the formation of lanes of uniform walking direction, formed because of the self-organization of the pedestrians [15]. Aside from the self-organizing behavior of the crowd, obstacles were also observed to both positively and negatively contribute to the flow of the traffic. During escape panic of large crowds, sev- eral behavioral phenomena were observed [17]: build up of pressure, clogging effects at bottlenecks, jamming at room widening areas, faster-is-slower effect, inefficient use of alter- native exits, initiation of panics by counter flows, and impa- tience. It was observed that the main contributing behaviors in these situations is a mixture of individual and grouping behavior. While macroscopic models describe the dynamics of the col- lective crowd, microscopic models describe the speed and position of each crowd member in a given time. Historically, Reynolds [31] is known to have pioneered the simulation of crowd in this manner. In his work, he simulated a flock of birds where each simulated bird is implemented as an independent individual that navigates according to its local perception of the dynamic environment. In a similar study in the years that follow, Tu and Terzopoulos [36] simulated schools of artificial fishes using perception to drive each fish’s Figure 1: An example output frame of the system developed by Ngoho and Pabico [29]. The upper left part contains the input image of student pedestrians with trackers (bounding boxes). The upper right part contains the trajectory of the pedestrians. The lower part contains the distance vs. time graph of the trajectory. This figure is in color in the electronic copy of this paper printed with permission from Ngoho and Pabico [29] and the Philippine Society of Information Technology Educators Foundation, Inc. (PSITE). locomotion while being constrained by underwater physics. The then novel programming of the fish’s motion and envi- ronment resulted into observed real-world school of fishes’ behaviors. Because of the success of microsimulation in exhibiting observed behaviors in animals, researchers in various fields of science (e.g., transportation science, computer gaming, artificial intelligence, and industrial robotics) adopted the programming method to simulate the motion and envi- ronment of human pedestrians in a crowd. The main purpose is to aid them in understanding some emergent crowd behaviors so that they can design better and safer walkways, or build realistic gaming elements, environments, and scenarios. During the advancement of the method itself, several enhancements to the model, as well as alter- natives, have been suggested. For instance, Kapadia et al. [22] proposed the use of what they called an “egocentric affordance field” to obtain the speed and the direction of each simulated pedestrian. On the other hand, Varas et al. [37] and Perez et al. [30] introduced cellular automata (CA) models to direct the motion of the simulated crowd member. Both works located their crowd members at cells of a grid wherein their respective coordinates are updated at discrete time intervals. The implementation of the CA model was so simple that Teknomo [33] implemented both the CA simulation and visualization using only a desktop electronic spreadsheet program. The simplicity of CA prompted other researchers [14, 26, 32] to further improve it using the prob- abilistic and statistical functions used in lattice-gas models to direct the motion of the simulated member in a cell. The lattice-gas model is known to be a special case of CA. Although the above models were already successful in sim- ulating observed phenomena in crowds, the behavior of the crowd member as an individual has not been fully inte- grated into these models. Because of this, Lerner et al. [24] proposed an approach for fitting human behaviors, such as talking on the phone, combing hair, and moving the head, into the simulated individual. In an attempt to simulate the collective behavior of humans in groups, Musse and Thalmann [27] used a hierarchical crowd organization with autonomous groups to simulate the grouping or flocking behavior in pedestrians. To be able to model and generate low-level behaviors for the simulated individuals, Golden- stein et al. [12] presented non-linear dynamic systems to model the interaction of the simulated individuals to its environment. Such low-level behaviors are real-time target tracking and obstacle avoidance in a highly dynamic envi- ronment. Later, the authors [13] used a hybrid low-level system integrating the non-linear dynamical systems with kinetic data structures to simulate a complex crowd envi- ronment that includes three-dimensional steering, crowding, and flocking within moving and non-moving obstacles. Recently, established works in the field of social psychology have been successfully incorporated into microsimulation models of the crowd, resulting into a much more realistic simulated collective behavior. Example of works go as far as 1895 such as that of Le Bon [23], and the newer ones such as that of Allport [1], Blumer [4], and Berk [3]. In these works, the authors discussed that people in crowds act similarly to one another, usually in a coordinated fashion that when viewed from a distance, it would appear that they are being choreographed by a single entity. However, in reality, the coordinated motion is achieved with little or no verbal communication at all. This somewhat homogeneous motion by the crowd is explained by Le Bon [23] as being directed by two processes, imitation and contagion that we discussed earlier in Section 1. The coordinated motion occurs in a “circular reaction” process, a type of interstimu- lation between two individuals a and b where the response of b to the stimuli from a reproduces a stimulation that is reflected back to a. The reflected stimulation reinforces a’s stimuli in return, thus resulting into a circular reaction process [4]. Following the adage that “similar people act in similar ways,” Allport [1] explained that the somewhat homogeneous motion of the crowd is the product of like- minded individuals. The likemindedness of the individuals in the crowd has been given a game-theoretic perspective by Berk [3]. Here, he said that the coordinated behavior of the crowd is consistent with that of an individual applying a minimax strategy. Given an action A being performed by the greatest number of individuals in a crowd, an indi- vidual c will see A as the action with least cost, and thus will perform it also. The individual that selects the action of the majority has to pay the least cost compared to that which follows the minority. One of the most recent models that incorporates works in social psychology, including that of cognitive sciences, is that of Fridman and Kaminka [10] where they used the ideas of SCT by Festinger [8]. In their work, they proposed an algo- rithmic framework for SCT which resulted into simulated crowd behavior that is more in tune with observed real-world crowd behavior compared to those techniques that do not incorporate the cognitive models. Because of this, we have decided to start with the SCT, but hybridized it with our own algorithm for mapping the perceived trajectory of an individual’s goal, limited by the individual’s field of vision (i.e., our “field of desire”) [5]. Our hybrid model was able to exhibit important crowd behaviors: arching, clogging, and bursty exit rates. We have also observed a new phenomenon we called double arching, which we will discuss in detail in another paper and forum. In this paper, we will explore the interactive roles of c and w to the formation of the arch using the hybridized microsimulation approach in the hope of exploiting the outcomes for designing better and safer walkways under various crowd egress situations. 3. METHODOLOGY 3.1 Programming the Individual In our approach, each simulated individual was programmed with the following general behaviors which were described originally by Wooldridge and Jennings [39], and then later on extended and explained further by Epstein [7], Franklin and Graesser [9], Torrens [34] and Macal and North [25]: 1. Autonomy: The individuals are autonomous units, and thus are not directed by any centralized control. They are capable of processing information and exchanging this information with other individuals in order to make independent decisions. They may freely interact with others in a few given situations, and this freedom does not affect their being autonomous. With this, we say that the individuals are active rather than purely passive (see discussion of “active” below). 2. Heterogeneity: We believe that individuals should be heterogeneous to allow the microsimulation to develop autonomous individuals. 3. Active: Being active means that the individuals influ- ence a simulation independently. The following active features can be identified: (a) Pro-active or goal-directed (b) Reactive or perceptive (c) Bounded rationality (d) Interactive and communicative Aside from the abovementioned general behaviors, we infuse the SCT by following the algorithmic approach of Fridman and Kaminka [10], which we describe here for clarity. The SCT follows a subset of axioms as follows: 1. When individuals can not objectively decide what to do next, they compare their current state with that of others. Each individual Pi is characterized by a tuple with k attributes P = haP k i, where each attribute ai j of individual Pi, ∀1 ≤ j ≤ k, corresponds to a point in a k-dimensional space. Each dimension corresponds to an attribute such as the coordinates of Pi in two-dimensional space, heading, height, gender, etc. 2 , . . . , aP 1 , aP 2. Individuals compare their state to individuals that are similar to them. We defined a similarity function Sim(Pi, Pj) which measures the similarity between the ith and jth individuals (i 6= j). The similarity function is given in Equation 1. For an individual Pi, we selected the individual with the highest similarity. 3. If there are differences in their state with the object of comparison, the individuals take steps to reduce the differences. If the similarity of the selected individual is below a given value, then Pi is triggered to do some- thing to reduce the difference. We measured the similarity between Px and Py indepen- dently along each dimension, such that the similarity in dimension ai is a function Sai (ax i ) : ai × ai → [0, 1]. A value of 1 indicates full similarity, while a 0 means dis- similarity. For example, one attribute of Px is its distance from Py, thus a value of 1 means they are positioned in the same location, while a 0 means they are far apart. i , ay Sim(Px, Py) = k X j=1 Saj (ax j , ay j ) × wj (1) Because our simulated individuals exhibit these behaviors, our simulation approach is more akin to modeling humans and objects in very realistic ways than other modeling approaches, such as those that aggregate mathematical equations to explain the dynamics of pedestrian movement (i.e., macrosimulations). Thus, our simulation can be used with higher confidence to perform what-if scenarios to aid, for example, university administrators and decision makers with regards to management policies, as well as infrastruc- ture development plans, for safer learning environments for the students and constituents (see for example Figure 2). Figure 2: Snapshots of example what-if scenario simulations on the effect of exit door configurations to student egress (students are shown as black circles with pointers to visualize heading) in a very large lecture hall: (a) Current exit door configuration which opens outward but hinders student flow; (b) Effect of moving the door hinge such that the exit door opens behind the student flow; and (c) Effect of increasing the effective width of the exit door by 50% and replacing the swinging door by a sliding door. This figure is in color in the electronic copy of this paper printed with permission from Castro and Pabico [6] and the Philippine Society of Information Technology Educators Foundation, Inc. (PSITE). 3.2 The Goal of the Individual In a two-dimensional flat world, let G be the goal of Pj (∀j) with set coordinates {(x1, y1), (x2, y2), . . . , (xn, yn)} repre- senting a line segment. This set is usually the location of a fixed-width exit door. Since all P aim to reach any one of the exit coordinates, a given Pi will face towards the direction of the nearest exit coordinates from its current coordinates. The Pi will scan its field of vision for the closest free space and move towards that space with its gaiting speed, which is currently set at one pace per simulation time step. In the future, we wish to vary this gaiting speed depending on the height of Pi. A free space is a location in the flat world that is not a wall, a Pj where j 6= i, or any other object. If other individuals are blocking Pi’s space within its field of vision, Pi stops at its current coordinates. It is possible that the individual might move away from the target exit coordinate if the chosen free space was at the edge of its field of vision. When that happens, the individual will still move to the free space, but after moving, it will redirect its heading towards the possibly new nearest exit coordinates. When the distance of the individual to the nearest exit coor- dinates is < 1, it considers itself as already exited and will move to the edge of the simulation world. Until this dis- tance has not been achieved, the individual will just repeat its decision making process as described above. 3.3 Interactive Effects of Crowd Size and Exit Width on Arching Profile We conducted experiments along a fixed-width hallway to determine the interactive effects of c and w to the time T of the onset of formation of the arch, as well as the dimen- sion of the arch’s profile at T . The arch profile is a half ellipse, halved across the major axis, and is measured with respect to the length of the minor and major axes. We experimented on five levels of c (200, 300, 350, 400, and 450 individuals) and seven levels of w (1, 3, 5, 7, 9, 11, and 13), resulting into 35 c–w combinations. We replicated each combination three (3) times to take into considera- tion random variabilities brought about by external factors that can not be modeled, such as the implementation of the random number generator of our host operating system and other systemic variabilities such as the occurrences of hard- ware interrupts during simulation that might have affected our measurements. The exit width is defined as the number of individuals in a straight line that can concurrently and comfortably yet tightly fit along the axis of the exit width. Similarly, the fixed-width hallway, set at 19, is also defined as the number of individuals along a straight line that can comfortably yet tightly fit in a straight line parallel to the corridor width. 4. RESULTS AND DISCUSSION Figure 3 shows an example profile of a crowd that exhibits arching near the exit door. We see here that the arch resem- bles a half ellipse with the ellipse’s major axis running along the direction of the crowd flow and perpendicular to the exit width, while the minor axis is parallel to the exit width. This result is a confirmation of the results of Epstein [7], Franklin and Graesser [9] and Helbing et al. [18]. 4.1 Arching Time Figure 4 shows the response of the onset of arching time T to the combination effect of varying c and w. We can see here that across different c, there is a recurrent pattern that the T decreases as the w is increased, but the amount of decrease is not the same between any c. To better quan- tify this observed decrease in T at increasing w, we con- ducted a regression analysis between T and w for each value of c. Equations 2 through 6 show these regressions, respec- tively for c = {450, 400, 350, 300, 200}. All coefficients are significantly different from zero at a confidence level of α = 0.01, while all R2 values are significantly different from zero at α = 0.05. This means that even though Equations 5 and 6 have low R2 values, these values are non zero. Looking at Figure 3: Profile of the crowd before the exit door as it forms an arch. This is a snapshot of the microsim- ulation when the c = 400 at w = 7. Notice that the major axis runs along the axis of the corridor, while the minor axis is parallel with the axis of the exit door. the respective slopes of the regression, we see here that as the c is decreased, the slope gets to zero. This means that at lower c values, the onset of the formation of the arch is becoming independent of the w. The w has a significant effect on T at higher c. Generally, however, T ∝ 1 c·w . T450 = −1.80w + 44.72, R2 = 0.97 T400 = −2.04w + 48.25, R2 = 0.92 T350 = −0.54w + 37.89, R2 = 0.95 T300 = −0.32w + 35.49, R2 = 0.66 T200 = −0.18w + 34.39, R2 = 0.34 (2) (3) (4) (5) (6) 4.2 Arch Profile Figure 5 shows the response of the lengths of the major and minor axes at varying values of c and w. Except for c at 450 and 400 individuals, the length of the major axis decreases as the c is decreased, and as the w is increased, while the length of the minor axis decreases as the c is decreased, and increases as the w is increased. Mathematically, M ∝ c w , while m ∝ c · w. We observed here from the figure that when the w = 9, the c = 400 plateaued at m = 19, while with M , the slope becomes positive. Similar behavior was observed for c = 450 but the plateauing and the sudden change of direction of the slope for m happened at w = 7. At these c values, the individuals have already reached the edge of the corridor that m cannot expand anymore. Since m is constrained to 19, the arch must accommodate more individuals, thus translating into the increase in M . However, had the hallway width been way more than 19, the same pattern would have been observed for all c. We confirmed this assertion when we rerun the simulation for c = {400, 450} and w = {9, 11, 13} at the same number of replicates. When we plotted the mean of each c × w combination across replicates, they fall reasonably well within the regression line that we obtained earlier, of course within some acceptable variance (i.e., the respective means’ ±1× standard deviations intersect with the regression line, see Figure 6). 5. SUMMARY AND CONCLUSION In this paper, we present the behavior for a simulated indi- vidual used in microsimulation of pedestrian crowd. The behavior is based on the SCT and “field of desire” hybrid technique in our attempt to provide cognitive behaviors to the simulated individual so that the crowd can exhibit crowd patterns that are observed in real-life, such as that of arching. With this programming, we experimented on the combination effect of c and w on T , and the arch pro- file described by M and m. Here, we found out that the T ∝ 1 w , and the m ∝ c × w. These results suggest that, intuitively, the w should be wide enough at higher c so that T will happen at the least possible time, and at the same time M will be shorter, and m will be longer (but still M > m). We believe, though, that when we find the values for c and w where M = m, then this is the combination when T will be the fastest. Thus, we will investigate this hypothesis in the future. c×w , the M ∝ c 6. ACKNOWLEDGMENT This research effort is under the research program Multi- agent Simulation of Student Egress from the ICS Mega Lecture Hall Under Normal, Controlled Emer- gency, and Panic Situations funded by the Institute of Computer Science, University of the Philippines Los Banos. References [1] Floyd Henry Allport. Social Psychology. Houghton Mif- flin, Boston, 1924. [2] Thomas Arns. Video films of pedestrian crowds, 1993. [3] Richard A. Berk. A gaming approach to crowd behavior. American Sociological Reviews, 39:355–373, 1974. [4] Herbert G. Blumer. Collective behavior. In A.M. Lee, editor, Principles of Sociology, pages 67–121. 1939. [5] Francisco Enrique Vicente G. Castro and Jaderick P. Pabico. Microsimulations of arching, clogging, and bursty exit phenomena in crowd dynamics. In Allan L. Sioson, editor, Proceedings of the 10th National Con- ference on Information Technology Education (NCITE 2012). Philippine Society of IT Educators Foundation, Inc., 2012. Figure 4: T as affected by c and w combination: (a) c = 450; (b) c = 400; (c) c = 350; (d) c = 300; and (e) c = 200. The blues squares are means of T at each c–w combination, averaged across all replicates. The blue lines are the respective regression lines. Figure 5: (a) M and (b) m as affected by c and w combination. Colored shapes are means while lines are regression fit at each c. Orange/triangle represents c = 450, blue/square c = 400, green/inverted triangle c = 350, red/triangle pointing right c = 300, and purple/diamond c = 200. Figure 6: (a) M and (b) m as affected by c and w combination when W = 35. [6] Francisco Enrique Vicente G. Castro and Jaderick P. Pabico. Microsimulations of arching, clogging, and bursty exit phenomena in crowd dynamics. Philippine Information Technology Journal, 6(1):11–16, 2013. [16] Dirk Helbing and Peter Molnar. Social force model for pedestrian dynamics. Physical Review E: Journal for Research in Statistical, Nonlinear and Soft-matter Physics, 51(5):4282–4286, 1995. [7] Joshua M. Epstein. Agent-based computational models and generative social science. Complexity, 4(5):41–60, 1999. [17] Dirk Helbing, Illes J. Farkas, and Tamas Vicsek. Simu- lating dynamical features of escape panic. Nature, 407 (6803):487–490, 2000. [8] Leon Festinger. A theory of social comparison pro- cesses. Human Relations, 7:117–140, 1954. [9] Stan Franklin and Art Graesser. Is it an agent, or just a program? A taxonomy for autonomous agents. In Proceedings of the Third International Work- shop on Agent Theories, Architectures, and Languages. Springer-Verlag, 1996. [10] Natalie Fridman and Gal A. Kaminka. Modeling pedes- trian crowd behavior based on a cognitive model of social comparison theory. Computational and Mathe- matical Organization Theory, 16:348–372, 2010. [11] Joseph Ganem. A behavioral demonstration of Fermat´s principle. The Physics Teacher, 36, 1998. [12] Siome Goldenstein, Edward Large, and Dimitris Metaxas. Non-linear dynamical system approach to behavior modeling. The Visual Computer, 15(7):349– 364, 1999. [13] Siome Goldenstein, Menelaos Karavelas, Dimitris Metaxas, Leonidas Guibas, Eric Aaron, and Ambarish Goswami. Scalable nonlinear dynamical systems for agent steering and crowd simulation. Computers and Graphics, 25(6):983–998, 2001. [14] R.Y. Guo and H.J. Huang. A mobile lattice gas model for simulating pedestrian evacuation. Physica A: Statis- tical Mechanics and its Applications, 387(2–3):580–586, 2008. [15] Dirk Helbing. Safety Management at Large Events: The Problem of Crowd Panic. Institute for Transport and Economics, Dresden University of Technology, 2006. [18] Dirk Helbing, Peter Molnar, Illes J. Farkas, and Kai Bolay. Self-organizing pedestrian movement. Environ- ment and Planning B: Planning and Design, 28:361– 383, 2001. [19] Colin M. Henein and Tony White. Agent-based mod- eling of forces in crowds. In Multi-agent and Multi- agent-based Simulation: Lecture Notes in Computer Science, volume 3415, pages 173–184. 2005. [20] Ling Huang, S.C. Wong, Mengping Zhang, Chi-Wang Shu, and William H.K. Lam. Revisiting hughes’ dynamic continuum model for pedestrian flow and the development of an efficient solution algorithm. Trans- portation Research Part B: Methodological, 43(1):127– 141, 2009. [21] Roger L. Hughes. A continuum theory for the flow of pedestrians. Transportation Research Part B: Method- ological, 36(6):507–535, 2002. [22] Mubbasir Kapadia, Shawn Singh, Billy Hewlett, and Petros Faloutsos. Egocentric affordance fields in pedes- trian steering. In Proceedings of the 2009 Symposium on Interactive 3D Graphics and Games, pages 215–223, 2009. [23] Gustave Le Bon. The Crowd: A Study of the Popular Mind. N.S. Berg, Dunwoody, Georgia, USA, 1968 edi- tion, 1895. [24] Alon Lerner, Eitan Fitusi, Yiorgos Chrysanthou, and Fitting behaviors to pedestrian Daniel Cohen-Or. simulations. In Proceedings of the 2009 ACM SIG- GRAPH/Eurographics Symposium on Computer Ani- mation, pages 199–208, 2009. [38] U. Weidmann. Transportation technique for pedes- trians. In Publication Series of the Institute of Trans- portation, Traffic Engineering, Roads and Railways, number 90. ETH Zurich, Switzerland, 1993. Translated from German. [39] Michael Wooldridge and Nicholas R. Jennings. Intel- ligent agents: Theory and practice. Knowledge Engi- neering Review, 10(2):115–152, 1995. [25] Charles M. Macal and Michael J. North. Tutorial on agent-based modelling and simulation. Journal Of Sim- ulation, 4:151–162, 2010. [26] Masakuni Muramatsu, Tunemasa Irie, and Takashi Nagatani. Jamming transition in pedestrian counter flow. Physica A: Statical Mechanics and its Applica- tions, 267(3–4):487–498, 1999. [27] Soraia R. Musse and Daniel Thalmann. Hierarchical model for real time simulation of virtual human crowds. IEEE Transactions on Visualization and Computer Graphics, 7(2):152–164, 2001. [28] Soraia R. Musse, Claudio R. Jung, Jr. Julio C.S. Jacques, and Adriana Braun. Using computer vision to simulate the motion of virtual agents. Com- puter Animation and Virtual Worlds, 18(2):83–93, 2007. [29] Louie Vincent A. Ngoho and Jaderick P. Pabico. Capturing the dynamics of pedestrian traffic using a machine vision system. Philippine Information Tech- nology Journal, 2(2):1–11, 2009. [30] Gay Jane Perez, Giovanni Tapang, May Lim, and Caesar Saloma. Streaming, disruptive interference and power-law behavior in the exit dynamics of confined pedestrians. Physica A: Statistical Mechanics and its Applications, 312(3–4):609–618, 2002. [31] Craig Reynolds. Flocks, herds and schools: A dis- tributed behavioral model. Computer Graphics (SIG- GRAPH), 21(4):25–34, 1987. [32] Yusuke Tajima and Takashi Nagatani. Scaling behavior of crowd flow outside a hall. Physica A: Statis- tical Mechanics and its Applications, 292(1–4):445–554, 2001. [33] Kardi Teknomo. Multi-agent mesoscopic pedestrian simulation. In Henry Adorna and Allan Sioson, editors, Proceedings of the 11th Philippine Computing Science Congress, 2011. [34] Paul Morrison Kevin Torrens. Simulating Sprawl: A Dynamic Entity-Based Approach to Modelling North American Suburban Sprawl Using Cellular Automata and Multi-Agent Systems. PhD thesis, University Col- lege London, London, 2004. [35] Adrien Treuille, Seth Cooper, and Zoran Popovi. Con- tinuum crowds. In Proceedings of the Annual Confer- ence on Computer Graphics and Interactive Techniques (SIGGRAPH 06), pages 1160–1168, 2006. locomotion, perception, behavior. [36] Xiaoyuan Tu and Demetri Terzopoulos. Artificial fishes: Physics, In Pro- ceedings of the 21st Annual Conference on Computer Graphics and Interactive Techniques (SIGGRAPH 94), pages 43–50, 1994. [37] A. Varas, M. Cornejo, D. Mainemer, B. Toledo, Cellular J. Rogan, V. Munoz, and J. Valdivia. automaton model for evacuation process with obstacles. Physica A: Statistical Mechanics and its Applications, 382(2):631–642, 2007.
1904.05429	1	1904	2019-04-08T12:24:47	A Multi-Agent based Approach for Simulating the Impact of Human Behaviours on Air Pollution	[ "cs.MA" ]	This paper presents a Multi-Agent System (MAS) approach for designing an air pollution simulator. The aim is to simulate the concentration of air pollutants emitted from sources (e.g. factories) and to investigate the emergence of cooperation between the emission source managers and the impact this has on air quality. The emission sources are controlled by agents. The agents try to achieve their goals (i.e. increase production, which has the side effect of raising air pollution) and also cooperate with others agents by altering their emission rate according to the air quality. The agents play an adapted version of the evolutionary N-Person Prisoners' Dilemma game in a non-deterministic environment; they have two decisions: decrease or increase the emission. The rewards/penalties are influenced by the pollutant concentration which is, in turn, determined using climatic parameters. In order to give predictions about the Plume Dispersion) model and an ANN (Artificial Neural Network) prediction model. The prediction is calculated using the dispersal information and real data about climatic parameters (wind speed, humidity, temperature and rainfall). Every agent cooperates with its neighbours that emit the same pollutant, and it learns how to adapt its strategy to gain more reward. When the pollution level exceeds the maximum allowed level, agents are penalised according to their participation. The system has been tested using real data from the region of Annaba (North-East Algeria). It helped to investigate how the regulations enhance the cooperation and may help controlling the air quality. The designed system helps the environmental agencies to assess their air pollution controlling policies.	cs.MA	cs	A Multi-Agent based Approach for Simulating the Impact of Human Behaviours on Air Pollution Informatica 39 (2015) 501 -- 505 501 Sabri Ghazi Laboratoire de gestion électronique du document, Computer Science department, University Badji Mokhtar, PO-Box 12, 23000,Annaba, Algeria E-mail: [email protected] Julie Dugdale University Grenoble Alps. LIG. France. Email: [email protected] Tarek Khadir Laboratoire de gestion électronique du document, Computer Science department, University Badji Mokhtar, PO-Box 12, 23000,Annaba, Algeria E-mail: [email protected] Keywords: Air Pollution, Multi-Agent Simulation, Agent-Based Simulation, Air Quality, N-persons Prisoners' Dilemma, Game Theory Received: [Enter date] This paper presents a Multi-Agent System (MAS) approach for designing an air pollution simulator. The aim is to simulate the concentration of air pollutants emitted from sources (e.g. factories) and to investigate the emergence of cooperation between the emission source managers and the impact this has on air quality. The emission sources are controlled by agents. The agents try to achieve their goals (i.e. increase production, which has the side effect of raising air pollution) and also cooperate with others agents by altering their emission rate according to the air quality. The agents play an adapted version of the evolutionary N-Person Prisoners' Dilemma game in a non-deterministic environment; they have two decisions: decrease or increase the emission. The rewards/penalties are influenced by the pollutant concentration which is, in turn, determined using climatic parameters. In order to give predictions about the concentration of pollutants: Particulates Matter (PM10), Sulphur Oxide and Dioxide (SOx), Nitrogen Oxides (NOx) and Ozone: (O3), a two stage prediction method is used, a GPD (Gaussian Plume Dispersion) model and an ANN (Artificial Neural Network) prediction model. The prediction is calculated using the dispersal information and real data about climatic parameters (wind speed, humidity, temperature and rainfall). Every agent cooperates with its neighbours that emit the same pollutant, and it learns how to adapt its strategy to gain more reward. When the pollution level exceeds the maximum allowed level, agents are penalised according to their participation. The system has been tested using real data from the region of Annaba (North-East Algeria). It helped to investigate how the regulations enhance the cooperation and may help controlling the air quality. The designed system helps the environmental agencies to assess their air pollution controlling policies. 1 Introduction The question about how humans should moderate their exploitation of environmental resources has occupied researchers for decades [1]. Promoting social and economic growth without affecting the environmental equilibrium is important for maintaining sustainable development. This paper addresses the relation between human behaviours and their impact on air quality in socio-environmental systems. Air pollution is a major concern in many cities in the world, especially in developing countries. It has a direct influence on our health and quality of life [2]. The degradation in air quality should be estimated before the establishment or the expansion of urban or industrial activities. Air pollution simulation and decision support tools can help decision-makers to establish policies for environmental management and to predict the impact of their decisions on the environment and ecosystem. Many modelling approaches have been proposed to study air pollution. Most of them, ([3], [4], [5]) to cite a few, are mainly focused on the physical and chemical aspects of air pollution; the concentration and dispersal of pollutant. These models do not take into consideration human- decision factors. Air pollution is by nature distributed and includes the interaction of individuals involved in the exploitation of a dynamic ecological resource which is the air. The anthropogenic activities (road traffic, industrial and agricultural activities) are among the major sources of air pollution. All of these activities are 502 Informatica 39 (2015) xxx -- yyy controlled by humans; therefore, including the human decision factors in the modelling of air pollution is essential. MAS (Multi-Agent System) based models are an appropriate method for modelling socio-environmental issues [6]. They allow us to model the behaviours of human actors sharing the exploitation of environmental resources. [7] presents a review of recent MAS models used to investigate socio-environmental problems. The models are classified according to: the decision making mechanism, the use or not of real data, the objective of the simulation, and the space and time representation. [8], [9] used a MAS approach to investigate the air pollution emission resulting from road activities; they used a traffic flow simulation and linked it to emission calculation. [10] used the same approach to study the effect of transport regulation on air pollution emission. [11] present a MAS designed for monitoring air quality in Athens, Greece. The MAS is a set of agents that control a network of sensors installed in an urban region. They verify and collect the data measured by sensors. [12] present a MAS to find the dispersion of air pollution in urban region. The pollution sources (polluters) are represented by homogeneous agents that emit pollution in their respective areas. Each agent pollutes with an emission rate. As the simulation runs, clusters are formed with different values of pollution concentration. At the end, a single cluster is formed, thus, the dispersion of pollution is estimated. The managers of emission sources share the exploitation of the air by emitting pollutants. We aim to simulate their different personalities (e.g. eco-friendly, selfish) and investigate the relationship between the emergence of cooperation and its impact on air quality. The main questions addressed in this paper are: How do emission source controllers cooperate, are they able to achieve their goals while preserving a reasonable air quality? What regulatory rules should be adopted to enhance the cooperation and sustain the air quality? To investigate these questions we have designed a MAS simulation tool that helps to investigate the emergence of cooperation and its effects on air quality. The proposed simulator models the population of emission source controllers as a network of agents playing an Evolutionary NPPD (N- Person Prisoners' Dilemma) game. Evolutionary NPPD has been widely used for studying the emergence of cooperative behaviour among a population of selfish agents, how agents exhibit altruistic behaviours and under which condition cooperation will be sustained. NPPD is a mathematical model, which models the conflict between players sharing the use of a common resource. Initially, it was formulated for two players where each one has to take two possible actions (defect or cooperate), and then receive a payoff according to their joint actions. A version for N-Persons has been proposed [13] where the payoff is calculated according to the number of agents choosing to cooperate; the payoff function is given in (1) the agent s ϵ{0,1} (where 0 means cooperate and 1 means defect), ncp is the number of players who chose to cooperate, N is the size of the player population and b is the defection temptation, the constant c is used to ensure that the cooperation reward is less than the defection reward. [14], [15], studied the emergence of cooperation in a NPPD. The authors used different agent personalities and neighbourhoods in order to investigate their impact on the evolution of the game outcome. The experiment used different agent types with different initial co-operators ratio; this showed that for the case where all agents are Pavlovians (repeating actions that give them more satisfaction), the aggregate outcome of the game can be predicted for any number of agents and any payoff function. The choice of the agent's neighbours also has a big influence on the game equilibrium. [16] investigated the effect of social welfare preference on the emergence of cooperation among agents placed on a BA [17] network. The authors proposed a model where some of the agents also take into consideration social welfare and not only their payoff received from the game. Agents do not only care about their own payoff, but also the payoff of their neighbours. Sabri Ghazi; Julie Dugdale; Tarek Khadir 𝑢(𝑛𝑐𝑝)= !∗!"#! −𝑐 𝑖𝑓 𝑠=0 !∗!"#! 𝑖𝑓 𝑠=1 (1) With 𝑏>𝑐>0 𝑎𝑛𝑑 𝑐>!!, s is the action taken by [18] describes the use of a NPPD game to investigate the cooperation in a socio geographic community. The use of NPDD for environmental modelling has proved to be suitable since the exploitation of a shared ecological resource can be formulated as a tragedy of the common [19]. Each actor tends to maximise its profits by exploiting a shared ecological resource. Thus, a tragedy of the common arises. [20] uses a PD model to review Porters' hypothesis, which studies the relationship between productivity and eco-friendly technologies. The work models how strict environmental regulations can enhance innovation for a less polluting technology. Firms have two actions which are: to invest in a new less polluting process or to continue using the old one and be penalised according to the governmental regulations. [21] used a version of NPPD to investigate the cooperation in international environmental negotiation to reduce CO2 emissions. [22] presents an evolutionary game theory approach to study the influence of the ecological dynamic and payoff structures over the emergence of cooperative behaviour between landowners. The landowners are modelled as selfish agents aiming to maximize their profit by managing the number of deer on their lands. The main novelty of our approach is the inclusion of human decisions as a key element for simulating the air pollution evolution. We model the managers of the emission sources of pollutants as autonomous agents. These agents aim to maximise their own profit and we investigate this effect on air quality. The designed system helps investigating the efficiency of the regulatory rules used by air pollution controlling agencies for maintaining the air quality. This is very A Multi-Agent based Approach for Air Pollution Informatica 39 (2015) 501 -- 505 503 important because it helps the environmental agencies to assess their air pollution controlling policies. simulation socio-environmental The paper is organised as follows: The methodology is presented in section (2) that describes a MAS approach for designing an air pollution agent based simulator. Subsection (2.1) presents the representation of space and time. Subsection (2.2) describes the dispersion model and the Artificial Neural Network (ANN) prediction model. Many conceptualizations have been proposed to represent a socio-environmental system [22]. Generally, a system can be represented as an interconnection of three components (or subsystems); each one is represented by a set of variables (attributes) forming its state at time t. The ecological component models the biotic (living) and abiotic (non-living) parts. The economic component represents the economic view point and groups the economic variables. The social component represents the human social networks such as decision-makers, firms, government agencies and consumers. A change in the state variable of each component affects other systems' state variables. For example, the increase in demand for a certain kind of fish, leads fishermen to intensify their exploitation; this in turn results in changes to the biodiversity. We present a generic formalization of a socio-environmental system. A coupled social and environmental system can be expressed as a set of economic, social and ecologic state variables. The state of the system at time step t can be formulated as (2). 𝐸𝑆𝑡=<𝐸𝑐𝑡,𝑆𝑐𝑡,𝐸𝑛𝑣𝑐𝑡> (2) 𝐸𝑐!=<𝐸𝑐!,!,…,𝐸𝑐!,!>,𝑆𝑐!=<𝑆𝑐!,!,…,𝑆𝑐!,!> 𝐸𝑛𝑣𝑐!=<𝐸𝑛𝑣𝑐!,!,…,𝐸𝑛𝑣𝑐!,!>(3) are: 𝐸𝑛𝑣𝑐!=<𝑐!,!,..𝑐!,!,𝑊𝑆!,𝐻𝑢!,𝑇!,𝑅𝐹!>(4) 𝑆!=<𝑒𝑟!,𝑡𝑐,𝑋,𝑌,𝑍> (5) ci,t is the concentration of the pollutant i, WS: wind speed, T: temperature, Hu: humidity and RF represents the rainfall, at time t. Assuming that the source of pollution at time t is modelled as: Where Ec, Sc and Envc represent, the sets of economic, social and environmental state variables, respectively: In our case, the environment state variables at time step t The source produces the pollutant tc with the rate er at the geo-position (X,Y,Z). Sources are controlled by agents. Every agent has to make a decision on which action to choose among all possible actions according to the state of the environment ES and its internal state at t. of an action is the change in the emission rate of the pollutant from the controlled source. We can define this as a function that takes the agent's action and as a result gives the new emission rate (6). Let A be the set of actions 𝐴= 𝑎!,𝑎!,…,𝑎! , the result Let 𝜋! be the action vector done by N agents at time t: 𝜋!=<𝐴!!,..𝐴!"> (7) 𝑄={𝑣𝑒𝑟𝑦_𝑏𝑎𝑑,𝑏𝑎𝑑,𝑎𝑣𝑒𝑟𝑎𝑔𝑒,𝑔𝑜𝑜𝑑,𝑣𝑒𝑟𝑦_𝑔𝑜𝑜𝑑}, 𝐹:𝐴→ℝ (6). Let Q be the set of possible air quality index values: each of these indexes has its numerical equivalent in terms of pollutant concentration, as shown in table 1. The agent decision-making mechanism is given in (2.3). A test scenario is presented in section (3). Results are detailed and discussed in section (4). The paper ends with conclusions and the possible further directions of our work. 2 Model Approach and architecture SOx µg 0 -- 30 30-60 60-125 125- 250 >250 Indices Category 1 2 3 4 5 Very Good Good Average Bad Very Bad NOx µg 0-45 O3 µg 0-45 PM10 µg 0-20 45-80 80-200 200- 400 >400 45-80 80-150 150- 270 >270 20-40 40-100 100-200 >200 Table 1: Air pollution quality 𝑇𝜋!,𝐸𝑆!,𝑐𝑢𝑟𝑟𝑒𝑛𝑡_𝑞 ⟶𝑛𝑒𝑤_𝑞, 𝑐𝑢𝑟𝑟𝑒𝑛𝑡_𝑞 𝑎𝑛𝑑 𝑛𝑒𝑤_𝑞∈𝑄 (8) The air quality can be modelled as a graph with T as transition function: T takes as arguments the state of the system ESt and the set of actions done by N agents and accordingly it moves the system from the current state (current_q) to a new state (new_q). Under some conditions current_q may be equal to new_q, which means that actions of the agents do not change the air quality under some climatic conditions. Our simulation approach can be schematised as shown in figure 1. Agents' actions affect the emission rate of the sources. Then the dispersion algorithm is used to compute the dispersion, the aggregated value of pollutant concentration is used with climatic parameters to forecast the next 2 hours air pollution concentration and air quality. According to these forecasts, agents are rewarded or penalised. Agents then adapt their strategies to earn more reward and reduce penalties. Dispersion models Emission rates Pollutants concentration Weather Parameters Prediction models point sourcei Emis sion sourcei Polluants Concentration & Air Quality Agenti Actions Rewards/ Penalties Environment Figure 1: The simulation process, using the dispersion model and the prediction model. 2.1 The spatial and temporal scale of the simulation model: The simulation uses a discrete representation of time where each simulation step represents by default 6 hours of real time. The simulation's duration is defined in the Informatica 39 (2015) xxx -- yyy 504 interface and depends on the objective of the simulation (short or long term prediction). Our model is based on the hypothesis that the action of the emission controllers (reducing or increasing emissions) has an impact within k time-steps. k is a parameter whose value is provided by the user according to the scenario and available data. Since the simulation step k can be changed, we can represent a long term simulation horizon by giving k a higher value. So for example we can represent 1 step as 24 hours meaning that industrial polluters can take several days to adjust their production volumes. Setting k even higher, such as 2 weeks or 1 month would require data, which is not available, to see the evolution of the AQ. represented as: 𝐵𝑋={𝑏𝑥!,…,𝑏𝑥!}, every box The environment is modelled as a set of 3D boxes, each one represents one km3 (see figure 3). It can be is localised in the geo-position point gp(x,y,z) and has attributes representing the concentration of air pollutants (cp0,…,cpv) and air quality, These attributes are used when agents are penalized according to the pollution level in the box where they are situated. In this case the position of the emission source in a box is relevant. Sources located in the same box are considered to be neighbours. 2.2 Dispersion and prediction models is frequently used The dispersion model helps to measure how the pollutant will spread in the air. It is calculated according to the distance from the point source, the wind speed and direction. We used a GPD (Gaussian Plum Dispersion model), which in atmospheric dispersion [24]. The dispersion model is run in a steady way, which means that no parameter (wind speed, emission rate and wind direction) is changed during the simulation step. This provides a series of snapshots of the situation at each step. These snapshots are then fed into the ANN model to obtain a prediction about the concentration. Since we cannot combine the two models in a continuous way our solution of taking a series of snapshots and feeding it to the ANN mimics a continuous process. The GPD simulates the dispersal from a point source emission according to the emission rate (9). !!!! !!!!∗𝑒! !!!∗!!!∗ (𝑒!(!!!! 𝐶𝑥,𝑦,𝑧,𝐻 = !"!,!∗! !!!! ))+ (𝑒!(!!!! !!!! )) (9) This means that the concentration of the pollutant at point (x,y,z) is calculated according to : eri,t: the emission rate in kilograms per hour of the source i in time step t. Ui: the wind speed in metres per second at time step t, σyσz: the standard deviation of the concentration distributions in a crosswind in a vertical direction, these two parameters are chosen according to the stability class 'C' in the Guifford-Pasquill scale [25], and H is the height of the source from the ground. The decay term D is given in [26] and computed according to (10). 𝐷= 𝑒 !/(!∗!!),𝑖𝑓 𝑅>0 1,𝑖𝑓 𝑅=0 , (10) Sabri Ghazi; Julie Dugdale; Tarek Khadir Where x is the downwind distance, u is the wind speed and R is the decay coefficient. The values for R are adopted from [27] for NOx (0.45 h-1) and SOx (0.31 h-1), PM10 is not considered (R=0 and D=1). For simplification and due to lack of wind direction data, we assume that the wind direction does not change during the simulation step. The resulting pollution level from each source is aggregated and the average of each box is computed. Then the dispersion value of the pollutant is passed to an ANN prediction model as described in [28]. The ANN prediction models are designed to give a forecast of the five air pollutants and the air quality. This includes an uncertainty aspect caused by the weather conditions [29]. The ANN predictor uses the aggregated air pollution concentration value given by the dispersion model of each source and the four climatic parameters: wind speed, humidity, temperature and rain fall. These parameters greatly influence the pollutant concentration in the air [30]. from O3 is a secondary pollutant, which means that it is the not emitted by photochemical interaction between SOx, COx, and organic components. Therefore, we used SOx and COx dispersion information to predict the O3 concentration. sources, but results For each pollutant a RBF (Radial Basis Function) network is designed and trained. The RBF is composed of three layers, the first layer is connected to the input of the network and its output is connected to the hidden layer, the neurones in the hidden layer have the RBF as the activation function. The outputs of the hidden layer are linearly combined to obtain the output of the network. Using a training data set, the objective is to find the optimal combination between the number of neurons in the hidden layer and the weight of each input. By increasing the number of the neurons in the hidden layer the algorithm [31] gives the optimal topology of the network. This is why many topologies are tested and only the best of them are taken. During the training step, each network receives as input a vector of the climatic condition parameters and the pollutant at time t. Each network generates the desired output that is the value of pollutant concentration at time t+PredictionHorizon. The forecast given by each network is passed as input to predict the air quality index using a MLP (Multi-Layered Perceptron). The MLP network is trained using the local air quality standards as shown in table 1. Air quality predictions for the different pollutants are obtained on a t+12 hours basis and give the most probable air quality category. The MLP model receives the predicted values of the five pollutants, COx, NOx, O3, PM10, and SOx, and predicts the index values for air quality ranging from 1 to 5, with 1 being very good and 5 being very bad. To train the MLP network we used back-propagation algorithm. The MLP network final topology, obtained after several trials, is: 5 neurons in the first hidden layer, 10 in the second and finally a linear neuron for the output layer. the concentration of Levenberg-Marquardt a A Multi-Agent based Approach for Air Pollution Informatica 39 (2015) 501 -- 505 505 The accuracy of the ANN models are given in table 2, and is calculated using one year's worth of data according to RMSE (Rooted Mean Squared Error) formulated in (11): 𝑅𝑀𝑆𝐸= (!!!!!)! !"#!!!!"# (11). Where lng is the length of the vectors, P and R are the predicted and measured values, respectively. The performances are computed using a validation data set that was not used in the training of the ANN models. Model PM10 SOX NOX COx O3 Topology [# of input neuron -# of hidden neuron] [10-320] [10-90] [10- 105] [10-45] [10-180] Validation error ( RMSE) 16.1945 µg/m 3 3.1618 µg/m3 9.7277 µg/m3 0.1220 µg/m3 39.8238 µg/m3 Table 2: Validation error of the ANN prediction models using the validation data set. 2.3 Decision-making mechanism Based on its internal state and the state of the environment, an agent has to choose an action to perform among all possible actions in order to reach its goals. This process is called decision-making. [32] presents a review of methods used for modelling decision-making in a coupled environmental and social system. Our system supports two cooperation strategies (centralized and evolutionary game) each one defines a decision- making mechanism. The centralized strategy (CS) is based on defining a central agent that represents the air pollution controlling agency. The central agent takes decisions according to the current air pollution level. The second strategy is based on an evolutionary game, where agents are rewarded and penalized according to the pollution levels; they make decisions according to their rewards. In our system, the cooperation strategy is defined within the simulation parameters. 2.3.1 Centralized Strategy (CS): The task of maintaining the air quality is assigned to an agent, which represents the air pollution control agency. It uses the GPD and ANN models to predict the air quality and pollutant levels, and according to the predictions it sends a reduce emission message to the emission agents. Then it will check the air quality. It will continue doing this until the air quality is improved to reach the air quality index goal. The central agent has absolute authority and its orders are executed by the emission source controllers. Agents communicate their emission rate at each simulation step to the central agent. This strategy is based on the communication between agents. We assume that agents are rational and have an environmental-responsible personality; this means they favour air quality improvement over their own interests and communicate their exact emission rate to the central agent. 2.3.2 Evolutionary Game Cooperating Strategy: each step of the simulation. A global goal GG can be In the EG strategy, every agent has its own goals (earning more rewards and keeping its emission rate high) and shares a global goal of maintaining air quality with other agents. The appreciation function defined as: 𝑎𝑝𝑝:𝑄→𝑅, allows comparing the air quality state at defined as (12). This means finding a set of actions 𝜋! to system to move to a new state of air quality 𝑞!!! that is 𝐺𝐺!={𝑇(𝜋!,𝑞!,𝑞!!!),𝑎𝑝𝑝𝑞!!! >𝑎𝑝𝑝𝑞!} (12) be performed by agents at time t, which allows the better than the current state. is Where: wi the weighting parameter where the payoff for the agent i for the time step t. An agent participates with other agents in the NPPD game, its own goal is to maximise its reward earned from the game. We adopted the approach of [33], where agents keep traces of their L previous steps (actions, rewards and its neighbours' rewards). To update the probabilities to increase or decrease the emission, we used [18] method. At each time step t the agent computes its weighted payoff, according to (13), and tries to maximise it (as its utility function) by taking it into consideration when computing its probability to increase or decrease its emission rate, respectively according to (14) and (15). (13) 𝑊𝑃!(𝑡)= 𝑤!∗𝑀!(𝑡) !!!! =1 and ∀ 𝑖,𝑗 (𝑖<𝑗→𝑤!>𝑤!), Mi(t) is 𝑤!"# !!!!"#!! 𝑃𝑐!𝑡+1 =𝑃𝑐!𝑡 +(1−𝑃𝑐!𝑡 ∗𝛼!𝑡 ,𝑖𝑓 𝑆!=0 𝑎𝑛𝑑 𝑊𝑃!(𝑡)>0 ∗𝑃𝑐!𝑡, 𝑖𝑓 𝑆!=0 𝑎𝑛𝑑 𝑊𝑃!(𝑡)≤0 𝑃𝑐!𝑡+1 = 1−𝛼!(𝑡) 𝑃𝑑!𝑡+1 =𝑃𝑑!𝑡 +(1−𝑃𝑑!𝑡 ∗𝛼!𝑡 ,𝑖𝑓 𝑆!=1 𝑎𝑛𝑑 𝑊𝑃!(𝑡)>0 𝑃𝑑!𝑡+1 = 1−𝛼!(𝑡) ∗𝑃𝑑!𝑡, 𝑖𝑓 𝑆!=1 𝑎𝑛𝑑 𝑊𝑃!(𝑡)≤0 0 𝑖𝑓 𝑋!,!=𝑋!,!!! 1 𝑖𝑓 𝑋!,!≠𝑋!,!!! Where: Pci and Pdi are respectively the probability to decrease (s=0) and increase (s=1) emissions for agent i, αi(t) is the learning rate of agent i at time step t, s is the strategy played at time t. The learning rate is updated according to (17): 𝐷!= 𝛼!𝑡+1 =𝛼!𝑡 +0.015 𝑖𝑓 𝐷!=0 𝛼!𝑡+1 =𝛼!𝑡 +0.010 𝑖𝑓 𝐷!>𝐿−1 𝛼!𝑡+1 =𝛼!𝑡 −0.010 𝑖𝑓 𝐷!≤𝐿−1 (17) (16) !!!!!! (15) (14) Where Di is the i-th agent actions homogeneity indicator, at time step t, Xi,j is j-th action of the agent i. Di is used to compare the last L actions of the agent. This is used to keep the agent from changing its actions. Agents are influenced by their neighbours, at each time; the average reward of the neighbours is calculated according to (18). 𝑛𝑃!(𝑡) = 𝑀!(𝑡))/𝑛𝑢𝑚𝑏𝑒𝑟𝑂𝑓𝑛𝑒𝑖𝑔ℎ𝑏𝑜𝑢𝑟𝑠! !"#$%&'(!%)*!!"#$%! !!! ( Where Mj(t) is the payoff of the neighbour j and numberOfneighboursi is the number of neighbours for the i-th agent. We keep a trace of the nP of the L previous (18) Informatica 39 (2015) xxx -- yyy Sabri Ghazi; Julie Dugdale; Tarek Khadir 506 (19) simulation steps and we compute their average in avgnP. The agent then uses the probabilities Pc, Pd and the average reward of its neighbours to choose an action according to (19): 𝑖𝑓 𝑆!𝑡 =0,𝑆!𝑡+1 = 1 ,𝑖𝑓 𝑊𝑃!<𝑎𝑣𝑔𝑁𝑃! 𝑎𝑛𝑑 𝑃𝑑!𝑡+1 >𝑃𝑐!𝑡+1, 𝑖𝑓 𝑆!𝑡 =1,𝑆!𝑡+1 = 0 ,𝑖𝑓 𝑊𝑃!<𝑎𝑣𝑔𝑁𝑃! 𝑎𝑛𝑑 𝑃𝑐!𝑡+1 >𝑃𝑑!𝑡+1, 0,𝑒𝑙𝑠𝑒 1,𝑒𝑙𝑠𝑒 At each simulation step, every agent gets a reward or penalty according to its actions and according to the pollution level. We have adopted the payoff curve (1) with b=2 and c=-0.5, but in the general case these parameters can be defined by the user. When the pollution level is higher than the maximum allowed value, the participation of the agent to the current level of the pollution 𝜎!(𝑡) is computed according to (20). !"!!!!"!"#,𝑃𝐿!𝑡 >𝑃𝐿!"# (20) 𝜎!𝑡 = !"!(!) Where, ERi(t) is the emission rate of the i-th agent at time t, PLe(t) is the pollution level of the pollutant e at time t and PLmax is the maximum allowed value for the pollutant level according to the regulation and local standards. The penalty for agent i at time step t is calculated according to (21): 𝐸𝑐𝑜𝐹𝑎𝑐𝑡𝑜𝑟!𝑡 =𝐸𝑐𝑜𝐹𝑎𝑐𝑡𝑜𝑟!𝑡−1 +(1− !!!!!) (21) Two penalising strategies were used; the first uses (21) and is a cumulative penalty. This means that the penalties from each step are kept and the agent is penalised as long as it continues to increase its emission. The second penalising method is not cumulative, and agents are penalised just according to the current simulation step. The reward of agent i, at the current time step t is computed according to (22), we compute the number of agents who choose to decrease their emission denoted ncp, after that we compute u use as defined in equation (1). 𝑀!(𝑡)= 𝑢𝑛𝑐𝑝 𝑖𝑓 𝑠!=0 𝑢𝑛𝑐𝑝 −𝐸𝑐𝑜𝐹𝑎𝑐𝑡𝑜𝑟!𝑡 𝑖𝑓 𝑠!=1 (22) 3 Simulation scenarios using data from the region of Annaba Annaba is a very industrialized region specialising in steel industries. The steel complex of Hadjar is located 12 kilometres south of the city of Annaba. The air pollution spreads over a radius of 6 km. According to [34], the complex annually releases into the atmosphere: 36890 tons of particles, 845 t of NOX, 30895 t of COx, 2260 t of SOx and 3093 t of NOx. The petrochemical station (ASMIDAL) produces fertilizers and pesticide products that have a big influence on air quality. 5 industrial zones, that contain hundreds of factories, are very close to the urban area and have a large impact on air pollution. The seaport is located in the centre of the city and attracts a lot of heavy transport, which also leads to deterioration in the air quality. The local pollution agency network provided hourly to from 01/01/2003 two-year period for a data parameters: Wind Speed 31/12/2004. The concentrations of air pollutants that have been continuously monitored are: Ozone (O3), Particulate Matter (PM10), Nitrogen Oxides (NOx), and Sulphur Oxides (SOx). The dataset also includes four meteorological (WS), Temperature (T) and relative Humidity (H). Daily rainfall measurements (RF) were also provided by the water management agency. The 2003 dataset was used for training the ANN and the 2004 dataset was used for validation; this helped us to assess the performance of the model. The pollutant concentration measurements are in microgram/m3 and they have been normalised using equation (23). 𝑉′!= !! (!"# (!!)!!"# (!!) (23) Where Vp is a parameter vector, min and max are functions that return the minimum and maximum values of the vector. Negative values, resulting from faulty measurements, were replaced using the mean of the previous and next values. It is impossible to discard faulty values since gaps in the time series will result in a data shift that affects the ANN training process leading to poor generalisation properties. Similarly, faulty (blank) measures for pollutants and weather parameters were replaced by an average of the v-q and u+q previous and future values respectively, with u being the faulty sample and q the number of values to take into consideration. This ensures the continuity and consistency of the time series and allows efficient the ANN predictors. Table 3 presents the statistical properties of the available data for different pollutants and weather parameters, for some parameters data are not available (N/A). training of We defined a simulation scenario for the Annaba region using the parameters in table 4. The goal levels for pollutants concentration were fixed according to the air quality index goal. For this scenario we aimed to reach a very good air quality level (Goal air quality index=1). The initial values (at t=0) for pollutant concentration and climatic parameters were fixed according to the dataset. Table 3: Statistical properties of the used dataset. (agents choosing For the EG strategies we fixed the initial proportion of cooperating agents to decrease emissions) to 0.5, this means that 50% of the agents decrease their emission at t=0. The value of this parameter was chosen following the work of [14] and [15]. The proportion will change during the simulation according to the game outcome. The prediction was for 2003 Parameter mean PM10 µg/m3 51.70 NOx µg/m3 14.50 SOx µg/m3 7.60 CO µg/m3 1.31 O3 µg/m3 N/A Wind Speed 2.65 µg/m3 Humidity (%) 63.52 Temperature 18.96 (°C) Rainfall (mm) N/A 2004 mean 27.76 N/A N/A N/A 42.27 2.12 71.92 16.82 2.96 2004 2003 STD STD 26.38 51.66 25.01 N/A 14.78 N/A N/A 0.52 N/A 64.58 1.27 1.78 14.33 16.50 6.30 7.76 N/A 9.27 Max value 508 435.0 190.0 12.2 688.0 9.6 93.0 42.1 73.9 A Multi-Agent based Approach for Air Pollution Informatica 39 (2015) 501 -- 505 507 to the random values used in the initialisation of some variables (neighbours rewards, first chosen action, weights, k last actions and rewards). The comparison is done according to the air quality index. Results are expressed in terms of the number of occurrences of air quality index as illustrated in figure 2, for example the number of times the air quality index equals 1 (very good). Figure 3, shows the evolution of the air quality index over time. The CS gives the best performance. With the CS the air quality index moves rapidly from bad to average and then to good and finally stabilises at very- good (which is the goal fixed in the simulation scenario). The EG-CP moves the index from bad to average, when the equilibrium is reached it stabilises in good and never reaches a very-good index. The EG-NCP strategy moves the air quality from bad to average and never improves. When penalties are not used (EG-NP) the air quality stabilises at bad. When cooperation is not used (NC), agents act selfishly and do not care about the pollution, therefore, the air quality oscillates between bad and very- bad. As the agents reach their maximum emission rate we can observe an oscillation, which is caused by the climatic conditions. The only thing that affects the pollutant concentration is the climatic conditions (the emission rate is constant); these have a big influence and are captured with the ANN model. 100 100 100 100 2000 (gram/hour). 20 µ gram/m3 30 µ gram/m3 45 µ gram/m3 45 µ gram/m3 0.5 the next 2 hours, the same as the simulation step. Each source emits according to its emission rate which cannot be higher than the maximum level defined in the simulation scenario. The position of sources was randomly generated and many sources are located in the same box. Value Parameter Name Polluting activities and Policy parameters Number of PM10 sources Number of SOX sources Number of NOx sources Number of CO sources Max emission rate Goal PM10 level Goal SOx level Goal NOx level Goal O3 level Number of memory steps (L) 4 steps Initial proportion of cooperating agents Environment parameters Number of boxes Temperature at t=0 Humidity at t=0 Wind Speed t=0 PM10 at t=0 SOX at t=0 NOX at t=0 CO at t=0 O3 at t=0 Air Quality at t=0 Total simulation time K Simulation step Prediction horizon 20 12.7 (°C) 71.0 % 2.4 m/s 13.0 µ gram/m3 17.0 µ gram/m3 2.0 µ gram /m3 0.5 µ gram /m3 29.0 µgram /m3 2 ( Good) 4900 hours 1 step = 2 hours Next 2 hours Table 4: Parameter values of the simulation scenario. 4 Results and discussion: We have built a simulator using the approach described above. We used the JADE agent framework [35] and ANN models from Encog [36]. We have defined 5 strategies: EG-CP (Evolutionary Game with Cumulative Penalty), EG-NCP (Evolutionary Game with No Cumulative Penalties), EG-NP (Evolutionary Game with No Penalty), CS (Centralized Strategy) and NC (No-Cooperation). The last one is used for comparison purposes. Using the parameters shown in table 4, we chose a strategy and ran the simulation 16 times. We then changed the strategy and ran the simulation again 16 times; since we have 5 strategies we obtain 80 simulations. The most explicative results are presented. For the CS and NC cases the simulator showed similar results for each run. For the EG strategies there were small differences between runs, especially concerning the proportion of cooperating agents. These changes are due Figure 2: Air quality index using 5 different cooperation strategies Figure 3: Air Quality index for 4900 hours. Informatica 39 (2015) xxx -- yyy Sabri Ghazi; Julie Dugdale; Tarek Khadir 508 Figure 4: Concentration of PM10 for the four tested cooperation strategies compared with the no-cooperation strategy. Figure 6: Proportions of cooperating agents for EG-CP, EG-NCP and EG-NP. Figure 5 : Concentration of SOX (a), NOx (b) and O3 (c) using the four strategies and the no-cooperation scenario. Figure 4 shows the evolution of the PM10 concentration during the simulation time. The PM10 concentration shows many peaks compared with the other pollutants under the same climatic conditions. This is due to the dry nature of the weather in the Annaba area, with wildfires, and sandstorms coming from the great Sahara desert. These events have a big effect on the PM10 concentration but not on the other pollutants. The CS strategy takes less time to control the pollution level and keep it below the goal level defined in the simulation parameters. All of the EG strategies take longer, keeping it close to the goal level, but without ever reaching it. The penalising regulations have a big effect on the PM10 level. As illustrated, the EG-CP (cumulative penalising method) controls the pollution better than the non-cumulative one, and both methods perform better than the no- penalising strategy. The no-cooperation is presented in order to show the impact of cooperation on the PM10 level. Figure 5 shows the evolution of the SOx, NOx and Ozone concentrations during the simulation time using four different cooperation strategies. The CS strategy gives the best performance since the pollution concentration rapidly decreases. The EG strategies show the same performance as for PM10 and the pollution level is widely influenced by the selected penalising method. The CP strategy appears to be the best one followed by the NCP. The pollution slowly decreases, but not enough to reach the goal level if penalisation is not used. Figure 7: The proportion of cooperating agents according to the emitted pollutant for the EG-CP strategy. A Multi-Agent based Approach for Air Pollution Informatica 39 (2015) 501 -- 505 509 5 Conclusions Figure 8, The proportion of cooperating agents for the EG-NCP strategy. Figure 9: The proportions of cooperating agents, for the 4 groups of agents, when penalisation is not used. The NC strategy gives the worst levels; when all agents are emitting pollution using their maximum emission rate, the pollutant level reaches alarming values and peak periods occur. Figure 6 shows how the penalising method affects the proportion of cooperating agents. For the case of EG-CP the game equilibrium is reached at time step 387 and stabilises when the proportion of cooperating agents is between 0.93 and 0.95. The EG-NCP strategy stabilises early at time step 196 and oscillates between a cooperation ratio of 0.73 and 0.80, after which the equilibrium is fixed at 0.77. The EG-NP is the slowest; the equilibrium is reached at time step 808 with a cooperation ratio of 0.57. This happens because the agents are not penalised since the strategy does not include penalising methods. Figures 7, 8 and 9 show, respectively, the proportion of cooperating agents according to the pollutant for the three penalising strategies EG-CP, EG-NCP and EG-NP. The PM10 agents gives the highest cooperation ratio. This is because many peaks occur with this pollutant and the others pollutants cooperating ratio (NOx and SOx) are influenced by the O3 concentration. The more the pollutant exceeds the allowed level, the greater the proportion of co-operators. The equilibriums are disturbed by the pollution level, because, when the pollution level reaches a level above the goal level, agents start being penalised, and thus they tend to cooperate more. Anthropogenic activities are among the main causes of pollution and environmental problems. These activities have to be included in the simulation models. Modelling the interaction between social and ecological components is a very important aspect. A MAS approach allows us to model the social network of human-beings sharing the exploitation of common environmental resources. Manipulating the behaviour at an individual and group level helps to gain more knowledge about the impact of human decision- making on pollution and makes the simulation more realistic. Studies treating air pollution are usually concerned with the physical aspects (concentration and dispersion of pollutant), and do not include human-decision factors on the emission sources.. In our approach, we model the decision-making activity of the air pollution emission source managers. This helps to investigate the conditions and regulations that may enhance and maintain the air quality. (CS) showed the strategy We used a two stage air pollution modelling method: a GPD dispersion model and an ANN forecasting model. The ANN predictor uses climatic parameters and dispersal information provided by the GPD model to make predictions. This helped to introduce the effect of uncertainty caused by the weather and made the simulation more realistic. Five cooperation strategies were tested. The centralized cooperation best performance, surpassing the reward/penalty strategies. However, the CS strategy needs an effective communication network between emission sources controllers and the regulation agency. Also, we assume that emission controllers communicate exactly their emission rate, which is not always the case. The reward/penalty strategies seem to be more realistic; penalising the polluting agents according to their participation during peak periods has a big influence on their behaviours. As shown in the simulation results, it helps reducing the pollution level and affects the evolution of the pollutant. Thus, air pollution regulations have a big impact on pushing the emission source controllers to take their polluting activity seriously; this is especially important during the peak periods where climatic conditions cause the pollutants to stagnate. To summarise our study helps to: (1) Model and introduce concerning emission sources and the process of simulating air pollution evolution. the possible cooperation between the actors concerned in managing the air quality. (3) Have a prediction about the efficiency of regulation rules for preserving the air quality. (4) Investigate the impact on air quality of the decision to expand or establish a new emission points. Our work aims to provide a decision-making tool to the air pollution control agencies that will help them evaluate the regulations and policies concerning air pollution control. The current version of the system deals only with point emission sources. In future decision-making (2) Evaluate human Informatica 39 (2015) xxx -- yyy Sabri Ghazi; Julie Dugdale; Tarek Khadir 510 and versions we aim to include line and area sources. Line sources model the road activities, whereas area sources model the waste management and agricultural sources of pollution. If data becomes available in future it could be interesting to experiment with different time representations. Fortunately the multi- agent system approach allows us to easily change to a different scale of time representation in the same simulation. We can envisage using one time representation for another monitoring. The first can help us to see long term impacts (e.g. investing in less polluting activities), and the second can help to see the short term changes. for decisions The simulator may also be enhanced by including topographic aspects of regions since this has a big influence on the dispersion of air pollutants. In addition, including more agent personalities and exploring other cooperation strategies are also among our future plans. Our system is designed in a generic way and it could be adapted for other types of pollution such as water pollution. This could be done by changing the current dispersion and the prediction models to a water pollution dispersion model. 6 References [1] Folke, Carl. Resilience: The emergence of a perspective systems analyses. Global environmental change, 16.3, 2006. pp. 253-267. social -- ecological [2] WHO (World Health Organisation), Ecosystems and Human Well-being: Health Synthesis . WHO Library Cataloguing-in-Publication Data, 2005. for [3] Kryza, M., Netzel, P., Drzeniecka-Osiadacz, A., Werner, M., & Dore, A. J. (2014). Forecasting the Ozone Concentrations with WRF and Artificial Neural Network Based System. In Air Pollution Modeling and its Application XXII Springer Netherlands, 2014, pp. 605-609. [4] Holmes, NS and Morawska, L., A Review of Dispersion Modelling and its application to the dispersion of particles: An overview of different dispersion models available. Atmospheric Environment, 40(30), 2006. pp. 5902-5928. [5] Lushi, Enkeleida, and John M. Stockie. An inverse Gaussian plume approach for estimating atmospheric pollutant emissions from multiple point sources. Atmospheric Environment, 44.8, 2010. pp. 1097-1107. [6] Aulinas Montse, Turon Clàudia and Sànchez- Marrè Miquel. Agents as a Decision Support Tool in Environmental Processes:The State of the Art. Whitestein Series in Software Agent Technologies and Autonomic Computing. 2009. pp. 5 -- 35. [7] Ghazi, Sabri, Khadir, Tarek, Julie Dugdale. Multi-Agent Based Simulation of Environmental Pollution Issues: A Review. Highlights of Practical Applications of Heterogeneous Multi- Agent Systems. The PAAMS Collection. Springer International Publishing, 2014. pp. 13- 21. [8] Hülsmann, F., Gerike, R., Kickhöfer, B., Nagel, K., & Luz, R. . Towards a multi-agent based modeling approach for air pollutants in urban regions. In Proceedings of the Conference on "Luftqualität an Strassen, 2011. pp. 144-166. [9] Kickhöfer, B., Hülsmann, F., Gerike, R., & Nagel, K. Rising car user costs: comparing aggregated and geo-spatial impacts on travel demand and air pollutant emissions. Smart Transport Networks: Decision Making, Sustainability and Market structure, NECTAR Series on Transportation and Communications Networks Research, 2012, pp.180-207. [10] Kickhöfer, B., & Nagel, K. Towards high- resolution first-best air pollution tolls. Networks and Spatial Economics, 2013. pp. 1-24. [11] Papaleonidas Andonis, Iliadis Lazaros. Hybrid and Reinforcement Multi Agent Technology for Real Time Air Pollution Monitoring. Artificial Intelligence Applications and Innovations IFIP Advances in Information and Communication Technology Volume 381, 2012, pp. 274-284. [12] Ahat M., Ben Amor S., Bui M., Lamure M., Marie-Françoise Courel. Pollution Modeling and Simulation with Multi-Agent and Pretopology. Complex Sciences, Lecture Notes of the Institute for Computer Sciences, Social Informatics and Telecommunications Engineering Volume 4, 2009, pp.225-231. [13] Boyd, Robert, Richerson, Peter J. The evolution of reciprocity in sizable groups. Journal of Theoretical Biology, vol. 132, no 3, 1988, pp. 337-356. [14] Szilagyi Miklos N., An Investigation of N-person Prisoners' Dilemmas. Complex Systems, 14, 2003. pp. 155 -- 174. [15] Szilagyi Miklos N., Somogyi Iren, The Shape of Payoff Functions in N-person Games, Journal of Game 1(3), doi: 10.5923/j.jgt.20120103.02. 2012,10-14 Theory. [16] Xianyu, Bo, and Ping Chen. Does Social Welfare Preference Always Promote Cooperation on Barabási and Albert Networks ?. Computational Economics, 37.3, 2011, pp. 249-266. [17] Barabási, Albert-László, and Réka Albert. in random networks. Emergence of scaling science 286.5439, 1999: 509-512. [18] Conrad Power, A Spatial Agent-Based Model of N-Person Prisoner's Dilemma Cooperation in a Socio-Geographic Community, of Artificial Societies and Social Simulation vol. 12, 2009, http://jasss.soc.surrey.ac.uk/12/1/8.html. Journal [19] Hardin, Garrett. The tragedy of the commons. 18, no. Science, 162, 1968. pp. 1243-1248. [20] Osang, Thomas, and Arundhati Nandy. Environmental regulation of polluting firms: Porter's hypothesis revisited. Brazilian Journal of Business Economics. 3.3, 2003. pp. 128-148. A Multi-Agent based Approach for Air Pollution Informatica 39 (2015) 501 -- 505 511 memory, learning and dynamic preferential selection in spatial prisoner's dilemma game. In Journal of Physics: Conference Series, IOP Publishing, Vol. 96, No. 1, 2008, p. 012107. du lapin rejetées par [34] Tadjine, A., Djebar, H., & Courtois, A. Toxicité des poussières le complexe sidérurgique d'Annaba sur quelques paramètres hématologiques Europeus. Environnement, Risques & Santé, 7(3), 2008, 209-215. [35] Bellifemine, Fabio, POGGI, Agostino, et RIMASSA, Giovanni. JADE -- A FIPA-compliant agent framework. In : Proceedings of PAAM. 1999. p. 33. [36] Heaton, J. ,Encog: Library of Interchangeable Machine Learning Models for Java and C#. Journal of Machine Learning Research, 16, 2015., 1243-1247. [21] Lange Andreas, , Vogt Carsten, , Cooperation in international environmental negotiations due to a preference Journal of Public Economics, Volume 87, Issues 9 -- 10, September 2003, pp. 2049 -- 2067. for equity, [22] Touza, J., Drechsler, M., Smart, J. C., & Termansen, M. Emergence of cooperative behaviours the management of mobile ecological resources. Environmental Modelling & Software, 45, 2013. pp. 52-63. [23] De Vries, HJM Bert. Environmental Modelling. : Principles of Environmental Sciences. In Springer Netherlands, 2009. pp. 345-373. in [24] Stockie, John M. The mathematics of atmospheric dispersion modeling., Siam Review, 53.2, 2011. pp. 349-372. [25] Hanna, Steven R., Gary A. Briggs, and Rayford P. Hosker Jr. Handbook on atmospheric diffusion. No. DOE/TIC-11223. National Oceanic and Atmospheric Administration, Oak Ridge, TN (USA). Atmospheric Turbulence and Diffusion Lab., 1982. [26] User's guide for the industrial source complex (ISC3) dispersion models volume II -description of model algorithms, U.S. Environmental Protection 1995, http://www.epa.gov/scram001/userg/regmod/isc3 v2.pdf (visited on 11/22/2014). Agency, [27] Behera, Sailesh N. , Sharma, Mukesh. Degradation of SO2, NO2 and NH3 leading to formation of secondary inorganic aerosols: An environmental chamber study. Atmospheric environment, vol. 45, no 24, 2011 pp. 4015- 4024. in for [28] Ghazi, Sabri,, Khadir Med Tarek. Combination of artificial neural network models for air quality predictions for the region of Annaba, Algeria. International Journal of Environmental Studies, 69.1, 2012.pp. 79-89. [29] Pelliccioni, Armando et Tirabassi, Tiziano. Air dispersion model and neural network: A new perspective the simulation of complex situations. Environmental modelling & software, vol. 21, no 4, (2006). pp. 539-546. integrated models [30] Khedairia, S., Mohamed Khadir T.. Impact of clustered meteorological parameters on air pollutants concentrations region of Annaba, Algeria, Atmospheric Research 113, 2012. pp. 89-101. [31] Chen, S., Cowan, C.F., and Grant, P. M., Orthogonal Least Squares Learning Algorithm for Radial Basis Function Networks, IEEE Transactions on Neural Networks, 2(2), 1991, 302-309. [32] Li An, Modeling human decisions in coupled human and natural systems: Review of agent- based models, Ecological Modelling 229, 2012, 25 -- 36. [33] Zhao-Han Sheng., Hou, Y. Z., Wang, X. L., & Du, J. G. ,The evolution of cooperation with the in
1812.01130	1	1812	2018-12-03T23:14:59	A Unified Approach to Dynamic Decision Problems with Asymmetric Information - Part I: Non-Strategic Agents	[ "cs.MA", "math.OC" ]	We study a general class of dynamic multi-agent decision problems with asymmetric information and non-strategic agents, which includes dynamic teams as a special case. When agents are non-strategic, an agent's strategy is known to the other agents. Nevertheless, the agents' strategy choices and beliefs are interdependent over times, a phenomenon known as signaling. We introduce the notions of private information that effectively compresses the agents' information in a mutually consistent manner. Based on the notions of sufficient information, we propose an information state for each agent that is sufficient for decision making purposes. We present instances of dynamic multi-agent decision problems where we can determine an information state with a time-invariant domain for each agent. Furthermore, we present a generalization of the policy-independence property of belief in Partially Observed Markov Decision Processes (POMDP) to dynamic multi-agent decision problems. Within the context of dynamic teams with asymmetric information, the proposed set of information states leads to a sequential decomposition that decouples the interdependence between the agents' strategies and beliefs over time, and enables us to formulate a dynamic program to determine a globally optimal policy via backward induction.	cs.MA	cs	A Unified Approach to Dynamic Decision Problems with Asymmetric Information - Part I: Non-Strategic Agents Hamidreza Tavafoghi, Yi Ouyang, and Demosthenis Teneketzis Abstract We study a general class of dynamic multi-agent decision problems with asymmetric information and non-strategic agents, which includes dynamic teams as a special case. When agents are non-strategic, an agent's strategy is known to the other agents. Nevertheless, the agents' strategy choices and beliefs are interdependent over times, a phenomenon known as signaling. We introduce the notions of private information that effectively compresses the agents' information in a mutually consistent manner. Based on the notions of sufficient information, we propose an information state for each agent that is sufficient for decision making purposes. We present instances of dynamic multi-agent decision problems where we can determine an information state with a time-invariant domain for each agent. Furthermore, we present a generalization of the policy-independence property of belief in Partially Observed Markov Decision Processes (POMDP) to dynamic multi-agent decision problems. Within the context of dynamic teams with asymmetric information, the proposed set of information states leads to a sequential decomposition that decouples the interdependence between the agents' strategies and beliefs over time, and enables us to formulate a dynamic program to determine a globally optimal policy via backward induction. 8 1 0 2 c e D 3 ] A M . s c [ 1 v 0 3 1 1 0 . 2 1 8 1 : v i X r a A preliminary version of this paper will appear in the Proceeding of the 57th IEEE Conference on Decision and Control is with the Department of Mechanical Engineering at (CDC), Miami Beach, FL, December 2018 [1]. H. Tavafoghi the University of California, Berkeley (e-mail: [email protected]). Y. Ouyang is with Preferred Networks America, Inc. (e-mail: [email protected]). D. Teneketzis is with the Department of Electrical Engineering and Computer Science at the University of Michigan, Ann Arbor (e-mail: [email protected]) This work was supported in part by the NSF grants CNS-1238962, CCF-1111061, ARO-MURI grant W911NF-13-1-0421, and ARO grant W911NF-17-1-0232. November 23, 2018 DRAFT 2 A. Background and Motivation I. INTRODUCTION Dynamic multi-agent decision problems with asymmetric information have been used to model many situations arising in engineering, economic, and socio-technological applications. In these applications many decision makers/agents interact with each other as well as with a dynamic system. They make private imperfect observations over time, and influence the evolution of the dynamic system through their actions that are determined by their strategies. An agent's strategy is defined as a decision rule that the agent uses to choose his action at each time based on his realized information at that time. In this paper, we study a general class of dynamic decision problems with non-strategic agents. We say an agent is non-strategic if his strategy (not his specific action) is known to the other agents. In a companion paper [2] we study dynamic decision problems with strategic agents where an agent's strategy is his private information and not known to the other agents. We consider an environment with controlled Markovian dynamics, where, given the agents' actions at every time, the system state at the next time is a stochastic function of the current system state. The instantaneous utility of each agent depends on the agents' joint actions as well as the system state. At every time, each agent makes a private noisy observation that depends on the current system state and past actions of all agents in the system. Therefore, agents have asymmetric and imperfect information about the system history. Moreover, each agent's information depends on other agents' past actions and strategies; this phenomenon is known as signaling in the control theory literature. In such problems, the agents' decisions and information are coupled and interdependent over time because (i) an agent's utility depends on the other agents' actions, (ii) the evolution of the system state depends, in general, on all the agents' actions, (iii) each agent has imperfect and asymmetric information about the system history, and (iv) at every time an agent's information depends, in general, on the agents' (including himself) past actions and strategies. There are two main challenges in the study of dynamic multi-agent decision problems with asymmetric information. First, because of the coupling and interdependence among the agents' decisions and information over time, we need to determine the agents' strategies simultaneously for all times. Second, as the agents acquire more information over time, the domains of their strategies grow. November 23, 2018 DRAFT 3 In this paper, we propose a general approach for the study of dynamic decision problems with non-strategic agents and address these two challenges. We propose the notion of sufficient information and provide a set of conditions sufficient to characterize a compression of the agents' private and common information in a mutually consistent manner over time. We show that such a compression results in an information state for each agent's decision making problem. We show that restriction to the set of strategies based on this information state entails no loss of generality in dynamic decision problems with non-strategic agents. We identify specific instances of dynamic decision problems where we can discover a set of information states for the agents that have time-invariant domain. Within the context of dynamic teams, we further demonstrate that the notion of sufficient information leads to a sequential decomposition of dynamic teams. This sequential decomposition results in a dynamic program the solution of which determines the agents' globally optimal strategies. B. Related Literature The Partially Observed Markov Decision Processes (POMDPs), i.e. centralized stochastic control problems, present the simplest form of dynamic decision problems with single agent [3], [4]. To analyze and identify properties of optimal strategies in POMDPs the notion of information state is introduced as the agent's belief about the current system state conditioned on his information history. The information state provides a way to compress the agent's information over time that is sufficient for the decision-making purposes. When the agent has perfect recall, this information state is independent of the agent's strategies over time; this result is known as the policy-independence belief property [3]. Dynamic multi-agent decision problems with non-strategic agents are considerably more dif- ficult compared to their centralized counterparts. This is because, due to signaling, they are (in general) non-convex functional optimization problems (see [5] -- [8]). The difficulties present in these problems were first illustrated by Witsenhausen [9], who showed that in a simple dynamic team problem with Gaussian primitive random variables and quadratic cost function where signaling occurs, linear strategies are suboptimal (contrary to the corresponding centralized problem where linear strategies are optimal). Subsequently, many researchers investigated control problems with various specific information structures such as: partially nested ( [10] -- [15] and references therein), stochastic nested [16], randomized partially nested [17], delayed sharing ( [11], [18] -- [20] and references therein), information structures possessing the i-partition property November 23, 2018 DRAFT 4 or the s-partition property [21], the quadratic invariance property [22], and the substitutability property [23]. Currently, there are three approaches to the analysis of dynamic multi-agent decision problems with non-strategic agents: the agent-by-agent approach [24], the designer's approach [25], and the common information approach [26]. We provide a brief discussion of these approaches here. We discuss them in details in Section VI-B, where we compare them with the sufficient information approach we present in this paper and show that our approach is distinctly different from them. The agent-by-agent approach [24], is an iterative method. At each iteration, we pick an agent and fix the strategy of all agents except that agent, and determine the best response for that agent and update his strategy accordingly. We proceed in a round robin fashion among the agents until a fixed point is reached, that is, when no agent can improve his performance by unilaterally changing his strategy. The designer's approach [25], considers the decision problem from the point of view of a designer who knows the system model and the probability distribution of the primitive random variables, and chooses the control strategies for all agents without having an information about the realization of the primitive random variables. The common information approach [26], assumes that at each time all agents possess private information and share some common information; it uses the common information to coordinate the agents' strategies sequentially over time. C. Contribution We develop a general methodology for the study and analysis of dynamic decision problems with asymmetric information and non-strategic agents. Our model includes problems with non- classical information structures [19] where signaling is present. We propose an approach that effectively compresses the agents' private and common information in a mutually consistent manner. As a result, we offer a set of information states for the agents which are sufficient for decision making purposes. We characterize special instances where we can identify an information state with a time-invariant domain. Based on the proposed information state, we provide a sequential decomposition of dynamic teams over time. We show that the methodology developed in this paper generalizes the existing results for dynamic teams with non-classical information structure. Our results in this paper, along those appearing in the companion paper [2] present a set of information states sufficient for decision making in strategic and non-strategic November 23, 2018 DRAFT 5 settings. Therefore, we provide a unified approach to decision making problems that can be used to study dynamic games and dynamic teams as well as dynamic games among teams of agents. D. Organization The rest of the paper is organized as follows. In Section II, we describe the model and present few examples. In Section III, we discuss the main challenges that are present in dynamic multi-agent decision problems with non-strategic agents. We present the sufficient information approach in Section IV. We present the main results of the paper in Section V. We discuss an open problem associated with the sufficient information approach in Section VI-A. In Section VI-B, we compare the sufficient information approach with the existing approaches in the literature. We provide a generalization of the sufficient information approach in Section VII. We present an extension of our results to infinite-horizon dynamic multi-agent decision problems with non- strategic agents in Section VIII. We conclude in Section IX. The proofs of all the theorems and lemmas appear in the Appendix. Notation t t t t t , . . . ,X N (resp. functions f 1 t , . . . ,X N t t (·), . . . ,f N t ) (resp. ft(·) := (f 1 Random variables are denoted by upper case letters, their realizations by the corresponding lower case letters. In general, subscripts are used as time index while superscripts are used to index agents. For t1≤ t2, Xt1:t2 (resp. ft1:t2(·)) is the short hand notation for the random variables (Xt1,Xt1+1, ...,Xt2) (resp. functions (ft1(·), . . . ,ft2(·))). When we consider a sequence of random variables (resp. functions) for all time, we drop the subscript and use X to denote X1:T (resp. f (·) to denote f1:T (·)). For random variables X 1 t (·)), t (·))) to denote the vector of the set we use Xt := (X 1 of random variables (resp. functions) at t, and X−n t ) (resp. := (X 1 t (·))) to denote all random variables (resp. functions) f−n at t except that of the agent indexed by n. P(·) and E(·) denote the probability and expectation of an event and a random variable, respectively. For a set X , ∆(X ) denotes the set of all beliefs/distributions on X . For random variables X,Y with realizations x,y, P(xy) := P(X = xY = y) and E(Xy) := E(XY = y). For a strategy g and a belief (probability distribution) π(·)) to indicate that the probability (resp. expectation) depends on the π, we use Pg choice of g and π. We use 1{X=x} to denote the indicator function for event X = x. For sets A and B we use A\B to denote all elements in set A that are not in set B. For random variables X and Y we write X dist.= Y when X and Y have an identical probability distribution. (·) := (f 1 t (·), . . . ,f n−1 π(·) (resp. Eg (·),f n+1 (·), . . . ,f N t (·), . . . ,f N t , . . . ,X n−1 t ,X n+1 , . . . ,X N t November 23, 2018 DRAFT II. MODEL 6 1) System dynamics: Consider N non-strategic agents who live in a dynamic Markovian world over a horizon T :={1,2, ...,T}, T <∞. Let Xt ∈Xt denote the state of the world at t∈T . At time t, each agent, indexed by i∈N :={1,2, ...,N}, chooses an action ai t denotes the set of available actions to him at t. Given the collective action profile At := (A1 t ), the state of the world evolves according to the following stochastic dynamic equation, t, where Ai t∈Ai t , ...,AN Xt+1 = ft(Xt, At, W x t ), (1) 1:T−1 is a sequence of independent random variables. The initial state X1 is a random where W x variable that has a probability distribution η ∈ ∆(X1) with full support. At every time t ∈ T , before taking an action, agent i receives a noisy private observation t ∈ Y i Y i t of the current state of the world Xt and the action profile At−1, given by Y i t = Oi t(Xt, At−1, W i t ), (2) 1:T , i ∈ N , are sequences of independent random variables. Moreover, at every t ∈ T , where W i all agents receive a common observation Zt ∈ Zt of the current state of the world Xt and the action profile At−1, given by Zt = Oc t (Xt, At−1, W c t ), (3) 1:T , and W i 1:T , i ∈ N are mutually independent. 1:T , is a sequence of independent random variables. We note that the agents' actions where W c At−1 is commonly observable at t if At−1 ⊆ Zt. We assume that the random variables X1, W x agents' aggregate information is given by Ht :=(cid:81) 1:T−1, W c 2) Information structure: Let Ht ∈ Ht denote the aggregate information of all agents at time t. Assuming that agents have perfect recall, we have Ht = {Z1:t, Y 1:N 1:t−1}, i.e. Ht denotes the set of all agents' past observations and actions. The set of all possible realizations of the τ <t Ai τ. At time t ∈ T , the aggregate information Ht is not fully known to all agents; each agent may have asymmetric information about Ht. Let Ct := {Z1:t} ∈ Ct denote the agents' common information about Ht and P i t denote agent i's private information about t and Ct denote the set of all possible realizations of agent i's private and common Ht, where P i information at t, respectively. In this paper, we discuss several instances of information structures τ≤t Zτ ×(cid:81) τ ×(cid:81) 1:t−1}\Ct∈ P i 1:t i∈N(cid:81) i∈N(cid:81) t :={Y i 1:t,Ai , A1:N τ≤t Y i November 23, 2018 DRAFT that can be captured as special cases of our general model. 7 3) Strategies and Utilities: Let H i i at t, where Hi i's strategy gi := {gi that determine agent i's action Ai t :={Ct,P i t}∈ Hi t denote the information available to agent t denote the set of all possible realizations of agent i's information at t. Agent t), t ∈ T , t, t ∈ T }, is defined as a sequence of mappings gi t : Hi t → ∆(Ai t of his history at t∈ T . t∈ Hi t for every realization hi Agent i's instantaneous utility at t depends on the state of the world Xt and the collective t(Xt, At). Therefore, agent i's total utility over the horizon action profile At and is given by ui T is given as U i(X1:T , A1:T ) := (cid:88) t∈T ui t(Xt, At). (4) We assume that agents are non-strategic. That is, each agent's, say i's, i∈N , strategy choice gi is known to other agents. We note that these non-strategic agents may have different utilities over time. Therefore, the model includes a team of agents sharing the same utilities (see Sections V) as well as agents with general non-identical utilities. In [2] we build on our results in this paper to study dynamic decision problems with strategic agents where an agent may deviate privately from the commonly believed strategy, and gain by misleading the other agents. To avoid measure-theoretic technical difficulties and for clarity and convenience of exposition, we assume that all the random variables take values in finite sets. Assumption 1. (Finite game) The sets Xt, Zt, Y i t, Ai t, i ∈ N , t ∈ T , are finite. Special Cases: We present several instances of dynamic decision problems with asymmetric information that are special cases of the general model described above. 1) Real-time source coding-decoding [27]: Consider a data source that generates a random sequence {X1,...,XT} that is k-th order Markov, i.e. for every sequence of realizations x1:T , P{Xt+k:T = xt+k:Tx1:t+k−1} =P{Xt+k:T = xt+k:Txt:t+k−1} for t≤ T − k. There exists an encoder (agent 1) who observes Xt at every time t; the encoder has perfect recall. At every time t, based on his available data {X1,...,Xt}, the encoder transmits a signal Mt ∈ Mt through a noiseless channel to a decoder (agent 2), where Mt denotes the transmission alphabet. At the receiving end, at every time t, the decoder wants to estimate the value of Xt−1−δ (with delay δ) as Xt−1−δ based on his available data M1:t−1; we assume that the decoder has perfect recall. The November 23, 2018 DRAFT 8 total distortion function given by (cid:80)T encoder and decoder choose their joint coding-decoding policy so as to minimize the expected t=2+δ dt(Xt, Xt), where dt(·,·) denotes the instantaneous distortion function. To capture the above-described model within the context of our model, we need to define an augmented system state Xt that includes the last max(k, δ + 1) states realizations as Xt :={Xt−max(k,δ+1)+1,...,Xt}. Moreover, the encoder's (agent 1's) observation is given by Y 1 t = t ) = (Mt, Xt−1−δ). The encoder's and decoder's instantaneous t ( Xt, At−1) = Xt and the decoder's (agent 2's) observation is given by Y 2 t = O1 t ( Xt,At−1) = Mt−1, where (A1 O2 utility are given by a distortion function uteam t ,A2 t ( Xt,At) = dt(Xt−1−δ, Xt−1−δ). 2) Delayed sharing information structure [18] -- [20], [28]: Consider a N-agent decision prob- lem where agents observe each others' observations and actions with d-step delay. We note that in our model we assume that the agents' common observation Zt at t is only a function of Xt and and At−1. Therefore, to describe the decision problem with delayed sharing information structure within the context of our model we need to augment our state space to include the agents' last d observations and actions as part of the augmented state. Define Xt := {Xt, M 1 t } t , ..., M d t := {At−i, Yt−i} ∈ At−i × Yt−i, i ∈ N ; that is, M i as the augmented system state where M i serves as a temporal memory for the agents' observations Yt−i and actions At−i at t−i. Then, we have Xt+1 = {Xt+1, M 1 } and Zt ={M d t } ={Yt−d, At−d}. t+1} = {ft(Xt, At, W x t , ..., M d−1 t ), (Yt, At), M 1 t+1, ..., M d t+1, M 2 t , M 2 t t 3) Real-time multi-terminal communication [29]: Consider a real-time communication system with two encoders (agents 1 and 2) and one receiver (agent 3). The two encoders make distinct t of a Markov source. The encoders' observation are conditionally observations X 1 independent Markov chains. That is, there is an unobserved random variable variable R such that P{X 1 t and X 2 1,X 2 1,R} =P{X 1 P{X 1 1R}P{X 2 t+1X 1 1R}P{R}, and t , R} =P{X 1 t , X 2 t+1, X 2 t+1X 1 t , R}P{X 2 t+1X 2 t , R}. Markov source X 1 t , X t 2 X 1 t X 2 t Encoder 1 1:t,M 1 g1 t(X 1 1:t−1) Encoder 2 1:t,M 2 g1 t(X 2 1:t−1) M 1 t M 2 t Channel 1 t M 1 t (Y 1 t−1) Q1 Channel 2 t M 2 t (Y 2 t−1) Q2 Y 1 t Receiver Xt Y 2 t g3 t(Y 1:2 1:t−1) Each encoder encodes, in real-time, its observations into a sequence of discrete symbols and November 23, 2018 DRAFT 9 t(··), i = 1, 2. sends it through a memoryless noisy channel characterized by a transition matrix Qi The receiver wants to construct, in real time, an estimate Xt of the state of the Markov source based on the channels' output Y 1 1:t. All agents have the same instantaneous utility given by a distortion function dt(Xt, Xt). 1:t,Y 2 4) Optimal remote and local controller [30], [31]: Consider a decentralized control problem for a Markovian plant with two controllers, a local controller (agent 1) and a remote controller (agent 2). A2 t Plant ft(Xt, A1 t , A2 t ) A1 t Remote Controller gt(Y1:t, A2 1:t−1) Xt Local Controller Yt Xt gt(X1:t, Y1:t, A1 1:t−1) The local controller observes perfectly the state Xt of the Markov chain, and sends his obser- vation through a packet-drop channel to the remote controller. The transmission is successful, i.e. Yt = Xt, with probability p > 0 and is not successful, i.e. Yt =∅, with probability 1−p≥ 0. We assume that the local controller receives an acknowledgment every time the transmission is successful. The controllers' joint instantaneous utility is given by a uteam (Xt,A1 t t ). t ,A2 III. STRATEGIES AND BELIEFS In a dynamic decision problem with asymmetric information agents have private information about the evolution of the system, and they do not observe the complete history {Ht, Xt}, t∈T . Therefore, at every time t∈T , each agent, say agent i∈N , needs to form (i) an appraisal about the current state of the system Xt and the other agents' information H−i (appraisal about the history), and (ii) an appraisal about how other agents will play in the future (appraisal about the future), so as to evaluate the performance of his strategy choices. t When agents are non-strategic, the agents' strategies g1:N 1:T are known to all agents. Therefore, agent i ∈ N can form these appraisals by using his private information H i t along with the commonly known strategies g−i. Specifically, agent i can utilize his own information H i t at t ∈ T , along with (i) the past strategies g1:t−1 and (ii) the future strategies gt:T to form these appraisals about the history and the future of the overall system, respectively. As a result, the November 23, 2018 DRAFT outcome of decision problems with non-strategic agents can be fully characterized by the agents' strategy profile g.2 10 However, we need to know the entire strategy profile g for all agents and at all times to t, at any time form these appraisals so as to evaluate the performance of an arbitrary strategy gi t∈T and for any agent i∈N . Therefore, we must work with the strategy profile g as a whole irrespective of the length of the time horizon T . Consequently, the computational complexity of determining a strategy profile that satisfies certain conditions (e.g. an optimal strategy profile in teams) grows doubly exponentially in T since the domain of agents' strategy (i.e. Hi t) and the number of temporally interdependent decision problems (one for each time instance) grows with T . As a result, the analysis of such decision problems is very challenging in general [32]. An alternative conceptual approach for the analysis of decision problems is to define a belief system µ along with the strategy profile g. For every agent i ∈ N , at every time t ∈ T , t } conditioned on the realization of hi t) as the agent i's belief about {Xt,P −i define µi t, t(hi hi t)(xt,p−i) := Pg1:t−1{Xt = xt,P −i t}. The belief µi t provides an intermediate that is, µ(hi instrument that encapsulates agent i's appraisal about the past. Therefore, agent i can evaluate the t) along with the future strategy profile t using only the belief µi performance of any action ai t)(xt,p−i) is dependent on g1:t−1 in general since the probability gt:T . However, the belief µ(hi t} depends on g1:t−1. Therefore, the introduction of a hi distribution Pg1:t−1{Xt = xt,P −i t = p−i belief system offers an equivalent problem formulation that does not necessarily break the inter- temporal dependence between g1:t−1 and gt:T and does not simplify the analysis of decision problems. t = p−i t t t(hi Nevertheless, the definition of a belief system has been shown to be suitable for the analysis of single-agent decision making problems (POMDP) for the following reasons. First, in POMDPs, under perfect recall, the probability distribution Pg1:t−1{Xt = xtht} is independent of g1:t−1; this is known as the policy-independence property of beliefs in stochastic control. Second, the complexity of the belief function does not grow over time since at every time t the agent only needs to form a belief about Xt, which has a time-invariant domain. As a result, we can sequentially decompose the problem over time to a sequence of static decision problems with time-invariant complexity; such a decomposition leads to a dynamic program. At each stage 2We discuss the decision problems with strategic agents in the companion paper [2].When agents are strategic each agent may have incentive to deviate an any time from the strategy the other agents commonly believe he uses if it is profitable to him (see [2] for more discussion). November 23, 2018 DRAFT t ∈ T of the dynamic program, we specify gt by determining an action for each realization of the belief µt(·) fixing the future strategies gt+1:T . Therefore, the computational complexity of the analysis is reduced from being exponential in T to linear in T . 11 Unfortunately, the above approach for POMDPs does not generalize to decision problems with many agents. This is because of three reasons. First, with many agents, currently in the literature, there exists no information state for each agent that provides a compression of the agent's information, in a mutually consistent manner among the agents, that is sufficient for decision making purposes. Therefore, an agent's, say agent i's, strategy gi t has a growing domain over time. Second, at every time t∈T , each agent i∈N needs to form a belief about the system state Xt as well as the other agents' private information P −i that has a growing domain. Therefore the complexity of belief functions grows over time. Third, in decision problems with many agents, the policy-independence property of belief does not hold in general and the agents' beliefs at (·) are every time t depend on the past strategy profile g1:t−1. Therefore, the agents' beliefs µ1:N correlated with one another. This correlation depends on g1:t−1, and thus, it is not known a priori. Consequently, if we follow an approach similar to that of POMDP to sequentially decompose the problem, we need to solve the decision problem at every stage for every arbitrary correlation among the agents' belief functions, and such a problem is not tractable.3 Hence, the methodology proposed for the study of POMDPs is not directly applicable to decision problems with many agents and non-classical information structures. t t In the sequel, we propose a notion of sufficient private information and sufficient common information as a mutually consistent compression of the agents' information for decision making purposes. Therefore, we address (partially) the first two problems on the growing domain of the agents' beliefs and strategies. We provide instances of decision problems where we can discover time-invariant information state for each agent. We then utilize the agents' sufficient common information as a coordination instrument, and thus, capture the implicit correlation among the agents' beliefs over time. Accordingly, we present a sequential decomposition of the original decision problems such that at every stage the complexity of the decision problem is similar to that of a static decision multi-agent problem and the size of state variable at every stage is proportional to the dimension of the sufficient private information; thus, we (partially) address the third problem discussed above. 3Alternatively, one can consider arbitrary correlation among the agents' information rather than their beliefs. This is the main idea that underlies the designer's approach proposed by Winstenhausen [25]. Please see Section VI-B for more discussion. November 23, 2018 DRAFT 12 IV. SUFFICIENT INFORMATION We present the sufficient information approach and characterize an information state that results from compressing the agents' private and common information in a mutually consistent manner. Therefore, we introduce a class of strategy choices that are simpler than general strategies as they require agents to keep track of only a compressed version of their information over time. We proceed as follows. In Section IV-A we provide conditions sufficient to determine the subset of private information an agent needs to keep track of over time for decision making purposes. In Section IV-B, we introduce the notion of sufficient common information as a compressed version of the agents' common information that along with sufficient private information provides an information state for each agent. We then show, in Section V, that this compression of the agents' private and common information provides a sufficient statistic in dynamic decision problems with non-strategic agents. In Section VII, we provide a generalization of sufficient information approach presented here. A. Sufficient Private Information The key ideas for compressing an agent's private information appear in Definitions 1 and 2 below. To motivate these definitions we first consider the decision problem with single agent, that is, a Partially Observed Markov Decision Process (POMDP), which is a special case of the model described in Section II where N = 1, H 1 t and Ct = ∅ for all t ∈ T . t = P 1 In a POMDP, the agent's belief about the system state Xt conditioned on his history realization t is an information state. We highlight the three main proprieties that underlie the definition of hi information state in POMDP (see [33], [34]): (1) the information state can be updated recursively, that is, at any time t the information state at t can be written as a function of the information state at t − 1 and the new information that becomes available at t, (2) the agent's belief about the information state at the next time conditioned on the current information state and action is independent of his information history, and (3) at any time t and for any arbitrary action the agent's expected instantaneous utility conditioned on the information state is independent of his information history. We generalize the key properties of information state for POMDPs, described above, to decision problems with many agents. We propose a set of conditions sufficient to compress the agents' private information in two steps. First, we consider a decision problem with many agents where there is no signaling among them. Motivated by the definition of information state November 23, 2018 DRAFT 13 in POMDPs, we describe conditions sufficient to determine a compression of the agents' private information (Definition 1). Next, we build on Definition 1 as an intermediate conceptual step, and consider the case where agents are aware of possible signaling among them. Accordingly, we present a set of conditions sufficient to determine a compression of the agents' private information in decision problems with many agents (Definition 2) . Therefore, we first characterize subsets of an agent's private information that are sufficient for the agent's decision making process when there is no signaling among the agents. Definition 1 (Private payoff-relevant information). Let P i,pr t (P i that agent i ∈ N forms at t ∈ T based on his private information P i Ct. We say P i,pr profile (A1:N t , Ct) denote a private signal t and common information is a private payoff-relevant information for agent i if, for all open-loop strategy 1:T ) and for all t ∈ T , 1:T = a1:N t = ¯ζ i t (i) it can be updated recursively as (ii) for all realizations {ct, pi t(P i,pr t−1 , H i t\H i t−1) if t(cid:54)= 1, P(A1:N 1:T =a1:N pi,pr t+1 t,ct,at =P(A1:N 1:T =a1:N t = ¯φi P i,pr t} it satisfies 1:T )(cid:110) 1:t )(cid:110) (cid:111) (cid:12)(cid:12)(cid:12)pi (cid:12)(cid:12)(cid:12)ct,pi (cid:111) (cid:12)(cid:12)(cid:12)pi,pr t pi,pr t+1 1:T )(cid:110) 1:t)(cid:110) 1:T =a1:N , (cid:111) ,ct,at t} > 0, (cid:12)(cid:12)(cid:12)ct, pi,pr t (cid:111) . , at (iii) for all realizations {ct, pi t} ∈ Ct × P i t such that P(A1:N 1:T ){ct, pi E(A1:N 1:t−1=a1:N ui t(Xt,At) t,at = E(A−i 1:t−1=a−i ui t(Xt,At) By assuming that all other agents play open-loop strategies we remove the interdependence between agents −i's strategy choices and agent i's information structure, thus, we eliminate signaling among the agents. Fixing the open-loop strategies of agents −i, agent i faces a , t ∈ T , is a private payoff- centralized stochastic control problem. Definition 1 says that P i,pr relevant information for agent i if (i) it can be recursively updated, (ii) P i,pr includes all t+1 and (iii) agent i's instantaneous conditional expected information in P i utility at any t ∈ T is only a function of Ct,P i,pr t at t. These three conditions are similar to properties (1)-(3) for an information state in POMDP, but they concern only agent t} of his private and common i's private information P i t t that is relevant to P i,pr instead of the collection H i , and his action Ai t = {Ct,P i t t t November 23, 2018 DRAFT 14 information.4 While the definition of private payoff-relevant information suggests a possible way to compress the information required for an agent's decision making process, it assumes that other agents play open-loop strategies and do not utilize the information they acquire in real-time for decision making purposes (i.e. no signaling). However, open-loop strategies are not in general optimal for agents −i. As a result, to evaluate the performance of any strategy choice gi agent i needs also to form a belief about the information that other agents utilize to make decisions. Definition 2 (Sufficient private information). We say Si sufficient private information for the agents if, t = ζ i t (P i t , Ct; g1:t−1), i ∈ N , t ∈ T , is (i) it can be updated recursively as Si t = φi t(Si t−1, H i t\H i t−1; g1:t−1) for t ∈ T \{1}, (5) (ii) for any strategy profile g and for all realizations {ct, pt, pt+1, zt+1, at} ∈ Ct×Pt×Pt+1×Zt+1 of positive probability, Pg1:t (cid:110) (cid:12)(cid:12)(cid:12)pt,ct,at (cid:111) (cid:110) (cid:12)(cid:12)(cid:12)st,ct,at (cid:111) st+1,zt+1 =Pg1:t st+1,zt+1 , (6) where s1:N τ = ζ 1:N τ (p1:N τ ,cτ ;g1:τ−1) for τ ∈ T ; (iii) for every strategy profile g of the form g :={gi t : S i t∈T ; Eg1:t−1 (cid:110) for all realizations {ct,pi for τ ∈ T ; ui t(Xt,At) t}∈Ct×P i t), i∈N,t∈T } and at∈At, (cid:111) (cid:12)(cid:12)(cid:12)ct,pi =Eg1:t−1 t,at t ×Ct → ∆(Ai (cid:110) ui t(Xt,At) (cid:12)(cid:12)(cid:12)ct,si (cid:111) t,at , t of positive probability where s1:N τ = ζ 1:N τ (p1:N τ (7) ,cτ ; g1:τ−1) (iv) given an arbitrary strategy profile g of the form g := {gi t), i∈N , t∈T }, i∈N , and t∈T , Pg1:t−1 (cid:110) (cid:111) (cid:12)(cid:12)(cid:12)pi s−i t t,ct =Pg1:t−1 s−i t t,ct , t : S i t ×Ct → ∆(Ai (cid:111) (cid:110) (cid:12)(cid:12)(cid:12)si (8) 4We note that we interpret a centralized control problem as a special case of our model where N =1, H 1 t =Pt and Ct =∅ for all t∈T , Definition 1 coincides with the definition of information state for the single agent decision problem. We would like to point out that conditions (i)-(iii) can have many solutions including the trivial solution P i,pr t =P i t . 5 November 23, 2018 DRAFT for all realizations {ct,pi for τ ∈ T . t}∈Ct×P i t of positive probability where s1:N τ = ζ 1:N τ (p1:N τ 15 ,cτ ; g1:τ−1) There are four key differences between the definition of sufficient private information and that of private payoff relevant information. First, we allow that the definition and the update rule t to depend on the agents' strategies g1:t−1. Second, comparing to of sufficient information Si part (ii) of Definition 1, part (ii) of Definition 2 requires that sufficient information St includes all information relevant to the realization of Zt+1 in addition to the information relevant to the realization of St+1. As we discuss further in Section VI, this is because when signaling occurs in a multi-agent decision problems agents need to have a consistent view about future commonly observable events. Third, comparing part (iii) of Definition 2 to part (iii) of Definition 1, we note that the probability measures in Definition 2 depend on the strategy profile g instead of the ope-loop strategy profile (A1:N 1:T ). Fourth, in part (iv) of Definition 2 there is an additional t must be rich enough so that condition requiring that agent i's sufficient private information Si he can form beliefs about agents −i's sufficient private information S−i ; such a condition is absent in Definition 1. 1:T = a1:N t In general, the notion of sufficient private information S1:N is more restrictive than that of , t ∈ T , needs to satisfy the private payoff relevant information P 1:N,pr additional condition (iv), and furthermore, open-loop strategies are a strict subset of closed loop strategies. Definition 2 provides (sufficient) conditions under which agents can compress their private information in a "mutually consistent' manner. We would like to point out that conditions (i)-(iv) of Definition 2 can have many solutions including the trivial solution Si . This is because, S1:N t t t t .6 t = P i B. Sufficient Common Information Based on the characterization of sufficient private information, we present a statistic (com- pressed version) of the common information Ct that agents need to keep track of over time for decision making purposes. Fix a choice of sufficient private information S1:N t to be the set of all t. Given the agents' strategy profile g, let γt : Ct → possible realizations of Si ∆(Xt × St) denote a mapping that determines a conditional probability distribution over the i=1 S i t , t ∈ T . Define S i t, and St :=(cid:81)N 6We do not discuss the possibility of finding a minimal set of sufficient private information in this chapter, and leave it for future research as such investigation is beyond the scope of this chapter. November 23, 2018 DRAFT 16 system state Xt and all the agents' sufficient private information St conditioned on the common information Ct at time t as γt(ct)(xt, st) = Pg1:t−1{Xt = xt, St = stct}, (9) for all ct ∈ Ct, xt ∈ Xt, st ∈ St. We call the collection of mappings γ := {γt, t ∈ T } a sufficient information based belief system (SIB belief system). Note that γt is only a function of the common information Ct, and thus, it is computable by all agents. Let Πγ t := γt(Ct) denote the (random) common information based belief that agents hold under belief system γ at t. We can interpret Πγ t as the common belief that each agent holds about the system state Xt and all the agents' (including himself) sufficient private information St at time t. We call the SIB belief Πt a sufficient common information for the agents. In the rest of the paper, we write Πt and drop the superscript γ whenever such a simplification in notation is clear. Moreover, we use the terms sufficient common information and SIB belief interchangeably. C. Sufficient Information based Strategy t The combination of sufficient private information S1:N and sufficient common information (the SIB belief) Πt offers a mutually consistent compression of the agents' private and common t) for information. Consider a class of strategies that are based on the information given by (Πt, Si each agent i∈N at time t ∈ T . We call the mapping σi t) a Sufficient Information Based (SIB) strategy for agent i at time t. A SIB strategy σi t determines a probability t). A SIB strategy is distribution for agent i's action Ai a strategy where agents only use the sufficient common information Πt = γt(Ct) (instead of complete common information Ct), and the sufficient private information Si t , Ct; g1:t−1) 1:T} is (instead of complete private information P i called a SIB strategy profile σ. The set of SIB strategies is a subset of general strategies, defined in Section II, as we can define, t = ζ i t ). A collection of SIB strategies {σ1 t at time t given his information (Πt, Si t : ∆(Xt × St)× S i t (P i 1:T , ..., σN t → ∆(Ai g(σ,γ),i t (hi t) := σi t(πγ t , si t) ∀t∈T (10) We note that from Definition 2 and (9), the realizations πt and s1:N Therefore, strategies g(σ,γ),i for t = 1, g(σ,γ),i 1) = σi at t only depends on g1:t−1. , defined above via (10) needs to be determined iteratively as follows; )); ...; 1 , ζ i 1, C1)); for t = 2, g(σ,γ),i 2, C2; g(σ,γ) 2) = σi 2(πγ 1(πγ 1(P i 2(P i 2 , ζ i (hi 1 t t (hi 2 1 November 23, 2018 DRAFT for t = T , g(σ,γ),i t∈T and i∈N . t (hi t) = σi t(πγ t , ζ i 2(P i t , Ct; g(σ,γ) t−1 )). Therefore, strategy g(σ,γ),i t 17 is well-defined for all D. Sufficient Information based Update Rule When the agents play a SIB strategy profile σ, it is possible to determine the SIB belief Πt recursively over time based on Πt−1 and the new common information Zt via Bayes' rule. Let : ∆(Xt−1×St−1)×Zt → ∆(Xt×St) describe such a update rule for time t + 1 ∈ T /{1} so ψσt−1 that t Πt = ψσt−1 t (Πt−1, Zt). (11) We note that the SIB update rule ψσt−1 depends on the SIB strategy profile σt−1 at t−1. In the rest of the paper, we drop the superscript σ whenever such a simplification in notation is clear. t E. Special Cases We consider the special cases (1)-(3) of the general model we presented in Section II, 1:T ; we discuss the application of sufficient and identify the sufficient private information S1:N information approach to special case (4) in Section VII. t = {X1:t} and P 2 2) Delayed sharing information structure: We have P i t = X ={Xt−max(k,δ+1)+1,...,Xt} and S2 1) Real-time source coding-decoding: The encoder's and decoders' private information are t = { X1:t−1−δ}, respectively. The agents' common information is given by P 1 t = ∅ satisfy given by Ct ={M1:t−1}. We can verify that S1 the conditions of Definition 2 ; this is similar to the structural results in [27, Sections III and VI]. Consequently, the common information based belief is Πt =Pg{Xt−max(k,δ+1)+1:tM1:t−1}. t−d+1:t} and Ct ={Y1:t−d, A1:t−d}. Since we do not assume any specific structure for the system dynamics and the agents' observations, agent i's complete private information P i is payoff-relevant for him. Therefore, t t . Consequently, we have Πt = Pg{Xt,Yt−d+1:t,At−d+1:tY1:t−d,A1:t−d}. The above we set Si sufficient information appears in the first structural result in [18]. t ={X 1 t = (X 1 1:t−1}, 1:t−1}), t ; this sufficient information corresponds to 1:t,M 1 t ,P{RX 1 1:t,M 2 1:t−1M 1 1:t−1}, P 2 3) Real-time multi-terminal communication: We have P 1 t ={Y 1 1:t, X1:t−1}, and Ct =∅. It is easy to verify that S1 P 3 S2 t = (X 2 the structural results that appear [29]. 1:t,Y 2 t ,P{RX 2 1:t−1}), and S3 1:t},P{Y 2 1:t−1M 2 t = P 3 1:t},P{Y 1 t ={X 2 t ={Y i t−d+1:t,Ai t = P i November 23, 2018 DRAFT 18 V. MAIN RESULTS In this section, we present our main results for the analysis of dynamic decision problems with asymmetric information and non-strategic agents using the notion of sufficient information. We first provide a generalization of the policy-independence property of beliefs to decision problems with many agents (Theorem 1). Second, we show that the set of SIB strategies are rich enough so that restriction to them is without loss of generality (Theorem 2). That is, given any strategy profile g, there exists a SIB strategy profile σ such that every agent gets the same flow of utility over time under σ as the one under g. Third, we consider dynamic team problems with asymmetric information. We show that using the SIB strategies, we can decompose the problem sequentially over time, formulate a dynamic program, and determine a globally optimal policy via backward induction (Theorem 3). Theorem 1 (Policy-independence belief property). (i) Consider a general strategy profile g. If agents −i play according to strategies g−i, then for every strategy gi that agent i plays, Pg(cid:110) xt, p−i t = Pg−i(cid:110) xt, p−i t . (12) (ii) Consider a SIB strategy profile σ along with the associated update rule ψ. If agents −i play according to SIB strategies σ−i, then for every general strategy gi that agent i plays, t (cid:111) (cid:12)(cid:12)(cid:12)hi (cid:12)(cid:12)(cid:12)hi (cid:111) t (cid:110) (cid:110) t (cid:111) (cid:12)(cid:12)(cid:12)hi (cid:12)(cid:12)(cid:12)hi (cid:111) t . Pσ−i,gi ψ xt, p−i t =Pσ−i ψ xt, p−i t (13) Theorem 1 provides a generalization of the policy-independence belief property for the central- ized stochastic control problem [3] to multi-agent decision making problems. Part (i) of Theorem 1 states that, under perfect recall, agent i's belief is independent of his actual strategy gi. Part (ii) of Theorem 1 refers to the case where agents −i play SIB strategies σ−i and update their SIB belief according to SIB update rule ψ. The update rule ψ is determined based on (σ−i, σi) via Bayes' rule, where σi denotes the SIB strategy that agents −i assume agent i utilizes. Equation (13) states that even if agent i unilaterally and privately deviates from his SIB strategy, his belief is independent of his actual strategy gi, and only depends on the other agents's strategy σ−i as well as the other agents' assumption about the SIB strategy σi (or equivalently the SIB update November 23, 2018 DRAFT 19 rule ψ).7 In POMDPs it is shown that restriction to Markov strategies is without loss of optimality. We provide a generalization of this result to decision problems with many agents. We show that restriction to SIB strategies is without loss of generality in non-strategic settings given that the agents have access to a public randomization device. We say that the agents have access to a public randomization device if at every time t ∈ T they observe a public random signal ωt that is completely independent of all events and primitive random variables in the decision problem and is uniformly distributed on [0,1], and is independent across time. As a result, in general, at every t∈T , all agents can condition their actions on the realization of ωt as well as their own information. In other words, a public randomization device enables the agents to play correlated randomized strategies. We denote by σi t, ωt) agent i's SIB strategy using the t(Πt, Si public randomization device for every i ∈ N and t ∈ T . Theorem 2. Assume that the non-strategic agents have access to a public randomization device. Then, for any strategy profile g there exists an equivalent SIB strategy profile σ that results in the same expected flow of utility, i.e. (cid:40) T(cid:88) Eg (cid:41) (cid:40) T(cid:88) (cid:41) ui τ (g1:N τ (H 1:N τ ),Xτ) = Eσ u1:N τ (σi τ(Πτ,S1:N τ , ωτ ),Xτ) , (14) τ =t for all i∈N and t∈T . τ =t We provide an intuitive explanation for the result of Theorem 2 below. For every agent i ∈ N , his complete information history H i t at any time t ∈ T consists of two components: (i) one component captures his information about past events that is relevant to the continuation decision problem; and (ii) another component that, given the first component, captures the information about past events that is irrelevant to the continuation decision problem. We show that the combination of sufficient private information Si t and sufficient common information Πt contains the first component. Nevertheless, in general, the agents can coordinate their action by incorporating the second component into their decision since their information about the past T that is not captured by events is correlated. Let Ri t )} t } are jointly independent of {(Πt, Si t denote the part of agent i's information H i t), ..., (Πt, SN (Πt, Si t). We show that the set of {R1 t , ..., RN 7The results of Theorem 1 provides a crucial property for the analysis of decision problems with strategic agents. This is because it ensures that an agent's unilateral deviation does not influence his belief (see the companion paper [2] for more details). November 23, 2018 DRAFT 20 t , ..., RN t }. Using the signals { R1 t , ..., RN t , ..., RN (Lemma 2 in the Appendix). Therefore, at every time t ∈ T , we can generate a set of signals t }, one for each agent, using the public randomization device ω so that they are { R1 identically distributed as {R1 t } along with the information t) for every agent i ∈ N , we can thus recreate a (simulated) history that is identically state (Πt, Si distributed to H i t. This implies that, given a public randomization device ω, it is sufficient for each agent i ∈ N to only keep track of (Πt, Si t, and play a SIB strategy σi to achieve an identical (in distribution) sequence of outcomes per stage as those under the strategy profile g. t) instead of his complete history H i The result of Theorem 2 states that the the class of SIB strategies characterizes a set of simpler strategies where the agents only keep track of a compressed version of their information rather than their entire information history. Moreover, the restriction to the class of SIB strategies is without loss of generality. Thus, along with results appearing in the companion paper [2], the result of Theorem 2 suggests that the sufficient information approach proposed in this paper presents a unified methodology for the study of decision problems with many non-strategic or strategic agents and asymmetric information. We would like to discuss the implication of Theorem 2 for two special instances of our model. First, when N = 1, there is no need for a public randomization device since the single decision maker does not need to correlate the outcome of his randomized strategy with any other agent. Therefore, the result of Theorem 2 states that the restriction to Markov strategies in POMDPs is without loss of generality. Second, when N > 1 and the agents have identical utilities, i.e. dynamic teams, utilizing a public randomization device does not improve the performance. This is because, in dynamic teams a randomized strategy profile is optimal if and only if it is optimal for every realization of the randomization. Therefore, the restriction to SIB strategies in dynamic teams is without loss of optimality. Using the result of Theorem 2, we present below a sequential decomposition of dynamic teams over time. We formulate a dynamic program that enables us to determine a globally optimal strategy profile via backward induction. Theorem 3. A SIB strategy profile σ is a globally optimal solution to a dynamic team problem with asymmetric information if it solves the following dynamic program: VT +1(πt+1) := 0, ∀πt+1,∀i ∈ N ; (15) DRAFT November 23, 2018 at every t ∈ T , and for every πt, (πt,·)∈ arg max σ1:N t α1:N :S1:N t ) t →∆(A1:N Eπt max t →∆(A1:N ) t α1:N :S1:N Eπt (cid:8)uteam (cid:8)uteam t t t (πt,α1:N,Zt+1))(cid:9), t (πt,α1:N,Zt+1))(cid:9). 21 (16) (17) (Xt,α1:N (S1:N )) + Vt+1(ψσt t Vt(πt) := (Xt,α1:N (S1:N )) + Vt+1(ψσt t The results of Theorems 2 and 3 extend the results of [18], [26] for the study of dynamic teams in two directions. First, they state that restriction to the set of SIB strategies is without loss of generality, while the results of [18], [26] only state that this restriction is without loss of optimality. Second, the definition of Common Information Based strategies, first presented in t , i ∈ N (or all their private [18], [26], requires the agents to use all of their private information P i memory that is a predetermined function of their private information if they do not have perfect recall); the result of Theorem 3 holds for SIB strategies where the agents' private information is effectively compressed , thus, it generalizes/extends the definition of CIB strategies proposed in [18], [26]. A. Constructive algorithm VI. DISCUSSION The sufficient information approach described in Sections IV and V, presents a generalization of the notion of information state to dynamic multi-agent decision problems with non-classical information structure. Nevertheless, we would like to point out that our approach does not address all the issues present in the study of dynamic multi-agent decision problems. We discuss the main limitation of our approach below. In POMDPs, an information state with time-invariant domain can be determined by forming the probability distribution over the system state conditioned on the current information. Our approach does not offer an explicit constructive algorithm that determines a mutually-consistent set of information states, one for each agent, with time-invariant domains in dynamic multi-agent decision problems. Specifically, Definition 2 describes only a set of sufficient conditions that one can use to evaluate whether a specific compression of agents' private information is sufficient for decision making purposes; it does not offer a constructive algorithm to determine a compression of the agents' private information that leads to an information state with time-invariant domain. Given a set of sufficient private information with time-invariant domain for the agents, we achieve, through the formation of SIB beliefs, a compression of the agents' common information November 23, 2018 DRAFT 22 that results in a set of information states with time-invariant domains. In Sections II and IV, we presented instances of multi-agent decision problems where we can discover a set of information states with time-invariant domains. Nonetheless, it is not clear if such a set of mutually-consistent information states with time-invariant domains exist for every dynamic multi-agent decision problem. Therefore, an interesting, but challenging, future direction would be to identify classes of dynamic decision problems with non-classical information structure where we can guarantee the existence of a set of mutually-consistent information states with time-invariant domains, and prescribe a constructive methodology for their identification. Moreover, we would like to point out that the sufficient information approach presented here provide sufficient conditions that can be used to evaluate an educated-guess one may have for specific multi-agent problems. B. Comparison with other Approaches The sufficient information approach proposed in this paper shares similarities and also has differences with existing conceptual approaches to the study of dynamic multi-agent decision problems. Below, we briefly discuss these approaches and compare them with the sufficient information approach. 1) Comparison with Agent-by-Agent Approach: The agent-by-agent approach proceeds as follows: start with an initial guess of a strategy profile g for all agents. At each iteration, select one agent, say agent i. and update his strategy to a best response strategy given the strategy g−i of all other agents. Repeat the process until a fixed point is reached, that is, when no agent can improve performance by unilaterally changing his strategy. If the above-described iterative process converges, the resulting strategy profile determines an agent-by-agent optimal strategy profile; however, such an agent-by-agent optimal strategy profile, in general, is not a globally optimal strategy profile [24]. This is because the multi-agent decision problems are, in general, not convex in the agents' strategies [5]. Therefore, the above-described iterative process does not necessarily converge, or it may converge to a locally optimal strategy profile that is not a globally optimal strategy profile. In contrast to agent-by-agent approach, the sufficient information approach determines a globally optimal strategy profile for multi-agent decision problems with non-strategic agents. The agent-by-agent approach can be used to discover qualitative properties of optimal strate- gies. Specifically, we fix the strategies of all agents except one, say agent i, to an arbitrary set of strategies g−i, and solve for agent i's best response; to determine agent i's best response we November 23, 2018 DRAFT 23 need to solve a POMDP, where the system state and system dynamics, in general, depend on g−i. If agent i's best response possesses a property that holds for every choice of g−i, then a globally optimal strategy for agent i possesses the same property. In contrast to the agent-by- agent approach where one need to solve a POMDP parameterized by g−i, to discover qualitative properties of a globally optimal strategy profile using the sufficient information approach we only need to check the set of conditions appearing in Definition 2 (or equivalently a more general Definition 3 that will appear in Section VII). Moreover, using the sufficient information approach we can discover qualitative properties of optimal strategies that cannot be discovered by the agent-by-agent approach. For instance, consider the following example. Example. Consider a team problem with two agents and observable actions, where agent 2's t,Wt). Each agent action A2 t does not affect the evolution of Xt for all t, i.e. Xt+1 = ft(Xt,A1 t ). An i, i = 1, 2, has an imperfect private observation of state Xt at t given by Y i t = Oi t for agent i at t depends, in general, on his complete information arbitrary choice of strategy gi 1:t, A1:t−1}. Therefore, following the agent-by-agent approach, if agent i's history given by {Y i τ for some τ, 1 ≤ τ ≤ t − 1, then agent j's, j (cid:54)= i, best response strategy depends on A2 also depends on A2 1:t−1 as irrelevant information for decision making purposes for agents 1 and 2. However, using the sufficient information approach we can simply show that a globally optimal strategy profile depend only on P{XtY i τ. Consequently, the agent-by-agent approach fails to characterize A2 t(Xt, W i 1:t} for agent i. 1:t,A1 2) Comparison with the Designer's Approach: The designer's approach was originally pro- posed by Witsenhausen in [25], and was further investigated in [35]. This approach considers the decision problem from the point of view of a designer (she) who knows the system model and the probability distribution of the primitive random variables, and chooses control/decision strategies for all agents; she chooses these strategies without having any observation/knowledge about the realizations of primitive random variables (i.e. she chooses these strategies before the system evolution starts). Therefore, the designer effectively solves a centralized panning problem. The designer's approach proceeds by: (i) formulating the centralized planning problem as a multi- stage, open-loop stochastic control problem in which the designer's decision at each time is a set of control strategies for all agents; (ii) using the standard techniques in centralized stochastic control to obtain a dynamic programming decomposition of the decision problem. Each step of the resulting dynamic program is a functional optimization problem. November 23, 2018 DRAFT 24 The designer's approach breaks the interdependencies between the agents' decision and infor- mation over time by transferring all the complexity that arises due to non-classical information structure and signaling to a larger information state which at each time is given by a probability distribution on Ht, the domain of which increases with time as agents have perfect recall.8 There- fore, the sequential decomposition resulting from the designer's approach is not, in general, very practical for the study of multi-agent dynamic decision problems with asymmetric information. In contrast to the designer's approach, the sufficient information approach provides a sequential decomposition of the decision problem over time where at each time t each agent makes decision based on only a compression of his information H i t. Therefore, it leads to a dynamic program where the state variable at each step of the program is a probability distribution on St instead of a probability distribution on Ht in the designer's approach. 3) Comparison with the Common Information Approach: The common information approach, proposed in [18], [26], addresses some of the drawbacks of designer's approach by modeling the decision problem as a closed-loop centralized planning problem (POMDP) in which a coordinator observes perfectly the common information Ct at each time t and, based on this knowledge, chooses a set of partial control strategies/prescriptions that determine how each agent takes an action based on his private information at time t. The coordinator's information state at time t is his belief on (Xt, Pt) conditioned Ct. As shown in [26], the dynamic programming decomposition achieved by the common information approach is simpler than that achieved by the designer's approach. In the common information approach the agents' private information remains intact. Therefore, the resulting decomposition is not very practical whenever the agents' private information grows in time (see special cases 1,3 and 4 in Section II). Furthermore, the common information approach becomes identical to the designer's approach whenever the agents do not share any common information over time (see special case 3). In the sufficient information approach, we provide conditions sufficient to identify mutually- consistent compressions of the agents' private information that are sufficient for decision making purposes and do not result in any loss in system performance. Thus, the sufficient information approach gives rise to a dynamic program that is simpler than the one resulting from the common information approach. As we show in Section VII, these conditions are the core of sufficient information approach; they are generalized by Definition 3 to captures a mutually-consistent 8An instance where the domain of the control law is time-invariant is presented in [35]. November 23, 2018 DRAFT 25 joint compressions of the agents' private and common information. Moreover, in the model of Section II, we do not assume that the agents share a common objective. Therefore, we do not reformulate the original multi-agent decision problem as a centralized planning problem from the coordinator's point of view when signaling occurs. Alternatively, we provide conditions sufficient to identify compression of the agents' information in a mutually-consistent manner on individual level. As a result, our approach is applicable to both strategic and non-strategic settings (see our companion paper [2] for strategic settings). VII. GENERALIZATION In the sufficient information approach presented in Section IV, we treat the agents' private information and common information separately. This is because the main challenge in the study of dynamic decision problems with non-strategic agents is due to the presence of the agents' private information. Nevertheless, such a separate treatment of private and common information is not necessary. Using the same rationale that leads to Definition 2, we present below a set of conditions sufficient to characterize a mutually consistent compression of agents' information, without separating private and common components, that is sufficient for decision making purposes. Definition 3 (Sufficient information). We say Li sufficient information for the agents if, t = ζ i t (P i t , Ct, g1:t−1) ∈ Li t, i ∈ N , t ∈ T , is (i) it can be updated recursively as t = φi Li t(Li t−1, H i t\H i t−1, g1:t−1) for t ∈ T \{1}, (18) (ii) for any strategy profile g and for all realizations {ct, pt, pt+1, zt+1, at} ∈ Ct×Pt×Pt+1×Zt+1 with positive probability, Pg1:t (cid:110) lt+1 (cid:12)(cid:12)(cid:12)pt,ct,at (cid:111) (cid:110) (cid:12)(cid:12)(cid:12)lt,at (cid:111) =Pg1:t lt+1 where l1:N τ = ζ 1:N τ (p1:N τ ,cτ ;g1:τ−1) for τ ∈ T ; (iii) for every strategy profile g of the form g := {gi , (19) t), i ∈ N,t ∈ T } and at ∈ At, (cid:111) (cid:12)(cid:12)(cid:12)ct,pi t : Li t → ∆(Ai (cid:110) (cid:111) (cid:12)(cid:12)(cid:12)li ui t(Xt,At) =Eg1:t−1 t,at ui t(Xt,At) t,at , (cid:110) t∈T ; Eg1:t−1 November 23, 2018 (20) DRAFT for all realizations {ct,pi for τ ∈ T ; t}∈Ct×P i t of positive probability where l1:N τ = ζ 1:N τ (p1:N τ 26 ,cτ ; g1:τ−1) (iv) given an arbitrary strategy profile g of the form g := {gi t → ∆(Ai t), i ∈ N , t ∈ T }, t with positive probability where l1:N τ = ζ 1:N τ (p1:N τ (21) ,cτ ; g1:τ−1) t : Li (cid:12)(cid:12)(cid:12)li (cid:111) , t (cid:110) (cid:111) (cid:12)(cid:12)(cid:12)pi (cid:110) t}∈Ct×P i Pg1:t−1 l−i t t,ct =Pg1:t−1 l−i t i∈N , and t∈T , for all realizations {ct,pi for τ ∈ T . The conditions of Definition 3 are similar to those of Definition 2, but they concern agents' private and common information rather than just their private information. Throughout the paper, we do not make any assumption that the agents' private observations are necessarily disjoint. t = ∅, for all i ∈ N and t ∈ T , in which case Therefore, one can define P i Definition 3 would be the same as Definition 2. Consequently, all the results appearing in this paper (Theorems 1-6) also hold for sufficient information characterized by Definition 3. t and C i t = H i We show below that the set of information states (Si t, Πt), i ∈ N} proposed in Section IV satisfies the conditions of Definition 3. Therefore, Definition 3 provides a generalization of the sufficient information approach presented in Section IV as it does not require to compress the agents' private and common information separately. Theorem 4. The set of information states Li t := (Si t, Πt), i ∈ N , t ∈ T , satisfies Definition 3. Compared to Definition 2, Definition 3 provides conditions sufficient for a mutually-consistent joint compression of the agents' private and common information. However, similar to the discussion in Section VI-A, it does not provide a constructive algorithm to determine a set of sufficient information Li t, i ∈ N , t ∈ T , with time-invariant domain. Remark 1. In view of Definition 3, one can replace condition (ii) of Definition 2 with a weaker one that requires that St include all the information necessary to form a belief about the realizations (of parts) of Zt+1 only if (those parts of) Zt+1 affect the realization of Πt+1 given Πt. Using Definition 3 we identify a set of sufficient information for special case 4 described in Section II. November 23, 2018 DRAFT 27 Special Case: 4) Optimal remote and local controller: We have Ct ={Y1:t}, P 1 1:t−1}\Ct, and P 2 t = {A2 1:t−1}. Let τ ≤ t denote the last time the data transmission was successful between the local and remote controllers. We can restrict attention, without loss of optimality, to the class of pure t ={Xt,{Pg{Xt = xtXτ},∀xt∈Xt}} strategies for both controllers. Therefore, one can show that L1 t ={Pg{Xt = xtXτ},∀xt∈Xt} satisfy the conditions of Definition 3; this is similar to the and L2 structural results in [30], [31]. t ={X1:t,A1 VIII. EXTENSION TO INFINITE HORIZON In the model of Section II, we assume that the horizon T is finite. We present a model similar to that of Section II with infinite horizon, i.e. T = ∞, and provide the extensions of our results to dynamic decision problems with infinite horizon. Infinite Horizon Dynamic Decision Problem: There are N non-strategic agents who live in a dynamic Markovian world over an infinite horizon. Consider a time-invariant model where the system state, actions, and observations spaces are finite and time-invariant, i.e. X∞ = Xt, A∞ = At, Z∞ = Zt, and Y∞ = Yt for all t ∈ N. Let Xt ∈ X∞ denote the system state at t ∈ N. Given the agents' actions At at t, the system state evolution is given by Xt+1 = f∞(Xt, At, W x t ), (22) where {W x t , t ∈ N} is a sequence of independent and identically distributed random variables. The initial state X1 is a random variable with probability distribution η ∈ ∆(X∞) with full support that is common knowledge among the agents. At every time t ∈ N, each agent i ∈ N , receives a noisy observation Y i t given by t = Oi∞(Xt, At−1, W i Y i t ), (23) t , t ∈ N, i ∈ N} is a sequence of independent and identically distributed random where {W i variables. In addition, at every t ∈ N all agents receive a common observation Zt ∈ Z∞ given by Zt = Oc∞(Xt, At−1, W c t ), (24) DRAFT November 23, 2018 28 ∞(cid:88) t , t ∈ N} is a sequence of independent and identically distributed random variables; t , t ∈ N, i ∈ N} and the initial state X1 are t , t ∈ N}, and {W i t , t ∈ N}, {W c where {W c the sequences {W x mutually independent. Similar to the model of Section II, let P i t and Ct denote agent i's, i ∈ N , private and common information at t ∈ N, respectively. Agent i has a time-invariant instantaneous utility function δt−1ui∞(Xt, At), and his total discounted utility is given by U i in(X, A) := δt−1ui∞(Xt, At), (25) where δ denotes the discount factor. t=1 We provide an extension of our results to infinite horizon dynamic decision problems with non- strategic agents. For that matter, we first present a generalization of the definition of sufficient private information to infinite horizon decision problems. t, i ∈ N , t ∈ N, is a Definition 4 (Time-invariant sufficient private information). We say Si time-invariant sufficient private information if it is a sufficient private information and has a time-invariant domain denoted by S i∞, i ∈ N . We note that for the special cases presented in Section IV, the characterized sufficient private information is time-invariant. Following an argument similar to the one presented in Section V, we extend the result of Theorem 2 to infinite horizon dynamic decision problems with non-strategic agents. Theorem 5. Consider an infinite horizon dynamic decision problem with non-strategic agents having access to a public randomization device. Then, for any arbitrary strategy profile g there exists an equivalent stationary SIB strategy profile σ that results in the same expected flow of utility, i.e., (cid:40) ∞(cid:88) Eg (cid:41) (cid:40) ∞(cid:88) (cid:41) δt−1ui∞(g1:N τ (H 1:N τ ),Xτ) = Eσ∞ δt−1ui∞(σ1:N τ (Πτ,S1:N τ , ωτ ),Xτ) , (26) τ =t for all i ∈ N and t ∈ N. τ =t Next, we consider the case where agents share the same objective ui∞(·,·) = uteam∞ (·,·) for all i ∈ N ., i.e. an infinite horizon dynamic team problem. It is shown that in infinite horizon November 23, 2018 DRAFT 29 POMDPS we can restrict attention, without loss of generality, to stationary Markov policies [3]. We provide a generalization of this results to dynamic multi-agent decision problems below. Given a set of time-invariant sufficient private information, let Πt ∈ ∆(X∞×S∞) denote the SIB belief about (Xt, St) at time t. We call the mapping σi∞ : ∆(X∞ × S∞) × S i∞ → ∆(Ai∞) t, i ∈ N , t ∈ N, is a time-invariant sufficient private a stationary SIB strategy for agent i if Si information. Similarly, given a stationary SIB strategy profile σ∞, we define a stationary SIB update rule as a time-invariant mapping ησ∞∞ : ∆(X∞ × S∞) × Z∞ → ∆(X∞ × S∞), that recursively determines the SIB belief via Bayes' rule for all t ∈ N. Similarly, let σi∞(Πt, Si t, ωt) denote agent i's stationary SIB strategy using the public randomization device for every i∈N and t∈T , when the agents have access to a public randomization device ωt for every t∈T . We provide a sequential decomposition similar to that of Theorem 3 for infinite horizon dynamic teams below. Theorem 6. A stationary SIB strategy profile σ∞ is an optimal solution to an infinite horizon dynamic team problem with asymmetric information if it solves the following Bellman equation: (cid:8)uteam∞ (Xt,α1:N (S1:N t )) + V∞(η∞(πt,α1:N,Zt+1))(cid:9), Eπ V∞(πt) := max α1:N :S1:N∞ →A1:N∞ (27) for all πt ∈ ∆(X∞ × S∞). The result of Theorem 6 provide a generalization of Bellman equation for POMDPS (see [3, Ch. 8]) to decision problems with many agents and asymmetric information. IX. CONCLUSION We presented a general approach to study a general class of dynamic multi-agent decision making problems with non-strategic agents. We proposed the notion of sufficient information that enables us to compress effectively the agents' (private and common) information in a mutually consistent manner for decision making purposes. We showed that the restriction to the class of SIB strategies are without loss of generality. Accordingly, we provided a sequential decomposition of dynamic decision problems with non-strategic agents, and formulated a dy- namic program to determine a globally optimal strategy profile in dynamic teams. The proposed sufficient information approach presented in this paper generalizes a set of existing results in the literature for the study of dynamic multi-agent decision making problems with non-strategic agents. Our results in this paper, along with those appearing in the companion paper [2], provide November 23, 2018 DRAFT 30 a unified appraoch to study dynamic decision problems with non-strategic agents (teams) and strategic agents (games). For future directions, we will investigate the problem of determining a constructive algorithm that enables us to identify sufficient (private) information in a systematic way. REFERENCES [1] H. Tavafoghi, Y. Ouyang, and D. Teneketzis, "A sufficient information approach to decentralized decision making," in 57th IEEE Conference on Decision and Control (CDC), 2018. [2] H. Tavafoghi, Y. Ouyang, and D. Teneketzis, "A unified approach to dynamic multi-agent decision problems with asymmetric information - part i: Strategic agents," working paper, 2018. [3] P. Kumar and P. Varaiya, Stochastic Systems: Estimation Identification and Adaptive Control. Prentice-Hall, Inc., 1986. [4] D. P. Bertsekas, Dynamic Programming and Optimal Control, vol. 1. Belmont, MA: Athena Scientific, 1995. [5] A. Mahajan, N. C. Martins, M. C. Rotkowitz, and S. Yuksel, "Information structures in optimal decentralized control," in 51st IEEE Conference on Decision and Control (CDC), pp. 1291 -- 1306, 2012. [6] A. A. Kulkarni and T. P. Coleman, "An optimizer's approach to stochastic control problems with nonclassical information structures," IEEE Transactions on Automatic Control, vol. 60, no. 4, pp. 937 -- 949, 2015. [7] L. Lessard and S. Lall, "Convexity of decentralized controller synthesis," IEEE Transactions on Automatic Control, vol. 61, no. 10, pp. 3122 -- 3127, 2016. [8] S. Yuksel and N. Saldi, "Convex analysis in decentralized stochastic control and strategic measures," in 55th IEEE Annual Conference on Decision and Control (CDC), pp. 6050 -- 6055, 2016. [9] H. S. Witsenhausen, "A counterexample in stochastic optimum control," SIAM Journal of Optimal Control, vol. 6, no. 1, pp. 131 -- 147, 1968. [10] Y.-C. Ho and K.-C. Chu, "Team decision theory and information structures in optimal control problems -- part i," IEEE Transactions on Automatic Control, vol. 17, no. 1, pp. 15 -- 22, 1972. [11] A. Lamperski and J. C. Doyle, "On the structure of state-feedback lqg controllers for distributed systems with communication delays," in 50th IEEE Conference on Decision and Control and European Control Conference (CDC- ECC), pp. 6901 -- 6906, 2011. [12] L. Lessard and A. Nayyar, "Structural results and explicit solution for two-player LQG systems on a finite time horizon," in 52nd IEEE Conference on Decision and Control (CDC), pp. 6542 -- 6549, 2013. [13] P. Shah and P. Parrilo, "H2-optimal decentralized control over posets: A state-space solution for state-feedback," vol. 58, pp. 3084 -- 3096, Dec. 2013. [14] A. Nayyar and L. Lessard, "Structural results for partially nested LQG systems over graphs," in American Control Conference (ACC), 2015, pp. 5457 -- 5464, 2015. [15] L. Lessard and S. Lall, "Optimal control of two-player systems with output feedback," IEEE Transactions on Automatic Control, vol. 60, no. 8, pp. 2129 -- 2144, 2015. [16] S. Yuksel, "Stochastic nestedness and the belief sharing information pattern," IEEE Transactions on Automatic Control, vol. 54, no. 12, pp. 2773 -- 2786, 2009. [17] Y. Ouyang, S. M. Asghari, and A. Nayyar, "Stochastic teams with randomized information structures," in 56th IEEE Conference on Decision and Control (CDC), 2017. November 23, 2018 DRAFT 31 [18] A. Nayyar, A. Mahajan, and D. Teneketzis, "Optimal control strategies in delayed sharing information structures," IEEE Transactions on Automatic Control, vol. 56, no. 7, pp. 1606 -- 1620, 2011. [19] H. Witsenhausen, "Separation of estimation and control for discrete time systems," Proceedings of the IEEE, vol. 59, no. 11, pp. 1557 -- 1566, 1971. [20] P. Varaiya and J. Walrand, "On delayed sharing patterns," IEEE Transactions on Automatic Control, vol. 23, no. 3, pp. 443 -- 445, 1978. [21] T. Yoshikawa, "Decomposition of dynamic team decision problems," IEEE Transactions on Automatic Control, vol. 23, no. 4, pp. 627 -- 632, 1978. [22] M. Rotkowitz and S. Lall, "A characterization of convex problems in decentralized control," IEEE Transactions on Automatic Control, vol. 50, no. 12, pp. 1984 -- 1996, 2005. [23] S. M. Asghari and A. Nayyar, "Dynamic teams and decentralized control problems with substitutable actions," 2016. [24] Y. Ho, "Team decision theory and information structures," Proceedings of the IEEE, vol. 68, no. 6, pp. 644 -- 654, 1980. [25] H. S. Witsenhausen, "A standard form for sequential stochastic control," Mathematical Systems Theory, vol. 7, no. 1, pp. 5 -- 11, 1973. [26] A. Nayyar, A. Mahajan, and D. Teneketzis, "Decentralized stochastic control with partial history sharing: A common information approach," IEEE Transactions on Automatic Control, vol. 58, no. 7, pp. 1644 -- 1658, 2013. [27] H. Witsenhausen, "On the structure of real-time source coders," The Bell System Technical Journal, vol. 58, no. 6, pp. 1437 -- 1451, 1979. [28] B. Kurtaran, "Corrections and extensions to" decentralized stochastic control with delayed sharing information pattern"," IEEE Transactions on Automatic Control, vol. 24, no. 4, pp. 656 -- 657, 1979. [29] A. Nayyar and D. Teneketzis, "On the structure of real-time encoding and decoding functions in a multiterminal communication system," IEEE Transactions on Information Theory, vol. 57, no. 9, pp. 6196 -- 6214, 2011. [30] Y. Ouyang, S. Asghari, and A. Nayyar, "Optimal local and remote controllers with unreliable communication," in 55th IEEE Conference on Decision and Control (CDC), pp. 6024 -- 6029, 2016. [31] S. M. Asghari, Y. Ouyang, and A. Nayyar, "Optimal local and remote controllers with unreliable uplink channels," IEEE Transactions on Automatic Control, forthcoming. [32] D. S. Bernstein, R. Givan, N. Immerman, and S. Zilberstein, "The complexity of decentralized control of markov decision processes," Mathematics of operations research, vol. 27, no. 4, pp. 819 -- 840, 2002. [33] A. Mahajan and M. Mannan, "Decentralized stochastic control," Annals of Operations Research, vol. 241, no. 1-2, pp. 109 -- 126, 2016. [34] S. Yuksel and T. Bas¸ar, Stochastic Networked Control Systems: Stabilization and Optimization under Information Constraints. Springer Science & Business Media, 2013. [35] A. Mahajan and D. Teneketzis, "Optimal design of sequential real-time communication systems," IEEE Transactions on Information Theory, vol. 55, no. 11, pp. 5317 -- 5338, 2009. November 23, 2018 DRAFT APPENDIX 32 Proof of Theorem 1. We prove the result of part (i) by induction. For t = 1 the result holds since the agents have not taken any action yet. Suppose that (12) holds for t− 1. Then, Pg(cid:8)xt, xt−1, h−i t hi hi (cid:88) (cid:88) (cid:9) = Pg(cid:8)xt, xt−1, h−i (cid:9) = xt,at−1}Pg(cid:8)xt,xt−1,h−i xt,at−1}P{xtxt−1,at−1}Pg(cid:8)xt−1,h−i t−1,a−i t−1hi t−1,ai xt−1 xt−1 t t (cid:9) t−1hi t−1,yi t,zt t−1,a−i (cid:9)(cid:105) t−1, a−i t−1, y−i t hi t−1, ai t−1, yi t, zt (cid:9) xt, at−1}P{xtxt−1, at−1}g−i t−1(h−i t−1)(a−i xt, at−1}P{xtxt−1, at−1}g−i t−1(h−i t−1)(a−i t−1) t−1, yi t, zthi t−1, ai t−1 t−1,yi t,zt t−1,ai t−1)Pg(cid:8)xt−1, h−i Pg(cid:8)xt−1, h−i Pg(cid:8)yi t−1hi t, zt t−1, yi t−1, ai t, zthi t−1, ai t−1, ai t−1 (cid:9) t−1 t−1, yi t, zthi (cid:9)(cid:105) (cid:9) (cid:105) (cid:9) in the nominator of the expression above. We . (28) t−1 t−1}P{xtxt−1, a−i t−1, ai t−1}g−i t−1(h−i t−1)(a−i t−1}P{xtxt−1, a−i t−1, ai t−1}g−i t−1(h−i t−1)(a−i t−1)Pg(cid:8)xt−1, h−i t−1)Pg−i(cid:8)xt−1, h−i t−1hi t−1hi t−1, ai t−1 t−1, ai t−1 t t t t = = = = xt−1 xt−1 xt−1 P{y−i Pg(cid:8)xt,h−i (cid:88) (cid:104)P{y−i (cid:88) (cid:104)P{y−i (cid:88) (cid:104)P{y−i (cid:88) Consider the term Pg(cid:8)xt−1, h−i Pg(cid:8)xt−1, h−i (cid:104)P{yi (cid:88) (cid:104)P{yi (cid:88) =Pg−i(cid:8)xt−1, h−i t, zthi t−1, yi t, ztxt, a−i t, ztxt, a−i t, zthi t−1, ai t−1, ai t−1, ai t−1, ai t−1, yi a−i t−1,xt a−i t−1,xt have, xt−1 = = t (cid:9) (cid:9) t−1 h−i t−1,xt November 23, 2018 where the second equality follows from the induction hypothesis (12) for t−1. Consequently, we also have, Pg(cid:8)yi t,zthi (cid:88) Pg−i(cid:110) (cid:9) = (cid:88) Pg(cid:110) h−i t−1,xt t−1,ai t−1 t, zt, xt−1, h−i yi t−1hi t−1, ai t−1 by (29) = (cid:111) =Pg−i(cid:8)yi t,zt,xt−1,h−i yi t−1hi t−1,ai t−1 t,zthi t−1,ai t−1 (cid:111) (cid:9) (cid:9)(cid:105) (cid:9)(cid:105) (29) (30) DRAFT Substituting (29) and (30) in (28), t Pg(cid:8)xt,h−i (cid:104)P{y−i (cid:88) =Pg−i(cid:8)xt,h−i xt−1 = t t (cid:9) (cid:9) hi t t hi xt, at−1}P{xtxt−1, at−1}g−i t−1(h−i t−1)(a−i t−1) 33 (cid:9) (cid:105) Pg−i(cid:8)xt−1,h−i Pg−i(cid:8)yi t−1,yi t, zthi t,zthi t−1, ai t−1,ai t−1 t−1 (cid:9) which establishes the induction step for t. Given the result of part (i), the result of part (ii) follows directly from the definition of SIB strategies (10) and SIB update rule (11). To provide the proof for Theorem 2, we need the following result. Lemma 1. Given a SIB strategy profile σ and update rule ψ consistent with σ, ψ{st+1, πt+1pt, ct, at} =Pσ Pσ ψ{st+1, πt+1st, πt, at}, (31) for all st+1, πt+1, st, πt, at. Proof of Lemma 1. Let gσ denote the strategy profile, given by (10), that corresponds to SIB strategy profile σ. We have, = zt+1 using update rule (11) = Pσ ψ{st+1,πt+1pt,ct,at,πt} (cid:105) ψ{st+1, zt+1pt, ct, at, πt}1{πt+1=ψt+1(πt,zt+1)} (cid:105) ψ{st+1,πt+1pt,ct,at} πt=γt(ct) Pσ (cid:104)Pσ (cid:88) (cid:104)Pσ (cid:88) (cid:104)Pσ (cid:88) ψ{st+1, zt+1st, ct, at, πt}1{πt+1=ψt+1(πt,zt+1)} by (6) = zt+1 = ψ{st+1, zt+1, yt+1, xt+1, xtst, ct, at, πt}1{πt+1=ψt+1(πt,zt+1)} (cid:105) yt+1,xt+1,xt,zt+1 by system dynamics (1) and (2) = November 23, 2018 DRAFT (cid:88) (cid:104)Pσ ψ{st+1st, ct, at, πt, zt+1, yt+1, xt+1, xt} yt+1,xt+1,xt,zt+1 (cid:104)(cid:16)(cid:89) by (5) = 1{sj t+1=φj t+1(sj t ,{yj t+1,zt+1,aj t};gσ)} (cid:17) yt+1,xt+1,xt,zt+1 j P{zt+1, yt+1at, xt+1}P{xt+1xt, at}Pσ{xtst, ct, at, πt}1{πt+1=ψt+1(πt,zt+1)} 34 (cid:105) (cid:105) (cid:88) (cid:88) (cid:88) 1{sj t ,{yj P{zt+1, yt+1at, xt+1}P{xt+1xt, at}Pσ ψ{xtst, ct}1{πt+1=ψt+1(πt,zt+1)} t};gσ)} by Bayes' rule t+1=φj t+1(sj t+1,zt+1,aj (cid:104)(cid:16)(cid:89) (cid:17) ψ{xt, stct} P{zt+1, yt+1at, xt+1}P{xt+1xt, at}Pσ ψ{stct} 1{πt+1=ψt+1(πt,zt+1)} Pσ (cid:17) (cid:104)(cid:16)(cid:89) (cid:80) P{zt+1, yt+1at, xt+1}P{xt+1xt, at} πt(xt, st) 1{πt+1=ψt+1(πt,zt+1)} t+1=φj t+1(sj t+1,zt+1,aj t};gσ)} t ,{yj 1{sj = = (cid:105) (cid:105) πt(xt, st) xt yt+1,zt+1,xt+1,xt j yt+1,zt+1,xt+1,xt j = ψ{st+1, πt+1st, πt, at}. Pσ Proof of Theorem 2. Consider an arbitrary strategy profile g. We prove the existence of SIB strategy profile that is equivalent to g by construction. With some abuse of notation, let σi(Πt, Si t, ωt) denote agent i's strategy using the public randomization device ωt. We construct a SIB strategy profile σt that has the following properties: (a) the induced distribution on {Πt+1, St+1} under σ coincides with one under g, i.e. Pσ1:t {πt+1, st+1} = Pg1:t {πt+1, st+1} . (32) (b) the continuation payoff for all the agents under σ is the same as that under g, i.e. for all November 23, 2018 DRAFT i ∈ N , (cid:40) T(cid:88) τ =t Eg (cid:41) ui τ (Xτ , gτ (Hτ )) (cid:40) T(cid:88) τ =t = Eσ 35 (cid:41) ui τ (Xτ , στ (Πτ , Sτ , ωτ )) . (33) We prove condition (a) by forward induction and condition (b) by backward induction. We note that condition (a) is satisfied for t = 1, since at t = 1 no action has been taken. Moreover, condition (b) is satisfied for t = T + 1 since there is no future. Assume that condition (a) is satisfied from 1 to t, t ∈ T . We construct σt below such that condition (a) is satisfied at t+1. To construct σt, we first define below a random vector R1:N , and (ii) H i every i∈N , (i) R1:N is independent of Πt and S1:N t. along with Πt and Si t t t based on H 1:N , such that for t can be reconstructed using Ri t t We proceed as follows. For every time t ∈ T , let (πt, s1:N sufficient common information and private information, respectively. Let Hi denote the set of all histories of agent i at time t, where Hi ble realizations of agent i's history at time t. Conditioned on the realization of (πt, si {p(hi,k for agent i. Define the random variable Ri ) denote the realization of the agents' t , ..., hi,Hi } t denote the number of possi- t), let t that leads to (πt, si t) t)} denote the probability mass function on Hi t), 1 ≤ k ≤ H i t := {hi,1 t on [0, 1] as follows: πt, si t t t t 1) and conditioned on the event 0≤ Ri t t≤ p(hi,1 t≤p(hi,1 t πt,si t) (cid:111) 0≤Ri πt,si t) = p(hi,1 t πt,si t), , Ri t is uniformly distributed on [0,p(hi,1 t P(cid:110) (cid:110) (cid:111) (34) πt,si t)]. (35) t, 2) For 1 < k ≤ Hi P(cid:110) k−1(cid:88) (cid:104)(cid:80)k−1 j=1 t≤ k(cid:88) (cid:111) t ≤(cid:80)k j=1 p(hi,j t πt,si t) (cid:105) t) ≤ Ri t πt,si . t) (cid:110)(cid:80)k−1 t),(cid:80)k j=1 p(hi,j t πt,si j=1 p(hi,j and conditioned on the event j=1 p(hi,j distributed on t πt,si p(hi,j t πt,si t)≤ Ri = p(hi,k t πt,si t), (cid:111) j=1 p(hi,j t πt,si t) , Ri t is uniformly Therefore, Ri t is uniformly distributed on [0, 1] and is independent of (Πt, Si t). Furthermore, November 23, 2018 DRAFT for any realization (πt, si t, ri t) we can uniquely determine hi,l t where t ≥ k−1(cid:88) l := min{k : ri p(hi,j t πt, si t)}. 36 (36) j=1 Therefore, the random variable Ri t takes finite values. and (ii) when H i t defined above, satisfies the mentioned-above conditions (i) We show below that Ri t is independent of St. Lemma 2. The random variable Ri Proof of Lemma 2. Consider an arbitrary realization (h1 t ,πt, rN ...,(sN responds to hi t )) denote the realization of ((S1 t , Πt, R1 t as it is defined above for every i ∈ N . For every i ∈ N we have, t, i ∈ N , is independent of Πt and St for all t ∈ T . t ). Let ((s1 t, πt, ri t ,...,H N t )) where (si t ) of (H 1 t , Πt, RN t,...,hN t ), ..., (SN t,πt, r1 t ), T ) cor- Pg{ri tπt, st} = Pg{rtπt, si t, s−i T } = Pg{ri t, s−i Pg{s−i πt, si t} t} = πt, si t t Pg{s−i t ri t, πt, si Pg{s−i t}Pg{ri πt, si t} t tπt, si t} (πt, si replace t) by hi t, ri t = Pg{s−i t tπt, si t}Pg{ri hi t} Pg{s−i πt, si t} t (37) The last equality holds because H i (34)-(36). Moreover, t is uniquely determined by (Πt, Si t, Ri t) and vice versa; see Pg{s−i t ht} by (8) = Pg{s−i t st, ct} = Pg{s−i tct} , si tct} = Pg{si t (cid:80) t (s−i πg , si t) t t (s−i πg , si t) s−i t t = P{s−i t st, πg t }. (38) Combining (37) and (38) Pg{ri tπt, st} = Pg{s−i t si t, πg Pg{s−i t }Pg{ri tπt, si t} t} πt, si t = Pg{ri tπt, si t} = Pg{ri t} (39) where the last equality is true since by definition Ri t is independent of Πt and St for all i ∈ N . (39), Ri t is independent of (Πt, Si t). Therefore, by Using the result of Lemma 2, we prove that for every i∈N , (i) R1:N is independent of Πt t along with Πt and Si t. t can be reconstructed using Ri , and (ii) H i and S1:N t t In the following, we construct a SIB strategy profile σt equivalent to gt as follows. Let R1:N (ωt) denote a random vector the agents construct using the public randomization device ωt that has t November 23, 2018 DRAFT an identical joint cumulative distribution to that of R1:N of R1:N is independent of St and Πt. t t 37 . Note that by Lemma 2, the distribution Define, Then, σi t(Πt, Si t, ωt) := gi t(F −1 tSi Ri t,Πt ( Ri t(ωt), Πt, Si t)). (40) Pg1:t{πt+1, st+1Ht} =Pg1:t{πt+1, st+1Πt, St, Rt} distribution= Pg1:t{πt+1, st+1Πt, St, Rt} =Pσ1:t{πt+1, st+1Πt, St, Rt}. Taking the expectation of the left and right hand sides with respect to ωt and Rt, respectively, and using the fact that Rt(ωt) and Rt are independent of St and Πt (Lemma 2), we obtain Pσ1:t {πt+1, st+1Πt, St} = Pg1:t {πt+1, st+1Πt, St} w.p.1. (41) By the induction hypothesis, we have Pσ1:t−1 {πt, st} = Pg1:t−1 {πt, st}. Therefore, taking the expectation of both sides of (41) with respect to Πt, St, we establish that condition (a) holds for time t + 1. Next, we prove condition (b) by backward induction. We have, Eg{ui t(Xt, At)Ht} =Eg{ui distribution= Eg{ui =Eσ{ui t(Xt, At)Πt, St, Rt} t(Xt, At)Πt, St, Rt} t(Xt, At)Πt, St, Rt} (42) Using (42) for t = T , we have condition (b) is satisfied for t = T . Now we assume that condition (b) is satisfied from t + 1 to T , t ∈ T . We prove that condition (b) is satisfied at t. Using condition (a) at time t, i.e Pσ1:t−1{st, πt} = Pg1:t−1{st, πt}, the induction hypothesis on condition (b) for t + 1 along with equation (42) for t, and the fact that Rt and Rt are identically distributed and independent of Πt and St, we obtain (cid:41) (cid:40) T(cid:88) Eg (cid:41) (cid:40) T(cid:88) ui τ (Xτ , gτ (Hτ )) = Eσ ui τ (Xτ , στ (Πτ , Sτ , ωτ )) . τ =t τ =t November 23, 2018 DRAFT 38 Proof of Theorem 3. By the result of Theorem 2, we can restrict attention to SIB strategies with public randomization device without loss of generality. Moreover, since by Assumption 1 all space are finite, we can restrict attention to SIB strategies (with no public randomization device) without loss of generality. The proof of Theorem 3 then follows from an argument identical to the one given for dynamic programming for POMDP (see [3, Ch. 6.7]). t t ,S 1:N The dynamic program described by (15-17) can be viewed as a solution to the following decision problem that is equivalent to the original dynamic team problem. Consider a "super agent" that knows the functional forms of system dynamics and the agents' utilities, and the set of spaces Xt,A1:N for all t. The super agent coordinates the agents' decisions at each time as follows. The super agent observes πt (which is common knowledge among all agents) but of the agents' sufficient private information. Based on his does not know the realizations s1:N (πt,·), one information, the super agent chooses a joint set of prescriptions/partial functions σ1:N t) for ∀t, i. The for each agent, that determine agent i's action for every realization si dynamic program described by (15-17) determines an optimal solution for the above-described super agent, and thus, equivalently, determine the optimal strategy for the original dynamic team problem.9 t as σi t(πt, si t t Proof of Theorem 4. We show below that Li (i)-(iv) of Definition 3. t := (Si t, Πt), i ∈ N , t ∈ T satisfies conditions and Πt can be updated recursively via update rules φ1:N t Condition (i) is satisfied since both S1:N and ψt, respectively, for every t ∈ T . Condition (ii) is satisfied by Lemma 1. To prove condition (iii), we have t P{xtct, st} = Therefore, (cid:80) P{xt, stct} P{xt, stct} = P{xtπt, st}. xt (43) 9The above interpretation of the dynamic program from the point of view of a super agent is similar to the coordinator problem formulated in [18], [26]. November 23, 2018 DRAFT 39 Eg−i 1:t−1 ui t(Xt,At) t,at = Eg−i 1:t−1 ui t(Xt,At) t,at ui t(Xt,At) t, at (cid:111) (cid:12)(cid:12)(cid:12)ct,si (cid:41) (cid:110) (cid:111)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)ct, si (cid:12)(cid:12)(cid:12)Xt,at (cid:111)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)πt, si (cid:12)(cid:12)(cid:12)Xt,at (cid:12)(cid:12)(cid:12)πt,pi (cid:111) t,at (cid:41) t, at (cid:110) (cid:40) (cid:40) Eg−i 1:t−1 Eg−i 1:t−1 (cid:111) by (7) = (cid:12)(cid:12)(cid:12)ct,pi (cid:110) 1:t−1 Eg−i by (43) = Eg−i 1:t−1 ui t(Xt,At) = Eg−i 1:t−1 ui t(Xt,At) (cid:110) s−i t (cid:110) (cid:111) by (7) (cid:111) (cid:110) (cid:110) (cid:12)(cid:12)(cid:12)pi (cid:12)(cid:12)(cid:12)si t Condition (iv) holds since, (cid:110) (cid:111) (cid:12)(cid:12)(cid:12)pi Pg−i 1:t−1,gi 1:t−1 l−i t t,ct = Pg−i 1:t−1 t,ct = Pg−i 1:t−1 s−i t t,ct = Pg−i 1:t−1,gi 1:t−1{stct} 1:t−1{si Pg−i tct} by (43) = Pg−i 1:t−1 s−i t,πt = Pg−i 1:t−1 l−i t (cid:111) (cid:110) (cid:110) (cid:12)(cid:12)(cid:12)si (cid:12)(cid:12)(cid:12)li (cid:111) t Proof of Theorem 5. Consider the SIB strategy σt constructed in the proof of Theorem 2 for every t ∈ N. We show below that σt satisfies (26). By the proof of Theorem 2, condition (32) holds for all t ∈ N. To prove (26), we show that (cid:41)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) ≤ δt−1ui∞(g1:N τ (H 1:N τ ),Xτ) − Eσ∞ δt−1ui∞(σ1:N∞ (Πτ,S1:N τ ),Xτ) (44) τ =t τ =t Let M = maxat,xt,i ui∞(xt, at). For every > 0, choose T ∈ N such that δT 1−δ M ≤ 2. Then, (cid:41) (cid:40) ∞(cid:88) under strategy σt, t ∈ N, we have (cid:40) ∞(cid:88) (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)Eg for all > 0. for any arbitrary strategy g,(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)Eg (cid:40) ∞(cid:88) τ =T (cid:41)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) ≤ 2 δt−1ui∞(g1:N τ (H 1:N τ ),Xτ) . (45) Therefore, for every t < T , condition (44) is satisfied by (45) and the triangle inequality. November 23, 2018 DRAFT For t > T , consider a finite decision problem with horizon T resulting by the truncation of the original infinite-horizon decision problem at T . Then, by Theorem 2, (cid:40) T(cid:88) Eg (cid:41) (cid:40) T(cid:88) 40 (cid:41) τ (g1:N ui τ (H 1:N τ ),Xτ) = Eσ u1:N τ (σi τ(Πτ,S1:N τ ),Xτ) , (46) τ =t τ =t for all i∈N and t∈T . Combining (46) with the result for t > T , we show that (26) is satisfied for t. Proof of Theorem 6. By Theorem 5, we can restrict attention to stationary SIB strategies with public randomization device without loss of generality. Moreover, since by Assumption 1 all space are finite, we can restrict attention to SIB strategies (with no public randomization device) without loss of optimality. Consequently, following the same rationale as the one given in the proof of Theorem 3, the result of Theorem 6 follows from an argument identical to the one given for dynamic programming in infinite-horizon Markovian Decision Processes (see [3, Ch. 8.2 and Ch.8.3]). November 23, 2018 DRAFT
1709.04622	4	1709	2018-08-23T04:20:34	An Agent-based Modelling Framework for Driving Policy Learning in Connected and Autonomous Vehicles	[ "cs.MA", "cs.RO" ]	Due to the complexity of the natural world, a programmer cannot foresee all possible situations, a connected and autonomous vehicle (CAV) will face during its operation, and hence, CAVs will need to learn to make decisions autonomously. Due to the sensing of its surroundings and information exchanged with other vehicles and road infrastructure, a CAV will have access to large amounts of useful data. While different control algorithms have been proposed for CAVs, the benefits brought about by connectedness of autonomous vehicles to other vehicles and to the infrastructure, and its implications on policy learning has not been investigated in literature. This paper investigates a data driven driving policy learning framework through an agent-based modelling approaches. The contributions of the paper are two-fold. A dynamic programming framework is proposed for in-vehicle policy learning with and without connectivity to neighboring vehicles. The simulation results indicate that while a CAV can learn to make autonomous decisions, vehicle-to-vehicle (V2V) communication of information improves this capability. Furthermore, to overcome the limitations of sensing in a CAV, the paper proposes a novel concept for infrastructure-led policy learning and communication with autonomous vehicles. In infrastructure-led policy learning, road-side infrastructure senses and captures successful vehicle maneuvers and learns an optimal policy from those temporal sequences, and when a vehicle approaches the road-side unit, the policy is communicated to the CAV. Deep-imitation learning methodology is proposed to develop such an infrastructure-led policy learning framework.	cs.MA	cs	An Agent-based Modelling Framework for Driving Policy Learning in Connected and Autonomous Vehicles … … Varuna De Silva, Xiongzhao Wang, Deniz Aladagli, Ahmet Kondoz, Erhan Ekmekcioglu Institute for Digital Technologies, Loughborough University, London, UK Email: [email protected], [email protected], [email protected], and [email protected] framework exchanged with other vehicles and Abstract -- Due to the complexity of the natural world, a programmer cannot foresee all possible situations, a connected and autonomous vehicle (CAV) will face during its operation, and hence, CAVs will need to learn to make decisions autonomously. Due to the sensing of its surroundings and information road infrastructure, a CAV will have access to large amounts of useful data. While different control algorithms have been proposed for CAVs, the benefits brought about by connectedness of autonomous vehicles to other vehicles and to the infrastructure, and its implications on policy learning has not been investigated in literature. This paper investigates a data driven driving policy learning through an agent-based modelling approaches. The contributions of the paper are two-fold. A dynamic programming framework is proposed for in-vehicle policy learning with and without connectivity to neighboring vehicles. The simulation results indicate that while a CAV can learn to make autonomous decisions, vehicle-to-vehicle (V2V) communication of this capability. Furthermore, to overcome the limitations of sensing in a CAV, the paper proposes a novel concept for infrastructure-led policy learning and communication with autonomous vehicles. In infrastructure-led policy learning, road-side infrastructure senses and captures successful vehicle maneuvers and learns an optimal policy from those temporal sequences, and when a vehicle approaches the road-side unit, the policy is communicated to the CAV. Deep-imitation learning methodology is proposed to develop such an infrastructure-led policy learning framework. information improves Keywords -- Agent-based learning; reinforcement driving policy; data driven control; imitation learning learning; I. INTRODUCTION The emergence of connected and autonomous vehicles (CAVs) marks a new phase of innovation in automotive and transportation industries since the development of personal automobiles [1]. Once technologically matured, CAVs would be able to move autonomously without the aid of a human driver and be able to communicate with other vehicles and traffic infrastructure. The promised benefits of CAVs are numerous, such as, reduced congestion, better fuel efficiency, reduced environmental pollution, reduced number of traffic related casualties and increased personal independence. The continuing evolution of sensor technologies, high speed communication infrastructure, machine learning and artificial intelligence will eventually bring us the above benefits. A CAV is composed of four major technological components. The first is the perception system, which is responsible for sensing the environment to understand its surroundings. The second component is the localization and mapping system that enables the vehicle to know its current location. The third component is responsible for the driving policy. The driving policy refers to the decision making capability of a CAV under various situations, such as negotiating at roundabouts, giving way to vehicles and pedestrians, and overtaking vehicles. Finally, the CAVs will be connected. It is expected that the CAVs will be connected to the surrounding vehicles: vehicle to vehicle connectivity (V2V), to the infrastructure: Vehicle to Infrastructure (V2I) and to anything else such as the internet: Vehicle to Anything (V2X), through wireless communications links [2]. While there are many challenges still to be addressed for high speed wireless connectivity for vehicular applications [3], the IEEE802.11p Wireless Access in Vehicular Environments (Wave) is considered the most relevant standard that currently caters to the requirements of such applications [4]. to The connectedness of the CAV's can be useful for many important functions related intelligent mobility. One important use case is to exchange sensor data between vehicles for improved perception of the surroundings [5], which enables to reduce the accidents. Connected vehicles also enables centralized traffic control to ease congestion in smart city applications [6]. For example in Compass4D, a pilot project funded by the EU, it was demonstrated that 15% reductions in fuel efficiency can be achieved through V2I communication to control the flow of traffic [7]. Another interesting application of connectivity is in "vehicle platooning", where a group of vehicles with common interests maintain a small and constant distance to each other [8]. Connectivity is also a key enabler of intelligent mobility applications such as ubiquitous taxi services. Furthermore, various infotainment applications such as music/video streaming or in-car WiFi connectivity, are expected to be delivered to the CAVs through high speed wireless internet connectivity [4]. All the above examples lead to a significant amount of data that are been collected and shared between CAVs. This leads to the question, whether CAV could learn interesting patterns from these data? Could a CAV autonomously evolve to make better decisions over time? Data driven approaches are emerging within the scope of CAV research to utilize this data in effective ways. In [9], authors propose a data-driven control algorithm derivation for .. 1 P a g e connected vehicle to account for unknown system dynamics such as human factors. In this paper we investigate opportunities that arise from the proliferation of data within a CAV. In particular, this paper focuses on optimal policy learning in autonomous vehicles over its life time. For example, how could it make better decisions to reach its destination while avoiding crashes with other objects or road side infrastructure? We consider a use case where a CAV is able to make driving decisions based on information shared by other CAVs. The proposed framework is based on agent based learning paradigm where a CAV is modelled as a rational agent that tries to maximize its utility. Furthermore, an learning and communicating framework is proposed based on deep- imitation learning. infrastructure-led policy The rest of this paper is organized as follows: Section II provides an overview of the agent based learning paradigm; Section III describes the experimental setup and the results; and Section IV concludes the paper with reference to some future work. II. RELATED WORK In this section we will discuss work related to driving policy learning in connected autonomous vehicles. This section is organized as follows: Section A describes work related to policy learning in CAVs, including applications of Vehicle-to- Infrastructure (V2I) communication. Section B describes the foundations of imitation learning and its applications in driverless vehicle technology. A. Policy Learning in CAVs Policy learning is the process by which an autonomous vehicle comes to learn which action to take given a particular situation. The situation or the state in which the CAV is operating is identified through sensing the environment and understanding it. the cases of Different types of control problems were formulated for CAVs considering range-limited V2V communication and input saturation to minimize the errors of distance and velocity and to optimize the fuel usage [9]. In [9] authors employed an adaptive dynamic programming technique, to derive the optimal controllers without relying on the knowledge of system dynamics. The effectiveness of the proposed approaches in [9] was demonstrated via the online learning control of the connected vehicles in platform for traffic microsimulation. Reinforcement learning was utilized for longitudinal control of autonomous vehicles [10]. A deep Reinforcement learning algorithm based on neural network-based Q-function approximation was proposed for ramp merging process for autonomous vehicles [11]. The ramp merging process involves interactions with other vehicles whose behaviors dynamic and varied and which influences the actions of the merging vehicles. Similarly, reinforcement learning based controllers were proposed for autonomous vehicles for tasks such as lane changing [12] and overtaking decisions[13]. While policy learning in CAV is an emerging area of research, agent-based models are utilized for various control … … functions in vehicle control. A learning based cruise control algorithm was proposed in [14]. B. Applications of Deep Imitation Learning Imitation learning is the process of learning by imitating certain set of actions. Imitation learning from human drivers for the task of lane keeping assistance in highway and country roads using grayscale images from a single front view camera was proposed in [15]. The employed method in [15] utilizes convolutional neural networks (CNN) to analyze the images to develop an appropriate policy to adhere to the lanes. Deep imitation learning was proposed as a method to model the defensive behavior of football players [16]. In [16], Multiview video based player tracking data are collected and the team behavior is learnt. For this purpose, a recurrent neural network, specifically a Long-Short Term Memory (LSTM) network is trained with sequences of player tracking data. C. Contributions of this Paper While different aspects of control systems have been developed for CAVs, the advantages of connectedness of vehicles and its implications for policy learning has not been investigated in literature. The contributions of this paper are twofold: firstly, we propose a dynamic programming approach for policy learning in an autonomous vehicle and compare it with a situation where the vehicles are connected to each other to share sensor information. Secondly, we propose a framework for policy learning at to be communicated to approaching vehicles. This way road-side units at which many examples of successful vehicle maneuverers can be observed, could be utilized for the safety of CAVs. road-side units, and for policies III. AGENT BASED MODELLING FRAMEWORK FOR CAVS In this section we will discuss the agent based learning framework that will be used for algorithm development in this paper. A. Foundations of Reinforcement Learning In the recent past, reinforcement learning has been utilized in attractive research projects. DeepMind, a modern technique research organization hosted in London, which concentrates on data mining and machine learning, has made landmark progress on reinforcement learning research. The world champion in 'Go', Lee Sedol was beaten by the deep mind developed program called AlphaGo [17]. Fig. 1 illustrates the fundamental process of reinforcement learning. Actions taken in the current environment would transfer the agent to a different state, and a reward would be given based on how good the action was. A utility function that assigns a different reward value for each possible actions in the same state, could be used to evaluate the best action in that particular state. This interaction scenario gives a general idea of reinforcement learning and builds the foundation cooperating with Markov Decision Processes (MDP) and Bellman equations [18]. .. 2 P a g e … … IMMEDIATE REWARDS FUNCTION TABLE I. Condition (State/Action) Alive or Goal Shift to right or left Crash or Bump Speed less than limit and alive or goal Speed less than limit with bump or crash Speed over limit B. Micro-simulation Framework for CAV Policy Learning Value 0.1 -0.1 -10 speed/10 - speed - 2×speed The CAV is simulated as a car that acquires sensor data of its surroundings and exchanges speed information with other vehicles in its surroundings through V2V communication. Each vehicle is simulated as a point on a 2D grid and ignores any spatial parameters such as the size of the vehicle. Fig. 2 illustrates the environmental scanner implemented on the CAV, which simulates a 7 directional scanner. the structure of At a given state, the agent can take nine (3x3) actions, which control the movement direction and speed. They are the three direction actions: maintain current direction, which mean no change in the status of vehicle, move left and move right and the three speed actions: maintain, increase and reduce the speed of agent. Other decisions, such as making turns at intersections or stop at target point, will not be considered in the current study. The immediate reward function for the agent is summarized in Table I. The prototype environment simulates a straight road that is 66 units long with two lanes of 1 unit width. The movement of each agent is simulated as a maze walk, where they move number of units at each time step according to the current speed. There would be a number of other randomly located vehicles, acting as obstacles for the intelligent agent, trying to move to the same destination. In this experiment we consider only one intelligent agent that adapts its decisions over time and the rest of the vehicles will move at a constant speed over a single simulation. The simulation set the lane above, which is the left-hand lane as a normal speed route, vehicles would move at one speed unit, and the other lane is for overtaking, where vehicles would have one speed unit higher. The intelligent agent would be able to change directions and speeds as discussed before. To simulate V2V communication, if the radar detects an obstacle vehicle through its environmental scanner, the agent will request for the speed of the vehicle. The obstacle vehicle would give its speed-reading to the agent, which enables the agent to use additional information to define the state of the environment. Fig. 1. The reinforcement learning interaction between the agent and its environment. to Markov decision processes are fundamental the development of active reinforcement learning algorithms. The state transition in RL can be described as a Markov process, where the future states are dependent on the current state only and are independent of the states in the past. An active reinforcement learning approach known as Q-value iteration is used to find the optimal reward function and the transition model. Q-value iteration uses direct rewards value as samples, and update the corresponding Q-value by a proportional summation with the initial value, which depends on learning rate 𝛼𝛼 as given below: 𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠=𝑅𝑅(𝑠𝑠,𝑠𝑠,𝑠𝑠′)+𝛾𝛾∙𝑠𝑠𝑠𝑠𝑚𝑚𝑎𝑎′ 𝑄𝑄(𝑠𝑠′,𝑠𝑠′). 𝑄𝑄(𝑠𝑠,𝑠𝑠)←(1−𝛼𝛼)∙𝑄𝑄(𝑠𝑠,𝑠𝑠)+𝛼𝛼∙𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠. to maximize its discounted future reward. 𝛾𝛾 is the discount Where R(s,a,s') is the immediate reward value given when an action a is taken at state s, which moves the agent to state s'. Q(x,y) in (1) and (2) refers to the expected future reward if action y is taken at state x. The agent acts in a rational manner factor. A dynamic programming approach is used to solve the above equations in an iterative manner. (1) (2) Fig. 2. The structure of the environmental scanner on the simulated CAV. Physical infrastructure with sensing Scene Perception and positive example identification Deep imitation learning for Data-driven driving policy Policy uploading to vehicles approaching infrastructure Fig. 3. System architecture for proposed Data-driven Vehicle-Infrastructure Policy Communication Scheme. .. 3 P a g e … … Fig. 4. An example sequence of states when a vehicle is trying to join a motorway/freeway. C. Deep Imitation Learning for Infrastructure-led Policy Learning The above sections described how a CAV with sensing and V2V communication ability could be trained to learn behavior over time. While such a model is suitable for simple scenarios like straight line driving, CAV's often come to deal with more tricky scenarios such as negotiations at roundabouts, joining motorways (freeways) and giving way to pedestrians etc. Moreover, most of such negotiating situations could be unique in its appearance with a unique local context attached to it. E.g. there could be tricky points such as give-way situations, which may go unnoticed to the driver / CAV if it has not come across it before. Could the road-side infrastructure located at such points, look at the traffic scenarios and learn the optimal behavior of a vehicle that is approaching it? If this is possible, then, when a vehicle is approaching the tricky point of negotiation, the infrastructure could instruct the CAV of the best policy. To handle such situations, in this section, we propose a novel algorithm for infrastructure-led policy learning. The system architecture for the proposed scheme for policy learning and communication is illustrated in Fig. 3. According to Fig. 3, the cameras and sensors based at key junctions and intersections will capture the passing vehicles and pedestrians. The captured data will be analyzed to recognize vehicle, pedestrians and other objects of interest. A sequential data classifier will classify positive negotiations from negative negotiations. The positive negotiations are temporal sequences of events without any casualties or collisions. Once positive sequences of events are identified, the event sequences are utilized to learn an optimal policy for vehicle control. The optimal policy is learnt through imitation learning. Recurrent neural networks provide a framework to efficiently learn patterns in sequential data streams. In particular, Long-Short Term Memory (LSTM) networks are suitable variation of recurrent neural networks that does not suffer from the vanishing gradient problem associated with vanilla recurrent neural networks. An example sequence of a positive outcome is depicted in Fig. 4. The sequence of steps taken by the marked car is a successful sequence of actions that a driver could follow to join the motorway. Our proposal is that, similar outcomes can be collected and the sequential pattern can be learnt. To learn the sequential patterns, and to come up with an optimal policy, the state of the environment will be encoded in a vector that represents the distance and speed of each vehicle in its surrounding, and fed in to the LSTM network. The output of the LSTM network will be the speed and angle of the vehicle at each time step. The pattern learnt, will yield in an optimal policy. Now when a CAV that has not experience this setting in the past comes to negotiate at this junction, the learnt policy can be transferred to it. This way we can overcome the requirement for a CAV to experience situations multiple times to learn a policy. Fig. 5. A sequence sample of agent actions. D. Simulation Framework for Infrastructure-led Policy Learning To illustrate the benefits of the proposed infrastructure-led policy learning scheme, we develop an experimental testbed based on a micro simulation framework and a python/Keras based neural network setup. While we do not have in possession a suitable data set, it is assumed that through sensing and perception an output similar to figure 4 can be obtained. Therefore, for the purpose of this simulation, the Simulation of Urban Mobility (SUMO) open source platform is utilized [19]. The output of the SUMO platform is dumped in XML format, which is processed to identify appropriate sequences of events, surrounding vehicles, and speed, and position of vehicles. Several positive outcomes are collected, and encoded, and fed to the LSTM network as training data. Once the network is trained, it will be used to predict the sequential actions of a CAV that is intending to join the motorway. .. 4 P a g e IV. EXPERIMENTAL RESULTS AND DISCUSSIONS testbed, the experimental This section describes simulation results and the discussion of the results. A. Experimental Conditions the For evaluating dynamic programming-based policy learning method for CAV, a simulation environment has been implemented in Matlab 2017a environment. Three different experiments are performed to test the fundamental process of decision making. The learning rate α and discount factor γ is set to 0.4 and 0.95, respectively. The maximum speed of the agent is set at 3 units per time step. (a) (b) (c) Fig. 6. Simulation results for comparison between the agent performance with and without V2V communication. (a) Average time to reach destination. (b) Amount of crashes (c) Frequency of quick finish (reach destination<40 time units). … … The infrastructure-led policy learning scheme is simulated with SUMO micro-simulation platform (version 0.32.0) and the LSTM network is implemented based on the Keras/python package. B. Simulation Results Fig. 5 shows a sequence of changes between agent and obstacle vehicles at different time steps. The orange square represents the location of agent; blue squares are the moving obstacle vehicles. The numbers above each picture are demonstrating the current speed of agent, seven readings from each direction, and the action it made to move it from previous states. From the initial five pictures, it shows that the agent tried to increase the speed and move to overtake other objects. From the last three pictures and changing speed values, it can be seen that a decision was taken to decrease the speed and follow the other vehicles, as there is no possible way to overtake them. In the next set of results, we compare between the two situations where the agent learns with just the environment scanner and when the agent has access to speed information of the through V2V communication. Firstly, let's consider the average time cost, i.e. the number of time steps that the agent took to get from the starting point to its target successfully. surrounding exchanged vehicles As the graph in Fig. 6(a) illustrates the time consumption to get to the destination gradually decreases as the number of iterations of experiment increases. This means that the agent self learns to make better decisions as it experiences more situations. It takes Approximately 25 units of time in average to reach the destination after 100,000 iterations of Q-learning. The pattern of the agent with vehicle communication, to share speed information of the surrounding neighbors, demonstrate to converge quickly than the scenario without the V2V communication. At the same time the time consumption to reach the destination gradually decreases too. In Fig. 6(b), the probability of crashing illustrates the rate of agent having crashes with other vehicle or bumping to the wall. It clearly present that after a period of Q-learning, the agent would have the ability to avoid obstacles, while its moving and reduces the possibility of having crashes with others. Fig. 6(b) also presents that with V2V communication, the agent keeps a lower possibility of crashing, compared to agent without V2V communication, which has a higher potential to crash Lastly, Fig. 6(c) shows the percentage of time an agent takes a faster route, which costs less than 40 units of time. The initial results of the infrastructure-led policy learning framework can be illustrated as given in Fig. 7(a)-(c). Fig. 7(a) illustrates the LSTM network architecture that is trained for the scenario. The Fig. 7(b)-(c) illustrate two different instances of a vehicle trying to join a motorway. In the two different instances illustrated, we compare the variation of speed of the vehicle as given by the micro-simulation platform and the speed predicted by the proposed LSTM network at different time steps. .. 5 P a g e (a) (b) (c) Fig. 7. Simulation results for infrastructure-led policy learning of a vehicle joining a motorway: (a) The LSTM network architecture and the variation of speed as approaching and joining the motorway. (b) test scenario 1. (c) test scenario 2. … … C. Discussion of Results The above results illustrate a significant improvement and an efficiency of learning after the first two thousand iterations. In the following iterations, the pattern of changing is slowing down. While the process of decision-making in the agent improves over the iterations, the agent with vehicle speed communication seems to have a better performance than the agent with only the environmental scanner. the additional parameters obtained In summary, agent with V2V communication would have a better efficiency in learning to make better driving policies. With through V2V communication, the agent could perform better as it can explore of more states, due to unknown policies of the states it never reached before. Thus, the speed information shared by the vehicles acts as an important additional dimension which significantly improves the decision making capability of the simulated CAV. Considering with the complexity of state locating, the agent with vehicle speed communication has approximately twice the number of dimensions to represent its state of the environment. As the simulation conditions become complex, such as for V2I communication or when considering real CAVs, this would lead to bottlenecks in terms of computational complexity. Deep reinforcement learning promises to tackle such situations of very high dimensions. For the infrastructure-led policy learning framework, while the results illustrated in Fig. 7 are encouraging, the results could be improved further by tuning the structure of the LSTM network and increasing the number of training data sequences. It has to be noted, that only a simple LSTM network was setup in this case to prototype the idea. V. CONCLUSIONS AND FUTURE WORK The connected autonomous vehicles (CAVs) operate by sensing its surroundings and exchanging information from other vehicles and infrastructure. This paper investigated the possibility for CAVs to learn from these data sources and derive its own safe driving policies. A reinforcement learning framework is presented in the paper, which simulates the self- evolution of a CAV over its lifetime. The results indicated that overtime the CAVs are able to learn useful policies to avoid crashes and achieve its objectives in more efficient ways. Vehicle to vehicle communication enables additional useful information to be acquired by CAVs, which in turn enables CAVs to learn driving policies more efficiently. Furthermore, the paper proposed an infrastructure-led policy learning and communication framework based on deep imitation learning. The initial validation results based on a microsimulation platform and a LSTM network were encouraging. The future work will investigate complex driving policies such as roundabout negotiations, cooperative learning between CAVs and deep reinforcement learning to traverse larger state spaces. Furthermore, we will validate the infrastructure-led policy learning framework with real world datasets on vehicle tracking. .. 6 P a g e REFERENCES [1] T. Koslowski, "Your Connected Vehicle Is Arriving," MIT Technology Available: Review. https://www.technologyreview.com/s/426523/your-connected-vehicle- is-arriving/. [Accessed: 31-Aug-2017]. [Online]. [2] G. Pau, "Quickly Home Please: How Connected Vehicles Are Revolutionizing Road Transportation," IEEE Internet Computing, vol. 17, no. 1, pp. 80 -- 83, Jan. 2013. [3] N. Lu, N. Cheng, N. Zhang, X. Shen, and J. W. Mark, "Connected Vehicles: Solutions and Challenges," IEEE Internet of Things Journal, vol. 1, no. 4, pp. 289 -- 299, Aug. 2014. [4] R. Coppola and M. Morisio, "Connected Car: Technologies, Issues, Future Trends," ACM Comput. Surv., vol. 49, no. 3, p. 46:1 -- 46:36, Oct. 2016. [5] B. Reichelt, H. Ussler, O. Michler, J. Holfeld, S. Eckelmann, and T. Trautmann, "Model-based generation and validation of different sensor information contributing to a fusion algorithm in connected vehicles," in 2017 5th IEEE International Conference on Models and Technologies for Intelligent Transportation Systems (MT-ITS), 2017, pp. 339 -- 344. J. Hu, L. Kong, W. Shu, and M.-Y. Wu, "Scheduling of connected autonomous vehicles on highway lanes," in 2012 IEEE Global Communications Conference (GLOBECOM), 2012, pp. 5556 -- 5561. [6] [7] R. King, "Traffic management in a connected or autonomous vehicle environment," in Autonomous Passenger Vehicles, 2015, pp. 1 -- 20. [8] D. Jia, K. Lu, J. Wang, X. Zhang, and X. Shen, "A Survey on Platoon- Based Vehicular Cyber-Physical Systems," IEEE Communications Surveys Tutorials, vol. 18, no. 1, pp. 263 -- 284, Firstquarter 2016. [9] W. Gao, Z. P. Jiang, and K. Ozbay, "Data-Driven Adaptive Optimal Control of Connected Vehicles," IEEE Transactions on Intelligent Transportation Systems, vol. 18, no. 5, pp. 1122 -- 1133, May 2017. [10] Z. Huang, X. Xu, H. He, J. Tan, and Z. Sun, "Parameterized Batch Reinforcement Learning for Longitudinal Control of Autonomous Land … … Vehicles," IEEE Transactions on Systems, Man, and Cybernetics: Systems, pp. 1 -- 12, 2017. [11] P. Wang and C. Y. Chan, "Formulation of deep reinforcement learning architecture toward autonomous driving for on-ramp merge," in 2017 IEEE 20th International Conference on Intelligent Transportation Systems (ITSC), 2017, pp. 1 -- 6. [12] F. Liu, Y. Li, L. Zuo, and X. Xu, "Research on Lane Change Decision for Autonomous Vehicles Based on Multi-Kernels Least Squares Policy Iteration," in 2017 9th International Conference on Intelligent Human- Machine Systems and Cybernetics (IHMSC), 2017, vol. 1, pp. 385 -- 389. [13] X. Li, X. Xu, and L. Zuo, "Reinforcement learning based overtaking decision-making for highway autonomous driving," in 2015 Sixth International Conference on Intelligent Control and Information Processing (ICICIP), 2015, pp. 336 -- 342. [14] J. Wang, X. Xu, D. Liu, Z. Sun, and Q. Chen, "Self-Learning Cruise Control Using Kernel-Based Least Squares Policy Iteration," IEEE Transactions on Control Systems Technology, vol. 22, no. 3, pp. 1078 -- 1087, May 2014. [15] C. Innocenti, H. Lindén, G. Panahandeh, L. Svensson, and N. Mohammadiha, "Imitation learning for vision-based lane keeping assistance," in 2017 IEEE 20th International Conference on Intelligent Transportation Systems (ITSC), 2017, pp. 425 -- 430. [16] H. M. Le, P. Carr, Y. Yue, and P. Lucey, "Data-Driven Ghosting using Analytics Conference, 2017. Deep Imitation Learning," in MIT Sloan Sports [17] "DeepMind's work Available: [Accessed: 01-Sep-2017]. in 2016: a round-up," DeepMind. [Online]. https://deepmind.com/blog/deepmind-round-up-2016/. [18] R. S. Sutton and A. G. Barto, Reinforcement Learning. MIT Press, 1998. [19] "DLR - Institute of Transportation Systems - SUMO -- Simulation of Available: Urban http://www.dlr.de/ts/en/desktopdefault.aspx/tabid-9883/16931_read- 41000/. [Accessed: 10-Apr-2018]. MObility." [Online]. .. 7 P a g e
1604.04727	1	1604	2016-04-16T11:37:50	Tasks for agent-based negotiation teams: Analysis, review, and challenges	[ "cs.MA", "cs.AI" ]	An agent-based negotiation team is a group of interdependent agents that join together as a single negotiation party due to their shared interests in the negotiation at hand. The reasons to employ an agent-based negotiation team may vary: (i) more computation and parallelization capabilities, (ii) unite agents with different expertise and skills whose joint work makes it possible to tackle complex negotiation domains, (iii) the necessity to represent different stakeholders or different preferences in the same party (e.g., organizations, countries, and married couple). The topic of agent-based negotiation teams has been recently introduced in multi-agent research. Therefore, it is necessary to identify good practices, challenges, and related research that may help in advancing the state-of-the-art in agent-based negotiation teams. For that reason, in this article we review the tasks to be carried out by agent-based negotiation teams. Each task is analyzed and related with current advances in different research areas. The analysis aims to identify special challenges that may arise due to the particularities of agent-based negotiation teams.	cs.MA	cs	Tasks for Agent-Based Negotiation Teams: Analysis, Review, and Challenges Victor Sanchez-Anguixa, Vicente Juliana, Vicente Bottia, Ana Garc´ıa-Fornesa aDepartamento de Sistemas Inform´aticos y Computaci´on Universitat Polit`ecnica de Val`encia Cam´ı de Vera s/n, 46022, Valencia, Spain Abstract An agent-based negotiation team is a group of interdependent agents that join together as a single negotiation party due to their shared interests in the nego- tiation at hand. The reasons to employ an agent-based negotiation team may vary: (i) more computation and parallelization capabilities; (ii) unite agents with different expertise and skills whose joint work makes it possible to tackle complex negotiation domains; (iii) the necessity to represent different stake- holders or different preferences in the same party (e.g., organizations, countries, married couple, etc.). The topic of agent-based negotiation teams has been re- cently introduced in multi-agent research. Therefore, it is necessary to identify good practices, challenges, and related research that may help in advancing the state-of-the-art in agent-based negotiation teams. For that reason, in this ar- ticle we review the tasks to be carried out by agent-based negotiation teams. Each task is analyzed and related with current advances in different research areas. The analysis aims to identify special challenges that may arise due to the particularities of agent-based negotiation teams. Keywords: Negotiation teams, automated negotiation, agreement technologies, multiagent systems. 1. Introduction According to Pruitt [1], negotiation can be defined as a process in which a joint decision is made by two or more parties. The parties first verbalize contra- dictory demands and then move towards agreement by a process of concession- making or search for new alternatives. Therefore, automated negotiation con- sists of an automated search process for an agreement between two or more parties where participants exchange proposals. Email addresses: [email protected] (Victor Sanchez-Anguix), [email protected] (Vicente Julian), [email protected] (Vicente Botti), [email protected] (Ana Garc´ıa-Fornes) Preprint submitted to Elsevier April 19, 2016 The work in automated negotiation has taken two alternative and comple- mentary paths: game-theoretic approaches and heuristic approaches. Game theory researchers focus on reaching solutions at equilibrium under varied as- sumptions like unbounded computational resources, complete information, or some statistical information regarding the strategies and preferences of the other parties [2, 3, 4]. Unfortunately, some of these assumptions may not be present in some real world applications. For instance, computational resources are of extreme im- portance for agents since they may be scarce and shared among different tasks (e.g., time, memory usage, etc.). Thus, negotiation should not always assume unbounded computational resources. Additionally, since agents are heteroge- neous, not all of the agents know the same strategies. Identifying which sets of strategies are known by each agent may be a hard task that can only be success- ful after several negotiations. The same goes for the knowledge regarding the opponents' preferences, reservations values, and so forth. Only after several in- teractions and negotiations with the opponents, the agent may be able to come with an approximate model of the other agents. Hence, models that tackle un- certainty and limit the use of computational resources are mandatory for some situations. Heuristic models usually avoid the aforementioned assumptions. In result, they usually do not find equilibriums, but they obtain satisfactory re- sults [5, 6]. From this point, this work assumes the heuristic perspective when computational models for automated negotiation are mentioned. Most of the research in automated negotiation has concentrated on bilateral bargaining and multi-party negotiations where parties are composed of single individuals. However, most real life negotiations, especially in business and politics, involve negotiation teams [7, 8, 9]: multi-individual parties that act together as a single negotiation party. The use of negotiation teams in auto- mated negotiation and multi-agent systems is relatively new [10, 11, 12, 13, 14]. Hence, a proper review needs to be done in order to identify the tasks and spe- cial challenges that may arise in negotiation teams' settings. The main goal of this article is analyzing the tasks to be carried out by agent-based negotiation teams, reviewing how current works may contribute to the consecution of such tasks, and identifying special challenges that may arise due to the particularities of agent-based negotiation teams. We also hope that this present work may help new researchers in the area to have a global perspective on the problems that may arise in agent-based negotiation teams. The remainder of this paper is organized as follows. In Section 2 we start by introducing negotiation teams from the point of view of the social sciences. We identify the main aspects that characterize human negotiation teams and, after that, we relate those aspects with the topic of agent-based negotiation teams. Then, we present an analysis on the tasks to be carried out by agent-based negotiation teams in Section 3. Finally, we conclude by overviewing the work presented in this article. 2 2. Motivation As a first introspective to negotiation teams, we describe the phenomenon from the point of view of the social sciences. Then, we relate the essential aspects identified in the social sciences with agent-based negotiation teams. 2.1. Human Negotiation Teams A negotiation team is a group of two or more interdependent persons who join together as a single negotiating party because their similar interests and objectives relate to the negotiation, and who are all present at the bargaining table [7, 9]. Hence, a negotiation team is a negotiation party that is formed by multiple individuals instead of just one individual. As a negotiation party, the team negotiates with other parties in order to reach a final agreement. In what kind of scenarios may a negotiation be involved? There are several scenarios and their importance range from day to day negotiations to crucial negotiation in real life like business and politics. For instance, we can think of the following negotiation cases where teams usually participate in real life: • In one scenario, a human organization like a company desires to sell a product line to another company [8]. It is usual for each company to send a negotiation team composed of different experts coming from different organizational departments. This team is entrusted with the task of un- derstanding the complex scenario at hand and obtaining the most in the final agreement. • Similarly to the scenario mentioned above, negotiations in politics also involve negotiation teams. We could for instance think of the negotiations carried out between Cambodia, Laos, Thailand and Vietnam for promot- ing cooperation on water resources [15]. In these negotiations, each na- tional party formed a negotiation team. Each team was formed of different specialists whose common goal was the quality of the final agreement for its country. • Imagine that a married couple wants to purchase a car [8]. For that matter, the couple has to negotiate with a car seller the purchasing conditions like price, payment method, and extras. Clearly, this is an agreement that is signed between two parties: the couple, and the car seller. However, one of the parties is clearly composed of two individuals (i.e., the couple is composed of the husband and the wife) that share the same goal (i.e., buy a car). • Suppose that a group of four friends decides to travel together. If a travel needs to be arranged, the group of friends needs to find an adequate destination, some nice accommodation, and flights that take the group to the city of their holidays. Additionally, it may even be interesting to include some pre-arranged social activities like visits to museums, visits to monuments, some sport activities, and so forth. There may be several 3 travel agencies that offer such services, and the group of friends may need to negotiate with some of them to get a travel package that satisfies their needs. As in the previous scenarios, the group of friends is one single negotiation party that is composed of multiple individuals that share a common objective (i.e., go on a trip together). Thus, it can be appreciated that negotiation teams are quite common in real life negotiations. Despite their importance in real negotiations, teams have not been studied by social sciences to the same extent as dyadic negotiations [16, 17, 18]. However, what are the reasons to send a negotiation team to the negotiation table instead of a single negotiator? The first reason that may come to our minds is that the more individuals the better task performance in negotiation. Thompson et al. [7] showed in several experiments involving human negotiation teams that as long as one of the parties is a negotiation team, better joint outcomes (i.e., integrative outcomes) were obtained. This is partially explained due to the fact that when teams are present at the negotiation table, parties are more inclined to exchange information [7]. However, as pointed out by Mannix [19], greater numbers may not pay off unless coordination is present among team members. Another reason to employ a negotiation team is skill distribution and infor- mation distribution [8, 19, 20]. With this, we mean that different team members may have different and complementary profiles needed to tackle properly the ne- gotiation problem. Thompson [20] recommends that managers should recruit negotiation teams composed of experts in negotiation, experts in the subject to be negotiated, and individuals with a variety of interpersonal skills. Mannix [19] states that negotiation teams require a diverse set of knowledge, abilities, or expertise in complex negotiations, and the correct assessment of such skills is pointed out as one of the keys for success in a negotiation. Skill distribution and complementary skills are of vital importance when using some classic team negotiation tactics like the good cop/bad cop persuasion tactic [21]. Finally, other authors consider that another reason to send a negotiation team to the negotiation table is the party being formed by different stakeholders [19, 22]. For instance, Mannix [19] points out union negotiation as an example of negotiations where parties are formed by different interests that have to be represented in the negotiation table. Halevy [22] also remarks the importance that despite negotiation teams being a single negotiation party, they are hardly ever a unitary player. In fact, a negotiation team is usually a multi-player party with different and possibly conflicting interests. 2.2. Agent-Based Negotiation Teams Up to this point, we have strictly considered human negotiation teams. How- ever, are negotiation teams feasible and needed for automated negotiation and electronic applications? We argue that the answer to such question is positive. Agent-based systems are not alien to negotiation scenarios where it may be in- teresting to employ negotiation teams. For instance, imagine a tourism e-market 4 application. It is usual for groups of friends/families, or even strangers, to or- ganize their holidays as a group. Nevertheless, travelers usually have different preferences regarding trip conditions (e.g., cities to visit, hotel location, leisure activities, number of days to spend, budget limitations, etc.). Humans may be slow at coming with a proper negotiated deal that accounts for everyone's preferences. This task may not be only slow, but also tedious since conflict may be present. On top of that, human negotiators have been documented to be content even with deals that are far from the optimal solution. This phe- nomenon is what is known in the literature as leaving money on the table [20]. As a way for overcoming these problems, agents representing each traveler could form a team in an electronic market to quickly obtain a good trip package for the group. Software agents should be quicker than humans as long as they can handle the negotiation domain and they should prevent human negotiators from carrying out tedious tasks. The application of negotiation teams is not limited to the aforementioned example. It can be extrapolated to other domains such as electronic farming cooperatives, customer coalitions, organizational merging be- tween virtual organizations, negotiation support systems for labor negotiations, and so forth. A trusted mediator who can perfectly aggregate preferences can be thought of as a possible mechanism to coordinate a negotiation team in the tourism e- market example. Nevertheless, there are several reasons that preclude us from aligning ourselves with this kind of coordination mechanisms. The first reason is that privacy is usually a concern among users in electronic applications. In fact approximately 90% of the users in electronic applications care to some degree about the amount of information that they filtrate in electronic applications, and only 10% do not care about letting others manage their information [23]. Hence, one cannot expect that every team member may be willing to share its full preferences to a mediator. The other important issue is the fact that even though there may be some degree of cooperation among team members, one should not forget that the team is a multi-player party and opportunistic behavior may be present. In that case, preference aggregation is a dangerous mechanism since it may be quite prone to being manipulated and exaggerated for one's own benefits. Therefore, new negotiation models are needed to coordinate agent-based negotiation teams. Making an analogy with the human counterpart, we can define an agent- based negotiation team as a group of two or more interdependent agents that join as a single negotiation party because of their shared interests in the nego- tiation. The reasons to use an agent-based negotiation team are also analogous to the human case. First, more agents in the team may mean more computa- tion capabilities and, thus, more extensive and parallelized exploration of the negotiation space. Second, we can also assume that different and heterogeneous agents may have different experiences, they may offer different services/skills, and they may implement different algorithms, which in the end may result in the team being able to tackle complex negotiation problems more efficiently. Third and lastly, the team may really represent a multi-player party whose preferences need to be satisfied as much as possible by the final agreement. 5 Another related topic is one-to-many and many-to-many negotiations [24, 25, 26, 27, 28, 29, 30]. One-to-many and many-to-many negotiations are concepts that address the number of parties involved in the negotiation. In one-to-many negotiations, one party negotiates simultaneously, in with multiple parties. It can be a party negotiating in parallel negotiation threads for the same good with different opponent parties [26, 27, 29, 30] or a party that negotiates simultane- ously with multiple parties like in the Contract-Net protocol, and the English and Dutch auction [24, 25, 28]. Many-to-many negotiations consider the fact that many parties negotiate with many parties, the double auction being the most representative example [28]. Differently to the aforementioned concepts, negotiation teams are not related with the cardinality of the parties but the nature of the party itself. When addressing a negotiation team, we consider a negotiation party that is formed by more than a multiple individuals whose preferences have to be represented in the final agreement. This complex negoti- ation party can participate in bilateral negotiations, one-to-many negotiations, or many-to-many negotiations. The reason to model this complex negotiation party instead of as multiple individual parties is the potential for cooperation. Despite having possibly different individual preferences, a negotiation team usu- ally exists because there is a shared common goal among team members which is of particular importance (e.g., going on a trip together in the case of the group of travelers, selling the product line in the case of the human organization, purchasing a car in the case of the married couple, and obtaining a beneficial agreement for their country/region in the case of politics). Nevertheless, human negotiation teams do not always guarantee a better outcome than individuals. The performance of the team is directly related to coordination among team members. A team that is not capable of achiev- ing such coordination may fail at the negotiation. In fact, Behfar et al. [18] study the causes that pose problems for human negotiation teams: logistics and communication problems (e.g., communications inefficiencies), substantive dif- ferences (e.g., confusion about goals, conflicting interests), inter-personal and personality differences (e.g., different negotiation styles), and confusion about team roles (e.g., unclear decision rights). The same authors also identify those strategies that help to overcome the aforementioned problems and lead teams towards success: time and logistics management (e.g., coordinating strategies during negotiation by stepping away from the table), team communications (e.g., preparing with teammates), within-team negotiations (e.g., team problem solving, managing conflicting interests), and defining leadership and team roles (e.g., defining decision rights). To put it briefly, communications, coordination, intra-team negotiation, and clear rules of the game lead human negotiation teams to success. We believe that those key elements are also important in agent-based negotiation teams. For that reason, we identify and describe the communications, coordination and negotiation tasks that may help agent-based negotiation teams to perform successfully. Solutions for each task are reviewed, and special challenges that may arise are also pointed out. This present work can be categorized as a review paper from the point of view of those areas that may be relevant to agent-based negotiation teams. In 6 the automated negotiation literature, one can find different descriptive studies and reviews that have appeared in the last few decades. Kraus [5] presents a re- view of negotiation and cooperation in multi-agent systems by analyzing works in game-theory, large scale mechanics, operation research, and heuristics. Beam et al. [31] presented another point of view of the research in automated negotia- tion. Their study focused on challenges for electronic commerce and, therefore, the article contained state-of-the-art sections concerning negotiation support systems, automated intelligent agents in negotiation, auction mechanisms, and electronic marketplaces. Later, Guttman et al. [32] overview the use of negoti- ating agents and agent technologies in electronic marketplaces while putting a special emphasis on some of the most important tasks needed in electronic mar- ketplaces: product brokering, merchant brokering, and negotiation. Similarly to Kraus [5], Jennings et al. [6] offer a review of some of the most important works in negotiation, advances, and challenges in game-theoretic models, heuris- tic models, and argumentation frameworks. Another review regarding agents in electronic commerce is presented by Lomuscio et al. [33]. The authors present a taxonomy to classify negotiation in electronic commerce attending to crite- ria like the cardinality of the negotiation (i.e., number of parties involved), the characteristics of the agents (e.g., role, rationality, knowledge, bidding strategy, etc.), the characteristics of the environment, the type of information exchanged, and event parameters. The taxonomy is applied to several well-known negotia- tion frameworks in the literature. He et al. [34] review agent-based approaches to electronic commerce from the point of view of business-to-consumer (B2C) and business-to-business (B2B) markets. In the former case, the authors review different approaches in one of the following tasks: need identification, product brokering, buyer coalition formation, merchant brokering, and negotiation. In the latter case, different approaches are reviewed in tasks such as partnership formation, brokering, and negotiation. In [35], Lopes et al. propose the nec- essary tasks for the complete implantation of a general negotiation framework. Later, Luo et al. [36] propose a software methodology for the design of negoti- ation models in different scenarios. In the article, they review and standardize different high impact negotiation models. All of the aforementioned works have reviewed different advances and particularities of the research in automated negotiation. However, none of them has focused on the specific problem of negotiation teams and the special challenges associated to the topic. 3. Tasks for Negotiation Teams In this section, we describe the different tasks that an agent-based negotia- tion team may need to carry out in different scenarios. For each task, a review of related literature is carried out and the special challenges that are present in the negotiation team setting are highlighted. Some authors [35] have suggested the following tasks for negotiating agents: Identify social conflict, identify negotiation parties, structuring personal infor- mation, analysis of the opponents, define a protocol and select a negotiation strategy, negotiation (i.e., exchange of offers, argumentation, learning, etc.), 7 and re-negotiation. We argue that the tasks proposed by other authors are the correct steps to be followed by negotiating parties. However, some additional tasks, mainly those involving team coordination, are introduced to cope with the team scenario. It should be noted that depending on the application domain, some of the tasks described may not be necessary. The tasks have been organized in a workflow as it can be observed in Fig- ure 1. In the schema, we distinguish between tasks that are carried out with opponents (task with opp.), tasks that mainly concern interactions with team members (team task), tasks that only involve one individual agent (individual task), and tasks that involve team members, opponents, and the individual. The workflow is therefore divided into Identify Problem (individual task), Team For- mation (team task), Opponent Selection (team task), Understand Negotiation Domain (team task), Agree Negotiation Issues (Task with Opponents), Plan Ne- gotiation Protocol (team task), Agree External Negotiation Protocol (task with opponents), Decide Intra-team Strategy (team task), Select Individual Strategy (individual task), and Negotiation & Adaptation (team, individual and oppo- nents). The workflow aims to be as general as possible so that it could be adapted to a wide variety of domains. Hence, some of the tasks proposed may not be necessary and they can be skipped in some of the domains (e.g., select- ing an opponent when there is one single opponent and it is known from the domain). Next, we describe all of the tasks in detail. 3.1. Identify Problem The first step consists of identifying a situation that requires negotiation. The agent has to look in its environment for partners, and, in the process, determine whether a conflict exists or not. As stated by Lopes et al. [35], most artificial intelligence researchers have focused on how to reach an agreement, but very few have studied the problem of detecting conflict. The final product of this task is a list of potential team members, a list of potential opponents, and, possibly, a list of potential competitors. Next, we review different solutions in the literature to this problem. 3.1.1. Domain Specific Conflict Detection In multi-agent literature, one can identify works where conflict detection mechanisms are designed for specific domains like cooperative planning or air traffic management [37, 38, 39, 40]. In [37, 40], virtual cylinders delimit the protected and alert zone of each aircraft. Conflict is detected whenever the cylinders of two different aircrafts overlap. Chu et al. [38] propose a conflict detection mechanism for collaborative planning between pairs of agents. The agents work towards a shared plan by exchanging plan proposals and beliefs. One agent is able to detect whether or not a conflict exists by checking proposals and beliefs with respect to its own belief base. Similarly, Barber et al. [39] pro- pose a conflict detection mechanism when agents integrate their individual plans into a global plan. More specifically, conflict is detected at the goal level (i.e., incompatible goals) or planning level (i.e., threats between actions) by means of 8 Figure 1: A general workflow of tasks for agent-based negotiation teams. the use of PERT diagrams. In other applications like electronic commerce, con- flict/negotiation comes inherently linked to the role of the agent: buyers want to buy at lower prices, while sellers want to maximize their profit. However, these mechanisms are specific and cannot be adapted to other potential domains. 3.1.2. Domain Independent Conflict Detection Research into domain independent mechanisms that allow conflict detection is a topic that needs further research. This research is especially important if one attempts to design general negotiators for different domains. Some re- searchers like Lopes et al. [41] have employed libraries of axioms (i.e., libraries of rules) that allow agents to compare their own plans and intentions with those expected plans and intentions of other agents. A conflict is detected when one's own intentions are not compatible with predicted intentions of other agents. Similarly, it may be possible to express axioms for detecting potential partners. The main disadvantage of this approach is that libraries of axioms are static, usually provided by experts, and they cannot learn from new situations. 9 INDIVIDUAL TASKTEAM TASKTASK WITH OPP.IDENTIFY PROBLEMTEAM FORMATIONOPPONENT SELECTIONUNDERSTAND NEGOTIATION DOMAINAGREE NEGOTIATION ISSUESPLAN NEGOTIATION PROTOCOLAGREE NEGOTIATION PROTOCOLDECIDE INTRA-TEAM STRATEGYSELECT INDIVIDUAL STRATEGYNEGOTIATION & ADAPTATION3.1.3. Looking for Partners and Opponents Unless the multi-agent system is a closed one, it is necessary to look for potential partners and opponents. The search process may be a necessary step for the aforementioned conflict detection mechanisms. One interesting technol- ogy is searching/communicating in social networks and markets [42, 43]. Sarne et al. [42] present an efficient algorithm for cooperative search in electronic markets. An agent or a group of agents look for goods in the marketplace to fulfill different goals. The algorithm takes into account purchase opportunities (i.e., the probability of finding a certain good) and the cost associated with finding new opportunities. For that matter, the agents employ distributions over opportunities in the electronic market. The search is capable of being partitioned in autonomous searches that aim to cover the different goals of the agent/group. While the search algorithm is interesting due to the fact that it considers search costs, obtaining distributions over partners/opponents for a specific task/negotiation may be difficult, especially if the negotiation task is new for the agent. Del Val et al. [43] present a decentralized service discovery system in social networks. The algorithm for service discovery is not based on previous information or statistics that require training, but similarity between agents. The similarity measure is calculated considering the semantic descrip- tions of the agents. Similarity measures may be employed to look for negotiation partners and opponents. However, it requires that agents make public certain semantic information regarding their profiles. Special Challenges Domain independent conflict resolution is a topic that needs further research. Mechanisms proposed in the literature have the disadvantage of being generated according to the knowledge of experts. Thus, they may not be robust to unex- pected scenarios unless learning mechanisms are used. One way to address this issue may be case-based reasoning [44]. Instead of using axioms, the agent may have a database of past conflict situations. If the current situation is similar enough to a past conflict situation, it can be considered that conflict may exist in the present. Once unexpected scenarios arise, it is possible for the agent to store the current situation as a new case. In future events, the new information may be used to predict new similar scenarios. The same mechanism may be used for detecting potential partners. 3.2. Team Formation Once the agent has studied which agents may be considered as potential partners, and which agents may be considered as opponents, the agent faces the challenge of determining whether or not benefits arise from forming a negotiation team. In some situations, it may be mandatory for the agent to be part of a negotiation team. In fact, the team may even be static (i.e., a married couple negotiating with a seller over an apartment). If that is the case, identifying negotiation partners, and forming a negotiation team are tasks that can be skipped. Nevertheless, some scenarios may be less rigid and the agent may 10 be able to form a negotiation team from the list of potential partners. Thus, the agent should analyze which team he expects to be the optimal negotiation team according to the list of potential partners. If it is expected that no team reports more benefits for the agent than negotiating individually, the agent should decide to negotiate as a single individual party. Next, we review those mechanisms that have been classically used to distribute agents into groups. 3.2.1. Coalition Formation Allocating agents into optimal groups has been a field of study for coali- tion formation [45, 46, 47, 48, 49]. A coalition is a short-lived group of agents that joins together for the accomplishment of a certain goal. The problem of coalition formation is usually divided into three main problems [50]: coalitional value calculation, coalition structure generation, and payoff distribution. The first problem, coalitional value calculation, consists of computing every possible coalition of agents and its associated value. Once every possible coalitional value is computed, the next task consists of dividing agents into disjoint coalitions that maximize the social welfare of the system. Finally, the payoff generated by the coalition is divided among the agents that are part of the group. In a team negotiation process, payoffs may be difficult to anticipate since the outcome of a negotiation is uncertain. On top of that, the result of the negotiation may be an object whose payoff may not be divided among team members. For instance, in the case of the traveling friends, the final result of the negotiation is the travel. Even though the cost of the travel may be divided among team members, how do you expect to divide the benefits of other factors of the negotiation such as the payment method, arranged foods, hotel location, and so forth? Thus, those works that focus on optimally dividing payoffs [46, 47, 48] may not be so interesting from the point of view of agent- based negotiation teams. On the other hand, efficiently calculating coalitional values and coalition structures [45, 51, 49] may come handy in negotiation team formation. This is especially true if the formation task is supported by some kind of centralized au- thority (i.e., a mediator, the electronic commerce site, etc.). Sandhlom et al. [45] present a coalition formation algorithm for problems where computing the cost of the coalitional value is exponential and agents have to pay for resources asso- ciated to the computation. Thus, as less computational resources are dedicated to the calculation of the coalitional value, less optimal solutions are obtained. The setting is similar in negotiation team formation. The exact value of the coalitional value cannot be calculated since it is uncertain and the actual value depends on the result of the negotiation. Therefore, approximated methods or heuristics should be used to approximately assess the quality of a coalition. In [11], results suggest that team performance is usually improved as team mem- bers' preferences are more similar. Thus, methods like similarity heuristics may be useful to calculate coalitional values when negotiation teams are concerned. Rahwan et al. [51] present a distributed anytime algorithm for computing coali- tional structures. The main advantage of this approach comes from the fact that despite being an anytime algorithm, it can guarantee a bounded solution with 11 respect to optimality. This feature makes it applicable for negotiation teams in marketplaces, where the number of potential partners may be too large for non-anytime methods. Another relevant work for agent-based negotiation teams is constrained coalition formation [49]. In this framework, not every coalition is possible due to constraints imposed by the application domain. Therefore, one should avoid generating and using coalitions which are not allowed. For instance, in a group travel electronic market, some agents may not be willing to travel with other agents (e.g., due to past experiences). Coalition formation algorithms are usually (not necessarily always) employed in a semi-centralized way, even when calculations can be distributed. These algorithms are usually applied by a centralized authority that manages all of the coalitional needs of the agents in the system. The authority distributes all the agents in coalitions with the aim of maximizing the social welfare of the system, which may be different to the social welfare of individual agents. Thus, this kind of mechanism is usually employed in settings where such a central authority exists (e.g., an electronic marketplace). 3.2.2. Buyer Coalitions Another trend of research in coalition formation are buyer coalitions [52, 53, 54, 55, 56]: groups of buyers that join together in order to take advantage from volume discounts. In [52], the authors show the incentive for buyers to form coalitions that take advantage of volume discounts. Additionally, they discuss the advantages and disadvantages of different protocols for the formation of customer coalitions (i.e., pre-negotiation and post-negotiation). Yamamoto et al. [53] present an algorithm for customer coalition formation in marketplaces where agents are able to make bids on different items. A me- diator gathers the bids (i.e., price) and groups together buyers that prefer the same item. Then, the surplus is divided fairly inside the coalition so that the coalition is stable (i.e., agents do not have incentive to leave the coalition and join another one). Later, Li et al. [55] study the problem of forming customer coalitions and distributing the surplus. They study the problem from the point of view of incentive compatibility (i.e., whether buyers have incentive to express their real reservation prices) and coalition stability. Several mechanisms are proposed for the problem, and experiments show a positive correlation between incentive compatibility and coalition stability. Ito et al. [54] present customer coalition models for agents that participate in a marketplace where buyers with the same needs are grouped together and they express their purchase desires in terms of reservation prices. On the other hand, sellers have deadlines to sell their products and they can cooperate with other sellers in case that stock requirements are not met. When a seller's deadline is reached, the seller negotiates with the coalition and buyers can decide whether or not to purchase the product. In [56], the authors propose a customer coalition formation mechanism for marketplaces where agents can buy bundles of items. An empty group is vir- tually formed for each seller. Each buyer agent searches for the optimal bundle that it can get by itself. The agent can register in those groups that are part of 12 its optimal bundle. Then, buyer agents negotiate among them to form coalitions that take the most from the purchase options. In general, the relationship between buyer coalitions and negotiation teams is strong: both are groups that engage in some kind of negotiation as a single party. The main disadvantage of current buyer coalition approaches is the fact that most works in group buying have focused on single attribute transactions where only price is involved or coalitions of buyers where bundles of items are involved. In most of these approaches, payoffs can be split and shared, and they do not consider qualitative attributes like cities, types of accommodation, and so forth. Therefore, complex multi-attribute negotiations are not supported by current group buying approaches. Additionally, it should be noted that even though every buyer coalition may be considered a type of negotiation team at some point, not every negotiation team is a buyer coalition. For instance, let us imagine the negotiation between a union and the manager of an enterprise. The union may send a negotiation team formed by different experts or different stakeholders (i.e., representatives for different types of workers) to negotiate with managers. In this case, the goal is not obtaining volume discounts as group buyers' case. Moreover, buyer coalitions may be highly dynamic formations that may change when better coalitional options arise. That is not the case of the group of traveling friends or the union, where once the team has been formed; it usually remains static during the negotiation process. 3.2.3. Cooperative Team Formation Another field relevant to this task is cooperative team formation [57, 58, 59, 60]. When teams are formed, agents with different skills, necessary for achiev- ing the common and shared goal, are sought. Team members with different skills/expertise may be desirable for complex negotiation domains. Neverthe- less, cooperative team literature in multi-agent systems usually assumes that team members share the same goal and all of them work towards the achieve- ment of such common goal. For that purpose, team members share an overall plan that they follow together, share knowledge of the environment, share in- tentions, and share skills needed to achieve the common goal [61]. Cohen et al. [57] present a framework where teams are groups of agents with a shared mental state that are committed to accomplish a joint goal. Speech acts are used as a mechanism to form and disband teams. If an agent requests another agent to perform a certain action, and it is followed by an accept mes- sage, both agents are considered to follow a joint goal and they cooperate for the realization of such goal. Dignum et al. [58] present a team formation framework based on persuasive dialogues and speech acts. An initiator agent forms an abstract plan. In the first stage, the agent gathers information on agents that may contribute to the plan attending to their abilities. Then, the initiator agent engages in persuasive dialogues with a subset of agents willing to work together towards the joint goal. The objective of the persuasion dialogue is to ensure that other agents are committed to the joint goal. Forming teams according to agents' skills should be a useful approach in situations where negotiation teams need different agent 13 profiles to tackle the negotiation domain (e.g., experts in different areas, different services, etc.). Persuasion is also useful, especially in open environments where agents do not necessarily hold the same beliefs, goals, and intentions. In [59], a framework for dynamic team formation in a network of agents is presented. Each agent has a certain skill, and the agents are connected with a limited number of agents, forming a network. In this scenario, teams have to be formed dynamically in order to carry out tasks that are needed for the global goal of the network. Those agents that are not active can start the team formation process and look for teammates. However, an agent can only be part of a team if it is connected with at least one agent of the team. Networks are realistic structures in open systems, where agents do not know every single agent in the system, but a set of agents with whom it usually interacts. Thus, mechanisms such as the one presented in [59] offer realistic group formation principles for negotiation teams in open multi-agent systems. Badica et al. [60] present a framework for the formation of dynamic work- flows of experts and services in environmental management applications. The organization of experts into workflows is carried out by multi-issue negotiations that take into account the fact that experts may have previous commitments that limit their availability for team tasks. Negotiations are cooperative in the sense that the goal of all of the agents is a successful management in an envi- ronmental incident. In most of the aforementioned frameworks, team members are usually fully cooperative with each other. Once agents commit to a joint goal, they cooper- ate and coordinate towards the consecution of that goal. Team members in a negotiation team may have different sub-goals or preferences despite sharing a common goal. Hence, team members may not be fully cooperative with other fellow team members, since other individual sub-goals may be at stake during the negotiation (e.g., trip conditions in the traveling friends' scenario). It should be highlighted that team formation algorithms are usually (not necessarily always) initiated by a single agent that asks for help to other agents on a specific goal/task. Hence, these mechanisms may be more appropriate in applications where a centralized authority does not exist and agents are more willing to fully cooperate. Special Challenges The ideal formula for negotiation team formation may depend on the type of application that needs to be constructed. In any case, we argue that the afore- mentioned approaches may need to consider the following issues when forming a negotiation team: • Electronic commerce has given more social power to consumers, which now can find new sellers at a relatively low cost [62] and access informa- tion more easily. Not only that, but reputation models [63] and gossiping [64, 65] (i.e., exchanging information among known agents) may give an additional coercive power to consumers over sellers. In both cases, an opin- ion is formed about sellers which is based on the experiences/knowledge of 14 the different agents in the society. This may produce sellers that are more willing to act cooperatively. Otherwise, sellers may be negatively rated by buyers. Thus, it is expected that, the larger the negotiation team, the greater social power it will be able to exert, and the more cooperative the seller will be. • Even though it seems intuitive that the larger a negotiation team is, the better, this may not be always true. If the preferences of the team mem- bers are compatible and very similar, adding new team members to a negotiation team may only result in greater social power. As stated, if the preferences of the team members are very dissimilar, adding new team members may result in lower team performance [11]. Thus, despite the fact that larger teams may be able to bring together more social power, intra-team conflict may deteriorate the quality of the final agreement to a point that greater social power does not compensate. Generally, it is not possible to exactly know the preferences of potential team members prior to the negotiation itself. Nevertheless, past experiences with negotiation partners [66] and recommender techniques like collaborative filtering [67] may help to accomplish the task of assessing which negotiation partners are more similar. In the former case, the agents use the history of past negotiations (negotiation outcome, and one's own utility function) with another agent to build a possibility distribution over the space of negotia- tion outcomes. Then, the expected utility of the negotiation with the agent is calculated and the agent negotiates with those agents whose expected utilities are higher. The same mechanism can be employed by one agent to determine which prospective teammates are the most similar. In the latter case, collaborative filtering consists of recommendation techniques for very large data sets. The rationale behind this technique is that if two agents have the same opinion on an issue, they are more likely to have a more similar opinion in other issues than randomly chosen agents. Thus, based on a history of past purchases/deals, it may be possible to deter- mine whether or not two agents are similar. However, this information is usually provided/used by an authorized third party (e.g., an electronic marketplace). • Related to the previous issue, more team members or stakeholders may bring additional negotiation attributes to the negotiation table. The first effect over the negotiation is that the negotiation domain becomes larger and possibly computationally harder to work with. Despite this computa- tional disadvantage, it may introduce negotiation attributes that are only interesting to a sub-group of the agents that participate in the negotia- tion. If the opponent is not interested in these attributes, it may make trade-offs easier. In contrast, if a negotiation attribute is introduced to satisfy a sub-group of the team members and the issue is important for the opponent, it may result in higher preference conflict with the opponent (i.e., a new issue which is important for the opponent and whose team preferences on issue values are opposite to those of the opponent). Hence, 15 it may be more difficult to find an agreement. In general, additional is- sues in the negotiation may be a double-edged sword that can report both benefits and disadvantages. • One of the problems that negotiation teams may face is tackling negoti- ation domains that are inherently complex. This means that the nature of the domain is hard to understand and it requires the expertise of dif- ferent agents. For instance, when an organization negotiates in a complex negotiation, it sends a negotiation team composed of different experts. These experts may come from the different departments of the organiza- tion (e.g., marketing, human resources, research & development, etc.) and have different backgrounds that enrich the understanding of the problem. Information regarding agent roles and/or agent identities [68] may come handy to determine which potential team members are more fit for the negotiation problem. • Different agents may provide different social relationships to the team's social network. Social networks may directly impact upon the performance of teams [69] since it can provide with extra information for the team. They may provide new information and team members that are beneficial for the team. 3.3. Opponent Selection Once the team has been formed, it is necessary to find suitable opponents from the list of prospective opponents. The team should decide which opponents they are going to face. If enough computational resources are available, all of the opponents can be selected and negotiations can be carried out in parallel. However, if computational resources are scarce, a subset of the opponents has to be selected. 3.3.1. Opponent Evaluation Mechanisms As introduced earlier, each individual agent can calculate the expected util- ity of a negotiation with an opponent based on past outcomes and situations [66]. Another classic method to evaluate opponents is trust and reputation [63]. Trust is built as an evaluation mechanism based on one's own past transac- tions/interactions, while reputation is built based on the aggregated opinions of agents in the society. The rationale behind trust and reputation is that op- ponents that respect trade conditions (e.g., dispatch dates, payments, etc.) are positively evaluated, whereas agents that violate trade conditions (e.g., delays) or misbehave are evaluated negatively. Generally, agents should select oppo- nents that are evaluated positively by trust and reputation mechanisms. 3.3.2. Discussion in Groups The expected utility of negotiating with an opponent and trust evaluations are calculated from the point of view of one agent. Since the team is composed by more than a single agent, the opinions of different agents may need to be 16 aggregated. In fact, there may be intra-team conflict since teammates' opinions on opponents may vary due to different experiences and preferences. In case of conflict, agents need to resort to other mechanisms for aggregating the opinions of teammates. One of these classic mechanisms is voting [70]. Classic voting schemes like majority voting and Borda count have the advantage of aggregating opinions with a few interactions. However, they may not be expressive enough to determine the best option for the negotiation team. For instance, if one of the teammates has bad expe- riences with certain vendor but other teammates have good and more recent experiences, the team may determine to discard the opinion of the former team member. Deliberation and persuasion in groups can discover and clear these differences [71, 72]. Deliberation is a discussion to decide a course of action among a group of agents. McBurney et al. [71] introduce the first formalization for deliberation dialogues in multi-agent systems. The interaction protocol for agents usually starts when agents have to decide on a course of action. Then, agents inform about the goals of the dialogue and the criteria that are relevant to the goals of the deliberation. After that, agents in the deliberation can enter a cycle where they are allowed to propose actions to be taken by the group, comment on actions proposed in the dialogue from different perspectives, revise the deliberation goals, recommend on actions that should be taken by the group, and confirm the acceptance of recommended actions. Heras et al. [72] propose a persuasion framework for groups of agents that have to make a decision. In the framework, agents argue using case-based reasoning with two purposes. The first purpose consists of learning new solutions to domain problems, which are later used in dialogues to propose course of actions. The second purpose is learning how to argue from previous dialogues. Special Challenges Despite the fact that mechanisms exist that allow agents to evaluate oppo- nents and discuss as a group on a course of action, as far as we are concerned it has not been applied to the evaluation of negotiation opponents. Comparing voting schemes and persuasion/deliberation, computation time may have a key on deciding on which technology should be chosen. If time is limited, voting schemes may be more appropriate when comparing a large number of negoti- ation opponents since they may require fewer messages. On the other part, if time is not of extreme importance in the process, deliberation and argumenta- tion are more expressive and allow discovering scenarios like the one commented above. 3.4. Understand the Negotiation Domain Understanding together the negotiation domain is a task of extreme im- portance. Not only does it allow team members to get a grasp of other team members' preferences, but it makes it possible for team members to tackle cor- rectly the negotiation when the domain is complex and requires expertise in different knowledge areas. 17 Even assuming that team members have similar backgrounds, it is still im- portant to understand the negotiation domain together. Let us imagine that a group of friends (e.g., Alice, Bob and Charlie) decides to go on a travel together and have fun. What is the meaning of "having fun"? Clearly, it may be different meaning for each friend: Alice thinks that if the city offers lots of adventure sports then the travel is fun, Bob thinks that having fun also involves finding a place with considerable night life, whereas Charlie is happy with any plan as long as it does not involve much money. From this situation, it can be inferred that the price, adventure activities included in the travel package, and the night life activities are relevant negotiation issues for the team in the negotiation at hand. 3.4.1. Discussion in groups The technologies that can give support to these processes are varied. As stated, persuasion and deliberation in groups can help to manage conflict when deciding on a course of action [71, 72]. The object of discussion in this phase is those issues that are relevant for the team during the negotiation. Differently to the deliberation regarding opponent selection, in this case the available options and possible negotiation issues are not fully known beforehand. Thus, they have to be discovered during the argumentation process. 3.4.2. Belief Merging Other technologies like belief merging [73, 74, 75] may also prove useful for obtaining a shared model of the negotiation domain. Belief merging consists of joining together the knowledge that comes from different sources. If no incon- sistency is present among the different data sources, information can be easily aggregated. However, if some piece of information is inconsistent among differ- ent data sources, special mechanisms have to be employed. Booth et al. [73] present a belief merging algorithm for information com- ing from multiple data sources. It is assumed that one of the data sources is completely reliable (e.g., trivial information) with respect to the others. Con- sistent information is added as new knowledge, while inconsistent information undergoes a contraction process where it is incrementally reduced to pieces of consistent information by means of negotiation. For the application of this so- lution in a negotiation team, the sources of information can be considered the different agents in the team. The use of negotiation results convenient since in a negotiation team, each data source corresponds to a different party. The result of belief merging is information which can be used as shared mental model for the team. Qi et al. [74] present another framework that merges different sources of information by means of negotiation among inconsistent sources. The main difference between this framework and the previous one is that the authors consider the fact that knowledge bases may have different priorities. These priorities represent different degrees of confidence in the different sources of information. This mechanism can represent the fact that in a negotiation team, the information of some team members may be more reliable in some areas. 18 However, it should be noted that most belief merging methods are not strategy-proof [75, 76]. A belief merging method is strategy-proof when it is robust against attempts of manipulation by agents. An agent may try to ma- nipulate the belief merging process if it expects to increase its utility by acting that way. In [75], the authors study the restricted conditions under which several families of belief merging operators are strategy-proof. Special Challenges Identifying negotiation issues that are relevant for the team is not the only task necessary to completely understand the negotiation domain. Identifying which attributes are predictable and compatible for the team, and which at- tributes are not predictable is also crucial. On the one hand, a negotiation issue is predictable and compatible among team members if the preferences of all of the team members over issue values are known and compatible. For instance, in a team of buyers, it is logical that all team members prefer low prices over high prices. In this type of negotiation issues, there is full potential for cooper- ation among team members since increasing the utility for a team member (i.e., decreasing the price) results in other team members staying at the same utility level or increasing their utility. On the other hand, a negotiation issue is not predictable among team members if nothing can be inferred about which issue values are preferred by team members. The issue may be compatible among team members (i.e., same ranking of preferences over issue values) or not, but it is not possible to know the nature of the negotiation issue unless team members are willing to share information. Using information regarding which issues are predictable and unpredictable among team members may be useful for deciding on which negotiation strategy is used among team members. For instance, in [12] the authors present an agent-based negotiation team model that, as long as the negotiation domain is exclusively composed by compatible and predictable issues, it is capable of guaranteeing unanimity regarding team decisions at each round of the negotiation. Another interesting issue is how team members analyze if they have incentive to share all of the information regarding the negotiation domain. An agent may be willing to share a piece information only if it expects that it is going to report higher utility than hiding the piece of information. Such et al. [77] propose disclosure mechanisms based on human psychology that allow agents to determine when sharing private information results in greater utility. However, research in this topic is still in early stages and it has not been applied to negotiation teams. 3.5. Agree Negotiation Issues Since the previous stage produced a list of issues which are relevant for the team members, the next stage consists of agreeing a final list of negotiation issues with the opponent. The opponent may have its own list of issues relevant to the negotiation. Thus, a final list of issues to be negotiated should be agreed between both parties. 19 From the initial set of negotiation issues proposed by the team, some of the issues may not be negotiable since the opponent does not offer that service. For example, if the team members had originally concluded that negotiating packages of adventure activities is a relevant issue to the team but a travel agency does not work with such packages, the issue cannot be included in the negotiation. Additionally, some negotiation issues that were not included in the list proposed by the team may be included in the final list since they are relevant to the opponent. As for those negotiation issues present in the lists proposed by team members and the opponent, it may also be necessary to agree on the issue domain (i.e., the values that the negotiation issue can take). Similarly to the agreement on the list of issues, the final domain value may not contain all of the values proposed by both parties (i.e., Rome cannot be a value for the city of destination if the travel agency does not offer flights to Rome). Special Challenges Despite being an important process in the pre-negotiation, very little atten- tion has been paid to agreeing negotiation issues between parties. In fact, most researchers in negotiation assume that the list of negotiation issues and their domains are already agreed in their negotiation models. Faratin [78] mentions in his thesis the possibility of adding and removing non-core issues during the negotiation. While core negotiation issues remain static during the negotiation process, involved parties may be able to add or remove non-core negotiation issues as the negotiation process advances. However, the list of non-core is- sues is assumed to be known by both parties and the development of an issue- manipulation algorithm was appointed as future work. We acknowledge that this is a process that needs to be researched in the future. 3.6. Plan Negotiation Protocol After the list of issues for the negotiation process is set, the different parties have to agree a negotiation protocol unless the protocol is already set by a third party or by the system. In that case, this task and the next task can be skipped. There are different negotiation protocols that may be applied for a specific situation. For instance, if the negotiation team engages with an opponent in a bilateral negotiation, both parties could employ the classical alternating offers protocol [79], extensions of such protocol like the k-alternating offers proto- col [80], protocols based on fuzzy tradeoffs [81], two-stage protocols based on optimization and concession [82], or complex mediated protocols like [83, 84]. The team as a whole may have different opinions and knowledge about the available protocols. In fact, some of the team members may not even know some of such protocols. In that case, those protocols cannot be used by the team since some of its players do not know the rules and decision making strategies to face such games. 20 3.6.1. Discussion in Groups In this phase, if more than an applicable protocol is known by all of the team members, they should decide as a group which protocols are preferred by the team (i.e., a ranking of the known protocols). Again, the team may resort to voting [70], and argumentation and deliberation [71, 72]. Special Challenges The main challenge in this task is how agents evaluate the different protocols available in the literature. As far as we are concerned, there is not an exhaustive study that analyzes which protocols are more convenient in different situations. Such a study would be necessary to provide agents with the mechanisms to make informed decisions on the negotiation protocol. 3.7. Agree Negotiation Protocol Considering that the team has already decided on which negotiation pro- tocols are preferred by team members (e.g., some ranking over the negotiation protocols), they should agree with the opponents on the negotiation protocols that are to be followed for interacting. Again, opponents may not know how to play some games, making some of the options not feasible. Some protocols known by the opponent may not be known by all of the team members. Over the list of protocols that are known by both parties, both parties may have differ- ent preferences and knowledge regarding the different protocols. This decision between both parties is going to involve some kind of simple negotiation (i.e., we do not expect the number of possibilities to be large) or discussion among both parties. In some cases, besides the negotiation protocol, some parameters of the protocol have to be decided also by both parties (i.e., who is the initiat- ing party in the alternating offers protocol [79], the number of offers allowed in the k-alternating offers protocol [80], who acts as trusted mediator in mediated protocols like [83, 84], etc.). Despite not being explicitly covered by any work, we consider that most negotiation models can be easily adapted to this task. 3.8. Decide Intra-team Strategy Negotiation protocols define the rules of interaction to be followed by the different parties. In a single player party, the decisions are individually taken by one agent. However, when the party is formed by multiple individuals, it is necessary to decide on how, when, and what decisions are taken, and who takes those decisions. This is what is termed as an intra-team strategy or negotiation team dynamics. For example, in the case of the alternating offers protocol [79], each party should decide on which offer is sent to the opponent party, whether or not to ac- cept the offer proposed by the other party, and when one should withdraw from the negotiation process. Thus, any intra-team strategy for teams participating in the alternating offers protocol should decide on those issues. Obviously, for different negotiation protocols, different decisions have to be taken and an intra-team strategy that has been proposed for a particular negoti- ation protocol may not be directly applied to other type of negotiation protocol. 21 Thus, usually negotiation protocols and intra-team strategies are tightly cou- pled. 3.8.1. Simulation Studies One of our hypotheses is that there is not a single intra-team strategy that is capable of outperforming the rest of intra-team strategies for every possible scenario. Depending on the goal of the team (e.g., social choice performance measure), and depending on the conditions of the negotiation some intra-team strategies may perform better than others. By conditions of the negotiation we refer to those factors, either external (e.g., the number of competitors, the number of prospective negotiation partners, etc.) or internal (e.g., team size, similarity among team members' preferences, deadline length, etc.) that can affect the negotiation. Usually, the number of variables is so large that it makes a theoretical analysis not feasible. Thus, some researchers have employed ex- tensive simulations to assess which strategies would work better in certain con- ditions [85, 86, 87, 88]. Oliver et al. [85], Matos et al. [86], Tu et al. [87], and Gerding et al. [88] propose evolutionary strategies in bilateral negotiations that converge towards reasonably good strategies in different environments. Never- theless, some specific conditions of the negotiation have to be known beforehand to evolve appropriate strategies. 3.8.2. Role Allocation It may be the case that some intra-team strategies additionally require role/task assignment (e.g., information retrieval, monitoring the market, etc.). Nair et al. [89] propose a framework that is capable of evaluating different role allocation and reallocation policies in cooperative teams by means of distributed Partially Observable Markov Decision Processes. This approach may be valid for intra-team strategies that require different roles. However, rewards steaming from role allocation may be uncertain in a negotiation team since agents are het- erogeneous and they may not know each other. Additionally, it is a cooperative approach that does not take into account the preferences of agents over tasks. Hoogendorn et al. [90] present a negotiation framework for the allocation of tasks between agents. An agent can place requests for task distribution among other agents. Then, agents bid for being assigned to the tasks of their interest. A negotiated approach is more convenient when agents are heterogeneous and they have different preferences with respect to roles. However, in this approach only bids are considered to assign tasks to agents. It does not take into account other factors like past experiences, trust and reputation, etc. Special Challenges Despite the fact that extensive simulations have been carried out in nego- tiations with single individual parties, no extensive study exists for the team case. The only exception is the work presented in [14, 11], where intra-team strategies are studied under varied negotiation conditions (i.e., deadline length, team member similarity, etc.). The work focuses on intra-team strategies where 22 only predictable and compatible issues among team members are considered. Thus, it is acknowledged that more work is required in these lines. With respect to role allocation, it should be pointed out that team mem- bers may also have their own individual goals. Teams with mixed motives have not been extensively studied in multi-agent literature, with the exceptions of [91] present a framework where agents have to reconcile [91, 92]. Grosz et al. conflict between team commitments and individual actions. If an individual action reports benefits and it is inconsistent with a committed team action, the agent can choose to decommit and pursue its own self-interest. Authors propose the use of social norms, with associated punishments and rewards, to make team commitments prevail. However, social norms are difficult to deploy in a negotiation team unless a central authority exists. Even so, it may not be possible to determine whether or not team members are collaborating since other agents' preferences are not known. Paruchuri et al. [92] present an ex- tension of Partially Observable Markov Decision Processes to the team case. However, instead of modeling a single reward for the team, individual rewards are also modeled. This way, they can take into account agents' selfishness. The frameworks attempts to calculate optimal group policies that maximize the re- wards of team members as much as possible. Nevertheless, rewards steaming from role allocation are difficult to anticipate. Therefore, it is necessary to fur- ther study to what extent team members would fully collaborate in negotiation team's tasks in spite of their own utility. For example, how interesting is for a team member to look for new outside options for the team when current ones report high utility for himself? A self-interested team member may decide to neglect its search tasks and continue with present outside options if it considers that new outside options will not increase its current welfare. We consider that this is a promising line of work that needs more attention. 3.9. Select Individual Strategy Each team member should plan its individual strategy before heading into the negotiation. An intra-team strategy defines mechanisms for team decision- making but they do not define how individual team members behave when playing those mechanisms. It is up to the agent to decide how to act inside it can be more or less cooperative. The agent should also decide the team: its attitude with the opponent. The two aforementioned factors will define the initial negotiation strategy of each team member. Generally, the selection of the initial negotiation strategy is based on what is expected about the opponent and teammates. As stated in the previous section, one of our hypotheses is that the conditions of the negotiation environment play a key role in selecting which intra-team strategies are more appropriate for each specific situation. Thus, team members should also decide on their individual strategy based on the knowledge about conditions of the negotiation. We consider that the mechanisms employed in other domains involving single player parties are also applicable in this task. Hence, no special challenge arises. 23 3.10. Negotiation & Adaptation The final phase is the negotiation itself. During this phase, team members should follow the planned intra-team strategies, individual strategies, and ne- gotiation protocols. However, negotiation is a dynamic process that may not work as planned (e.g., opponents not behaving as one initially thought, team members performing below/above one's expectations, team members leaving the team, etc.). Therefore, it may be necessary that each team member adapts its own negotiation strategy, and that the team replans some of the aspects related to team composition and team dynamics. Special Challenges We argue that some of the special challenges that arise in a negotiation team during the negotiation are: • Team membership: As stated, team membership may be dynamic. In fact, how dynamic a negotiation team is may depend on the domain ap- plication. Domains where team members are more self-interested, and less bonds exist between team members (e.g., team of buyers) may result in more dynamic negotiation teams than domains where team members are more cooperative and there are human bonds behind team membership (e.g., group of travelers, human organizations, etc.). In any case, in both situations the problem of dynamic membership may arise. For instance, new buyers may appear in the electronic market that may join the team to take advantage of larger price discounts. Similarly, a new traveler de- cides to travel during holidays and his user states the desire of joining the pre-existing group of travelers. Cases of team member's withdrawal are also possible. For instance, one of the buyers participates in other buyers' coalitions and decides to close a deal, making its membership in the rest of the buyers' coalitions no longer necessary. • Negotiation issues: As stated, both parties agreed to negotiate over some initial issues. For some reasons (e.g., computational issues, computational tractability, etc.), they may have decided to leave some less relevant issues out of the negotiation table. However, at some points of the negotiation an impasse [93, 94] may occur in the negotiation. A negotiation impasse occurs when the parties are unable to reach an agreement and the per- spectives of reaching one are very negative. They are in a deadlock. A possible solution for such problematic situation is what is known as issue linkage [95, 96, 97]. Basically, when parties negotiate on one issue, adding another issue and linking its value to the value of the initial issue can increase the probability of finding an agreement. The new issue may be added to reduce intra-team conflict (e.g., how costs are split in the team), or they may be added to reduce conflict with the opponent (e.g., include a payment method issue and maximize the preferences of the opponent in the new issue). This adaptation heuristic during the negotiation may be positive for cases that are prone to fail. However, as suggested by [96, 97], 24 issue linkage may also have negative effects since it may also reduce the agreement space. As of today, issue linkage is an area that has not been widely studied in automated negotiation, where it has been assumed that issues remain static during the negotiation process. Hence, it is an area that requires further exploration, especially for the team case since conflict may appear at the team level and the opponent level. • Intra-team strategy and individual strategy adaptation: As stated, intra- team strategies define what decisions are taken by the team, how decisions are taken, and when those decisions are taken. In case that some nego- tiation conditions change, it may be wise for team members to change their current intra-team strategy to match accordingly the new changes. Obviously, changes in the intra-team strategy and environment's condi- tion also call for an adaptation in team members' individual strategies. In this sense, there have been some works that advocate for a change in individual agents' strategies in automated negotiation. We can dis- tinguish between works where individual agents adapt their behavior at- tending to environmental conditions like outside options and competitors [98, 99, 100, 101, 102, 103] and works where individual agents adapt their behavior during the negotiation attending to the attitude of the opponent [104, 105, 106, 107, 108]. Sim et al. [98] introduce market driven agents. These agents adapt their concession strategies based on market conditions like eagerness, trading time, trading opportunities and trading competition. Later, Ren et al. [101] improve market driven agents by considering not only current market options, but also future opportunities. Differently from other approaches, in [103] the authors propose bilateral negotiation strategies for multi-issue negotiations that account for market conditions. A stochastic Markov chain is proposed in An et al. [100] with the goal of modeling market dynamics in single-issue negotiations. In their model, they assume that agents have a probability distribution over future trading opportunities, reservation prices, and deadlines. A sit-and-wait strategy is also assumed so at the first negotiation step both partners offer their highest utility value and no new bid is offered until the first private deadline, where the agent offers its reservation value. Opponent offers are accepted based on the prediction of new and better offers in the future. Li et al. [99] present a bilateral negotiation framework where the reservation price is updated based on the information of current trading opportunities and future trading opportunities. In [102], the authors propose a bilateral negotiation model with decommitment for leasing resources on the cloud. The negotiation strategy of buyers takes in consideration environmental factors like the demand ratio of resources, the cost of resources, and market competition for resources. Robu et al. [104, 105] present a learning mechanism used by sellers to adapt to the preferences of buyers. The preferences are represented by utility graphs, which model interdependencies between binary issues. 25 Narayanan et al.[106] present a negotiation framework where pairs of agents negotiate over a single issue (i.e., price). The authors assume that the environment is non-stationary in the sense that agents' strate- gies may change over time. Non-stationary Markov chains and Bayesian learning are employed to tackle the uncertainty in this domain, and even- tually converge towards the optimal negotiation strategy. Hindriks et al. [107] present a negotiation framework for bilateral multi-issue negotiations where agents' preferences are represented by means of additive utility func- tions. The main goal of this work is learning a model of the opponent's preferences (i.e., issue value rankings, and issue weights), and Bayesian learning is used for this purpose. Williams et al. [108] present a strategy that is capable of adapting its concession strategy based on the predicted behavior of the opponent. The previous works adapt different elements of the negotiation (e.g., pref- erence models, strategies, reservation prices, etc.). However, they usually focus on modeling aspects of the environment where the negotiation is being carried out or the opponent. Both opponent's models and envi- ronment's models can be applied to the team case. However, there is an additional layer of negotiation/interaction, which in this case is carried out within teammates. One agent should also consider the actions and behaviors of its teammates in order to act accordingly in the negotiation. Thus, new learning and adaptation mechanisms based on the information that is generated inside the team by team members are needed. 4. Discussion & Conclusions Agent-based negotiation teams is a fairly novel topic in the area of multi- agent systems and artificial intelligence. More specifically, an agent-based ne- gotiation team is a group of interdependent agents that join together as a single negotiation party because of their shared interests at the negotiation process. The reasons to employ these multi-individual parties are: (i) more computation and parallelization capabilities; (ii) unite agents with different expertise and skills whose joint work makes possible to tackle complex negotiation domains; (iii) the necessity to represent different stakeholders or different preferences in the same party (e.g., organizations, countries, married couple, etc.). In this article, we have analyzed and reviewed the tasks that may help agent-based negotiation teams to perform successfully. Being a special type of automated negotiation where at least one of the par- ties is a group of agents, it inherits some of the challenges from the classic topic of automated negotiation like opponent selection, typical pre-negotiation tasks (e.g., setting negotiation issues, agreeing negotiation protocol, etc.), strategy selection, and so forth. Nevertheless, an additional layer of difficulty is added to those challenges due to the fact that team members may have different expertise and different preferences (i.e., intra-team conflict). Hence, those decisions that once were taken alone or with the opponent parties should now be discussed 26 and agreed within the negotiation team. Moreover, some unique challenges like negotiation team formation and deciding the intra-team strategy (i.e., team dy- namics) that should be employed in the negotiation also arise with the new topic. In this article we have organized, in a workflow, the tasks that a negotiation team should take. The workflow aims to be general and adaptable to a wide variety of domains and applications. For each of these tasks we have highlighted the aforementioned challenges that arise with the field of agent-based negotiation teams and we have pointed out to related fields of study that may hold the right answer for such challenges. Next, we briefly describe the workflow and the challenges that we have identified throughout this article. • Identify Problem: The first step consists of being able to perceive when conflict is present, and who may be involved in the conflict (i.e., opponents, prospective teammates, and competitors). For that matter, it is necessary to analyze the environment and look for potential partners. The literature has largely focused on domain dependant conflict detection and search mechanisms in networks and markets, leaving conflict detection in open systems almost unexplored. This latter aspect is relatively important for making general negotiators in a wide variety of domains. Hence, this area of work remains largely unexplored. • Team Formation: If the agent thinks that it may be beneficial to form a negotiation team, it should attempt to select its teammates. Closely re- lated research areas are coalition formation, buyer coalitions, and cooper- ative team formation. On the one hand, coalitional approaches (including buyer coalitions) have focused on forming optimal groups of agents and how to divide the payoff generated by the group task. However, the result of the negotiation with the opponent may be difficult to anticipate mak- ing it complicated to estimate the value of coalitions. Moreover, while the payoff generated by some negotiation issues like price may be naturally di- visible, the payoff generated by other issues is not divisible (e.g., payment method). On the other hand, cooperative team formation aims to form teams for a certain task based on complementary skills. The usual trend in the literature is to consider these team members as fully cooperative and committed to the team goal. Nevertheless, in agent-based negotia- tion teams, team members may not be fully cooperative since they may additionally have their own and possibly conflicting individual interests. Special challenges identified in this task are considering (i) the relation- ship between team size and social power; (ii) the relationship between team size, team similarity, intra-team conflict and conflict with the oppo- nent; (iii) requiring different knowledge expertise; (iv) the social network provided by each team member. • Opponent Selection: The next step consists of selecting the opponents with whom the team will negotiate. Classic approaches to this task have relied on trust and reputation, or the negotiation outcome in previous 27 interactions with the opponent. Since the aforementioned perspectives are usually individual, it is necessary to form an opinion for the team. In this case, we pointed out that social choice and argumentation/deliberation in groups can play this role. The main challenge in this workflow task is adapting the aforementioned mechanisms to the opponent evaluation task. • Understand the Negotiation Domain: The general idea behind this task is creating a shared model of the negotiation domain at hand. It in- cludes identifying relevant negotiation issues, merging different points of views and expertise, clarifying team goals that may be abstract in essence, and identifying the nature of prospective negotiation issues (e.g., compati- ble and predictable, unpredictable, etc). Discussion in groups can certainly help to discover these issues during argumentation processes. Addition- ally, belief merging is also a potential candidate to solve this task as long as team members are truthful or the belief merging method is strategy-proof. The main challenges in this task are identifying the nature of negotiation issues among team members, and taking into account opportunistic be- haviors (e.g., manipulating belief merging, hiding relevant information for one's own interest, etc.). • Agree Negotiation Issues: The next part consists of agreeing with the opponent which negotiation issues should be considered in the negotia- tion. Depending on the participating parties, it is possible that some of the issues/issue values are not available for negotiation (e.g., the service provider cannot offer such issue). Additionally, it is also possible that some issues/issue values that were not initially considered by the team are included due to opponent parties' requests. Despite its importance, negotiation models usually assume that the negotiation domain is given and they do not provide mechanisms that allow forming or negotiating the domain of each negotiation. This is an interesting research challenge since, as far as we know, there is almost no related literature. • Plan Negotiation Protocol: Given a specific situation, there may be different negotiation protocols that may be used to negotiate with the op- ponent. Team members should argue about which protocols are preferred according to their experiences, known strategies , and so forth. Although the problem has not been explicitly covered by the literature, its solution may not pose exceptional efforts compared with classic argumentation and social choice problems. Nevertheless, it is also acknowledged that there is a lack of works comparing the advantages and disadvantages of available negotiation protocols under different conditions. • Agree Negotiation Protocol: Similarly, once team members have dis- cussed about the available negotiation protocols, they should negotiate a proper negotiation protocol and its parameters with the opponent. Cur- rent negotiation technologies may be able to support this phase. 28 • Decide Intra-team Strategy: Intra-team strategies define team dy- namics for a specific negotiation protocol. They define the coordination and negotiation protocol carried out within the team to decide on the steps to be carried out in the negotiation with the opponent. Team members may employ information regarding the current environment state (e.g., deadline length, number of competitors, team size, beliefs regarding the opponent, etc.) in order to decide on the most appropriate intra-team strategy. For this matter, it may be necessary to carry out simulation studies in order to discover under what circumstances some intra-team strategies perform better than others. The scarce research in negotiation teams has gone in that direction. We have identified that even though some studies exist that identify good practices and good strategies for sin- gle individual parties, the area remains largely unexplored for intra-team strategies. Moreover, if the intra-team strategy requires role differenti- ation, techniques from role/task allocation may be employed. However, it should be considered that agents may not be fully cooperative, which is the usual approach in role/task allocation. Thus, the problem slightly differs from classic role/task allocation. • Select Individual Strategy: Intra-team strategies define team dynam- ics, but they do not define the individual behavior of team members per se. The next step consists of each team member individually deciding the most appropriate individual behavior for the negotiation at hand. This task may not pose additional difficulties compared with the selection of the individual negotiation strategy in classic negotiation literature. • Negotiation & Adaptation: The final task consists of negotiating and adapting in order to face unexpected events properly. We have identified three special challenges. The first of them is team membership adaptation. The second challenge that needs some consideration is adapting negotia- tion issues. Parties can solve impasses in the negotiation and better off other parties by including other issues that were not initially included in the negotiation. As far as we are concerned, this problem has not been widely studied in multi-agent literature. The third and final challenge that we consider is the adaptation of the intra-team strategy and the individ- ual strategy. Usually, team members have planned on using an intra-team strategy and an individual strategy based on some initial prediction of the negotiation environment, opponents, and teammates' behavior. How- ever, based on new evidence, initial predictions may prove wrong and adjustments need to be done in order to properly tackle the negotiation. In automated negotiation, some works exist that allow single individual parties to adapt themselves to changes in the negotiation environment and new predictions. However, these mechanisms need to be adapted for the team case since interactions among team members also influence the negotiation. Therefore, it is observable that some issues remain open and they need to be 29 solved before deploying full-fledged applications that make use of agent-based negotiation teams. In this article, we have observed that despite these open issues, technology exists that can be adapted in the near future to solve the aforementioned challenges. We expect that, due to inherent social component of negotiating as a group, the deployment of applications supporting agent- based negotiation teams can specially benefit from social networks. For instance, travel services are classically offered to individuals in Web sites. Even though it is possible to purchase travel services for multiple individuals, usually the process only involves interactions with one single user, leaving the intra-team negotiation, if any, to human users. Human negotiations may be time consuming and costly in cognitive terms. A social network for travelers that introduces computational models for negotiation teams may help to avoid these problems. In this social network, users may indicate to their personal agents their desire to go on a travel together. Then, the agents help users from the initial stages of forming a group of travelers (i.e., a negotiation team) to the later stages where the group of agents looks for appealing travel agreements with different travel agencies. The general scheme of this social application can be extrapolated to other domains involving purchases in group. Another future application that may be supported by agent-based nego- tiation teams is management of complex democractic organizations. For ex- ample, agricultural cooperatives are supposed to be democratic institutions where groups of farmers join together to save resources for the distribution of their products. One of the main problems of agricultural cooperatives is the principal-agent problem [109]. Basically, despite being democratic institutions, agricultural cooperatives are managed by a board of directors who take decisions on behalf of the democratic institution. It has been reported in the literature [109] that dissatisfaction in cooperatives comes from the fact that the goals of members are not aligned with those of the managers. Agent-based negotiation teams may provide support for part of the processes that are carried out by cooperatives. For instance, the negotiations between agricultural cooperatives and distributors may be supported by an electronic market where the agricul- tural cooperative is modeled as an agent-based negotiation team. Each member may be represented by an electronic and personal agent that participates in the negotiation team according to the preferences of its owner. For this kind of application, it is specially important to provide scalable models for agent-based negotiation teams. We hope that with this article, the reader can be introduced to the novel topic of agent-based negotiation teams and that it facilitates their work towards building successful agent-based negotiation teams. As of today, our work has focused on designing new intra-team strategies for negotiation teams. We plan to extend our present research with new intra-team strategies, and additional phases of the proposed workflow like team formation based on agents' prefer- ences. 30 Acknowledgements This work is supported by TIN2011-27652-C03-01 and TIN2012-36586-C03- 01 of the Spanish government. References [1] D. G. Pruitt, Negotiation Behavior, Academic Press, 1981. [2] S. S. Fatima, M. J. Wooldridge, N. R. Jennings, An agenda-based frame- work for multi-issue negotiation, Artificial Intelligence 152 (1) (2004) 1–45. [3] Y. In, R. Serrano, Agenda restrictions in multi-issue bargaining (ii): un- restricted agendas, Economics Letters 79 (3) (2003) 325–331. [4] N. Gatti, Game theoretic models for strategic bargaining, in: Negotiation and Argumentation in Multi-Agent Systems, Bentham Science, 2011. URL http://home.dei.polimi.it/ngatti/ [5] S. Kraus, Negotiation and cooperation in multi-agent environments, Ar- tificial Intelligence 94 (1-2) (1997) 79–97. [6] 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. [7] L. Thompson, E. Peterson, S. Brodt, Team negotiation: An examina- tion of integrative and distributive bargaining, Journal of Personality and Social Psychology 70 (1) (1996) 66–78. [8] E. Peterson, L. Thompson, Negotiation teamwork: The impact of infor- mation distribution and accountability on performance depends on the relationship among team members, Organizational Behavior and Human Decision Processes 72 (3) (1997) 364–383. [9] S. Brodt, L. Thompson, Negotiating teams: A levels of analysis, Group Dynamics 5 (3) (2001) 208–219. [10] V. Sanchez-Anguix, V. Julian, V. Botti, A. Garcia-Fornes, Towards agent- based negotiation teams, in: Working Conference on Human Factors and Computational Models for Negotiation at Group Decision and Negotiation (HuCom@GDN2010), 2010, pp. 328–331. [11] V. Sanchez-Anguix, V. Julian, V. Botti, A. Garcia-Fornes, Studying the Impact of Negotiation Environments on Negotiation Teams' Performance, Information Sciences 219 (2013) 17–40. [12] V. Sanchez-Anguix, V. Julian, V. Botti, A. Garcia-Fornes, Reaching Unanimous Agreements within Agent-Based Negotiation Teams with Lin- ear and Monotonic Utility Functions, IEEE Transactions on Systems, Man and Cybernetics, Part B 42 (3) (2012) 778–792. 31 [13] V. Sanchez-Anguix, T. Dai, Z. Semnani-Azad, K. Sycara, V. Botti, Mod- eling power distance and individualism/collectivism in negotiation team dynamics, in: 45 Hawaii International Conference on System Sciences (HICSS-45), 2012, pp. 628–637. [14] V. Sanchez-Anguix, R. Aydogan, V. Julian, C. M. Jonker, Analysis of Intra-Team Strategies for Teams Negotiating Against Competitor, Match- ers, and Conceders, in: The 5th International Workshop on Agent-based Complex Automated Negotiations (ACAN 2012), 2012, pp. 1–8. [15] G. Browder, An analysis of the negotiations for the 1995 mekong agree- ment, International Negotiation 5 (2) (2000) 237–261. [16] K. O'Connor, Groups and solos in context: The effects of accountability on team negotiation, Organizational Behavior and Human Decision Processes 72 (3) (1997) 384–407. [17] M. J. Gelfand, J. M. Brett, L. Imai, H. Tsai, D. Huang, Team negotiation across cultures: When and where are two heads better than one?, in: Proceedings of 18th Annual Conference for the International Association for Conflict Management (IACM-2005), 2005. URL http://dx.doi.org/10.2139/ssrn.735003 [18] K. Behfar, R. A. Friedman, J. M. Brett, The team negotiation challenge: Defining and managing the internal challenges of negotiating teams, in: Proceedings of the 21st Annual Conference for the International Associa- tion for Conflict Management (IACM-2008), 2008. URL http://dx.doi.org/10.2139/ssrn.1298512 [19] E. Mannix, Strength in numbers: Negotiating as a team, Negotiation 8 (5) (2005) 3–5. [20] L. Thompson, The Mind and heart of the negotiator, Prentice Hall Press, Upper Saddle River, NJ, USA, 2005. [21] S. E. Brodt, M. Tuchinsky, Working together but in opposition: An exam- ination of the good-cop/bad-cop negotiating team tactic, Organizational Behavior and Human Decision Processes 81 (2) (2000) 155 – 177. [22] N. Halevy, Team negotiation: Social, epistemic, economic, and psycholog- ical consequences of subgroup conflict, Personality and Social Psychology Bulletin 34 (2008) 1687–1702. [23] H. Taylor, Most people are privacy pragmatists who, while concerned about privacy, will sometimes trade it off for other benefits, The Harris Poll 17 (2003) 19. [24] R. G. Smith, The contract net protocol: High-level communication and control in a distributed problem solver, Computers, IEEE Transactions on 100 (12) (1980) 1104–1113. 32 [25] T. Sandholm, An implementation of the contract net protocol based on marginal cost calculations, in: Proceedings of the eleventh national con- ference on Artificial intelligence, AAAI Press, 1993, pp. 256–262. [26] T. Nguyen, N. Jennings, Coordinating multiple concurrent negotiations, in: Proceedings of the Third International Joint Conference on Au- tonomous Agents and Multiagent Systems, IEEE Computer Society, 2004, pp. 1064–1071. [27] B. An, K. Sim, L. Tang, S. Li, et al., Continuous-time negotiation mech- anism for software agents, Systems, Man, and Cybernetics, Part B: Cy- bernetics, IEEE Transactions on 36 (6) (2006) 1261–1272. [28] Y. Shoham, K. Leyton-Brown, Multiagent systems: Algorithmic, game- theoretic, and logical foundations, Cambridge University Press, 2009. [29] C. R. Williams, V. Robu, E. H. Gerding, N. R. Jennings, Negotiating concurrently with unknown opponents in complex, real-time domains, in: 20th European Conference on Artificial Intelligence, Vol. 242, 2012, pp. 834–839. [30] K. Mansour, R. Kowalczyk, A meta-strategy for coordinating of one-to- many negotiation over multiple issues, in: Foundations of Intelligent Sys- tems, Vol. 122, Springer Berlin / Heidelberg, 2012, pp. 343–353. [31] C. Beam, A. Segev, Automated negotiations: A survey of the state of the art, Wirtschaftsinformatik 39 (3) (1997) 263–268. [32] R. H. Guttman, A. G. Moukas, P. Maes, Agent-mediated electronic com- merce: a survey, Knowledge Engineering Review 13 (2) (1998) 147–159. [33] A. R. Lomuscio, M. Wooldridge, N. R. Jennings, A classification scheme for negotiation in electronic commerce, Group Decision and Negotiation 12 (2003) 31–56. [34] M. He, N. R. Jennings, H. Leung, On agent-mediated electronic commerce, IEEE Trans on Knowledge and Data Engineering 15 (4) (2003) 985–1003. [35] F. Lopes, M. Wooldridge, A. Novais, Negotiation among autonomous com- putational agents: principles, analysis and challenges, Artificial Intelli- gence Review 29 (1) (2008) 1–44. [36] X. Luo, C. Miao, N. R. Jennings, M. He, Z. Shen, M. Zhang, Kemnad: A knowledge engineering methodology for negotiating agent development, Computational Intelligence 28 (1) (2012) 51–105. [37] C. Tomlin, G. Pappas, S. Sastry, Conflict resolution for air traffic man- agement: A study in multiagent hybrid systems, IEEE Transactions on Automatic Control 43 (4) (1998) 509–521. 33 [38] J. Chu-Carroll, S. Carberry, Conflict resolution in collaborative planning dialogs, International Journal of Human-Computer Studies 53 (6) (2000) 969 – 1015. [39] K. Barber, T. Liu, S. Ramaswamy, Conflict detection during plan integra- tion for multi-agent systems, IEEE Transactions on Systems, Man, and Cybernetics, Part B: Cybernetics 31 (4) (2001) 616–628. [40] S. Wollkind, J. Valasek, T. Ioerger, Automated conflict resolution for air traffic management using cooperative multiagent negotiation, in: AIAA guidance, navigation, and control conference, 2004, pp. 1–11. [41] F. Lopes, N. Mamede, A. Novais, H. Coelho, A negotiation model for au- tonomous computational agents: Formal description and empirical evalu- ation, Journal of Intelligent & Fuzzy Systems 12 (3) (2002) 195–212. [42] D. Sarne, E. Manisterski, S. Kraus, Multi-goal economic search using dynamic search structures, Autonomous Agents and Multi-Agent Systems 21 (2010) 204–236. [43] E. D. Val, M. Rebollo, V. Botti, Self-organized service management in social systems, in: Proceedings of the 45 Hawaii International Conference on System Sciences (HICSS 2012), 2012, pp. 810–817. [44] R. Lopez de Mantaras, Case-based reasoning, in: Machine Learning and Its Applications, Vol. 2049 of Lecture Notes in Computer Science, Springer Berlin / Heidelberg, 2001, pp. 127–145. [45] T. Sandhlom, V. Lesser, Coalitions among computationally bounded agents, Artificial Intelligence 94 (1) (1997) 99–137. [46] V. Conitzer, T. Sandholm, Computing shapley values, manipulating value division schemes, and checking core membership in multi-issue domains, in: Proceedings of the 19th national conference on Artifical intelligence (AAAI 2004), AAAI Press, 2004, pp. 219–225. [47] N. Ohta, V. Conitzer, Y. Satoh, A. Iwasaki, M. Yokoo, Anonymity-proof shapley value: extending shapley value for coalitional games in open en- vironments, in: Proceedings of the 7th international joint conference on Autonomous agents and multiagent systems, International Foundation for Autonomous Agents and Multiagent Systems, 2008, pp. 927–934. [48] Y. Zick, E. Elkind, Arbitrators in overlapping coalition formation games, in: Proceedings of the 10th International Conference on Autonomous Agents and Multiagent Systems (AAMAS'11), International Foundation for Autonomous Agents and Multiagent Systems, 2011, pp. 55–62. [49] T. Rahwan, T. Michalak, E. Elkind, P. Faliszewski, J. Sroka, M. Wooldridge, N. Jennings, Constrained coalition formation, in: Twenty Fifth AAAI Conference on Artificial Intelligence (AAAI), 2011, pp. 719– 725. 34 [50] T. Rahwan, Algorithms for coalition formation in multi-agent systems, Ph.D. thesis, University of Southampton (2007). [51] T. Rahwan, S. D. Ramchurn, N. R. Jennings, A. Giovannucci, An anytime algorithm for optimal coalition structure generation, Journal of Artificial Intelligence Research 34 (1) (2009) 521–567. [52] M. Tsvetovat, K. Sycara, Y. Chen, J. Ying, Customer coalitions in elec- tronic markets, in: Agent-Mediated Electronic Commerce III, Vol. 2003 of Lecture Notes in Computer Science, Springer Berlin / Heidelberg, 2001, pp. 121–138. [53] J. Yamamoto, K. Sycara, A stable and efficient buyer coalition forma- tion scheme for e-marketplaces, in: Proceedings of the fifth international conference on Autonomous agents (AGENTS'01), ACM, New York, NY, USA, 2001, pp. 576–583. [54] T. Ito, H. Ochi, T. Shintani, A group-buy protocol based on coalition formation for agent-mediated e-commerce, International Journal of Com- puter and Information Science 3 (1) (2002) 11–20. [55] C. Li, U. Rajan, S. Chawla, K. Sycara, Mechanisms for coalition formation and cost sharing in an electronic marketplace, in: Proceedings of the 5th international conference on Electronic commerce (ICEC'03), ACM, New York, NY, USA, 2003, pp. 68–77. [56] L. He, T. Ioerger, Combining bundle search with buyer coalition forma- tion in electronic markets: A distributed approach through explicit ne- gotiation, Electronic Commerce Research and Applications 4 (4) (2006) 329–344. [57] P. Cohen, H. Levesque, I. Smith, On team formation, in: Contemporary Action Theory. Synthese, Kluwer Academic Publishers, 1999, pp. 87–114. [58] F. Dignum, B. Dunin-Keplicz, R. Verbrugge, Agent theory for team for- mation by dialogue, in: Proceedings of the 7th International Workshop on Intelligent Agents VII. Agent Theories Architectures and Languages (ATAL'00), Springer-Verlag, London, UK, 2001, pp. 150–166. [59] M. E. Gaston, M. desJardins, Agent-organized networks for dynamic team formation, in: Proceedings of the fourth international joint conference on Autonomous agents and multiagent systems (AAMAS'05), ACM, New York, NY, USA, 2005, pp. 230–237. [60] C. Badica, S. Ilie, M. Kamermans, G. Pavlin, A. Penders, M. Scafes, Multi-agent systems, ontologies and negotiation for dynamic service com- position in multi-organizational environmental management, in: Software Agents, Agent Systems and Their Applications, 2012, pp. 286–306. 35 [61] K. Sycara, G. Sukthankar, Literature review of teamwork models, CMU- RI-TR-06-50. [62] B. Rezabakhsh, D. Bornemann, U. Hansen, U. Schrader, Consumer power: A comparison of the old economy and the internet economy, Journal of Consumer Policy 29 (2006) 3–36. [63] J. Sabater, C. Sierra, Review on computational trust and reputation mod- els, Artificial Intelligence Review 24 (1) (2005) 33–60. [64] S. Younger, Reciprocity, sanctions, and the development of mutual obli- gation in egalitarian societies, Journal of Artificial Societies and Social Simulation 8 (2). URL http://jasss.soc.surrey.ac.uk/8/2/9.html [65] A. Perreau de Pinninck, C. Sierra, M. Schorlemmer, Distributed norm en- forcement: Ostracism in open multi-agent systems, in: Computable Mod- els of the Law, Vol. 4884 of Lecture Notes in Computer Science, Springer Berlin / Heidelberg, 2008, pp. 275–290. [66] J. Brzostowski, R. Kowalczyk, Experimental evaluation of possibilistic mechanism for negotiation partners selection, in: Rational, Robust, and Secure Negotiations in Multi-Agent Systems, Vol. 89, Springer Berlin / Heidelberg, 2008, pp. 127–145. [67] J. Schafer, D. Frankowski, J. Herlocker, S. Sen, Collaborative filtering recommender systems, in: The Adaptive Web, Vol. 4321 of Lecture Notes in Computer Science, Springer Berlin / Heidelberg, 2007, pp. 291–324. [68] J. M. Such, A. Espinosa, A. Garcia-Fornes, V. Botti, Partial identities as a foundation for trust and reputation, Engineering Applications of Artificial Intelligence 24 (7) (2011) 1128–1136. [69] S. Koc-Menard, Team performance in negotiation: a relational approach, Team Performance Management 15 (7-8) (2009) 357–365. [70] H. Nurmi, Voting systems for social choice, Handbook of Group Decision and Negotiation (2010) 167–182. [71] P. McBurney, D. Hitchcock, S. Parsons, The eightfold way of deliberation dialogue, International Journal of Intelligent Systems 22 (1) (2007) 95– 132. [72] S. Heras, J. Jord´an, V. Botti, V. Julian, Argue to Agree: A Case-Based Argumentation Approach, International Journal of Approximate Reason- ing 54 (1) (2013) 82–108. [73] R. Booth, Social contraction and belief negotiation, Information Fusion 7 (1) (2006) 19–34. 36 [74] G. Qi, W. Liu, D. Bell, Combining multiple prioritized knowledge bases by negotiation, Fuzzy Sets and Systems 158 (23) (2007) 2535–2551. [75] P. Everaere, S. Konieczny, P. Marquis, et al., The strategy-proofness land- scape of merging, Journal of Artificial Intelligence Research 28 (1) (2007) 49–105. [76] S. Konieczny, R. Pino P´erez, Logic based merging, Journal of Philosoph- ical Logic 40 (2) (2011) 239–270. [77] J. M. Such, A. Espinosa, A. Garcia-Fornes, C. Sierra, Self-disclosure De- cision Making based on Intimacy and Privacy, Information Sciences 211 93–111. [78] P. Faratin, Automated service negotiation between autonomous compu- tational agents, Ph.D. thesis, University of London (2000). [79] A. Rubinstein, Perfect equilibrium in a bargaining model, Econometrica 50 (1982) 155–162. [80] G. Lai, K. Sycara, C. Li, A decentralized model for automated multi- attribute negotiations with incomplete information and general utility functions, Multiagent and Grid Systems 4 (1) (2008) 45–65. [81] X. Luo, N. Jennings, N. Shadbolt, H. Leung, J. Lee, A fuzzy constraint based model for bilateral, multi-issue negotiations in semi-competitive en- vironments, Artificial Intelligence 148 (1) (2003) 53–102. [82] L. Pan, X. Luo, X. Meng, C. Miao, M. He, X. Guo, A two-stage win– win multiattribute negotiation model: Optimization and then concession, Computational Intelligence (2012) In Press. URL http://dx.doi.org/10.1111/j.1467-8640.2012.00434.x [83] M. Klein, P. Faratin, H. Sayama, Y. Bar-Yam, Negotiating complex con- tracts, Group Decision and Negotiation 12 (2003) 111–125. [84] T. Ito, M. Klein, H. Hattori, A multi-issue negotiation protocol among agents with nonlinear utility functions, Multiagent and Grid Systems 4 (1) (2008) 67–83. [85] J. R. Oliver, On artificial agents for negotiation in electronic commerce, in: The 29 Hawaii International Conference on System Sciences (HICSS'96), Vol. 4, IEEE Computer Society, Washington, DC, USA, 1996, pp. 337– 346. [86] N. Matos, C. Sierra, N. R. Jennings, Determining successful negotiation strategies: An evolutionary approach, in: The 3rd International Con- ference on Multi Agent Systems (ICMAS '98), IEEE Computer Society, Washington, DC, USA, 1998, pp. 182–189. 37 [87] T. Tu, E. Wolff, W. Lamersdorf, Genetic algorithms for automated nego- tiations: A fsm-based application approach, in: Proceedings of the 11th International Conference on Database and Expert Systems (DEXA 2000), IEEE, 2000, pp. 1029–1033. [88] E. Gerding, D. Van Bragt, H. La Poutr´e, Multi-issue negotiation processes by evolutionary simulation, validation and social extensions, Computa- tional Economics 22 (1) (2003) 39–63. [89] R. Nair, M. Tambe, S. Marsella, Role allocation and reallocation in multi- agent teams: Towards a practical analysis, in: Proceedings of the second international joint conference on Autonomous agents and multiagent sys- tems, ACM, 2003, pp. 552–559. [90] M. Hoogendoorn, C. Jonker, Formation of virtual organizations through negotiation, in: Multiagent System Technologies, Vol. 4196 of Lecture Notes in Computer Science, Springer Berlin / Heidelberg, 2006, pp. 135– 146. [91] B. J. Grosz, S. Kraus, D. G. Sullivan, S. Das, The influence of social norms and social consciousness on intention reconciliation, Artificial Intelligence 142 (2) (2002) 147 – 177. [92] P. Paruchuri, M. Tambe, S. Kapetanakis, S. Kraus, Between collaboration and competition: An initial formalization using distributed pomdps, in: Proceedings of the GTDT workshop at AAMAS-2003, 2007. URL http://citeseerx.ist.psu.edu/viewdoc/summary?doi=10.1.1. 2.5244 [93] B. Spector, Creativity heuristics for impasse resolution: Reframing in- tractable negotiations, The Annals of the American Academy of Political and Social Science 542 (1) (1995) 81–99. [94] L. Babcock, G. Loewenstein, Explaining bargaining impasse: The role of self-serving biases, The Journal of Economic Perspectives 11 (1) (1997) 109–126. [95] R. Tollison, T. Willett, An economic theory of mutually advantageous issue linkages in international negotiations, International Organization 33 (4) (1979) 425–449. [96] J. Sebenius, Negotiation arithmetic: adding and subtracting issues and parties, International Organization 37 (2) (1983) 281–316. [97] T. Morgan, Issue linkages in international crisis bargaining, American Journal of Political Science 34 (2) (1990) 311–333. [98] K. M. Sim, C. Y. Choi, Agents that react to changing market situations, IEEE Transactions on on Systems, Man, and Cybernetics, Part B: Cyber- netics 33 (2) (2003) 188 – 201. 38 [99] C. Li, J. Giampapa, K. Sycara, Bilateral negotiation decisions with uncer- tain dynamic outside options, IEEE Transactions on Systems, Man, and Cybernetics, Part C: Applications and Reviews 36 (1) (2006) 31–44. [100] B. An, K. M. Sim, C. Y. Miao, Z. Q. Shen, Decision making of negotiation agents using markov chains, Multiagent and Grid Systems 4 (1) (2008) 5– 23. [101] F. Ren, M. Zhang, K. Sim, Adaptive conceding strategies for automated trading agents in dynamic, open markets, Decision Support Systems 46 (3) (2009) 704–716. [102] B. An, V. Lesser, D. Irwin, M. Zink, Automated negotiation with de- commitment for dynamic resource allocation in cloud computing, in: Pro- ceedings of the 9th International Conference on Autonomous Agents and Multiagent Systems, 2010, pp. 981–988. [103] F. Ren, M. Zhang, X. Luo, D. Soetanto, A parallel, multi-issue negoti- ation model in dynamic e-markets, in: AI 2011: Advances in Artificial Intelligence, Springer, 2011, pp. 442–451. [104] V. Robu, D. J. A. Somefun, J. A. La Poutr´e, Modeling complex multi- issue negotiations using utility graphs, in: Proceedings of the fourth inter- national joint conference on Autonomous agents and multiagent systems (AAMAS'05), ACM, New York, NY, USA, 2005, pp. 280–287. [105] 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: Rational, Robust and Secure Negotiations, Vol. 89 of Computational Intelligence, Springer, 2008. [106] V. Narayanan, N. R. Jennings, Learning to negotiate optimally in non- stationary environments, in: Cooperative Information Agents X, 10th In- ternational Workshop, CIA 2006, Edinburgh, UK, September 11-13, 2006, Proceedings, Vol. 4149 of Lecture Notes in Computer Science, Springer, 2006, pp. 288–300. [107] K. V. Hindriks, D. Tykhonov, Opponent modelling in automated multi- issue negotiation using bayesian learning, in: The 7th International Joint Conference on Autonomous Agents and Multiagent Systems (AAMAS'08), International Foundation for Autonomous Agents and Multiagent Sys- tems, Richland, SC, 2008, pp. 331–338. [108] C. Williams, V. Robu, E. Gerding, N. Jennings, Using gaussian processes to optimise concession in complex negotiations against unknown oppo- nents, in: Proceedings of the 22nd International Joint Conference on Ar- tificial Intelligence (IJCAI'11), 2011, pp. 432–438. [109] G. Ortmann, R. King, Agricultural cooperatives i: history, theory and problems, Agrekon 46 (1) (2007) 18–46. 39
1710.03354	1	1710	2017-10-10T00:00:03	The Role of Data-driven Priors in Multi-agent Crowd Trajectory Estimation	[ "cs.MA" ]	Trajectory interpolation, the process of filling-in the gaps and removing noise from observed agent trajectories, is an essential task for the motion inference in multi-agent setting. A desired trajectory interpolation method should be robust to noise, changes in environments or agent densities, while also being yielding realistic group movement behaviors. Such realistic behaviors are, however, challenging to model as they require avoidance of agent-agent or agent-environment collisions and, at the same time, seek computational efficiency. In this paper, we propose a novel framework composed of data-driven priors (local, global or combined) and an efficient optimization strategy for multi-agent trajectory interpolation. The data-driven priors implicitly encode the dependencies of movements of multiple agents and the collision-avoiding desiderata, enabling elimination of costly pairwise collision constraints and resulting in reduced computational complexity and often improved estimation. Various combinations of priors and optimization algorithms are evaluated in comprehensive simulated experiments. Our experimental results reveal important insights, including the significance of the global flow prior and the lesser-than-expected influence of data-driven collision priors.	cs.MA	cs	The Role of Data-driven Priors in Multi-agent Crowd Trajectory Estimation Gang Qiao1, Sejong Yoon2, Mubbasir Kapadia1, Vladimir Pavlovic1 1Rutgers University, 2The College of New Jersey {gq19, mk1353, vladimir}@cs.rutgers.edu, [email protected] 7 1 0 2 t c O 0 1 ] A M . s c [ 1 v 4 5 3 3 0 . 0 1 7 1 : v i X r a Abstract Trajectory interpolation, the process of filling-in the gaps and removing noise from observed agent trajectories, is an essen- tial task for the motion inference in multi-agent setting. A desired trajectory interpolation method should be robust to noise, changes in environments or agent densities, while also being yielding realistic group movement behaviors. Such re- alistic behaviors are, however, challenging to model as they require avoidance of agent-agent or agent-environment col- lisions and, at the same time, seek computational efficiency. In this paper, we propose a novel framework composed of data-driven priors (local, global or combined) and an effi- cient optimization strategy for multi-agent trajectory interpo- lation. The data-driven priors implicitly encode the depen- dencies of movements of multiple agents and the collision- avoiding desiderata, enabling elimination of costly pairwise collision constraints and resulting in reduced computational complexity and often improved estimation. Various combina- tions of priors and optimization algorithms are evaluated in comprehensive simulated experiments. Our experimental re- sults reveal important insights, including the significance of the global flow prior and the lesser-than-expected influence of data-driven collision priors. Introduction In a multi-agent scenario, a tracking system deployed on an observing, moving or stationary agent (camera), needs to estimate complete trajectories of other moving agents (people, robots, crowds) c.f., (Bera and Manocha 2014; Bera, Kim, and Manocha 2015). However, direct tracker out- put of the observing agent is often insufficient to reconstruct accurate trajectories because of factors such as noise, the environmental configuration, density of agents in a crowd, or hardware failures. For instance, during the movement of multiple agents in a complex scenario, there are inevitable inter-agent occlusions and agent-obstacle occlusions from the perspective of the observing agent, resulting in observed trajectories that are both noisy and incomplete. To obtain high-quality complete trajectories, tracking sys- tems typically adopt a multi-step strategy. First, the sensor deployed on the observing agent detects and tracks objects in its neighboring area to obtain local tracklets, continuous but short trajectories, of other moving agents. The tracklets of specific agents are then linked together by applying track- let similarity measurements and re-identification algorithms (e.g., min-cost flow). Next, a trajectory interpolation algo- rithm is applied to fill in the gaps between the linked track- lets, in a way that the filled portions present desired realis- tic properties such as containing few potential collisions and saving energy. Finally, trajectory extrapolation could be ap- plied to the reconstructed trajectories so that the observing agent can further plan its own motion. While several trajectory interpolation approaches have been proposed to-date (Rodriguez et al. 2011; Sharma, Huang, and Nevatia 2012; Bera and Manocha 2014), few have proven to work robustly in a general setting, often exhibiting slow performance and inability to tackle com- plex scenarios. In practice, many real-time tracking sys- tems typically trade-off accuracy largely for speed (Bera and Manocha 2014). In addition, few approaches attempt to in- tegrate realistic data-driven priors into the estimation frame- work except for few very recent attempts in simulation and content generation community (Bera, Kim, and Manocha 2016), resulting in interpolated trajectories that fail to match typical realistic crowd behaviors. In this work we aim to develop a multi-agent trajectory interpolation method of low complexity that results in realis- tic agent trajectory estimates. To that end, we propose a new multi-agent trajectory interpolation framework composed of a data-driven prior (either local, global or the combination thereof) and an optimization algorithm. The data prior im- plicitly encodes the movement dependencies of multiple agents and thus can decouple individual agent trajectories, resulting in reduced computational complexity while main- taining or improving the quality of estimation. We evaluate different combinations of prior representations in simulated experiments and demonstrate the essential role of these pri- ors to accomplish the low-complexity, high-accuracy multi- agent trajectory estimation task. Prior Work Modeling, simulation and analysis of crowd motion have at- tracted significant interest in research community over the past years, e.g. (Jacques Junior, Raupp Musse, and Jung 2010; Ali et al. 2013; Kapadia, Pelechano, and Allbeck 2015). We refer readers to survey papers for important works in this area (Zhan et al. 2008; Li et al. 2015). In the follow- manathan, and Fei-Fei 2014) proposed origin-destination priors to enhance trajectory estimation performance in the context of global optimization-based data-driven methods. They introduced a handcrafted feature called social affin- ity map (SAM) to capture relative adjoining positions of a number of people within a limited area, by computing the histograms of a number of agents in that region. In contrast, our global optimization framework allows to describe the motion patterns of crowds in a complex scenario within the visibility area of an agent that incorporates much richer in- formation (e.g., the configuration of the environmental ob- stacles, the shapes and velocities of the agents, rather than a histogram of agent numbers in 2D space) than SAM fea- ture. Obtaining high-quality simulation data of crowd mo- tion is an important aspect in the context of data-driven crowd trajectory estimation, with many solutions proposed in the graphics community (Kapadia, Pelechano, and All- beck 2015). Since crowd simulation, tracking and trajectory estimation are closely related, crowd simulators (Helbing and Moln´ar 1995; van den Berg, Lin, and Manocha 2008) may be utilized as the source of motion priors or training data for motion prior models necessary to improve the track- ing accuracy (Lerner, Chrysanthou, and Lischinski 2007). Neural Network-based Methods. The work in (Long, Liu, and Pan 2017) introduced an approach for data-driven collision-minimizing motion planning. In this approach a neural network (NN) is trained from examples of collision- free behaviors, aimed at cloning the Optimal Reciprocal Collision Avoidance (ORCA) policy (van den Berg et al. 2011), which provides a sufficient condition for avoiding collisions if the agents are not densely packed, otherwise ORCA has to select a reasonable velocity. However, this method can not mimic the movement behavior well without a large amount of training samples. Even with large train- ing sets, emphasizing realistic behaviors such as collision minimization is still difficult, since behavior cloning gen- erates only rough imitations while ignoring some detailed properties without environmental interactions to reinforce those properties. In contrast, our approach can embed var- ious priors, including not only the neural network based lo- cal prior, but also globally informed probabilistic velocity priors (see Sec. ) that are critical for proper behaviors, into a trajectory interpolation framework. This leads to an advan- tage over either ORCA or neural network-based behavior cloning, resulting in a framework that can handle large por- tions of missing trajectories. Notation t; si t; yi We use the following notation to define the framework. Typ- ical tracker output at a specific time point is represented with location and a timestamp. For 2D trackers, the triple t] specifies the x-y location and the corresponding [xi timestamp of agent i at video frame (time step) t. In this work, we assume a uniform temporal grid with a fixed tem- poral gap between each consecutive pair of frames of a tra- jectory. Under this assumption, we can omit the timestamp t from the triple. si Figure 1: Overview of the proposed framework. A tracker generates noisy and missing trajectories due to occlusions or sensor failure. Our approach takes such observations and applies a global optimization-based trajectory interpolation framework, which incorporates data-driven priors from a model trained on trajectories generated by crowd simulator. Outcome of the optimization (updated trajectories) can be used as new observations to obtain new priors for iteratively improving the tracking performance. ing, discuss related work on crowd motion modeling and es- timation in context of the trajectory interpolation problem. Crowd Motion and Tracking. It is common to introduce a motion prior to improve pedestrian tracking accuracy (Hu, Ali, and Shah 2008; Ali and Shah 2008; Rodriguez, Ali, and Kanade 2009). The Social Force model (Helbing and Moln´ar 1995) has been one particularly popular choice (Pellegrini et al. 2009; Pellegrini et al. 2010), that represents individ- ual dynamic behavior as a combination of different types of forces that characterize both attractive and repulsive ef- fects between pairs of agents or between an agent and an obstacle (an obstacle could be viewed as a static agent with its own shape). Extensions include the multi-target tracking model (Pellegrini and Gool 2013) that incorporates several different priors such as appearance, physical constraints and social behavior of pedestrians. Bera and Manocha (Bera and Manocha 2014; Bera, Kim, and Manocha 2015) introduced a real-time algorithm for trajectory estimation, which works in medium density crowds using particle filtering. Their ap- proach relies on a multi-agent motion model called velocity obstacle (van den Berg, Lin, and Manocha 2008) that delim- its a space of relative velocity within which collision free behavior is guaranteed for a certain period of time. Simi- lar to these model-based approaches (Pellegrini et al. 2009; Bera, Kim, and Manocha 2015), our approach also relies on a synthetic simulator to provide motion priors for crowd trajectory estimation, but it utilizes the prior as a term within a modified global optimization framework of (Yoon et al. 2016), which is different from the aforementioned ap- proaches. In this way, complex dependencies of movements are implicitly encoded in the prior to decouple trajectories while the optimization provides flexibility of incorporating desired properties in the objective to capture more realistic behaviors. Crowd Trajectory Estimation. Alahi et al. (Alahi, Ra- t (cid:9)T (cid:9)T i=1, where Xi = (cid:8)xi i=1, Oi =(cid:8)oi Our goal is to estimate a set of trajectories for N agents {Xi}N t=1 is the desired but unob- served trajectory of agent i. We will be estimating {Xi}N i=1 from the corresponding observed fragments of trajectories {Oi}N t=1, with noisy and missing portions. We use the term tracklet to refer an observed continuous fragment of a trajectory, and we assume known identifica- tion (i.e., correspondences between measurements and esti- mates) of all tracklets. Let ∆t denote the sampling period between two consecutive samples t and t − 1, and vi t de- note the true average velocity of agent i over time interval [t − 1, t], thus xi t. Index t = 0 indicates known initial position of each agent, xi 0. t−1 + ∆t · vi t = xi t Data Driven Priors We consider two types of data driven priors: local collision avoidance priors and global flow priors. We next described these priors and how they will be used in the global opti- mization framework. Local Collision Avoidance Prior The goal of utilizing local collision avoidance priors is to effectively replace the collision constraints in (Yoon et al. 2016). We will use the data driven regression approach that aims to estimate the velocity vi of agent i at space-time point xi as vi t+1 = fN N t,N N , vi t (1) (cid:12)(cid:12) θN N (cid:1) , (cid:0) oi where oi t,N N is the local measurement which encodes local visibility of the state space of agent i and θN N is the param- eter of the model. Similar to (Long, Liu, and Pan 2017), the ot,N N generally includes the following three components: (a) 2D desired ve- locity: we assume that each agent receives a global velocity guidance signal vi t. This signal is typically a velocity vector pointing toward the goal position of each agent, while dis- regarding local environment geometry or other agents. (b) Local range/occupancy map: we assume that each agent is equipped with a 360-dimensional distance scanner to col- lect distances to surface points of other agents and obstacles within a certain range. Within the distance map relative po- sitions and shapes of the environmental configurations and other agents are explicitly encoded as distances along a 360 degree circularly sampled grid. (c) 360 × 2 dimensional ve- locity map: in addition to aforementioned distances to other agents or obstacles, we also assume the local velocity mea- surements of neighboring agents or obstacles. Global Flow Prior In a multi-agent setting, individual agent movement typi- cally follows a flow pattern that depends on the environ- ment and obstacles, other agents and their density, as well as the global movement goal. This pattern can be encoded in a global flow-field. While one could obtain it from a path plan- ning algorithms, we instead use a data driven approach to capture the global flow field patterns. Specifically, we model the flow field as a random walk encoded by a Gaussian Pro- cess (GP) prior: t + δi(cid:0)xi t (cid:1) , where (cid:0)xi t δi x xi t+∆t = xi (cid:1) ∼ GP(cid:0)δxi t, st t, Xtrain, θGP (cid:1) . (2) (3) t) is the "global" velocity of agent i at frame t, and δi x(xi t denotes the predicted standard deviation of the GP. The st GP will now model the data-prior velocity field in (x, y, s) space, similar to (Kim, Lee, and Essa 2011). Xtrain denotes the trajectory data used to train the GP model and θGP the model hyper-parameter. Embedding Priors into Multi-agent Optimization Framework The essence of the multi-agent optimization framework is to estimate (interpolate) the trajectories of a set of N agents X from some set of observations O, given a partially observ- able environment Z, P r (XO, Z). In (Yoon et al. 2016), this problem is formulated as a MAP estimation problem by minimizing the following Gibbs energy u(xi, vioi) + Ei Ei,j p (xi, vi, xj, vj), (4) (cid:88) (cid:88) i(cid:54)=j i (cid:88) i u is the energy term related to an individual agent i where Ei p denotes the pairwise energy term de- (i.e., unary) and Ei,j scribing dependencies of a pair of agents (i, j). The unary term includes energies that model kinematic constraint, maximum velocity constraint as well as the compatibility between the estimated trajectories and the measurements oi. The pairwise term is responsible for avoiding collisions be- tween agents. However, joint optimization of X is a chal- lenging task, in part because of the existence of the pairwise terms as well as the lack of strong motion priors. We next describe some of the weak but frequently used priors and then suggest a way to combine our data-driven priors while keeping the computational complexity under control. Existing Unary Priors In the work (Yoon et al. 2016), the following three unary energy terms are defined to model individual behavior of an agent: tracker output, kinetic energy, and maximum veloc- ity constraint. The Tracker Output term seeks to keep the estimated trajectory close to the measured trajectories while taking into account the amount of observation uncertainty t: ui gt(xioi) = Ei t(cid:107)xi ui t − oi t(cid:107)2. (5) For instance, ui t = 0 indicates that the measurement is miss- ing. The Kinetic Energy term ensures that the total traveled distance is minimized: Ei kn(xi) = Ckn (cid:107)xi t − xi t−1(cid:107)2. (6) (cid:88) t (cid:88) t Parameter Ckn can be interpreted as the mass of an agent, which typically could be set as 1. Finally, the Maximum Ve- locity Constraint certifies that each agent's speed not exceed a physically feasible velocity, given as Cmv: Ei mv(xi) = if (cid:107)vi t(cid:107) ≤ Cmv ∞ otherwise ,∀t = 1..T. (7) (cid:26)0 Combining Local and Global Priors The key challenge in the aforementioned multi-agent opti- mization framework arises from the existence of the pair- wise terms, implying coupling trajectories between agents, the highly nonlinear nature of the coupling (collision) con- straints and the expensive computations due to this cou- pling. If such coupling were to be eliminated, the optimiza- tion of trajectories for each agent could be solved indepen- dently from other agents. However, their elimination would result in infeasible motion with possibly many collisions. We therefore propose to replace the computationally costly pairwise terms with stronger global motion and local data- driven collision priors. Specifically, we propose to modify the unary objectives by augmenting them as follows: dg(vioi) = Ei (cid:1)(cid:13)(cid:13)vi 1 t, θGP (cid:0)oi (cid:13)(cid:13)vi σ2 GP 1 σ2 N N + t+∆t − µGP (cid:0)oi t+∆t − fN N N NθN N (cid:1)(cid:13)(cid:13)2 (cid:0)oi t, θGP (cid:1)(cid:13)(cid:13)2 (8) Here, µGP is the predictive mean and σGP is the predictive standard deviation given by the learned GP and σN N is the standard deviation of the NN regression model. Integrating all unary terms above, we obtain the final global objective, which we seek to minimize in the estima- tion process: (cid:88) X = arg min X i u(xi, vioi) = arg min Ei X +Ei kn(xi) + Ei mv(xi) + Ei (cid:88) i gt(xioi) Ei dg(xioi) (9) Optimization of the Global Objective In the following, we introduce three optimization ap- proaches to solve the optimization problem in Eq. 9. In gen- eral, our optimization framework is iterative and outlined in Alg. 1. The iterative nature of our algorithm stems from the coupling between the essential collision-avoiding NN term in Eq. 1 and the solution to the optimization prob- lem. Namely, the NN term's range observations require the knowledge of the agents' locations and velocities, which are the variable we are solving for. To mitigate this effect, we propose the alternating optimization scheme where the agent's trajectories from a previous iteration are used as the proxies for measurements in Eq. 1. The choice of the minimizer in Step-5 of this algorithm is important but less essential. We consider three methods for optimizing optimizing Eq. 9: a message-passing algo- rithm (MPA) of (Bento et al. 2013; Yoon et al. 2016), a gen- eral interior point method (IPM), and an unscented Kalman Algorithm 1: Proposed Optimization Framework Input : O, (θN N , σN N ), (θGP , µGP , σGP ), Ckn, Cmv Output: X 1 Initialize X; 2 repeat 3 4 5 6 until X converges; 7 X = X; Compute NN prior velocities of X using Eq. 1; Compute GP prior velocities of X using Eq. 3; Find X by minimizing Eq. 9; smoother (UKS) that exploits the sequential nature of each (independent) agent's trajectory optimization task while ap- plying a nonlinear Rauch-Tung-Striebel smoother (Rauch, Striebel, and Tung 1965). Further details of some of the se- lected approaches are provided in the Supplement. Experiments To evaluate the proposed framework, we consider 6 exper- imental settings similar to those in (Yoon et al. 2016): 3 different settings of bottlenecks (each contains a challeng- ing egress in evacuation-like scenarios), concentric circle (agents are symmetrically placed along a circle and aim to reach their antipodal positions), two-way and four-way hall- ways (the environment is divided by two or four building blocks and agents move along the regulated ways). The con- figuration of each scenario and the details of training can be found in the supplementary. We considered five different prior velocity predictors: Gaussian Process (GP), Neural Network (NN), a linear com- bination of the NN and GP (LinComb GP+NN), GP-driven NN (GP-fed-NN), and GP-driven ORCA (GP-fed-ORCA). For GP, NN, and LinComb GP+NN, we used the training split of the trajectories to train the data-driven priors we described in Eq. 1, Eq. 3, and Eq. 8, respectively. For GP- fed-NN, we used the outputs of the trained GP (the velocity (mean) and the variance) as two additional input branches to a neural network model, besides the local observation. For GP-fed-ORCA, we used the trained GP's velocity (mean) as the preferred velocity for ORCA (van den Berg et al. 2011), a common local collision avoiding framework. When training the models, GP was trained within each scenario while the NN was trained across the scenarios, due to dis- tinct model complexities. Experimental Setup. The ground truth trajectories are obtained by running SteerSuite (Singh et al. 2009a) library with social force AI (Helbing and Moln´ar 1995), and are split into a training and testing sets. In testing set, each tra- jectory contains a challenging missing segment around 30% points, which are initially inferred with linear interpolation. See the illustration in Fig.2. We consider two evaluation strategies. In the first basic strategy, we evaluate our trained models on the test sets from the same agent density setting. Namely, both the training and the test scenarios contain the matching (identical) number (a) bottleneck evacuation (b) bottleneck evacuation 2 (c) bottleneck squeeze (d) concentric circles (e) hallway-two-way (f) hallway-four-way Figure 3: DTW measurements over 6 scenarios using UKS optimization for extended evaluation. Horizontal axis de- notes the number of agents (density). Bounding box indi- cates the case where training and test densities are the same. which means this metric is strict. Note that for simplic- ity a collision does not change the velocities of involved agents. We also measure the time-to-completion as a proxy for the computational complexity of each approach. Indices are measured after the 5th optimization loop. Experimental Results. Results of experiments in the first, matching-agent-density setting, are summarized in Ta- ble 1, 2, 3. The average rankings of different methods can be used to ascertain relative performance and are presented in Table 4. Table 5 shows the computational time of dif- ferent evaluated approaches. Finally, results of evaluations across different train-test agent densities are shown in Fig- ures 3, 4 and 5. We only show evaluations for one of the optimization approaches, the UKS, for brevity and because other approaches follow similar trends. Discussion For the basic evaluations where the agent density of the test set matches that of the training set, we can see from Ta- ble 1, 2 and 3 that (i) for DTW score, GP-fed-NN, GP and GP-fed-ORCA perform better than other priors. This illus- trates that GP is important for regulating the trajectory to roughly follow the movement (flow) pattern. (ii) In terms of the number of agent-agent collisions, GP-fed-NN, LinComb GP+NN perform best, followed by NN. This demonstrates Figure 2: Visualization: the first two rows show the bot- tleneck evacuation and bottleneck squeeze scenario respec- tively. In these two rows the left figure denotes the initial lin- ear interpolation and the right figure denotes the optimized trajectories, shown in yellow portions and zoomed in corre- spondingly in the third row. Best viewed by zooming in. of agents. In the second extended strategy, we consider test scenarios where the agent density varies compared to train- ing setting to evaluate the generalization ability of our ap- proaches. Details of these evaluations are described below. The following methods are evaluated: explicit colli- sion avoidance local optimization using message-passing ADMM (MPA) (Yoon et al. 2016) that includes pair-wise constraints, our framework with various prior velocity pre- dictor settings (GP, NN, LinComb GP+NN, GP-fed-NN, and GP-fed-ORCA), while the optimizer is either IPM or UKS. We set the parameter as: ut = 1 if the point is actually observed, otherwise ut = 0; Ckn = 1, Cmv = 2.6m/s, ∆t = 1.5s and λ = 1/(σ2 N N ∆t2) ≈ 108.0. We employ three evaluation scores. The similarity be- tween the ground truth trajectory and reconstructed tra- jectory is measured with dynamic time warping distance (DTW). We also measure the number of agent-agent col- lisions and agent-obstacle collisions. An collision occurs when the distance between the centers of two agents is strictly less than the sum of their radii during their con- tinuous movements, and it could be checked by solving a quadratic equation provided with locations of two agents at consecutive time points. The number of collisions is accu- mulated by counting all collisions along every time step, Table 1: Missing 30% of frames, Average Relative DTW Distance in Percentage for basic evaluation. Red color indicates the best method in each scenario, followed by the second best in blue. Best viewed in color. Scenario bottleneck-evacuation bottleneck-evacuation-2 bottleneck-squeeze concentric-circles hallway-two-way hallway-four-way MPA 22.2 20.9 19.2 47.2 16.9 18.0 NN IPM UKS 18.1 11.4 16.3 18.5 4.4 3.7 14.8 42.5 13.2 17.4 25.0 17.3 GP-fed-NN IPM UKS 12.0 9.2 12.7 14.4 2.3 3.1 11.2 46.2 11.2 13.7 15.5 14.9 LinComb GP+NN UKS 10.4 13.1 2.8 36.6 13.5 14.1 IPM 10.2 11.4 2.6 17.2 14.3 15.0 GP IPM UKS 9.0 9.1 9.3 11.1 3.1 3.1 35.8 17.1 14.1 15.3 12.0 12.6 GP-fed-ORCA IPM 9.0 12.3 3.3 37.3 14.8 12.9 UKS 9.3 14.1 3.1 34.0 12.6 12.3 Table 2: Missing 30% of frames, Agent-Agent Collisions for basic evaluation. Red color indicates the best method in each scenario, followed by the second best in blue. Best viewed in color. Scenario bottleneck-evacuation bottleneck-evacuation-2 bottleneck-squeeze concentric-circles hallway-two-way hallway-four-way MPA 10.6 147.0 132.0 63.4 53.6 35.8 NN IPM UKS 15.4 16.2 140.4 159.8 46.6 49.0 18.4 52.4 29.6 30.4 13.6 13.6 GP-fed-NN IPM UKS 17.2 15.0 132.4 157.0 47.0 47.4 19.6 71.0 27.8 29.0 12.2 13.0 LinComb GP+NN UKS 12.6 140.8 47.0 44.0 27.6 12.4 IPM 14.6 139.4 46.6 26.0 29.8 10.6 GP IPM UKS 13.8 14.2 139.6 144.8 47.0 47.0 26.0 44.0 29.6 29.8 13.6 13.4 GP-fed-ORCA IPM 16.0 158.6 47.4 67.0 29.0 15.0 UKS 14.2 143.4 47.4 54.4 30.0 15.4 Table 3: Missing 30% of frames, Average-Obstacle Collisions for basic evaluation. Red color indicates the best method in each scenario, followed by the second best in blue. Best viewed in color. Scenario bottleneck-evacuation bottleneck-evacuation-2 bottleneck-squeeze hallway-two-way hallway-four-way MPA 39.2 46.2 28.0 0.0 0.0 NN IPM UKS 61 52.4 83.2 75.4 32.6 30.8 0.4 0.4 0.0 0.0 GP-fed-NN IPM UKS 60.2 45.4 88.0 76.4 7.8 28.8 0.4 0.4 0.0 0.0 LinComb GP+NN UKS 54.6 87.6 18.8 0.4 0.0 IPM 46.6 88.8 8.2 0.4 0.0 GP IPM UKS 37.2 46.2 70.4 73.8 6.6 18.6 0.4 0.4 1.8 0.0 GP-fed-ORCA IPM 39.4 89.4 13.2 0.4 3.2 UKS 49.4 81.4 16.8 0.4 0.4 Table 4: Average ranks of different combinations over scenarios for basic evaluation. Lower rank is better. Red color indicates the best method for each evaluation approach, followed by the second best in blue. Best viewed in color. Evaluation DTW Agent-Agent Agent-Obstacle MPA 10.67 8.50 3.30 NN IPM UKS 8.50 8.33 8.17 5.50 8.00 6.60 GP-fed-NN IPM UKS 6.67 4.00 5.67 5.17 7.80 4.40 LinComb GP+NN UKS 5.33 3.50 7.00 IPM 4.83 3.58 6.00 GP IPM UKS 3.75 3.67 5.67 4.41 4.10 5.00 GP-fed-ORCA IPM 5.58 8.25 7.10 UKS 4.67 7.58 6.70 Table 5: Computational time, in seconds, for basic evaluation. Note that in this table, the time of IPM and UKS is sequentially accumulated over trajectories, while in practice, trajectories could be optimized in parallel when using IPM and UKS. Scenario bottleneck-evacuation bottleneck-evacuation-2 bottleneck-squeeze concentric-circles hallway-two-way hallway-four-way MPA 1343.6 521.1 738.8 29.5 876.3 1135.9 NN IPM UKS 304.4 887.9 201.4 290.1 283.9 746.8 143.8 173.0 841.5 293.9 319.1 901.9 GP-fed-NN IPM UKS 352.7 887.5 233.8 321.2 329.6 769.3 162.0 191.6 876.4 342.6 369.5 941.4 LinComb GP+NN UKS 337.9 222.4 313.4 148.5 324.9 350.6 IPM 922.2 308.7 637.5 178.0 885.8 944.4 GP IPM UKS 353.4 881.1 234.7 320.0 331.2 615.7 162.2 192.1 880.7 343.7 368.8 944.4 GP-fed-ORCA IPM 681.1 174.7 475.5 85.2 694.7 729.7 UKS 144.4 87.7 136.4 56.3 143.2 147.2 (a) bottleneck evacuation (b) bottleneck evacuation 2 (a) bottleneck evacuation (b) bottleneck evacuation 2 (c) bottleneck squeeze (d) concentric circles (c) bottleneck squeeze (d) hallway-two-way (e) hallway-two-way (f) hallway-four-way Figure 4: Number of agent-agent collisions over 6 scenarios using UKS optimization for extended evaluation. See Fig. 3 for additional caption details. that pure GP alone is not sufficient to guarantee the desired collision avoidance, and has to be augmented by another pol- icy learner like NN to achieve the least number of agent- agent collisions. Surprisingly, for this metric GP-fed-ORCA does not give satisfactory results, which might stem from our complex training/testing datasets, containing densely packed agents (especially for bottleneck-evacuation-2 sce- nario) such that the permitted set of ORCA is prone to be empty under such agent density. (iii) For the number of agent-obstacle collisions, MPA and GP perform better than others, indicating the importance of the global flow patterns in order to avoid stationary environmental obstacles. Tab. 4 suggests that, as the optimization approach, IPM is slightly better than UKS and MPA for DTW; on the other hand, UKS achieves the fewest agent-agent collisions, lower than IPM and MPA. MPA yields the fewer agent-obstacle collisions than IPM and UKS. However, in terms of com- putational complexity, Tab. 5 shows that MPA is the most expensive approach since it includes pair-wise constraint. In addition, UKS is cheaper than IPM, in that UKS exploits fast forward-backward computations specific to the trajec- tory domain. The lack of conformability in the above obser- vations implies that there might not exist a dominant opti- mization algorithm and one might need to trade off when choosing the optimizations. Figure 5: Number of agent-obstacle collisions over 4 sce- narios using UKS optimization for extended evaluation. See Fig. 3 for additional caption details. Evaluation across different agent densities, shown in Fig 3, 4 and 5, provides an insight into the important general- ization ability and robustness of proposed approaches. We can see that (i) for DTW, GP-fed-ORCA performs best ex- cept for concentric-circle scenario, where NN presents least DTW distance with respect to the ground truth test trajecto- ries. The reason might be that in concentric-circle scenario, agents are symmetrically placed along a circle, leading to observation patterns similar across all agents and, hence, lo- cal NN policy patterns that can be learned by reusing (shar- ing) the data across agents. On the other hand, global GP fails to reuse the data, leading to a weaker generalization model. (ii) There is a general increasing monotonic trend in the number of collisions across densities. This is an intu- itive outcome, however without a clear winner in terms of collisions except for concentric-circle scenario. For this sce- nario, again, the ability to reuse data across agents, due to symmetry, may lead to better generalization of the local NN approach. (iii) Models in the matching density settings out- perform those in the mismatched density settings in terms of the agent-agent collisions only for the concentric-circle scenario, while in terms of the agent-obstacle collisions this happens for the bottleneck-squeeze scenario. These results suggest that varying density impact the performance of dif- ferent models in a reasonable and predictable manner. Overall, the above observations emphasize the importance of the global flow priors, embodied in the GP model, in our multi-agent trajectory optimization framework. Local colli- sion avoidance priors, manifested through the NN model, have lesser than expected yet still measurable impact. We plan to further investigate these factors as well as the gener- alization across environments and considerations of exploit- ing advanced trajectory learning algorithms where the dy- namics of the trajectory and the cost of making an inference could be implicitly learned, in our future work. Supplemental Materials In this supplementary material, we describe details on the optimization methods we used in the main manuscript and provide some additional details on the experiments. We use the same notation that was used in the main manuscript. Global Objective gt(xioi) = Ei We define our global objective function as a combination of local and global priors we described in the main manuscript. First, we summarize the local and global prior terms we pre- sented in the main manuscript. t − oi t(cid:107)xi t(cid:107)2 (cid:88) ui t − xi (cid:107)xi t−1(cid:107)2 t(cid:107) ≤ Cmv if (cid:107)vi (cid:13)(cid:13)xi (cid:88) (cid:26)0 (cid:88) Ei mv(xi) = dg(xioi) = λ Ei ∞ otherwise t − xi t−1 kn(xi) = Ckn ,∀t = 1..T (S1) (S2) (S3) Ei t t (cid:0)oi (cid:0)oi (cid:1)(cid:1)(cid:13)(cid:13)2 (S4) −∆t · fN N N N ∆t2). Combining all unary prior terms, N N , µGP t, θGP where λ = 1/(σ2 we obtain the final global objective: t (cid:88) (cid:8)Ei X = arg min X i gt(xioi) + Ei kn(xi) dg(xioi)(cid:9) . +Ei mv(xi) + Ei (S5) Optimization of the Global Objective In the following, we introduce two optimization approaches to solve Eq. S5. In general, our optimization framework is iterative in nature, like the algorithm described in Alg. 1 in the main manuscript with different options for optimiz- ing Eq. S5. First, we explain the message passing algo- rithm (MPA) and then we consider the unscented Kalman smoother (UKS). Message Passing Algorithm (MPA) The first optimization method, message-passing (Bento et al. 2013), can be utilized to find X. In this algorithm, we con- sider each energy term that minimizes each point in a trajec- tory as an independent minimizer node, and the minimizer nodes are connected to one another via another type of nodes called equality nodes, building a bipartite graph. Minimizer nodes will minimize each variable, while equality nodes make sure that all minimizer nodes are up-to-date about de- pendent variables optimized by other minimizer nodes. Mes- sages here will be agent location in different time step. The prior work (Yoon et al. 2016) had to link inter-agent min- imizer nodes to deal with Ei,j p . However, we only need to connect nodes within the same agent this time. More specifically, we first define messages from the equality nodes to the minimizer nodes as ni t for i-th agent at time step t. This message conveys the expected location of the agent predicted by the other agents in the previous iteration of the optimization. Then we denote xi t as the opti- mized trajectory variable stored in the equality nodes, and xi t as the corresponding optimized trajectory variable stored in the minimizer node. A general strategy here, is to encode lo- t to be equal across all minimizers that are cally optimized xi referencing xi t estimates, and penalize the error when performing the optimization in the minimizer nodes. t by iteratively passing errors in xi Then, how can we optimize each minimizer node? For toi t), the term is a simple squared equation and Ei gt(xi since it is not dependent on any other nodes, we can solve it in closed form for the quadratic equation. For t), they are dependent on two variables, so Ekn(xi we need two squared penalty terms to optimize, e.g., in case of Ekn(·), we need to solve t), Emv(xi (cid:34) (cid:88) t−1 − ni t kn(·) = Ei arg min t−1,xi xi t Ckn (cid:107)xi t − xi t−1(cid:107)2 (cid:107)xi + ρ 2 t−1(cid:107)2 + (cid:107)xi t − ni ρ 2 t(cid:107)2(cid:105) , (S6) where ρ is the penalty weight. Original algorithm (Bento et al. 2013) uses dynamically changing weight value, but one can consider this as a fixed parameter, reducing the algo- rithm to the standard alternating direction of method of mul- tipliers (ADMM). To solve this, we need to take first order derivative and solve for the variables xi t. In case of Emv(·), we can utilize KKT conditions to find the closed form solution, as explained in (Bento et al. 2013). It is worth dg(·), from each minimizer's perspective, noting that, in Ei t are constant in terms of all other terms except the xi optimization, so the minimizer solution should be identical to the Ekn(·) with minor difference in coefficients. t−1, xi t−1, xi Specifically, Eq. S4 can be optimized as t−1 − F(cid:13)(cid:13)2 (cid:88) (cid:13)(cid:13)xi t − xi t−1 − ni t ρ 2 + (cid:107)xi t−1(cid:107)2 + (cid:107)xi t − ni t(cid:107)2 (S7) ρ 2 dg(xioi) = λ Ei where λ = 1/(σ2 constant term. Taking derivative for each t-th term, DN N ∆t2) and F to denote the nonlinear ∂Ei ∂Ei toi) dg(xi ∂xi t−1 toi) dg(xi ∂xi t = −2λ(xi t − xi t−1 − F ) + ρ(xi t−1 − ni t−1) (S8) = 2λ(xi t − xi t−1 − F ) + ρ(xi t − ni t), (S9) Then, by setting the above as zero and solving for xi t and t−1, we get xi 0 = −2λ(xi 0 = 2λ(xi t − xi t − xi (2λ + ρ)ρni t−1 − F ) + ρ(xi t−1 − F ) + ρ(xi t−1 + 2λρni t−1 − ni t − ni t), t + 2λF (2λ + ρ − 1) t−1), (S10) (S11) (4λ + ρ)ρ (S12) (2λ + ρ)ρni t + 2λρni t−1 + 2λF (2λ + ρ + 1) (4λ + ρ)ρ (S13) xi t−1 = xi t = Unscented Kalman Smoother (UKS) In this subsection, we approximate the constrained optimiza- tion problem Eq. S5 with a dynamical system, and solve it efficiently using unscented Kalman smoother. First, consider the following minimization problem: J(x) = 1 2 (x − a)(cid:62)A(x − a) + 1 2 s.t. (cid:107)v(cid:107) ≤ Cmv where (x − b)(cid:62)B(x − b) (S14) (S15) x = xt+1 a = xt A = 2CknI b = xt + ∆t · fN N−GP (ot) B = 2λI. (S16) (S17) (S18) (S19) (S20) Note that we suppressed the superscript i since we will op- timize each agent's trajectory independently. Substituting Eq. S16-Eq.S20 back into Eq. S14, it is clear that J(xt+1) = Cknxt+1 − xt2 + λxt+1 − xt − ∆t · fN N−GP (ot)2, (S21) which is the summation of the kinetic energy term and the newly introduce prior energy term. If we ignore the veloc- (cid:107)v(cid:107) ≤ Cmv, setting the first derivative of ity constraint Eq. S14 to zero immediately gives the solution of the mini- mization problem x = (A + B)−1(Aa + Bb). Substituting Eq. S16-Eq.S20 back into the solution, we get: xt+1 = xt + ∆t · λ · fN N−GP (ot) λ + Ckn ≈ xt + ∆t · fN N−GP (ot) + = T (xt), (S22) prior velocity model for a time period ∆t, with error . The error measures the uncertainty of the prior velocity model about the unknown, true velocity. Thus the error could be modeled as a Gaussian noise. In this way, Eq. S22 implies a transition function from xt to xt+1 of a dynamical system, and denotes the process noise, which is Gaussian. Next, let us consider the constraint of the above minimiza- tion problem. The velocity in this case is 'noisy', expressed as v = fN N−GP (ot) + 1 ∆t . One can show that, if we di- rectly pass the 2-dimensional velocity v to a hard magnitude limiter (cid:26)v f (v) = Cmv if (cid:107)v(cid:107) ≤ Cmv otherwise , (S24) the dynamical system will become explosively complex along transitions. Instead, we apply the magnitude limiter to each of the velocity component as vx < Cmv, vy < Cmv with a soft magnitude limiter: f (v) = 2Cmv · (s(v) − 1/2) where v is either vx or vy and s(·) the sigmoid function. The key point is that we do not directly compute f (v) given v. Instead, first-order Taylor approximation is consid- ered: f (v) ≈ f (a) + f(cid:48)(a)(v − a) (S25) where a is a constant (a is a moving point such that it is close to the current v) given v. Therefore, f (v) is a linear transfor- mation of v, hence Gaussian if v is Gaussian. This recursive property makes the dynamical system always a Gaussian, avoiding being explosively complex along transitions. Thus, if we use [f (vx), f (vy)]T rather than v to conduct the transi- tion in Eq. S22, the maximum-velocity constraint is applied. Furthermore, if the measurement function of the dynami- cal system is defined as: ot = xt + rt = H(xt), (S26) where the Gaussian measurement noise rt adaptively con- trols the tracker output energy term Egt(xt, ot) = utxt − ot2 = utrt2. For instance, we know that which points are visible and which points are missing given the incom- plete trajectories. If ut = 1, denoting that ot is an actually observed point, the covariance of rt could be set as the iden- tity matrix scaled by a small number, indicating high confi- dence and hence low energy. On the other hand, if ut = 0, denoting that ot is only from inference, the covariance of rt would be large when estimated from training data, indicat- ing low confidence and high energy. Overall, applying unscented Kalman filter followed by Rauch-Tung-Striebel smoother (Rauch, Striebel, and Tung 1965) that calculates "smoothed" sequence from the given filter output optimizes the approximated objective accumu- lated over all time steps. where ∼ N (0, (A + B)−1) = N (0, 1 2(Ckn + λ) I). (S23) The interpretation of Eq. S22 is that the true location of an agent xt+1 at time t + 1, starting from the true location xt at time t, is determined by the average velocity from a Experiments To evaluate the proposed framework for various combina- tions of priors and optimization methods, we prepare sim- ilar experimental settings as in (Yoon et al. 2016). We prepare 6 scenarios (environmental configurations): 3 dif- ferent settings of bottlenecks (each contains a challenging Table S1: List of scenarios we have tested and the envi- ronment width and height in each scenario (in meters), and number of frames subsampled. Scenario bottleneck-evacuation bottleneck-evacuation2 bottleneck-squeeze concentric-circles hallway-two-way hallway-four-way Width × Height 200 × 160 100 × 80 200 × 200 20 × 20 200 × 200 200 × 200 #frames 101 53 94 37 101 101 and use the same 1,000 spatio-temporal points to fit the sec- ond GP regression model for the y component of the ve- locity. We treat the two velocity components as two inde- pendent Gaussian processes and the final velocity is simply their superposition. We conduct data normalization and auto relevance determination, and choose the Matern 5/2 kernel. Note that even though there are distinct destinations of tra- jectories in the two-way hallway scenario and in the four- way hallway, one set of GPs are sufficient to model multiple modalities of movements in a single scenario. NN regressor is trained using around 800,000 samples across all 6 scenarios, and tested in test simulations involv- ing around 130,000 test samples. There are several branches in NN and all of them are fully connected followed by ReLU activations and dropout layers. If we adopt the linear com- bination of NN and GP, there are three input branches for NN: 360 dimensional distance map, 360 × 2 dimensional velocity map and 2 dimensional global desired velocity. For GP-fed NN, there are two additional branches for NN: 2 di- mensional GP local guidance velocity and 2 dimensional GP standard deviation. These branches merge at a higher level (at the 6-th layer), which is still fully connected, with ReLU activation and dropout layers. The output layer is a 2 dimen- sional dense layer for velocity regression. The width of the network is 1,024 while the depth is 10. Mean squared error is the loss function and RMSprop is the learner. With early stop, we set learning rate to be 0.0001, batch size 64 and dropout rate 0.2. In addition, similar to the work in (Long, Liu, and Pan 2017), we also apply a clamping to regulate the output of the NN. References A.; 2014. Ra- [Alahi, Ramanathan, and Fei-Fei 2014] Alahi, Socially-aware manathan, V.; and Fei-Fei, L. In Computer Vision and large-scale crowd forecasting. Pattern Recognition (CVPR), 2014 IEEE Conference on, 2211–2218. [Ali and Shah 2008] Ali, S., and Shah, M. 2008. Floor Fields for Tracking in High Density Crowd Scenes. In ECCV. [Ali et al. 2013] Ali, S.; Nishino, K.; Manocha, D.; and Shah, M. 2013. Modeling, Simulation and Visual Analysis of Crowds: A Multidisciplinary Perspective. Springer Publish- ing Company, Incorporated. [Bento et al. 2013] Bento, J.; Derbinsky, N.; Alonso-Mora, Figure S1: The optimized trajectories in concentric-circles scenario in (top) 2D view and (bottom) normalized time axis. Note the rotating behavior of agents in this scenario. exit to go through for evacuation), concentric circle (agents are symmetrically placed along a circle and aim to reach their antipodal positions), two-way and four-way hallways (the environment is divided by two or four building blocks and agents move along the regulated ways). These sce- narios were previously introduced in (Singh et al. 2009b) and are representative settings for studying crowd behav- iors. For each scenario, we simulate 3,000 frames of 30- 40 agents except for the concentric circle scenario, where there are 20 agents, followed by sub-sampling those frames. Detailed statistics of the scenarios are summarized in Ta- ble S1. An example visualization of the scenarios is shown in Fig. S1. The ground truth trajectories were obtained by randomly setting the agents' initial locations and running SteerSuite (Singh et al. 2009a) library driven by social force AI (Helbing and Moln´ar 1995). Then the trajectories were split into training set and testing set with 6:1 ratio. GP regressor is trained per-scenario. More specifically, for each scenario, we use 1,000 spatio-temporal points which are three dimensional denoted by (x, y, t) to fit one GP regression model for the x component of the velocity, stochastic pedestrian motion model. In International Work- shop on Socially Intelligent Surveillance and Monitoring (SISM). [Rauch, Striebel, and Tung 1965] Rauch, H.; Striebel, C.; and Tung, F. 1965. Maximum likelihood estimates of linear dynamic systems. AIAA Journal 3(8):1445–1450. [Rodriguez, Ali, and Kanade 2009] Rodriguez, M.; Ali, S.; and Kanade, T. 2009. Tracking in Unstructured Crowded Scenes. In ICCV. [Rodriguez et al. 2011] Rodriguez, M.; Sivic, J.; Laptev, I.; and Audibert, J.-Y. 2011. Data-driven Crowd Analysis in Videos. In IEEE International Conference on Computer Vi- sion. [Sharma, Huang, and Nevatia 2012] Sharma, P.; Huang, C.; and Nevatia, R. 2012. Unsupervised incremental learning In 2012 IEEE for improved object detection in a video. Conference on Computer Vision and Pattern Recognition, 3298–3305. [Singh et al. 2009a] Singh, S.; Kapadia, M.; Faloutsos, P.; and Reinman, G. 2009a. An Open Framework for Develop- ing, Evaluating, and Sharing Steering Algorithms. In Pro- ceedings of the 2nd International Workshop on Motion in Games, MIG '09, 158–169. Berlin, Heidelberg: Springer- Verlag. [Singh et al. 2009b] Singh, S.; Kapadia, M.; Faloutsos, P.; and Reinman, G. 2009b. SteerBench: a benchmark suite for evaluating steering behaviors. Computer Animation and Virtual Worlds 9999(9999):n/a+. [van den Berg et al. 2011] van den Berg, J.; Guy, S. J.; Lin, M. C.; and Manocha, D. 2011. Reciprocal n-Body Col- In International Symposium on Robotic lision Avoidance. Research (ISRR). [van den Berg, Lin, and Manocha 2008] van den Berg, J.; Lin, M. C.; and Manocha, D. 2008. Reciprocal Velocity Obstacles for Real-Time Multi-Agent Navigation. In Pro- ceedings of the IEEE International Conference on Robotics and Automation (ICRA). [Yoon et al. 2016] Yoon, S.; Kapadia, M.; Sahu, P.; and Pavlovic, V. 2016. Filling in the blanks: Reconstructing mi- croscopic crowd motion from multiple disparate noisy sen- sors. In 2016 IEEE Winter Applications of Computer Vision Workshops, WACV 2016 Workshops, 1–9. 25% contribution. [Zhan et al. 2008] Zhan, B.; Monekosso, D.; Remagnino, P.; Velastin, S.; and Xu, L.-Q. 2008. Crowd analysis: a survey. Machine Vision Applications 19:345?–357. J.; and Yedidia, J. S. 2013. A message-passing algorithm for multi-agent trajectory planning. In Burges, C. J. C.; Bottou, L.; Welling, M.; Ghahramani, Z.; and Weinberger, K. Q., eds., Advances in Neural Information Processing Systems 26. Curran Associates, Inc. 521–529. [Bera and Manocha 2014] Bera, A., and Manocha, D. 2014. Realtime Multilevel Crowd Tracking using Reciprocal Ve- locity Obstacles. In 22nd International Conference on Pat- tern Recognition. [Bera, Kim, and Manocha 2015] Bera, A.; Kim, S.; and Manocha, D. 2015. Efficient Trajectory Extraction and Parameter Learning for Data-Driven Crowd Simulation. In Graphics Interface. [Bera, Kim, and Manocha 2016] Bera, A.; Kim, S.; and Manocha, D. 2016. Online parameter learning for data- driven crowd simulation and content generation. Computers and Graphics. [Helbing and Moln´ar 1995] Helbing, D., and Moln´ar, P. 1995. Social force model for pedestrian dynamics. Phys. Rev. E 51:4282–4286. [Hu, Ali, and Shah 2008] Hu, M.; Ali, S.; and Shah, M. 2008. Learning Motion Patterns in Crowded Scenes Using Motion Flow Field. In ICPR. [Jacques Junior, Raupp Musse, and Jung 2010] Jacques Ju- nior, J.; Raupp Musse, S.; and Jung, C. 2010. Crowd anal- ysis using computer vision techniques. Signal Processing Magazine, IEEE 27(5):66–77. M.; [Kapadia, Pelechano, and Allbeck 2015] Kapadia, Pelechano, N.; and Allbeck, J. 2015. Virtual Crowds: Steps Toward Behavioral Realism. MORGAN & CLAYPOOL. [Kim, Lee, and Essa 2011] Kim, K.; Lee, D.; and Essa, I. 2011. Gaussian Process Regression Flow for Analysis of Motion Trajectories. In IEEE International Conference on Computer Vision, (ICCV). [Lerner, Chrysanthou, and Lischinski 2007] Lerner, Chrysanthou, Y.; and Lischinski, D. example. Comput. Graph. Forum 26(3):655–664. [Li et al. 2015] Li, T.; Chang, H.; Wang, M.; Ni, B.; Hong, R.; and Yan, S. 2015. Crowded scene analysis: A survey. IEEE Transactions on Circuits and Systems for Video Tech- nology 25(3):367–386. [Long, Liu, and Pan 2017] Long, P.; Liu, W.; and Pan, J. 2017. Deep-learned collision avoidance policy for dis- tributed multiagent navigation. IEEE Robotics and Automa- tion Letters 2(2):656–663. [Pellegrini and Gool 2013] Pellegrini, S., and Gool, L. V. 2013. Tracking with a mixed continuous-discrete condi- tional random field. Computer Vision and Image Under- standing. [Pellegrini et al. 2009] Pellegrini, S.; Ess, A.; Schindler, K.; and van Gool, L. 2009. You'll Never Walk Alone: Mod- In IEEE eling Social Behavior for Multi-target Tracking. International Conference on Computer Vision. [Pellegrini et al. 2010] Pellegrini, S.; Ess, A.; Tanaskovic, M.; and Gool, L. V. 2010. Wrong turn - no dead end: a A.; 2007. Crowds by
1702.07934	1	1702	2017-02-25T18:10:37	A decentralized algorithm for control of autonomous agents coupled by feasibility constraints	[ "cs.MA", "eess.SY" ]	In this paper a decentralized control algorithm for systems composed of $N$ dynamically decoupled agents, coupled by feasibility constraints, is presented. The control problem is divided into $N$ optimal control sub-problems and a communication scheme is proposed to decouple computations. The derivative of the solution of each sub-problem is used to approximate the evolution of the system allowing the algorithm to decentralize and parallelize computations. The effectiveness of the proposed algorithm is shown through simulations in a cooperative driving scenario.	cs.MA	cs	A decentralized algorithm for control of autonomous agents coupled by feasibility constraints Ugo Rosolia, Francesco Braghin, Andrew G. Alleyne, Stijn De Bruyne and Edoardo Sabbioni 7 1 0 2 b e F 5 2 ] A M . s c [ 1 v 4 3 9 7 0 . 2 0 7 1 : v i X r a Abstract- In this paper a decentralized control algorithm for systems composed of N dynamically decoupled agents, coupled by feasibility constraints, is presented. The control problem is divided into N optimal control sub-problems and a communication scheme is proposed to decouple computations. The derivative of the solution of each sub-problem is used to approximate the evolution of the system allowing the algorithm to decentralize and parallelize computations. The effectiveness of the proposed algorithm is shown through simulations in a cooperative driving scenario. I. INTRODUCTION In the last decade researchers have focused on automa- tion in several application fields and in the near future autonomous mechatronic systems will be part of our ev- eryday lives. Under this scenario, the need for cooperative control algorithms able to manage the interactions among autonomous agents is increasing. Despite the advance in computational power allowing for solving complex tasks for a single autonomous agent in real time, it is far more challenging to control the interaction among autonomous agents [1]. Indeed, when two or more agents have to interact, there could be communication limitations or the dimension of the problem could increase exponentially, and consequently the computational burden. In this paper we focus on dynamically decoupled systems subjected to coupling constraints. This could be the case for UAV flight formation, air traffic control, power management and several other applications [1]–[10]. Early works in the field did not explicitly take into account the coupling con- straints [11], [12]. For example in UAV flight control, the collision avoidance constraints are usually enforced using barrier functions, which do not guaranty safety. In [8] a decentralized control strategy able to take into account hard constraints was proposed. However, the problem is solved sequentially and each decentralized optimization has to wait until the previous one is completed. Thus, for large scale systems, this approach could prove infeasible for real-time. In order to overcome this issue, the authors in [9] proposed a strategy to parallelize computations. The problem is divided U. Rosolia is with the Department of Mechanical Engineering, University of California at Berkeley , Berkeley, CA 94701, USA ugo.rosolia@ berkeley.edu. Braghin Francesco are with Mechanical Engineering Department, Politecnico di Milano, {francesco.braghin, edoardo.sabbioni} @polimi.it. the Italy and Edoardo Sabbioni A. G. Alleyne is with the Mechanical Science and Engineering, IL 61801 USA Illinois at UrbanaChampaign, Urbana, University of [email protected]. S. De Bruyne is with McKinsey & Company stijn.debryune @gmail.com. into N sub-problems, which are solved in parallel when the agents are not coupled. In [13] a robust distributed MPC which allows the authors to decouple the computation is presented. A robust tube is constructed for each i-th agent and a local feedback controller is used to keep the agent into the tube. Therefore, communication between agents is only required to update the tube. This work proposes a decentralized and parallelized algo- rithm to compute a nearly optimal solution of a specific class of non-linear non-convex problems under the assumption of no delay or loss of communication. The optimization is divided into N sub-problems, similar to [9]. The main contribution of this paper is to propose a communication scheme which allows for independent computation of the solution of each sub-problem. This scheme is inspired by the GMRES\Continuation method [14]–[16], where the time evolution of a nonlinear algebraic system is traced by its derivative. The proposed scheme uses the derivative of the optimal solution to decouple the sub-problems, namely each autonomous agent approximates the behavior of the system based on its derivative. Continuations methods do not con- verge to the solution when the evolution of the system is discontinuous [14]. Unfortunately, the coupling inequality constraints introduce a discontinuity, as shown in [17]. Thus, we use a relaxed approach to deal with inequality constraints which allow us to use a continuation method. These relaxed conditions could be used also for explicit fixed time step algorithms. This paper is organized as follow: Section II the cen- tralized system is expressed as the summation of N de- centralized optimal control sub-problems. In Section III the control algorithm is described and the proposed conditions to deal with inequality constraints are derived. Section IV provides additional details on the algorithm and its range of applicability. Finally, in Section V the proposed control logic is tested on simulations of a cooperative driving scenario. Section VI provides final remarks. II. PROBLEM FORMULATION In this section the centralized control problem is intro- duced. Afterwards, the relaxation method used to guarantee continuity of the optimal solution to the control problem is described. Finally, we present the decoupling strategy. A. System description The proposed algorithm aims to compute the trajectories of a system composed by N dynamically decoupled agents. The dynamics of each agent have the following non-linear state space representation xi(t) = fi(xi(t), ui(t)), (1) with xi ∈ Rni being the state vector and ui ∈ Rnui the control action related with the i-th agent. Thus, the dynamic of the overall system can be written as X(t) = [f1(x1(t), u1(t)),··· , fN (xN (t), uN (t))]T = (2) = F (X(t), U (t)), where X(t) = [x1(t),··· , xN (t)] ∈ Rn1×···×nN is the state vector and U (t) = [u1(t),··· , uN (t)] ∈ Rnu1×···×nuN the input vector. The optimal control problem consists in the minimization of N decoupled cost functions over a moving time interval with fixed duration T : (cid:90) t0+T N(cid:88) t0 i=1 (cid:124) hi(xi(t), ui(t)) dt, (3a) (cid:123)(cid:122) (cid:125) Running Cost J∗ c (X(t0)) = inf U (t) s.t. X(t) = F (X(t), U (t)) I(X(t)) = [C1(X(t)),··· , Cnc(X(t))]T ≥ 0 (3b) (3c) = C. Relaxation method When the inequality constraint Cj(X(t)) in Equation (5) is tightly satisfied after a period where it was not, the optimal solution has a discontinuity [17]. Unfortunately, continuation methods cannot be used to compute the solution at discontinuity points [19]. Thus, continuation methods are not suitable to compute the optimal solution when optimality is described by the KKT conditions (Eq. (5)). In order to overcome this issue, we introduce a set of slack variables to convert the inequality constraints into equality constraints,  E1(X(t)) ...  = Enc(X(t))  C1(X(t)) − z2 Cnc(X(t)) − z2 ... 1(t) nc(t)  = 0. It is clear, (8) that when the equality constraints (Eq. 8) hold also the inequality feasibility constraints (Eq. (3c)) are satisfied. Moreover, this problem formulation provides for the removal of the KKT conditions (Eq. 5), which introduced a discontinuity. The relaxed optimal control problem is defined as the E(X(t)) = minimization of the cost function J(X(t), U (t)) = (cid:90) t+T t N(cid:88) (cid:124) (cid:124) i=1 hi(xi(t), ui(t) + Λ(t)E(t) + (cid:123)(cid:122) Running cost (cid:125) (cid:123)(cid:122) (cid:125) Enforcing feasibility constraint (cid:124) (cid:123)(cid:122) nc(cid:88) (cid:124) j=1 Wz zj(t) (cid:123)(cid:122) (cid:125) Slack varible effect on optimality dt (cid:125) Running cost of the equivalent relaxed unconstrained centralized problem (9) subject to the dynamic constraint (2). We underline that the effect of the slack variable is to add a safety margin which is determined by the tuning parameter Wz, and that the optimal solution of the relaxed problem does not saturates the feasibility constraints. Therefore, the solution of the relaxed problem is suboptimal for the original problem, described in Section II.B. D. Decoupling strategy The centralized control problem could be written as the summation of N Pi optimal control problems. Each Pi problem is related to the i-th agent and is defined as, (cid:90) t0+T t0 J∗ i (X(t0)) = inf ui(t) s.t Ji(X(t), ui(t))dt (10a) (10b) (10c) xi(t) = f (xi(t), ui(t)) Ji(X(t), ui(t)) = with (cid:90) t+T t (cid:88) j∈Ai + (cid:125) (cid:124) (cid:123)(cid:122) Running cost hi(xi(t), ui(t) [λj(t)Ej(X(t)) + (cid:123)(cid:122) (cid:125) Enforcing feasibility constraint (cid:123)(cid:122) Wz zj(t) (cid:124)(cid:123)(cid:122)(cid:125) Slack varible effect on optimality ]dt (cid:125) (11) Equivalent running cost of the Pi problem where the nc feasibility constraints in (3c) may couple the agents. B. The optimal control problem The feasibility constraints in Equation (3c) can be enforced through the cost function [17], [18]. Given a vector of time varying Lagrange multipliers, defined as Λ(t) = [λ1(t),··· , λj(t),··· λnc(t)] (cid:40)(cid:54)= 0 If Cj(X(t)) = 0 = 0 If Cj(X(t)) > 0 . where λj(t) = The centralized optimal control problem consists of the minimization of the augmented cost function Jo(X(t), U (t)) = hi(xi(t), ui(t) + Λ(t)I(t) dt (cid:90) t+T t N(cid:88) (cid:124) (cid:124) i=1 (cid:124) (cid:123)(cid:122) (cid:125) Enforcing feasibility constraint (cid:123)(cid:122) Running cost (cid:125) (cid:123)(cid:122) Running cost of the equivalent unconstrained problem (4) (5) (cid:125) (6) and is defined as J∗ 0 (X(t0)) = inf U (t) s.t (cid:90) t0+T t0 Jo(X(t), U (t))dt X(t) = F (X(t), U (t)) = (7a) (7b) (7c) (cid:124) (cid:124) where Ai is the set of subscripts of the inequalities Cj(X(t)) involving the i-th agent. It is clear that if at time t the global optimal solutions of N −1 agents are known, the Pi problem could be solved independently and its solution is globally optimal for the centralized relaxed problem. III. ALGORITHM In this section a variation to the GMRES\Continuation methods, which allows to parallelized and decentralize com- putations, is presented. Moreover, we suggest a numerical strategy to handle the feasibility constraints based on their effect on optimality. A. Decentralized algorithm The GMRES\Continuation method uses the derivative of the optimal solution to trace its behavior in time. For details on the numerical implementation and accuracy of continuation methods we refer to [15], [16], [20]–[22]. The proposed algorithm uses the derivative of the optimal solution to approximate the optimal trajectories of N − 1 agents, enabling independent solution of each Pi problem. To initialize the algorithm the derivative is computed with a centralized optimization method. After the initialization, the derivative is computed on-board on each i-th agent and communicated to the others. Table I illustrates the algorithms steps. It is interesting to notice that the proposed algorithm does not introduced further numerical approximation with respect to the centralized algorithm based on continuation methods, as shown in the result section. TABLE I ALGORITHM SCHEME Initialization Step 1) Step 2) Iteration k Step 3) Step 4) Step 5) Step 6) Compute the optimal solution and the optimal derivative with a centralized algorithm Communicate the optimal solution and its derivative to all the agents Each i-th agent integrates numerically the trajectories of the other agents Each i-th agent solves its Pi problem to compute the optimal solution and its derivative Each i-th agents communicates the optimal solution and its derivative to all the other agents k = k + 1 go to step 3) B. Handling coupling constraints In this section the effect of the coupling feasibility con- straints on optimality is analyzed, and the relaxed approach to deal with inequality constraints is introduced. As the solution to the relaxed problem is similar to the original one, when the feasibility inequality constraint is satisfied, the relaxed equality constraint does not influence optimality. In order to verify this statement, it is possible to compute the relationship between the slack variable and the Lagrange multiplier related to the j-th constraint. If zj(t) is optimal, the following relationship for the derivative of the equivalent running cost (Eq. 11) holds ∂((cid:80)N i=1(h(xi(t), ui(t) + Λ(t)E(t) +(cid:80)nc j=1 ∂zj(t) Combining Equation (8) and Equation (12), relation between λj(t) and the state can be written as Wz j (t) ) z2 = 0. (12) the explicit λj(t) = Wz z4 j (t) = Wz C 2 j (X(t)) . (13) Equation (13) shows that when the inequality constraints Cj(X(t)) is safely satisfied, the Lagrange multiplier related with the relaxed constraint is small in magnitude. Therefore, the effect of the related relaxed feasibility constraint on op- timality is negligible. When this condition occurs, we would like not to consider the unnecessary feasibility constraint, reducing the dimensions of the i-th optimization problem, Pi, in Step 4) of Table I. Namely, we set a threshold value, Hlim, for which the j-th Lagrange multiplier of Equation (9) is set to zero λj(t) = If λ2 0 (cid:54)= 0 If λ2 j (t) ≤ H 2 j (t) > H 2 lim lim . (14) Substituting in Equation (14) the relationship between the Lagrange multiplier and the system state (Eq. 13), a threshold value for which the relaxed feasibility constraint has to be enforced to the problem is obtained: (cid:40) (cid:40) λj(t) = If C 2 0 (cid:54)= 0 If C 2 j (X(t)) ≥ Wz Hlim j (X(t)) < Wz Hlim . (15) It is interesting to notice that these conditions (Eq. 15) are similar to the KKT conditions in Equation (5), but these are suitable to apply continuation methods and fixed time step algorithms. IV. ALGORITHM ANALYSIS A. Minimum principle properties The MGRES/Continuation method is based on the opti- mality conditions stated by the minimum principle [16], [23], [24]. The minimum principle provides necessary conditions for global optimality and it is not always sufficient to com- pute the optimal solution [18]. Therefore, it is important to analyze the algorithm to understand which class of problems could be solved with the proposed control logic. Firstly we define the difference between weak and strong minima. Given a general optimal control problem, J∗ g (x(0)) = inf u(t) g(x(t), u(t))dt (16a) s.t (16b) (16c) A trajectory x∗(t) is a weak minima if it minimizes the functional (Eq. (16a)) over all the trajectories ¯x(t) close to x(t) = f (x(t), u(t)). (cid:90) T 0 x∗(t) in the sense of the 1-norm, meaning that x∗(t) − ¯x(t)1 := max 0≤t≤T x∗(t) − ¯x(t)+ x∗(t) − ¯x(t) ≤ . (17) Conversely, a trajectory x∗(t) is a strong minima if it minimizes the functional (Eq. (16a)) over all the trajectories ¯x(t) close to x∗(t) in the sense of 0-norm, + max 0≤t≤T x ∗ (t) − ¯x(t)0 := max 0≤t≤T x ∗ (t) − ¯x(t) ≤ . (18) The optimality conditions stated by minimum principle are satisfied for strong minima and not for weak minima [18]. Therefore, the proposed algorithm is suitable to solve non- convex problems with respect to the 1-norm, if those are convex with respect to the 0-norm. For example, say that our control problem is to find the trajectory closest to zero, outside an unfeasible region as shown in Figure 1. From Figure 1 is clear that the trajectory on the left and the one the right are far in the sense of 1-norm (the two trajectories have non-infinitesimal derivatives different in sign) and for this reason those could represent two weak minima for this problem. However, the two trajectories are close in the sense of the 0-norm, thus the problem has just one strong minima which can be correctly computed with the optimality conditions of the minimum principle. C. Non-convex problem Combining the properties of the minimum principle and the continuation methods we are able to solve a particular type of non-linear non-convex problem. Indeed the algorithm is able to take non-convex decisions if the candidate trajecto- ries are close in the sense of the 0-norm. This property has a key importance in control problems where the optimization is performed on a moving time interval. In Figure 2 the domain of an optimal control problem similar to the one in Section IV.A is shown. Here the objective is to compute, on a moving time interval, the feasible trajectory closest to zero. In Figure 2 the unfeasible region is outside the opti- mization window, thus the problem is convex. When the optimization windows moves in time, as soon as it encounters the unfeasible region, the problem becomes non-convex. Indeed there are two weak minima as shown in Figure 3. Fig. 2. Optimal control problem on a time moving window. The objective is to compute the trajectory closest to zero, outside the unfeasible region. Fig. 1. Domain of the feasible trajectories of the optimal control problem. The objective is to compute the trajectory x(t) closest to zero which does not cross the unfeasible region. B. Continuation method properties The algorithm in Section IV.C is based on a continuation method, meaning that at each time instant the optimal solution is given and the algorithm computes its derivative. This derivative is used at the next time instant to approximate the optimal solution. Therefore if there are more trajectories satisfying the min- imum principle, the algorithm would compute the evolution in time of the given trajectory. However, there could be issues at bifurcation points where the optimal solution has two possible derivatives. This particular situation is discussed in the next section. Fig. 3. Representation of a bifurcation point. As soon as the unfeasible region enters the optimization window the there are two weak local minima. Figure 3 depicts bifurcation point mentioned in Section IV.B. As soon as the unfeasible region enters the optimization windows, the global optimal solution bifurcates into two local optimal solutions with respect to the 1-norm. However, as shown in Figure 3, the two trajectories are close with respect to the 0-norm, and the optimality conditions of the minimum principle allow to compute the optimal derivative for the unique strong minima. Therefore, the algorithm correctly choses the global optimal trajectory. Afterwards, when the unfeasible region is almost, com- pletely inside the optimization window, the continuation algorithm follows the trajectory which was globally optimal at the bifurcation point (Fig. 4). Fig. 4. entered the optimization window. Evolution of the optimal solution after the unfeasible region has V. RESULTS The algorithm is tested on a cooperative driving scenario, where autonomous vehicles are driving on the same roadway at different target speeds. In particular, the algorithm is used to compute the collision free-trajectories of each autonomous vehicles. The vehicles are modeled with a simplified sys- tem; this choice for the trajectory planning phase is well- established in literature [25]–[27]. It is important to note that this problem is suitable to test our algorithm as each vehicle has to take a non-convex decision during overtaking maneuvers. Moreover, we assume that no safety maneuvers are needed to guaranty the existence of the derivative re- quired in Section III.A. Simulation was performed on a Windows computer fea- turing an Intel CORE i5 processor using Matlab 2013b. In order to measure the computational time, a stand-along executable mex-function has been compiled for each agent. This function could be used on Linux PCs and experimental results are envisaged for the future. A. Comparison between decentralized and centralized ap- proach The agents in section III.A represent autonomous vehicles and are modeled using a Single Point Mass Model in a curvilinear abscissa reference frame, for more details [28], [29]. The cost function of each vehicle is designed for lane keeping at a cruise velocity: hi(xi(t), ui(t) = W1i(y1−yitarget)2 + W2i y2 i + W3i( si − Vitarget)2 + W4i θ2 i (19) where si represents the distance traveled along the roadway mid-line, and yi the lateral distance between the vehicles center of gravity and the roadway mid-line. The inputs, Vi and θi, are the velocity and the heading angle, respectively. Wji, ∀j ∈ [1,··· , 4] are the weighting parameters. More details on this curvilinear reference frame are given in [28], [29] and [30]. Finally, the feasibility constraints in Section II.C are expressed as ellipses C(si, yi, sj, yj) = (si − sj)2 2l (yi − yj)2 2w + − 1 (20) where the axes are chosen accordingly with vehicle length, l, and width, w: 4 and 2 meters, respectively. In this example, for each problem Pi the distance between the i-th agent and the j-th agent is given by zj(t). From Equation (13) and form our choice of Wz = 7, when zj(t) = 60m then λj(t) (cid:39) 0.5· 10−7. Therefore we picked the threshold Hlim = 0.5 · 10−7 so that, when the relative distance between two agents is greater that 60m, λj(t) is set to zero and the agents are decoupled. Note that 60m is the threshold distance used in commercial blind spot detection system. B. Simulation Results 1) Comparison with a centralized algorithm: In this sec- tion two simulations with the same boundary conditions are carried out. The first one uses the proposed decentral- ized algorithm and the second one uses a centralized GM- RES\Continuation algorithm. The solutions are compared to test optimality, as the solution computed with the centralized method is optimal for the relaxed problem. As shown in Figure 5 two agents are traveling on the same straight path at different target velocities; therefore the faster agent overtakes the slower one. During the overtaking maneuver the agents move sideways from the centerline, so that the overall derivative of the steering angle and lateral velocity are minimized. Coefficients and boundary conditions used in the simulation for the two agents (i = 1 and i = 2) can be found in Tables II and III. Fig. 5. Trajectories of two agents traveling on the same target path at different target speed: 30m/s the agent in red, 24m/s the agent in green. Figures 6 shows the lateral difference between the tra- jectories of the two agents computed with the proposed decentralized control algorithm and the centralized one. The maximum difference, between the trajectories computed 0100200300400500600−1.5−1−0.500.511.52X axis [m]Y axis [m] Trajectory Agent1Trajectory Agent2Reference pathTABLE II SIMULATION COEFFICIENTS W1i W2i W3i W4i Wz 0.55 0.05 9 145 7 T ∆T Hlim 2 5 · 10−7 20 TABLE III AGENTS'S TARGET VELOCITY AND LATERAL OFFSET Agent yitarget Vitarget i = 1 0.1 30 i = 2 0 24 i = 3 0 24 i = 4 i = 5 −0.1 −0.1 18 18 [units] [m] [m/s] with the centralized algorithm and the decentralized one, is 0.095m which is 5.59% of the maximum lateral displace- ment. Thus, this proposed algorithm does not introduce fur- ther approximation with respect to a centralized continuation algorithm and it is able to compute a nearly optimal solution for the relaxed problem. Finally, it is important to analyze the computational cost. The centralized control strategy takes on average 12.9ms to compute the solution, while the decentralized one just 4.2ms, as show in Figure 7. Fig. 7. decentralized algorithms. Comparison between computational cost of the centralized and In Figure 8 the trajectories of the five agents are shown. Agent1 travels at the highest cruise velocity and its starting position is the closest to the Y axis. Therefore, during the simulation it overtakes the slower agents that it encounters on the path. Fig. 6. Lateral difference between the trajectories of the agents computed with the proposed decentralized algorithm and with a centralized one. C. Communication method When the number of agents increases, a decentralized algorithm is necessary to limit the computational burden. In this section a simulation involving five agents is carried out and the computational time is analyzed. In this scenario, the proposed relaxed method to deal with inequality constraints (Eq. 15) plays a crucial role. Coefficients and boundary conditions used in the simulation for the five agents can be found in Tables II, III and IV. TABLE IV AGENTS' INITIAL CONDITIONS Agent sistart i = 1 2 i = 2 20 i = 3 50 i = 4 680 i = 5 480 [units] [m] Fig. 8. Trajectories of the five agents. Agent1 has the highest cruise velocity and it overtakes the others four agents. Agent2 and Agent3 have an intermediate cruise velocity and they overtake Agent4 and Agent5. Concluding, during the simulation there is a total of 8 overtaking maneuver. It is clear that the feasibility constraints, which couple Agent1 with the others, should be enforced to the P1 just during the overtaking maneuvers. In Figure 9, a Boolean variable with values 1 and 0 is used to indicate, respectively, if the i-th constraint is enforced or is not enforced to P1. Figure 10 shows the trajectory of Agent1, and those of the other agents when the related feasibility constraints are enforced to P1. Here it is possible to see that the relaxed constraints are correctly enforced to the problem just during the overtaking maneuvers. 0100200300400500600700−0.06−0.04−0.0200.020.040.060.080.1X axis [m]Lateral difference [m] Agent 1Agent 2010020030040050060000.0020.0040.0060.0080.010.0120.0140.0160.0180.02Computational time [s]X axis Centralized (2 Agents)Decentralized (2 Agents)010002000300040005000−1.5−1−0.500.511.5X axis [m]Y axis [m] Agent1Agent2Agent3Agent4Agent5Reference PathFig. 9. Coupling between Agent1 and the i-th agent in function of the traveled distance on the X axis. When the Boolean variable is set to one the Lagrange multiplier is different from zero. Fig. 10. Trajectory of Agent1 with its prospective view. The trajectories of the others agents are reported when the related feasibility constraint are enforced to the P1 problem, namely when the Boolean variable of Figure 9 is set to 1. D. Complete Simulation Finally the algorithm is tested in the worst case scenario, where all the feasibility constraints has to be enforced to the P1 problem (A video of the simulation can be found at http://youtu.be/wTfb5M1YH44 ). Figure 11 shows the behavior of the computational cost as a function of the Fig. 11. Trajectory of Agent1 with its prospective view. The trajectory of the others agents are reported when the related feasibility constraint are enforced to the P1 problem, namely when the Boolean variable of Figure 10 is set to 1. relaxed constraints. In particular, Figure 11 is divided in five zones, numbered from 0 to 4, to indicate the number of enforced constraints. The minimum computational cost, 3.85ms, is achieved when no constrains are enforced to the problem. Furthermore, the maximum computational cost, 6.02ms, is reached when all the four constraints are enforced to the P1 problem. Thus, the increment in computational cost, between the unconstrained problem and the one where all four constrains are enforced, is 2.17ms. This increment 01000200030004000500000.51X axis [m]Boolean VariableEnforcement of the constraint between the Agent1 and Agent2 to the P1 problem01000200030004000500000.51X axis [m]Boolean VariableEnforcement of the constraint between the Agent1 and Agent3 to the P1 problem01000200030004000500000.51X axis [m]Boolean VariableEnforcement of the constraint between the Agent1 and Agent4 to the P1 problem01000200030004000500000.51X axis [m]Boolean VariableEnforcement of the constraint between the Agent1 and Agent5 to the P1 problem010002000300040005000−1.5−1−0.500.511.5X axis [m]Y axis [m] Agent1Agent2Agent3Agent4Agent5Reference Pathis small when compared with the centralized approach which took 12.5ms to solve a problem involving two agents. VI. CONCLUSIONS In this paper a decentralized control algorithm for dynam- ically decoupled system, coupled by feasibility constraint, is presented. The algorithm, similarly to continuation methods, uses the derivative of the optimal solution to approximate the behavior of the system. This strategy allows to decouple and to parallelize computations. Moreover, a relaxed approach to deal with inequalities constraints is introduced. This approach allows one to elimi- nate the discontinuity introduced by the KKT conditions; but it is able to recognize when an inequality constraint does not influence optimality and thus should not be enforced on the problem. The algorithm has been successfully tested in simulation in a cooperative driving scenario. The control logic is able to compute a solution near the global optimal with a decen- tralized strategy. The size of the problem is reduced when the coupling between agents is not relevant, thus the com- putational burden in reduced. Finally, the computational cost of a simulation involving five coupled agents is compared with a centralized control problem involving two agents. This comparison underlines the advantage of the decentralized control strategy which took, on average, 50% less time to solve the optimal control problem, though the dimension of the problem is four times larger. VII. ACKNOWLEDGMENTS The authors thank Dr. Jacopo Guanetti from UC Berkeley from for his feedback on the manuscript presented. REFERENCES [1] A. Bemporad and D. Barcelli, "Decentralized model predictive con- trol," in Networked control systems. Springer, 2010, pp. 149–178. [2] A. Vahidi and W. Greenwell, "A decentralized model predictive control approach to power management of a fuel cell-ultracapacitor hybrid," in American Control Conference, 2007. ACC'07. IEEE, 2007, pp. 5431–5437. [3] A. Mohammadi and M. B. Menhaj, "Formation control and obstacle avoidance for nonholonomic robots using decentralized mpc," in Net- working, Sensing and Control (ICNSC), 2013 10th IEEE International Conference on. IEEE, 2013, pp. 112–117. [4] Z. Ma, I. Hiskens, and D. Callaway, "A decentralized mpc strategy for charging large populations of plug-in electric vehicles," IFAC Proceedings Volumes, vol. 44, no. 1, pp. 10 493–10 498, 2011. [5] S. Devasia, D. Iamratanakul, G. Chatterji, and G. Meyer, "Decoupled conflict-resolution procedures for decentralized air traffic control," IEEE Transactions on Intelligent Transportation Systems, vol. 12, no. 2, pp. 422–437, 2011. [6] V. R. Desaraju and J. P. How, "Decentralized path planning for multi- agent teams in complex environments using rapidly-exploring random trees," in Robotics and Automation (ICRA), 2011 IEEE International Conference on. IEEE, 2011, pp. 4956–4961. [7] D. Raimondo, L. Magni, and R. Scattolini, "Decentralized mpc of nonlinear systems: An input-to-state stability approach," International Journal of Robust and Nonlinear Control, vol. 17, no. 17, pp. 1651– 1667, 2007. [8] A. Richards and J. How, "Decentralized model predictive control of cooperating uavs," in Decision and Control, 2004. CDC. 43rd IEEE Conference on, vol. 4. IEEE, 2004, pp. 4286–4291. [9] T. Keviczky, F. Borrelli, K. Fregene, D. Godbole, and G. J. Balas, "Decentralized receding horizon control and coordination of au- tonomous vehicle formations," IEEE Transactions on Control Systems Technology, vol. 16, no. 1, pp. 19–33, 2008. [10] P. Trodden and A. Richards, "Cooperative distributed mpc of linear systems with coupled constraints," Automatica, vol. 49, no. 2, pp. 479– 487, 2013. [11] P. Ogren, M. Egerstedt, and X. Hu, "A control lyapunov function approach to multi-agent coordination," in Decision and Control, 2001. Proceedings of the 40th IEEE Conference on, vol. 2. IEEE, 2001, pp. 1150–1155. [12] A. Bicchi and L. Pallottino, "On optimal cooperative conflict res- olution for air traffic management systems," IEEE Transactions on Intelligent Transportation Systems, vol. 1, no. 4, pp. 221–231, 2000. [13] P. Trodden and A. Richards, "Distributed model predictive control of linear systems with persistent disturbances," International Journal of Control, vol. 83, no. 8, pp. 1653–1663, 2010. [14] S. L. Richter and R. A. Decarlo, "Continuation methods: Theory and applications," IEEE Transactions on Systems, Man, and Cybernetics, no. 4, pp. 459–464, 1983. [15] T. Ohtsuka, "Continuation/gmres method for fast algorithm of non- linear receding horizon control," in Decision and Control, 2000. Proceedings of the 39th IEEE Conference on, vol. 1. IEEE, 2000, pp. 766–771. [16] --, "A continuation/gmres method for fast computation of nonlinear receding horizon control," Automatica, vol. 40, no. 4, pp. 563–574, 2004. [17] A. E. Bryson, Applied optimal control: optimization, estimation and control. CRC Press, 1975. [18] D. Liberzon, Calculus of variations and optimal control theory: a concise introduction. Princeton University Press, 2012. [19] C. Kelley, "Iterative methods for linear and nonlinear equations, siam, philadelphia, 1995," MR 96d, vol. 65002. [20] Y. Saad and M. H. Schultz, "Gmres: A generalized minimal residual algorithm for solving nonsymmetric linear systems," SIAM Journal on scientific and statistical computing, vol. 7, no. 3, pp. 856–869, 1986. [21] T. Hashimoto, Y. Yoshioka, and T. Ohtsuka, "Receding horizon control with numerical solution for spatiotemporal dynamic systems," in Decision and Control (CDC), 2012 IEEE 51st Annual Conference on. IEEE, 2012, pp. 2920–2925. [22] D. A. Knoll and D. E. Keyes, "Jacobian-free newton–krylov methods: a survey of approaches and applications," Journal of Computational Physics, vol. 193, no. 2, pp. 357–397, 2004. [23] T. Hashimoto, Y. Yoshioka, and T. Ohtsuka, "Receding horizon control for nonlinear parabolic partial differential equations with boundary control inputs," in Decision and Control (CDC), 2010 49th IEEE Conference on. IEEE, 2010, pp. 6920–6925. [24] T. Ohtsuka and H. A. Fujii, "Real-time optimization algorithm for nonlinear receding-horizon control," Automatica, vol. 33, no. 6, pp. 1147–1154, 1997. [25] A. Gray, Y. Gao, T. Lin, J. K. Hedrick, H. E. Tseng, and F. Borrelli, "Predictive control for agile semi-autonomous ground vehicles using motion primitives," in American Control Conference (ACC), 2012. IEEE, 2012, pp. 4239–4244. [26] Y. Gao, T. Lin, F. Borrelli, E. Tseng, and D. Hrovat, "Predictive control of autonomous ground vehicles with obstacle avoidance on slippery roads," in ASME 2010 dynamic systems and control conference. American Society of Mechanical Engineers, 2010, pp. 265–272. [27] P. Falcone, F. Borrelli, J. Asgari, H. Tseng, and D. Hrovat, "Low complexity mpc schemes for integrated vehicle dynamics control prob- lems," in 9th international symposium on advanced vehicle control, 2008. [28] U. Rosolia, F. Braghin, A. Alleyne, and E. Sabbioni, "Nlmpc for real time path following and collision avoidance," SAE International Journal of Passenger Cars-Electronic and Electrical Systems, vol. 8, no. 2015-01-0313, pp. 401–405, 2015. [29] U. Rosolia, S. De Bruyne, and A. G. Alleyne, "Autonomous vehicle control: A nonconvex approach for obstacle avoidance," IEEE Trans- actions on Control Systems Technology, 2016. [30] A. Micaelli and C. Samson, "Trajectory tracking for unicycle-type and two-steering-wheels mobile robots," Ph.D. dissertation, INRIA, 1993.
0905.0747	1	0905	2009-05-06T06:11:29	Self-stabilizing Determinsitic Gathering	[ "cs.MA" ]	In this paper, we investigate the possibility to deterministically solve the gathering problem (GP) with weak robots (anonymous, autonomous, disoriented, deaf and dumb, and oblivious). We introduce strong multiplicity detection as the ability for the robots to detect the exact number of robots located at a given position. We show that with strong multiplicity detection, there exists a deterministic self-stabilizing algorithm solving GP for n robots if, and only if, n is odd.	cs.MA	cs	Self-stabilizing Deterministic Gathering Yoann Dieudonn´e1 Franck Petit2 1 MIS CNRS, Universit´e de Picardie Jules Verne Amiens, France 2 INRIA, LIP UMR 5668, Universit´e de Lyon / ENS Lyon, France Abstract. In this paper, we investigate the possibility to deterministi- cally solve the gathering problem (GP) with weak robots (anonymous, autonomous, disoriented, deaf and dumb, and oblivious). We introduce strong multiplicity detection as the ability for the robots to detect the exact number of robots located at a given position. We show that with strong multiplicity detection, there exists a deterministic self-stabilizing algorithm solving GP for n robots if, and only if, n is odd. Keywords: Distributed Coordination, Gathering, Mobile Robot Net- works, Self-stabilization. 1 Introduction The distributed systems considered in this paper are teams (or swarms) of mobile robots (sensors or agents). Such systems supply the ability to collect (to sense) environmental data such as temperature, sound, vibration, pressure, motion, etc. The robots use these sensory data as an input in order to act in a given (sometimes dangerous) physical environment. Numerous potential applications exist for such multi-robot systems, e.g., environmental monitoring, large-scale construction, risky area surrounding, exploration of an unknown area. All these applications involve basic cooperative tasks such as pattern formation, gathering, scatter, leader election, flocking, etc. Among the above fundamental coordination tasks, we address the gathering (or Rendez-Vous) problem. This problem can be stated as follows: robots, ini- tially located at various positions, gather at the same position in finite time and remain at this position thereafter. The difficulty to solve this problem greatly de- pends on the system settings, e.g., whether the robots can remember past events or not, their means of communication, their ability to share a global property like observable IDs, sense of direction, global coordinate, etc. For instance, assuming that the robots share a common global coordinate system or have (observable) IDs allowing to differentiate any of them, it is easy to come up with a deter- ministic distributed algorithm for that problem. Gathering turns out to be very difficult to solve with weak robots, i.e., devoid of (1) any (observable) IDs al- lowing to differentiate any of them (anonymous), (2) any central coordination mechanism or scheduler (autonomous), (3) any common coordinate mechanism or common sense of direction (disoriented), (4) means of communication allow- ing them to communicate directly, e.g., by radio frequency (deaf and dumb), and (5) any way to remember any previous observation nor computation performed in any previous step (oblivious). Every movement made by a robot is then the result of a computation having observed positions of the other robots as a only possible input. With such settings, assuming that robots are points evolving on the plane, no solution exists for the gathering problem if the system contains two robots only [19]. It is also shown in [15] that gathering can be solved only if the robots have the capability to know whether several robots are located at the same position (multiplicity detection). Note that a strong form of such an ability is that the robot are able to count the exact number of robots located at the same position. A weaker form consists in considering the detector as an abstract device able to say if any robot location contains either exactly one or more than one robot. In this paper, we investigate the possibility to deterministically solve the gathering problem with weak robots (i.e., anonymous, autonomous, disoriented, deaf and dumb, and oblivious). This problem has been extensively studied in the literature assuming various settings. For instance, the robots move either among the nodes of a graph [11,13], or in the plane [1,2,4,12,14,15,19], their visibility can be limited (visibility sensors are supposed to be accurate within a constant range, and sense nothing beyond this range) [12,17], robots are prone to faults [1,7]. In this paper, we address the stabilization aspect of the gathering problem. A deterministic system is (self-)stabilizing if, regardless of the initial states of the computing units, it is guaranteed to converge to the intended behavior in a finite number of steps [9]. To our best knowledge, all the above solutions assume that in the initial configuration, no two robots are located at the same position. So, effectively, as already noticed in [6,8], this implies that none of them is "truly" self-stabilizing -- initial configurations where robots are located at the same positions are avoided. Note that surprisingly, such an assumption prevents to initiate the system where the problem is solved, i.e., initially all the robots occupy the same position. In this paper, we study the gathering problem assuming any arbitrary ini- tial configurations, that is in which some robots can share the same positions. Clearly, assuming weak multiplicity detection (each robot location contains ei- ther exactly one or more than one robot), the problem cannot be solved de- terministically. Informally, if all the robots are at exactly two positions, then there is no way to maintain a particular position as an invariant. So, there are some executions where the system behaves as if it contains exactly two robots, leading to the impossibility result in [19]. We introduce the concept of strong multiplicity detection -- the robot are able to count the exact number of robots located at the same position. Even with such capability, the problem cannot be solved deterministically, if the number of robots is even. The proof is similar as above: If initially the robots occupy exactly two positions, then there is no way to maintain a particular position as an invariant. Again, the impossibility result in [19] holds. By contrast, we show that with an odd number of robots, the problem is solvable. Our proof is constructive, as we present and prove a deterministic algorithm for that problem. The proposed solution has the nice property of being self-stabilizing since no initial configuration is excluded. In the next section (Section 2), we describe the distributed system and the problem we consider in this paper. Our main result with its proof is given in Section 3. We conclude this paper in Section 4. Due to the lack of space, some proofs have been moved in the Annexes section. 2 Preliminaries In this section, we define the distributed system and the problem considered in this paper. 2.1 Distributed Model. We adopt the semi-synchoronous model introduced in [18], below referred to as SSM . The distributed system considered in this paper consists of n robots r1, r2, · · · , rn -- the subscripts 1, . . . , n are used for notational purpose only. Each robot ri, viewed as a point in the Euclidean plane, moves on this two-dimensional space unbounded and devoid of any landmark. It is assumed that two or more robots may simultaneously occupy the same physical location. Any robot can observe, compute and move with infinite decimal precision. The robots are equipped with sensors enabling to detect the instantaneous po- sition of the other robots in the plane. In particular, we distinguish two types of multiplicity detection : weak multiplicity detection and strong multiplicity de- tection. Definition 1 (Weak multiplicity detection). [4,10] The robots have weak multiplicity detection if, for every point p, their sensors can detect if there is no robot, there is one robot, or there are more than one robot. In the latter case, the robot might not be capable of determining the exact number of robots. Definition 2 (Strong multiplicity detection). The robots have strong mul- tiplicity detection if, for every point p, their sensors can detect the number of robots on p. Each robot has its own local coordinate system and unit measure. The robots do not agree on the orientation of the axes of their local coordinate system, nor on the unit measure. They are uniform and anonymous, i.e, they all have the same program using no local parameter (such that an identity) allowing to differentiate any of them. They communicate only by observing the position of the others and they are oblivious, i.e., none of them can remember any previous observation nor computation performed in any previous step. Time is represented as an infinite sequence of time instants 0, 1, . . . , j, . . . Let P(t) be the set of the positions in the plane occupied by the n robots at time t. For every t, P(t) is called the configuration of the distributed system in t. Given any point p, p denotes the number of robots located on p. Note that, if the robots do not have the multiplicity detection then p ≤ 1 for all the robots. P(t) expressed in the local coordinate system of any robot ri is called a view.At each time instant t, each robot ri is either active or inactive. The former means that, during the computation step (t, t + 1), using a given algorithm, ri computes in its local coordinate system a position pi(t + 1) depending only on the system configuration at t, and moves towards pi(t + 1) -- pi(t + 1) can be equal to pi(t), making the location of ri unchanged. In the latter case, ri does not perform any local computation and remains at the same position. In every single activation, the distance traveled by any robot r is bounded by σr. So, if the destination point computed by r is farther than σr, then r moves toward a point of at most σr. This distance may be different between two robots. The concurrent activation of robots is modeled by the interleaving model in which the robot activations are driven by a fair scheduler. At each instant t, the scheduler arbitrarily activates a (non empty) set of robots. Fairness means that every robot is infinitely often activated by the scheduler. 2.2 Specification The Gathering Problem (GP) is to design a distributed protocol P for n mobile robots so that the following properties are true : -- Convergence: Regardless of the initial positions of the robots on the plane, all the robots are located at the same position in finite time. -- Closure: Starting from a configuration where all the robots are located at the same position, all the robots are located at the same position thereafter. 3 Gathering with strong multiplicity detection In this section, we prove the following theorem : Theorem 1. With strong multiplicity detection, there exists a deterministic self-stabilizing algorithm solving GP for n robots if, and only if, n is odd. As mentionned in the introduction, even with strong multiplicity detection there do not exist any deterministic algorithm solving GP for an even number of robots. So, to prove Theorem 1 we first give a deterministic self-stabilizing algorithm solving GP for an odd number of robots having the strong multiplicity detection. Then, we prove the correctness of the algorithm. 3.1 Deterministic Self-stabilizing Algorithm for an odd number of robots. In this subsection, we give a deterministic self-stabilizing algorithm solving GP for an odd number of robots. We first provide particular notations, basic defi- nitions and properties that we use for symplifying the design and proofs of the protocol. Next, the protocol is presented. Notations, Basic Definitions and Properties. Given a configuration P, M axP indicates the set of all the points p such that p is maximal. In other terms, ∀pi ∈ M axP and ∀pj ∈ P, we have pi ≥ pj. M axP will be the cardinality of M axP. Remark 1. Since the robots have the strong multiplicity detection, then they are able to compute p for every point p ∈ P. In particular, all the robots can determine M axP(t) at each time instant t. Given three distinct points r, r′ and c in the plane, we say that the two half-lines [c, r) and [c, r′) divide the plane into two sectors if and only if -- either r, r′ and c are not colinear, -- or r, r′ and c are colinear and c is between r and r′ on the segment [r, r′]. If it exists then this pair of sectors is denoted by {rcr′, rcr′} and we assume that the two half-lines [c, r) and [c, r′) do not belong to any sector in {rcr′, rcr′} . Note that, if the three points r, r′ and c are not colinear then one of two sectors is convex (angle centered at c between r and r′ ≤ 180o) and the other one is concave (angle centered at c between r and r′ > 180o). Otherwise, the three points r, r′ and c are colinear and the two sectors are convex and more precisely they are straight (both conjugate angles centered at c between r and r′ are equal to 180o). Definition 3 (Smallest enclosing circle). [6] Given a set P of n ≥ 2 points p1, p2, · · · , pn on the plane, the smallest enclosing circle of P , called SEC(P), is the smallest circle enclosing all the positions in P. It passes either through two of the positions that are on the same diameter (opposite positions), or through at least three of the positions in P. When no ambiguity arises, SEC(P) will be shortly denoted by SEC and SEC(P) ∩ P will indicate the set of all the points both on SEC(P) and P. Besides, we will say that a robot r is inside SEC if, and only if, there is not located on the circumference of SEC. In any configuration P, SEC is unique and can be computed in linear time [3]. Given a set P of n ≥ 2 points p1, p2, · · · , pn on the plane and SEC(P) its smallest enclosing circle, Rad(SEC(P)) will indicate the length of the radius of SEC(P). The next lemma contains a simple fact. Lemma 1. Let P1 be an arbitrary configuration of n points. Let P2 be a con- figuration obtained by pushing inside SEC(P1) all the points which are in P1 ∩ SEC(P1). We have Rad(SEC(P2)) < Rad(SEC(P1)). Let S and C be respectively a sector in {pcp′, pcp′} and a circle centered at c. We denote by arc(C, S) the arc of the circle C inside S. Given a set P of n ≥ 2 points p1, p2, · · · , pn on the plane and SEC(P) its smallest enclosing circle centered at c, we say that p and p′ are adjacent on SEC(P) if, and only if, p and p′ are in P and there exists one sector S ∈ {pcp′, pcp′} such that there is no point in arc(SEC(P), S) ∩ P. The following property is fundamental about smallest enclosing circles Property 1. [5] Let P and c be respectively a set of n ≥ 2 points p1, p2, · · · , pn on the plane and the center of SEC(P). If p and p′ are adjacent on SEC(P) then, there does not exist a concave sector S in {pcp′, pcp′} such that there is no point in arc(SEC(P), S) ∩ P. Property 2 is more general than Property 1 Property 2. Let P and c be respectively a set of n ≥ 2 points p1, p2, · · · , pn on the plane and the center of SEC(P). If p and p′ are in P then, there does not exist a concave sector S in {pcp′, pcp′} such that there is no point in S ∩ P. Proof. Assume by contradiction that p and p′ are in P and, there exists a concave sector S in {pcp′, pcp′} such that there is no point in S ∩ P. So, there is no point in arc(SEC(P), S) ∩ P. We deduce that there exists a concave sector S ′ in {qcq′, qcq′} such that q and q′ are adjacent on SEC(P) and there is no point in arc(SEC(P), S ′) ∩ P. Contradiction with Property 1. Figure 1 illustrates Property 2. r i c C 1 kr r j C 2 Fig. 1. C2 is an enclosing circle for the three points ri, rj and rk. However, there is no point in the intersection between C2 and the concave sector formed by ri, rj and the center c of C2. So, C2 can be replace by a smaller enclosing circle, here C1, even if all the points are on the circumference of C2. Observation 1 Given three colinear points, c,r,r'. If c is on the segment [r, r′], then c cannot be on the circumference of a circle enclosing r and r′. Definition 4 (Convex Hull). [16] Given a set P of n ≥ 2 points p1, p2, · · · , pn on the plane, the convex hull of P, denoted H(P) , is the smallest polygon such that every point in P is either on an edge of H(P) or inside it. Informally, it is the shape of a rubber-band stretched around p1, p2, · · · , pn. The convex hull is unique and can be computed with time complexity O(n log n) [16]. When no ambiguity arises, H(P) will be shortly denoted by H and H(P) ∩ P will indicate the set of the positions both on H(P) and P. From Definition 4, we deduce the following property : Property 3. Let P be respectively a set of n ≥ 2 points that are not on the same line and let H(P) be a convex hull. The two following properties are equivalent 1. Any point c, not necessarily in P, is located on H (either on a vertice or an edge) 2. there is a concave or a straight sector S in {rcr′, rcr′} such that r and r′ are in P and there exists no point ∈ P ∩ S. The relationship between the smallest enclosing circle and the convex hull is given by the following property Property 4. [3] Given a set P of n ≥ 2 points on the plane. We have SEC(P) ∩ P ⊆ H(P) ∩ P . The Algorithm Based on the definitions and basic properties introduced above, we are now ready to present a deterministic self-stabilizing algorithm that allows n robots (n odd) to gather in a point, regardless of the initial po- sitions of the robots on the plane. The idea of our algorithm is as follows : It consists in transforming an arbitrary configuration P into one where there is exactly one point pmax ∈ M axP. When such a configuration is reached, all the robots which are not located at pmax move towards pmax avoiding to create another point q than pmax such that q ≥ pmax. When M axP 6= 1, we will distinguish two cases : M axP = 2 and M axP ≥ 3. If M axP = {pmax1; pmax2}, then each robot which is not located neither on pmax1 nor pmax2 moves towards its closest position ∈ M axP by avoiding to create an adding maximal point. Since the number of robots is odd, we have eventually either pmax1 > pmax2 or pmax1 > pmax2 and then, M axP = 1. For the case M axP ≥ 3, our strategy consists in trying to create a unique maximal point inside SEC. To reach such a configuration, we distinguish three subcases : 1. If there is no robot inside SEC, then all the robots are allowed to move towards the center of SEC. 2. If all the robots inside SEC are located at the center of SEC, then only the robots located in SEC ∩ M axP are allowed to move towards the center of SEC. 3. If some robots inside SEC are not located at the center of SEC, then only the robots inside SEC are allowed to move towards the center of SEC. The main algorithm is shown in Algorithm 1. In Algorithm 1, we use two subroutines : move to caref ully(p) and choose closest position(p1, p2). The for- mer allows a robot r, located at q, to move towards p only if there is no robot on the segment [q, p] except the robots located on p or the robots located on q. The latter one returns the closest position to r among {p1, p2}. If the distance between r and p1 is equal to the distance between r and p2 then the function returns p1. Algorithm 1 Gathering for an odd number of robots, executed by each robot. P := the set of all the positions; M axP := the set of all the points p ∈ P such that p is maximal; if M axP = 1 then pmax := the unique point in M axP; I am not on pmax; if then move to caref ully(pmax); endif endif if M axP = 2 then pmax1 := the first point in M axP; pmax2 := the second point in M axP; if then q := choose closest position(pmax1, pmax2); I am not neither on pmax1 nor pmax2 move to caref ully(q); endif endif if M axP ≥ 3 then SEC := the smallest circle enclosing all the points in P; c := the center of SEC; Boundary := SEC ∩ P; Inside := P \ Boundary; if Inside 6= ∅ then if All the robots ∈ Inside are located at c then if I am in (Boundary ∩ M axP) then move to(c); endif else if I am in Inside then move to(c); endif endif else move to(c); endif endif Proof of closure Lemma 2 (Closure). According to Algorithm 1, if all the robots are located at the same position p, then all the robots are located at the same position thereafter. Proof of convergence Cases M axP = 1 and M axP = 2. Lemma 3. Let P be an arbitrary configuration for an odd number of n robots. According to Algorithm 1, if M axP = 1 then all the robots are located at the same position in finite time. Lemma 4. Let P be an arbitrary configuration for an odd number of n robots. According to Algorithm 1, if M axP = 2 then all the robots are located at the same position in finite time. Case M axP ≥ 3. In this paragraph, we prove that starting from a configuration where M axP ≥ 3, all the robots are located at the same position in finite time. More precisely, we consider the case where there exists at least one robot inside SEC(P(t)) ( refer to Lemma 7) and the case where there is no robot inside SEC(P(t)) ( refer to Lemma 8). In order to prove Lemma 7, we use Lemmas 5 and 6. In particular, Lem- mas 5 shows that, under specific conditions, the center of SEC(P(t)) is inside SEC(P(t + 1)) even if SEC(P(t)) 6= SEC(P(t + 1)) or the center of SEC(P(t)) is not the center of SEC(P(t + 1)). The proof of Lemma 5 is organized in two parts. In the former one, we consider the case where the center of SEC is also on the convex hull (see Figure 2.c). In the latter one, we consider the case where the center of SEC is not on the convex hull. Lemma 5. Let P(t) be a configuration such that M axP ≥ 3 and there exists at least one robot inside SEC(P(t)). According to Algorithm 1, if both conditions are true : 1. some robots ∈ P(t) ∩ SEC(P(t)) move in straight line toward the center c of SEC(t) and, 2. for every p ∈ P(t) ∩ SEC(P(t)) there exists at least one robot in p which does not reach c at time t + 1 then, the center of SEC(P(t)) is inside SEC(P(t + 1)) at time t + 1. Proof. Let c be the center of SEC(t) at time t. We consider two cases, depending on whether c is on the convex hull H(P(t)) or not, at time t. -- c is on H(P(t)) at time t. From Property 3, there exists a concave or a straight sector S in {xcy, xcy} such that x and y are in P(t) and there is no point ∈ P(t) ∩ S. However, from Property 2, we know that there do not exist two points x and y in P(t) such that there exists a concave sector S (5) (2) (1) (3) (2) C (5) (a) (2) (1) (4) (3) C (3) (1) (4) (b) Fig. 2. The numbers between parenthesis indicate the multiplicity. In Figure a, we have a configuration P(t) where the center c of SEC(P(t)) is inside the convex hull. Figure b, we have configuration P(t + 1) where some robots have moved toward c and c is inside the new convex hull. in {xcy, xcy} and P(t) ∩ S = ∅. So, there exists only a straight sector S in {xcy, xcy} such that x and y are in P(t) and there is no point ∈ P(t) ∩ S. Consequently, c is on the segment [x, y] at time t. Since the robots move in straight line towards c and since there exist some robots located at x and some robot located at y which do not reach c at time t + 1 then, c is on the segment [r, s] at time t + 1 with r and s ∈ P(t + 1). From Observation 1, we deduce that c is inside SEC(P(t + 1)) at time t + 1. -- c is not on H(P(t)) at time t. In this case, all the points in P(t) are not on the same line otherwise c would have been on H(P(t)). So, from Property 3 we know that there does not exist a concave or a straight sector S in {xcy, xcy} such that x and y are in P(t) and there is no point ∈ P(t) ∩ S. Since the robots move in straight line towards c and since for each point p ∈ P(t) there exists at least one robot located on p which does not reach c at time t+1 then, we deduce that there does not exist a concave or a straight sector S in {rcs, rcs} such that r and s are in P(t + 1) and there is no point ∈ P(t + 1) ∩ S (Figures 2.a and 2.b illustrate this fact). So, from Property 3 c is inside H(P(t + 1)) at time t + 1, and from Lemma 4 we deduce that c is inside SEC(P(t + 1)). Lemma 6. Let P(t) be a configuration such that M axP ≥ 3 and there exists at least one robot inside SEC(P(t)). If any robot r is inside SEC(P(t)) and r is located on the boundary of SEC(P(t + 1)) then M axP(t + 1) ≤ 2. Proof. By contradiction assume that r is inside SEC(P(t)) and r is located on the boundary of SEC(P(t + 1)) and M axP(t + 1) > 2. Let c be the cen- ter of SEC(P(t)) at time t. From assumption, some robots on the boundary of SEC(P(t)) have moved toward the center of SEC(P(t)). According to Algo- rithm 1, that implies that all the robots inside SEC(P(t)), notably r, are located at the center of SEC(P(t)) at time t. So, c is on the boundary of SEC(P(t+ 1)). From Lemma 5, we deduce that there exists a point p ∈ P(t) ∩ SEC(P(t)) such that all the robots in p have reached c at time t + 1. However, according to Algorithm 1 only the robots located in ∈ M axP(t) ∩ SEC(P(t)) are allowed to move at time t. Therefore, for every point p 6= c we have c > p at time t + 1. Hence, M axP(t + 1) = {c} i.e., M axP(t + 1) = 1. A contradiction. Lemma 7. Let P(t) a configuration such that M axP ≥ 3 and there exists at least one robot inside SEC(P(t)). According to Algorithm 1, all the robots are located at the same position in finite time. Proof. Assume by contradiction M axP ≥ 3 forever. From Lemma 6, we know that the robots inside SEC(P(t)) are inside SEC(P(t + 1)). So, by induction we deduce that the robots inside SEC(P(t)) are inside SEC(P(ti)) for all ti ≥ t. From Lemma 1, fairness and because of the fact that each robot r can move to at least a constant distance σr > 0 in one step, we know that there exists a time instant tk where the number of robots at the center of SEC(P(tk)) will be greater than the number of robots not located at the center of SEC(P(tk)). So, M axP = 1 : contradiction. So, M axP ≤ 2 in finite time and from Lemmas 3 and 4 all the robots are located at the same position in finite time. Lemma 8. Let P(t) be a configuration such that M axP ≥ 3 and there exists no robot inside SEC(P(t)). According to Algorithm 1, all the robots are located at the same position in finite time. Proof. According to Algorithm 1, all the robots may decide to move toward the center of SEC. Since each robot r can move to at least a constant distance σr > 0 in one step, if all the robots are always on the boundary of SEC(P) then, by fairness, the gathering problem is solved in finite time. Otherwise, -- either there exists tk > t such that M axP(tk) ≥ 3 and there exists at least one robot inside SEC(P(t)) : From Lemma 7, we deduce that all the robots are located at the same position in finite time, -- or there exists tk > t such that M axP(tk) ≤ 2 : from Lemmas 3 and 4 all the robots are located at the same position in finite time. 4 Conclusion Assuming strong multiplicity detection, we provided a complete characterization (necessary and sufficient conditions) to solve the gathering problem. Note that we do not know whether strong multiplicity detection is a necessary condition to solve the gathering problem. In future works, we would like to study the weakest minimal multiplicity detection that solves this problem and under which conditions. Note that the gathering problem seems to be the only positioning problem that can be deterministically and self-stabilizing solved. Indeed, since initially the robots can share the same positions, there exists no deterministic algorithm to scatter them in the plane [8]. References 1. Noa Agmon and David Peleg. Fault-tolerant gathering algorithms for autonomous mobile robots. SIAM J. Comput., 36(1):56 -- 82, 2006. 2. H Ando, Y Oasa, I Suzuki, and M Yamashita. A distributed memoryless point convergence algorithm for mobile robots with limited visibility. IEEE Transaction on Robotics and Automation, 15(5):818 -- 828, 1999. 3. P Chrystal. On the problem to construct the minimum circle enclosing n given points in a plane. In the Edinburgh Mathematical Society, Third Meeting, page 30, 1885. 4. M Cieliebak, P Flocchini, G Prencipe, and N Santoro. Solving the robots gath- ering problem. In Proceedings of the 30th International Colloquium on Automata, Languages and Programming (ICALP 2003), pages 1181 -- 1196, 2003. 5. Mark Cieliebak. Gathering non-oblivious mobile robots. In LATIN, pages 577 -- 588, 2004. 6. X Defago and A Konagaya. Circle formation for oblivious anonymous mobile robots with no common sense of orientation. In 2nd ACM International Annual Workshop on Principles of Mobile Computing (POMC 2002), pages 97 -- 104, 2002. 7. Xavier D´efago, Maria Gradinariu, St´ephane Messika, and Philippe Raipin Parv´edy. Fault-tolerant and self-stabilizing mobile robots gathering. In DISC, pages 46 -- 60, 2006. 8. Yoann Dieudon´e and Franck Petit. Scatter of robots. Parallel Processing Letters, 19(1):175 -- 184, 2009. 9. S. Dolev. Self-Stabilization. The MIT Press, 2000. 10. Paola Flocchini, David Ilcinkas, Andrzej Pelc, and Nicola Santoro. Remembering without memory: Tree exploration by asynchronous oblivious robots. In SIROCCO, pages 33 -- 47, 2008. 11. Paola Flocchini, Evangelos Kranakis, Danny Krizanc, Nicola Santoro, and Cindy Sawchuk. Multiple mobile agent rendezvous in a ring. In LATIN 2004: Theoretical Informatics, 6th Latin American Symposium, volume 2976 of Lecture Notes in Computer Science, pages 599 -- 608. Springer, 2004. 12. Paola Flocchini, Giuseppe Prencipe, Nicola Santoro, and Peter Widmayer. Gath- ering of asynchronous robots with limited visibility. Theor. Comput. Sci., 337(1- 3):147 -- 168, 2005. 13. Ralf Klasing, Euripides Markou, and Andrzej Pelc. Gathering asynchronous obliv- ious mobile robots in a ring. Theor. Comput. Sci., 390(1):27 -- 39, 2008. 14. G Prencipe. Corda: Distributed coordination of a set of autonomous mobile robots. In Proceedings of the Fourth European Research Seminar on Advances in Dis- tributed Systems (ERSADS 2001), pages 185 -- 190, 2001. 15. Giuseppe Prencipe. Impossibility of gathering by a set of autonomous mobile robots. Theor. Comput. Sci., 384(2-3):222 -- 231, 2007. 16. Franco P. Preparata and S. J. Hong. Convex hulls of finite sets of poin ts in two and three dimensions. Commun. ACM, 20(2):87 -- 93, 1977. 17. Samia Souissi, Xavier D´efago, and Masafumi Yamashita. Using eventually consis- tent compasses to gather memory-less mobile robots with limited visibility. ACM Trans. Auton. Adapt. Syst., 4(1):1 -- 27, 2009. 18. I Suzuki and M Yamashita. Agreement on a common x-y coordinate system by a group of mobile robots. Intelligent Robots: Sensing, Modeling and Planning, pages 305 -- 321, 1996. 19. I Suzuki and M Yamashita. Distributed anonymous mobile robots - formation of geometric patterns. SIAM Journal of Computing, 28(4):1347 -- 1363, 1999. Annexes Lemma 2. According to Algorithm 1, if all the robots are located at the same position p, all the robots are located at the same position thereafter. Proof. If all the robots are located at the same position p, then M axP = 1 and all the robots are located at the unique position p ∈ M axP. According to Algorithm 1, in the case M axP = 1 the robots located on p remains idle. So, all the robots are located at the same position forever. Lemma 3. Let P be an arbitrary configuration for an odd number of n robots. According to Algorithm 1, if M axP = 1 then all the robots are located at the same position in finite time. Proof. Let pmax be the unique point in M axP(t). According to Algorithm 1, the robots located on pmax during step (t, t+1) remains idle. Moreover, according to Algorithm 1 and Function move to caref ully(), if two robots ri and rj are not at the same point at time t, i.e., pi(t) 6= pj(t) then pi(t + 1) 6= pj(t + 1) at time t + 1 unless they have reached pmax. Hence, pmax remains the unique point in M axP(tk), for all tk ≥ t. So, according to Algorithm 1 and by fairness, we deduce that pmax = n in finite time. Lemma 4. Let P be an arbitrary configuration for an odd number of n robots. According to Algorithm 1, if M axP = 2 then all the robots are located at the same position in finite time. Proof. The proof is organized as follows : First, we prove that there exists tk ≥ t such that M axP(tk) 6= 2. Then, we prove that there does not exist any time tk ≥ t such that M axP(tk) ≥ 3. Finally, we deduce that Lemma 3 holds. 1. Assume by contradiction that there does not exist any time tk ≥ t such that M axP(tk) 6= 2. Consequently, for every tk ≥ t, M axP(tk) = 2. Let pmax1 and pmax2 be the two points in M axP(t) at time t. According to Algorithm 1, the robots located either on pmax1 or on pmax2 during step (t, t + 1) remains idle. Moreover, according to Algorithm 1 and Function move to caref ully(), if two robots ri and rj are not at the same point at time t, i.e., pi(t) 6= pj(t) then pi(t + 1) 6= pj(t + 1) at time t + 1 unless either ri and rj have reached pmax1 or ri and rj have reached pmax2. So, by induction we deduce that pmax1 and pmax2 remains the only positions in M axP(tk) for every tk ≥ t. By fairness, we deduce that, all the robots are either at pmax1 or at pmax2 in finite time. However, since the number of robots is odd then, we have either pmax1 > pmax2 or pmax1 < pmax2. Hence, M axP(tk) = 1 : Contradiction. 2. Assume by contradiction that there exists tk ≥ t such that M axP(tk) ≥ 3. Without lost of generality, we assume that tk is the first time for which M axP(tk) ≥ 3. Clearly, there does not exist any time tl such that t < tl < tk and M axP(tk) = 1 : Indeed from Lemma 2 and the proof of Lemma 3, once there exist a unique point pmax then, it remains the unique point in M axP forever and that would be a contradiction. Hence, M axP(tk − 1) = 2. Let pmax1 and pmax2 be the two points in M axP(tk − 1) at time tk − 1. According to Algorithm 1, the robots located either on pmax1 or on pmax2 during step (t, t + 1) remains idle. Besides, according to Algorithm 1 and Function move to caref ully(), if two robots ri and rj are not at the same point at time tk − 1, i.e., pi(tk − 1) 6= pj(tk − 1) then pi(k) 6= pj(k) at time tk unless either ri and rj have reached pmax1 or ri and rj have reached pmax2. So, M axP(tk) ≤ 2 at time tk. A contradiction. From above, we deduce that if M axP(t) = 2 at time t then, according to Algorithm 1 there exists tk, tk > t such that M axP(tk) = 1. So, from Lemma 3, we know that all the robots will be located at the same position in finite time.
1107.3225	1	1107	2011-07-16T12:47:54	An Agent-based Strategy for Deploying Analysis Models into Specification and Design for Distributed APS Systems	[ "cs.MA" ]	Despite the extensive use of the agent technology in the Supply Chain Management field, its integration with Advanced Planning and Scheduling (APS) tools still represents a promising field with several open research questions. Specifically, the literature falls short in providing an integrated framework to analyze, specify, design and implement simulation experiments covering the whole simulation cycle. Thus, this paper proposes an agent-based strategy to convert the 'analysis' models into 'specification' and 'design' models combining two existing methodologies proposed in the literature. The first one is a recent and unique approach dedicated to the 'analysis' of agent-based APS systems. The second one is a well-established methodological framework to 'specify' and 'design' agent-based supply chain systems. The proposed conversion strategy is original and is the first one allowing simulation analysts to integrate the whole simulation development process in the domain of distributed APS.	cs.MA	cs	IJCSI International Journal of Computer Science Issues, Vol. 8, Issue 3, No. 1, May 2011 ISSN (Online): 1694-0814 www.IJCSI.org An Agen t-b ased S tra tegy for Dep loy ing Ana lys is Mode ls into Spec i f icat ion and Des ign for D is tr ibu ted APS Sys tems 7 Luis Antonio de Santa-Eulalia1, Sophie D’Amours2 and Jean-Marc Frayret3 1 Téluq, Université du Québec à Montréal Quebéc City, Québec, Canada 2 Université Laval Quebéc City, Québec, Canada 3 École Polytechnique de Montréal Montréal, Québec, Canada Abstract Despite the extensive use of the agent technology in the Supply Chain Management field, its integration with Advanced Planning and Scheduling (APS) tools still represents a promising field with several open research questions. Specifically, the literature falls short in providing an integrated framework to analyze, specify, design and implement simulation experiments covering the whole simulation cycle. Thus, this paper proposes an agent-based strategy to convert the ‘analysis’ models into ‘specification’ and ‘design’ models combining two existing methodologies proposed in the literature. The first one is a recent and unique approach dedicated to the ‘analysis’ of agent-based APS systems. The second one is a well-established methodological framework to ‘specify’ and ‘design’ agent-based supply chain systems. The proposed conversion strategy is original and is the first one allowing simulation analysts to integrate the whole simulation development process in the domain of distributed APS. Keywords: Advanced Planning and Scheduling (APS), Agent- Based Simulation, Methodological Framework, Analysis, Specification and Design, FAMASS. 1. Introduction Advanced Planning and Scheduling (APS) systems comprise a set of techniques for the supply chain planning over short, intermediate, and long-term time periods. They employ advanced mathematical algorithms or logic to perform optimization or simulation on finite capacity scheduling, sourcing, capital planning, resource planning, forecasting, demand management, and other. APS simultaneously considers a range of constraints and business rules to provide real-time planning and scheduling, decision support, available-to-promise, and capable-to-promise capabilities. In addition, these systems often generate and evaluate multiple ‘what-if’ scenarios [1]. The use of these sophisticated optimization approaches in complex real-life supply chain situations has recently become possible mainly due to the increased computing power of companies [2]. Despite the contribution of APS systems to the supply chain planning domain, some criticism exists in this area [3]. Traditional APSs are basically monolithic systems that cannot model and take into account the complex everyday interactions and information exchanges between partners. For example, APS systems are deficient in handling that allow interaction mechanisms sophisticated the delegation implementation of and coordination approaches, which are methodologies based on negotiation, and cooperation strategies [4, 5]. As a result, the focus on relationships in a multi-tier environment has only recently been claimed by the APS community [6]. To cope with this problem, recent advances in supply chain planning have arisen in the area of agent technology. This technology is able to capture the distributed nature of supply chain entities (e.g. customers, manufacturers, logistics operators etc.) and mimic their business behaviours (e.g. making advanced production decisions and negotiating with other supply chain members), thus supporting their collaborative planning process. Because of these abilities, among several others described in the IJCSI International Journal of Computer Science Issues, Vol. 8, Issue 3, No. 1, May 2011 ISSN (Online): 1694-0814 www.IJCSI.org 8 literature, agent-based supply chain systems have great potential for simulating complex and realistic scenarios [7, 4; 9, 10, 11]. Distributed APS systems employing agent- based technology are referred to in this paper as distributed APS systems [12]. Distributed APS systems are normally developed through the use of modelling and simulation frameworks and, usually, these frameworks provide principles, steps, methods and tools for creating a model. They help people understand the simulation problem to be modelled and translate it into a computing model normally used in simulation experiments in the supply chain planning area. In order to create such models, these frameworks guide simulation modellers through one or several development steps [13]. The first modelling step is analysis, where one defines an abstract description of the modelled supply chain planning system containing functional and non- functional requirements. Next, during specification, the information derived from the analysis is translated into a formal model. As the analysis phase does not necessarily allow obtaining a formal model, the specification examines the analysis requirements and builds a model based on a formal approach. After, in the design phase one creates a data-processing model that describes the specification model in more detail. In the case of an agent- based system, design models are close to how agents operate. Finally, during implementation, the design model is translated into a specific software platform or it is programmed [13]. The problem behind these modelling frameworks is that normally simulation systems are implemented as directed by pre-stated requirements with little explicit focus on system analysis, specification, design and implementation in an integrated manner [14]. According to a recent literature review [15], to the best of our knowledge there are no integrated modelling approaches covering the whole developed process in this area. Moreover, there is one unique ‘analysis’ modelling, the FAMASS (FORAC Architecture for Modelling Agent-based Simulations for Supply chain planning) framework, dedicated to the distributed APS domain, and which was proposed by us recently [21, 22, 23]. Despite its contribution to the literature, FAMASS is limited to the identification and mapping of functional requirements of distributed APS simulations, i.e. the ‘analysis’ phase only. If the simulation analysts desire to go further in the modelling process, they have to employ another ‘specification’ and ‘design’ methodology. This can be laborious, since analysts need to thoroughly master FAMASS and another methodology. In order to facilitate FAMASS analysts in converting their analysis models into specification and design models, this paper proposes an agent-based deployment strategy. This strategy enlarges the FAMASS scope to the other modelling phases, thus covering the entire modelling cycle. By doing so, analysis can go smoother and quicker through this cycle. To do so, we were inspired by the specification and design principles of the Labarthe et al. [9] framework, a recent and field of the in largely cited development methodological agent-oriented framework for supply chain management simulation. Since the focus of this framework is on supply chain management as a general concept (and not specialized in APS systems), we had to perform some minor adaptations to this approach. Despite these adaptations, the main ideas of Labarthe et al. [9] are explicitly considered in the deployment strategy. The Labarthe et al. framework is adopted here because it covers the specification and design phases properly at the business and agent levels, just as FAMASS does, which facilitates the deployment process. This deployment strategy demonstrates that the analysis phase of FAMASS can be integrated with other existing approaches specialized in specification and design modelling. Furthermore, it allows us to avoid the research effort needed to develop a totally new specification and design methodology for the domain, although it would be suitable (and even desirable) for future research initiatives. This paper is organized as follows: a literature review in modelling and simulation for distributed APS systems is presented in Section 2. Section 3 introduces the FAMASS approach, while Section 4 summarizes the Labarthe et al. [9] framework. Next, the deployment process is explained in Section 5. Finally, Section 6 outlines some final remarks and suggests future work. 2. Modelling and Simulation Frameworks for distributed APS The use of agent technology in Supply Chain Management is a fruitful field. From the inaugural work of Fox et al. [16] until today, a large variety of works have appeared to propose different ways of encapsulating supply chain entities and performing simulation experiments. Two types of modelling approaches can be identified in the literature. The first type proposes generic approaches for modelling agent-based supply chain systems in general terms, while the second type proposes a modelling framework into consideration takes that specifically Advanced Planning and Scheduling (APS) tools when IJCSI International Journal of Computer Science Issues, Vol. 8, Issue 3, No. 1, May 2011 ISSN (Online): 1694-0814 www.IJCSI.org 9 planning, i.e. the incorporate optimization procedures or finite capacity planning models when performing supply chain planning. APS systems emerged in the last decade to provide a suite of planning and scheduling modules for the firm’s internal supply chain, from the raw materials source to the consumers and covering decisions ranging from the strategic to the operational level [17]. In the first type of approach (general agent-based models), examples of relevant contributions include Labarthe et al. [9], Van der Zee, and Van der Vorst [18], MaMA-S [13]. One of the most cited works in the domain is Labarthe et al. [9], which propose a methodological framework for modelling customer-centric supply chains in the context of mass customization. They define a conceptual model for supply chain modelling and show how the multi-agent system can be implemented using predefined agent platforms. Van der Zee and Van der Vorst [18] propose an agent framework derived from an object-oriented approach to explicitly model control structures of supply chains. MaMA-S [13] provides a multi-agent methodology for a distributed industrial system, which is divided into five main phases and two support phases. The authors propose formal methods the specification, design and for implementation phases, but the analysis phase is not tackled by them. This second type of modelling approach provides more sophisticated models of supply chains by incorporating Advanced Planning and Scheduling routines [12]. These approaches, sometimes called d-APS systems (for distributed APS), are composed of semi-autonomous APS tools, each dedicated to a specialized planning area and that can act together in a collaborative manner employing sophisticated interaction schemas. Examples of this kind of work are Egri et al. [19], Lendermann et al. [20] and Swaminathan et al. [11]. Egri et al. [19] is a Gaia-based approach for modelling advanced distributed supply chain planning for mass customization. They develop a model for representing roles and interactions of agents based on the SCOR (Supply-Chain Operations Reference) model. Lendermann et al. [20] developed an approach to couple discrete-event simulation and APS for collaborative supply chain optimization, based on the HLA (High Level Architecture) technology for distributed simulation synchronization. Swaminathan et al. [11] provide a supply chain modelling framework containing a library of modular and reusable software components, which represents different kinds of supply chain agents, their constituent control elements and their interaction protocols. These simulation and modelling approaches have greatly contributed to the domain, however, in spite of these advances, there exists a relevant gap in this field related to the initial developing step of such simulation systems, the analysis phase [12]. Most of the researched works in the literature suggest approaches for specification and design, and some for implementation, but the analysis phase is not explicitly treated [12, 13, 14, 21]. Most of these works suppose that the analysis phase furnishes the necessary information and concentrate their discussions on further phases, mainly specification and design. The first work dedicated to the analysis of distributed APS systems using the agent-based paradigm is FAMASS [21]. Despite its contribution to the agent-based modelling of distributed APS systems, FAMASS does not cover the specification and design phases of the development process. This is an interesting research gap in the literature. Section 3 details the FAMASS approach for the analysis phase, while Section 4 presents a frequently cited method for specification and design of agent-based supply chain systems from Labarthe et al. [9]. Next, Section 5 combines these two approaches in order to create a deployment strategy to translate analysis models into specification and design. 3. The FAMASS Approach The FAMASS (FORAC Architecture for Modelling Agent-based Simulation for Supply chain planning) is the first and unique modelling approach dedicated to the analysis phase of distributed APS simulations [21, 22, 23]. This approach was recently tested in Santa-Eulalia et al. [24]. It is organized into two abstraction levels: Supply chain: refers to the supply chain planning problem, i.e. the business viewpoint; Agent: the supply chain domain problem is translated into an agent-based view (Figure 1). At these two abstraction levels, four modelling approaches are proposed, namely the General Problem Analysis (GPA), the Distributed Planning Analysis (DPA), the Social Agent Organization Analysis (SAOA) and the Individual Agent Organization Analysis (IAOA), as schematized in Fig. 1. !"#$%&'&()#&+( !"#"$%&' ($)&"+',#%&-./.' 0!(,1' 2/.3$/43"5' (&%##/#6',#%&-./.' 02(,1' 7)8/%&',6"#3' 9$6%#/:%;)#' ,#%&-./.' 07,9,1' <#5/=/54%&' ,6"#3' 9$6%#/:%;)#' ,#%&-./.' 0<,9,1' ' # / % @ ? ' - & > > 4 7 ' 3 # " 6 , Fig. 1: Four main modelling approaches proposed for analysis of supply chain and agent levels [23]. IJCSI International Journal of Computer Science Issues, Vol. 8, Issue 3, No. 1, May 2011 ISSN (Online): 1694-0814 www.IJCSI.org These four modelling approaches are explained in the following subsections. are proposed in the framework of Fig. 2, which is called the supply chain planning cube. 10 3.1 General Problem Analysis (GPA) GPA is the first modelling effort where simulation analysts have to think about the simulation problems. The GPA is based on Santa-Eulalia et al. [12], in which a discussion about the simulation objective and the problem structure is provided. Basically, the simulation that the GPA proposes analysis has to take two main aspects into consideration: general aspects and experimental aspects. General aspects represent macro definitions of the simulation problem, including the object and environment to be simulated, the and simulation objectives. hypotheses questions, Experimental aspects are related to the design of experiments, where one defines the factors, uncertainties and key performance indicators of the simulation. These elements refer to the general definition of the simulation problem, according to what is desired to be studied, and it will guide the whole development process. This general definition is then organized through some formalisms from SysML (Systems Modeling Language) [25]. In this case, some Requirements Diagrams help the analysts organize the GPA. An example of how this can be done is provided in [23]. 3.2 Distributed Planning Analysis (DPA) The DPA identifies what the desired supply chain planning entities are, as well as their roles. These entities are identified according to their mission in the supply chain and their planning functions at different decision levels. To identify the main supply chain planning entities, FAMASS employs the concepts of supply chain integration proposed by Shapiro [26]. The author states that supply chain management refers to integrated planning relying on three basic dimensions: i) Intertemporal dimension: refers to different decision levels, i.e. strategic, tactical and operational decision levels; ii) Functional dimension: stands for different planning functions in a supply chain, which can be related to procurement, manufacturing, distribution and sales; iii) Spatial dimension: refers to the fact that supply chains are composed of geographically dispersed units of analysis. This gives rise to the notion of a Supply Chain Block. A Supply Chain Block can be defined as a supply chain planning entity, which is a functional unit capable of: performing part of the supply chain planning decisions or their totality; or performing the execution of the supply chain decisions (part of them or their totality). These entities have a certain degree of autonomy and are able to interact with each other. Possible Supply Chain Blocks for covering the integrated supply chain planning dimensions #&1':" 9):$%);4,/0" 7&048&$4%)0(" 5%/4%'6'0$" !" !" @"#4.1A" ?B&)0"C1/D" #$%&$'()" +&," -.'%&,/0&1" 23'4,/0"" <'0=/%:" >&)1),':" ?1)'0$:" ?/0:46'%:" +B'"#4.1A"?B&)0"?4;'" Fig. 2: Supply Chain Planning Cube [23]. A vertical slice of the supply chain planning cube for one spatial unit of analysis (e.g. facilities) is similar to the planning matrix proposed by Meyr and Stadtler [27], except for the execution level. The supply chain planning cube is an evolution of the planning matrix, due to the fact that it represents the possibility of collaboration among different includes It also traditional APS systems. execution entities. Based on the supply chain cube, one has to perform requirements determination for the simulation aspects. This cube serves as a metamodel to help simulation analysts identify their simulation requirements. For example, the analysts decide which kind of Supply Chain Blocks will be needed in their simulation experiments, providing the basic architectural aspects of the simulation system. Then, their requirements are organized through UML-based use cases and requirements diagrams from SyML. An example of the DPA is provided in Santa-Eulalia et al. [23]. 3.3 Social Agent Organization Analysis (SAOA) So far, the concept of Supply Chain Block has been used to represent entities responsible for part of the supply chain planning. Together they compose a population of entities interacting with each other, having a collective co- existence within the planning system. When these entities incorporate attitudes, orientations and behaviours comprising the interests, needs or intentions of other Supply Chain Blocks, they can be seen as social entities. They can exhibit complex actions that take into account the collectivity. A way to represent social entities is to model them as agents, thus creating multi-agent societies. The general logic indicated that a Supply Chain Block can be directly translated into agents by adding agent abilities to them. This is based on the agentification definition of Shen et al. [28], who explain that the agentification process can be functional-based (i.e. white Supply Chain Block) or physical-based (i.e. gray Supply Chain Block). IJCSI International Journal of Computer Science Issues, Vol. 8, Issue 3, No. 1, May 2011 ISSN (Online): 1694-0814 www.IJCSI.org However, in some situations a Supply Chain Block can be transformed into more than one agent, for example when specialization is required, in which case a planning agent can be specialized according to certain generic responsibility orientations, such as products, processors, processes or projects, to obtain faster or more precise responses for certain given situations. In other situations, apart from agents proceeding from the supply chain planning cube, different intermediary agents can be created to perform activities related to, e.g. the coordination of the agents’ society. In addition, the agentification process can also include the representation of information sources, interfaces and other services. The importance of this discussion relies on the notion that agentification is the basis for two mutually dependent aspects in agent-based systems which define the metamodel for the SAOA: • Social structures: system the agent represent architecture [24] characterizing the blueprint of relationships, giving a high level view of how groups solve problems and the role each agent plays within the structure. There are diverse types of social structures, such as hierarchical, federated and autonomous. Social protocols: are agents’ abilities concerning social aspects, normally related to cooperation principles (i.e. agents have to cooperate in order to plan the entire supply chain). Diverse abilities can be considered, like and multiplication, communication, grouping coordination, collaboration by sharing tasks and resources and conflict resolution through negotiation and arbitration. Different social structures and protocols are provided in Santa-Eulalia [22]. Similar to the DPA, these two aspects of the SAOA serve as a metamodel to help simulation analysis identify their requirements for the simulation model. For example, different social protocols can be tested in the simulation. Then, requirements can be organized through agent-based use cases from AUML (Agent Unified Modelling Language) and requirements diagrams from SysML. An example of the SAOA is provided in Santa-Eulalia et al. [23]. • 3.4 Individual Agent Organization Analysis (IAOA) As mentioned by Ferber [29], the task of assigning roles to every individual agent is normally regarded as the last phase in constructing an organization. The logic is that as soon as one knows what the functions to be assigned are, one defines individual specializations. These local assignments influence social protocols functioning inside their respective social structures. In addition, it also influences the local performance of the supply chain planning entities. This is the main idea of the IAOA. 11 At the individual level, agents can be organized according to different internal architectures but there is little consensus on how to conceive the internal architectures of agents [30] in the literature. In order to cope with this, the metamodel for the IAOA proposes that whatever the state of mind of an agent is (cognitive, reactive or hybrid), and whatever internal architecture an agent employs, an agent can be described simply according to its ‘abilities’. This is the central point when performing simulation. An ‘ability’ can be defined as the quality of being able to perform an action, or facilitate the action’s accomplishment. ‘Abilities’ allow for the implementation of actions and the determination of the system’s behaviour, as well as the determination of its related performance. Based on this notion, the metamodel defines two elements: • The Response Space: stands for a collection of general abilities available for the agents, including very simple reactive abilities or sophisticated cognitive ones. For example, one agent can have a simple ability to monitor the inventory levels of the supply chain, or a complex ability to perform production planning employing an optimization method. • Capacity to Produce an Adapted Response: represents the aptitude to choose which abilities have to be transformed into actions at a given time to respond to a given situation. This capacity can vary from elementary to complex. The simplest possible capacity is related to a reactive ‘if-then’ mechanism, where no cognition is necessary. For example, if the inventory level drops to a given threshold, the agent uses its procurement ability to start a procurement action. As the agent becomes more intelligent, more complex responses can be made for some given situations. For example, the linear “if-then” logic can be substituted by more complex approaches based on action optimization and learning. Based on these two elements of the metamodel, one can carry out requirements determination for the simulation model, selecting the desired requirements in terms of agents’ abilities. Similar to the SAOA, the IAOA’s requirements are organized through agent-based use cases from AUML and requirements diagrams from SysML [23]. FAMASS is detailed in Santa-Eulalia et al. [21, 22, 23]. An application of this approach is presented in Santa- Eulalia et al. [24]. 4. Labarthe et al.’s Methodological Framework The Labarthe et al. [9] framework is schematized in Fig. 3 and is briefly described afterwards. IJCSI International Journal of Computer Science Issues, Vol. 8, Issue 3, No. 1, May 2011 ISSN (Online): 1694-0814 www.IJCSI.org 12 ' + ' $ 2 & 0 % / + . + * ) - " & ( " , ' + ' $ 2 & & " % + 5 + * $ ) 4 " * ( . 3 ' & " 5 $ / & * # % 4 * . ; : !"#$%&'(")+' ,"&-./0$+'12&/'(")+' 3.4$5"&$+'12&/'(")+' (0+5612&/'789/#' :;.4%#&/$+'<+$&' ' F E D D C B ' A + $ ' / ' * @ / 4 $ = $ ? ' > " ' & " 5 0 = % 4 / & " , Fig. 3: Summarizing the Labarthes et al. [9] framework. The authors propose the modelling steps indicated in Fig. 3. Their contribution corresponds to two abstraction levels: conceptual modelling and operational modelling. Conceptual modelling is performed in two steps, the Domain Model and the Conceptual Agent Model. 4.1 Domain Model (DM) The Domain Model (DM) creates an abstraction of the supply chain. Inspired from the NetMAN approach [31, 32], Labarthe et al. [9] create two sub-models: a Structural Model and a Dynamic Model. The Structural Model, which is based on responsibility networks [33], defines the structure of the supply chain, i.e. its ‘actors’ and their related responsibilities, and it also depicts the material flows among all ‘actors’. The Dynamic Model complements the Structural Model by defining the behaviour of each ‘actor’ and its related interaction modes. 4.2 Conceptual Agent Model (CAM) The Conceptual Agent Model (CAM) remodels the Domain Model guided by the agentification process. From the Structural and Dynamic models, a unique agent model is created. A Conceptual Agent Model specifies the ‘agents’, the ‘objects’ transacted between them and the nature of the agent’s interactions (‘physical interactions’ and ‘informational interactions’). In this case, each ‘actor’ specified in the Structural Model produces a specific agent. Also, any activity of an actor generates a specific agent in close interaction with the agent associated to the actor concerned, which is regrouped in the same partition. In addition, any exchange of information from the Dynamic Model generates a message-based informational interaction; and any material flow from the dynamic model leads to a physical type interaction. After, at the Operational Level, Labarthe et al. [9] proposes the Operational Agent Model (OAM). 4.3 Operational Agent Model (OAM) The Operational Agent Model (OAM) is based on the Conceptual Agent Model, and it aims to build a computer model of the studied supply chain which will be later implemented on a simulation platform. First, the Operational Agent models the software architecture (at the social level). Next, it models the internal agent architecture (individual level), dealing with knowledge, behaviours and interactions among agents. After creating the Domain Model, the Conceptual Agent Model and the Operational Agent Model, a Multi-Agent System is implemented at the Exploitation level and a set of Experimental Plans supports the realization of simulation experiments (the black modelling approaches shown in Fig. 3). The author illustrated the Exploitation level though the implementation of a case study in a simulation environment. This is only a summarized review of Labarthe et al. [9]’s work. For further details about this framework and its applications, the reader is referred to Labarthe et al. [9, 35, 36] and Labarthe [34]. 5. The Deployment Process As explained in the introduction, the original framework of Labarthe et al. [9] had to be slightly adapted to be suitable to the distributed APS domain. The first adaptation occurs at the Domain Modelling. The main reason for not strictly employing the Labarthe et al. [9] Domain Model is because it is based on the responsibility network [33], which uses the definition of centre, i.e. a business entity – a decisional one – linked at the physical level by material flow. Centres do not correspond exactly to our semi-autonomous units, the Supply Chain Blocks (defined in subsection 3.2), which are based on the supply chain cube. We adapted the Labarthe et al. [9] model and thus proposed a modelling approach where the ‘centres’ are substituted by Supply Chain Blocks. Another relevant difference refers to the fact that we separate the Operating System (i.e. the Execution layer) and the Decision System (i.e. the Strategic, Tactical and Operational layers) in the Domain Model, which is not done in the Labarthe et al. [9] Domain Model. They distinguish these two layers later, in the Operational Agent Model. We decided to separate them earlier because both IJCSI International Journal of Computer Science Issues, Vol. 8, Issue 3, May 2011 ISSN (Online): 1694-0814 www.IJCSI.org 13 systems have to be identified in regard to the supply chain cube introduced in subsection 3.2. If we did not consider entities of the Operating System at this step, the Domain Model would be incomplete for a distributed APS, according to the definition of the supply chain cube. !"#"$%&' ($)&"+',#%&-./.' 0!(,1' 2/.3$/43"5' (&%##/#6',#%&-./.' 02(,1' 7)8/%&',6"#3' 9$6%#/:%;)#' ,#%&-./.' 07,9,1' <#5/=/54%&' ,6"#3' 9$6%#/:%;)#' ,#%&-./.' 0<,9,1' n i a h C y l p p u S t n e g A !"#$%&'(")+' /"&0123$+'4-&2' (")+' 516$7"&$+'4-&2' (")+' The following subsection discusses the Domain Model generation. 5.1 Domain Model (DM) The objective of the Domain Model is to identify what is to be modelled in the supply chain. As seen in Fig. 4, the Distributed Problem Analysis (DPA) can be translated directly into the Domain Model. Table 1 and Table 2 provide a translation strategy to create FAMASS Structural and Dynamic Models based on Labarthe et al. [9]. Specification Design Element Table 1: Structural Model. Labarthe et al. FAMASS Counterpart Analysis Analysis ,-&).' 67484"12('%9$%"&%()(+,%(:1;17+,%<'(=)>%2( !"#$%"&%'()(+,%(-./.00(."123'4'(5,1'%( Fig. 4: Deploying process. Fig. 4 depicts the general idea of the deployment process. From the analysis phase, the Distributed Planning Analysis models are the basis for the creation of the Domain Model. The Domain Model represents the supply chain under study and how advanced planning decisions are articulated. Next, the Conceptual Agent Model is naturally created from the Domain Model, but the Social Agent Organization Analysis is also used as an important reference. The Social Agent Organization Analysis provides the Social Structures for the Conceptual Agent Model and it reflects the agentification process used during the Social Agent Organization Analysis. Finally, the Operational Agent Model is created from the Conceptual Agent Model. However, relevant information about social protocols requirements comes from the Social Agent Organization Analysis, while requirements concerning the agents’ abilities come from the Internal Agent Organization Analysis. It is interesting to note, in Fig. 4, that the Domain Model and Conceptual Agent Model roughly correspond to the specification phase, while the Operational Agent Model can be considered equivalent to the design phase. The Domain Model and Conceptual Agent Model are the first formal models to describe the supply chain and the agent domain. The Operational Agent Model is closely related to how agents operate. To sum up, FAMASS proposes a set of abstract notions for distributed APS systems, while Labarthe et al. [9] provide a formal and detailed description of how the system should work. Central elements Main element: a network of (roles and Centres [33] responsibilities) and their interactions. Roles: Processor, producer, assembler, fulfiller, distributor, retailer, transporters, customer. Roles define the nature of the responsibility set. Responsibilities: examples, packing, grouping, sales, etc. Organizational level: supply chain, enterprise, business unit, cells, resources. Modelling formalism Responsibility networks of Montreuil and Lefrançois [33]. Modelling process Identify decision elements of the supply chain and the physical interactions among them. Main element: a network of Supply Chain Blocks and their interactions (interactions are simple representations of Supply relations). A Chain Block’s Supply Chain Block is used instead of centres. Roles: From the “spatial” axis of the supply chain cube (subsection 3.2), we identify the Supply Chain Blocks and their roles: vendors, facilities, clients and consumers. Responsibilities: can one identify responsibilities from the ‘functional’ axis of the supply (subsection 3.2): chain cube procurement, manufacturing, distribution and sales. Organizational levels: strategic, tactic, operational, execution for facilities, clients and vendors, consumers (i.e., the intertemporal axis). Class diagrams and class tables (from AUML – Agent Unified The Language). Modelling concept is the same as for responsibility network, but it is represented using AUML formalisms. Centres are classes; roles are roles in each class; responsibilities are operations in each class; organizational levels are stereotypes of the classes; business processes are operations in each class. We identify the elements from the execution and decision systems and we add only the interactions. physical Informational interactions are the dynamic model added in (later on in the modelling process). IJCSI International Journal of Computer Science Issues, Vol. 8, Issue 3, No. 1, May 2011 ISSN (Online): 1694-0814 www.IJCSI.org 14 Element Table 2: Dynamic Model. FAMASS Counterpart Labarthe et al. Central elements Modelling formalism Modelling process same Describes the elements, but with the to add more possibility information based on different experimental definitions, i.e. different configurations of the Supply Chain Blocks, and different performance indicators and uncertainties. Class diagrams and class tables (AUML). All flows are represented by arrows. The decoupling point is represented the class in name. Centre models are represented by arrows as well. Stock holding (raw material, work-in-process or final products) is represented the in operations of each class. are same The flows identified, as as well inventory positions and decoupling point position. They are described in the class tables. Describes (in time) the system behaviour and that elements the compose it. Uses the responsibility network to recognize the [33] coordination modes by the physical identifying and informational relations used according to the environmental stimulus. 32] [31, NetMan a plus approach the representation of decoupling point position. The decoupling point position is mentioned here because it is an the in issue important Labarthe et al. [9] framework. Apart from the physical flow identified previously, the modelling process describes the informational flow exchanged according to the dynamics of the environment. Four flow informational types for coordination are needs i) identified: offers expression; ii) iii) expression; about information coordination; and iv) information sharing by models exchanges. In addition, the decoupling point is positioned and inventories are mapped (raw material, work-in- process and final product). It identifies two models (for models exchange): the network model and the centre model. The most important difference between Labarthe et al. [9] and FAMASS is the use of centre for the former and the use of Supply Chain Block for the latter. Supply Chain Block is used instead of centres in FAMASS because decision entities are central elements. Labarthe [34, p.119] explains that a centre represents a decision process, but centre definitions are closely associated to physical entities of the execution system, i.e. there is a direct relation between a centre and an entity of the execution system. Later in the Labarthe et al. [9] modelling process, the decision system is introduced more formally in the Operational Agent Model. We separate the decision system from the execution system in the Domain Model, since we know that they are relevant for experimental definitions in distributed APS systems. Another difference is related to the fact that we employ a unique modelling formalism based on an AUML approach, coherent with the analysis phase of FAMASS, which employs only UML- inspired formalisms. The next sub-section transforms the Domain Model into a Conceptual Agent Model. 5.2 Conceptual Agent Model (CAM) The Conceptual Agent Model represents the agentification process of the Labarthe et al. [9] approach. The agentification process defines the agent society based on the Domain Model, i.e. which agents are created from the centres (in our case, Supply Chain Block) and how they are organized. Labarthe et al. [9] propose rules for creating agents (i.e., each centre becomes an actor-agent and each centre activity becomes an activity-agent). As discussed before, FAMASS converts each Supply Chain Block into an agent. It also verifies whether some agents are extinguished (e.g. merged with another agent) or whether new agents are introduced (e.g. a mediator). This information the Social Agent is obtained during Organization Analysis (SAOA). As indicated in Fig. 4, the Conceptual Agent Model is generated from the Domain Model and the SAOA (in this case, the social structures). Using Labarthe et al. [9] rules, the Domain Model provides the basic classes’ definition and, using the SAOA, it can be verified if new agent classes are derived from the Domain Model and if different social structures have to be tested and considered in the Conceptual Agent Model. Social Protocols from SAOA are not used in Conceptual Agent Modelling. The Strategy for creating a Conceptual Agent Model is shown in Table 3. IJCSI International Journal of Computer Science Issues, Vol. 8, Issue 3, May 2011 ISSN (Online): 1694-0814 www.IJCSI.org 15 Element Central elements Table 3: Conceptual Agent Models. Labarthe et al. FAMASS Counterpart Actor-agent: agents representing an organizational unit of the supply chain facilities, clients or (i.e. vendors, customers), related to the ‘spatial’ axis. Actor-agents group several other agents, the activity-agents. Activity-agent: agents the from representing decision system, the processes procurement, of manufacturing distribution and sales. These agents are at three different decision levels and they are related to the ‘functional’ axis. Objects: defined products. This is the first time products are specified. Information interactions: they come from the Domain Model. Physical Interactions: from the Domain Model. Actor-agent: centre. Activity-agent: represents a process of transformation, distribution, or stock keeping. Object: products. Informational interaction: same as in Domain Model. Physical interaction: same as in Domain Model. they come Modelling formalism A modelling graphical formalism [34] that models the two types of agents and their interactions. The CAM model is derived from the DM model. Adapted class diagrams, tables and package diagrams. The adaptation of the class diagrams refers to the insertion of objects (products), represented by simple square boxes in the link between two classes. Modelling process Similar process, with the following differences: - Actor-agents and activity-agents: in the classes, use role definitions to indicate if it is an actor-agent or an activity-agent; - Interactions: links between classes. 1. From centre to actor- agent: each centre creates an actor-agent. interactions 2. Physical between actor-agents: physical flow is specified by an linking agents and arrow indicating their respective exchanged objects. 3. Informational interactions between actor-agents: similar to 2, but for information flow. frontiers 4. Organizational definition: establishes the organization frontiers for the actor-agents and places the physical the flows between organizations. 5. Definition of the activity- agents: each activity of a centre is transformed into an activity-agent. Physical 6. interactions activity-agents: between the physical specify flow the activity-agents between and their related objects exchanged. 7. Informational interactions between activity-agents: same as 6, plus the interaction between actor-agents and activity-agents. It is important to note that an actor-agent coordinates a population of other activity-agents in the Labarthe et al. [9] approach. In the case of FAMASS, we decided to use the notion of actor-agent only as an aggregation of agents inside the same organization using a package diagram. The next sub-section transforms the Conceptual Agent Model into an Operational Agent Model. 5.3 Operational Agent Model (OAM) According represents the OAM [34], to Labarthe implementable models. These models involve a choice between two different agent architectures, i.e. the cognitive and the reactive architectures. We believe that most of the time it is not possible to completely distinguish cognitive agents from deliberative agents, meaning that normally agents can be seen as a hybrid state within the cognitive-reactive continuum. In Labarthe et al. [9]’s work, agents from the decision system assume a cognitive agent architecture, composing a cognitive agent society. Based on this society, the author then creates a reactive society responsible for the transformation process (execution system), linked with the cognitive society. As we believe that the agents from the decision system can also assume reactive behaviours (see subsection 3.2), we prefer not to use this agent architecture notation for the Operational Agent Model. Instead, we create two societies (decision agents and execution agents) from the Conceptual Agent Model and start to define all agents’ behaviours and agents’ protocols in detail, as done by Labarthe et al. [9], which is not contradictory to Labarthe et al.’s [9] work. As explained before, instead of separating into decision and execution societies at the Operational Agent Model, our approach does it at the beginning of the specification phase, i.e. at the Domain Model. In sum, our Operational Agent Model is generated from the Conceptual Agent Model, the Social Agent Organization Analysis and the Internal Agent Organization Analysis, as illustrated in Fig. 5. 5'36%)+,#(-* !$,%(67%&'(* +(%)8969* 05+!+1* :(/6;6/4%)* +,#(-* !$,%(67%&'(* +(%)8969* 0:+!+1* 2'(3#"-4%)+,#(- .'/#)02+.1 !"#$%&'(%)+,#(- .'/#)0!+.1 Fig. 5: Creating an Operational Agent Model. From the Conceptual Agent Model we represent two societies, the decision agents and the execution agents. This is the starting point of the Operational Agent Model. After, we obtain requirements about agent protocols from the Social Agent Organization Analysis, and we obtain requirements about agent abilities from the Internal Agent Organization Analysis. IJCSI International Journal of Computer Science Issues, Vol. 8, Issue 3, No. 1, May 2011 ISSN (Online): 1694-0814 www.IJCSI.org 16 Table 4 summarizes the deployment strategy for the Operational Agent Model. going faster and smoother through the whole modelling process. Element Table 4: Operational Agent Models. Labarthe et al. FAMASS Counterpart Central elements Modelling formalism Modelling process Multi-agent system architecture: a cognitive and a reactive agent society are represented. A cognitive agent, together with its corresponding reactive agent, form the ‘agent-actor’. It is a generic architecture to represent entities capable of taking their own decisions and acting accordingly. Specification of the software agent: knowledge, behaviour and interactions of each agent are defined. For the behaviours, the following entities are defined: a) external event: concerning the communication aspect with external entities of the multi-agent system; b) internal event: concerning internal activities of an agent; c) passive state: waiting state; d) active state, being an elementary action or a composite action. system the multi-agent For Labarthe architecture, [34] graphical his proposes own modelling formalism. For the specification of the software agent for cognitive behaviours, the Agent Behaviour Representation (ABR) formalism [37] is used. For reactive agent behaviours, AUML formalisms are used, specifically state charts. For interactions, protocol diagrams from AUML are used. 1. Create a society of cognitive agents. Incorporate the informational flow. 2. Create a society of reactive agents. the physical Incorporate flow and the related exchanged objects (products). 3. Define the responsibility links between cognitive and reactive agents. 5. Specify agent behaviour of the cognitive society using the Agent Behaviour Representation (ABR) formalism. 6. Specify agent’s behaviour of the reactive society using statecharts. 7. Specify agents’ interactions through protocol diagrams. system Multi-agent architecture: cognitive agents are seen as decision agents the decision (from system); reactive agents are represented by execution agents (from the execution system). the Specification of agent: software same elements, knowledge, i.e. behaviour and interactions. used We adapted only diagrams from AUML. For behaviours and knowledge representation, we employ Activity Diagrams. For interactions, we use Protocol Diagrams. Same process, but with different formalisms from AUML. The next sub-section provides some final remarks and conclusions about the proposed deployment strategy. 6. Final Remarks and Future Works This paper presents a conversion strategy from the FAMASS analysis models into specification and design models inspired by the methodological agent-based framework of Labarthe et al. [9]. This strategy facilitates the FAMASS analysts in converting their models and In addition, this deployment strategy demonstrates that the analysis phase of FAMASS can be integrated with other existing approaches specialized in specification and design modelling. With this as an impetus, other methodological frameworks could be inspected in the future so as to verify that FAMASS concepts adhere to other frameworks. Furthermore, the proposed strategy allows us to avoid the research effort needed to develop a totally new specification and design methodology for the domain, although it would be suitable and desirable for future research initiatives. With regard to this, a forthcoming research effort will work on extending the FAMASS analysis approach, so as to cover the whole FAMASS life- cycle from analysis to simulation. In this way the proposed deploying strategy launches the basis for this FAMASS- extended version of a complete architecture to deal with agent-based simulations in the context of distributed APS systems. Future versions of the FAMASS approach are to be published shortly. References [1] APICS, “The Association of Operations Management - Online Dictionary”, Retrieved June, 2008, from www.apics.org. [2] K. Kumar, "Technology for supporting supply chain management", Communications of the ACM, Vol. 44, Bo. 6, 2001, pp. 58-61. [3] M. Van Eck, Advanced planning and scheduling: is logistics everything? Working Paper, Vrije Universiteit Amsterdam, Amsterdam, 2003. [4] J.-M. Frayret, S. D'Amours, A. Rousseau, S. Harvey, S., and J. Gaudreault, "Agent-based supply chain planning in the forest products industry", International Journal of Flexible Manufacturing Systems, Vol. 19, No. 4, 2007, pp. 358-391. [5] A. L. Azevedo, C. Toscano, J. P. Sousa, and A. L. Soares, "An advanced agent-based order planning system for dynamic networked enterprises", Production Planning & Control, Vol. 15, No. 2, 2004, pp. 133–144. [6] L. Cecere, "A changing technology landscape", Supply Chain Management Review, Vol. 10, No. 1, 2006. [7] J.-H. Lee, and C.-O. Kim, "Multi-agent systems applications in manufacturing systems and supply chain management: a review paper" International Journal of Production Research, Vol. 46, No. 1, 2008, pp. 233-265. [8] W. Shen, Q. Hao, H. L. Yoon, and D. H. Norrie, "Applications of agent-based systems in intelligent manufacturing: An review", Advanced updated Engineering Informatics, Vol. 20, No. 4, 2006, pp. 415- 431. [9] O. Labarthe, B. Espinasse, A. Ferrarini, A., and B. Montreuil, "Toward a methodological framework for agent-based modelling and simulation of supply chain in a IJCSI International Journal of Computer Science Issues, Vol. 8, Issue 3, No. 1, May 2011 ISSN (Online): 1694-0814 www.IJCSI.org 17 mass customization context", Simulation Modelling Practice and Theory, Vol. 15, No. 2, 2007, pp. 113-136. [10] H. Baumgaertel, and U. John, “Combining agent-based supply net simulation and constraint technology for highly efficient simulation of supply networks using APS systems”, in 2003 Winter Simulation Conference, 2003. [11] J. M. Swaminathan, S. F. Smith, and N. M. Sadeh, "Modeling supply chain dynamics: a multiagent approach", Decision Sciences, Vol. 29, No. 3, 1998, pp. 607-632. [12] L. A. Santa-Eulalia, S. D’Amours, and J.-M. Frayret, "Essay on conceptual modeling, analysis and illustration of agent-based simulations for distributed supply chain planning", INFOR Information Systems and Operations Research Journal, Vol. 46, No. 2, 2008, pp. 97-116. [13] S. Galland, F. Grimaud, P. Beaune, and J. P. Campagne, "MAMA-S: an introduction to a methodological approach for the simulation of distributed industrial systems", International Journal of Production Economics, No. 85, 2003, pp. 11–31. [14] R. Govindu, and R.B. Chinnam, "A software agent- component based framework for multi-agent supply chain modelling and simulation", International Journal of Modelling and Simulation, Vol. 30, No. 2, 2010. [15] L. A. Santa-Eulalia, G. Halladjian, S. D’Amours, and J.-M. Frayret, “Integrated methodological frameworks for modelling agent-based APS systems: a systematic literature review”, Working Paper CIRRELT-2011-50, CIRRELT – Interuniversity Research Centre on Enterprise Networks, Logistics and Transportation, 2011, available at www.cirrelt.ca. [16] M. S. Fox, J. F. Chionglo, and M. Barbuceanu, "The integrated supply chain management system", Internal Report - Department of Industrial Engineering, University of Toronto, Canada, from www.eil.utoronto.ca/iscm- descr.html, 1993. [17] H. Stadtler, and C. Kilger, Supply chain management and advanced planning: concepts, models, software and case studies, Berlin: Springer, 2004. [18] D. J. Van der Zee, and J.G.A.J. Van der Vorst, "A Modeling framework for supply chain simulation: opportunities for improved decision making", Decision Sciences, Vol. 36, No. 1, 2005, pp. 65-95. [19] P. Egri, and J. Vancza, “Cooperative planning in the supply network – a multiagent organization model”, in CEEMAS 2005 - 4th International Central and Eastern European Conference on Multi-Agent Systems, Budapest, Hungary, Springer Verlag, 2005. [20] P. Lendermann, B. P. Gan, and L. F. McGinnis, “Distributed simulation with incorporated APS procedures for high-fidelity supply chain optimization’, in 2001 Winter Simulation Conference, Arlington, 2001. [21] L. A. Santa-Eulalia, S. D'Amours, and J.-M. Frayret, “Modeling Agent-Based Simulations for Supply Chain Planning: the FAMASS Methodological Framework”, in 2010 IEEE International Conference on Systems, Man, and Cybernetics, Collaborative on Session Special Manufacturing and Supply Chains, Istanbul, 10-13 October 2010. [22] L. A. Santa-Eulalia, “Agent-based for simulations advanced supply chain planning: a methodological framework for requirements analysis and deployment”, Ph.D. Thesis, Faculté des Sciences et Génie, Université Laval, Canada, 2009. 387p. [23] L. A. Santa-Eulalia, S. D’Amours, and J.-M. Frayret, “Agent-Based Simulations for Advanced Supply Chain Planning: The FAMASS Methodological Framework for Requirements Analysis and Deployment”. Working Paper CIRRELT-2011-22, CIRRELT – Interuniversity Research Centre on Enterprise Networks, Logistics and Transportation, 2011, available at www.cirrelt.ca. [24] L. A. Santa-Eulalia, D. Aït-Kadi, S. D’Amours, J.-M. Frayret, and S. Lemieux, “Agent-based experimental investigations about the robustness of tactical planning and control policies in a softwood lumber supply chain”, Production Planning & Control, Special Issue on Applied Simulation, Planning and Scheduling Techniques in Industry, 1366-5871, first published on February 10th 2011 (iFirst). [25] OMG, “OMG Systems Modeling Language (OMG SysML™)”, Object Management Group Specification Report, June 2010. [26] J. F. Shapiro, “Modeling the supply chain”, Duxbury: Pacific Grove, 2000. [27] H. Meyr, and H. Stadtler, “Types of supply chain”, in Supply chain management and advanced planning: concepts, models, software and case studies, Berlin: Springer, 2004. [28] W. Shen, F. Maturana, and D. H. Norrie, "MetaMorph II: intelligent an agent-based architecture for distributed design and manufacturing", Journal of Intelligent Manufacturing, No. 11, 2000, pp. 237-251. [29] J. Ferber, Multi-agent systems: an to introduction distributed artificial intelligence, Harlow: Addison-Wesley, 1999. [30] L. Sanya, and W. Hongwei, "Agent architecture for agent- based supply chain integration & coordination", Software Engineering Notes, Vol. 28, No. 4, 2003. [31] J.-M. Frayret, S. D'Amours, B. Montreuil, and L. Cloutier, , "A network approach to operate agile manufacturing systems", International Journal of Production Economics, Vol. 74, No. 1-3, 2001, pp. 239-259. [32] B. Montreuil, J.-M. Frayret, and S. D'Amours, "A strategic framework for networked manufacturing", Computers in Industry, Vol. 42, No. 2-3, 2000, pp. 299-317. [33] B. Montreuil, and P. Lefrançois, "Organizing factories as responsibility networks", Progress in Material Handling Research, 1996, pp. 375-411. [34] O. Labarthe, “Modélisation et simulation orientées agents de chaînes logistiques dans un contexte de personnalisation de masse : modèles et cadre méthodologique”, Ph.D. Thesis, Université Laval (Canada) and Université Paul Cézanne (France), 2006. [35] O. Labarthe, B. Montreuil, A. Ferrarini, and B. Espinasse, “Modelisation multi-agents pour la simulation de chaines logistiques de type personnalisation de masse”, in 5e Conférence Francophone de MOdélisation et SIMulation: Modélisation et simulation pour l’analyse et l’optimisation des systèmes industriels et logistiques, MOSIM’04. Nantes, 2004. [36] O. Labarthe, E. Tranvouez, A. Ferrarini, B. Espinasse, and B. Montreuil, “A heterogeneous multi-agent modelling for IJCSI International Journal of Computer Science Issues, Vol. 8, Issue 3, No. 1, May 2011 ISSN (Online): 1694-0814 www.IJCSI.org 18 distributed simulation of supply chains” in HoloMAS 2003, pp. 134-145, 2003. [37] E. Tranvouez, “IAD et ordonnancement : une approache coopérative du réordonnancement par systèmes multi- agents”, Ph.D. Thesis, Université de Valenciennes et du Hainaut-Cambrésis, 2001. Luis Antonio de Santa-Eulalia is a professor at Téluq-UQAM (Université du Québec à Montréal), Canada, a member of the innovation board of Axia Value Chain (North America division), and a researcher of the NSERC Strategic Research Network on Value Chain Optimization (VCO). He holds a Ph.D. Industrial in Engineering from Université Laval, Canada, an MSc. and BSc. both in Industrial Engineering respectively from the University of São Paulo, Brazil, and Federal University of São Carlos, Brazil. He has worked as a researcher and consultant in the domains of production planning and control, supply chain management, and simulations. His current research interests are related to novel business models and technology for sustainable value chain management. Sophie DʼAmours holds a Ph.D. in Applied Mathematics and Industrial Engineering from the École Polytechnique de Montréal, as well as a MBA and a BSc in Mechanical Engineering from Université Laval. She is currently a scientific director of the NSERC Strategic Research Network on Value Chain Optimization (VCO) and she holds a Canada Research Chair in Planning Sustainable Forest Value Networks as well as an NSERC Industrial Chair. She is also Director of the FORAC Research Consortium. Sophie is a full professor at the Faculty of Science and Engineering, Department of Mechanical Engineering, at Université Laval. Her research interests are in supply chain management and planning, web-based applications, and forest sector. Jean-Marc Frayret the École is Associate Professor at Polytechnique de Montréal, Québec, Canada. He holds a Ph.D. in Mechanical and Industrial Engineering from Université Laval, Canada. He is a member of the CIRRELT, a research centre dedicated to the study of network organizations and logistics. He is also a researcher of the NSERC Strategic Research Network on Value Chain Optimization (VCO) and of the FORAC Research Consortium. His research interests include agent-based and distributed manufacturing systems, supply chain management and interfirm collaboration. Dr. Frayret has published several articles in these fields in various journals and international conferences.
1802.03427	1	1802	2018-02-09T19:40:01	Structural matrix algebras, generalized flags and gradings	[ "math.RA" ]	We show that a structural matrix algebra $A$ is isomorphic to the endomorphism algebra of an algebraic-combinatorial object called a generalized flag. If the flag is equipped with a group grading, an algebra grading is induced on $A$. We classify the gradings obtained in this way as the orbits of the action of a double semidirect product on a certain set. Under some conditions on the associated graph, all good gradings on $A$ are of this type. As a bi-product, we obtain a new approach to compute the automorphism group of a structural matrix algebra.	math.RA	math	STRUCTURAL MATRIX ALGEBRAS, GENERALIZED FLAGS AND GRADINGS F. BES¸LEAG A AND S. D ASC ALESCU Abstract. We show that a structural matrix algebra A is isomorphic to the endomorphism algebra of an algebraic-combinatorial object called a generalized flag. If the flag is equipped with a group grading, an algebra grading is induced on A. We classify the gradings obtained in this way as the orbits of the action of a double semidirect product on a certain set. Under some conditions on the associated graph, all good gradings on A are of this type. As a bi-product, we obtain a new approach to compute the automorphism group of a structural matrix algebra. 8 1 0 2 b e F 9 ] . A R h t a m [ 1 v 7 2 4 3 0 . 2 0 8 1 : v i X r a 1. Introduction and preliminaries Flags and flag varieties play a key role in Algebraic Geometry, Representation Theory, Alge- braic Groups and Combinatorics, see [6]. In this paper we consider a more general concept of flag and we give some applications. This new kind of flag arises as follows. Let k be a field. A structural matrix algebra is a subalgebra of a full matrix algebra over k, consisting of all matrices with zero entries on certain prescribed positions, and allowing anything on the remaining positions. An important number of examples and counterexamples use such algebras. These algebras were called structural matrix algebras in [10], but they had already been considered in [8]. Particular examples of structural matrix algebras are upper triangular matrix algebras, and more generally, upper block triangular matrix algebras, which are of fundamental importance in Linear Algebra, the study of linear groups, the study of numerical invariants of PI algebras, etc. A structural matrix subalgebra A of Mn(k) is associated with a preorder relation (i.e. reflexive and transitive) ρ on the set {1, . . . , n}; A consists of all matrices (aij)1≤i,j≤n such that aij = 0 whenever (i, j) /∈ ρ. We denote A = M (ρ, k); in other terminology, this is the incidence algebra over k associated with ρ, see [9]. Automorphisms of structural matrix algebras are of interest, and the problem to understand them gets more complicated if we take k to be just a ring. If k is a field, the description of the automorphism group of M (ρ, k) was done in [5]. A general problem in Ring Theory is to describe and classify all group gradings on various matrix algebras. In the case of full matrix algebras over a field k, the problem was solved in [1] and [2] for algebraically closed k, and descent theory was used in [4] to approach the case of an arbitrary k. For subalgebras of a full matrix algebra, in particular for structural matrix algebras, it is much more complicated to describe all gradings. The aim of our paper is to construct and classify a certain class of gradings on structural matrix algebras. We show in Section 2 that in the same way the full matrix algebra Mn(k) is the endomorphism group of a vector space of dimension n, a structural matrix algebra M (ρ, k) is isomorphic to the endomorphism algebra of a certain algebraic-combinatorial structure F, which we call a ρ-flag. In Section 5 we explain that if F is additionally equipped with a G-grading, where G is a group, then its endomorphism algebra End(F) gets an induced G-graded algebra structure; we denote by END(F) the obtained G-graded algebra. This grading transfers to a G-grading on 2010 Mathematics Subject Classification. 16W50, 16W20, 16S50, 06A06. Key words and phrases. structural matrix algebra, preorder relation, flag, group grading, automorphism group. 1 2 F. BES¸LEAG A AND S. D ASC ALESCU M (ρ, k) via the isomorphism mentioned above. The gradings produced in this way on M (ρ, k) are good gradings, i.e. all the matrix units present in M (ρ, k) are homogeneous elements. It is an interesting question whether all good gradings are obtained like this. This is a problem of independent interest, and it can be formulated in simple terms related to the graph Γ associated with ρ: if G is a group, and on each arrow of Γ we write an element of G as a label, such that for any two paths starting from and terminating at the same points the product of the labels of the arrows is the same for both paths, does the set of labels arise from a set of weights on the vertices of Γ, in the sense that an arrow starting from v1 and terminating at v2 has label g1g−1 2 , where g1 and g2 are the weights of v1 and v2? This problem was considered in [8] in the case where G is abelian, and it was showed that the answer is positive if and only if the cohomology group H 1(∆, G) = 0, where ∆ is a certain simplicial complex associated with ρ. Also, for a given ρ, the answer to the above question is positive for any abelian group G if and only if the homology group H1(∆) = 0. We show that the answer is positive for any arbitrary group G if and only if the normal closure of two certain subgroups A(Γ) ⊆ B(Γ) of the free group generated by the arrows of Γ coincide; A(Γ) and B(Γ) are defined in terms of cycles of the un-directed graph obtained from Γ. This parallels the result in the abelian case, where H1(∆) = B/A for similar subgroups A and B in a free abelian group associated with Γ. In fact we use slightly different A and B, by working with a different graph. In order to classify G-gradings on the structural matrix algebra M (ρ, k), we first look at the isomorphisms between the algebras End(F) and End(F ′), where F and F ′ are ρ-flags under the vector spaces V and V ′. An equivalence relation ∼ on {1, . . . , n} arises from ρ, where i ∼ j whenever iρj and jρi. Then ρ induces a partial order on the set C of equivalence classes with respect to ∼. We show in Section 3 that the End(F)-submodules of V are in a bijective correspondence with the antichains of C; let A(C) be the lattice structure on the set of all such antichains, induced via this bijection. Then an algebra isomorphism ϕ : End(F) → End(F ′) induces a linear isomorphism γ : V → V ′ which is a ϕ′-isomorphism for a certain deformation of ϕ. The new algebra isomorphism ϕ′ is obtained from ϕ by using a transitive function on ρ with values in k∗. Since ϕ′ is an algebra isomorphism, γ induces an isomorphism between the lattices of End(F)-submodules of V and of End(F ′)-submodules of V ′, and this lattice isomorphism reduces in fact to an automorphism of the lattice A(C). Such an automorphism is completely determined by an automorphism g of the poset C. Moreover, we explain that ϕ can be recovered from g, the deformation constants producing ϕ′ from ϕ, and a matrix of γ in a fixed pair of bases. Thus we obtain that the set of algebra isomorphisms from End(F) to End(F ′) is in a bijective correspondence with the equivalence classes of a set involving the invertible matrices of M (ρ, k), the automorphisms of C preserving the cardinality of elements, and the transitive functions on ρ, with respect to an equivalence relation. In particular, if F ′ = F, the automorphism group of End(F) is described as a factor group of a double semidirect product. As a bi-product, we obtain a descriptive presentation of the automorphism group of a structural matrix algebra. This automorphism group was computed in [5], and we show how the presentation in [5] can be derived from ours. For classifying G-gradings arising from graded flags, we consider two G-graded ρ-flags F and F ′, and we look at the isomorphisms between the graded algebras END(F) and END(F ′). Using the structure of isomorphisms between End(F) and End(F ′), which we already know by now, and adding the additional information about gradings, we obtain in Section 6 that END(F) ≃ END(F ′) if and only if the connected components of F and F ′ are pairwise isomorphic up to a permutation, some graded shifts and an automorphism of C. Using this result, we show in Section 7 that the isomorphism types of graded algebras of the form END(F) are classified by 3 the orbits of the action of a certain group, which is a double semidirect product of a Young subgroup of Sn, a certain subgroup of automorphisms of C, and Gq, where q is the number of connected components of C, on the set Gn. We use the standard terminology on gradings, see for example [7]. 2. Structural matrix algebras as endomorphism algebras Let k be a field, n a positive integer and ρ a preorder relation on {1, . . . , n}. Let M (ρ, k) be the structural matrix algebra associated with ρ. Let ∼ be the equivalence relation on {1, . . . , n} associated with ρ, i.e. i ∼ j if and only if iρj and jρi, and let C be the set of equivalence classes. Then ρ induces a partial order ≤ on C defined by i ≤ j if and only if iρj, where i denotes the equivalence class of i. For any α ∈ C, let mα be the number of elements of α. Definition 2.1. A ρ-flag is an n-dimensional vector space V with a family (Vα)α∈C of subspaces Bα with the property that Bα = mα such that there is a basis B of V and a partition B = [ and [ α∈C Bβ is a basis of Vα for any α ∈ C. If F = (V, (Vα)α∈C) and F ′ = (V ′, (V ′ α)α∈C) are ρ- β≤α flags, then a morphism of ρ-flags from F to F ′ is a linear map f : V → V ′ such that f (Vα) ⊂ V ′ α for any α ∈ C. If ρ is just the usual ordering relation on {1, . . . , n} (which corresponds to M (ρ, k) being the algebra of upper triangular matrices), then a ρ-flag is just an usual flag on an n-dimensional vector space. More generally, if ρ is such that C = {α1, . . . , αr} is totally ordered, say α1 < . . . < αr, and αi = mi for any 1 ≤ i ≤ r, then a ρ-flag is a flag of signature (m1, . . . , mr), and M (ρ, k) is the algebra   Mm1 (k) Mm1,m2(k) 0 . . . 0 Mm2(k) . . . 0 . . . Mm1,mr (k) . . . Mm2,mr (k) . . . . . . Mmr (k) . . .   of upper block triangular matrices, with diagonal blocks of size m1, . . . , mr. For any i, j ∈ {1, . . . , n} with iρj, let eij be the matrix in Mn(k) having 1 on the (i, j)-position and 0 elsewhere. The set of all such eij's is a basis of M (ρ, k). Now we present M (ρ, k) as an algebra of endomorphisms. Proposition 2.2. Let F = (V, (Vα)α∈C) be a ρ-flag. Then the algebra End(F) of endomorphisms of F (with the map composition as multiplication) is isomorphic to M (ρ, k). Proof. Let B be a basis of V as in Definition 2.1. We can choose some set of indices such that B = (vi)1≤i≤n and for any α ∈ C the basis Bα of Vα is just {vii ∈ α}. If i, j ∈ {1, . . . , n}, then iρj if and only if i ∈ α and j ∈ β for some α, β ∈ C with α ≤ β; in this case let Eij ∈ End(F) be defined by Eij(vt) = δjtvi for any t (δij is Kronecker's delta). Then it is easy to see that EijEpq = δjpEiq for any i, j, p, q and {Eij iρj} is a basis of End(F). Hence the linear map Eij 7→ eij for any i, j with iρj, is an algebra isomorphism between End(F) and M (ρ, k). (cid:3) We associate with ρ another useful combinatorial object. Let Γ = (Γ0, Γ1) be the graph whose set Γ0 of vertices is the set C of equivalence classes. The set Γ1 of arrows is constructed as follows: if α, β ∈ C, there is an arrow a from α to β (we write s(a) = α, t(a) = β) if α < β and there is no γ ∈ C with α < γ < β. Clearly, if α, β ∈ C, then α ≤ β if and only if there is a 4 F. BES¸LEAG A AND S. D ASC ALESCU path in Γ starting from α and ending at β (recall that paths of length zero are just vertices of Γ). We denote by Γu the undirected graph obtained from Γ when we forget the orientation of arrows. 3. The lattice of End(F)-submodules of V Let F = (Vα)α∈C be a ρ-flag on the space V . The aim of this section is to describe the lattice LEnd(F )(V ) of submodules of the End(F)-module V , where the action of End(F) on V is just the restriction of the usual End(V )-action on V . We also determine the automorphisms of this lattice. If D is a subset of C, we denote by VD = X α∈D Vα. By convention V∅ = 0. Proposition 3.1. The End(F)-submodules of V are the subspaces of the form VD, where D is a subset of C. {Vi there exists α ∈ D such that i ≤ α}, and Epqvi = δq,ivp α∈D nX i=1 Vα = for any p, q with p ≤ q, and any i, it is clear that VD is an End(F)-submodule of V . Proof. Since VD = X Conversely, let X be an End(F)-submodule of V . If v = Pn for some i0, then for any j with bj ≤ bi0 we have Eji0v = ai0vj, so vj ∈ X, thus V bi0 i=1 aivi ∈ X − {0} with ai0 6= 0}, we obtain that VD = X D = {bi0 there exists v = Pn i=1 aivi ∈ X − {0}, and ai0 6= 0 ⊂ X. If Obviously X ⊂ VD, so then X = VD. α∈D Vα ⊂ X. (cid:3) Remark 3.2. If D ⊂ C and Dmax is the set of maximal elements of D (with respect to the partial order of C), it is clear that VD = VDmax. As Dmax is an antichain in C (i.e. a subset of C whose any two different elements are not comparable with respect to ≤), and for any antichain D we have Dmax = D, we conclude that the End(F)-submodules of V are VD with D an antichain in C. If we denote by A(C) the set of all antichains of C, we have a bijection between A(C) and LEnd(F )(V ), given by D 7→ VD. Its inverse is X 7→ Dmax, where D is a subset of C such that X = VD. This bijection induces a lattice structure on A(C) from the lattice LEnd(F )(V ). Since for D, E ∈ A(C) we have {Vα there exist β1 ∈ D, β2 ∈ E such that α ≤ β1 and α ≤ β2} VD ∩ VE = X α∈C and we see that the infimum and the supremum in A(C) are given by VD + VE = VD∪E , D ∧ E = {α ∈ C there exist β1 ∈ D, β2 ∈ E such that α ≤ β1 and α ≤ β2}max, D ∨ E = (D ∪ E)max for any D, E ∈ A(C). Note that (D ∪ E)max may be strictly contained in D ∪ E, since D ∪ E is not necessarily an antichain. The partial order relation on A(C) is D ≤ E ⇐⇒ VD ⊂ VE ⇐⇒ for any α ∈ D there exists β ∈ E such that α ≤ β. 5 Note that this is not the inclusion, thus A(C) is not in general a sub-poset of the power set (P(C), ⊆). The next result describes the automorphisms of the lattice A(C). Proposition 3.3. If g is an automorphism of the poset (C, ≤), then the map fg : A(C) −→ A(C), fg(D) = g(D) = {g(α) α ∈ D} is an automorphism of the lattice A(C). Moreover, for any lattice automorphism f of A(C) there exists an automorphism g of the poset (C, ≤) such that f = fg. Proof. It is straightforward to check the first part. Now let f be a lattice automorphism of A(C). Let L0(C) be the set of all minimal elements of C, L1(C) be the set of all minimal elements of C \ L0(C) (or equivalently, elements of height 1 in C) and recurrently we define Lh(C) for any h, as the set of all minimal elements of C \ (L0(C) ∪ · · · ∪ Lh−1(C)) (or equivalently, elements of height h in C). The set L0 of minimal elements of A(C) consists of all singletons {α}, with α ∈ L0(C). As f is a lattice automorphism, we have f (L0) = L0, and this induces a bijection g0 : L0(C) −→ L0(C). We note that g0 (or f L0) uniquely determines the value of f at any non-empty D ⊂ L0. Next A(C) \ P(L0) is a poset with the order inherited from A(C), such that for any two elements their supremum exists. Moreover, f induces by restriction an isomorphism of posets f : A(C) \ P(L0) −→ A(C) \ P(L0) (there is no harm if we also denote it by f ), such that f (x ∨ y) = f (x) ∨ f (y) for any x, y. The set L1 of minimal elements of A(C) \ P(L0) consists of all singletons {α}, where α is minimal in C \ L0(C), i. e. α ∈ L1(C). Since f (L1) = L1, f induces a bijection g1 : L1(C) −→ L1(C), and g0 and g1 uniquely determine the value of f at any antichain D with D ⊂ L0(C) ∪ L1(C). We continue recurrently, by considering for any h the set Lh of minimal elements in A(C) \ P(L0 ∪ · · · ∪ Lh−1). This consists of all singletons {α}, where α ∈ Lh(C). As above, f induces an automorphism of the poset A(C) \ P(L0 ∪ · · · ∪ Lh−1), hence a bijection gh : Lh(C) −→ Lh(C). As Lh(C) = C, the coproduct (i.e. disjoint union) g of all gh's is an automorphism of the poset [ h C, and it is clear that f = fg. (cid:3) 4. Isomorphisms between endomorphism algebras of flags We consider the set Aut0(C, ≤) = {g ∈ Aut(C, ≤) mα = mg(α) for any α ∈ C}, For any g ∈ Aut0(C) we define a bijection which is a subgroup of Aut(C). g : {1, . . . , n} → {1, . . . , n} as follows: if α = {i1, . . . , ir} with i1 < . . . < ir, and g(α) = {j1, . . . , jr} with j1 < . . . < jr, then g(i1) = j1, . . . , g(ir) = jr. If g, h ∈ Aut0(C), we clearly have fhg = hg, thus g 7→ g is an embedding of Aut0(C) into the symmetric group Sn. If A ∈ Mn(k), let Ag ∈ Mn(k) be the matrix whose (i, j)-entry is the (i, g(j))-entry of A, and let gA ∈ Mn(k) be the matrix whose (i, j)-entry is the (g(i), j)-entry of A. If g, h ∈ Aut0(C), then (Ag)h = Agh, h(gA) = ghA, (gA)h = g(Ah) for any A ∈ Mn(k). Also (AB)g = ABg, g(AB) = (gA)B, (Ag)B = A(g−1 B) 6 F. BES¸LEAG A AND S. D ASC ALESCU for any A, B ∈ Mn(k) and any g ∈ Aut0(C). It follows that if A is invertible, then so is Ag, and its inverse is g(A−1). Let T = {(aij)iρj ⊂ k∗ aijajr = air for any i, j, r with iρj, jρr}. Using the terminology of Section 5, T can be identified with the set of transitive k∗-valued functions on ρ. Multiplication on positions (i.e. pointwise multiplication of functions) makes T a group. Let F be a ρ-flag on the space V , and F ′ be a ρ-flag on the space V ′. We keep the notation of Section 2 for F, the basis of V , and the associated Eij's. Thus we fix a basis (vi)i of V such i i ≤ α} is a that {vi i ≤ α} is a basis of Vα, and similarly a basis (v′ basis of V ′ ij)iρj be the basis of End(F ′) associated with (v′ i)i of V ′ such that {v′ α for any α; let (E′ i)i. Define F : U (M (ρ, k)) × Aut0(C) × T −→ Isoalg(End(F), End(F ′)) as follows. If A ∈ U (M (ρ, k)), g ∈ Aut0(C) and (aij)iρj ∈ T , let (wi)1≤i≤n be the basis of V ′ defined by (. . . , wi, . . .) = (. . . , v′ i, . . .) Ag. Then for any i, j with iρj let Fij ∈ End(V ′) be such that Fij(wj) = aijwi for any iρj, and Fij(wr) = 0 for any r 6= j. Clearly FijFjr = Fir for iρj, jρr. Lemma 4.1. With the above notation, let A = (λpq)p,q and A−1 = (λpq)p,q. Then Fij = aij X sρg(i) g(j)ρt λsg(i)λg(j)tE′ st. In particular Fij ∈ End(F ′). Proof. Since (. . . , v′ i, . . .) = (. . . , wi, . . .)(Ag)−1 = (. . . , wi, . . .) g(A−1), we see that v′ t = X p λg(p)twp for any t. Then Fij(v′ t) = X p λg(p)tFij(wp) = aijλg(j)twi = aij X s λg(j)tλsg(i)v′ s. The coefficient of v′ thus it is zero unless s ≤ t. This shows that s in the last sum may be non-zero only if s ≤ dg(i) = g(i) and g(j) = dg(j) ≤ t, λsg(i)λg(j)tE′ st. Fij = aij X sρg(i) g(j)ρt Now define F (A, g, (aij )iρj) = ϕ, where ϕ : End(F) −→ End(F ′) is the linear map such that ϕ(Eij) = Fij for any iρj. Clearly ϕ is an algebra isomorphism. Proposition 4.2. F is surjective. Proof. Let ϕ : End(F) −→ End(F ′) be an algebra isomorphism. Denote Fij = ϕ(Eij) for any i, j with iρj. Then (Fij)iρj is a basis of End(F ′), FijFjr = Fir for any i, j, r with iρj and jρr, and (Fii)i is a complete set of orthogonal idempotents of End(F ′). (cid:3) It is easy to see that V ′ = nM i=1 Qi, where Qi = ImFii 6= 0, so dim Qi = 1 for any i. Choose 7 some non-zero wi ∈ Qi for any i. Then (wi)i is a basis of V ′. Since Fij = ϕ(Eij) 6= 0, Fij(Qr) = FijFrr(V ′) = 0 for any r 6= j, and Fij(Qj) = FiiFij(Qj) ⊂ Qi, we see that Fij(wj) = aijwi for some aij ∈ k∗. Since FijFjr = Fir, we must have aijajr = air for any i, j, r with iρj and jρr. In particular, aii = 1 for any i. Let γ : V −→ V ′ be the linear isomorphism such that γ(vi) = wi for any i. Regard V as a left End(F)-module, and V ′ as a left End(F ′)-module with the usual action of the endomorphism algebra. If iρj we have γ(Eijvt) = γ(δjtvi) = δjtwi, and Fijγ(vt) = Fijwt = δjtaijwi for any t. We get (1) ϕ(Eij )γ(vt) = aijγ(Eijvt) for any iρj and any t. Thus in general γ is not a ϕ-isomorphism, the obstruction being the scalars aij. However, γ is a ϕ′-isomorphism for a deformation ϕ′ of ϕ. Indeed, the linear map θ : End(F) −→ End(F) defined by θ(Eij) = a−1 ij Eij for any iρj, is an algebra automorphism, and (1) shows that (2) γ(Eijvt) = (ϕθ)(Eij )γ(vt) for any iρj, and any t. Thus γ is a ϕ′-isomorphism, where ϕ′ = ϕθ : End(F) −→ End(F ′) is also an algebra isomorphism. Then the lattice of End(F)-submodules of V is isomorphic to the lattice of End(F ′)-submodules of V ′ via the map γ : LEnd(F )(V ) −→ LEnd(F ′)(V ′), γ(X) = γ(X) for any End(F)-submodule X of V. By Remark 3.2, there is an isomorphism of lattices Φ : LEnd(F )(V ) −→ A(C), given by Φ(X) = Dmax, where X = VD; its inverse is Φ−1(D) = VD for any D ∈ A(C). Similarly, there is an isomorphism of lattices Φ′ : LEnd(F ′)(V ′) −→ A(C). Let f : A(C) −→ A(C) be the isomorphism of lattices such that the diagram LEnd(F )(V ) γ ∼ / / LEnd(F ′)(V ′) Φ ∼ A(C) f ∼ Φ′ / A(C) is commutative. By Proposition 3.3, f = fg for some automorphism g of the poset (C, ≤). Then for any α ∈ C γ(Vα) = γ(Vα) = (Φ′−1fgΦ)(Vα) = (Φ′−1fg)({α}) = Φ′−1({g(α)}) = V ′ g(α). This shows that dim Vα = dim V ′ g(α) = dim Vg(α) for any α ∈ C. Since dim Vα = X mβ, we see β≤α by induction on the length of α that mα = mg(α) for any α ∈ C, so g ∈ Aut0(C). / 8 F. BES¸LEAG A AND S. D ASC ALESCU Let M be the matrix of γ in the bases (vi)i and (v′ g(α), M may have non-zero entries only on positions (i, j) with i ≤ g(j), i.e. on blocks (α, β) with α ≤ g(β). We see that A = M g−1 ∈ M (ρ, k) and M = Ag. Moreover, A is invertible and A−1 = g−1 i)i. Since γ(Vα) = V ′ (M −1). Now we have ϕ = F (A, g, (aij )iρj), and this ends the proof. (cid:3) We consider the relation ≈ on U (M (ρ, k)) × Aut0(C) × T defined by (A, g, (aij )iρj) ≈ ij = did−1 (B, h, (bij )iρj) if and only if g = h and there exist d1, for any iρj, and Bg = Ag diag(d1, . . . , dn) (or equivalently AB−1 = g−1 where D = diag(d1, . . . , dn)); here diag(d1, . . . , dn) denotes the diagonal matrix with diagonal en- tries d1, . . . , dn. This is clearly an equivalence relation. ..., dn ∈ k∗ such that aijb−1 (D−1)g−1 j Theorem 4.3. With the above notation, F (A, g, (aij )iρj) = F (B, h, (bij )iρj) if and only if (A, g, (aij )iρj) ≈ (B, h, (bij)iρj ). Thus F induces a bijection F : U (M (ρ, k)) × Aut0(C) × T ≈ −→ Isoalg(End(F), End(F ′)). Proof. Denote F (A, g, (aij )iρj) = ϕ and F (B, h, (bij )iρj) = ψ. Thus ϕ(Eij) = Fij, where Fij(wj) = aijwi; here (. . . , wi, . . .) = (. . . , u′ j) = bijw′ i and (. . . , w′ for any iρj. Now j = djwj for some dj ∈ k∗, for any j. ImFij =< wj > and ImF ′ Thus (. . . , w′ i, . . .) = (. . . , wi, . . .)diag(d1, . . . , dn), showing that Bh = Ag diag(d1, . . . , dn). Since ϕ = ψ, then Fij = F ′ ij i, . . .)Ag. Also ψ(Eij) = F ′ j >, so we must have w′ i, . . .) = (. . . , u′ ij, where F ′ i, . . .)Bh. ij =< w′ ij(w′ The (α, β)-block of Bh may be non-zero only if α ≤ h(β), and the (h(β), β)-block of Bh is non- zero (since the (β, β)-block of B is non-zero). Similarly, the (α, β)-block of Ag may be non-zero only if α ≤ g(β); the same holds for Ag diag(d1, ..., dn). Since Bh = Ag diag(d1, ..., dn), the (h(β), β)-block of Ag must be non-zero, so then h(β) ≤ g(β). Similarly, since the (g(β), β)-block of Ag is a non-zero, we obtain that g(β) ≤ h(β). Thus g = h. On the other hand, F ′ ij(wj) = F ′ ij(d−1 j w′ j) = d−1 j bijw′ i = did−1 j bijwi, so Fij = F ′ ij requires aij = did−1 j bij for any iρj. We conclude that (A, g, (aij )iρj ) ≈ (B, h, (bij )iρj). Conversely, the above computations show that the F ′ ij's associated to each of the two triples (cid:3) are the same. Let us fix some notation. If G and A are groups, a right action of G on A is a mapping A × G −→ A, (a, g) 7→ a · g, such that (ab) · g = (a · g)(b · g) and (a · g) · h = a · (gh) for any a, b ∈ A, g, h ∈ G. This is equivalent to giving a group morphism ϕ : G −→ Aut(A), where the multiplication of Aut(A) is the opposite map composition. Indeed, one can take ϕ(g)(a) = a · g. In this case, the right crossed product G ⋉ A is the cartesian product G × A of sets, with multiplication (h ⋉ b)(g ⋉ a) = hg ⋉ (b · g)a for any g, h ∈ G, a, b ∈ A; here we denote the pair (g, a) by g ⋉ a. The group Aut0(C) acts to the right on T by (aij)iρj · g = (ag(i)g(j))iρj, thus we can form a right crossed product Aut0(C) ⋉ T . Similarly, a left action of G on A is a mapping G × A −→ A, (g, a) 7→ g · a, such that g · (ab) = (g · a)(g · b) and g · (h · a) = (gh) · a; this is the same with giving a group morphism G −→ Aut(A), where this time the multiplication of Aut(A) is just the map composition. The left crossed product A ⋊ G is the set A × G with the multiplication (a ⋊ g)(b ⋊ h) = a(g · b) ⋊ gh. Now if G ⋉ A is a right crossed product, and B is a group, then a left G ⋉ A-action on B is a pair consisting of a left G-action on B and a left A-action on B, such that (3) a · (g · x) = g · ((a · g) · x) for any a ∈ A, g ∈ G, x ∈ B. 9 The action of G ⋉ A on B is (g ⋉ a) · x = g · (a · x). Indeed, this easily follows from the fact that A, respectively G, embeds into G ⋉ A by a 7→ 1 ⋉ a, respectively g 7→ g ⋉ 1, and (1 ⋉ a)(g ⋉ 1) = g ⋉ (a · g) = (g ⋉ 1)(1 ⋉ (a · g)). For later use (in Section 7), we note that if G ⋉ A is a right crossed product, and B is just a set, then a right action of G ⋉ A on B is a pair consisting of a right action of G on B and a right action of A on B satisfying the compatibility condition (4) (x · a) · g = (x · g) · (a · g) for any x ∈ B, a ∈ A, g ∈ G. Also for later use, we mention that if A ⋊ G is a left semidirect product, and both A and G act to the right on a set B, then x · (a ⋊ g) = (x · a) · g defines a right action of A ⋊ G on B, provided that (5) (x · g) · a = (x · (g · a)) · g for any x ∈ B, a ∈ A, g ∈ G. We use some of these facts in the following particular situation. The group T acts to the left if (aij)iρj ∈ T and A = (αij)1≤i,j≤n ∈ U (M (ρ, k)), then on the group U (M (ρ, k)) as follows: (aij)iρj · A is the matrix whose (i, j)-spot is αijaij if iρj, and 0 elsewhere. We note that (aij)iρj · A ∈ U (M (ρ, k)), its inverse being just (aij)iρj · A−1. The group Aut0(C) . The two actions are compatible in the Ag−1 also acts to the left on U (M (ρ, k)) by g · A = g−1 sense of (3), since (6) (aij)iρj · g−1 Ag−1 = g−1 ((ag(i)g(j))iρj · A)g−1 for any (aij)iρj ∈ T and g ∈ Aut0(C). Indeed, if A = (αij)1≤i,j≤n, it is easy to check that both sides in (6) have aijαg−1(i)g−1(j) on the (i, j)-spot for any iρj. We obtain that Aut0(C) ⋉ T acts to the left on U (M (ρ, k)) by ((aij)iρj · A)g−1 (g ⋉ (aij)iρj) · A = g−1 , and we can form the left crossed product U (M (ρ, k)) ⋊ (Aut0(C) ⋉ T ). Its multiplication is given by (7) (B ⋊ (h ⋉ (bij)iρj)) · (A ⋊ (g ⋉ (aij)iρj )) = (B h−1 ) ⋊ (hg ⋉ (bg(i)g(j)aij)iρj). Now if we apply the construction of the map F to the case where F ′ = F, we obtain a surjective map ((bij)iρj · A)h−1 F : U (M (ρ, k)) ⋊ (Aut0(C) ⋉ T ) −→ Aut(End(F)). Theorem 4.4. F is a morphism of groups, and it induces a group isomorphism where D = {diag(d1, . . . , dn) ⋊ (Id ⋉ (d−1 U (M (ρ, k)) ⋊ (Aut0(C) ⋉ T ) ≃ Aut(End(F)), D i dj)iρj) d1, . . . , dn ∈ k∗}. Proof. Let ϕ = F (A ⋊ (g ⋉ (aij)iρj)), where A = (λij)i,j; denote A−1 = (λij)i,j. Let also ψ = F (B ⋊ (h ⋉ (bij)iρj)), with B = (µij)i,j and B−1 = (µij)i,j. We will show that (8) ψϕ = F ( (B ⋊ (h ⋉ (bij)iρj)) · (A ⋊ (g ⋉ (aij)iρj)) ) and this will prove that F is a group morphism. We know that ϕ(Eij) = aij X sρg(i) g(j)ρt λsg(i)λg(j)tEst, 10 F. BES¸LEAG A AND S. D ASC ALESCU λsg(i)λg(j)tψ(Est) so then ψϕ(Eij ) = aij X = aij X sρg(i) g(j)ρt sρg(i) g(j)ρt X pρh(s) h(t)ρq bstλsg(i)λg(j)tµph(s)µh(t)qEpq = aij X X pρhg(i) hg(j)ρq h−1(p)ρsρg(i) g(j)ρtρh−1(q) = aijbg(i)g(j) X ( X pρhg(i) hg(j)ρq h−1(p)ρsρg(i) bsg(i)bg(i)g(j)bg(j)tλsg(i)λg(j)tµph(s)µh(t)qEpq µph(s)bsg(i)λsg(i))( X g(j)ρtρh−1(q) bg(j)tλg(j)tµh(t)q)Epq. But the first bracket in the last row above is the product of the p-th row in Bh and the i-th column in ((bij )iρj ·A)g, thus it is the (p, i)-position in the matrix Bh((bij)iρj ·A)g, or equivalently, the (p,fhg(i))-position in Denote B h−1 (Bh((bij )iρj · A)g)(hg)−1 = Bh((bij)iρj · A)h−1 = B h−1 ((bij)iρj · A)h−1 . ((bij )iρj · A)h−1 = (νij)i,j, and the inverse of this matrix by (ν ij)i,j. A similar argument shows that the second bracket above is just νhg(j)q. Then ψϕ(Eij ) = aijbg(i)g(j) X νphg(i)ν hg(j)qEpq, pρhg(i) hg(j)ρq showing that ψϕ = F (B h−1 ⋊ (hg ⋉ (bg(i)g(j)aij)iρj)), which using (7) is just (8). Now z = A ⋊ (g ⋉ (aij)iρj) is in the kernel of F if and only if it is equivalent (via ≈) to the (cid:3) identity element. In view of Theorem 4.3, this is the same with z ∈ D. ((bij )iρj · A)h−1 Now we explain how the description of the automorphism group of M (ρ, k) given in [5] can be deduced from Theorem 4.4. We recall a few basic things about semidirect products. We first note that if a group G acts to the right on a group A, with action denoted by a · g for a ∈ A and g ∈ G, then there is also a left action of G on A, defined by g · a = a · g−1, and the associated right and left semidirect products are isomorphic. Indeed, φ : A ⋊ G → G ⋉ A, φ(a ⋊ g) = g ⋉ (g−1 · a), is a group isomorphism, with inverse φ−1(g ⋉ a) = a · g−1 ⋊ g. Secondly, if we have a double semidirect product B ⋊ (A ⋊ G), then the left action of A ⋊ G on B induces (via the usual embeddings of A and G) left actions of A and G on B. Moreover, G acts to the left on the group B ⋊ A by g · (x ⋊ a) = g · x ⋊ g · a for any g ∈ G, x ∈ B, a ∈ A, and the map γ : B ⋊ (A ⋊ G) → (B ⋊ A) ⋊ G, γ(x ⋊ (a ⋊ g)) = (x ⋊ a) ⋊ g, is an isomorphism of groups. Finally, if A ⋊ G is a left semidirect product, and N is a normal subgroup of A which is invariant to the action of G, then N ⋊ 1 is a normal subgroup of A ⋊ G and A⋊G ⋊ G, where the action of G on A N is the one induced by the action of G on A. N ⋊G ≃ A N Using these remarks, we see that U (M (ρ, k)) ⋊ (Aut0(C) ⋉ T ) ≃ U (M (ρ, k)) ⋊ (T ⋊ Aut0(C)) ≃ (U (M (ρ, k)) ⋊ T ) ⋊ Aut0(C) 11 and in the third double semidirect product the action of Aut0(C) on U (M (ρ, k)) ⋊ T is given by g · (A ⋊ (aij)iρj) = g−1 Ag−1 ⋊ (aij)iρj · g−1. The image of D through these isomorphisms in (U (M (ρ, k)) ⋊ T ) ⋊ Aut0(C) is D0 ⋊ 1, where D0 = {diag(d1, . . . , dn) ⋊ (d−1 i dj)iρj d1, . . . , dn ∈ k∗}. Clearly D0 is a normal subgroup in U (M (ρ, k))⋊T , since D is a normal subgroup of U (M (ρ, k))⋊ (Aut0(C) ⋉ T ), and it is easy to check that D0 is invariant under the action of Aut0(C). Thus we obtain that U (M (ρ, k)) ⋊ (Aut0(C) ⋉ T ) D ≃ ≃ (U (M (ρ, k)) ⋊ T ) ⋊ Aut0(C) D0 ⋊ 1 U (M (ρ, k)) ⋊ T D0 ⋊ Aut0(C). Following [5], we consider an undirected graph ∆, whose vertices are all elements i ∈ {1, . . . , n} such that the equivalence class i is not an isolated point in the poset (C, ≤); we say that α ∈ C is an isolated point if any of α ≤ β and β ≤ α implies that β = α. If i and j are vertices of ∆, there is an edge connecting i and j if and only if i and j are not equal, but they are in relation ≤ (in any possible way). If ∆1, . . . , ∆z are the connected components of ∆, choose a tree Tℓ in each ∆ℓ (note that there are several possible such choices). Now let G be the subgroup of T consisting of all (aij)iρj 's for which aij = 1 whenever iρj and i, j are vertices of ∆, joined by an edge of some Tℓ, or when both i and j lie in an isolated equivalence class in C. Also, let I be the group of inner automorphisms of M (ρ, k), i.e. I = {CA A ∈ U (M (ρ, k))}, where CA(X) = AXA−1 for any X ∈ M (ρ, k). Then there is a left action of G on I given by (aij)iρj · CA = C(aij )iρj ·A. We prove that (9) and this will show that U (M (ρ, k)) ⋊ T D0 ≃ I ⋊ G Aut(M (ρ, k)) ≃ U (M (ρ, k)) ⋊ (Aut0(C) ⋉ T ) D ≃ (I ⋊ G) ⋊ Aut0(C), thus recovering the description of the automorphism group of the structural matrix algebra given in [5]. For proving (9), define Ψ : I ⋊ G → U (M (ρ, k)) ⋊ T D0 , Ψ(CA ⋊ (aij)iρj) = A ⋊ (aij)iρj, where y denotes the class of y modulo D0. We note that Ψ is well defined since CA = CB if and only if A−1B lies in the center of M (ρ, k), and this is the set of diagonal matrices constant on all ith diagonal positions with i in the same connected component of C. Then B = Adiag(d1, . . . , dn) for such a central diagonal matrix diag(d1, . . . , dn). Since d−1 i dj = 1 for any iρj, we have that diag(d1, . . . , dn) ⋊ (1)iρj ∈ D0. But B ⋊ (aij)iρj = (diag(d1, . . . , dn) ⋊ (1)iρj )(A ⋊ (aij)iρj), so B ⋊ (aij)iρj = A ⋊ (aij)iρj. We show that Ψ is injective. Indeed, if Ψ(CA ⋊ (aij)iρj) is trivial, then A = diag(d1, . . . , dn) and aij = d−1 i dj for any iρj, where d1, . . . , dn are some non-zero scalars. Since aij = 1 for any i, j joined by an edge of some Tℓ, and Tℓ is a tree, we see that di must be the same when i runs through the vertices of a fixed ∆ℓ. Also, aij = 1 for i, j in the same isolated equivalence class α ∈ C, so di is constant for i in such a class. We conclude that A is a central element, so CA is the identity, and all aij's with iρj are equal to 1. 12 F. BES¸LEAG A AND S. D ASC ALESCU j To prove that Ψ is surjective, it is enough to show that for any A ⋊ (aij)iρj ∈ U (M (ρ, k)) ⋊ T there exists B ⋊ (bij)iρj ∈ I ⋊ G with A ⋊ (aij)iρj = B ⋊ (bij)iρj. We first show that there are d1, . . . , dn ∈ k∗ such that aij = did−1 for any i, j joined by an edge of some Tℓ, and also for any i, j in the same isolated equivalence class. Indeed, for the tree Tℓ we can fix some vertex i0, set di0 = 1, and then define di for any other vertex in Tℓ by induction on the distance from i0 to i in the tree Tℓ, using at each step the desired condition aij = did−1 . For an isolated equivalence class, say {i1, . . . , ip}, set dip = 1, and ai1,ip = di1 , . . . , aip−1,ip = dip−1. Then air,is = air,ipaip,is = air,ipa−1 is,ip (diag(d1, . . . , dn) ⋊ (d−1 and (d−1 i dj)iρj)(A ⋊ (aij)iρj) = diag(d1, . . . , dn)((d−1 i dj)iρj · A) ⋊ (d−1 i djaij)iρj ∈ G, so we are done. = dir d−1 is i djaij)iρj . Now j 5. Graded flags and associated gradings of matrices Let G be a group. A G-graded vector space is a vector space V with a decomposition Vg are called the homogeneous Vg, where each Vg is a subspace. The elements of [ V = M elements of V . Each v ∈ V is uniquely written as v = X vg, vg ∈ Vg. g∈G g∈G g∈G Definition 5.1. A G-graded ρ-flag is a ρ-flag (V, (Vα)α∈C) such that V is a G-graded vector space, and the basis B from Definition 2.1 consists of homogeneous elements. If F = (V, (Vα)α∈C) and F ′ = (V ′, (V ′ α)α∈C) are G-graded ρ-flags, then a morphism of graded flags from F to F ′ is a morphism f : V → V ′ of ρ-flags, which is also a morphism of graded vector spaces. Note that in a graded flag any Vα is a graded vector subspace, as it has a basis of homogeneous elements. If F = (V, (Vα)α∈C) is a G-graded ρ-flag and σ ∈ G, define End(F)σ = {f ∈ End(F) f (Vg) ⊆ Vσg for any g ∈ G}. Proposition 5.2. End(F) = M σ∈G algebra. End(F)σ, and this decomposition makes End(F) a G-graded Proof. It is clear that End(F)σ can be non-zero only for σ ∈ (supp V ) · (supp V )−1, where supp V = {g ∈ G Vg 6= 0}. Thus only finitely many End(F)σ are nonzero. End(F)σ is a direct sum, choose some f σ ∈ End(F)σ for each σ ∈ G. In order to see that X If X f σ = 0, then 0 = (X σ∈G σ∈G f σ)(Vg) = X σ∈G σ∈G f σ(Vg) for any g ∈ G. Since f σ(Vg) ⊂ Vσg, this shows that f σ(Vg) = 0 for any σ and any g, and we get that f σ = 0 for any σ. Now let f ∈ End(F). For any σ ∈ G define the linear maps fσ : V −→ V such that (fσ)(vg) = f (v)σg for any g ∈ G, vg ∈ Vg. Then fσ(Vα) = X fσ((Vα)g) ⊂ X g∈G g∈G f (Vα)σg ⊂ Vα for any α ∈ C, so fσ ∈ End(F)σ. Moreover, M σ∈G End(F)σ. it is clear that f = X σ∈G fσ, so End(F) = Obviously, End(F)σEnd(F)τ ⊂ End(F)στ for any σ, τ ∈ G, so End(F) is a G-graded algebra. (cid:3) 13 We will denote by END(F) the algebra End(F), regarded with the G-grading defined in Proposition 5.2. This grading transfers via the isomorphism defined in the proof of Proposition 2.2 to a G-grading on the structural matrix algebra M (ρ, k). If g1, . . . , gn are the degrees of the basis elements v1, . . . , vn, then each matrix unit eij with iρj is a homogeneous element of degree gig−1 in this grading. j Definition 5.3. A G-grading on the algebra M (ρ, k) is called a good grading if eij is a homo- geneous element for any i, j with iρj. Gradings on M (ρ, k) arising from graded flags are good gradings. We note that giving a good G-grading on M (ρ, k) is equivalent to giving a family (uij)iρj of elements of G such that uijujr = uir for any i, j, r with iρj and jρr. If we regard this family as a function u : ρ → G, defined by u(i, j) = uij for any i, j with iρj, then u is just a transitive function on ρ with values in G, in the terminology of [8]. j Examples of a transitive functions on ρ can be obtained as follows. Let g1, . . . , gn ∈ G, and let uij = gig−1 for any i, j with iρj. Then (uij)iρj is a transitive function on ρ. A transitive function on ρ is called trivial if it is obtained in this way. Clearly, the good G-gradings corresponding to trivial transitive functions on ρ are precisely the gradings obtained from graded flags as above. It is an interesting question whether all good gradings arise from graded flags, or equivalently Question. Let ρ be a preorder relation. Is it true that for any group G all transitive functions u : ρ → G are trivial? Several variations of the above questions can be formulated, for example to determine for a fixed ρ all groups G such that any transitive function on ρ with values in G is trivial. The following shows that the problem posed in the question above reduces to answering it for the associated poset (C, ≤). We also give an equivalent formulation involving the associated graph Γ. The equivalence between (1) and (2) in the next Proposition was proved in [8]. Proposition 5.4. Let G be a group. The following are equivalent: (1) Any transitive function u : ρ → G is trivial. (2) Any transitive function w : ≤ → G is trivial, where ≤ is the partial order on C. (3) For any function v : Γ1 → G such that v(a1) . . . v(ar) = v(b1) . . . v(bs) for any paths a1 . . . ar and b1 . . . bs in Γ with s(a1) = s(b1) and t(ar) = t(bs), there exists a function f : Γ0 → G such that v(a) = f (s(a))f (t(a))−1 for any a ∈ Γ1. Proof. Denote C = {α1, ..., αh} and pick i1 ∈ α1, ..., ih ∈ αh. (1)⇒(2) Let w : ≤ → G be a transitive function. Then u : ρ → G defined by u(i, j) = w(i, j) for any i, j with iρj, is a transitive function, so there exists a function f : {1, . . . , n} → G such that u(i, j) = f (i)f (j)−1 for any i, j with iρj. Now define f : C → G, f (αq) = f (iq) for any 1 ≤ q ≤ h. Then if αq ≤ αp, we have w(αq, αp) = u(iq, ip) = f (iq)f (ip)−1 = f (αq)f (αp)−1, so w is trivial. (2)⇒(3) Let v : Γ1 → G be as in (3). Define w : ≤ → G as follows. If α < β, let a1 . . . ar be a path in Γ starting at α and ending at β; we define w(α, β) = v(a1) . . . v(ar), and we note that the definition does not depend on the path, taking into account the property satisfied by v. We also define w(α, α) = e, the neutral element of G, for any α ∈ C. Then w is a transitive 14 F. BES¸LEAG A AND S. D ASC ALESCU function, so there exists f : C → G such that w(α, β) = f (α)f (β)−1 for any α ≤ β. In particular v(a) = f (s(a))f (t(a))−1 for any a ∈ Γ1. (3)⇒(1) Let u : ρ → G be a transitive function. Define v : Γ1 −→ G as follows: if a ∈ Γ1 with s(a) = αp, t(a) = αq, then v(a) = u(ip, iq). If a1 . . . ar and b1 . . . bs are paths in Γ with s(a1) = s(b1) = αp and t(ar) = t(bs) = αq, it is clear that v(a1) . . . v(ar) = v(b1) . . . v(bs) = u(ip, iq). Then there exist z1, ..., zn ∈ G such that v(a) = zpz−1 for any arrow a with s(a) = αp, t(a) = αq, 1 ≤ p, q ≤ h. q Now define the family (gi)1≤i≤n of elements of G by gi = u(i, ip)zp for any i, where p is such that i ∈ αp (note that u(i, ip) makes sense since i, ip ∈ αp). Then if i and j are such that iρj, let i ∈ αp, j ∈ αq. We know that gi = u(i, ip)zp, gj = u(j, iq)zq and u(ip, iq) = zpz−1 . Then q u(i, j) = u(i, ip)u(ip, iq)u(iq, j) q u(j, iq)−1 p zpz−1 q zqg−1 j = giz−1 = giz−1 = gig−1 j . (cid:3) The next result gives an easy way to check whether a function v as in Proposition 5.4 (3) arises from a function f : Γ0 → G, by looking at the cycles of the undirected graph associated with Γ. We consider the graph Γ, constructed from Γ by 'doubling the arrows'. It has the same vertices as Γ thus Γ0 = Γ0. For any arrow a ∈ Γ1 from α to β, we consider an arrow a from β to α, and define Γ1 = Γ1 ∪ {a a ∈ Γ1}. If G is a group and v : Γ1 → G is a function, we denote by v : Γ1 → G the function whose restriction to Γ1 is v, and such that v(a) = v(a)−1 for any a ∈ Γ1. Proposition 5.5. Let G be a group, and let v : Γ1 −→ G such that v(a1) . . . v(ar) = v(b1) . . . v(bs) for any paths a1 . . . ar and b1 . . . bs in Γ with s(a1) = s(b1) and t(ar) = t(bs). Then the following are equivalent. (1) There exists a function f : Γ0 → G such that v(a) = f (s(a))f (t(a))−1 for any a ∈ Γ1. (2) For any cycle z1...zm in Γ, with z1, ..., zm ∈ Γ1 (this corresponds to a cycle in the undirected graph Γu associated with Γ) one has v(z1) . . . v(zm) = 1. Proof. (1)⇒(2) For any a ∈ Γ1 starting from α and ending at β, we have v(a) = v(a) = f (α)f (β)−1 and v(a) = f (β)f (α)−1. Thus v(z) = f (s(z))f (t(z))−1 for any z ∈ Γ1. Now it is clear that v(z1) ... v(zm) = 1 for any cycle z1 ... zm in Γ. (2)⇒(1) We construct a function f : Γ0 −→ G satisfying the desired property. It is clear that we can reduce to the case where Γ is connected (and putting together the functions constructed for the connected components). Choose some ω ∈ Γ0 and take f (ω) be an arbitrary element of G. If α ∈ Γ0, let z1 . . . zd be a path from α to ω in Γ. Define f (α) = v(z1) . . . v(zd)f (ω); this does not depend on the path z1 . . . zd, since for another path y1 . . . yh from α to ω, z1 . . . zdy−1 is a cycle in Γ, where y−1 denotes the arrow opposite to y (thus a−1 = a, and a−1 = a for any a ∈ Γ1). It is clear that v(y−1) = v(y)−1 for any y ∈ Γ1. Then v(z1) . . . v(zd)v(y−1 1 ) = 1, so v(z1) . . . v(zd) = v(y1) . . . v(yh). Now if a ∈ Γ1 is an arrow from α to β, let z1 ... zd be a path from β to ω in Γ. Then f (β) = v(z1) ... v(zd)f (ω), and f (α) = v(a)v(z1) ... v(zd)f (ω) = v(a)f (β) = v(a)f (β), so v(a) = f (α)f (β)−1. (cid:3) h ) . . . v(y−1 h . . . y−1 1 Let F (Γ) be the free group generated by the set Γ1 of arrows of Γ. Let A(Γ) be the subgroup 1 , where a1 . . . ar and b1 . . . bp are of F (Γ) generated by all elements of the form a1 . . . arb−1 p . . . b−1 15 two paths (in Γ) starting from the same vertex and terminating at the same vertex. We also consider the subgroup B(Γ) of F (Γ) generated by all elements of the form a1aε2 m , where a1, . . . , am are arrows forming in this order a cycle in the undirected graph obtained from Γ when omitting the direction of arrows, and εi = 1 if ai is in the direction of the directed cycle given by a1, and εi = −1 otherwise. Clearly A(Γ) ⊆ B(Γ), since any generator of A(Γ) lies in B(Γ). Now we can give an answer to the question posed above in terms of these groups. We recall that for a group X and a subgroup Y of X, the normal closure Y N of Y is the smallest normal subgroup of X containing Y . The elements of Y N are all products of conjugates of elements of Y . 2 . . . aεm Proposition 5.6. With notation as above, the following are equivalent. (1) For any group G, any transitive function u : ρ → G is trivial. (2) A(Γ)N = B(Γ)N . (3) Any generator b of B(Γ) can be written in the form b = g1x1g−1 m for some 1 positive integer m, some g1, . . . , gm ∈ F (Γ) and some x1, . . . , xm among the generators in the construction of A(Γ). . . . gmxmg−1 Proof. By Propositions 5.4 and 5.5, (1) is equivalent to the fact that for any group G and any group morphism f : F (Γ) → G such that f (A(Γ)) = 1, we also have f (B(Γ)) = 1. Indeed, giving a function v : Γ1 → G is the same with giving a group morphism f : F (Γ) → G; moreover, v satisfies v(a1) . . . v(ar) = v(b1) . . . v(bs) for any paths a1 . . . ar and b1 . . . bs in Γ with s(a1) = s(b1) and t(ar) = t(bs), if and only if f (A(Γ)) = 1. On the other hand, such a v is trivial if and only if f (B(Γ)) = 1. Now if (1) holds, then the projection F (Γ) → F (Γ)/A(Γ)N is trivial on A(Γ), so it must be trivial on B(Γ). This shows that B(Γ) ⊆ A(Γ)N , or A(Γ)N = B(Γ)N . Conversely, if (2) holds, then any group morphism f : F (Γ) → G which is trivial on A(Γ) is also trivial on A(Γ)N , and then also on B(Γ), so (1) holds. The equivalence between (2) and (3) follows from the description of the normal closure of a (cid:3) subgroup, and the fact that the given set of generators of A(Γ) is closed under inverse. Example 5.7. Assume that ρ is a preorder relation such that the associated graph Γ is of the form am−1 <①①①①①①①① m−1• m• bp+1 b❋❋❋❋❋❋❋❋ •m+1 2• b❊❊❊❊❊❊❊❊❊ a1 1• •m+p <②②②②②②②②② b1 for some integers m ≥ 3 and p ≥ 1. Thus there are two paths from 1 to m, these are a1 . . . am−1 and b1 . . . bp+1. Then for any group G, any transitive function u : ρ → G is trivial. Indeed, if v is a function as in Proposition 5.5, then v(a1) . . . v(am−1) = v(b1) . . . v(bp+1), and then the condition in (2) of the mentioned Proposition is obviously satisfied, as it is clear that the associated undirected graph has just one cycle. < b b < 16 F. BES¸LEAG A AND S. D ASC ALESCU Alternatively, A(Γ) is the cyclic group generated by x = a1 . . . am−1b−1 p+1 . . . b−1 generators of B(Γ) are all conjugates of x or x−1. For example, a2 . . . am−1b−1 a−1 1 xa1. Thus A(Γ)N = B(Γ)N . 1 , while the 1 a1 = p+1 . . . b−1 Example 5.8. Assume that ρ is a preorder relation such that the associated graph Γ is of the form • ⑦⑦⑦⑦⑦⑦⑦ _❅❅❅❅❅❅❅ • • • • ❅❅❅❅❅❅❅ ?⑦⑦⑦⑦⑦⑦⑦ • Thus the un-directed graph Γu associated to Γ is cyclic, and in Γ there are at least two vertices where both adjacent arrows terminate (equivalently, Γu is cyclic and Γ is not of the type in Example 5.7). Then for any non-trivial group G, there exist transitive functions u : ρ → G that are not trivial. Indeed, it is enough to show that the condition (3) in Proposition 5.4 does not hold. Since Γ does not have two different paths starting from the same vertex and ending at the same vertex, we only have to see that not any function v : Γ1 → G is given by v(a) = f (s(a))f (t(a))−1, a ∈ Γ1, for some function f : Γ0 → G. This is clear, for instance we can take v to be the identity of G on all but one arrows. We note that A(Γ) is trivial, since there are no two different paths starting from the same vertex and terminating at the same vertex. On the other hand, B(Γ) is not trivial, since Γu has a cycle. The simplest example of such a graph is • ⑦⑦⑦⑦⑦⑦⑦ _❅❅❅❅❅❅❅ • • ❅❅❅❅❅❅❅ ?⑦⑦⑦⑦⑦⑦⑦ • and the corresponding structural matrix algebra, whose not all good gradings arise from graded flags, is   k 0 k k 0 k k k 0 0 k 0 0 0 0 k   . Example 5.9. If we construct a graph ∆ by taking a graph Γ as in Example 5.8, adding a vertex v, and adding an arrow from each vertex in Γ where both adjacent arrows terminate to v, as in the picture below _ ? _ ? 17 v• • • ~⑥⑥⑥⑥⑥⑥⑥⑥ `❇❇❇❇❇❇❇❇ • • • ❅❅❅❅❅❅❅ ?⑦⑦⑦⑦⑦⑦⑦ • then all transitive functions (on the corresponding preordered set) are trivial. For simplicity, we explain this in the case where ∆ is x • • a ⑦⑦⑦⑦⑦⑦⑦ _❅❅❅❅❅❅❅ c • • y b ❅❅❅❅❅❅❅ ?⑦⑦⑦⑦⑦⑦⑦ d • but the argument is the same in general. In this case, A(∆) is generated by g = axy−1b−1 and h = cxy−1d−1, while B(∆) is generated by certain conjugates of g and h (as explained in Example 5.8), ac−1db−1, and certain conjugates of the latter. But ac−1db−1 = g(db−1)−1h−1(db−1), showing that A(∆)N = B(∆)N . 6. Isomorphisms between graded endomorphism algebras If V = ⊕g∈GVg is a G-graded vector space, then for any σ ∈ G the right σ-suspension of V is the G-graded vector space V (σ) which is just V as a vector space, with the grading shifted by σ as follows: V (σ)g = Vgσ for any g ∈ G. Also, the left σ-suspension of V , denoted by (σ)V , is the vector space V with the grading given by ((σ)V )g = Vσg. For any σ, τ ∈ G one has (V (σ))(τ ) = V (τ σ), (σ)((τ )V ) = (τ σ)V and ((σ)V )(τ ) = (σ)(V (τ )). The fact that there are two types of suspensions for a graded vector space V can be explained by the fact that V is an object in the category of left graded modules, and also an object in the category of right graded modules over the algebra k, regarded with the trivial G-grading. Then V (σ) and (σ)V are just the suspensions when V is regarded in these categories. If V and W are G-graded vector spaces and σ ∈ G, we say that a linear map f : V → W is a morphism of left degree σ if f (Vg) ⊆ Wσg for any g ∈ G; this means that f is a morphism of graded vector spaces when regarded as f : V → (σ)W . Similarly, f is a morphism of right degree σ if f (Vg) ⊆ Wgσ for any g ∈ G. If F = (V, (Vα)α∈C) is a G-graded ρ-flag and σ ∈ G, then the right suspension of F is F(σ) = (V (σ), (Vα)α∈C), and the left suspension of F is (σ)F = ((σ)V, (Vα)α∈C). It is clear that END(F)σ is just the space of morphisms of graded flags from F to (σ)F. Let F = (V, (Vα)α∈C) be a G-graded ρ-flag, with a homogeneous basis B = [ Bα of V α∈C providing the flag structure. ~ A A h h ` ? K K n n _ ? 18 F. BES¸LEAG A AND S. D ASC ALESCU Let C = C1 ∪ . . . ∪ Cq be the decomposition of C in disjoint connected components; these correspond to the connected components of the undirected graph Γu. For each 1 ≤ t ≤ q, let ρt α, by restricting ρ. be the preorder relation on the set [ If V t = X V = M α∈Ct 1≤t≤q Vα, then F t = (V t, (Vα)α∈Ct ) is a G-graded ρt-flag with basis [ Bα. Obviously, α∈Ct V t. In a formal way we can write F = F 1 ⊕ . . . ⊕ F q, where F is a G-graded ρ-flag, α∈Ct and F t is a G-graded ρt-flag for each 1 ≤ t ≤ q. Definition 6.1. Let ρ and µ be isomorphic preorder relations (i.e. the preordered sets on which ρ and µ are defined are isomorphic). Let C and D be the posets associated with ρ and µ, and let g : C → D be an isomorphism of posets. We say that a ρ-flag F = (V, (Vα)α∈C)) is g- isomorphic to a µ-flag G = (W, (Wβ)β∈D)) if there is a linear isomorphism u : V → W such that u(Vα) = Wg(α) for any α ∈ C. If F and G are G-graded flags, we say that they are g-isomorphic as graded flags if there is such an u which is a morphism of graded vector spaces. Lemma 6.2. With notation as in Definition 6.1, if the G-graded flags F and G are g-isomorphic, then END(F) and END(G) are isomorphic as G-graded algebras. Proof. If u : V → W is a g-isomorphism between the graded flags F and G, then Φ : END(F) → END(G), Φ(φ) = uφu−1 is an isomorphism of G-graded algebras. (cid:3) Now we will consider another G-graded ρ-flag F ′ = (V ′, (V ′ α)α∈C). As we did for F, we also V ′t and F ′ = F ′1 ⊕ . . . ⊕ F ′q, where F ′t is a G-graded ρt-flag for each 1 ≤ t ≤ q. have V ′ = M 1≤t≤q α)α∈C) be G-graded ρ-flags. Then the Theorem 6.3. Let F = (V, (Vα)α∈C) and F ′ = (V ′, (V ′ following assertions are equivalent: (1) END(F) and END(F ′) are isomorphic as G-graded algebras. (2) There exist g ∈ Aut0(C), σ1, . . . , σq ∈ G and a g-isomorphism γ : V → V ′ between the (ungraded) ρ-flags F and F ′, such that γV ′g(t) : V t → V ′g(t) is a linear isomorphism of left degree σt for any 1 ≤ t ≤ q, where g ∈ Sq is the permutation induced by g, i.e. g(Ct) = Cg(t). (3) There exists a permutation τ ∈ Sq, an isomorphism gt : Ct → Cτ (t) for each 1 ≤ t ≤ q, and σ1, . . . , σq ∈ G, such that F t(σt) is gt-isomorphic to F ′τ (t) for any 1 ≤ t ≤ q. V t Proof. (1)⇒(2) Let φ : END(F) → END(F ′) be an isomorphism of G-graded algebras. We follow the line of proof of Proposition 4.2 and its notation, adding the additional information related to the graded structure. Let gi = deg vi for any i. Then Eij is a homogeneous element of degree gig−1 Now Qi = Im Fii is a graded subspace of V ′, V ′ = M of END(F), and Fij = φ(Eij) also has degree gig−1 Qi and each Qi is 1-dimensional. For in END(F ′). j j 1≤i≤n each i pick wi ∈ Qi \ {0}, which is a homogeneous element. Then for any i, j with iρj we have Fij(wj) = aijwi for some aij ∈ k∗, and aijajr = air for any i, j, r with iρj and jρr. Since (Fij)Qi Qj : Qj → Qi is a linear isomorphism of degree gig−1 : Qi → Qj), we obtain that Qi ≃ (gjg−1 )(gj )Qj for any i, j with iρj, so then (gi)Qi ≃ (gj)Qj as graded vector spaces. This implies that (gi)Qi has the same isomorphism type for all i ∈ α with α lying in a connected component of C. (with inverse (Fji)Qj Qi )Qj ≃ (g−1 j i i 19 On the other hand, if Ri = kvi for any i, then (gi)Ri and (gj)Rj are isomorphic graded vector spaces for any i, j. Then there are σ1, . . . , σq ∈ G such that for any 1 ≤ t ≤ q we have Qi ≃ Ri(σt) for any i ∈ α with α ∈ Ct. Indeed, fix some t and pick α0 ∈ Ct and i0 ∈ α0. Since Qi0 and Ri0 are 1-dimensional graded vector spaces, there exists σt ∈ G such that Qi0(σt) ≃ Ri0. Then for any α ∈ Ct and any i ∈ α one has Qi(σt) ≃ ((gi0 g−1 i )Qi0)(σt) = (gi0g−1 i )(Qi0(σt)) ≃ (gi0 g−1 i )Ri0 ≃ Ri Thus we obtain that (10) deg vi = (deg wi)σt for any i ∈ α with α ∈ Ct As in the proof of Proposition 4.2, the linear map γ : V → V ′ with γ(vi) = wi for any i, is a φ′-isomorphism for a certain algebra isomorphism φ′ : END(F) → END(F ′), and then there exists g ∈ Aut0(C) with γ(Vα) = V ′ g(α) for any α ∈ C. By (10), γV t is a linear morphism of left degree σt, and we are done. (2)⇒(3) Take τ = g, and let gt : Ct → Cτ (t) be the restriction and corestriction of g for each t. Then the restriction and corestriction of γ to V t and V ′τ (t) gives a gt-isomorphism of graded flags F t(σt) ≃ F ′τ (t). (3)⇒(1) It is clear that END(F) ≃ END(F 1) × . . . × END(F q) and END(F ′) ≃ END(F ′1) × . . . × END(F ′q) as G-graded algebras. By Lemma 6.2, we see that END(F t(σt)) ≃ END(F ′τ (t)) for any 1 ≤ t ≤ q. As it is obvious that END(F t(σt)) = END(F t), we get END(F t) ≃ END(F ′τ (t)) for any 1 ≤ t ≤ q. We conclude that END(F) ≃ END(F ′). (cid:3) 7. Classification of gradings arising from graded flags The aim of this section is to classify G-gradings on M (ρ, k) arising from graded flags by the orbits of a certain group action. We keep all the notations of Section 6. We first consider three group actions on the set Gn. • Aut0(C) acts to the right on Gn by (hi)1≤i≤n ← g = (hg(i))1≤i≤n for any (hi)1≤i≤n ∈ Gn and g ∈ Aut0(C). • Gq acts to the right on Gn by where for each i we define h′ (hi)1≤i≤n ← (σt)1≤t≤q = (h′ i)1≤i≤n i = hiσp, where p is such that i ∈ Cp. • For each α ∈ C let S(α) be the symmetric group of α (regarded as a subset of {1, . . . , n}. We consider the group Qα∈C S(α), which is a Young subgroup of Sn (isomorphic to Qα∈C Smα). Then Qα∈C S(α) acts to the right on Gn by (hi)1≤i≤n ← (ψα)α∈C = (h′ i)1≤i≤n with h′ i defined by h′ i = hψα(i), where α = i, for each i. 20 F. BES¸LEAG A AND S. D ASC ALESCU Now there is a right action of the group Aut0(C) on the group Gq defined by (σt)1≤t≤q ← g = (στ (t))1≤t≤q where τ ∈ Sq is the permutation induced by g. Then we have a right semidirect product Aut0(C) ⋉ Gq. Moreover, the compatibility condition (4) holds for these actions, i.e. (11) ((hi)1≤i≤n ← (σt)1≤t≤q) ← g = ((hi)1≤i≤n ← g) ← ((σt)1≤t≤q ← g) Indeed, it is easy to see that both sides of equation (11) have on the ith position hg(i)στ (p), where p is such that i ∈ Cp. We conclude that Aut0(C) ⋉ Gq acts to the right on the set Gn by (hi)1≤i≤n ← (g ⋉ (σt)1≤t≤q) = ((hi)1≤i≤n ← g) ← (σt)1≤t≤q On the other hand, it is straightforward to check that Aut0(C) acts to the left on the group Qα∈C S(α) by g → (ψα)α∈C = (ψ′ α)α∈C, where for any α ∈ C, ψ′ α is defined by ψ′ α(i) = g(ψg−1(α)(g−1(i))) for any i ∈ α. Moreover, one can check that (12) ((hi)1≤i≤n ← g) ← (ψα)α∈C = ((hi)1≤i≤n ← (g → (ψα)α∈C)) ← g The left action of Aut0(C) on Qα∈C S(α) induces a left action of Aut0(C) ⋉ Gq on Qα∈C S(α), with 1 ⋉ Gq acting trivially on Qα∈C S(α), and then we can consider the left semidirect product Qα∈C S(α) ⋊ (Aut0(C) ⋉ Gq). The right actions of Qα∈C S(α) and Gq on Gn commute, i.e. (13) ((hi)1≤i≤n ← (ψα)α∈C) ← (σt)1≤t≤q = ((hi)1≤i≤n ← (σt)1≤t≤q) ← (ψα)α∈C Indeed, it is easily checked that on the ith position of each side one finds the element hψα(i)σp, where α = i ∈ Cp. Now (12) and (13) show that the compatibility relation (5) is satisfied for the right actions of Aut0(C) ⋉ Gq and Qα∈C S(α) on Gn. In conclusion, the group Qα∈C S(α) ⋊ (Aut0(C) ⋉ Gq) acts to the right on the set Gn by (hi)1≤i≤n ← ((ψα)α∈C ⋊ (g ⋉ (σt)1≤t≤q)) = (((hi)1≤i≤n ← (ψα)α∈C) ← g) ← (σt)1≤t≤q. Now we can prove the following. Theorem 7.1. The isomorphism types of G-gradings of the type END(F), where F is a G-graded ρ-flag, are classified by the orbits of the right action of the group Qα∈C S(α) ⋊ (Aut0(C) ⋉ Gq) on the set Gn. Proof. We first need some simple remarks. If V and V ′ are G-graded vector spaces with homo- n}, then V and V ′ are isomorphic as G-graded geneous bases {v1, . . . , vn}, respectively {v′ vector spaces if and only if dim Vg = dim V ′ g for any g ∈ G, and this is also equivalent to the 1, . . . , deg v′ fact that the n-tuple (deg v1, . . . , deg vn) of elements of G is obtained from (deg v′ n) by a permutation. 1, . . . , v′ Also, if σ ∈ G, then V (σ) ≃ V ′ if and only if ((deg v1)σ−1, . . . , (deg vn)σ−1) is obtained from n) by a permutation; this follows from the fact that a homogeneous element of 1, . . . , deg v′ (deg v′ degree g of V has degree gσ−1 in V (σ). Let F be a ρ-flag with basis B = {v1, . . . , vn} as in Definition 2.1. A G-graded structure on F is given by assigning arbitrary degrees h1, . . . , hn ∈ G to v1, . . . , vn. Thus G-graded structures on F are given by elements (h1, . . . , hn) of Gn. 21 Let F and F ′ be G-graded ρ-flags given by n-tuples (h1, . . . , hn) and (h′ The homogeneous basis of V and V ′ are denoted by B = [ Bα and B′ = [ 1, . . . , h′ B′ α. n) as above. α∈C α∈C We show that END(F) ≃ END(F ′) as G-graded algebras if and only if (h1, . . . , hn) and 1, . . . , h′ n) are in the same orbit of Gn with respect to the right action ofQα∈C S(α)⋊(Aut0(C)⋉ (h′ Gq), and this will finish the proof. By Theorem 6.3, END(F) ≃ END(F ′) if and only if there exists g ∈ Aut0(C) such that F t(σt) is gt-isomorphic as a graded flag to F ′g(t) for any 1 ≤ t ≤ q, where g ∈ Sq is the permutation such that for any t, g(Ct) = Cg(t), and gt : Ct → Cg(t) is the isomorphism of posets induced by g via restriction and corestriction. Using the considerations above, we obtain that END(F) ≃ END(F ′) if and only if there exists g ∈ Aut0(C) such that for each 1 ≤ t ≤ q and any α ∈ Ct, the degrees of the elements of Bα multiplied to the right by σ−1 are obtained by a permutation from the elements of B′ g(α). But this is equivalent to t ((hi)1≤i≤n ← (σ−1 )1≤t≤q) ← (ψα)α∈C = (h′ t i)1≤i≤n ← g for some (ψα)α∈C ∈ Qα∈C S(α). Since the right actions of Gq and Qα∈C S(α) on Gn commute, this is the same with which is the same to (hi)1≤i≤n and (h′ (hi)1≤i≤n = (((h′ Qα∈C S(α) ⋊ (Aut0(C) ⋉ Gq) on Gn. i)1≤i≤n ← g) ← (σt)1≤t≤q) ← (ψ−1 α )α∈C i)1≤i≤n lying in the same orbit of the right action of (cid:3) Example 7.2. As particular cases of our results we obtain the following. (1) Let A = Mn(k), the full matrix algebra. If G is a group, then the good G-gradings on A are all isomorphic to gradings of the form END(V ), where V is a G-graded vector space of dimension n. These gradings are classified by the orbits of the biaction of the groups Sn (by usual permutations of the elements) and G (by right translations) on Gn. Indeed, in this case C is a singleton, so obviously Aut0(C) is trivial. This result appears in [4]. (2) Let A be the algebra of upper block triangular matrices of type m1, . . . , mr, where n = m1 + . . . + mr. Then C is isomorphic to the poset {1, . . . , r} with the usual order, so any good grading on A is of the type END(F), where F is a graded (usual) flag of signature (m1, . . . , mr). Again, Aut0(C) is trivial and C is connected, so the isomorphism types of good G-gradings on A are classified by the orbits of the biaction of a Young subgroup Sm1 × . . . × Smr (by permutations) and G (by translations) on Gn. This result appears in [3]. References [1] Yu. A. Bahturin, S. K. Sehgal, M. V. Zaicev, Group gradings on associative algebras, J. Algebra 241 (2001), 677 -- 698. [2] Yu. A. Bahturin, M. V. Zaicev, Group gradings on matrix algebras, Canad. Math. Bull. 45 (2002), 499 -- 508. [3] M. Barascu, S. Dascalescu, Good gradings on upper block triangular matrix algebras, Comm. Algebra 41 (2013), 4290-4298. [4] S. Caenepeel, S. Dascalescu, C. Nastasescu, On Gradings of Matrix Algebras and Descent Theory, Comm. Algebra 30 (2002), 5901 -- 5920. [5] S. P. Coelho, The automorphism group of a structural matrix algebra, Linear Alg. Appl. 195 (1993), 35-58. [6] V. Lakshmibai and J. Brown, Flag varieties. An interplay of geometry, combinatorics and representation theory, Hindustan Book Agency, 2009. [7] C. Nastasescu and F. van Oystaeyen, Methods of graded rings, Lecture Notes in Math., vol. 1836 (2004), Springer Verlag. [8] A. Nowicki, Derivations of special subrings of matrix rings and regular graphs, Tsukuba J. Math. 7 (1983), 281-297. 22 F. BES¸LEAG A AND S. D ASC ALESCU [9] E. Spiegel and C. J. O'Donnell, Incidence algebras, Pure and Appl. Math. 206 (1997), Marcel Dekker, New York. [10] L. Van Wyk, Maximal left ideals in structural matrix rings, Comm. Algebra 16 (1988), 399-419. University of Bucharest, Facultatea de Matematica, Str. Academiei 14, Bucharest 1, RO- 010014, Romania e-mail: filoteia [email protected], [email protected]
1510.05621	1	1510	2015-10-19T19:00:32	A classification of torsors over Laurent polynomial rings	[ "math.RA" ]	Let R\_n be the ring of Laurent polynomials in n variables over a field k of characteristic zero and let K\_n be its fraction field.Given a linear algebraic k-group $G$, we show that a K\_n-torsor under G which is unramified with respect to X=Spec(R\_n)extends to a unique toral R\_n-torsor under G. This result, in turn, allows us to classify all G-torsors over R\_n.	math.RA	math	A CLASSIFICATION OF TORSORS OVER LAURENT POLYNOMIAL RINGS V. CHERNOUSOV, P. GILLE, AND A. PIANZOLA Abstract. Let Rn be the ring of Laurent polynomials in n variables over a field k of characteristic zero and let Kn be its fraction field. Given a linear algebraic k -- group G, we show that a Kn-torsor under G which is unramified with respect to X = Spec(Rn) extends to a unique toral Rn -- torsor under G. This result, in turn, allows us to classify all G-torsors over Rn. Keywords: Reductive group scheme, torsor, multiloop algebra. MSC 2000: 14F20, 20G15, 17B67, 11E72. Contents Introduction 1. 2. Unramified functors 3. Non-abelian cohomology 3.1. Some terminology 3.2. Torsion bijection 3.3. Acyclicity Theorem 3.5. Grothendieck -- Serre's conjecture 3.10. Rational torsors everywhere locally defined 4. Proof of Theorem 1.2 5. Applications 5.1. A disjoint union decomposition for Rn -- torsors 5.6. The case of orthogonal groups 5.9. The projective linear case 5.14. Applications to Rn -- Lie algebras References 2 4 6 6 6 7 7 8 9 11 11 13 14 16 16 Date: July 28, 2021. V. Chernousov was partially supported by the Canada Research Chairs Program and an NSERC research grant. P. Gille was supported by the Romanian IDEI project PCE−2012-4-364 of the Ministry of National Education CNCS-UEFISCIDI. A. Pianzola wishes to thank NSERC and CONICET for their continuous support. 1 2 V. CHERNOUSOV, P. GILLE, AND A. PIANZOLA 1. Introduction Torsors are the algebraic analogues of the principal homogeneous spaces that one encounters in the theory of Lie groups. While the latter are (under natural assumptions) locally trivial (in the usual topology), this is not the case for the former: a torsor E → X under a group scheme G over X is not necessarily trivial (i.e., isomorphic to G with G acting on itself by right multiplication) when restricted to any non-empty Zariski open subset U of X. The reason for this is that the Zariski topology is too coarse. The path out of this serious obstacle was initiated by J.-P. Serre (with what is now called the finite ´etale topology) and then implemented in full generality (together with an accompanying descent theory) by A. Grothendieck. The idea is to have certain morphisms U → X (e.g. ´etale or flat and of finite presentation) replace open immersions as trivializing local data. Torsors have played an important role in number theory (Brauer groups, Tate-Shafarevich group, Manin obstructions) and in the Langlands program (Ngo's proof of the Fundamental Lemma). Somehow surprisingly torsors have been used over the last decade to solve difficult problems in infinite dimensional Lie theory (see [GP3] for an extensive list of references. See also [CGP2] and [KLP]). The way that torsors arise in this context is the following. The infinite dimensional object L under consideration (for example, the centreless core of an extended affine Lie algebra [EALA] or a Lie superconformal algebra) has an invariant called the centroid (essentially the linear endomorphisms of the object that commute with their multiplication). These centroids are Laurent polynomial rings k[t±1 n ] in finitely many variables over a base field k of characteristic 0. We will denote this ring by Rn, or simply by R if no confusion is possible. 1 , · · · , t±1 The object L is naturally an R-module and it inherits the algebraic struc- ture of L (for example L is a Lie algebra over R). It is when L is viewed as an object over R that torsors enter into the picture. One can for ex- ample classify up to R-isomorphism the objects under consideration using non-abelian ´etale cohomology. Of course at the end of the day one wants to understand the problem under consideration (say a classification) over k and not R. Thankfully there is a beautiful theory, known as the "centroid trick", that allows this passage. The above discussion motivates why one is interested in the classification of torsors over R under a smooth reductive R-group scheme. We believe that the understanding of such torsors is of its own interest. This is the purpose of the present work. There is an important class of torsors under G called loop torsors that appear naturally in infinite dimensional Lie theory. (Loop torsors are defined over an arbitrary base in [GP3]. They have already caught the attention of researchers in other areas. See for example [PZ].) It is shown in [GP3] that if E is a torsor over R under G, then E being a loop torsor is equivalent to E being toral, i.e. that the twisted R-group scheme TORSORS OVER LAURENT POLYNOMIAL RINGS 3 EG admits a maximal torus. This is a remarkable property of the ring R: loop and toral torsors coincide. Toral torsors under reductive group schemes were completely classified in our paper [CGP2] with the use of Bruhat-Tits theory of buildings (see the Acyclicity Theorem 3.4). We will use toral torsors in what follows (but the reader is asked to keep in mind that these are precisely the torsors that arise in infinite dimensional Lie theory). Given a smooth reductive group scheme G over Rn we want to classify/describe all the isomorphism classes of Rn-torsors under G. Since G is smooth reductive they are in natural one-to-one correspondence with elements of the pointed set H 1 ´et(Rn, G). We have of course by definition a natural inclusion H 1 ´et,toral(Rn, G) ⊆ H 1 ´et(Rn, G) where H 1 toral G-torsors. From now on by default our topology will be ´etale; particular we will denote H 1 main results is the following. ´et,toral(Rn, G) is the subset consisting of (isomorphism) classes of in toral. One of our ´et by H 1 and H 1 ´et,toral by H 1 1.1. Theorem. Under the above notation there is a natural bijection H 1(Rn, G) ←→ H 1 Zar(Rn, E G). G [E]∈H 1 toral(Rn,G) The content of the above equality could be put in words as follows. Given a torsor E′ over Rn under G there exists a unique toral torsor E such that E′ is locally isomorphic (in the Zariski topology) to E. The proof of this result is achieved by a careful analysis (of independent interest) of the ramification of the torsors under consideration. We denote by Kn = k(t1, . . . , tn) the fraction field of Rn and set Fn = k((t1)) . . . ((tn)). The precise statement of our other main result is the following. 1.2. Theorem. Let G be a smooth affine Rn -- group scheme. Assume that either (i) G is reductive and admits a maximal Rn-torus (equivalently G is "loop reductive" [GP3, cor. 6.3]); or (ii) there exists a linear (smooth, not necessary connected) algebraic group G over k and a loop torsor E under G ×k Rn such that G = E(G ×k Rn). Then we have natural bijections H 1 toral(Rn, G) ∼−→ H 1(Kn, G)Rn−unr ∼−→ H 1(Fn, G) where H 1(Kn, G)Rn−unr stands for the subset of the Galois cohomology set H 1(Kn, G) consisting of (isomorphism) classes of G-torsors over Kn ex- tending everywhere in codimension one (see §3.10). We need to explain briefly the assumptions. In both cases (i) and (ii) we consider loop (=toral) group schemes because they play a central role in the 4 V. CHERNOUSOV, P. GILLE, AND A. PIANZOLA classification of Rn-torsors (see Theorem 1.1). Also, even though reductive group schemes are the main interest in this paper, in case (ii) we include group schemes over Rn which are not necessary "connected". Far from being a trivial generalization, this case is absolutely essential for applications to infinite dimensional Lie theory. Indeed one is forced to understand twisted forms of Rn-Lie algebras g ⊗k Rn, where g is a split finite dimensional simple Lie algebra over k. This leads to torsors under the group G ×k Rn where G is the linear algebraic k-group Aut(g). Many interesting infinite dimensional Lie objects over k, including extended affine Lie algebras (a particular case of them are the celebrated affine Kac -- Moody Lie algebras) and Lie super- conformal algebras, follow under the above considerations. Note that the special case G = PGLd (i.e. Rn -- Azumaya algebras) was already quite understood by Brauer group techniques when the base field is algebraically closed [GP2, §4.4]. Note also that Theorem 1.2 refines our ∼−→ H 1(Fn, G) ([CGP2, acyclicity theorem, i.e. the bijection H 1 th. 14.1] in case (i) (resp. [GP3, th. 8.1] in case (ii)). toral(Rn, G) The structure of the paper is as follows. In section 2, we establish useful generalities about unramified functors. Section 3 discusses unramified non- abelian cohomology. In section 4, we prove Theorem 1.2. Section 5 is devoted to applications including a disjoint union decomposition for the set H 1(Rn, G) (Theorem 1.1). Two important particular cases are considered in detail: the cases of orthogonal groups and projective linear groups. This illustrates that our main result, which may look rather abstract and remote in appearance, can lead to new very concrete classifications/descriptions of familiar objects. 2. Unramified functors We follow essentially the setting of [CT]. Let S be a scheme. If X is an integral S-scheme we denote by κ(X) the fraction field of X. Let F be an S-functor, that is a contravariant functor X 7→ F(X) from the category of S -- schemes into the category of sets. If X is integral normal one defines the following two subsets of F(κ(X)): F(κ(X))X−loc := \ x∈X Im(cid:16)F(OX,x) → F(κ(X))(cid:17) and F(κ(X))X−unr := \ x∈X (1) Im(cid:16)F(OX,x) → F(κ(X))(cid:17). The first subset F(κ(X))X−loc is called the subset of local classes with re- spect to X and the second one F(κ(X))X−unr is called the subset of unram- ified classes with respect to X. Obviously, we have the inclusions F(κ(X))X−loc ⊆ F(κ(X))X−unr ⊆ F(κ(X)). TORSORS OVER LAURENT POLYNOMIAL RINGS 5 2.1. Lemma. Let Y is an integral normal scheme over S. Let f : Y → X be a dominant morphism of S-schemes. Consider the map F(f ∗) : F(κ(X)) → F(κ(Y )) induced by the comorphism f ∗ = f ∗ (1) F(f ∗)(F(κ(X))X−loc) ⊆ F(κ(Y ))Y −loc. (2) If f is flat then F(f ∗)(F(κ(X))X−unr ) ⊆ F(κ(Y ))Y −unr. κ(X) : κ(X) → κ(Y ). Then. Proof. (1) The comorphism f ∗ allows us to view κ(X) as a subfield κ(X) ֒→ κ(Y ) of the field κ(Y ). Let γ ∈ F(κ(X))X−loc. We want to show that its image γκ(Y ) := F(f ∗)(γ) ∈ F(κ(Y )) under the base change is local with respect to Y . Let y ∈ Y and put x = f (y). The commutative square κ(X) ⊂ κ(Y ) ∪ ∪ OX,x ⊂ OY,y induces a commutative diagram F(κ(X)) / F(κ(Y )) F(OX,x) F(OY,y). Since γ ∈ Im(cid:0)F(OX,x) → F(κ(X))(cid:1), it follows that γκ(Y ) is contained in Im(cid:0)F(OY,y) → F(κ(Y ))(cid:1). Thus γκ(Y ) ∈ F(κ(Y ))Y −loc. (2) Assume now that γ ∈ F(κ(X))X−unr and let y ∈ Y (1). As above, we set x = f (y). Without loss of generality we may assume that and that X = Spec(A) = Spec(OX,x) Y = Spec(B) = Spec(OY,y) where B is a DVR. Let v be the discrete valuation on κ(Y ) corresponding to the valuation ring B. For brevity we denote by K (resp. L) the fraction field of A (resp. B). If v(K ×) = 0, then K ⊂ B and therefore γκ(Y ) ∈ Im(cid:0)F(OY,y) → F(κ(Y ))(cid:1). Assume now that v(K ×) 6= 0. Then mAB 6= B so that A → B is a local morphism. By [EGAI, 0.6.6.2], since B is flat over A the ring B is a faithfully flat A -- module. It follows that A = B ∩ K (apply [Bbk, § I.3, §5, prop. 10] with F = K and F ′ = A. An alternative proof can be given by appealing to [EGAIV, 2.1.13]). Let Av = {x ∈ K × v(x) ≥ 0} be the valuation ring of vK. Then Av = K ∩ B = A, so that A is a DVR. This implies that γ ∈ Im(cid:0)F(OX,x) → F(κ(X))(cid:1), and the commutative diagram above yields that γκ(Y ) ∈ Im(cid:0)F(OY,y) → F(κ(Y ))(cid:1). Thus γκ(Y ) ∈ F(κ(Y ))Y −unr. (cid:3) / / / O O O O 6 V. CHERNOUSOV, P. GILLE, AND A. PIANZOLA We shall discuss next the case of non-abelian cohomology functors, but we remark that this technique and considerations can be applied to various interesting functors such as Brauer groups, Witt groups, unramified Galois cohomology... 3. Non-abelian cohomology ´et(X, G), H 1 3.1. Some terminology. Let X be a scheme and let G be an X -- group scheme. The pointed set of non-abelian Cech cohomology on the flat (resp. ´etale, Zariski) site of X with coefficients in G, is denoted by H 1 f ppf (X, G) (resp. H 1 Zar(X, G)). These pointed sets measure the isomor- phism classes of sheaf torsors over X under G with respect to the chosen topology (see [M, Ch. IV §1] and [DG] for basic definitions and references). If X = Spec(R), following customary usage and depending on the context, we also use the notation H 1 f ppf (X, G). Similarly for the ´etale and Zariski sites. f ppf (R, G) instead of H 1 If G is in addition affine, by faithfully flat descent all of our sheaf torsors are representable. They are thus torsors in the usual sense. For a G-torsor E we denote by EG the twisted form of G by inner automorphisms; it is an affine group scheme over X. Furthermore, if G is smooth all torsors are lo- cally trivial for the ´etale topology. In particular, H 1 f ppf (X, G). These assumptions on G hold in most of the situations that arise in our work. Also, as we mentioned in the introduction, by default our topology will be ´etale so that instead of H 1 ´et(X, G) we will write H 1(X, G). ´et(X, G) = H 1 Given an X -- group G and a morphisms Y → X of schemes, we let GY denote the Y -- group G ×X Y obtained by base change. For convenience, we will denote H 1(Y, GY ) by H 1(Y, G). Assuming G affine of finite type, a maximal torus T of G is a subgroup X-scheme T of G such that T ×X κ(x) is a maximal κ(x) -- torus of G ×X κ(x) for each point x ∈ X [SGA3, XII.1]. Here κ(x) denotes an algebraic closure of κ(x). A G -- torsor E is toral if the twisted group scheme E G admits a maximal X-torus. We denote by H 1 f ppf (X, G) consisting of (isomorphism) classes of toral X -- torsors under G. f ppf,toral(X, G) the subset of H 1 3.2. Torsion bijection. If E is an X -- torsor under G (not necessarily toral), according to [Gi, III.2.6.3.1] there exists a natural bijection τE : H 1 f ppf (X, E G) → H 1 f ppf (X, G), called the torsion bijection, which takes the class of the trivial torsor under EG to the class of E. It is easy to see that its restriction to classes of toral torsors induces a bijection H 1 f ppf,toral(X, E G) → H 1 f ppf,toral(X, G). TORSORS OVER LAURENT POLYNOMIAL RINGS 7 3.3. Acyclicity Theorem. The following theorem is the main tool for proving our main results. 3.4. Theorem. Let G be a smooth affine Rn -- group scheme. Assume that either (i) G is reductive and admits a maximal Rn-torus (equivalently G is "loop reductive" [GP3, cor. 6.3]); or (ii) there exists a linear (smooth, not necessary connected) algebraic group G over k and a loop torsor E under G ×k Rn such that G = E(G ×k Rn). Then a natural map H 1 toral(Rn, G) −→ H 1(Fn, G) is bijective. Proof. See [CGP2, th. 14.1] in case (i) and [GP3, th. 8.1] in case (ii). (cid:3) 3.5. Grothendieck -- Serre's conjecture. The following conjecture is due to Grothendieck -- Serre ([Se, Remarque 3], [Gr, Remarque 1.1.a]). 3.6. Conjecture. Let R be a regular local ring with fraction field K. If G a reductive group scheme over R then the natural map H 1(R, G) → H 1(K, G) has trivial kernel. If R contains an infinite field k (of any characteristic), the conjecture has been proven by Fedorov -- Panin [FP, PSV]. When G = G ×k R for some reductive k-group G, the so called "constant" case, was established before by Colliot-Th´el`ene and Ojanguren [CTO]. For our considerations we need a similar result for group schemes which are not necessary "connected". 3.7. Lemma. Let R be a regular local ring with fraction field K. Let G be an affine smooth group scheme over R which is an extension of a finite twisted constant group scheme F over R by a reductive group scheme G0 over R. Assume that Grothendieck -- Serre's conjecture holds for G0. Then the map H 1(R, G) → H 1(K, G) has trivial kernel. Proof. We consider the commutative exact diagram of pointed sets F(R) −−−−→ H 1(R, G0) −−−−→ H 1(R, G) −−−−→ H 1(R, F) ≀y ηy νy λy F(K) −−−−→ H 1(K, G0) −−−−→ H 1(K, G) −−−−→ H 1(K, F). Since λ is injective, an easy diagram chase shows that Ker η surjects onto Ker ν. But by hypothesis Ker η vanishes, so the assertion follows. (cid:3) As a corollary of Fedorov -- Panin's theorem we get the following facts. 8 V. CHERNOUSOV, P. GILLE, AND A. PIANZOLA 3.8. Corollary. Let R be a regular local ring containing an infinite field with fraction field K. Let G be an affine smooth group scheme over R which is an extension of a finite twisted constant group scheme over R by a reductive group scheme over R. Then the natural map H 1(R, G) → H 1(K, G) has trivial kernel. 3.9. Corollary. Let X be an integral smooth affine variety over an infinite field with function field K. Let G be an affine smooth group scheme over X which is an extension of a finite twisted constant group scheme over X by a reductive group scheme over X. Then the sequence of pointed sets 1 −→ H 1 Zar(X, G) −→ H 1(X, G) −→ H 1(K, G) is exact. 3.10. Rational torsors everywhere locally defined. For a smooth affine group scheme G over an integral normal scheme X Colliot-Th´el`ene and Sansuc [CTS, §6] introduced the following sets: DG(X) := Im(cid:16)H 1(X, G) → H 1(κ(X), G)(cid:17), and H 1(κ(X), G)X−loc := \ H 1(κ(X), G)X−unr := \ x∈X x∈X (1) DG(OX,x) ⊆ H 1(κ(X), G), DG(OX,x) ⊆ H 1(κ(X), G). Clearly, we have the inclusions DG(X) ⊆ H 1(κ(X), G)X−loc ⊆ H 1(κ(X), G)X−unr . In our terminology introduced in § 2 the two last sets are nothing but the local and unramified subsets with respect to X attached to the functor F given by F(Y ) = H 1(Y, G) for each X-scheme Y . Unramified classes have the following geometrical characterization. 3.11. Lemma. Let γ ∈ H 1(κ(X), G). Then γ ∈ H 1(κ(X), G)X−unr if and ´et(U, G) such only if there exists an open subset U of X and a class eγ ∈ H 1 (i) γ = (eγ)κ(X); (ii) X (1) ⊂ U . that Proof. In one direction the statement is obvious.The other one was treated in [GP2, cor. A.8]. (cid:3) A special case of Lemma 2.1 is then the following. 3.12. Lemma. Let f : Y → X be a dominant morphism of integral and normal S-schemes. Let F(f ∗) : H 1(κ(X), G) → H 1(κ(Y ), GY ) be the map induced by the comorphism f ∗ : κ(X) → κ(Y ). Then. TORSORS OVER LAURENT POLYNOMIAL RINGS 9 (1) F(f ∗)(H 1(κ(X), G)X−loc) ⊆ H 1(κ(Y ), G)Y −loc. (2) If f is flat then F(f ∗)(H 1(κ(X), G)X−unr ) ⊆ H 1(κ(Y ), G)Y −unr. 3.13. Remark. If X is regular and the "purity conjecture" holds for G and local rings of X, i.e. DG(OX,x) = H 1(κ(X), G)OX,x−unr for all points x ∈ X, then assertion (1) implies that (2) holds without flatness assumption for f . We now combine earlier work by Colliot-Th´el`ene/Sansuc and Nisnevich theorem [N] on Grothendieck-Serre conjecture for reductive groups over DVR. 3.14. Proposition. Assume that X is a regular integral scheme and that G is an extension of a finite twisted constant group scheme by a reductive group scheme G0. Then (1) DG defines a contravariant functor for the category of regular integral X-schemes. (2) If X = Spec(k), then H 1( , G)loc defines a contravariant functor for the category of smooth integral k -- varieties. Proof. Nisnevich's theorem states that if A is a DVR and G is a reductive group over A then the natural map H 1(R, G) → H 1(K, G), where K is a fraction field of A, is injective. By Lemma 3.7, it holds more generally for a group G which is an extension of a finite twisted constant group by a reductive group. In particular, if γ1, γ2 ∈ H 1(A, G) have the same image in H 1(K, G), they have the same specialization modulo the maximal ideal of A. By [CTS, 6.6.1], this specialization property holds more generally over an arbitrary valuation ring. Then the assertions follow from [CTS, 6.6.3] and [CT, proposition 2.1.10]. (cid:3) 4. Proof of Theorem 1.2 Let G be a smooth affine group scheme over Rn satisfying condition (i) or (ii) in Theorem 1.2. Clearly, Im [H 1(Rn, G) → H 1(Kn, G)] ⊂ H 1(Kn, G)Rn−unr, so that we have the factorization (∗) H 1 toral(Rn, G) φ −→ H 1(Kn, G)Rn−unr ψ −→ H 1(Fn, G). The Acyclicity Theorem 3.4 states that the composite map ψ ◦ φ is bijective; in particular, φ is injective and ψ is surjective. 4.1. Lemma. The following are equivalent. (i) φ is bijective. (ii) The map Eψ : H 1(Kn, E G)Rn−unr −→ H 1(Fn, E G) has trivial kernel for all toral Rn -- torsors E under G. 10 V. CHERNOUSOV, P. GILLE, AND A. PIANZOLA Proof. (i) =⇒ (ii): Assume that φ is bijective. The above factorization (∗) and the bijectivity of ψ ◦ φ yield that we have bijections H 1 toral(Rn, G) ∼−→ H 1(Kn, G)Rn−unr ∼−→ H 1(Fn, H). Let now E be a toral Rn -- torsor under G. Recall that the torsion bijection map induces a bijection τE : H 1(Rn, E G) ∼−→ H 1(Rn, G) H 1 toral(Rn, E G) ∼−→ H 1 toral(Rn, G) (see § 3.2). The commutative diagram of torsion bijections H 1 toral(Rn, G) ∼ / H 1(Kn, G)Rn−unr ∼ / H 1(Fn, G) ≀ ≀ ≀ H 1 toral(Rn, E G) H 1(Kn, E G)Rn−unr Eψ H 1(Fn, E G) shows that Eψ is bijective, so a fortiori has trivial kernel. (ii) =⇒ (i): We have noticed above that φ is injective, hence it remains to prove its surjectivity only. Let [E′] ∈ H 1(Kn, G)Rn−unr. Since ψ ◦ φ is bijective there exists a class [E] ∈ H 1(Rn, G) such that ψ([E′]) = ψ(φ([E])). It follows from the above commutative diagram that under the torsion bi- jection H 1(Kn, EG)Rn−unr → H 1(K, G)Rn−unr the class [E′] corresponds to an element in H 1(Kn, EG)Rn−unr lying in the kernel of Eψ. Since by our hypothesis Ker(Eψ) = 1, this implies that the class [E′] corresponds to the trivial one in H 1(Kn, E G)Rn−unr or equivalently φ([E]) = [E′]. (cid:3) Note that hypothesis (i) and (ii) in Theorem 1.2 are stable with respect to twisting by a toral Rn-torsor under G. Therefore the above lemma reduces the proof of Theorem 1.2 to showing that for all group schemes H over Rn satisfying conditions (i) or (ii) in Theorem 1.2 a natural map ψ : H 1(Kn, H)Rn−unr −→ H 1(Fn, H) has trivial kernel. To prove this fact we proceed by induction on n ≥ 1 by allowing the base field k to vary. n = 1: Since we are in dimension one, by Lemma 3.11 the map H 1(R1, H) −→ H 1(K1, H)R1−unr is onto. Therefore, Lemma 4.1 applied to the group scheme G = H and the trivial torsor E = 1 shows that ψ has trivial kernel. n ≥ 2: Consider the following field tower: Kn ⊂ Fn−1(tn) ⊂ Fn. / / / / O O / / O O O O TORSORS OVER LAURENT POLYNOMIAL RINGS 11 Let γ ∈ Ker(ψ) and let γ′ be its image in H 1(Fn−1(tn), H). Since the mor- phism of affine schemes Spec(Fn−1[t±1 n ]) → Spec(Rn) is flat and dominant, by Lemma 3.12 (2) we have γ′ ∈ H 1(cid:0)Fn−1(tn), H(cid:1)Fn−1[t±1 n ]−unr. Since Fn = Fn−1((tn)) and γFn = 1, we then conclude that γ′ ∈ Ker(cid:16)H 1(cid:0)Fn−1(tn), H(cid:1)Fn−1[t±1 n ]−unr → H 1(cid:0)Fn−1((tn)), H(cid:1)(cid:17). But according to the case n = 1 (applied to the base field Fn−1) the last kernel is trivial. Thus γ′ = 1, i.e. (∗∗) γ ∈ Ker(cid:16)H 1(cid:0)Kn, H)Rn−unr → H 1(Fn−1(tn), H(cid:1)(cid:17). Now, we observe that the field Fn−1(tn) = k((t1)) . . . ((tn−1))(tn) embeds into k(tn)((t1)) . . . ((tn−1)) = k′((t1)) . . . ((tn−1)) with k′ = k(tn), so that we have a commutative diagram H 1(cid:0)Kn, H(cid:1)Rn−unr ∩ −−−−→ H 1(cid:0)Fn−1(tn), H(cid:1) y H 1(cid:0)k′(t1, ..., tn−1), H(cid:1)Rn−1⊗kk′−unr −−−−→ H 1(cid:0)k′((t1)) . . . ((tn−1)), H(cid:1), where the left vertical inclusion again is due to Lemma 3.12 (2). By the induction hypothesis applied to the base field k′, the bottom horizontal map has trivial kernel. Therefore the top horizontal map has trivial kernel as well. By (∗∗), this implies γ = 1. 5. Applications 5.1. A disjoint union decomposition for Rn -- torsors. We shall use now our Theorem 1.2 and Fedorov-Panin's theorem to prove Theorem 1.1. In fact we will prove a little bit more general result by allowing G to be any group scheme satisfying conditions of Theorem 1.2. 5.2. Theorem. Let G be as in Theorem 1.2. Then there is a natural bijection G [E]∈H 1 toral(Rn,G) H 1 Zar(Rn, E G) Θ ∼−→ H 1(Rn, G). Proof. Recall first that the torsion bijection τE (see § 3.2) allows us to embed H 1 Zar(Rn, E G) ֒→ H 1(Rn, EG) τE ∼−→ H 1(Rn, G). and this, in turn, induces a natural map Θ : G [E]∈H 1 toral(Rn,G) H 1 Zar(Rn, E G) → H 1 ´et(Rn, G). Surjectivity of Θ. Assume [γ] ∈ H 1(Rn, G). Since the generic class γKn ∈ H 1(Kn, G) is Rn -- unramified by Theorem 1.2 there is a unique toral class 12 V. CHERNOUSOV, P. GILLE, AND A. PIANZOLA [E] ∈ H 1(Rn, G)toral such that [E]Kn = γKn. Consider the following com- mutative diagram H 1(Rn, G) −−−−→ H 1(Kn, G) τEx≀ τEx≀ H 1 Zar(Rn, E G) ι ֒→ H 1(Rn, E G) −−−−→ H 1(Kn, E G) with an exact bottom horizontal line (see Corollary 3.9). the diagram that that τ −1 η ∈ H 1 Injectivity of Θ. Let E, E′ be two toral torsors under G and let It follows from E (γ) ∈ Im ι. Hence there exists a unique class Zar(Rn, E G) such that η = τ −1 E (γ). By construction, Θ(η) = γ. η ∈ H 1 Zar(Rn, E G), η′ ∈ H 1 Zar(Rn, E ′ G) be such that Θ(η) = Θ(η′), i.e. τE(η) = τE ′(η′) ∈ H 1(Rn, G). Since η is locally trivial in the Zariski topology we conclude that and similarly for E′. It follows that τE(η)Kn = τE(1)Kn = [E]Kn [E]Kn = [E′]Kn ∈ H 1(Kn, G)Rn−unr and therefore, by Theorem 1.2, we have [E] = [E′]. But τE = τE ′ is bijective. Therefore, η = η′. (cid:3) 5.3. Example. Let G be a reductive k -- group with the property that all of its semisimple quotients are isotropic (for example, k-split). By a result of Raghunathan [R, th. B] one has H 1 Zar(An k , G) = 1 for all n ≥ 0. According to [GP2, prop. 2.2], any Zariski locally triv- ial G-torsor over Rn can be extended to a G-torsor over An k . Therefore H 1 Zar(Rn, G) = 1. This implies that the map H 1(Rn, G) → H 1(Kn, G) has trivial kernel. In other words, rationally trivial Rn -- torsors under G are trivial. We expect that the similar result holds in a more general case. 5.4. Conjecture. Let G be a loop reductive group scheme over Rn. If all semisimple quotients of G are isotropic, then H 1 Zar(Rn, G) = 1. 5.5. Remark. Note that Artamonov's freeness result [A] as well as Pari- mala's result [P] for quadratic forms over R2 are particular special cases of this conjecture Theorem 5.2 gives a classification of all G-torsors that involves first the description of all its toral torsors and then to studying locally trivial in Zariski topology torsors under its twisted toral forms. In the following two subsections we show how our theorem works in particular cases of orthogonal groups and projective linear groups. TORSORS OVER LAURENT POLYNOMIAL RINGS 13 5.6. The case of orthogonal groups. Let qsplit be a split quadratic form over k of dimension d ≥ 1 and let O(qsplit) be the corresponding split orthog- onal group. It is well-known that H 1(Rn, O(qsplit)) classifies non-singular quadratic Rn -- forms of dimension d [DG, III.5.2]. This allows us to iden- tify classes of torsors under G = O(qsplit,Rn) with classes of d-dimensional quadratic forms over the ring Rn. 5.7. Proposition. For each subset I ⊆ {1, ..., d}, we put tI = Qi∈I ti ∈ R× toral(Rn, O(qsplit)) contains a unique Rn -- quadratic form with the convention t∅ = 1. (1) Each class in H 1 of the shape n M I⊆{1,...,d} htI i ⊗ qI,Rn where all qI's with I 6= ∅ are non-singular anisotropic quadratic forms over k such that d = M I⊆{1,...,d} dim(qI ). (2) Let q be a non -- singular quadratic Rn -- form of dimension d. Then there exists a unique quadratic Rn-form qloop as in (1) such that q is a Zariski Rn -- form of qloop. Furthermore, q is isometric to qloop if and only if q is diagonalizable. Proof. (1) By Acyclicity Theorem 3.4 it suffices to compute H 1(Fn, O(qsplit)) or equivalently isometry classes of d-dimensional quadratic forms over Fn. Let q be such quadratic form. We want to show that it is as in (1). By the Witt theorem we may assume without loss of generality that q is anisotropic. We proceed by induction on n ≥ 0. The case n = 0 is obvious. Assume that n ≥ 1. Note that Fn = Fn−1((tn)). Springer's decomposition [EKM, §19], then shows that q ∼= q′ ⊕ htni q′′ where q′ and q′′ are (unique) anisotropic quadratic forms over Fn−1. By induction on n, q′ and q′′ are of the required form, hence we the assertion for q follows. The unicity is clear. (2) The first assertion follows from Theorem 5.2. If q is isometric to qloop, then q is diagonalizable since so is qloop. Conversely, assume that q is diag- onalizable: q = hb1, . . . , bd i. Since R× n /(R× n )2 ∼−→ k×/(k×)2 × htǫ1 1 . . . tǫn n ǫ1, . . . , ǫn = 0, 1 i. all coefficients of q are of the shape bi = ai tIi with Ii ⊆ {1, . . . , d} and ai ∈ k×. Then we can renumber b1, . . . , bd in such a way that q is as in (1) and we are done. (cid:3) 5.8. Corollary. Toral classes in H 1(Rn, O(qsplit)) correspond to diagonal- izable Rn-quadratic forms. 14 V. CHERNOUSOV, P. GILLE, AND A. PIANZOLA 5.9. The projective linear case. Let G = PGLd. The set H 1(Rn, PGLd) classifies Azumaya Rn-algebras of degree d [CF, prop. 2.5.3.13]. Recall also that if A is an Azumaya algebra over Rn of degree d, then there is a one- to-one correspondence between commutative ´etale Rn -- subalgebras of A of dimension d and maximal Rn -- tori of the group scheme PGL1(A). (This is a general well-known fact [SGA3, XIV.3.21.(b)]. Indeed, if S is a commutative ´etale Rn -- subalgebra of A of degree d then T = RS/Rn(Gm)/Gm is a maximal Rn -- torus of PGL1(A). Conversely, to a maximal Rn-subtorus T of PGL1(A) one associates the Rn -- subalgebra AT of fixed points under the natural action of T. This subalgebra AT has the required properties locally and hence globally). From the above discussion it follows that H 1 toral(Rn, PGLd) consists of isomorphism classes of Azumaya Rn -- algebras A of degree d having ´etale commutative Rn -- subalgebras of dimension d. We pass to description of locally trivial torsors under PGL1(A). 5.10. Lemma. Let A be an Azumaya algebra over Rn. Then the natural map H 1 Zar(Rn, PGL1(A)) is bijective. Zar(Rn, GL1(A)) → H 1 Proof. The exact sequence 1 → Gm → GL1(A) → PGL1(A) → 1 gives rise to a commutative diagram with exact rows (of pointed sets) 1 −−−−→ H 1(Rn, GL1(A)) −−−−→ H 1(Rn, PGL1(A)) −−−−→ Br(Rn) φ H 1(Kn, GL1(A)) = 1 −−−−→ H 1(Kn, PGL1(A)) −−−−→ Br(Kn). Every class in H 1(Rn, GL1(A) is rationally trivial, hence by Fedorov -- Panin's result [FP, PSV] it is locally trivial in the Zariski topology. In other words H 1(Rn, GL1(A)) = H 1 Zar(Rn, GL1(A)). Clearly φ(H 1 Zar(Rn, GL1(A))) ⊆ H 1 Zar(Rn, PGL1(A)). Conversely, let γ ∈ H 1 injective, we obtain γ ∈ Im(φ). Zar(Rn, PGL1(A)). Since ψ1(γ) = 1 and since ψ2 is Finally, it remains to note that the above diagram shows that φ has trivial kernel and this is true for all Azumaya algebras over Rn. Then the standard twisting argument enables us to conclude that φ is injective. (cid:3) Thus, the disjoint union decomposition of the set of isomorphism classes of torsors under PGLd becomes G [A]∈H 1 toral(Rn,PGLd) H 1 Zar(Rn, GL1(A)) Θ ∼−→ H 1(Rn, PGLd). y ψ1y ψ2y TORSORS OVER LAURENT POLYNOMIAL RINGS 15 In general we can't say much about the subset H 1 Zar(Rn, GL1(A)). Recall only that it classifies right invertible A-modules (so that the above decom- position is coherent with [GP2, prop. 4.8]). More precisely, if A is a toral Azumaya Rn -- algebra and L is a right invertible A-module, then the class [L] corresponds to the class of the Azumaya algebra EndA(L) under the map Θ. However, when the base field k is algebraically closed, toral Azumaya algebras over Rn are easy to classify explicitly, and also much more infor- mation about Zariski trivial torsors is available due to Artamonov's freeness statements [A]. More precisely, let k be an algebraically closed field. Choose a coherent system of primitive roots of unity (ζn)n≥1 in k. Given integers r, s satisfying 1 ≤ r ≤ s, we let A(x, y)s r denote the Azumaya algebra of degree s over the Laurent polynomial ring k[x±1, y±1] defined by a presentation X s = x, Y s = y, Y X = ζ r s XY. 5.11. Lemma. Let s1, . . . , sm, r1, . . . rm be positive integers such that (si, ri) = 1 for all i = 1, . . . , m. Then the Azumaya algebra A = A(t1, t2)s1 r1 ⊗ A(t3, t4)s2 r2 ⊗ · · · A(t2m−1, t2m)sm rm is a division algebra. Proof. Indeed, using the residue method it is easy to see that it is a division algebra even over the field F2m = k((t1)) . . . ((t2m)). (cid:3) We next recall that the group GLd(Z) acts in a natural way on the ring Rn, hence it acts on H 1(Rn, PGLd) (for details see [GP3, 8.4]). 5.12. Theorem. Assume that k is algebraically closed. (1) H 1 the following shape: toral(Rn, PGLd) consists of GLd(Z) -- orbits of classes of Rn -- algebras of A = Ms0(Rn) ⊗Rn A(t1, t2)s1 r1 ⊗ A(t3, t4)s2 r2 ⊗ · · · A(t2m−1, t2m)sm rm where the integers m, s0, s1, r1, . . . , sm, rm satisfy the following conditions: (i) 0 ≤ 2m ≤ n; 1 ≤ ri ≤ si and (ri, si) = 1 for all i = 1, ..., m; (ii) s0 ≥ 1 and s0s1 . . . sm = d. (2) A ⊗Rn Fn is division if and only if s0 = 1. (3) Let A be an Azumaya algebra as in (1). If s0 ≥ 2 we have H 1 Zar(Rn, GL1(A)) = 1. Proof. (1) If A ⊗Rn Fn is division, this is [GP3, th. 4.7]. Assume now that A ⊗Rn Fn is not division. By Wedderburn's theorem there exists an integer s ≥ 1 and a central division Fn -- algebra A′ such that A ⊗Rn Fn ∼= Ms(A′). The acyclicity theorem provides a toral Rn -- Azumaya algebra A′ such that A′ ⊗Rn Fn ∼= A′. Since Ms(A′) and A are isomorphic over Fn, the acyclicity theorem again shows that A ∼= Ms(A′). It remains to note that A′ being a division algebra over Rn is of the required form by the first case. 16 V. CHERNOUSOV, P. GILLE, AND A. PIANZOLA (2) The assertion follows from Lemma 5.11. (3) Write A = Ms(A′) with A′ is division and s ≥ 2. Morita equivalence pro- vides a one-to-one correspondence between invertible A -- modules and finitely generated projective A′ -- modules of relative rank s. Artamonov's result [A] states that those A′ -- modules are free [A, cor. 3] since s ≥ 2, so that invert- ible A -- modules are free as well. This implies H 1 (cid:3) Zar(Rn, GL1(A)) = 1. 5.13. Remarks. (a) The third statement shows that for each invertible A -- module P the module P ⊕ A is free. (b) The third statement refines Steinmetz's results in the 2-dimension case (n = 2) [St, th. 4.8] where the case s ≥ 3 was considered only. Note that [St] provides some other cases for classical groups when all Zariski locally trivial torsors are trivial. 5.14. Applications to Rn -- Lie algebras. We next consider the special case G = Aut(g) where g is a split simple Lie algebra over k of finite dimension. For such group the set H 1 ´et(Rn, G) classifies Rn-forms of the Lie algebra g ⊗k Rn and H 1 toral(Rn, G) classifies loop objects, i.e. those which arise from loop cocycles. More precisely, by [GP3, §6] we have Im(cid:16)H 1(cid:0)π1(Rn, 1), G(ks)) → H 1(Rn, G)(cid:17) = H 1 toral(Rn, G). Theorems 1.2 and 5.2 have the following consequences. 5.15. Corollary. (1) Let eL be a Kn -- form of the Lie algebra g ⊗k Kn. If eL is unramified, i.e. extends everywhere in codimension one, then eL is isomorphic to the generic fiber of a unique multiloop Lie algebra L. (Of course L, being a multiloop algebra, is a twisted from of the Rn -- Lie algebra g ⊗k Rn). (2) Let L be any Rn -- form of g⊗R Rn. Then there exists a (unique up to Rn -- isomorphism) multiloop Lie algebra Lloop over Rn such that L is a Zariski Rn -- form of Lloop. References [A] V. A. Artamonov, Projective modules over crossed products, J. Algebra 173 (1995), 696 -- 714. [Bbk] N. Bourbaki, Alg`ebre commutative, Ch. 1 `a 4, Springer. [CF] B. Calm`es, J. Fasel, Groupes classiques, to appear in Autour des sch´emas en groupes, vol II, Panoramas et Synth`eses (2015). [CGP1] V. Chernousov, P. Gille and A. Pianzola, Torsors over the punctured affine line, American Journal of Mathematics 134 (2012), 1541 -- 1583. [CGP2] V. Chernousov, P. Gille and A. Pianzola, Conjugacy theorems for loop reductive group schemes and Lie algebras, Bulletin of Mathematical Sciences 4 (2014), 281 -- 324. [CT] J. -- L. Colliot -- Th´el`ene, Birational invariants, purity and the Gersten conjecture, K- theory and algebraic geometry: connections with quadratic forms and division al- gebras (Santa Barbara, CA, 1992), 1 -- 64, Proc. Sympos. Pure Math. 58.1 (1995), AMS. TORSORS OVER LAURENT POLYNOMIAL RINGS 17 [CTO] J. -- L. Colliot -- Th´el`ene, M. Ojanguren, Espaces principaux homog`enes localement triviaux, I.H.´E.S. Publ. Math. 75 (1992), 97 -- 122. [CTS] J. -- L. Colliot -- Th´el`ene, J. -- J. Sansuc, Fibr´es quadratiques et composantes connexes r´eelles, Math. Annalen 244 (1979), 105 -- 134. [DG] M. Demazure et P. Gabriel, Groupes alg´ebriques, Masson (1970). [EKM] R. Elman, N. Karpenko, A. Merkurjev, The algebraic and geometric theory of qua- dratic forms, American Mathematical Society Colloquium Publications, 56. American Mathematical Society, Providence, RI, 2008. [EGAI] A. Grothendieck (avec la collaboration de J. Dieudonn´e), El´ements de G´eom´etrie Alg´ebrique I: Le langage des sch´emas, Publications math´ematiques de l'I.H.´E.S. no 4 (1960). [EGAIV] A. Grothendieck (avec la collaboration de J. Dieudonn´e), El´ements de G´eom´etrie Alg´ebrique IV: ´Etude locale des sch´emas et des morphismes de sch´emas, Publications math´ematiques de l'I.H.´E.S. no 20, 24, 28 and 32 (1964-1967). [FP] R. Fedorov, I. Panin, A proof of Grothendieck-Serre conjecture on principal bun- dles over regular local rings containing infinite fields, to appear in Publications Math´ematiques de l'I.H.´E.S. [GP1] P. Gille and A. Pianzola, Galois cohomology and forms of algebras over Laurent polynomial rings, Mathematische Annalen 338 (2007), 497 -- 543. [GP2] P. Gille and A. Pianzola, Isotriviality and ´etale cohomology of Laurent polynomial rings, J. Pure Appl. Algebra 212 (2008), 780 -- 800. [GP3] P. Gille and A. Pianzola, Torsors, reductive group schemes and extended affine Lie algebras, Memoirs of AMS 1063 (2013). [Gi] J. Giraud, Cohomologie non ab´elienne, Die Grundlehren der mathematischen Wis- senschaften 179 (1971), Springer-Verlag. [Gr] A. Grothendieck, Le groupe de Brauer. II, Th´eorie cohomologique (1968), Dix Ex- pos´es sur la Cohomologie des Sch´emas, 67 -- 87, North-Holland. [KLP] V. Kac, M. Lau and A. Pianzola. Differential conformal superalgebras and their forms, Advances in Mathematics 222 (2009), 809 -- 861. [M] J. S. Milne, Etale cohomology, Princeton University Press (1980). [N] Y. A. Nisnevich, Espaces homog`enes principaux rationnellement triviaux et arithm´etique des sch´emas en groupes r´eductifs sur les anneaux de Dedekind, C. R. Acad. Sci. Paris S´er. I Math. 299 (1984), 5 -- 8. [PSV] I. Panin, A. Stavrova and N. Vavilov, On Grothendieck-Serres conjecture concerning principal G-bundles over reductive group schemes: I, Compositio Math. 151 (2015), 535 -- 567. [PZ] G. Pappas, X. Zhu, Local models of Shimura varieties and a conjecture of Kottwitz, Inventiones Mathematicae 194 (2013), 147 -- 254. [P] R. Parimala, Quadratic spaces over Laurent extensions of Dedekind domains, Trans. Amer. Math. Soc. 277 (1983), 569 -- 578. [R] M. S. Raghunathan, Principal bundles on affine space and bundles on the projective line, Math. Ann. 285 (1989), 309-332. [Se] J.-P. Serre, Espaces fibr´es alg´ebriques, S´eminaire Claude Chevalley 3 (1958), Expos´e no. 1, 37 p. [SGA1] S´eminaire de G´eom´etrie alg´ebrique de l'I.H. ´E.S., Revetements ´etales et groupe fondamental, dirig´e par A. Grothendieck, Lecture Notes in Math. 224 (1971), Springer. [SGA3] S´eminaire de G´eom´etrie alg´ebrique de l'I.H. ´E.S., 1963-1964, sch´emas en groupes, dirig´e par M. Demazure et A. Grothendieck, Lecture Notes in Math. 151-153 (1970), Springer. [St] A. Steinmetz-Zikesch, Alg`ebres de Lie de dimension infinie et th´eorie de la descente, M´emoire de la SMF 129 (2012). 18 V. CHERNOUSOV, P. GILLE, AND A. PIANZOLA Department of Mathematics, University of Alberta, Edmonton, Alberta T6G 2G1, Canada E-mail address: [email protected] UMR 5208 Institut Camille Jordan - Universit´e Claude Bernard Lyon 1 43 boulevard du 11 novembre 1918 69622 Villeurbanne cedex - France Institute of Mathematics Simion Stoilow of the Romanian Academy, Calea Grivitei 21, RO-010702 Bucharest, Romania. E-mail address: [email protected] Department of Mathematics, University of Alberta, Edmonton, Alberta T6G 2G1, Canada. Centro de Altos Estudios en Ciencia Exactas, Avenida de Mayo 866, (1084) Buenos Aires, Argentina. E-mail address: [email protected]
1902.04430	2	1902	2019-02-15T12:31:53	Ore extensions for function algebras	[ "math.RA" ]	In this article we consider the Ore extension Algebra for the algebra $\mathcal{A}$ of functions with finite support on a countable set. We derive explicit formulas for twisted derivations on $\mathcal{A}.$ We give a description for the centralizer of $\mathcal{A},$ and the center of the Ore extension algebra when the derivation is zero.	math.RA	math	Ore extensions of functional algebras Alex Behakanira Tumwesigye, Johan Richter, Sergei Silvestrov 9 1 0 2 b e F 5 1 ] . A R h t a m [ 2 v 0 3 4 4 0 . 2 0 9 1 : v i X r a Abstract In this article we consider the Ore extension algebra for the algebra A of functions with finite support on a countable set. We derive explicit formulas for twisted derivations on A , give a description for the centralizer of A , and the center of the Ore extension algebra under specific conditions. 1 Introduction A topic of interest in the field of operator algebras is the connection between prop- erties of dynamical systems and algebraic properties of crossed products associated to them. More specifically the question when a certain canonical subalgebra is max- imal commutative and has the ideal intersection property, that is, each non-zero ideal of the algebra intersects the subalgebra non-trivially. For a topological dy- namical systems (X,σ), one may define a crossed product C-algebra C(X) ⋊ σ Z where σ is an automorphism of C(X) induced by σ. It turns out that the property known as topological freeness of the dynamical system is equivalent to C(X) being a maximal commutative subalgebra of C(X) ⋊ σ Z and also equivalent to the con- dition that every non-trivial closed ideal has a non-zero intersection with C(X). An excellent reference for this correspondence is [19]. For analogues, extensions and applications of this theory in the study of dynamical systems, harmonic analysis, Alex Behakanira Tumwesigye Department of Mathematics, College of Natural Sciences, Makerere University, Box 7062, Kam- pala, e-mail: [email protected] Johan Richter Division of Applied Mathematics, UKK, Malardalens University, Box 883, Vasteras e-mail: [email protected] Sergei Silvestrov Division of Applied Mathematics, UKK, Malardalens University, Box 883, Vasteras e-mail: [email protected] 1 2 Alex Behakanira Tumwesigye, Johan Richter, Sergei Silvestrov quantum field theory, string theory, integrable systems, fractals and wavelets, see [1, 2, 4, 5, 6, 10, 11, 19]. For any class of graded rings, including gradings given by semigroups or even filtered rings (e.g. Ore extensions), it makes sense to ask whether the ideal intersec- tion property is related to maximal commutativity of the degree zero component. For crossed product-like structures, where one has a natural action, it further makes sense to ask how the above mentioned properties of the degree zero component are related to properties of the action. These questions have been considered recently for algebraic crossed products and Banach algebra crossed products, both in the tra- ditional context of crossed products by groups as well as generalizations to crossed products by groupoids and general categories [3, 9, 12, 13, 17, 18, 21, 22]. Ore extensions constitute an important class of rings, appearing in extensions of differential calculus, in non-commutative geometry, in quantum groups and algebras and as a uniting framework for many algebras appearing in physics and engineering models. An Ore extension of a ring R is an overring with a generator x satisfying xr = σ(r)x + ∆(r), r ∈ R, for some endomorphism σ and a σ-derivation ∆. This article aims at giving a description of the centralizer and the center of the coefficient subalgebra A in the Ore extension algebra A [x, σ,∆], where A is the algebra of functions with finite support on a countable set X and σ : A → A is an automorphism of A that is induced by a bijection σ : X → X . A number of studies on centralizers in Ore extensions have been carried out before in [14, 15, 20], but this is in completely different settings the one here. In section 2, we recall some notation and basic facts about Ore extensions used through out the rest of the article. In section 3, we give a description of twisted derivations on the algebra of functions on a finite set from which it is observed that there are no non trivial derivations on Rn. In section 4, we give the description of the centralizer of the coefficient algebra A and the center of the Ore extension A [x, σ, 0]. In section 5, we turn to the case when A is the algebra of functions on a countable set with finite support, give a description for the centralizer and the center of the Ore extension A [x, σ, 0]. Sections 6 and 7 are devoted to the special case of the skew power series and the skew Laurent rings respectively. 2 Ore extensions. Basic preliminaries For general references on Ore extensions, see for example, [7, 16]. For the conve- nience of the reader, we recall the definition. Let R be a ring , σ : R → R a ring endomorphism (not necessarily injective) and ∆ : R → R a σ-derivation, that is, ∆(a + b) = ∆(a) + ∆(b) and ∆(ab) = σ(a)∆(b) + ∆(a)b for all a, b ∈ R. Ore extensions of functional algebras 3 Definition 1. The Ore extension R[x,σ,∆] is defined as the ring generated by R and an element x /∈ R such that 1, x, x2, . . . form a basis for R[x,σ,∆] as a left R-module and all r ∈ R satisfy xr = σ(r)x + ∆(r). (1) Such a ring always exists and is unique up to isomorphism [7]. From ∆(1 · 1) = σ(1) · 1 + ∆(1) · 1 we get that ∆(1) = 0, and since σ(1) = 1 we see that 1R will be the multiplicative identity for R[x,σ,∆] as well. If σ = idR, then we say that R[x, idR,∆] is a differential polynomial ring. If instead ∆ ≡ 0, then we say that R[x,σ, 0] is a skew polynomial ring. The reader should be aware that some authors use the term skew polynomial ring to mean Ore extensions. An arbitrary non-zero element P ∈ R[x,σ,∆] can be written uniquely as P = aixi for some n ∈ Z≥0, with ai ∈ R for i ∈ {0, 1, . . . , n} and an 6= 0. The degree of n ∑ i=0 P will be defined as; deg(P) := n. We set deg(0) := −∞. Definition 2. A σ-derivation ∆ is said to be inner if there exists some a ∈ R such that ∆(r) = ar − σ(r)a for all r ∈ R. A σ-derivation that is not inner is called outer. Given a ring S we denote its center by Z(S). The centralizer C(T ), of a subset T ⊆ S is defined as the set of elements of S that commute with every element of T . If T is a commutative subring of S and the centralizer of T in S coincides with T , then T is said to be a maximal commutative subring of S. 3 Derivations on algebras of functions on a finite set Let X = [n](= {1, 2, · · · , n}) be a finite set and let A = { f : X → R} denote the algebra of real-valued functions on X with respect to the usual pointwise operations, that is, pointwise addition, scalar multiplication and pointwise multiplication. By writing fn := f (n), then we can identify A with Rn. Here, Rn is equipped with the usual operations of pointwise addition, scalar multiplication and multiplication defined by xy = (x1y1, x2y2, · · · , xnyn) for every x = (x1, x2, · · · , xn) and y = (y1, y2, · · · , yn). Now, let σ : X → X be a bijection such that A is invariant under σ, (that is σ is a permutation on X ) and let σ : A → A be the automorphism induced by σ, that is σ( f ) = f ◦ σ−1 (2) for every f ∈ A . We would like to consider the Ore extension A [x; σ,∆] where ∆ is a σ− derivation on A and x is an indeterminate. Recall that ∆ is a σ derivation on A if it is R−linear and for every f , g ∈ A , ∆( f g) = σ( f )∆(g) + ∆( f )g. 4 Alex Behakanira Tumwesigye, Johan Richter, Sergei Silvestrov Since A can be identified with Rn, then ∆ is an operator on Rn which can be repre- sented by a matrix, that is; [∆] = [∆(e1)∆(e2) · · · ∆(en)], (3) where {e1, e2, · · · , en} is the standard basis of Rn. In the following Theorem we give the description of the matrix [∆] in (3) above. Theorem 1. Let σ : X → X be a bijection and let ∆ be a σ−derivation whose stan- dard matrix [∆] = [ki j] is as given by (3). Then 1. kli = 0 if l 6∈ {i,σ(i)}, 2. k ji = −k j j for all i 6= j. Proof. 1. If σ(i) = j, then σ(ei) = e j, where σ is as defined by (2) and {ei, i = 1, 2, · · · , n} is the standard basis for Rn. Therefore from the definition of ∆, ∆(e2 i ) = σ(ei)∆(ei) + ∆(ei)ei = e j∆(ei) + ∆(ei)ei = ∆(ei)(e j + ei). Now e2 i = ei and hence ∆(ei) = ∆(ei)(ei + e j). The term on the right side is zero except in the ith and jth component. Therefore, ∆(ei) =  k1i k2i ... kni   where kli = 0 whenever l 6= i, j. 2. For i 6= j, ∆(eie j) = ∆(0) = 0 and hence 0 = ∆(eie j) = σ(ei)∆(e j) + ∆(ei)e j = e j∆(e j) + ∆(ei)e j = e j(∆(e j) + ∆(ei)), where the right hand side term is zero except in the jth component. Looking at the jth component we get ⊓⊔ k j j + k ji = 0 or k ji = −k j j. Corollary 3.1 There are no non zero derivations on A . Ore extensions of functional algebras 5 Proof. If σ = id, that is, σ(ei) = ei, i = 1, 2, · · · , n, then from Theorem 1, it follows that if [∆] = [ki j], then ki j = 0 for all i 6= j, and from kii = −ki j we have kii = 0 for all i = 1, 2, · · · , n. Therefore, ∆ = 0. ⊓⊔ Example 3.1 Let n = 3 and let σ : [3] → [3] be a permutation such that σ(1) = 2, σ(2) = 3 and σ(3) = 1. Then σ( f )(1) = f(cid:0)σ−1(1)(cid:1) = f (3) σ( f )(2) = f(cid:0)σ−1(2)(cid:1) = f (1) σ( f )(3) = f(cid:0)σ−1(3)(cid:1) = f (2) Therefore σ(x1, x2, x3) = (x3, x1, x2). Let k11 k12 k13 k21 k22 k23 k31 k32 k33 [∆] =    be the standard matrix for ∆ and let x = (x1, x2, x3) and y = (y1, y2, y3) be arbitrary vectors in R3. Then x3 x1 x2     σ(x)∆(y) =  =      k11y1 k12y2 k13y3 k21y1 k22y2 k23y3 k31y1 k32y2 k33y3 k11x3y1 k12x3y2 k13x3y3 k21x1y1 k22x1y2 k23x1y3 k31x2y1 k32x2y2 k33x2y3 y1 y2 y3     k11x1 k12x2 k13x3 k21x1 k22x2 k23x3 k31x1 k32x2 k33x3 ∆(x)y =  =  σ(x)∆(y) + ∆(x)y = k31(x2y1 + x1y3) k32(x2y2 + x2y3) k33(x2y3 + x3y3)   k11(x1y1 + x3y1) k12(x2y1 + x3y2) k13(x3y1 + x3y3) k21(x1y2 + x1y1) k22(x1y2 + x2y2) k23(x1y3 + x3y2) k11x1y1 k12x2y1 k13x3y1 k21x1y2 k22x2y2 k23x3y2 k31x1y3 k32xxy3 k33x3y3     Also, . 6 Alex Behakanira Tumwesigye, Johan Richter, Sergei Silvestrov ∆(xy) =  =  k11 k12 k13 k21 k22 k23 k31 k32 k33 x1y1 x2y2 x3y3       k11x1y1 k12x2y2 k13x2y3 k21x1y1 k22x2y2 k23x3y3 k31x1y1 k32x2y2 k33x3y3  . ∆(xy) = σ(x)∆(y) + ∆(x)y Therefore, from we get k11x3y1 + k12(x2y1 + x3y3 − x2y2) + k13x3y1 = 0, from which we obtain, k13 = −k11 and k12 = 0. Similarly, we obtain k21 = −k22, k23 = 0, and k31 = 0, k32 = −k33 which is agreement with the assertions of Theorem 1. Setting k11 = s, k22 = t and k33 = u we obtain the matrix of ∆ as; [∆] =  s −t 0 −s 0 t 0 −u u  . In the next Theorem, we prove that if σ : Rn → Rn is an automorphism of Rn and ∆ : Rn → Rn is an operator on Rn that satisfies the conditions of Theorem 1, then ∆ is a σ− derivation. Theorem 2. Let σ : X → X be a bijection on X and let σ : Rn → Rn be the auto- morphism induced by σ. Let ∆ : Rn → Rn be an operator whose standard matrix [∆] = [ki j] has the following properties 1. kli = 0 if l 6∈ {i,σ(i)}, 2. k ji = −k j j. Then ∆ is a σ − derivation. Proof. We first do the proof for the standard basis {e1, e2, · · · , en} of Rn. Recall that if σ(i) = j then σ(ei) = e j. From the definition of ∆, ∆(ei) =  where kli = 0 for all l 6= i, j and k ji = −k j j. Now, for i 6= j, ∆(eie j) = ∆(0) = 0. And k1i k2i ... kni   σ(ei)∆(e j) + ∆(ei)e j = e j∆(e j) + ∆(ei)e j = e j(∆(e j) + ∆(ei)). Ore extensions of functional algebras 7 All the components in the above vector are zero except the jth component which is given by k j j + k ji = k j j − k j j = 0. Also, ∆(e2 i ) = ∆(ei), (since e2 i = ei), and σ(ei)∆(ei) + ∆(ei)ei = e j∆(ei) + ∆(ei)ei = ∆(ei)(ei + e j), where all the components in the above vector are zero except the ith and jth compo- nent. That is (assuming i < j) σ(ei)∆(ei) + ∆(ei)ei = = ∆(ei). 0 ... kii 0 ... k ji 0 ... 0     Therefore, ∆ is a σ−derivation on the standard basis vectors. Now let x = (x1, x2, · · · , xn) and y = (y1, y2, · · · , yn) be arbitrary vectors in Rn. Then and x = n ∑ i=1 xiei and y = n ∑ j=1 y je j xy = n ∑ i, j=1 xiy jeie j. Using the fact that both σ and ∆ are R−linear we have; 8 Alex Behakanira Tumwesigye, Johan Richter, Sergei Silvestrov ∆(xy) = ∆ n ∑ i, j=1 xiy jeie j! = = = = n ∑ i, j=1 n ∑ i, j=1 n ∑ i, j=1 n ∑ i, j=1 ∆(xiy jeie j) xiy j∆(eie j) xiy j( σ(ei)∆(e j) + ∆(ei)e j) xiy j σ(ei)∆(e j) + n ∑ i, j=1 xiy j∆(ei)e j xi σ(ei)! n ∑ j=1 y j∆(e j)! + n ∑ i=1 ∑ i=1 = n = σ n ∑ i=1 xiei!! ∆ n ∑ j=1 y je j!! + ∆ n ∑ i=1 y je j! xi∆(ei)! n ∑ j=1 y je j! xiei!! n ∑ j=1 = σ(x)∆(y) + ∆(x)y. Therefore ∆ is a σ−derivation on Rn. ⊓⊔ 4 Centralizers in Ore extensions for functional algebras Consider the Ore extension A [x, σ,∆], that is, the algebra A [x, σ,∆] :=( m ∑ k=0 fkxk, : fk ∈ A) with the operations of pointwise addition, scalar multiplication, and multiplication given by the relation x f = σ( f )x + ∆( f ) for every f ∈ A . Our interest is to give a description of the centralizer C(A ), of A in the Ore extension A [x, σ,∆] where σ and ∆ are as described before. Using the notation introduced in [20], we define functions πl 0 = id. If m, n are nonzero integers such that m > n or atleast one of the integers is negative, then πn k : A → A , for k, l ∈ Z as follows; π0 m = 0. For the other remaining cases, m = σ ◦ πn−1 πn m−1 + ∆ ◦ πn−1 m . Ore extensions of functional algebras 9 It has been proven in [20] that an element tralizer of A if and only if m ∑ k=0 fkxk ∈ A [x, σ,∆] belongs to the cen- g fk = m ∑ j=k f jπ j k (g) (4) holds for all k ∈ {0, 1, · · · , m} and all g ∈ A . Observe that since A is commutative, then the centralizer C(A ) of A is also commutative and hence a maximal commutative subalgebra of A [x, σ,∆]. 4.1 Centralizer for the case ∆ = 0 In this section we treat the simplest case when ∆ = 0. Recall that σ acts like a permutation of the elements of A and since [n] is a finite set, then σ is of finite order. Before we give the description, we need the following definition. Definition 3. For any nonzero n ∈ Z, set Sepn(X) = {x ∈ X x 6= σn(x)}, Pern(X) = {x ∈ X x = σn(x)}. Observe that σn(h)(x) 6= h(x) if and only if σn(x) 6= x for every x ∈ X and every h ∈ A . We give the description of the centralizer in the following theorem. Theorem 3. The centralizer C(A ), of A in the Ore extension A [x, σ,∆] is given by C(A ) =( m ∑ k=0 fkxk such that fk = 0 on Sepk(X)). Proof. From equation (4) and the fact that ∆ = 0, we see that an element A [x, σ,∆] belongs to the centralizer of A if and only if m ∑ k=0 fkxk ∈ for every k = 0, 1, · · · , m and every g ∈ A . That is, g fk = fk σk(g) g(x) fk(x) = fk(x) σk(g)(x) (5) for every x ∈ X. Since A is commutative, then equation (5) will hold if and only if x ∈ Perk(X) or fk = 0. Therefore, the centralizer C(A ), of A will be given by C(A ) =( m ∑ k=0 fkxk : fk ∈ A where fk = 0 on Sepk(X)). ⊓⊔ 10 Alex Behakanira Tumwesigye, Johan Richter, Sergei Silvestrov 4.2 Centralizer for the case ∆ 6= 0 Now, suppose σ 6= id is of order j ∈ Z>0, that is σ j = id but σk 6= id for all k < j. In the next Theorem, we state a necessary condition for an element in the Ore extension to belong to the centralizer. Theorem 4. If an element of degree m, izer of A , then fm = 0 on Sepm(X). m ∑ k=0 fkxk ∈ A [x, σ,∆] belongs to the central- Proof. As already stated an element of A if and only if for every g ∈ A m ∑ k=0 fkxk ∈ A [x, σ,∆] belongs to the centralizer g fk = m ∑ j=k f jπ j k (g) for k = 0, 1, · · · , m. Looking at the leading coefficient we have g fm = fmπm m (g) = fm σm(g) (6) Since A is commutative, then equation (6) holds on Perm(X) and on Sepm(X), equation (6) holds if and only if fm = 0. ⊓⊔ The above condition is not sufficient to describe all the elements that belong to the centralizer of A . In the next example we show that conditions satisfied by all elements in the centralizer of A are actually quite complicated even for the case when n = 2. Example 4.1 Let n = 2 and let σ : [n] → [n] be a bijection on [n]. We already know that if σ = id, then ∆ = 0 so we will consider the case σ 6= id, that is σ(e1) = e2 and σ(e2) = e1. In this case, ∆ has a standard matrix given by [∆] =(cid:20) s −s t (cid:21) −t for some s,t ∈ R with s,t 6= 0. A direct calculation shows that an element f x ∈ A [x, σ,∆] belongs to the cen- tralizer of A if and only if f = 0. So we consider a monomial of degree 2. f2(cid:19)x2 ∈ A [x, σ,∆] be an element in the centralizer of A . Then, if Let f =(cid:18) f1 g2(cid:19) ∈ A , we have, g =(cid:18)g1 g f =(cid:18)g1 g2(cid:19)(cid:18) f1 f2(cid:19)x2 =(cid:18)g1 f1 g2 f2(cid:19)x2. Ore extensions of functional algebras 11 On the other hand, using the fact that for every g ∈ A , x2g = σ2(g)x2 + [∆( σ(g)) + σ(∆(g))]x + ∆2(g) and since σ2 = id we have, −t f g =(cid:18) f1 g1(cid:19) + σ(cid:18)(cid:18) s −s f2(cid:19)(cid:20)(cid:18) s −s t (cid:19)(cid:18)g2 g2(cid:19)x2 +(cid:18) f1 f2(cid:19)(cid:18)g1 f2(cid:19)(cid:18) s2 + st −(s2 + st) g2(cid:19). st + t 2 (cid:19)(cid:18)g1 +(cid:18) f1 f2g2(cid:19)x2 +(cid:20)(cid:18)− f1(s + t)(g1 − g2) =(cid:18) f1g1 f2(s + t)(g1 − g2) (cid:19)(cid:21)x +(cid:20) f1(s2 + st)(g1 − g2) − f2(st + t 2)(g1 − g2)(cid:21) g2(cid:19)(cid:19)(cid:21)x −(st + t 2) −t t (cid:19)(cid:18)g1 Solving f g = g f and looking at the coefficient of x2, we get that f1, f2 are free variables and hence the centralizer C(A ) is non trivial. In the more general case, we have the following. As already seen, an element A if and only if for every g ∈ A m ∑ k=0 fkxk ∈ A [x, σ,∆] belongs to the centralizer of g fk = m ∑ j=k f jπ j k (g) for k = 0, 1, · · · , m. Looking at the constant term and using the fact that π j 0(g) = ∆ j(g) we have; g f0 = m ∑ j=0 f jπ j 0(g) = f0g + m ∑ j=1 f j∆ j(g) m ∑ j=1 f j∆ j(g) = 0. from which we obtain that Now, for any g ∈ A , g =(cid:18)g1 Therefore, from equation (7), we have g2(cid:19), we have g2(cid:19) =  ∆ j(g) = ∆ j(cid:18)g1 (g1 − g2) ∑ j −(g1 − g2) ∑ j k=0(cid:18) j − 1 k=0(cid:18) j − 1 k (cid:19)s j−kt k k (cid:19)skt j−k (7) .   12 Alex Behakanira Tumwesigye, Johan Richter, Sergei Silvestrov   (g1 − g2) ∑m j=1 f j1 ∑ −(g1 − g2) ∑m j=0 f j2 ∑ j k=0(cid:18) j − 1 k=0(cid:18) j − 1 k (cid:19)s j−kt k k (cid:19)skt j−k j   =(cid:20)0 0(cid:21). Since equation (7) should hold for every g ∈ A , then we have (8) Since s,t 6= 0, then from equation (8), we have j=1 f j1 ∑ j=0 f j2 ∑ j k=0(cid:18) j − 1 k=0(cid:18) j − 1 j ∑m ∑m     k=0(cid:18) j − 1 ∑m ∑m ∑ j j=1 f j1 ∑ j=0 f j2 ∑ j−1 k=0(cid:18) j − 1 k=0(cid:18) j − 1 j−1 k (cid:19)s j−k−1t k = =(cid:20)0 0(cid:21). k (cid:19)s j−kt k k (cid:19)skt j−k    k (cid:19)s j−k−1t k  k (cid:19)skt j−k−1 k (cid:19)skt j−k−1. k=0(cid:18) j − 1 =(cid:20)0 0(cid:21). ∑ j Observe that Therefore we get a matrix equation of the form (cid:20) f11 f21 f31 · · · f12 f22 f32 · · ·  fm2(cid:21) fm1  1 s + t s2 + 2st + t 2 ... k=0 (cid:18)m − 1 k (cid:19)sm−k−1tt ∑m−1 =(cid:20)0 0(cid:21) (9)   which always has nontrivial solutions if m > 2. 4.3 Center of the Ore extension algebra In the following section, we give a description of the center of our Ore extension algebra. We will give the description for the case ∆ = 0. Theorem 5. The center of the Ore extension algebra A [x, σ, 0] is given by Z(A [x, σ, 0]) =( m ∑ k=0 fkxk : where fk = 0 on Sepk(X) and σ( fk) = fk). Ore extensions of functional algebras 13 Proof. Let f = m ∑ k=0 fkxk be an element in Z(A [x, σ, 0]), then f ∈ C(A ), that is fk(x) = 0 for every x ∈ Sepk(x). Since the Ore extension A [x, σ, 0] is associative, it is enough to derive conditions under which x f = f x. Now f x = m ∑ k=0 fkxk!x = m ∑ k=0 fkxk+1. On the other hand, m x f = x fkxk ∑ k=0 m x fkxk = = ∑ k=0 m ∑ k=0 σ( fk)xk+1. From which we obtain that x f = f x if and only if σ( fk) = fk. Therefore, Z(A [x, σ, 0]) =( m ∑ k=0 fkxk : where fk = 0 on Sepk(x) and σ( fk) = fk). ⊓⊔ 5 Infinite dimensional case Let J be a countable subset of R and let A be the set of functions f : J → J such that f (i) = 0 for all except finitely many i ∈ J. Then A is a commutative non-unital algebra with respect to the usual pointwise operations of addition, scalar multipli- cation and multiplication. For i ∈ J, let ei ∈ A denote the characteristic function of i, that is ei( j) = χ{i}( j) =(1 if i = j 0 if i 6= j. Then every f ∈ A can be written in the form f = ∑ i∈J fiei (10) where fi = 0 for all except finitely many i ∈ J. Let σ : J → J be a bijection and let σ : A → A be the automorphism of A induced by σ, that is, σ( f ) = f ◦ σ−1 14 Alex Behakanira Tumwesigye, Johan Richter, Sergei Silvestrov for every f ∈ A . We can still construct the non-unital Ore extension A [x, σ,∆] as follows A [x, σ,∆] :=( m ∑ k=0 fkxk where fk ∈ A) with addition and scalar multiplication given by the usual pointwise operations and multiplication determined by the relation ( f x)g = σ(g) f x + ∆(g) where ∆ is a σ−derivation on A . In the following Theorem, we state the necessary and sufficient conditions for ∆ : A → A to be a σ−derivation on A . Theorem 6. Let σ : J → J be a bijection and let σ : A → A be the automorphism induced by σ. A linear map ∆ : A → A is a σ−derivation on A , if and only if, for every i ∈ J 1. ∆(ei) = −∆(cid:0)eσ(i)(cid:1) and 2. ∆(ei)(k) = 0 if k 6∈ {i,σ(i)} Proof. Suppose ∆ is a σ−derivation on A and letσ(i) = j, then σ(ei) = e j. If i 6= j, then, ∆(eie j) = ∆(0) = 0. On the other hand, That is, for every k ∈ J, ∆(eie j) = σ(ei)∆(e j) + ∆(ei)e j = e j(∆(e j) + ∆(ei)). ∆(eie j)(k) = [e j(∆(ei) + ∆(e j))] (k) =(∆(ei) + ∆(e j) 0 if k = j if k 6= j. Therefore, since ∆(eie j) = 0, then ∆(ei) = −∆(e j). Also, for any k ∈ J ∆(e2 i )(k) = ( σ(ei)∆(ei) + ∆(ei)ei)(k) = ∆(ei)(e j + ei)(k) = 0 k 6∈ {i, j}. if Conversely, suppose ∆ : A → A is a R−linear map which satisfies conditions (1) and (2) for some bijection σ : J → J of J. We prove that ∆ is σ−derivation on A . Suppose σ(i) = j and consider the characteristic functions ei, e j for i 6= j. Since ∆ is a linear map, Ore extensions of functional algebras 15 On the other hand, ∆(eie j) = ∆(0) = 0. σ(ei)∆(e j) + ∆(ei)e j = e j∆(e j) + ∆(ei)e j = e j(∆(e j) + ∆(ei)) = 0. Therefore ∆(eie j) = σ(ei)∆(e j) + ∆(ei)e j (11) for i 6= j and the same holds for i = j. The fact that equation (11) holds for every f , g ∈ A follows from linearity of both σ and ∆ and equation (10). Therefore ∆ is a σ−derivation. ⊓⊔ Remark 5.1 It can be seen from Theorem 6 above that, if σ(i) = i for all i ∈ J, then ∆ = 0. That is, there are no non trivial derivations on A . 5.1 Centralizer for A In this section, we give a description of the centralizer C(A ) of A , in the Ore exten- sion A [x, σ,∆] and the center of the Ore extension. Since A is commutative, then by [20, Proposition 3.3], if σ is of infinite order, then A is maximal commutative, that is, C(A ) = A . Therefore we will focus on the case when σ is of finite order. We do this for two cases. 5.1.1 The case ∆ = 0 The following Theorem gives the description of the centralizer of A in the skew- polynomial ring A [x, σ, 0]. Theorem 7. The centralizer C(A ), of A in the Ore extension A [x, σ,∆] is given by C(A ) =( m ∑ k=0 fkxk such that fk = 0 on Sepk(X)) where Sepk(X) is as defined in definition 3. Proof. Let f = m ∑ k=0 fkxk ∈ A [x, σ, 0] be an element of degree m which belongs to C(A ). Then f g = g f should hold for every g ∈ A . Now, g f = g m ∑ k=0 fkxk = m ∑ k=0 g fkxk. 16 Alex Behakanira Tumwesigye, Johan Richter, Sergei Silvestrov On the other hand, f g = m ∑ k=0 fkxk!g = m ∑ k=0 fk(cid:16)xkg(cid:17) = m ∑ k=0 fk σk(g)xk. Therefore, g f = f g if and only if g fk = fk σk(g) for all k = 0, 1, · · · , m. Since A is commutative, then the above equation holds on Perk(X). Therefore, C(A ) =( m ∑ k=0 ⊓⊔ fkxk such that fk = 0 on Sepk(X)). 5.1.2 The case ∆ 6= 0 Now, suppose σ 6= id is of order j ∈ Z>0, that is σ j = id but σk 6= id for all k < j. In the next Theorem, whose proof is similar to Theorem 4, we state a necessary condition for an element in the Ore extension to belong to the centralizer. Theorem 8. Let σ : A → A be an automorphism on A . If an element of order m, m ∑ k=0 fkxk ∈ A [x, σ,∆] belongs to the centralizer of A , then fm = 0 on Sepm(X). 5.2 Center of A [x, σ,∆] when ∆ = 0 In the following section, we give a description of the center of our Ore extension algebra. We will give the description for the case ∆ = 0. Theorem 9. Then the center of the Ore extension algebra A [x, σ, 0] is given by Z(A [x, σ, 0]) =( m ∑ k=0 fkxk : where fk = 0 on Sepk(X) and σ( fk) = fk). Proof. Observe that since A is not unital, the proof of Theorem 5 does not work for Theorem 9, since the element x 6∈ A [x, σ, 0]. Therefore, we adopt the following proof. Denote A [x, σ,∆] by R and let f = fk = 0 on Sepk(X). Now let g = n ∑ l=0 m ∑ k=0 fkxk ∈ Z(R). Then f ∈ C(A ), that is glxl be an arbitrary element in R. Then Ore extensions of functional algebras 17 f g = m = ∑ = ∑ k,l k,l ∑ k=0 glxl! ∑ l=0 fkxk! n fk(xkgl)xl! fk σk(gl)xk+l! In the same way, it can be shown that g f = n ∑ l=0 glxl! m ∑ k=0 fkxk! = ∑ k,l gl σl( fk)xk+l. It follows that f g = g f if and only if fk σk(gl) = gl σl( fk). (12) for all k = 0, 1, · · · , m and all l = 0, 1, · · · , n. Now, fk = 0 on Sepk(X) and on Perk(X), we have σk = id. Therefore, equation (12) holds if and only if fkgl = gl σl( fk) for all l = 0, 1, · · · , n. Since A is commutative, we conclude that (12) holds iff σ( fk) = fk. Therefore Z(A [x, σ, 0]) =( m ∑ k=0 fkxk : where fk = 0 on Sepk(X) and σ( fk) = fk). ⊓⊔ 6 The skew power series ring As before, we let X = [n](= {1, 2, · · · , n}) be a finite set and let A = { f : X → R} denote the unital algebra of real-valued functions on X with respect to the usual pointwise operations. Let σ : X → X be a bijection such that A is invariant under σ, (that is σ is a permutation on X ) and let σ : A → A be the automorphism induced by σ, that is σ( f ) = f ◦ σ−1 (13) for every f ∈ A . Consider the skew ring of formal power series over A , A [x; σ]; that is the set ( ∞ ∑ n=0 fnxn such that fn ∈ A) 18 Alex Behakanira Tumwesigye, Johan Richter, Sergei Silvestrov with pointwise addition and multiplication determined by the relations x f = σ( f )x. That is, if f = ∑∞ n=0 fnxn and g = ∑∞ n=0 gnxn are elements of A [x; σ], then f + g = ∞ ∑ n=0 ( fn + gn)xn and f g = ∞ ∑ n=0 fnxn! ∞ ∑ n=0 gnxn! = ∞ ∑ n=0 n ∑ k=0 fk σk(gn−k)!xn. 6.1 Centralizer of A in the skew power series ring A [x; σ] In the next Theorem, we give the description of the centralizer of A in the skew power series ring A [x; σ]. Theorem 10. The centralizer C(A ), of A in the skew power series ring A [x; σ] is given by C(A ) =( ∑ n∈Z fnxn such that fn = 0 on Sepn(X)) where Sepk(X) is as given in definition 3. Proof. Let f = fnxn ∈ A [x; σ] be an element which belongs to C(A ). Then ∞ ∑ n=0 f g = g f should hold for every g ∈ A . Now, g f = g ∞ ∑ n=0 fnxn = ∞ ∑ n=0 g fnxn. On the other hand, f g = ∞ ∑ n=0 fnxn!g = ∞ ∑ n=0 fn(xng) = ∞ ∑ n=0 fn σn(g)xn. Therefore, g f = f g if and only if g fn = fn σn(g) for all n ∈ N. Since A is commutative, then the above equation holds on Pern(X). Therefore, Ore extensions of functional algebras 19 C(A ) =( ∑ n∈Z ⊓⊔ fnxn such that fn = 0 on Sepn(X)). 6.2 The center of the skew power series ring The next Theorem gives the description of the center for the skew power series ring A [x; σ]. Theorem 11. The center of the skew power series ring A [x; σ] is given by Z(A [x; σ]) =( ∞ ∑ k=0 fnxn : where fn = 0 on Sepn(X) and σ( fn) = fn). ∞ ∑ n=0 Proof. Let f = fnxn be an element in Z(A [x; σ]). Then f ∈ C(A ), that is fn(x) = 0 for every x ∈ Sepn(x). Since A [x; σ] is associative, it is enough to de- rive conditions under which x f = f x. Now f x = ∞ ∑ n=0 fnxn!x = ∞ ∑ n=0 fnxn+1. On the other hand, x f = x fnxn ∞ ∑ n=0 ∞ ∑ n=0 x fnxn = = ∞ ∑ n=0 σ( fn)xn+1. From which we obtain that x f = f x if and only if σ( fn) = fn. Therefore, Z(A [x; σ]) =( ∞ ∑ n=0 fnxn : where fn = 0 on Sepn(x) and σ( fn) = fn). ⊓⊔ 20 Alex Behakanira Tumwesigye, Johan Richter, Sergei Silvestrov 7 The skew-Laurent ring A [x, x−1; σ] The fact that σ is an automorphism of A naturally leads us to the consideration of the skew-Laurent ring A [x, x−1; σ]. Definition 4. Let R be a ring and σ an automorphism of R. By a skew-Laurent ring R[x, x−1;σ] we mean that 1. R[x, x−1;σ] is a ring, containing R as a subring, 2. x is an invertible element of R[x, x−1;σ], 3. R[x, x−1;σ] is a free left R−module with basis(cid:8)1, x, x−1, x2, x−2, · · ·(cid:9), 4. xr = σ(r)x,(cid:0)and x−1r = σ−1(r)x−1(cid:1) for all r ∈ R. As before, we let X = [n](= {1, 2, · · · , n}) and let A = { f : X → R} denote the unital algebra of real-valued functions on X with respect to the usual pointwise operations. Let σ : X → X be a bijection such that A is invariant under σ, (that is σ is a permutation on X ) and let σ : A → A be the automorphism induced by σ, as defined by equation (13). Consider the skew-Laurent ring A [x, x−1; σ], that is the set ( ∑ n∈Z fnxn : fn ∈ A and fn = 0 for all except finitely many n) with pointwise addition and multiplication determined by the relations x f = σ( f )x and x−1 f = σ−1x−1. In the next Theorem, we give the description of the centralizer of A in the skew- Laurent ring A [x, x−1; σ]. Theorem 12. The centralizer of A in the skew-Laurent extension A [x, x−1; σ] is given by C(A ) =( ∑ n∈Z fnxn : fn = 0 on Sepn(X)) where Sepk(X) is as given in Definition 3. Proof. Let f = ∑ n∈Z f g = g f should hold for every g ∈ A . Now, fnxn ∈ A [x, x−1; σ] be an element which belongs to C(A ). Then g f = g ∑ n∈Z fnxn = ∑ n∈Z g fnxn. On the other hand, f g = ∑ n∈Z fnxn!g = ∑ n∈Z fn(xng) = ∑ n∈Z fn σn(g)xn. Ore extensions of functional algebras 21 Therefore, g f = f g if and only if g fn = fn σn(g) for all n ∈ Z. Since A is commutative, then the above equation holds on Pern(X). Therefore, C(A ) =( ∑ n∈Z fnxn : fn = 0 on Sepn(X)). ⊓⊔ Acknowledgement This research was supported by the Swedish International Development Coopera- tion Agency (Sida), International Science Programme (ISP) in Mathematical Sci- ences (IPMS), Eastern Africa Universities Mathematics Programme (EAUMP). Alex Tumwesigye is also grateful to the research environment Mathematics and Applied Mathematics (MAM), Division of Applied Mathematics, Malardalen Uni- versity for providing an excellent and inspiring environment for research education and research. References 1. Archbold, R. J., Spielberg, J. S., (1994). Topologically free actions and ideals in descrete C-dynamical systems, Proc. Edinburgh Math. Soc. (2) 37, no. 1, 119-124. 2. Carlsen, T. M., and Silvestrov, S. D., (2009). On the Excel crossed product of topological covering maps, Acta Appl. Math. 108, no. 3, 573-583. 3. de Jeu, M., Svensson, C., Tomiyama, J., (2012). On the Banach -algebra crossed product associated with a topological dynamical system, J. Funct. Anal. 262, no. 11, 4746-4765 4. Dutkay, D. E., Jorgensen, P.E.T., Silvestrov, S., (2012). Decomposition of wavelet represen- tations and Martin boundaries, J. Funct. Anal. 262, no. 3, 1043-1061 5. Dutkay, D. E., Jorgensen, P.E.T., (2007). Martingales, endomorphisms and covariant sys- tems of operators in Hilbert space, J. Operator Theory. Anal. 58, no. 2, 269-310 6. Exel, R., Vershik, A., (2006). C-algebras of irreversible dynamical systems, Canad. J. Math. 58, no. 1, 39-63 7. Goodearl, K., R., Warfield, R., B., (2004). An introduction to non commutative Noetherian rings. Second edition. London Mathematical Society Student Texts, 61. Cambridge Univer- sity Press, Cambridge. 8. Li, B-R., Introduction to operator algebras, World Sci., Singapore, 1992. 9. Lundstrom, P., Oinert, J., (2012). Skew category algebras associated with partially defined dynamical systems, Internat. J. Math. 23, no. 4, pp. 16 10. Mackey, G. W., (1989). Unitary Group Representations in Physics, Probability and Number Theory, xxviii+402pp. Second edition. Advanced book classics. Addison-Wesley Publish- ing Company, Advanced Book Program, Redwood city. 11. Ostrovskyi, V., Samoilenko, Yu., (1999). Introducion to the Theory of Representations of Finitely Presented *-Algebras, I, iv+261 pp. Representations by bounded operators. Re- views in Mathematics and Mathematical Physics, 11, pt.1. Harwood Academic Publishers, Amsterdam. 22 Alex Behakanira Tumwesigye, Johan Richter, Sergei Silvestrov 12. Richter J., Silvestrov, S. D., Tumwesigye, B. A., (2016). Commutants in crossed product algebras for piece-wise constant functions In: Silvestrov S., Ranci´c M., (eds.), Engineering Mathematics II, Springer Proceedings in Mathematics and Statistics, vol 179. Springer, Cham. 13. Richter, J., Silvestrov S. D. and Ssembatya, V. A. Tumwesigye, A. B, (2016). Crossed prod- uct algebras for piece-wise constant functions, Silvestrov S., Ranci´c M., (eds.), Engineering Mathematics II, Springer Proceedings in Mathematics and Statistics, Springer, pp. 18. 14. Richter J. (2016). Centralizers and Pseudo-Degree Functions. In: Silvestrov S., Ranci´c M. (eds) Engineering Mathematics II. Springer Proceedings in Mathematics and Statistics, vol 179. Springer, Cham 15. Richter, J., Silvestrov, S. (2014). Centralizers in Ore extensions over polynomial rings, In- ternational Electronic Journal of Algebra, 15, 196207. 16. Rowen, L., H., (1988). Ring theory, Vol. I. Pure and Applied Mathematics, 127. Academic Press, Inc., Boston, MA. 17. Svensson, C., Silvestrov, S. D., de Jeu, M., (2009). Dynamical systems associated with crossed products, Acta Appl. Math. 108 no. 3, 547-559. 18. Svensson, C., Silvestrov ,S. D. and De Jeu, M., (2007). Dynamical Systems and Commutants in Crossed Products, Internat. J. Math. 18 455 -- 471 2007. 19. Tomiyama, J., (1987). Invitation to C∗−algebras and topological dynamics, World Sci., 20. 21. 22. Singapore, New Jersey, Hong Kong. Oinert, J., Richter, J., and Silvestrov, S. D. (2013). Maximal commutative subrings and simplicity in ore extensions, Journal of Algebra and Its Applications, 12, 1250192. Oinert, J., Silvestrov, S., (2009). Commutativity and ideals in Pre-crystalline graded rings. Acta Appl. Math. 108, no. 3, 603-615. Oinert J., and Silvestrov S. D., (2008). Commutativity and ideals in algebraic crossed prod- ucts, J. Gen. Lie Theory Appl. 2, no.4, 287-302.
1708.08873	2	1708	2019-05-03T00:48:37	Testing isomorphism of graded algebras	[ "math.RA" ]	We present a new algorithm to decide isomorphism between finite graded algebras. For a broad class of nilpotent Lie algebras, we demonstrate that it runs in time polynomial in the order of the input algebras. We introduce heuristics that often dramatically improve the performance of the algorithm and report on an implementation in Magma.	math.RA	math	TESTING ISOMORPHISM OF GRADED ALGEBRAS PETER A. BROOKSBANK, E. A. O'BRIEN, AND JAMES B. WILSON Abstract. We present a new algorithm to decide isomorphism between finite graded algebras. For a broad class of nilpotent Lie algebras, we demonstrate that it runs in time polynomial in the order of the input algebras. We introduce heuristics that often dramatically improve the performance of the algorithm and report on an implementation in Magma. ] . A R h t a m [ 2 v 3 7 8 8 0 . 8 0 7 1 : v i X r a 1. Introduction It is possible to decide if algebraic objects A and B of order n are isomorphic by fixing a generating sequence a1, . . . , ad for A and searching through all sequences b1, . . . , bd in B until we find an identification ai 7→ bi that extends to an isomorphism A → B. This process takes nd steps: for groups and algebras, where d can be as large as log n, the resulting complexity is not polynomial in the orders of the input objects. Despite significant progress over the years on various isomorphism problems, asymptotic improvements over "brute force" for substantial classes of objects are rare. We introduce a general strategy for testing isomorphism of finite graded algebras. It is particularly effective for nilpotent matrix Lie algebras, and we describe a class of such algebras for which our isomorphism test runs in time polynomial in the order of the input algebra. We have also implemented a version in Magma [3]. While graded algebras are natural structures in their own right, they also arise from the study of other algebraic structures. For example, given a ring R, one can compute its Jacobson radical, J, and consider the graded algebra gr R = R/J ⊕ L∞ i=1 J i/J i+1. Similarly, the intersection Op(G) of the Sylow p-subgroups of a finite group G for a prime p dividing G is normal and nilpotent, so gr G = Z[G/Op(G)]⊕ L∞ i=1 ηi/ηi+1 is a graded Lie Fp-algebra, where η1 := Op(G) and ηi+1 = [ηi, η1]ηp i . Isomorphism tests for associated graded structures work with individual graded components, often exploiting the power of linear algebra. Existing uses of graded algebras within isomorphism testing proceed sequentially through the grading; see [12], for example. Starting with the first graded compo- nent, one considers all possible isomorphisms between corresponding components, and uses the graded product to decide which of them induces an isomorphism between subsequent components. While this iterative approach usually offers im- provements over brute force, a single large homogeneous component may create a bottleneck. Our approach is not constrained by the need to process the components sequentially. Instead, it identifies sections of the two graded algebras between which This work was supported in part by the Marsden Fund of New Zealand via grant UOA 1626, by NSF grants DMS-1620454 and DMS-1620362, and by the Simons Foundation #281435. We thank the referee for helpful comments. 1 2 PETER A. BROOKSBANK, E. A. O'BRIEN, AND JAMES B. WILSON the list of possible maps is small and can be computed quickly. It then determines which of these maps between sections lift to isomorphisms of the algebras. To state our main result, we require a few preliminaries. Let K be a finite field. A K-algebra is a K-module, A, equipped with a (possibly nonassociative) K-bilinear product ◦ : A × A ֌ A. If, as a K-module, ∞ A = Ms=0 As, where As ◦ At 6 As+t, then A is N-graded. We assume that A is generated in degree 1 in the sense that, for all s > 0, As =Ps−1 j=1 Aj ◦ As−j. An isomorphism between graded algebras that maps each graded component of one algebra to the corresponding component of the other is a graded isomorphism. For each s > 1, one can restrict the product on A to produce a bilinear map A1 × As ֌ As+1. The ring of adjoints of this bilinear map, denoted Ms, is the largest ring R faithfully represented on A1 ⊕ As such that the bilinear map A1 × As ֌ As+1 factors through the tensor product space A1 ⊗R As. Defined explicitly as operators in (2.7), Ms has group of units M× s . Our main result, proved in Section 4, is the following. Theorem 1.1. There is a deterministic algorithm that, given two finite graded K- algebras A and B generated in degree 1, decides whether or not there is a graded isomorphism A → B. The algorithm has complexity O(min s {M× s · Aut(As+1)} · (dim A)2ω · log2 K), where s runs over the grading and 2 6 ω < 3 is the exponent of matrix multiplication over K. By comparison, the sequential approach has complexity O(K(dim A)2 ). While the estimate in Theorem 1.1 is not asymptotically better in all cases, the flexibility to choose which component to process first may lead to dramatic improvements. As one illustrative example, define Fq(d1, . . . , dℓ) to be the smallest class of graded Lie algebras containing the Lie subalgebras L∗ 6 gld1+...+dℓ(q) satisfying [L∗, L∗] = 6 L∗ 6 where 0m,n denotes the zero (m × n)-matrix, and 0n = 0n,n. (A class of algebras is a collection closed under isomorphism; in particular, no specific representation is assumed as part of the input.) An analysis of the algorithm in Theorem 1.1 applied to F (d1, . . . , dℓ) establishes the following. Theorem 1.2. Isomorphism testing in Fq(d1, . . . , dℓ) is in time O(Lm2/ε), where didi+s) . The number of isomorphism classes in F (d1, . . . , dℓ) is qO((d1+···+dℓ)2). ε = X16i<j6ℓ 2(ℓ−s Xi=1 m = min s61+ d didj, 0d1     0d1,d2 . . . ∗ . . . . . . ··· . . . . . . 0dℓ−1 ∗ ... ∗ 0dℓ 0dℓ−1,dℓ     ∗ . . . ··· . . . . . . ∗ ... ∗0dℓ 0d1     ,     TESTING ISOMORPHISM OF GRADED ALGEBRAS 3 The complexity in Theorem 1.2 is polynomial in the size of the algebras when both m2, ε ∈ O((d1 + . . . + dℓ)2), a condition that holds, for instance, when di is bounded, and in many other cases. One of the motivations of this work is to develop practical tools to decide iso- morphism within families of finite groups and algebras. From this viewpoint, we are concerned with developing algorithms that perform well as a function of the length of standard encodings of the input algebras, such as by generating sets, or as bases with structures constants. Despite the significant improvements offered by the strategy underlying our main results, we have encountered situations where the necessary exhaustive search is still intractably large. To address this concern, in Section 7 we introduce heuristics that use local in- variants to deduce global restrictions on the possible automorphisms arising from A1 × As ֌ As+1, thereby reducing substantially the ensuing exhaustive search. More precisely, we design a labeling of the points and lines of the projective space on As+1 that is invariant under isomorphism. The labeling offers sufficient variabil- ity that the resulting constraints often reduce intractable searches to the practical realm. We also revisit another "local-to-global" process called fingerprinting that was introduced in [13]. In Section 8 we report on our implementation of the al- gorithms in Magma, and show for some families of examples that our techniques have significant practical impact. We prove a more general version of Theorem 1.1 in which algebras are not nec- essarily generated in degree 1, and may be graded using an arbitrary monoid. This general treatment allows our algorithms to take as input more refined gradings on the given algebras that often decompose them into smaller pieces. The develop- ment of such refined gradings is an emerging area of study that may lead to faster isomorphism tests; see [16], for example. 2. Preliminaries Throughout the paper K denotes a finite field, and all K-vector spaces are finite- dimensional. The K-dual of a K-vector space V is denoted by V †. If f : U → V is a linear map, then f † : V † → U † denotes the dual map. 2.1. Graded algebras. For convenience in our exposition we assume that an al- gebra A is specified by a basis {a1, . . . , ad} together with structure constants [αk ij] defined by ai · aj =Xk αk ijak. Although the structure constant model of an algebra is both less compact and less efficient than alternatives -- such as matrix algebras specified by generating sets, multivariate polynomials, and more general quotients of free algebras -- it provides a convenient, uniform starting point. Let M be a commutative monoid with pre-order ≺, where a ≺ b if there exists c ∈ M such that a + c = b. We assume that, relative to ≺, every nonempty subset has a minimal element. We also assume that 0 is a minimal element of M so that M is conical. This ensures that we can perform (Noetherian) induction on the indices in M : if S ⊂ M has the property that for every s ∈ S there exists t ∈ S \ {s} with t ≺ s, then S = ∅. The monoids M = Nc satisfy the necessary conditions. 4 PETER A. BROOKSBANK, E. A. O'BRIEN, AND JAMES B. WILSON An algebra A∗ is M -graded if A∗ =Ls∈M As and for all s, t ∈ M , As◦At ⊂ As+t. We assume, for each s, that As is the K-linear span of As ∩ {a1, . . . , ad}. We say that A is generated in degrees T if (2.1) (∀s) s ∈ M \ T =⇒ Xs = s1 + s2, s1, s2 /∈ {0, s} As1 ◦ As2 = As. As M is conical, if A∗ is generated in degrees T , then 0 ∈ T . The following observation is a direct consequence of the induction principle. Lemma 2.2. If A∗ is an M -graded algebra generated in degrees T then, for each s ∈ M , As consists of linear combinations of products of elements in St∈T At. If M = N, then the generating degrees are 0 and 1. Here, A0 contributes no shift in the grading; it is customary to ignore 0 and regard A∗ as generated in degree 1. The main results in the introduction were formulated for this special case. As an infinite monoid can grade a finite algebra, repetitions As = At for s 6= t are common. To avoid redundancy, one can truncate all rays in the conical monoid M in the first place where all are stable. This is always possible but may produce a less familiar monoid: for each c ∈ N, associate to M the ℓ-truncated cyclic monoid Nℓ = {0, . . . , ℓ}, where x ⊞ y := x + y if x + y < ℓ, and x ⊞ y := ℓ otherwise. Proposition 2.3. Let M = hm1, . . . , mdi, let A∗ be an M -graded algebra, and let ℓ be the smallest positive integer satisfying ∀i ∈ {1, . . . , d}, ∀u ∈ M \ {0} Aℓmi = Aℓmi+u. Then A∗ is naturally (Nℓ)d -- graded where A(r1,...,rd) = Ar1m1+···+rdmd . In particu- lar, we may assume for fixed d that M ∈ O((dim A∗)d). 2.2. Bimaps. A bi-additive map (or just bimap) is a tuple U∗ = hU2, U1, U0,◦i where the Ui are abelian groups and ◦ : U2 × U1 ֌ U0 is a function satisfying the following two-sided distributive law: (u2 + u′ 2) ◦ u1 = u2 ◦ u1 + u′ 2 ◦ u1 u2 ◦ (u1 + u′ 1) = u2 ◦ u1 + u2 ◦ u′ 1. We always work with K-bimaps: the Ui are K-modules and ◦ is K-bilinear. The dimension of U∗ is the sum of the dimensions of its component spaces, namely dim U∗ := dim U2 + dim U1 + dim U0. A homotopism f∗ : U∗ → V∗ is a tuple (fi : Ui → Vi i ∈ {0, 1, 2}) satisfying (2.4) (u2 ◦ u1)f0 = u2f2 ◦ u1f1. (∀u2)(∀u1) Denote by Hom(U∗, V∗) the set of all homotopisms f∗ : U∗ → V∗. The class of bimaps with homotopisms forms the homotopism category. There are various natural morphisms on classes of bimaps, such as adjoint-morphisms [22], so we name the categories after the morphisms rather than the objects. We are interested primarily in isotopisms, namely homotopisms whose constituent maps are all isomorphisms. The autotopism group of a bimap U∗ is (2.5) Aut(U∗) = Hom(U∗, U∗) ∩ Aut(Ui). 2 Yi=0 TESTING ISOMORPHISM OF GRADED ALGEBRAS 5 2.3. Shuffling bimaps. Since a bimap U∗ = hU2, U1, U0,◦i consists of multiple components, it is tedious, both in proofs and in algorithms, to specify individual components. Accordingly, we often "shuffle" indices in our bimaps. To do this, however, we must ensure that autotopisms are unaffected by the process, so we now define precisely what we mean by shuffling. Related categorical subtleties are considered in [22]; observe that our treatment differs from that of [14]. Given a bimap U∗ = hU2, U1, U0,◦i and a permutation σ ∈ Sym({2, 1, 0}), we define a new bimap U σ ∗ . It suffices to consider just transpositions, as these generate all possible permutations, but the definition depends on the particular transposi- tion. First consider σ = (2, 1), the transposition swapping 2 and 1. Set U σ 2 := U1, 1 := U2, U σ U σ 2 and u1 ∈ U σ 1 , set 0 as follows: given u2 ∈ U σ 0 := U0, and define ◦σ : U σ 2 × U σ 1 ֌ U σ 2 , U σ We define U σ 2 := f1, f σ f σ ∗ : U σ ∗ → V σ also called the transpose of a bimap. ∗ to be hU σ ∗ where f σ 1 , U σ u2 ◦σ u1 := u1 ◦ u2. 0 ,◦σi. A homotopism f∗ : U∗ → V∗ is sent to In the literature this is 1 := f2, f σ 0 := f0. For σ = (1, 0) and σ = (2, 0) we need a slightly more elaborate construction. We 1 := U † 0 , 1 . We switch to Greek letters when working with elements in a dual vector 1 (so consider just (1, 0) since (2, 0) works in the same way. Set U σ 0 := U † U σ space. Define ◦σ : U σ ν1 : U1σ → K), define (u2 ◦σ ν1) = ν0 ∈ U σ 0 as follows: given u2 ∈ U σ 0 = Hom(U1, K) as follows: 2 and ν1 ∈ U σ 2 := U2, U σ 2 × U σ 1 ֌ U σ ν0 : u0σ 7→ (u2 ◦ u0σ)ν1. A complication arises with homotopisms, since interchanging f1 and f0 is no longer meaningful. Instead, we put f σ 0 and f σ ∗ is not a homotopism but satisfies the following condition: 1 . In particular, f σ 2 := f2, f σ 1 := f † 0 := f † (∀u2)(∀ν1) (u2 ∈ U σ 2 ) ∧ (ν1 ∈ V σ 1 ) ⇒ u2f σ 2 ◦σ ν1 = (u2 ◦σ ν1f σ 1 )f σ 0 . This illustrates the inherent delicacy of the otherwise trivial idea of re-indexing ∗ = variables. (f2, (f † In the special case where f∗ is an isotopism we observe that f σ 1 )−1) is an isotopism U σ 0 )−1, (f † ∗ → V σ ∗ . 2.4. Weakly Hermitian bimaps. A bimap is weakly Hermitian if it is isotopic to its transpose, namely to its σ = (2, 1) shuffle. The familiar symmetric and alternat- ing bimaps are examples of weakly Hermitian bimaps, but the notion is substantially more general. Associated to weakly Hermitian bimaps are group invariants that respect their symmetry. If we fix an isotopism g∗ : U∗ → U † ∗ , then ΨIsom(U∗) = {f∗ ∈ Aut(U∗) : f∗g∗ = g∗f † ∗} is the group of pseudo-isometries of U∗ and does not depend on the choice of g∗. 2.5. Algebras that operate on bimaps. Consider the following algebras deter- mined by a bimap U∗. (2.6) (2.7) (2.8) (2.9) L(U∗) = {(f †, h†) ∈ End(U2) × End(U0) : (f †u2) ◦ u1 = h†(u2 ◦ u1)}, M(U∗) = {(f, g†) ∈ End(U2) × End(U1) : (u2f ) ◦ u1 = u2 ◦ (g†u1)}, R(U∗) = {(g, h) ∈ End(U1) × End(U0) : u2 ◦ (u1g) = (u2 ◦ u1)h}, and T (U∗) = L(U∗) ⊕ M(U∗) ⊕ R(U∗). 6 PETER A. BROOKSBANK, E. A. O'BRIEN, AND JAMES B. WILSON Each Ui is a natural module under each of these rings, and is thus a T (U∗)-module. Although the action is non-unital -- for example, the representation of L(U∗) on U1 is trivial -- it is more convenient to think of each of U2, U1, U0 as a module over a common ring than continually to clarify that the action on U1 is by T (U∗)/L(U∗), and so on. We also use the following algebra: End(Ui) : (u2f2) ◦ u1 = u2 ◦ (u1f1) = (u2 ◦ u1)f0) . C(U∗) =(f∗ ∈ (2.10) We need one final notion. Fix bimaps U∗ and V∗. As defined in [22], an adjoint- morphism (f, g) : U∗ → V∗ is a pair of maps f : U2 → V2 and g : V1 → U1 satisfying 2 Yi=0 (∀u2)(∀v1) (u2 ∈ U2) ∧ (v1 ∈ V1) ⇒ (u2f ) ◦ v1 = u2 ◦ (v1g). The set of all such pairs (f, g) is denoted Adj(U∗, V∗). This defines another category on bilinear maps distinct from those already discussed: it is an abelian category and plays a role similar to modules of rings [22, Theorem 2.27]. In particular, observe that Adj(U∗, U∗) is simply the ring M(U∗). As adjoint-bimap categories are not equivalent to module categories [22, Theorem 2.10], however, we must adapt some established results in module theory to suit our purpose. 3. Testing isotopism of bimaps In this section we consider the problem of deciding if two bimaps are equivalent under isotopisms. This is an essential step in our isomorphism test for graded algebras, but it is also a problem of independent interest. We require an efficient solution to the following problem. IsotopismCoset Given: K-bilinear maps U∗ and V∗. Return: the coset Iso(U∗, V∗) of isotopisms U∗ → V∗. Here, we present a basic algorithm to solve this problem; in Section 7 we introduce heuristics to speed up the construction. Note, if f∗ : U∗ → V∗ is an arbitrary isotopism, then Iso(U∗, V∗) = Aut(U∗)f∗, so the output can be encoded compactly using generators for Aut(U∗). If U∗ = V∗, then the output is simply Aut(U∗). 3.1. Principal autotopisms. The difficulty of IsotopismCoset stems from hav- ing to find solutions to quadratic polynomials in multiple variables: namely, we solve for (f2, f1, f0) where the parameters f2 and f1 occur in a product u2f2 ◦ u1f1. Quadratic varieties are as complex as arbitrary varieties, but fixing any one of the fi renders the problem substantially more tractable. Thus, we consider first the following restricted version of the autotopism group problem; in Section 3.2 we handle the coset version. PrincipalAutotopismGroup Given: a K-bilinear map U∗ and i ∈ {2, 1, 0}. Return: generators for Aut(U∗)(i) := {f∗ ∈ Aut(U∗) : fi = 1Ui}. The following observation leads to an efficient solution to this problem. TESTING ISOMORPHISM OF GRADED ALGEBRAS 7 Proposition 3.1. For a bimap U∗, the following hold: (i) Aut(U∗)(2) ∼= R(U∗)×; (ii) Aut(U∗)(1) ∼= L(U∗)×; and (iii) Aut(U∗)(0) ∼= M(U∗)×. Proof. Following our discussion in Section 2.3, we can assume i = 0 after a possible shuffling of the variables. (We stress once more that reindexing requires some re- adjustment of the resulting isotopisms.) If f∗ ∈ Aut(U∗) and f0 = 1, then 1 ))f0 = u2 ◦ (u1f −1 1 ), 1 , 1U0) ∈ Aut(U∗). (u2f2) ◦ u1 = (u2f2) ◦ (u1f −1 1 f1) = (u2 ◦ (u1f −1 1 ) ∈ M(U∗). If (f2, f1) ∈ M(U∗) then (f2, f −1 so (f2, f −1 (cid:3) Algorithm 1 Principal Autotopism Group Input: a K-bilinear map U∗ and i ∈ {2, 1, 0}. Output: generators for Aut(U∗)(i). 1: Choose a permutation σ on {2, 1, 0} with iσ = 0. 2: Solve a system of linear equations to find a basis for M(U σ ∗ ). 3: Use [6, Theorem 2.3] to compute generators X for the group of units of M(U σ ∗ ). 4: Set G = h(f, g−1, 1)σ : (f, g) ∈ Xi 6Q2 5: return G. i=0 Aut(Ui). Proposition 3.2. Algorithm 1 solves PrincipalAutotopismGroup. As a de- terministic algorithm it runs in time O((dim U∗)2ω log2 K + char K); a Las Vegas variant runs in time O((dim U∗)2ω log2 K). Proof. Since the correctness of the algorithm is clear from Proposition 3.1 and the mechanics of shuffling variables, we focus on the complexity. Line 2 involves solv- ing a system of (dim U2)(dim U1)(dim U0) linear equations in (dim U2)2 + (dim U1)2 variables, which can be done in time O((dim U∗)2ω log2 K). Line 3 invokes the algorithm of [6, Theorem 2.3], which depends on the ability to factor polynomials over K. The algorithm runs in Las Vegas polynomial-time O((dim A)2ω log2 K) if we use Las Vegas polynomial factorization routines such as that of [9]. A determin- istic algorithm is known when the ground field of K can be listed: in this case Line 3 runs in time O((dim A)2ω log2 K + char K). The remaining steps of Algorithm 1 have negligible influence on the timing, so the result follows. (cid:3) 3.2. Extending to isotopisms. Our next objective is to solve a single instance of isotopism. We focus first on principal isotopisms, and assume that i = 0 by shuffling coordinates. PrincipalIsotopism Given: K-bilinear maps U∗ and V∗ and a map fi : Ui → Vi for fixed Return: an isotopism f∗ : U∗ → V∗ extending fi. i ∈ {2, 1, 0}. Just as the construction of the principal autotopism group is a problem in rings, the construction of a principal isotopism resembles a problem in modules. As we indicated in Section 2, however, it is not precisely a module problem that we solve. 8 PETER A. BROOKSBANK, E. A. O'BRIEN, AND JAMES B. WILSON Definition 3.3. An orthogonal decomposition of a bimap U∗ is a pair of direct Each Uij is an orthogonal factor. decompositions U2 = Lj U2j and U1 = Lk U1k such that U2j ◦ U1k = 0 if j 6= k. For example, if U∗ = hK 2, K 3, K,◦i where u2 ◦ u1 = u2(cid:20)1 1 0(cid:21) u† 0 0 then U2 = K(1, 0) ⊕ K(0, 1) and U1 = K(1, 0, 0) ⊕ K(0, 1, 0) ⊕ K(0, 0, 1) is an orthogonal decomposition. More generally, in terms of structure constants this implies that A(k) = [ak ij ] is block diagonal, the blocks coinciding with the (U2j, U1j) pairs; see [22, Section 2.4] for details. 0 1 Our plan is to imitate the algorithm of [4], which builds a module isomorphism one direct summand at a time. Both that algorithm and our adaptation rely on the following construction. If X ⊆ EndK(V ), then let X be the semigroup generated by X and let KhXi be the K-linear span of X. (Contrary to the usual notion of enveloping algebra, KhXi need not be unital: this occurs if, and only if, the identity can be written as a linear combination of elements of X.) Theorem 3.4 ([4, Corollary 2.5]). There is a polynomial-time algorithm that, given X ⊆ EndK(V ), decides if KhXi is nilpotent and, if not, returns a product of elements in X that is not nilpotent. The algorithm in [4] uses non-nilpotent elements to decompose the modules into direct summands. Instead of module isomorphisms we construct principal isotopisms; instead of direct summands we use orthogonal factors. We capture the key recursive step with the following technical definition. Definition 3.5. Fix f0 : U0 → V0. A partial f0-isotopism of bimaps U∗ and V∗ is an isotopism g∗ = (g2, g1, g0) defined on the restriction to some orthogonal factors of U∗ and V∗, and such that g0 = f0. A partial f0-isotopism is maximal if it is not a restriction to proper subspaces of another partial f0-isotopism. The idea is to build a (possibly nilpotent) ring from two sets of adjoint-morphisms. If this ring contains an invertible element, then we find the desired principal iso- topism. Following [4], we propose Algorithm 2 to construct a maximal partial isotopism. ∗ )); Y ← Basis(Adj(V f0 Algorithm 2 Partial Principal Isotopism Input: bimaps U∗ and V∗, and an isomorphism f0 : U0 → V0. Output: a maximal partial f0-isotopism. 1: X ← Basis(Adj(U∗, V f0 ∗ , U∗)). 2: A ← Khxy : x ∈ X , y ∈ Yi ⊂ End(U2) × End(U1)op. 3: For i = 1, 2, U − 4: while A has z = xyw not nilpotent with x ∈ X and y ∈ Y do i = ker zn ⊕ im zn. fi ← fi ⊕ resim zn (x). 1 , f0 : U0 → V0). fi ← 0. Find n > 0 such that, for i = 1, 2, U − i ← ker zn; U + U − Restrict A to ker zn. 7: 8: return f∗ = (f2 : U + i ⊕ im zn; 2 , f1 : U + i ← U + 2 → V + i ← Ui; U + 1 → V + i ← 0; 5: 6: TESTING ISOMORPHISM OF GRADED ALGEBRAS 9 Proposition 3.6. Algorithm 2 is deterministic and constructs a maximal f0-isotopism in polynomial time O((dim U∗)2ω log2 K). Proof. The objective of the algorithm is to find an invertible element of Adj(U∗, V f0 ∗ ). To do this, we first create a (possibly non-unital) algebra A in Line 2 by composing the sets of homomorphisms created in Line 1. Observe that composition in the second variable is in the op-ring End(U1)op. First, note that A ⊂ M(U∗). Secondly, if f0 extends to an isotopism f∗ : U∗ → V∗ then (f2, f −1 ∗ , U∗); in particular, A contains units. However, finding a unit of A does not guarantee that we can extract an invertible element of Adj(U∗, V f0 1 ) ∈ Adj(U∗, V f0 2 ) ∈ Adj(V f0 ∗ ) and (f1, f −1 ∗ ). We claim that the loop starting in Line 4 maintains the following invariants: for is an orthogonal decomposition of Ui; and (f2, f1, f0) is i ∈ {2, 1}, Ui = U + a partial f0-isotopism. Clearly, this is true at the start. i ⊕ U − i By its construction in Line 4, clearly z ∈ M(U∗), and so zn ∈ M(U∗). Note, for every b ∈ M(U∗), the decomposition Ui = ker b ⊕ im b is orthogonal. Therefore we maintain throughout an orthogonal decomposition Ui = U + i . Furthermore, by Fitting's lemma, zn is invertible on im zn. The guard of the loop in Line 4 is a call to Theorem 3.4, which provides x ∈ Adj(U∗, V f0 ∗ ) such that z = xyw. As zn is invertible on im zn, and zn = x··· , it follows that x is injective on im zn. Since all spaces are finite, this injection is a bijection. Therefore the extension of fi by the restriction of x to the image of zn remains a partial f0-isotopism, as required. i ⊕ U − Finally, the loop continues while A contains a non-nilpotent element. Thus, the partial f0-isotopism is maximal and the output is correct. The major work is solving the system of linear equations in Line 1; this results (cid:3) in the complexity stated in the theorem. Remark. In many settings, invertible elements of Adj(U∗, V f0 ∗ ) may be found by random search with high probability, but there are examples that require an ex- ponential number of samples to return an invertible element. Nevertheless, once Adj(U∗, V f0 ∗ ) is constructed, it is sensible to test a small number of random ele- ments. The following is now immediate. Theorem 3.7. There is a deterministic, polynomial-time algorithm to solve Prin- cipalIsotopism. We are finally ready to present Algorithm 3, the main result of this section. Viewing the group G in Line 4 of this algorithm as a parameter, the following is an immediate consequence of the results and algorithms of this section. Theorem 3.8. Algorithm 3 solves IsotopismCoset and runs deterministically in time O(G maxi{(dim Ui)2ω} log2 K), where G is the group in Line 4. 4. Testing isomorphism of graded algebras Our algorithm to decide isomorphism between graded algebras proceeds under the assumption that an isomorphism exists. If this does not occur then the test is aborted. A standard mechanism to do this is to raise an exception. This means that all further steps are aborted and the algorithm backtracks to the nearest place that can handle the exception. Recall from Proposition 2.3 that we assume an algebra A∗ is graded by a monoid M that has size polynomial in dim A∗. 10 PETER A. BROOKSBANK, E. A. O'BRIEN, AND JAMES B. WILSON Algorithm 3 Isotopism coset Input: bimaps U∗ and V∗. Output: the coset Iso(U∗, V∗) of isotopisms U∗ → V∗. 1: Choose a permutation σ on {2, 1, 0} with dim U0σ minimized. 2: if dim U0σ 6= dim V0σ then return ∅. 3: Choose any isomorphism f0σ : U0σ → V0σ. 4: Choose G such that Aut(U∗)U0σ 6 G 6 Aut(U0σ). 5: I ← ∅. 6: for all g ∈ G do /* Algorithms 1 & 2 / Find the coset C of isotopisms h∗ : U∗ → V∗ with h0σ = gf0σ. I ← I ∪ C. 7: 8: 9: 10: return I / see Section 7 / 4.1. Extending isotopisms to graded isomorphisms. The isomorphism algo- rithm proceeds by attempting to extend isotopisms between the bimaps obtained by restricting the given products to certain fixed homogeneous components. We there- fore begin by considering the necessary extension problem; our solution is summa- rized in Algorithm 4. As A∗ and B∗ are generated in degrees T , we assume that we have a homotopism whose restriction to T , namely fT = {(ft : At → Bt) : t ∈ T}, is defined. Our task is to extend fT to a graded algebra homomorphism f∗ : A∗ → B∗. For s ∈ M \ T , define A⊗s := Ms = s1 + s2, si /∈ {0, s} As1 ⊗ As2 ⊂ A∗ ⊗ A∗. Given s ∈ M , we can construct {(s1, s2) ∈ M × M : s = s1 + s2} in at most M2 6 (dim A∗)2 steps. Define f⊗s := Ms = s1 + s2, si /∈ {0, s} fs1 ⊗ fs2 ∈ Hom(A⊗s, B⊗s), where fs1 ⊗ fs2 is defined component-wise on As1 ⊗ As2 . For s ∈ M , let ≺s = {u ∈ M : u ≺ s}. For R ⊂ M , write R ≺ s if r ≺ s for every r ∈ R. If A∗ is generated in degrees T , then setting s =Pt∈T t implies that T ≺ s. O(M(dimA∗)2ω log2 K). Proposition 4.1. Algorithm 4 is correct and runs deterministically in time Proof. Consider first the correctness of the algorithm. Observe that f∗ : A∗ → B∗ is a graded homomorphism if, and only if, for all s and t, (fs, ft, fs+t) is a homotopism from the bimap As × At ֌ As+t to Bs × Bt ֌ Bs+t. We claim that the loop starting on Line 2 has the following invariants: (i) R is an interval closed set of indices; (ii) for all r ∈ R, the map fr : Ar → Br is defined; and (iii) for all r, r′ ∈ R, if r + r′ ∈ R, then (fr, fr′, fr+r′) is an isotopism from The loop terminates when R = M , so f∗ : A∗ → B∗ is defined and is consequently a graded algebra isomorphism. Ar × Ar′ ֌ Ar+r′ to Br × Br′ ֌ Br+r′. TESTING ISOMORPHISM OF GRADED ALGEBRAS 11 Algorithm 4 Extending Homotopisms Input: finite M -graded K-algebras A∗ and B∗ generated in degrees T ⊂ M , and fT = {(ft : At → Bt) : t ∈ T} such that for every t, r ∈ T with t + r ∈ T , the triple (ft, fr, ft+r) is a homotopism from Ar × At ֌ Ar+t to Br × Bt ֌ Br+t. Output: an algebra homomorphism f∗ : A∗ → B∗ extending fT , or raise an excep- tion if fT does not extend to an algebra homomorphism. 1: R ← T . 2: while R 6= M do 3: 4: 5: 6: 7: 8: 9: Choose s ∈ M \ R such that {r ∈ M : r ≺ s, r 6= s} ⊂ R. Induce πs,A∗ : A⊗s → As and likewise πs,B∗ : B⊗s → Bs. Compute ker πs,A∗ and ker πs,B∗ . if (ker πs,A∗ )f⊗s 6 ker πs,B∗ then s,A∗ · f⊗s · πs,B∗ ∈ Hom(As, Bs). fs ← π−1 R ← R ∪ {s}. raise exception no extension exists at s. else 10: 11: return f∗. Consider any s selected in Line 3. Since A∗ is generated in degrees T , and T ⊂ R, it follows that πs,A∗ is surjective and As ∼= A⊗s/ ker πs,A∗ . The same holds for B∗. By our choice of s, for every s1, s2 /∈ {0, s} with s = s1 + s2, fs1 and fs2 are defined, so f⊗s is defined. If (ker πs,A∗ )f⊗s 6 ker πs,B∗ , then we may induce fs : As → Bs on the generators of As as follows: for as1 ∈ As1 and as2 ∈ As2 , (as1 ◦ as2 )fs := (as1 ⊗ as2 )f⊗s ≡ as1fs1 ◦ as2 fs2 mod ker πs,B∗ . Conversely, if f≺s extends to a graded isomorphism f∗ : A∗ → B∗, then fs is defined as above. Thus, if (ker πs,A∗ )f⊗s 66 ker πs,B∗ we conclude that {ft : t ∈ T} does not extend, and raise an exception to abort all subsequent steps. Next, we analyze the timing. The loop executes at most M ∈ (dim A∗)O(1) iterations, and the timing in each one is dominated by the computation of ker πs,A∗ and ker πs,B∗ and the subsequent membership test in the latter. Each requires solving systems of linear equations in (dim As)2 variables. In total this takes (cid:3) O(Ps(dim As)2ω log2 K) steps, as stated. 4.2. The isomorphism test. For an M -graded algebra A∗ and S ⊂ M , define AS =Ls∈S As; for a graded homomorphism f∗, define fS =Ls∈S fs. Algorithm 5 is our isomorphism test for graded algebras. The mechanism for selecting S in Line 1 is discussed in Section 4.3. Proposition 4.2. Algorithm 5 is correct. Proof. If f∗ : A∗ → B∗ is a graded isomorphism, then the restriction  Mt∈T ft, Ms∈S fs, Ms∈S,t∈T fs+t  12 PETER A. BROOKSBANK, E. A. O'BRIEN, AND JAMES B. WILSON Algorithm 5 Graded Isomorphism Coset Input: finite M -graded algebras A∗ and B∗ generated in degrees T ⊂ M . Output: the coset of graded isomorphisms A∗ → B∗, or ∅ if A 6∼= B. 1: Choose ∅ 6= S ⊂ M . 2: Restrict multiplication in A∗ to obtain the bimap AT × AS ֌ AS+T . 3: Similarly, obtain the bimap BT × BS ֌ BS+T . 4: I ← Iso(AT × AS ֌ AS+T , BT × BS ֌ BS+T ), using Algorithm 3. 5: Γ ← {f∗ : A∗ → B∗ : f∗ extends some (fT , fS, fS+T ) ∈ I}, using Algorithm 4. 6: return Γ. / see Section 4.3 */ is an isotopism from AT ×AS ֌ AS+T to BT ×BS ֌ BS+T . Since A∗ is generated in degrees T , for each s ∈ M \ T , fs is determined by fT =Lt∈T ft. In particular, f∗ = 1 if, and only if, fT = 1. Hence, the mapping fs+t  f∗ 7→ Mt∈T ft,Ms∈S fs, Ms∈S,t∈T is injective. The algorithm constructs the inverse image of this injection, and so returns the coset of graded isomorphisms A∗ → B∗. (cid:3) Proof of Theorem 1.1. We analyze the complexity of Algorithm 5. If the algorithm discovers that A∗ 6∼= B∗, then it terminates. Hence, the case that dominates com- plexity is A∗ ∼= B∗. Using Algorithm 3, I = Iso(U∗, V∗) is constructed in time O(min{ Aut(AT ), Aut(AS), Aut(AS+T )}(dim A∗)2ω log2 K), and hence in time O( Aut(AS+T )(dim A∗)2ω log2 K). The remaining time to construct Iso(A∗, B∗) is O(I). Observe that I = Aut(U∗) 6 M(U∗)× · Aut(AS+T ). The complexity stated in Theorem 1.1 now follows by substituting T = {1} and S = {s}, where s is the largest positive integer such that As+1 6= 0. 4.3. Selecting optimal indices. We now discuss the issue left open in Line 1 of Algorithm 5: how to choose the subset, S, of optimal indices. Our aim is to predict the order of (cid:3) Aut (AT × AS ֌ AS+T ) without computing it. This allows us in Line 1 of Algorithm 5 to sample several subsets S to find a selection whose estimated work is either minimal, or below an acceptable threshold. If U∗ is the bimap AT × AS ֌ AS+T , then by definition there is an exact sequence (4.3) (4.4) from which we immediately obtain the bound 1 → M(U∗)× → Aut(U∗) → Aut(U0), Aut(U∗) 6 M(U∗)× · q(dim U0)2 . This bound suffices to prove our main theorems, but more precise bounds on Aut(U∗) can be obtained with additional work. We include a brief discussion here both because our implementation uses the better bounds, and also because future analyses of the complexity of our algorithm for specific families of inputs may require them. TESTING ISOMORPHISM OF GRADED ALGEBRAS 13 In [7, Theorem 3.2] a property of autotopisms is given which leads to a general bound on Aut(U∗). However, a better bound using the rings defined in Section 2.5 may be derived from the exact sequences in [23, Theorem 1.2]. Let Out(R) be the group of outer automorphisms of a ring R. If R is equipped with an involution a 7→ ¯a, then R# = {a ∈ R : a¯a = 1} denotes its group of unitary elements, and Out#(R) is the subgroup of Out(R) commuting with the involution. Proposition 4.5. For a bimap U∗ = (U2, U1, U0,◦) the following holds: 6 Aut(U∗) 6 T (U∗)× · Out(T (U∗)) · min{AutT (U∗)(Ui) : 0 6 i 6 2}. If U∗ is weakly Hermitian then T (U∗) and C(U∗) are rings with involutions, and T (U∗)× C(U∗)× T (U∗)# C(U∗)# 6 ΨIsom(U∗) 6 T (U∗)# · Out#(T (U∗)) · min{AutT (U∗)(Ui) : 0 6 i 6 2}. Each bound can be computed in time polynomial in P2 5. Proof of Theorem 1.2 i=0 dim Ui. The efficiency of our test for isomorphism between graded algebras A∗ and B∗ depends critically on two conditions: first, we can find a homogeneous component AS+T of moderate size; secondly, the order of Aut(AT × AS ֌ AS+T ) is manage- able. In this section, we consider a natural family of nilpotent Lie algebras whose basic parameters illustrate the performance of our isomorphism test. In doing so we prove Theorem 1.2. Every nilpotent matrix Lie algebra L∗ 6 gl(V ) has a nontrivial 0-eigenspace V1. Recursively, for i > 1, let Vi+1 6 V so that Vi+1/Vi is the 0-eigenspace of the representation of L∗ on V /Vi. The flag 0 < V1 < . . . < Vℓ = V is denoted F (L∗). Conversely, associated to each flag F of V is a unique maximal nilpotent Lie subalgebra P(F )∗ 6 gl(V ) such that F (P(F )∗) = F . We say L∗ is dense if [L∗, L∗] = [P(F (L∗))∗, P(F (L∗))∗]. 5.1. Comparing results. Before presenting the proof of Theorem 1.2, we pause to compare its complexity to that of other algorithms for algebra isomorphism, and to other well known computational problems. It is helpful to use L-notation which we define in terms of logarithms to base q = K: Ln[α, c] = q(c+o(1))(log n)α(log log n)δ(0,α) , where δ(0, α) =(cid:26) 0 α = 0 1 α > 0 . This function interpolates between polylogarithms Ln[0, c] ∈ O((log n)c), polyno- mials Ln[1, c] ∈ O(nc), quasi-polynomials Ln[2, c] ∈ O(nc ln n), and so forth. For context, the heuristic estimates (in base e) for the cost to factor an integer n are Ln[ 1 2 , 1] for the quadratic sieve, and Ln[ 1 Assuming the size of the field is constant, the complexity of graded algebra 3 , 3p64/9] for the number field sieve. isomorphism in Theorem 1.1 is Ln[2, 1 2 ]. 14 PETER A. BROOKSBANK, E. A. O'BRIEN, AND JAMES B. WILSON We compare this to the general isomorphism test for Lie algebras presented in [11]. That algorithm must select the first homogeneous component L1. In the worst case -- where L∗ = L1 ⊕ L2 and dim L1 = r dim L∗ for some constant r 6 1 -- it exhaustively searches Aut(L1), resulting in a complexity of Ln[2, 1]. For cases in which dim L1 = r√dim L∗, however, the complexity improves to Ln[1, c]. This shows how the Hilbert series Pi(dim Li)xi of the input influences the complexity of [11]; as we explain below, the Hilbert series exerts an influence over the com- plexity of our algorithm that is both more subtle and more emphatic. Another recent approach to isomorphism [7, 23] can be applied to graded Lie algebras of class 2. This exploits invariant algebras of bimaps and is particularly effective when there are large automorphism groups. This method also has com- plexity ranging from a worst case of Ln[2, 1] down to nearly optimal run times of Ln[0, 2ω] for inputs such as generalized Heisenberg Lie algebras. Now, let us consider the complexity of our current algorithm as it appears in Theorem 1.2. First, ε measures the "area" occupied by the dense algebra L∗. The presence of a few large blocks around the middle of F (L∗) makes ε comparatively small, and hence slows down the performance of our algorithm. Secondly, m is the dimension of the smallest Ls+1 we encounter, subject to the product L1×Ls ֌ Ls+1 being nondegenerate (the condition s 6 1 + d/2 in the formula for m ensures this). As illustrated in Figure 1, small values of m2/ε correspond to super block diagonal layers near the middle that are as thin as possible. The dimensions of the possible layers are determined by the Hilbert series of the input. Figure 1. Contrasting the work to construct automorphisms Observe that the work to construct automorphisms by lifting from a fixed shaded layer is roughly qa2 , where q is the size of the field and a is the area of the shaded region (the dimension of the corresponding homogeneous component). The shaded area on the left of Figure 1 is roughly 3 times that on the right. Thus, if w is the work needed to search for automorphisms using the shaded region on the left, then using that on the right decreases the work to w1/9. 5.2. Multiplication tables. We work with L∗ as matrices relative to a basis ex- hibiting F (L∗), so elements of L∗ are block upper-triangular matrices with 0's on the diagonal. Although the specific entries in each block determine the structure of the algebra, much can be deduced simply by studying the block structure. For x ∈ L∗, denote by xst the block of x in (block) row s and (block) column t. For TESTING ISOMORPHISM OF GRADED ALGEBRAS 15 x, y ∈ L∗, the following is the general formula for the product on L∗:   X16s<t6d xst, X16u<v6d yuv  = X16s<u<t6d xsuyut − ysuxut. (5.1) Each product xsuyut − ysuxut may be specified by structure constants depending solely on (s, t); we demonstrate this below for d = 6. (As L∗ is generated in degree 1, we display only the structure constants for each product L1 × Ln ֌ Ln+1 for n = 1, . . . , 5.) x12 . x12 x23 −a† 1 x34 x45 x56 . . . x23 a1 . −a† 2 . . x34 x45 x56 x13 . a2 . −a† 3 . . . a3 . −a† 4 . . . a4 . . . −b† 1 . . x24 b1 . . −b† 2 . x35 x46 x14 . b2 . . −b† 3 . . b3 . . . . . −c† 1 . x25 c1 . . . c† 2 x36 x15 . c2 . . . . . . . −d† 1 x26 d1 . . . . x16 . . . . . This example illustrates two competing tensions that determine our success. On one hand, we want to select s such that Aut(L1 × Ls ֌ L1+s) is small. This occurs when the corresponding product has the most constraining equations, which happens generically when s is small. On the other hand, our method to compute Aut(L1 × Ls ֌ L1+s) requires that we list Aut(L1+s), and the order of this group tends to decrease as s increases. This explains why the best choice of s is typically near the middle. 5.3. Estimating Aut(L1 × Ls ֌ L1+s). To prove Theorem 1.2 it suffices to consider the bound on Aut(L1 × Ls ֌ L1+s) arising from the exact sequence (4.3). In particular we show that if s 6 1 + d/2, then the choice U∗ = (L1, Ls, L1+s, [, ]) of homogeneous component in Theorem 1.1 leads to the complexity stated in The- orem 1.2. Put U2 := L1, U1 := Ls, and U0 := L1+s. It suffices to establish the bounds (5.2) M(U∗)× 6 qd2 6 L1/ε, Aut(U0) 6 qm2d2 6 Lm2/ε. We first make an observation. If X i As suggested in Section 4.3 we can derive more subtle bounds by applying Propo- sition 4.5, but the analysis offers little insight into the complexity of our algorithm. 0,◦i) (i = 1, 2) are two bimaps on a common domain X2×X1, but with possibly different codomains and products, then we can define a new bimap Y∗ = X 1 ∗ as follows. Put Y2 := X2, Y1 := X1, Y0 := X 1 ∗ ∩ X 2 0 , and define ◦ : Y2 × Y1 ֌ Y0 by ∗ = (X2, X1, X i 0 ⊕ X 2 y2 ◦ y1 := (y2 ◦1 y1) ⊕ (y2 ◦2 y1). ∗ ∩ X 2 Note that M(X 1 U∗ as Ti U i To compute M(U∗) we solve equations of the form F A = AG†, where A is a ∗, thereby making it simpler to compute the ring of adjoints. ∗ ), so selecting a basis for U0 decomposes ∗ ) = M(X 1 ∗ )∩M(X 2 matrix of structure constants of U∗ of the form . (5.3) A =   a1 . . . . −a† 1 . . . . . . . . . . −a† ℓ .   aℓ . 16 PETER A. BROOKSBANK, E. A. O'BRIEN, AND JAMES B. WILSON We caution that our illustration shows the structure constants for a typical con- figuration but changing the numbers of blocks both changes the number of boxes and alters their positions. To solve this system of equations, we decompose U∗ into terms U i ∗ using the block structure, as follows: a1 . . . . −a† 1 . . . .   . . . . . . . −a† ℓ aℓ .   = −a† 1 .   a1 . . . . . . . . . . . . . . .   . . ∩ ··· ∩ . . . . . . . . . . . . . . . . . −a† ℓ aℓ .   .   This is not precisely a decomposition into the blocks U i ∗ because each term on the right hand side is padded with zeros, thereby defining a degenerate extension U i i ) is straightforward: ∗ of the desired U i . . . ∗. However, the relationship between M( U i 1)⊥(cid:1) × Hom(cid:0)U i  ,   M( U i ∗) = M(U i ∗) + Hom(cid:0)U i =     M       Crucially, the positions of the zeros in the matrices of the M(U i Recalling that M(U∗) =Ti M(U i ∗ . ∗ . −γ† ∗ ∗) we discover that   2, (U i . α . . γ . . β . . δ . ∗ . ∗ ∗ . ∗ ∗ . ∗ ∗ . ∗ ∗ . ∗ ∗ . ∗ ∗ . ∗ ∗ . ∗ . δ† . −a† i ai . . . . . . . . . 1, (U i ∗) and M(U ∗ 2)⊥(cid:1) , and       ∗ . −β† ∗ . ∗ . α† . ∗ . ∗ . ∗ . ∗) shift as i changes. α0 β1 α1 M −a† 1     a1 . . . . . . . . . . . . . . −a† ℓ ∼=     aℓ .     Here, αi is an (ei × ei)-matrix where ei = dim U i is a (eℓ−1 × eℓ)-matrix. Hence, 1, β1 is a (e1 × e2)-matrix, and β2 . . . αℓ−1 β2 αℓ .     e2 i . ℓ Xi=1 (5.4) dimM(U∗) 6 e1 · e2 + eℓ−1 · eℓ + We remark that the specific entries αi are further constrained, since they are ad- joints of the individual bimaps U i ∗. Finally, we observe that ei = dim Vi − dim Vi−1, where 0 = V0 < V1 < ··· < Vℓ = V is the fixed point flag for L∗, from which the desired bound M(U∗)× 6 L∗1/ε in (5.2) follows. The bound on Aut(U0) is straight-forward, since dim U0 = dim L1+s = didi+s+1. s Xi=1 Thus, Aut(U∗) 6 L∗1/ε·L∗m2/ε, so Theorem 1.2 now follows from Theorem 1.1. 5.4. Number of isomorphism types. An immediate bound to the number of isomorphism types in Fq(d1, . . . , dℓ) is provided by the number of partitions, which grows exponentially in the dimension. Having fixed the partition, the number of isomorphism types among dense subalgebras is also large. To see this, observe that blocks yield numerous characteristic subalgebras. We can, for example, remove TESTING ISOMORPHISM OF GRADED ALGEBRAS 17 rows and columns resulting in characteristic quotients. If P(F ) is symmetric along the anti-diagonal, then we must remove both rows and columns; otherwise we can remove these independently. By removing sufficiently many rows/columns, we may assume that the partition has three parts. The number of isomorphism types for algebras based on such partitions may be estimated using arguments similar to those of Higman and Sims [2, Chapter 2] and suffice for the bound stated in Theorem 1.2. 6. Pseudo-isometries The occasional presence of symmetry provides an opportunity to improve the complexity of our isomorphism test for graded algebras. For example, if A∗ is a commutative algebra generated in degrees T , then the map AT × AT ֌ AT +T is symmetric; if A is a Lie algebra, then AT×AT ֌ AT +T is alternating. In such cases, the group of graded automorphisms of A∗ embeds in ΨIsom(AT × AT ֌ AT +T ), the group of pseudo-isometries defined in Section 2.4; typically this is a proper subgroup of Aut(AT × AT ֌ AT +T ). To take advantage of this observation, we must lift elements of Aut(AT +T ) directly to ΨIsom(AT × AT ֌ AT +T ) rather than to Aut(AT × AT ֌ AT +T ). Ivanyos and Qiao [14] devised an algorithm to do this. Our objective is to solve the following problem. PseudoIsometryCoset Given: weakly Hermitian bimaps U∗ and V∗. Return: the coset of all pseudo-isometries of f∗ : U∗ → V∗. Once again, we first consider the restricted version of the problem that arises by insisting that maps induce the identity on the codomain. The analogue of a principal isotopism in this weakly Hermitian context is an isometry. Definition 6.1. Let U∗ and V∗ be weakly Hermitian bimaps. If h : U0 → V0 is a linear isomorphism, then U∗ and V∗ are h-isometric if there is a pseudo-isometry f∗ : U∗ → V∗ with f0 = h. If U∗ = V∗ and h = 1, then Isom(U∗) is the isometry group of U∗. Consider the following problem and its coset analogue. IsometryGroup Given: a weakly Hermitian bimap U∗. Return: generators for Isom(U∗). IsometryCoset Given: weakly Hermitian bimaps U∗, V∗ and a linear isomorphism h : U0 → V0. Return: the coset of all isometries from U∗ → V∗ extending h. Polynomial-time solutions to IsometryGroup and IsometryCoset when K has odd characteristic appear in [8, Theorem 1.2] and [14], respectively. We combine these and an adaption of Algorithm 3 to solve PseudoIsometryCoset. 18 PETER A. BROOKSBANK, E. A. O'BRIEN, AND JAMES B. WILSON 7. Heuristics The family of graded Lie algebras described in Section 5 illustrates decisively the importance of the overarching "layer selection" philosophy of our approach. In this section, we assume that we have chosen a bimap U∗ = (U2, U1, U0,◦), and consider the problem of constructing its group of autotopisms. In Line 4 of Algorithm 3, we have a bimap U∗ = (U2, U1, U0,◦) and we must choose a group G with Aut(U∗)U0 6 G 6 Aut(U0). Without further information, we are forced to choose G = Aut(U0), a group which is often too large to search exhaustively. In this section we introduce heuristics to cut down the order of G. One idea is to compose U2 × U1 ֌ U0 with projections π : U0 → K c and record ∗ : U2 × U1 ֌ K c. Thus, we label isotopism invariants of the resulting bimap U π subspaces of the dual space U † 0 = HomK(U0, K). There are O(Kc(dim U0)) sub- spaces of dimension c, by contrast to the O(K(dim U0)2 ) elements of Aut(U0) we would otherwise be forced to list. It is therefore reasonable to list and label the former for small values of c. One then chooses G as the subgroup of Aut(U0) that preserves labels. In practice, we treat the projective geometry of U † 0 as a complete, colored graph, where vertices and edges are colored according to isotopism invariants. We then construct G as the automorphism group of the colored graph. Since no polynomial- time algorithm is known to construct the automorphism group of a graph, we cannot satisfactorily bound the complexity of this task. In practice, the construction of the graph is the more expensive task since the NAUTY algorithm [15] is extremely fast on generic graphs. Of course, there is no guarantee that G is a proper subgroup of Aut(U0). We discuss this matter further in Section 8. We now describe our labels for the points (vertices) and lines (edges) of U † 0 . We also examine another heuristic technique called fingerprinting that was first introduced in [13]. 7.1. Vertex labels. For each epimorphism π : U0 → K, we define a K-bilinear In particular, if we fix bases for U2 ∼= K a×1 and U1 ∼= K 1×b, then form U π ∗ . ∗ : K a × K b ֌ K is represented by its Gram matrix D ∈ K a×b defined by U π (ei ∗ ej)π = eiDej. Base changes to U2 and U1 leave the rank of the Gram matrix unchanged. We use the rank of D to label hπi ∈ P(U † 7.1.1. Examples. To illustrate subtleties arising from vertex labels, we consider two alternating bimaps K 4 × K 4 ֌ K 3 specified by systems of alternating forms. Let 0 ); this can be computed using [21]. 0 0 0  D =  0 1 0 −1 0 0    0 0 1 0 −1 0 0  E =     0 1 0 −1 0 0 0 0 1 0 −1 0 0 0 0 0 , 0 0 0 0  0 −1 0 0 , 0 0 0 1  0 −1 0 0 0 0 0 0 0 0 1 0 0 0 0 0     0 0 0 ,   0 0 0 0 0 1    0 0 0 0 −1 0 0 ,   0 0    0 −1 0 0 0 0 0 0 0 0 1 0 0 0 , . 0 0 0 0 TESTING ISOMORPHISM OF GRADED ALGEBRAS 19 In each case the dual space Hom(K 3, K) determines a projective plane. For each point P = (x1 : x2 : x3) in the plane we obtain matrices (mod K ×), (mod K ×). MD(P ) ≡ x1D[1] + x2D[2] + x3D[3] ≡  ME(P ) ≡ x1E[1] + x2E[2] + x3E[3] ≡  0 −x1 0 0 0 −x1 0 −x3 x1 0 0 x2 −x2 0 −x3 −x1 0 x1 0 x2 −x2 0 −x1 0 0 x3 x1 0 x3 0 x1 0     As indicated, these matrices are unique up to nonzero scalars, and scalars do not affect their ranks. Since MD(P ) and ME(P ) are nonzero and alternating, the only possible ranks are 2 and 4, distinguished by whether the determinants 1x2x3 + x2 det(MD(P )) = x4 1, 2x2 3 det(ME(P )) = x4 1 + 2x2 are zero or nonzero. Hence, the points of rank 2 for D lie on the hyperplane at infinity, namely (0 : 0 : 1) + (0 : 1 : 0), while all other points have rank 4. For E, the only rank 2 points with x1 = 0 satisfy x2x3 = 0, so there are just two, namely (0 : 1 : 0) and (0 : 0 : 1). The remaining rank 2 points have the form (1 : a : −1/a) for a ∈ K ×. Hence, there are also K + 1 total points of rank 2, but they clearly do not lie on a line. We illustrate the situation for the field of order 3 in Figure 2. It follows that the alternating bimaps defined by the systems D and E are not pseudo-isometric, a fact that will be transparent when we consider line labels. 7.2. Line labels. For each epimorphism λ : U0 → K 2, we consider the composition U λ ∗ : U2 × U1 ֌ U0 → K 2. Choices of bases for U2, U1, U0 determine pairs of matrices associated to U λ ∗ , and the label of hλi is a polynomial invariant derived from this pair. Characterizations of indecomposable pairs of matrices defined over the complex numbers were given by Kronecker and later for arbitrary fields by Dieudonn´e [10]. Using [10], Scharlau [17] characterized indecomposable skew-symmetric pairs. The reduction to indecomposable pairs is accomplished by computing a fully-refined orthogonal decomposition of U λ ∗ using an algorithm of Wilson [20]. Based on Scharlau's characterization, the aforementioned polynomial invariant for alternat- ing pairs was first introduced by Vishnivitskiı [19] for prime fields. It was later shown in [5, Theorem 3.22] to determine alternating pairs up to pseudo-isometry over all finite fields. The invariant may be adapted to determine all U λ ∗ up to autotopism. There are two consequences. First, the best possible labels for the lines of our projective geometry can be computed efficiently using the algorithm of [5]. Sec- ondly, there is a far greater variety of possible line labels than there are point labels, and so the study of lines may reveal significant global constraints. 7.2.1. Examples, revisited. Although the bimaps D and E in Section 7.1.1 are too small to illustrate great variability, we already begin to see differences in behaviour. We illustrate this for the field of order 3 in Figure 2. For bimap D there are only two line labels. The first (colored green) labels the single line at infinity 20 PETER A. BROOKSBANK, E. A. O'BRIEN, AND JAMES B. WILSON 001 012 011 102 112 122 101 111 121 010 100 110 120 2 4 2 2 4 4 4 4 4 4 2 4 4 4 2 2 4 4 4 4 4 4 2 4 2 4 Figure 2. On top is the geometry P G(2, 3) labeled by homoge- neous coordinates. The labeled and colored versions corresponding to bimaps D and E are shown on the left and right, respectively. (0 : 0 : 1) + (0 : 1 : 0), and the other lines are all labeled the same (colored red). For the bimap E there are three labels (colored blue, green, and red). 7.3. Fingerprinting. This process was introduced in [13] in the context of con- structing automorphism groups of p-groups. Here, we examine it in more detail and generalize it to our setting. First, consider a linear map f : V → W . Every subspace of V is mapped to a subspace of W , and the preimage of a subspace of W is a subspace of V . This cor- respondence is order-preserving and allows us to compare the projective geometries of V and W . Next, consider a bilinear map U∗ : U2 × U1 ֌ U0. We compare subspaces of U2 ⊗ U1 and U0. Since most of the elements of U2 ⊗ U1 are not pure tensors, the preimages that are pure tensors provide a largely ad hoc distribution of subspaces. This presents an opportunity to discover properties of U∗ that can be used to distinguish between bimaps. Definition 7.1. The left idealizer of S 6 U0 is λ(S) = {u2 ∈ U2 : u2 ◦ U1 6 S} and the right idealizer is ρ(S) = {u1 ∈ U1 : U2 ◦ u1 6 S}. These are subspaces of U2 and U1, respectively. We collect some basic properties of idealizers (stated just for left idealizers). Lemma 7.2. For subspaces S and T of U0, the following hold: (i) S 6 T =⇒ λ(S) 6 λ(T ); (ii) λ(S ∩ T ) = λ(S) ∩ λ(T ); and TESTING ISOMORPHISM OF GRADED ALGEBRAS 21 (iii) λ(S) + λ(T ) 6 λ(S + T ). Since the subspaces λ(S) + λ(T ) and λ(S + T ) are typically distinct, there is an opportunity to discover hidden invariants of a bimap. Using λ and ρ, uniform incidences in the projective geometry PG(U0) may be lifted to create generically irregular structures known as subspace arrangements. These are more general con- figurations than the better known hyperplane arrangements studied by Bjorner [1] and others. This can help to break symmetries, allowing us to see differences in bimaps that have no other obvious distinctions. Let us define more carefully the objects we use. An (affine) subspace arrangement is a set of subspaces of a vector space. We define the idealizer arrangements of U∗ = (U2, U1, U0,◦) as follows: A2 = {λ(P ) : P ∈ PG(U0)} A1 = {ρ(P ) : P ∈ PG(U0)}. We can build other arrangements for U∗ by shuffling coordinates. One way to discover Aut(U∗)-invariant substructures is to compute the intersection numbers of the arrangement up to some rank. This is essentially what is described as a fingerprint in [13] and the algorithm described there is polynomial in U0. One could, however, use more refined tools such as the Mobius function µ : PG(U0) × PG(U0) → Z, where µ(P, Q) =  1 λ(P ) = λ(Q), −Pλ(P )6X<λ(Q) µ(λ(P ), X) λ(P ) < λ(Q), otherwise. 0 8. Implementation and performance We have implemented our algorithms in Magma. The implementation utilizes various packages -- all publicly available on GitHub -- that have been developed by the authors and their collaborators [24]. The examples in Sections 7.1.1 and 7.2.1 show how our local heuristics expose global structures. Here we briefly report on experiments with our implementation that demonstrate their impact. 8.1. Twisted Heisenberg groups. Fix a prime p and odd integer k > 1. Put q := pk, A := Fq, and choose σ ∈ Gal(Fq). Define a product · on A, where x · y = xy + ixσyσ2 for x, y ∈ A. This turns A into a twisted Albert algebra, a nonassociative finite division algebra. Finally, define a bimap • : A2 × A2 ֌ A by u • v = u(cid:20) 0 1 −1 0 (cid:21) v†. p × F2k Over the ground field Fp, this gives a bimap F2k p, which can be encoded for computation as a system of k alternating forms of degree 2k. These bimaps arise from the twisted Heisenberg groups p ֌ Fk H(A) =    1 α γ 1 β 0 0 0  : α, β, γ ∈ A 1   . For various choices 1 < e 6 k, we now compose • with random ϕ ∈ Hom(F pk , Fpe) to obtain bimaps ◦ : F2k p, where u◦ v := u•ϕ v = (u• v)ϕ. The objective is to construct the group ΨIsom(◦). The default is to use the algorithm outlined in p × F2k p ֌ Fe 22 PETER A. BROOKSBANK, E. A. O'BRIEN, AND JAMES B. WILSON Section 6 which requires us to list GL(e, p). An alternative is to build the projective geometry PG(e, p) and label its points and lines, as described in Section 7. Table 1 records information from experiments with our implementation. For dif- ferent choices of p, k and e, it records: the numbers of points and lines in PG(e, p); the approximate order of GL(e, p); the order of the subgroup Ω of GL(e, p) pre- serving the point and line labels; and finally the order of the subgroup Ψ of Ω that lifts to pseudo-isometries of ◦. For each choice of parameters, we performed 10 random trials and recorded the Ω and Ψ that occurred most often. (In fact, these numbers were the same for all runs.) p 3 3 3 5 5 k 5 5 6 5 5 e # points 3 4 4 3 4 13 40 40 31 156 # lines 13 130 130 31 806 GL(e, p) ≈ 104 ≈ 108 ≈ 108 ≈ 107 ≈ 1011 Ω 16 10 1 1 20 Ψ 2 10 1 1 20 Table 1. Experiments for the twisted Heisenberg groups We remark that other weaker (but more efficiently computed) invariants can be used to label the lines in our projective geometries. Often these suffice to discover structures that must be preserved by pseudo-isometries (or by autotopisms in more general applications). However, the weaker invariants do not distinguish quotients of twisted Heisenberg groups: all resulting line labels are identical. But, by using the invariants described in Section 7.2, the problem breaks completely. 8.2. Random nilpotent Lie algebras of class 2. If U∗ = (U2, U1, U0,◦) is a generic bimap -- which we loosely define to be one specified by an arbitrary selection of structure constants -- then the group induced by Aut(U∗) on U0 is usually very small, and is often trivial. In such cases, local invariants (even the weaker ones) almost always cut down to an overgroup containing the correct group as a subgroup of very small index. In Table 2, we mimic the experiments and reporting of Section 8.1, but this time we generate alternating bimaps ◦ : Fd p by selecting e random skew- symmetric d × d matrices over Fp. Once again, we select the most commonly occurring Ω and Ψ from 10 runs with each choice of parameters (d, p, e). p × Fd p ֌ Fe d 10 20 20 10 20 10 10 p 3 3 3 3 5 5 5 e # points # lines 13 3 13 3 130 4 5 1210 31 3 4 806 20306 5 13 13 40 121 31 156 781 GL(e, p) ≈ 104 ≈ 104 ≈ 108 ≈ 1011 ≈ 106 ≈ 1011 ≈ 1017 Ω 1 1 1 1 1 1 1 Ψ 1 1 1 1 1 1 1 Table 2. Experiments for Lie algebras of class 2 TESTING ISOMORPHISM OF GRADED ALGEBRAS 23 These tests suggest that, for generic bimaps, ΨIsom(◦) acts trivially on its codomain. The variation of the dimension, d, of the domain space has little impact on the outcome, but increasing e introduces more constraints and therefore makes it increasingly likely that ΨIsom(◦) acts trivially. As we see, the local invariants usually detect when this is the case. 8.3. Do the heuristics always work? The striking practical success of the local invariants raises the obvious question of whether we can strengthen the theoretical complexity of Theorem 1.1. If we label just points and lines, the answer is no: there exist alternating bimaps U∗ for which all points and lines are labeled identically, yet the group induced by ΨIsom(U∗) on its codomain U0 is a proper subgroup of Aut(U0); one such example appears in [18]. References [1] Anders Bjorner, Subspace arrangements, First European Congress of Mathematics, Vol. I (Paris, 1992), Progr. Math., vol. 119, Birkhauser, Basel, 1994, pp. 321 -- 370. MR1341828 [2] Simon R. Blackburn, Peter M. Neumann, and Geetha Venkataraman, Enumeration of finite groups, Cambridge Tracts in Mathematics, vol. 173, Cambridge University Press, Cambridge, 2007. MR2382539 [3] W. Bosma, J. Cannon, and C. Playoust, The Magma algebra system. I. The user language, J. Symbolic Comput. 24 (1997), no. 3-4, 235 -- 265. MR1484478 [4] Peter A. Brooksbank and Eugene M. Luks, Testing isomorphism of modules, J. Algebra 320 (2008), no. 11, 4020 -- 4029. MR2464805 (2009h:16001) [5] Peter A. Brooksbank, James B. Wilson, and Joshua Maglione, A fast isomorphism test for groups whose Lie algebra has genus 2, J. Algebra 473 (2017), 545 -- 590. MR3591162 [6] Peter A. Brooksbank and E. A. O'Brien, Constructing the group preserving a system of forms, Internat. J. Algebra Comput. 18 (2008), no. 2, 227 -- 241. MR2403820 (2009g:20020) [7] Peter A. Brooksbank and James B. Wilson, Groups acting on tensor products, J. Pure Appl. Algebra 218 (2014), no. 3, 405 -- 416. MR3124207 [8] , Computing isometry groups of Hermitian maps, Trans. Amer. Math. Soc. 364 (2012), no. 4, 1975 -- 1996. MR2869196 [9] David G. Cantor and Hans Zassenhaus, A new algorithm for factoring polynomials over finite fields, Math. Comp. 36 (1981), no. 154, 587 -- 592. [10] Jean Dieudonn´e, Sur la r´eduction canonique des couples de matrices, Bull. Soc. Math. France 74 (1946), 130 -- 146. MR0022826 (9,264f) [11] Bettina Eick, Computing the automorphism group of a solvable Lie algebra, Linear Algebra Appl. 382 (2004), 195 -- 209. MR2050106 [12] , Computing automorphism groups and testing isomorphisms for modular group al- gebras, J. Algebra 320 (2008), no. 11, 3895 -- 3910. [13] Bettina Eick, C. R. Leedham-Green, and E. A. O'Brien, Constructing automorphism groups of p-groups, Comm. Algebra 30 (2002), no. 5, 2271 -- 2295. MR1904637 (2003d:20027) [14] G´abor Ivanyos and Youming Qiao, Algorithms based on ∗-algebras, and their applications to isomorphism of polynomials with one secret, group isomorphism, and polynomial identity testing, SIAM Journal on Computing. [15] Brendan D. McKay, Practical Graph Isomorphism, Congr. Numer. 30 (1981), 45 -- 87. [16] Joshua Maglione, Efficient characteristic refinements for finite groups, J. Symbolic Comput. 80 (2017), 511 -- 520. MR3574524 [17] Rudolf Scharlau, Paare alternierender Formen, Math. Z. 147 (1976), no. 1, 13 -- 19. MR0419484 (54 #7505) [18] Libero Verardi, Semi-extraspecial groups of exponent p, Ann. Mat. Pura Appl. 148 (1987), no. 4, 131 -- 171. MR0932762 (89h:20033) [19] A. L. Vishnevetskiı, A system of invariants of certain groups of class 2 with commutator subgroup of rank two, Ukrain. Mat. Zh. 37 (1985), no. 3, 294 -- 300, 403. MR795568 (86k:20033) [20] James B. Wilson, Finding central decompositions of p-groups, J. Group Theory 12 (2009), no. 6, 813 -- 830. MR2582050 24 [21] [22] [23] PETER A. BROOKSBANK, E. A. O'BRIEN, AND JAMES B. WILSON , Optimal algorithms of Gram-Schmidt type, Linear Algebra Appl. 438 (2013), no. 12, 4573 -- 4583. MR3039211 , Division, adjoints, and dualities of bilinear maps, Comm. Algebra 41 (2013), no. 11, 3989 -- 4008. MR3169502 , On automorphisms of groups, rings, and algebras, Comm. Algebra 45 (2017), no. 4, 1452 -- 1478. MR3576669 [24] Peter A. Brooksbank, Joshua Maglione, E.A. O'Brien, and James B. Wilson, isomorphism tests, for multilinear TheTensor.Space: Repositories https://github.com/thetensor-space , 2019. algebra and Brooksbank, Department of Mathematics, Bucknell University, Lewisburg, PA 17837, USA O'Brien, Department of Mathematics, University of Auckland, Private Bag 92019, Auckland, New Zealand Wilson, Department of Mathematics, Colorado State University, Fort Collins, CO 80523, USA
1601.00207	1	1601	2016-01-02T18:51:26	Subrings of $\mathbb{C}$ Generated by Angles	[ "math.RA" ]	Consider the following inductively defined set. Given a collection $U$ of unit magnitude complex numbers, and a set initially containing just 0 and 1, through each point in the set, draw lines whose angles with the real axis are in $U$. Add every intersection of such lines to the set. Upon taking the closure, we obtain $R(U)$. We investigated for which $U$, $R(U)$ is a ring. Our main result holds for when $1 \in U$ and $\|U\| \ge 4$. If $P$ is the set of real numbers in $R(U)$ generated in the second step of the construction, then $R(U)$ equals the module over $\mathbb{Z}[P]$ generated by the set of points made in the first step of the construction. This lets us show that whenever the pairwise products of points made in the first step remain inside $R(U)$, it is closed under multiplication, and is thus a ring.	math.RA	math	Subrings of C Generated by Angles Juniper Bahr, Arielle Roth July 27, 2021 Abstract Consider the following inductively defined set. Given a collection U of unit magnitude complex numbers, and a set initially containing just 0 and 1, through each point in the set, draw lines whose angles with the real axis are in U . Add every intersection of such lines to the set. Upon taking the closure, we obtain R(U ). We investigated for which U , R(U ) is a ring. Our main result holds for when 1 ∈ U and U ≥ 4. If P is the set of real numbers in R(U ) generated in the second step of the construction, then R(U ) equals the module over Z[P ] generated by the set of points made in the first step of the construction. This lets us show that whenever the pairwise products of points made in the first step remain inside R(U ), it is closed under multiplication, and is thus a ring. 1 Introduction Suppose we are given a collection U of unit length elements of C. If we have some collection of points in C, we can draw lines through each of them with every angle in U (with respect to the real axis). In this way we can construct intersections of lines and repeat the process. Specifically, if we start with 0 and 1 in the complex plane and continue this construction forever until it is completed, when is the resulting collection of points a subring of the complex numbers? Note that even though we are drawing lines, only the intersection points are considered to be constructed. In [1], Buhler et al. motivated this construction with a discussion of origami where two folds can intersect to create a reference point. Definition 1.1. Let p, q, α, β ∈ C with α = β = 1. Define Lα(p) to be In other words, Lα(p) := p + Rα. Define the line through p with angle α. Iα,β(p, q) := Lα(p) ∩ Lβ(q) when α 6= ±β so that an intersection exists. Definition 1.2. Let U be a set of unit magnitude complex numbers. Set S0 = {0, 1}. For each n ∈ N, set Sn+1 = {Iα,β(p, q) α, β ∈ U , p, q ∈ Sn, and α 6= ±β}. We then define R(U ) = Sn∈N Sn. 1 Definition 1.3. T = {z ∈ C z = 1} which is viewed as a group under complex multiplication. T /{±1} will be used for the collection of angles, since α and β are considered equivalent iff α = ±β. Unless otherwise specified, U ⊆ T /{±1}. Definition 1.4. Given U ⊆ T /{±1}, we define all elements z ∈ R(U ) of the form Iα,β(0, 1) to be elementary monomials, i.e., length 1 monomials. Next, if m is a length k monomial, then Iα,β(0, m) ∈ R(U ) is a length k + 1 monomial. In this way we inductively define monomials. Proposition 1.5 ([1]). We can calculate Iα,β (p, q) as follows for p, q ∈ C and α 6= β ∈ T /{±1}. Iα,β(p, q) = [β, q] [β, α] α [α, p] [α, β] β + where [x, y] = x¯y − y ¯x and ¯z is the complex conjugate of z. Proposition 1.6 ([1]). We list some properties of Iα,β (p, q) below for w ∈ T /{±1} and r ∈ R. (Symmetry) Iu,v(p, q) = Iv,u(q, p) (Reduction) Iu,v(p, q) = Iu,v(p, 0) + Iv,u(q, 0) (Linearity) Iu,v(rp + q, 0) = rIu,v(p, 0) + Iu,v(q, 0) (Rotation) For w ∈ T /{±1}, wIu,v(p, q) = Iwu,wv(wp, wq). Lemma 1.7 ([1]). Let U ≥ 3 with 1 ∈ U . Then, R(U ) is closed under addition and additive inverses. Theorem 1.8 ([1]). Let U ≥ 3. R(U ) is the collection of integer linear com- binations of monomials. Remark. Since whenever U ≥ 3, R(U ) is a group under addition, we need only check closure under multiplication to ensure that R(U ) is a ring. The authors of [1] then studied the case when U is a group. Specifically, they took the set of unit magnitude complex numbers T (i.e., the unit circle) and considered it to be a group under complex multiplication. Then they took the quotient of T by {−1, +1}. The result can be viewed as the top half of the unit circle. By convention, whenever we use U , we will refer to U ⊆ T /{±1} where the elements are viewed as complex numbers. Theorem 1.9 ([1]). Let U be a subgroup of T /{−1, +1} with U ≥ 3. Then, R(U ) is a ring. In their paper, Buhler et al. observed that R(U ) maybe be a ring even when U is not a group. They left the question of what properties U must satisfy exactly for R(U ) to be a group open. 2 2 Three Angles In order to understand R(U ), first we looked at U = 3 with 0 ∈ arg(U ). We found that R(U ) has the structure of a lattice and can be understood in terms of one of the elementary monomials. Lemma 2.1. Let U = {1, u, v}. We claim that R(U ) is a lattice in C with the form R(U ) = Z + Iu,v(0, 1)Z. Proof. Set x = Iu,v(0, 1). From Lemma 1.7, we know that R(U ) is a subgroup of C with addition. Since 1 ∈ R(U ) and x ∈ R(U ), we clearly see that R(U ) ⊇ Z + xZ. We will prove the other containment with induction. We know that S1 = {x, 1 − x, 0, 1} ⊆ Z + xZ. Let p, q ∈ Sn, which is assumed to be in Z + xZ. Let α, β ∈ U . We claim that z = Iα,β(p, q) ∈ Z + xZ. Since Iα,β(p, q) = Iα,β(p, 0) + Iβ,α(q, 0), it suffices to prove that Iα,β (a + bx, 0) ∈ Z + xZ. Further note that Iα,β(a + bx, 0) = Iα,β (a, 0) + Iα,β(bx, 0) = aIα,β(1, 0) + Iα,β (bx, 0). by linearity. Iα,β(1, 0) ∈ S1, so Iα,β(1, 0) = 1, 0, x, or 1− x. There are only four choices since if one of the angles is 0 radians, the resulting point is 0 or 1. If α, β 6= 1, then there are two choices left, α = u, β = v or α = v, β = u. One of these yields the point x and the other yields (by the parallelogram law) 1 − x. Thus Iα,β(a, 0) ∈ Z + xZ. Next, note that Iα,β(bx, 0) = bIα,β(x, 0). Thus it suffices to prove that Iα,β(x, 0) ∈ Z + xZ. We have 6 cases. (u, v) Since x = ru for some r ∈ R, Iu,v(x, 0) = rIu,v(u, 0) = 0 ∈ Z + xZ. (v, u) Iv,u(x, 0) is the projection of x on to the line ru in the direction of v, but x ∈ Ru, so Iv,u(x, 0) = x. (u, 1) Iu,1(x, 0) is the projection of x on to the real axis in the direction of u. It is easy to see that this must be 0, since the line from 0 (which is on the real axis) extending in the u direction intersects x. (v, 1) Iv,1(x, 0) = 1, for a similar reason. The line extending from 1 (which is on the real axis) in the v direction intersects x. (1, u) I1,u(x, 0) is the line crossing through x + s and ru for s, r ∈ R, but since x ∈ Ru, this intersection is clearly at x. (1, v) I1,v(x, 0) is at x − 1 which is demonstrated by the fact that I1,v(x, 0) + Iv,1(x, 0) = x and Iv,1(x, 0) = 1. All of these points line in Z + xZ, so we have shown that R(U ) for U = 3 is of the form Z + xZ where x = Iu,v(0, 1). 3 Remark. Given U = {1, u, v}, if we find u′, v′ such that Iu′,v′ (0, 1) = m + Iu,v(0, 1) for m ∈ Z and set U ′ = {1, u′, v′}, by the above structural result R(U ) = R(U ′). Theorem 2.2 expands on this remark and show exactly when U and U ′ of size three generate the same lattice. Theorem 2.2. Let Iu,v(0, 1) = x and let Iu′,v′ (0, 1) = y. Let x = a + bi and y = c + di. Set U = {1, u, v} and U ′ = {1, u′, v′}. We claim that R(U ) = R(U ′) if and only if b = ±d and a ∓ c ∈ Z. Proof. Z + xZ = Z + yZ means that {m + nx m, n ∈ Z} = {p + qy p, q ∈ Z}. For arbitrarily m, n ∈ Z, m + nx ∈ {p + qy p, q ∈ Z} holds iff nx ∈ Z + yZ, which is equivalent to na + nbi = p + qc + qdi for some p, q ∈ Z. In order for this to hold, the imaginary parts must equal: nbi = qdi (for any n, there is some q). Thus d b (using n = 1). We can make the same argument swapping x and y, which tells us that b d, so b = ±d and thus n = ±q. Also, the real parts must be equal: na − qc = p (for any n there are such p, q). Above we determined that n = ±q, so n(a ∓ c) = p. Such a p exists for any n, so a ∓ c ∈ Z. We showed that if Z + xZ = Z + yZ, then b = ±d and a ∓ c ∈ Z. Now, if we assume that b = ±d and a ∓ c ∈ Z, then for any Z + xZ = m + na + nbi, we have m + na + nbi = m + n(k ± c) + n(±d)i = (m + nk) ± nc ± ndi ∈ Z + yZ. This shows that Z + xZ ⊆ Z + yZ. Likewise, Z + yZ ⊆ Z + xZ. and only if b = ±d and a ∓ c ∈ Z, so Z + xZ = Z + yZ. Since R(U ) = Z + xZ and R(U ′) = Z + yZ, we have that R(U ) = R(U ′) if Now that we understand what form R(U ) has for U = 3 with 0 ∈ arg(U ), we can easily show exactly when R(U ) is a ring. The only point that gives any difficulty is x, one of the two elementary monomials off of the real line. If we can square this point and the result lies in R(U ), then R(U ) = Z + xZ must be closed under multiplication. Now we characterize all U with 0 ∈ arg(U ) and U = 3 such that R(U ) is a ring. Theorem 2.3. Let U = {1, u, v} and let Iu,v(0, 1) = x. R(U ) is a ring if and only if x is a (non-real) quadratic integer, i.e., x is the root of some monic integer quadratic polynomial. Proof. First we will prove that if x is a quadratic integer, then R(U ) is a ring. Note that R(U ) = Z + xZ where x = Iu,v(0, 1). Since R(U ) is already a group, we need to show closure under multiplication. We write (a + bx)(c + dx) = ac + (bc + ad)x + bdx2. Since x is a quadratic integer, x2 = λx + µ for some 4 λ, µ ∈ Z. Then, (a + bx)(c + dx) = ac + (bc + ad)x + bd(λx + µ) = (ac + bdµ) + (bc + ad + bdλ)x so in fact R(U ) is closed under multiplication. Now assume that R(U ) is closed under multiplication. Then (a+bx)(c+dx) ∈ Z + xZ, but we can expand this: (a + bx)(c + dx) = ac + (bc + ad)x + bdx2 ∈ Z + xZ Since ac + (bc + ad)x ∈ Z + xZ, we know that bdx2 ∈ Z + xZ for every b, d ∈ Z. In particular, this holds for b = d = 1, so x2 ∈ Z + xZ. In other words, x must be a quadratic integer. Also, if x ∈ R, then our R(U ) is degenerate, so we need x /∈ R. We can compute the intersection point x in terms of arg(u) and arg(v) and rephrase Theorem 2.3. Corollary 2.4. Let arg(U ) = {0, θ, φ} with φ < θ. Then R(U ) is a ring if and only if tan θ + tan φ tan θ tan θ − tan φ is a quadratic integer. tan θ − tan φ Proof. We can see from the following figure that (1 + w) tan φ = h = w tan θ so w = tan φ tan θ−tan φ . i x h φ 1 θ w Immediately, we see also that h = tan φ tan θ tan θ−tan φ . Thus, x = tan θ tan θ − tan φ + tan φ tan θ tan θ − tan φ Remark. In [2], Nedrenco independently characterized R(U ) where U = 3, describing R(U ) = Z + xZ and generalized to when 0 /∈ arg(U ). In the same paper, Nedrenco also noted that R(U ) is dense when U = 4. We present what we found independently. 5 3 Four or More Angles Since we understood R(U ) for U = 3 in terms of an elementary monomial, we wish to understand R(U ) for U ≥ 4 in terms of elementary monomials. Because R(U ) is now dense in the complex plane, we cannot hope for an integral basis. By linearity if we have some p ∈ R ∩ R(U ), then Iα,β(0, p) = pIα,β(0, 1). This means we can scale points. This motivates our interest in "projections" on to the real axis. Proposition 3.1. Let U = {1, u, v, w} with arg(u) < arg(v) < arg(w) < π. There are at most eight length-two monomials on the real axis. There are at most five length-two monomials constructed from elementary monomials of the form Iα,β(0, 1) with arg(α) < arg(β). They are 0, 1, x, 1/x, x/(x − 1) where x = Iv,1(Iu,w(0, 1), 0). Proof. With the exception of 0 and 1, the only way to construct a length-two monomial on the real axis is to intersect a line through an elementary monomial and the line passing through 0 and 1. For any given elementary monomial, there are already two lines passing through the point: one passes through 0 and one passes through 1. Thus there can be at most 6 extra length-two monomials on the real axis, at most three of which created from z1, z2, z3 in the form described in the claim, and at most three of which created from 1− z1, 1− z2, 1− z3 which are of the opposite form. z3 z1 z2 1 p2 p6 p5 p3 0 1 − z2 p4 p1 1 − z1 1 − z3 Note that p1 = 1 − p4, p2 = 1 − p5, and p3 = 1 − p6. As proof, we calculate I1,α(0, Iβ,γ(0, 1)) = I1,α(0, 1 − Iγ,β(0, 1)) = I1,α(0, 1) − I1,α(0, Iγ,β(0, 1)) = 1 − I1,α(0, Iγ,β(0, 1)) Now we will show that the projections have the described form. Set x = p1. 1 = z1 . z2 = z1 . z2 Note that the triangle 0− p1− z1 is similar to the triangle 0− 1− z2, so p1 Also, the triangle 0 − 1 − z1 is similar to the triangle 0 − p2 − z2, so 1 Thus, p2 = 1/x. p2 6 z2 z3 z1 p3 0 p1 1 p2 z1 Next, the triangle 0 − p1 − z1 is similar to the triangle p3 − 0 − z3, so p3 . Also, the triangle 0− 1− z1 is similar to the triangle p3 − 1− z3, z3−p3 = p1 so z3−p3 = 1 z1 1−p3 . We conclude = x p3 1 1 − p3 p3 − 1x = p3 x p3 = (cid:12)(cid:12)(cid:12)(cid:12) (cid:12)(cid:12)(cid:12)(cid:12) x − 1 To remove the absolute value signs, we note that since arg(z3) > arg(z1), the line through z3 with angle arg(z1) must intersect the negative real axis, so x−1 < 0, so we deduce that p3 = x/(x − 1). p3 < 0. Furthermore, since x < 1, x Now that we understand a small amount of R ∩ R(U ), we can quickly con- struct an entire ring inside R ∩ R(U ) with the scaling mentioned earlier. Later we will show that what we construct next is exactly R ∩ R(U ) Proposition 3.2. Let 0 ∈ arg(U ) with U ≥ 4. Let P be the set of length-two monomials on the real axis. For any x ∈ R(U ) and any p ∈ P , px ∈ R(U ). As a result, the ring Z[P ] is constructible, i.e., Z[P ] ⊆ R(U ). Proof. Let p be a projection. Since R(U ) is the collection of finite linear com- binations of monomials, it suffices to construct pm for a given monomial m, i=1 cimi for ci ∈ Z The proof that pm ∈ R(U ) follows from linearity. Formally, we rely on since if we have x ∈ R(U ), we can simply represent x = Pn and then write px = Pn i=1 ci(pmi). induction. Base Case: The length of m is one, so m = Iα,β(0, 1) for some α, β ∈ U . Then, pm = Iα,β(0, p) by linearity, which is in R(U ) since p ∈ R(U ). Inductive Step: Suppose every length n− 1 monomial satisfies the claim. Let m be of length n. Then, m = Iα,β(0, q) for some length n − 1 monomial q. By linearity, pm = Iα,β(0, pq) which is constructible since pq ∈ R(U ) by the inductive hypothesis. 7 Thus every monomial can be arbitrarily multiplied by projections, so in fact everything in R(U ) can be arbitrarily multiplied by projections. In particular, so can the projections themselves. This means that arbitrary powers of projections are in R(U ). Furthermore, since R(U ) is a group under addition, Z[P ] ⊆ R(U ). Remark. Since the above result does not rely on the previous two results, this holds even when U > 4. Our current goal is to characterize all monomials in terms of Z[P ] and el- ementary monomials. By Theorem 1.8, if the monomials have a nice enough form, we will be able to understand all of R(U ). Characterizing all monomials starts with the length two monomials. First, however, we need a quick lemma. Lemma 3.3. Let 0, α, β ∈ arg(U ). Let p, q ∈ R(U ), and let x = Iα,β(p, q) and y = Iβ,α(p, q). Then, x = p + q − y. Proof. Since the lines from x to q and from p to y are parallel, and also the lines from x to p and from q to y are parallel, this forms a parallelogram. It is clear that 0, x− q, p− q, and y− q form a parallelogram and that x− q + y− q = p− q, so x + y − q = p. p y α x β q Lemma 3.4. Let U ≥ 4 and let 0 ∈ arg(U ). Let P be the set of projections from the elementary monomials to the real axis along angles in U . Every length two monomial is a Z[P ]-linear combination of elementary monomials. Proof. Let z = Iα,β(0, 1) for some α, β ∈ U and let our length two monomial m = Iγ,δ(0, z). We will prove that m is a Z[P ]-linear combination of elementary monomials by cases. (δ = 1): Note that Iγ,0(0, z) + I0,γ(0, z) = z, so Iγ,δ(0, z) = z − I0,γ(0, z). Since I0,γ(0, z) ∈ P , m is a Z[P ]-linear combination of elementary monomials. 8 (δ = α): Since the line through z = Iα,β(0, 1) with angle arg(α) passes through the origin, m = Iγ,α(0, z) = 0. This is trivially a Z[P ]-linear combination of elementary monomials. (δ = β): Since the line through z = Iα,β(0, 1) with angle arg(β) passes through 1, m = Iγ,β(0, z) = Iγ,β(0, 1), which is an elementary monomial. (δ ∈ U \ {1, α, β}): Let p = I0,γ(0, z) be the projection from z to the real axis in the direction of γ. Note that Iγ,δ(0, p) = pIγ,δ(0, 1) by linearity. Set x = Iγ,δ(0, p). We that x + z − p = m, and since x = pIγ,δ(0, 1), this is enough to prove that m is a Z[P ]-linear combination of elementary monomials. Restated, the claim is that Iγ,δ(0, I0,γ(0, z)) + z − I0,γ(0, z) = Iγ,δ(0, z) To prove this, we will show that Iγ,δ(x, z) = m. This follows by the fact that x ∈ Rγ, so the line through x with angle arg δ also passes through 0 and thus Iγ,δ(x, z) = Iγ,δ(0, z) = m. Furthermore, Iδ,γ(x, z) = p. To see this, first note that Iγ,0(z, 0) = p. Also, Iδ,0(x, 0) = p, because Iδ,0(x, 0) = Iδ,0(Iγ,δ(0, p), 0) and both x and p lie along the same line through p with angle arg(δ) (by construction of x). This means that x and z lie on opposite corners of a parallelogram which has a corner at p through the real axis and another corner through m. Thus, 0, (x − p), (z − p), and (m − p) form the corners of a parallelogram and (x − p) + (z − p) = m − p so x + z − p = m, concluding the proof. Since in all cases m is a Z[P ]-linear combination of elementary monomials, we know that every length two monomial is of this form. Now that we understand length two monomials, we can apply induction to characterize all monomials, and thus all of R(U ). Theorem 3.5. Let 0 ∈ arg(U ). Let P be the set of projections of elementary monomials along lines with angles from arg(U ) on to the real axis. Then, ev- ery monomial in R(U ) is a Z[P ]-linear combination of elementary monomials. Indeed, R(U ) is the set of Z[P ]-linear combinations of elementary monomials. Proof. We will prove this by induction on the length of the monomial. Length one monomials are already elementary and length two monomials follow from the above theorem. Let m be length n and suppose that all length n−1 monomials 9 are of this form. Then, m = Iα,β(0, m′) k = Iα,β(0, cizi) X i=1 = = k X i=1 k X i=1 ciIα,β(0, zi)  ci ℓ X j=1 dixi   length two monomials are of this form. using linearity and the fact that all The ci and di are in Z[P ] and the xi and zi are elementary monomials. After simplification, it is easy to see that m is in fact a Z[P ]-linear combination of elementary monomials. Since everything in R(U ) is an integer linear combination of monomials, everything in R(U ) is a Z[P ]-linear combination of elementary monomials. Furthermore, since Z[P ] is constructible by Proposition 3.2, and pR(U ) ⊆ R(U ) for all p ∈ P , we can construct every Z[P ]-linear combination of ele- mentary monomials. Thus, R(U ) equals the set of Z[P ]-linear combinations of elementary monomials. Remark. We can alternatively say that R(U ) is a Z[P ]-module in C generated by the elementary monomials. As in the three-angle case, understanding the structure of R(U ) led us to understand when R(U ) is a ring in terms of products of elementary monomials. In fact Theorem 2.3 could probably be seen as a special case of the following theorem. Theorem 3.6. Let U with U ≥ 4 and 0 ∈ arg(U ) and let P represent the collection of projections. R(U ) is a ring if and only if every pairwise product of elementary monomials is a Z[P ]-linear combination of elementary monomials. Proof. First note that R(U ) equals the collection of Z[P ]-linear combinations of elementary monomials. We that the Z[P ]-linear combinations of elementary monomials are closed under multiplication if and only if every pairwise product of elementary monomials is a Z[P ]-linear combination of elementary monomi- als. Assume that every pairwise product of elementary monomials is as above. j=1 djyj for i=1 cixi and y = Pm Then, for any x, y ∈ R(U ), we write x = Pn ci, dj ∈ Z[P ] and xi, yj elementary monomials. mentary monomials, so is xy. Thus R(U ) is a ring. Then, xy = Pi,j cidjxiyj. Since xiyj is a Z[P ]-linear combination of ele- Now, suppose that R(U ) is a ring. It must be closed under multiplication, so the pairwise product of elementary monomials must be in R(U ), but R(U ) is the Z[P ]-linear combinations of elementary monomials, so the claim holds. 10 Since we have at least one projection p ∈ (0, 1), we can construct points close to zero. Because elements of R(U ) scaled by p are still in R(U ) and R(U ) is a group, it is actually dense in C as we will prove below. Theorem 3.7. If 1 ∈ U and U ≥ 4, then R(U ) is dense in C. Proof. Since R(U ) is the set of Z[P ]-linear combinations of elementary mono- mials, if z is a non-real elementary monomial and p ∈ Z[P ] ∩ (0, 1), we can construct pn and pnz which go to zero from, two different directions. Let ε > 0 and let x ∈ C. Since R(U ) is a group under addition, we can construct apN1 + bpN2z for all N1, N2 ∈ N. Since p ∈ (0, 1), we can find N2 such that (cid:12)(cid:12)Im(z)pN2(cid:12)(cid:12) < ε/2. To simplify the following expression, write θ = Im(z)pN2. Then there exists a unique b ∈ Z such that So we can show that b − 1 ≤ Im(x) θ ≤ b (cid:12)(cid:12)b Im(z)pN2i − Im(x)i(cid:12)(cid:12) = bθ − Im(x) ≤ ε/2 Likewise we can find a, N1 such that (cid:12)(cid:12)apN1 − (cid:0)Re(x) − bpN2 Re(p)(cid:1)(cid:12)(cid:12) < ε/2. Once we have such a ∈ Z and N1 ∈ N, we have (cid:12)(cid:12)apN1 + bpN2z − x(cid:12)(cid:12) = (cid:12)(cid:12)apN1 + bpN2 Re(z) − Re(x) + bpN2 Im(z)i − Im(x)i(cid:12)(cid:12) ≤ (cid:12)(cid:12)apN1 + bpN2 Re(z) − Re(x)(cid:12)(cid:12) + (cid:12)(cid:12)bpN2 Im(z) − Im(x)(cid:12)(cid:12) Since apN1 + bpN2z ∈ R(U ), and this holds for any x ∈ C and for every ε > 0, we can always find a point in R(U ) arbitrarily close to any point of C. Thus, R(U ) is dense in C. < ε 4 Some U for Which R(U ) Is a Ring Now we can use Theorem 3.6 to prove that R(U ) is a ring for a particular example of U . Example 1. Let U = {1, eiπ/6, eiπ/3, eiπ/2}. R(U ) is a ring. Proof. It suffices to show that all products of elementary monomials are Z[P ]- linear combinations of elementary monomials. Our elementary monomials are 0, 1, z1, z2, z3, 1 − z1, 1 − z2, 1 − z3, where 2√3 3 eiπ/6 z1 = z2 = √3eiπ/6 z3 = 2eiπ/3 11 Z[2/3, 3/2] = Z[1/3, 1/2] = Z[1/6]. First we calculate the projections and get 2/3, 3/2,−2. Note that Z[2/3, 3/2,−2] = We calculate all pairwise products of z1, z2, z3, since calculating more would be redundant, as the others are either 0, 1, or an integer linear combination of {1, z1, z2, z3}. 4 3 z2 1 = eiπ/3 = 2 3 z1z2 = 2eiπ/3 = z3 z3 z1z3 = = 4(z1 − 1) z3 = z3 eiπ/2 = 4i √3 2 3 4 √3 z2 3 2 z2 2 = z2 1 z2 2 1 z2 z1z3 = 6(z1 − 1) z1 9 4 · = z2z3 = z2 3 = z1z2z3 = 6(z2 1 − z1) = 4z3 − 6z1 These are all in R(U ), so R(U ) is closed under multiplication and is a ring. Remark. We suspected that perhaps any subset U of a finite group containing a generator for that finite group would result in a ring. The following example shows that this cannot be necessary. Example 2. Let U = {1, eiπ/6, eiπ/4, eiπ/3}. R(U ) is a ring. Proof. As above, it suffices to show that the products of all elementary mono- mials are Z[P ]-linear combinations of elementary monomials. We go by the con- vention that z1 = Ieiπ/6,eiπ/2 (0, 1), z2 = Ieiπ/6,eiπ/3(0, 1), and z3 = Ieiπ/3,eiπ/2(0, 1) and that p1, p2, p3 are projections from z1, z2, z3 to the real axis. We calculated z1z2 = p3(1 − z3) z1z3 = −p1z2 − (p2p3)z3 + 2p3 z2z3 = −p3z2 − (p2p3)z3 + 2p2p3 z2 1 = p1p3(1 − z3) z2 2 = p2p3(1 − z3) z2 3 = −6z2 − 3p2p3z3 + 3p3 Remark. We then suspected that any subset of a finite group might result in a ring. Our next result shows this too cannot be necessary. Example 3. Let U = {1, ei, e2i, e3i}. R(U ) is a ring. This example is a special case of Theorem 4.1. 12 Remark. We strongly suspect that R({1, eiπ/5, eiπ/4, eiπ/3}) is not a ring, so we suspect that it is not sufficient for U to just be a subset of a finite group. Theorem 4.1. Let U = {1, α, α2, α3}. R(U ) is a ring. Proof. Set z1 = Iα,α3 (0, 1), z2 = Iα,α2 (0, 1), and z3 = Iα2,α3(0, 1). Since the only elementary monomials are 0, 1, z1, z2, z3, 1 − z1, 1 − z2, 1 − z3, it suffices to check pairwise products of {z1, z2, z3}. Set p1 = I1,α2 (0, z1), p2 = I1,α3 (0, z2), and p3 = I1,α(0, z3). Then Z[P ] = Z[p1, p2, p3], since the other projections are 0, 1, 1 − p1, 1 − p2, and 1 − p2. First we claim that z1z2 = z3. We will prove this by calculation. z1z2 = = = [1, α3] [1, α2] [α, α3] [α, α2] e−3iθ − e3iθ e−2iθ − e2iθ [1, α3] α2 = [α, α2] α2 e−2iθ − e2iθ e−iθ − eiθ [1, α3] α2 [α2, α3] = z3 Next we claim that z1/z2 = p1 and z2/z1 = p2. These can also be calculated but a geometrical figure makes it clear. The first claim follows from the fact that the triangles 0−p1−z1 and 0−1−z2 are similar. The second claim follows from the similarity of the triangles 0−1−z1 and 0 − p2 − z2. z3 z2 z1 p3 0 p1 1 p2 So far we can construct the following pairwise products of elementary mono- mials. z2 1 = z1z2 z1 z2 = p1z3 z1z2 = z3 z2 2 = p2 2p1z3 = p2z3 13 We need only construct z2 3 = p2 First we show z2 formula given in Proposition 1.5 and obtained 3 and z2z3 since z1z3 = p1z2z3. 3(z3 − z2) algebraically. We calculated z2 3 using the z2 3 = 1 + 2α2 + 3α4 + 2α6 + α8 which is exactly what we found by calculating p2 equal. 3(z3 − z2), so the two must be Likewise, we calculated z2z3 to be z2z3 = 1 + 2α2 + 2α4 + α6 which precisely equals p3(1 − z3). elementary monomials, so R(U ) is a ring. Thus all 6 pairwise products of {z1, z2, z3} are Z[P ]-linear combinations of This characterization of R(U ) makes finding examples of rings R(U ) a mat- ter of verifying that finitely many products are contained in R(U ). However, finding counterexamples is more difficult. Some U that are difficult to work with, like {1, ei, e2i, e3i}, yield rings. Other U that are nicer to work with, such as {1, eiπ/5, eiπ/4, eiπ/3} are suspected to not yield rings. 5 Open Questions Some open questions we considered in research are posed below. 1. How does 1 /∈ U affect our current results? Can we still express R(U ) as a module over some ring generated by elementary monomials? 2. When exactly are the products of elementary monomials Z[P ]-linear com- binations of elementary monomials? 3. Is R({1, eiπ/5, eiπ/4, eiπ/3}) a ring? 4. What subrings of C are of the form R(U ) for some U ? 5. Given p ∈ C, for which U is p ∈ R(U )? [u,v] v + [v,q] 6. We can write Iu,v(p, q) = [u,p] [v,u] u where [x, y] = x¯y − y ¯x. Note that [x, y] is an alternating bilinear map. If V is some vector space equipped with [·,·], an alternating bilinear map into R and we have some S ⊆ V of allowable "angles", we can define I : S2 × V 2 → V via Iu,v(p, q) = [u, p] [u, v] v + [v, q] [v, u] u Do similar results hold for this generalization? Perhaps we could require V to be a normed vector space and say that S is the sphere of radius one. 14 References [1] Joe Buhler, Steve Butler, Warwick de Launey and Ron Graham, 'Origami rings', Journal of the Australian Mathematical Society 92 (June 2012), 299- 311. [2] D. Nedrenco, 'On origami rings', arXiv:1502.07995 (Feburary 2015). 15
1603.09709	1	1603	2016-03-31T18:44:04	Finite-dimensional algebras are (m>2)-Calabi-Yau-tilted	[ "math.RA", "math.RT" ]	We observe that over an algebraically closed field, any finite-dimensional algebra is the endomorphism algebra of an m-cluster-tilting object in a triangulated m-Calabi-Yau category, where m is any integer greater than 2.	math.RA	math	FINITE-DIMENSIONAL ALGEBRAS ARE (m > 2)-CALABI-YAU-TILTED SEFI LADKANI Abstract. We observe that over an algebraically closed field, any finite-dimensional algebra is the endomorphism algebra of an m-cluster-tilting object in a triangulated m-Calabi-Yau category, where m is any integer greater than 2. 6 1 0 2 r a M 1 3 ] . A R h t a m [ 1 v 9 0 7 9 0 . 3 0 6 1 : v i X r a 1. Introduction Cluster categories were introduced by Buan, Marsh, Reineke, Reiten and Todorov [8] as a means to model the combinatorics of cluster algebras with acyclic skew-symmetric exchange matrices within the framework of quiver representations (see also the work of Caldero, Chapoton and Schiffler [12] for the case of An quivers). The cluster category of an acyclic quiver is the orbit category of the bounded derived category of its path algebra with respect to the autoequivalence F = νΣ−2 where ν denotes the Serre functor and Σ is the suspension functor. By a result of Keller [20], the cluster category is a triangulated 2-Calabi-Yau category. Particular role is played by the 2-cluster-tilting objects within this category (the precise definitions will be given in Section 2.1 below) which model the clusters in the corresponding cluster algebra. As already shown in [20], by replacing Σ−2 by Σ−m for m > 2 and considering the orbit category with respect to the autoequivalence νΣ−m, one gets an m-Calabi-Yau triangulated category with m-cluster-tilting objects. The categories obtained in this way, called m-cluster categories, were the subject of many investigations, see [5, 26, 28, 29]. The endomorphism algebras of 2-cluster-tilting objects in cluster categories are known as cluster-tilted algebras and they possess many remarkable representation-theoretic and homological properties [4, 9, 10, 23]. More generally, consider an m-Calabi-Yau-tilted algebra, i.e. an algebra A = EndC(T ) where C is a K-linear, triangulated, Hom-finite, m-Calabi-Yau category over a field K and T is an m-cluster-tilting object in C for some positive integer m. Keller and Reiten have shown the following results in the case m = 2 (for the first two points, see Sections 2 and 3 of [23] and for the third one, see [24, §2]): • A is Gorenstein of dimension at most m − 1 (i.e. idA A ≤ m − 1 and pdA DA ≤ m − 1, where D(−) = HomK (−, K)); • The stable category of Cohen-Macaulay A-modules is (m + 1)-Calabi-Yau; • If K is algebraically closed, C is algebraic and A ∼= KQ for an acyclic quiver Q, then C is triangle equivalent to the m-cluster category of Q. In addition, they have shown that these results hold also in the case m > 2 provided that one imposes an additional condition on the m-cluster-tilting object T stated in This work was partially supported by the Center for Absorption in Science, Ministry of Aliyah and Immigrant Absorption, State of Israel. 1 2 SEFI LADKANI terms of the vanishing of some of its negative extensions, namely HomC(T, Σ−iT ) = 0 for any 0 < i < m − 1, (⋆) see [24, §4]. Note that by the m-Calabi-Yau property of C this condition is equivalent to the vanishing of the positive extensions HomC(T, ΣiT ) for all m < i < 2m − 1, whereas these extensions for 0 < i < m always vanish since T is m-cluster-tilting. In the terminology of [6], the condition (⋆) means that T is (m − 2)-corigid, and in [6, Theorem B] Beligiannis presents a more refined result connecting the corigidity property of an m-cluster-tilting object with the Gorenstein property of its endomorphism algebra and the Calabi-Yau property of its stable category of Cohen-Macaulay modules. We note also that a condition analogous to (⋆), stated for m-cluster-tilting subcategories inside bounded derived categories of modules and abbreviated "vosnex", appears in the works of Amiot-Oppermann [3, Definition 4.9] and Iyama-Oppermann [18, Notation 3.5]. Whereas for m = 2 the condition (⋆) is empty and hence automatically holds, this is no longer the case for m ≥ 3. In particular, there are examples of triangulated m-Calabi- Yau categories C and m-cluster-tilting objects T ∈ C for which the vosnex condition (⋆) does not hold and moreover the endomorphism algebra EndC(T ) is a path algebra of a connected acyclic quiver, see Iyama and Yoshino [19, Theorem 9.3] for m = 3 and [19, Theorem 10.2] for m odd, and also [24, Example 4.3]. Another example, where the algebra EndC(T ) is not Gorenstein, is given in [23, Example 5.3]. The purpose of this note is to extend this class of examples by showing that given m ≥ 3, any finite-dimensional algebra over an algebraically closed field is the endomorphism algebra of an m-cluster-tilting object in an m-Calabi-Yau triangulated category. Hence, without any further assumptions on C and T , one cannot say too much about the algebras EndC(T ). To this end we invoke the construction of generalized cluster categories due to Amiot [1] in the case m = 2 and generalized by Guo [17] to the case m > 2. This construction produces an m-Calabi-Yau triangulated category with an m-cluster-tilting object from any dg-algebra which is homologically smooth, bimodule (m + 1)-Calabi-Yau and sat- isfies additional finiteness conditions. A rich source of such dg-algebras is provided by the deformed Calabi-Yau completions defined and investigated by Keller [22]. Given a basic finite-dimensional algebra A, we choose a dg-algebra B with two prop- erties; firstly, the underlying graded algebra of B is the path algebra of a graded quiver whose arrows are concentrated in degrees 0 and −1, and secondly, H0(B) ∼= A. Such dg-algebra can be constructed from any presentation of A as a quotient of a path algebra of a quiver by an ideal generated by a finite sequence of elements. Conversely, any such dg-algebra arises in this way. It turns out that for any m ≥ 2, the (m+1)-Calabi-Yau completion of B is a Ginzburg dg-algebra Γ of a graded quiver with homogeneous superpotential of degree 2 − m which can be written explicitly in terms of the quiver and the sequence of elements. In the case m = 2, the zeroth homology H0(Γ) is a split extension of A, whereas when m > 2 it is isomorphic to A, hence Γ satisfies the finiteness conditions required in the construction of [1, 17] and thus gives rise to a Hom-finite m-Calabi-Yau triangulated category with an m-cluster-tilting object whose endomorphism algebra is isomorphic to A. The m-cluster-tilting object we get almost never satisfies the vosnex condition (⋆). More precisely, that condition holds if and only if A is the path algebra of an acyclic FINITE-DIMENSIONAL ALGEBRAS ARE (m > 2)-CALABI-YAU-TILTED 3 quiver and B is chosen such that B = A. Moreover, the flexibility in the choice of the dg-algebra B allows to construct, for certain algebras A and any m > 2, inequivalent triangulated m-Calabi-Yau categories C 6≃ C′ with m-cluster-tilting objects T ∈ C and T ′ ∈ C′ such that EndC(T ) ∼= EndC(T ′) ∼= A. 2. Recollections 2.1. Notations. We recall the definitions of Calabi-Yau triangulated categories and cluster-tilting objects. Throughout, we fix a field K. Definition 2.1. Let C be a K-linear triangulated category with suspension functor Σ. (a) We say that C is Hom-finite if the spaces HomC(X, Y ) are finite-dimensional over K for any X, Y ∈ C. (b) Let m ∈ Z. We say that C is m-Calabi-Yau if C is Hom-finite and there exist functorial isomorphisms HomC(X, Y ) ∼= D HomC(Y, ΣmX) for any X, Y ∈ C, where D denotes the duality D(−) = HomK(−, K). For an object X in an additive category C, denote by add X the full subcategory of C whose objects are finite direct sums of direct summands of X. For a subcategory Y of C, let ⊥Y = {X ∈ C : HomC(X, Y ) = 0 for any Y ∈ Y}. Definition 2.2. Let C be a triangulated m-Calabi-Yau category for some integer m ≥ 1. An object T of C is m-cluster-tilting if: (i) HomC(T, ΣiT ) = 0 for any 0 < i < m; and (ii) if X ∈ C is such that HomC(X, ΣiT ) = 0 for any 0 < i < m, then X ∈ add T . Equivalently, add T =T0<i<m ⊥Σi(add T ). Definition 2.3. Let m ≥ 1. A K-algebra Λ is m-Calabi-Yau-tilted (m-CY-tilted for short) if there exist a triangulated m-Calabi-Yau category C and an m-cluster-tilting object T of C such that Λ ∼= EndC(T ). 2.2. Generalized cluster categories. In this section we briefly review the con- struction of triangulated m-Calabi-Yau categories with m-cluster-tilting object due to Amiot [1] (for the case m = 2) and its generalization by Guo [17] (for the case m > 2). Let Γ be a differential graded (dg) algebra over a field K. Denote by D(Γ) its derived category and by per Γ its smallest full triangulated subcategory containing Γ and closed under taking direct summands. Let Dfd(Γ) denote the full subcategory of D(Γ) whose objects are those of D(Γ) with finite-dimensional total homology. Definition 2.4. Γ is said to be homologically smooth if Γ ∈ per Γe, where Γe = Γop ⊗K Γ. Let Γ be homologically smooth and let Ω = RHomΓe(Γ, Γe). By definition, Ω ∈ D((Γe)op). We can view Ω as an object of D(Γe) via restriction of scalars along the morphism τ : Γe ∼−→ (Γe)op given by τ (x ⊗ y) = y ⊗ x. By [22, Lemma 3.4] (see also [21, Lemma 4.1] and [17, Lemma 2.1]) one has Dfd(Γ) ⊆ per Γ and (2.1) HomD(Γ)(L ⊗Γ Ω, M ) ∼= D HomD(Γ)(M, L) for any L ∈ D(Γ), M ∈ Dfd(Γ). 4 SEFI LADKANI Definition 2.5. Let m ∈ Z. bimodule m-Calabi-Yau. If RHomΓe(Γ, Γe) ∼= Σ−mΓ in D(Γe) we say that Γ is If Γ is homologically smooth and bimodule m-Calabi-Yau then the triangulated cat- egory Dfd(Γ) is m-Calabi-Yau. More precisely, (2.1) yields functorial isomorphisms HomD(Γ)(Σ−mL, M ) ∼= D HomD(Γ)(M, L) for any L ∈ D(Γ), M ∈ Dfd(Γ). Theorem 2.6 ([1, §2], [17, §2]). Let m ≥ 1 and let Γ be a dg-algebra satisfying the following conditions: (i) Γ is homologically smooth; (ii) Hi(Γ) = 0 for any i > 0; (iii) dimK H0(Γ) < ∞; (iv) Γ is bimodule (m + 1)-Calabi-Yau. Consider the triangulated category CΓ = per Γ/Dfd(Γ). Then: (a) CΓ is Hom-finite and m-Calabi-Yau. (b) For any i ∈ Z, set D≤i = {X ∈ D(Γ) : Hp(X) = 0 for all p > i} and let F = D≤0 ∩ ⊥D≤−m ∩ per Γ. The restriction of the canonical projection π : per Γ → CΓ to F induces an equivalence of K-linear categories F (c) The image πΓ of Γ in CΓ is an m-cluster-tilting object and EndCΓ (πΓ) ∼= H0(Γ). ∼ −→ CΓ. The object πΓ occurring in part (c) of the theorem is called the canonical m-cluster- tilting object in CΓ. The statement of the next lemma concerning negative extension groups of the canonical m-cluster-tilting object is implicit in the proof of [17, Corol- lary 3.4]. Lemma 2.7. Let Γ be a dg-algebra satisfying the conditions of Theorem 2.6 and let T = πΓ be the canonical m-cluster-tilting object in CΓ. Then HomC(T, Σ−iT ) ∼= H−i(Γ) for any 0 ≤ i ≤ m − 1. Proof. As observed in [17], the objects Γ, ΣΓ, . . . , Σm−1Γ are in the fundamental domain F. Therefore, for any 0 ≤ i ≤ m − 1, HomCΓ(πΓ, Σ−iπΓ) ∼= HomCΓ(ΣiπΓ, πΓ) = HomCΓ(πΣiΓ, πΓ) ∼= HomD(Γ)(ΣiΓ, Γ) ∼= HomD(Γ)(Γ, Σ−iΓ) ∼= H−i(Γ). (cid:3) 2.3. Ginzburg dg-algebras. Ginzburg dg-algebras were introduced by Ginzburg in [16]. We recall their definition in the case of a graded quiver with homogeneous superpotential and quote the result of Keller [22] that they are homologically smooth and bimodule Calabi-Yau (see also the paper [27] by Van den Bergh). We note that a graded version of quivers with potentials (and their mutations) has also been introduced in [2] and [13], however in these papers one still implicitly considers Ginzburg dg-algebras which are bimodule 3-Calabi-Yau. In contrast, in the setting described below the degree of the superpotential affects the Calabi-Yau dimension of its Ginzburg dg-algebra. A quiver is a finite directed graph. More precisely, it is a quadruple Q = (Q0, Q1, s, t), where Q0 and Q1 are finite sets (of vertices and arrows, respectively) and s, t : Q1 → Q0 FINITE-DIMENSIONAL ALGEBRAS ARE (m > 2)-CALABI-YAU-TILTED 5 are functions specifying for each arrow its starting and terminating vertex, respectively. A quiver Q is graded if we are given a grading · : Q1 → Z. A path p in Q is a sequence of arrows α1α2 . . . αn such that s(αi+1) = t(αi) for all 1 ≤ i < n. For a path p we denote by s(p) its starting vertex s(α1) and by t(p) its terminating vertex t(αn). A path p is a cycle if it starts and ends at the same vertex, i.e. s(p) = t(p). Any vertex i ∈ Q0 gives rise to a cycle ei of length zero with s(ei) = t(ei) = i. The path algebra KQ has a basis consisting of the paths of Q, and the product of two paths p and q is their concatenation pq if s(q) = t(p) and zero otherwise. The path algebra is graded if Q is graded, with the degree of a path being the sum of the degrees of its arrows. The degree of a homogeneous element x of KQ will be denoted by x. Consider the K-bilinear map [−, −] : KQ × KQ → KQ whose value on a pair of homogeneous elements x, y is given by their supercommutator [x, y] = xy − (−1)xyyx. Denote by [KQ, KQ] the linear subspace of KQ spanned by all the supercommutators. The quotient KQ/[KQ, KQ] has a basis consisting of cycles considered up to cyclic permutation "with signs". Definition 2.8. A superpotential on Q is a homogeneous element in KQ/[KQ, KQ]. Any arrow α ∈ Q1 gives rise to a linear map ∂α : KQ/[KQ, KQ] → KQ (called cyclic derivative with respect to α) whose value on any cycle p is given by (2.2) ∂αp = (−1)α Xp=uαv (−1)u(α+v)vu where the sum runs over all possible decompositions p = uαv with u, v paths of length ≥ 0. More explicitly, if p = α1α2 . . . αn has degree w, then (−1)(w−1)(α1+···+αℓ−1)αℓ+1 . . . αnα1 . . . αℓ−1 ∂α(α1α2 . . . αn) = (−1)α Xℓ : αℓ=α is homogeneous of degree w − α. Definition 2.9. Let Q be a graded quiver and let m ∈ Z. Let Q be the graded quiver whose vertices are those of Q and its set of arrows consists of • the arrows of Q (with their degree unchanged); • an arrow α∗ : j → i of degree 1 − m − α for each arrow α : i → j of Q; • a loop ti : i → i of degree −m for each vertex i ∈ Q0. Let W be a superpotential on Q of degree 2 − m. The Ginzburg dg-algebra of (Q, W ), denoted Γm+1(Q, W ), is the dg-algebra whose underlying graded algebra is the path algebra KQ and the differential d is defined by its action on the generators as • d(α) = 0 and d(α∗) = ∂αW for each α ∈ Q1; • d(ti) = ei(Pα∈Q1[α, α∗])ei for each i ∈ Q0. Remark 2.10. Note that for each α ∈ Q1 the element ∂αW is homogeneous of degree 2 − m − α and the supercommutator [α, α∗] is homogeneous of degree 1 − m, hence the definition of the differential d makes sense. Moreover, as [α, α∗] = αα∗ − (−1)αα∗α∗α for any α ∈ Q1, by using the sign conventions in (2.2) one verifies that d2(ti) = 0 for any i ∈ Q0, so that d is indeed a differential. Note also that in [22] the differential of ti is (−1)m−1 times the one given here, but of course by replacing each ti by (−1)m−1ti one sees that the dg-algebras are isomorphic. 6 SEFI LADKANI Theorem 2.11 ([22, Theorem 6.3]). Let m ∈ Z and let W be a superpotential of degree 2 − m on a graded quiver Q. Then Γm+1(Q, W ) is homologically smooth and bimodule (m + 1)-Calabi-Yau. We record two useful observations. Let Q be a graded quiver, m ∈ Z an integer and W a homogeneous superpotential on Q of degree 2 − m. Lemma 2.12. Suppose that α : i → j is an arrow in Q such that no term of W contains α. Define a graded quiver Q′ by Q′ 1 = Q1 \ {α} ∪ {α∗} where α∗ : j → i has degree 1 − m − α. Then W can be naturally viewed as a superpotential W ′ on Q′ and Γm+1(Q, W ) ∼= Γm+1(Q′, W ′). 0 = Q0 and Q′ Proof. Since no term of W contains α, we can view W as an element W ′ in the path algebra KQ′. The graded quivers Q′ and Q are isomorphic by the map ϕ : Q′ → Q sending (α∗)∗ to α and fixing all other arrows. Moreover, since ∂αW = 0 = ∂α∗ W ′ and ∂βW = ∂βW ′ for any β ∈ Q1 ∩ Q′ 1, the map ϕ induces an isomorphism of the Ginzburg dg-algebras Γm+1(Q′, W ′) ∼= Γm+1(Q, W ). (cid:3) For a subset of arrows Ω ⊆ Q1, let QΩ be the subquiver of Q with (QΩ)0 = Q0 and (QΩ)1 = Ω. For the definition of Calabi-Yau completion, see [22, §4]. Lemma 2.13. Suppose that Ω ⊆ Q1 is a set of arrows such that W = Pβ∈Q1\Ω βωβ with ωβ ∈ KQΩ for each β 6∈ Ω. Consider the subquiver Q′ of Q defined by Q′ 1 = Ω ∪ {β∗ : β ∈ Q1 \ Ω}. Let d be the differential on Γm+1(Q, W ). Then: and Q′ (a) d(KQ′) ⊆ KQ′, hence B = (KQ′, d) is a sub-dg-algebra of Γm+1(Q, W ). (b) Γm+1(Q, W ) is isomorphic to the (m + 1)-Calabi-Yau completion of B. 0 = Q0 Proof. The first claim holds since d(α) = 0 for any α ∈ Ω ⊆ Q1 and d(β∗) = ∂βW = (−1)βωβ ∈ KQΩ ⊆ KQ′ for any β 6∈ Ω. This argument also shows that the differential d on KQ′ satisfies the condition in [22, §3.6] via the filtration ∅ ⊆ Ω ⊆ Q′ of the set of arrows. Hence we can use [22, Proposition 6.6] to compute the (m + 1)-Calabi-Yau completion of B and get that it is isomorphic to (KQ′, d′) with the differential d′ given on the generators by d′(α) = ∂α∗W ′ = 0, d′((β∗)∗) = ∂β ∗ W ′ = 0, d′(α∗) = ∂αW ′, d′(β∗) = ∂(β ∗)∗W ′ [γ, γ∗])ei, where W ′ ∈ KQ′ is the element (−1)β ∗(β∗)∗d(β∗) = Xβ∈Q1\Ω (−1)m−1(β∗)∗ωβ. Finally, the isomorphism ϕ : KQ′ → KQ defined on the generators by and d(ti) = (−1)m+1ei(Pγ∈Q′ 1 W ′ =Xα∈Ω (−1)αα∗d(α) + Xβ∈Q1\Ω ϕ(γ) = (−1)m−1γ β γ if γ = α∗ for some α ∈ Q1, if γ = (β∗)∗ for some β ∈ Q1 \ Ω, otherwise induces an isomorphism (KQ′, d′) ∼= Γm+1(Q, W ) of dg-algebras. (cid:3) FINITE-DIMENSIONAL ALGEBRAS ARE (m > 2)-CALABI-YAU-TILTED 7 2.4. Non-positively graded Ginzburg dg-algebras. In this section we restrict at- tention to Ginzburg dg-algebras which are concentrated in non-positive degrees and quote the construction of Guo [17] (generalizing that of Amiot in [1] for the case m = 2) of the generalized cluster category associated to a quiver with superpotential. of the arrows of degree w, in other words, Q(w) Let Q be a graded quiver and let w ∈ Z. Denote by Q(w) the subquiver of Q consisting 1 = {α ∈ Q1 : α = w}. Now fix a graded quiver Q and let m ∈ Z. Let Q denote the quiver constructed in Definition 2.9. Fix a homogeneous superpotential W on Q of degree 2 − m and let Γ = Γm+1(Q, W ) be the Ginzburg dg-algebra. We immediately observe: 0 = Q0 and Q(w) Remark 2.14. Γi = 0 for any i > 0 if and only if all the arrows of Q have non-positive degrees. This condition implies that Hi(Γ) = 0 for any i > 0. Lemma 2.15. Assume that all the arrows of Q have non-positive degrees. Then H0(Γ) ∼= KQ /(dα : α ∈ Q (0) (−1) 1 ). Proof. Since there are no arrows of Q of positive degree, the graded piece Γ0 equals the and the graded piece Γ−1 is spanned by the elements of the form path algebra of Q (0) uαv where u, v ∈ KQ (cid:3) . As d(uαv) = u(dα)v, the claim follows. (−1) and α ∈ Q 1 (0) Lemma 2.16. The following conditions are equivalent: (a) Each arrow of Q has non-positive degree; (b) m ≥ 0 and for any α ∈ Q1 one has 1 − m ≤ α and α ≤ 0. Proof. If m < 0 then the each of the loops ti in Q has positive degree −m. In addition, each other arrow of Q is either α or α∗ for some arrow α ∈ Q1, and its degree is α or 1 − m − α, respectively. These observations imply the statement of the lemma. (cid:3) In the next two examples we discuss the cases m = 0 and m = 1. Example 2.17. Assume that m = 0. Lemma 2.16 implies that the arrows of Q have non-positive degrees if and only if Q has no arrows. In this case Q is a disjoint union of graded quivers of the form • t with t = 0 and dt = 0. Hence Γ is concentrated in degree 0 and it is a finite direct product of polynomial rings k[t]. The algebra k[t] is 1-Calabi-Yau, see [21, §4.2] Example 2.18. Assume that m = 1. Lemma 2.16 implies that the arrows of Q have non-positive degrees if and only if all the arrows of Q are in degree 0, so we can regard Q as an ungraded quiver. Since the superpotential on Q is homogeneous of degree 1, In this case all the arrows α and α∗ of Q are in degree 0 and the it must vanish. differential of each loop ti, whose degree is −1, is given by d(ti) = ei(Pα∈Q1[α, α∗])ei. By Lemma 2.15, H0(Γ) is isomorphic to the preprojective algebra of the quiver Q. When m ≥ 2, the next lemma shows that we may assume that Q(1−m) has no arrows. Lemma 2.19. Assume that m ≥ 2 and that 1 − m ≤ α ≤ 0 for any α ∈ Q1. Then there exist a graded quiver Q′ with 2 − m ≤ α′ ≤ 0 for each α′ ∈ Q′ 1 and a homogeneous superpotential W ′ on Q′ of degree 2 − m such that Γm+1(Q, W ) ∼= Γm+1(Q′, W ′). 8 SEFI LADKANI Proof. We define the graded quiver Q′ as a subquiver of Q; we set Q′ 0 = Q0 and 1 =(Q(0) Q(w) 1 Q′(w) 1 ∪ {α∗ : α ∈ Q(1−m) 1 } if w = 0, if 1 − m < w < 0, 1 with Q′(w) empty for any other w ∈ Z. Then 2 − m ≤ α′ ≤ 0 for any arrow α′ ∈ Q′ 1 by construction. Moreover, since the superpotential W is of degree 2 − m and all the arrows of Q have non-positive degrees, no term of W can contain any arrows of Q(1−m). The result now follows by iterated application of Lemma 2.12 for each of the arrows in Q(1−m). (cid:3) In particular, when m = 2 one can always reduce to the classical setting of an ungraded quiver with potential. Corollary 2.20. Let Q be a graded quiver such that −1 ≤ α ≤ 0 for any α ∈ Q1 and let W be a homogeneous superpotential on Q of degree 0. Then there exist a quiver Q′ concentrated in degree 0 and a superpotential W ′ on Q′ such that Γ3(Q, W ) ∼= Γ3(Q′, W ′). The next lemma generalizes [25, Lemma 2.11]. Lemma 2.21. Assume that m ≥ 2 and 2 − m ≤ α ≤ 0 for any α ∈ Q1. Then H0(Γm+1(Q, W )) ∼= KQ(0)/(∂αW : α ∈ Q(2−m) 1 ). Proof. Observe that since the arrows of Q have non-positive degrees and W is of degree 2 − m, the cyclic derivative ∂αW with respect to any arrow α of degree 2 − m lies in the path algebra of Q(0) and the quotient in the right hand side makes sense. The arrows of Q have non-positive degrees by Lemma 2.16, hence we may apply ). Observe that 1 ∪{α∗ : α ∈ Q(2−m) } (cid:3) Lemma 2.15 to deduce that H0(Γm+1(Q, W )) ∼= KQ Q with d(α) = 0 and d(α∗) = ∂αW for any α ∈ Q1, hence the claim follows. /(dα : α ∈ Q (−1) 1 = Q(−1) is empty by assumption and Q = Q(0) since Q(1−m) (−1) 1 (0) (0) 1 1 Theorem 2.22 ([17, Theorem 3.3]). Let m ≥ 1 be an integer, let Q be a graded quiver, let W be a superpotential on Q of degree 2 − m and denote by Γ = Γm+1(Q, W ) the Ginzburg dg-algebra. Assume that: (i) 1 − m ≤ α ≤ 0 for all α ∈ Q1; (ii) H0(Γ) is finite-dimensional. Then the triangulated category C(Q,W ) = per Γ/Dfd(Γ) is Hom-finite and m-Calabi- Yau. Moreover, the image of Γ in C(Q,W ) is an m-cluster-tilting object whose endomor- phism algebra is isomorphic to H0(Γ). Proof. The condition (i) implies that Hi(Γ) = 0 for all i > 0. Now Theorem 2.11 and condition (ii) imply that Γ satisfies the conditions of Theorem 2.6 and the result is now a consequence of that theorem. (cid:3) 3. The construction 3.1. Superpotentials from quivers with relations. In this section we construct, given a quiver Q, a finite sequence R of relations on Q and an integer m ≥ 2, a graded quiver with homogeneous superpotential of degree 2 − m. The construction generalizes FINITE-DIMENSIONAL ALGEBRAS ARE (m > 2)-CALABI-YAU-TILTED 9 that of Keller in [22, §6.9] for the case where m = 2 and KQ/(R) has global dimension 2. The idea of adding, for each relation, an arrow in the opposite direction appears already in the description of relation-extension algebras by Assem, Brustle and Schiffler [4]. Definition 3.1. Let Q be a quiver and let r be the ideal of KQ generated by all the arrows. A relation on Q is an element of eirej for some i, j ∈ Q0. In other words, a relation is a linear combination of paths of positive lengths starting at i and ending at j. We start with some preparations concerning split extensions of algebras which will be needed for the case m = 2. Definition 3.2. An algebra eA is a split extension of an algebra A if there exist algebra homomorphisms ι : A → eA and π : eA → A such that πι = idA. Let Q be a quiver and let eQ be a quiver such that eQ0 = Q0 and Q1 ⊆ eQ1 (in other words, eQ is obtained from Q by adding arrows). Then the path algebra KeQ is a split extension of the path algebra KQ. Indeed, there are algebra homomorphisms π eQ,Q−−−→ KQ KQ ιQ, eQ−−−→ KeQ π eQ,Q(α) =(α if α ∈ Q1, 0 if α ∈ eQ1 \ Q1. whose values on the generators are given by ιQ, eQ(α) = α (α ∈ Q1) and the algebra KQ/(R). Denote by r′ the ideal of KeQ generated by the arrows in the set eQ1 \ Q1. Let R be a set of relations in KQ and let eR be a set of relations in KeQ such that R ⊆ eR (via the natural embedding ιQ, eQ : KQ ֒→ KeQ). Lemma 3.3. Assume that eR \ R ⊆ r′. Then the algebra KeQ/(eR) is a split extension of ιQ, eQ−−−→ KeQ ։ KeQ/(eR). Since R ⊆ eR, the image π eQ,Q−−−→ KQ ։ KQ/(R) and let ρ ∈ eR. KeQ/(eR). Consider now the composition KeQ that eR \ R ⊆ r′ implies that π eQ,Q(ρ) = 0, so its image vanishes as well. Hence we get a well defined algebra homomorphism π : KeQ/(eR) → KQ/R. The composition πι maps the image of any arrow α ∈ Q in KQ/(R) to itself, therefore it is the identity on KQ/(R). (cid:3) Proof. Consider the composition KQ of any relation ρ ∈ R vanishes and we get an algebra homomorphism ι : KQ/(R) → If ρ ∈ R, then its image obviously vanishes. Otherwise, ρ 6∈ R and our assumption Any quiver with a finite sequence of relations gives rise to a dg-algebra whose under- lying graded algebra is the path algebra of a graded quiver with arrows concentrated in degrees 0 and −1. The details are given in the construction below. Construction 3.4. Let (Q, R) be a pair where Q is a quiver and R = Si,j∈Q0 Ri,j, where each Ri,j is a finite sequence of relations inside eirej and R is the concatenation of these sequences (there may be repetitions inside each sequence Ri,j and moreover the zero element can appear inside several such sequences). For a relation ρ ∈ Ri,j, set s(ρ) = i and t(ρ) = j. 10 SEFI LADKANI We define a graded quiver Q′ as follows: • The set of vertices of Q′ equals that of Q; • The set of arrows consists of -- the arrows of Q, with their degree set to 0; -- an arrow ηρ : s(ρ) → t(ρ) of degree −1 for each relation ρ ∈ R; and denote by B(Q, R) the dg-algebra whose underlying graded algebra is the path algebra KQ′ with the differential acting on the generators by • d(α) = 0 for any α ∈ Q1; • d(ηρ) = ρ for any ρ ∈ R. Obviously, B(Q, R) is concentrated in non-positive degrees and H0(B(Q, R)) is isomor- phic to KQ/(R), but in general B(Q, R) is not quasi-isomorphic to its zeroth homology. Remark 3.5. If B = (KQ′, d) is any dg-algebra whose underlying graded algebra is the path algebra of a graded quiver with arrows concentrated in degrees 0 and −1 and the image of the differential d lies in the ideal generated by the arrows, then B ∼= B(Q, R) for some (Q, R). Indeed, we can take Q = Q′(0) and for each i, j ∈ Q0 let Ri,j be the list of d(α) where α runs over the arrows in Q′(−1) starting at i and ending at j. 1 It turns out (see Lemma 3.10 below) that for any m ≥ 2, the (m + 1)-Calabi-Yau completion of B(Q, R) is a Ginzburg dg-algebra of a graded quiver with homogeneous superpotential of degree 2 − m whose construction is described below. Construction 3.6. Let (Q, R, m) be a triple where (Q, R) is as in Construction 3.4 and m ≥ 2 is an integer. We construct a graded quiver eQ with homogeneous superpotential W of degree 2 − m as follows. -- the arrows of Q, with their degree set to 0; -- an arrow ερ : t(ρ) → s(ρ) of degree 2 − m for each relation ρ ∈ R; of the sequence R and not only on the two-sided ideal (R) it generates in KQ, as we shall see in Example 3.30. • The set of vertices of eQ equals that of Q; • The set of arrows of eQ consists of • The superpotential W is the image of the elementPρ∈R ερρ in KeQ/[KeQ, KeQ]. We denote by Γ(Q, R, m) the Ginzburg dg-algebra Γm+1(eQ, W ). Remark 3.7. The quiver with superpotential (eQ, W ) depends on the particular choice sequence of relations on Q and m ≥ 2. We denote by (eQ, W ) the graded quiver with Γ = Γ(Q, R, m) = Γm+1(eQ, W ) its Ginzburg dg-algebra. We denote the elements of R degree 2 − m in eQ corresponding to the relation ρk, so that W is the image ofPn modulo [KeQ, KeQ]. Similarly, denote by ηk the arrow ηρk and let B = B(Q, R) be the We start by describing the graded quiver eQ underlying Γ occurring in Definition 2.9. by ρ1, ρ2, . . . , ρn and write R = n. For simplicity, we denote by εk the arrow ερk of k=1 εkρk superpotential of degree 2 − m associated to (Q, R, m) as in Construction 3.6, and by For the rest of this section, we fix a triple (Q, R, m) where Q is a quiver, R is a finite dg-algebra of Construction 3.4. FINITE-DIMENSIONAL ALGEBRAS ARE (m > 2)-CALABI-YAU-TILTED 11 arrow degree differential α ε∗ k εk α∗ ti 0 −1 2 − m 1 − m −m 0 (−1)mρk 0 ∂αW d(ti) (α ∈ Q1) (1 ≤ k ≤ n) (1 ≤ k ≤ n) (α ∈ Q1) (i ∈ Q0) Table 1. The arrows of the graded quiver eQ and their differentials. are as given in Table 1. Lemma 3.8. The arrows of the graded quiver eQ, their degrees and their differentials Proof. The description of the arrows and their degrees is evident from Definition 2.9. For the differentials, note that since none of the arrows εk occur in any ρ ∈ R, we have ∂εk W = ∂εk(cid:16) nXℓ=1 for any 1 ≤ k ≤ n. εℓρℓ(cid:17) = ∂εk (εkρk) = (−1)εkρk = (−1)mρk (cid:3) 1 ). Lemma 3.9. Consider the algebras A = KQ/(R) and eA = H0(Γ). (a) If m > 2 then eA ∼= A. (b) If m = 2 then eA is a split extension of A. Proof. By Lemma 2.21, eA ∼= KeQ(0)/(∂αW : α ∈ eQ(2−m) If m > 2 then eQ(0) = Q, the arrows of eQ(2−m) are ε1, ε2, . . . , εn and ∂εk W = (−1)mρk according to Table 1. Hence eA ∼= KQ/(ρ1, ρ2, . . . , ρn). This shows part (a). If m = 2 then all the arrows of the quiver eQ have degree 0, and we can think of eQ as an words, eQ = eQ(0) and Q ⊆ eQ with eQ \ Q = {ε1, ε2, . . . , εn}. Moreover, eA = KeQ/(eR) for eR = {∂αW : α ∈ eQ1}. Now R ⊆ eR since ∂εk W = ρk for each 1 ≤ k ≤ n. addition, for any α ∈ Q1 the element ∂αW =Pn by ε1, ε2, . . . , εn in eQ. Part (b) now follows from Lemma 3.3. In k=1 ∂α(εkρk) lies in the ideal generated (cid:3) ungraded quiver consisting of the arrows of Q and the arrows εk for 1 ≤ k ≤ n. In other The next two lemmas relate the dg-algebra B and the Ginzburg dg-algebra Γ. For a similar result in the case m = 2, see [22, §6.7 and Proposition 6.8]. Lemma 3.10. Γ is the (m + 1)-Calabi-Yau completion of B. Proof. This is a consequence of Lemma 2.13. eQ1 \ Ω = {ε1, . . . , εn} and the superpotential W = Pn Indeed, we may take Ω = Q1 so that k=1 εkρk has the required form. We only note that the differential of Γ restricts to the path algebra of the quiver with arrows Q1 ∪ {ε∗ n} and the resulting dg-algebra is isomorphic to B by mapping each arrow ε∗ (cid:3) Lemma 3.11. H−i(Γ) ∼= H−i(B) for any i < m − 2. k to (−1)mηρk and sending each α ∈ Q1 to itself. 1, . . . , ε∗ 12 SEFI LADKANI Proof. From Table 1 we see that Γ−i ∼= B−i for any i < m − 2, hence H−i(Γ) ∼= H−i(B) for any i < m − 3. The graded piece Γ2−m can be decomposed as Γ2−m ∼= B2−m ⊕ F , where the space F is spanned by the elements of the form uεkv where 1 ≤ k ≤ n and u, v are paths in Q. Since the differential of each such element vanishes, one has d(F ) = 0, hence d(Γ2−m) ∼= d(B2−m) and therefore H3−m(Γ) ∼= H3−m(B) as well. (cid:3) Lemma 3.12. If R is empty then H−i(Γ) = 0 for any 1 ≤ i ≤ m − 2. Proof. If R is empty, then by Lemma 3.8 the arrows in the graded quiver eQ have degrees 0, 1 − m or −m. Hence the graded piece Γ−i vanishes for each 1 ≤ i ≤ m − 2 and the claim follows. (cid:3) The next lemma provides a partial converse to Lemma 3.12. Lemma 3.13. If m > 2 and R ⊆ r2, then dimK H2−m(Γ) ≥ R. Proof. The graded piece Γ2−m is spanned by two types of elements: (1) u1ε∗ k1 u2ε∗ k2 u3 . . . ε∗ km−2 um−1, where u1, u2, . . . , um−1 are paths in Q and 1 ≤ kj ≤ n for each 1 ≤ j ≤ m − 2; (2) uεkv, where u, v are paths in Q and 1 ≤ k ≤ n; As a K-vector space, we may thus decompose Γ2−m into a direct sum E ⊕ E′, where E is the n-dimensional subspace E = {λ1ε1 + λ2ε2 + · · · + λnεn : (λ1, λ2, . . . , λn) ∈ K n} and E′ is spanned by all the elements of type (1) and those of type (2) such that at least one of u, v has positive length. We will show that d(Γ1−m) ⊆ E′ and hence no non-zero element in E lies in the image of the differential d acting on the graded piece Γ1−m. Since d vanishes on E, this will yield an n-dimensional subspace inside H2−m(Γ). If m > 3, the graded piece Γ1−m is spanned by three types of elements: (1) u1ε∗ km−1 um, where u1, u2, . . . , um are paths in Q and 1 ≤ kj ≤ n k2u3 . . . ε∗ k1u2ε∗ for each 1 ≤ j ≤ m − 1; (2) uεkvε∗ (3) uα∗v, where u, v are paths in Q and α ∈ Q1. l w and uε∗ kvεlw, where u, v, w are paths in Q and 1 ≤ k, l ≤ n; If m = 3, in addition to the elements above there is a fourth type (4) uεkvεlw, where u, v, w are paths in Q and 1 ≤ k, l ≤ n. It suffices to prove that the differential of any of these elements belongs to E′. This is clear for the elements of the type (1) since d(u1ε∗ k1u2ε∗ k2u3 . . . ε∗ km−1um) = ±u1ε∗ k1u2 . . . ε∗ kj−1ujρjuj+1ε∗ kj+1 . . . ε∗ km−1um m−1Xj=1 and none of the arrows ε1, ε2, . . . , εn can appear in the right hand side. This is also clear for the elements of type (4) since their differential vanishes. Consider an element of type (2). Then d(uεkvε∗ l w) = uεkvρlw, hence the differential is spanned by elements of the form u′εkv′ where u′, v′ are paths in Q and v′ has positive length. The case of d(uε∗ Consider an element of type (3). Then d(uα∗v) = u(∂αW )v =Pn k=1 u∂α(εkρk)v. Our assumption that R ⊆ r2 implies that each ∂α(εkρk) is a linear combination of terms u′′εkv′′ where at least one of the paths u′′, v′′ has positive length. (cid:3) kvεlw) is similar. FINITE-DIMENSIONAL ALGEBRAS ARE (m > 2)-CALABI-YAU-TILTED 13 Corollary 3.14. If m > 2 and R ⊆ r2, then H2−m(Γ) = 0 if and only if R is empty. Consider the triangulated category C(Q,R,m) = per Γ(Q, R, m)/Dfd(Γ(Q, R, m)). Theorem 3.15. Let Q be a quiver, let R be a finite sequence of relations on Q such that the algebra A = KQ/(R) is finite-dimensional and let m > 2 be an integer. Then: (a) The category C = C(Q,R,m) is Hom-finite and m-Calabi-Yau. (b) The image T of Γ(Q, R, m) in C is an m-cluster-tilting object with EndC(T ) ∼= A. (c) If R ⊆ r2 then dimK HomC(T, Σ−(m−2)T ) ≥ R. superpotential of degree 2 − m associated to the triple (Q, R, m) as in Construction 3.6. Proof. Recall that Γ(Q, R, m) = Γm+1(eQ, W ) where (eQ, W ) is the graded quiver with We claim that (eQ, W ) satisfies the conditions of Theorem 2.22. Indeed, condition (i) holds since the degree of any arrow in eQ is either 0 or 2 − m, and condition (ii) holds since by Lemma 3.9, H0(Γ) ∼= A and the algebra A is assumed to be finite-dimensional. Hence, by Theorem 2.22, the triangulated category C = C(Q,R,m) = C( eQ,W ) is Hom- finite, m-Calabi-Yau and the image T of Γ = Γ(Q, R, m) under the canonical projection per Γ → C is an m-cluster-tilting object whose endomorphism algebra is EndC(T ) ∼= H0(Γ) ∼= A. The last assertion follows from Lemma 2.7 and Lemma 3.13. (cid:3) Remark 3.16. The m-Calabi-Yau category C of Theorem 3.15 depends on Q and R and not only on the algebra A. There exist quivers Q and sequences of relations R and R′ on Q such that the algebras KQ/(R) and KQ/(R′) are finite-dimensional and isomorphic but for any m > 2 the categories C(Q,R,m) and C(Q,R′,m) are not equivalent, see Example 3.30 below. Remark 3.17. Keep the notations and assumptions of Theorem 3.15 and write A as A = ⊕i∈Q0Pi where Pi = eiA are the indecomposable projective right A-modules. The decomposition Γ = ⊕i∈Q0eiΓ for Γ = Γ(Q, R, m) induces a decomposition T = ⊕i∈Q0Ti with Ti being the image of eiΓ under the canonical projection to C. Hence for any finitely generated projective A-module P = ⊕i∈Q0P ei such that ei > 0 for all i ∈ Q0 there exists i an m-cluster-tilting object TP = ⊕i∈Q0T ei i in C with EndC(TP ) ∼= EndA(P ). The next result shows that the m-cluster category of any acyclic quiver can be realized as a category of the form C(Q, R, m). Recall that a quiver is acyclic if it has no cycles of positive length. We refer to [17, Corollary 3.4] for a related result. Proposition 3.18. Let Q be a quiver, let R be a finite sequence of relations on Q such that R ⊆ r2 and the algebra KQ/(R) is finite-dimensional, and let m > 2 be an integer. Then the following conditions are equivalent, where C denotes the m-Calabi-Yau category C(Q,R,m) and T is the canonical m-cluster-tilting object in C. (a) The quiver Q is acyclic and the sequence R is empty; (b) The dg-algebra B(Q, R) is concentrated in degree 0 and has finite total dimension; (c) HomC(T, Σ−iT ) = 0 for any 0 < i < m − 1; (d) HomC(T, Σ−(m−2)T ) = 0. Moreover, if any of these equivalent conditions holds and the field K is algebraically closed, then C is triangle equivalent to the m-cluster category of Q. 14 SEFI LADKANI Proof. The equivalence of (a) and (b) is clear. Let Γ = Γ(Q, R, m). For the implication (a) ⇒ (c), note that if Q is acyclic and R is empty then HomC(T, Σ−iT ) ∼= H−i(Γ) = 0 for any 0 < i < m − 1 by Lemma 2.7 and Lemma 3.12. The implication (c) ⇒ (d) is clear. For the implication (d) ⇒ (a), note that by Lemma 2.7 and Lemma 3.13 dimK HomC(T, Σ−(m−2)T ) = dimK H2−m(Γ) ≥ R, hence if HomC(T, Σ−(m−2)T ) vanishes R must be empty and then Q is acyclic by our assumption that KQ/(R) is finite-dimensional. If any of these conditions holds, then C is an algebraic Hom-finite, m-Calabi-Yau triangulated category with an m-cluster-tilting object T such that EndC(T ) ∼= KQ for an acyclic quiver Q and HomC(T, Σ−iT ) = 0 for any 0 < i < m − 1. By the characterization of higher cluster categories of Keller and Reiten [24, Theorem 4.2]) if K is algebraically closed, then C is triangle equivalent to the m-cluster-category of the quiver Q. (cid:3) 3.2. Systems of relations. Let Q be a quiver and let r be the two-sided ideal of KQ generated by the arrows of Q. An ideal I of KQ is admissible if there exists some N ≥ 2 such that rN ⊆ I ⊆ r2. Definition 3.19 ([7]). Let I be an ideal of KQ. A system of relations for I is a set R of relations such that R, but no proper subset of it, generates I as a two-sided ideal. The following statement is well-known. Lemma 3.20. If I is admissible then there exists a finite system of relations for I. Proof. By assumption, there is some N ≥ 2 such that rN ⊆ I. The algebra KQ/rN is finite-dimensional, as it is spanned by all the paths of Q of length smaller than N . Therefore the space I/rN is also finite-dimensional. For each i, j ∈ Q0, choose a basis of ei(I/rN )ej and choose a set Ri,j inside eiIej whose image modulo rN equals that basis. Then I = (R) for the finite set R given by R = {the paths of length N in Q} ∪ [i,j∈Q0 Ri,j. If R is not a system of relations for I then there is a proper subset R′ of R such that I = (R′). In this way we can repeatedly remove elements and still have a set generating I. Since R is finite, this process must terminate and we eventually end with a system of relations for I. (cid:3) Under some conditions, another approach to the construction of systems of relations involves lifting of basis elements of the space I/(I r + rI), see the discussion in [11, §7]. Lemma 3.21. If I is admissible, then I/(I r + rI) is finite-dimensional. Proof. Let N ≥ 2 be such that rN ⊆ I. Then rN +1 ⊆ (I r + rI) and we have an inclusion and a surjection KQ/r N +1 ⊃ I/r N +1 ։ I/(I r + rI). The claim now follows since the quotient KQ/rN +1 is finite-dimensional. (cid:3) Lemma 3.22. Let I be an ideal of KQ and let R be a set of relations inside I. (a) If I = (R) then the image of R modulo I r + rI spans the vector space I/(I r + rI). FINITE-DIMENSIONAL ALGEBRAS ARE (m > 2)-CALABI-YAU-TILTED 15 (b) Assume that the ideal (R) is admissible and the image of R modulo I r + rI spans the vector space I/(I r + rI). Then I = (R). (c) Assume that the ideal (R) is admissible and the image of R modulo I r + rI is a basis of the vector space I/(I r + rI). Then R is a system of relations for I. Proof. For part (a), observe that since each ρ ∈ R is a relation, multiplying it from the left or from the right by an element of the formPi∈Q0 λiei gives a scalar multiple of ρ. For part (b), we slightly modify the argument in [9, Lemma 3.6]. Let N ≥ 2 such that rN ⊆ (R) and let x ∈ I. By assumption, we can write kr′ x = λ1ρ1 + · · · + λnρn + x′ 1 + · · · + x′ 1r′ k + r′′ 1 x′′ (3.1) 1, . . . , x′′ with scalars λ1, . . . , λn ∈ K, relations ρ1, . . . , ρn ∈ R and x′ ℓ ∈ I, 1, . . . , x′ 1, . . . , r′ r′ ℓ using (3.1) and repeating this process N times, we conclude that x = ρ + r where ρ ∈ (R) and r ∈ rN , hence x ∈ (R). ℓ ∈ r. Writing each of the elements x′ ℓ x′′ k, x′′ 1, . . . , x′′ 1, . . . , x′ k, x′′ 1 + · · · + r′′ 1 , . . . , r′′ k, r′′ ℓ Finally, part (c) follows from (a) and (b). Moreover, R is finite by Lemma 3.21. (cid:3) Example 3.23. The assumption in parts (b) and (c) that the ideal (R) is admissible cannot be dropped. For example, consider the quiver Q given by α • β and let I = (α2 − βαβ, β2 − αβα, α2β). One can check that r5 ⊆ I ⊆ r2 hence the ideal I is admissible. The 8-dimensional algebra KQ/I is an algebra of quaternion type in the sense of Erdmann [14]. When K is algebraically closed of characteristic 2, this algebra is isomorphic to the group algebra of the quaternion group. The image of α2β in I r + rI vanishes, as the following calculation shows: α2β = (α2 − βαβ)β + βα(β2 − αβα) + βα2βα ∈ I r + rI, hence I/(I r + rI) is spanned by the images of the elements α2 − βαβ and β2 − αβα. Nevertheless, the ideal I ′ = (α2 − βαβ, β2 − αβα) is not equal to I. Indeed, by letting α and β act on K as the identity we get a one-dimensional module over KQ/I ′ with a non-zero action of α2β, hence α2β 6∈ I ′. 3.3. Finite-dimensional algebras are (m > 2)-CY-tilted. In this section we assume that the field K is algebraically closed. A finite-dimensional algebra A over K is called basic if AA ∼= P1 ⊕ · · · ⊕ Pr where P1, . . . , Pr are representatives of the isomorphism classes of the indecomposable projective right A-modules. Theorem 3.24 (Gabriel). Let A be a basic, finite-dimensional algebra over K. Then there exist a quiver Q and an admissible ideal I of KQ such that A ∼= KQ/I. Proof. See [15, §4.3]. (cid:3) Let A be a basic, finite-dimensional algebra. By Theorem 3.24, we can write A = KQ/I for a quiver Q and an admissible ideal I of Q. We denote by Si the simple A-module corresponding to a vertex i ∈ Q0 and consider the A-module S =Li∈Q0 Si. Lemma 3.25. If R is a system of relations for I then R ≥ dimK Ext2 A(S, S). 9 9 e e 16 SEFI LADKANI Proof. We have R ≥ dimK I/(I r + rI) = dimK Ext2 a consequence of Lemma 3.22(a) and the right equality is [7, Corollary 1.1]. A(S, S) where the left inequality is (cid:3) Note that Example 3.23 shows that the inequality in Lemma 3.25 can be strict. Theorem 3.26. Let A be a basic finite-dimensional algebra. Then the set of pairs (Q, R) consisting of a quiver Q and a sequence of relations R ⊆ r2 such that A ∼= KQ/(R) is not empty. For any such pair (Q, R) and any integer m > 2, the triangulated category C = C(Q,R,m) is Hom-finite, m-Calabi-Yau and its canonical m-cluster-tilting object T satisfies EndC(T ) ∼= A and dimK HomC(T, Σ−(m−2)T ) ≥ dimK Ext2 Proof. The first claim is a consequence of Theorem 3.24 and Lemma 3.20. The second claim is a consequence of Theorem 3.15 and Lemma 3.25. (cid:3) A(S, S). Corollary 3.27. A finite-dimensional algebra over an algebraically closed field is m- CY-tilted for any m > 2. Proof. We can write A ∼= End ¯A(P ) for a finite-dimensional, basic algebra ¯A and a finitely generated projective ¯A-module P containing as direct summands all the inde- composable projective ¯A-modules. Hence the claim is a consequence of Theorem 3.26 and Remark 3.17. (cid:3) Remark 3.28. If gldim A ≥ 2, then for any of the m-Calabi-Yau categories C of Theo- rem 3.26, the small negative extension HomC(T, Σ−(m−2)T ) of the canonical m-cluster- tilting object T cannot vanish. Had it vanished, Proposition 3.18 would then imply that A is the path algebra of an acyclic quiver, a contradiction. 3.4. Examples. Our first example is similar in spirit to [17, Example 3.5]. Example 3.29. Consider the algebra A = KQ/I where Q is the left quiver • α ?⑦⑦⑦⑦⑦⑦ ❅❅❅❅❅❅ γ • • β ❅❅❅❅❅❅ ?⑦⑦⑦⑦⑦⑦ δ • • α ?⑦⑦⑦⑦⑦⑦ ❅❅❅❅❅❅ γ • η • β ❅❅❅❅❅❅ ?⑦⑦⑦⑦⑦⑦ δ / • • α ?⑦⑦⑦⑦⑦⑦ ❅❅❅❅❅❅ γ • ε • β ❅❅❅❅❅❅ ?⑦⑦⑦⑦⑦⑦ δ • and I = (R) for the system of relations R = {αβ − γδ}. The algebra A has global dimension 2, hence it cannot be 2-CY-tilted by [23, Corollary 2.1]. Let B = B(Q, R) be the dg-algebra of Construction 3.4. Its graded quiver is shown in the middle; the arrows α, β, γ and δ have degree 0 while η has degree −1 and d(η) = αβ − γδ. Observe that the dg-algebra B is quasi-isomorphic to A ∼= H0(B). of the arrows α, β, γ, δ are 0, that of ε is 2− m and the superpotential is W = ε(αβ − γδ). Let m ≥ 2 and let (eQ, W ) be the graded quiver with homogeneous superpotential of degree 2−m of Construction 3.6. The graded quiver eQ is shown on the right; the degrees The Ginzburg dg-algebra Γ = Γm+1(eQ, W ) is the (m + 1)-Calabi-Yau completion of B (Lemma 3.10). Since B is quasi-isomorphic to A and A is derived equivalent to the path algebra of the Dynkin quiver D4, the Morita invariance of Calabi-Yau completions [22, Proposition 4.2] implies that Γ is Morita equivalent to Γm+1(D4, 0), hence by [17, Corol- lary 3.4] (or Proposition 3.18) the generalized m-cluster category CΓ is triangle equivalent to the m-cluster category of type D4. ? ? ? / ? ? o o ? FINITE-DIMENSIONAL ALGEBRAS ARE (m > 2)-CALABI-YAU-TILTED 17 If m = 2 then H0(Γ) ∼= KeQ/(eR), where eR = {αβ −γδ, εα, βε, εγ, δε}. This is a cluster- tilted algebra [9] of type D4, which is the relation-extension [4] of the tilted algebra A. If m > 2 then H0(Γ) ∼= A and A is m-CY-tilted by Theorem 3.15. Example 3.30. Let K be algebraically closed and consider the algebra A = K whose quiver Q is •. Consider the two sequences of relations R = {} and R′ = {0} on Q. Let m > 2. The Ginzburg dg-algebras Γ = Γ(Q, R, m) and Γ′ = Γ(Q, R′, m) are given by the following graded quivers with differentials t=−m d(t)=0 • t ε∗ ε ε∗=−1, ε=−(m−2), t=−m • t d(ε)=d(ε∗)=0 d(t)=εε∗−(−1)mε∗ε A basis for the graded piece Γ′i is given by (ε∗)i if 0 < i < m − 2; (ε∗)m−2, ε if i = m − 2; (ε∗)m−1, εε∗, ε∗ε if i = m − 1 and m > 3; and (ε∗)m−1, εε∗, ε∗ε, ε2 if i = m − 1 and m = 3, hence one computes (3.2) dimK H−i(Γ) =(1 0 if i = 0, if 0 < i < m, dimK H−i(Γ′) = if 0 ≤ i < m − 2, 1 2 if i = m − 2, 2 + δ3,m if i = m − 1. Let C = C(Q, R, m) and C′ = C(Q, R′, m′) be the corresponding m-Calabi-Yau cate- gories and let T ∈ C and T ′ ∈ C′ be the canonical m-cluster-tilting objects. We have EndC(T ) ∼= EndC ′(T ′) ∼= K. By Proposition 3.18, the category C is triangle equivalent to the m-cluster category of Q. Hence the m-cluster-tilting objects in C are the indecomposable objects ΣjT for 0 ≤ j < m. Since HomC(T, Σ−iT ) = 0 for any 0 < i < m, this remains true if we replace T by any of the other m-cluster-tilting objects ΣjT in C. On the other hand, T ′ is an m-cluster-tilting object in C′ with HomC ′(T ′, Σ−iT ′) 6= 0 for any 0 < i < m by (3.2), hence C′ cannot be triangle equivalent to C. The previous example can be generalized as follows. A Dynkin quiver is a quiver obtained by orienting the edges of a Dynkin diagram of type An (n ≥ 1), Dn (n ≥ 4) or En (n = 6, 7, 8). Proposition 3.31. Let Q be a Dynkin quiver and let m > 2. There exists an m-Calabi- Yau triangulated category C′ with an m-cluster-tilting object T ′ such that EndC ′(T ′) ∼= KQ but C′ is not triangle equivalent to the m-cluster category of Q. Proof. Let C be the m-cluster category of Q. Since Q is Dynkin, the category C has only finitely many indecomposable objects, hence the number of m-cluster-tilting objects T in C such that EndC(T ) ∼= KQ is finite. Therefore there exists an integer n such that dimK HomC(T, Σ−(m−2)T ) < n for any such T . Let R = {0, 0, . . . , 0} be a sequence consisting of n zero elements (it does not matter which starting and ending vertex we assign to each zero element) and let C′ = C(Q,R,m). By Theorem 3.15, the category C′ is Hom-finite and m-Calabi-Yau with an m-cluster- tilting object T ′ satisfying EndC ′(T ′) ∼= KQ and dimK HomC ′(T ′, Σ−(m−2)T ′) ≥ n. e e e e M M 18 SEFI LADKANI If F : C′ ≃ C were a triangulated equivalence, then T = F T ′ would be an m-cluster- tilting object in C with EndC(T ) ∼= KQ and dimK HomC(T, Σ−(m−2)T ) ≥ n, a contra- diction. (cid:3) References 1. Claire Amiot, Cluster categories for algebras of global dimension 2 and quivers with potential, Ann. Inst. Fourier (Grenoble) 59 (2009), no. 6, 2525 -- 2590. 2. Claire Amiot and Steffen Oppermann, Cluster equivalence and graded derived equivalence, Doc. Math. 19 (2014), 1155 -- 1206. 3. , Higher preprojective algebras and stably Calabi-Yau properties, Math. Res. Lett. 21 (2014), no. 4, 617 -- 647. 4. I. Assem, T. Brustle, and R. Schiffler, Cluster-tilted algebras as trivial extensions, Bull. Lond. Math. Soc. 40 (2008), no. 1, 151 -- 162. 5. Karin Baur and Robert J. Marsh, A geometric description of m-cluster categories, Trans. Amer. Math. Soc. 360 (2008), no. 11, 5789 -- 5803. 6. Apostolos Beligiannis, Relative homology, higher cluster-tilting theory and categorified Auslander- Iyama correspondence, J. Algebra 444 (2015), 367 -- 503. 7. Klaus Bongartz, Algebras and quadratic forms, J. London Math. Soc. (2) 28 (1983), no. 3, 461 -- 469. 8. Aslak Bakke Buan, Robert Marsh, Markus Reineke, Idun Reiten, and Gordana Todorov, Tilting theory and cluster combinatorics, Adv. Math. 204 (2006), no. 2, 572 -- 618. 9. Aslak Bakke Buan, Robert J. Marsh, and Idun Reiten, Cluster-tilted algebras of finite representation type, J. Algebra 306 (2006), no. 2, 412 -- 431. 10. 11. Igor Burban, Osamu Iyama, Bernhard Keller, and Idun Reiten, Cluster tilting for one-dimensional , Cluster-tilted algebras, Trans. Amer. Math. Soc. 359 (2007), no. 1, 323 -- 332 (electronic). hypersurface singularities, Adv. Math. 217 (2008), no. 6, 2443 -- 2484. 12. P. Caldero, F. Chapoton, and R. Schiffler, Quivers with relations arising from clusters (An case), Trans. Amer. Math. Soc. 358 (2006), no. 3, 1347 -- 1364. 13. Louis de Thanhoffer de Volcsey and Michel Van den Bergh, Some new examples of nondegenerate quiver potentials, Int. Math. Res. Not. IMRN (2013), no. 20, 4672 -- 4686. 14. Karin Erdmann, Blocks of tame representation type and related algebras, Lecture Notes in Mathe- matics, vol. 1428, Springer-Verlag, Berlin, 1990. 15. Peter Gabriel, Auslander-Reiten sequences and representation-finite algebras, Representation theory, I (Proc. Workshop, Carleton Univ., Ottawa, Ont., 1979), Lecture Notes in Math., vol. 831, Springer, Berlin, 1980, pp. 1 -- 71. 16. Victor Ginzburg, Calabi-Yau algebras, arXiv:math.AG/0612139,. 17. Lingyan Guo, Cluster tilting objects in generalized higher cluster categories, J. Pure Appl. Algebra 215 (2011), no. 9, 2055 -- 2071. 18. Osamu Iyama and Steffen Oppermann, Stable categories of higher preprojective algebras, Adv. Math. 244 (2013), 23 -- 68. 19. Osamu Iyama and Yuji Yoshino, Mutation in triangulated categories and rigid Cohen-Macaulay modules, Invent. Math. 172 (2008), no. 1, 117 -- 168. 20. Bernhard Keller, On triangulated orbit categories, Doc. Math. 10 (2005), 551 -- 581. 21. , Calabi-Yau triangulated categories, Trends in representation theory of algebras and related topics, EMS Ser. Congr. Rep., Eur. Math. Soc., Zurich, 2008, pp. 467 -- 489. 22. , Deformed Calabi-Yau completions, J. Reine Angew. Math. 654 (2011), 125 -- 180, With an appendix by Michel Van den Bergh. 23. Bernhard Keller and Idun Reiten, Cluster-tilted algebras are Gorenstein and stably Calabi-Yau, Adv. Math. 211 (2007), no. 1, 123 -- 151. 24. , Acyclic Calabi-Yau categories, Compos. Math. 144 (2008), no. 5, 1332 -- 1348, With an appendix by Michel Van den Bergh. 25. Bernhard Keller and Dong Yang, Derived equivalences from mutations of quivers with potential, Adv. Math. 226 (2011), no. 3, 2118 -- 2168. FINITE-DIMENSIONAL ALGEBRAS ARE (m > 2)-CALABI-YAU-TILTED 19 26. Hugh Thomas, Defining an m-cluster category, J. Algebra 318 (2007), no. 1, 37 -- 46. 27. Michel Van den Bergh, Calabi-Yau algebras and superpotentials, Selecta Math. (N.S.) 21 (2015), no. 2, 555 -- 603. 28. Anette Wralsen, Rigid objects in higher cluster categories, J. Algebra 321 (2009), no. 2, 532 -- 547. 29. Yu Zhou and Bin Zhu, Cluster combinatorics of d-cluster categories, J. Algebra 321 (2009), no. 10, 2898 -- 2915. Department of Mathematics, University of Haifa, Mount Carmel, Haifa 31905, Israel E-mail address: [email protected]
1209.6544	4	1209	2015-06-17T15:22:01	Two-generated algebras and standard-form congruence	[ "math.RA" ]	Matrix congruence can be used to mimic linear maps between homogeneous quadratic polynomials in $n$ variables. We introduce a generalization, called standard-form congruence, which mimics affine maps between non-homogeneous quadratic polynomials. Canonical forms under standard-form congruence for three-by-three matrices are derived. This is then used to give a classification of algebras defined by two generators and one degree two relation. We also apply standard-form congruence to classify homogenizations of these algebras.	math.RA	math	TWO-GENERATED ALGEBRAS AND STANDARD-FORM CONGRUENCE JASON GADDIS Abstract. Matrix congruence can be used to mimic linear maps between homogeneous quadratic polynomials in n variables. We introduce a general- ization, called standard-form congruence, which mimics affine maps between non-homogeneous quadratic polynomials. Canonical forms under standard- form congruence for three-by-three matrices are derived. This is then used to give a classification of algebras defined by two generators and one degree two relation. We also apply standard-form congruence to classify homogenizations of these algebras. Keywords: Matrix congruence; Isomorphism problems; Two-generated algebras; Automorphism groups; Skew polynomial rings; Homogenization 1. Introduction Let k be an algebraically closed field of characteristic zero. All algebras are k- algebras and all isomorphisms are as k-algebras. We denote by Mn(k) the ring of n × n matrices over k. We denote the center of an algebra A by Z(A). Our interest is in algebras A defined as a factor of the free algebra on two degree one generators by a single degree two relation, i.e., A = khx, y fi, deg(f ) = 2. (1.1) In case f is homogeneous, the classification of such algebras is well-known (see, e.g., [14]). The polynomial f can be represented by a 2 × 2 matrix and matrix congru- ence corresponds to linear isomorphisms between homogeneous algebras. Hence, canonical forms for matrices in M2(k) give a maximal list of algebras to consider. One must verify that there are no non-linear isomorphisms between the remaining algebras. This can be accomplished by considering ring-theoretic properties, re- sulting in four types of algebras: the quantum planes Oq(k2), the Jordan plane J , Ryx, and Rx2 . We give a method for extending this idea to algebras in which f is not necessarily homogeneous. In Section 3, we develop a modified version of matrix congruence called standard-form congruence. Canonical forms in M3(k) under standard-form congruence are determined in Section 4. These forms are in near 1-1 correspondence with isomorphism classes of algebras of the form (1.1). This leads to the following theorem. 2000 Mathematics Subject Classification. Primary 15A21, 16S37; Secondary 16S36. 1 2 JASON GADDIS Theorem 1.1. Suppose A ∼= khx, y fi where f is a polynomial of degree two. Then A is isomorphic to one of the following algebras: Oq(k2), f = xy − qyx (q ∈ k×), J , f = yx − xy + y2, U, f = yx − xy + y, Rx2, f = x2, Ryx, f = yx, Aq 1(k), f = xy − qyx − 1 (q ∈ k×), J1, f = yx − xy + y2 + 1, k[x], f = x2 + y, Rx2−1, f = x2 − 1, S, f = yx − 1. Furthermore, the above algebras are pairwise non-isomorphic, except (k). Many of these algebras are well-known. The algebras Aq Oq(k2) ∼= Oq−1 (k2) and Aq 1(k) ∼= Aq−1 1(k) are the quantum Weyl algebras, U the enveloping algebra of the non-abelian two-dimensional solvable Lie algebra, and J1 the deformed Jordan plane. This list slightly contradicts that given in [14] since S and J1 both have Gelfand-Kirillov (GK) dimension two. We prove this theorem in Section 5. 1 We define one additional algebra, V = khx, y yx − xy + y2 + xi. This algebra is not included in Theorem 1.1 because it is isomorphic to U (Propo- sition 5.1). As a second application of standard-form congruence, we consider a related class of algebras, H = khx, y, z xz − zx, yz − zy, fi, f ∈ khx, y, zi, deg(f ) = 2, f /∈ k[z], f homogeneous. (1.2) Algebras of form (1.2) may be regarded as homogenizations of those of form (1.1). In Section 6, we prove (Theorem 6.1) that canonical forms under sf-congruence are in 1-1 correspondence with isomorphism classes of algebras of the form (1.2). This result may be framed in terms of (Artin-Schelter) regular algebras. We refer the reader to [13] for undefined terms. If H is a global dimension three regular algebra, then H is associated to a point scheme C ⊂ P2 and an automorphism σ of C. Suppose C contains a line fixed by σ (Type S ′ 1 regular algebras). By [13], Proposition 1.2, H may be twisted so that it is isomorphic to a k-algebra on generators x, y, z with defining relations where h is one of the following polynomials xz = zx, yz = zy, h = 0, (I) yx − xy + y2 + z(αx + βy + γz), (II) xy − qyx + z(αx + βy + γz), q ∈ k×, for some α, β, γ ∈ k. Hence, Theorem 6.1 provides a refinement on this classifica- tion. 2. Congruence Let f = ax2 + bxy + cyx + dy2, a, b, c, d ∈ k. By a slight abuse of notation, f =(cid:0)x y(cid:1)(cid:18)a b c d(cid:19)(cid:18)x y(cid:19) . TWO-GENERATED ALGEBRAS AND STANDARD-FORM CONGRUENCE 3 Hence, we can represent any homogeneous quadratic polynomial by an element of M2(k). If A is of the form (1.1), then f is called a defining polynomial for A and the matrix corresponding to f is called a defining matrix for A. The map φ given by x 7→ p11x + p12y and y 7→ p21x + p22y, pij ∈ k, with p11p22− p12p21 6= 0 corresponds to a linear isomorphism between the algebras with defining polynomials f and φ(f ). Similarly, M, M ′ ∈ Mn(k) are said to be congruent and we write M ∼ M ′ if there exists P ∈ GLn(k) such that P T M P = M ′. Matrix congruence is an equivalence relation on the set Mn(k). A canonical form under congruence is a distinguished representative from an equivalence class. When two defining matrices are congruent there is an linear map between the polynomials that they determine. In turn, the algebras with these defining poly- nomials are isomorphic. On the other hand, if there is a linear map between two defining polynomials, then the corresponding algebras are isomorphic. However, two such algebras can still be isomorphic even if there is no linear map between the defining polynomials. Thus, canonical forms for congruent matrices give us a maximal list of algebras to consider and we are then left to determine whether there are any other isomorphisms. The Horn-Sergeichuk forms depend on three block-types which we henceforth refer to as HS-blocks, Jn(λ) = Γn =     0 1 λ 1 λ . . . . . . 0 0 1 λ   −1 1 1 −1 −1 1 , J1(λ) =(cid:0)λ(cid:1) , (−1)n+1 (−1)n . . . . . .   , Γ1 =(cid:0)1(cid:1) , 0 H2n(µ) =(cid:18) 0 Jn(µ) In 0(cid:19) , H2(µ) =(cid:18)0 µ 0(cid:19) . 1 Theorem 2.1 (Horn, Sergeichuk [9]). Each square complex matrix is congruent to a direct sum, uniquely determined up to permutation of summands, of canonical matrices of the three types Jn(0), Γn, and H2n(µ), µ 6= 0, (−1)n+1. Moreover, H2n(µ) is determined up to replacement of µ by µ−1. As a consequence of the previous theorem, there are four HS-block types in dimension two: Γ1 ⊕ J1(0), J2(0), Γ2, and H2(µ). We choose to use J2(0)T in place of J2(0) and let q = µ in H2(µ). These matrices are given explicitly as (cid:18)1 0 0 0(cid:19) ,(cid:18)0 1 0 0(cid:19) ,(cid:18)0 −1 1 (cid:19) ,(cid:18)0 −1 0 (cid:19) , q ∈ k×. 1 q (2.1) We denote these matrices by Mx2, Myx, MJ , and Mq, respectively. It follows from Theorem 2.1, or a quick computation, that Mq ∼ Mq−1 . Moreover, Mp ∼ Mq if and only if p = q±1 (see Corollary 2.3). 4 JASON GADDIS As it will be useful in the general (non-homogeneous) case we compute the stabi- lizer groups for the matrices in (2.1). In general, these stabilizer groups correspond to some orthosymplectic group but, because some of the forms are degenerate, there are shifts in the dimension. Any M ∈ Mn(k) admits a unique decomposition M = A+ S where A, S ∈ Mn(K) with S symmetric and A antisymmetric. Because congruence preserves symmetry (resp. antisymmetry), then the stabilizer group of M is equal to the intersection of the stabilizer groups for A and S. Proposition 2.2. Let M be one of the matrices in (2.1). The group is described below. Stab(M ) = {P ∈ GL2(k) P T M P = M} r 0 s(cid:19)(cid:12)(cid:12)(cid:12)(cid:12) Stab (Mx2) =(cid:26)(cid:18)±1 0 0 r−1(cid:19)(cid:12)(cid:12)(cid:12)(cid:12) Stab (Myx) =(cid:26)(cid:18)r 1(cid:19)(cid:12)(cid:12)(cid:12)(cid:12) Stab (MJ ) =(cid:26)±(cid:18)1 0 r−1(cid:19)(cid:12)(cid:12)(cid:12)(cid:12) Stab (Mq) =(cid:26)(cid:18)r Stab (M−1) =(cid:26)(cid:18)r 0 r−1(cid:19) ,(cid:18) 0 0 0 0 r Stab (M1) = SL2(k). s−1 r, s ∈ k×(cid:27) , r ∈ k×(cid:27) , r ∈ k×(cid:27) , r ∈ k×(cid:27) (q ∈ k×, q 6= ±1), r, s ∈ k×(cid:27) , s 0(cid:19)(cid:12)(cid:12)(cid:12)(cid:12) Proof. Throughout, let P ∈ Stab(M ) and write P =(cid:18)a b d(cid:19). alternating form. Thus, its stabilizer is Sp(2) ∼= SL2(k). c The matrix M1 corresponds to the standard basis non-degenerate dimension two The matrix M−1 corresponds to a non-standard basis non-degenerate dimension two symmetric form. We have P T M−1P =(cid:18) 2ac ac + bd ac + bd 2bd (cid:19) =(cid:18)0 1 1 0(cid:19) . Hence, either a = d = 0 or b = c = 0 and the result follows. In the case of Mx2 we have P T Mx2P =(cid:18)a2 ab b2(cid:19) =(cid:18)1 0 0 0(cid:19) . ab Then a = ±1 and b = 0. We have MJ =(cid:18)0 0 0 1(cid:19) +(cid:18)0 −1 0 (cid:19) . 1 Finally, for q 6= ±1, Thus, Stab(MJ ) = Stab(Mx2)T ∩ Stab(M1). 2 (cid:18)0 −1 0 (cid:19) + Thus, Stab(Mq) = Stab(M1) ∩ Stab(M−1). Mq = q + 1 1 2 (cid:18)0 q − 1 1 1 0(cid:19) . The case of Myx may be seen from the previous computation by letting q = 0. (cid:3) TWO-GENERATED ALGEBRAS AND STANDARD-FORM CONGRUENCE 5 Corollary 2.3. Let p, q ∈ k×. Then Mp ∼ Mq if and only if p = q±1. Proof. Sufficiency is provided by Theorem 2.1. Suppose Mp ∼ Mq and choose P ∈ GL2(k) such that Mp = P T MqP . Write P =(cid:18)a b c P T MqP =(cid:18)(q − 1)ac qad − bc d(cid:19). Then, (q − 1)bd(cid:19) =(cid:18)0 −1 0 (cid:19) . qbc − ad q Observe that, if q = 1, then Mp = P T M1P = (ad − bc)M1, and so p = 1. In general, we see by comparing P T MqP to Mp in (2.2) that ac = bd = 0. Thus, either b = c = 0 or a = d = 0. In the first case, Mp = P T MqP = (ad)Mq, and so p = q. In the second case, Mp = P T MqP = (−qbc)Mq−1, and so p = q−1. (cid:3) 3. Standard Form Congruence In the non-homogeneous case, we write f = ax2 + bxy + cyx + dy2 + αx + βy + γ, a, b, c, d, α, β, γ ∈ k. We can represent f by a 3 × 3 matrix via the rule (2.2) We extend the terms defining polynomial and defining matrix as one would expect. However, our choice of defining matrix is not unique. One could define f by Hence, it is necessary to fix a standard form for the defining matrices of non- homogeneous polynomials. We restrict our attention to the following set, f =(cid:0)x y f =(cid:0)x y 1(cid:1)  1(cid:1)  (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) c  a b α c d β 0 0 γ    x y 1  . 0 0 b d a c x y 1 α β γ   .   b2(cid:19) 6= 0  (cid:18)a1 a2 b1 Every degree two polynomial has a unique corresponding matrix in G3. Consider the matrix ⊂ M3(k). a1 a2 a3 b1 b3 0 b2 0 G3 =    M =  m11 m12 m13 m21 m22 m23 m31 m32 m33   ∈ M3(k). This corresponds to the polynomial f = m11x2 + m12xy + m13x + m21yx + m22y2 + m23y + m31x + m32y + m33 = m11x2 + m12xy + m21yx + m22y2 + (m13 + m31)x + (m23 + m32)y + m33, which in turn corresponds to the matrix   . Hence, we define a k-linear map sf : M3(k) → G3 by m11 m12 m13 + m31 m21 m22 m23 + m32   m33 0 0 m11 m12 m13 m21 m22 m23 m31 m32 m33     7→  m11 m12 m13 + m31 m21 m22 m23 + m32 0 0 m33   . 6 JASON GADDIS Let pij ∈ k and define a k-linear map by φ(x) = p11x + p12y + p13, φ(y) = p21x + p22y + p23, φ(1) = 1. (3.1) If p11p22 − p12p21 6= 0, then φ defines an affine isomorphism between khx, y fi and khx, y φ(f )i. Thus, the matrices corresponding to affine isomorphisms of these algebras should be contained in the set P3 =(cid:26)(cid:18)P1 P2 1(cid:19) ∈ M3(k) P1 ∈ GL2(k), P2 ∈ k2(cid:27) . mial and adds the linear parts. We write M ∈ Mn(k) in block form 0 In general, we want a map that fixes the degree two part of a quadratic polyno- M =(cid:26)(cid:18) M1 M2 3 m(cid:19) M1 ∈ Mn−1(k), M2, M3 ∈ kn−1, m ∈ k(cid:27) . M T (3.2) We call M1 the homogeneous block of M . Define the set Gn =(cid:26)(cid:18)M1 M2 0 m(cid:19) ∈ Mn(k) 0 6= M1 ∈ Mn−1(k), M2 ∈ kn−1, m ∈ k(cid:27) . Then define the map sf : Mn → Gn by (cid:18) M1 M2 3 m(cid:19) 7→(cid:18)M1 M2 + M3 m (cid:19) , M T 0 (3.3) where the matrix is written according to (3.2). The matrices corresponding to affine isomorphisms of these algebras should be contained in the set Pn =(cid:26)(cid:18)P1 P2 1(cid:19) ∈ Mn(k) P1 ∈ GLn−1(k), P2 ∈ kn−1(cid:27) . 0 Proposition 3.1. Pn is a group. Proof. That Pn contains the identity matrix is clear. Let P, P ′ ∈ Pn. Then P P ′ =(cid:18)P1 P2 1(cid:19)(cid:18)P ′ 0 1 P ′ 2 0 1(cid:19) =(cid:18)P1P ′ 0 Since P1 ∈ GLn−1(k), then we can set P ′ It is now clear from the above that P ′ = P −1. 1 = P −1 1 P1P ′ 2 + P2 1 (cid:19) ∈ Pn. 1 ∈ GLn−1(k) and P ′ 2 = −P −1 1 P2. (cid:3) Under ordinary matrix congruence, two matrices which are scalar multiples of each other are always congruent. However, if we restrict to Pn, that is no longer the case. Hence, in our modified definition of congruence, we set scalar multiple matrices to be congruent to each other. Definition 3.2. We say M, N ∈ Mn(k) are standard-form congruent (sf- congruent) and write M ∼sf N if there exist P ∈ Pn and α ∈ k× such that sf(M ) = α · sf(P T N P ). The next proposition shows that sf-congruence is a true extension of congruence. Proposition 3.3. Let M, N ∈ Mn(k) with homogeneous blocks M1, N1, respec- tively. If M ∼sf N , then M1 ∼ N1. TWO-GENERATED ALGEBRAS AND STANDARD-FORM CONGRUENCE 7 Proof. By hypothesis, sf(M ) = α · sf(P T N P ) for some P ∈ Pn, α ∈ k×. Then 0 m(cid:19) = sf(M ) = α · sf(P T N P ) (cid:18)M1 M2 1 P T 2 1 N1P1 = α · sf(cid:18)(cid:18)P T = α · sf(cid:18)(cid:18)P T 1 N1P1, so M1 ∼ N1. ∗ 0 1(cid:19)(cid:18)N1 N2 0 m(cid:19)(cid:18)P1 P2 ∗(cid:19)(cid:19) =(cid:18)α · P T ∗ 0 1(cid:19)(cid:19) 1 N1P1 0 ∗(cid:19) . ∗ (cid:3) Thus, M1 = α · P T The following may be regarded as a sort of converse to Proposition 3.3. Corollary 3.4. Let M, N ∈ Mn(k) with the same homogeneous block L. If P ∈ Pn is such that sf(M ) = α · sf(P T N P ) for some α ∈ k×, then P1 = γQ1 for some Q1 ∈ Stab(L) and γ ∈ k×. Proof. This is an immediate consequence of the computation in the previous propo- sition. We have L = α · P T γ = √α−1. 1 LP1. Let Q1 = √αP1, then Q1 ∈ Stab(M1) and (cid:3) In the next section we will determine equivalence classes in M3(k) under sf- congruence. By Proposition 3.3 and Corollary 3.4, we may immediately divide the matrices into distinguished classes depending on the homogeneous blocks. The next proposition will allow us to show sf-congruence between matrices with the same homogeneous block, but whose column vector M2 or constant m are scalar multiples. Proposition 3.5. Let M ∈ Gn and γ ∈ k×. Then γ−1M2 γ−2m(cid:19) . M ∼sf (cid:18)M1 0 Proof. Let I be the (n − 1) × (n − 1) identity matrix. Let P ∈ Pn with P1 = γI and P2 = 0. Then P T M P =(cid:18)γ2M1 0 γM2 m (cid:19) = γ2 ·(cid:18)M1 0 γ−1M2 γ−2m(cid:19) . (cid:3) Proving that standard-form congruence is an equivalence relation requires the following technical lemmas. Lemma 3.6. If M ∈ Mn(k) and P ∈ Pn, then sf(P T M P ) = sf(P T sf(M )P ). 8 JASON GADDIS Proof. We have, 1 P T 2 sf(P T M P ) = sf(cid:18)(cid:18)P T = sf(cid:18)(cid:18) =(cid:18)P T =(cid:18)P T = sf(cid:18)(cid:18)P T 1 P T 2 0 0 1 M1P1 0 1(cid:19)(cid:18)M1 0 = sf(P T sf(M )P ). 1 M1P1 P T 1 M1P2 + P T 0 1(cid:19)(cid:18) M1 M2 3 m(cid:19)(cid:18)P1 P2 1(cid:19)(cid:19) M T 0 P T 1 M1P1 P T 2 M1P1 + M T 3 P1 P T 2 M1P2 + P T 1 M2 + (P T P T 1 M1P2 + P T P T 2 M1P2 + P T 2 M2 + M T P T 1 M1P2 + P T 1 M2 2 M2 + M T 2 M1P1 + M T 3 P2 + m 1 P2 + P T 3 P2 + m(cid:19)(cid:19) 3 P1)T (cid:19) 2 M1P2 + P T P T M2 + M3 1 M2 + P T 1 M T 2 M2 + P T m (cid:19)(cid:18)P1 P2 0 1 M3 2 M3 + m (cid:19) 1(cid:19)(cid:19) (cid:3) Proposition 3.7. Standard-form congruence defines an equivalence relation. Proof. Reflexivity is obvious. Now suppose M ∼sf M ′, so sf(M ) = α · sf(P T M ′P ) for some α ∈ k× and P ∈ Pn. By Lemma 3.6, (P −1)T sf(M )(P −1) = α · (P −1)T sf(P T M ′P )(P −1) sf(cid:0)(P −1)T sf(M )(P −1)(cid:1) = α · sf(cid:0)(P −1)T sf(P T M ′P )(P −1)(cid:1) α−1 · sf(cid:0)(P −1)T M (P −1)(cid:1) = sf(cid:0)(P −1)T P T M ′P (P −1)(cid:1) α−1 · sf(cid:0)(P −1)T M (P −1)(cid:1) = sf(M ′). Hence, M ′ ∼sf M , so symmetry holds. P, Q ∈ Pn such that Finally, suppose M ∼sf M ′ and M ′ ∼sf M ′′. Then there exists α, β ∈ k× and sf(M ) = α · sf(P T M ′P ) and sf(M ′) = β · sf(QT M ′′Q). By two additional applications of Lemma 3.6, sf(M ) = α · sf(P T M ′P ) = α · sf(cid:0)P T sf(M ′)P(cid:1) = α · sf(cid:0)P T (cid:0)β · sf(QT M ′′Q)(cid:1) P(cid:1) = (αβ) · sf((QP )T M ′′(QP )). Thus, M ∼sf M ′′, so transitivity holds as well. (cid:3) 4. Canonical Forms Canonical forms for these equivalence classes are presented in Theorem 4.2. In this section, we determine equivalence classes for M3(k) under sf-congruence. If M ∼sf N , then M1 ∼ N1 by Proposition 3.3. Thus, we may assume that M ∈ G3 and M1 is one of (2.1). By Proposition 2.2, it is left only to determine which pairs (M2, m) determine distinct forms. Corollary 4.1. Let p, q ∈ k×. Let Wp and Wq be the defining matrices for Ap and Aq Proof. That Wp ∼sf Wq if p = q±1 is an easy check and we omit it. The converse now follows by Corollary 2.3 and Proposition 3.3. 1(k), respectively. Then Ap 1(k) if and only if p = q±1. 1(k) ∼sf Aq 1(k) (cid:3) TWO-GENERATED ALGEBRAS AND STANDARD-FORM CONGRUENCE 9 Our last step is to determine, for each canonical form in M2(k), which pairs (M2, m) give sf-congruent matrices. Theorem 4.2. Canonical forms for M3(k) under sf-congruence are given below: 0 0 1 0 0 0 0 0 0 0 0 0(cid:17)(cid:16) 1 0 0 (cid:16) 1 0 0 0 0 −1(cid:17)(cid:16) 1 0 0 0 0 0(cid:17) 0 0 0(cid:17)(cid:16) 0 0 0 (cid:16) 0 0 0 0 0 −1(cid:17) 1 0 0 1 0 0 1 1 0 1 1 0 1 1 0 0 0 0(cid:17)(cid:16) 0 −1 0 (cid:16) 0 −1 0 0 0 0(cid:17)(cid:16) 0 −1 0 (cid:16) 0 −1 0 0 0 1(cid:17)(cid:16) 0 −1 1 0 0 0(cid:17) 0 0 1(cid:17)(cid:16) 0 −1 0 0 0 0(cid:17) . q 0 0 1 0 1 q 0 0 Moreover, the forms involving q are determined up to replacement by q−1. Proof. Suppose M ∈ M3(k). We perform necessary congruence operations to put M1 in canonical form. Thus, M is sf-congruent to a block matrix of the form N =(cid:18)L N2 n (cid:19) 0 1(cid:19) ∈ P3. where L is one of (2.1), N2 =(cid:0)u v(cid:1)T By Corollary 3.4, we may assume P1 ∈ Stab(L). Write P1 as in Proposition 2.2 and P2 =(cid:0)e ±1. Both cases are similar and we only consider the positive case below, ∈ k2, and n ∈ k. Let P =(cid:18)P1 P2 (Case 1: L = Mx2) There are two cases for the stabilizer of L corresponding to f(cid:1)T ∈ k2. 0 sf(P T N P ) =  1 0 0 0 0 2e + u + rv 0 e2 + eu + f v + n  . sv Because det(P ) 6= 0, then s 6= 0. Thus, sv = 0 if and only if v = 0. In case v 6= 0 we set e = 0, s = v−1, r = −uv−1, and f = −nv−1. This is the defining 2 u. The bottom right entry matrix of k[x]. becomes − 1 4 u2, then we have the defining matrix of Rx2 and otherwise, by Proposition 3.5, that of Rx2−1. In case v = 0, then we set e = − 1 4 u2 + n. Thus, if n = 1 (Case 2: L = Myx) We have sf(P T N P ) =  0 1 0 0 0 r(u + f ) r−1(e + v) 0 f e + eu + f v + n  . Setting f = −u and e = −v gives a bottom right entry of n − uv. Thus, there are two cases corresponding to n = uv and n 6= uv. In the former case we arrive at the defining matrix of Ryx and in the other case, by Proposition 3.5, that of S. (Case 3: L = MJ ) There are two cases for the stabilizer. We consider only the positive case, which gives, sf(P T N P ) = 0 −1 1 1  0 0 u 2f + ru + v f 2 + eu + f v + n  . 2 v and the (3, 3)-entry becomes n − 1 Setting f = − 1 2 (ru + v) allows us to make the (2, 3)-entry zero. If u = 0, we let f = − 1 4 v2 we have the defining matrix for J and otherwise that for J1. If u 6= 0, then by Proposition 3.5 we can assume u = 1. Thus, f 2 + f v = 1 4 (v2− r2)− n so that the (3, 3)-entry is zero, giving the defining matrix for V. 4 (r2− v2), and so we take e = 1 4 v2. Thus, in case n = 1 10 JASON GADDIS (Case 4a: L = M1) Then sf(P T N P ) =  0 −1 0 1 0 0 au + cv bu + dv eu + f v + n  . Suppose u = v = 0. If n = 0, then we have the defining matrix for k[x, y] and otherwise we have the matrix for A1(k). Suppose u = 0 and v 6= 0. Setting a = v, c = 0, d = v−1, and f = −nv−1 gives the defining matrix for U. Similarly for the case v = 0 and u 6= 0. Finally, suppose u, v 6= 0. We choose e, f such that eu + f v = −n. Because det(P1) 6= 0, we can choose a, b, c, d such that au + dv = 0 and bu + dv = 1, giving the defining matrix for U. (Case 4b: L = Mq, q 6= ±1) We note that, in case q = −1, there are additional matrices in the stabilizer group than those considered here. However, they are not needed in this result. We have sf(P T N P ) =  0 −1 0 q 0 0 r(u + (q − 1)f ) r−1(v + (q − 1)e) f e(q − 1) + eu + f v + n  . Set f = u(1 − q)−1 and e = v(1 − q)−1. Then the bottom right entry becomes n − uv(q − 1)−1. Thus, if n = uv(q − 1)−1, then this form corresponds to the defining matrix for Oq(k2) and otherwise, by Proposition 3.5, it corresponds to that of Aq 1(k). Combining this with Corollaries 2.3 and 4.1 completes the result. (cid:3) 5. Classification We wish to show that the list in Theorem 1.1 is complete with no isomorphic repetitions. The observant reader may have noticed a discrepancy in the Theorem 4.2 and Theorem 1.1. Proposition 5.1. The algebras U and V are isomorphic. Proof. Let X, Y be the generators for U and let x, y be the generators for V. Define a map Φ : U → V by Φ(X) = −y, Φ(Y ) = x + y2. This map extends to an algebra homomorphism since Φ(Y )Φ(X) − Φ(X)Φ(Y ) + Φ(Y ) = (x + y2)(−y) − (−y)(x + y2) + (x + y2) = yx − xy + x + y2. We also define Ψ : V → U by Ψ(x) = Y − X 2, Ψ(y) = −X. This map also extends to an algebra homomorphism since Ψ(y)Ψ(x) − Ψ(x)Ψ(y) + Ψ(x) + Ψ(y)Ψ(y) = (−X)(Y − X 2) − (Y − X 2)(−X) + (Y − X 2) − (−X)2 = 0. It is readily checked that Ψ(Φ(X)) = X and Ψ(Φ(Y )) = Y so that Ψ = Φ−1. (cid:3) This is the one case where two algebras are isomorphic even though their defining matrices are not sf-congruent. This makes sense as the map Φ constructed above is not an affine isomorphism. The relationship between U and V is explored further in [5]. In particular, U is a Poincar´e-Birkhoff-Witt (PBW) deformation of k[x, y] while V is a PBW deformation of J . In Section 6, we will show that the respective homogenizations of U and V are not isomorphic. TWO-GENERATED ALGEBRAS AND STANDARD-FORM CONGRUENCE 11 One can divide the remaining algebras into two classes: the domains and non- domains. The non-domains can be distinguished using well-known ring-theoretic results. For details, we refer the interested reader to [7]. Proposition 5.2. The algebras Ryx, Rx2 , Rx2−1 and S are all non-isomorphic. Proof. The algebras Rx2, Rx2−1 and S are prime while Ryx is not. We have gld Rx2−1 = gldS = 1 whereas gld Rx2 = ∞. Finally, gkS = 2 whereas gk Rx2−1 = ∞. (cid:3) The domains can be further subdivided into differential operator rings, quantum Weyl algebras, and quantum planes. We review each of those classes here. Proposition 5.3. Let p, q ∈ k×. Then Aq Proof. When p and q are not roots of unity, this was proved in [2], Corollary 3.11 (c). This result follows in full from [6], Propositions 6.1, 6.3, and 6.4. See also [15] for a more general result that also applies here. (cid:3) 1(k) if and only if p = q±1. 1(k) ∼= Ap The corresponding result for quantum planes is [7], Corollary 4.2.12. However, as a more general result is useful in Section 6, we review it here. We say q = (qij) ∈ Mn(k×) is multiplicatively antisymmetric if qii = 1 and qij = q−1 for all i 6= j. Let Sn be the symmetric group on n letters. If A ∈ Mn(k×) ji is multiplicatively antisymmetric and σ ∈ Sn, then σ acts on A by σ.A = [Aσ(i)σ(j)]. We say p is a permutation of q if there exists σ ∈ Sn such that p = σ.q. For q ∈ Mn(k×) multiplicatively antisymmetric, quantum affine n-space Oq(kn) is defined as the algebra with generating basis {xi}, 1 ≤ i ≤ n, subject to the relations xixj = qij xjxi for all 1 ≤ i, j ≤ n. Theorem 5.4 ([7], Theorem 4.2.11). Op(kn) ∼= Oq(kn) if and only if p is a permutation of q. Proof. (Sketch) Let {xi} be a generating basis for Op(kn) and {yi} that for Oq(kn). If p = σ.q, then one easily constructs an isomorphism Φ : Op(kn) → Oq(kn) given by Φ(xi) = yσ(i). Conversely, any isomorphism determines a permutation of the degree one ele- ments. One then checks that such a permutation must also permute the parameters accordingly. (cid:3) We now review isomorphisms between the rings U, J , and J1. Let S be a ring. Given σ ∈ Aut(S), a k-linear map δ : S → S is said to be a σ-derivation if it satisfies the twisted Leibniz rule, δ(ab) = σ(a)δ(b) + δ(a)b for all a, b ∈ S. The skew polynomial ring R = S[x; σ, δ] is the overring of S with commutation given by xa = σ(a)x + δ(a) for all a ∈ S. If δ = 0, then we write R = S[x; σ]. If σ = idS, then we write R = S[x; δ] and R is said to be a differential operator ring. The algebras U, J , and J1 all have this form with δ(y) = y, y2, and y2 + 1, respectively. Proposition 5.5 (Alev, Dumas, [4], Proposition 3.6). Let k[y][x; δ] and k[Y ][X; d] be differential operator rings with δ(y) = f ∈ k[y] and d(Y ) = g ∈ k[Y ]. Then k[y][x; δ] ∼= k[Y ][X; d] if and only if there exists λ, α ∈ k× and β ∈ k such that f (y) = λg(αy + β). 12 JASON GADDIS Corollary 5.6. The algebras A1(k), U, J and J1 are all non-isomorphic. Proof. The algebra A1(k) is simple and therefore distinct. The algebra U is not isomorphic to J and J1 since deg(xy − yx) = 1. If J1 ∼= J is an isomorphism, then by Proposition 5.5 there exists α, β ∈ k and λ ∈ k× such that (y2 + 1) = λ(αy + β)2 = λ(α2y2 + 2αβy + β2). Comparing coefficients of y we get that α = 0 or β = 0, a contradiction. (cid:3) Combining with results obtained previously, we can now complete the proof of our main theorem. Proof of Theorem 1.1. Let A and A′ be of the form (1.1) with defining matrices If M ∼sf M ′, then A ∼= A′. By Theorem 4.2 M, M ′ ∈ M3(k), respectively. and Proposition 5.1, we need only show that there are no additional isomorphisms between the algebras in the present theorem. The non-domains Ryx, Rx2 , Rx2−1 and S are all non-isomorphic by Proposition 5.2. The algebra with defining polynomial x2 − y is isomorphic to k[x] via the map x 7→ x and y 7→ x2. This algebra, along with O1(k2) ∼= k[x, y] are the only commutative rings considered, and are distinct from one another. As noted above, A1(k) is simple and therefore distinct from the other domains considered. The domains can be divided, as above, into one of three classes: quantum planes, quantum Weyl algebras, and differential operator rings. By Corollary 5.6, Propo- sition 5.3, and Theorem 5.4, the algebras belonging to each class are distinct from one another with the exceptions Oq(k2) ∼= Oq−1 (k2) and Aq (k). Proving that an algebra belongs to exactly one of these classes requires a study of their automorphism groups. 1(k) ∼= Aq−1 1 If q 6= ±1, then Aut(Aq 1(k)) ∼= (k×) and Aut(Oq(k2)) ∼= (k×)2. On the other 1(k)) ∼= k× ⋊ {ω} and Aut(Oq(k2)) ∼= (k×)2 ⋊ {ω} hand, if q = −1, then Aut(Aq where ω is the involution switching the generators x and y (see [1] and [3]). By counting subgroups of order four, it follows that Ap 1(k) ≇ Oq(k2) for all p, q ∈ k×. In particular, k× has one subgroup of order four and (k×)2 has four. On the other hand, in k× ⋊ {ω} there are two subgroups of order four and in (k×)2 ⋊ {ω} there are eight. As a consequence of Proposition 5.5, the automorphism groups of U, J , or J1 are non-abelian semidirect products of a subgroup of k× by k[y] (see also [4], Proposition 3.6). Thus, these rings are distinct from Oq(k2) and Aq 1(k) except perhaps in the case that q = −1. However, in this case, x2 is central and Oq(k2) and Aq 1(k) are not primitive by [11], Proposition 3.2, whereas differential operator rings over k[y] are always primitive (see, e.g., [8]). (cid:3) Our results can be summed up succinctly in the following theorem. Theorem 5.7. Let A and A′ be of the form (1.1) with defining matrices M, M ′ ∈ M3(k), respectively. If M ∼sf M ′, then A ∼= A′. Conversely, if A ∼= A′, then M ∼sf M ′ unless A ∼= U and A′ ∼= V (or vice-versa). 6. Homogenizations Our goal in this section is to show that sf-congruence applies to algebras of the form (1.2) and prove a result analogous to Theorem 1.1 for these algebras. TWO-GENERATED ALGEBRAS AND STANDARD-FORM CONGRUENCE 13 Theorem 6.1. Let H be for the form (1.2). Then H is isomorphic to one of the following algebras, with one representative of f given in each case: 1(k)), f = xy − qyx − z2 (q ∈ k×), H(Oq(k2)), f = xy − qyx (q ∈ k×), H(Aq H(J ), f = yx − xy + y2, H(U), f = yx − xy + yz, H(Rx2), f = x2, H(Ryx), f = yx, H(k[x]), f = x2 + yz. H(J1), f = yx − xy + y2 + z2, H(V), f = yx − xy + y2 + xz, H(Rx2−1), f = x2 − z2, H(S), f = yx − z2, Furthermore, the above algebras are pairwise non-isomorphic, except H(Oq(k2)) ∼= H(Oq−1 (k2)) and H(Aq 1(k)) ∼= H(Aq−1 1 (k)). The key difference in this situation, versus that in the case of two-generated algebras, is that H(U) and H(V) are non-isomorphic (see Proposition 6.7). Proposition 6.2. The algebras H(k[x]), H(Rx2 ), H(Ryx), H(Rx2−1), and H(S) are not domains. Proof. In the case of H(Rx2) and H(Ryx), this is obvious. In H(k[x]), xyz = x(−x2) = (−x2)x = yzx = yxz ⇒ (xy − yx)z = 0. (6.1) If H(k[x]) is a domain, then (6.1) implies H(k[x]) is commutative. This is impossible since H(k[x])/(z) ∼= Rx2. In H(S), y(xy − z2) = yxy − yz2 = (yx)y − z2y = (yx − z2)y = 0. Finally, in H(Rx2−1), (x + z)(x − z) = x2 − xz + zx − z2 = x2 − z2 = 0. (cid:3) Let H of the form (1.2). Consider the linear map given by, φ(x) = p11x + p12y + p13z, φ(y) = p21x + p22y + p23z, φ(z) = z, (6.2) for pij ∈ k with p11p22 − p12p21 6= 0. One should compare this to the maps (3.1). It is clear that this defines an isomorphism khx, y, z xz − zx, yz − zy, fi → khx, y, z φ(xz − zx), φ(yz − zy), φ(f )i. Moreover, it follows that, since z ∈ Z(H) then z ∈ Z(φ(H)). As in the two- dimension general case, we do not assume these constitute all isomorphisms between algebras of the form (1.2). However, a consequence of Theorem 6.1 is that if two such algebras are isomorphic, then there exists an isomorphism of the above form. We can represent H by a triple of matrices, (X, Y, Z) =    0 0 0 1 0 0 −1 0 0  ,  0 0 0 0 0 −1 0 1 0  ,  m11 m12 m13 m21 m22 m23 0 m33 0    ,  m21 m22 ) not the zero matrix. This representation follows by letting ~x = (6.3) with ( m11 m12 (x y z)T so that xz − zx = ~xT X~x, yz − zy = ~xT Y ~x, f = ~xM ~xT . 14 JASON GADDIS As in Section 3, the matrix M is not uniquely determined for f unless we fix a standard form for M . Let H ′ be another algebra of form (1.2) and suppose φ : H → H ′ is a linear iso- morphism given by (6.2). The algebra H ′ is also defined by a triple, say (X, Y, M ′). Since φ(z) = z, then we let P ∈ P3 be the matrix of φ. It is too much to ask that P T XP = X and P T Y P = Y . We can still hope to preserve those relations up to linear combination. Proposition 6.3. Let P ∈ P3 and let X, Y be as in (6.3). The matrices X and Y are linear combinations of P T XP and P T Y P . Proof. Write Let b2 0 1  . a1 a2 a3 b1 b3 0 P =  0  and V = P T Y P =  a1 a2 0 0 0 0 −b1 −b2 b1 b2 0  . U = P T XP =  0 0 0 0 −a1 −a2 We require r, s ∈ k such that rU +sV = X. That is, ra1 +sb1 = 1 and ra2 +sb2 = 0. Since det P 6= 0, then this system has a solution. Similarly, we can find r′, s′ ∈ k such that r′U + s′V = Y . Hence, P fixes the commutation relations for z. (cid:3) As a consequence of the previous proposition, we have that standard form con- gruence preserves the commutation relations for z. Thus, we extend in a natural way the map (3.3) to this case. Moreover, there is no loss in referring to the ma- trix M as the defining matrix of H. If M ∼sf M ′, then the triples (X, Y, M ) and (X, Y, M ′) define isomorphic algebras. Therefore, if H is of the form (1.2), then H is isomorphic to an algebra given by one of the canonical forms from Theorem 4.2. One could now prove Theorem 6.1 in a manner analogous to Theorem 1.1, that is, by considering ring-theoretic properties of the algebras. However, we take a different approach here by considering prime ideals. This will allow us to apply Theorem 1.1 to this situation. Just as each domain of the form (1.1) can be represented as a skew polynomial ring over k[y], so can each domain of the form (1.2) be represented as a skew polynomial ring over k[y, z] (or k[x, z]). In these cases we can completely determine the prime ideals. In the case that A is not a domain, we can partially describe the prime ideals of H. The following is an immediate corollary of [5], Proposition 2.5. Corollary 6.4. Z(H(J )) = Z(H(J1)) = Z(H(U)) = Z(H(V)) = k[z]. Z(H(Oq(k2))) =(k[xn, yn, z] 1(k))) =(k[xn, yn, z] Z(H(Aq k[z] k[z] if q is a primitive nth root of unity otherwise. if q 6= 1 is a primitive nth root of unity otherwise. We say an algebra H of the form (1.2) has trivial center in case Z(H) = k[z]. The proof of Theorem 6.1 will be easier in this case. TWO-GENERATED ALGEBRAS AND STANDARD-FORM CONGRUENCE 15 Lemma 6.5. Suppose H is of the form (1.2). If P is a prime ideal of H with P ∩ k[z] 6= 0, then P contains z − α for some α ∈ k. Proof. Let g ∈ P ∩ k[z] be nonzero. If g is not irreducible in k[z], then g = g1g2 for some nonconstant g1, g2 ∈ k[z]. Because k[z] is central, then g1Hg2 = g1g2H ⊂ P . The primeness of P implies g1 ∈ P or g2 ∈ P . Hence, P contains az − b for some a, b ∈ k, a 6= 0, and so contains a−1(az − b) = z − a−1b. (cid:3) Before proceeding to the main theorem, we need one additional definition. Let J be an ideal in a ring R and σ ∈ Aut(R). Then J is σ-cyclic if J = J1∩···∩Jn where the Ji are distinct prime ideals of R such that σ−1(Ji+1) = Ji and σ−1(J1) = Jn. Let R be a commutative ring. The radical of an ideal I in R is √I = {a ∈ R an ∈ I for some n}. It is not difficult to see that the radical of an ideal is again an ideal in R. The ideal I is said to be primary if ab ∈ I implies a ∈ I or bn ∈ I for all a, b ∈ R and some n ∈ N. If I is primary, then √I is prime. Suppose H is a domain of form (1.2). If z is not a zero divisor in H, then we can localize at the set C = k[z]\{0}. We refer to this ring as HC . Theorem 6.6. Let H be a domain of the form (1.2). If P is a nonzero prime ideal in H, then one of the following holds: (1) z ∈ P and P corresponds to a prime of H/(z); (2) z − α ∈ P , α ∈ k×, and P corresponds to a prime of H/(z − α); (3) xy − yx ∈ P ; (4) P ∩ k[y, z] = (g1 ··· gn) where the gi are irreducible polynomials in k[y, z] (5) P ∩ k[x, z] = (g1 ··· gn) where the gi are irreducible polynomials in k[x, z] disjoint from k[z]; disjoint from k[z]. Proof. First, suppose P ′ = P ∩ k[z] 6= 0. Then P ′ is a prime ideal of k[z] and so, by Lemma 6.5, z − α ∈ P for some α ∈ k. Now assume P ∩ k[z] = 0. In this case, P extends to a prime ideal in HC . Let R = k(z)[y], then R has Krull dimension one. Let I = P ∩ R. By [10], Theorem 7.2, one of the following must hold: • HC /P is commutative; • I is σ-cyclic for some σ ∈ Aut(R); • I is primary with σ(√I) = √I. If HC /P is commutative, then xy − yx ∈ P . If I is σ-cyclic, then I = P1 ∩ ··· ∩ Pn for distinct prime ideals of R. But the prime ideals of R are exactly extensions of prime ideals of k[y, z] disjoint from k[z]. Therefore, I = (gσ(g)··· σn−1(g)) for some irreducible g ∈ k[y, z], g /∈ k[z], σn(g) = g. Otherwise, I is primary. Since √I is prime, then √I = (g) where g ∈ R is irreducible. We claim I = (gn). Because R is a principal ideal domain, I = (h) for some h ∈ R. Write h = tgn where n is maximal such that g does not divide t. We claim t is a constant. Suppose otherwise, then gm /∈ I for any m > 0. This contradicts g ∈ √I, and so the claim holds and I = (gn). For the remaining case, we need only observe that we can rewrite H as a skew polynomial ring with base ring k[x, z] and repeat. (cid:3) Note that the (4) and (5) occur if and only if H = H(Oq(k2)) or H = H(Aq 1(k)) with q a primitive nth root of unity. This is because in this case the center is nontrivial (Corollary 6.4). By Theorem 6.6, it remains only to consider those prime ideals P whose intersection with the base ring is zero. In this case, it suffices to 16 JASON GADDIS localize the base ring k[y, z] to Q = k(y, z) and consider the extension Q[x; σ, δ]. Then we can appeal to [12], Corollary 2.3. Thus, the prime ideals lying over zero in Q are of the form (g) where g ∈ k[xn, yn, z] is irreducible and such that g /∈ k[yn, z] and g /∈ k[xn, z]. Proposition 6.7. Let A and A′ be of the form 1.1 such that Z(A) = Z(A) = k. Let H = H(A) and H ′ = H(A′), respectively. If H ∼= H ′ then either A ∼= A′ and gr(A) ∼= gr(A′), or else A ∼= gr(A′) and A′ ∼= gr(A). Proof. Suppose Φ : H → H ′ is the given isomorphism. Let I and J be the ideal in H generated by z and z − 1, respectively, and let Φ(I) = I ′, Φ(J) = J ′. Since I and J are generated by a central element in H, then I ′ ∩ k[z] 6= 0 and similarly for J ′. By Proposition 6.5, I ′ = (z − α) and J ′ = (z − β) for some α, β ∈ k. This gives, Φ(H/I) ∼= Φ(H)/Φ(I) ∼= H ′/I ′. Similarly, Φ(H/J) ∼= H ′/J ′. (cid:3) We observe that the previous proposition holds so long as the homogenizing element is regular in H and H ′. We can now show that the non-domains are non-isomorphic. Details are available in [7]. Proposition 6.8. The algebras H(Rx2), H(Rx2−1), H(Ryx), H(S), and H(k[x]) are non-isomorphic. Proof. The algebras H(Rx2), H(Rx2−1) and H(S) are prime whereas H(k[x]) and H(Ryx) are not. Moreover, gk(H(Ryx)) = 2 while H(k[x]) has infinite GK dimen- sion. Similarly, gk(H(S)) = 2 whereas gk(H(Rx2)) = gk(H(Rx2−1)) = ∞. (cid:3) Proof of Theorem 6.1. By Proposition 6.8, we need only consider the domains. In this case, the algebras H(J ), H(J1), H(U), H(V) as well as H(Oq(k2)) and H(Aq 1(k)) for q ∈ k× not a root of unity, are all distinct as a consequence of Proposition 6.7 and then an application of Theorem 1.1. This leaves only the root of unity case for H(Oq(k2)) and H(Aq 1(k)), as well as isomorphisms between those. That H(Op(k2)) ∼= H(Oq(k2)) if and only if p = q±1 1(k)) ∼= follows from Theorem 5.4. Similar methods can be applied to show H(Ap 1(k)) if and only if p = q±1 and H(Op(k2)) ≇ H(Aq H(Aq 1(k)) for all p, q ∈ k×. The reader is referred to [7] for more details. (cid:3) We now summarize our results in a similar manner as to Theorem 5.7. The statement, however, is slightly more satisfying. Theorem 6.9. Let H and H ′ be of the form (1.2) with defining matrices M, M ′ ∈ M3(k), respectively. Then M ∼sf M ′ if and only if H ∼= H ′. Acknowledgements The author would like to thank his advisor, Allen Bell, for suggesting this topic and for his patient guidance and support throughout this project. TWO-GENERATED ALGEBRAS AND STANDARD-FORM CONGRUENCE 17 References [1] J. Alev and M. Chamarie, D´erivations et automorphismes de quelques alg`ebres quantiques, Comm. Algebra 20 (1992), no. 6, 1787 -- 1802. [2] J. Alev and F. Dumas, Sur le corps des fractions de certaines alg`ebres quantiques, J. Algebra 170 (1994), no. 1, 229 -- 265. [3] J. Alev and F. Dumas, Rigidit´e des plongements des quotients primitifs minimaux de q(sl(2)) dans l'alg`ebre quantique de Weyl-Hayashi, Nagoya Math. J. 143 (1996), 119 -- 146. [4] J. Alev and F. Dumas, Invariants du corps de Weyl sous l'action de groupes finis, Comm. U Algebra 25 (1997), no. 5, 1655 -- 1672. [5] J. Gaddis, PBW deformations of Artin-Schelter regular algebras, Journal of Algebra and its Applications (to appear) (2013), available at arXiv:1210.0861. [6] J. Gaddis, Isomorphisms of some quantum spaces, Ring theory and its applications, 2014, pp. 107 -- 116. MR3204354 [7] J. D. Gaddis, PBW deformations of Artin-Schelter regular algebras and their homogeniza- tions, ProQuest LLC, Ann Arbor, MI, 2013. Thesis (Ph.D.) -- University of Wisconsin - Mil- waukee. [8] K. R. Goodearl and R. B. Warfield Jr., An introduction to noncommutative Noetherian rings, Second, London Mathematical Society Student Texts, vol. 61, Cambridge University Press, Cambridge, 2004. [9] R. A. Horn and V. V. Sergeichuk, Canonical forms for complex matrix congruence and *congruence, Linear Algebra Appl. 416 (2006), no. 2-3, 1010 -- 1032. [10] R. S. Irving, Prime ideals of Ore extensions over commutative rings. II, J. Algebra 58 (1979), no. 2, 399 -- 423. [11] E. Kirkman and J. Kuzmanovich, Primitivity of Noetherian down-up algebras, Comm. Alge- bra 28 (2000), no. 6, 2983 -- 2997. [12] A. Leroy and J. Matczuk, Prime ideals of Ore extensions, Comm. Algebra 19 (1991), no. 7, 1893 -- 1907. [13] M. Nafari, M. Vancliff, and J. Zhang, Classifying quadratic quantum P2s by using graded skew Clifford algebras, J. Algebra 346 (2011), 152 -- 164. [14] S. P. Smith, Quantum groups: an introduction and survey for ring theorists, Noncommutative rings (Berkeley, CA, 1989), 1992, pp. 131 -- 178. [15] M. Su´arez-Alvarez and Q. Vivas, Automorphisms and isomorphism of quantum generalized Weyl algebras (2013), available at arXiv:1206.4417. Wake Forest University, Department of Mathematics, P. O. Box 7388, Winston- Salem, NC 27109 E-mail address: [email protected]
1801.06474	2	1801	2018-06-20T15:26:11	Some Minimal Rings Related to 2-Primal Rings	[ "math.RA" ]	In a paper on the taxonomy of 2-primal rings, examples of various types of rings that are related to commutativity such as reduced, symmetric, duo, reversible and PS~I were given in order to show that the ring class inclusions were strict. Many of the rings given in the examples were infinite. In this paper, where possible, examples of minimal finite rings of the various types are given. Along with the rings in the previous taxonomy, NI, abelian and reflexive rings are also included.	math.RA	math	SOME MINIMAL RINGS RELATED TO 2-PRIMAL RINGS STEVE SZABO Abstract. In a paper on the taxonomy of 2-primal rings, exam- ples of various types of rings that are related to commutativity such as reduced, symmetric, duo, reversible and PS I were given in order to show that the ring class inclusions were strict. Many of the rings given in the examples were infinite. In this paper, where possible, examples of minimal finite rings of the various types are given. Along with the rings in the previous taxonomy, NI, abelian and reflexive rings are also included. Keywords: finite rings, reversible rings, semicommutative rings, abelian rings, symmetric rings, minimal rings, NI rings, reflexive rings, reduced rings, duo rings, 2-primal rings 1. Introduction and Overview As finite structures, in particular finite rings, become more and more prevalent and useful in various disciplines, having minimal examples be- comes more important. In coding theory for instance, the most general class of rings that are useful are finite Frobenius rings. This was justi- fied by Wood in [17]. So, it is important to be aware of small Frobenius rings. Another reason for minimal examples of various types of rings is simply to have more tangible examples for understanding. In the case of 2-primal rings, it is helpful to know that U2(F2) is 2-primal but M2(F2) is not. In [12], the taxonomy of 2-primal rings, Marks states that, in reac- tion to a question by T. Y. Lam, he provided an example of a finite reversible nonsymmetric ring, namely F2Q8, the F2 group algebra over the quaternion group. This prompted his "funny little problem" as he puts it. Is F2Q8 a minimal reversible nonsymmetric ring? In [15], the present author shows that F2Q8 is indeed such a ring. It is also shown that F2 hu, vi hu3, v3, u2 + v2 + vu, vu2 + uvu + vuvi is a minimal reversible nonsymmetric ring as well. The two stated examples differ in that fact that F2Q8 is right duo but the other ring is 1 2 STEVE SZABO not. Hence minimality is independent of being duo. Finding minimal rings with a given property is many times not a trivial matter. Since there is no full characterization of finite rings and as of yet there is no way of listing rings of a particular order, finding rings with particular properties especial minimal ones is a tedious process. Minimal rings of various types have been found in recent years: in [14] minimal commutative Frobenius nonchain rings, in [20] minimal noncommutative right duo rings, in [18] minimal noncommutative semi- commutative, in [8] noncommutative reversible and reflexive rings. In- terestingly enough, all such minimal rings are of order 16. Some other well known minimal rings are for instance, F2[x] hx2i and Z4, the smallest chain rings that are not fields, F4[x;σ] hx2i , the smallest noncommutative chain ring, U2(F2), the smallest noncommutative ring. A finite ring with identity is a direct sum of rings of prime power order for distinct primes. Representations of finite rings can be found in [16]. There have been full classifications of finite rings of order pn where p is prime and n ≤ 5: p3 in [5], commutative p4 in [14], noncommutative p4 in [3], nonlocal p5 in [1] and local p5 in [2]. The general classification for n ≥ 6 is still open although there was some initial work done in [7] for n = 6. In [13] there is an extensive set of references provided for the various types of rings covered in the taxonomy. For background on NI rings the reader is refered to [6, 11] and for background on reflexive rings to [8, 9]. Although names for most of the ring types mentioned have been settled in the literature, semicommutative rings appear under al- ternative names. For instance in [13], Marks refers to semicommutative rings as rings having the S I property. They have also been called zero insertive rings. In Section 2, definitions of the various ring types discussed are given as well as the ring class inclusions. It also covers some basic results on these ring types some known minimal rings needed. Section 3 has the main results on minimal rings of various types. It is split between into two subsections, one on reflexive rings and one on nonreflexive rings. 2. Preliminaries In this paper a ring has unity unless otherwise stated. Given a ring R, J(R) is the Jacobson radical of R, N(R) is the set of nilpotent elements of R, Nil∗(R) is the lower nil radical of R (the prime radical of R, the intersection of all prime ideals of R) and Nil∗(R) is the upper nil radical of R (the sum of all nil ideals of R). It is well known that SOME MINIMAL RINGS RELATED TO 2-PRIMAL RINGS 3 Nil∗(R) is a nil ideal and that Nil∗(R) is the unique largest nil ideal of R. So, Nil∗(R) ⊂ Nil∗(R) ⊂ N(R). Definition 2.1. A ring R is ... (1) reduced if N(R) = 0 (equivalently, for all a ∈ R, a2 = 0 implies a = 0). (2) symmetric if for all a, b, c ∈ R, abc = 0 implies bac = 0. (3) reversible if for all a, b ∈ R, ab = 0 implies ba = 0. (4) semicommutative if for all a, b ∈ R, ab = 0 implies aRb = 0. (5) reflexive if for all a, b ∈ R, aRb = 0 implies bRa = 0. (6) right (resp. left) duo if for all a, b ∈ R, ba ∈ aR (ba ∈ Rb) (equivalently, every right (left) ideal of R is 2-sided). A ring that is both right and left duo is simply a duo ring. (7) abelian if each idempotent of R is central. (8) said to satisfy (PS I) if for every a ∈ R, R/annR r (aR) is 2- primal. (9) 2-primal if N(R) = Nil∗(R). (10) NI if N(R) = Nil∗(R) (equivalently, N(R) ⊳ R). Of course a search for minimal rings only needs to consider finite rings. It turns out that in the class of finite rings, there is no distinction between the subclasses of PS I, 2-primal and NI rings. Lemma 2.2. Let R be a finite ring. Then Nil∗(R) = Nil∗(R) = J(R) ⊂ N(R). Furthermore, the following are equivalent. (1) R satisfies PS I (2) R is 2-primal (3) R is NI Proof. It is well known that Nil∗(R) is a nil ideal and that Nil∗(R) is the unique largest nil ideal of R. So, Nil∗(R) ⊂ Nil∗(R) ⊂ N(R). From the definitions it is clear then that PS I implies 2-primal which implies NI. By Proposition 10.27 in [10], Nil∗(R) = Nil∗(R) = J(R) ⊂ N(R). So, if R is NI then R is 2-primal. Since R is semilocal (R/J(R) is semisimple) and J(R) is nil (J(R) ⊂ N(R)), by Proposition 3.15 of [13] which says a 2-primal semilocal ring with nil jacobson radical satisfies PS I, if R is 2-primal it satisfies PS I. (cid:3) From now on, NI rings will be used when any of the 3 types in the previous lemma are needed. Lemma 2.3 (Proposition 3 in [19]). A finite ring is right duo if and only if it is left duo. Lemma 2.4. A finite reduced ring is commutative. 4 STEVE SZABO Proof. From Theorem 12.7 in [10], a reduced ring is subdirect product of domains. Since a finite domain is a field, a finite reduced ring is commutative. (cid:3) Lemma 2.5. A semicommutative ring is abelian. Proof. Let R be a semicommutative ring and e an idempotent of R. Then e(1 − e) = 0 and (1 − e)e = 0. Since R is semicommutative, eR(1 − e) = 0 and (1 − e)Re = 0. So, for a ∈ R, ea = ea(1 − e + e) = eae = (1 − e + e)ae = ae and e is central. (cid:3) Lemma 2.6. Let R be a local ring. If R/J(R) is a prime field and J(R)3 = 0 then R is semicommutative. Proof. Assume J(R)3 = 0. Let a, b ∈ R and assume ab = 0. Then a, b ∈ J(R). Let r ∈ R. Now, r = α + r′ where α ∈ R/J(R) and r′ ∈ J(R). So, arb = a(α + r′)b = aαb + ar′b = αab + ar′b = 0 since α ∈ Z(R) and ar′b ∈ J(R)3 = 0. Hence, R is semicommutative. (cid:3) Lemma 2.7. A ring is reversible if and only if it is semicommutative and reflexive. Proof. Clearly, a reversible ring is semicommutative and reflexive. Let R be semicommutative and reflexive. Let a, b ∈ R and assume ab = 0. By semicommutativity, aRb = 0. By reflexivity, bRa = 0. So, ba = 0 and R is reversible. (cid:3) In [1] it was shown that given a finite ring R there exists an orthog- onal set of idempotents {e1, . . . , em} ⊂ R such that e1 + · · · + em = 1, Ri = eiRei is a primary ring (a full matrix ring over a local ring) and for i 6= j, Mij = eiRej is an (Ri, Rj)-bimodule contained in J(R). Fur- thermore, letting S = ⊕iRi and M = ⊕i6=jMij then R = S ⊕ M, S is a subring of R and M is an additive subgroup of J(R) which is an S-bimodule. For the purposes of the work herein, this will simply be called the decomposition of the finite ring R. Clearly, J(R) = J(S)⊕M. Notice, if M 2 = 0 then multiplication in R can be modeled as (2.1) (s, u)(t, v) = (st, su + vt). Lemma 2.8. Let R be a finite ring with ring decomposition {e1, . . . , em}, Ri = eiRei, Mij = eiRej, S = ⊕iRi and M = ⊕i6=jMij. Then (1) R is abelian if and only if M = 0 and Ri is local for i ∈ {1, . . . , m}. (2) R is NI if and only if Ri is local for i ∈ {1, . . . , m}. SOME MINIMAL RINGS RELATED TO 2-PRIMAL RINGS 5 (3) If R is abelian then R is NI. (4) If M 2 = 0 and M 6= 0 then R is nonreflexive. Proof. (1) This follows directly from the definition of an abelian ring. (2) Assume R1 is not local. Since R1 is a primary ring, R1 = Mn(T ) for n > 1 and some local ring T . Let eij be the (i, j) matrix unit of R1. Then e12 and e21 are nilpotent but e12 + e21 is not. So, R in this case is not NI. Now assume Ri is local for i ∈ {1, . . . , m}. Then N(S) = J(S). By Lemma 2.2, J(R) ⊂ N(R). Let r ∈ N(R). Then r = s + u for some s ∈ S and u ∈ M ⊂ J(R) and for some n, 0 = (s + u)n = sn + t for some t ∈ J(R) since M ⊂ J(R). So, sn = −t and then sn ∈ J(R)∩S = N(S). Hence, s ∈ N(S) = J(S) and r ∈ J(R) showing N(R) = J(R) and R is NI. (3) Follows from (1) and (2). (4) Assume M 2 = 0 and M 6= 0. Renumbering if necessary, it may be assumed that M12 6= 0. Since M 2 = 0, multiplication in R can be modeled as in 2.1. Let d ∈ M12 \ 0 and (s, u) ∈ R. Then (0, d)(s, u)(e1, 0) = (0, d)(se1, ue1) = (0, 0). showing (0, d)R(e1, 0) = 0. But, (e1, 0)(e1, 0)(0, d) = (0, d) showing (e1, 0)R(0, d) 6= 0. Hence, R is nonreflexive. (cid:3) Figure 2 has the ring class inclusions of the types of finite rings being considered. The implications not covered in the lemmas above follow directly from the definitions. red. 3 comm. symm. 3 rev. ◆◆◆◆◆◆ ◆◆◆◆◆◆ refl. + <qqqqqqq qqqqqqq ▼▼▼▼▼▼▼ ▼▼▼▼▼▼▼ duo 3 s.comm. 3 abel. 3 NI Figure 1. Finite Rings It is easily seen that if * is one of the ring types being considered, a finite ring is of type * if and only if every ring direct summand is of type . This is why when searching for minimal rings of these types, only indecomposable rings need to be checked. The only exception is a non-NI nonreflexive ring (see Theorem 3.7). Since a finite ring with identity is a direct sum of rings of prime power order for distinct primes, the search is further limited to rings of prime power order. + + 3 " + " * 4 k s + + + 6 STEVE SZABO Proposition 2.9 ([4]). For prime p, there are no noncommutative rings of order p or p2 and the only noncommutative ring of order p3 is U2(Fp). Proposition 2.10 ([3] or Corollary 1.10 in [1]). Let p be prime. The noncommutative nonlocal indecomposable rings of order p4 are 0 • F • (cid:18) Z p[x] hx2i Fp Fp ! Fp Fp (cid:19) 0 • M2(Fp) p2 p Fp p[x] hx2i 0 p2 Fp 0 Fp ! • F • (cid:18) Z Fp (cid:19) • (cid:18) Fp F2 0 Fp (cid:19) •     a 0 d 0 0 a 0 0 0 0 b 0 0 c 0 b   a, b, c, d ∈ Fp  p2 and M2(Fp) ⊕ F Proposition 2.11. Let p be prime. There are no non-NI rings of order p, p2 or p3. The only non-NI ring(s) of order p4 is M2(Fp), of order p5 p[x] is M2(Fp) ⊕ Fp and of order p6 are M2(Fp) ⊕ Fp ⊕ Fp, M2(Fp) ⊕ F hx2i , p2. The minimal non-NI ring is M2(F2). M2(Fp) ⊕ Z Proof. Since commutative rings are NI, by Lemma 2.8(2) and Propo- sition 2.9, all rings of order p, p2 or p3 are NI. By Lemma 2.8(2) and Proposition 2.10, the only ring of order p4 that is not NI is M2(Fp) and of order p5 is M2(Fp) ⊕ F2. So, the minimal non-NI ring is M2(F2). Knowing that F2 ⊕ F2, F2[x] p2 are the rings of order 4, by Lemma 2.8(2), the rings M2(Fp) ⊕ F2 ⊕ F2, M2(Fp) ⊕ F2[x] hx2i , M2(Fp) ⊕ Z4 p2 are the only decomposable non-NI rings of order p6. and M2(Fp) ⊕ F The only possible indecomposable non-NI ring of order p6 would be a ring with components R1 = M2(Fp) and R2 = Fp and M = p in its ring decomposition. But, the minimal submodules of M2(Fp) are isomorphic to F2 (cid:3) hx2i , Z4 and F p so no such ring exists. Remark 2.12. The set of noncommutative local rings of order 16 can be deduced from work in [1] or [3]. It turns out that they are precisely the minimal noncommutative semicommutative rings which is the sub- ject of Proposition 2.13. Note σ denotes the Frobenius automorphism on F4. Proposition 2.13 ([18] Theorem 8). A minimal noncommutative semi- commutative ring has order 16. The complete list of such rings is SOME MINIMAL RINGS RELATED TO 2-PRIMAL RINGS 7 • F4[x;σ] hx2i • F2hu,vi hu3,v3,vu,u2−uv,v2−uvi • • • Z4hu,vi hu3,v3,vu,u2−uv,v2−uv,2−uv,2u,2vi F2hu,vi hu3,v2,vu,u2−uvi Z4hu,vi hu3,v2,vu,u2−uv,2−uv,2u,2vi . Two other results on rings of order 16 which are necessary for our classification are provided here. Proposition 2.14 ([20] Theorem 3). A minimal noncommutative duo ring has order 16. The complete list of such rings is • F4[x;σ] hx2i • • F2hu,vi hu3,v3,vu,u2−uv,v2−uvi hu3,v3,vu,u2−uv,v2−uv,2−uv,2u,2vi . Z4hu,vi Proposition 2.15 ([8] Theorem 5). A minimal noncommutative re- flexive ring has order 16. The complete list of such rings is • F4[x;σ] hx2i • M2(F2). 3. Minimal Rings In this section the minimal rings of the various types according to the diagram in Figure 2 are identified. This will be done first for nonreflexive rings and then for reflexive rings. 3.1. Minimal Nonreflexive Rings. In the context of finite nonre- flexive rings, the ring class inclusions are duo 3 semicommutative 3 abelian 3 NI Theorem 3.1 (Minimal Nonreflexive Duo). A minimal nonreflexive duo ring has order 16. The complete list of such rings is • • F2hu,vi hu3,v3,vu,u2−uv,v2−uvi hu3,v3,vu,u2−uv,v2−uv,2−uv,2u,2vi . Z4hu,vi Proof. Follows from Propositions 2.14 and 2.15. (cid:3) Theorem 3.2 (Minimal Nonreflexive Semicommutative Nonduo). A minimal nonreflexive semicommutative nonduo ring has order 16. The complete list of such rings is • F2hu,vi hu3,v2,vu,u2−uvi + + + 8 STEVE SZABO • hu3,v2,vu,u2−uv,2−uv,2u,2vi . Z4hu,vi Proof. Follows from Propositions 2.13, 2.14 and 2.15. (cid:3) From Remark 2.12 and Proposition 2.9, any noncommutative abelian ring of order less than 32 is local semicommutative. Therefore, a non- semicommutative abelian ring is of order at least 32. The next propo- sition shows that such a ring must actually be at least of order 64. Proposition 3.3. A local ring of order less than 64 is semicommuta- tive. Proof. Commutative rings are semicommutative. Let p be a prime. By Proposition 2.9, there are no noncommutative local rings of order p, p2 or p3. By Remark 2.12, there are no nonsemicommutative local rings of order p4. So, a local ring of order less than 32 is semicommutative. For rings of order 32, by Lemma 2.6, only the local noncommutative rings where the cube of its jacobson radical is nonzero need to be checked for semicommutativity. Such rings can be found in the classification of rings of order 32 in [2] which are given in cases 1.2.1, 2.2.b and 2.2.c. There are 7 such rings which are listed here. (1) (2) (3) (4) (5) (6) (7) F2hu,vi hu4,uv,vu−u3,v2i F2hu,vi hu4,uv,vu−u3,v2−u3i Z4hu,vi hu4,uv,vu−u3,v2,u3−2i Z4hu,vi hu4,uv,vu−u3,v2−u3,u3−2i Z4hu,vi hu4,uv,vu−u3,v2,u2−2i Z4hu,vi hu4,uv,vu−u3,v2−u3,u2−2i hu4,uv,vu−u3,v2,u2−2−2ui. Z4hu,vi Let R be any of the rings above. Then R is local, J(R) = F2u+F2v+J 2, J 2 = F2u2 + J 3, J 3 = F u3 and J 4 = 0. In a finite local ring, the jacobson radical is the set of zero divisors. Let a, b ∈ J(R) and c ∈ R. Then for some αi, βi, γi, δi, r ∈ F2, a = α1u + β1v + γ1u2 + δ1u3, b = α2u + β2v + γ2u2 + δ2u3 and c = r + α3u + β3v + γ3u2 + δ3u3. Assume ab = 0. Then 0 = ab = (α1u + β1v + γ1u2 + δ1u3)(α2u + β2v + γ2u2 + δ2u3) = α1α2u2 + x for some x ∈ J(R)3 showing α1α2 = 0. Now, acb = arb+α1α3α2u3 = 0. So, if ab = 0 then aRb = 0 and R is semicommutative. (cid:3) SOME MINIMAL RINGS RELATED TO 2-PRIMAL RINGS 9 Theorem 3.4 (Minimal Nonreflexive Abelian Nonsemicommutative). A minimal nonreflexive abelian nonsemicommutative ring is of order 64 an example of which is F2 hu, vi hu2, v2, uvu − vuvi Proof. By Proposition 3.3, an abelian ring of order less than 64 is semi- commutative. Let R = F2 hu, vi hu2, v2, uvu − vuvi . R is local, J(R) = F2u + F2v + J 2, J 2 = F2uv + F2vu + J 3, J 3 = F2uvu and J 4 = 0. So, R = 64. Notice u2 = 0 but uvu 6= 0 showing R is nonsemicommutative. To see that R is nonreflexive, notice, for α, β, γ, δ, ǫ, ζ ∈ F , u(α + βu + γv + δuv + ǫvu + ζuvu)uv = αu2v = 0 but 0 6= uvu ∈ uvRu. So, uRuv = 0 but uvRu 6= 0 showing R is nonreflexive. (cid:3) Remark 3.5. The ring F2 hu, vi hu2, v2, uvu − vuvi was shown to be a minimal nonreflexive abelian nonsemicommutative ring. From the proof it can be seen that it is actually minimal among all nonsemicommutative rings, reflexive or not. Theorem 3.6 (Minimal Nonreflexive NI Nonabelian). The minimal NI nonabelian nonreflexive ring is U2(F2) and is of order 8. Proof. By Proposition 2.15, U2(F2) is nonreflexive. Clearly U2(F2) is the minimal nonabelian ring since it is not abelian and is the unique minimal noncommutative ring. By Lemma 2.8(2), U2(F2) is NI. (cid:3) Theorem 3.7 (Minimal Non-NI Nonreflexive). A minimal non-NI nonreflexive ring has order 128. The complete list of such rings is M2(F2) (cid:18) F2 0 F2 0 (cid:19) 0 F2 R =    , Rop and M2(F2) ⊕ U2(F2). Proof. By Proposition 2.11, the only non-NI indecomposable rings of order less than 128 are M2(F2) and M2(F3). But, these are both re- flexive. By Theorem 3.6, the smallest nonreflexive ring is U2(F2) which is unique. So, M2(F2) ⊕ U2(F2) is the unique decomposable minimal non-NI nonreflexive of order 128. 10 STEVE SZABO An indecomposable example would have to have at least one of the component rings in its ring decomposition to be nonlocal by Lemma 2.8. Since the minimal right or left modules of M2(F2) are isomorphic to F2 2, the only indecomposable non-NI non-reflexive rings of order 128 are and R =  Rop =  M2(F2) (cid:18) F2 0 F2 0 (cid:19) 0 F2   0 (cid:19) F2   , 0 M2(F2) (cid:18) F2 F2 0 which is clearly not isomorphic to R. By the uniqueness of the compo- nents, it is clear that there are no other such rings. (cid:3) 3.2. Minimal Reflexive Rings. In this section minimal reflexive rings are identified. The following diagram shows the ring class inclusions for finite reflexive rings. reduced 3 commutative symmetric ❘❘❘❘❘❘ ❘❘❘❘❘❘ 9❦❦❦❦❦❦❦ ❦❦❦❦❦❦❦ ❚❚❚❚❚❚❚❚❚❚ ❚❚❚❚❚❚❚❚❚❚ reversible 9❧❧❧❧❧❧❧❧❧ ❧❧❧❧❧❧❧❧❧ duo 3 abelian 3 NI Of course F2 is the minimal reduced ring and the minimal nonreduced commutative rings are the indecomposable rings of order 4 that are not fields namely F2[x] hx2i and Z4. Theorem 3.8 (Minimal Symmetric Duo Noncommutative). The min- imal symmetric duo noncommutative ring is F4[x;σ] hx2i and is of order 16. Proof. Let R = F4[x;σ] hx2i . By Propositions 2.13 and 2.15, R is the minimal semicommutative reflexive duo ring. By Lemma 2.7, R is reversible. In [15] it was shown that a minimal reversible nonsymmetric ring is of order 256. Since R = 16, R is symmetric. The symetricity of R can easily be shown directly. (cid:3) Theorem 3.9 (Minimal Symmetric Nonduo). A minimal symmetric nonduo ring is of order 32 an example of which is F2 hu, vi hu3, v2, u2 + uv + vu, uvui . % - + % - 1 + + 1 SOME MINIMAL RINGS RELATED TO 2-PRIMAL RINGS 11 Proof. By Remark 2.12, the rings listed in Proposition 2.13 are the noncommutative local rings of order 16. By inspection, it is clear that none of these rings are reversible hence they are nonsymmetric. For a prime p, by Proposition 2.9, any local ring of order p, p2 or p3 is com- mutative. So, any symmetric ring of order less than 32 is commutative and therefore duo. Let R = F2 hu, vi hu3, v2, u2 + uv + vu, uvui . First note that any third degree monomial in u and v is 0. It is easy to see then that R is an F2-algebra with basis {1, u, v, uv, vu} and R = 32. Next, R is not left duo since v(u + v) = vu /∈ {av + buva, b ∈ F2} = Rv. By Lemma 2.3, R is not duo. To show R is reversible, assume (a1 + b1u + c1v + d1uv + e1vu)(a2 + b2u + c2v + d2uv + e2vu) = 0 for ai, bi, ci, di, ei ∈ F2. Since R is local, a1 = a2 = 0. Then 0 = (b1u + c1v + d1uv + e1vu)(b2u + c2v + d2uv + e2vu) = (b1c2 + b1b2)uv + (c1b2 + b1b2)vu showing (b1c2 + b1b2) = 0 and (c1b2 + b1b2) = 0. So, 0 = (b1c2 + b1b2)vu + (c1b2 + b1b2)uv = (a2 + b2u + c2v + d2uv + e2vu)(a1 + b1u + c1v + d1uv + e1vu) and R is reversible. From [15], a reversible ring of order less than 256 is symmetric. Since R = 32, R is symmetric. Hence, R is a minimal symmetric nonduo ring. (cid:3) Theorem 3.10 (Minimal Reversible Nonsymmetric Right Duo). A minimal reversible nonsymmetric right duo ring is of order 256 an ex- ample of which is F2Q8. Proof. One of the main results of [15]. (cid:3) Theorem 3.11 (Minimal Reversible Nonsymmetric Nonduo). A min- imal reversible nonsymmetric nonduo ring is of order 256 an example of which is F2 hu, vi hu3, v3, u2 + v2 + vu, vu2 + uvu + vuvi . Proof. One of the main results of [15]. (cid:3) 12 STEVE SZABO Remark 3.12 (Reflexive Abelian Nonsemicommutative). A minimal reflexive abelian nonsemicommutative ring has turned out to be more elusive than the rest. By Proposition 3.3, a minimal reflexive abelian nonsemicommutative ring is of order at least 64. At the moment, there is no known finite local reflexive nonsemicommutative ring. In [13], Lemma 5.3 states that there exists an infinite local ring R that contains a prime ideal P that is not completely prime. So, R/P is a local ring that is prime but not completely prime i.e. a domain, meaning R/P is a prime ring with zero divisors. It is easy to show that a prime ring is reflexive and that prime ring with zero divisors is not semicommutative. Hence, R/P is a local reflexive nonsemicommutative ring. So, such rings do exist. It is important to note that according to [10] Exercise 13.3, a finite prime ring is a matrix ring over a finite field. Hence, a local finite prime ring is a finite field. This shows there does not exist a finite local prime ring with zero divisors. So, if there does exist a finite reflexive abelian nonsemicommutative ring, it is not a prime ring with zero divisors. Example 3.13. The following ring is a nonabelian NI reflexive ring of order 64. On the additive group R = F2[x] hx2i ⊕ F2[x] hx2i ⊕ F2 ⊕ F2 define multiplication as (α1 + β1x, γ1 + δ1x, ǫ1, ζ1)(α2 + β2t, γ2 + δ2x, ǫ2, ζ2) = (cid:0)ǫ1ζ2x + (α1 + β1x)(α2 + β2x), ǫ2ζ1x + (γ1 + δ1x)(γ2 + δ2x), α1ǫ2 + γ2ǫ1, γ1ζ2 + α2ζ1(cid:1) for αj, βj, γj, δj, ǫj, ζj ∈ F2. It is straight forward to check that R is a ring with 1 = (1, 1, 0, 0). Furthermore, e1 = (1, 0, 0, 0) and e2 = (0, 1, 0, 0) are idempotents which give the ring decomposition of R where the direct summands are R1, R2, M12 and M21 respectively. First, R is nonabelian since M 6= 0. Secondly, since R1 and R2 are local, by Lemma 2.8(2), R is NI. Finally, it will be shown that R is also reflexive. Let a, b ∈ R and assume aRb = 0. So a = (a1, a2, a12, a21) and hx2i and a12, a21, b12, b21 ∈ F2. b = (b1, b2, b12, b21) for a1, a2, b1, b2 ∈ F2[x] Then 0 = a(1, 0, 0, 0)b = (a1b1, a21b12x, a1b12, a21b1), SOME MINIMAL RINGS RELATED TO 2-PRIMAL RINGS 13 0 = a(0, 1, 0, 0)b = (a12b21x, a2b2, a12b2, a2b21), 0 = a(0, 0, 1, 0)b = (a1b21x, a21b2x, a1b2, 0) and 0 = a(0, 0, 0, 1)b = (a12b1x, a2b12x, 0, a2b1). This shows that a1b1 = a2b2 = 0 and a21b12, a12b21, a1b21, a21b2, a12b1, a2b12, a1b12, a21b1, a12b2, a2b21, a1b2, a2b1 ∈ hxi . So, for r = (r1, r2, r12, r21) ∈ R, bra = (b1, b2, b12, b21)(r1, r2, r12, r21)(a1, a2, a12, a21) = (b1r1a1 + b1r12a21x + b12r2a21x + b12r21a1x, b2r2a2 + b2r21a12x + b21r1a12x + b21r12a2x, b1r1a12 + b1r12a2 + b12r2a2, b2r2a21 + b2r21a1 + b21r1a1) = 0. Hence, R is reflexive. Theorem 3.14 (Minimal NI Nonabelian Reflexive). A minimal NI nonabelian reflexive ring is of order 64 an example of which is the additive group F2[x] hx2i with multiplication as in Example 3.13. F2[x] hx2i R = ⊕ ⊕ F2 ⊕ F2 Proof. For a prime p, Proposition 2.9 shows that the only indecompos- able nonabelian ring of order p, p2 or p3 is U2(Fp). But this ring is nonreflexive (see Theorem 3.6). From Lemma 2.8(4) and Proposition 2.10, the only indecomposable nonabelian reflexive ring of order p4 is M2(Fp) since all the others have M 6= 0 but M 2 = 0 in their ring de- composition. But M2(Fp) is non-NI by Proposition 2.11. In [1], it was shown that there are only 2 rings of order 32 which have a ring decom- position with M 2 6= 0 (see Lemma 1.5 in [1]). In either one of those two rings, (0, 0, 1, 0)R(0, 0, 0, 1) 6= 0 but (0, 0, 0, 1)R(0, 0, 1, 0) = 0 so they are nonreflexive. By Lemma 2.8, a nonabelian NI ring with M 2 = 0 (nonabelian NI guarantees M 6= 0) is nonreflexive. So, all indecompos- able nonabelian rings of order 32 are nonreflexive. Hence, there are no NI nonabelian reflexive rings of order less than 64. R was shown to be an NI nonabelian reflexive ring of order 64 in Example 3.13. (cid:3) Theorem 3.15 (Minimal Non-NI Reflexive). The minimal non-NI re- flexive rings is M2(F2). 14 STEVE SZABO Proof. By Proposition Proposition 2.15, there are two minimal non- commutative reflexive rings, F4[x;σ] hx2i and M2(F2). By Proposition 2.13, F4[x;σ] is semicommutative and therefore NI. By Lemma 2.8(2), M2(F2) (cid:3) hx2i is not NI. 4. Conclusion Minimal rings of all but one of the types of rings under consideration were found. Reflexive abelian nonsemicommutative, proved more elu- sive and is left for future consideration. From Proposition 3.3, such a ring is at least of order 64. There is no known example of a finite local reflexive nonsemicommutative ring. It was however shown in Remark 3.12 that an infinite ring of this type exists. References 1. B. Corbas and G. D. Williams, Rings of order p 5. I. Nonlocal rings, J. Algebra 231 (2000), no. 2, 677 -- 690. MR 1778165 2. , Rings of order p 5. II. Local rings, J. Algebra 231 (2000), no. 2, 691 -- 704. MR 1778166 3. J. B. Derr, G. F. Orr, and Paul S. Peck, Noncommutative rings of order p Pure Appl. Algebra 97 (1994), no. 2, 109 -- 116. MR 1312757 (95m:16012) 4, J. 4. K. E. Eldridge, Orders for finite noncommutative rings with unity, Amer. Math. Monthly 75 (1968), 512 -- 514. MR 0230772 (37 #6332) 5. Robert Gilmer and Joe Mott, Associative rings of order p 3, Proc. Japan Acad. 49 (1973), 795 -- 799. MR 0369422 6. Seo Un Hwang, Young Cheol Jeon, and Yang Lee, Structure and topological conditions of NI rings, J. Algebra 302 (2006), no. 1, 186 -- 199. MR 2236599 7. P. Karimi Beiranvand, R. Beyranvand, and M. Gholami, Classification of finite 6 by generators and relations, J. Math. (2013), Art. ID 467905, rings of order p 8. MR 3096801 8. Byung-Ok Kim and Yang Lee, Minimal noncommutative reversible and reflexive rings, Bull. Korean Math. Soc. 48 (2011), no. 3, 611 -- 616. MR 2827769 9. Nam Kyun Kim and Yang Lee, Extensions of reversible rings, J. Pure Appl. Algebra 185 (2003), no. 1-3, 207 -- 223. MR 2006427 10. T. Y. Lam, A first course in noncommutative rings, second ed., Graduate Texts in Mathematics, vol. 131, Springer-Verlag, New York, 2001. MR MR1838439 (2002c:16001) 11. Greg Marks, On 2-primal Ore extensions, Comm. Algebra 29 (2001), no. 5, 2113 -- 2123. MR 1837966 12. 13. , Reversible and symmetric rings, J. Pure Appl. Algebra 174 (2002), no. 3, 311 -- 318. MR 1929410 , A taxonomy of 2-primal rings, J. Algebra 266 (2003), no. 2, 494 -- 520. MR 1995125 14. Edgar Mart´ınez-Moro and Steve Szabo, On codes over local Frobenius non- chain rings of order 16, Noncommutative Rings and Their Applications (Steven Dougherty, Alberto Facchni, Andr´e Leroy, Edmund Puczylowski, and Patrick SOME MINIMAL RINGS RELATED TO 2-PRIMAL RINGS 15 Sol´e, eds.), Contemp. Math., vol. 634, Amer. Math. Soc., Providence, RI, 2015, pp. 227 -- 243. 15. Steve Szabo, Minimal reversible nonsymmetric rings, submitted (2017). 16. Robert S. Wilson, Representations of finite rings, Pacific J. Math. 53 (1974), 647 -- 649. MR 0369423 17. Jay A. Wood, Duality for modules over finite rings and applications to cod- ing theory, Amer. J. Math. 121 (1999), no. 3, 555 -- 575. MR MR1738408 (2001d:94033) 18. Liqiong Xu and Weimin Xue, Structure of minimal non-commutative zero- insertive rings, Math. J. Okayama Univ. 40 (1998), 69 -- 76 (2000). MR 1755920 (2001b:16036) 19. Weimin Xue, On strongly right bounded finite rings, Bull. Austral. Math. Soc. 44 (1991), no. 3, 353 -- 355. MR 1138010 20. , Structure of minimal noncommutative duo rings and minimal strongly bounded nonduo rings, Comm. Algebra 20 (1992), no. 9, 2777 -- 2788. MR 1176837 (93i:16003) Department of Mathematics and Statistics, Eastern Kentucky Uni- versity, Richmond, KY 40475 E-mail address: [email protected]
1708.01963	2	1708	2017-08-24T12:35:58	Classification of three-dimensional Jordan superalgebras	[ "math.RA" ]	In this paper, we classify Jordan superalgebras of dimension up to three over an algebraically closed field of characteristic different of two. Our main motivation to obtain such classification comes out from the intention to give an answer to the problem of determining the minimal dimension of exceptional Jordan superalgebras. Our strategy to provide a lower bounded for this dimension is to determine the complete list of Jordan superalgebras of small dimensions and verify which ones are special or exceptional.	math.RA	math	Classification of three-dimensional Jordan superalgebras aInstituto de Matemática e Estatística. Universidade de São Paulo. R. Matão 1010, 05508-090. São Paulo, SP, Brazil. M. E. Martin1,a, Abstract In this paper, we classify Jordan superalgebras of dimension up to three over an algebraically closed field of characteristic different of two. We also prove that all such superalgebras are special. Keywords: Jordan superalgebras, classification of superalgebras, speciality 2010 MSC: 17C27, 17C40, 17C70. Introduction The goal of this paper is to classify Jordan superalgebras of dimension up to three over an algebraically closed field of characteristic different of 2. Our main motivation to obtain such classification comes out from the intention to give an answer to the problem of determining the minimal dimension of exceptional Jordan superalgebras posed in [11]. This problem is analogous to the problem presented in [3] by H. Petersson and A. M. Slin'ko for ordinary Jordan algebras. Our strategy to provide a lower bounded for this dimension is to determine the complete list of Jordan superalgebras of small dimensions and verify which ones are special or exceptional. The classification problem of the algebraic structures of given dimension, i.e. find the list of all non-isomorphic objects in this structure, has been extensively studied. But in the context of graded algebras the results are relatively scarce. Most of the known results focus on the theory of graded Lie algebras. In [8], V. G. Kac gave a brief account of the main features of the theory of finite dimensional Lie superalgebras. The author announced in [6] the complete solution of the classification problem of simple Lie superalgebras under the stimuli of the physicists' interest in the subject. In the context of nonsimple Lie superalgebras, in 1978 N. Backhouse gave a classification of real Lie superalgebras, which are not Lie algebras, up to dimension four (see [1] and [2]). He also noted that none of them are simple. In 1999, A. Hegazi in [4] proved a version of the Kirillov lemma for nilpotent Lie superalgebras and applied its to the classification of nilpotent Email address: [email protected] (M. E. Martin) 1The author was supported by CAPES, PNPD Program, IME-USP. Lie superalgebras up to dimension five. Classifications of five dimensional Lie superalgebras over the complex and real numbers are contributions of N. L. Matiadou and A. Fellouris, see [15, 2005] and [16, 2007] respectively. As expected, there was a similar theory for graded Jordan algebras. The paper [10] of I. Kaplansky, lays the foundations for the study of the Jordan superalgebras case. In 1977, V. G. Kac developed the Kantor's method and applied it to obtain the classification of finite di- mensional simple Jordan superalgebras over an algebraically closed field of characteristic 0 from the classification of simple graded Lie algebras (see [7]). There was one missing case in this classification that was later described by I. L. Kantor in [9]. More than two decades later in [18], M. L. Racine and E. I. Zel'manov took advantage of Jordan theoretic methods, like Peirce decomposition and representation theory, that are less sensitive to the characteristic of the base field to obtain a classification of simple Jordan superalgebras over fields of charac- teristic different from 2 whose even part is semisimple. The remaining case, namely simple Jordan superalgebras with nonsemisimple even part, was tackled in [13] by C. Martinez and E. Zel'manov. However, nonsimple Jordan superalgebras are very plentiful and as yet do not have re- ceived enough attention. In what follows we present the classification of Jordan superalgebras of small dimensions and we determine in each case whether the superalgebras obtained are special or exceptional. Regarding special or exceptional Jordan superalgebras we highlight some central results. In [14] the authors construct universal associative enveloping algebras for a large class of simple Jordan superalgebras showing they are special. On the other hand, in [17] has been proved that the 10-dimensional simple Kac superalgebra is exceptional and showed some new exceptional simple superalgebras appear in characteristic 3. As for exceptional Jordan superalgebras that are not simple, I. Shestakov and others in [11, 2000] have constructed an example of an exceptional Jordan superalgebra of dimension 7 over an algebraically closed field of characteristic different of 2. So far this is the smallest known example. The paper is organized as follows. In Section 1, we recall the basic concepts, examples and some results for finite-dimensional Jordan superalgebras. In Section 2, we describe all non- isomorphic Jordan superalgebras of dimension less or equal to 3 and, finally, we show that all superalgebras are special. 1 Preliminaries In what follows k will represent a field of characteristic not 2. A Jordan algebra J is a commutative k-algebra that satisfies the Jordan identity a2 · (b · a) = (a2 · b) · a (1.1) or, equivalently, the linearization 2 (a·b)·(c·d) + (a·c)·(b·d) + (a·d)·(b·c) = ((a·b)·c)·d + ((a·d)·c)·b + ((b·d)·c)·a for any a, b, c, d ∈ J. In [12] all Jordan algebras up to dimension four over an algebraically closed field were described. To fix notation we present the indecomposable Jordan algebras of dimension less or equal to 3 in Table 1. This algebras will be used in the development we present below. Table 1: Indecomposable Jordan algebras of dimension 6 3 J U1 U2 B1 B2 B3 T1 T2 T3 T4 T5 T6 T7 T8 T9 T10 Multiplication table dim e2 1 = e1 e2 1 = 0 e1 · e2 = e2 e1 · e2 = 1 2 e2 e1 · e2 = 0 e2 2 = 0 e2 2 = 0 e2 2 = 0 e2 1 = e1 e2 1 = e1 e2 1 = e2 e2 1 = e1 e2 2 = e3 e1 · e2 = e2 e1 · e3 = e3 e2 1 = e1 e2 2 = 0 e1 · e2 = e2 e1 · e3 = e3 e2 1 = e2 e2 2 = 0 e1 · e2 = e3 e1 · e3 = 0 e2 1 = e2 e2 2 = 0 e1 · e2 = 0 e1 · e3 = e2 e2 3 = 0 e2 · e3 = 0 e2 3 = 0 e2 · e3 = 0 e2 3 = 0 e2 · e3 = 0 e2 3 = 0 e2 · e3 = 0 e2 1 = e1 e1 · e2 = 0 e2 2 = e2 e1 · e3 = 1 e2 3 = e1 + e2 2 e3 e2 2 = 0 e2 3 = 0 e2 · e3 = 1 2 e3 e1 · e3 = e3 e2 · e3 = 0 e2 2 = 0 e1 · e3 = 1 e2 3 = 0 2 e3 e2 3 = 0 e2 · e3 = 0 e2 2 = e3 e1 · e3 = 0 e2 · e3 = 0 e2 2 = e3 e1 · e3 = e3 e2 3 = 0 e2 · e3 = 0 e2 2 = e2 e1 · e3 = 1 2 e3 e2 3 = 0 e2 · e3 = 1 2 e3 e2 1 = e1 e1 · e2 = 1 2 e2 e2 1 = e1 e1 · e2 = 1 2 e2 e2 1 = e1 2 e2 e2 1 = e1 2 e2 e2 1 = e1 e1 · e2 = 0 e1 · e2 = 1 e1 · e2 = 1 1 1 2 2 2 3 3 3 3 3 3 3 3 3 3 A superalgebra A is a Z2-graded algebra, that is, A = A0 + A1 where A0 and A1 are subspaces of A verifying Ai · Aj ⊆ Ai+j for any i, j ∈ Z2 (here + is a direct sum of vector spaces). Therefore, A0 (called the even part of A) is a subalgebra of A and A1 (the odd part) is a bimodule over A0. If a ∈ Ai, then we say that the element a is homogeneous and we will 3 denote by a the parity of a, that is, a = i. A subspace V of a superalgebra A is called sub-superspace if V = (V ∩ A0) + (V ∩ A1). By a sub-superalgebra (or ideal) of a superalgebra A we mean any sub-superspace that is a subal- gebra (or ideal) of A as an algebra. Given two superalgebras A and B and a homomorphism of algebras φ : A → B, we say φ is a homomorphism of superagebras if it preserves the grading, i.e., φ(Ai) ⊆ Bi for i = 0, 1. A superalgebra A is called simple if A2 6= 0 and it does not contain any non-trivial (graded) ideals. A superalgebra J = J0 + J1 is a Jordan superalgebra if it satisfies the supercommutativity a · b = (−1)abb · a (1.2) and the super Jordan identity (a · b) · (c · d) + (−1)bc(a · c) · (b · d) + (−1)bd+cd(a · d) · (b · c) = ((a · b) · c) · d + (−1)cd+bc((a · d) · c) · b + (−1)ab+ac+ad+cd((b · d) · c) · a (1.3) for any a, b, c, d ∈ J. The even part J0 is a Jordan algebra and the odd part J1 is an J0-bimodule with an operation J1 × J1 → J1. Note that a superidentity is obtained from the corresponding identity following the Ka- plansky rule: any alteration in the order of odd elements causes a change in sign. An associative superalgebra is just an associative algebra with a Z2-graduation but a Jordan superalgebra is not necessarily a Jordan algebra. Example 1. Consider the three-dimensional simple Jordan superalgebra due to I. Kaplansky: K3 = ke + (kx + ky) with multiplication given by e2 = e, x2 = y2 = 0, e · x = x · e = 1 2 x, e · y = y · e = 1 2 y and x · y = −y · x = e. Take a = e + x and b = y in identity (1.1) then (a2 · b) · a − a2(b · a) = a 6= 0. Indeed, a Jordan superalgebra is a Jordan algebra with a Z2-graduation if and only if J2 1 = 0. On the other hand, the same Jordan algebra can have two or more non-isomorphic Z2- graduations. Example 2. The non-associative three-dimensional Jordan algebra T6 (see Table 1) admits four Z2-graduations that, as we will see in Theorem 11, determine four non-isomorphic Jordan superalgebras: S3 12 and the trivial one Ts 6 . 4, S3 9, S3 In fact, an algebra A admits a non-trivial Z2-graduation if and only if there exists an au- tomorphism ϕ ∈ Aut(A) such as ϕ2 = Id and ϕ 6= Id. The relation between ϕ and the Z2-graduation is given by A0 = {a ∈ A ϕ(a) = a} and A1 = {a ∈ A ϕ(a) = −a}. Example 3. Consider the three dimensional associative Jordan algebra T1. We observe that Aut(T1) = {ϕ : T1 → T1 ϕ(e1) = e1; ϕ(e2) = αe2 + βe3; ϕ(e3) = α2e3; with α, β ∈ k} then 4 (T1)0 = ke1 + ke3 and (T1)1 = ke2 is the only (up to graded isomorphism) non-trivial Z2- graduation on T1. Then it is an associative superalgebra which is not a Jordan superalgebra since it does not satisfy the supercommutativity. If A = A¯0 + A¯1 is an associative superalgebra, we can get a Jordan superalgebra structure with the same underlying vector space, A(+), by defining a new product a ⊙ b = 1 2 (ab + (−1)abba) for any a, b ∈ A. A Jordan superalgebra J = J0 + J1 is said to be special is J is a sub-superalgebra of A(+), for some associative superalgebra A. Otherwise, it is called exceptional. We call A an associative envelope for J. The universal associative envelope U(J) of the Jordan superalgebra J is the unital associative superalgebra satisfying the following universal property, which implies that U(J) is unique up to isomorphism: there is a morphism of Jordan superalgebras α : J → U(J)(+) such that for any unital associative superalgebra A and any morphism of Jordan superalgebras β : J → A(+), there is a unique morphism of associative superalgebras γ : U(J) → A satisfying β = γ◦α. Example 4. The Kaplansky superalgebra K3 of Example 1 is special, in fact it is a sub-super- algebra of M(+) 11 (W1) where W1 = alg hξ, η : ξ · η − η · ξ = 1i is the Weyl algebra. The even part and the odd part are respectively, (cid:12)(cid:18) a 0 0 b (cid:19) a, b ∈ W1(cid:13) and (cid:12)(cid:18) 0 c d 0 (cid:19) c, d ∈ W1(cid:13) , then the application of K3 in M(+) ξηe21 is an homomorphism. 11 (W1) such as e 7→ e11, x 7→ 2e12 + 2ξe21 and y 7→ ηe12 + Example 5. (Superalgebra of a superform) Let V = V0 ⊕ V1 be a vector superspace over a field k, and let f : V ×V → k be a supersymmetric bilinear form, i.e., f V0 is symmetric, f V1 is skew- symmetric and f(Vi, Vj) = 0 if i 6= j. Then J(V, f) = k · 1 ⊕ V is a Jordan superalgebra with J0 = k·1⊕V0, J1 = V1 and the product given by (α1+v)·(β1+w) = (αβ+f(v, w))·1+(αw+ βv). Moreover, suppose that f is non-degenerate, dim V0 = n and dim V1 = 2m. It follows from [14] that J(V, f) is an special Jordan superalgebra. In fact J(V, f) ⊆ (Cl(n)e⊗kWm)(+), where Cl(n) = h1, e1, . . . en eiej + ejei = 0, i 6= j, e2 i = 1i is the Clifford superalgebra and Wm = h1, xi, yi, 1 6 i 6 m [xi, yj] = δij, [xi, xj] = [yi, yj] = 0i is the Weyl superalgebra with ei = xj = yk = 1. Example 6. Algebra T6 in Table 1 is special as an ordinary Jordan algebra. Its universal 5 associative envelope is the five-dimensional algebra given by U(T6) 1 e1 e2 e3 e4 e1 e1 e1 1 1 e1 e2 e2 − e4 e3 e4 e3 0 e2 e3 e4 e2 e3 e4 e4 e3 e4 0 0 0 0 0 0 0 0 0 with homomorphism ei 7→ ei for i = 1, 2, 3. Now, consider the Z2-graduation in T6 which result in the superalgebra S3 12 (see Table 6) with dim J0 = 2 and dim J1 = 1. Then the graduation induced on U(T6) = A0 + A1 where A0 = {1, e1, e2, e4} and A1 = {e3} is an associative envelope for S3 12) is the superalgebra with 12 but is not the universal one. In this case, U(S3 basis (cid:8)1, e1, e2, e1e2, ok 1 , k > 1(cid:9) quotient with the ideal: 2, e2o1, o1e2, e2 (cid:10)e2 1 − e1, e1o1 − o1, o1e1 − o1, e2e1 − e2 + e1e2(cid:11) . Example 7. In [7] the 10-dimensional Kac superalgebra K was defining. It is given by K0 = ha1, . . . , a6i and K1 = hξ1, . . . , ξ4i with a1 · ai = ai, i=1,...,5 a1 · a6 = 0, a1 · ξi = a2 · a3 = a1, a2 · ξ3 = ξ1, a2 · ξ4 = ξ2, a3 · ξ1 = a4 · a5 = a1, a4 · ξ2 = ξ1, a4 · ξ4 = ξ3, a5 · ξ1 = a2 6 = a6, a6 · ξi = 1 2 ξi, i=1,...,4; 1 2 1 2 1 2 ξi, i=1,...,4; ξ3, a3 · ξ2 = ξ2, a5 · ξ3 = 1 2 1 2 ξ4; ξ4; ξ1 · ξ2 = a2, ξ2 · ξ3 = a1 + a6, ξ1 · ξ3 = a4, ξ1 · ξ4 = a1 + a6; ξ3 · ξ4 = a3; ξ2 · ξ4 = a5, and products which are not obtained from these transpositions of factors are equal to 0. In [17] the authors proved that K is exceptional. If char k = 3 this superalgebra is not simple, but it has a simple subalgebra of dimension 9, called the degenerated Kac superalgebra. There are two other examples of exceptional simple Jordan superalgebras in char k = 3, see [18]. Example 8. An example of non-simple exceptional Jordan superalgebra was obtained in It is the 7-dimensional superalgebra with even part J0 = T7, odd part generated by [11]. {o1, o2, o3, o4} and multiplication given by e1 · oi = oi · e1 = 1 2 oi for i = 2, 3 n1 · oj = oj · n1 = oj+1 for j = 1, 3 n2 · o1 = o1 · n2 = o3 n2 · o2 = o2 · n2 = −o4 o1 · o2 = −o2 · o1 = n2 with the other product being zero. 6 1.1. Peirce decomposition for Jordan superalgebras Suppose that a Jordan superalgebra J possess an idempotent element then it has an homo- geneous idempotent e such as e ∈ J0. Hence (1.3) gives 0 = ((a · e) · e) · e + ((a · e) · e) · e + a · e − (a · e) · e − (a · e) · e − (a · e) · e 0 = 2((a · e) · e) · e − 3(a · e) · e + a · e 0 = [2R3 e − 3R2 e + Re](a) for any element a ∈ J, then 2R3 e + Re = 0 (where Re : J → J denote the right multiplication by e). Using the standard argument for the Peirce decomposition for Jordan algebras (see [5]), we obtain: e − 3R2 Theorem 9. Let e ∈ J be an homogeneous idempotent element, then J is a direct sum of vector spaces J = P0 ⊕ P 1 2 ⊕ P1, where Pi = {x ∈ J x · e = e · x = ix} with i = 0, 1 2 , 1 which have the following multiplicative properties: P2 1 ⊆ P1, P1 · P0 = (0), P2 , P2 ⊆ P 1 , P1 · P 1 ⊆ P 1 2 2 2 P0 · P 1 2 0 ⊆ P0, ⊆ P0 + P1. 1 2 Indeed, put b = d = e in (1.3) and take a ∈ Pi and c ∈ Pj to obtain (1 − 2i)(a · c) · e = (−1)acj(c · a) − 2ij(a · c). = (−1)2acj(a · c) − 2ij(a · c) = j(1 − 2i)(a · c) (1.4) (1.5) (1.6) (1.7) (1.8) If i = j = 0 we get (a · c) · e = 0 and if i = j = 1 we have (a · c) · e = a · c. Also i = 0, j = 1 and i = 1, j = 0 give (a · c) · e = (a · c) and (a · c) · e = 0, respectively. Hence 0 = (a · c) · e = (−1)ac(c · a) · e = (−1)ac(c · a), thus c · a = 0 = a · c and (1.5) holds. Similarly, i = 0, 1 and j = 1 for i = 0, 1. Now, putting c = d = e and taking a, b ∈ P 1 2 in (1.7) give Pi · P 1 in (1.3) we obtain ⊆ P 1 2 2 2 (a · b) · e + 1 2 (a · b) = ((a · b) · e) · e + Writing a · b = c0 + c1 + c 1 a · b ∈ P0 + P1. Thus (1.6) holds. 2 1 4 a · b + (−1)ab 1 4 where ci ∈ Pi we have(cid:0) 1 2 − 1 4(cid:1) c 1 2 b · a = ((a · b) · e) · e + 1 2 a · b. = 0. Hence c 1 2 = 0 and More generally, we have the refined Peirce decomposition: Theorem 10. If e = Pn into a direct sum of subspaces i=1 ei, is a sum of pairwise orthogonal idempotents, then J decomposes J = M06i6j6n Pij 7 (1.9) where P00 = {x ∈ J x · ei = 0 for all i} , Pii = {x ∈ J x · ei = x, x · ek = 0 for k 6= i and i 6= 0} , P0i = (cid:12)x ∈ J x · ei = Pij = (cid:12)x ∈ J x · ei = x · ej = 1 2 x, x · ek = 0 for k 6= i and i 6= 0(cid:13) , 1 2 x i 6= j, x · ek = 0 for 0 6= i < j and k 6= i, j(cid:13) Note that P00 = ∩n P 1 2 (ei) ∩ P 1 2 (ej) ∩n k=1 k6=i,j i=1P0(ei), Pii = P1(ei) ∩n P0(ek) for 0 6= i < j and P00 = P0i = 0 if e = 1 is unity in J. Then using P0(ek), P0i = P 1 P0(ek), Pij = (ei) ∩n k=1 k6=i 2 k=1 k6=i the previous decomposition relative to one idempotent we have that the Peirce components Pij satisfy the following relations: P2 ii ⊆ Pii, Pij · Pii ⊆ Pij, P2 ij ⊆ Pii + Pjj, Pij · Pjk ⊆ Pik, Pii · Pjj = Pii · Pjk = Pij · Pkl = (0) where the indices i, j, k, l are all distinct. Peirce decompositions are inherited by ideals or by subalgebras containing e1, . . . , en. The latter is the case of J0 since the idempotents are supposed to be homogeneous. On the other hand, note that this also happens in J1. Indeed, taking x ∈ J1 we can write x = Σpij with pij ∈ Pij , 0 6 i 6 j 6 n. Multiplying successively x by appropriate idempotents ei and using ij = J1 ∩ Pij. the fact that ei · J1 ⊆ J1 we get pij ∈ J1. Therefore J1 =L06i6j6n P1 ij where P1 2 Classification of Jordan superalgebras In the next sections we will present the complete classification of Jordan superalgebras of dimensions 1, 2 and 3. From now on k will be understood as an algebraically closed field of char 6= 2. Henceforth, for convenience we drop · and denote the multiplication in a superal- gebra J simply as ab. Products that we do not specify are understood to be 0 or be determined by commutativity or anticommutativity. We will say that J = J0 + J1 is a Jordan superalgebra of type (n, m) if n = dim J0 and m = dim J1. We will denote by ei the elements in J0 and by oi the ones which belong to J1. It follows from the supercommutativity (1.2) that o2 i = 0 for i = 1, . . . , m. Every indecompos- able superalgebra obtained in the classification will be denoted by Si j where the superscript i represent its dimension. The description of Jordan superalgebras will be organized according to all possible values of n and m. For example, we will look at the most trivial cases. Suppose that n = 0, then {o1, . . . , om} is a basis for J = J1 an m-dimensional Jordan superalgebra of type (0, m). Since oioj ∈ J0 = 0 for any 1 6 i, j 6 m then there exists only one such J and it is a direct sum of m copies of S1 1 = 0. On the other hand, the Jordan superalgebras of type (n, 0) 1 = ko1 where o2 8 correspond to the ordinary n-dimensional Jordan algebras with the trivial Z2-graduation. We denote trivial indecomposable superalgebras by Us k with multiplication given, re- spectively, by Ui, Bj and Tk in Table 1 for i = 1, 2, j = 1, 2, 3 and 1 6 k 6 10. Henceforward, we will suppose n and m both not null. j and Ts i , Bs 2.1. Jordan superalgebras of dimension 2 Suppose that J = J0 + J1 is a Jordan superalgebra of type (1, 1) and let {e1, o1} be a basis of J. According to Table 1, we have two possibilities for the even part J0 = U1 or J0 = U2. Since o2 1 = 0 we only have to determinate β ∈ k such as e1o1 = βo1. 1 ⊕ S1 1, S2 2 ⊕ S1 1. Writing the Jordan superidentity (1.3) for the basis in the first case, i.e. J0 = U1, we obtain only one equation (β − 1)β(2β − 1)o1 = 0 with solutions β = 0, 1 2, 1 which defines, respec- tively, three non-isomorphic superalgebras: Us 2 (see Table 2). Proceeding in the same way in case of J0 = U2 we still obtain one equation 2β3o1 = 0, then necessarily β = 0 and the superalgebra is Us 1 and S2 Every Jordan superalgebra of dimension two satisfies J2 1 = 0 so they are ordinary Jordan algebras. Since non-isomorphic Jordan algebras have non-isomorphic Z2-graduations, super- algebras in different lines in Table 2 are non-isomorphic. On the other hand, isomorphism of superalgebras preserves the grading then superalgebras in different columns in Table 2 are also non-isomorphic. Consequently, all two-dimensional Jordan superalgebras here presented are pairwise non-isomorphics. Table 2: Two-dimensional Jordan superalgebras Algebra B1 B2 B3 U1 ⊕ U1 U1 ⊕ U2 U2 ⊕ U2 Type (0, 2) Superalgebras Type (1, 1) S2 2 : e2 S2 1 : e2 1 = e1 e1o1 = o1 1 = e1 e1o1 = 1 2 o1 1 ⊕ S1 S1 1 Us Us 1 ⊕ S1 1 2 ⊕ S1 1 Type (2, 0) Bs 1 Bs 2 Bs 3 1 ⊕ Us 1 1 ⊕ Us 2 2 ⊕ Us 2 Us Us Us 2.2. Jordan superalgebras of dimension 3 2.2.1. Jordan superalgebras of type (1, 2) Let J be a three-dimensional Jordan superalgebra of type (1, 2). The even part, J0, is one of the one-dimensional Jordan algebras J0 = U1 or J0 = U2. If {¯e1, ¯o1, ¯o2} is a basis of J, we must determine the products ¯e1 ¯o1 = α ¯o1 + β ¯o2, ¯e1 ¯o2 = γ ¯o1 + δ ¯o2 ¯o1 ¯o2 = ε¯e1. 1. Suppose that J0 = U2. We claim that J is one of the superalgebras in Table 3. Indeed, the 9 Jordan superidentity (1.3) for the basis provides the system of equations 2(cid:0)α3 + 2αβγ + βγδ(cid:1) ¯o1 + 2β(cid:0)α2 + αδ + βγ + δ2(cid:1) ¯o2 = 0 ε(cid:0)(α − δ)2 + 4βγ(cid:1) ¯e1 = 0 2γ(cid:0)α2 + αδ + βγ + δ2(cid:1) ¯o1 + 2(cid:0)αβγ + 2βγδ + δ3(cid:1) ¯o2 = 0 ε(cid:0)α2 − αδ + 2βγ(cid:1) ¯o1 + βε(α + δ) ¯o2 = 0 ε(δ(δ − α) + 2βγ) ¯o2 + γε(α + δ) ¯o1 = 0 2ε(αδ − βγ)¯e1 = 0 with the following solutions: (a) γ = − α2 β and δ = −α. Suppose ε 6= 0 then J ≃ S3 1: if α 6= 0 the isomorphism is given by ¯e1 7→ e1, ¯o1 7→ − ¯e1 7→ e1, ¯o1 7→ − 1√βε o1 and ¯o2 7→ −q β with change of basis ¯e1 7→ e1, ¯o1 7→ − β ¯e1 7→ e1, ¯o1 7→ o1 and ¯o2 7→ βo2 if α = 0. √β α√ε o2 and ¯o2 7→ α√βε o1 +q β ε o2, but if α = 0 then ε o2. In case of ε = 0, we obtain J ≃ S3 3 α2 o2 and ¯o2 7→ o1 + β α o2 for α 6= 0 or (b) α = β = γ = δ = 0. If we suppose ε 6= 0, the superalgebra J is S3 2 with isomor- 1 ⊕ S1 1. phism ¯e1 7→ εe1, ¯o1 7→ o1 and ¯o2 7→ o2. Otherwise, J ≃ Us 2 ⊕ S1 (c) α = β = δ = 0. If γ = 0 we are in the previous case, so we can consider γ 6= 0. 1 with isomorphism ε o1. But if ε = 0, then again we Suppose further that ε 6= 0 then we obtain the superalgebra S3 given by ¯e1 7→ e1, ¯o1 7→ − i√γε o2 and ¯o2 7→ −ip γ 3 with change of basis ¯e1 7→ e1, ¯o1 7→ o2 and ¯o2 7→ γo1. get J ≃ S3 Table 3: Jordan superalgebras of type (1, 2) with J0 = U2. J Multiplication table Observation S3 1 S3 2 S3 3 e1o1 = o2 o1o2 = e1 o1o2 = e1 e1o1 = o2 2 ⊕ S1 1 ⊕ S1 1 Us associative associative associative 2. Suppose that J0 = U1, then J has an even homogeneous idempotent element. Using the and the corresponding decomposition 2, 1. Then J is completely i to which o1 and o2 belong. Peirce decomposition (1.4) we get J = P0⊕P1⊕P 1 of the odd part J1 = P1 defined by the product o1o2 = εe1 and the Peirce subspace P1 In this way we obtain eight non-isomorphic superalgebras given in Table 4. i = J1 ∩ Pi, i = 0, 1 where P1 0 ⊕ P1 1 ⊕ P1 1 2 2 J S3 4 Table 4: Jordan superalgebras of type (1, 2) with J0 = U1. Multiplication table Observation S2 1 ⊕ S1 1 1 = e1 e1o1 = o1 e1o2 = 1 e2 2 o2 10 o1 ∈ P0, o2 ∈ P 1 o1 ∈ P1, o2 ∈ P 1 2 2 Us S2 2 ⊕ S1 1 1 ⊕ S1 1 ⊕ S1 1 1 = e1 e1o1 = 1 2 o1 e1o2 = 1 e2 2 o2 e2 1 = e1 e1o1 = o1 e1o2 = o2 1 = e1 e1o1 = 1 e2 2 o2 o1o2 = e1 e2 1 = e1 e1o1 = o1 e1o2 = o2 o1o2 = e1 2 o1 e1o2 = 1 S3 5 S3 6 S3 7 S3 8 o1 ∈ P0, o2 ∈ P1, associative o1, o2 ∈ P0 associative o1, o2 ∈ P 1 ε = 0 2 o1, o2 ∈ P1 ε = 0 associative o1, o2 ∈ P 1 ε 6= 0 o1, o2 ∈ P1 ε 6= 0 2 Up to permutation of o1 and o2, we have six possibilities: . If ε = 0 we get the superalgebra S3 (a) o1, o2 ∈ P 1 thus J ≃ S3 7. 2 (b) o1, o2 ∈ P1. Analogously, if ε = 0 we obtain S3 and we get the superalgebra S3 8. 5, but if ε 6= 0 we can choose ε = 1 6, otherwise we can choose ε = 1 (c) o1, o2 ∈ P0, thus it follows from the multiplicative properties of Peirce components 0 ⊆ P0 = J1 but o1o2 ∈ J0 then o1o2 = 0 and we obtain the 1 ⊕ S1 1. (1.5) that o1o2 ∈ P2 1 ⊕ S1 superalgebra Us (d) o1 ∈ P0, o2 ∈ P1, then o1o2 ∈ P0P1 = 0 and J ≃ S2 ⊆ P 1 (e) o1 ∈ P0, o2 ∈ P 1 1 ⊕ S1 1. (f) o1 ∈ P1, o2 ∈ P 1 , it follows that o1o2 ∈ P0P 1 , analogously o1o2 ∈ P1P 1 get J ≃ S2 ⊆ P 1 2 2 2 2 2 2 ⊕ S1 1. = ko2 thus o1o2 = 0 and we = ko2 then o1o2 = 0 and we obtain the superalgebra S3 4. 2 2.2.2. Jordan superalgebras of type (2, 1) Let J = J0 + J1 be a three-dimensional Jordan superalgebra of type (2, 1) and let {e1, e2, o1} be a basis of J. In this case, the even part J0 is one of the six two-dimensional Jordan algebras given in Table 1 and J1 = S1 1. All that remains is to determine the action of J0 in o1, namely, to found scalars α, β ∈ k such as e1o1 = αo1 and e2o1 = βo1. 1. Suppose J0 = B1, so we have one idempotent e1 relative to which J1 = P1 0 ⊕ P1 1 ⊕ P1 1 2 with e1, e2 ∈ P1. We claim that J is one of the superalgebras in Table 5. Table 5: Jordan superalgebras of type (2, 1) with J0 = B1. J S3 9 S3 10 Multiplication table Observation Bs 1 ⊕ S1 1 1 = e1 e1e2 = e2 e1o1 = 1 e2 2 o1 e2 1 = e1 e1e2 = e2 e1o1 = o1 o1 ∈ P0, associative o1 ∈ P 1 2 o1 ∈ P1, associative Indeed, according to the Peirce subspace Pi to which o1 belongs we have: (a) o1 ∈ P0 then e2o1 ∈ P1P0 = 0 and we obtain the superalgebra Bs (b) o1 ∈ P 1 , writing the Jordan superidentity (1.3) for the basis we get the equations 1 ⊕ S1 1. 2 −β2o1 = 0 and 2β3o1 = 0, then necessarily β = 0 and J ≃ S3 9. (c) o1 ∈ P1, proceeding in the same way the equation 2β3o1 = 0 gives β = 0, thus J ≃ S3 10. 2. If J0 = B2, analogously we have one idempotent e1 relative to which J1 = P1 0 ⊕ P1 1 ⊕ P1 1 2 11 with e1 ∈ P1, e2 ∈ P 1 . Whatever the Peirce subspace that o1 belongs to, it follows from the multiplicative properties (1.5) and (1.6) and the fact that e2o1 ∈ J1 that the product e2o1 is always zero. So, we obtain three non-isomorphic Jordan superalgebras given in Table 6. 2 Table 6: Jordan superalgebras of type (2, 1) with J0 = B2. J S3 11 S3 12 Multiplication table Observation Bs 2 ⊕ S1 1 1 = e1 e1e2 = 1 e2 1 = e1 e1e2 = 1 e2 2 e2 e1o1 = 1 2 o1 2 e2 e1o1 = o1 o1 ∈ P0 o1 ∈ P 1 o1 ∈ P1 2 3. Consider J0 = B3, since B3 is a nilpotent algebra J does not have any idempotent ele- ment, therefore Peirce decomposition is not available. Then, computing the superiden- tity in the basis elements we get the system α(cid:0)2α2 − 3β(cid:1) o1 = 0 β(cid:0)2α2 − β(cid:1) o1 = 0 2αβ2o1 = 0 2β3o1 = 0 whose solution is α = β = 0 (still when char k = 3). Consequently, we obtain the associative (then Jordan) superalgebra Bs 3 ⊕ S1 1. 4. Now, if J0 = U1 ⊕U1 then J has two even homogeneous orthogonal idempotent elements which implies in the following decomposition of J1 in Peirce subspaces J1 = P1 00 ⊕ P1 01 ⊕ P1 02 ⊕ P1 11 ⊕ P1 12 ⊕ P1 22. Thus J is completely defined by the Peirce component P1 provides us the four non-isomorphic superalgebras: Us if o1 ∈ P01 (or, analogously, o1 ∈ P02), S2 and the indecomposable one S3 the case of o1 ∈ P12. ij to which o1 belongs. This 1 ⊕ Us 1 1 if o1 ∈ P11 (or, analogously, o1 ∈ P22) 2 o1 in 2 = e2, e1o1 = 1 1 if o1 ∈ P00, S2 2o1, e2o1 = 1 13 given by e2 1 = e1, e2 1 ⊕ Us 2 ⊕ Us 1 ⊕ S1 5. Assume J0 = U1 ⊕ U2, then J has an idempotent e1 such as J1 = P1 with e1 ∈ P1 and e2 ∈ P0. necessarily β has to be zero for the respective superalgebras to satisfy the Jordan superidentity: 2β3o1 = 0 in the first case and β2o1 = 0, 2β3o1 = 0 in the second one. We obtain the superalgebras Us 2 ⊕ S1 1 and S2 2 , respectively. Lastly, if o1 ∈ P1 then e2o1 ∈ P0P1 = 0 and the resulting superalgebra is S2 If o1 ∈ P0 or o1 ∈ P 1 1 ⊕ Us 1 ⊕ Us 0 ⊕ P1 1 ⊕ P1 1 2 2 2 ⊕ Us 2 . 6. Finally, if J0 is the null algebra U2 ⊕ U2 then again Peirce decomposition is not available. Computing the Jordan superidentity in the basis we have: 2α3o1 = 0 2αβ2o1 = 0 2α2βo1 = 0 2β3o1 = 0 which has as solution α = β = 0. Hence, we obtain the associative superalgebra Us 2 ⊕ 12 Us 2 ⊕ S1 1. Until now, we have proved that any three-dimensional Jordan superalgebra J over an alge- braically closed field of characteristic not 2 is isomorphic to one of the superalgebras presented in this subsection. The following theorem completes the algebraic classification. Theorem 11. All superalgebras in Section 2.2 are pairwise non-isomorphic. Proof. Firstly, consider the superalgebras such as J2 = 0, namely those algebras in Table 7. 1 Following an analogous reasoning to that done in dimension two and taking into account that decomposable superalgebras are guaranteed to be non-isomorphics, we only have to verify whether S3 12 ? ≃ S3 9. The result follows from (S3 It remains to analyze superalgebras with J2 8. They all have type (1, 2), the first two have even part U2 while J0 = U1 in the remaining two superalgebras. S3 2 is an associative superalgebra while S3 1 is not. On the other hand, the Jordan superalgebra 8 has a unity e1 whereas S3 S3 7 is K3 the Kaplansky superalgebra, one of the few examples of finite dimensional simple nonunital Jordan superalgebras. 7 has not, indeed S3 12)0 = B2 6≃ B1 = (S3 1 6= 0, namely S3 1, S3 9)0. 2, S3 7 and S3 Table 7: Three-dimensional Jordan superalgebras with J2 1 = 0 Algebra Superalgebras (0, 3) (1, 2) (2, 1) (3, 0) T1 T2 T3 T4 T5 T6 T7 T8 T9 T10 B1 ⊕ U1 B1 ⊕ U2 B2 ⊕ U1 B2 ⊕ U2 B3 ⊕ U1 B3 ⊕ U2 S3 10 S3 12 S3 9 S3 11 S3 13 S2 2 ⊕ Us 1 S2 2 ⊕ Us 2 1 ⊕ S1 Bs 1 S2 1 ⊕ Us 1 S2 1 ⊕ Us 2 2 ⊕ S1 Bs 1 Bs 3 ⊕ S1 1 S3 6 S3 3 S3 4 S3 5 S2 2 ⊕ S1 1 S2 1 ⊕ S1 1 13 Ts 1 Ts 2 Ts 3 Ts 4 Ts 5 Ts 6 Ts 7 Ts 8 Ts 9 Ts 10 Bs Bs 1 ⊕ Us 1 1 ⊕ Us 2 Bs Bs 2 ⊕ Us 1 2 ⊕ Us 2 Bs Bs 3 ⊕ Us 1 3 ⊕ Us 2 U1 ⊕ U1 ⊕ U1 U1 ⊕ U1 ⊕ U2 U1 ⊕ U2 ⊕ U2 U2 ⊕ U2 ⊕ U2 S1 1 ⊕ S1 1 ⊕ S1 1 Us 1 ⊕ Us 1 ⊕ Us 2 ⊕ Us 1 ⊕ S1 2 ⊕ S1 2 ⊕ S1 1 Us 1 Us 1 Us Us 1 ⊕ Us 1 ⊕ Us 1 ⊕ Us 2 ⊕ Us 1 ⊕ Us 1 1 ⊕ Us 2 2 ⊕ Us 2 2 ⊕ Us 2 Us Us 1 ⊕ S1 2 ⊕ S1 1 ⊕ S1 1 ⊕ S1 1 Us 1 Us Theorem 12. All Jordan superalgebras of dimension up to 3 are special. Proof. It follows from [19] that ordinary Jordan algebras up to dimension 3 are special, then it is clear that Jordan superalgebras up to dimension 3 such as J2 1 = 0 are also special since the graduation on the associative enveloped algebra is induced by the graduation of J. Associative Jordan superalgebras are trivially special, it is the case of S3 2. Furthermore, we have already proved in Example 4 that K3 ≃ S3 7 is also special. Now, observe that S3 8 is the simple Jordan superalgebra of the non-degenerate supersym- metric bilinear form f : V × V → k where V = V1 is a vector superspace over k such as dim V1 = 2, given by f(oi, oi) = 0 for i = 1, 2 and f(o1, o2) = 1 = −f(o2, o1). Therefore, the speciality follows from Example 5. Finally, superalgebra S3 1 is exactly the algebra supercommutative, solvable of index 2 and nonnilpotent described in Example 1 of [20] and proved to be special by exhibiting an embed- ding in the superalgebra A(+) for the matrix superalgebra A = M1,2 3 (k), where A0 = (cid:12)(cid:18) a 0 0 b (cid:19) , a ∈ k, b ∈ M2(k)(cid:13) and A1 = (cid:12)(cid:18) 0 x y 0 (cid:19) , x ∈ M1,2(k), y ∈ M2,1(k)(cid:13) putting e1 = e23 ∈ A0, o1 = e13 + 4e31 ∈ A1, o2 = e1 ⊙ o1. As a consequence, we obtain a bound for the minimal dimension of exceptional Jordan superalgebras as posed in [11]. We observe that, up to our knowledge, the dimension 7 for the Example 8 is the least dimension for the known examples of exceptional Jordan superalgebras. Corollary 13. The minimal dimension d of an exceptional Jordan superalgebra satisfies 4 6 d 6 7. 14 References References [1] N. Backhouse, A classification of four-dimensional Lie superalgebras, J. Math. Phys., 19 (1978), pp. 2400–2402. [2] , On the construction of graded Lie algebras, Springer Berlin Heidelberg, Berlin, Hei- delberg, 1978, pp. 364–366. [3] V. T. Filippov, V. K. Kharchenko, and I. Shestakov, eds., Dniester notebook: unsolved problems in the theory of rings and modules, Mathematics Institute, Russian Academy of Sciences Siberian Branch, Novosibirsk, 4 ed., 1993. [4] A. S. Hegazi, Classification of nilpotent Lie superalgebras of dimension five II, International Journal of Theoretical Physics, 38 (1999), pp. 2681–2693. [5] N. Jacobson, Structure and representations of Jordan algebras, vol. 39 of American Math- ematical Society Colloquium Publications, American Mathematical Society, Providence, 1968. [6] V. G. Kac, Classification of simple Lie superalgebras, Functional Analysis and Its Applica- tions, 9 (1975), pp. 263–265. [7] , Classification of simple Z-graded Lie superalgebras and simple Jordan superalgebras, Communications in Algebra, 5 (1977), pp. 1375–1400. [8] , Lie superalgebras, Advances in Mathematics, 26 (1977), pp. 8 – 96. [9] I. L. Kantor, Connection between Poisson brackets and Jordan and Lie superalgebras, Lie Theory, Differential Equations and Representation Theory, (1990), pp. 213 – 225. [10] I. Kaplansky, Graded Jordan algebras I, preprint, (1975), pp. 1–14. [11] M. C. López-Díaz, I. Shestakov, and S. Sverchkov, On speciality of Bernstein Jordan algebras, Communications in Algebra, 28 (2000), pp. 4375–4387. [12] M. E. Martin, Four dimensional Jordan algebras, International Journal of Mathematics, Game Theory and Algebra, 20 (2013), pp. 41–59. [13] C. Martínez and E. Zel'manov, Simple finite-dimensional Jordan superalgebras of prime characteristic, Journal of Algebra, 236 (2001), pp. 575 – 629. [14] , Specializations of Jordan superalgebras, RACSAM, 95 (2001), pp. 1–6. [15] N. L. Matiadou and A. Fellouris, A classification of the five-dimensional Lie superalge- bras over the complex numbers, Mathematica Balkanica, 19 (2005), pp. 143–154. 15 [16] , Classification of the five-dimensional Lie superalgebras over the real numbers, Inter- national Journal of Theoretical Physics, 46 (2007), pp. 451–470. [17] Y. Medvedev and E. Zel'manov, Some counter-examples in the theory of Jordan algebras, Nonassociative Algebraic Models, Nova Sci. Publ., (1992), pp. 1–16. [18] M. Racine and E. Zel'manov, Simple Jordan superalgebras with semisimple even part, Journal of Algebra, 270 (2003), pp. 374–444. [19] H. Sherlukov, Jordan algebras in small dimensions, masters dissertation, University of Novosibirsk, 1989. [20] I. Shestakov, Superalgebras and counterexamples, Siberian Mathematical Journal, 32 (1991), pp. 1052–1060. 16
1012.5505	1	1012	2010-12-26T09:40:26	Diameters of commuting graphs of matrices over semirings	[ "math.RA" ]	We calculate the diameters of commuting graphs of matrices over the binary Boolean semiring, the tropical semiring and an arbitrary nonentire commutative semiring. We also find the lower bound for the diameter of the commuting graph of the semiring of matrices over an arbitrary commutative entire antinegative semiring.	math.RA	math	DIAMETERS OF COMMUTING GRAPHS OF MATRICES OVER SEMIRINGS DAVID DOLZAN, DAMJANA KOKOL BUKOVSEK, POLONA OBLAK Abstract. We calculate the diameters of commuting graphs of matrices over the bi- nary Boolean semiring, the tropical semiring and an arbitrary nonentire commutative semiring. We also find the lower bound for the diameter of the commuting graph of the semiring of matrices over an arbitrary commutative entire antinegative semiring. 1. Introduction Definition. A semiring is a set S equipped with binary operations + and · such that (S, +) is a commutative monoid with identity element 0, and (S, ·) is a monoid with identity element 1. In addition, operations + and · are connected by distributivity and 0 annihilates S. A semiring is commutative if ab = ba for all a, b ∈ S. A semiring S is called antinegative, if a + b = 0 implies that a = b = 0. Antinegative semirings are also called antirings (or zero-sum-free semirings). A semiring is entire (or zero-divisor-free) if ab = 0 implies that a = 0 or b = 0. A semiring is a division semiring if all nonzero elements have multiplicative inverses. The simplest example of an antinegative semiring is the binary Boolean semiring, the set {0, 1} with 1 + 1 = 1 · 1 = 1. We will denote the binary Boolean semiring by B. Moreover, the set of nonnegative integers (or reals) with the usual operations of addition and multiplication, is a commutative entire antinegative semiring, and all distributive lattices are antinegative semirings. On the set R ∪ {−∞}, we define operations a ⊕ b = max{a, b} and a ⊙ b = a + b, where −∞ + a = a + (−∞) = −∞. It is easy to verify that (R ∪ {−∞}, ⊕, ⊙) is a semiring. It is denoted by T and called the tropical semiring, sometimes also the max-plus semiring. Tropical semiring is a commutative entire antinegative division semiring and it is closely related to the max algebra, i.e. the semiring of nonnegative reals R+, where the addition is the same as in the tropical semiring and the multiplication is the ordinary multiplication on reals. Recently, commuting matrices in max algebra were studied in [10]. Over past decades, the tropical semirings and other tropical structures were widely investigated. Let us mention only a few pioneering works [4, 5, 13] that connect the theory of matrices over classical and tropical worlds. We denote by Mn(S) the semiring of all n × n matrices over a semiring S and by GLn(S) the group of all invertible matrices in Mn(S). By In ∈ Mn(S), we denote the Date: July 19, 2018. 2010 Mathematics Subject Classification. 15A27, 16Y60, 05C50, 05C12. Key words and phrases. semiring, Boolean semiring, tropical semiring, commuting graph, diameter. 1 identity matrix and by 0n ∈ Mn(S) the zero matrix. The matrix with the only nonzero entry 1 in the i-th row and the j-th column will be denoted by Ei,j. If X is a subset of S, let denote the centralizer of X in S. For x ∈ S we also define CS(x) = CS({x}). CS(X) = {s ∈ S; sx = xs for all x ∈ X} For any subset T of a semiring S, we denote by Γ(T ) the commuting graph of T . The vertex set V (Γ(T )) of Γ(T ) is the set of elements in T \CS(T ). An unorderded pair of vertices x − y is an edge of Γ(T ) if x 6= y and xy = yx. The sequence of edges x0 − x1, x1 − x2, ..., xk−1 − xk is called a path of length k and is denoted by x0 − x1 − . . . − xk. The distance between two vertices is the length of the shortest path between them. The diameter of the graph is the longest distance between any two vertices of the graph Γ and will be denoted by diam(Γ). For example, the commuting graph of the set of all 2×2 nilpotent matrices over an entire antinegative semiring S, is a disconnected graph with two components (corresponding to the strictly upper, and strictly lower triangular matrices), where both components are isomorphic to a complete graph on S − 1 vertices. Recently, the commuting graphs of matrix rings and semirings ([1, 2, 3, 6, 7, 8, 11]) and commuting graphs of various algebraic structures ([9, 12]) have been studied. They give an illustrative way of describing centralizers of elements. It was proved in [2, Cor. 7] that the diameter of the commuting graph of the full matrix ring over an algebraically closed field is equal to 4. For rings and fields that are not algebraically closed, the commuting graph of the full matrix ring might not be connected at all, or if it is connected, its diameter can be larger than 4. (See e.g. [6, Ex. 2. 15], where it has been proven that 5 ≤ diam(Γ(M9(Z2))) < ∞.) In this paper, we continue with the investigation of the diameters of commuting graphs In [7, Prop. 10] it has been shown that the diameter of the of full matrix semirings. commuting graph of the full matrix semiring over B is bounded between 3 and 4. In the second section, we prove that it is equal to 4. This implies that the diameter of the commuting graph of Mn(S) is at least 4 for every commutative entire antinegative semiring S. Using this, we prove in Theorem 3.2 that for the tropical semiring T the diameter of Γ(Mn(T)) is equal to 4. In Section 4, we prove that diam(Γ(Mn(S))) = 3 for every nonentire commutative semiring S and all n ≥ 2, thus generalizing [8, Thm. 1.1], where a similar result has been proven for Mn(Zm), where m is not a prime number. 2. The commuting graph of the full matrix semiring over B In this section, we prove that the diameter of the commuting graph of the full n × n matrix semiring over Boolean semiring B is equal to 4 for all n ≥ 3. We start with a general lemma, which can be easily proved by a straightforward calcu- lation. Denote by Jn the nilpotent n × n matrix E1,2 + E2,3 + . . . + En−1,n. 2 Lemma 2.1. If S is a semiring and n ≥ 2, then the centralizer of Jn is equal to CMn(S)(Jn) = {a0In + a1Jn + a2J 2 ; ai ∈ S}, and the centralizer of J T n is equal to CMn(S)(J T ; bi ∈ S}. (cid:4) n + . . . + an−1J n−1 n ) = {b0In + b1J T n n(cid:1)2 n + b2(cid:0)J T + . . . + bn−1(cid:0)J T n(cid:1)n−1 The following Theorem is the main result of this section. Theorem 2.2. If B is a binary Boolean semiring, then diam(Γ(M2(B))) = ∞ and diam(Γ(Mn(B))) = 4 for n ≥ 3. Proof. It was proved in [7, Prop. 10] that diam(Γ(M2(B))) = ∞. Let n ≥ 3 and let E ∈ Mn(B) be a matrix of all ones. It was also proved in [7, Prop. 10] that the distance between any matrix A ∈ Mn(B) and matrix E is at most 2. It follows that diam(Γ(Mn(B))) ≤ 4 for n ≥ 3. To prove the equality, we have to find two matrices with the distance between them at least 4. Let n = 3 and let It is easy to check, that for any pair of matrices C ∈ A and D ∈ B, C and D do not commute, so that the distance between A and B is at least 4 and thus diam(Γ(M3(B))) = 4. Now, let n ≥ 4. Let A, B ∈ Mn(B) be A = 0 0 · · · 0 1 ... 1 J T n−1   0 1   , 0 B =(cid:20)1 0 Jn−1(cid:21) . Note that the centre of Mn(B) consists only of 0n and In. Suppose that the distance between A and B is at most 3. Then there exist nonscalar matrices C, D ∈ Mn(B) such that A − C − D − B is a path in Γ(Mn(B)). Observe that C is not a diagonal matrix, otherwise all of its diagonal entries are equal, since C commutes with A. So, C + I is not in the centre and it commutes with A and D. Since C and C + I have the same centralizer, we can therefore assume that all the diagonal entries of C are equal to 1. 3 A straight-forward calculation shows that the set of all matrices with distance 1 to matrix A is . , 1 0 0 0 0 1 0 0 1 0 0 0 1 0 1 0 1 0 0 0 0 B = 1 1 0 A =     A =   , 1 1 1 , 0 0 1 , 0 0 1 , 1 1 1 1 1 1             B =   0 0 1 , 0 0 1 , 0 0 0 , 0 0 1 , 0 0 0             1 1 1 0 0 0 1 1 1 0 1 0 1 1 0 0 1 0 1 1 0 0 0 0 1 0 0 0 0 0 1 0 0 0 1 1 0 0 0 0 0 1 0 0 0 0 1 0 0 0 0 0 1 1 . and the set of all matrices with distance 1 to matrix B is Now, suppose that D is diagonal. Since it commutes with B, it has the form 0 D =(cid:20)1 0 0n−1(cid:21) or 0 D =(cid:20)0 0 In−1(cid:21) . In both cases, since C and D commute, C has the form 0 C =(cid:20)1 0 C1(cid:21) . Since C commutes with A, C1 also commutes with J T n−1 and the last row of C1 is equal D is not diagonal and again we can assume that all diagonal entries of D are equal to 1. 0 1(cid:3). So, C1 = In−1 and C is in the centre, which is a contradiction. Thus, to (cid:2)0 · · · Now, since D and B commute, we have 1 c2,1 ... cn,1 C =  c1,2 1 ... cn,2 · · · · · · . . . · · · c1,n c2,n ... 1   and D = where di 6= 0 for some i ∈ {2, 3, . . . , n − 1}. 0 1 0 0 1 d2 0 0 ... ... 0 0 0 0 1 0 0   · · · · · · . . . . . . · · · · · · 0 0 dn−2 dn−1 dn−2 ... d2 1 . . . 1 0 ,   Since A and C commute, we have for every i = 2, ..., n − 1 that 0 = (AC)2,i = (CA)2,i = c2,i+1 and 0 = (AC)2,n = (CA)2,n = c2,1, so (1) c2,i = 0 for every i 6= 2. Also, for every i = 2, ..., n − 1 we have (AC)1,i = (CA)1,i and thus (2) cn,i = c1,i+1. Further, for every i = 3, ..., n − 1 we have c1,i = c1,i + c2,i = (AC)3,i = (CA)3,i = c3,i+1 and c1,n = (AC)3,n = (CA)3,n = c3,1. For i ≥ j ≥ 3 and using (3) c1,i + cj−1,i = (AC)j,i = (CA)j,i = cj,i+1, we prove by induction that (4) and (5) cj,i+1 = c1,i + c1,i−1 + ... + c1,i−j+3 for every i ≥ j ≥ 3 cj,1 = c1,n + c1,n−1 + ... + c1,n−j+3 for every j ≥ 3. Let 3 ≤ i < j ≤ n − 1. First cj,3 = (CA)j,2 = (AC)j,2 = c1,2 + cj−1,2 and cj,i = (CA)j,i−1 = (AC)j,i−1 = c1,i−1 + cj−1,i−1 and by induction we have cj,i = c1,2 + ... + c1,i−1 + cj−i+2,2 for every 3 ≤ i < j ≤ n − 1. Since D is not diagonal, at least one of d2, ..., dn−1 is nonzero. Let k be the greatest index, such that dk 6= 0, so dk = 1. Note that the second row of D is therefore equal to · · · dk−1 1 0 · · · (cid:2)0 1 d2 Suppose first that k = n − 1. Since C and D commute, we have 0 = (CD)2,1 = (DC)2,1 = c2,1 + d2c3,1 + ... + dn−1cn,1 = c2,1 + d2c3,1 + ... + dn−2cn−1,1 + cn,1, thus cn,1 = 0 and by (5), we have c1,3 + ... + c1,n = cn,1 = 0. It follows that c1,i = 0 for all i ≥ 3. Furthermore, 0 = c1,n = (DC)1,n = (CD)1,n = c1,2dn−1 = c1,2 so c1,2 = 0 and by (4) it follows that C is lower-triangular. 0(cid:3). 4 Suppose now that k < n − 1. Since C and D commute, we have 0 = (CD)2,1 = (DC)2,1 = c2,1 + d2c3,1 + ... + dkck+1,1 = c2,1 + d2c3,1 + ... + dk−1ck,1 + ck+1,1, thus ck+1,1 = 0 and by (5) it follows that c1,n−k+2 + ... + c1,n = 0. Therefore, c1,i = 0 for every i ≥ n − k + 2. Furthermore, 0 = c1,n = (DC)1,n = (CD)1,n = c1,n−k+1dk = c1,n−k+1. By induction, c1,n−k−i+2 = c1,n−k−i+2dk = (CD)1,n−i+1 = (DC)1,n−i+1 = c1,n−i+1 = 0 for every i = 1, ..., n − k, so that c1,j = 0 for every j ≥ 2 and by (4), C is lower-triangular. Now, we have cj,i = 0 for every i > j and by applying (3) it follows that (6) cj,i = cj−1,i−1 = . . . = cj−i+2,2 for every 3 ≤ i < j ≤ n − 1. Since C is lower-triangular, (2) implies that cn,i = 0 for i ≥ 2 and cn,1 = 0 by (5). Therefore we have for every i = 2, ..., n − 2 that 0 = cn,i+1 = (CA)n,i = (AC)n,i = cn−1,i, which by (6) implies that cn−i+1,2 = 0, so again by (6), cj,i = 0 for all 3 ≤ i < j ≤ n − 1. Together with (1) we have that C = In, a contradiction. So, we have proved that the distance between A and B cannot be less than 4 and (cid:4) therefore diam(Γ(Mn(B))) = 4. Corollary 2.3. If S is a commutative entire antinegative semiring, then diam(Γ(M2(S))) = ∞ and diam(Γ(Mn(S))) ≥ 4, for n ≥ 3. Proof. For a matrix A ∈ Mn(S), let us denote by supp(A) ∈ Mn(B) the unique (0, 1)- matrix with the property Ai,j 6= 0 if and only if (supp(A))i,j 6= 0 for all 1 ≤ i, j ≤ n. Since S is a commutative entire antinegative semiring, AB = BA for A, B ∈ Mn(S) implies that supp(A)supp(B) = supp(B)supp(A) for supp(A), supp(B) ∈ Mn(B). Thus, it follows that diam(Γ(Mn(S))) ≥ diam(Γ(Mn(B))). Now, the statement follows by Theorem 2.2. (cid:4) 3. The commuting graph of the full matrix semiring over T The two operations ⊕ and ⊙ in T naturally induce the matrix addition and multi- plication on the semiring (Mn(T), ⊕, ⊙), namely for A = [ai,j], B = [bi,j] ∈ Mn(T) we have (A ⊕ B)i,j = ai,j ⊕ bi,j, and (A ⊙ B)i,j = ai,1 ⊙ b1,j ⊕ ai,2 ⊙ b2,j ⊕ . . . ⊕ ai,n ⊙ bn,j. Let In be the identity matrix in Mn(T), i.e. matrix with zeros on the diagonal and −∞ offdiagonal, and let E be the matrix with all entries equal to 0 ∈ R. For a ∈ T and A = [ai,j] ∈ Mn(T) we also define the matrix aA in the natural way, i.e. (aA)i,j = a ⊙ aij = a + aij. It was proved in [7, Cor. 11] that diam(Γ(Mn(T))) ≥ 3 for all n ≥ 3. Here, we will prove that diam(Γ(Mn(T))) = 4 for all n ≥ 3. 5 Lemma 3.1. The centralizer CMn(T)(E) consists of exactly all matrices A = [ai,j], such that there exists some element a ∈ T, with the property max j {ai,j} = a for all i = 1, 2, . . . , n and max i {ai,j} = a for all j = 1, 2, . . . , n. Proof. For the matrix A = [ai,j] we denote a = maxi,j=1,...,n{ai,j}. If A commutes with E, then max{ai,1, ai,2, ..., ai,n} = ai,1 ⊕ ai,2 ⊕ ... ⊕ ai,n = (A ⊙ E)i,1 = (E ⊙ A)i,1 = a1,1 ⊕ a2,1 ⊕ ... ⊕ an,1 = max{a1,1, a2,1, ..., an,1} for every i = 1, ..., n. So all rows of A have the same maximum a, thus a appears in every row of A. Similarly, a appears in every column of A. Conversely, if a appears in every row and column of A, then clearly E ⊙ A = aE = A ⊙ E. (cid:4) Theorem 3.2. For the tropical semiring T, we have diam(Γ(M2(T))) = ∞ and diam(Γ(Mn(T))) = 4 for n ≥ 3. Proof. By Corollary 2.3 we have that diam(Γ(M2(T))) = ∞ and diam(Γ(Mn(T))) ≥ 4. Suppose now n ≥ 3. If D = [dk,l] is a diagonal matrix, where di,i = dj,j for some i, j, then let us define the n × n matrix F = [fk,l], such that fk,l =(0, −∞, otherwise . k = l or (k, l) = (i, j) or (k, l) = (j, i), Now, (F ⊙ D)j,i = (F ⊙ D)i,j = dj,j = di,i = (D ⊙ F )i,j = (D ⊙ F )j,i. Furthermore, (F ⊙ D)k,k = dk,k = (D ⊙ F )k,k for all k and (F ⊙ D)k,l = −∞ = (D ⊙ F )k,l for all k 6= l, and {k, l} 6= {i, j}. Thus D − F − E is a path in Γ(Mn(T)) by Lemma 3.1. If A = [ai,j] is a nondiagonal matrix, let a = max{ai,j}, and then, A − A ⊕ aIn − E is a path in Γ(Mn(T)) by Lemma 3.1. We have thus proved that d(A, E) ≤ 2 for all matrices A except for the diagonal matrices with all of their diagonal entries distinct. We will now prove that d(D, A) ≤ 4 for a diagonal matrix D with all diagonal entries distinct and an arbitrary A ∈ Mn(T). If A is diagonal as well, it is clear that d(D, A) = 1. So, suppose A = [ai,j] is nondiagonal and let µ = max{ai,j} and ε = min{ai,j; ai,j 6= −∞}. Denote by B = [bi,j] the n × n matrix, defined by bi,j =(µ, ε, i = j, i 6= j . Note that ε + ai,j ≤ ε + µ ≤ ak,l + µ for all i, j, k, l and thus for every 1 ≤ i, j ≤ n we have (A ⊙ B)i,j = ai,j + µ = (B ⊙ A)i,j. For the matrix it is easy to see that B ⊙ C = C ⊙ B. Therefore, −∞ µ ε ε µ −∞ µIn−2 C =  ,   D −(cid:20)d1,1I2 −∞ d2,2In−2(cid:21) − C − B − A −∞ 6 is a path in Γ(Mn(T)) and thus d(D, A) ≤ 4. (cid:4) As a Corollary of [7, Thm. 7] we have the following. Proposition 3.3. For the tropical semiring T, we have diam(Γ(GLn(T))) =(5, ∞, n prime, otherwise. Remark. Note that [2, Thm. 14, 15] state that for an infinite commutative division ring R, we have the equality diam(Γ(Mn(R))) = diam(Γ(GLn(R))) for all n ≥ 2. With Theorem 3.2 and Proposition 3.3, we proved that this result cannot be generalized to commutative division semirings. 4. The commuting graph of the full matrix semiring over a nonentire commutative semiring In this section, we prove that the diameter of the commuting graph of the full matrix semiring over a nonentire commutative semiring is always equal to 3. This generalizes the result from [8, Thm. 1.1], where it has been proved that the diameter of the commuting graph of the full matrix ring over the ring Zm for a composite number m is always equal to 3. First, we find the lower bound for the diameter of the commuting graph of matrices over an arbitrary commutative semiring. Proposition 4.1. If S is a commutative semiring, then diam(Γ(Mn(S))) ≥ 3 for n ≥ 2. Proof. By Lemma 2.1, the only matrices in the intersection of the centralizer of Jn and the centralizer of J T n are scalar matrices, but these are of course central. Thus, the distance between matrices Jn and J T (cid:4) n in Γ(Mn(S)) is at least 3. We now prove that this bound is sharp, since the diameter is equal to 3 for all commuting graphs of matrices over nonentire commutative semirings. Theorem 4.2. If S is a nonentire commutative semiring, then diam(Γ(Mn(S))) = 3 for n ≥ 2. Proof. Since S is not entire, there exist nonzero x, y ∈ S such that xy = yx = 0. (Note that it may happen that x = y.) Choose arbitrary noncentral matrices A, B ∈ Mn(S). Case 1: Suppose xA and yB are noncentral matrices in Mn(S). Then A − xA − yB − B is a path of length 3 in Γ(Mn(S)). 7 Case 2: Suppose xA is central and yB is a noncentral matrix in Mn(S). Then AxE1,2 = xAE1,2 = E1,2xA = xE1,2A, so A − xE1,2 − yB − B is a path of length 3 in Γ(Mn(S)). Similarly, we treat the case when yB is central and xA is not. Case 3: Suppose xA and yB are central matrices in Mn(S). Then A − xE1,2 − yE1,2 − B is a path of length 3 in Γ(Mn(S)). The three cases considered show that the distance between any two matrices in Mn(S) is at most 3, and therefore diam(Γ(Mn(S))) ≤ 3. By Proposition 4.1, it follows that diam(Γ(Mn(S))) = 3. (cid:4) References [1] A. Abdollahi, Commuting graphs of full matrix rings over finite fields, Linear Algebra Appl., 428 (2008), 2947 -- 2954. [2] S. Akbari, A. Mohammadian, H. Radjavi, P. Raja: On the diameters of commuting graphs, Linear Algebra Appl., 418 (2006), 161 -- 176. [3] S. Akbari, P. Raja: Commuting graphs of some subsets in simple rings, Linear Algebra Appl., 416 (2006), 1038 -- 1047. [4] M. Akian, S. Gaubert, A. Guterman: Linear independence over tropical semirings and be- yond, Tropical and idempotent mathematics, 138, Contemp. Math., 495, Amer. Math. Soc., Providence, RI, 2009. [5] M. Develin, F. Santos, B. Sturmfels: On the rank of a tropical matrix, Combinatorial and computational geometry, Math. Sci. Res. Inst. Publ., 52, Cambridge Univ. Press, Cambridge, 2005, 213-242. [6] G. Dolinar, B. Kuzma, P. Oblak: On maximal distances in a commuting graph, preprint. [7] D. Dolzan, P. Oblak: Commuting graphs of matrices over semirings, Linear Algebra Appl., doi: 10.1016/j.laa.2010.04.014. [8] M. Giudici, A. Pope: The diameters of commuting graphs of linear groups and matrix rings over the integers modulo m, Australasian Journal of Combinatorics, 48 (2010), 221 -- 230. [9] A. Iranmanesh, A. Jafarzadeh: On commuting graph associated with the symmetric and al- ternating groups, J. Alg. Appl., 7 (2008), no. 1, 129 -- 146. [10] R. D. Katz, H. Schneider, S. Sergeev: On commuting matrices in max algebra and in classical nonnegative algebra, Linear Algebra Appl., doi: 10.1016/j.laa.2010.08.027. [11] A. Mohammadian, On commuting graphs of finite matrix rings, Commun. Algebra, 38 (2010), 988 -- 994. [12] Y. Segev: The commuting graph of minimal nonsolvable groups, Geom. Dedicata, 88 (2001), 55 -- 66. [13] I. Simon: On semigroups of matrices over the tropical semiring, RAIRO Inform. Thor. Appl., 28 (1994), no. 3-4, 277 -- 294. D. Dolzan: Department of Mathematics, Faculty of Mathematics and Physics, University of Ljubljana, Jadranska 19, SI-1000 Ljubljana, Slovenia; e-mail: [email protected] D. Kokol Bukovsek: Department of Mathematics, Faculty of Mathematics and Physics, University of Ljubljana, Jadranska 19, SI-1000 Ljubljana, Slovenia; e-mail: [email protected] lj.si P. Oblak: Faculty of Computer and Information Science, Trzaska 25, SI-1000 Ljubljana, Slovenia; e-mail: [email protected] 8
1602.01627	2	1602	2017-02-14T12:46:39	Covers of the arithmetic site	[ "math.RA" ]	We give an explicit description of the Barr- and Diaconescu covers of the arithmetic site, which are relevant to cohomology. Further, we construct the arithmetic site as the commutative shadow of a non-commutative topological space.	math.RA	math	COVERS OF THE ARITHMETIC SITE LIEVEN LE BRUYN b e F 4 1 ] . A R h t a m [ 2 v 7 2 6 1 0 . 2 0 6 1 : v i X r a Abstract. We give an explicit description of the Barr- and Diaconescu covers of the arithmetic site, which are relevant to cohomology. Further, we construct the arithmetic site as the commutative shadow of a noncommutative topolog- ical space. 1. The arithmetic site The arithmetic site, introduced and studied by A. Connes and C. Consani in [3] and [4], is an algebraic geometric object deeply related to the non-commutative geometric approach to the Riemann hypothesis. It involves two elaborate math- ematical concepts: the notion of topos in geometry and that of characteristic 1 structures in algebra. topos of sheaves on the small category C consisting of one object • with monoid of endomorphisms isomorphic to the multiplicative semigroup N× + of strictly posi- It is defined as a topos bC endowed with a structure sheaf O. Here, bC is the tive integers and equipped with the chaotic Grothendieck topology. Thus,bC is the category of contravariant functors from C to Sets, the category of sets. Any such functor Cop ✲ Sets is fully determined by a set X = X(•), equipped with a commuting family of endomorphisms Ψn : X ✲ X for all n ∈ N× +, with Ψ1 the identity morphism. That is, bC is equivalent to the category Sets − N× + of all sets equipped with a right action by the monoid N× + and action preserving maps as morphisms. The connection with characteristic one structures comes from the fact that this action is fully determined by the commuting endomorphisms Ψp for p ∈ P (the set of all prime numbers), which can be viewed as Frobenius morphisms. automorphisms of the monoid C (such as arbitrary permutations of the prime num- However, there is a large group of automorphisms of the topos bC arising from bers). For this reason, Connes and Consani equipbC with a structure sheaf, turning the presheaf toposbC in the arithmetic site (bC, Zmax). Here, Zmax is the fundamental Two topos-theoretic objects associated tobC carry a lot of arithmetic information. semiring in characteristic 1. That is, Zmax = (Z ∪ {−∞}, max, +), which means that ordinary addition is replaced by x⊕ y = max(x, y) and ordinary multiplication by x ⊗ y = x + y, see [3] and [4]. The points of the topos can be identified with the finite ad`ele classes [4] and the subobject-classifier is a complete Heyting algebra on the set of all (right) ideals of N× +. 1.1. The points ofbC : Recall that a point p of the toposbC is a geometric morphism from Sets to bC, see [11, VII.5]. Using Grothendieck's description via filtering 1 2 LIEVEN LE BRUYN subgroup is of the form functors, see [11, VII.6], a point p corresponds to a set P having a free left- N× +- action such that P is of rank one meaning that for all x, y ∈ P there exists an element z ∈ P and numbers n, m ∈ N× + such that x = n.z and y = m.z. As such, points correspond to additive subgroups of Q. If Af is the ring of finite ad`eles of Q, that is Af = Q ⊗Qp∈PbZp, then it is proved in [4, Prop. 2.5] that any such additive where bZ∗ is the multiplicative group of invertible elements in the ring of profinite integers bZ =Qp∈PbZp acting by (right) multiplication on Af . Moreover, the map Pa = {q ∈ Q a.q ∈bZ} for some a ∈ Af /bZ∗ a 7→ Pa induces a canonical bijection pts(bC) = Q× +\Af /bZ∗ = [Af ] between the isomorphism classes of points of the arithmetic site and finite ad`ele classes [Af ]. see [4, Prop. 2.5]. Alternatively one can describe additive subgroups of Q by supernatural numbers as in [2]. Recall that a supernatural number (also called a is a formal product s =Qp∈P psp where p runs over all prime numbers P and each sp ∈ N ∪ {∞}. The set S of all supernatural numbers forms a multiplicative semigroup with multiplication defined by exponent addition and the multiplicative semigroup N× + embeds in S via unique factorization. If s ∈ S, the corresponding additive subgroup is Ps = {q ∈ Q s.q ∈ S} In [2] it is shown that isomorphism classes of additive subgroups of Q correspond to equivalence classes on S defined by s ∼ s′ iff (sp = ∞ ⇔ s′ sp = s′ p = ∞ p for all but at most finitely many p [Af ] = pts(bC) = [S] [S] of supernatural numbers. The connection between the two descriptions description via supernatural numbers and will denote the isomorphism class of the point Hs by the corresponding equivalence class [s]. Note also from [10, 2.2] or That is, we have natural identifications between points ofbC and equivalence classes is induced via the identifications S = bZ/bZ∗. In this note we will always use the [4, §8] that we can view pts(bC) as the moduli space of noncommutative spaces 1.2. The subobject classifier Ω : The subobject-classifier Ω of the toposbC co- incides by with the set of all sieves, that is with all right ideals in the monoid N× +, see [11, I.4]. Any sieve S ∈ Ω is of the form corresponding to UHF-algebras. S =[i∈I niN× + for some (finite or infinite) set of positive integers ni which are incomparable with respect to the ordering by division. As a subobject-classifier in a topos, Ω is a (complete) Heyting algebra. Recall from [11, I.8] that a Heyting algebra H is a partially ordered set under ≤, with unique minimal element 0 and unique maximal element 1 and equipped COVERS OF THE ARITHMETIC SITE 3 with 2-ary operations ∧ and ∨ and an additional 2-ary operation → satisfying the following conditions for all x, y, z ∈ H: (1) x ∧ (y ∧ z) = (x ∧ y) ∧ z = x ∧ y ∧ z (2) x ∨ (y ∨ z) = (x ∨ y) ∨ z = x ∨ y ∨ z (3) x ∧ y = y ∧ x and x ∨ y = y ∨ x (4) x ∧ x = x and x ∨ x = x (5) 1 ∧ x = x and 0 ∨ x = x (6) x ∧ (y ∨ x) = x = (x ∧ y) ∨ x (7) x ∧ (y ∨ z) = (x ∧ y) ∨ (x ∧ z) and x ∨ (y ∧ z) = (x ∨ y) ∧ (x ∨ z) (8) x ≤ y iff x = x ∧ y iff y = x ∨ y (9) (x ∧ y) ≤ z iff x ≤ (y → z) In an Heyting algebra one can also define the 1-ary 'negation' operation ¬ by ¬x = (x → 0) and one always has x ≤ ¬¬x. A Heyting algebra is called Boolean if for all x ∈ H we have ¬¬x = x, or equivalently that x ∨ ¬x = 1. We will now make these operations explicit for the subobject-classifier Ω of the presheaf toposbC using [11, Prop. I.8.5]. Assume S, T ∈ Ω with mjN× + niN× then niN× mjN× + lcm(ni, mj)N× + + and T = [j∈J + ∪ [j∈J S =[i∈I S ∨ T = S ∪ T =[i∈I S ∧ T = S ∩ T = [(i,j)∈I×J [∀i:lcm(e,ni)∈T ¬S =(0 if S 6= 0 + and 0 = ∅ 1 if S = 0 1 = N× S → T = eN× + 1.3. The Grothendieck topologies on C. In addition to the Heyting algebra structure on Ω there is also a right N× S ⊙ n = n−1.S ∩ N× gcd(ni, n) N× + +-action ⊙ on Ω. ni + =[i∈I Recall from [11, III.2] that a Grothendieck topology on C is a subset G ⊂ Ω satisfying the following properties (1) 1 ∈ G (2) (stability) if S ∈ G, then S ⊙ N× (3) (transitivity) if S ∈ G and if R ∈ Ω such that R ⊙ S = {R ⊙ s s ∈ S} ⊂ G, + = {S ⊙ n n ∈ N× +} ⊂ G then R ∈ G. Observe that it follows that if S ∈ G and S ⊂ S' in Ω, then also S' ∈ G. The coarsest Grothendieck topology on C, with Gch = {1}, is called the chaotic topology on the category C. There are uncountable many different Grothendieck topologies on C. 4 LIEVEN LE BRUYN Proposition 1. Let P ⊂ P be a subset of prime numbers, then GP = {S ∈ Ω ∃m : mN× + ⊂ S and all prime divisors of m belong to P } is the smallest Grothendieck topology on C containing the sieves S(p) = pN× p ∈ P . + for all Proof. One uses the third property of Grothendieck topologies to deduce that S(m) = mN× + must belong to GP if all prime divisors of m belong to P using S(m) ⊙ S(p) =(S(m) if p ∤ m p ) if p m S( m But then all sieves S of the required form belong to GP and one verifies easily that these sieves indeed form a Grothendieck topology. (cid:3) But there are plenty of other Grothendieck topologies. If G is a Grothendieck topology on C and if ∪i∈I niN× + ∈ G for all possible prime divisors pi of ni. Hence, to G one can associate a set of subsets of prime numbers {Pj : j ∈ J} with the Pj minimal with respect to the property that ∪p∈Pj pN× + ∈ P. + ∈ G, then also all ∪i∈I piN× 2. Topos theoretic covers of the arithmetic site +- • 7→ {∗} 1 : C ✲ Sets We have seen that the subobject classifier ofbC is Ω with its natural right N× action. The terminal object of the toposbC is given by the functor +-action. Open objects inbC corre- where {∗} is the singleton with trivial right N× spond to subobjects of 1, that is to right N× be a sieve fixed under the N× N× +-maps {∗} ✲ Ω so its image must +-action. There are only two such sieves: ∅ = 0 and + = 1. That is Opens(bC) = {0, 1}, or equivalently, the toposbC is two-valued. 2.1. The Diaconescu cover. For every Grothendieck topos E there exists a locale X and an open surjective geometric morphism Sh(X) ✲ E, [11, IX.9 Thm 1]. The locale X is then called a Diaconescu cover of E. In this section we will determine the Diaconescu covers of Sh(C, G) of the arithmetic site equipped with a Grothendieck topology. As the topos of sheaves of sets on a site with underlying small category a partially ordered set is a locale by [11, IX.5 Thm 1], we first introduce the poset corresponding to the monoid category C, and call it the big cell. Definition 1. The big cell D is the category with one object [k] for every strictly positive integer k ∈ N+ and with morphisms induced by the partial ordering by reverse division [k] ≤ [l] if and only if lk That is, there is exactly one arrow [k] ✲ if and only if [k] ≤ [l]. There is a π ✲ C sending each object [k] to • and every morphism natural covariant functor D [k] / [l] to the endomorphism corresponding to k l . The terminology 'big cell' s due to John Conway in [6]. In order to under- stand groups commensurable with the modular group Γ occurring in moonshine, he looked at (projective classes of) lattices commensurable with the standard lat- tice < ~e1, ~e2 > and showed that they have a canonical form L =< M~e1 + g h~e2, ~e2 > / COVERS OF THE ARITHMETIC SITE 5 with M ∈ Q+ and g introduced a hyper-distance h ∈ Q/Z. For any other such lattice L′ =< N~e1 + i j ~e2, ~e2 > he δ(L, L′) = det(αDLL′) ∈ Z where α is the minimal strictly positive rational number such that αDLL′ ∈ M2(Z) with DLL′ =(cid:20)M g 1(cid:21)−1 1(cid:21) .(cid:20)N i 0 0 h j The hyperdistance is symmetric and depends only on the projective classes of lat- tices. We can turn the set of all classes of lattices into a graph by connecting lattices having prime hyperdistance. Conway calls the resulting graph the big picture and proved that is the product, over all prime numbers p, of the subgraphs of lattices with hyperdistance p, which is itself an infinite p + 1-valent tree, [6]. The big cell is the subgraph consisting of the lattices LM with M ∈ N+, and is therefore the product of an infinite number of graphs of type A∞. A Grothendieck topology G on C can be extended, in a canonical way, to a Grothendieck topology Gc on D by taking for each object [k] in D the set of sieves Gc([k]) = G where we mean for each sieve S ∈ Ω that S([k]) = { [l] / [k] : l k ∈ S}. With respect to these Grothendieck topologies on D and C the functor π satisfies the covering lifting property (clp) as well as the property of preserving covers of [11, p. 507]. It follows that the conditions of [11, Prop. IX.8.1] are satisfied and we obtain the following result. Proposition 2. For any Grothendieck topology G on C, the induced map from π on the corresponding sheaf toposes f : Sh(D, Gc) ✲ Sh(C, G) is an open, surjective geometric morphism. As D is a poset, the topos Sh(D, Gc) is localic, that is, Sh(D, Gc) is a Diaconescu cover of Sh(C, G). Recall that a point of a topos is a geometric morphisms from Sets to the topos. As composition of geometric morphisms remain geometric, the above functor f defines a map on the level of points We will now investigate this map in case G is the chaotic topology, that is, Sh(C, G) = pts(D, Gc) ✲ pts(C, G) + closed under taking divisors and least common multiples. As such they correspond naturally to super- PSh(C) =bC in which case Sh(D, Gc) = PSh(D) which topos we will denote bybD. Proposition 3. The points ofbD correspond to subsets S ⊂ N× natural numbers: pts(bD) = S. Proof. Recall from [11, VII,§5-6] that a point in the presheaf topos correspond to covariant functors satisfying the following conditions • There is some [k] such that P ([k]) 6= ∅ P : D ✲ Sets / as lcm(k, l)m it follows that also lcm(k, l) ∈ SP . To a subset S ⊂ N× number s ∈ S with s =Yp∈P psp 6 LIEVEN LE BRUYN • Given elements a ∈ P ([k]) and b ∈ P ([l]), there is an [m] with maps / [l] and an element c ∈ P ([m]) such that u(c) = a u v [k] [m] and v(c) = b. As D does not have parallel arrows, the third condition of [11, p. 386] is vacuous. Note that for all [k] we have that either P ([k]) is the empty set or a singleton. For, suppose a 6= b in P ([k]), apply the second condition to k = l. As there is a unique map from [m] to [k] it must follow that a = u(c) = b. By the first condition we have some [k] such that P ([k]) = {∗k}. If lk there is a unique map [k] u / [l] and therefore u(∗k) ∈ P ([l]) 6= ∅. Hence the set SP = {k ∈ N× + P ([k]) 6= ∅} is closed under division. It is also closed under taking least common multiples, for if k, l ∈ SP then by the second condition there is an m such that km and lm such that for the unique maps u and v from [m] to resp. [k] and [l] we have u(∗m) = ∗l and v(∗m) = ∗l + satisfying these conditions one can associate a supernatural where sp is the maximal exponent d such that pd ∈ S. Conversely, if s ∈ S then the subset is closed under division and least common multiples. (cid:3) S = {n ∈ N× + ns} / [l] 7→ id{∗} 1 : D ✲ Sets are described in the following result. ([k] 7→ {∗} [k] 2.2. The locale S. The terminal object inbD is given by the contravariant functor By definition, the set of opens in the toposbD correspond to subfunctors of 1, which Proposition 4. Opens(bD) correspond to subsets of N× between the opens inbD and sieves in C. Proof. A subfunctor S ∈ Sub(1) assigns to each [k] a set Sk which is either {∗} or the emptyset. If kl we must have the commuting diagram ples, that is, there is a natural one-to-one correspondence Opens(bD) ↔ Ω + closed under taking multi- 1 : S : {∗} id / {∗} Sk / Sl idSk so the set of k ∈ N× right ideal of the monoid N× + such that Sk = {∗} is closed under taking multiples, so is a (cid:3) +, that is, a sieve in C. o o / / / / O O / O O COVERS OF THE ARITHMETIC SITE 7 With operations as before, Ω is a frame, that is, a lattice with finite meets, finite and infinite joins, satisfying the infinite distributive law. More precisely, we have for all S, T, Ti ∈ Ω: (1) S ≤ T iff S ⊂ T. (2) S ∧ T = S ∩ T. (3) S ∨ T = S ∪ T. (4) S ∧ (∨iTi) = ∨i(S ∩ Ti). (5) 0 = ∅ and 1 = N× +. If cat is a poset category, thendcat is generated by the subobjects of the terminal object, and is called a localic topos as dcat ≃ Sh(X), the category of sheaves on a locale X, where the category of locales is defined to be the opposite of the category of frames, see for example [11, Chp. IX]. The locale X has a frame isomorphic to the frame of open sets of a topological space X if and only if X has enough points, that is, if elements of the corresponding frame can be distinguished by points of X, [11, Prop. IX.3.3]. Here, a point of X is a frame morphism p−1 : O(X) = Open(dcat) ✲ {0, 1} Theorem 1. The localic toposbD is equivalent to the category of sheaves of sets on the topological space S of all supernatural numbers, with open sets corresponding to S = ∪iniN× + ∈ Ω where bD = Sh(S) Xl(S) = Xl(∪iniN× +) = {s ∈ S ∃i : nis} Proof. As D is a poset category,bD ≃ Sh(X) for some locale X. A point of X is a frame morphism or equivalent, a map satisfying the following conditions: p−1 : O(X) = Open(bD) = Ω ✲ {0, 1} +) = 1 (1) p−1(N× (2) p−1(S ∧ S') = 0 iff p−1(S) = 0 or p−1(S') = 0 (3) p−1(∨Si) = 0 iff p−1(Si) = 0 for all i If Ps is the point of bD corresponding to the supernatural number s ∈ S, that is, : ks}, then there is a corresponding point ps of the locale X determined Ps = {[k] by the map p−1 s = Ω ✲ {0, 1} p−1(∪iniN× +) =(0 1 iff ∀i : ni ∤ s iff ∃i : nis as one easily verifies that this map satisfies the required properties. But then, the locale X has enough points. For any two distinct elements S, S′ ∈ Ω = O(X), there is some point p such that p−1(S) 6= p−1(S′). Indeed, let k ∈ S − S′ then p−1 k (S′) = 0. To finish the proof we have to show that pts(X) = S. k (S) = 1 whereas p−1 Let p−1 : Ω ✲ {0, 1} be a point of X and consider the set Dp = {n ∈ N+ p−1(nN× + it follows from the second condition on p−1 that Dp is closed under divisors and least common multiples, whence corresponds to a supernatural number s ∈ S. (cid:3) +) = 1}. As nN× + = lcm(m, n)N× + ∧ mN× 8 LIEVEN LE BRUYN This localic topology on S has a countable basis of open sets of the form Xl(n) = {s ∈ S : ns} Observe that these opens are clopen in the compact Hausdorff topology on S via 2.3. The projection map. We have already seen that the functor π : D ✲ C the identification withQp∈P(N ∪ {∞}) (see also the next section). induces an open surjective geometric morphismbD that the corresponding map between the points f ✲✲ bC. We now want to verify is the natural map, that is sending a supernatural number to its equivalence class. This requires a chase through the defining functors. pts(bD) = S ✲ pts(bC) = [S] corresponding to s ∈ S is defined by the covariant flat functor We have identified points ofbD with the set of supernatural numbers S. The point Ps : D ✲ Sets [k] 7→({∗k} ∅ if ks if k ∤ s This functor induces a geometric morphism p(s) : Sets ✲ bD determined by the adjoint functors p(s)∗ Sets p(s)∗ p(s)∗ = HomD(Ps, −) p(s)∗ = − ⊗D Ps see [11, VII.5 Thm 2, VII.2 Thm 1]. Here, HomD(Ps, S) is the contravariant functor (presheaf) if lks if ls but k ∤ s if k, l ∤ s {∗k} ✲ {∗l} ∅ ✲ {∗k} ∅ ✲ ∅ 7→ )bD [k] 7→(HomSets({∗k}, S) = S HomSets(∅, S) = {ik} if ks if k ∤ s [k] / [l] D ✲ Sets and with [k] / [l] mapping to the composition if lks {∗k} >⑥ ⑥ ⑥ ⑥ ⑥ ⑥ ⑥ ⑥ S ik ∅ >⑤ ⑤ ⑤ ⑤ ⑤ ⑤ ⑤ ⑤ {∗l} {∗l} S ∅ ∅ S @✁ ✁ ✁ ✁ ✁ il ✁ ✁ ✁ and to the canonical maps S 7→ {ik} if ls but k ∤ s and {il} ✲ {ik} if k, l ∤ s. In general, a presheaf G ∈ bD is a contravariant functor G : D ✲ Sets and : Gl ✲ Gk is defined by assigning a set Gk to all [k], and a morphism gk,l whenever lk. Working through the definition of [11, p. 356] we find that G ⊗D Ps is the set of equivalence classes of the set Gk with gk ∼ gl iff gk,l(gl) = gk Gks / ) j j / / / > / / > / / @ COVERS OF THE ARITHMETIC SITE 9 The functor π : D ✲ C induces by [11, VII.2 Thm 2] a geometric morphism π∗ π∗ ( C bD where π∗ assigns to a N× sition +-set M , that is a functor M : Cop ✲ Sets, the compo- π∗(M ) : Dop πop ✲ Cop M ✲ Sets Concretely, π∗(M ) is the presheaf inbD with all Gk = M and the maps gk,l are given l on M whenever lk. by the right action of k Theorem 2. The functor π : D ✲ C induces a map sending a supernatural number s to its equivalence class [s]. π : S = pts(bD) ✲ pts(C) = S/ ∼ Proof. In view of the foregoing, we have to prove that the composition p(s)∗ ◦ π∗ Sets p(s)∗ p(s)∗ )bD −⊗CQ+(s) π∗ π∗ C is naturally isomorphic to the functor − ⊗C Q+(s) where Q+(s) = { n k n ∈ N+, ks} is the set with left N× +-action by multiplication corresponding to the flat functor C ✲ Sets determining the point [s] ∈ pts(C). Let M be a set with a right N× +-action, then p(s)∗(π∗(M )) is the set of equivalence classes of M with [m′]k ∼ [m]1 iff m′ = m.k Gks that is, we should view m ∈ M in component k as m. 1 the set M ⊗C Q+(s). k and hence we recover indeed (cid:3) 2.4. The Barr cover. Barr's theorem [11, Thm. IX.9.2] asserts that for every Grothendieck topos E there exists a complete Boolean algebra B and a surjective geometric morphism Sh(B) ✲✲ E. The purpose of this theorem is to generalize the Godement resolution by flabby sheaves to the context of cohomology of arbitrary toposes. We will now describe the Barr cover of the arithmetic site. Theorem 3. Consider the complete Boolean algebra then there are surjective geometric morphisms That is, B is the Barr cover of the arithmetic site. {0, 1} B = 2Ω = YS∈Ω Sh(B) ✲✲ Sh(S) ✲✲ bC ( h h ) ( ( j j h h a a 10 LIEVEN LE BRUYN Proof. We have already seen that the locale S is a topological space with the frame of opens Open(S) ↔ Ω. This allows us to determine the Heyting algebra structure on Open(S) as in [11, I.8] and verify that it coincides with the Heyting algebra structure on Ω given before. In particular we have for each non-empty S that its negation ¬S = 0 = ∅. Following the construction of the proof of [11, Lemma IX.9.3], we have to deter- mine the frame of ¬¬-fixed points of the frame Open(S − S) of the closed sublocale S − S of S. Now, by definition Open(S − S) = {T ∈ Ω S ≤ T} ∪ {0} and by the above remark on the negation in Open(S) we have that the double- negation frame Open(S − S)¬¬ ≃ {0, 1} The claims now follow from the proof of [11, Lemma IX.9.3]. (cid:3) 3. Noncommutative covers of the arithmetic site Rather than studying Grothendieck topologies on C and their associated toposes of sheaves, we will consider in this section a number of ordinary topologies on the set of points pts(bC) = [Af ] = [S]. We then construct a noncommutative topolog- ical space, in the sense of [1], with commutative shadow [S] and a corresponding noncommutative Heyting algebra Θ, in the sense of [7], with commutative quotient the subobject classifier Ω. These results are inspired by a remark of A. Connes in [5, §5] where he asserts that the arithmetic site is but a semiclassical shadow of a still mysterious structure dealing with compactifications of Spec(Z). For this reason we first consider three topologies on the points [S] of the arithmetic site and the corresponding continuous maps from Spec(Z). We then construct a noncommutative frame Θ, elements of which can be viewed as constructible truth fluctuations on open sets in the sieve topology on the finite ad`ele classes. 3.1. The arithmetic topology: The locally compact topology on the finite ad`eles Af induces a weak remnant on the set of points of the arithmetic site, which we call the arithmetic topology. If we identify S = Qp∈P(N ∪ {∞}, view each of the factors as a point point compactification of the discrete topology on N then the arithmetic topology is the product topology turning S into a compact Hausdorff. The arithmetic topology is the induced topology on the equivalence classes [S] and was described in [10, Thm. 1]. It has a countable basis of open sets, corresponding to finite subsets P = {p1, . . . , pk} of prime numbers, defined by Xa(P ) = {[s] ∈ [S] ∀p ∈ P : sp 6= ∞} (note that this does not depend on the choice of representative in [s]). There is an obvious connection with the Zariski topology on the prime spectrum Spec(Z) which is P ∪ {0} with open sets of the form XZ (n) = {p ∈ P p ∤ n} ∪ {0}. We can map a prime number p to the class determined by the element (1, . . . , 1, 0, 1, . . .) ∈bZ with a zero in the factor bZp and ones in the other factors. A direct calculation proves the following result. Proposition 5. The inclusion map Spec(Z) ⊂ i ✲ pts(bC) = [S] given by p 7→ [p∞] COVERS OF THE ARITHMETIC SITE 11 is continuous with respect to the Zariski topology on Spec(Z) and the arithmetic topology on [S] as i−1(Xa(P )) = XZ (Qp∈P p). As a consequence, the direct image sheaf i∗OZ has as sections over the arithmetic topology on [S] Γ(i∗OZ, Xa(P )) = Z[ 1 ] Qp∈P p The stalk of i∗OZ in a point [s] of the arithmetic site is then (i∗OZ)[s] = { a b (b, p) = 1 ∀p ∈ P : sp = ∞} so, in particular, the stalk in [p∞] is the local ring Zp. 3.2. The sieve topology: In [10] we introduced another topology on the set of points of the arithmetic site, called the sieve topology as the basic open sets corre- spond to the sieves S ∈ Ω and are defined as Xs(S) = {[s] ∈ [S] ∃n1, n2, . . . ∈ S : nis} ∞Yi=1 (this definition does not depend on the choice of representative in [s]). The 'morale' behind this definition is that Xs(S) are precisely the points of the arithmetic site which are also points in the monoid category of S ∪ {id}. In [10, Thm. 6] it was shown that this sieve topology has some properties one might expect of the mythical space Spec(Z): it is compact, does not admit a countable basis of opens, every non-empty open set is dense and it satisfies the T1-separation property on incomparable points. Further note that the frame of open sets in the sieve topology is isomorphic to the lattice structure of the Heyting algebra Ω. For the connection with the arithmetic topology, note that if P = {p1, . . . , pk} we have Xa(P ) = [S] − Xs(p1N× + ∪ . . . ∪ pkN× +) That is, basic opens of the arithmetic topology are closed in the sieve topology. Still, we have a nice connection with Spec(Z). This time we map the prime ideal (p) to the class of the element (0, . . . , 0, 1, 0, . . .) ∈ bZ with a one in the factor corresponding to bZp and zeroes elsewhere. In this case we calculate: Proposition 6. The inclusion map Spec(Z) ⊂ given by j ✲ pts(bC) = [S] p 7→ [Yq6=p q∞] is continuous with respect to the Zariski topology on Spec(Z) and the sieve topology on [S] as for each sieve S = ∪i∈I niN× + we have with gcd(S) = gcd(ni, i ∈ I) j−1(Xs(S)) = Xa(gcd(S)) As a consequence the direct image sheaf j∗OZ has as sections with respect to the sieve topology Γ(j∗OZ, Xs(S)) = Z[ 1 gcd(S) ] Therefore, the stalk of j∗OZ in a point [s] is equal to (j∗OZ)[s] = Z[ 1 p : p∞s] 12 LIEVEN LE BRUYN or, in terms of the additive subgroups of Q (j∗OZ)[s] = \[t]=[s] Pt In particular, the stalk of j∗OZ at the image j(p) is the localization Zp. This result shows that the sieve topology and its induced structure sheaf from the connection with Spec(Z) is close to the intended structure sheaf of the arith- metic site as defined by Connes and Consani in [3] and [4]. The stalk of the structure sheaf O at a point [s] ∈ pts(bC), which as we have seen corresponds to the (isomorphism class) of the additive subgroup Ps of Q, will be the semiring Ps,max = (Ps ∪ {−∞}, max, +) as sub-semiring of Qmax. Clearly, if one ever wants to have a concrete realization of this "stalk" as a genuine stalk over some topology on the set of all points of the arithmetic site one has to dispose of the ambiguity that different semirings Ps,max defining the same point [s] are only isomorphic as semirings, and not equal. Replacing Ps.max by the intersection ∩[t]=[s]Pt,max is an obvious choice, which gives us the exact stalk of the sheaf j∗OZ in the sieve topology. 3.3. The patch topology: At this moment we have two topologies on pts(bC) = [S] and two natural continuous maps from Spec(Z) to [S] such that in both cases at least the stalks of the direct image structure sheaf give the expected ring Zp. The arithmetic topology might be better to transport arithmetic (ad`elic) information to the arithmetic site whereas the sieve topology, by its very nature, is better at encoding topos-theoretic information. In order to have the best of both worlds, we consider a common refinement of the two topologies, the patch topology (or constructible topology) with respect to the sieve topology. That is, open sets in the patch topology are unions of locally closed subsets in the sieve topology on pts(bC). That is, a basic open set in the patch topology is of the form Xp(S, T) = Xs(S) ∩ Vs(T) where S, T ∈ Ω and Vs(T) = [S] − Xs(T). As we have seen that the basic open sets in the arithmetic topology are closed in the sieve topology we have: Proposition 7. The patch topology on the points of the arithmetic site is a com- mon refinement of both the arithmetical and the sieve topology. It is the topology generated by both the open and closed sets of the sieve topology. 3.4. A noncommutative topology with shadow Ω. Our aim is to equip the set of points E = [S] × {0, 1} with a noncommutative topology such that its classical commutative shadow gives us the sieve topology on [S]. Here, a noncommutative topology has a corresponding 'frame' of open sets Θ which will be a skew-lattice on which the operations ∧ and ∨ are no longer commutative. By the shadow property we will mean that any skew-lattice map to a (commutative) lattice must factor through the lattice Ω of all opens of the sieve topology on [S]. It turns out, see for example [1] or [7], that in passing to the noncommutative setting we need to sacrifice either the top element 1 or the bottom element 0 in the definition of a skew-lattice. In this section we choose to sacrifice 1 as our construction will involve taking local sections of sheaves and whereas there is always a unique such section over 0 = ∅ there are usually lots of global sections. COVERS OF THE ARITHMETIC SITE 13 Definition 2. A noncommutative frame is a set Θ with 2-ary operations ∧ and ∨ (intersection resp. union of open sets) and a minimal element 0 such that for all x, y, z ∈ Θ the following conditions are satisfied: (1) x ∧ (y ∧ z) = (x ∧ y) ∧ z = x ∧ y ∧ z (2) x ∨ (y ∨ z) = (x ∨ y) ∨ z = x ∨ y ∨ z (3) x ∧ x = x (4) x ∨ x = x (5) x ∧ (x ∨ y) = x = x ∨ (x ∧ y) (6) (y ∨ x) ∧ x = x = (y ∧ x) ∨ x (7) x ∧ 0 = 0 = 0 ∧ x (8) x ∨ 0 = x = 0 ∨ x (9) x ∧ (y ∨ z) = (x ∧ y) ∨ (x ∧ z) (10) (y ∨ z) ∧ x = (y ∧ x) ∨ (z ∧ x) (11) x ∧ y ∧ x = x ∧ y (12) x ∨ y ∨ x = y ∨ x (13) x ∧ y ∧ z = x ∧ z ∧ y Here, (1)-(4) state that the operations are associative and idempotent, (5)-(6) are called the absorption properties, (9)-(10) are the strongly distributive properties and (11)-(12) are the left-handed properties, see [1] for more details. Θ becomes a partially ordered set via x ≤ y ⇔ x ∧ y = x = y ∧ x ⇔ x ∨ y = y = y ∨ x and (7)-(8) state that 0 is a minimal element with respect to this partial order. Condition (13) follows from the others and states that (Θ, ∧) is a left-normal semi- group. One defines an equivalence relation on the elements of a noncommutative frame Θ by x ∼ y ⇔ x ∧ y ∧ x = x and y ∧ x ∧ x = y A result of Jonathan Leech, see for example [1, Thm. 2.1] asserts that the induced operations turn the equivalence classes [Θ] into a (commutative) lattice without top element, and moreover, any noncommutative frame morphism Θ ✲ Λ where Λ is a lattice factors through [Θ]. The main result of Bauer et al. [1] on noncommutative Priestly duality tells us that any noncommutative topology Θ with commutative semiclassical shadow [Θ] ≃ Ω comes from the local sections of a sheaf with respect to the patch topology for Ω. We will now construct the simplest such sheaf in the case of interest to us, that is, when Ω is the commutative frame of opens of the sieve topology on the points of the arithmetic site. Definition 3. The sheaf Oc of constructible truth fluctuations has as sections over the open set Xs(S) for S ∈ Ω Γ(Oc, Xs(S)) = {x : Xs(S) ✲ B = {0, 1} continuous} Here, continuous is with respect to the discrete topology on the Boolean semifield B = {0, 1} and the induced patch topology on Xs(S). Alternatively, one can think of x as the characteristic function of an open subset of Xs(X) in the patch topology. 14 LIEVEN LE BRUYN Using the sheaf of constructible truth fluctuations Oc we will now construct the noncommutative frame Θ which is the set of 'opens' of a noncommutative topology on E = [S] × B. Theorem 4. Let Θ be the set of all pairs (S, x) where S ∈ Ω and x ∈ Γ(Xs(S), Oc). On Θ we define 2-ary operations ∧ and ∨ as follows: (S, x) ∧ (T, y) = (S ∧ T, xXs(S)∩Xs(T)) (S, x) ∨ (T, y) = (S ∨ T, yXs(T) ∪ xXs(S)−Xs(T)) With respect to these operations, Θ is a noncommutative frame with semiclassical shadow [Θ] ≃ Ω. there is an embedding of noncommutative frames Ω ⊂ ✲ Θ given by sending S to (S, 1Xs(S)). Further, (S, x) commutes with (T, y) for ∧ or ∨ if and only if xXs(S)∩Xs(T) = yXs(S)∩Xs(T). To each (S, x) ∈ Θ we associate the 'open' set in E = [S] × B Xnc(S, x) = {([s], x([s])) : [s] ∈ Xs(S)} The natural projection om the first factor makes E into an ´etale space over [S] with respect to the patch topology on [S]. Proof. All claims follow from computations, or by invoking [1]. Here, we will merely make the partial ordering explicit as well as the equivalence relation. Let (S, x) and (T, y) be elements of Θ. Then, if (S, x) ∧ (T, y) = (S ∧ T, xXs(S)∩Xs(T)) = (S, x) this implies that S ≤ T. The reverse condition (T, y) ∧ (S, x) = (T ∧ S, yXs(T)∩Xs(S)) = (S, x) asserts that S ≤ T and that yXs(S) = x. That is, the partial ordering on Θ (S, x) ≤ (T, y) ⇔ S ≤ T and yXs(S) = x is given by extension of local sections. Using this, we also have that if (S, x) ∧ (T, y) ∧ (S, x) = (S ∧ T, xXs(S)∩Xs(T)) = (S, x) is equivalent to S ≤ T whereas the other condition (T, y) ∧ (S, x) ∧ (T, y) = (S ∧ T, yXs(T)∩Xs(S)) = (T, y) is equivalent to T ≤ S. That is, (S, x) ∼ (T, y) iff S = T. As a consequence, [Θ] is just projection on the first factor so is isomorphic to the lattice Ω. (cid:3) We can, of course, repeat this construction starting from more involved sheaves with respect to the sieve topology on [S]. A natural choice would be to consider F to be the disjoint union of all the stalks of the structure sheaf j∗OZ. That is, F is equal to the following subset of [S] × Qmax: (j∗OZ)[s] = {([s], q) ∈ [S] × Qmax ∀[t] = [s] : q.t ∈ S} ∪ {([s], −∞) [s] ∈ S} G[s]∈S One can then consider the sheaf of constructible functions O having as its sections over Xs(S) {x : Xs(S) ✲ Qmax : Graf (x) ⊂ F and locally constant wrt. patch topology} Using the same operations as above we can then define a noncommutative frame structure on the collection of all (S, x) with S ∈ Ω and x ∈ Γ(O, Xs(S)) having as its semiclassical shadow Ω. COVERS OF THE ARITHMETIC SITE 15 3.5. A noncommutative Heyting algebra with shadow Ω. In [7] K. Cvetko- Vah introduced the notion of a skew-Heyting algebra. The starting point is again a skew-lattice but, in view of the crucial role of the top 1 as truth in logic, this time with a top element but without bottom element, to ensure noncommutativity of ∧ and ∨. The additional requirements for a skew-lattice with 1 to a a skew-Heyting algebra are by [7, Thm. 3.2] that for all x, y, z and u we have (1) (x → y) = (y ∨ x ∨ y → y) (2) (x → x) = 1 (3) (x ∧ (x → y) ∧ x) = x ∧ y ∧ x (4) (y ∧ (x → y)) = y and ((x → y) ∧ y) = y (5) (x → u ∨ (y ∧ z) ∨ u) = (x → (u ∨ y ∨ u)) ∧ (x → (u ∨ z ∨ u)) If we exchange the definition of ∨ with that of ∧ in theorem 4, reverse the partial ordering and define 1 = ∅ one can extend the resulting skew-lattice with top to a skew-Heyting algebra as this is an instance of the partial maps with poset domain example of [7, 4.4]. However, this is rather artificial and we want to keep the definitions of ∨ and ∧ of theorem 4 on Θ. For this reason we proceed differently and immediately define the 2-ary operation → on Θ (using the corresponding operation on Ω) as follows (S, x) → (T, y) = (S → T, y ∪ 1Xs(S→T)−Xs(T)) Observe that in the Heyting algebra Ω we always have that T ≤ (S → T) so the extension of the function y to Xs(S → T) makes sense. Of course, as we do not have a top element in Θ we cannot hope to satisfy condition (2) of a skew-Heyting algebra. That is, in the noncommutative world we have to sacrifice absolute truth. Based on the example below, K. Cvetko-Vak [8] formalized a non-commutative Heyting algebra to be a noncommutative frame (with bottom 0) with a 2-ary op- eration → satisfying the following axioms. Here, t is a distinguished element of T , the set of all top class elements of the poset. (NH1) x → y = (y ∨ (t ∧ x ∧ t) ∨ y) → y, (NH2) x → x = x ∨ t ∨ x, (NH3) x ∧ (x → y) ∧ x = x ∧ y ∧ x, (NH4) y ∧ (x → y) = y and (x → y) ∧ y = y, (NH5) x → (t ∧ (y ∧ z) ∧ t) = (x → (t ∧ y ∧ t)) ∧ (x → (t ∧ z ∧ t)). Theorem 5. With the above definition of (S, x) → (T, y), the noncommutative frame Θ becomes a noncommutative Heyting algebra. Moreover, any morphism of noncommutative Heyting algebras Θ ✲ Λ where Λ is a Heyting algebra, factors through the Heyting algebra Ω. Proof. Let x = (S, x), y = (T, y), z = (V, z) and t = (1, t). Using the definitions x∧y = (S∧T, x), x∨y = (S∨T, y∪xXs(S)−Xs(T), x → y = (S → T, y∪1Xs(S→T)−Xs(T) one verifies that both sides of the equalities in (NH1)-(NH5) are equal, see also [8, Example 3.1]. (cid:3) Acknowledgement: I like to thank K. Cvetko-Vah for pointing out that the example above did not satisfy the axioms given in a previous version of this paper and for providing the corrected axioms (NH1)-(NH5). 16 LIEVEN LE BRUYN References [1] A. Bauer, K. Cvetko-Vah, M. Gehrke, S.J. Van Gook and G. Kudryavtseva, A non- commutative Priestly duality, arXiv:1206.5848 (2012) [2] Ross A. Beaumont and H. S. Zuckerman, A characterization of the subgroups of the additive rationals, Pacific J. Math. 1, 169-177 (1951) [3] Alain Connes and Caterina Consani, The Arithmetic Site, Le Site Arithm´etique, arXiv:1405.4527 (2014) [4] Alain Connes and Caterina Consani, Geometry of the Arithmetic Site, arXiv:1502.05580 (2015) [5] Alain Connes, An essay on the Riemann hypothesis, arXiv:1509.05576 (2015) [6] John H. Conway, Understanding groups like Γ0(N ), In "Groups, difference sets, and the Monster", Walter de Gruyter (1996) [7] Karin Cvetko-Vah, On skew Heyting algebras, arXiv:1410.8780 (2014) [8] Karin Cvetko-Vah, Non-commutative frames, preprint (2017) [9] Lieven Le Bruyn, Supernatural numbers and a new topology on the Arithmetic Site, arXiv:1407.5538 (2014) [10] Lieven Le Bruyn, The sieve topology on the arithmetic site, Journal of Algebra and Applica- tions (2016) [11] Saunders Mac Lane and Ieke Moerdijk, Sheaves in Geometry and Logic, a first introduction to topos theory, Universitext, Springer-Verlag (1992) Department of Mathematics, University of Antwerp, Middelheimlaan 1, B-2020 Antwerp (Belgium), [email protected]
1902.09354	1	1902	2019-02-25T15:16:21	Centrosymmetric nonnegative realization of spectra	[ "math.RA" ]	A list $\Lambda =\{\lambda _{1},\lambda _{2},\ldots ,\lambda _{n}\}$ of complex numbers is said to be realizable if it is the spectrum of an entrywise nonnegative matrix. In this paper we intent to characterize those lists of complex numbers, which are realizable by a centrosymmetric nonnegative matrix. In particular, we show that lists of nonnegative real numbers, and lists of complex numbers of Suleimanova type (except in one particular case), are always the spectrum of some centrosymmetric nonnegative matrix. For the general lists we give sufficient conditions via a perturbation result. We also show that for $n=4,$ every realizable list of real numbers is also realizable by a nonnegative centrosymmetric matrix.	math.RA	math	Centrosymmetric nonnegative realization of spectra∗ b e F 5 2 ] . A R h t a m [ 1 v 4 5 3 9 0 . 2 0 9 1 : v i X r a Ana I. Julio, Oscar Rojo, Ricardo L. Soto† Departamento de Matem´aticas, Universidad Cat´olica del Norte Casilla 1280, Antofagasta, Chile. Abstract A list Λ = {λ1, λ2, . . . , λn} of complex numbers is said to be real- izable if it is the spectrum of an entrywise nonnegative matrix. In this paper we intent to characterize those lists of complex numbers, which are realizable by a centrosymmetric nonnegative matrix. In particular, we show that lists of nonnegative real numbers, and lists of complex numbers of Suleimanova type (except in one particular case), are al- ways the spectrum of some centrosymmetric nonnegative matrix. For the general lists we give sufficient conditions via a perturbation result. We also show that for n = 4, every realizable list of real numbers is also realizable by a nonnegative centrosymmetric matrix. AMS classification: Key words: Inverse eigenvalue problem, nonnegative matrix, centrosym- 15A18, 15A29 metric matrix. 1 Introduction The nonnegative inverse eigenvalue problem (hereafter NIEP) is the prob- lem of finding necessary and suffcient conditions for the existence of a n × n entrywise nonnegative matrix with prescribed complex spectrum Λ = ∗Supported by conicyt-fondecyt 1170313, Chile; conicyt-pai 79160002, 2016, Chile. †Corresponding author, [email protected], (Ana I Julio), [email protected] (Oscar Rojo), [email protected] (Ricardo L. Soto) 1 {λ1, λ2, . . . , λn}. If there exists an n × n nonnegative matrix A with spec- trum Λ we say that Λ is realizable and that A is the realizing matrix. A complete solution for the NIEP is known only for n ≤ 4, which shows the difficulty of the problem. Throughout this paper, if Λ = {λ1, λ2, . . . , λn} is realizable by a nonnegative matrix A, then λ1 = ρ(A) = max{λi , λi ∈ Λ} is the Perron eigenvalue of A. We shall denote the transpose of a matrix A by AT , ⌊x⌋ and ⌈x⌉ denote the largest integer less than or equal to x and the smallest integer greater than or equal to x, respectively. J will denote the counteridentity matrix, that is, J = [en en−1 · · · e1]. Then it is clear that J T = J, J 2 = I. Observe that multiplying a matrix A by J from the left results in reversing its rows, while multiplying A by J from the right results in reversing its columns. A vector x is called symmetric if Jx = x. The set of all n× n real matrices with constant row sums equal to α ∈ R will be denote by CS α. It is clear that e = [1, 1, . . . , 1]T is an eigenvector of any matrix in CS α, corresponding to the eigenvalue α. The relevance of matrices with constant row sums is due to the well known fact that the problem of finding a nonnegative matrix A with spectrum Λ = {λ1, . . . , λn} is equivalent to the problem of finding a nonnegative matrix B ∈ CS λ1 with spectrum Λ, that is, A and B are similar if the Perron eigenvalue is simple and they are cospectral otherwise (see [4]). In this paper we study the NIEP for centrosymmetric matrices. Centrosym- metric matrices appear in many areas: physics, communication theory, dif- ferential equations, numerical analysis, engineering, statistics, etc. Now we state the definition and certain properties about centrosymmetric matrices. Definition 1.1 A matrix C = (cij) ∈ Mm,n is said to be centrosymmetric, if its entries satisfy the relation ci,j = cm−i+1,n−j+1. The definition means that a centrosymmetric matrix C can be written as JmCJn = C, where Jn is the n × n counteridentity matrix. That is, Jn = [en en−1 · · · e1] . For m = n, we shall write J instead Jn. 2 The following properties and results on centrosymmetric matrices are easy to verify (see [2]). 1 , if exists, ii) C T Lemma 1.1 [2] Let C1 and C2 be centrosymmetric matrices, and α1, α2 ∈ R. Then, i) C−1 1 , iii) α1C1 ± α2C2, iv) C1C2 are all centrosymmetric. Lemma 1.2 [2] Let C be an n×n centrosymmetric nonnegative matrix. If v is an eigenvector of C corresponding to the eigenvalue λ, then Jv is also an eigenvector of C corresponding to λ. Moreover, if λ1 is the Perron eigenvalue of C, then there is a nonnegative eigenvector x such that Jx = x. Theorem 1.1 [2] Let C be an n × n centrosymmetric matrix. i) If n = 2m, then C can be written as C =(cid:20)A JBJ B JAJ(cid:21) , where A, B and J are m × m matrices. ii) If n = 2m + 1, then C can be written as A x JBJ yT yT J B Jx JAJ c C =    , where A, B and J are m × m matrices, x and y are m-dimensional vectors, and c is a real number. Theorem 1.2 [2] Let C be an n × n centrosymmetric matrix. B JAJ(cid:21), then C is orthogonally similar to the matrix i) If C =(cid:20)A JBJ (cid:20)A + JB A − JB(cid:21) . Moreover, if C is nonnegative with the Perron eigenvalue λ1, then λ1 is the Perron eigenvalue of A + JB. ii) If C is written as C =  A x JBJ yT yT J B Jx JAJ c 3  , then C is orthogonally similar √2yT to the matrices  c √2x A + JB  A + JB √2x  √2yT c  Perron eigenvalue λ1, then λ1 is the Perron eigenvalue of (cid:20) c  . Moreover, if C is a nonnegative matrix with the √2yT √2x A + JB(cid:21).   and A − JB A − JB The following theorem, due to R. Rado and published by H. Perfect in [6], will be used to throughout the paper to show some of our results. Rado's theorem show how to modify r eigenvalues of an n × n matrix, r < n, via a rank-r perturbation, without changing any of the remaining n−r eigenvalues (see [3, 9] for the way in which Rado's result has been applied to the NIEP). The case r = 1, is the well known Brauer's theorem [1], which states that If A has eigenvalues λ1, λ2, . . . , λn, then A + vqT , has eigenvalues λ1, . . . , λk−1, λk + vT q, λk+1, . . . , λn, where Av = λkv and q ∈ Cn. (1) Theorem 1.3 Rado [6] Let M be an n × n arbitrary matrix with spectrum Λ = {λ1, λ2, . . . , λn}. Let X = [x1 · · · xr] be such that rank(X) = r and Mxi = λixi, i = 1, . . . , r, r < n. Let C be an r×n arbitrary matrix. Then M + XC has eigenvalues µ1, . . . , µr, λr+1, . . . , λn, where µ1, . . . , µr are eigenvalues of the matrix Ω + CX with Ω = diag{λ1, . . . , λr}.   In [10] the authors prove that a realizable list of complex numbers of Suleimanova type Λ = {λ1, λ2, . . . , λn}, that is, with λ1 being the Perron eigenvalue, and λi ∈ F = {λi ∈ C : Reλi ≤ 0,Reλi ≥ Imλi}, i = 2, . . . , n, is in particular realizable with spectrum Λ and with prescribed diagonal entries ω1, ω2, . . . , ωn if and only if ωi = λi. This result, which we shall use in Section 3, follows directly from the following lemma and (1), which we set here for sake of completeness: n Pi=1 n Pi=1 4 n Pi=2 n n n n Pi=2 Lemma 1.3 Let Λ′ = {− λi, λ2, . . . , λn} be a realizable list of complex numbers, with Perron eigenvalue − λi, the list Λ = {λ1, λ2, . . . , λn} is the spectrum of a nonnegative matrix with prescribed diagonal entries Γ = {ω1, ω2, . . . , ωn}, if and only if Proof. Since Λ′ is realizable, there exists a nonnegative matrix B ∈ CS β, λi, with spectrum Λ′ and tr(B) = 0. Let qT = [ω1, . . . , ωn] . where β = − Then from (1), A = B + eqT is a nonnegative matrix with spectrum Λ and with diagonal entries ω1, ω2, . . . , ωn. Pi=2 λi. Then for any λ1 ≥ − Pi=1 Pi=1 Pi=2 ωi = λi. n 2 Real centrosymmetric matrices with pre- scribed spectrum In this section we show that a list of complex numbers Λ = {λ1, λ2, . . . , λn}, with Λ = Λ, is always the spectrum of a real centrosymmetric matrix (not necessarily nonnegative). That is, the real centrosymmetric inverse eigen- value problem for a list of conjugate complex numbers has always a solution. Theorem 2.1 Let Λ = {λ1, λ2, . . . , λn} be a list of complex numbers, with Λ = Λ, n ≥ 3. Then Λ is the spectrum of an n × n real centrosymmetric matrix. Proof. We shall distinguish two cases: Case 1. Let n be even. First we consider Λ with only two real numbers, that is, Λ = {λ1, λ2, z1, z1, . . . , zm, zm}, zj = aj +ibj , aj ∈ R, bj > 0, j = 1, . . . , m. If m is even, we take the partition Λ = Λ1 ∪ Λ2 with Λk = {λk1, zk1, zk1, . . . , zk m 2 , zk m 2 }, k = 1, 2, λ11 = λ1, λ21 = λ2. Then, from Theorem 1.2 the (m + 1) × (m + 1) matrices a1j (cid:21) and A − JB = λ2 Mj=1(cid:20)a1j −b1j A + JB = λ1 b1j m 2 m 2 Mj=1(cid:20)a2j −b2j a2j (cid:21) b2j 5 have spectrum Λ1 and Λ2 respectively. Then A = 1 and B = 1 2(A + JB + (A − JB)) 2 J(A + JB − (A − JB)). Therefore C =(cid:20)A JBJ B JAJ(cid:21) is real centrosymmetric with spectrum Λ. If m is odd, we take the partition Λ = Λ1 ∪ Λ2 with Λ1 = {λ11, λ12, z11, z11, . . . , z1⌊ m 2 ⌋, z1⌊ m Λ2 = {z21, z21, . . . , z2⌈ m 2 ⌉}. 2 ⌉, z2⌈ m 2 ⌋}, λ11 = λ1, λ12 = λ2 Then, A + JB = λ1 ⊕ λ2 A − JB = ⌊ m 2 ⌋ b1j Mj=1(cid:20)a1j −b1j a1j (cid:21) and a2j (cid:21) Mj=1(cid:20)a2j −b2j ⌈ m 2 ⌉ b2j have the spectra Λ1 and Λ2, respectively and the proof follows as above. It is clear that if Λ has r real numbers, r, m even, we take the partition Λ = Λ1∪Λ2 as above, with Λk, k = 1, 2, having m + r 2 numbers, m complex numbers and r 2 real numbers. If m is odd and Λ has r real numbers, r even, then Λ1 will have ⌊ m real numbers, while Λ2 will have ⌈ m 2 ⌉ complex numbers plus r−2 real numbers. Then the proof follows as before. 2 ⌋ complex numbers plus r+2 2 2 Case 2: Let n be odd. First we consider Λ with only one real number, that is, Λ = {λ1, z1, z1, . . . , zm, zm}, zj = aj + ibj , aj ∈ R, bj > 0, j = 1, . . . , m. If m is even we take the partition Λ = Λ1 ∪ Λ2 with Λ1 = {λ11, z11, z11, . . . , z1 m Λ2 = {z21, z21, . . . , z2 m , z2 m , z1 m 2 }. 2 2 2 }, λ11 = λ1, Then the matrices λ1 m 2 Mj=1(cid:20)a1j −b1j a1j (cid:21) =(cid:20)λ1 b1j A + JB(cid:21) and 6 A − JB = m 2 Mj=1(cid:20)a2j −b2j a2j (cid:21) , b2j have spectrum Λ1 and Λ2, respectively. Then A = 1 and B = 1 2 J(A + JB − (A − JB)). Hence from Theorem 1.2 2 (A + JB + (A − JB)) A 0 JBJ 0 λ1 B 0 JAJ 0     is real centrosymmetric with spectrum Λ. On the other hand, if m is odd, a partition Λ = Λ1 ∪ Λ2 must be of the form Λ1 = {z11, z11, . . . , z1⌈ m Λ2 = {λ21, z21, z21, . . . , z2⌊ m 2 ⌉, z1⌈ m 2 ⌉}, 2 ⌋, z2⌊ m 2 ⌋}, λ21 = λ1, k = 1, 2 Then the matrices ⌈ m 2 ⌉ Mj=1(cid:20)a1j −b1j a1j (cid:21) = b1j a11 −b11 0 b11 0 0 ... 0   0 · · · 0 A + JB   A − JB = λ1 ⌊ m 2 ⌋ a2j (cid:21) Mj=1(cid:20)a2j −b2j b2j and have spectrum Λ1 and Λ2 respectively. Then A = 1 and B = 1 2 J(A + JB − (A − JB)). Hence 2(A + JB + (A − JB)) A x JBJ yT a11 yT J B Jx JAJ     where x = 1√2[b11, 0 · · · , 0]T and yT = 1√2 [−b11,· · · , 0], is a real centrosym- metric matrix with spectrum Λ. It is easy to see that the list Λ = {λ1, z1, z1, . . . , zm, zm} can be extended 7 to a list Λ′ = {λ1, . . . , λp, z1, z1, . . . , zm, zm} with p real numbers, p odd, which always admit a partition in two self-conjugate lists Λ1, Λ2 such that Λ1 has ⌈ 2m+p 2 ⌉ eigenvalues and Λ2 has the remaining eigenvalues, and a real centrosymmetric matrix with spectrum Λ′ can be obtained as above. 3 Centrosymmetric nonnegative inverse eigen- value problem In this section we study the NIEP for centrosymmetric matrices. First, we show that lists of real nonnegative numbers are always realizable for a cen- trosymmetric nonnegative matrix. Second, we show that lists of complex numbers of Suleimanova type [11] are also realizable for centrosymmetric nonnegative matrices, except if the list has only one real number and m pairs of conjugated complex numbers, with m being odd. Third, for the gen- eral lists, we give sufficient conditions for the existence of a centrosymmetric nonnegative matrix with prescribed complex spectrum via a perturbation re- sult. Finally, we study the centrosymmetric realizability of lists of complex numbers of size n = 4 with prescribed diagonal entries. We start by showing that a list of real nonnegative numbers Λ = {λ1, . . . , λn} is always realizable by a centrosymmetric matrix. In addition, if λ1 is simple, Λ is realizable by a centrosymmetric positive matrix. Theorem 3.1 Let Λ = {λ1, λ2, . . . , λn} be a list of nonnegative real numbers with λ1 ≥ λ2 ≥ · · · ≥ λn ≥ 0. Then Λ is realizable by an n × n centrosym- metric matrix. Proof. For even n = 2m, we take the partition Λ = Λ1 ∪ Λ2, where Λ1 = {λ1, . . . , λm}, Λ2 = {λm+1, . . . , λn}. Then for A + JB = diag{λ1, . . . , λm} and A − JB = diag{λm+1, . . . , λn} we have that A = 1 2 diag{λ1 + λm+1, . . . , λm + λn}, B = 1 2 Jdiag{λ1− λm+1, . . . , λm − λn}, are nonnegative matrices, and a solution matrix is of the form B JAJ(cid:21) . C =(cid:20)A JBJ 8 For odd n = 2m + 1, we take the partition Λ = Λ1 ∪ Λ2, where Then for Λ1 = {λ1, . . . , λ⌈ n 2 ⌉+1, . . . , λn}. 2 ⌉}, Λ2 = {λ⌈ n c (cid:21) = diag{λ1, . . . , λ⌈ n 2 ⌉} (cid:20)A + JB √2x √2yT and A − JB = diag{λ⌈ n c = λ⌈ n 2 ⌉. Therefore 2 ⌉+1, . . . , λn}, we have A + JB = diag{λ1, . . . , λ⌊ n 2 ⌋}, A = B = 1 2 1 2 diag{λ1 + λ⌈ n Jdiag{λ1 − λ⌈ n 2 ⌉+1, . . . , λ⌊ n 2 ⌉+1, . . . , λ⌊ n 2 ⌋ + λn}, 2 ⌋ − λn}, are nonnegative matrices and a solution matrix is 0 A 0 λ⌈ n 2 ⌉ 0 B JBJ 0 JAJ C =    . In [5] Perfect introduces the n × n matrix · · · 1 1 1 · · · 1 −1 1 · · · −1 0 1 ... ... ... · · · 1 −1 · · · 0 0 1 1 1 ... P =     , (2) and she proves that if D = diag{λ1, λ2, . . . , λn}, with λ1 > λ2 ≥ · · · ≥ λn ≥ 0, then A = P DP −1 is a positive matrix in CS λ1. As a consequence, we have the following result, which gives a very simple way to compute a centrosymmetric positive matrix with prescribed nonnegative spectrum. Theorem 3.2 Let Λ = {λ1, λ2, . . . , λn} be a list of nonnegative real numbers with λ1 > λ2 ≥ · · · ≥ λn ≥ 0. Then Λ is realizable by an n × n centrosym- metric positive matrix. 9 Proof. For even n = 2m, we take the partition Λ = Λ1 ∪ Λ2, with Λ1 = {λ1, . . . , λm}, Λ2 = {λm+1, . . . , λn}. We set A+JB = P DP −1 > 0, where D = diag{λ1, . . . , λm}, P is the Perfect matrix in (2), and A − JB = diag{λm+1, . . . , λn}. Then A = 1 2 (P DP −1 + diag{λm+1, . . . , λn}) > 0, B = 1 2 J(P DP −1 − diag{λm+1, . . . , λn}). and Therefore, C =(cid:20)A JBJ B JAJ(cid:21) (3) is a centrosymmetric matrix with spectrum Λ. It remains to prove that C is positive. In fact, if djj, j = 1, 2, . . . , m, are the diagonal entries of P DP −1, we must show that It is enough to show that B is positive. djj > λm+j, for all j = 1, 2, . . . , m. It is easy to see that the diagonal entries of the matrix P DP −1 are given by: 1 2m−1 λ1 + 1 2m−2 λ1 + 1 2m−1 λ2 + 1 2m−2 λ2 + 1 2m−2 λ3 + · · · + 1 2m−3 λ3 + · · · + 1 22 λm−1 + 1 22 λm−2 + 1 2 1 2 λm = d22 λm−1 d11 = d33 = ... ... dm−1,m−1 = dmm = 1 2 λ3 1 1 22 λ2 + 22 λ1 + 1 1 λ1 + λ2. 2 2 For j = 1, d11 > λm+1. In fact: since 1 2m−1 λ1 > 1 2m−1 λm+1, 1 2m−1 λ2 ≥ 1 2m−1 λm+1, . . . , 1 2 λm ≥ 1 2 λm+1, 10 then by adding the inequalities we have d11 > ( 1 2m−1 + 1 2m−2 + · · · + 1 1 22 + 2m−1 + 1 + 1 + 2 + 22 + · · · + 2m−3 + 2m−2 1 + (2m−1 − 1) 2m−1 λm+1 = = 1 2 )λm+1 λm+1 2m−1 = λm+1. Since d11 = d22 and λm+1 ≥ λm+2, then d22 > λm+2. By proceeding in the same way we have djj > λm+j, j = 3, 4, . . . , m. There- fore B > 0, and the matrix C in (3) is centrosymmetric positive. If n = 2m + 1 is odd, then we take the partition Λ = Λ1 ∪ Λ2 with Then for (cid:20)A + JB √2x √2yT A − JB = diag{λ⌈ n Λ1 = {λ1, . . . , λ⌈ n 2 ⌉}, Λ2 = {λ⌈ n c (cid:21) = P DP −1 =(cid:20)A11 2 ⌉+1, . . . , λn}, we have bT 2 ⌉+1, . . . , λn}. am+1,m+1(cid:21) > 0, and a A = 1 2 B = (A11 + diag{λ⌈ n 1 2 J(A11 − diag{λ⌈ n 2 ⌉+1, . . . , λn}) > 0, 2 ⌉+1, . . . , λn}). B is positive from the same argument as in the even case. Therefore the matrix C =  A 1√2a 1√2bT am+1,m+1 B J 1√2 a JBJ 1√2bT J JAJ ,   is centrosymmetric positive with spectrum Λ. Next we shall see an anomalous case. That is, we shall prove that if a list realizable Λ has only one real positive number and m pairs of conjugated complex numbers, with m being odd, then Λ cannot be the spectrum of a centrosymmetric nonnegative matrix. 11 Theorem 3.3 Let Λ = {λ1, z1, z1, . . . , zm, zm} be a realizable list of complex numbers, with odd m, zj = aj + ibj j = 1, . . . , m, a ∈ R, bj > 0. Then Λ cannot be the spectrum of a centrosymmetric nonnegative matrix. Proof. Suppose that Λ is the spectrum of a centrosymmetric nonnegative matrix C of order 2m + 1. Then C is of the form C =  c A x JBJ yT J yT B Jx JAJ  , where A and B are m × m nonnegative matrices, x, y are nonnegative m × 1 matrices, and c is a nonnegative real number. From Theorem 1.2 C is orthogonally similar to the matrix  A − JB  √2yT c √2x A + JB   , √2yT where (cid:20) c √2x A + JB(cid:21) is an (m + 1) × (m + 1) nonnegative matrix with m + 1 eigenvalues of C including the Perron eigenvalue. That is, C has m (an odd number) complex eigenvalues, which is a contradiction. Now we consider the centrosymmetric realizability of lists of Suleimanova type. We start with the following simple result: Lemma 3.1 Let Λ = {λ1, λ2, . . . , λn} be the spectrum of a centrosymmetric nonnegative matrix and let ǫ > 0. Then {λ1 + ǫ, λ2, . . . , λn} is also the spectrum of a centrosymmetric nonnegative matrix. Proof. Let C be a centrosymmetric nonnegative matrix with spectrum Λ. Then from Lemma 1.2, there exists v ≥ 0, Cv = λ1v, with Jv = v, and from Brauer's Theorem (see (1)) the matrix C + ǫ vT v vvT , is centrosymmetric nonnegative with spectrum {λ1 + ǫ, λ2, . . . , λn}. It is well known that lists of complex numbers Λ = {λ1, λ2, . . . , λn} of 12 Suleimanova type are realizable by a nonnegative matrix if only ifPn i=1 λi ≥ 0. We prove that a list of this type is in particular realizable by a centrosym- metric nonnegative matrix, except if the list is as in the Theorem 3.3, that is, it has only one real positive number and m pairs of conjugated complex numbers, with m being odd. The following result show that any realizable list of complex numbers of Suleimanova type, with two real numbers and m pairs of complex numbers is always the spectrum of a centrosymmetric nonnegative matrix of order n = 2m + 2. As a consequence, all realizable lists of complex numbers of Suleimanova type, with an even number of elements, are always realizable by centrosymmetric matrices. Lemma 3.2 Let Λ = {λ1, λ2, z1, z1, . . . , zm, zm}, with λ2, zj, zj ∈ F , zj = aj + ibj, aj ∈ R, bj > 0, j = 1, . . . , m, be a realizable list of complex numbers. Then Λ is the spectrum of a centrosymmetric nonnegative matrix. Proof. Since Λ is realizable if only if Pn Λ′ = {λ′1, λ2, z1, z1, . . . , zm, zm}, with λ′1 = −λ2 − 2 i=1 λi ≥ 0, we take the list aj, and aj = Rezj. m Xj=1 For even m we take the partition Λ = Λ1 ∪ Λ2 with Λ1 = {λ11, z11, z11, . . . , z1 m Λ2 = {λ21, z21, z21, . . . , z2 m 2 2 , z1 m , z2 m 2 }, λ11 = λ′1, 2 }, λ21 = λ2. Then, A − JB = λ2 m 2 a2j (cid:21) Mj=1(cid:20)a2j −b2j b2j has spectrum Λ2, and from Lemma 1.3 we can always compute a nonnegative matrix A + JB with spectrum Λ1 and diagonal entries −λ2,−a21,−a21, . . . ,−a2 m 2 ,−a2 m 2 . Since Rezj ≥ Imzj , j = 1, . . . , m, both matrices, A = 1 2 (A + JB + A − JB) and B = 1 2 J((A + JB) − (A − JB)). 13 are nonnegative and C′ =(cid:20)A JBJ B JAJ(cid:21) is centrosymmetric nonnegative with spectrum Λ′. Finally, from Lemma 3.1, C = C′ + Pn i=1 λi vT v vvT , where v is the Perron eigenvector of C′, is centrosym- metric nonnegative with spectrum Λ. For odd m we take the partition Λ = Λ1 ∪ Λ2 with 2 ⌋, z1⌊ m 2 ⌉}, λ11 = λ′1, λ12 = λ2, Λ1 = {λ11, λ12, z11, z11, . . . , z1⌊ m Λ2 = {z21, z21, . . . , z2⌈ m 2 ⌉, z2⌈ m 2 ⌋}, and the proof follows as above. More generally, if Λ has r ≥ 2 real numbers, with even r, we have the following result: Corollary 3.1 Let Λ = {λ1, . . . , λr, z1, z1, . . . , zm, zm}, with zj, zj ∈ F , zj = aj + ibj, aj ∈ R, bj > 0, j = 1, . . . , m; λj ∈ F , j = 2, . . . , r, be a realizable list of complex numbers. Then Λ is the spectrum of a centrosym- metric nonnegative matrix. Proof. We consider the list Λ′ = {λ′1, λ2, . . . , λr, z1, z1, . . . , zm, zm}, with λ′1 = − λi − 2 m r aj, where aj = Rezj, j = 1, . . . , m. Xi=2 Xj=1 Then, for even m we take the partition of Λ′ = Λ1 ∪ Λ2 with: Λ1 = {λ11, λ12, . . . , λ1 r Λ2 = {λ21, λ22, . . . , λ2 r 2 2 , z11, z11, . . . , z1 m 2 , z21, z21, . . . , z2 m 2 while for odd m we take , z1 m , z2 m 2 }, λ11 = λ′1, 2 }, Λ1 = {λ11, λ12, . . . , λ1, r Λ2 = {λ21, λ22, . . . , λ2, r 2 +1, z11, z11, . . . , z1⌊ m 2−1, z21, z21, . . . , z2⌈ m 2 ⌋, z1⌊ m 2 ⌉, z2⌈ m 2 ⌋}, λ11 = λ′1 2 ⌉}, and the proof follows as the proof of Lemma 3.2. Now, we consider lists of complex numbers of Suleimanova type with an odd number of elements. We start with the following result: 14 Lemma 3.3 Let Λ = {λ1, z1, z1, . . . , zm, zm} be a list of complex numbers with zj, zj ∈ F , zj = aj + ibj, aj ∈ R, bj > 0, j = 1, . . . , m, m even, which is realizable. Then Λ is the spectrum of a centrosymmetric nonnegative matrix. Proof. We consider the list Λ′ = {−2 m Xj=1 aj, z1, z1, . . . , zm, zm}, with the partition Λ′ = Λ1 ∪ Λ2, where m Xj=1 Λ1 = {−2 Λ2 = {z21, z21, . . . , z2 m 2 aj, z11, z11, . . . , z1 m 2 } , z1 m 2 , z2 m Then 2 }. Mj=1(cid:20)a2j −b2j a2j (cid:21) is an m × m real matrix with spectrum Λ2, and aT A − JB = √2yT b2j m 2 √2x A + JB(cid:21) =(cid:20)0 (cid:20) c b A22(cid:21) ≥ 0, of order m + 1, has spectrum Λ1 and diagonal entries 0,−a21,−a21, . . . ,−a2 m 2 ,−a2 m 2 , 1 which there exists from Lemma 1.3. Then A = 1 2 J(A22 − (A − JB)), are both nonnegative and the matrix C′ =  b√2 A aT 0 √2 B J b√2 JBJ aT J √2 JAJ 2(A22 + A − JB) and B =   is centrosymmetric nonnegative with spectrum Λ′. Thus, C = C′+ Pn i=1 λi vT v vvT , where v is the Perron eigenvector of C′, is centrosymmetric nonnegative with spectrum Λ. More generally, if Λ has p real numbers, with odd p, we have the follow- ing result: 15 Corollary 3.2 Let Λ = {λ1, . . . , λp, z1, z1, . . . , zm, zm}, with zj, zj ∈ F , zj = aj + ibj, aj ∈ R, bj > 0, j = 1, . . . , m; λk ∈ F , k = 2, . . . , p, be a realizable list of complex numbers. Then Λ is the spectrum of a centrosym- metric nonnegative matrix. Proof. Case even m. We consider the list Λ′ = {− p Xk=2 λk − 2 m Xj=1 aj, λ2, . . . , λp, z1, z1, . . . , zm, zm}, and take the partition Λ′ = Λ1 ∪ Λ2 with, p Λ1 = {− Xk=2 Λ2 = {λ⌈ p λk − 2 n Xj=1 aj, λ2, . . . , λ⌈ p 2⌉, z11, z11, . . . , z1 m 2 2 } , z1 m 2⌉+1, . . . , λp, z21, z21, . . . , z2 m 2 , z2 m 2 }. Then A − JB = diag{λ⌈ p 2⌉+1, . . . , λp} m 2 a2j (cid:21) Mj=1(cid:20)a2j −b2j b2j is a real matrix with spectrum Λ2, and there exists a nonnegative matrix A + JB with spectrum Λ1 and diagonal entries 0,−λ⌈ p 2⌉+1, . . . ,−λp,−a21,−a21, . . . ,−a2 m 2 ,−a2 m 2 . Then the proof follows as before. Case odd m: We consider a partition Λ = Λ1 ∪ Λ2, in which the list Λ1 has the Perron eigenvalue and only one more element than the list Λ2. Both lists, Λ1 and Λ2 must be self-conjugated. Partitions of this type always exist if p ≥ 3. The construction of a nonnegative centrosymmetric matrix follows as in the proof of Lemma 3.3. Remark 3.1 According to the results in this section, it is clear that a real- izable list of real numbers of Suleimanova type is in particular realizable by a centrosymmetric nonnegative matrix. 16 We conclude this section by presenting a sufficient condition for the existence and construction of centrosymmetric nonnegative matrices with prescribed spectrum. Let Λ = {λ1, λ2, . . . , λn} be a realizable list of complex numbers. In order to construct a centrosymmetric nonnegative matrix A with spectrum Λ we consider the following partition of Λ : Λ = Λ0 ∪ Λ1 ∪ · · · ∪ Λ p0 2 ∪ Λ p0 2 ∪ · · · ∪ Λ1, for even p0, Λ = Λ0 ∪ Λ1 ∪ · · · ∪ Λ⌊ p0 2 ⌋ ∪ Λ⌈ p0 2 ⌉ ∪ Λ⌊ p0 2 ⌋ ∪ · · · ∪ Λ1, for odd p0, with Λ0 = {λ01, λ02, . . . , λ0p0}, λ01 = λ1, Λk = {λk1, λk2, . . . , λkpk}, k = 1, 2, . . . , p0 2 (⌈ p0 2 ⌉ for odd p0), (4)   where some of the lists Λk can be empty. For each list Λk, k 6= ⌈ p0 associate the list 2 ⌉, we Γk = {ωk, λk1, λk2, . . . , λkpk}, 0 ≤ ωk ≤ λ1, k = 1, 2, . . . , p0 2 ⌋ for odd p0), (5) which is realizable by a (pk+1)×(pk+1) nonnegative matrix Ak. In particular, Ak can be chosen as Ak ∈ CS ωk . To the sub-list Λ⌈ p0 2 ⌉ we associate the list Γ⌈ p0 2 ⌉, which is the spectrum of a centrosymmetric nonnegative matrix A⌈ p0 2 ⌉ of order p⌈ p0 2 ⌉+1. Then, the block diagonal matrices p0 2 (⌊ A = diag{A1, A2, . . . , A p0 A = diag{A1, A2, . . . , A⌊ p0 , JA p0 2 ⌋, A⌈ p0 2 2 J, . . . , JA2J, JA1J}, for even p0, and 2 ⌋J, . . . , JA2J, JA1J}, for odd p0, ) 2 ⌉, JA⌊ p0 (6) are centrosymmetric nonnegative with spectrum Γ1 ∪ Γ2 ∪ · · · ∪ Γ p0 Γ1 ∪ Γ2 ∪ · · · ∪ Γ⌊ p0 2 ⌋ ∪ Γ⌈ p0 2 ∪ Γ p0 2 ∪ · · · ∪ Γ2 ∪ Γ1, even p0 and 2 ⌋ ∪ · · · ∪ Γ2 ∪ Γ1, odd p0, 2 ⌉ ∪ Γ⌊ p0 respectively. Of course, the matrices J in (6) have the same order pk + 1 of the corresponding matrices Ak, k = 1, 2, , . . . ,⌊ p0 2 ⌋. Let X be an n × p0 matrix whose columns are nonnegative eigenvectors of 17 the matrix A in (6), that is, x1 . . . X =   x p0 2 x p0 2 . . . x1   , for even p0, (7) and X = x1 . . . x⌊ p0 2 ⌋ y⌈ p0 2 ⌉ x⌊ p0 2 ⌋ . . . x1     , for odd p0. (8) where, xk = eT = [1, 1, . . . , 1] is the Perron eigenvector of Ak in (6), k = 1, 2, . . . , p0 2 ⌉ = Jy⌈ p0 2 ⌉ ≥ 0 (symmetric vector). Observe that 2 ⌉ is the Perron eigenvector of A⌈ p0 2 ⌋), and y⌈ p0 2 ⌉, with y⌈ p0 2 (⌊ p0 if Ake = ωke then JAkJe = ωke. Now we state the main result of this section. First, we consider the even p0 case: Theorem 3.4 Let Λ = {λ1, λ2, . . . , λn} be a list of complex numbers with λi ≥ 0. Suppose there exists a Λ = Λ, λ1 ≥ λi , i = 2, 3, . . . , n, and partition of Λ (as defined in (4)), n Pi=1 2 ∪ · · · ∪ Λ1, with even p0, Λ = Λ0 ∪ Λ1 ∪ · · · ∪ Λ p0 Λ0 = {λ01, λ02, . . . , λ0p0}, λ01 = λ1 Λk = {λk1, λk2, . . . , λkpk}, k = 1, 2, . . . , 2 ∪ Λ p0 p0 2 , 18 where some of the lists Λk can be empty, such that the following conditions are satisfied: i) For each k = 1, 2, . . . , p0 2 , there exists a nonnegative matrix with spectrum Γk = {ωk, λk1, λk2, . . . , λkpk}, 0 ≤ ωk ≤ λ1, ii) There exists a centrosymmetric nonnegative matrix of order p0, with spec- trum Λ0 and diagonal entries ω1, ω2, . . . , ω p0 Then Λ is realizable by an n × n centrosymmetric matrix. Proof. From i) let Ak be a (pk + 1) × (pk + 1) nonnegative matrix with spectrum Γk, k = 1, 2, . . . , p0 2 . We may assume that Ak ∈ CS ωk. Then Ake = ωke, The matrix , . . . , ω2, ω1. , ω p0 2 2 A = A1 ⊕ A2 ⊕ · · · ⊕ A p0 2 ⊕ JA p0 2 J ⊕ · · · ⊕ JA2J ⊕ JA1J is centrosymmetric nonnegative with spectrum Γ = Γ1 ∪ · · · ∪ Γ p0 2 ∪ Γ p0 2 ∪ · · · ∪ Γ1. Let X be the matrix in (7). Then X is centrosymmetric nonnegative of zeros and ones. Now, from ii) and Theorem 1.3 let B = Ω+CX centrosymmetric nonnegative matrix of order p0 with spectrum Λ0 and diagonal entries ω1, . . . , ω p0 2 , ω p0 2 , . . . , ω1. b p0 2 2 · · · Jp0b1i is centrosymmetric 2 , are the columns of B − Ω. Jp0b1]. (9) · · · Jp0b p0 Then CX = B − Ω =hb1 nonnegative, where bk and Jp0bk, k = 1, . . . , p0 Therefore the p0 × n matrix C can be obtained as follows 2 · · · 0 . . . 0 {z } p1−times C = [b1 0 . . . 0 {z } p1−times {z } 0 . . . 0 p0 · · · b p0 0 . . . 0 p0 {z } 2 −times 2 −times Jp0b p0 2 It is easy to see that C is centrosymmetric nonnegative. Then the pertur- bation A + XC, with A, X,C as in (6),(7),(9) respectively, is nonnegative centrosymmetric and by Theorem 1.3, A + XC have the spectrum Λ. Now we consider the odd p0 case: 19 n Theorem 3.5 Let Λ = {λ1, λ2, . . . , λn} be a list of complex numbers with Λ = Λ, λ1 ≥ λi , i = 2, 3, . . . , n, and λi ≥ 0. Suppose there exists a partition of Λ (as defined in (4)), Pi=1 Λ = Λ0 ∪ Λ1 ∪ · · · ∪ Λ⌊ p0 2 ⌉ ∪ Λ⌊ p0 2 ⌋ ∪ · · · ∪ Λ1, with odd p0, Λ0 = {λ01, λ02, . . . , λ0p0}, λ01 = λ1 2(cid:25), Λk = {λk1, λk2, . . . , λkpk}, k = 1, 2, . . . ,(cid:24) p0 2 ⌋ ∪ Λ⌈ p0 where some of the lists Λk can be empty, such that the following conditions are satisfied: i) For each k = 1, 2, . . . ,⌊ p0 2 ⌋, there exists a nonnegative matrix with spectrum Γk = {ωk, λk1, λk2, . . . , λkpk}, 0 ≤ ωk ≤ λ1, while for k = ⌈ p0 2 ⌉ there exists a centrosymmetric nonnegative matrix with 2 ⌉. spectrum Γ⌈ p0 ii) There exists a centrosymmetric nonnegative matrix of order p0, with spec- trum Λ0 and diagonal entries ω1, ω2, . . . , ω⌊ p0 2 ⌉, ω⌊ p0 Then Λ is realizable by an n × n centrosymmetric matrix. 2 ⌉ is realizable by a centrosymmetric nonnegative Proof. Case 1. The list Γ⌈ p0 matrix of odd order: In this case the proof follows as the proof of Theorem 3.4, with a perturbation of the form A+ XC, with A, X as in (6), (8) respectively, and 2 ⌋, . . . , ω2, ω1. 2 ⌋, ω⌈ p0 · · · b⌊ p0 Jp0b⌊ p0 ⌈ 0 · · · 0 {z } p0 2 ⌉ 2 −times Jp0b1]. b⌈ p0 2 ⌉ 0 · · · 0 {z } p0 2 ⌉ 2 −times ⌈ C = [b1 0 · · · 0 {z } p1−times 0 · · · 0 {z } ⌊ p0 2 ⌋−times 2 ⌋ 0 · · · 0 {z } ⌊ p0 2 ⌋−times 2 ⌋ · · · 0 · · · 0 {z } p1−times Case 2. The list Γ⌈ p0 2 ⌉ is realizable by a centrosymmetric nonnegative matrix of order even: In this case we consider A as in (6), X as in (8) with all its columns normalized, and in ii) we take B = Ω + C. Then C = B − Ω, is centrosymmetric nonnegative, and by Lemma 1.1 and Theorem 1.3, the perturbation A+XCX T is centrosymmetric nonnegative with spectrum Λ. 20 To apply Theorem 3.4 and Theorem 3.5, we need to know conditions un- der which there exists a p0 × p0 centrosymmetric nonnegative matrix with spectrum Λ0 and diagonal entries ω1, ω2, . . . , ω p0 ω1, ω2, . . . , ω⌊ p0 , ω p0 2 ⌋, ω⌈ p0 , . . . , ω2, ω1, for even p0, or 2 ⌉, ω⌊ p0 2 ⌋, . . . , ω2, ω1 for odd p0. 2 2 In [8, Theorem 3.1] (see also [7]) it has been proved that the spectrum of a 3 × 3 nonnegative centrosymmetric matrix is real. Then, since bisymmet- ric matrices are centrosymmetric, we may use the necessary and sufficient conditions given in [3], for the existence of a 3 × 3 centrosymmetric nonneg- ative matrix with real eigenvalues λ1, λ2, λ3 and diagonal entries ω1, ω2, ω1. For n = 4 we have, in the Theorems 3.7 and 3.8 below, sufficient conditions for the existence and construction of a centrosymmetric nonnegative matrix with prescribed eigenvalues and diagonal entries. The following result show that for lists of four real numbers, both prob- lems, the nonnegative inverse eigenvalue problem and the centrosymmetric nonnegative inverse eigenvalue problem, are equivalent. Theorem 3.6 Let Λ = {λ1, λ2, λ3, λ4} be a list realizable of real numbers. Then Λ is in particular realizable by a centrosymmetric nonnegative matrix. Proof. Case 1: If λ1 ≥ λ2 ≥ λ3 ≥ λ4 ≥ 0, the result follows from Theorem 3.1 when n = 4. Case 2: If λ1 > 0 ≥ λ2 ≥ λ3 ≥ λ4, that is, if the list is of real Suleimanova type, then the result follows from Remark 3.1. Case 3: If λ1 ≥ λ2 ≥ λ3 ≥ 0 > λ4, we have the following general result: A list realizable of real numbers with only one negative eigenvalue is also real- izable by a centrosymmetric nonnegative matrix. In fact, let Λ = {λ1, . . . , λn}, with λ1 ≥ λ2 ≥ · · · ≥ λn−1 ≥ 0 > λn. If n = 2m is even we take the partition Λ = Λ1 ∪ Λ2, with Λ1 = {λ1, . . . , λm}, Λ2 = {λm+1, . . . , λn}, and we set A + JB = diag{λ1, λ2, . . . , λm}, A − JB = diag{λn, λm+1, . . . , λn−1}, 21 and the proof follows as before. If n = 2m + 1 is odd we take the partition Λ = Λ1 ∪ Λ2, with 2 ⌉+1 . . . , λn}. Λ1 = {λ1, . . . , λ⌈ n 2 ⌉}, Λ2 = {λ⌈ n and we set (cid:20)A + JB √2x √2yT c (cid:21) = diag{λ1, λ2, . . . , λ⌊ n 2 ⌉+1, . . . , λn−1}. A − JB = diag{λn, λ⌈ n 2 ⌋, λ⌈ n 2 ⌉}, Then, the proof follows as before. Case 4: If λ1 ≥ λ2 ≥ 0 > λ3 ≥ λ4, with λ2 + λ3 > 0; we take the partition Λ = Λ1 ∪ Λ2, with Λ1 = {λ1, λ2}, Λ2 = {λ3, λ4}, and we set A + JB = diag{λ1, λ2}, A − JB = diag{λ4, λ3}. Then, the proof follows as before. If λ1 ≥ λ2 ≥ 0 > λ3 ≥ λ4, with λ2 + λ3 < 0; we take the partition Λ1 = {λ1, λ2}, Λ2 = {λ3, λ4} and we set (−λ3)(λ1 + λ2 + λ3) − λ1λ2 λ1 + λ2 + λ3(cid:21) A − JB = diag{λ3, λ4}. A + JB =(cid:20) −λ3 1 Then A = B = 1 0 −(λ2 + λ3)(λ1 + λ3) λ1 + λ2 + λ3 + λ4(cid:21) , 2(cid:20) −(λ2 + λ3)(λ1 + λ3) λ1 + λ2 + λ3 − λ4(cid:21) J(cid:20) −2λ3 1 2 1 1 are both nonnegative matrices, and C =(cid:20)A JBJ B JAJ(cid:21) , is a 4 × 4 centrosymmetric nonnegative matrix. Now we give a sufficient condition for the existence and construction of a 4 × 4 centrosymmetric nonnegative matrix with prescribed eigenvalues and diagonal entries. We start with lists of real numbers. 22 2(λ1 + λ2 + λ3 + λ4), Theorem 3.7 Let Λ = {λ1, λ2, λ3, λ4} be a list of real numbers withP4 i=1 λi ≥ 0, λ1 ≥ λj, j = 2, 3, 4. Let {ω1, ω2, ω2, ω1} be a list of nonnegative numbers such that the following conditions are satisfied: i) 0 ≤ ωk ≤ λ1 k = 1, 2, ii) ω1 + ω2 = 1 iii) ωk ≥ λk+2 k = 1, 2, iv) (2ω1 − λ3)(2ω2 − λ4) ≥ λ1λ2. Then, there exists a centrosymmetric nonnegative matrix with espectrum Λ and diagonal entries ω1, ω2, ω2, ω1. Proof. Let Λ be particioned as Λ = Λ1 ∪ Λ2, with Λ1 = {λ1, λ2} and Λ2 = {λ3, λ4}.We set (2ω1 − λ3)(2ω2 − λ4) − λ1λ2 2ω2 − λ4(cid:21) , A−JB = diag{λ3, λ4}, A+JB =(cid:20) 2ω1 − λ3 1 with spectrum Λ1 and Λ2 respectively. Then ω1 2((2ω1 − λ3)(2ω2 − λ4) − λ1λ2) ω2(cid:21) , 2((2ω1 − λ3)(2ω2 − λ4) − λ1λ2) ω2 − λ4 1 2 1 A =(cid:20) B =(cid:20) 1 are both nonnegative matrices and (cid:21) 1 ω1 − λ3 B JAJ(cid:21) C =(cid:20)A JBJ is centrosymmetric nonnegative matrix with spectrum Λ and diagonal entries ω1, ω2, ω2, ω1. Note that in the above theorem, if Λ is a list of real numbers of Suleimanova type, then the conditions iii) and iv) are always satisfied. Then i) and ii) become necessary and sufficient conditions. In [8, Theorem 3.1] the authors show that λ1+λ2−2a ≥ 0 and λ1−λ2−2b ≥ 0 are necessary and sufficient conditions for the list {λ1, λ2, a + ib, a − ib}, with λ1, λ2, a ∈ R, b > 0 to be the spectrum of a normal centrosymmetric nonnegative matrix. Now, we give a sufficient condition for the existence and construction of a centrosymmetric nonnegative matrix with prescribed spectrum and diagonal entries. 23 Theorem 3.8 Let Λ = {λ1, λ2, a + ib, a − ib} be a list of complex numbers with λ1, λ2, a ∈ R, b > 0, λ1 + λ2 − 2a ≥ 0 and λ1 − λ2 − 2b ≥ 0. Let {ω1, ω2, ω2, ω1} be a list of nonnegative numbers, such that the following con- ditions are satisfied: i) 0 ≤ ωk ≤ λ1, k = 1, 2, ii) 1 iii) (2ω1 − a)(2ω2 − a) ≥ λ1λ2 + b2, iv) ωk ≥ a, k = 1, 2. Then, there exists a centrosymmetric nonnegative matrix with spectrum Λ and diagonal entries ω1, ω2, ω2, ω1. 2(λ1 + λ2 + 2a) = ω1 + ω2, Proof. Let Λ be particioned as Λ = Λ1 ∪ Λ2 with Λ1 = {λ1, λ2}, Λ2 = {a + ib, a − ib}. We set −b2 a(cid:21) , A + JB =(cid:20) (2ω1 − a)(2ω2 − a) − λ1λ2 2ω2 − a(cid:21) , A − JB =(cid:20) a 2ω1 − a 1 1 with spectrum Λ1 and Λ2, respectively. Then 1 ω1 A =(cid:20) B =(cid:20) 1 2 ((2ω1 − a)(2ω2 − a) − λ1λ2 − b2) ω2(cid:21) , (cid:21) , 2((2ω1 − a)(2ω2 − a) − λ1λ2 + b2) ω2 − a ω1 − a 1 0 are both nonnegative matrices, and C =(cid:20)A JBJ B JAJ(cid:21) , is nonnegative centrosymmetric matrix with spectrum Λ and diagonal entries ω1, ω2, ω2, ω1. Note that in the above theorem, if Λ is a list of complex numbers of Suleimanova type, then the conditions λ1 + λ2 − 2a ≥ 0 and λ1 − λ2 − 2b ≥ 0 are always satisfied. Hence, for n = 4, a list of complex numbers of Suleimanova type is always realizable by a centrosymmetric matrix. Moreover, conditions iii) and iv) are always satisfied for this kind of lists. Then, conditions i) and ii) become necessary and sufficient conditions. 24 4 Examples Example 4.1 Let Λ = {20,−1,−2,−3,−2 + 2i,−2− 2i,−3 + i,−3− i,−1 + i,−1 − i}. We shall use Corollary 3.1 to construct a centrosymmetric non- negative matrix with spectrum Λ. Then we consider the list Λ′ = {18,−1,−2,−3,−2 + 2i,−2 − 2i,−3 + i,−3 − i,−1 + i,−1 − i}, partitioned as Λ′ = Λ1 ∪ Λ2 with Λ1 = {18,−1,−2,−2 + 2i,−2 − 2i}, Λ2 = {−3,−3 + i,−3 − i,−1 + i,−1 − i}. Then A − JB =  −3  −3 −1 1 −3 −1 −1 1 −1   has spectrum Λ2. Now, from Lemma 1.3, we construct the nonnegative matrix A + JB = = (cid:2)3 3 3 1 1(cid:3) 1 1 1 1 1     + ,     0 1 2 2 2 1 0 2 2 2 2 1 0 2 2 4 1 2 0 0 0 1 2 4 0 3 4 5 3 3 4 3 5 3 3 5 4 3 3 3 7 4 5 1 1 3 4 5 5 1     with spectrum Λ1 and diagonal entries 3, 3, 3, 1, 1. Then A = 1 2   0 4 5 3 3 4 0 4 3 3 5 5 0 3 3 7 4 5 0 0 3 4 5 6 0   , B = 1 2   3 4 5 4 2 7 4 5 2 2 5 3 6 3 3 4 6 6 3 3 6 4 5 3 3   , 25 and  0 4 5 3 3 3 3 5 4 6 4 0 4 3 3 3 3 6 6 4 5 5 0 3 3 3 3 6 3 5 7 4 5 0 0 2 2 5 4 7 3 4 5 6 0 2 4 5 4 3 3 4 5 4 2 0 6 5 4 3 7 4 5 2 2 0 0 5 4 7 5 3 6 3 3 3 3 0 5 5 4 6 6 3 3 3 3 4 0 4 6 4 5 3 3 3 3 5 4 0 is centrosymmetric nonnegative with spectrum Λ′. Then  C′ = 1 2   C = C′ + 2 eT e eeT = 1 10   22 27 17 17 17 17 27 22 32 2 2 22 17 17 17 17 32 32 22 22 17 17 17 17 32 17 27 27 27 2 12 12 27 22 37 2 37 22 27 2 2 17 22 27 32 12 22 27 22 17 32 27 22 17 2 17 22 27 22 12 27 22 37 37 22 27 12 12 2 2 27 27 27 17 32 17 17 17 17 2 22 32 32 17 17 17 17 22 2 22 2 32 22 27 17 17 17 17 27 22   is centrosymmetric nonnegative with spectrum Λ. Example 4.2 Let Λ = {10, 3, 1 + i, 1− i,−2 + 2i,−2− 2i,−2 + 2i,−2− 2i}. Now we use Theorem 3.4 and Theorem 3.8 to construct a centrosymmetric nonnegative matrix with spectrum Λ. We take the partition Λ = Λ0 ∪ Λ1 ∪ Λ2 ∪ Λ2 ∪ Λ1, with Λ0 = {10, 3, 1 + i, 1 − i}, Λ1 = {−2 + 2i,−2 − 2i}, Λ2 = ∅, and consider the associated realizable lists Γ1 = {4,−2 + 2i,−2 − 2i}, Γ2 = { 7 2}, with realizing matrices A1 =  0 0 4 2 0 2 0 4 0   , A2 =(cid:2) 7 2(cid:3) = JA2J, JA1J =  0 4 0 2 0 2 4 0 0   . 26   and X = 7 2 7 2 0 4 0 2 0 2 4 0 0 1 1 1 1 1 1 1 1   .   From Theorem 3.8 we compute the centrosymmetric nonnegative matrix with spectrum Λ0 and diagonal entries {4, 7 2 , 4} 2 , 7 Then A = 0 0 4 2 0 2 0 4 0   Then B =  4 11 2 13 2 3 1 0 5 7 2 2 7 5 2 2 0 1 . 3 13 2 11 2 4   C =  0 11 2 13 2 3 0 0 1 0 0 0 0 0 0 5 0 0 5 0 0 0 1 0 0 3 2 0 0 13 2 0 0 0 11 2 0 2   is centrosymmetric nonnegative. Therefore by Theorem 3.4 0 2 0 11 2 13 2 3 3 3   A + XC = References 0 4 1 0 0 0 0 2 1 0 0 0 4 0 1 0 0 0 0 0 7 0 0 5 0 0 0 1 0 4 0 0 0 1 2 0 0 0 0 1 4 0 3 3 3 2 0 0 13 2 0 0 11 2 0 2 0 7 2 5 2 2 is the desired matrix.   [1] A. Brauer, Limits for the characteristic roots of a matrix IV. Applica- tions to stochastic matrices, Duke Math. J. 19 (1952) 75-91. 27 [2] A. Cantoni, P. Butler, Eigenvalues and eigenvectors of symmetric cen- trosymmetric matrices, Linear Algebra Appl. 13 (1976) 275-288. [3] A. I. Julio, R. L. Soto. Persymmetric and bisymmetric nonnegative in- verse eigenvalue problem, Linear Algebra Appl. 469 (2015) 130-152. [4] A. I. Julio, C. Mariju´an, M. pisonero, R. L. Soto, On universal realiz- ability of spectra, Linear Algebra Appl. 563 (2019) 353-372. [5] H. Perfect, Methods of constructing certain stochastic matrices. Duke Math. J. 20 (1953) 395-404. [6] H. Perfect, Methods of constructing certain stochastic matrices II, Duke Math. J. 22 (1955) 305-311. [7] S. Somphotphisut, K. Wiboonton, On the bisymmetric nonnegative in- verse eigenvalue problem, Linear Algebra Appl. 543 (2018) 42-71. [8] S. Somphotphisut, K. Wiboonton On the normal centrosymmetric non- negative inverse eigenvalue problem, arXiv: 1710.08757v1. [9] R. L. Soto, O. Rojo, C. B. Manzaneda, On nonnegative realization of partitioned spectra, Electron.J. Linear Algebra 22 (2011) 557-572. [10] R. L. Soto, A. I. Julio, M. Collao, Brauer's theorem and nonnegative matrices with prescribed diagonal entries, submited (2018). [11] H. R. Suleimanova, Stochastic matrices with real characteristic values, Dokl. Akad. Nauk SSSR 66 (1949) 343-345. 28
1210.3520	1	1210	2012-10-12T14:07:27	Distributive lattices determined by weighted double skeletons	[ "math.RA" ]	Related to his S-glued sum construction, the skeleton S(L) of a finite lattice L was introduced by C. Herrmann in 1973. Our theorem asserts that if D is a finite distributive lattice and its second skeleton, S(S(D)), is the trivial lattice, then D is characterized by its weighted double skeleton, introduced by the second author in 2006. The assumption on the second skeleton is essential.	math.RA	math	DISTRIBUTIVE LATTICES DETERMINED BY WEIGHTED DOUBLE SKELETONS G ´ABOR CZ´EDLI, JOANNA GRYGIEL, AND KATARZYNA GRYGIEL Abstract. Related to his S-glued sum construction, the skeleton S(L) of a finite lattice L was introduced by C. Herrmann in 1973. Our theorem asserts that if D is a finite distributive lattice and its second skeleton, S(S(D)), is the trivial lattice, then D is characterized by its weighted double skeleton, introduced by the second author in 2006. The assumption on the second skeleton is essential. ] . A R h t a m [ 1 v 0 2 5 3 . 0 1 2 1 : v i X r a 1. Introduction Let L be a finite modular lattice. Then, according to Herrmann [11], L is the union of its maximal complemented (equivalently, atomistic) intervals, which are glued together along a lattice S = S(L), the skeleton of L. His construction of S(L) makes sense even without modularity, so we drop this assumption until otherwise stated. It appeared somewhat later that S(L) is a factor lattice of L by a tolerance relation in the sense of the first author [3]. Define S0(L) := L and Si+1(L) := S(cid:0)Si(L)(cid:1). Then there is a smallest n such that Sn(L) = 1, which we will call the Herrmann rank of L. We say that L is H n-irreducible if its Herrmann rank is at most n. Equivalently, L is H n-irreducible iff Sn(L) = 1. H 1-irreducibility was previously called H-irreducibility by the second author in the monograph [9] and in many of her papers, including [9] and [10]. The skeleton of L does not tell too much on L. Indeed, the second author [9, Corollary 3.2.6] proved that each finite lattice S is the skeleton of infinitely many pairwise non-isomorphic finite distributive lattices. The weighted double skeleton S wd(L) of L, introduced by the second author in [10] and to be defined in the present paper soon, carries much more information on the initial lattice. Let K be a class of finite distributive lattices, and let L ∈ K. If for any L′ ∈ K such that S wd(L′) is isomorphic to S wd(L) the lattice L′ is isomorphic to L, then we say that L is determined by its weighted double skeleton in the class K. As usual, the partially ordered set (in short, the poset, in other words, the order) of all non-zero join-irreducible elements of L is denoted by Ji(D) = (Ji(D), ≤). The sets {1, 2, 3, . . . } and {0, 1, 2, 3, . . . } are denoted by N and N0, respectively. The Date: October 27, 2011; revised August 21, 2012; September 5, 2012; October 12, 2012. 2000 Mathematics Subject Classification. Primary 06B99, Secondary 06C15, 06D05. Key words and phrases. S-glued sum, skeleton of a lattice, double skeleton, weighted double skeleton, glued tolerance, lattice tolerance, Herrmann rank. This research of the first author was supported by the NFSR of Hungary (OTKA), grant numbers K77432 and K83219. The second author was supported by the NSC of Poland, grant number 2011/01/B/HS1/00944. The third author was supported by the NSC of Poland, grant number 2011/01/B/HS1/00944, and by the Polish Ministry of Science and Higher Education, grant number NN206 376137. 1 2 G. CZ´EDLI, J. GRYGIEL, AND K. GRYGIEL length of a finite poset Q is length(Q) := max{n ∈ N0 : Q has an n + 1-element chain}. A nontrivial lattice is a lattice that has at least two elements. Postponing the rest of definitions to the next section, our main result reads as follows. Theorem 1.1. Let L be a finite nontrivial lattice. (i) If L is modular and H 2-irreducible, then length(Ji(L)) ≤ 1. (ii) If L is distributive and H 2-irreducible, then L is determined by its weighted double skeleton in the class of finite distributive lattices. (iii) If L is distributive and length(Ji(L)) ≤ 1, then L is determined by its weighted double skeleton in the class of finite distributive lattices D satisfying the in- equality length(Ji(D)) ≤ 1. Notice that S wd(L) determines some properties of a modular L even if H 2- irreducibility is not assumed. Namely, S wd(L) clearly determines length(L), and see Lemma 3.1 for further properties in the distributive case. However, we will soon prove the following remark, which indicates that Theorem 1.1 is optimal in some sense. Remark 1.2. There exist H 3-irreducible finite distributive lattices L1 and L2 such that S wd(L1) is isomorphic to S wd(L2) but L1 is not isomorphic to L2. Also, there is a finite distributive lattice L such that length(Ji(L)) ≤ 1 but L is not H 2-irreducible. A well-known economic way of describing a finite distributive lattice D by a little amount of data is to consider Ji(D) = (Ji(D), ≤). The next remark outlines a more economic way for certain distributive lattices. Remark 1.3. Let D be a finite distributive lattice with length(Ji(D)) ≤ 1. Assume that D is the union of few maximal boolean intervals but D is large. Then S wd(D) constitutes an economic description of D. 2. Basic concepts and statements For the basic concepts of Lattice Theory the reader is referred to Gratzer [7]. By a tolerance of a lattice L we mean a reflexive, symmetric, compatible relation of L. Equivalently, a tolerance of L is the image of a congruence by a surjective lattice homomorphism onto L, see the first author and Gratzer [4]. Let R be a tolerance of L. If X ⊆ L is a maximal subset with respect to the property X × X ⊆ R, then X is called a block of R. Blocks are convex sublattices by Bandelt [1] and Chajda [2]. Let α and β be blocks of R. As it follows immediately from Zorn's Lemma, there are blocks γ and δ of R such that (1) {x ∨ y : x ∈ α, y ∈ β} ⊆ γ =: α ∨ β, {x ∧ y : x ∈ α, y ∈ β} ⊆ δ =: α ∧ β. The first author [3] proved that γ and δ are uniquely determined, and the set L/R of all blocks of R with the join and meet defined by (1) is a lattice. This lattice, also denoted by L/R, is called the factor lattice (or quotient lattice) of L modulo R. Notice that there is an alternative way, which does not rely on the axiom of choice (and, therefore, on Zorn's Lemma), to define L/R in an order-theoretic way and to prove that it is a lattice, see Gratzer and Wenzel [8]. In the rest of the paper, all lattices will be assumed to be finite. Then the blocks of a tolerance R are intervals. So if α is a block of R, then α equals the interval DISTRIBUTIVE LATTICES DETERMINED BY WEIGHTED DOUBLE SKELETONS 3 Figure 1. Non-isomorphic lattices with isomorphic double skeletons [0α, 1α] of L. It was proved in [3] and [6] that, for all α, β ∈ L/R, (2) 0α ∨ 0β = 0α∨β, 1α ∧ 1β = 1α∧β, α ≤ β (in L/R) ⇐⇒ 0α ≤ 0β ⇐⇒ 1α ≤ 1β. 1α ∨ 1β ≤ 1α∨β, 0α ∧ 0β ≥ 0α∧β, The most important particular case of L/R, under the name skeleton, was discov- ered by Herrmann [11] much earlier; we survey it partly and only for the finite case. A tolerance R of (a finite lattice) L is called a glued tolerance, see Reuter [12], if its transitive closure R∗ is the total relation L2. The (unique) smallest glued tolerance of L is called the skeleton tolerance of L, and it is denoted by Θ(L). There are two easy ways to see that Θ(L) exists. Firstly, we know from the second author [9], and it is routine to check, that for any tolerance R of a finite lattice L, (3) R is a glued tolerance ⇐⇒ (x, y) ∈ R for all x ≺ y ∈ L. This clearly implies that the intersection of all glued tolerances of L is a glued tolerance again, whence it is the skeleton tolerance of L. Secondly, we know from [5] that the transitive closure of lattice tolerances commutes with their (finitary) intersections, which also implies the existence of Θ(L). The factor lattice L/Θ(L) is called the skeleton S(L) of L. We claim that length(S(L)) < length(L). Indeed, assume that k = length(S(L)) and α0 ≺ α1 ≺ · · · ≺ αk is a maximal chain in S(L) = L/Θ(L). It follows from (2) and (3) that 0α0 < 0α1 < · · · < 0αk < 1αk , showing that length(S(L)) = k < k + 1 ≤ length(L). The inequality length(S(L)) < length(L) shows that each finite lattice has a Herrmann rank. It is clear from (2) that both {0α : α ∈ S(L)} and {1α : α ∈ S(L)}, as sub- posets of L, are order isomorphic to S(L). Their union carries a lot of information on L provided we equip it with an appropriate structure. Following the second author [10], a structure (P, ≤, K, η0, η1, w) will be called an abstract weighted double skeleton if (P, ≤) is a finite poset, K is a lattice, η0 : K → P is a join-preserving (and, therefore, order-preserving) embedding, η1 : K → P is a meet-preserving order-embedding, P = η0(K) ∪ η1(K), η0(x) ≤ η1(x) holds for all x ∈ K, and w is a mapping of the covering relation {(a, b) ∈ P 2 : a ≺ b} into N. The underlying set of (P, ≤, K, η0, η1, w) is P , and we often denote the structure (P, ≤, K, η0, η1, w) simply by P . 4 G. CZ´EDLI, J. GRYGIEL, AND K. GRYGIEL Figure 2. length(Ji(L)) ≤ 1 does not imply H 2-irreducibility Let (P ′, ≤′, K ′, η′ 0, η′ 1, w′) be another abstract weighted double skeleton, and let (ψ, κ) be a pair of bijective mappings. We say that (ψ, κ) : (P, ≤, K, η0, η1, w) → (P ′, ≤′, K ′, η′ 0, η′ 1, w′) is an isomorphism if ψ : (P, ≤) → (P ′, ≤′) is an order isomorphism, κ : K → K ′ is a lattice isomorphism, ψ(cid:0)ηi(x)(cid:1) = η′ i(cid:0)κ(x)(cid:1) for all x ∈ K and i ∈ {0, 1}, and w(x, y) = w′(ψ(x), ψ(x)) for all x, y ∈ P such that x ≺ y. If there is such a (ψ, κ), then the two abstract weighted double skeletons are called isomorphic. By the (concrete) weighted double skeleton of L we mean the structure (4) S wd(L) := (cid:0){0α : α ∈ S(L)} ∪ {1α : α ∈ S(L)}, ≤, S(L), η0, η1, w(cid:1) where ≤ is the ordering inherited (restricted) from L, η0(α) := 0α and η1(α) := 1α for all α ∈ S(L), and w(x, y) := length([x, y]L) for any x ≺Swd(L) y. For example, consider Lj given in Figure 1 for j ∈ {1, 2}. Then S(Lj) is the three-element chain {ζ0 ≺ α ≺ β}, and S wd(Lj) is depicted in the middle of the figure. For x ≺ y in S wd(Lj), the edge x ≺ y of the diagram is labeled by w(x, y). Since L1 6∼= L2 and Sn(Lj) is the (4 − n)-element chain for n ∈ {1, 2, 3} and j ∈ {1, 2}, Figure 1 together with the self-explanatory Figure 2 proves Remark 1.2. We have defined all the concepts Theorem 1.1 is based on. The rest of the paper is devoted to proofs, including some auxiliary statements. 3. Proofs and auxiliary statements 3.1. The number of join-irreducible elements in a block. Unless otherwise stated, by a block of a lattice L we mean a block of its skeleton tolerance Θ(L), that is, a member of the skeleton S(L). Throughout this subsection, L denotes a finite modular lattice. We are going to extend the weight function w of S wd(L), see (4), to a function w∗ : S wd(L) × S wd(L) → N0. If x, y ∈ S wd(L) and x 6< y, then we let w∗(x, y) := 0. If x, y ∈ S wd(L) and x < y, then take a maximal chain z0 = x ≺Swd(L) z1 ≺Swd(L) · · · ≺Swd(L) zt = y in S wd(L), and define w∗(x, y) := t−1 X i=0 w(zi, zi+1). The lattice theoretical Jordan-Holder theorem applies on L and we conclude that w(x, y) := length([x, y]L) for any x ≤Swd(L) y. This guarantees that w∗(x, y) does not depend on the maximal chain chosen. Given a poset (Q, ≤), the Mobius function DISTRIBUTIVE LATTICES DETERMINED BY WEIGHTED DOUBLE SKELETONS 5 µQ : Q2 → Z is defined recursively as follows: µQ(x, y) = 1, − X 0, x≤z<y µQ(x, z), if x = y; if x < y; if x 6≤ y.   If L is a finite modular lattice, then every α ∈ S(L) is an atomistic lattice by Herrmann [11]. In this case, Ji(α) stands for the set of join-irreducible elements of α that are distinct from 0α. The set of atoms of α is denoted by At(α). We let (5) Jα := Ji(L) ∩ Ji(α) = Ji(L) ∩ At(α) = Ji(L) ∩ (α \ {0α}). Since S wd(L) determines S(L), the next lemma, based on the notation above, im- plies that Jα is determined by S wd(L), provided L is distributive. Lemma 3.1. Let L be a finite distributive lattice. Then for each α ∈ S(L), (6) Jα = X β≤α, β∈S(L) µS(L)(β, α) · w∗(0α, 1β). Proof. For k ∈ N and a poset K, let Jik(K) denote the set of elements of K with exactly k lower covers. Notice that Ji(L) = Ji1(L). Reuter [12, Corollary 3] asserts, even when L is only modular, that for each α ∈ S(L) and k ∈ N, (7) Jik(L) ∩ Jik(α) = X β≤α, β∈S(L) µS(L)(β, α) · Jik(α ∩ β). For k = 1, the lefthand side of (6) equals that of (7). Hence it suffices to show that w∗(0α, 1β) = Ji1(α ∩ β) holds for β ≤ α. This is obvious if α ∩ β = ∅ since then 0α 6≤ 1β follows from (2). Hence we assume that α ∩ β 6= ∅. Then, again by (2), α ∩ β = [0α, 1β] is a principal ideal of α, whence α ∩ β is a boolean sublattice of the boolean interval α. Thus, Ji1(α ∩ β) = Ji([0α, 1β]) = At([0α, 1β]) = length([0α, 1β]) = w∗(0α, 1β). (cid:3) 3.2. More about blocks. Although the following lemma requires a proof in the present setting, it is a part of the original definition of S(L) given by Herrmann [11]. For the reader's convenience, we present an easy proof. Lemma 3.2 (Wille [13, Proposition 9]). Let L be a finite lattice. If α ≺S(L) β, then α ∩ β 6= ∅. Proof. Assume that α ≺S(L) β, and let a := 1α. Since 1α < 1β by (2), we can take an element b ∈ L such that a = 1α ≺ b ≤ 1β. Since L = Sν∈S(L) ν, there is a γ1 ∈ S(L) containing b. By (1), γ2 := α ∨ γ1 contains b = a ∨ b, and γ3 := β ∧ γ2 also contains b = 1β ∧ b. Clearly, α ≤ γ3 ≤ β. Taking b ∈ γ3 \ α and α ≺S(L) β into account, we conclude that b ∈ γ3 = β. Next, {a, b}2 ⊆ Θ(L) since a ≺ b. Hence there is a block γ ∈ S(L) such that {a, b} ⊆ γ. Using (1) repeatedly, we obtain that a = a ∨ a ∈ α ∨ γ, b = a ∨ b ∈ α ∨ γ, a = a ∧ b ∈ (α ∨ γ) ∧ β and b = b ∧ b ∈ (α ∨ γ) ∧ β. That is, {a, b} ⊆ (α ∨ γ) ∧ β. On the other hand, α ≤ (α ∨ γ) ∧ β ≤ β together with α ≺S(L) β yields that α = (α ∨ γ) ∧ β or β = (α ∨ γ) ∧ β. Hence {a, b} ⊆ α or {a, b} ⊆ β, and we conclude that α ∩ β 6= ∅. (cid:3) 6 G. CZ´EDLI, J. GRYGIEL, AND K. GRYGIEL For the reader's convenience, we also prove the following lemma. Due to the forthcoming formula (8), the present approach is slightly simpler than the original one of the second author [9]. Lemma 3.3 ([9, Theorem 2.2.10]). Let L be a finite lattice. Assume that α, β ∈ S(L) such that (α, β) ∈ Θ(S(L)). Then α ∩ β 6= ∅. Proof. A straightforward induction based on (1) shows that, for any lattice term p(x1, . . . , xn) and for any ν1, . . . , νn ∈ S := S(L), (8) {p(x1, . . . , xn) : x1 ∈ ν1, . . . , xn ∈ νn} ⊆ p(ν1, . . . , νn). Next, let S = {σ1, . . . , σt}, and let κ1 ≺S λ1, . . ., κk ≺S λk be a list of all covering pairs of S. Since the skeleton tolerance Θ(S) of S is generated by {(κ1, λ1), . . . , (κk, λk)}, it coincides with the subalgebra of S2 generated by {(κ1, λ1), . . . , (κk, λk), (λ1, κ1), . . . (λk, κk), (σ1, σ1), . . . , (σt, σt)}. Hence there exists a (2k + t)-ary lattice term p such that (α, β) = p(cid:0)(κ1, λ1), . . . , (κk, λk), (λ1, κ1), . . . , (λk, κk), (σ1, σ1), . . . , (σt, σt)(cid:1) = (cid:0)p(κ1, . . . , κk, λ1, . . . , λk, σ1, . . . , σt), p(λ1, . . . , λk, κ1, . . . , κk, σ1, . . . , σt)(cid:1). It follows from Lemma 3.2 that there are x1, . . . , xk, y1, . . . , yt ∈ L such that xi ∈ κi ∩ λi = λi ∩ κi for i = 1, . . . , k and yj ∈ σj for j = 1, . . . , t. Hence the above expression for (α, β) together with (8) yields that p(x1, . . . , xk, x1, . . . , xk, y1, . . . , yt) ∈ α ∩ β. (cid:3) 3.3. More about join-irreducible elements in blocks. Notation 3.4. Let Ji+(L) := Ji(L) \ At(L). Let ζ0 = [0, z0] be the least element of S(L), and let S+(L) := S(L) \ {ζ0}. For x ∈ L, domin(x) := {y ∈ At(L) : y ≤ x} is called the set of atoms dominated by x. Similarly, for α ∈ S+(L), domin(α) := {y ∈ At(L) : y ≤ 0α} is the set of atoms dominated by α. The next lemma is easy. Having no reference at hand, we will give a proof. Lemma 3.5. Let L be a finite modular lattice. Then [0,Wx∈At(L) x] = ζ0. Further- more, if α ∈ S(L) such that Jα ∩ At(L) 6= ∅, then α = ζ0. Proof. Let z0 := Wx∈At(L) x. Since (0, x) ∈ Θ(L) for all x ∈ At(L), we obtain that (0, z0) ∈ Θ(L). Hence we can extend {0, z0} to a block α = [0, y] of Θ(L). Obviously, z0 ≤ y. We know from Herrmann [11] that α is an atomistic lattice. Hence y = Wx∈At(α) x ≤ Wx∈At(L) x = z0, and we conclude that [0, z0] = [0, y] = α ∈ S(L). It is the smallest element of S(L), that is ζ0, by (2). Finally, if an atom a ∈ At(L) belongs to Jα, then 0α ≺ a implies 0α = 0 = 0ζ0, whence α = ζ0 by (2). (cid:3) Lemma 3.6. Let L be a finite modular lattice. Then (i) Ji(L) = S{Jα : α ∈ S(L)}; (ii) for all α, β ∈ S(L), if α 6= β, then Jα ∩ Jβ = ∅; (iii) Ji+(L) = S{Jα : α ∈ S+(L)} and At(L) = Jζ0 = At(ζ0). DISTRIBUTIVE LATTICES DETERMINED BY WEIGHTED DOUBLE SKELETONS 7 Proof. Assume that a ∈ Ji(L), and let a− stand for its unique lower cover. Then {a−, a} ⊆ Θ(L) since Θ(L) is a glued tolerance. We can extend {a−, a} to a block α ∈ S(L). Then a ∈ Ji(L) ∩ (α \ {0α}) = Jα. This proves that Ji(L) ⊆ S{Jα : α ∈ S(L)}. The reverse inclusion in part (i) is trivial. Assume that x ∈ Jα ∩ Jβ. Then, by (5), x ≻ 0α and x ≻ 0β. Hence x ∈ Ji(L) implies that 0α = 0β. This together with (2) yields α = β, which proves part (ii). Finally, parts (i) and (ii) together with Lemma 3.5 imply part (iii). (cid:3) Lemma 3.7. Let α, β and ζ be distinct blocks of a finite modular lattice L such that ζ < α, ζ < β and {α, β, ζ}2 ⊆ Θ(cid:0)S(L)(cid:1). Then a k b holds for all a ∈ Jα and b ∈ Jβ. Proof. Assume that a is comparable with b. It follows from Lemma 3.6(ii) that a 6= b. Hence we can assume that a < b. We infer α ∩ ζ 6= ∅ and β ∩ ζ 6= ∅ by Lemma 3.3. This together with ζ < α and ζ < β yields that 0ζ < 0α ≤ 1ζ and 0ζ < 0β ≤ 1ζ. Since a ∈ Jα ⊆ At(α) by (5), we have that 0α ≺ a. Similarly, 0β ≺ b. Since b ∈ Jβ ⊆ Ji(L), 0β is the only lower cover of b. This together with a < b implies that a ≤ 0β. Therefore, 0ζ < 0α ≺ a ≤ 0β ≤ 1ζ. This means that a ∈ ζ but, since Ji(ζ) = At(ζ), a is join-reducible in ζ, whence it is also join-reducible in L. This contradicts a ∈ Jα ⊆ Ji(L). (cid:3) Proof of Theorem 1.1(i). Let us assume for a contradiction that there exist a1, a2, a3 ∈ Ji(L) such that a1 < a2 < a3. By Lemma 3.6(i), we can choose α1, α2, α3 ∈ S(L) such that ai ∈ Jαi , for i ∈ {1, 2, 3}. Since Jαi ⊆ At(αi) is an antichain for i = 1, 2, 3, the blocks α1, α2, α3 are pairwise distinct. Hence at least two of them, say αj and αk, are distinct from the smallest element ζ0 of S(L). Since S2(L) = S(L)/Θ(cid:0)S(L)(cid:1) is the singleton lattice, Θ(S(L)) is the full relation on S(L). Therefore, applying Lemma 3.7 to αj, αk and ζ0, we obtain that aj k ak, a contradiction. (cid:3) Clearly, Ji+(L) 6= ∅ iff length(Ji(L)) ≥ 1. Hence the next lemma would (vacu- ously) also hold if length(Ji(L)) was 0. Notation 3.4 will be in effect. Lemma 3.8. Let L be a finite distributive lattice such that length(Ji(L)) = 1. Let α, α1, . . . , αn ∈ S+(L) such that none of Jα, Jα1 , . . . , Jαn is empty. Then (i) domin(α) 6= ∅; (ii) α1 ≤ α2 iff domin(α1) ⊆ domin(α2); (iii) domin(α1) ∪ . . . ∪ domin(αn) = w∗(0, 0α1∨···∨αn ). Proof. Since 0 6= 0α, part (i) is trivial. If α1 ≤ α2, then 0α1 ≤ 0α2 by (2), whence domin(α1) ⊆ domin(α2). To prove the reverse implication of (ii), assume that α1 6≤ α2. Then (2) implies 0α1 6≤ 0α2, which yields an x ∈ Ji(L) such that x ≤ 0α1 but x 6≤ 0α2. Since Jα1 6= ∅ by the assumption, there is a y ∈ Ji(L) such that x ≤ 0α1 < y. This together with length(Ji(L)) = 1 shows that x ∈ At(L). Hence x ∈ domin(α1) \ domin(α2), proving part (ii). Next, we claim that (9) domin(α1) ∪ . . . ∪ domin(αn) = {x ∈ At(L) : x ≤ 0α1∨···∨αn}. The "⊆" inclusion is an evident consequence of (2). To prove the converse inclusion, assume that x belongs to the righthand side of (9). Then, by (2) and distributivity, x = x ∧ 0α1∨···∨αn = x ∧ (0α1 ∨ · · · ∨ 0αn ) = (x ∧ 0α1) ∨ · · · ∨ (x ∧ 0αn). 8 G. CZ´EDLI, J. GRYGIEL, AND K. GRYGIEL By the join-irreducibility of x, there exists an i ∈ {1, . . . , n} such that x = x ∧ 0αi. Hence x ≤ 0αi implies that x ∈ domin(αi), proving (9). For i ∈ {1, . . . , n}, 0αi is the join of some (possibly only one) join-irreducible elements of L. These elements are necessarily atoms since length(Ji(L)) = 1 and Jαi 6= ∅. Therefore all the 0αi belong to ζ0, and so does their join, which is 0α1∨···∨αn by (2). Since ζ0 is a boolean lattice, the number of atoms below 0α1∨···∨αn , that is the size of the set given in (9), is length([0, 0α1∨···∨αn ]) = w∗(0, 0α1∨···∨αn ). This together with (9) proves part (iii). (cid:3) 3.4. A lemma on bipartite graphs. In order to formulate a statement that we need in the proof of Theorem 1.1(ii)-(iii), we have to associate a number-valued function with bipartite graphs. By a finite directed bipartite graph we shall mean a structure G = (U, X, E) where U and X are finite nonempty sets, referred to as upper and lower vertex sets, and E ⊆ U × X is an arbitrary relation. The power set, that is the set of all subsets, of U is denoted by P (U ), and P (X) has the analogous meaning. Let P ∞(U ) := P (U ) ∪ {∞} where ∞ is a symbol not in P (U ). For V ∈ P (U ), we let dG(V ) := {x ∈ X : there is a v ∈ V such that (v, x) ∈ E}. This set is called the set of (lower) vertices dominated by V . (We shall not use the word "covered" in this context since we want to avoid any confusion with the order- theoretic covering relation.) We define dG(∞) := X. Let d♯ G(V ) stand for dG(V ); if V = {v}, then we write d♯ G, called the domination function associated with G, is a P ∞(U ) → N0 mapping. If ϕ : U → U ′ is a bijection and V ∈ P ∞(U ), then ϕ(V ) := {ϕ(v) : v ∈ V } for V ∈ P (U ) while ϕ(∞) := ∞ ∈ P ∞(U ′). G({v}). This way d♯ G(v) rather than d♯ Lemma 3.9. Let G = (U, X, E) and G′ = (U ′, X ′, E′) be finite directed bipartite graphs. Then these two graphs are isomorphic iff there is a bijection ϕ : U → U ′ that preserves the domination function, that is, d♯ G(V ) holds for all V ∈ P ∞(U ). G′(ϕ(V )) = d♯ Proof. In order to prove the non-trivial direction of the lemma, assume that ϕ : U → U ′ is a bijection that preserves the domination function. We associate two addi- tional mappings with G as follows: sG : P (U ) → P (X), eG : P (U ) → P (X), V 7→ {x ∈ X : (v, x) ∈ E for all v ∈ V }, V 7→ {x ∈ X : (v, x) ∈ E for all v ∈ V and (u, x) /∈ E for all u ∈ U \ V }. The corresponding number-valued functions are denoted by s♯ G, that is, s♯ G(V ) := sG(V ) and e♯ G(V ) := eG(V ). These functions will be called the strong domination function and the exact domination function, respectively. Replacing G by G′, we obtain the definition for s♯ G′. Usually, we will elaborate our formulas only for G since, sometimes implicitly, we will rely on the fact that the analogous formulas hold for G′ as well. G′ and e♯ G and e♯ G′(ϕ(V )) = s♯ Firstly, we prove that ϕ preserves the strong domination function. Let V ∈ P (U ); we show s♯ G(∅) = X = d♯ G(∞), and the desired equality follows easily. The case V = 1, say V = {v}, is even easier since s♯ G(v). Next, assume that 1 < n ∈ N, V = {v1, . . . , vn} is an n-element subset of U , and the desired equality holds for all subsets of U G(V ) by induction on V . If V = 0, then s♯ G(v) = d♯ DISTRIBUTIVE LATTICES DETERMINED BY WEIGHTED DOUBLE SKELETONS 9 with less than n elements. Based on the inclusion-exclusion principle, also called (logical) sieve formula, T1 ∪ · · · ∪ Tn = n X i=1 (−1)i−1 X I⊆{1,...,n} I=i \ j∈I (cid:12)(cid:12)(cid:12) , Tj)(cid:12)(cid:12)(cid:12) and using that sG agrees with dG on singleton sets and satisfies the identity sG(V1)∩ sG(V2) = sG(V1 ∪ V2), we can compute as follows. d♯ G(V ) = dG(V ) = dG(v1) ∪ · · · ∪ dG(vn) = sG(v1) ∪ · · · ∪ sG(vn) = (−1)n−1(cid:12)(cid:12)(cid:12) n \ j=1 sG(vj )(cid:12)(cid:12)(cid:12) n−1 + X i=1 (−1)i−1 X I⊆{1,...,n} I=i \ j∈I (cid:12)(cid:12)(cid:12) sG(vj)(cid:12)(cid:12)(cid:12) = (−1)n−1s♯ G(V ) + n−1 X i=1 (−1)i−1 X I⊆{1,...,n} I=i s♯ G({vj : j ∈ I}). Now, ϕ preserves all summands but (−1)n−1s♯ G(V ) in the previous line by the induction hypothesis. Since d♯ G(V ) is pre- served either, completing the induction. Thus, ϕ preserves the strong domination function. G(V ) is also preserved, we conclude that s♯ G′(ϕ(V )) = e♯ Next, to show that ϕ preserves the exact domination function, let V ∈ P (U ). We want to show that e♯ G(U ) = s♯ G(U ) and the strong domination function is preserved. Hence we can assume that V 6= U . Let W := U \ V . It is a k-element set for some k ∈ N, so we can write W = {w1, . . . , wk}. For i = 1, . . . , k, let Ai := {x ∈ sG(V ) : (wi, x) ∈ E}. Notice that Ai = sG(V ∪ {wi}). Using the inclusion-exclusion principle again, we obtain that G(V ). This is clear if V = U since e♯ k e♯ sG(V ) \ G(V ) = eG(V ) = (cid:12)(cid:12)(cid:12) X = sG(V ) + k i=1 [ Ai(cid:12)(cid:12)(cid:12) (−1)i X i=1 I⊆{1,...,k} = s♯ G(V ) + = s♯ G(V ) + I=i k X i=1 k X i=1 (−1)i X I⊆{1,...,k} I=i (−1)i X I⊆{1,...,k} I=i k i=1 [ Ai(cid:12)(cid:12)(cid:12) Aj(cid:12)(cid:12)(cid:12) = sG(V ) − (cid:12)(cid:12)(cid:12) (cid:12)(cid:12)(cid:12) \ sG(cid:0)V ∪ {wj : j ∈ I}(cid:1)(cid:12)(cid:12)(cid:12) j∈I (cid:12)(cid:12)(cid:12) s♯ G(V ∪ {wj : j ∈ I}(cid:1). Therefore, since ϕ preserves the strong domination function, it preserves the exact domination function either. Now, we are ready to define an isomorphism (ϕ, ξ) : G → G′. That is, ϕ was originally given, and we intend to define a bijection ξ : X → X ′ such that E′ = 10 G. CZ´EDLI, J. GRYGIEL, AND K. GRYGIEL {(ϕ(u), ξ(x)) : (u, x) ∈ E}. Clearly, (10) eG(V1) ∩ eG(V2) = ∅ for all V1 6= V2 ∈ P (U ), and X = [ V ∈P (U) eG(V ); and the analogous assertion holds for G′. Notice at this point that the elements of X with degree 0 belong to eG(∅). For each V ∈ P (U ), let us fix a bijection ξV : eG(V ) → eG′(ϕ(V )); this is possible since e♯ G′(ϕ(V )). Then ξ := SV ∈P (U) ξV is an X → X ′ bijection by (10). Observe that the role of (G, ϕ, ξ) and that of (G′, ϕ−1, ξ−1) can be interchanged. Hence, in order to prove that (ϕ, ξ) is an isomorphism, it suffices to show that (ϕ, ξ) sends edges to edges. To do so, assume that (u, x) ∈ E. Let V := {v ∈ U : (v, x) ∈ E}. Then u ∈ V and x ∈ eG(V ). Hence ϕ(u) ∈ ϕ(V ) and ξ(x) = ξV (x) ∈ eG′(ϕ(V )). By the definition of eG′, this implies that (ϕ(u), ξ(x)) belongs to E′, as desired. (cid:3) G(V ) = e♯ 3.5. The end of the proof. Based on the auxiliary statements given so far, we are now in the position to complete the proof of the main result. Proof of Theorem 1.1 (ii) and (iii). Assume that L and L′ are finite distributive lattices and (ψ, κ) : (cid:0)S wd(L), ≤, S(L), η0, η1, w(cid:1) → (cid:0)S wd(L′), ≤′, S(L′), η′ 0, η′ 1, w′(cid:1) is an isomorphism between their weighted double skeletons. Remember that η0 and η1 were defined right after (4), the meaning of η′ 1 is analogous, and the diagram 0 and η′ (11) commutes for i ∈ {0, 1}. Observe that (12) length(Ji(L)) ≤ 1 and length(Ji(L′)) ≤ 1. Indeed, in case of part (iii) this is assumed. In case of part (ii), the assump- tion together with the meaning of κ implies that L′ is also H 2-irreducible, whence length(Ji(L)) ≤ 1 and length(Ji(L′)) ≤ 1 follow from part (i). Firstly, assume that S(L) = 1. Then S(L′) = κ(S(L)) = 1. It is well-known, and follows from Lemmas 3.5 and 3.6(iii), that L is boolean iff 1L = W At(L). There- fore, we obtain from Lemma 3.5 that S(L) = 1 iff L is boolean iff length(Ji(L)) = 0, and the same holds for L′. Therefore length(L) = w∗(0L, 1L) = w∗′(ψ(0L), ψ(1L)) = w∗′(0L′, 1L′) = length(L′), and we conclude that L ∼= L′. Observe that the role of L and that of L′ in the above argument can be interchanged, whence we also conclude that length(Ji(L)) = 0 iff length(Ji(L′)) = 0. In the rest of the proof, we assume that length(Ji(L)) = 1. Then, by the previous paragraph and (12), length(Ji(L′)) = 1 also holds. We are going to define some auxiliary sets and structures associated with S wd(L); their "primed" counterparts associated with S wd(L′) are understood analogously. DISTRIBUTIVE LATTICES DETERMINED BY WEIGHTED DOUBLE SKELETONS 11 Let U := Ji+(L), X := At(L), and E := {(u, x) ∈ U × X : u > x}. Notice that none of U , X and E is empty since length(Ji(L)) = 1. Obviously, the directed bipartite graph G := (U, X, E) determines the poset Ji(L). Therefore, G and G′ determine L and L′, respectively, up to isomorphism. Consequently, by Lemma 3.9, it suffices to find a bijection ϕ : U → U ′ such that (13) d♯ G′(ϕ(V )) = d♯ G(V ) holds for all V ∈ P ∞(U ). It follows from the commutativity of (11) that, for γ ∈ S(L), (14) ψ(0γ) = ψ(η0(γ)) = η′ 0(κ(γ)) = 0κ(γ) and, similarly, ψ(1γ) = 1κ(γ). For α ∈ S+(L), Lemma 3.1 and (14) imply that Jα = X β≤α, β∈S(L) µS(L)(β, α) · w∗(0α, 1β) = X β≤α, β∈S(L) = X β≤α, β∈S(L) µS(L′)(κ(β), κ(α)) · w∗′(ψ(0α), ψ(1β)) µS(L′)(κ(β), κ(α)) · w∗′(0κ(α), 1κ(β)) = X β ′≤κ(α), β ′∈S(L′) µS(L′)(β′, κ(α)) · w∗′(0κ(α), 1β ′) = Jκ(α). This allows us to fix a bijection ϕα : Jα → Jκ(α). (Notice that if Jα happens to be empty, then ϕα = ∅ is the empty mapping.) Let ϕ be the union of all these ϕα, α ∈ S+(L). It follows from Lemma 3.6(ii) that ϕ is a mapping. Since the union of the corresponding Jα, α ∈ S+(L), is Ji+(L) = U by Lemma 3.6(iii), and the analogous assertion holds for U ′, ϕ : U → U ′ is a bijective mapping. It follows from Lemma 3.6(ii)-(iii) that for each u ∈ U , there is a unique α(u) ∈ S+(L) such that u ∈ Jα(u). Similarly, for each u′ ∈ U ′, there is a unique α′(u′) ∈ S+(L′) such that u′ ∈ Jα′(u′). The definition of ϕ implies that (15) α′(cid:0)ϕ(u)(cid:1) = κ(cid:0)α(u)(cid:1), for all u ∈ U . Assume that u ∈ U . Then u is not an atom of L, and its only lower cover in L is 0α(u). Hence, for any a ∈ At(L), we have that u > a ⇐⇒ 0α(u) ≥ a ⇐⇒ a ∈ domin(α(u)). This yields that, for any V ∈ P (U ), dG(V ) = Su∈V domin(α(u)). Therefore, taking the meaning of η0 into account and using Lemma 3.8, (16) G(V ) = w∗(cid:16)η0(ζ0), η0(cid:0) _ d♯ u∈V α(u)(cid:1)(cid:17). Indicating the referenced formulas or their "primed version" at the equation signs and using that (ψ, κ) preserves the extended weight function, we obtain that G′(ϕ(V )) =(16)′ d♯ w∗′(cid:16)η′ 0(ζ′ 0), η′ =(15) w∗′(cid:16)η′ 0(κ(ζ0)), η′ = w∗′(cid:16)η′ 0(κ(ζ0)), η′ u∈V 0(cid:0) _ 0(cid:0) _ 0(cid:16)κ(cid:0) _ α′(ϕ(u))(cid:1)(cid:17) κ(α(u))(cid:1)(cid:17) α(u)(cid:1)(cid:17)(cid:17) u∈V u∈V 12 G. CZ´EDLI, J. GRYGIEL, AND K. GRYGIEL =(11) w∗′(cid:16)ψ(η0(ζ0)), ψ(cid:16)η0(cid:0) _ = w∗(cid:16)η0(ζ0), η0(cid:0) _ α(u)(cid:1)(cid:17) =(16) d♯ u∈V α(u)(cid:1)(cid:17)(cid:17) u∈V G(V ). This proves (13) for V ∈ P (U ). We are left with the case V = ∞ ∈ P ∞(U ). Then, using Lemma 3.6(iii), the validity of (13) is obtained as follows: G(∞) = At(L) = At(ζ0) = w∗(0ζ0 , 1ζ0) = w∗′(ψ(0ζ0 ), ψ(1ζ0 )) d♯ =(14) w∗′(0κ(ζ0), 1κ(ζ0)) = w∗′(0ζ ′ = At(ζ′ 0) = At(L′) = d♯ 0) 0, 1ζ ′ G′(∞) = d♯ G′(ϕ(∞)). (cid:3) References [1] Bandelt, H.J.: Tolerance relations of lattices. Bulletin of the Australian Math. Soc. 23, 367 -- 381 (1981) [2] Chajda, I.: Algebraic Theory of Tolerances. Palacky University Olomouc (1991) [3] Cz´edli, G.: Factor lattices by tolerances. Acta Sc. Math. (Szeged) 44, 35 -- 42 (1982). [4] Cz´edli, G., Gratzer, G.: Lattice tolerances and congruences. Algebra Universalis 66, 5 -- 6 (2011) [5] Cz´edli, G, Horv´ath, E.K., Radeleczki, S.: On tolerance lattices of algebras in congruence modular varieties. Acta Math. Hungar. 100, 9 -- 17 (2003) [6] Cz´edli, G., Klukovits, L.: A note on tolerances of idempotent algebras. Glasnik Matematicki (Zagreb) 18 (38), 35 -- 38 (1983) [7] Gratzer, G.: Lattice Theory: Foundation. Birkhauser/Springer, Basel (2011) [8] Gratzer, G., Wenzel G.H.: Notes on tolerance relations on lattices. Acta Sci. Math. (Szeged) 54, 229 -- 240 (1990) [9] Grygiel, J.: The concept of gluing for lattices. Wydawnictwo WSP, Cz¸estochowa (2004) [10] Grygiel, J.: Weighted double skeletons. Bulletin of the Section of Logic 35, 37 -- 47 (2006) [11] Herrmann, Ch.: S-verklebte Summen von Verbanden. Math. Z. 130, 255 -- 274 (1973) [12] Reuter, K.: Counting formulas for glued lattices. Order 1, 265 -- 276 (1985) [13] Wille, R.: Complete tolerance relations of concept lattices (1983, preprint) E-mail address: [email protected] URL: http://www.math.u-szeged.hu/~czedli/ University of Szeged, HUNGARY 6720 E-mail address: [email protected] URL: http://www.j.grygiel.eu/ Institute of Mathematics and Computer Science, Jan D lugosz University of Cze¸stochowa, Poland E-mail address: [email protected] URL: http://tcs.uj.edu.pl/Grygiel Theoretical Computer Science Department, Faculty of Mathematics and Computer Science, Jagiellonian University, ul. Prof. Lojasiewicza 6, 30-348 Krak´ow, Poland
1912.10855	1	1912	2019-12-20T12:10:12	Inverse eigenproblems and approximation problems for the generalized reflexive and antireflexive matrices with respect to a pair of generalized reflection matrices	[ "math.RA" ]	A matrix $P$ is said to be a nontrivial generalized reflection matrix over the real quaternion algebra $\mathbb{H}$ if $P^{\ast }=P\neq I$ and $P^{2}=I$ where $\ast$ means conjugate and transpose. We say that $A\in\mathbb{H}^{n\times n}$ is generalized reflexive (or generalized antireflexive) with respect to the matrix pair $(P,Q)$ if $A=PAQ$ $($or $A=-PAQ)$ where $P$ and $Q$ are two nontrivial generalized reflection matrices of demension $n$. Let ${\large \varphi}$ be one of the following subsets of $\mathbb{H}^{n\times n}$ : (i) generalized reflexive matrix; (ii)reflexive matrix; (iii) generalized antireflexive matrix; (iiii) antireflexive matrix. Let $Z\in\mathbb{H}^{n\times m}$ with rank$\left( Z\right) =m$ and $\Lambda=$ diag$\left( \lambda_{1},...,\lambda_{m}\right) .$ The inverse eigenproblem is to find a\ matrix $A$ such that the set ${\large \varphi }\left( Z,\Lambda\right) =\left\{ A\in{\large \varphi}\text{ }\|\text{ }AZ=Z\Lambda\right\} $ nonempty and find the general expression of $A.$\newline In this paper, we investigate the inverse eigenproblem ${\large \varphi}\left( Z,\Lambda\right) $. Moreover, the approximation problem: $\underset{A\in{\large \varphi}}{\min\left\Vert A-E\right\Vert _{F}}$ is studied, where $E$ is a given matrix over $\mathbb{H}$\ and $\parallel \cdot\parallel_{F}$ is the Frobenius norm.	math.RA	math	Inverse eigenproblems and approximation problems for the generalized reflexive and antireflexive matrices with respect to a pair of generalized reflection matrices1 Haixia Chang aSchool of Statistics and Mathematics, Shanghai Lixin University of Accounting and Finance, Shanghai 201209, P.R. China Abstract A matrix P is said to be a nontrivial generalized reflection matrix over the real quaternion algebra H if P ∗ = P 6= I and P 2 = I where ∗ means conjugate and trans- pose. We say that A ∈ Hn×n is generalized reflexive (or generalized antireflexive) with respect to the matrix pair (P, Q) if A = P AQ (or A = −P AQ) where P and Q are two nontrivial generalized reflection matrices of demension n. Let ϕ be one of the following subsets of Hn×n : (i) generalized reflexive matrix; (ii)reflexive matrix; (iii) generalized antireflexive matrix; (iiii) antireflexive matrix. Let Z ∈ Hn×m with rank(Z) = m and Λ = diag(λ1, ..., λm) . The inverse eigenproblem is to find a matrix A such that the set ϕ (Z, Λ) = {A ∈ ϕ AZ = ZΛ} nonempty and find the general expression of A. In this paper, we investigate the inverse eigenproblem ϕ (Z, Λ). Moreover, the approxi- is studied, where E is a given matrix over H and k · kF mation problem: min kA − EkF A∈ϕ is the Frobenius norm. Keywords inner inverse of a matrix; reflexive inverse of a matrix; generalized reflexive matrix, generalized antireflexive matrix, Frobenius norm, approximation problem 2000 AMS Subject Classifications 15A24, 15A33, 15A57, 15A09 1. Introduction Throughout Hm×n denotes the set of all m × n matrices over the real quaternion algebra H = {a0 + a1i + a2j + a3k i2 = j2 = k2 = ijk = −1 and a0, a1, a2, a3 are real numbers}; the symbols I, A∗, r(A), A(1) A+, k A kF , stand for the identity matrix with the appropriate size, the conjugate transpose, the rank, the inner inverse, the reflective inverse of a matrix A ∈ Hm×n and Frobenius norm of A ∈ Hm×n, respectively; define inner product, hA, Bi = trace (B ∗A) for all A, B ∈ Hm×n, then Hm×n is a Hilbert inner product space and the norm that is generated by this inner product is Frobenius norm. The two matrices LA and RA stand for LA = I − A+A, RA = I − AA+ where A+ is an any but fixed reflexive inverse of A. Vn denotes the n × n backward identity matrix whose elements along the southwest-northeast diagonal are ones and whose remaining elements are zeros. The real quaternion matrices play an important role in computer science, quantum physicis, and so on ( [20]-[22]). The more interests of quaternion matrices have been witnessed recently (e.g., [1], [23]). As H is noncommutative for its mutiplication, i.e., ab 6= ba, for a, b ∈ H in general, it bring some obstacles in the study of quatenion matrices. Many authors has used the method of embeddings of the quaternions into complex vector spaces, through symplectic representations, such as D.R. Farenick and B. A.F. Pidkowich in [3] and homotopy theory, such as Zhang in [1]. 1This research was supported by the Natural Science Foundation of China (11601328), the research fund of Shanghai Lixin University of Accounting and Finance (AW-22-2201-00118). 1 2 Since left and right scalar multiplications are different, it force us to consider Ax = λx, and Ax = xλ separately. A quaternion λ is said to a left (right) eigenvalue if Ax = λx (Ax = xλ) . In 1985, Wood proposed that every n × n quaternion matrix has at least one left eigenvalue in H and Zhang in [1] give its proof with the homotopy theory method. But the problem that how many left eigenvalues a square quaternion matrix has has not been solved yet. While the right eigenvalues of a square quaternion matrix has been well studied, Zhang in [1] explicitly refer to any n × n quaternion matrix A has exactly n right eigenvalues which are complex numbers with nonnegative imaginary parts (standard eigenvalues of A). Weighing the above cases, we only investigate the right inverse eigenproblems of a square quaternion matrix in this paper. Reflexive and antireflexive matrices with respect to a generalized reflection matrix have been widely [10]-[15]. But the generalized reflexive and generalized antireflexive matrices with respect studied, e.g. to the generalized reflection matrix pair have not been widely studied yet. In 1998, Chen in [4] defined that A ∈ Hm×n is generalized reflexive (or generalized antireflexive) and give two special subsets of the space Cm×n. The application of the generalized reflexive (or generalized antireflexive) matrix is very wide in engineering and science, such as the altitude estimation of a level network, electric network and structural analysis of trusses and so on, see [4]. Definition 1.1. A matrix A ∈ Hn×n is generalized reflexive with respect to the matrix pair (P, Q) if A = P AQ and A ∈ Hn×n is generalized antireflexive with respect to the matrix pair (P, Q) if A = −P AQ,where P and Q are two nontrivial generalized reflection matrices of demension n, i.e., P ∗ = P 6= I, P 2 = I, and Q∗ = Q 6= I, Q2 = I. The generalized reflexive and the generalized antireflexive matrix are further generalization of reflexive and antireflexive, respectively. From this definition, it is clear that A is a reflexive martix if generalized reflection matrices P = Q, i.e., A = P AP and a centrosymmetric matrix if P = Q = Vn, i.e., A = VnAVn. Centrosymmetric and centroskew matrices was investigated by many authors ,such as Andrew [5],[8], Weaver [9], Boullion and Atchison [6], Wang [2], and the others. Define two special subspaces of Hn×n Hn×n r Hn×n a (P, Q) =(cid:8)A A ∈ Hn×n and A = P AQ(cid:9) , (P, Q) =(cid:8)A A ∈ Hn×n and A = −P AQ(cid:9) , where P and Q are two generalized reflection matrices of dimension m and n, respectively. Hn×n If A ∈ (P, Q)) , then A is a generalized reflexive (generalized antireflexive) matrix with (P, Q) (∈ Hn×n r a respect to nontrivial generalized reflection matrix pair (P, Q) . The inverse eigenproblems of quaternion matrices: given matrices Z ∈ Hn×m and Λ =diag(λ1, ..., λm) , find A ∈ Hn×n satisfying AZ = ZΛ, where λ1, ..., λm are complex numbers with nonnegative imaginary parts. The inverse problems and inverse eigenproblems of matrices has been wide topics, (e.g., [14], [16]-[19]) with the method involing Moore-Penrose inverse or singular value decompositions of matrices related to (Z, Λ) , respectively. In this paper, we investigate the inverse (right) eigenproblem of (Z, Λ) , find A ∈ Hn×n r (P, Q) (∈ Hn×n (P, Q)) a (P, Q) and present the general expression of A. We make use of the structure properties of A ∈ Hn×n (∈ Hn×n (P, Q)) to deal with the inverse eigenproblem. Moreover, the approximation problem: min kA − EkF a r , where E is a given matrix over H and ϕ is one of Hn×n (P, Q). The explicit expression of A is also presented. Futhermore, it is pointed that some results in recent papers are special cases of this paper. (P, Q) and Hn×n a r A∈ϕ 2. The solution of the inverse eigenproblems of (Z, Λ) 3 In this section, we first give a structure property of generalized reflexive matrix and a generalized antireflexive matrix. Then we make use of the structure property to derive the general expression of A ∈ Hn×n (P, Q)) satisfying AZ = ZΛ. (P, Q) (∈ Hn×n r a By Proposition 2.1 in ([3]), we know that R ∈ Hn×n has infinitely many nonreal right eigenvalues if R has a nonreal right eigenvalue. Because if (λ, ξ) is a right eigenvalue, then (cid:0)w−1λw, ξw(cid:1) is also a right eigenpair of R. Then we introduce the orbit θ(λ), i.e., a class from a partition of H. For example, a similarity orbit θ(λ) of λ: and a conjugacy class θ(λ) of λ: θ(λ) =(cid:8)w−1λw w ∈ H, w 6= 0(cid:9) θ(λ) = { ¯wλw w ∈ H, w = 1} . Lemma 2.1. (See Theorem 3.3 in [3]) If R ∈ Hn×n is normal, then there are matrices D, U ∈ Hn×n such that : (1) U is unitary matrix, D is a diagonal matrix, and U ∗RU = D; (2) each diagonal entry of D is a complex number contained in the closed upper halfplane C+. (3) q ∈ H is a right eigenvalue of R if and only if q ∈ θ(λ) for some diagonal element λ of D. Lemma 2.2. Suppose that P ∈ Hn×n is a nontrivial generalized reflection matrix; then there exists a unitary matrix U ∈ Hn×n such that P = U" Ir1 0 −In−r1 # U ∗. 0 (2.1) Proof. A matrix P ∈ Hn×n is a nontrivial generalized reflection matrix, i.e., P ∗ = P 6= I, P 2 = I. Obviously, P is a normal matrix. According to the minimal polynomial of P, we know 1, -1 are all right engivalues (or standard eigenvalues ) of P. By Lemma 2.1, there exists a unitary matrix U satisfying (2.1). (cid:3) If a matrix Q ∈ Hn×n is a nontrivial generalized reflective matrix, by Lemma 2.2, we get Q = V " Ir2 0 −In−r2 # V ∗, 0 (2.2) where V is a unitary matrix. We can use a staightforward method to obtain U and V mentioned above. In fact, by P 2 = I, i.e., (I − P ) (I + P ) = 0 = 0 or (I + P ) (I − P ) = 0, we know the column vectors of I + P and I − P are the eigenvector belonging to the right eigenvalues λ = 1, λ = −1 of the matrix P, respectively. We can only apply the Gram-Schmidt process to the columns of I + P and I − P,respectively. Let and in (2.1) I + P Gram − Schmidt −−−−−−−−−−−−−→ I − P Gram − Schmidt −−−−−−−−−−−−−→ U1 U2 where U ∗ 1 U1 = Ir1 , U ∗ 2 U2 = Im−r1 , U ∗ 1 U2 = 0. Similarily, we can get the factorizaton of V in (2.2) U =h U1, U2 i V =h V1, V2 i (2.3) (2.4) where V ∗ 1 V1 = Ir2 , V ∗ 2 V2 = In−r2 , V ∗ 1 V2 = 0. By (2.1)-(2.4), we obtain where and 0 P =h U1, U2 i" Ir1 Q =h V1, V2 i" Ir2 1 U ∗ 0 −In−r1 #" U ∗ 2 # , 0 −In−r2 #" V ∗ 2 # , 1 V ∗ 0 U ∗ 1 U1 = Ir1 , U ∗ 2 U2 = Im−r1 , U ∗ 1 U2 = 0 1 V1 = Ir2 , V ∗ V ∗ 2 V2 = In−r2 , V ∗ 1 V2 = 0. 4 (2.5) (2.6) (2.7) (2.8) Theorem 2.3. A matrix A ∈ Hn×n is generalized reflexive matrix with respect to the nontrivial gener- alized reflection matrix pair (P, Q) if and only if A can be expressed A =h U1, U2 i" A11 0 where 0 A22 #" V ∗ 1 V ∗ 2 # , A11 = U ∗ 1 AV1, A22 = U ∗ 2 AV2 and U1, U2, V1, V2 are defined as (2.3) and (2.4). Proof. By the definition of A ∈ Hn×n r (P, Q) , (2.5) and (2.6), we have A = P AQ =h U1, U2 i" Ir1 0 0 −In−r1 #" U ∗ 1 U ∗ 2 # Ah V1, V2 i" Ir2 0 0 −In−r2 #" V ∗ 1 V ∗ 2 # , (2.9) (2.10) (2.11) (2.12) (2.13) i.e., Put 0 1 U ∗ " U ∗ 2 # Ah V1, V2 i =" Ir1 0 −In−r1 #" U ∗ " U ∗ A21 A22 # , 2 # Ah V1, V2 i =" A11 A12 2 # Ah V1, V2 i" Ir2 1 U ∗ 1 U ∗ 0 0 −In−r2 # where A11 ∈ Hr1×r2 , A12 ∈ Hr1×(n−r2), A21 ∈ H(n−r1)×r2 , A22 ∈ H(n−r1)×(n−r2). Substituting (2.13) into (2.12), we have " A11 A12 A21 A22 # =" Ir1 0 −In−r1 #" A11 A12 A21 A22 #" Ir2 0 −In−r2 # =" A11 −A12 −A21 A22 # , 0 0 yielding A12 = 0, A21 = 0. (2.11) becomes (2.9). By (2.9), (2.7) and (2.8), it is easy to get (2.10). Conversly, if (2.9) holds it is easy to verify A is generalized reflexive matrix with respect to the nontrivial generalized reflection matrix pair (P, Q). (cid:3) Similarly, we can obtain the factorization of generalized antireflexive matrix with respect to the gen- eralized reflection matrix pair (P, Q). We have the following Theorem. Theorem 2.4. A matrix A ∈ Hn×n a (P, Q) if and only if A can be expressed as A =h U1, U2 i" 0 A21 where A12 0 #" V ∗ 1 V ∗ 2 # , A12 = U ∗ 1 AV2, A21 = U ∗ 2 AV1, 5 and U1, U2, V1, V2 are defined as (2.3) and (2.4). Now we discuss the solution of the inverse eigenproblems of (Z, Λ) . Lemma 2.5. Given B ∈ Hm×t X ∈ Hs×t. Then the inverse problem is consistent if and only if AX = B BX +X = B. In that case, the general form of A ∈ Hn×n is where W is an arbitrary matrix over H with appropriate size. A = BX + + W RX Theorem 2.6. Given Z ∈ Hn×m, Λ =diag(λ1, ..., λm). Let U1, U2, V1, V2 be defined as (2.3), (2.4), 0 Ψ # , Z =h V1X1, V2X2 i =h U1Y1, U2Y2 i , Λ =" Φ 0 (2.14) where 0 < k < m, X1 ∈ Hr1×k, X2 ∈ H(n−r1)×(m−k), Y1 ∈ Hr2×k, Y2 ∈ H(n−r2)×(m−k), Φ =diag(λ1, ..., λk) , Ψ =diag(λk+1, ..., λm) . Then the set is nonempty if and only if ϕ1 (Z, Λ) =(cid:8)A ∈ Hn×n r (P, Q) AZ = ZΛ(cid:9) Y1ΦX1X + 1 = Y1Φ, Y2ΨX2X + 2 = Y2Ψ. In that case, the general expression of A is A =h U1, U2 i" Y1ΦX + 0 1 + W1RX1 0 2 + W2RX2 #" V ∗ 1 V ∗ 2 # , Y2ΨX + (2.15) (2.16) (2.17) where W1, W2 are arbitrary matrices over H with appropriate sizes. Proof. Since A ∈ Hn×n into AZ = ZΛ. We get r i.e., (P, Q) , A has the form of (2.9). By (2.3) and (2.4), substituting (2.9) and (2.14) " A11X1 0 0 A22X2 # =" Y1Φ 0 0 Y2Ψ # , A11X1 = Y1Φ, A22X2 = Y2Ψ. The set ϕ (Z, Λ) is nonempty if and only if the inverse problem A11X1 = Y1 and A22X2 = Y2Ψ are consistent. By Lemma 2.5, the invese problem A11X1 = Y1 and A22X2 = Y2Ψ are consistent if and only if (2.16) holds and their general solution are A11 = Y1ΦX + A22 = Y2ΨX + 1 + W1RX1 , 2 + W2RX2 , (2.18) respectively. Then substituting (2.18) into (2.9), we get (2.17). Conversly, if (2.16) holds and A has the expression of (2.9), it is easy to verify A ∈ ϕ (Z, Λ) . (cid:3) If A ∈ Hn×n r (P, P ) , i.e., A is a reflexive matrix with respect to a generalized reflection matrix P. We have the following Corollary. Corollary 2.7. Given Z ∈ Hn×m, Λ =diag(λ1, ..., λm) . Let U1, U2, V1, V2 be defined as (2.3) and (2.4), 6 0 Ψ # , Z =h U1X1, U2X2 i , Λ =" Φ 0 where 0 < k < m, X1 ∈ Hr1×k, X2 ∈ H(n−r1)×(m−k), Φ =diag(λ1, ..., λk) , Ψ =diag(λk+1, ..., λm) . Then the set is nonempty if and only if ′′ ϕ 1 (Z, Λ) =(cid:8)A ∈ Hn×n r (P, P ) AZ = ZΛ(cid:9) (2.19) In that case, the general expression of A is X1ΦX1X + 1 = X1Φ, X2ΨX2X + 2 = X2Ψ. A =h U1, U2 i" X1ΦX + 0 1 + W1RX1 0 2 + W2RX2 #" U ∗ 1 U ∗ 2 # , X2ΨX + where W1, W2 are arbitrary matrices over H with appropriate sizes. Theorem 2.8. Given Z ∈ Hn×m, Λ =diag(λ1, ..., λm). Let U1, U2, V1, V2 be defined as (2.3), (2.4), 0 Ψ # , Z =h V2X1, V1X2 i =h U1Y1, U2Y2 i , Λ =" Φ 0 where 0 < l < m, X1 ∈ H(n−r2)×l, X2 ∈ Hr2×(m−l), Y1 ∈ Hr2×l, Y2 ∈ H(n−r2)×(m−l), Φ =diag(λ1, ..., λl) , Ψ =diag(λl+1, ..., λm) . Then the set ϕ (Z, Λ) = {A ∈ Hn×n a ϕ2 (Z, Λ) =(cid:8)A ∈ Hn×n a (P, Q) AZ = ZΛ(cid:9) (P, Q) AZ = ZΛ} is nonempty if and only if In that case, the general expression of A is Y1ΦX1X + 1 = Y1Φ, Y2ΨX2X + 2 = Y2Ψ. (2.20) A =h U1, U2 i" 0 Y1ΦX + 1 + W1RX1 Y2ΨX + 2 + W2RX2 0 #" V ∗ 1 V ∗ 2 # , where W1, W2 are arbitrary matrices over H with appropriate sizes. Proof. Similar to the proof of the Theorem 2.6. (cid:3) If A ∈ Hn×n a (P, P ) , i.e., A is a antireflexive matrix with respect to a generalized reflection matrix P. We have the following Corollary. Corollary 2.9. Given Z ∈ Hn×m, Λ =diag(λ1, ..., λm) . Let U1, U2 be defined as (2.3), 0 Ψ # , Z =h U1X1, U2X2 i , Λ =" Φ 0 where 0 < l < m, X1 ∈ H(n−r1)×l, X2 ∈ Hr1×(m−l), Φ =diag(λ1, ..., λl) , Ψ =diag(λl+1, ..., λm) . Then the set is nonempty if and only if ′′ ϕ 2 (Z, Λ) =(cid:8)A ∈ Hn×n a (P, P ) AZ = ZΛ(cid:9) (2.21) X1ΦX1X + In that case, the general expression of A is 1 = X1Φ, X2ΨX2X + 2 = X2Ψ. A =h U1, U2 i" 0 X1ΦX + 1 + W1RX1 X2ΨX + 2 + W2RX2 0 #" U ∗ 1 U ∗ 2 # , 7 where W1, W2 are arbitrary matrices over H with appropriate sizes. Remark 2.1. We find it is not necessary that P and Q are hermitian involutionary matrices in the process of dealing with inverse eigenproblems. So we can only consider that P and Q are involutionary matrices and obtain some corresponding conclusions. Furthermore, when P = Q are only involutionary matrix, we can get some results of inverse eigenproblems of [14]. 3. The approximation problem to the solution to the inverse eigenproblems In this section, we consider the approximation problem: min kA − EkF A∈ϕ matrix over H and ϕ is either Hn×n r (P, Q) or Hn×n a (P, Q) . , where E ∈ Hm×n is a given Lemma 3.1. Suppose that E ∈ Hm×n, Γ ∈ Hm×m and ∈ Hn×n where Γ2 = Γ = Γ∗ and 2 = = ∗. Then the nearness problem min kΓX − EkF is consistent if and only if X ∈Hm×n In that case, Proof. Note that Then Γ(X − E) = 0. min X ∈Hm×n kΓX − EkF = kΓE − EkF ΓX − E = (ΓE − E) + Γ(X − E). (ΓX − E)∗ (ΓX − E) = (ΓE − E)∗(ΓE − E) + [Γ(X − E)]∗ Γ(X − E) + G + G∗ (3.1) where G = (ΓE − E)∗Γ(X − E) = (E ∗Γ − E ∗)Γ(X − E) = ( − I) E ∗Γ(X − E) Suppose (λ, x) is a right eigenpair for the matrix G ∈ Hn×n, i.e., Gx = xλ with λ 6= 0. Since 2 = , i.e., ( − I) = 0, we obtain Gx = ( − I) E ∗Γ(X − E)x = xλ = 0 i.e. x = 0. yielding x = 0. Hence, G has nonzero right eigenvalues. Consequently, taking traces in (3.1) yields kΓX − Ek2 F = kΓX − Ek2 F + kΓ(X − E)k2 F . min X ∈Hm×n kΓX − EkF is equivalent to min Clearly, when X = E + LΓY R, with appropriate size. Hence kΓ(X − E)kF . kΓ(X − E)kF = 0 where Y is an arbitrary matrix over H X ∈Hm×n min X ∈Hm×n min X ∈Hm×n kΓX − EkF = kΓE − EkF . Theorem 3.2. Given a matrix E ∈ Hn×n; " U ∗ 2 # Eh V1, V2 i =" E11 E12 E21 E22 # . 1 U ∗ (cid:3) (3.2) where U1, U2, V1, V2 are defined as (2.3) and (2.4) and E11 ∈ Hr1×r2 , E12 ∈ Hr1×(n−r2), E21 ∈ H(n−r1)×r2 , E22 ∈ H(n−r1)×(n−r2). Then the approximation problem: min kA − EkF has a unique solu- A∈ϕ1 tion 8 Ar =h U1, U2 i" Y1ΦX + 0 1 + E11RX1 0 2 + E22RX2 #" V ∗ 2 # 1 V ∗ Y2ΨX + (3.3) where ϕ1 are defined as (2.15). Proof. A ∈ ϕ1, by Theorem 2.6 and (3.2), kA − Ek2 2 F 0 0 0 0 1 U ∗ Y2ΨX + 1 + W1RX1 1 + W1RX1 F =(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) 2 # Eh V1, V2 i(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) " Y1ΦX + 2 + W2RX2 # −" U ∗ =(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) E21 E22 #(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) " Y1ΦX + 2 + W2RX2 # −" E11 E12 =(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) 2 + W2RX2 − E22 #(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) " Y1ΦX + =(cid:13)(cid:13)W1RX1 −(cid:0)E11 − Y1ΦX + F +(cid:13)(cid:13)W2RX2 −(cid:0)E22 − Y2ΨX + 1 (cid:1)(cid:13)(cid:13) 2 (cid:1)(cid:13)(cid:13) F is consistent is equivalent to 1 + W1RX1 − E11 kA − Ek2 Y2ΨX + Y2ΨX + −E12 −E21 2 F 2 F 2 Therefore, min A∈ϕ1 2 F + kE12k2 F + kE21k2 F . min(cid:13)(cid:13)W1RX1 −(cid:0)E11 − Y1ΦX + 1 (cid:1)(cid:13)(cid:13) min(cid:13)(cid:13)W2RX2 −(cid:0)E22 − Y2ΨX + 2 (cid:1)(cid:13)(cid:13) 2 F 2 F , (3.4) (3.5) = RXi = R∗ Xi , i = 1, 2, it foillows from Lemma 3.1 that (3.4) is consistent if min(cid:13)(cid:13)W1RX1 −(cid:0)E11 − Y1ΦX + 1 (cid:1) −(cid:0)E11 − Y1ΦX + 1 (cid:1) RX1(cid:13)(cid:13)F are consistent. Since R2 Xi and only if and and (3.5) is consistent if and only if and (cid:2)W1 −(cid:0)E11 − Y1ΦX + 1 (cid:1)(cid:3) RX1 = 0 1 (cid:1)(cid:13)(cid:13)F =(cid:13)(cid:13)(cid:0)E11 − Y1ΦX + 1 − Y1ΦX + =(cid:13)(cid:13)E11X1X + (cid:2)W2 −(cid:0)E22 − Y2ΨX + 2 (cid:1)(cid:3) RX2 = 0 2 (cid:1)(cid:13)(cid:13)F =(cid:13)(cid:13)(cid:0)E22 − Y2ΨX + 2 − Y2ΨX + =(cid:13)(cid:13)E22X2X + 2 (cid:13)(cid:13)F 1 (cid:13)(cid:13)F W1 = E11 − Y1ΦX + W2 = E22 − Y2ΨX + 1 + T1X1X + 1 , 2 + T2X2X + 2 , (3.6) (3.7) (3.8) (3.9) min(cid:13)(cid:13)W2RX2 −(cid:0)E22 − Y2ΨX + 2 (cid:1) −(cid:0)E22 − Y2ΨX + . 2 (cid:1) RX2(cid:13)(cid:13)F It is easy to know the solutions of (3.6) and (3.7) are respectively, where T1, T2 are arbitary matrices with appropriate sizes. Substituting (3.8) and (3.9) into (2.17), we easily get (3.3). (cid:3) Corollary 3.3. Given a matrix E ∈ Hn×n; " U ∗ 2 # Eh U1, U2 i =" E11 E12 E21 E22 # . 1 U ∗ 9 where U1, U2 are defined as (2.3) and E11 ∈ Hr1×r1 , E12 ∈ Hr1×(n−r1), E21 ∈ H(n−r1)×r1 , E22 ∈ H(n−r1)×(n−r1). Then the approximation problem: min A∈ϕ kA − EkF has a unique solution ′′ 1 Ar =h U1, U2 i" X1ΦX + 0 1 + E11RX1 0 2 + E22RX2 #" U ∗ 2 # 1 U ∗ X2ΨX + ′′ where ϕ 1 are defined as (2.19). Theorem 3.4. Given a matrix E ∈ Hn×n; Suppose " U ∗ 2 # Eh V1, V2 i =" E11 E12 E21 E22 # , 1 U ∗ where U1, U2, V1, V2 are defined as (2.3) and (2.4) and E11 ∈ Hr1×r2 , E12 ∈ Hr1×(n−r2), E21 ∈ H(n−r1)×r2 , E22 ∈ H(n−r1)×(n−r2). Then the approximation problem: Aa = arg min kA − EkF has a A∈ϕ2 unique solution Aa =h U1, U2 i" 0 Y1ΦX + 1 + E12RX1 Y2ΨX + 2 + E21RX2 0 #" V ∗ 2 # 1 V ∗ where ϕ2 are defined as (2.20). Proof. Combing Theorem 2.7, similarily to the proof of Theorem 3.2, we easily complete the proof the Theorem. (cid:3) Corollary 3.5. Given a matrix E ∈ Hn×n; Suppose " U ∗ 2 # Eh U1, U2 i =" E11 E12 E21 E22 # , 1 U ∗ where U1, U2 are defined as (2.3) and (2.4) and E11 ∈ Hr1×r1 , E12 ∈ Hr1×(n−r1), E21 ∈ H(n−r1)×r1 , E22 ∈ H(n−r1)×(n−r1). Then the approximation problem: Aa = arg min kA − EkF has a unique solution A∈ϕ ′′ 2 Aa =h U1, U2 i" 0 X1ΦX + 1 + E12RX1 X2ΨX + 2 + E21RX2 0 #" U ∗ 2 # 1 U ∗ ′′ where ϕ 2 are defined as (2.21). 4. Conclusion We in this paper used the matrix decomposition method to give the properties of the generalized re- flexive and antireflexive matrices with respect to a pair of generalized reflection matrices over quaternion algebra. We studied the inverse eigenproblems and the approximation problems of the generalized reflex- ive and antireflexive matrices with respect to a pair of generalized reflection matrices. We also showed the explicit expressions and formulas. References 10 [1] F. Zhang, Quaternions and matrices of quaternions, Linear Algebra Appl. 251 (1997): 21-57. [2] Q.W. Wang, Bisymmetric and centrosymmetric solutions to systems of real quaternion matrix equations, Comput. Math. Appl., 49/5-6(2005), 641-650. [3] D.R. Farennick, B.A.F. Pidkowich, The spectral theorem in quaternions, Linear Algebra Appl. 371 (2003): 75-102. [4] H.C. Chen, Generalized reflexive matrices: special properties and application, SIAM J. Matrix Anal. Appl. 19 (1998): 140-153. [5] A.L. Andrew, Centrosymmetric matrices, SIAM Rev. 40(1998): 697-698. [6] W.C. Pye, T.L. Boullion, T.A. Atchison, The pseudoinverse of a centrosymmetric matrix, Linear Algebra Appl. 6 (1973): 201-204. [7] I.J. Good, The inverse of a centrosymmetric matrix, Technometrics 12 (1970): 925-928. [8] A.L. Andrew, Solution of equations involing centrosymmetric matrices, Technometrics 15 (1973): 151-162. [9] J.R. Weaver, Centrosymmetric (cross-symmetric) matrices, their basic problems, eigenvalues, and eigenvectors, Amer. Math. Monthly 92 (1985): 711-717. [10] Z.-Y. Peng, X.Y. Hu, The reflexive and anti-reflexive solutions of matrix equation AX = B, Linear Algebra Appl. 375 (2003): 147-155. [11] C.J. Meng, X.Y. Hu, The skew-symmetric orthogonal solutions of the matrix equation AX = B, Linear Algebra Appl. 402 (2005): 303-318. [12] F.Z. Zhou, X.Y. Hu and L. Zhang, The solvability conditions for the inverse problems of symmetric ortho-symmetric matrices, Appllied Math. and Comput. 154(2004):153-166. [13] W.F. Trench, Characterization and properties of matrice with generalized symmetry or skew symmetry, Linear Algebra Appl. 377 (2004): 207-218. [14] W.F. Trench, Inverse eigenproblems and associated approximation problems for matrices with generalized symmetry or skew symmetry, Linear Algebra Appl. 380 (2004): 199-211. [15] W.F. Trench, Hermitian, hermitian R−symmetric, and hermitian R−skew symmetric Procrustes problem, Linear Algebra Appl. 387 (2004): 83 -- 98. [16] Z.J. Bai, R.H. Chan, Inverse eigenproblem for centrosymmetric and centroskew matrices and their approximation, Theor. Comput. Sin., in press. [17] D. Xie, Y. Sheng, Inverse eigenproblem of anti-symmetric and persymmetric matrices and its approximation, Inverse Problems, 19 (2003): 217 -- 225. [18] F.Z. Zhou, X.Y. Hu, L. Zhang, The solvability conditions for the invese eigenvalue problems of centro-symmetric matrices, Linear Algebra Appl. 364 (2003): 147 -- 160. [19] D.X., X.Y. Hu, L. Zhang, The solvability conditions for the invese eigenvalue problems of symmetric and anti- persymmetric matrices and its approximation, Numer. Linear Algebra Appl. 10 (2003): 223 -- 234. [20] J.W. Shuai, The quaternion NN model: The identification of colour images (in Chinese), Acta Comput. Sinica 18(5), 372-379, (1995). [21] A. Rezon and L.P. Horwitz, Uniqueness of the scalar product in the tensor product of quaternion Hilbert modules, J. Math. Physical Sci.33 (9), 3098-3104, (1992). [22] S.L. Adler, Quaternionic Quantum Mechanics and Quantum Fields, Oxford University Press, Oxford, 1995. [23] D.I. Merino, V. Sergeichuk, Littlewood's algorithm and quaternion matrices, Linear Algebra Appl. 298 (1999): 193-208.
1610.04425	3	1610	2018-05-08T12:03:41	Verbally prime T-ideals and graded division algebras	[ "math.RA" ]	Let $F$ be an algebraically closed field of characteristic zero and let $G$ be a finite group. We consider graded Verbally prime $T$-ideals in the free $G$-graded algebra. It turns out that equivalent definitions in the ordinary case (i.e. ungraded) extend to nonequivalent definitions in the graded case, namely verbally prime $G$-graded $T$-ideals and strongly verbally prime $T$-ideals. At first, following Kemer's ideas, we classify $G$-graded verbally prime $T$-ideals. The main bulk of the paper is devoted to the stronger notion. We classify $G$-graded strongly verbally prime $T$-ideals which are $T$-ideal of affine $G$-graded algebras or equivalently $G$-graded $T$-ideals that contain a Capelli polynomial. It turns out that these are precisely the $T$-ideal of $G$-graded identities of finite dimensional $G$-graded, central over $F$ (i.e. $Z(A)_{e}=F$) which admit a $G$-graded division algebra twisted form over a field $k$ which contains $F$ or equivalently over a field $k$ which contains enough roots of unity (e.g. a primitive $n$-root of unity where $n = ord(G)$).	math.RA	math	VERBALLY PRIME T-IDEALS AND GRADED DIVISION ALGEBRAS ELI ALJADEFF AND YAAKOV KARASIK Abstract. Let F be an algebraically closed field of characteristic zero and let G be a finite group. We consider graded Verbally prime T -ideals in the free G-graded algebra. It turns out that equivalent definitions in the ordi- nary case (i.e. ungraded) extend to nonequivalent definitions in the graded case, namely verbally prime G-graded T -ideals and strongly verbally prime T - ideals. At first, following Kemer's ideas, we classify G-graded verbally prime T -ideals. The main bulk of the paper is devoted to the stronger notion. We classify G-graded strongly verbally prime T -ideals which are T -ideal of affine G-graded algebras or equivalently G-graded T -ideals that contain a Capelli polynomial. It turns out that these are precisely the T -ideal of G-graded iden- tities of finite dimensional G-graded, central over F (i.e. Z(A)e = F ) which admit a G-graded division algebra twisted form over a field k which contains F or equivalently over a field k which contains enough roots of unity (e.g. a primitive n-root of unity where n = ord(G)). 1. introduction Let F be an algebraically closed field of characteristic zero. The study of PI algebras over F is closely related to the study of T -ideals of the noncommutative free associative algebra F hXi where X is a countable set of variables. Since any T -ideal Γ appears as the ideal of polynomial identities of a suitable algebra A, it is natural to consider special types of T -ideals and try to understand what algebras have them as their ideal of identities. In the ungraded case two types of primeness on T -ideals are considered, namely prime T -ideals and verbally prime T -ideals. By definition a T -ideal I is prime if for any two ideals J1, J2 in F hXi such that J1J2 ⊆ I, either J1 ⊆ I or J2 ⊆ I whereas I is verbally prime if a similar condition holds whenever the ideals J1 and J2 are T -ideals. We adopt the definition from T -ideals and say that an algebra A is verbally prime if its T -ideal of identities Id(A) is verbally prime. As for primeness, we recall that an algebra A is prime if for ideals I and J in A with IJ = 0 we must have I = 0 or J = 0, hence we cannot just define primeness in terms of the corresponding T -ideal of identities. Nevertheless, by structure theorems of Amitsur ([5]) and Posner ([10], section 1.11) we have Theorem 1.1. Let A be a PI algebra. Then it is PI equivalent to a prime algebra B if and only if its T-ideal of polynomial identities Id(A) is a prime T-ideal. Fur- thermore, A is a prime PI algebra if and only if it is PI equivalent to Mn(F ) for some n > 0. Key words and phrases. graded algebras, polynomial identities, verbally prime, graded division algebras. 1 VERBALLY PRIME T-IDEALS AND GRADED DIVISION ALGEBRAS 2 As for the classification of verbally prime PI algebras recall the following funda- mental result which is due to Kemer [12]. Theorem 1.2. A is a verbally prime PI algebra if and only if it is PI equivalent to the Grassmann envelope of a finite dimensional Z/2Z−simple F -algebra A. It follows that if A is also affine then it is PI equivalent to a prime PI algebra. Posner's Theorem was recently extended to the G-graded setting by Karasik (see [11]). As a corollary he obtains a generalization of Theorem 1.1 for G-graded algebras (see Theorem 1.5 below). Our goal in this paper is to extend Theorem 1.2 to the G-graded setting where G is an arbitrary finite group. Somewhat more precisely, we wish to classify G−graded algebras (up to graded PI equivalence) which are PI as ungraded algebras and whose T -ideal of G-graded identities satisfy conditions which are analogue to verbally prime (see below for the precise definitions). Remark 1.3. A key ingredient in Kemer's proof of Theorem 1.2 is his own celebrated Representability Theorem which states that every PI algebra is PI equivalent to the Grassmann envelope of a finite dimensional Z/2Z− graded algebra. Also in the G- graded setting an important tool in our proofs is the Representability Theorem for G-graded algebras (see Theorem (A) below or ([1], Theorems 1.1 and 1.3)). Since the Representability Theorem makes sense only for G-graded algebras that are ungraded PI, we restrict our discussion to algebras of that type. It may seem that the generalization to G− graded algebras is a natural extension of Kemer's original work. Surprisingly, this is not the case as we shall soon explain. It is well known that for ordinary T -ideals (i.e. ungraded) the property of a T -ideal I being verbally prime may be expressed also in terms of polynomials, that is, I is verbally prime if and only if the following holds: if f, g ∈ F hXi are polynomials defined on disjoint sets of variables with f g ∈ I, then either f or g is in I. Note that if I is the T -ideal of identities of an algebra A, then the above condition on f and g says that if f g is an identity of A then either f or g is an identity of A. Most of the time we shall work with that terminology. It turns out that the equivalence of the two conditions mentioned above for verbally primeness of T -ideals is not longer valid when considering G−graded T - ideals and hence we obtain two primeness conditions on G−graded T -ideals. One finds that the corresponding condition for G-graded T -ideals expressed in terms of polynomials is strictly stronger and hence we will say that G-graded T -ideals (or G-graded algebras) satisfying that property are strongly verbally prime. For future reference we record below the following definitions. Definition 1.4. Let Γ be a G-graded T -ideal which contains a nonzero PI. (1) We say Γ is G-graded prime if for any G-graded ideals I and J with IJ ⊆ Γ, then I ⊆ Γ or J ⊆ Γ. (2) We say Γ is G-graded verbally prime if for any G-graded T -ideals I and J with IJ ⊆ Γ, then I ⊆ Γ or J ⊆ Γ. (3) We say Γ is G-graded strongly verbally prime if for any G-graded homoge- neous polynomials f and g defined on disjoint sets of variables with f g ∈ Γ, then f ∈ Γ or g ∈ Γ. VERBALLY PRIME T-IDEALS AND GRADED DIVISION ALGEBRAS 3 As mentioned above, Posner's theorem was generalized to the G-graded setting by Karasik ([11], Corollary 4.7). In particular it gives the classification of G-prime T -ideals. Theorem 1.5. Let A be a G-graded algebra which is PI as an ungraded algebra. Then its T -ideal of identities is prime if and only if A is G-graded PI equivalent to a finite dimensional G-simple F -algebra. The classification of G-graded verbally prime T -ideals (or algebras) was obtained by Berele and Bergen in [7] and independently by Di Vincenzo in [8] for the impor- tant case where G = Z2. The precise result is as follows. Theorem 1.6. Let A be a G−graded algebra which is also PI as an ungraded algebra. Then A is verbally prime if and only if it is G−graded PI equivalent to the Grassmann envelope of a finite dimensional G2 = Z/2Z × G−simple F -algebra. If A is also affine, then A is G-verbally prime if and only if it is G-prime, that is, if and only if A is G−graded PI equivalent to a finite dimensional G−simple F -algebra. The classification of G-graded T -ideals satisfying the stronger condition, namely G-strongly verbally prime, appears to be more involved. One needs to translate the language of G−graded polynomial identities into the algebraic structure of G- simple algebras. In this paper we provide a full answer in the affine case, namely we will classify up to G−graded PI equivalence G−strongly verbally prime algebras which are affine. At this point, the nonaffine case remains open. From Theorem 1.6 and the fact that strongly verbally primeness implies verbally primeness we have that in order to classify the G-graded affine algebras which are strongly verbally prime we need to determine the finite dimensional G-simple alge- bras A which satisfy the following condition: if f and g are G-graded homogeneous polynomials (with respect to the G-grading) with disjoint sets of graded variables and such that f g is an identity of A, then f or g is an identity of A. In order to state our main results we recall the structure of finite dimensional G-graded algebras which are G-simple. This result is key and is due to Bahturin, Sehgal and Zaicev [6]. Theorem 1.7. Let A be a f.d. G−simple algebra over an algebraically closed field F of zero characteristic. Then there is a subgroup H of G, a two cocycle c ∈ Z 2(H, F ∗) (the H−action on F ∗ is trivial) and g = (g1, . . . , gm) ∈ G(m), such that A ∼= F cH ⊗ Mm(F ), where (by this identification) the G−grading on A is given by Ag = spanF(cid:8)uh ⊗ ei,jg = g−1 i hgj(cid:9) . It is easy to see that up to a G-graded isomorphism the algebra A does not depend on the 2-cocycle c but only on its cohomology class [c] ∈ H 2(H, F ∗). We therefore denote the G-grading on the algebra A by PA = (H, [c], g). We also say that (H, [c], g) is a presentation of the G-graded algebra A. Note that if an algebra A is G-graded then it is U -graded in a natural way whenever G is a subgroup of U (just by putting Au = 0 for u ∈ U − G). It is clear in that case that A is strongly verbally prime as a G-graded algebra if and only if A is strongly verbally prime as a U -graded algebra. We therefore assume for the rest of the paper that the grading is connected. This means that the support of the grading generates the group G. Here is one of the two main results of the paper. VERBALLY PRIME T-IDEALS AND GRADED DIVISION ALGEBRAS 4 Theorem 1.8. Let Γ be a G-graded T -ideal where G is a finite group. Suppose Γ contains an ungraded Capelli polynomial cn for some integer n (equivalently, Γ is the T -ideal of G-graded identities of an affine PI algebra). Then the ideal Γ is G-graded strongly verbally prime if and only if it is the T -ideal of G-graded identities of a finite dimensional G-simple algebra A with presentation PA = (H, α, g = (g1, . . . , gk)) satisfying the following conditions: (1) The group H is normal in G. (2) The different cosets of H in G are equally represented in the k-tuple g = (g1, . . . , gk). In particular the integer k is a multiple of [G : H]. (3) The cohomology class α ∈ H 2(H, F ∗) is G-invariant. Here G acts on H 2(H, F ∗) via conjugation on H and trivially on F ∗. The following special cases are of interest. Corollary 1.9. Let A be a G-simple algebra with presentation PA = (H, α, g). Suppose H is normal and H-cosets representatives are equally represented in g. If α is trivial then A is strongly verbally prime. Corollary 1.10. Suppose G is finite abelian and let A be a G-simple algebra with presentation PA = (H, α, g). Then A is strongly verbally prime if and only if the H-cosets representatives are equally represented in g. In particular we have Corollary 1.11. Let A be a finite dimensional Z2 = {e, σ}-simple algebra over F with presentation PA = (H, α, g). Then A is strongly verbally prime if and only if H ∼= Z2 or else H = {e} and g = (e, . . . , e, σ, . . . , σ) where e and σ have the same multiplicity. One of the main features of finite dimensional strongly verbally prime algebras is its intimate relation to finite dimensional G-graded division algebras, that is finite dimensional G-graded algebras whose nonzero homogeneous elements are invertible (e.g. the group algebra F G). Naturally, these algebras will be defined over fields K which are usually nonalgebraically closed. Nonetheless, for our purposes, we'll only consider fields which contain F , and hence any algebra we consider in this paper is by default an F -algebra. Let A be a finite dimensional G-simple F -algebra and suppose further that It is easy to see that if A Z(A)e = F 1A where e is the identity element of G. is a G-graded division algebra, then necessarily A is G-graded strongly verbally prime. Considerably more generally, if an algebra A admits a G-graded K-form BK (in the sense of descent ) for which every nonzero homogeneous element is in- vertible then again A is strongly verbally prime. We say that A has a G-graded K-form division algebra. Using Theorem 1.8 we obtain the second main result of the paper. Notation as above. Theorem 1.12. Let A be a finite dimensional algebra G−simple F -algebra where Z(A)e = F 1A. Then A is strongly verbally prime if and only if it has a G-graded division algebra form. That is, there exists a finite dimensional G-graded division algebra B over K where Z(B)e = K1B, and a field E which extends K (and F ) such that BE and AE are G-graded isomorphic algebras over E. Combining Theorems 1.8 and 1.12 we obtain the main corollary of the paper. VERBALLY PRIME T-IDEALS AND GRADED DIVISION ALGEBRAS 5 Corollary 1.13. Suppose F is an algebraically closed field of characteristic zero and let A be a finite dimensional G-graded F -algebra. Then A admits a G-graded K-form division algebra B, some field K, if and only if A is G-simple whose pre- sentation PA = (H, α, g = (g1, . . . , gk)) satisfies conditions (1) − (3) in Theorem 1.8. Once again, it is immediate that if A is finite dimensional and has a G-graded K-form division algebra then A is strongly verbally prime (i.e. if f and g are G- homogeneous polynomials on disjoint sets of variables such that f g ∈ IdG(A) then either f or g is in IdG(A)). However, it turns out that in that case one can remove the condition on the disjointness of sets of the variables. In fact we have Theorem 1.14. Notation as above. Let A be a G−graded F -algebra (possibly infi- nite dimensional). Then A is G−PI equivalent to a finite dimensional G−division algebra over some field K if and only if for any two G-homogeneous polynomials f and g such that f g ∈ IdG(A) then either f ∈ IdG(A) or g ∈ IdG(A). Conse- quently, any of these conditions holds if and only if A is G−strongly verbally prime and G−PI equivalent to a finite dimensional algebra. Here is the outline of the paper. In section §2 we introduce some basic results and notation in graded PI theory. As mentioned earlier, Theorem 1.6 appears in [7] although there is some confusion with the different definitions of verbally primeness. For completeness of this article we devote a short section (section §3) for its proof. In section §4 we discuss the strongly verbally prime condition and show the conditions of Theorem 1.8 are necessary. In the same section we discuss also their sufficiency however the proof of Theorem 1.8 is completed only in section §5 after we discuss G-graded division algebra twisted forms and prove Theorems 1.12 and 1.14. 2. Preliminaries 2.1. G−graded PI. We fix for the rest of the article a finite group G and an algebraically closed field of characteristic zero F . We are considering algebras A having a G−graded structure, that is A = ⊕g∈GAg, where Ag is an F −vector space contained in A, and AgAh ⊆ Agh for every g, h ∈ G. An element in the set ∪g∈GAg is said to be homogenous. Sometimes we will say that a ∈ Ag is an element of degree g and denote dega = g. A G−graded polynomial is an element of the noncommutative free algebra F hXGi generated by the variables of XG = {xi,gi ∈ N, g ∈ G}. It is clear that F hXGi is itself G−graded, where F hXGig is the span of all monomials xi1,g1 · · · xin,gn , where g1 · · · gn = g. For elements in F hXGi we adopt the following terminology. A G-graded poly- nomial p is said to be G-homogeneous if it belongs to F hXGig for some g ∈ G. if any variable xig appears the same Moreover, we say p is multihomogeneous number of times in each monomial of p. More restrictive, we say p is multilinear if it is multihomogeneous and any variable xig appears at most once in each monomial of p. Note that unless G is abelian, multihomogeneous or multilinear polynomials are not necessarily G-homogeneous. We say that a G−graded polynomial f (xi1,g1 , . . . , xin,gn ) is a G−graded identity of a G−graded algebra A if f (a1, . . . , an) = 0 for all a1 ∈ Ag1 , . . . , an ∈ Agn . The set of all such polynomials is denoted by IdG(A) and it is evident that it is an ideal VERBALLY PRIME T-IDEALS AND GRADED DIVISION ALGEBRAS 6 of F hXGi which is G-graded. In fact it is a G−graded T-ideal, that is, an ideal which is closed under G−graded endomorphisms of F hXGi. It is often convenient to view any ungraded polynomial as G−graded via the embedding F hXi → F hXGi , sense that if f ∈ Id(A) (the T −ideal of ungraded identities of A) then f ∈ IdG(A). As in the ungraded case, the set of G−homogeneous multilinear identities of A where xi →Pg∈G xi,g. This identification respects the G−graded identities in the  Xσ∈Λn,g ⊆Sn ασxiσ(1) ,gσ(1) · · · xiσ(n),gσ(n) ∈ IdG(A) g ∈ G, ασ ∈ F, i1, . . . , in are all distinct T (G−)generates IdG(A). Here, Λn,g denotes the set of permutations which yield monomials of G-degree g. Let f ∈ F hXGi, A a G−graded algebra and S ⊆ ∪g∈GAg. We write f (S) for the subset of A consisting of all graded evaluations of f on elements of S. More generally, if S ⊆ A (i.e. not necessarily homogeneous elements) we set f (S) := f (S∩(∪g∈GAg)). Note that if f is G-graded multilinear then f is a graded identity of A if and only if f (S) = 0 whenever S contains a basis of A consisting of homogeneous elements. 2.2. Grassmann envelope. Let V be a countably generated F −vector space. Denote by T (V ) the unital tensor F −algebra of V . That is T (V ) = F ⊕ V ⊗k ∞Mk=1 with the obvious addition and multiplication. The unital Grassmann algebra E is defined to be E = T (V )/hT0(V )2i, where hSi is the ideal generated by S and T0(V ) = V ⊗k. ∞Mk=1 In other words, if e1, e2, . . . is a basis of V , we may see E as the unital F −algebra generated by 1, e1, e2, ... and satisfies the relations eiej = −ejei (so e2 i = 0). In fact, the elements ai1,..,ik := ei1 · · · eik , where k = 0, 1, ... and i1 < i2 < · · · < ik ∈ N, form a basis of E (for k = 0 we just mean a = 1). It is easy to see that E is C2 = Z/2Z−graded: E = E0 ⊕E1, where E0 is spanned by elements ai1,...,ik where k is even and E1 is spanned by elements ai1,...,ik where k is odd. Moreover, it is easy to check that E0 is the center of E. Suppose that A is a G2 = C2 × G−graded algebra. We denote by E(A), the Grassmann envelope of A, E(A) =Mg∈G(cid:0)A(0,g) ⊗ E0(cid:1) ⊕(cid:0)A(1,g) ⊗ E1(cid:1) . which may be viewed as a G-graded algebra and also as a G2−graded algebra. A basic property of the Grassmann envelope is the following (proof is omitted). Proposition 2.1. IdG2(A) ⊆ IdG2(B) ⇐⇒ IdG2 (E(A)) ⊆ IdG2(E(B)). VERBALLY PRIME T-IDEALS AND GRADED DIVISION ALGEBRAS 7 3. G−verbally prime T -ideals In this section we prove Theorem 1.6. In order to proceed, let us recall the G-graded analog of Kemer's Representability theorem (see [1]). As usual, all algebras are defined over F , an algebraically closed field of characteristic zero. Theorem (A). The following hold. (1) Let W be an affine G−graded algebra which is PI as an ungraded alge- bra. Then W is G−graded PI equivalent to a finite dimensional G−graded algebra A. (2) Let W be a G−graded algebra which is PI as an ungraded algebra, then W is G−graded PI equivalent to the Grassmann envelope of an affine G2 = C2 × G−graded algebra W0. Thus, by part 1, W is G−graded PI equivalent to the Grassmann envelope of a finite dimensional G2−graded algebra A. We start with the proof of Theorem 1.6. Lemma 3.1. Let A = A1 × · · · × Ak be a direct product of G−graded algebras. If A is verbally prime then IdG(A) = IdG(Ai0 ) for some i0. Moreover, if A = A1 × · · · × Ak is a product of G2−graded algebras such that E(A) is G−verbally prime, then IdG(E(A)) = IdG(E(Ai0 )) for some i0. Proof. If IdG(A1) is contained in IdG(A), then IdG(A) = IdG(A1) and we are done. Moreover, if IdG(A′ = A2 × · · · × Ak) ⊆ IdG(A), then IdG(A′) = IdG(A) and so if the theorem holds for A′ it surely holds for A. Therefore, by induction we may assume that k > 1 and IdG(A1), IdG(A′) * IdG(A) and IdG(A1)IdG(A′) ⊆ IdG(A1) ∩ IdG(A′) = IdG(A). This contradicts A being verbally prime, so k = 1 as desired. The second part follows from the first part and from the equality E(A) = E(A1) × · · · × E(Ak). (cid:3) The following definition is nonstandard. It is used only in this section. Definition 3.2. A finite dimensional algebra A over F is said to be is [G : G]− minimal (or just G-minimal) if it is G-graded and for any finite dimensional G- graded algebra B we have IdG(A) = IdG(B) =⇒ dimF A ≤ dimF B. We say that a finite dimensional algebra A is [G2 : G]− minimal if it is G2-graded and for any finite dimensional G2-graded algebra B we have IdG(E(A)) = IdG(E(B)) =⇒ dimF A ≤ dimF B. Clearly, any finite dimensional G-graded algebra is G-PI equivalent to some [G : G]-minimal algebra. Similarly, the Grassmann envelope E(A) of any finite dimensional G2−graded algebra is G-graded PI equivalent to E(B) for some [G2 : G]−minimal algebra B. It is therefore sufficient to characterize T -ideals by minimal algebras. Lemma 3.3. The following holds. VERBALLY PRIME T-IDEALS AND GRADED DIVISION ALGEBRAS 8 (1) If A is G−minimal and A is G−verbally prime, then A is semisimple. (2) If A is [G2 : G]-minimal and E(A) is G−verbally prime, then A is semisim- ple. Proof. We show the Jacobson radical J of A is zero. Since J is G-graded (resp. G2−graded)(see [9]), the algebra Ass = A/J is semisimple G-graded (resp. G2−graded). Suppose J 6= 0. Then, by minimality of A, we have IdG(A) IdG(Ass) (resp. IdG(E(A)) IdG(E(Ass))). Let p be any multilinear polynomial in IdG(Ass) \ IdG(A) (resp. IdG(E(Ass)) \ IdG(E(A))). If we consider evaluations on the span- ning set Ass ∪ J (resp. E(Ass) ∪ E(J)), we see that these evaluations are nonzero only if at least one of its variables is evaluated on an element of J (resp. E(J)). Thus, p(A) ⊆ J (resp. p(E(A)) ⊆ E(J)). Since J (resp. E(J)) is a nilpotent ideal, there is some integer 2 ≤ m ≤ the nilpotency index of J (resp. E(J)), such that p(X)p(X) · · · p(X) zgp(Y ) ∈ /∈ IdG(A) (resp. IdG(E(A))) but p(X)p(X) · · · p(X) m−1 {z } IdG(A) (resp. IdG(E(A))). m−1 {z } The following lemma shows that A (resp. E(A)) is not verbally prime. Contra- (cid:3) diction. Lemma 3.4. A G−graded algebra A is not verbally prime if and only if the fol- lowing holds: There are two G−graded polynomials p(X), q(Y ) defined on disjoint sets of variables which are not in IdG(A), but pzgq ∈ IdG(A) = Γ for every g ∈ G. Equivalently, A is G-graded verbally prime if and only if for any two polynomials p and q defined on disjoint sets of variables, nonidentities of A, the product pzgq is not a G-graded identity for some g ∈ G. Here zg is an extra variable, g-homogeneous, not in X ∪ Y . Proof. Suppose A is verbally prime and take two polynomials p(X), q(Y ), with disjoint sets of variables, such that pzgq ∈ Γ for every g ∈ G. We show p or q is in Γ. Since every element of hpiT hqiT is T −generated by pzgq, where g runs over the elements of G, we get that hpiT hqiT ⊆ Γ and the result follows by definition of verbally prime. Now suppose that A is not verbally prime and let I, J be two G−graded T −ideals not contained in Γ such that IJ ⊆ Γ. Choose p ∈ I − Γ and q ∈ J − Γ. Since pzg ∈ I, for every g ∈ G, we have (by our assumption) that pzgq ∈ Γ. But clearly, we may assume that p is defined on a set of variables disjoint from the ones of q and zg and so we are done. (cid:3) We can now prove Theorem 1.6. Proof. Suppose Γ contains a Capelli polynomial (or equivalently, Γ is the T -ideal of an affine G-graded algebra which is ungraded PI (see[1])). Then by part 1 of Theorem A, Γ = IdG(A), where A is finite dimensional G−graded algebra. Assume also that A is G−minimal. By Lemma 3.3 we obtain that A is semisimple. Since every semisimple G−graded algebra is a direct product of G−simple algebras, invoking Lemma 3.1 we can replace A by a G−simple algebra. In the general case, by part 2 of Theorem A, Γ = IdG(E(A)), where A is finite dimensional G2−graded algebra. We assume that A is [G2 : G]−minimal and continue the proof as above. This proves one direction of each one of the statements in the theorem. For the converse, suppose A is a finite dimensional G−graded simple algebra. Since the evaluation of any T −ideal I on A is a G−graded ideal I of A, in case VERBALLY PRIME T-IDEALS AND GRADED DIVISION ALGEBRAS 9 verbally prime. I * Γ we get I = A. So if I, J * Γ we have fIJ = I J = A, so IJ * Γ. Thus, A is Finally, we prove that if A is a G2−simple finite dimensional algebra, then E(A) is G−verbally prime. Suppose I is a G−graded T −ideal not contained in Γ = IdG(E(A)). Denote IA = {a ∈ Ab ⊗ a = f (x1, . . . , xn), b 6= 0, f ∈ I ⊆ F hXGi} Here, f is multilinear and G-homogeneous in F hXGi and xi ∈ E(A) is of the form ei1ei2 · · · eik ⊗ ai. Note that SpanF (IA) is a nonzero G2−graded ideal of A. Thus SpanF (IA) = A. Hence, for any two G−graded T −ideals I, J not contained in Γ, we have SpanF (IA)SpanF (JA) = A and hence I(A)J(A) 6= 0 which says that IJ is not contained in Γ. This completes the classification of G−graded verbally prime ideals. (cid:3) 4. strongly verbally prime ideals As mentioned in the introduction, we shall give a complete characterization of G-graded strongly verbally prime ideals Γ in case Γ is the T -ideal of an affine G-graded algebra which is ungraded PI. The first step is to show Corollary 4.1. If Γ is G-graded strongly verbally prime, then it is G-graded verbally prime. Proof. Suppose not and consider p, q as in the second part of the proof of Lemma 3.4. Since q /∈ Γ, pzg ∈ Γ for every g ∈ G. Similarly, since p /∈ Γ, zg ∈ Γ for all g ∈ G. We get that Γ is equal to the whole free algebra and so p (and q) are in Γ. Contradiction. (cid:3) In case G is trivial it is not hard to prove (see [10], Lemma 1.3.10) that a vector space T −generated by a single polynomial p is in fact a Lie T −ideal, so any one- sided T −ideal must be a two-sided T −ideal. Hence, pze is in the left T −ideal hpiT , so pzeq ∈ hpqiT . In other words pzeq is an identity if pq is such. Thus, in the ungraded case the opposite of Corollary 4.1 holds. In contrast to the ungraded case, we shall see that if the group G is nontrivial then there exist algebras which are verbally prime but not strongly verbally prime. For instance, it follows easily from our classification that the algebra of 3 × 3−matrices over F with the elementary Z2−grading given by the vector (e, e, σ) where Z2 = hσi is verbally prime but not strongly verbally prime. Before embarking into the proof of Theorem 1.8 we need some more preparations. Let Γ be a G-graded strongly verbally prime T -ideal which contains a Capelli polynomial. By Kemer's theory for G-graded algebras (see [1]) we have that Γ is the T -ideal of identities of a G-graded finite dimensional algebra A over F . Furthermore, by Theorem 1.6 and the fact that strongly verbally primeness implies verbally primeness (Corollary 4.1) we may assume the algebra A is G-simple. Our goal in the first part of the proof is to find necessary conditions on presentations of the G-simple algebra A so that IdG(A) is G-graded strongly verbally prime. The word the in the last sentence needs justification. It is trivial to construct nonisomorphic finite dimensional algebras which are PI equivalent. On the other hand, PI equivalent finite dimensional simple algebras over an algebraically closed VERBALLY PRIME T-IDEALS AND GRADED DIVISION ALGEBRAS 10 field F are F −isomorphic by e.g. the Amitsur-Levitzki theorem. This is true also for finite dimensional G-simple algebras over F . Indeed, it was proved by Koshlukov and Zaicev that if finite dimensional G-simple algebras A and B, G abelian, are G-graded PI equivalent, then they are G-graded isomorphic [13]. This result was extended to arbitrary groups by the first named author of this article and Haile [3]. In particular, note that if A is G -graded verbally prime, the minimal algebra B with IdG(B) = IdG(A) is uniquely determined up to a G-graded isomorphism. Indeed, as noted in the proof of Theorem 1.6 any G-minimal algebra and G-verbally prime is necessary G-simple. We record this in the following Proposition 4.2. G-minimal, G-graded verbally prime algebras B and B′ are G- graded PI equivalent if and only if they are G-graded isomorphic. As mentioned in the introduction, the G-graded structure of a finite dimensional G-simple algebra is presented by a triple (H, [c], g). We recall that a G-graded alge- bra may admit different presentations, that is algebras with presentations (H, [c], g) and (H ′, [c′], g′) (different parameters) may be G-graded isomorphic. In fact, in [3] the authors present three type of moves on presentations, M1, M2, M3 (see below) such that (1) Mi((H, [c], g)) ⋍ G (H ′, [c′], g′) if and only if (H ′, [c′], g′) may be obtained from (2) (H, [c], g)) ⋍ G (H, [c], g), i = 1, 2, 3 (G−isomorphic) (H, [c], g) by composing a sequence of such moves. For the reader convenience let us recall the moves M1, M2, M3. M1 : g′ = (gσ(1), . . . , gσ(m)), where σ ∈ Sm, and H = H ′, c = c′. M2 : g′ = (h1g1, . . . , hmgm), where h1, . . . , hm ∈ H, and H = H ′, c = c′. M3 : For g ∈ G. g′ = (gg1, . . . , ggm), H ′ = gHg−1 and c′ ∈ Z 2(H ′, F ∗) is defined by c′(gh1g−1, gh2g−1) = c(h1, h2). Applying these moves we obtain easily the following normalization on the pre- sentation of a G-simple algebra (see also [3]). Corollary 4.3. Any finite dimensional G-simple algebra has a presentation with ) ∈ Gm, where the gi's are different representatives of g = (g1, . . . , g1 , . . . , gk, . . . , gk Let PA = (H, [c], g) be a presentation of a G-simple algebra A. Suppose the presentation is normalized as above. We denote by e(i,j) r,s the element em0+···mi−1+r,m0+···mj−1+s ∈ Mm(F ) where 1 ≤ i, j ≤ k, m0 = 0 and r ∈ {1, . . . , mi}, s ∈ {1, . . . , mj). In case i = j we slightly simplify the notation and replace e(i,i) r,s by e(i) r,s. For a fixed i we refer to the linear span of {ue ⊗ e(i) r,s}r,s as the ith e−block of A. Note that the linear span of all e−blocks (i = 1, . . . , k) is equal to Ae, the e−component of A. Suppose A is G-graded strongly verbally prime. Notation as above. Lemma 4.4. The following hold. (1) k = [G : H] m1 {z } mk {z } right H−cosets, 0 < m1 ≤ · · · ≤ mk and g1 = e. VERBALLY PRIME T-IDEALS AND GRADED DIVISION ALGEBRAS 11 (2) m1 = mk and so m1 = · · · = mk Proof. Let us show first all right H-cosets are represented in g. If not, it fol- lows from Lemma 3.3 in [2] that the G-grading on A is degenerate (recall that a G-grading on an algebra A is nondegenerate if the product of homogeneous compo- nents Ag1 Ag2 · · · Ags is nonzero, for any s > 0 and any s-tuple (g1. . . . , gs) ∈ G(s); equivalently, the T -ideal of G-graded identities has no G-graded multilinear mono- mials). It follows that IdG(A) contains multilinear monomials. Since the grading is assumed to be connected, we have monomials of the form xg1 xg2 · · · xgn ∈ IdG(A) where gi ∈ SuppG(A) and hence xgi /∈ IdG(A). Clearly, this shows that A is not G-graded strongly verbally prime and the first part of the lemma is proved. Next, suppose by contradiction that m1 < mk (mi are positive integers). For i = 1, . . . , k consider the following monomial of e-variables Zi = y(i) 0 x(i) 1 y(i) 1 · · · x(i) m2 i y(i) m2 i and denote Z = Z1w1Z2 · · · wk−1Zk, where wi is a g−1 i gi+1−graded variable. Finally let p = AltX Z, where X is the set of all x variables. We refer to the y variables as frames and the w variables as bridges. Let us show that p is a nonidentity of A. It is well known that one can find an ordering on the elementary matrices er1,s1, . . . , erm2 i of Mmi(F ) such that ,sm2 i = 1. i i − 1. 0 = x(i) 1 := x(i) if t = 0. This yields Zi = ue ⊗ e(i) (1) r1 = sm2 (2) st = rt+1 for t = 1, . . . , m2 Evaluate x(i) t = ue ⊗ e(i) t 's by e-homogeneous elements such that Zi 6= 0, namely y(i) rt,st. Notice that there is a unique evaluation of the y(i) st,st if t > 0 t and y(i) 1,1. Finally, we set wi = ue ⊗ e(i,i+1) and obtain Z = ue ⊗ e(1,k) 1,1 . We claim that any nontrivial alternation of the x's gives zero. Indeed, if the alternation sends an x corresponding to Zi to x′ corresponding to Zj (i 6= j), we get zero since the frames of Zj are evaluated on elements of the j−block and these do not match with the indices of x. On the other hand if the alternation sends some x of Zi to a different x′ of Zi we also get zero since x and x′ are evaluated on different elementary matrices and again the indices of the frames do not match. All in all, we get that p = Z 6= 0. t = ue ⊗ e(i) 1,1 Next, we claim that p(A) ⊆ F cH ⊗ R, where R is spanned by e(i,j) r,s , where mi = m1 and mj = mk (in particular i < j). Indeed, since p is multilinear it is enough to consider only evaluations on elements of the form uh ⊗ ei,j. Consider such an evaluation: x, y, w. Obviously, we may assume that p 6= 0 and so without loss of generality we may assume that Z 6= 0. Note that since the x, y (alternating variables and frame variables) have been chosen to be homogeneous of degree e, all variables of each Zi must be evaluated on elements of the same e−block. Moreover, since the x variables are alternating in p, their evaluation must consist of linearly independent elements, thus consists of a basis of Ae. Thus, the x variables of each Zi are evaluated on a basis of some e−block of Ae where for i 6= j the evaluation is on different e-blocks. It follows that each Zi must be evaluated on elements of an e−block of dimension m2 i . Consequently Z(A) ⊆ F cH ⊗ R (evaluations of the monomial Z on A). In order VERBALLY PRIME T-IDEALS AND GRADED DIVISION ALGEBRAS 12 to show that p(A) ⊆ F cH ⊗ R we proceed as follows. Any evaluated monomial fZ ′ resulting from a permutation of the evaluated x variables of Z is in fact equal to Z after reevaluation of the x variables. The previous argument shows that Z ′(A) ⊆ F cH ⊗ R and hence also p(A) ⊆ F cH ⊗ R. Remark 4.5. In fact we can strengthen the statement in the last part of the proof above. We claim that if the following conditions hold (1) All variables are evaluated on elements of A of the form uh ⊗ ei,j (2) The variables of Zi are homogeneous of degree e. (3) Z(A) 6= 0 then for any monomial Z ′ in p obtained from Z by a nontrivial permutation of the the polynomial p (i.e. an evaluation of the x's, the y's and the w's) which yields a nonzero value of the monomial Z, we have that the value of the x's is determined by the value of the y's (frames). Since Z ′ is obtained from Z by permuting nontrivially x-variables we havefZ ′ = 0. Indeed, as noted above, for any nonzero evaluation of the x's and keeping the y's, its valuefZ ′ must be zero. We complete now the proof of the lemma. Since m1 < mk, we have (F cH ⊗ R)2 = 0. It follows that pq ∈ IdG(A) where q is the polynomial p but with a disjoint set of variables - contradiction to A being strongly verbally prime. We let gset = {g1, . . . , gk} be a set of representatives of the right cosets of H in G. Denote mblock = m1 = · · · = mk. Proposition 4.6. gset = {g1, . . . , gk} normalizes H (i.e. H is normal in G). Proof. First let us define an equivalence relation ∼ on gset: gi ∼ gj iff g−1 i Hgj ∩ H 6= ∅, (cid:3) and denote by [g] the equivalence class of g ∈ gset. Clearly, the equivalence rela- tion among the gi's determines an equivalence relation among the corresponding e-blocks. Notice that the claim of the proposition holds if and only if the equivalence classes are of cardinality 1. Let g∼ = {γ1, . . . , γk} ⊆ gset, k ≤ k, be a full set of representatives of ∼. Changing representatives we may write {z } {z } m1 {z } mk {z } ([g1], . . . , [g1] , . . . , [gk], . . . , [gk] ) = ([γ1], . . . , [γ1] , . . . , [γk], . . . , [γk] ), mblock mblock where mi = mblock ·[γi]. Here [γi] denotes the cardinality of the equivalence class in gset represented by γi. Moreover, since e ∈ G normalizes H, we may assume (1) [γ1] = [e](= {e}). (2) 0 < m1 ≤ · · · ≤ mk. We make similar conventions to the ones we made before Lemma 4.4. Denote the element e m0+··· mi−1+r, m0+··· mj−1+s ∈ Mm(F ), where m0 = 0 and is in the r,s = uh ⊗ e[γi,γi] by e[i,j] r,s r ∈ {1, . . . , mi}, s ∈ {1, . . . , mj}, 1 ≤ i, j ≤ k. We say that uh ⊗ e[i,j] r,s [i, j]−block. Moreover, we denote by uh ⊗ e[i] and call its block the [i]−block. r,s the element uh ⊗ e[i,i] r,s VERBALLY PRIME T-IDEALS AND GRADED DIVISION ALGEBRAS 13 The idea of the proof is somewhat similar to the one above. We construct a "Capelli type" polynomial p(X), obtained by alternating some variables of a suitable G-graded monomial Z. For i = 1, . . . , k, consider the monomial of e-variables Zi = y(i) 0 x(i) 1 y(i) 1 · · · x(i) m2 i y(i) m2 i where mi = mblock and let Z = Z1w1Z2 · · · wk−1Zk, where wj is a g−1 j hjgj+1−graded variable and hj ∈ H to be chosen, j = 1, . . . , k −1. If gj and gj+1 are nonequivalent we know that g−1 j hjgj+1 is not in H for any hj ∈ H and so we take hj = e for simplicity. On the other hand if gj and gj+1 are equivalent we choose hj = hj ∈ H such that g−1 hjgj+1 ∈ H. Note that by our construction w1, w2, . . . , w m1/mblock−1 are elements in H, w m1/mblock /∈ H, then the next m2/mblock − 1 variables w's have homogeneous degrees in H, followed by w m1+ m2/mblock whose homogeneous degree is not in H and so on. j Let us introduce an equivalence relation on submonomials: we say that sub- monomials Zi and Zj are equivalent if and only if i = j or they are separated by w's whose homogeneous degrees is in H. We denote the equivalent class of Zi by [Zi]. Note that we have constructed precisely k equivalence classes. Let p(X) be the polynomial obtained by alternating all xr's variables of Z. Lemma 4.7. The polynomial p(X) is a graded nonidentity of A. Furthermore, evaluating the polynomial p on A yields elements in F cH ⊗ R, where R is spanned by elements e[i,j] r,s with i and j satisfying mi = m1 and mj = mk. Using the lemma, since [F cH ⊗ R]2 = 0, by taking the product of two such polynomials as p(X) with disjoint sets of variables, we see that A is not strongly verbally prime and so the proof of the proposition is complete. Note that with this we have proved the necessity of conditions 1 and 2 in Theorem 1.8. Proof. (of the lemma) By the construction the monomial Z has a nonzero value on A as follows. We evaluate the x's in each segment Zi of the monomial Z on r,s, the y's (frames) on e-elements of the form ue ⊗ e[i] e-elements of the form ue ⊗ e[i] r,r and finally the variables wi (bridges), i = 1, . . . , k − 1, on uhi ⊗ e[i,i+1] where hi = e if gi and gi+1 are nonequivalent and hi = hi if gi and gi+1 are equivalent. Other monomials of p(X) obtained from Z by alternating the x's yield a zero value since the values on the y's and the x's do not match. This shows p(X) is a nonidentity of A. 1,1 In order to prove the second claim of the lemma, note that by the multilinearity of p(X) it is sufficient to check evaluations on basis elements of the form uh ⊗ e[i,j] r,s . From now on we fix an evaluation of p(X) of that kind. We'll show the value is in F cH ⊗ R. Clearly, we may assume the value is nonzero and so there exists a monomial Z whose value is nonzero. In that case, by the alternation of the x's in p(X), their values must consist of the full basis of the e-homogeneous component of A. Since the y's are homogeneous of degree e, their value is determined uniquely by the x's and the nonvanishing of the monomial Z. Therefore as above, any monomial Z ′ obtained from Z by a nontrivial permutation vanishes. In order to complete the proof we must show the value of the monomial Z is in F cH ⊗R. As noted above, the VERBALLY PRIME T-IDEALS AND GRADED DIVISION ALGEBRAS 14 evaluation of the monomial Z establishes a one to one correspondence between the submonomials Zi and e-blocks. Furthermore, equivalent classes of submonomials [Zi] must be evaluated on equivalent classes of e-blocks such that the sizes of the equivalence classes coincide. The proof of the lemma now follows easily and so the proof of Proposition 4.6 is now complete. (cid:3) (cid:3) We pause for a moment and summarize what we have so far. We showed that if Γ is a G-graded T -ideal which contains a Capelli set and is G-graded strongly verbally prime then it is the T -ideal of identities of a finite dimensional G-simple algebra. Of course, there may be other algebras (finite dimensional or not) whose T -ideal of G-graded identities is Γ. However, if we restrict our attention to finite dimensional G-simple algebras A whose IdG(A) = Γ, the algebra A is uniquely determined up to a G-graded isomorphism. We know by Bahturin-Sehgal-Zaicev's theorem that any G-simple algebra and hence the algebra A, is represented by a triple (H, [c], g), where H is a subgroup of G, [c] ∈ H 2(H, F ∗) and g = (g1, . . . , gm) ∈ Gm. We know the presentation (H, [c], g) of A is unique up to 3 type of "moves" M1, M2, M3. We showed that the strongly verbally prime condition on A implies the group H is normal in G and H cosets of G are represented equally in g = (g1, . . . , gm) ∈ Gm. In particular m/[G : H] is an integer. The above considerations lead to the following characterization of G-simple crossed products. We recall Definition 4.8. A G-graded algebra B is said to be a G-crossed product if Bg has an invertible element for any g ∈ G. Note that this implies dimF (Bg) = dimF (Bg′ ) for any g, g′ ∈ G. Proposition 4.9. Let A be a G−simple algebra over an algebraically closed field of characteristic zero. Then A is a G-crossed product if and only if A has a pre- sentation (H, [c], g) where H is a subgroup of G (not necessarily normal) and the right H cosets of G are equally represented in g = (g1, . . . , gm) ∈ Gm. Proof. Suppose A has a presentation as above where right H-cosets are equally represented. We know the elements of the form uh⊗ei,j ⊗E, where i, j ∈ {1, . . . , [G : H]}, E ∈ Mr(F ) are homogeneous of degree g−1 i hgj and span A. We note the homogeneous components are concentrated in block permutation matrices which means that for any g ∈ G and any i ∈ {1, . . . , [G : H]} there exist a unique h = h(i) and unique j = j(i) such that for every E ∈ Mr(F ) the element uh ⊗ ei,j ⊗ E has homogeneous degree g. It follows that the element Xi∈{1,...,[G:H]} uh(i) ⊗ ei,j(i) ⊗ Ir, where Ir is the identity matrix of rank r, is invertible and homogeneous of degree g. The inverse is Xi∈{1,...,[G:H]} c−1(h(i), h(i)−1)uh(i)−1 ⊗ ej(i),i ⊗ Ir. As noted above, in this case the dimensions of the homogeneous components of A coincide and equal [G : H]r2 = dimF (A)/ G . VERBALLY PRIME T-IDEALS AND GRADED DIVISION ALGEBRAS 15 For the converse let r1, . . . , r[G:H] be the multiplicities of the right H-cosets in g and suppose are not all equal. Clearly r1 +· · ·+r[G:H] = m. Moreover, by definition of the G-grading on A, dimF (Ae) = r2 [G:H] which strictly exceeds dimF (A)/ G . This implies the dimension of the homogeneous components are not equal and hence A is not a G-crossed product. (cid:3) 2 + · · · + r2 1 + r2 In particular we have Corollary 4.10. Let A be a finite dimensional G-simple algebra with presenta- tion (H, [c], g). Suppose H is normal and multiplicities of H-coset representatives are equal (= r). Then the homogeneous components consist of block permutation matrices. Our next step is to put restrictions (i.e. necessary conditions) on the cohomology class [c] provided the G-simple algebra A is strongly verbally prime. This will lead to a necessary condition on [c] which together with the conditions on H and on g mentioned above turns out to be also sufficient for strongly verbally primeness. In order to proceed we recall briefly the theory and general terminology which relate the cohomology group H 2(H, F ∗) with H-graded PI theory (see [4], [3]). Let H be a finite, c : H × H → F ∗ a 2-cocycle and [c] the corresponding cohomology class. Let F cH be the corresponding twisted group algebra and let {uh}h∈H be a set of representatives for the H-homogeneous components in F cH. We have uh1uh2 = c(h1, h2)uh1h2 for hi ∈ H. Let n be a positive integer. For h = (h1, . . . , hn) ∈ H n and σ ∈ Sn consider Zh,σ = xi.hi − nYi=1 xσ(i),hσ(i) , nYi=1 c(h) c(hσ) αh,σ {z} where c(h) ∈ F ∗ (slightly abusing notation) is the element determined by uh1 · · · uhn = c(h)uh1···hn and hσ = (hσ(1), . . . , hσ(n)). We are interested in the binomials Zh,σ which are H-graded identities of F cH. This happens precisely when h1 · · · hn = hσ(1) · · · hσ(n). Note that if h1 · · · hn = hσ(1) · · · hσ(n), the polynomialQn is an H-graded identity if and only if λ = αh,σ. i=1 xi.hi −λ·Qn We refer to these binomials as the binomial identities of the twisted group algebra F cH. It is known that the binomial identities generate the T -ideal of H-graded identities of F cH (see [3] ,Theorem 4). Furthermore, one sees easily that the scalars αh,σ depend on the cohomology class [c] but not on the representative c. i=1 xσ(i),hσ(i) Next it is convenient to introduce into our discussion presentations of the group H. Let F be the free group generated by Ω = {xi,h i ∈ N , h ∈ H} and consider the exact sequence 1 / R / F / H / 1 where the map F → H is given by xi,h 7→ h. This induces the central extension 1 / R/[F , R] / F /[F , R] / H / 1 . Consider the central extension 1 / F ∗ / Γ / H / 1 / / / / / / / / / / / / VERBALLY PRIME T-IDEALS AND GRADED DIVISION ALGEBRAS 16 which corresponds to the cocycle c (i.e. {uh h ∈ H} is a set of representatives in Γ and uh1uh2 = c(h1, h2)uh1h2). The map ψc : F /[F , R] → Γ, given by ψc(xi,h) = uh, induces a map of extensions and its restriction ψc : [F , F ]∩R/[F , R] = M (H) → F ∗ is independent of the presentation and corresponds to [c]. That is, Φ([c]) = ψc where Φ : H 2(G, F ∗) → HomZ(M (H), F ∗) is the well known isomorphism of the Universal Coefficient Theorem. Here, M (H) is the Schur multiplier of the group H and the equality with [F , F ] ∩ R/[F , R] is the well known Hopf formula. It follows that Φ([c])x1,h1 · · · xn,hn (xσ(1),hσ(1) · · · xσ(n),hσ(n) )−1  = uh1 · · · uhn(uhσ(1) · · · uhσ(n) )−1 } {z z = c(h) c(hσ) = αh,σ Now, it is clear that z ∈ [F , F ] and moreover, by the condition we imposed on h and σ ∈ Sn, z ∈ R and hence z ∈ R ∩ [F , F ]. It is not difficult to show that modulo [F , R] all elements of M (H) are of this form (see [4], Lemma 11), and so, we readily see that the binomial identities of the twisted group algebra F cH determine uniquely the cohomology class [c]. We are ready to prove the necessity of the third condition in Theorem 1.8 for A to be strongly verbally prime, namely the cohomology class [c] is G-invariant. Recall that G acts on H 2(H, F ∗) via conjugation, that is for [β] ∈ H 2(H, F ∗) (g · β) (h1, h2) = β(gh1g−1, gh2g−1). Applying the isomorphism Φ one checks the cohomology class [c] is G−invariant if and only if (4.1) ∀g ∈ G : αh,σ = αghg−1,σ, for all αh,σ corresponding to the binomial identities {Zh,σ}. Lemma 4.11. Suppose A is strongly verbally prime with presentation (H, [c], g), g = ((g1, . . . , g1), (g2, . . . , g2), . . . , (gk, . . . , gk)) ∈ (Gmblock )k ≃ Gm. Then the class [c] is invariant under the action of G, i.e. 4.1 holds. Proof. First note that since 4.1 holds for g ∈ H (indeed, it is well known that the inner action of H induces a trivial action on H 2(H, F ∗)) it is sufficient to check the invariance of [c] for the action of elements from gset (a full set of H-coset representatives in G). Suppose this is not the case. Note that by using elementary moves on the presentation of A we may assume the tuple g is ordered so that (1) Equal cosets representatives are adjacent. (2) Elements gi which normalize the cohomology class [c] "stand on the left of the tuple" whereas gj's which do not normalize "stand on the right", that is there is 1 ≤ t0 < k such that gj normalizes [c] if 1 ≤ j ≤ t0 and gj does not normalize [c] if t0 < j ≤ k. Note that since e normalizes [c] we may assume g1 = e whereas (by our ordering) gk does not normalize [c]. Now, we know that the set of Binomial Identities of F γH determines the co- homology class γ (indeed, the Binomial identities generate the T -ideal of graded VERBALLY PRIME T-IDEALS AND GRADED DIVISION ALGEBRAS 17 identities and the latter determines γ). Moreover, if α and γ are different coho- mology classes of H 2(H, F ∗) we have IdH (F γH) * IdH (F αH). Let us fix for any such pair (α, γ) a binomial polynomial B(¯α, γ) which is an H-graded identity of F γH but a nonidentity of F αH. Furthermore, for any β ∈ H 2(H, F ∗) we let R(β) = Qγ6=β(B( ¯β, γ))), where the binomial polynomials are taken with disjoint sets of variables implying R(β) is multilinear. Note that since any twisted group algebra is a H-graded division algebra the polynomial R(β) is a nonidentity of F βH but an H-identity of F γH for any γ 6= β. block block block block , Ym2 , Ym2 Now, we proceed as in [AH]. Let p(Xm2 ) denote the central Regev polynomial on 2m2 block variables. Fixing h ∈ H, construct an H-graded polynomial ) by replacing one of ) which is obtained from p(Xm2 ph(Xm2 its X variables, say x1, by a variable x1,h homogeneous of degree h, and all other variables xi, 2 ≤ i ≤ m2 block are replaced by variables of degree e. Finally, we replace all variable of R(β) by Regev central polynomials. More precisely, we replace any variable in R(β) of degree h by a multilinear poly- nomial ph(Xm2 ) (as usual, different variables of R(β) are substituted by multilinear polynomial with disjoint sets of variables). We obtain a multilinear polynomial which we denote by R(β, mblock). block and yj, 1 ≤ j ≤ m2 , Ym2 , Ym2 block block block block For i = 1, . . . , k, consider the monomial of e-variables Zi = y(i) 0 x(i) 1 y(i) 1 · · · x(i) m2 block y(i) m2 block and let Z = Z1w1Z2 · · · wk−1Zk, where wi is a g−1 i higi+1−graded variable and hi ∈ H to be chosen (in fact any ele- ment of H will do). For i = 1, . . . , k, we insert in Z the polynomial R(gi([c]), mblock) on the right of Zi. We let Z = Z1R(g1([c]), mblock)w1Z2R(g2([c]), mblock) · · · wk−1ZkR(gk([c]), mblock). Finally, we let Θ be the multilinear polynomial obtained by alternating the set of e-variables X = {x(i) j } in Z. We claim (1) Θ is a G-graded nonidentity of A. (2) All evaluations of the G-graded polynomial Θ on A yield elements in R = 1 gi and span{F [c]H ⊗ e(i,j) g−1 k gj normalize the cohomology class [c]. Note that this is equivalent to r,s } where (i, j) satisfies the following condition: g−1 R(gi([c]), mblock) = R(g1([c]), mblock) and R(gj([c]), mblock) = R(gk([c]), mblock). (3) R2 = 0. Note that the lemma follows at once from the claim. Indeed, if we duplicate the polynomial Θ with different variables and denote them by Θ1 and Θ2, we obtain two homogeneous G-graded nonidentities with different variables whereas their product is an identity. To see the first part of the claim consider the evaluation of the monomials Zi, i = 1, . . . , k on the ith e-block of A. We then evaluate the polynomials R(gi([c]), mblock), i = 1, . . . , k, on the same block of A (the ith block). Finally, the variables wi, i = 1, . . . , k − 1 are evaluated at elements that bridge the ith and ith+1 diagonal VERBALLY PRIME T-IDEALS AND GRADED DIVISION ALGEBRAS 18 blocks. By construction the value of Z is nonzero. The polynomial Θ is obtained from Z by alternating the variables of X. Since any nontrivial permutation of these variables yields a zero value (the values of {yi j} do not match) we get a nonzero evaluation of Θ on A as desired. j} and {xi For the second part of the claim note that by the alternation the elements of X must be evaluated precisely on all e-elements of A as long as the value is nonzero. Moreover, since the variables yi j are also e-variables, the elements of Zi (and in particular the x's variables of Zi) must be evaluated on the same e-block of A, again as long as the evaluation is nonzero. But the presence of the polynomials R(gj([c]), mblock) in Θ implies the evaluation of Θ on A vanishes unless the mono- mial Zj, any j, is evaluated on a e-block whose index is j for which gj([c]) = gj([c]). In particular the monomials Z1 and Zk must be evaluated on e-blocks indexed by 1th and kth respectively. This proves part 2 of the claim. The 3rd part of the claim follows from the fact that 1 < k. This completes the proof of the claim and hence of the lemma. (cid:3) In order to complete the proof of Theorem 1.8 we need to show that if A is a finite dimensional G−simple algebra with presentation (H, [c], g), where (1) H is normal, (2) g = ((g1, . . . , g1), (g2, . . . , g2), . . . , (gk, . . . , gk)) ∈ (Gmblock )k ≃ Gm (3) [c] ∈ H 2(H, F ∗) is G invariant then A is strongly verbally prime. For the proof we shall replace the 3rd condition with an apparently stronger condition 3′ which we call the path property. After introducing that property, we will show firstly that 1 + 2 + 3′ is sufficient for strongly verbally prime and secondly that conditions 1 + 2 + 3 and 1 + 2 + 3′ are equivalent. For future reference let us state the following. Theorem 4.12. Let A be a finite dimensional G-simple algebra over F with pre- sentation PA = (H, α, g = (g1, . . . , gk)). Then condition 1 + 2 + 3 (Theorem 1.8) is equivalent to 1 + 2 + 3′ below. (1) The group H is normal in G. (2) The different cosets of H in G are represented exactly the same number of times in the k-tuple g = (g1, . . . , gk). In particular the integer k is a multiple of [G : H]. (3') The path property. Let us introduce the notation we need. Definition 4.13. (Good permutations and pure polynomials) For g ∈ G denote by g its image in G/H. We say the G−graded multilinear monomial Zσ = xσ(1),tσ(1) · · · xσ(n),tσ(n) is a good permutation of Z = x1,t1 · · · xn,tn if (1) (degZ =)t1 · · · tn = tσ(1) · · · tσ(n)(= degZσ). (2) For every 1 ≤ i ≤ n, t1 · · · ti = tσ(1) · · · tσ(σ−1(i))=i. A G-graded multilinear polynomial is said to be pure if all its monomials are good permutations of each other. Note 4.14. (1) Good permutation on monomials is an equivalence relation. VERBALLY PRIME T-IDEALS AND GRADED DIVISION ALGEBRAS 19 (2) If all variables degrees are in H, then the second condition says nothing. In that case good permutation reduces to Z, Zσ having the same G-degree. Our first step is to show that it is sufficient to consider pure polynomials when analyzing the T -ideal of G-graded identities of A. We assume the algebra A is G−simple with presentation (H, [c], g), where H is normal and g = ((g1, . . . , g1), (g2, . . . , g2), . . . , (gk, . . . , gk)) ∈ (Gmblock )k ≃ Gm. Lemma 4.15. Let p = Pσ∈Sn ασZσ be a multilinear G-graded polynomial. Let p = p1 + . . . + pq be the decomposition of p into its pure components. Then p is a G-graded identity of A if (and only if) pi is a G-graded identity of A for any i = 1, . . . , q. Proof. Since G-homogeneous components are linearly independent subspaces of A we may restrict our attention to monomials which have the same homogeneous degree so the first condition in the definition of good permutation is satisfied. We therefore let all monomials of p be homogeneous of degree g. Since the polynomial p is multilinear we can restrict evaluations to a spanning set of A and in particular to elements of the form bh,{i,j},E = uh ⊗ ei,j ⊗ E, where h ∈ H, i, j ∈ {1, . . . , [G : H]}, E ∈ Mr(F ). Note that the homogeneous degree of bh,{i,j},E is determined by uh ⊗ ei,j, namely g−1 i hgj, as the degree of E is trivial. Let xg be a variable of degree g. By our assumption, its admissible evaluations are elements bh,{i,j},E where g = g−1 i hgj. The following observations are important. (1) For any i ∈ {1, . . . , [G : H]} there exist h and j such that g = g−1 (2) h = h(i) and j = j(i) are uniquely determined. i hgj. This follows easily from the fact that {gj} is a full set of H-cosets representatives and gig ∈ ∪j Hgj. Let Z = xt1 xt2 · · · xtn . Following the above observations we choose any index i ∈ {1, . . . , [G : H]}. This determines h = h(i) and j = j(i) in any possible value of xt1 , that is xt1 is evaluated at bh(i),{i,j(i)},E, any E ∈ Mr(F ), where g−1 i h(i)gj(i) = t1 (i.e. the degree of bh(i),{i,j(i)},E is t1). It follows that if we insist on nonzero evaluations of the monomial Z, the evaluation of xt2 must be of the form bh′,{j(i),j ′},E ′ and so on. We see that the choice of i = i1 (for the evaluation of xt1 ) determines uniquely the triples (i1, j1, h1), . . . , (in, jn, hn). Note on the other hand that a monomial Z as above indeed admits nonzero evaluations with elements b1, . . . , bn where bν = bhν ,{iν ,jν },Eν for any choice i1 ∈ {1, . . . , [G : H]} (for instance any evaluation of the form where Et is invertible will do). Claim: Let Z = x1,t1 x2,t2 · · · xn,tn be a G-graded monomial as above and let (i1, j1, h1), . . . , (in, jn, hn) be n-triples such that the evaluation of Z with xtν ← bν = bhν ,{iν ,jν },Eν , ν = 1, . . . , n, is nonzero. Let Zσ = xσ(1),tσ (1)xσ(2),tσ (2) · · · xσ(n),tσ (n) be a monomial obtained by permuting the variables of Z with σ ∈ Sn. Suppose the total G- degrees of Z and Zσ coincide. Suppose also the evaluation above satisfies the following conditions (1) The evaluation does not annihilates Zσ (2) iσ(1) = i1 Then Z and Zσ are good permutations of each other. Consequently, jσ(n) = jn and eiσ(1),jσ(1) · · · eiσ(n),jσ(n) = ei1,j1 · · · ein,jn . VERBALLY PRIME T-IDEALS AND GRADED DIVISION ALGEBRAS 20 Let us complete the proof of the lemma using the claim. Suppose p1(xt1 , xt2 · · · xtn ) is a nonidentity. We need to show p(xt1 , xt2 · · · xtn ) is a nonindentity. Let xtν ← bν = bhν ,{iν ,jν },Eν , ν = 1, . . . , n be a nonzero evaluation of p1. The values of p1 are contained in one block, say block (i0, j(i0)) of the block permutation matrix which corresponds to g ∈ G. On the other hand, by the claim, the evaluations on ps, s > 1 are contained in blocks (i, j(i)), with i 6= io. This shows the evaluation of p is nonzero and we are done. Proof. (claim) Our setup implies the following: (1) j1 = i2, j2 = i3, . . . , jn−1 = in (2) jσ(1) = iσ(2), jσ(2) = iσ(3), . . . , jσ(n−1) = iσ(n) (3) i1 = iσ(1), jn = jσ(n) (4) t1 = g−1 Since we are assuming the monomials Z and Zσ have the same homogeneous degree in G, in order to prove the good permutation property we need to show the following condition holds. i2 hi2 gj2, . . . , tn = g−1 i1 hi1 gj1, t2 = g−1 hin gjn in If σ(l) = k then the values of the products of elements in G and i1 hi1 gj1 · g−1 g−1 i2 hi2 gj2 · · · g−1 ik−1 hik−1 gjk−1 g−1 iσ(1) hiσ(1) gjσ(1) · g−1 iσ(2) hiσ(2) gjσ(2) · · · g−1 iσ(l−1) hiσ(l−1) gjσ(l−1) coincide in G/H. Indeed, ignoring the elements of H and invoking the 1st and 2nd sets of equalities it is sufficient to show g−1 i1 gjk−1 = g−1 iσ(1) gjσ(l−1) . But g−1 i1 = g−1 iσ(1) by (3) and so it remains to show gjk−1 = gjσ(l−1) . By σ(l) = k, (1) and (2) above we have gjk−1 = gik = giσ(l) = gjσ(l−1) and we are done. This proves the claim and completes the proof of the lemma. (cid:3) Now we can introduce the path property. As usual we let A be a G-simple algebra with presentation (H, [c], g), where H is normal and g = ((g1, . . . , g1), (g2, . . . , g2), . . . , (gk, . . . , gk)) ∈ (Gmblock )k ≃ Gm. Let p(xt1 , . . . , xtn ) be a G-graded pure polynomial. Write (cid:3) p(xt1 , . . . , xtn ) = Z + Xe6=σ∈Λ(Z) ασZσ where Z = xt1 · · · xtn and Λ(Z) = {σ ∈ Sn : σ good permutation of Z}. Definition 4.16. (Path property). Notation as in lemma 4.15. (1) We say the polynomial p(xt1 , . . . , xtn ) satisfies the path property if the fol- lowing condition holds: If p vanishes whenever the evaluation of xt1 has the form uh(i1) ⊗ ei1,j(i1) ⊗ E for some index i1 ∈ {1, . . . , [G : H]} and E ∈ Mr(F ), then p is a G-graded identity of A. VERBALLY PRIME T-IDEALS AND GRADED DIVISION ALGEBRAS 21 (2) We say the G-simple algebra A satisfies the path property if any pure poly- nomial satisfies the path property. Roughly speaking, the condition says that if p is annihilated by choosing an "evaluation path", then it vanishes for all [G : H] paths and hence it is a G-graded identity. Proposition 4.17. Let A be a G-simple algebra with presentation (H, [c], g) as above (i.e. H is normal and g = ((g1, . . . , g1), (g2, . . . , g2), . . . , (gk, . . . , gk)) ∈ (Gmblock )k ≃ Gm.) Then A is strongly verbally prime if and only if satisfies the path property. Proof. Suppose the path property does not hold. Then there exists a g-homogeneous pure polynomial p which vanishes on all evaluations arising from one path, say i1 ∈ {1, . . . , [G : H]} but is a nonidentity of A. For any j ∈ {1, . . . , [G : H]} there exists a suitable pure polynomial pj obtained by changing the scalars coefficients of p (depending on the evaluation path j) such that pj is clearly a nonidentitity and vanishes on evaluations corresponding to the jth path (see remark below). Taking disjoint variables on different pj's we see that the product p1 · · · p[G:H] vanishes on all paths and hence is an identity whereas pj, j = 1, . . . , [G : H], is a nonidentity. This shows A is not strongly verbally prime. Remark 4.18. Since we don't really need the above direction details are omitted. For the converse suppose A satisfies the path property. We let p and q be mul- tilinear G-graded homogeneous polynomial nonidentities of A with disjoint sets of variables. We need to show p · q is a nonidentity. Claim 4.19. We may assume p and q are pure polynomials Proof. To see this, let p = p1 + · · ·+ pr and q = q1 + · · ·+ qs be the decomposition of p and q into its pure components. Observe that since p and q are defined on disjoint sets of variables, the decomposition pq =Pi,j pi · qj is precisely the decomposition of pq into its pure components. Now suppose the sufficiency in the proposition holds for pure polynomials and let p and q be multilinear homogeneous nonidentities. We show pq is a nonidentity of A. But by Lemma 4.15 this reduces to pure polynomials and the claim is proved. (cid:3) ασZσ and q(y1, . . . , ym) =Pτ ∈∆q So let p and q be multilinear, nonidentities pure polynomials where p(x1, . . . , xn) = βτ Wτ . Choose i ∈ {1, . . . , [G : H]} and consider evaluations of p of the form uh ⊗ ek,s ⊗ E corresponding to the path deter- mined by i. Since the monomials of p are good permutations of each other, for any σ ∈ ∆p there is a scalar γσ ∈ F ∗ such that all evaluations corresponding to the ith γσασZσ is a polynomial identity of Mr(F ). Now, we are assuming the algebra A satisfies the path property and hence, since p is a nonidentity, for any path, say the path deter- mined by i ∈ {1, . . . , [G : H]}, there exists an evaluations of the form uh ⊗ ek,s ⊗ E Pσ∈∆p path annihilate p if and only if the ungraded polynomial pi =Pσ∈∆p which does not annihilate p =Pσ ασZσ and hence pi is a nonidentity of Mr(F ). In order to show pq is a nonidentity we choose an evaluation of q which corresponds to the path j = j(i) determined by the chosen path for p and let qj be the corre- sponding polynomial. We obtain the product pq is a graded identity if and only if piqj is an ungraded identity of Mr(F ). Invoking the path property of A, we have VERBALLY PRIME T-IDEALS AND GRADED DIVISION ALGEBRAS 22 the polynomial qj is an ungraded nonidentity of Mr(F ) and so, being both pi and qj nonidentities and Mr(F ) verbally prime, the product piqj is a nonidentity and we are done. (cid:3) In order to complete the proof the of Theorem 1.8 it remains to show that the conditions 1 + 2 + 3 and 1 + 2 + 3′ (see paragraph before Definition 4.13) are equivalent. In fact, since we have seen already that a finite dimensional G-simple algebra A which satisfies the condition 1 + 2 + 3′ is strongly verbally prime, it remains to show 1 + 2 + 3 implies 1 + 2 + 3′. At first, we will consider a simpler case, namely we will assume A = F cH ⊗ M[G/H](F ), where H is normal in G, β is a G invariant cocycle and [G/H] stands for a transversal of H in G containing e (i.e. each element of [G/H] appears exactly once in g, the tuple which provides the elementary grading on A). We need to show condition 3′ holds. Claim: It is sufficient to show the following Proposition 4.20. Notation as above. For every Z = xg1 · · · xgn and every good permutation σ ∈ Sn, there is a scalar s ∈ F × such that Z − sZσ ∈ IdG(A) (Zσ = xgσ(1) · · · xgσ(n) ). Acknowledgment 4.21. We are thankful to Darrell Haile for his contribution to the proof of this important proposition. Let us prove condition 3′ assuming the proposition. Let p(xg1 , . . . , xgn ) = PσΛ βσZσ be a pure polynomial. Note that βσ may be zero. We need to show that if p vanishes on one path, it vanishes on every path and hence it is a G-graded iden- tity. Clearly, we may assume βe = 1. We proceed by induction on the number k0 of nonzero coefficients βσ in p and so by way of contradiction we assume p is a minimal counter example. Note that k0 > 1 since any multilinear monomial does not vanish βσ1 , βσ2 6= 0. Let s = sσ1 be such that Z − sZσ is a G-graded identity. It follows on any path. Suppose k0 ≥ 3 and let p = Z+βσ1Zσ1 +βσ2 Zσ2 +Pτ ∈Λ\{e,σ1,σ2} βτ Zτ , that (βσ1 − s)Zσ1 +Pτ ∈Λ\{e,σ1,σ2} βτ Zτ doesn't satisfy the path property. Contra- diction to the minimality of p. So we need to show the path property is satisfied by pure polynomials of the form p = Z + βσZσ where βσ 6= 0. But again, if βσ 6= s, p does not vanish on any evaluation path and the claim follows. Proof. (of Proposition) Write g(i) for [g1 · · · gi] and g(0) = e. Here [g] ∈ [G/H] is the corresponding element to g ∈ G. Let Z = xg1 · · · xgn and Z = xgσ(1) · · · xgσ(n) be good permutations of each other. Then gσ(1) · · · gσ(n) = g1 · · · gn and (4.2) for i = 1, . . . , n. By setting σ(0) = 0 we get equivalently (4.3) for i = 1, . . . , n. Fix t ∈ [G/H]. For i = 1, . . . , n we evaluate xgi by u[tg(i−1)]gi[tg(i)]−1⊗e[tg(i−1)],[tg(i)] and obtain that Z, under this substitution, is equal to g(σ(i)) =cgσ(i) := [gσ(1) · · · gσ(i)] g(σ(i) − 1) =cgσ(i − 1) u[tg(i−1)]gi[tg(i)]−1! ⊗ et,[tg(n)]. nYi=1 VERBALLY PRIME T-IDEALS AND GRADED DIVISION ALGEBRAS 23 Next, let αi,t ∈ H (i = 0, . . . , n) be such that [tg(i)] = αi,tt[g(i)]. So (4.4) hi,t := [tg(i − 1)]gi[tg(i)]−1 = = αi−1,t · t ·(cid:0)[g(i − 1)]gi[g(i)]−1(cid:1) } {z fi∈H · t−1 · α−1 i,t = αi−1,tf t i α−1 i,t , where α0,t = e and ασ(n),t = [tg(σ(n)][g(σ(n)]−1t−1 = [tg(n)][g(n)]−1t−1 = αn,t. Lemma 4.22. Every fi is independent of t and we have the equality f t 1 · · · f t n = f t σ(1) · · · f t σ(n). Proof. By 4.2 and 4.3, for i = 1, . . . , n. Hence, σ(1) · · · f t f t g(σ(i − 1)) =cgσ(i − 1) = g(σ(i) − 1) [g(σ(i) − 1)]gσ(i)[g(σ(i))]−1!t [g(σ(i − 1))]gσ(i)[g(σ(i))]−1!t σ(n) = nYi=1 = nYi=1 = (g1 · · · g−1 n · [g(n)−1])t = f t 1 · · · f t n. =(cid:0)gσ(1) · · · gσ(n) · [g(n)−1](cid:1)t Lemma 4.23. For i = 1, . . . , n (cid:3) (4.5) As a result, (4.6) ασ(i)−1,t = ασ(i−1),t hσ(i),t = ασ(i−1),tf t σ(i)α−1 σ(i),t Proof. We saw in the previous proof that g(σ(i − 1)) = g(σ(i) − 1) for i = 1, . . . , n. Thus, ασ(i−1),t = [tg(σ(i−1))][g(σ(i−1))]−1t−1 = [tg(σ(i)−1)][g(σ(i)−1)]−1t−1 = ασ(i)−1,t. Thus, hσ(i),t = ασ(i)−1,tf t σ(i)α−1 σ(i),t = ασ(i−1),tf t σ(i)α−1 σ(i),t. Next, for i = 1, . . . , n uhi,t = uαi−1,tf t i α−1 i,t = 1 c(αi−1,t, f t i )c(αi−1,tf t i , α−1 i,t ) (cid:3) · uαi−1,t · uf t i · uα−1 i,t . As a result, uh1,t · · · uhn,t = nYi=1 Using 4.6 and 4.4 we compute 1 c(αi−1,t, f t i )c(αi−1,tf t i , α−1 i,t )! nYi=1 c(αi,t, α−1 i,t )!·uf t 1 · · · uf t n·uα−1 n,t . VERBALLY PRIME T-IDEALS AND GRADED DIVISION ALGEBRAS 24 uσ(h1,t) · · · uσ(hn,t) = nYi=1 1 c(ασ(i−1),t, f t σ(i))c(ασ(i−1),tf t σ(i), α−1 σ(i),t)! nYi=1 c(ασ(i),t, α−1 σ(i),t)!·uf t · · · uf t σ(n) ·uα−1 n,t . σ(1) Hence, for every t ∈ [G/H], we have where s(f t 1, . . . , f t uh1,t · · · uhn,t = s(f t n) ∈ F × satisfies 1, . . . , f t n) · uσ(h1,t) · · · uσ(hn,t), uf t 1 · · · uf t n = s(f t 1, . . . , f t n) · uf t σ(1) · · · uf t ′ σ(n) . However, since we assumed that [c] is invariant under G, we must have that s(f t 1, . . . , f t n) = s(f1, . . . , fn). This completes the proof of the proposition and hence the implication 1+2+3 ⇒ (cid:3) 1 + 2 + 3′ in case r = 1. 5. G−graded division algebras In this section we prove 1) Theorem 1.12. 2) 1 + 2 + 3 ⇒ 1 + 2 + 3′ where r is arbitrary (Theorem 4.12). 3) Theorem 1.14. Proof. (of 1.12) One direction is easy. Suppose such B exists. Since A and B have the same G-graded identities it is sufficient to show that products of multilinear nonidentities of B with disjoint sets of variables is a nonidentity of B. But this is clear since p and q have nonzero (invertible) evaluations on B and hence also their product pq. We proceed now to show that such a B exists provided the algebra A is G-graded strongly verbally prime. We will show that the corresponding generic G-simple algebra is a G-graded division algebra. We consider an affine relatively free G-graded algebra UG = F hXGi/IdG(A) corresponding to A with IdG(UG) = IdG(A) (it is known that such algebra exists (see subsection 7.1, [1])). Here XG is a finite set of G-graded variables. We claim UG is semisimple, that is the Jacobson radical J = J(UG) is zero. Indeed, we know by Braun-Kemer-Razmyslov theorem J is nilpotent. Moreover, the Jacobson radical is G-graded (see [9]) and hence this would contradict the strongly verbally prime condition on A unless J = 0. Let Z(U ) denote the center of U and Z(U )e the central e-homogeneous elements. Note: the center Z(U ) is not graded in general. Lemma 5.1. 1) There exist e-homogeneous central polynomials, that is Z(U )e 6= 0. 2) Z(U )e has no nontrivial zero divisors as elements of U . Proof. Both statements were proved in [11] (see Theorem 3.12, Lemma 5.4). Nev- ertheless, we recall the proof of the second statement since a similar argument will be used later. Let p be a central e-homogeneous nonidentity and q an arbitrary homogeneous nonidentity. Note that we are not assuming further conditions on p and q, and in particular no assumption was made on the sets of variables of p and q. We claim, we may assume that p and q are multihomogeneous (each variable VERBALLY PRIME T-IDEALS AND GRADED DIVISION ALGEBRAS 25 appears in any monomial with the same multiplicity). To see this put an ordering on the variables which appear either in p or q. Counting variables multiplicities we order lexicographically the monomials in p and q respectively. Let p = p1 + . . . + pr and q = q1 + · · · + qs be the decomposition of p and q into its multihomogeneous constituents where p1 < . . . < pr and q1 < . . . < qs. Suppose pq is a G-graded identity. It follows that its homogeneous constituent are G-graded identities and in particular p1q1 is a G-graded identity. This proves the claim. Let Vp and Vq denote the set of variables appearing in p or q respectively. Let V = Vp ∪ Vq and let n = card(V). An evaluation of p (or q) is a choice of an element in An = F d, where d = dimF (A) × n. Let Ap = {z ∈ An : p(z) 6= 0} and Bq = {z ∈ An : q(z) = 0}. Note that Ap is an open nonempty (and hence dense) subset of An in the Zariski topology whereas Bq is a closed set 6= An. Since p take values in Z(A)e = F , pq being an identity is equivalent to Ap ⊆ Bq which is impossible. This proves the second part of the lemma. (cid:3) Consider the central localization AG = (Z(U )e − {0})−1UG. We obtain that Z(AG)e = F rac(Z(U )e) is a field and we denote it by K. Invoking [11], Corollary 4.7 (Posner's theorem for G-graded algebras), we obtain a finite dimensional algebra G-simple over K. We need to show the e-component (AG)e is an (ordinary) division algebra. Since the e-component is finite dimensional over K we need to show there are no nontrivial zero divisors in (AG)e. Suppose this is not the case and let p and q be e−polynomials, nonidentities, whose product is zero (obtained by clearing denominators). Lemma 5.2. We may assume the polynomials p and q are multihomogeneous. Proof. Same as above. (cid:3) Recall that {ue ⊗ ei,i ⊗ E}i∈{1,...,[G:H]},E∈Mr(F ) is a spanning set of the e- homogeneous component of A. We refer to Ae,i = spanF {ue ⊗ ei,i ⊗ E} as the ith component of Ae, i = 1, . . . , [G : H]. Claim: Evaluations of any nonidentity e-polynomial on the algebra A yield nonzero values on every component of Ae. More precisely, if we denote by Pi the projection Ae → Ae,i, then for any i the evaluation set of Pip on A is nonzero. Let us complete the proof of the Theorem using the claim. We let Ap,i = {z ∈ An : Pip(z) 6= 0} and Bq,i = {z ∈ An : Piq(z) = 0}. As Ap,i is nonempty and open, Bq,i 6= An and closed, we have Ap,i * Bq,i. This shows Pipq = Pip · Piq : An → Mr(F ) is a nonzero polynomial function and we are done. We turn now to the proof of the claim. We'll see that basically the claim follows from the path property which is known to hold for A by Proposition 4.17. Suppose first p is e-homogeneous and pure. We know that if p does not vanish on a certain path, say the i-th path, it does not vanish on any path j ∈ {1, . . . , [G : H]}. Furthermore, we saw that different paths yields values on different diagonal blocks and so the claim is clear in this case. Suppose now p is multilinear and let p = p1 + . . . + pq be the decomposition of p into its pure components. Let Z be a monomial of p1 say and suppose it does not vanish on a certain path evaluation. Let Zσ be a monomial in p which obtained from Z by the permutation σ ∈ Sn. If Zσ is a good permutation of Z then they are constituents of the same pure polynomial, the case we considered above. Otherwise, applying the claim in the proof of Lemma 4.15 we know Z and Zσ get values on different blocks (this includes the case the evaluation annihilates Zσ). This establishes the claim in case p is multilinear. VERBALLY PRIME T-IDEALS AND GRADED DIVISION ALGEBRAS 26 Finally, we suppose p is multihomogeneous and proceed by induction on its mul- tihomogeneous degree. Suppose p(x1, x2 . . . , xn) is multihomogeneous where vari- ables x1, . . . , xn appear in degree k1, . . . , kn respectively. Without loss of generality we may assume k1 ≥ k2 ≥ · · · ≥ kn. Clearly we may view the tuple (k1, . . . , kn) as a partition of m =Ps ks. Put x1 = t1 + · · · + tk1 and let ¯p(t1, . . . , tk1 , x2, . . . , xn) = p((t1 + . . . + tk1 ), x2, . . . , xn). Since Pip(x1, x2, . . . , xn) = 0, we have that Pi ¯p(t1, . . . , tk, x2, . . . , xn) = 0 and hence also Pi ¯p(tj , x2, . . . , xn) = 0, for j = 1, . . . , k1. It follows that Pi(¯p(t1, . . . , tk1 , x2, . . . , xn) − p(tj, x2, . . . , xn)) = 0 k1Xj=1 where ¯p(t1, . . . , tk, x2, . . . , xn) −Pj p(tj, x2, . . . , xn)) is a sum of multihomogeneous polynomials of degree m whose multihomogeneous degrees are strictly smaller than (k1, . . . , kn). The proof is completed by induction. Thus, we have proved the algebra AG = (Z(U )e − {0})−1UG is a finite dimen- sional G-graded division algebra over the field K = Z(AG)e. It remains to show the algebra AG is a twisted form of A. But this is clear since by construction IdG(AG) = IdG(A) = Γ and finite dimensional G-graded simple algebras over an algebraically closed field are determined up to a G-graded isomorphism by the T -ideal of G-graded identities. (cid:3) We still owe the reader the proof of the implication 1 + 2 + 3 ⇒ 1 + 2 + 3′ in the general case, namely in case A = F cH ⊗ M[G/H](F ) ⊗ Mr(F ), r ≥ 1. Interestingly, we complete the proof of that implication (2 → 3 below) applying Theorem 1.12. For convenience, we summarize our main results in the following Theorem. Theorem 5.3. Let Γ ⊂ F hXGi be a G-graded T -ideal which contains a Capelli polynomial. The following conditions are equivalent: (1) Γ is strongly verbally prime. (2) There exists a finite dimensional algebra G-simple algebra A = F cH ⊗ M[G/H](F ) ⊗ Mr(F ) which satisfies conditions 1 + 2 + 3 of Theorem 1.8 such that IdG(A) = Γ. (3) There exists a finite dimensional algebra G-simple algebra A = F cH ⊗ M[G/H](F ) ⊗ Mr(F ) which satisfies conditions 1 + 2 + 3′ of Theorem 4.12 such that IdG(A) = Γ. (4) There exists a field K ⊇ F and a finite dimensional G-graded division algebra B over K with IdG(B) = Γ. Furthermore (1) The G-simple algebra A in (2) and (3) is uniquely determined. (2) The finite dimensional G-graded division algebra B (in 3) is uniquely de- termined in the following sense: Any two such algebras are twisted forms of each other, that is, if B′ is a finite dimensional G-graded division algebra over K ′ with IdG(B′) = Γ, then there exists a field extension E of K and K ′ such that B ⊗K E ∼= B′ ⊗K ′ E as G-graded algebras. Proof. Note that we have shown that 1 → 2 in the first part of section 4 (Lemma 4.4, Proposition 4.6 and Lemma 4.11). The equivalence 3 ↔ 1 is stated and proved VERBALLY PRIME T-IDEALS AND GRADED DIVISION ALGEBRAS 27 in Proposition 4.17. Above in this section, we showed 1 ↔ 4 and in last part of section 4 (Proposition 4.20) we showed 2 → 3 in case r = 1. Let A = F cH ⊗ M[G/H](F ) ⊗ Mr(F ) and suppose it satisfies conditions 1 + 2 + 3 of Theorem 1.8. Clearly, the proof will be completed if we show that A admits a twisted form which is a G-graded division algebra. Consider the algebra A0 = F cH ⊗ M[G/H](F ). Clearly it satisfies 1 + 2 + 3 of Theorem 1.8 and since here r = 1 it admits a finite dimensional G-graded twisted form B0 over a field K0 which is a G-division algebra. All we need to do now is to find a field K and G-graded division algebra B over some field K which is a twisted form of A. Since B0 is a G-graded division algebra over K0 and since its e-component is commutative, we have that L0 = (B0)e is a finite field extension of K0. Clearly, we may extend scalars to K0(t, s) (the rational function field over K0 on ideterminates t and s) and obtain a G-graded division algebra (B0)K0(t,s) which is also a twisted form of A0. It remains to prove there exists an ungraded division algebra D of index r over K0(t, s) which remains a division algebra after extending scalars to L0(t, s). Since F is algebraically closed of characteristic zero, it contains a primitive rth root of unity µr and so we may consider the symbol algebra (t, s)r = hur = t, vr = s, vu = µruvi over K0(t, s). It is easy to prove that ((t, s)r)L0(t,s) is a division algebra and we are done. Details are left to the reader. Finally, for the uniqueness (second part of the theorem) we invoke once again the fact that any two G-graded simple algebras over an algebraically closed field F of characteristic zero are G-graded isomorphic if and only if they satisfy the same G-graded identities (see [3]). (cid:3) We close the paper with the proof of Theorem 1.14. Proof. Suppose A is a finite dimensional G-graded division algebra over a field K and let p and q be G-homogeneous polynomials such that p and q /∈ IdG(A). If pq ∈ IdG(A) we have that Ap = {z ∈ An : p(z) 6= 0} ⊆ Bq = {z ∈ An : q(z) = 0}. But this is impossible since Ap is nonempty and open whereas Bp 6= An and closed. For the opposite direction note that our condition on polynomials p and q implies A is G-prime and hence by Theorem 1.5 A is G-graded PI equivalent to a finite dimensional G-simple algebra. On the other hand the condition on p and q implies A is strongly G-graded verbally prime. Invoking Theorem 1.8 and Theorem 4.12 the result follows. (cid:3) References [1] E. Aljadeff, A. Kanel-Belov, Representability and Specht problem for G−graded algebras, Adv. Math. 225 (2010), no. 5, 2391–2428. 2, 7, 8, 9, 24 [2] E. Aljadeff and O. David, On group gradings on PI-algebras, Journal of Algebra 428 (2015) 403–424. 11 [3] E. Aljadeff and D. Haile, Simple G-graded algebras and their polynomial identities, Trans. Amer. Math. Soc. 366 (2014), 1749–1771. 10, 15, 27 [4] E. Aljadeff, D. Haile and M. Natapov, Graded identities of matrix algebras and the universal graded algebr, Trans. Amer. Math. Soc. 362 (2010), 3125–3147. 15, 16 [5] S.A. Amitsur, The T-ideals of the free ring. J. London Math. Soc. 30 (1955), 470-475. 1 [6] Y. Bahturin, S. K. Sehgal and M. Zaicev, Finite-dimensional simple graded algebras. (Rus- sian), Mat. Sb., 199(7), (2008), 21–40; translation, Sb. Math., 199(7), (2008), 965–983. 3 VERBALLY PRIME T-IDEALS AND GRADED DIVISION ALGEBRAS 28 [7] A. Berele and J. Bergen, P.I. algebras with Hopf algebra actions, J. Algebra 214 (1999), no. 2, 636–651. 3, 5 [8] O. M. Di Vincenzo, Z2-graded polynomial identities of verbally semiprime algebras. Comm. Algebra 27 (1999), no. 5, 2031–2044. 3 [9] M. Cohen, S. Montgomery, Group-graded rings, smash products, and group actions, Trans. Amer. Math. Soc. 282 (1) (1984) 237–258. 8, 24 [10] A. Giambruno, M. Zaicev, Polynomial Identities and Asymptotic methods, AMS Mathemat- ical Surveys and Monographs 122, Providence, R.I., 2005. 1, 9 [11] Y. Karasik G-graded central polynomials and G-graded Posner's theorem, arXiv:1610.03977 2, 3, 24, 25 [12] A. R. Kemer Varieties and Z2-graded algebras. (Russian) Izv. Akad. Nauk SSSR Ser. Mat. 48 (1984), no. 5, 1042–1059. 2 [13] P. Koshlukov and M. Zaicev, Identities and isomorphisms of graded simple algebras, Linear Algebra Appl. 432 (2010), no. 12, 3141–3148. 10 Department of Mathematics, Technion - Israel Institute of Technology, Haifa 32000, Israel E-mail address: aljadeff 'at' tx.technion.ac.il (E. Aljadeff), E-mail address: yaakov 'at' tx.technion.ac.il (Y. Karasik).
1007.4118	1	1007	2010-07-23T12:57:27	Hom-power associative algebras	[ "math.RA" ]	A generalization of power associative algebra, called Hom-power associative algebra, is studied. The main result says that a multiplicative Hom-algebra is Hom-power associative if and only if it satisfies two identities of degrees three and four. It generalizes Albert's result that power associativity is equivalent to third and fourth power associativity. In particular, multiplicative right Hom-alternative algebras and non-commutative Hom-Jordan algebras are Hom-power associative.	math.RA	math	HOM-POWER ASSOCIATIVE ALGEBRAS DONALD YAU Abstract. A generalization of power associative algebra, called Hom-power associative algebra, is studied. The main result says that a multiplicative Hom-algebra is Hom-power associative if and only if it satisfies two identities of degrees three and four. It generalizes Albert's result that power associativity is equivalent to third and fourth power associativity. In particular, multiplicative right Hom-alternative algebras and non-commutative Hom-Jordan algebras are Hom-power associative. 1. Introduction Power associative algebras were first systematically studied by Albert [1, 2]. These non-associative algebras play an important role in the studies of associative bilinear forms, loops, the Cayley-Dickson algebras, and Banach algebras, among many other subjects [3, 6, 14]. A power associative algebra can be defined conceptually as an algebra in which every element generates an associative sub-algebra. Equivalently, power associative algebras can be algebraically characterized as follows. For an element x in an algebra A, define the right powers for n ≥ 1. Then an algebra A is power associative if and only if x1 = x and xn+1 = xnx xn = xn−ixi (1.0.1) (1.0.2) for all x ∈ A, n ≥ 2, and i ∈ {1, . . . , n − 1}. Therefore, one can also think of (1.0.2) as the defining identities for power associative algebras. If A satisfies (1.0.2) for a particular integer n, then A is called nth power associative. A theorem of Albert [1] says that (in characteristic 0, which is assumed throughout this paper) an algebra is power associative if and only if it is third and fourth power associative, which in turn is equivalent to x2x = xx2 and x4 = x2x2 (1.0.3) for all elements x. Associative algebras are clearly power associative, but the class of power associative algebras in- cludes some important classes of non-associative algebras. For example, alternative algebras and Jordan algebras are all power associative because they satisfy (1.0.3). In fact, right-alternative and non-commutative Jordan algebras are power associative algebras (Corollary 5.7). Therefore, power associative algebras are common generalizations of right-alternative algebras and non-commutative Jordan algebras. Moreover, there exist power associative algebras that are neither alternative nor Date: September 4, 2018. 2000 Mathematics Subject Classification. 17A05, 17A15, 17A20, 17D15. Key words and phrases. Hom-power associative algebra, right Hom-alternative algebra, noncommutative Hom-Jordan algebra, sedenions. 1 2 DONALD YAU non-commutative Jordan. One such power associative algebra is the sedenions, which is the sixteen- dimensional algebra constructed from the octonions via the Cayley-Dickson process with the parameter −1 (Example 2.7). The purpose of this paper is to study a generalization of power associative algebras, called Hom-power associative algebras, in which the defining identities (1.0.2) are relaxed by a certain linear self-map, called the twisting map. Hom-type algebras appeared in [8] in the form of Hom-Lie algebras, in which the Jacobi identity is relaxed by a twisting map, to describe the structures on some deformations of the Witt and the Virasoro algebras. Hom-Lie and Hom-associative algebras are further studied in [11, 12, 15, 16] and the references therein. Hom-type generalizations of other algebraic structures, including quantum groups, the Yang-Baxter equations, Novikov algebras, and Lie and infinitesimal bialgebras, can be found in [17] - [29] and the references therein. One reason for considering Hom-power associative algebras is that the class of Hom-power associative algebras has a certain closure property with respect to twisting by self-morphisms that the class of power associative algebras does not have. Moreover, Hom-power associative algebras retain some important properties of power associative algebras. For example, the main result of this paper says that when the twisting map is multiplicative, Hom-power associativity, which involves infinitely many defining identities, is equivalent to the Hom-versions of the two identities (1.0.3). As a consequence of this result, multiplicative right Hom-alternative algebras and non-commutative Hom-Jordan algebras are Hom-power associative. The rest of this paper is organized as follows. In section 2 we define Hom-type analogues of the right powers and power associativity, called Hom- powers and Hom-power associativity. The closure property mentioned above is discussed in Theorem 2.4. It says that a Hom-power associative algebra gives rise to another one when its structure maps (the multiplication and the twisting map) are both twisted by any self-morphism. If this twisting construction is applied to a power associative algebra, regarded as a Hom-power associative algebra with identity twisting map, the result is a multiplicative Hom-power associative algebra (Corollary 2.6) that is usually not power associative. We illustrate this result in Example 2.7 with the sedenions. Given any Hom-algebra A, one can associate a family of Hom-algebras A(λ) with the same underlying vector space, where λ is any parameter in the ground field, in which the multiplication is an interpolation between those in A and its opposite algebra. We show that when A is Hom-power associative, the Hom-powers in A and A(λ) coincide (Theorem 2.9), and A(λ) is also Hom-power associative (Corollary 2.10). In preparation for the main result in section 5, in sections 3 and 4 we study third and fourth Hom-power associativity, respectively. In Theorem 3.2 several characterizations of third Hom-power associativity are given. One consequence of this result is that up to (n − 1)st Hom-power associativity implies certain commutative identities involving the Hom-powers xn−i and xi (Proposition 3.4). It is also observed that the intersection of the classes of third Hom-power associative algebras and of Hom- Lie admissible algebras is precisely the class of A3-Hom-associative algebras (Theorem 3.9). This last result is not used in later sections. In section 4 it is observed that multiplicative right Hom-alternative and non-commutative Hom- Jordan algebras are third and fourth Hom-power associative (Proposition 4.4). A conceptual difference between power associative algebras and Hom-power associative algebras is discussed (Examples 4.5 and HOM-POWER ASSOCIATIVE ALGEBRAS 3 4.6). Theorem 4.8 provides the linearized form of the identity x4 = α(x2)α(x2), which is part of fourth Hom-power associativity. This result is crucial to the proof of the main result. The main result of this paper (Corollary 5.2) is proved in section 5. It says that for a multiplicative Hom-algebra, Hom-power associativity is equivalent to up to fourth Hom-power associativity, which in turn is equivalent to two identities involving the first few Hom-powers. In particular, for a multiplicative Hom-flexible algebra, Hom-power associativity is equivalent to one identity (Corollary 5.4). Another consequence of the main result is that multiplicative right Hom-alternative and non-commutative Hom- Jordan algebras are Hom-power associative (Corollary 5.6). 2. Basic properties of Hom-power associative algebras In this section we introduce Hom-power associative algebras and provide some basic construction results and examples. 2.1. Notations. Throughout the rest of this paper, we work over a fixed field k of characteristic 0. If V is a k-module and µ : V ⊗2 → V is a bilinear map, then µop : V ⊗2 → V denotes the opposite map, i.e., µop = µτ , where τ : V ⊗2 → V ⊗2 interchanges the two variables. For x and y in V , we sometimes write µ(x, y) as xy. For a linear self-map α : V → V , denote by αn the n-fold composition of n copies of α, with α0 ≡ Id. The symmetric group on n letters is denoted by Sn. A multi-linear map f : V ⊗n → V is said to be totally anti-symmetric if f = (−1)ε(σ)f ◦ σ for all σ ∈ Sn, where ε(σ) is the signature of σ. Let us begin with the basic definitions regarding Hom-algebras. Definition 2.2. (1) By a Hom-algebra we mean a triple (A, µ, α) in which A is a k-module, µ : A⊗2 → A is a bilinear map (the multiplication), and α : A → A is a linear map (the twisting map). Such a Hom-algebra is often abbreviated to A. (2) A Hom-algebra (A, µ, α) is multiplicative if α is multiplicative with respect to µ, i.e., αµ = µα⊗2. (3) A Hom-algebra (A, µ, α) is commutative if its multiplication is commutative, i.e., µ = µop. (4) Let A and B be Hom-algebras. A morphism f : A → B of Hom-algebras is a linear map f : A → B such that f µA = µBf ⊗2 and f αA = αBf . If (A, µ) is any algebra, then we also regard it as a Hom-algebra (A, µ, Id) with identity twisting map. Next we define the Hom-versions of Albert's right powers and power associativity [1, 2]. Definition 2.3. Let (A, µ, α) be a Hom-algebra and x ∈ A. (1) Define the nth Hom-power xn ∈ A inductively by for n ≥ 2. x1 = x, xn = xn−1αn−2(x) (2.3.1) 4 DONALD YAU (2) We say that A is nth Hom-power associative if xn = αn−i−1(xi)αi−1(xn−i) (2.3.2) for all x ∈ A and i ∈ {1, . . . , n − 1}. (3) We say that A is up to nth Hom-power associative if A is kth Hom-power associative for all k ∈ {2, . . . , n}. (4) We say that A is Hom-power associative if A is nth Hom-power associative for all n ≥ 2. Some remarks are in order. First, if α = Id in Definition 2.3, then xn = xn−1x, which coincides with Albert's notion of right power [1, 2]. Moreover, in this case, nth Hom-power associativity reduces to nth power associativity, i.e., xn = xixn−i for 1 ≤ i ≤ n − 1. Therefore, a power associative algebra is equivalent to a Hom-power associative algebra whose twisting map is the identity map. Second, the first few Hom-powers are x2 = xx, x3 = x2α(x) = (xx)α(x), x1 = x, x4 = x3α2(x) = (x2α(x))α2(x) = ((xx)α(x))α2(x), x5 = x4α3(x) = (cid:0)(x2α(x))α2(x)(cid:1) α3(x) = (cid:0)((xx)α(x))α2(x)(cid:1) α3(x). Note that the i = n − 1 case of nth Hom-power associativity (2.3.2) is automatically true because it says xn = xn−1αn−2(x), which holds by definition. It is included in the definition for symmetry. Moreover, every Hom-algebra is automatically second Hom-power associative, since with n = 2 the condition (2.3.2) says x2 = xx, which holds by definition. Therefore, the first nontrivial case is third Hom-power associativity, which will be studied in the next section. Third, by definition nth Hom-power associativity (2.3.2) requires n − 1 identities. Under certain commutative conditions, the number of identities that need to be satisfied for nth Hom-power associa- tivity can be reduced by half. In fact, the two factors αn−i−1(xi) and αi−1(xn−i) on the right-hand side of (2.3.2) are related by an interchange of i and n − i. Therefore, if these two elements commute for all i ∈ {1, . . . , n − 1}, then nth Hom-power associativity reduces to n/2 or (n − 1)/2 identities, according to whether n is even or odd. These commutative conditions will appear in Proposition 3.4, which will be used in the proof of Theorem 5.1. Our first result says that the category of (nth) Hom-power associative algebras is closed under twisting by self-morphisms. Theorem 2.4. Let (A, µ, α) be a Hom-algebra and β : A → A be a morphism. Define the Hom-algebra Aβ = (A, µβ = βµ, βα). (2.4.1) Then the following statements hold. (1) If A is multiplicative, then so is Aβ. (2) We have (2.4.2) for all x ∈ A and n ≥ 1, where xn and (xn)′ are the nth Hom-powers in A and Aβ, respectively. (3) If A is nth Hom-power associative, then so is Aβ. Therefore, if A is Hom-power associative, (xn)′ = βn−1(xn) then so is Aβ. HOM-POWER ASSOCIATIVE ALGEBRAS 5 Proof. The first assertion regarding multiplicativity is obvious. The second assertion is proved by induction, with the n = 1 case being trivially true. For the induction step, suppose that (2.4.2) is true. Using the fact that β is a morphism, we have (xn+1)′ = µβ((xn)′, (βα)n−1(x)) = µβ(βn−1(xn), βn−1αn−1(x)) = βnµ(xn, αn−1(x)) = βn(xn+1). This finishes the induction and proves the second assertion. For the last assertion, assume that A is nth Hom-power associative and i ∈ {1, . . . , n − 1}. We must show (2.3.2) for Aβ. Using (2.4.2) and the fact that β is a morphism, we have (xn)′ = βn−1(xn) = βn−1µ(αn−i−1(xi), αi−1(xn−i)) = µβ(βn−2αn−i−1(xi), βn−2αi−1(xn−i)) = µβ (cid:0)(βα)n−i−1(βi−1(xi)), (βα)i−1(βn−i−1(xn−i))(cid:1) = µβ (cid:0)(βα)n−i−1(xi)′, (βα)i−1(xn−i)′(cid:1) . This shows that Aβ is nth Hom-power associative. (cid:3) Now we discuss some consequences of Theorem 2.4. The result below says that every multiplicative Hom-power associative algebra induces a derived sequence of multiplicative Hom-power associative algebras with twisted structure maps. Corollary 2.5. Let (A, µ, α) be a multiplicative Hom-power associative algebra. Then is a multiplicative Hom-power associative algebra for each n ≥ 1. An = (A, µ(n) = α2n −1µ, α2n ) Proof. That A1 is a multiplicative Hom-power associative algebra follows from the β = α case of Theorem 2.4. Then we use induction and the fact An+1 = (An)1 to finish the proof. (cid:3) The next result says that multiplicative Hom-power associative algebras can be constructed from power associative algebras and their self-morphisms. A result of this form first appeared in [16]. Corollary 2.6. Let A be a power associative algebra and β : A → A be an algebra morphism. Then is a multiplicative Hom-power associative algebra. Aβ = (A, µβ = βµ, β) Proof. This is the α = Id case of Theorem 2.4. (cid:3) Example 2.7 (Hom-Sedenions). In this example, let k be the field of real numbers. Let S be the 16-dimensional algebra constructed from k (whose involution is taken to be the identity map) by applying the Cayley-Dickson process four times with the parameter −1 in each step [13]. Elements in the non-commutative algebra S are called sedenions. Here we show how Corollary 2.6 can be applied to S to create Hom-power associative algebras. 6 DONALD YAU Let us first recall some basic properties of S [5, 9]. Unlike the 8-dimensional Cayley octonions, S is not an alternative algebra, but it is power associative. Let us provide proofs of these facts. From the Cayley-Dickson process itself, the algebra S has a standard basis {ei : 0 ≤ i ≤ 15}, with e0 = 1 a two-sided multiplicative identity. Moreover, for i 6= j in {1, . . . , 15}, we have The multiplication table for S using this basis is given below. e2 i = −1 and eiej = −ejei. 1 1 e1 e2 e3 e4 e5 e6 e7 e8 e9 1 e1 e2 e3 e4 e5 e6 e7 e8 e9 e1 e1 −1 −e3 e2 e2 e3 −1 e2 −e1 e3 e3 −e2 e1 −1 −e5 −e6 −e7 e4 e4 e5 e6 e7 −1 e4 e7 −e6 e4 e5 −e7 e6 −e1 −e5 −e2 e5 e5 e6 e6 −e4 −e7 e7 e7 e6 e7 −e4 −e5 −e6 e1 −1 e3 e5 e2 −e3 −1 −e4 e3 e2 −e1 −1 e4 −e3 −e2 e1 e8 e8 e9 e10 e11 e12 e13 e14 e15 −1 −e9 −e10 −e11 −e12 −e13 −e14 −e15 e8 −e11 e10 −e13 e12 −e14 −e1 e15 e12 −e9 −e14 −e15 −e15 e14 −e13 e13 e12 −e2 e3 −e3 −e2 −e6 −e7 e9 e9 e10 e10 −e8 −e11 e11 e11 e10 e12 e12 −e13 e13 e13 e12 e11 −e8 −e9 −e14 −e15 e14 e14 e15 e12 −e8 −e15 e14 −e13 e15 −e8 −e9 −e10 −e11 e15 e15 −e14 e13 e12 e7 −e6 e5 e4 −e2 e1 −1 −e8 e11 −e10 −e11 −e8 e9 e10 −e9 −e8 e5 e4 −e1 −1 −e3 e6 e7 e4 e3 −1 e6 −e5 −e2 −e3 e9 e10 e11 e4 −e5 −e7 −1 e1 e2 e3 −e10 e13 −e12 e9 e14 e15 −e14 −e15 −e12 e13 e14 −e13 −e12 e1 −1 e2 −e3 −1 e1 e6 e7 e3 e2 −e1 −1 e7 −e6 e10 e10 e11 e11 e11 −e10 e12 e12 e13 e13 e13 −e12 e8 e9 e14 e15 e8 e15 −e14 e14 e14 −e15 −e12 e13 e15 e15 e14 −e13 −e12 −e11 e8 e10 −e9 e9 e8 −e6 −e7 −e4 e5 −e7 e6 −e5 −e4 e2 −e1 −e9 −e10 −e11 −e4 e5 e8 e11 −e10 −e5 −e4 e8 e9 e10 e11 To see that S is not an alternative algebra, recall that an algebra is said to be left alternative if it satisfies (xx)y = x(xy) for all elements x and y. If x = e1 + e8 and y = e3 + e9 in S, then we have (xx)y = −2(e3 + e9), x(xy) = −e3 + e5 − 2e9 − e10 − e12. Therefore, S is not left alternative. To see that S is power associative, let x = P15 Then we have i=0 λiei be an arbitrary sedenion with each λi ∈ k. It follows that and x2 = a + bx with a = − 15 X i=0 λ2 i , b = 2λ0. x2x = ab + (a + b2)x = xx2 (x2x)x = (a2 + ab2) + (2ab + b3)x = x2x2. Therefore, by [1] (Theorem 2) (or, equivalently, Corollary 5.3 below), S is power associative. In order to apply Corollary 2.6 to the power associative algebra S, consider, for example, the algebra (iso)morphism α : S → S with α(1) = 1, α(e1) = e5, α(e2) = e7, α(e4) = e6, α(e8) = e15. The action of α on the other basis elements are then uniquely determined by the multiplication table above: α(e3) = e2, α(e5) = −e3, α(e6) = e1, α(e7) = −e4, α(e9) = −e10, α(e10) = −e8, α(e11) = e13, α(e12) = e9, α(e13) = −e12, α(e14) = −e14, α(e15) = e11. By Corollary 2.6 there is a multiplicative Hom-power associative algebra Sα = (S, µα = αµ, α), HOM-POWER ASSOCIATIVE ALGEBRAS 7 where µ is the multiplication in S. Observe that (S, µα) is not a power associative algebra. Indeed, with x = 1 + e1, we have µα(µα(x, x), x) = 2(−e3 + e6), µα(x, µα(x, x)) = 2(−e3 − e6). Therefore, (S, µα) is not third power associative, and hence not power associative. Other algebra isomorphisms on S can be constructed following the same pattern as in the previous paragraph. In more details, let us extend the terminology in [4] for the Cayley octonions and define a basic quadruple in S to be a quadruple (ei1 , ei2 , ei3, ei4 ) of distinct basis elements, none of which is 1, such that ei3 6= ±ei1ei2 and ei4 6∈ ±{ei1ei2 , ei1 ei3, ei2 ei3 , (ei1 ei2)ei3 }. The assumptions on a basic quadruple guarantee that it generates S as an algebra and that the other eleven non-identity basis elements can be written uniquely in terms of the given four. Given another basic quadruple (ej1 , ej2 , ej3, ej4 ), there is a unique algebra isomorphism β : S → S determined by β(eil ) = ejl for 1 ≤ l ≤ 4. Corollary 2.6 can then be applied to β, which yields a multiplicative Hom-power associative algebra Sβ = (S, µβ = βµ, β). In this context, the map α in the previous paragraph is the unique map that sends the basic quadruple (e1, e2, e4, e8) to the basic quadruple (e5, e7, e6, e15). (cid:3) In Corollary 2.5 it was shown that every multiplicative Hom-power associative algebra gives rise to a sequence of multiplicative Hom-power associative algebras whose structure maps are twisted by its own twisting map. In the rest of this section, we provide another construction of Hom-power associative algebras from a given one. Definition 2.8. Let (A, µ, α) be a Hom-algebra. Define the Hom-algebra A(λ) = (A, µλ = λµ + (1 − λ)µop, α), in which λ ∈ k is an arbitrary scalar. The multiplication µλ is an interpolation between µ and its opposite multiplication µop. In particular, we have A(1) = A and A(0) = (A, µop, α). Moreover, with λ = 1/2 we have A(1/2) = A+ = (A, ∗ = (µ + µop)/2, α), which is commutative. Also note that A is multiplicative if and only if A(λ) is multiplicative for all λ. The following result is the Hom-type analogue of [2] (Theorem 1, p.581). Theorem 2.9. Let (A, µ, α) be a Hom-algebra, and let k ≥ 2 be an integer. Then the following statements hold. (1) A is up to kth Hom-power associative if and only if A(λ) is up to kth Hom-power associative for all λ ∈ k. (2) If A is up to kth Hom-power associative, then the lth Hom-powers in A and A(λ) coincide for l ∈ {1, . . . , k}. Proof. The "if" part of the first assertion is obvious because A(1) = A. To prove the other direction, suppose A is up to kth Hom-power associative. Pick any λ ∈ k. Let us write xl and (xl)′ for the lth Hom-powers in A and A(λ), respectively. First we prove the second assertion, i.e., xl = (xl)′ (2.9.1) 8 DONALD YAU for l ∈ {1, . . . , k}. This is clearly true when l = 1. For the induction step, suppose 1 ≤ l < k. Then (xl+1)′ = µλ((xl)′, αl−1(x)) = µλ(xl, αl−1(x)) = λµ(xl, αl−1(x)) + (1 − λ)µ(αl−1(x), xl) (by induction hypothesis) = (λ + 1 − λ)xl+1 (by the (l + 1)st Hom-power associativity of A) = xl+1. This finishes the induction and proves (2.9.1). Now suppose 2 ≤ l ≤ k and 1 ≤ i ≤ l − 1. To prove that A(λ) is lth Hom-power associative, we compute as follows: µλ (cid:0)αl−i−1((xi)′), αi−1((xl−i)′)(cid:1) = λµ(cid:0)αl−i−1(xi), αi−1(xl−i)(cid:1) + (1 − λ)µ(cid:0)αi−1(xl−i), αl−i−1(xi)(cid:1) = (λ + 1 − λ)xl (lth Hom-power associativity of A) (by (2.9.1)) = xl = (xl)′ (by (2.9.1)). This shows that A(λ) is up to kth Hom-power associative, as desired. (cid:3) The following result is an immediate consequence of Theorem 2.9. Corollary 2.10. Let (A, µ, α) be a Hom-algebra. Then A is Hom-power associative if and only if A(λ) is Hom-power associative for all λ ∈ k. The next result is the special case of Corollary 2.10 with λ = 1/2. Corollary 2.11. Let A be a Hom-power associative algebra. Then A+ = (A, ∗ = (µ + µop)/2, α) is a commutative Hom-power associative algebra. 3. Third Hom-power associativity In this section we give several characterizations of third Hom-power associativity. Then we discuss some of their consequences, including the relationships between third Hom-power associative algebras and Hom-Lie admissible algebras. Let us first recall the Hom-type analogue of the associator and some classes of Hom-algebras from [11]. Definition 3.1. Let (A, µ, α) be a Hom-algebra. (1) Define the Hom-associator asA : A⊗3 → A by (2) Define the cyclic Hom-associator SA : A⊗3 → A by asA = µ ◦ (µ ⊗ α − α ⊗ µ). SA = asA ◦ (Id + π + π2), where π is the cyclic permutation (1 2 3). (3) A is called a Hom-associative algebra if asA = 0. HOM-POWER ASSOCIATIVE ALGEBRAS 9 (4) A is called a Hom-alternative algebra if asA is totally anti-symmetric [10]. (5) A is called a Hom-flexible algebra if asA(x, y, x) = 0 for all x, y ∈ A. In terms of elements x, y, z ∈ A, we have asA(x, y, z) = (xy)α(z) − α(x)(yz) (3.1.1) and SA(x, y, z) = asA(x, y, z) + asA(z, x, y) + asA(y, z, x). When there is no danger of confusion, we will omit the subscript A in the (cyclic) Hom-associator. It is immediate from the definitions that every Hom-associative algebra is a Hom-alternative algebra, and every Hom-alternative algebra is a Hom-flexible algebra. Also note that every commutative Hom- algebra is Hom-flexible. The following result provides several characterizations of third Hom-power associativity and is the Hom-type analogue of [1] (Lemma 1). One consequence of this result (Proposition 3.4) is crucial to the proof of the main result in section 5. We use the notation x • y = xy + yx (3.1.2) for elements x and y in a Hom-algebra A. Theorem 3.2. Let (A, µ, α) be a Hom-algebra. Then the following statements are equivalent. (1) A is third Hom-power associative. (2) The condition holds for all x ∈ A. (3) The condition holds. (4) The condition asA(x, x, x) = 0 asA ◦ σ = 0 X σ∈S3 (3.2.1) (3.2.2) [x • y, α(z)] + [z • x, α(y)] + [y • z, α(x)] = 0 (3.2.3) holds for all x, y, z ∈ A, where [, ] = µ − µop is the commutator. (5) The cyclic Hom-associator SA is totally anti-symmetric. Proof. Third Hom-power associativity (2.3.2) says x3 = x2α(x) = α(x)x2 for all x ∈ A. Since the first equality holds by the definition of x3, third Hom-power associativity is equivalent to 0 = (xx)α(x) − α(x)(xx) = as(x, x, x), which is the condition (3.2.1). To show tat (3.2.1) and (3.2.2) are equivalent, first observe that as(x, x, x) = 1 6 X σ∈S3 as ◦ σ(x, x, x), 10 DONALD YAU so (3.2.2) implies (3.2.1). For the converse, we linearize (3.2.1) by replacing x with x + λw for λ ∈ k and w ∈ A. Collecting the coefficients of the powers of λ, we obtain: 0 = as(x + λw, x + λw, x + λw) = as(x, x, x) + λA(x, w) + λ2A(w, x) + λ3as(w, w, w) = λA(x, w) + λ2A(w, x), where Setting λ = ±1 and subtracting the results, we infer that (3.2.1) implies A(x, w) = as(x, x, w) + as(x, w, x) + as(w, x, x). for all x, w ∈ A. Next we replace x with x+ γz for γ ∈ k and z ∈ A in (3.2.4). Collecting the coefficients of the powers of γ, we obtain: A(x, w) = 0 (3.2.4) 0 = A(x + γz, w) = as(x + γz, x + γz, w) + as(x + γz, w, x + γz) + as(w, x + γz, x + γz) = A(x, w) + γB(w, x, z) + γ2A(z, w) = γB(w, x, z), where B(x1, x2, x3) = X σ∈S3 as ◦ σ(x1, x2, x3). Setting γ = 1, we infer that (3.2.1) implies B = 0, which is the condition (3.2.2). To see that (3.2.2) and (3.2.3) are equivalent, apply (3.2.2) to a generic tensor x ⊗ y ⊗ z ∈ A⊗3. Expanding the left-hand side of (3.2.2) using (3.1.1), we obtain twelve terms. If we expand the left-hand side of (3.2.3) using [, ] = µ − µop, we obtain the same twelve terms. Therefore, (3.2.2) and (3.2.3) are equivalent. Finally, observe that we have as ◦ σ = S ◦ (Id + (1 2)) = S ◦ (Id + (2 3)) = S ◦ (Id + (1 3)). X σ∈S3 So (3.2.2) implies that S is anti-symmetric in any two of its three variables. Since S = S ◦ π = S ◦ π2, where π is the cyclic permutation (1 2 3), we infer that (3.2.2) implies that S is totally anti-symmetric. On the other hand, if S is totally anti-symmetric, then 0 = S(x, x, x) = 3as(x, x, x), so as(x, x, x) = 0, which is the condition (3.2.1). Since (3.2.1) is equivalent to (3.2.2), the proof is complete. (cid:3) We now discuss some consequences of Theorem 3.2. First we observe that the Hom-algebras in Definition 3.1 are all third Hom-power associative. Corollary 3.3. Every Hom-flexible algebra is third Hom-power associative. Proof. A Hom-flexible algebra A satisfies as(x, y, x) = 0 for all x and y in A, which in particular implies as(x, x, x) = 0. So A is third Hom-power associative by Theorem 3.2. (cid:3) HOM-POWER ASSOCIATIVE ALGEBRAS 11 The next result is the Hom-version of [1] (Lemma 2). It says that up to (n − 1)st Hom-power associativity implies part of nth Hom-power associativity. It is a prelude to Theorem 5.1 and will play a crucial role in the proof of that result. Proposition 3.4. Let (A, µ, α) be a multiplicative up to (n − 1)st Hom-power associative algebra for some n ≥ 4. Then [αλ−1(xn−λ), αn−λ−1(xλ)] = 0 (3.4.1) for all x ∈ A and λ ∈ {1, . . . , n − 1}, where [, ] = µ − µop is the commutator. Proof. Pick x ∈ A and define We must show that x(λ) = 0 for λ ∈ {1, . . . , n − 1}. First we claim that x(λ) = [αλ−1(xn−λ), αn−λ−1(xλ)]. for λ ∈ {1, . . . , n − 2}. To prove (3.4.2), note that, with the notation in (3.1.2), the up to (n − 1)st Hom-power associativity of A implies: x(λ + 1) = x(λ) + x(1) (3.4.2) 2xλ+1 = xλ • αλ−1(x), 2xn−λ = xn−λ−1 • αn−λ−2(x), 2xn−1 = αn−λ−2(xλ) • αλ−1(xn−λ−1). (3.4.3) Applying αn−λ−2 and αλ−1 to the first and the second equalities in (3.4.3), respectively, and using multiplicativity, we obtain: 2αn−λ−2(xλ+1) = αn−λ−2(xλ) • αn−3(x), 2αλ−1(xn−λ) = αλ−1(xn−λ−1) • αn−3(x). (3.4.4) In the condition (3.2.3) in Theorem 3.2, replace (x, y, z) by (cid:0)αn−λ−2(xλ), αn−3(x), αλ−1(xn−λ−1)(cid:1) and use (3.4.4) and the last equality in (3.4.3). The result is: 0 = [2αn−λ−2(xλ+1), αλ(xn−λ−1)] + [2xn−1, αn−2(x)] + [2αλ−1(xn−λ), αn−λ−1(xλ)] = −2x(λ + 1) + 2x(1) + 2x(λ), from which we obtain the desired condition (3.4.2). From (3.4.2) it follows that for λ ∈ {1, . . . , n − 1}. Indeed, (3.4.5) is trivially true when λ = 1. Then we obtain (3.4.5) by a finite induction, using (3.4.2) in the inductive step. x(λ) = λx(1) (3.4.5) Since the commutator is anti-symmetric, we have x(n − λ) = −x(λ). Combining the previous line with the λ = n − 1 case of (3.4.5), we have which implies −x(1) = x(n − 1) = (n − 1)x(1), x(1) = 0. Putting this back into (3.4.5) completes the proof of the Proposition. (cid:3) 12 DONALD YAU In the rest of this section, we discuss the relationships between third Hom-power associative algebras and Hom-Lie admissible algebras. These results are not used in later sections. Let us first recall the Hom-type analogues of Lie (admissible) algebras from [11]. Definition 3.5. Let (A, µ, α) be a Hom-algebra. (1) Define its Hom-Jacobian JA : A⊗3 → A as JA = µ ◦ (µ ⊗ α) ◦ (Id + π + π2), where π is the cyclic permutation (1 2 3). (2) A is called a Hom-Lie algebra if µ is anti-symmetric and if JA = 0 (Hom-Jacobi identity). (3) Define the commutator Hom-algebra as A− = (A, [, ], α), where [, ] = µ − µop is the com- mutator. (4) A is called a Hom-Lie admissible algebra if A− is a Hom-Lie algebra, i.e., JA− = 0. One way to understand Hom-Lie admissible algebras is through the classes of G-Hom-associative algebras, which we now recall from [11]. The classes of G-associative algebras were first defined in [7]. Definition 3.6. Let (A, µ, α) be a Hom-algebra and G be a subgroup of the symmetric group S3. Then A is called a G-Hom-associative algebra if it satisfies (−1)ε(σ)asA ◦ σ = 0, X σ∈G (3.6.1) where asA is the Hom-associator (Definition 3.1). Clearly Hom-associative algebras are exactly the {Id}-Hom-associative algebras. Let A3 be the subgroup {Id, π, π2} of S3 generated by the cyclic permutation π = (1 2 3). Then Hom-Lie algebras are exactly the A3-Hom-associative algebras whose multiplications are anti-symmetric. Moreover, as shown in [11], every G-Hom-associative algebra is a Hom-Lie admissible algebra. Conversely, a Hom-Lie admissible algebra is an S3-Hom-associative algebra. Examples of and construction results along the lines of Corollary 2.6 for G-Hom-associative algebras can be found in [16]. It follows from Corollary 3.3 that Hom-associative algebras are all third Hom-power associative. So some Hom-Lie admissible algebras are third Hom-power associative. This raises the question: Exactly which Hom-Lie admissible algebras are third Hom-power associative? We answer this question in Theorem 3.9. The answer turns out to be the class of A3-Hom-associative algebras. The following two preliminary results are used in the proof of Theorem 3.9. The next result describes A3-Hom-associative algebras in terms of its cyclic Hom-associator. Lemma 3.7. Let (A, µ, α) be a Hom-algebra. Then A is an A3-Hom-associative algebra if and only if SA = 0, where SA is the cyclic Hom-associator. Proof. Every element in the subgroup A3 = {Id, π, π2} of S3 is even. So we have (−1)ε(σ)asA ◦ σ = asA ◦ (Id + π + π2) = SA, X σ∈A3 which proves the Lemma. (cid:3) HOM-POWER ASSOCIATIVE ALGEBRAS 13 The next result describes the Hom-Jacobian of the commutator Hom-algebra in terms of the cyclic Hom-associator. Lemma 3.8. Let (A, µ, α) be a Hom-algebra and A− = (A, [, ], α) be its commutator Hom-algebra. Then where JA− is the Hom-Jacobian of A−. JA− = SA ◦ (Id − (2 3)), Proof. Expanding in terms of µ, a simple calculation shows that the twelve terms in JA− are exactly the same as the twelve terms in SA ◦ (Id − (2 3)). (cid:3) The following result says that, within the class of Hom-algebras, the intersection of the classes of third Hom-power associative algebras and of Hom-Lie admissible algebras is equal to the class of A3- Hom-associative algebras. Theorem 3.9. Let (A, µ, α) be a Hom-algebra. Then A is an A3-Hom-associative algebra if and only if it is both a third Hom-power associative algebra and a Hom-Lie admissible algebra. Proof. First assume that A is an A3-Hom-associative algebra. Then it is Hom-Lie admissible [11]. Moreover, by Lemma 3.7 we have SA = 0. In particular, SA is totally anti-symmetric, so Theorem 3.2 implies that A is third Hom-power associative. Conversely, assume that A is both a third Hom-power associative algebra and a Hom-Lie admissible algebra. Since A is third Hom-power associative, by Theorem 3.2, SA is totally anti-symmetric, which implies The previous line and Lemma 3.8 imply 1 2 SA = (SA − SA ◦ (2 3)) = 1 2 JA−. SA = −SA ◦ (2 3). Since A is Hom-Lie admissible, we have JA− = 0, and hence SA = 0. Lemma 3.7 now says that A is A3-Hom-associative. (cid:3) 4. Fourth Hom-power associativity The purpose of this section is to study fourth Hom-power associativity. Using certain Hom-alternative and Hom-Jordan algebras, we also demonstrate that in a Hom-power associative algebra, it is not necessarily the case that every element generates a Hom-associative sub-Hom-algebra. Let us first record some consequences of third Hom-power associativity regarding fourth Hom-power associativity. The following result is the special case of Proposition 3.4 with n = 4 and λ = 1. It says that up to third Hom-power associativity implies part of fourth Hom-power associativity. Corollary 4.1. Let (A, µ, α) be a multiplicative third Hom-power associative algebra. Then x4 = x3α2(x) = α2(x)x3 (4.1.1) for all x ∈ A. The following consequence of Corollary 4.1 provides a characterization of up to fourth Hom-power associativity in terms of two identities. 14 DONALD YAU Corollary 4.2. Let (A, µ, α) be a multiplicative Hom-algebra. Then the following statements are equiv- alent. (1) A satisfies for all x ∈ A. (2) A satisfies for all x ∈ A. x2α(x) = α(x)x2 and x4 = α(x2)α(x2) as(x, x, x) = 0 = as(x2, α(x), α(x)) (4.2.1) (4.2.2) (3) A is up to fourth Hom-power associative. Proof. The identity as(x, x, x) = 0 is equivalent to the first identity in (4.2.1) by the definition of the Hom-associator. Moreover, since A is multiplicative, the identity as(x2, α(x), α(x)) = 0 is equivalent to: 0 = as(x2, α(x), α(x)) = (x2α(x))α2(x) − α(x2)(α(x)α(x)) = x4 − α(x2)α(x2). Therefore, (4.2.1) and (4.2.2) are equivalent. Next we show the equivalence between the first and the last statements. The first equality in (4.2.1) is third Hom-power associativity, and the second equality is part of fourth Hom-power associativity. Conversely, suppose (4.2.1) is true. Then A is multiplicative third Hom-power associative. By Corollary 4.1 we have x4 = x3α2(x) = α(x2)α(x2) = α2(x)x3, which means that A is fourth Hom-power associative. (cid:3) In Corollary 5.2 below, we will see that the Hom-algebras in Corollary 4.2 are, in fact, Hom-power associative. Next we discuss some classes of fourth Hom-power associative algebras. Definition 4.3. Let (A, µ, α) be a Hom-algebra. (1) A is called a right Hom-alternative algebra [10] if it satisfies for all x, y ∈ A. (2) A is called a non-commutative Hom-Jordan algebra if A is Hom-flexible, i.e., as(x, y, x) = asA(y, x, x) = 0 (4.3.1) 0, and if for all x, y ∈ A. asA(x2, α(y), α(x)) = 0 (4.3.2) (3) A is called a Hom-Jordan algebra [19] if A is commutative and satisfies (4.3.2) for all x, y ∈ A. Construction results for these Hom-algebras along the lines of Theorem 2.4 and Corollaries 2.5 and 2.6 can be easily formulated and proved. Right-alternative and (non-commutative) Jordan algebras are equivalent to right Hom-alternative and (non-commutative) Hom-Jordan algebras with identity twisting maps. The conditions (4.3.1) and (4.3.2) are called the right Hom-alternative identity HOM-POWER ASSOCIATIVE ALGEBRAS 15 and the Hom-Jordan identity, respectively. They are the Hom-type generalizations of the usual right-alternative identity and the Jordan identity. Every Hom-alternative algebra is a right Hom-alternative algebra. Every Hom-Jordan algebra is a non-commutative Hom-Jordan algebra, since commutativity implies Hom-flexibility. Analogous to the classical case, it is proved in [19] that multiplicative Hom-alternative algebras are both Hom-Jordan admissible and Hom-Maltsev admissible. We should point out that the Hom-Jordan identity above is different from the one defined in [10]. Proposition 4.4. Let (A, µ, α) be a multiplicative Hom-algebra that is either a right Hom-alternative algebra or a non-commutative Hom-Jordan algebra. Then A is up to fourth Hom-power associative. Proof. By Corollary 4.2 we only need to show that A satisfies the two identities in (4.2.2). The identity as(x, x, x) = 0 is satisfied in both cases because of either right Hom-alternativity (4.3.1) or Hom- flexibility. The other identity in (4.2.2), as(x2, α(x), α(x)) = 0, can be obtained from either (i) the right Hom-alternative identity (4.3.1) by replacing (x, y) by (α(x), x2), or from (ii) the Hom-Jordan identity (4.3.2) by setting y = x. (cid:3) In Corollary 5.6 below, we will see that the Hom-algebras in Proposition 4.4 are, in fact, Hom-power associative. Using Hom-alternative and Hom-Jordan algebras, let us point out a major conceptual difference between power associative algebras and Hom-power associative algebras. Recall from the introduction that a power associative algebra can be equivalently defined as an algebra in which every element generates an associative sub-algebra. In contrast, in a general Hom-power associative algebra, it is not true that every element generates a Hom-associative sub-Hom-algebra. This latter condition is too restrictive, and it excludes some Hom-alternative and Hom-Jordan algebras, all of which are Hom-power associative by Corollary 5.6. In fact, in a Hom-algebra (A, µ, α), the sub- Hom-algebra generated by an element x includes αk(x) for all k ≥ 1. However, the Hom-associator (Definition 3.1) asA(x, α(x), α2(x)) = (xα(x))α3(x) − α(x)(α(x)α2(x)) (4.4.1) does not need to be equal to 0 in a Hom-alternative algebra or a Hom-Jordan algebra. When this par- ticular Hom-associator is not equal to 0, the sub-Hom-algebra generated by x is not a Hom-associative algebra. We illustrate this in the following two examples. Example 4.5. In this example, we exhibit a multiplicative Hom-alternative algebra (hence a Hom- power associative algebra by Corollary 5.6) in which the Hom-associator in (4.4.1) is not equal to 0 for some element x. The octonions O [4] is constructed from the field k of real numbers using the Cayley-Dickson pro- cess three times with the parameter −1 in each step. More concretely, the octonions O is an eight- dimensional alternative (but not associative) algebra with a basis {e0, . . . , e7} and the following multi- plication table, where µ denotes the multiplication in O. 16 DONALD YAU e6 e6 e4 e3 e2 e1 µ e0 e3 e4 e2 e0 e0 e1 e7 −e2 e4 e1 −e0 e1 e2 −e4 −e0 e2 e5 e3 −e7 −e5 −e0 e3 e4 e4 e5 e5 −e6 e6 e6 e7 e7 e7 e5 e5 e7 e6 −e5 −e3 e7 −e6 e1 e3 −e5 e1 e4 e2 e4 −e3 −e1 −e0 e5 −e4 −e2 −e0 e1 −e3 e6 e2 −e1 −e6 −e0 e2 −e4 e7 e3 −e2 −e7 −e0 e5 −e7 e3 e6 −e1 There is an algebra automorphism α : O → O given by α(e0) = e0, α(e1) = e5, α(e2) = e6, α(e3) = e7, α(e4) = e1, α(e5) = e2, α(e6) = e3, α(e7) = e4. (4.5.1) Using [10] (Theorem 3.1), which is the analogue of Corollary 2.6 for Hom-alternative algebras, we obtain a multiplicative Hom-alternative algebra with the following multiplication table. Oα = (O, µα = αµ, α) µα e0 e1 e2 e3 e4 e5 e6 e7 e6 e3 e4 e3 e2 e1 e0 e7 e1 e6 e0 e5 e4 −e6 e1 e5 −e0 e6 −e1 −e0 e2 e7 −e4 −e2 −e0 e1 e2 −e3 e3 e4 e7 e5 e2 e4 e3 −e2 −e7 e4 −e3 e5 e7 −e2 e5 e1 e6 e1 −e7 −e5 −e0 e2 −e1 −e6 −e0 e5 −e7 e3 e6 −e5 −e3 −e0 e6 −e1 e4 e7 −e6 −e4 −e0 e2 −e4 e7 e3 −e5 By Corollary 5.6 below, Oα is a Hom-power associative algebra. On the other hand, one can check that asOα(e1, α(e1), α2(e1)) = −2e1 6= 0. Therefore, the sub-Hom-algebra in the Hom-power associative algebra Oα generated by e1 is not Hom- associative. (cid:3) Example 4.6. In this example, we exhibit a multiplicative Hom-Jordan algebra (hence a Hom-power associative algebra by Corollary 5.6) in which the Hom-associator in (4.4.1) is not equal to 0 for some element x. Consider the 27-dimensional exceptional simple Jordan algebra M 8 3 [13]. The elements of M 8 3 are 3 × 3 Hermitian octonionic matrices, i.e., matrices of the form    a1 x  y y z a3 x a2 z X = with each ai ∈ k and x, y, z ∈ O (see Example 4.5). Here we are using the convention ai = aie0 for the diagonal elements. For an octonion x = P7 i=0 biei with each bi ∈ k, its conjugate is defined as the HOM-POWER ASSOCIATIVE ALGEBRAS 17 octonion x = b0e0 − 7 X i=1 biei. This k-module M 8 3 becomes a Jordan algebra with the multiplication where XY and Y X are the usual matrix multiplication. X ∗ Y = 1 2 (XY + Y X), Let α : O → O be the algebra automorphism in (4.5.1), which is both unit-preserving and conjugate- preserving, i.e., α(e0) = e0 and α(x) = α(x) for all x ∈ O. It extends entrywise to a linear automorphism α : M 8 3 . This extended map α respects matrix multiplication and the Jordan product ∗, i.e., α is an algebra automorphism on (M 8 3 , ∗). Using the Hom-Jordan analogue of Corollary 2.6 (= Theorem 5.8 in [19]), we obtain a multiplicative Hom-Jordan algebra 3 → M 8 (M 8 3 )α = (M 8 3 , ∗α = α ◦ ∗, α). By Corollary 5.6 below, (M 8 that 3 )α is a Hom-power associative algebra. On the other hand, one can check for the element as(M 8 3 )α(X, α(X), α2(X)) 6= 0 X = 1  −e2  −e3 e2 0 0 e3 0 0  .  Therefore, the sub-Hom-algebra in the Hom-power associative algebra (M 8 Hom-associative. 3 )α generated by X is not (cid:3) The rest of this section is devoted to studying the identity x4 = α(x2)α(x2), which is part of fourth Hom-power associativity. To state the result below, we introduce the following notations to represent the linearized forms of x4 = ((xx)α(x))α2(x) and α(x2)α(x2) = α(xx)α(xx). Definition 4.7. Let (A, µ, α) be a Hom-algebra, and let x, y, z, and w be elements in A. Regard the permutations in S4 as permutations of the letters (x, y, z, w). Consider the subsets IL = {(1 2 3)i(1 2 3 4)j : 0 ≤ i ≤ 2, 0 ≤ j ≤ 3}, IR = {Id, (1 3), (2 3), (1 4), (2 4), (1 3)(2 4)} of S4. Let PL and PR denote the sums indexed by the sets IL and IR, respectively. Define the elements FL(x, y, z, w) = X L FR(x, y, z, w) = X R ((x • y)α(z)) α2(w), α(x • y)α(z • w), F (x, y, z, w) = FL(x, y, z, w) − FR(x, y, z, w) in A, where x • y = xy + yx as in (3.1.2). Note that x • y = y • x and that there are twelve terms in PL and six terms in PR. The twelve terms in PL are the twelve permutations of the four variables in the expression ((x • y)α(z)) α2(w). Likewise, the six terms in PR are the six permutations of the four variables in the expression α(x • y)α(z • w). 18 DONALD YAU The following result is the Hom-version of [1] (Lemma 3). It gives a linearized form of the identity x4 = α(x2)α(x2). It is also a crucial ingredient in the proof of Theorem 5.1. Theorem 4.8. Let (A, µ, α) be a Hom-algebra. Then A satisfies x4 = α(x2)α(x2) for all x ∈ A if and only if for all x, y, z, w ∈ A. Proof. For x ∈ A, let F (x, y, z, w) = 0 (4.8.1) B(x) = (x2α(x))α2(x) − α(x2)α(x2). The "if" part follows from the identity and the definition x4 = (x2α(x))α2(x). F (x, x, x, x) = 24B(x) For the "only if" part, assume that B(x) = 0 for all x ∈ A. For λ ∈ k and y ∈ A, we have 0 = B(x + λy) = B(x) + λC(x, y) + λ2D(x, y) + λ3C(y, x) + λ4B(y) = λC(x, y) + λ2D(x, y) + λ3C(y, x). (4.8.2) Here C(x, y) is some expression involving x, y, and α, and D(x, y) is D(x, y) = (y2α(x))α2(x) + (x2α(y))α2(y) + ((x • y)α(x)) α2(y) + ((x • y)α(y)) α2(x) − α(x • y)α(x • y) − α(x2) • α(y2). Substituting λ = ±1 in (4.8.2) and adding the results, one infers that D(x, y) = 0 for all x, y ∈ A. Next, for η ∈ k and z ∈ A, we have 0 = D(x + ηz, y) = D(x, y) + ηE(x, y, z) + η2D(z, y) = ηE(x, y, z). Setting η = 1, this implies E(x, y, z) = 0 for all x, y, z ∈ A. A simple calculation shows that E(x, y, z) = (cid:0)y2α(x)(cid:1) α2(z) + (cid:0)y2α(z)(cid:1) α2(x) + ((x • z)α(y)) α2(y) + ((x • y)α(z)) α2(y) + ((z • y)α(x)) α2(y) + ((x • y)α(y)) α2(z) + ((z • y)α(y)) α2(x) − α(x • y) • α(z • y) − α(x • z) • α(y2). Finally, for ν ∈ k and w ∈ A, we have 0 = E(x, y + νw, z) = E(x, y, z) + νF (x, y, z, w) + ν2E(x, w, z) = νF (x, y, z, w), where F (x, y, z, w) is as defined in Definition 4.7. Setting ν = 1, this implies F (x, y, z, w) = 0 for all x, y, z, w ∈ A, as desired. (cid:3) HOM-POWER ASSOCIATIVE ALGEBRAS 19 5. Hom-power associativity The purpose of this section is to show that, in the presence of multiplicativity, up to fourth Hom- power associativity implies Hom-power associativity, generalizing a well-known result of Albert [1] about power associative algebras. The desired result is a consequence of the following result, which is the Hom-version of [1] (Lemma 4). Theorem 5.1. Let (A, µ, α) be a multiplicative up to (n − 1)st Hom-power associative algebra for some n ≥ 5. Then A is up to nth Hom-power associative. Before we give the proof of Theorem 5.1, let us record some of its consequences. Corollary 5.2. Let (A, µ, α) be a multiplicative Hom-algebra. Then the following statements are equiv- alent. (1) A satisfies for all x ∈ A. (2) A satisfies for all x ∈ A. x2α(x) = α(x)x2 and x4 = α(x2)α(x2) as(x, x, x) = 0 = as(x2, α(x), α(x)) (3) A is up to fourth Hom-power associative. (4) A is Hom-power associative. Proof. The equivalence of the first three statements is Corollary 4.2. On the other hand, if A is multiplicative up to fourth Hom-power associative, then we apply Theorem 5.1 repeatedly to conclude that A is up to nth Hom-power associative for all n ≥ 5, i.e., that A is Hom-power associative. (cid:3) Setting α = Id in Corollary 5.2, we recover the following result of Albert [1] (Theorem 2). Corollary 5.3. An algebra A is power associative if and only if it satisfies x2x = xx2 and (x2x)x = x2x2 for all x ∈ A. In the presence of Hom-flexibility (e.g., when the multiplication is commutative), Hom-power asso- ciativity requires only one identity. Corollary 5.4. Let (A, µ, α) be a multiplicative Hom-flexible algebra. Then A is Hom-power associative if and only if it satisfies x4 = α(x2)α(x2) for all x ∈ A. Proof. This is an immediate consequence of Corollaries 3.3 and 5.2. (cid:3) Setting α = Id in Corollary 5.4, we recover the following result of Albert [1] (Theorem 1). 20 DONALD YAU Corollary 5.5. Suppose A is a flexible (e.g., commutative) algebra. Then A is power associative if and only if it satisfies (x2x)x = x2x2 for all x ∈ A. Recall the two types of Hom-algebras defined in Definition 4.3. Corollary 5.6. Multiplicative right Hom-alternative and multiplicative non-commutative Hom-Jordan algebras are Hom-power associative. Proof. This is an immediate consequence of Proposition 4.4 and Corollary 5.2. (cid:3) Setting the twisting map to be the identity map in Corollary 5.6, we obtain the following well-known result for ordinary algebras [13]. Corollary 5.7. Right-alternative algebras and non-commutative Jordan algebras are power associative. The rest of this section is devoted to the proof of Theorem 5.1, which involves a series of Lemmas. To simplify the presentation, we use the notation xi,j = αj−1(xi)αi−1(xj ) (5.7.1) for an element x in a Hom-algebra (A, µ, α) and positive integers i and j, where xi and xj are the ith and the jth Hom-powers (2.3.1). With this notation, nth Hom-power associativity (2.3.2) can be stated as xn = xn−i,i (5.7.2) for all elements x and 1 ≤ i ≤ n − 1. The equality (5.7.2) is exactly what we need to prove for Theorem 5.1. Several equalities are used repeatedly below. In particular, in Theorem 5.1 the assumption of up to (n − 1)st Hom-power associativity of A means xm = xm−i,i for all x ∈ A, 2 ≤ m < n, and 1 ≤ i ≤ m − 1. Also, by the definition of the Hom-powers we have for all k ≥ 2. Moreover, the conclusion of Proposition 3.4 can be stated as xk = xk−1αk−2(x) = xk−1,1 xn−i,i = xi,n−i for all x ∈ A and 1 ≤ i ≤ n − 1. In the following Lemmas, we derive identities involving xn and xn−i,i for 2 ≤ i ≤ 5. Then we play these identities against each other to obtain the nth Hom-power associativity of A. Our first objective is to prove the special case of Theorem 5.1 when n = 5, which requires the following result. Lemma 5.8. Let (A, µ, α) be a multiplicative up to (n − 1)st Hom-power associative algebra for some n ≥ 5. Then xn−3,3 = 4xn−2,2 − 3xn (5.8.1) for all x ∈ A. HOM-POWER ASSOCIATIVE ALGEBRAS 21 Proof. Pick x ∈ A. The desired identity (5.8.1) is equivalent to the following special case of (4.8.1): F (cid:0)αn−4(x), αn−4(x), αn−4(x), xn−3(cid:1) = 0. (5.8.2) Indeed, using the assumptions on A, the definition (2.3.1) of the Hom-powers, the notations above and those in Definition 4.7, and the fact that three of the four arguments in F in (5.8.2) are the same, we have: FL (cid:0)αn−4(x), αn−4(x), αn−4(x), xn−3(cid:1) = 3(cid:0)(2αn−4(x)αn−4(x))αn−3(x)(cid:1) α2(xn−3) + 3(cid:0)(2αn−4(x)αn−4(x))α(xn−3)(cid:1) αn−2(x) + 3(cid:0)2(xn−3 • αn−4(x))αn−3(x)(cid:1) αn−2(x) = 6x3,n−3 + 6(cid:0)x2,n−3 + 2xn−2,1(cid:1) αn−2(x) = 6xn−3,3 + 6(3xn−1)αn−2(x) = 6xn−3,3 + 18xn. A similar calculation yields: FR (cid:0)αn−4(x), αn−4(x), αn−4(x), xn−3(cid:1) = 3α(cid:0)2αn−4(x)αn−4(x)(cid:1) α(cid:0)αn−4(x) • xn−3(cid:1) + 3α(cid:0)αn−4(x) • xn−3(cid:1) α(cid:0)2αn−4(x)αn−4(x)(cid:1) = 12x2,n−2 + 12xn−2,2 = 24xn−2,2. Therefore, (5.8.2) is equivalent to 6xn−3,3 + 18xn = 24xn−2,2, which in turn is equivalent to the desired equality (5.8.1). (cid:3) Using Lemma 5.8 we now prove the special case of Theorem 5.1 when n = 5. Lemma 5.9. Let A be a multiplicative up to fourth Hom-power associative algebra. Then A is up to fifth Hom-power associative. Proof. To show that A is fifth Hom-power associative, pick x ∈ A. By the n = 5 case of (3.4.1) we have The n = 5 case of Lemma 5.8 says that x5 = x4,1 = x1,4 and x3,2 = x2,3. This implies that and hence We conclude that x2,3 = 4x3,2 − 3x5. 3x5 = 4x3,2 − x2,3 = 3x3,2, x5 = x3,2 = x2,3. x5 = xi,5−i for 1 ≤ i ≤ 4, so A is fifth Hom-power associative, as desired. (cid:3) The following result relates xn−4,4 to xn−2,2 and xn. 22 DONALD YAU Lemma 5.10. Let (A, µ, α) be a multiplicative up to (n − 1)st Hom-power associative algebra for some n ≥ 5. Then xn−4,4 = 11xn−2,2 − 10xn (5.10.1) for all x ∈ A. Proof. A calculation similar to the proof of Lemma 5.8, applied to the special case of (4.8.1), yields F (cid:0)αn−5(x2), αn−4(x), αn−4(x), α(xn−4)(cid:1) = 0 6xn−4,4 + 6xn−2,2 + 12xn = 16xn−3,3 + 8xn−2,2. (5.10.2) (5.10.3) Indeed, the left and the right sides of (5.10.3) are, respectively, FL and FR applied to the same variables as in (5.10.2). Continuing (5.10.3) we have: 3xn−4,4 = 8xn−3,3 + xn−2,2 − 6xn = 8(cid:0)4xn−2,2 − 3xn(cid:1) + xn−2,2 − 6xn = 33xn−2,2 − 30xn. (by (5.8.1)) Dividing by three, we obtain the desired equality (5.10.1). (cid:3) In the next two Lemmas, we assume that n ≥ 6, which is sufficient in view of Lemma 5.9. In the following result, part of the nth Hom-power associativity of A is established. Lemma 5.11. Let (A, µ, α) be a multiplicative up to (n − 1)st Hom-power associative algebra for some n ≥ 6. Then xn = xn−2,2 (5.11.1) for all x ∈ A. Proof. Pick x ∈ A. A calculation similar to the proof of Lemma 5.8, applied to the special case of (4.8.1), yields Likewise, the special case of (4.8.1) yields F (cid:0)αn−5(x2), αn−5(x2), αn−4(x), α2(xn−5)(cid:1) = 0 3xn = −6xn−2,2 + 8xn−3,3 + 4xn−4,4 − 3xn−5,5. (5.11.2) F (cid:0)αn−6(x3), αn−4(x), αn−4(x), α2(xn−5)(cid:1) = 0 Subtracting (5.11.2) from (5.11.3) and using (5.8.1) and (5.10.1), we obtain: 6xn = 4xn−2,2 − 3xn−3,3 + 8xn−4,4 − 3xn−5,5. (5.11.3) 3xn = 10xn−2,2 − 11xn−3,3 + 4xn−4,4 = 10xn−2,2 − 11(cid:0)4xn−2,2 − 3xn(cid:1) + 4(cid:0)11xn−2,2 − 10xn(cid:1) = 10xn−2,2 − 7xn. This is equivalent to 10xn = 10xn−2,2, which yields the desired equality (5.11.1). (cid:3) The following result will be used in the induction step of the proof of Theorem 5.1. HOM-POWER ASSOCIATIVE ALGEBRAS 23 Lemma 5.12. Let (A, µ, α) be a multiplicative up to (n − 1)st Hom-power associative algebra for some n ≥ 6. Suppose k is an integer in the range 1 ≤ k ≤ n/2 − 2. Then 3xn−(k+2),k+2 = −2xn + 8xn−(k+1),k+1 − 3xn−k,k (5.12.1) for all x ∈ A. Proof. Pick x ∈ A. A calculation similar to the proof of Lemma 5.8, applied to the special case of (4.8.1), yields F (cid:0)αn−k−3(xk), αn−4(x), αn−4(x), αk−1(xn−k−2)(cid:1) = 0 3xn−(k+2),k+2 = −6xn + 8xn−(k+1),k+1 − 3xn−k,k + 4xn−2,2. To finish the proof, one applies Lemma 5.11 to replace xn−2,2 by xn in the previous line. (cid:3) We can now finish the proof of Theorem 5.1 Proof of Theorem 5.1. The n = 5 case of Theorem 5.1 was established in Lemma 5.9. So we may assume, without lost of generality, that n ≥ 6. Pick x ∈ A. We must prove the nth Hom-power associativity condition xn = xn−i,i (5.12.2) for 1 ≤ i ≤ n − 1. In view of Proposition 3.4, we only need to prove (5.12.2) when 1 ≤ i ≤ n/2. We prove this by a finite induction. The i = 1 case of (5.12.2) is true by the definition of the Hom-power xn. The i = 2 case of (5.12.2) is true by Lemma 5.11. For the induction step, suppose 1 ≤ k ≤ n/2 − 2 and that (5.12.2) is true for 1 ≤ i ≤ k + 1. We want to prove (5.12.2) when i = k + 2. By Lemma 5.12 and the induction hypothesis, we have: 3xn−(k+2),k+2 = −2xn + 8xn−(k+1),k+1 − 3xn−k,k = −2xn + 8xn − 3xn = 3xn. Dividing by three in the above calculation yields (5.12.2) when i = k + 2, finishing the induction and the proof of Theorem 5.1. (cid:3) References [1] A.A. Albert, On the power-associativity of rings, Summa Brasil. Math. 2 (1948) 21-32. [2] A.A. Albert, Power-associative rings, Trans. Amer. Math. Soc. 64 (1948) 552-593. [3] M. Arenas, An algorithm for associative bilinear forms, Linear Alg. Appl. 430 (2009) 286-295. [4] J.C. Baez, The octonions, Bull. Amer. Math. Soc. 39 (2002) 145-205. [5] K. Carmody, Circular and hyperbolic quaternions, octonions, and sedenions, Appl. Math. Comp. 28 (1988) 47-72. [6] R.E. Cawagas, Loops embedded in generalized Cayley algebras of dimension 2r , r ≥ 2, Int. J. Math. Math. Sci. 28 (2001) 181-187. [7] M. Goze and E. Remm, Lie-admissible algebras and operads, J. Algebra 273 (2004) 129-152. [8] J.T. Hartwig, D. Larsson, and S.D. Silvestrov, Deformations of Lie algebras using σ-derivations, J. Algebra 295 (2006) 314-361. [9] K. Imaeda and M. Imaeda, Sedenions: algebra and analysis, Appl. Math. Comp. 115 (2000) 77-88. [10] A. Makhlouf, Hom-alternative algebras and Hom-Jordan algebras, Int. Elect. J. Alg. 8 (2010) 177-190. [11] A. Makhlouf and S. Silvestrov, Hom-algebra structures, J. Gen. Lie Theory Appl. 2 (2008) 51-64. [12] A. Makhlouf and S. Silvestrov, Hom-algebras and Hom-coalgebras, arXiv:0811.0400. [13] R.D. Schafer, An introduction to nonassociative algebras, Dover Pub., New York, 1995. 24 DONALD YAU [14] M.V. Velasco, Spectral theory for non-associative complete normed algebras and automatic continuity, J. Math. Anal. Appl. 351 (2009) 97-106. [15] D. Yau, Enveloping algebras of Hom-Lie algebras, J. Gen. Lie Theory Appl. 2 (2008) 95-108. [16] D. Yau, Hom-algebras and homology, J. Lie Theory 19 (2009) 409-421. [17] D. Yau, The Hom-Yang-Baxter equation, Hom-Lie algebras, and quasi-triangular bialgebras, J. Phys. A 42 (2009) 165202 (12pp). [18] D. Yau, Hom-bialgebras and comodule Hom-algebras, Int. Elect. J. Alg. 8 (2010) 45-64. [19] D. Yau, Hom-Maltsev, Hom-alternative, and Hom-Jordan algebras, arXiv:1002.3944. [20] D. Yau, Hom-Novikov algebras, arXiv:0909.0726. [21] D. Yau, The Hom-Yang-Baxter equation and Hom-Lie algebras, arXiv:0905.1887. [22] D. Yau, The classical Hom-Yang-Baxter equation and Hom-Lie bialgebras, arXiv:0905.1890. [23] D. Yau, Infinitesimal Hom-bialgebras and Hom-Lie bialgebras, arXiv:1001.5000. [24] D. Yau, Hom-quantum groups I: quasi-triangular Hom-bialgebras, arXiv:0906.4128. [25] D. Yau, Hom-quantum groups II: cobraided Hom-bialgebras and Hom-quantum geometry, arXiv:0907.1880. [26] D. Yau, Hom-quantum groups III: representations and module Hom-algebras, arXiv:0911.5402. [27] D. Yau, On n-ary Hom-Nambu and Hom-Nambu-Lie algebras, arXiv:1004.2080. [28] D. Yau, On n-ary Hom-Nambu and Hom-Maltsev algebras, arXiv:1004.4795. [29] D. Yau, A Hom-associative analogue of n-ary Hom-Nambu algebras, arXiv:1005.2373. Department of Mathematics, The Ohio State University at Newark, 1179 University Drive, Newark, OH 43055, USA E-mail address: [email protected]
1610.07178	1	1610	2016-10-23T14:43:39	Zero product determined Lie algebras	[ "math.RA", "math.QA" ]	A Lie algebra $L$ over a field $\mathbb{F}$ is said to be zero product determined (zpd) if every bilinear map $f:L\times L\to \mathbb{F}$ with the property that $f(x,y)=0$ whenever $x$ and $y$ commute is a coboundary. The main goal of the paper is to determine whether or not some important Lie algebras are zpd. We show that the Galilei Lie algebra $\mathfrak{sl}_2\ltimes V$, where $V$ is a simple $\mathfrak{sl}_2$-module, is zpd if and only if $\dim V =2$ or $\dim V$ is odd. The class of zpd Lie algebras also includes the quantum torus Lie algebras $\mathcal{L}_q$ and $\mathcal{L}^+_q$, the untwisted affine Lie algebras, the Heisenberg Lie algebras, and all Lie algebras of dimension at most $3$, while the class of non-zpd Lie algebras includes the ($4$-dimensional) aging Lie algebra $\mathfrak {age}(1)$ and all Lie algebras of dimension more than $3$ in which only linearly dependent elements commute. We also give some evidence of the usefulness of the concept of a zpd Lie algebra by using it in the study of commutativity preserving linear maps.	math.RA	math	ZERO PRODUCT DETERMINED LIE ALGEBRAS MATEJ BRESAR, XIANGQIAN GUO, GENQIANG LIU, RENCAI L U, KAIMING ZHAO Abstract. A Lie algebra L over a field F is said to be zero product determined (zpd) if every bilinear map f : L × L → F with the property that f (x, y) = 0 whenever x and y commute is a coboundary. The main goal of the paper is to determine whether or not some important Lie algebras are zpd. We show that the Galilei Lie algebra sl2 ⋉ V , where V is a simple sl2-module, is zpd if and only if dim V = 2 or dim V is odd. The class of zpd Lie algebras also includes the quantum torus Lie algebras Lq and L+ q , the untwisted affine Lie algebras, the Heisenberg Lie algebras, and all Lie algebras of dimension at most 3, while the class of non-zpd Lie algebras includes the (4-dimensional) aging Lie algebra age(1) and all Lie algebras of dimension more than 3 in which only linearly dependent elements commute. We also give some evidence of the usefulness of the concept of a zpd Lie algebra by using it in the study of commutativity preserving linear maps. We say that a Lie algebra L over a field F is zero product determined (zpd for short) if for 1. Introduction every bilinear map f : L × L → F with the property (1.1) there exists a linear map Φ : [L, L] → F such that (1.2) [x, y] = 0 =⇒ f (x, y) = 0, ∀ x, y ∈ L, f (x, y) = Φ(cid:0)[x, y](cid:1), ∀ x, y ∈ L. Thus, borrowing the Lie algebra cohomology terminology, L is zpd if (1.1) implies that f is a coboundary. The question whether (1.1) implies (1.2) was first studied in [8], for the case where L is the Lie algebra of all n × n matrices. The answer, which is easily seen to be positive, was used as a crucial tool in describing commutativity preserving linear maps on matrix algebras. The concept of a zpd algebra was introduced slightly later, in [7], not only for Lie algebras but for general, not necessarily associative algebras (in the definition, one just replaces [· , · ] in (1.1) and (1.2) by the product in the algebra in question). So far, however, the focus was primarily on associative algebras -- not only in pure algebra where the concept is quite well understood, especially in finite dimensions [5], but also, and in fact more extensively, in functional analysis where one additionally assumes that the maps in question are continuous (see [1] and the subsequent series of papers on the so-called Banach algebras with property B). Not much has been known, so far, about which Lie algebras are zpd. This question was studied sporadically in some papers, including the seminal work [7], and more systematically in [11] and [20] (and [2] in the analytic context). In [20] it was shown that parabolic subalgebras of complex finite Key words and phrases. Zero product determined Lie algebra, zero product determined module, Galilei Lie algebra, quantum torus Lie algebra, Heisenberg Lie algebra, affine Lie algebra, commutativity preserving map. Mathematics Subject Classification. 17B05, 17B40, 17B65, 17B99. 1 2 MATEJ BRESAR, XIANGQIAN GUO, GENQIANG LIU, RENCAI L U, KAIMING ZHAO dimensional simple Lie algebras are zpd. This readily implies that complex finite dimensional semisimple Lie algebras are zpd, to point out one basic example. For infinite dimensional Lie algebras, as well as for other finite dimensional Lie algebras, the question is largely open. The aim of this paper is to remedy this situation. We will prove that various different Lie algebras are zpd, but also various different are not. Given a Lie algebra, it is usually not easy to guess whether it is zpd or not. This makes the problem challenging and, in our opinion, interesting in its own right. Another motivation for us for studying it is the striking similarity between the notion of a zpd Lie algebra and the notion of a group with trivial Bogomolov multiplier (see, e.g., [17]). Finally, as we will see at the end of the paper, knowing that a Lie algebra is zpd can be used for describing commutativity preserving linear maps. Hopefully other applications will be found later, just as they were found for zpd associative algebras (cf. [1, 5]). The paper is organized as follows. In Section 2 we survey some general facts on zpd Lie algebras, and also introduce and study zero product determined modules which play an important role in some of the subsequent sections. Sections 3-5 are rather easy; in Section 3 we show that every Lie algebra of dimension at most 3 is zpd, in Section 4 we show that a Lie algebra in which only linearly dependent commute is not zpd if its dimension is greater than 3, and in Section 5 we show that the Heisenberg Lie algebras are zpd. The next sections contain deeper results. In Section 6 we show that the Galilei Lie algebra sl2 ⋉ V , where V is a simple sl2-module, is zpd if and only if dim V = 2 or dim V is odd. We also find a 4-dimensional algebra, namely the aging algebra age(1), that is not zpd. In Section 7 we show that the quantum torus Lie algebras Lq and L+ q are zpd, and in Section 8 we show that the untwisted affine Lie algebras are zpd. The final section, Section 9, is devoted to applications of the zpd property to commutativity preserving linear maps. 2. General remarks on zpd Lie algebras In this section we gather together several elementary observations on zpd Lie algebras, some new and some already known. First a word on notation: throughout, F denotes a field, and all algebras, modules and vector spaces are assumed to be over F. Sometimes we will assume that F = C; if this is not assumed, then F is just an arbitrary field. By N and Z+ we denote the set of all positive integers and all nonnegative integers. We begin by giving some equivalent definitions of zpd Lie algebras (they are actually valid in greater generality, but we will formulate them for Lie algebras for convenience). First of all, note that the existence of a linear map Φ : [L, L] → F satisfying (1.2) is equivalent to the condition (2.1) Xi [xi, yi] = 0 =⇒ Xi f (xi, yi) = 0, ∀ xi, yi ∈ L. Indeed, if (2.1) holds, then the map Φ : [L, L] → X given by f (xi, yi) Φ(cid:16)Xi [xi, yi](cid:17) = Xi is well-defined and satisfies (1.2), and the converse is obvious. In the next lemma we show that bilinear maps from the definition may have their ranges in arbitrary vector spaces, not ZERO PRODUCT DETERMINED LIE ALGEBRAS 3 necessarily in the field of scalars (this is actually taken as the definition in the seminal paper [7]). The proof is very easy, but the fact itself is important for applications. Lemma 2.1. Let L be a zpd Lie algebra. If X is an arbitrary vector space and f : L× L → X is a bilinear map satisfying then there exists a linear map Φ : [L, L] → X such that [x, y] = 0 =⇒ f (x, y) = 0, ∀ x, y ∈ L, f (x, y) = Φ(cid:0)[x, y](cid:1), ∀ x, y ∈ L. Proof. Take a linear functional ξ on X. The map fξ := ξ◦f : L× L → F satisfies the condition that [x, y] = 0 implies fξ(x, y) = 0. Consequently, there exists a linear functional Φξ on [L, L] such that fξ(x, y) = Φξ([x, y]) for all x, y ∈ L. Thus, for all xi, yi ∈ L, Pi[xi, yi] = 0 implies Pi fξ(xi, yi) = 0, i.e., ξ(cid:0)Pi f (xi, yi)(cid:1) = 0. Since ξ is an arbitrary linear functional on X, this Given a Lie algebra L, denote by ML the kernel of the linear map L ⊗ L → L defined by actually shows that f satisfies (2.1). (cid:3) x ⊗ y 7→ [x, y], i.e., ML = nXi xi ⊗ yi ∈ L ⊗ L xi, yi ∈ L, Xi [xi, yi] = 0o. If L is finite dimensional, we see that dimML = (dim L)2 − dim[L, L]. Further, let us write KL = {x ⊗ y ∈ L ⊗ L x, y ∈ L, [x, y] = 0}. Obviously, we have We also remark that spanKL ⊆ ML. (2.2) x ⊗ y + y ⊗ x = (x + y) ⊗ (x + y) − x ⊗ x − y ⊗ y ∈ span KL, ∀ x, y ∈ L. The next lemma is essentially [9, Theorem 2.3]. Nevertheless, we give the proof for the sake of completness. Lemma 2.2. A Lie algebra L is zpd if and only if spanKL = ML. Proof. Suppose that span KL 6= ML. Then there exists a linear functional ξ on L⊗L such that ξ(KL) = {0} and ξ(ML) 6= {0}. The bilinear map f : L × L → F given by f (x, y) = ξ(x ⊗ y), therefore satisfies (1.1) but not (2.1). To prove the converse, assume that L is not zpd. Then there exists a bilinear map f : L × L → F that satisfies (1.1) but not (2.1). The linear functional ξ : L ⊗ L → F, ξ(x ⊗ y) = f (x, y), therefore vanishes on KL but not on ML. Accordingly, span KL 6= ML. (cid:3) It is even more convenient to consider the wedge product L ∧ L and define L = nXi M′ xi ∧ yi ∈ L ∧ L xi, yi ∈ L, Xi [xi, yi] = 0o, The following results are obvious. K′ L = {x ∧ y ∈ L ∧ L x, y ∈ L, [x, y] = 0}. Lemma 2.3. Let L be a Lie algebra. 4 MATEJ BRESAR, XIANGQIAN GUO, GENQIANG LIU, RENCAI L U, KAIMING ZHAO (a) L is zpd if and only if spanK′ (b) If the linear map L ∧ L → [L, L], x ∧ y 7→ [x, y] is a linear isomorphism, then L is zpd. (c) If dim L ∧ L = dim [L, L] < ∞, then L is zpd. (d) If dim (L ∧ L)/span {x ∧ yx, y ∈ L, [x, y] = 0} = dim [L, L] < ∞, then L is zpd. L = M′ L. Remark 2.4. The notion of a zpd Lie algebra may remind one on the notion of a centrally closed Lie algebra, i.e., a Lie algebra on which every 2-cocycle is a coboundary (equivalently, the second cohomology group H 2(L, F) is trivial). Let us therefore make some comment on this. Recall that a 2-cocycle on a Lie algebra L is a skew-symmetric bilinear map f : L×L → F satisfying f ([x, y], z) + f ([z, x], y) + f ([y, z], x) = 0, ∀ x, y, z ∈ L, and that a coboundary is a bilinear map f : L × L → F of the form (1.2) for some linear map Φ : [L, L] → F. A coboundary is obviously a 2-cocycle, while the converse is not always true. Setting CL = {x ⊗ y + y ⊗ x, [x, y] ⊗ z + [z, x] ⊗ y + [y, z] ⊗ x x, y, z ∈ L} one can show, just by following the proof of Lemma 2.2, that L is centrally closed ⇐⇒ spanCL = ML. The spaces spanCL and spanKL do not seem to be connected in some simple way, and the problem of showing that L is zpd is, in general, independent of the problem of showing that L is centrally closed. Given any Lie algebra L, it may be of interest to describe and compare the spaces spanCL, span KL, and ML. However, in this paper we will not touch upon this problem. The next lemma provides two conditions under which a homomorphic image of a zpd Lie algebra is again zpd. The first condition, (a), is known [9, Corollary 5.3], but we give the proof anyway. Lemma 2.5. Let L be a zpd Lie algebra. If an ideal I of L satisfies at least one of the following two conditions: (a) [L, I] = [L, L] ∩ I, (b) the Lie algebra L/I is centrally closed, then L/I is also a zpd Lie algebra. Proof. Let f : L/I × L/I → F be a bilinear map such that f (r, s) = 0 whenever r, s ∈ L/I commute. Then ¯f : L×L → F , ¯f (x, y) = f (x+I, y +I) satisfies ¯f (x, y) = 0 whenever x, y ∈ L commute. Since L is zpd it follows that (2.3) for some linear map ¯Φ : [L, L] → F . ¯f (x, y) = ¯Φ([x, y]) Suppose that (a) holds. We claim that the map Φ : [L/I, L/I] → F given by Φ(cid:16)Xi [xi + I, yi + I](cid:17) = Xi ¯Φ(cid:0)[xi, yi](cid:1) ZERO PRODUCT DETERMINED LIE ALGEBRAS 5 is well-defined. Indeed, suppose that Pi[xi, yi] ∈ I. Since this element belongs also to [L, L], (a) implies that Xi [xi, yi] = Xj [zj, uj] for some zj ∈ L and uj ∈ I. From ¯Φ([zj , uj]) = ¯f (zj, uj) = f (zj + I, uj + I) = f (zj + I, 0) = 0 we infer that Xi ¯Φ(cid:0)[xi, yi](cid:1) = ¯Φ(cid:16)Xi [xi, yi](cid:17) = ¯Φ(cid:16)Xj [zj, uj](cid:17) = Xj ¯Φ(cid:0)[zj, uj](cid:1) = 0, as desired. We therefore have f (x + I, y + I) = ¯f (x, y) = ¯Φ([x, y]) = Φ([x + I, y + I]), ∀ x, y ∈ L, proving that L/I is zpd. Assume now that (b) holds. Using (2.3) it follows immediately that that is, ¯f ([x, y], z) + ¯f ([z, x], y) + ¯f ([y, z], x) = 0, ∀ x, y, z ∈ L, Since (b) holds, the desired conclusion that f is a coboundary follows. f ([r, s], t) + f ([t, r], s) + f ([s, t], r) = 0, ∀ r, s, t ∈ L/I. (cid:3) We now extend the notion of a zpd Lie algebra to modules as follows. We will say that a module V over a Lie algebra L is zero action determined (zad for short) if for every bilinear map f : L × V → F with the property that (2.4) there exists a linear map Φ : LV → F such that (2.5) xv = 0 =⇒ f (x, v) = 0, ∀ x ∈ L, v ∈ V, f (x, v) = Φ(xv), ∀ x ∈ L, v ∈ V. Note that L is a zpd Lie algebra if and only if its adjoint module is zad. The notion of a zad module is perhaps interesting in its own right. In this paper, however, we will use it only as a tool for answering the question whether certain Lie algebras are zpd. Until the rest of this section we will consider some basic properties of zad modules. First we record the obvious analogue of Lemma 2.2. Setting MV = nXi xi ⊗ vi xi ∈ L, vi ∈ V, Xi xivi = 0o and KV = {x ⊗ v x ∈ L, v ∈ V, xv = 0}, we have Lemma 2.6. A module V over a Lie algebra L is zad if and only if span KV = MV . Let us only mention that Lemmas 2.1 and 2.5 (a) can also be easily extended to zad modules, but we do not need these results in this paper. The following elementary lemma, however, is needed. We leave the proof to the reader (cf. [9, Theorem 3.1] which states that the direct sum of Lie algebras is zpd if and only if every summand is zpd). 6 MATEJ BRESAR, XIANGQIAN GUO, GENQIANG LIU, RENCAI L U, KAIMING ZHAO Lemma 2.7. Let L be a Lie algebra and let {Vi i ∈ I} be a family of L-modules. Then their direct sum ⊕i∈I Vi is zad if and only if each Vi, i ∈ I, is zad. As an application of this lemma, we will now show that the zpd property is stable under tensoring with commutative associative unital algebras. Lemma 2.8. Let L be a zpd Lie algebra and let A be a commutative associative unital algebra. Then the Lie algebra L ⊗ A is zpd. Proof. From (x ⊗ 1 + y ⊗ a2) ⊗ (x ⊗ a1 + y ⊗ a1a2) ∈ KL⊗A, ∀ x, y ∈ L,∀ a1, a2 ∈ A, it easily follows that Now for every Pi(xi ⊗ ai) ⊗ (yi ⊗ bi) ∈ ML⊗A, we have (x ⊗ a1) ⊗ (y ⊗ a2) ≡ (x ⊗ 1) ⊗ (y ⊗ a1a2) (cid:0)mod span KL⊗A(cid:1). (xi ⊗ ai) ⊗ (yi ⊗ bi) ≡ Xi (xi ⊗ 1) ⊗ (yi ⊗ aibi) (cid:0)mod span KL⊗A(cid:1). Xi Since L is zpd, Lemma 2.7 implies that L ⊗ A is a zad module over L = L ⊗ 1. Consequently, Xi (xi ⊗ 1) ⊗ (yi ⊗ aibi) = Xj (zj ⊗ 1) ⊗ uj for some zj ∈ L, uj ∈ L ⊗ A with We have thus proved that [zj ⊗ 1, uj ] = 0, ∀j. Xi (xi ⊗ ai) ⊗ (yi ⊗ bi) ≡ 0 (cid:0)mod spanKL⊗A(cid:1), showing that L ⊗ A is a zpd Lie algebra. (cid:3) Recall that the semidirect product of a Lie algebra L and an L-module V is the Lie algebra L ⋉ V defined on the space L ⊕ V by [x + u, y + v] = [x, y] + xv − yu, ∀ x, y ∈ L,∀ u, v ∈ V. Lemma 2.9. Let L be a zpd Lie algebra and V be a zad L-module. Then L ⋉ V is a zpd Lie algebra. Proof. Take an element such that Xi (xi + ui) ⊗ (yi + vi) ∈ (L ⋉ V ) ⊗ (L ⋉ V ) [xi + ui, yi + vi] = 0. Xi Our goal is to show that this element lies in spanKL⋉V . We have Xi ([xi, yi] + xivi − yiui) = 0, ZERO PRODUCT DETERMINED LIE ALGEBRAS 7 which yields Xi [xi, yi] = 0 and Xi (xivi − yiui) = 0. Since L is a zpd Lie algebra and V is a zad L-module, we have and Xi xi ⊗ yi ∈ span KL Xi (xi ⊗ vi − yi ⊗ ui) ∈ span KV . Clearly, KL,KV ⊆ KL⋉V . Since, by (2.2), yi ⊗ ui + ui ⊗ yi ∈ span KL⋉V , and trivially ui ⊗ vi ∈ KL⋉V , it follows from Xi (xi + ui)⊗ (yi + vi) = Xi that Pi xi ⊗ yi +Xi (xi + ui) ⊗ (yi + vi) ∈ span KL⋉V . ui ⊗ vi (xi ⊗ vi − yi ⊗ ui) +Xi (yi ⊗ ui + ui ⊗ yi) +Xi (cid:3) The converse of the lemma is not true in general. Example 2.10. Let b be the noncommutative Lie algebra with a basis {h, e} and multiplication given by [h, e] = e. Let A be any commutative associative unital algebra. Then the Lie algebra b⊗ A = (h⊗ A) ⋉ (e⊗ A) is always zpd by Lemma 2.8. However, e⊗ A is a zad h⊗ A-module if and only if A is zpd as an associative algebra. Remark 2.11. The definition of a zad module obviously makes sense for associative algebras, too. If A is a unital associative zpd algebra, then every unital module V over A is automatically zad. Indeed, if f : A × V → F is a bilinear map satisfying (2.4), then for any fixed v ∈ V the map fv : A × A → F, fv(x, y) = f (x, yv), satisfies the condition that xy = 0 implies fv(x, y) = 0. Since A is zpd it follows that fv(x, 1) = fv(1, x) for every x ∈ A. This means that f (x, v) = f (1, xv) for all x ∈ A, v ∈ V , showing that V is zad. Example 2.10 shows that a module over a zpd Lie algebra may not be zad. Let us show that the converse of Lemma 2.9 does hold in the special case where L has the property that elements in L commute only when they are linearly dependent. A simple example is the Lie algebra sl2. Lemma 2.12. Let L be a Lie algebra such that any two commuting elements in L are linearly dependent. If V is an L-module, then the Lie algebra L ⋉ V is zpd if and only if L is zpd and V is zad. Proof. The sufficiency follows from Lemma 2.9. Now suppose L ⋉ V is zpd. From Lemma 2.5 (a) we see that L = (L ⋉ V )/V is zpd. Next we show that V is a zad L-module. Take any bilinear map f : L × V → F with property (2.4). We can extend f to a bilinear map F : (L ⋉ V ) × (L ⋉ V ) → F as follows F (x + u, y + v) = f (x, v) − f (y, u), ∀ x, y ∈ L, u, v ∈ V. First we show that F also satisfies (2.4). 8 MATEJ BRESAR, XIANGQIAN GUO, GENQIANG LIU, RENCAI L U, KAIMING ZHAO Take any x, y ∈ L, u, v ∈ V with [x + u, y + v] = 0. If x = 0, then we have [x + u, y + v] = −yu = 0. Consequently, f (y, u) = 0, which implies F (x + u, y + v) = 0 in this case. Now suppose x 6= 0. Then [x + u, y + v] = [x, y] + xv − yu = 0 implies [x, y] = 0 and xv − yu = 0. By the assumption, there exists a ∈ F such that y = ax. Consequently, xv − yu = x(v − au) = 0. Hence F (x + u, y + v) = f (x, v) − f (y, u) = f (x, v − au) = 0 by the assumption on f . Since we have assume that L ⋉ V is zpd, there is a linear map Ψ : L ⋉ V → F such that F (x + u, y + v) = Ψ([x + u, y + v]), ∀ x, y ∈ L, u, v ∈ V. In particular, Let Φ be the restriction of Ψ on LV . Then (2.5) holds and V is a zad L-module. f (x, v) = F (x + 0, 0 + v) = Ψ(xv), ∀ x ∈ L, v ∈ V. (cid:3) 3. Lie algebras of low dimensions In this section we will prove the following Proposition 3.1. All Lie algebras over any field F of dimension ≤ 3 are zpd. Proof. From Lemma 2.3 we see that all one or two dimensional Lie algebras are zpd. Now suppose L is a Lie algebra over any field F of dimension 3. If it is a direct sum of two nonzero ideals which have to be one or two dimensional Lie algebras, it follows that L is zpd. Now we assume that L cannot be a direct sum of two nonzero ideals. Case 1: [L, L] = L. It is easy to see that dim L ∧ L = 3. Then the linear map L ∧ L → [L, L], x ∧ y 7→ [x, y], is a linear isomorphism. From Lemma 2.3 (b) we know that L is zpd in this case. Case 2: dim[L, L] = 2. Let [L, L] = Fx ⊕ Fy and x, y, z form a basis for L. If [x, y] = 0 it follows that dim(cid:16)(L ∧ L)/span {u ∧ vu, v ∈ L, [u, v] = 0}(cid:17) = dim[L, L]. From Lemma 2.3 (d) we see that L is zpd. Next we may assume that [x, y] = y. If [z, Fx + Fy] = Fx, there is a nonzero cx + dy such that [z, cx + dy] = 0. If [z, Fx + Fy] 6= Fx, then [z, ax + by] = y for some a, b ∈ F. For all cases we have [z − b−1x, ax + by] = 0, if b = 0 it follows that [z + ax, y] = 0. If b 6= 0 we know that dim(cid:16)(L ∧ L)/span {u ∧ vu, v ∈ L, [u, v] = 0}(cid:17) = dim[L, L]. From Lemma 2.3 (d) we see that L is zpd. Case 3: dim[L, L] = 1. Let [L, L] = Fw and u, v, w form a basis for L. We may assume that [u, v] = aw, [u, w] = bw, [v, w] = cw, for some a, b, c ∈ F. ZERO PRODUCT DETERMINED LIE ALGEBRAS 9 If b = c = 0 we see that L is the 3-dimensional Heisenberg algebra which is zpd. This can be easily checked; on the other hand, we will prove this in Section 5. Now we assume that b 6= 0 or c 6= 0. By symmetry we may assume that b 6= 0. By re-choosing v we have [u, v] = 0, [u, w] = bw, [v, w] = cw, By re-choosing v again, we have [u, v] = 0, [u, w] = bw, [v, w] = 0. We see that L is a direct sum of two ideals Fv and span{u, w}, which contradicts the assump- tion on L. Thus L is zpd in this case. This completes the proof. (cid:3) From this proposition we know that the lowest dimension of a non-zpd Lie algebra is possibly 4. In Section 6 we will construct such examples. 4. Lie algebras in which only linearly dependent elements commute The purpose of this section is to describe zpd Lie algebras L such that any two commuting elements in L are linearly dependent. We first need a technical lemma: Lemma 4.1. Suppose L is a zpd Lie algebra such that any two commuting elements in L are linearly dependent. Then any subspace of L with dimension > 3 is a subalgebra of L. Proof. Since L is zpd, we see that every element in L ⊗ L of the form [x, y] ⊗ z + [z, x] ⊗ y + [y, z] ⊗ x can be written as a linear combination of elements of the form u ⊗ u, u ∈ L. Since elements of the latter form are fixed points of the linear transformation of L ⊗ L: r ⊗ s 7→ s ⊗ r, the same should be true for elements of the former form. Accordingly, [x, y] ⊗ z + [z, x] ⊗ y + [y, z] ⊗ x = z ⊗ [x, y] + y ⊗ [z, x] + x ⊗ [y, z] for all x, y, z ∈ L. If x, y, z are linearly independent, it follows that [x, y] lies in the linear span of x, y, z (see, e.g., [4, Lemma 4.11]) and hence the space spanned by x, y, z is a subalgebra of L. Now take any subspace L′ ⊆ L with dim L′ > 3. For any x, y ∈ L′ such that [x, y] 6= 0, there exists z ∈ L′ such that x, y, z are linearly independent. By the previous argument, we see that the subspace spanned by x, y, z is a subalgebra of L. Hence [x, y] ∈ span{x, y, z} ⊆ L′, which is a subalgebra of L as desired. (cid:3) Theorem 4.2. Let L be a zpd Lie algebra such that any two commuting elements in L are linearly dependent. Then L is isomorphic to a 3-dimensional simple Lie algebra, or the 2- dimensional noncommutative Lie algebra, or the 1-dimensional Lie algebra. Proof. By the assumption on L, we see that any bilinear map f on L satisfying the property (1.1) is equivalent to a linear map on L ∧ L. Hence L is zpd if and only if the canonical map π : L ∧ L → [L, L], x ∧ y 7→ [x, y], is a vector space isomorphism. If L is finite dimensional, we have dim L ∧ L = dim[L, L], which holds only if dim L 6 3. 10 MATEJ BRESAR, XIANGQIAN GUO, GENQIANG LIU, RENCAI L U, KAIMING ZHAO If dim L = 2, since L cannot be commutative, it has to be the noncommutative 2-dimensional Lie algebra. Now suppose dim L = 3. Then dim[L, L] = dim L ∧ L = 3. Suppose L is not simple and let I be a nonzero proper ideal of L. If dimI = 1, there exists x, y ∈ L such that L = I ⊕ Fx ⊕ Fy, which implies [L, L] ⊆ span{[x, y],I} and dim[L, L] 6 2, contradiction. If dimI = 2, there exists x ∈ L such that L = I ⊕ Fx, which implies [L, L] ⊆ I and dim[L, L] 6 dimI, contradiction. So L is a 3-dimensional simple Lie algebra. Next suppose that L is infinite dimensional. By Lemma 4.1, there is a 4-dimensional subalgebra L′ ⊂ L. We will prove that this is impossible. Clearly, any two commutative elements of L′ ⊆ L are also linearly dependent. For any linearly independent elements x, y, u, v ∈ L′, since {x, y, u} spans a 3-dimensional Lie algebra, and the same is true for {x, y, v}, we may assume that [x, y] = a1x + b1y + c1u, [x, y] = a2x + b2y + c2v, for some ai, bi, ci ∈ F. We deduce that c1 = c2 = 0. Thus, any two elements in L span a subalgebra. Since {x, y} can span a 2-dimensional Lie algebra, and {x, u} and {x, v} can also, we may assume that [x, y] = ax + by, [x, u] = a′x + b′u, [x, v] = a′′x + b′′v. Considering the subalgebras spanned by {x, y + u} and {x, u + v}, we deduce that b = b′ = b′′ 6= 0. By rechoosing x, y and u we may assume that [x, y] = y, [x, u] = u. Since x, y, u form a basis of the Lie subalgebra L′′ =span{x, y, u}, we see that the only eigenvalues of ad(x) on L′′ are 0 and 1. From [x, [y, u]] = [[x, y], u] + [y, [x, u]] = 2[y, u], we obtain that [y, u] = 0, or [y, u] ∈ Fx if char(F) = 2, contradicting the assumptions on L. This completes the proof. (cid:3) Corollary 4.3. An infinite dimensional Lie algebra in which any two commuting elements are linearly dependent is not zpd. Let us point out two special cases of the corollary. Corollary 4.4. A free Lie algebra in at least two variables is not zpd. Corollary 4.5. The first Witt algebra W1 is not zpd. Moreover, the same is true for every infinite dimensional subalgebra of W1. The fact that W1 is not zpd was already noticed in [11]. 5. Heisenberg Lie algebras Recall that a Heisenberg algebra h is a Lie algebra with basis {c, xi, x−i i ∈ I} where I is a nonempty finite or infinite set, such that [xi, x−i] = c,∀i ∈ I, and all other brackets of basis elements are 0. See [15]. In this section we will prove the following ZERO PRODUCT DETERMINED LIE ALGEBRAS 11 Proposition 5.1. All Heisenberg algebras over any field F are zpd. Proof. Let us also write L for the above defined Heisenberg algebra h. It is clear that spanK′ M′ L, so it is enough to show that M′ By u ∼= v we mean u − v ∈ span K′ L ⊆ span K′ L. L. We see that L ⊆ c ∧ h ∼= 0, x±i ∧ x±j ∼= 0,∀i 6= j ∈ I, xi ∧ x−i ∼= xj ∧ x−j,∀i, j ∈ I, since [xi + xj, x−i − x−j] = 0. Take an arbitrary Y = Pm j=1 yj ∧ y′ L. We may assume that j ∈ M′ yj = Xi∈P j = Xi∈P y′ ajixi + aj,−ix−i, a′ jixi + a′ j,−ix−i, where P is a finite subset of I, and aj,±i, a′ j,±i ∈ F. Then j,−i − a′ jiaj,−i) = 0. m Xj=1Xi∈P (ajia′ We see that Y = ∼= m Xj=1Xi∈P Xj=1Xi∈P m (ajia′ j,−i − a′ jiaj,−i)xi ∧ x−i (ajia′ j,−i − a′ jiaj,−i)x1 ∧ x−1 = 0. Thus h is zpd in this case. This completes the proof. (cid:3) Let sl2 be the 3-dimensional simple Lie algebra over C, i.e., the Lie algebra consisting of all 6. Galilei Lie algebras traceless 2 × 2 matrices. Let E = (cid:18) 0 1 0 0 (cid:19) , H = (cid:18) 1 0 0 −1 (cid:19) , F = (cid:18) 0 0 1 0 (cid:19) be the standard basis of sl2(C). Denote by ad the adjoint representation of sl2, i.e., (adA)(B) = [A, B] for all A, B ∈ sl2. An element A ∈ sl2 is called ad-nilpotent if adA is nilpotent; similarly, A is called ad-semisimple if sl2 can be decomposed as the sum of engeispaces with respect to the action of adA. For any nonzero A ∈ sl2(C), it is clear that A has eigenvalues ±p− det(A). Thus A is ad-nilpotent if det(A) = 0 and ad-semisimple if det(A) 6= 0. In particular, if A ∈ sl2(C) is ad-semisimple, then there exists P ∈GL2 such that P −1AP = det(A)H; if A ∈ sl2(C) is ad-nilpotent and nonzero, then there exists P ∈GL2 such that P −1AP = E. Lemma 6.1. Let V be a simple sl2-module. If V is zad, then dim V < ∞. 12 MATEJ BRESAR, XIANGQIAN GUO, GENQIANG LIU, RENCAI L U, KAIMING ZHAO Proof. It is straightforward to check that 0 6= H ⊗ F v− F ⊗ (H − 2)v ∈ MV . By the definition of zad modules, we see that spanKV = MV 6= 0. Claim 1. There exists a nonzero A ∈ sl2 such that A acts semisimply on V . is ad-semisimple (which is similar to a multiple of H) and acts semisimply on V . Choose any nonzero A ⊗ v ∈ span KV , A ∈ sl2, v ∈ V . Then Av = 0. If det(A) 6= 0, then A If det(A) = 0, then A is ad-nilpotent and thus similar to E. Without lost of generality, we may assume that 0 6= E ⊗ v ∈ spanKV , or equivalently, Ev = 0. Since the Casimir element, (H + 1)2 + 4F E ∈ U (sl2), of sl2 acts as a scalar on V , we see that (H + 1)2v = (cid:0)(H + 1)2 + 4F E(cid:1)v ∈ Cv. Thus, H has an eigenvector in V and hence acts semisimply on V , since V is a simple module. By Claim 1, we may assume that H acts semisimply on V by replacing E, F, H with P EP −1, P F P −1, P HP −1 for some invertible matrix P if necessary, that is, V is a weight module with respect to the Cartan subalgebra CH of sl2. Claim 2. V is finite dimensional. Suppose on the contrary that V is infinite dimensional. By the representation theory of sl2, either E or F acts injectively on V . Without lose of generality, we assume that E acts injectively on V . Take any nonzero A ⊗ v ∈ KV . Then Av = 0. Writing v as a sum of weight vectors and comparing the highest weights that occur in the expressions of v and Av respectively, we can easily see A ∈ CH ⊕ CF . In particular, KV ⊆ (CH + CF ) ⊗ V . On the other hand, we have H ⊗ Ev− E ⊗ (H + 2)v ∈ MV = KV for all nonzero v ∈ V , contradiction. So we must have dim V < ∞. (cid:3) Now we consider finite dimensional simple modules over sl2. Let m ∈ Z+ and V (m) be the simple sl2-module of dimension m + 1. Choose a standard basis {v0,··· , vm} of V (m) such that the actions of E, F, H can be expressed as F vi = vi+1, Evi = i(m + 1 − i)vi−1, Hvi = (m − 2i)vi, ∀ i = 0, 1,··· , m, where we have made the convention that vi = 0 whenever i 6= 0, 1,··· , m. Lemma 6.2. Let V be a simple sl2-module of even dimension. Then V is zad if and only if dim V = 2. Proof. Assume that V = V (m) is zad where m ∈ N is odd. As in the proof of the previous lemma, we know that span KV = MV 6= 0. Choose any nonzero A ∈ sl2 and v ∈ V such that A ⊗ v ∈ KV , that is, Av = 0. Noticing the fact that any ad-semisimple element acts bijectively on V , we have that A is ad-nilpotent and hence det(A) = 0. If A ∈ CE, it is clear that A ⊗ v ∈ C(E ⊗ v0). Now suppose that A 6∈ CE. By replacing A with its nonzero multiple we may assume that A = F + λH − λ2E for some l ∈ C. Assume v = Pm i=0 civi for some ci ∈ C. Then from Av = 0, ZERO PRODUCT DETERMINED LIE ALGEBRAS 13 we compute out that m 0 =(F + lH − l2E) m−1 = = Xi=0 Xi=1 m civi+1 + l ci−1vi + l m civi Xi=0 Xi=0 (m − 2i)civi − l2 Xi=0 (m − 2i)civi − l2 m m Xi=1 Xi=0 m−1 i(m + 1 − i)civi−1 (i + 1)(m − i)ci+1vi, l2mc1 = lmc0, cm−1 = lmcm ci−1 = l2(i + 1)(m − i)ci+1 − l(m − 2i)ci, ∀ i = 1,··· , m − 1. (6.1) which implies and (6.2) As a result, we deduce for all i = 0, 1,··· , m. Combining the previous result, we can obtain that ci = m! i! lm−icm span KV =spannE ⊗ v0, m Xi=0 m! i! lm−i(F + lH − l2E) ⊗ vi : l ∈ Co =spanni(i − 1)F ⊗ vi−2 + iH ⊗ vi−1 − E ⊗ vi : i = 0, 1,··· , m + 2o. In particular, we have dim spanKV = m + 3. On the other hand, we know that dimMV = (dim sl2 ⊗ V )2 − dim(sl2 · V ) = 2(m + 1). If m > 2, then dimMV > dim spanKV and V is not zad. So m = 1, i.e., dim V = 2. If V is of dimension 2, i.e., m = 1, from the above arguments we know that dim spanKV = 4 and dimMV = (dim sl2 ⊗ V )2 − dim(sl2 · V ) = 4. Thus dimMV = dim span KV and hence V is indeed zad. Lemma 6.3. If V is a simple sl2-module of odd dimension, then V is zad. Proof. Suppose V = V (m) where m = 2k > 0 for some k ∈ N. By a similar argument as in the previous lemma, solving Av = 0 for v ∈ V with A = F + lH − l2E for any l ∈ C, we deduce that (cid:3) B1 = (cid:8)i(i − 1)F ⊗ vi−2 + iH ⊗ vi−1 − E ⊗ vi, i = 0, 1,··· , m + 2(cid:9) ⊂ span KV . Recall that we have assumed that vi = 0 whenever i /∈ {0, 1,··· , m}. Taking A = H + 2λF for l ∈ C, and solving Av = 0 for v ∈ V we deduce that k Xi=0 li i! (H + 2lF ) ⊗ vk+i ∈ KV ,∀l ∈ C. 14 MATEJ BRESAR, XIANGQIAN GUO, GENQIANG LIU, RENCAI L U, KAIMING ZHAO Thus B2 = (cid:8)H ⊗ vk+i + 2iF ⊗ vk+i−1 i = 0, 1 . . . , k − 1(cid:9) ⊂ spanKV . Note that we did not include the elements for i = k and i = k + 1, since they have already occurred in B1. Similarly for A = H − 2λE, by solving Av = 0 for v ∈ V , we deduce that Xi=0 Thus k (k + i)! i!(k − i)! li(H − 2λE) ⊗ vk−i ∈ KV . B3 = (cid:8)(k − i + 1)(k + i)H ⊗ vk−i − 2iE ⊗ vk−i+1i = 1, . . . k − 1(cid:9) ⊂ span KV . Here we do not include the elements for i = 0, i = k and i = k + 1, since they have already occurred in B1 or B2. Now it is straightforward to check that B1 ∪ B2 ∪ B3 is a linearly independent set and has cardinality 2m + 2. We know that 2m + 2 ≤ dim spanKV ≤ dimMV . Since sl2 · V = V and dim sl2 ⊗ V = 3m + 3, we see that dimMV = 2m + 2. Thus dimMV = dim span KV and hence V is zad. (cid:3) Combining Lemma 2.12 and the above results in this section, we deduce the following Theorem 6.4. Suppose V is a simple module over sl2. Then the Galilei algebra sl2 ⋉ V is zpd if and only if dim V = 2 or dim V is odd. Let b = CE ⊕ CH be the Borel subalgebra of sl2. Next we consider a subalgebra of the Galilei algebra sl2 ⋉ V (m). Let b(m) = b ⋉ V (m) ⊂ sl2 ⋉ V (m). Corollary 6.5. Let m ∈ N. (a) The b-module V (m) is zad if and only if m = 2. (b) The Lie algebra b ⋉ V (m) is zpd if and only if m = 2. Proof. (a) For any m ∈ N, we see that the linear map b ⊗ V (m) → V (m), x ⊗ v 7→ xv,∀x ∈ b, v ∈ V (m), is onto. Then dimMV (m) = m + 1. MV (m), i.e., V (m) is not zad in this case. we see that First we assume that m is odd. It is easy to see that KV (m) = CE ⊗ v0. Thus KV (m) 6= Now we assume that m = 2k is even. From the computations in the proof of Lemma 6.3 B3 = (cid:8)(k − i + 1)(k + i)H ⊗ vk−i − 2iE ⊗ vk−i+1i = 0, 1, . . . k + 1(cid:9) is a basis of KV (m), i.e., dimKV (m) = k + 2. Then V (m) is zad if and only if MV (m) = KV (m), if and only if m + 1 = k + 2, if and only if m = 2. (b) follows from (a) and Lemma 2.12 since only linearly dependent elements in b commute. (cid:3) From this corollary we obtain a 4-dimensional non-zpd Lie algebra b(2) = b ⋉ V (2) where V (2) is the 2 dimensional module with basis {u, v} and action Hu = u, Hv = −v, Ev = u, Eu = 0. ZERO PRODUCT DETERMINED LIE ALGEBRAS 15 This 4-dimensional Lie algebra b(2) is the centerless 1-spacial aging algebra age(1). See [12]. For representations of Galilei algebras, see [16]. 7. Quantum torus Lie algebras Let q ∈ C, not a root of unity. Let Cq[t±1 t±1 1 , t±1 2 subject to the relation 1 , t±1 2 ] be the associative algebra generated by t1t−1 1 = t−1 1 t1 = t2t−1 2 = t−1 2 t2 = 1, t2t1 = qt1t2. See [3, 10]. Let Cq[t1, t2] be the subalgebra of Cq[t±1 (m1, m2) ∈ Z2, denote tm = tm1 1 , t±1 2 . Then it is clear that 1 tm2 2 ] generated by t1, t2. For any m = tmtn = qm2n1 tm+n, tmtn = qm2n1−n2m1 tntm,∀m, n ∈ Z2. Let Lq be the associated Lie algebra of Cq[t±1 1 , t±1 2 ]. The Lie bracket of Lq is given by [tm, tn] = (qm2n1 − qn2m1)tm+n. q be the associated Lie algebra of Cq[t1, t2]. Let L+ To show that Lq and L+ q are zpd Lie algebras, we need the following lemma: + such that n1, n2 are coprime and n1n2 6= 0, there + with (m1, m2) 6= (n1, n2), m1 ≤ n1, m2 ≤ n2 such that {m, n} Lemma 7.1. For any n = (n1, n2) ∈ Z2 exists m = (m1, m2) ∈ Z2 forms a Z-basis of Z2. Proof. Since n1, n2 ∈ N are coprime, there exists u, v ∈ Z such that un1 − vn2 = 1. Take k ∈ Z such that 0 6 v − kn1 < n1 and denote m1 = v − kn1, m2 = u − kn2. We have m2n1 − m1n2 = 1. Since 0 6 m1 < n1, we also have 0 < m2 6 n2. Theorem 7.2. If q ∈ C is not a root of unity, then the Lie algebra Lq is zpd. Proof. Let f : Lq × Lq → C be a bilinear map satisfying that f (x, y) = 0 whenever [x, y] = 0. Define a linear map Φ : [Lq,Lq] → C as follows: (cid:3) Φ(tm1,m2) = f (t(m1−1,m2), t(1,0))/(qm2 − 1), ∀ m2 6= 0; Φ(tm1,0) = f (t(m1,−1), t(0,1))/(1 − qm1), ∀ m1 6= 0. It is easy to check that f (t(1,0), tm) = Φ([t(1,0), tm]), ∀ m ∈ Z2, f (t(0,1), t(m,−1)) = Φ([t(0,1), t(m,−1)]), ∀ m ∈ Z. Define ϕ(x, y) = f (x, y) − Φ([x, y]) for all x, y ∈ Lq and denote T = (cid:8)m ∈ Z2 ϕ(tm, tn) = 0, ∀ n ∈ Z2(cid:9). We have (0, 0), (1, 0) ∈ T . To show that Lq is zpd, we need only to show that T = Z2. Claim 1. If m ∈ T , then jm ∈ T for all j ∈ Z+ and (7.1) ϕ(tn, tjm+kn) = 0, ∀ n ∈ Z2, k, j ∈ Z+. 16 MATEJ BRESAR, XIANGQIAN GUO, GENQIANG LIU, RENCAI L U, KAIMING ZHAO For any l ∈ C, k ∈ N and n ∈ Z2, we always have [ltm + tn, (ltm + tn)k] = 0, where k (ltm + tn)k = Xi=0 liai(q)tim+(k−i)n and each ai(q) is a polynomial function in q. By the assumption on f , we have f (ltm + tn, k Xi=0 In particular, liai(q)tim+(k−i)n) = 0 = Φ([ltm + tn, k Xi=0 liai(q)tim+(k−i)n]), ∀ l ∈ C. ϕ(ltm + tn, k Xi=0 liai(q)tim+(k−i)n) = 0, ∀l ∈ C. k Xi=0 liai(q)ϕ(tn, tim+(k−i)n) = 0, ∀l ∈ C, Since tm ∈ T , we see which implies ai(q)ϕ(tn, tim+(k−i)n) = 0, ∀ n ∈ Z2, i = 0, 1,··· , k. Taking i = k, we deduce ϕ(tn, tkm) = 0 forcing km ∈ T since ak(q) = q 6= 0. Taking i = 1, we deduce ϕ(tn, tm+(k−1)n) = 0 for all k ∈ N since a1(q) is a nonzero multiple of (1 + qm,n + ··· + qk−1 m,n) and hence nonzero, where qm,n is defined by tmtn = qm,ntntm. Replacing m with jm for j ∈ Z+ in the above argument, we get the result of this claim. [(ltm + µt−m + tn, (ltm + µt−m + tn)k] = 0. Suppose Now take arbitrary m, n ∈ Z2. For any l, µ ∈ C and k ∈ N, we consider the identity k(k−1)m1 m2 2 (ltm + µt−m + tn)k = k Xi,j=0 liµjaij(q)tim−jm+(k−i−j)n, where aij are some polynomials in q. By a similar argument as before, we can deduce (7.2) ϕ(cid:16)ltm + µt−m + tn, k Xi,j=0 liµjaij(q)t(i−j)m+(k−i−j)n(cid:17) = 0, ∀ l, µ ∈ C. Claim 2. If m ∈ T , then jm ∈ T for all j ∈ Z and ϕ(tn, tjm+kn) = 0, ∀ n ∈ Z2, j ∈ Z, k ∈ Z+. Taking m ∈ T in (7.2), we see k (7.3) Xi,j=0 liµjaij(q)ϕ(µt−m + tn, t(i−j)m+(k−i−j)n) = 0, ∀ l, µ ∈ C. ZERO PRODUCT DETERMINED LIE ALGEBRAS 17 Regarding the left hand side of (7.3) as a polynomial in l, µ and considering the coefficient of lk−1µ, we see that ak−1,1(q)ϕ(tn, t(k−2)m) + ak−1,0(q)ϕ(t−m, t(k−1)m+n) = 0. Taking k = 2 and noticing a1,0(q) = qm2n1 + qn2m1 6= 0, we see that ϕ(t−m, tm+n) = 0 for all n ∈ Z2, that is, −m ∈ T . Now by Claim 1, we get jm ∈ T and ϕ(n, jm + kn) = 0 for all j ∈ Z, k ∈ Z+. Claim 3. If m ∈ T , then ϕ(tn, tjm+kn) = 0, ∀ n ∈ Z2, j ∈ Z, k ∈ Z, k 6= −1. Exchanging m and n in (7.2), we get ϕ(cid:16)ltn + µt−n + tm, k Xi,j=0 liµjaij(q)t(i−j)n+(k−i−j)m(cid:17) = 0, ∀ l, µ ∈ C. Note that aij(q) has been changed since we have exchanged m and n. Since m ∈ T , we have (7.4) k Xi,j=0 liµjaij(q)ϕ(cid:16)ltn + µt−n, t(i−j)n+(k−i−j)m(cid:17) = 0, ∀ l, µ ∈ C. Considering the coefficient of lµk−1, we get a0,k−1(q)ϕ(tn, t(1−k)n+m) + a1,k−2(q)ϕ(t−n, t(3−k)n+m) = 0, ∀ k ∈ N. If k > 3, we already have ϕ(t−n, t(3−k)n+m) = 0 by Claim 2 and hence a0,k−1(q)ϕ(tn, t(1−k)n+m) = 0. m,n + ··· + q1−k It is straightforward to calculate that a0,k−1(q) is a nonzero multiple of 1 + q−1 and hence nonzero. We see that ϕ(tn, t(1−k)n+m) = 0 for k > 3. Replacing m with jm, j ∈ Z in the above arguments, we see ϕ(tn, t−kn+jm) = 0 for all k, j ∈ Z with k > 2. The claim follows from Claim 2. Claim 4. T = Z2. m,n Recalling that (1, 0) ∈ T , we have ϕ(t(0,1), t(j,k)) = 0 for j, k ∈ Z, k 6= −1 by Claim 3. On the other hand, by the definition of Φ, we have f (t(0,1), t(j,−1)) = Φ([t(0,1), t(j,−1)]) for all j ∈ Z and hence ϕ(t(0,1), t(j,−1)) = 0. As a result, (0, 1) ∈ T . By Claim 2, we also have (0, k), (k, 0) ∈ T for all k ∈ Z. Now take any n = (n1, n2) ∈ Z2 such that n1, n2 are coprime. We will prove n ∈ T by induction on n1n2. For n1n2 = 0, the result is true by the previous discussion. Now suppose n1n2 > 1. Without loss of generality we may assume that n ∈ Z2 +. By Lemma 7.1, there exists m = (m1, m2) ∈ Z2 + with (m1, m2) 6= (n1, n2), m1 ≤ n1, m2 ≤ n2, such that {m, n} forms a Z-basis of Z2. Denote m′ = (n1 − m1, n2 − m2) ∈ Z2 +. Then {m′, n} forms a basis of Z2 and (n1 − m1)(n2 − m2) < n1n2. By induction hypothesis, we have m, m′ ∈ T . Then Claim 3 indicates ϕ(tn, tjm+kn) = 0, ϕ(tn, tjm ′+kn) = 0, ∀ j, k ∈ Z, k 6= −1. 18 MATEJ BRESAR, XIANGQIAN GUO, GENQIANG LIU, RENCAI L U, KAIMING ZHAO For any n′ ∈ Z2, we can write n′ = j1m + k1n = j2m′ + k2n for some j1, k1, j2, k2 ∈ Z. ) = 0 and n ∈ T . By Noticing that k1 and k2 can not both be equal to −1, we have ϕ(tn, tn Claim 2, we see kn ∈ T for all k ∈ Z. The claim and hence the theorem follows. Using similar methods as in the previous theorem, we can also show that L+ q is a zpd Lie (cid:3) ′ algebra. Theorem 7.3. If q ∈ C is not a root of unity, then the Lie algebra L+ Proof. Note that [L+ satisfying that f (x, y) = 0 whenever [x, y] = 0. Define a linear map Φ : [L+ follows: 1tj 2 : i, j ∈ N}. Let f : L+ q ] =span{ti q × L+ q ,L+ q is zpd. Φ(t(m1,m2)) = f (t(m1−1,m2), t(1,0))/(qm2 − 1), ∀ m1, m2 ∈ N. For convenience, we denote q → C be a bilinear map q ] → C as q ,L+ for all x, y ∈ L+ q . Denote ϕ(x, y) = f (x, y) − Φ([x, y]) q is zpd, we need only to show ϕ(m, n) = 0 for all m, n ∈ Z2 +, or equivalently, + ϕ(tm, tn) = 0, ∀ n ∈ Z2 +(cid:9). T = (cid:8)m ∈ Z2 +. Note that (0, 0), (1, 0) ∈ T . To show that L+ T = Z2 Claim 1. If m ∈ T , then km ∈ T for all k ∈ Z+ and ϕ(tn, tjm+kn) = 0, ∀ n ∈ Z2 (7.5) +, j, k ∈ Z+. For any l ∈ C, k ∈ N and n ∈ Z2 +, we always have [ltm + tn, (ltm + tn)k] = 0, where (ltm + tn)k = k Xi=0 liai(q)tim+(k−i)n and each ai(q) is a function on q. By the assumption on f , we have f (ltm + tn, k Xi=0 In particular, liai(q)tim+(k−i)n) = 0 = Φ([ltm + tn, k Xi=0 liai(q)tim+(k−i)n]), ∀ l ∈ C. ϕ(ltm + tn, k Xi=0 liai(q)tim+(k−i)n) = 0 for all l ∈ C. Since m ∈ T , we see Xi=0 k liai(q)ϕ(tn, tim+(k−i)n) = 0 for all l ∈ C, which implies ai(q)ϕ(tn, tim+(k−i)n) = 0, ∀ i = 0, 1,··· , k. ZERO PRODUCT DETERMINED LIE ALGEBRAS 19 Taking i = k, we deduce ϕ(tn, tkm) = 0 forcing km ∈ T since ak(q) 6= 0. Taking i = 1, we deduce ϕ(tn, tm+(k−1)n) = 0 for all k ∈ N since a1(q) 6= 0. Replacing m with jm for j ∈ Z+ in the above argument, we get the result of this claim. Claim 2. T = Z2 +. By Claim 1, we see (0, 1) ∈ T by the fact that (1, 0) ∈ T . Hence (j, 0), (0, k) ∈ T for all j, k ∈ Z+. Now for any n ∈ Z2 + and l, µ ∈ C, the identity [tn + lt(j,0) + µt(0,k), (tn + lt(j,0) + µt(0,k))2] = 0, ∀ j, k ∈ Z+ implies ϕ(cid:0)tn, l(tnt(j,0) + t(j,0)tn) + µ(tnt(0,k) + t(0,k)tn) + lµ(t(j,0)t(0,k) + t(0,k)t(j,0))(cid:1) = 0, forcing ϕ(tn, t(j,k)) = 0 for all j, k ∈ Z+. The result follows. 8. Affine Lie algebras (cid:3) In this section we assume that F = C. Let g be a finite-dimensional simple Lie algebra with Cartan decomposition g = h ⊕ ⊕α∈∆gα, where ∆ is root system of g with repect to a Cartan subalgebra h and We write (x, y) for the Killing form on g. gα = {x ∈ g[h, x] = α(h)x,∀h ∈ h}. Let ∆∨ = {hαα ∈ ∆} be the coroot system of ∆, and {eαα ∈ ∆} be the Chevalley basis of g (See Chapter 7 in [13]). For any α ∈ ∆ we know that Since the Killing form is invariant, we have h−α = −hα, α(hα) = 2, [eα, e−α] = hα, [hα, e±α] = ±2e±α. (hα, hα) = ([eα, e−α], hα) = (eα, [e−α, hα]) = 2(eα, e−α),∀α ∈ ∆. The (untwisted) affine Lie algebra associated with g is defined as where c is the canonical central element and the Lie algebra structure is given by (see [14]) g = g ⊗ C[t, t−1] ⊕ Cc ⊕ Cd, [x ⊗ tn, y ⊗ tm] = [x, y] ⊗ tn+m + n(x, y)δn+m,0c, [d, x ⊗ tn] = nx ⊗ tn for x, y ∈ g and m, n ∈ Z. We will write x(n) for x ⊗ tn, and g = [g, g] = g ⊗ C[t, t−1] + Cz, h = h ⊗ C[t, t−1] + Cc. We first prove the following crucial formulae. Lemma 8.1. For any α ∈ ∆, h ∈ h and i, j ∈ Z we have (a) h(i) ⊗ eα(j) ≡ h(0) ⊗ eα(i + j) (cid:0)mod spanKg(cid:1); (b) eα(i) ⊗ e−α(−i) ≡ eα(0) ⊗ e−α(0) + 1 (c) eα(i) ⊗ eβ(j) ≡ eα(0) ⊗ eβ(i + j) (cid:0)mod spanKg(cid:1), if (α + β, i + j) 6= (0, 0); (d) d ⊗ eα(i) ≡ i 2 hα(i) ⊗ hα(−i) (cid:0)mod spanKg(cid:1); 2 hα(0) ⊗ eα(i) (cid:0)mod span Kg(cid:1); 20 MATEJ BRESAR, XIANGQIAN GUO, GENQIANG LIU, RENCAI L U, KAIMING ZHAO (e) d ⊗ hα(i) ≡ ieα(0) ⊗ e−α(i) (cid:0)mod span Kg(cid:1). Proof. (a) From we see that Hence we obtain (a). (b) From [e−α(0) + eα(i) + [h(i) + eα(i + j), h(0) + eα(j)] = 0 (h(i) + eα(i + j)) ⊗ (h(0) + eα(j)) ∈ Kg. √2 2 (hα(0) − hα(i)), e−α(−i) + eα(0) + √2 2 (hα(−i) − hα(0))] = 0 we see that (e−α(0) + eα(i) + √2 2 (hα(0) − hα(i))) ⊗ (e−α(−i) + eα(0) + √2 2 (hα(−i) − hα(0))) ∈ Kg. Using (a) we get (b). Further, (c) follows from and (d) follows from (eα(i) + eβ(i + j)) ⊗ (eβ(j) + eα(0)) ∈ Kg, (d − i 2 hα(0)) ⊗ eα(i) ∈ Kg. Finally, applying (a),(d) to (d + i 2 hα(0) − ieα(0)) ⊗ (hα(i) − eα(i) + e−α(i)) ∈ Kg we obtain (e). (cid:3) Theorem 8.2. The untwisted affine Lie algebras g and g are zpd. Proof. We will first prove that that g is zpd. Note that c ⊗ g ∈ Kg. Let X ∈ Mg. Using Lemma 8.1 (a)-(c), we see that We may assume that X ≡ X1 + X ′ i ∈ h ⊗ h. From 1 = Ps X ′ i ⊗ v′ i=1 u′ X ∈ (g ⊗ 1) ⊗ g + h ⊗ h + span Kg. 1 mod span(Kg) where X1 = Pr i=1 ui ⊗ vi ∈ (g⊗ 1)⊗ g and r s [ui, vi] + Xi=1 [u′ i, v′ i] = 0, Xi=1 and [g ⊗ 1), g] ∩ [h, h] = 0 we deduce that r s Xi=1 i.e., X1 ∈ ((g ⊗ 1) ⊗ g) ∩ Mg and X ′ [ui, vi] = 0, Xi=1 1 ∈ Mh. Hence [u′ i, v′ i] = 0, Mg = ((g ⊗ 1) ⊗ g + h ⊗ h + span Kg) ∩ Mg = (g ⊗ 1) ⊗ g) ∩ Mg + Mh + span Kg. ZERO PRODUCT DETERMINED LIE ALGEBRAS 21 However, h is zpd and g is a zad g-module. Thus ((g ⊗ 1) ⊗ g) ∩ Mg ⊆ span Kg and Mh = span Kh ⊆ span Kg. Therefore we have Mg = span Kg, as desired. fact c ⊗ g ∈ Kg we see that Next we prove that g is zpd. Let Y ∈ Mg. In view of Lemma 8.1 (d) and (e), using the Y ∈ g ⊗ g + spanKg. Noting that (g ⊗ g) ∩ Mg = Mg, and using the established result that Mg = spanKg we see that Y ∈ (g ⊗ g) ∩ Mg + spanKg = Mg + spanKg ⊂ spanKg + spanKg = spanKg. Thus Mg = spanKg, i.e., g is zpd. (cid:3) 9. Commutativity preserving maps Let ϕ be a linear map from one algebra into another. We say that ϕ preserves commuta- tivity if ϕ(x) and ϕ(y) commute whenever x and y commute. The problem of describing such maps has a long history in the context of associative algebras. It originated in linear algebra and operator theory, and later, in connection with the development of the theory of functional identities, moved to noncommutative ring theory. We refer the reader to [6, pp. 218-219] for historic details and references. One usually assumes that ϕ is bijective or at least surjective. In this case a natural possibility is that ϕ is a linear combination of a homomorphism or an antihomomorphism and a map having the range in the center of the target algebra. Without the surjectivity assumption the problem becomes much more involved as we have another natural possibility, i.e., maps with commutative range. The concept of a zpd Lie algebra actu- ally arose from the problem of describing not necessarily surjective commutativity preserving linear maps on finite dimensional central simple algebras [8]. The solution was based on first showing that Mn(F), viewed as a Lie algebra, is zpd. Although the notion of a commutativity preserving linear map is Lie-theoretic in nature, to the best of our knowledge it was studied by Lie algebra tools only in a few papers [18, 19, 21]. Our aim now is to show how it can be handled in zpd Lie algebras. More precisely, we will assume that the first Lie algebra is zpd and the target Lie algebra satisfies conditions suitable for using the theory of functional identities. Indeed functional identites are a standard tool in treating commutativity preservers, but results of such a type are nevertheless new. In light of examples of zpd Lie algebras that have been found in previous sections, they are now applicable to various concrete situations. Let us begin with a simple example indicating the delicacy of the problem. Example 9.1. Let L be a Lie algebra in which only linearly dependent elements commute. Then every linear map ϕ : L → L preserves commutativity. Indeed such Lie algebras usually are not zpd, but some of them, like sl2(F), are. This shows that one cannot expect that commutativity preserving maps between two zpd Lie algebras can be always nicely described. Some other conditions are needed, too. 22 MATEJ BRESAR, XIANGQIAN GUO, GENQIANG LIU, RENCAI L U, KAIMING ZHAO The condition that we will require is that the target Lie algebra is a 3-free subset of an associative unital algebra. To avoid making this section too lengthy, we only refer to the book [6] for the definition and basic properties of such sets, and give the proof of the next lemma without a detailed explanation of the results that are used. Lemma 9.2. Let V be a vector space over a field F with char(F) 6= 2, let A be a unital associative algebra over F, let ϕ : V → A be a linear map and B : V × V → A be a skew- symmetric bilinear map. Suppose that (9.1) [ϕ(x), B(y, z)] + [ϕ(z), B(x, y)] + [ϕ(y), B(z, x)] = 0, ∀ x, y, z ∈ V. If the range of ϕ is a 3-free subset of A, then there exist λ ∈ Z, the center of A, and a skew-symmetric bilinear map ν : V × V → Z such that B(x, y) = λ[ϕ(x), ϕ(y)] + ν(x, y), ∀ x, y ∈ V. Proof. Since the range of ϕ is 3-free it follows from [6, Theorem 4.13] that there exist elements λ, λ′ ∈ Z and maps µ, µ′ : V → Z, ν : V × V → Z such that B(x, y) = λϕ(x)ϕ(y) + λ′ϕ(y)ϕ(x) + µ(x)ϕ(y) + µ′(y)ϕ(x) + ν(x, y), ∀ x, y ∈ V. As B is bilinear, [6, Lemma 4.6, Remark 4.7] imply that µ, µ′ are linear maps and ν is bilinear. Further, since B(x, y) = −B(y, x) holds by assumption, it follows that (λ + λ′)ϕ(x)ϕ(y) + (λ + λ′)ϕ(y)ϕ(x) + (µ + µ′)(x)ϕ(y) + (µ + µ′)(y)ϕ(x) + ν(x, y) + ν(y, x) = 0. Now, [6, Lemma 4.4] implies that λ = −λ′, µ = −µ′, and ν is skew-symmetric. Thus, we have B(x, y) = λ[ϕ(x), ϕ(y)] + µ(x)ϕ(y) − µ(y)ϕ(x) + ν(x, y). Using this form in (9.1) we obtain 2µ(y)[ϕ(x), ϕ(z)] + 2µ(z)[ϕ(y), ϕ(x)] + 2µ(x)[ϕ(z), ϕ(y)] = 0. Since char(F) 6= 2 it follows from [6, Lemma 4.4] that µ = 0. (cid:3) We will now derive two theorems on commutativity preservers from this lemma. To un- derstand their meaning, it is important to mention that Lie ideals of associative algebras are their 3-free subsets under rather mild assumptions [6, Corollary 5.16]. The simplest, but very illustrative example is that sln(F) is a 3-free subset of Mn(F) provided that n ≥ 3. The n = 2 case is an exception and indeed the next theorem does not hold if L = L′ = sl2(F), cf. Example 9.1. Similarly, Lie algebras of skew-symmetric elements in associative algebras with involution are "usually" 3-free subsets [6, Corollary 5.18], and so are their Lie ideals [6, Corollary 5.19]. It seems reasonable to conjecture that there are many other types of Lie algebras that are 3-free subsets of some of their associative envelopes. However, only little is known about this at present. Theorem 9.3. Let L and L′ be a Lie algebras over a field F with char(F) 6= 2. Suppose that L is zpd, L′ is centerless, and L′ is Lie subalgebra of an associative unital algebra A whose center is equal to F1. If L′ is a 3-free subset of A, then every commutativity preserving bijective linear map ϕ : L → L′ is a scalar multiple of an isomorphism. ZERO PRODUCT DETERMINED LIE ALGEBRAS 23 Proof. The condition that ϕ preserves commutativity can be interpreted as that the map f (x, y) = [ϕ(x), ϕ(y)] satisfies the condition from the definition of the zpd property. Thus, since L is zpd it follows (from Lemma 2.1) that there exists a linear map Φ : [L, L] → L′ such that (9.2) [ϕ(x), ϕ(y)] = Φ([x, y]), ∀ x, y ∈ L. Applying the Jacobi identity this yields [ϕ(x), ϕ([y, z])] + [ϕ(z), ϕ([x, y])] + [ϕ(y), ϕ([z, x])] = 0, ∀ x, y, z ∈ L. This makes it possible for us to apply Lemma 9.2. Accordingly, there exist λ ∈ F and a skew-symmetric bilinear map ν : L × L → F1 such that (9.3) However, since the center of L′ is 0 by assumption it follows from (9.3) that ν = 0. Con- sequently, λϕ is a Lie algebra homomorphism. We only have to show that λ 6= 0. Assume, therefore, that λ = 0. Then ϕ([x, y]) = 0 for all x, y ∈ L, yielding that L is Abelian. Since ϕ preserves commutativity this implies that L′ is Abelian, too. However, this is impossible for L′ is a 3-free subset of A. (cid:3) ϕ([x, y]) = λ[ϕ(x), ϕ(y)] + ν(x, y), ∀ x, y ∈ L. The next, and the last theorem is similar. The difference is that we impose another condi- tions on L and less conditions on ϕ, A, and L′ -- in fact, L′ does not even appear since we do not need to assume that the range of ϕ is a Lie subalgebra of A. Theorem 9.4. Let L be a perfect zpd Lie algebra over a field F with char(F) 6= 2, and let A be an associative unital algebra over F whose center Z is a field. If ϕ : L → A is a commutativity preserving linear map whose range is a 3-free subset of A, then ϕ is of the form (9.4) where α ∈ Z, θ is a Lie algebra homomorphism from L into A and β is a linear map from L into Z. ϕ(x) = αθ(x) + β(x) Proof. Just as in the previous proof one derives that (9.2) holds for some linear map Φ : [L, L] → A and (9.3) holds for some λ ∈ Z and a skew-symmetric bilinear map ν : L× L → Z. We remark that the range of ϕ cannot be contained in Z for it is a 3-free subset of A. Since L is perfect, this implies that λ 6= 0. Define θ, β : L → A by θ(x) = λ2Φ(x) and Writing α for λ−1 we see that all that remains to show is that β maps into Z and that θ is a Lie algebra homomorphism. We have β(x) = ϕ(x) − λ−1θ(x). β([x, y]) = ϕ([x, y]) − λ−1θ([x, y]) = λ[ϕ(x), ϕ(y)] + ν(x, y) − λΦ([x, y]) = ν(x, y) ∈ Z. Since L is perfect this proves that β maps into in Z. Consequently, θ([x, y]) = λ2Φ([x, y]) = λ2[ϕ(x), ϕ(y)] = [λϕ(x), λϕ(y)] = [θ(x) + λβ(x), θ(y) + λβ(y)] = [θ(x), θ(y)], proving that θ is a homomorphism. (cid:3) 24 MATEJ BRESAR, XIANGQIAN GUO, GENQIANG LIU, RENCAI L U, KAIMING ZHAO It should be mentioned that we now know quite a few perfect zpd Lie algebras, for example, complex finite-dimensional semisimple Lie algebras, loop algebras, untwisted affine Lie alge- bras, quantum torus Lie algebras in Sect.7, and some Gallilei algebras. Theorem 9.4 therefore generalizes some results from [18, 19, 21]. Acknowledgement. The majority of this paper was conducted during the four authors' visit to Soochow University, China, in the August of 2016. M. Bresar is partially supported by ARRS grant P1 -- 0288; X.Guo is partially supported by NSF of China (Grant 11101380) and the Outstanding Young Talent Research Fund of Zhengzhou University (Grant 1421315071); G. Liu is partially supported by NSF of China (Grant 11301143) and the school fund of Henan University (2012YBZR031, yqpy 20140044); R.Lu is partially supported by NSF of China (Grant 11471233, 11371134); K. Zhao is partially supported by NSF of China (Grants 11271109, 11471233) and NSERC. References [1] J. Alaminos, M. Bresar, J. Extremera, A. R. Villena, Maps preserving zero products, Studia Math. 193 (2009), 131 -- 159. [2] J. Alaminos, M. Bresar, J. Extremera, A. R. Villena, Zero Lie product determined Banach algebras, preprint. [3] B. N. Allison, S. Azam, S. Berman, Y. Gao, A. Pianzola, Extended affine Lie algebras and their root systems, Mem. Amer. Math. Soc. 126 (1997), no. 603. [4] M. Bresar, Introduction to noncommutative algebra, Universitext, Springer, 2014. [5] M. Bresar, Finite dimensional zero product determined algebras are generated by idempotents, Expo. Math. 34 (2016), 130 -- 143. [6] M. Bresar, M. A. Chebotar, W. S. Martindale 3rd, Functional identities, Birkhauser Verlag, 2007. [7] M. Bresar, M. Grasic, J. Sanchez, Zero product determined matrix algebras, Linear Algebra Appl. 430 (2009), 1486 -- 1498. [8] M. Bresar, P. Semrl, On bilinear maps on matrices with applications to commutativity preservers, J. Algebra 301 (2006), 803 -- 837. [9] D. Brice, H. Huang, On zero product determined algebras, Linear Multilinear Algebra 63 (2015), 326 -- 342. [10] S. Eswara Rao, K. Zhao, Highest weight irreducible representations of rank 2 quantum tori, Math. Res. Lett. 11 (2004), 615 -- 628. [11] M. Grasic, Zero product determined classical Lie algebras, Linear Multilinear Algebra 58 (2010), 1007 -- 1022. [12] M. Henkel, S. Stoimenov, On non-local representations of the ageing algebra, Nuclear Physics B 847 (2011), 612 -- 627. [13] J. E. Humphreys, Introduction to Lie algebras and representation theory, Graduate Texts in Mathematics, no. 9, Springer-Verlag, New York-Berlin, 1978. [14] V. Kac, Infinite dimensional Lie algebras, 3rd edition, Cambridge Univ. Press, 1990. [15] V. Kac, A. Raina, Bombay lectures on highest weight representations of infinite dimensional Lie algebras, World Sci., Singapore, 1987. [16] R. Lu, V. Mazorchuk, K. Zhao, On simple modules over conformal Galilei algebras, J. Pure Appl. Algebra 218 (2014), 1885 -- 1899. [17] P. Moravec, Unramified Brauer groups of finite and infinite groups, Amer. J. Math. 134 (2012), 1679 -- 1704. [18] D. Wang, Invertible linear maps on Borel subalgebras preserving zero Lie products, J. Algebra 406 (2014), 321 -- 336. [19] D. Wang, Z. Chen, Invertible linear maps on simple Lie algebras preserving commutativity, Proc. Amer. Math. Soc. 139 (2011), 3881 -- 3893. [20] D. Wang, X. Yu, Z. Chen, A class of zero product determined Lie algebras, J. Algebra 331 (2011), 145 -- 151. ZERO PRODUCT DETERMINED LIE ALGEBRAS 25 [21] W. J. Wong, Maps on simple algebras preserving zero products, II: Lie algebras of linear type, Pacific J. Math. 92 (1981), 469 -- 488. M. B.: Faculty of Mathematics and Physics, University of Ljubljana, and Faculty of Natural Sciences and Mathematics, University of Maribor, Slovenia E-mail address: [email protected] X. G.: School of Mathematics and Statistics, Zhengzhou University, Zhengzhou, P. R. China E-mail address: [email protected] G. L.: School of Mathematics and Statistics, Henan University, Kaifeng, P. R. China E-mail address: [email protected] R. L.: Department of Mathematics, Soochow University, Suzhou, P. R. China E-mail address: [email protected] K. Z.: Department of Mathematics, Wilfrid Laurier University, Waterloo, Canada, and Col- lege of Mathematics and Information Science, Hebei Normal University, Shijiazhuang, P. R. China E-mail address: [email protected]
1503.00564	3	1503	2019-05-20T11:56:23	On Scalable Monoids	[ "math.RA" ]	This brief exposition presents some basic properties of scalable monoids and quantity spaces, introduced in arXiv:1408.5024	math.RA	math	ON SCALABLE MONOIDS DAN JONSSON Abstract. This brief exposition presents some basic properties of scalable monoids and quantity spaces, introduced in [Jon14a]. Let X be a non-empty set and R a commutative unital ring. An associative 1. Introduction R-algebra with X as carrier set combines three operations on X: (1) addition of elements of X, a binary operation + : X × X → X, written (x, y) 7→ x + y, such that (X, +) is an abelian group; multiplication of elements of X, a binary operation · : X × X → X, written (x, y) 7→ x · y or (x, y) 7→ xy, such that (X, ·) is a monoid; scalar multiplication of elements of X by elements of R, a monoid action R × X → X, written (α, x) 7→ αx, where the multiplicative monoid of R acts on X, so that 1x = x and α(βx) = (αβ)x. (2) (3) (b) (c) These structures are linked pairwise: (a) addition and multiplication are linked by the distributive laws x(y + z) = xy + xz and (x + y)z = xz + yz; one of these laws suffices when multi- plication in X is commutative; addition and scalar multiplication are linked by the distributive laws α(x + y) = αx + αy and (α + β)x = αx + βx; multiplication and scalar multiplication are linked by the laws α(xy) = (αx)y and α(xy) = x(αy); one of these laws suffices when multiplication in X is commutative. Related algebraic systems can be obtained from associative R-algebras by removing one of the three operations and hence the links between the removed operation and the two others. Two cases are very familiar. A unital ring has only addition and multiplication of elements, linked as described in (a). A module has only addition and scalar multiplication of elements, linked as described in (b). The question arises whether it makes sense to define an "algebra without an additive group", with only multiplication and scalar multiplication, linked as described in (c). Would such a notion be of mathematical interest and useful for applications? Some reasons to give an affirmative answer to that question are found in this article. 2. Basic notions It is tempting to call an "algebra without an additive group" an "alebra", but that may sound somewhat frivolous, and considering the nature of this algebraic struc- ture, "scalable monoid" [Jon14a], abbreviated "scaloid", might be a better name. This structure can be defined more or less broadly. In the definition below, multi- plication in R is assumed to be commutative, but multiplication in X is not. 1 Dan Jonsson 2 Definition 1. A scalable monoid, or scaloid, is an algebraic structure incorporating structures described by (2), (3) and (c) above. More explicitly, a scalable monoid X over a commutative unital ring R is a monoid X such that there is a function called scalar multiplication, such that for any α, β ∈ R and any x, y ∈ X we have σ : R × X → X, (α, x) 7→ αx, (1) 1x = x, (2) α(βx) = (αβ)x, (3) α(xy) = (αx)y = x(αy). For clarity, we may denote the unit element of the monoid X by 1X or 1 and the unit element of the ring R by 1R. (cid:3) An invertible element of a scaloid X is an element x ∈ X which has a (necessarily unique) inverse x−1 ∈ X such that xx−1 = x−1x = 1X . We can define positive powers of x, denoted xk, in the usual way; if x is invertible, negative powers of x . By convention, x0 = 1X . By associativity, can be defined by setting x−k =(cid:0)x−1(cid:1)k xkxℓ = x(k+ℓ). A product of invertible elements is clearly an invertible element. The following two simple facts about scaloids will be used repeatedly: Lemma 1. Let X be a scalable monoid over R. (1) (αx)(βy) = (αβ)(xy) for any x, y ∈ X and α, β ∈ R. (2) α(βx) = β(αx) and (αβ)x = (βα)x for any x ∈ X and α, β ∈ R. Proof. (1). (αx)(βy) = α(x(βy)) = α(β(xy)) = (αβ)(xy). (2). α(βx) = α(β(1x)) = α((β1)x) = (β1)(αx) = β(1(αx)) = β(α(1x)) = β(αx), so (αβ)x = (βα)x. (cid:3) Note that the assumption that R is commutative is not used in the proofs given in Section 3, only the fact that αβ and βα act in the same way on X, as stated in Lemma 1 (2). The assumption that αβ = βα in R was also not used in the proof of Lemma 1 (2). 3. Ideals and additive group operations By removing the additive group of an R-algebra, we remove the three group operations (x, y) 7→ x + y, (x) 7→ −x and () 7→ 0 (addition, taking the inverse and selecting the group identity). It turns out, however, that under natural assumptions the group structure can be partially recovered. Specifically, a scalable monoid over R can be partitioned into equivalence classes in each of which a group structure corresponding to the additive group structure in R can often be defined. 3.1. Commensurability and ideals. Definition 2. Let X be a scalable monoid over R, and let ∼ be a relation on X such that x ∼ y if and only if αx = βy for some α, β ∈ R. Then x and y are said to be commensurable, and the relation ∼ is called commensurability. (cid:3) As a trivial consequence of this definition, x ∼ λx for any x ∈ X and λ ∈ R, since λx = 1(λx). Scalable Monoids 3 Proposition 1. Commensurability is an equivalence relation. Proof. The commensurability relation ∼ is reflexive because 1x = 1x, it is sym- metric because if αx = βy then βy = αx, and it is transitive because if αx = βy and γy = δz then γ(αx) = γ(βy) and β(γy) = β(δz), so (γα)x = (βδ)z since γ(βy) = β(γy). (cid:3) Definition 3. Let X be a scalable monoid. An ideal in X is an equivalence class with respect to commensurability. We denote the ideal containing x ∈ X by [x]. (cid:3) Proposition 2. Let X be a scalable monoid. The commensurability relation ∼ is a congruence relation on X, so the corresponding set of equivalence classes can be made into an algebraic structure X/∼ with a binary operation ([x], [y]) 7→ [x][y] defined by for any x, y ∈ X. X/∼ is a monoid with [1] as unit element. [x][y] = [xy] Proof. If αx = α′x′ and βy = β ′y ′ then (αx)(βy) = (α′x′)(β ′y ′), so (αβ)(xy) = (α′β ′)(x′y ′). Thus, if x ∼ x′ and y ∼ y ′ then xy ∼ x′y ′, so the product in X/∼ given by [x][y] = [xy] is well-defined. Also, for any x ∈ X, 1x = x1 = x, so [1][x] = [x][1] = [x], so [1] is a unit element for X/∼. (cid:3) X/∼ is the image of X under the mapping h : x 7→ [x], and it is clear that h is a homomorphism of monoids. Note that a scaloid is akin to a tensor product of modules. Recall that the tensor product of u ∈ U and v ∈ V is a tensor u ⊗ v belonging to U ⊗ V ; this is a module in general distinct from U and V . Similarly, the product xy of x and y in a scaloid X does not in general belong to the same ideal as x or y; in general [x], [y] 6= [xy]. This suggests an application of scaloids: "quantities" such as 1 N, 2 m or 2 Nm can be regarded as elements of a scaloid, specifically quantities with different "sorts" or "dimensions", corresponding to different equivalence classes of such quantities, namely ideals. Pursuing the analogy with tensor products, certain additional assumptions al- low us to regard each ideal as a module, making it possible to add and subtract quantities as described below. 3.2. Recovering additive group operations. Definition 4. Let X be a scalable monoid over R, and consider an ideal D ⊂ X. A pivot for D is an element p ∈ D such that for every x ∈ D there is a unique λ ∈ R such that x = λp. (cid:3) Proposition 3. If p and p′ are pivots for D, so that x = αp = α′p′ and y = βp = β ′p′ for any x, y ∈ D, then (α + β)p = (α′ + β ′)p′. Proof. As p′ ∈ R, there is a unique λ ∈ R such that p′ = λp. Thus, (α′ + β ′)p′ = (α′ + β ′)(λp) = ((α′ + β ′)λ)p = (α′λ + β ′λ)p. Also, x = αp = α′p′ = α′(λp) implies α = α′λ, and y = βp = β ′p′ = β ′(λp) implies β = β ′λ. (cid:3) Hence, the sum of two elements of a scalable monoid can be defined as follows: Dan Jonsson 4 Definition 5. Let X be a scalable monoid over R, and let p be a pivot for an ideal D. If x = αp and y = βp, we set x + y = (α + β)p. (cid:3) Note that if x = αp and y = βp then βx = β(αp) = α(βp) = αy, so x + y can be defined only if x ∼ y -- one cannot add incommensurable quantities. As a trivial consequence of Definition 5, addition in an ideal is commutative. If p is a pivot for an ideal D and q ∈ D then 0q = 0(λp) = (0λ)p = 0p, so there is a unique element 0D ∈ D such that 0D = 0q for any q ∈ D, defined by 0D = 0p for any pivot p for D. 0D + x = x for any x ∈ D. It follows immediately from Definition 5 that x + 0D = A unital ring R has a unique additive inverse −1 of 1R, and we set −x = (−1)x for all x ∈ X. If D has a pivot, we also set x − y = x + (−y) for any x, y ∈ D, so that x − x = −x + x = 0D for any x ∈ D. 4. Free commutative scalable monoids A commutative scalable monoid X is one where xy = yx for all x, y ∈ X. In this section, only commutative scalable monoids will be considered. Definition 6. Let X be a commutative scalable monoid over R. A (finite) basis for X is a set B = {b1, . . . , bn} of invertible elements of X such that every x ∈ X has a unique (up to order of factors) expansion where µ ∈ R and k1, . . . , kn are integers. (cid:3) x = µ bki i , nYi=1 Any product of invertible quantities is invertible, so any product of basis elements is invertible. As X is commutative, we have µ nYi=1 i ! ν bki nYi=1 i ! = (µν) bℓi nYi=1 b(ki +ℓi) i . 4.1. Quantity spaces. Below, only scaloids over a field K will be considered. Definition 7. A (finite-rank) free commutative scalable monoid over a field K, or a quantity space over K, is a commutative scalable monoid over K which has a (finite) basis. (cid:3) With a view to applications, elements of a quantity space may be called quan- tities, and ideals in a quantity space may be called dimensions. As explained in [Jon14a], a basis for a quantity space can be interpreted as a system of fundamental units of measurement. Lemma 2. Let X be a quantity space over K with a basis {b1, . . . bn}, and consider i . The following conditions are equivalent: x = µQn i and y = νQn i=1 bki (1) x ∼ y. (2) ki = ℓi for i = 1, . . . , n. (3) νx = µy, or equivalently µB(y)x = µB(x)y. i=1 bℓi Scalable Monoids Proof. (1) ⇒ (2). α, β ∈ K. As the expansion of z is unique, ki = ℓi for i = 1, . . . , n. i=1 bki If x ∼ y then (αµ)Qn i = z = (βν)Qn The implications (2) ⇒ (3) and (3) ⇒ (1) are trivial. (cid:3) i=1 bℓi i 5 for some 4.2. The measure of a quantity. Definition 8. Let X be a quantity space over K, and let B = {b1, . . . , bn} be a basis for X. The uniquely determined scalar µ ∈ K in the expansion x = µ bki i nYi=1 is called the measure of x relative to B and will be denoted by µB(x). (cid:3) For example, µB(1) = 1 for any B, because 1 = 1 · 1 = 1Qn Relative to a fixed basis, measures of quantities can be used as proxies for the quantities themselves. For example, the measure of a product of quantities is equal to the product of the measures of these quantities. i for any B. i=1 b0 Proposition 4. Let B = {b1, . . . , bn} be a basis for a quantity space X over K. (1) For any x ∈ X and λ ∈ K, µB(λx) = λµB(x). (2) For any x, y ∈ X, µB(xy) = µB(x)µB(y). i , where µ, ν ∈ K. i , so µB(λx) = λµ = λµB(x). i=1 b(ki +ℓi) i , so µB(xy) = µν =µB(x)µB(y). Proposition 5. A quantity x ∈ X is invertible if and only if µB(x) 6= 0, and for (cid:3) i=1 bki i=1 bℓi i=1 bℓi i=1 bki i=1 bki i=1 bki i (cid:17) = (λµ)Qn i and y = νQn i (cid:17) = (µν)Qn Proof. We have x = µQn (1). λx = λ(cid:16)µQn i (cid:1)(cid:16)νQn (2). (cid:0)µQn any invertible x ∈ X we have µB(cid:0)x−1(cid:1) = 1/µB(x). Proof. If µB(x) 6= 0 and x = µB(x)Qn = µB(x) nYi=1 i=1 b−ki i µB(x) i=1 b−ki i µB(x) Qn i=1 bki then i bki = 1, i !Qn so x is invertible. i=1 b−ki i µB(x) x = xQn If, conversely, x has an inverse x−1 then µB(x)µB(cid:0)x−1(cid:1) = Proposition 6. Let X be a quantity space over K. For every x ∈ [1], µB(x) does not depend on B. µB(cid:0)xx−1(cid:1) = µB(1) = 1, so µB(x) 6= 0 and µB(cid:0)x−1(cid:1) = 1/µB(x). (cid:3) Proof. Let B = {b1, . . . , bn} and bB =nbb1, . . . ,bbmo be bases for X. The quantity 1 i and 1 = 1Qm has the expansions 1 = 1Qn i=1bb0 i(cid:17) = µ bB(x) · 1 = i = µ bB(x)(cid:16)1Qm x = µB(x)Qn i and x = µ bB(x)Qm i=1bb0 i=1bb0 i(cid:1) = µ bB(x)Qn µ bB(x)(cid:0)1Qn It is common to refer to a quantity x ∈ [1] as a "dimensionless quantity", although x is not really "dimensionless" -- it belongs to, or "has", the dimension [1]. The fact that the measure of any x ∈ [1] does not depend on a choice of basis -- that is, a choice of fundamental units of measurement -- is the foundation for dimensional analysis [Bri22, Bar96, Jon14a]. i . In view of Lemma 2, therefore, i , so µB(x) = µ bB(x). (cid:3) i=1 b0 i=1 b0 i=1 b0 i=1 b0 Dan Jonsson 6 4.3. Additive group operations in quantity spaces. Proposition 7. Let X be a quantity space over K. (1) If x, y ∈ X, x ∼ y and x is invertible then y = κx for some κ ∈ K. (2) If x ∈ X is invertible and λx = λ′x for some λ, λ′ ∈ K then λ = λ′. Proof. (1). µB(y)x = µB(x)y by Lemma 2, and µB(x) 6= 0 by Proposition 5, so y = x. µB(y) µB(x) i=1 bki i (2). Let x have the expansion x = µQn i = z = λ′µQn If λx = λ′x then λµQn i=1 bki of z is unique, and µ 6= 0 since x is invertible, so λ = λ′. (cid:3) relative to a basis {b1, . . . , bn} for X. i , so λµ = λ′µ since the expansion i=1 bki Thus, every invertible quantity is a pivot for the dimension to which it belongs. This has an important consequence: Corollary 1. Every dimension in a quantity space has a pivot. Proof. An equivalence class is not empty, so for every dimension D there is some quantity µQn i=1 bki i ∈ D, so there is an invertible element 1Qn In a quantity space, x + y is thus defined if and only if x ∼ y. Addition of i=1 bki i ∈ D. (cid:3) quantities in a dimension relates to addition of measures in the following way: Proposition 8. Let X be a quantity space over K. For any basis B for X and x, y ∈ X such that x ∼ y, we have µB(x) + µB(y) = µB(x + y). i=1 bki Proof. Let x = µB(x)Qn y relative to B = {b1, . . . .bn}. Qn nYi=1 x + y = µB(x) i and y = µB(y)Qn i=1 bki i i=1 bki i be the expansions of x and is invertible, and thus a pivot, so bki i + µB(y) bki i = (µB(x) + µB(y)) bki i , nYi=1 nYi=1 so we have obtained the unique expansion of x + y relative to B, and this expansion shows that µB(x + y) = µB(x) + µB(y). (cid:3) In words, the measure of a sum of quantities is equal to the sum of the measures of these quantities. Also, µB(λx) = λµB(x), so measures represent quantities in a given dimension in the same way that coordinates represent vectors. 4.4. Groups of dimensions; cardinality of bases. If X is commutative then X/∼ is also commutative, and in this section we prove that X/∼ is actually a free abelian group, not only a commutative monoid. Proposition 9. Let X be a quantity space over K. For every dimension [x] ∈ X/∼ there is a unique dimension [x]−1 ∈ X/∼ such that [x][x]−1 = [x]−1[x] = [1]. i=1 bki i be the unique expansion of x relative to the basis . (By Lemma 2, y de- pends on [x], but not on its representative x.) Then [x] = [y] and [y][z] = [z][y] = [zy] = [1], so [x][z] = [z][x] = [1]. Furthermore, if [x]−1[x] = 1 then Proof. Let x = µQn {b1, . . . , bn}, and set y = 1Qn [x]−1 = [x]−1([x][z]) =(cid:16)[x]−1[x](cid:17)[z] = [z]. (cid:3) i and z = 1Qn i=1 b−ki i=1 bki i Scalable Monoids 7 Recall that a (finite) basis for an abelian group G is a set {b1, . . . , bn} of elements i , and that a free (finitely generated) abelian group is an abelian group for which such a basis exists. of G such that every x ∈ G has a unique expansion x =Qn i=1 bki Proposition 10. Let X be a quantity space over K. (1) If B = {b1, . . . , bn} is a basis for X, then B ∗ = {[b1], . . . , [bn]} is a basis with the same cardinality for X/∼. (2) Conversely, if B ∗ = {[b1], . . . , [bn]}, where each bi is invertible, is a basis for X/∼, then B = {b1, . . . , bn} is a basis with the same cardinality for X. i=1 bki i i=1 bki i=1 bki Proof. The unique expansions of bi, bi′ ∈ B relative to B are bi = 1bi and bi′ = 1bi′, respectively, so µB(bi) = µB(bi′ ) = 1. Hence, [bi] = [bi′ ] implies bi = bi′ since implies µB(bi′ )bi = µB(bi)bi′ by Lemma 2, so the surjective mapping bi ∼ bi′ bi 7→ [bi] is injective as well. i=1[bi]ki , so B ∗ generates X/∼. i=1[bi]ℓi, then hQn for i = 1, . . . , n by Lemma 2. (2). Let x be an arbitrary quantity in X. As B ∗ generates X/∼, we have [x] = is invertible, Proposition 7 (1) implies (1). Let [x] be an arbitrary dimension in X/∼. As B generates X, x = µQn i i =hQn for some µ ∈ K and some integers k1, . . . , kn, so [x] =hµQn Qn Also, if [x] =Qn i=1[bi]ki = hQn Qn that there exists some κ ∈ K and integers k1, . . . , kn such that x = κQn i i = hνQn Finally, if x = µQn i i = hQn hQn i=1[bi]ki =Qn i i, and as Qn i = νQn i i, so Qn i=1[bi]ki = Qn i i = i i, so ki = ℓi i=1 bki i=1[bi]ℓi , so ki = ℓi for i = 1, . . . , n, since B ∗ is a basis for X/ ∼. Also, x = µy = νy, where y is invertible, so µ = ν by Proposition 7 (2). (cid:3) then hµQn i i =hQn i=1 bki i . i=1 bℓi i i, so i=1 bki i=1 bℓi i=1 bki i=1 bℓi i i=1 bki i=1 bki i=1 bki i i=1 bℓi Corollary 2. Let X be a quantity space over K. (1) X/∼ is a free abelian group. (2) Any two bases for X/∼ have the same number of elements, any basis for X has the same number of elements as any basis for X/∼, and any two bases for X have the same number of elements. Proof. (1) is immediate. To prove (2), it suffices to note that any two bases for a free abelian group have the same cardinality. (cid:3) 4.5. Concerning algebraic structures acting on monoids of quantities. In theoretical discussions about measurement, the measure of a physical quantity is usually considered to be a real number, so it is natural to let the field associated with a quantity space be the real numbers R. However, some physical quantities, such as a distance or a mass, can only have positive or non-negative measures. This suggests that the field acting on the monoid of quantities should be replaced by a more general structure. In [Jon14a], the concept of a scaloid over a field is replaced by the concept of a scaloid over a so-called scalar system. A scalar system can be conveniently defined as a subset of a field, inheriting addition and multiplication in the field, such that it is closed under addition and its non-zero elements constitute a group under multiplication. The real numbers, the non-negative real numbers, and the positive real numbers are obvious examples of scalar systems. Dan Jonsson 8 There is, however, a way to accommodate constrained sets of measures without modifying the notion of a quantity space: we let functions of the form µB : D → R be partial functions, defined on some subset S of D, so that, for example, µB(x) > 0 for all x ∈ S. This approach is quite flexible; it also works, for example, when we require µB(S) to be a discrete set of numbers. It is an elementary observation that physical quantities are ordered; three grams is more than two grams, and so on. This means that it should be possible to equip a quantity space with some kind of order relation, not considered so far in this article. Such an order relation can be induced by the order on an ordered field acting on the monoid of quantities. Specifically, let X be a quantity space over an ordered field K. To induce an order relation, based on the order on K, on X, we need to to choose an orientation for X. Analogous to how we proceed for vector spaces, we choose an orientation by fixing a basis B = {b1, . . . , bn} for X and stipulating that for any D ∈ X/∼ and x ∈ D we have x ≥ 0D if and only if there are integers k1, . . . , kn and µ ∈ K such i where µ ≥ 0. Moreover, we stipulate that x ≥ y if and only if i=1 bki that x = µQn (x − y) ≥ 0D. In view of this construction, the fact that R is an ordered field gives another reason for considering quantity spaces over R. 5. In conclusion As shown in the Introduction, scalable monoids complement rings and modules from an abstract mathematical point of view. From the point of view of appli- cations, a quantity space is a natural counterpart to a vector space: quantities and vectors are both fundamental notions in physics and quantitative sciences gen- erally. Recall that the transformation of vector space theory into an axiomatic, "coordinate-free" form was completed during the interwar period in Europe, more than three quarters of a century ago, but the corresponding formulation of an axiomatic, "coordinate-free" notion of quantities has not yet been completed, in my opinion, despite important contributions [Dro53, Whi68a, Whi68b]). Indeed, the research reported here started as an attempt to model quantities, not as an attempt to fill a mathematical lacuna. (e.g., A crucial feature of systems of quantities is that two quantities can be added if and only if they are "commensurable", so mathematical models of systems of quantities should reflect this peculiar property, not shared by numbers without sorts. An algebraic structure satisfying this requirement can of course be found if sufficiently complicated algebraic structures are considered. However, it would seem to be desirable that systems of quantities be modeled by a formal mathematical structure defined in a simple, direct manner, similar to the definition of a vector space, and in such a way that the crucial "commensurability" feature is a natural consequence of this formulation. The definitions given here would seem to pass this test. Applications of quantity spaces, considered only informally and superficially here, are discussed at some length in [Jon14a, Jon14b]. Scalable Monoids 9 References [Bar96] Barenblatt G.I. (1996). Scaling, Self-similarity and Intermediate Asymptotics. Cam- bridge University Press. [Bri22] Bridgman P.W. (1922). Dimensional Analysis. Yale University Press. [Dro53] Drobot S. (1953). "On The Foundations of Dimensional Analysis". Studia Mathematica, 14, 84 -- 99. [Jon14a] Jonsson D. (2014). Quantities, Dimensions and Dimensional Analysis. arXiv:1408.5024. [Jon14b] Jonsson D. (2014). Dimensional Analysis: A Centenary Update. arXiv:1411.2798. [Whi68a] Whitney H. (1968). "The Mathematics of Physical Quantities. Part I: Mathematical Models for Measurement", American Mathematical Monthly, 75, 115 -- 138. [Whi68b] Whitney H. (1968). "The Mathematics of Physical Quantities. Part II: Quantity Struc- tures and Dimensional Analysis", American Mathematical Monthly, 75, 227 -- 256 University of Gothenburg, SE-405 30 Gothenburg, Sweden E-mail address: [email protected]
1304.2451	1	1304	2013-04-09T02:41:43	Multihomogeneous Normed Algebras and Polynomial Identities	[ "math.RA" ]	In this paper we consider PI-algebras $A$ over $\R$ or $\C$. It is well known that in general such algebras are not normed algebras. In fact, there is a nilpontent commutative algebra which is not a normed algebra, see [1]. Here we address the question of whether it is possible to find a normed PI-algebra $B$ with the same polynomial identities as $A$, and moreover, whether there is some Banach PI-algebra with this property. Our main theorem provides an affirmative answer for this question and moreover we also show the existence of a Banach Algebra with the same polynomial identities as $A$. As a byproduct we prove that if $A$ is a normed PI-algebra and its completion is nil, then $A$ is nilpotent. By introducing the concept of multihomogeneous norm we obtain as an application of our main results that if $\FX$ is multihomogeneus normed algebra and $A$ is a PI-algebra such that the completion of the quotient space $\FX/Id(A)$ is nil, then $A$ is nilpotent. Both applications are extensions of the study initiated in [4].	math.RA	math	Multihomogeneous Normed Algebras and Polynomial Identities Leandro Cioletti∗ Jos´e Antonio Freitas† Departamento de Matem´atica, Universidade de Bras´ılia, 70910-900 Bras´ılia, DF, Brazil Dimas Jos´e Gon¸calves‡ Departamento de Matem´atica, Universidade Federal de Sao Carlos, 13565-905 Sao Carlos, SP, Brazil May 1, 2018 Abstract In this paper we consider PI-algebras A over R or C. It is well known In fact, there is a that in general such algebras are not normed algebras. nilpontent commutative algebra which is not a normed algebra, see [1]. Here we address the question of whether it is possible to find a normed PI-algebra B with the same polynomial identities as A, and moreover, whether there is some Banach PI-algebra with this property. Our main theorem provides an affirmative answer for this question and moreover we also show the existence of a Banach Algebra with the same polynomial identities as A. As a byprod- uct we prove that if A is a normed PI-algebra and its completion is nil, then A is nilpotent. By introducing the concept of multihomogeneous norm we obtain as an application of our main results that if F hXi is multihomogeneus normed algebra and A is a PI-algebra such that the completion of the quo- tient space F hXi/Id(A) is nil, then A is nilpotent. Both applications are extensions of the study initiated in [4]. Key words: PI-Algebras, Normed Algebras, Banach Algebras. 2010 Mathematics Subject Classification: 16R10, 16R40, 46H10. ∗[email protected] †[email protected]; Partially supported by grant from CNPq No. 478318/2010-3 ‡[email protected]; Partially supported by grant from CNPq No. 478318/2010-3 1 1 Introduction We begin this article by stating precisely some of our main results and then we proceed to introduce the concept of multihomogeneous norm. In order to be concise and objective, we will skip the precise definition of some of the basic concepts needed here, such as PI-algebra and Normed algebra. These, together with other additional background concepts, will appear in detail in Section 2. The proofs of the results stated here are found in the last section. Let F be the field R or C and F hXi the free non-unitary associative algebra, freely generated over F by the infinite set X = {x1, x2, . . .}. All the algebras considered in this paper will be non-unitary, associative and over the field F . Thus for convenience we will only use the term algebra. A good example to keep in mind is the algebra F hXi. For a normed algebra A we write C(A) to denote its completion. If A is a PI-algebra then we denote by Id(A) the set of all polynomial identities of A. The statement of our first result is: Proposition 1. If A is a normed PI-algebra, then Id(A) = Id(C(A)). In other words this proposition tell us that every normed PI-algebra A has the same polynomial identities that some Banach PI-algebra. Since not all PI-algebras are normed PI-algebras, see for example [1], a natural question to ask is: given a PI-algebra A, is there some Banach PI-algebra B with the same polynomial identities of A ? As we said before we give an affirmative answer for this question and we also show how to construct such algebra B. To explain the construction we introduce some definitions. Let f = f (x1, . . . , xn) ∈ F hXi be a polynomial, which will be written as f = Xd1≥0,...,dn≥0 f (d1,...,dn) , where f (d1,...,dn) = f (d1,...,dn)(x1, . . . , xn) is the multihomogeneous component of f with multidegree (d1, . . . , dn). Definition 2. A norm · in F hXi is called multihomogenous if f (d1,...,dn) ≤ f for all f = f (x1, . . . , xn) ∈ F hXi and all d = (d1, . . . , dn). If F hXi is a normed algebra with respect to a multihomegeneous norm, then we say that F hXi is a MN-algebra. An example of MN-algebras can be obtained as follows. Take f = Pm αmm, where αm ∈ F and m is a monomial, then F hXi with the norm is a MN-algebra. f = Xm αm. 2 In the sequel, we prove that if F hXi is a MN-algebra and if A is a PI-algebra, then Id(A) is a closed ideal of F hXi. Thus the multihomogeneous norm in F hXi induces a norm in the quotient algebra F hXi/Id(A) by f + Id(A) = inf{f + g : g ∈ Id(A)}, where f ∈ F hXi. We remark that the quotient F hXi/Id(A) is a normed algebra with this norm and using this fact we obtain our second main result: Theorem 3. Let F hXi be a MN-algebra. If A is a PI-algebra, then Id(A) = Id(cid:18)C (cid:18) F hXi Id(A)(cid:19)(cid:19) . In particular, a PI-algebra has the same polinomial identities that some Banach PI-algebra. As an application, we obtain similar results as in the Grabiner's paper. In [4] the author proves the following: Theorem 4. Let A be a Banach algebra. If A is nil then A is nilpotent. In the above theorem, the algebra A is required to be a Banach algebra. Here we investigate when the nilpotency of an algebra A can be obtained by hypothesis imposed on C(A). In this direction our first result is Corollary 5. Let A be a normed PI-algebra. If C(A) is nil, then A is nilpotent. Our second result relates the nilpotency of A to the completion of certain quotient space related to the polynomial identities of A. To be more precise we prove the following: Corollary 6. Let F hXi be a MN-algebra and let A be a PI-algebra. If Id(A)(cid:19) C (cid:18) F hXi is nil, then A is nilpotent. 2 Banach and PI-Algebras An algebra A is said to be normed if it satisfies the followings properties: a) A has a norm k · k; b) kabk ≤ kakkbk for all a, b ∈ A. A normed algebra A is called Banach algebra if A is a complete normed space. It's well known that every normed algebra A is contained in some Banach algebra C(A) such that A is dense in C(A). This algebra is the completion of A. Here we recall its construction: we first define the relation ∼ in the set of all Cauchy sequences of A by (an)n ∼ (bn)n ⇐⇒ lim n→∞ kan − bnk = 0. 3 Denote by C(A) the set of all equivalence classes. If (an)n is a Cauchy sequence in A, we denote by [(an)n] its equivalence class. The algebraic operations in C(A) are defined as usual: [(an)n] + [(bn)n] = [(an + bn)n], for any λ ∈ F , we put λ[(an)n] = [(λan)n] and [(an)n][(bn)n] = [(anbn)n]. Endowed with these operations and with the norm k[(an)n]k = lim n→∞ kank we have that C(A) is a Banach algebra. Note that an element a ∈ A is identified with the class [(a)n] ∈ C(A) of the constant sequence equal to a. For more details see [6, 5]. Let F hXi be the free non-unitary associative algebra, freely generated over F by the infinite set X = {x1, x2, . . .}. The elements of F hXi are called polynomials and a polynomial of the kind xi1xi1 . . . xin is called monomial. A polynomial f (x1, . . . , xm) ∈ F hXi is called a polynomial identity for an algebra A if f (a1, . . . , am) = 0, for all a1, . . . , am ∈ A. We denote by Id(A) the set of all polynomial identities of A. If Id(A) 6= {0} then we say that A is a PI-algebra. The set Id(A) is an ideal in F hXi and has the property f (g1, . . . , gm) ∈ Id(A) for all f (x1, . . . , xm) ∈ Id(A) and g1, . . . , gm ∈ F hXi. Thus we say that Id(A) is a T-ideal. For details, see [2, 3]. Let F hXi(d1,...,dm) be the vector subspace of F hXi spanned by all monomials u = xj1 . . . xjt, where the variable xi appears di times in u for all i = 1, . . . , m. If f (x1, . . . , xm) ∈ F hXi(d1 ,...,dm) then we say that f is multihomegenous of mul- tidegree (d1, . . . , dm). Note that if f = f (x1, . . . , xm) ∈ F hXi, we can always write f = Xd1≥0,...,dm≥0 f (d1,...,dm) where f (d1,...,dm) ∈ F hXi(d1 ,...,dm). The polynomials f (d1,...,dm) are called the mul- tihomogenous components of f . Now we recall an important result that will be used in the next section. Theorem 7. If f = f (x1, . . . , xm) is a polynomial identity for an algebra A, then every multihomogeneous component of f is a polynomial identity for A. Proof. See [3], Theorem 1.3.2. We call the reader's attention to the fact that the above result holds for every infinite field F , but it is not always true for finite fields. 3 Proofs of the Main Results In this section we give the proofs of the main results of the paper. Proof. (Proposition 1) Since A ⊆ C(A), it follows that Id(C(A)) ⊆ Id(A). Let f (x1, . . . , xm) ∈ Id(A) and a1, . . . , am ∈ C(A). By the construction of C(A) we have f (a1, . . . , am) = f ([(a1n)n], . . . , [(amn)n]) = [(f (a1n, . . . , amn))n] = [(0)n]. The last identity implies that f ∈ Id(C(A)). Therefore Id(A) ⊆ Id(C(A)). 4 Let g(x1, . . . , xt) a polynomial and let g(d1,...,dt) its multihomogeneous compo- nent of multidegree (d1, . . . , dt). If m < t and dm+1 = dm+2 = . . . = dt = 0, then we write If t < m, then we write g(d1,...,dm,...,dt) = g(d1,...,dm). g(d1,...,dt) = g(d1,...,dt,0,...,0), where the number of zeros is m − t. With this convention we have the following lemma: Lemma 8. Let F hXi be a MN-algebra and let f = f (x1, . . . , xm) be a polynomial. If (fn)n is a sequence in F hXi such that fn → f , then for all multidegree d = (d1, . . . , dm). f (d1,...,dm) n → f (d1,...,dm) Proof. Once k · k is a multihomogeneous norm we have the following inequality kf (d1,...,dm) n − f (d1,...,dm)k ≤ kfn − f k. Since fn → f it follows that f (d1,...,dm) n → f (d1,...,dm). Proposition 9. Let A be a PI-algebra. If F hXi is a MN-algebra, then Id(A) is a closed ideal. Proof. Let (fn)n ∈ Id(A) be a sequence of polynomials such that fn → f . We want to prove that f ∈ Id(A). Write f = f (x1, . . . , xm) = X(d1,...,dm) f (d1,...,dm). By the Lemma 8, we have that f (d1,...,dm) Note that f (d1,...,dm) n → f (d1,...,dm) for all multidegree (d1, . . . , dm). n ∈ F hXi(d1,...,dm) ∩ Id(A) by the Theorem 7. Since F hXi(d1 ,...,dm) is a finite-dimensional vector space follows that F hXi(d1,...,dm) ∩ Id(A) has also finite dimension. Since every finite-dimensional space is closed in the norm topology, we have that F hXi(d1 ,...,dm) ∩ Id(A) is closed. Thus f (d1,...,dm) ∈ F hXi(d1,...,dm) ∩ Id(A) and therefore f ∈ Id(A). If A is a PI-algebra and F hXi is MN-algebra, then by the above proposition we can define a norm in the quotient algebra F hXi/Id(A): kf + Id(A)k = inf{kf + gk : g ∈ Id(A)}, where f ∈ F hXi. With this norm we have that F hXi/Id(A) is a normed alge- bra. So we can see that the Theorem 3 describes the polynomial identities of the completion of this quotient algebra. Now we proceed to its proof. 5 Proof. (Theorem 3) By a classical result in PI-Algebra we have Id(A) = Id(cid:18) F hXi Id(A)(cid:19) , see [3]. So the proof of the theorem follows immediately from Proposition 1. Proof. (Corollary 5) If C(A) is nil, then by Theorem 4, we have that C(A) is nilpotent. Thus x1x2 . . . xn is a polynomial identity of C(A) for some n. Since by Proposition 1 we have Id(A) = Id(C(A)), follows that A is nilpotent. Proof. (Corollary 6) Let B = F hXi/Id(A). If C(B) is nil, then by Theorem 4 we have that C(B) is nilpotent. Thus x1x2 . . . xn is polynomial identity of C(B) for some n. Since by Theorem 3 we have Id(A) = Id(C(B)), follows that A is nilpotent. References [1] H. G. Dales, Norming Nil Algebras, Proc. Amer. Math. Soc., 83, Number 1, 71 -- 74, 1981. [2] V. Drensky, Free algebras and PI-algebras, Graduate Course in Algebra, Springer, Singapore, 1999. [3] A. Giambruno, M. Zaicev, Polynomial identities and asymptotic methods, Math. Surveys Monographs 122, AMS, Providence, RI, 2005. [4] S. Grabine, The nilpotency of Banach nil algebras, Proc. Amer. Math. Soc., 21, 510, 1969. [5] I. Kaplansky, Set Theory and Metric Spaces, Allyn and Bacon Series in Ad- vanced Mathematics, Boston, Mass., 1972. [6] M. A. Naimark, Normed algebra, Wolters-Noordhoff, 3 edition, 1972. 6
1504.04001	1	1504	2015-04-15T19:52:18	Triangulable Leibniz Algebras	[ "math.RA" ]	A converse to Lie's theorem for Leibniz algebras is found and generalized. The result is used to find cases in which the generalized property, called triangulable, is 2-recognizeable; that is, if all 2-generated subalgebras are triangulable, then the algebra is also. Triangulability joins solvability, supersolvability, strong solvability, and nilpotentcy as a 2-recognizeable property for classes of Leibniz algebras.	math.RA	math	TRIANGULABLE LEIBNIZ ALGEBRAS TIFFANY BURCH AND ERNIE STITZINGER ABSTRACT: A converse to Lie's theorem for Leibniz algebras is found and generalized. The result is used to find cases in which the generalized property, called triangulable, is 2-recognizeable; that is, if all 2-generated subalgebras are triangulable, then the algebra is also. Triangulability joins solvability, supersolvability, strong solvability, and nilpotentcy as a 2-recognizeable property for classes of Leibniz algebras. Simultaneous triangulation of a set of linear transformations related to Leibniz algebras will be considered in this work. A necessary condition is that the minimum polynomials of the linear transformations are the products of linear factors. If the matrices commute, then this condition is sufficient. Lie's theorem is a generalization of this result, a result that is field dependent even in the algebraically closed case. Lie's theorem extends to Leibniz algebras. It can also be useful to have a condition on the set which guarantees simultaneous triangulation when extending to the algebraic closure, K, of the field of scalars, F . We will call this property, borrowing from Bowman, Towers and Varea [3], triangulable. Thus a Leibniz algebra A over a field F is triangulable on module M if the induced representation of K N A on K N M admits a basis such that the representing matrices are upper triangular. We find a condition on the Leibniz algebra that is equivalent to being triangulable. The condition is independent of the field and includes Lie's theorem. The condition is on A although triangulable refers to the action in the algebra that is over K. As an application, we consider to what extent an algebra is triangulable if all 2-generated subalgebras have the property. Such a property is called 2-recognizable, a concept investigated in groups, Lie algebras and Leibniz algebras for classes such as solvable, supersolvable, strongly solvable and, trivially, nilpotent and abelian. For an introduction to Leibniz algebras, see [1], [6] and [7]. Let M be a module for the Leibniz algebra A. Let left and right multiplication by x ∈ A be denoted by Tx and Sx, respectively. If Tx is nilpotent, then Sx is also nilpotent and an Engle theorem holds; that is, A acts nilpotently on M [2]. Define A to be nil on M if this property holds. If I is an ideal of a Leibniz algebra A, then I is nil on A if and only if I is nilpotent. We will show that A is triangulable on itself if and only if A2 is nilpotent. When A2 is nilpotent, then A is called strongly solvable, a term introduced in [3]. Lemma 1. Let B be an irreducible submodule of the A-module M such that dim(B) = 1. Then A2 ⊆ CA(B) = {a ∈ AaB = 0 = Ba}. Proof. Let x, y ∈ A, b ∈ B with xb = αxb where αx is a scalar. Then (xy)b = x(yb) − y(xb) = x(αyb) − y(αxb) = αxαyb − αyαxb = 0. Since B is irreducible, it follows that b(xy) = 0 ([1]). Hence A2B = BA2 = 0 and the result holds. (cid:3) Theorem 1. Let A be a Leibniz algebra and M be an A-module. There is a basis for M such that the representing matrices are in upper triangular form if and only if the linear transformations in the representation have minimum polynomials with all linear factors and A2 is nil on M . 1 We then have Corollary 1. Let A be a Leibniz algebra over an algebraically closed field and M be an A-module. There is a basis for M such that the representing matrices are in upper triangular form if and only if A2 is nil on M . Proof of Theorem. The condition on the minimum polynomials is equivalent to being able to tri- angulate all the linear transformations. Suppose also that A2 is nil on M . Then Tx, and hence Sx, is nilpotent for all x ∈ A2 ([2]) and A2 acts nilpotently on M by Engel's theorem. Let B be an irreducible submodule of M . It is enough to show that dim(B) = 1 for then the result will follow by induction. Since the submodule BA2 + A2B is contained in B, BA2 + A2B is 0 or B. If the latter holds, then B = BA2 + A2B = (BA2 + A2B)A2 + A2(BA2 + A2B) = . . . , which is never 0, a contradiction. Hence BA2 + A2B = 0 and BA2 = 0 = A2B. Therefore, 0 = (xy)b = x(yb) − y(xb) for all x, y ∈ A, b ∈ B. Hence x(yb) = y(xb) and TxTy = TyTx. Since B is irreducible, either Sx = 0 for all x ∈ A or Sx = −Tx for all x ∈ A [1]. Then either TxSy = −TxTy = −TyTx = SyTx or TxSy = 0 = SyTx on B and SxSy = TxTy = TyTx = SySx or SxSy = 0 = SySx on B. Hence all Tx, Sy commute on B. Thus these transformations are simultaneously triangularizeable on B. Hence dim(B) = 1 and the result holds in this direction. Suppose that there is a basis for M such that all the transformations are triangularizeable on this basis. Then there is a one dimensional submodule B such that A2B = 0 = BA2 by the lemma. A2 is nil on M/B by induction and hence on M . Since each transformation is triangularizeable, the minimum polynomials have linear factors. (cid:3) We state these results when A is acting on itself. A is called strongly solvable if A2 is nilpotent ([3], [5]). A is supersolvable if there is an increasing chain of ideal of A, each of codimension 1 in the next. Theorem 2. Let A be a Leibniz algebra over an algebraically closed field. Then A is supersolvable if and only if A is strongly solvable. Proof. A is supersolvable if and only if A triangulable on itself. By Corollary 1, this is equivalent to A2 being nil on A which in turn is the same as A2 being nilpotent. The last statement is the definition of strongly solvable. (cid:3) The general case looks like the following. Theorem 3. A is triangulable if and only if A2 is nil on A. Proof. Let A∗ = K N A where K is the algebraic closure of the base field for A. If A2 is nil on A, then A2 is nilpotent as is (A2)∗ = (A∗)2. Hence A∗ is supersolvable by the last result and A is triangulable by definition. The converse holds by reversing the steps. (cid:3) A property is n-recognizeable in a class of algebras if whenever all n generated subalgebras have the property, then the algebra has the property. The concept when n = 2 has been considered in groups, Lie algebras and Leibniz algebras for classes that are solvable, supersolvable, nilpotent and 2 strongly solvable (see [4], [5], [8]). A version of this has also been considered for triangulable in the Lie case in [3]. We considered it as an application of our result. Theorem 4. In the class of solvable Leibniz algebras, triangulable is 2 recognizable. Proof. Let N be a 2-generated subalgebra of A. Since N is triangulable, N 2 is nil on N and, thus, N is strongly solvable. Therefore A is strongly solvable by [5] and A2 is nil on A. Then, using Theorem 3, A is triangulable. (cid:3) The result can be extended from solvable to all algebras in characteristic 0, again using a result in [5]. Corollary 2. For Leibniz algebras, triangulable is 2-recognizeable at characteristic 0. Proof. Let A be a minimal counterexample. Simple Leibniz algebras are Lie algebras. Simple Lie algebras are 2-generated, hence A is triangulable. If I is a proper ideal in A, then all 2-generated subalgebras of I and A/I are strongly solvable by Theorem 3 and I and A/I are strongly solvable by Theorem 3 of [5] and A is solvable. The result holds by Theorem 4. (cid:3) References [1] D. Barnes, Some theorems on Leibniz algebras, Comm. Alg. 39 (2011) 2463-2472 [2] L. Bosko, A. Hedges, J.T. Hird, N. Schwartz, K. Stagg, Jacobson's refinement of Engel's theorem for Leibniz algebras, Involve 4 (2011)293-296 [3] K. Bowman, D. Towers, V. Varea, Two generator subalgebras of Lie algebras, Linear and Multilinear Algebra 55 (2007) 429-438 [4] R. Brandl, Zur Theorie der untergruppenabgeschlossenen Formationen: endliche Varietten, J.Algebra 73 (1981) 1-22 [5] T. Burch, M. Harris, A. McAlister, E. Rogers, E. Stitzinger, S.M. Sullivan, 2-recognizeable classes of Leibniz algebras, J. Algebra. 423 (2015). 506-513 [6] I. Demir, K. Misra, E. Stitzinger, On some structures of Leibniz algebras, AMS Contemporary Math. 623 (2014) 41-54 [7] J.-L. Loday, Une version non commutative des algebres de Lie: les algebres de Leibniz, Enseign Math. (2) (1993) 39 (3-4) 269-293 [8] K. Moneyhun, E. Stitzinger, Some finite varieties of Lie algebras, J. Algebra 143 (1991) 173-17 3
1811.07165	2	1811	2018-12-16T17:17:34	Constructing large self-small modules	[ "math.RA" ]	We give a method for constructing (possible large) self--small modules via some special homomorphisms of rings, called here weak epimorphisms.	math.RA	math	CONSTRUCTING LARGE SELF -- SMALL MODULES GEORGE CIPRIAN MODOI Abstract. We give a method for constructing (possible large) self -- small mod- ules via some special homomorphisms of rings, called here weak epimorphisms. Various kind of smallness appear naturally in the study of situations in which the covariant or contravariant hom functor induces an equivalence, respectively a duality between some categories of modules. For example Morita theory says that if R is an arbitrary ring, P is a progenerator in the category Mod-R of right R-modules and E = EndR(P ) is its endomorphism ring, then the functor HomR(P, −) : Mod-R → Mod-E is an equivalence, with the inverse the tensor product − ⊗E P . In these conditions, P has to be small, that is HomR(P, −) has to commute with arbitrary direct sums. The smallness notion can be generalized in various ways, by imposing some restrictions to the class of direct sums which the covariant hom functor has to commute. In this note we deal with the following generalization: A self -- small R-module is a module M such that HomR(M, M (I)) ∼= Hom(M, M )(I), naturally for every set I. Self -- small abelian groups (that is, Z-modules) were introduced by Arnold an Murley in [2]. The relevance of the study of self -- small abelian groups is justified by many papers (see, for example, [1] and the references therein). In this note we want to construct a module which is self -- small but it is large in some sense. More precisely, we want this self -- small module to be not small. Because finitely generated modules are always small, the modules we are looking for have to be infinitely generated. The method is inspired by the construction of the abelian group of p-adic integers Jp, where p is a prime. In this case, Jp is uncountable, that is its cardinality is also larger than the cardinality of the ring of integers Z. Note that another way of constructing large self -- small modules can be found in [8]. More precisely, from [8, Example 2.7] we learn that the direct product Qp Z/p is self -- small, but the direct sum Lp Z/p is not, where p runs over all primes and Z/n = Z/nZ for every n ∈ N. More generally, for a ring R let denote by SR a representative set of simple modules. Then in [8, Theorem 2.5 and Corollary 1.3] we find some sufficient conditions for the direct product QS∈SR S to be, respectively to be not self -- small. In what follows we consider two rings with one R and J, and we denote by Mod-R and Mod-J the respective categories of modules (which by default are left modules). Let ϕ : R → J a unitary ring homomorphism. Thus J has a natural structure of R−R-bimodule and ϕ induces a pair of adjoint functors (the restriction Date: December 18, 2018. 1991 Mathematics Subject Classification. 16D99, 20K40. Key words and phrases. self -- small module, weak epimorphism of rings. 1 2 GEORGE CIPRIAN MODOI and the induction of the scalars): ϕ∗ = HomJ (J, −) : Mod-J ⇄ Mod-R : (− ⊗R J) = ϕ∗. The restriction functor ϕ∗ acts as follows: ϕ∗(M ) = M and ax = ϕ(a)x for all J-modules M and all x ∈ M and a ∈ R. Henceforth it is obviously faithful, since it sends a J-linear map in itself, but seen as R-linear. Recall that ϕ is called an epimorphism of rings, if for every two parallel homo- morphisms of rings ψ, ζ : J → J ′ we have ψ · ϕ = ζ · ϕ ⇒ ψ = ζ. This is the case, exactly if ϕ∗ is full too, by [7, Ch. XI, Proposition 1.2], therefore if we have HomR(M, N ) ∼= HomJ (M, N ) for all M, N ∈ Mod-J. Inspired by this, we call ϕ weak epimorphism if HomR(J, J) ∼= HomJ (J, J), that is HomR(J, J) ∼= J. Proposition 1. If ϕ : R → J is a weak epimorphism of rings, then J is self -- small as R-module. Proof. Let I be a set and denote by πi : J (I) → J the projection of the coproduct of copies of J into its i-th component (i ∈ I). If f : J → J (I) is an arbitrary R-linear map, then πif : J → J is R-linear for all i ∈ I. According to our hypothesis it is J-linear too, therefore it is determined by πif (1) ∈ J. Because πif (1) 6= 0 only for a finite number of i's, we conclude that πI f = 0 for almost all i ∈ I, hence f factors through a finite subcoprodct of J (I), what is the same as saying that HomR (cid:0)J, J (I)(cid:1) ∼= Hom(J, J)(I). (cid:3) Since epimorphisms of rings are obviously weak epimorphisms too we obtain: Corollary 2. If ϕ : R → J is an epimorphism of rings, then the R-module J is self -- small. Example 3. The inclusion Z → Q is known to be an epimorphism of rings, namely one which is not surjective. Therefore Corollary 2 above gives us a new proof that the abelian group Q is self -- small. In the sequel we assume that the ring R is commutative. Thus Mod-R coincide to the category of right R-modules, and HomR(M, N ) is an R-module for all M, N ∈ Mod-R. In Mod-R consider an ascending chain of submodules of the module (DS) Z1 µ1→ Z2 µ2→ Z3 → . . . , Z∞ = lim → Zn = lim → (Z1 µ1→ Z2 µ2→ Z3 → . . .) = [ Zn, the morphisms µn being inclusions. Relative to the above chain consider the fol- lowing conditions: (1) All modules Zm are finitely presented. (2) HomR(Zm, Zn) ∼= Zm naturally, for all 1 ≤ m ≤ n. (3) The R-module Z∞ is injective relative to all exact sequences 0 → Zm µn→ Zm+1 → Zm+1/Zm → 0, with m ≥ 1. CONSTRUCTING LARGE SELF -- SMALL MODULES 3 (4) Z1 is simple, and denote by U the annihilator of Z1 in R, that is U is a maximal ideal in R and there is a short exact sequence 0 → U → R → Z1 → 0. Moreover assume that Zm+1U = Zn, for all m ∈ N∗. (5) Zm ⊗ Z1 ∼= Z1 naturally, for all m ∈ N∗. Note that the condition (3) is automatically satisfied, if we know that the R- module Z∞ is injective. On the other hand we can replace (3) with a condition relative to the direct system (DS), rather than relative to its direct limit Z∞, as in the the following: (3') The R-module Zn is injective relative to all exact sequences 0 → Zm µn→ Zm+1 → Zm+1/Zm → 0, with 1 ≤ m < n. Lemma 4. If (1) and (3') are satisfied then (3) holds too. Proof. The condition (3') implies that the induced homomorphism HomR(Zm+1, Zn) → HomR(Zm, Zn) is surjective for all 1 ≤ m < n. The condition (1) says that all Zi, i ≥ 1 are finitely generated, and this means the functors HomR(Zi, −) commute with direct limits as we can see from [7, Ch. V, Proposition 3.4]. We deduce that the induced homomorphism HomR(Zm+1, Z∞) ∼= lim → HomR(Zm+1, Zn) HomR(Zm, Zn) ∼= HomR(Zm, Z∞) → lim → is also surjective, therefore (3) holds. (cid:3) Lemma 5. If (1) and (2) hold, we have for all m ≥ 1 a natural isomorphism: HomR(Zm, Z∞) ∼= Zm. Proof. Using again the property that HomR(Zi, −) commutes with direct limits, for all i ≥ 1, we get: HomR(Zm, Z∞) = HomR(Zm, lim → Zn) ∼= lim → HomR(Zm, Zn) ∼= Zm, where the last isomorphisms follows from the fact that (2) implies that the direct system {HomR(Zm, Zn)}n≥1 looks like HomR(Z1, Zm) → . . . → HomR(Zm−1, Zm) → Zm =→ Zm =→ Zm =→ . . . , that is it has a cofinal constant subsystem. (cid:3) Assume that (1) and (2) hold. For all n ≥ 1, we denote δn the composed homomorphism Zn+1 ∼=→ HomR(Zn+1, Z∞) (µn)∗→ HomR(Zn, Z∞) ∼=→ Zn, where the isomorphisms are coming from Lemma 5. We obtain an inverse system of R-modules (IS) Z1 δ1← Z2 δ2← Z3 ← . . . Let now denote J = EndR(Z∞). Thus J is naturally an R-algebra, and let ϕ : R → J denote the structure homomorphism of this algebra. 4 GEORGE CIPRIAN MODOI Lemma 6. If (1) and (2) hold, we have a natural isomorphism in Mod-R: J ∼= lim ← Zn = lim ← (Z1 δ1← Z2 δ2← Z3 ← . . .). Proof. The chain of isomorphisms (the last one coming from Lemma 5) J = HomR(Z∞, Z∞) ∼= HomR(lim → Zn, Z∞) ∼= lim ← HomR(Zn, Z∞) ∼= lim ← Zn. proves our lemma. (cid:3) For the inverse system (IS) we denote δjj = 1Zj and δji = δj . . . δi, for all 1 ≤ j ≤ i. With these notations, the inverse system is called Mittag -- Leffler if for each k ≥ 1 there is j > k such that Im(δki) = Im(δkj ) for all j ≤ i. In particular this is always true, provided that the homomorphisms δi are surjective, for all i ≥ 1. Lemma 7. If (1), (2) and (3) hold, then the inverse system (IS) is Mittag -- Leffler. Proof. The homomorphism (µn)∗ is surjective by (3), so the same property is true for δn, and the conclusion follows. (cid:3) From now on, we assume that all conditions (1)-(5) hold. Lemma 8. We have Zn+m/Zm ∼= Zn for all n, m ∈ N∗. Proof. First we will show that Zn+1/Zn ∼= Z1 for all n ∈ N∗. Indeed, applying the functor Zn+1 ⊗R − to the short exact sequence 0 → U → R → Z1 → 0, keeping in the mind that Zn = Zn+1U is the image of the map Zn+1 ⊗R U → Zn+1 ⊗R R and using condition (5) for the isomorphism in the last vertical arrow, we get a commutative diagram with exact rows Zn+1 ⊗R U Zn+1 ⊗R R Zn+1 ⊗R Z1 0 / Zn which proves our claim. ∼= / Zn+1 ∼= / Z1 0 / 0 Fix n ∈ N∗ and proceed by induction on m. For m = 1, we apply the functor HomR(−, Z∞) to the exact sequence from the second row of the last diagram. According to (3), we get an exact sequence too, which by Lemma 5 looks like: 0 → Z1 → Zn+1 → Zn → 0, proving our desired isomorphism Zn+1/Z1 ∼= Zn. Suppose now that Zn+m/Zn ∼= Zm. Then construct the diagram having exact rows and columns (the exactness of the rows is shown in the first part of this proof, the induction hypothesis gives exactness of the first column, and for the second / / / / / / / / / / CONSTRUCTING LARGE SELF -- SMALL MODULES 5 column it is obvious): 0 0 / Zm / Zm+1 / Zn+m / Zn+m+1 0 0 / Z1 / Z1 / 0 / 0 Zn 0 Zn+m+1/Zm+1 0 Now the Ker-Coker lemma gives us the isomorphism Zn+m+1/Zm+1 ∼= Zn. Remark 9. Puttig together above lemmas, we deduce that for all n, m ∈ N∗ we have the short exact sequences (cid:3) 0 → Zn → Zn+m → Zm → 0 and 0 → Zn → Zn+m → Zm → 0 and the functor HomR(−, Z∞) sends them to each other. Lemma 10. There is a short exact sequence 0 → J u→ J → Z1 → 0 such that Im u = U J. Proof. Consider the diagram with exact columns: 0 0 Z1 0 Z1 Z2 0 Z2 Z3 =o Z1 =o Z1 Z1 0 0 0 · · · · · · · · · Note that the involved inverse systems are Mittag Leffler by Lemma 7, therefore the their inverse limits are exact by [4, Theorem 5]. Therefore the inverse limit gives us the desired short exact sequence. By its construction the homomorphism u acts as follows: for all (x1, x2, x3, . . .) ∈ J (that is (x1, x2, x3, . . .) ∈ Qn≥1 Zn such that δn(xn+1) = xn, for all n) we have u(x1, x2, x3, . . .) = (0, x1, x2, . . .), so U Zn+1 = Zn, for all n ≥ 1 implies U J = Im u. (cid:3) / / / / / / / / o o o o o o o o o o o o o o o o 6 GEORGE CIPRIAN MODOI Lemma 11. The following sentences hold: (a) For all n ≥ 1 we have Zn ⊗R J ∼= Zn as (left) J-modules. (b) For all n ≥ 1 we have HomR(J, Zn) ∼= Zn as R-modules. Proof. Note first that Zn ∼= HomR(Zn, Z∞) is a left J = EndR(Z∞)-module. (a). We proceed by induction on n. For n = 1 we apply the functor Z1 ⊗R − to the short exact sequence 0 → U J → J → Z1 → 0 coming from Lemma 10. Since U is the annihilator of Z1 we deduce Z1 ⊗R U J = 0, so we get an isomorphism Z1 ⊗R J ∼=→ Z1 ⊗R Z1, so Z1 ⊗R J ∼= Z1. Now suppose Zn ⊗R J ∼= Zn. Starting from the short exact sequence 0 → Zn → Zn+1 → Z1 → 0 given by Lemma 8) we construct the commutative diagram with exact rows: Zn ⊗R J Zn+1 ⊗R J Z1 ⊗R J ∼= / Zn 0 / Zn+1 ∼= / Z1 0 / 0 whose vertical maps are obtained from the natural homomorphism − ⊗R J ∼= HomJ (J, −) ⊗R J = ϕ∗ · ϕ∗ → 1Mod-J , the last arrow coming from the adjunction. Then the middle vertical arrow is an isomorphism too, proving the conclusion. (b). Using first the (proof of the) point (a), and second the adjunction isomor- phism we obtain an isomorphism of R-modules HomJ (Zn, Zn) ∼= HomJ ((ϕ∗ · ϕ∗)(Zn), Zn) ∼= HomR(ϕ∗(Zn), ϕ∗(Zn)) = HomR(Zn, Zn) ∼= Zn. Combining it with the adjunction isomorphism between the functors HomJ (Zn, −) : Mod- ⇄ Mod-R : Zn ⊗R − and the isomorphism of part (a) we get the isomorphisms of R-modules: HomR(J, Zn) ∼= HomR(J, HomJ (Zn, Zn)) ∼= HomJ (Zn ⊗R J, Zn) ∼= HomJ (Zn, Zn) ∼= Zn concluding the proof. (cid:3) Theorem 12. With the notations above, if all conditions (1)-(5) are true, then ϕ : R → J is a weak epimorphism of rings. Consequently J is a self -- small R- module. Proof. Using the isomorphism from the point (b) of Lemma 11 we get HomR(J, J) = HomR(J, lim ← Zn) ∼= lim ← HomR(J, Zn) ∼= lim ← Zn ∼= J, therefore the ring homomorphism ϕ : R → J is a weak epimorphism. Then J is self -- small as R-module, by Proposition 1. (cid:3) Example 13. Let R = Z and let p be a prime. The direct system Z/p1 → Z/p2 → Z/p3 → . . . , / / / / / / / / / / CONSTRUCTING LARGE SELF -- SMALL MODULES 7 whose direct limit is the cocyclic abelian group Z/p∞, satisfies the conditions (1)-(5). Thus Theorem 12 gives a proof that the group of p-adic integers Jp = HomZ(Z/p∞, Z/p∞) is self -- small (for details, see also [5]). Example 14. Let R be a Dedekind ring, and let m be a maximal ideal. Put Zi = R/mi, for all i ≥ 1. Then S = Z1 is a simple R-module, and modules Zi are indecomposable, uniserial, with the composition series of the form 0 ⊆ Z1 ⊆ . . . ⊆ Zi−1 ⊆ Zi whose factors are all isomorphic to S. Moreover for every i ≥ 1 there is an exact sequence 0 → S → Zi+1 → Zi → 0. For more details concerning these modules we refer to [6, 1.4]. Then we obtain a direct system (DS) satisfying the conditions (1)-(5), so its inverse limit, the so called m-adic module, J = lim ← i is self -- small as R-module. R/m Acknowledgements. The author would like to thank Simion Breaz and Jan Zemlicka for a (quite old) discussion, in which they made him aware of the im- portance of constructing large self -- small modules, as a counterpart of their work [3], and to the first named of them for some (more recent) valuable comments and suggestions. Last but not least, the author is very indepted to Jan Stov´ıcek for pointing him out a very serious error made in a previous version of this paper. References [1] Albrecht U., Breaz S., Wickless W., Self -- small abelian groups, Bull. Aust. Math. Soc., 80 (2009), 205 -- 216. [2] Arnold, D. M., Murley, C. E., Abelian groups, A, such that Hom(A,-) preserves direct sums of copies of A, Pacific J. Math., 56, (1975), 7 -- 21. [3] Breaz S., Zemlicka J., When every self -- small module is finitely generated, J. Algebra 315 (2007), 885 -- 893. [4] Emmanouil I., Mittag -- Leffler condition and the vanishing of lim 1, Topology, 35(1996), 267 -- ← 271. [5] Fuchs L., Infinite Abelian Groups, vol I, Academic Press, 1970. [6] Ringel C. M., Algebraically Compact Modules Arising from Tubular Families. A Survey, Algebra Colloq., 11 (2004), 155 -- 172. [7] Stenstrom B., Rings of Quotients, Springer Verlag, Berlin, Heidelberg, New York, 1975. [8] Zemlicka J., When products of self -- small modules are self -- small, Comm. Alg. 36 (2008), 2570 -- 2576. Babes¸ -- Bolyai University, Faculty of Mathematics and Computer Science, 1, Mihail Kogalniceanu, 400084 Cluj -- Napoca, Romania E-mail address: [email protected]
1808.01407	1	1808	2018-08-04T01:57:36	Some notes on the ternary modular commutator	[ "math.RA" ]	We define a relation that describes the ternary commutator for congruence modular varieties. Properties of this relation are used to investigate the theory of the higher commutator for congruence modular varieties.	math.RA	math	SOME NOTES ON THE TERNARY MODULAR COMMUTATOR ANDREW MOORHEAD Abstract. We define a relation that describes the ternary commutator for congruence modular varieties. Properties of this relation are used to investigate the theory of the higher commutator for congruence modular varieties. 1. Introduction The topic of these notes is commutator theory. Specifically, we study higher com- mutators which are a higher arity generalization of the classical binary commutator for Mal'cev varieties that was discovered by Smith in [12]. Smith's commutator was extended to modular varieties by Hagemann and Hermann in [6]. The theory of the modular commutator is developed in detail by Freese and McKenzie in [4] and Gumm in [5]. Commutator theory has been investigated for varieties that are not modular, notably by Kearnes and Szendrei in [8] and Kearnes and Kiss in [7]. The definition of the binary commutator was generalized by Bulatov in [2] to a definition of a commutator of higher arity. The basic properties of this higher commutator were developed for congruence permutable varieties by Aichinger and Mudrinski in [1]. The author of [9] showed many of the basic properties true of the higher commutator in permutable varieties hold also for modular varieties. In [13], Wires develops several properties of higher commutators outside of the context of modularity. Our main objective here is to show that the development of the binary commuta- tor for congruence modular varieties is a special case of a more general development of a commutator of arity ≥ 2 for congruence modular varieties, although we present here only the ternary case. Specifically, the binary modular commutator is shown in [4] to be equivalently defined in three ways: (1) as the least congruence satisfying the term condition, (2) as the union of classes related to the diagonal by a special congruence called ∆, or (3) the greatest binary operation on all congruence lattices across a variety satisfying certain conditions. The higher commutator was defined by Bulatov as a higher arity operation obtained by generalizing the term condition. In [11], Oprsal develops the theory of the higher commutator for permutable varieties by finding the right way to define ∆ in the higher arity case. We extend this idea to the theory of the ternary commutator for a modular variety. Date: August 7, 2018. This material is based upon work supported by the National Science Foundation grant no. DMS 1500254. 1 2 ANDREW MOORHEAD The primary value of this treatment of the modular ternary commutator is that it involves a detailed and nontrivial example of a type of relation that we call a 'higher dimensional congruence.' Higher dimensional congruences are used to develop properties of the higher commutator for Taylor varieties in [10]. The structure of the note is as follows: In Sections 2-4 we develop notation, list the properties of modularity that are to be used and briefly review properties of ∆ for the binary case. In Section 5 we introduce the idea of a complex of matrices. In Section 6 these complexes of matrices are used to define a new type of centrality which is shown to be equivalent to the usual term condition centrality in a modular variety. In Section 7 we define ∆ and use it to characterize the ternary commutator. In Section 8 we show that the ternary commutator is related to the binary commutator by the inequality [[θ0, θ1], θ2] ≤ [θ0, θ1, θ2]. In the final section we show that the ternary commutator and 3-dimensional cube terms are related in a manner which is analogous to the relationship between the binary commutator and a Mal'cev operation. 2. Matrices and Centrality In this section we define matrices. Relations that are comprised of such matrices can be used to define the Bulatov or what we call the term condition commuta- tor. Although this note deals with the ternary commutator, the definitions given here are for arbitrary arity. 2.1. Dimension Two. We begin with (2)-dimensional matrices. The binary com- mutator is defined with the so-called term condition. Let A be an algebra and let θ0, θ1, δ ∈ A. We say that θ0 centralizes θ1 modulo δ if for all t ∈ Pol(A), a0 ≡θ0 b0 and a1 ≡θ1 b1, t(a0, a1) ≡δ t(a0, b1) =⇒ t(a0, b1) ≡δ t(b0, b1). This is the standard definition of the term condition, but the term condition can be formulated in terms of matrices, which we shall now define. The set 22 is a natural coordinate system for a square and we say that a relation R ⊆ A22 is (2)-dimensional. We will usually orient our squares like so that when we write (0, 1) (1, 1) , (0, 0) (1, 0) h = c a d b ∈ R for some (2)-dimensional relation R, we mean that h(0,0) = a, h(1,0) = b and so on. Now define the algebra of (θ0, θ1)-matrices to be  x  x y y M (θ0, θ1) := SgA22 : (cid:104)x, y(cid:105) ∈ θ0 3  .  : (cid:104)x, y(cid:105) ∈ θ1 The centrality condition C(θ0, θ1; δ) is equivalent to the assertion where the diagram in parenthesis is a picture representing the implication Similarly, C(θ1, θ0; δ) is equivalent to the assertion (cid:104)a, c(cid:105) ∈ δ =⇒ (cid:104)b, d(cid:105) ∈ δ.  y x  ∪  c δ a  c a y x  , d b  . d b c a ∀ d b ∈ M (θ0, θ1) c a ∀ d b ∈ M (θ0, θ1) Now, set (cid:94) [θ0, θ1] := δ {δ : C(θ0, θ1; δ)}. 2.2. Arbitrary Dimension. For higher arity commutators we prefer at the outset to define term condition centrality and the commutator in terms of matrices. The definition we give is easily seen to be equivalent to the usual one, see [9]. Let A be an algebra and let k ∈ N ≥1. The set of functions 2k is a coordinate system for the (k)-dimensional hypercube, where two functions are connected by an edge if and only if they differ in exactly one coordinate. An A-invariant relation is called a (k)-dimensional invariant relation. Members of such an R are what we call vertex labeled hypercubes, although we usually not be so formal. R ≤ A2k Take an m ∈ A2k . Notation is needed to single out factors of m that are lower dimensional matrices. We specify a partial function g : k → 2 by a pair of tu- ples, D = (d0, . . . , dl−1) and R = (g(d0), . . . , g(dl−1)), where the tuple D lists the elements of the domain of g in order and dom(g) = l. Let the map crsecR D : A2k → A2k\D be the projection onto those factors indexed by functions f ∈ 2k that extend g. Formally, (crsecR D(m))h = mf , where f = g ∪ h. In case D = k − 2 or D = k − 1 we denote the map crsecR D, respectively. When D = 1 the map will be called faceR D. Suppose D = k − 1 and let i ∈ k \ D. When R is not the constant 1 tuple D(m) an (i)-supporting line, and when R is the constant 1 tuple D(m) the (i)-pivot line. Similarly, suppose D = k − 2 and take we call lineR we call lineR D by squareR D or lineR 4 ANDREW MOORHEAD i (cid:54)= j ∈ k \ D. When R is not the constant 1 tuple we call squareR supporting square, and when R is the constant 1 tuple we call squareR (i, j)-supporting square. D(m) an (i, j)- D(m) the A picture of a matrix, pair of squares and a line is given in Figure 1. Notice the orientation of the cube with respect to the coordinates, as this orientation is used throughout this note. i (cubek i (a, b) ∈ A2k The term condition definition of higher centrality can now be formulated in terms i (a, b)) and i (a, b)) are constant matrices with value a and b, respectively. Now let of matrices. Let cubek be the matrix such that face0 face1 T = (θ0, . . . , θk−1) ∈ Con(A)k be a tuple of congruences and set i (cubek M (θ0, . . . , θk−1) = SgA2k ({cubek i (a, b) : i ∈ k and (cid:104)a, b(cid:105) ∈ θi}) We call M (θ0, . . . , θk−1) the algebra of (θ0, . . . , θk−1)-matrices. If no confusion can result, we will refer to M (θ0, . . . , θk−1) as M (T ), and call its elements T - matrices. Now we can define higher centrality, see [9] for the same definition. Definition 2.1. Let A be an algebra, k ≥ 2, T = (θ0, . . . , θk−1) ∈ Con(A)k, and δ ∈ Con(A). For j ∈ k, we say that T = (θ0, . . . , θk−1) is centralized at j modulo δ if the following property holds for all (θ0, . . . , θk−1)-matrices m: () If every (j)-supporting line of m is a δ-pair, then the (j)-pivot line of m is a δ-pair. We write C(T ; j; δ) to indicate that T is centralized at j modulo δ. Now, set (cid:94) {δ : C(T ; j; δ)}. [T ]j := 3. Day Terms The following classical results on congruence modular varieties are needed. For proofs see [3], [5] or [4]. Proposition 3.1 (Day Terms). A variety V is congruence modular if and only if there exist term operations me(x, y, z, u) for e ∈ n + 1 satisfying the following identities: (1) me(x, y, y, x) ≈ x for each 0 ≤ e ≤ n, (2) m0(x, y, z, u) ≈ x, (3) mn(x, y, z, u) ≈ u, (4) me(x, x, u, u) ≈ me+1(x, x, u, u) for even e, and (5) me(x, y, y, u) ≈ me+1(x, y, y, u) for odd e Proposition 3.2 (Lemma 2.3 of [4]). Let V be a variety with Day terms me for e ∈ n + 1. Take δ ∈ Con(A) and assume (cid:104)b, d(cid:105) ∈ δ. For a tuple (cid:104)a, c(cid:105) ∈ A2 the following are equivalent: (1) (cid:104)a, c(cid:105) ∈ δ (2) (cid:104)me(a, a, c, c), me(a, b, d, c)(cid:105) ∈ δ for all e ∈ n + 1 5 Figure 1 a=m(0,0,0)b=m(1,0,0)c=m(0,1,0)d=m(1,1,0)e=m(0,0,1)f=m(1,0,1)g=m(0,1,1)h=m(1,1,1)abcdefghface02face12line(1,1)(0,2),the(1)-pivotline6 ANDREW MOORHEAD 4. Binary Commutator and the ∆ relation The binary commutator of two congruences θ0, θ1 of A can be developed by examining a special congruence called ∆θ0,θ1 , which is a congruence of θ0, considered here to be a subalgebra of A2. The definition of ∆θ0,θ1 is as follows: ∆θ0,θ1 = Cgθ0({(cid:104)(cid:104)x, x(cid:105),(cid:104)y, y(cid:105)(cid:105) : (cid:104)x, y(cid:105) ∈ θ1}) It is informative to use the matrices and coordinates introduced earlier. With this perspective in mind, θ0-pairs become 1-dimensional matrices, which according to our convention are horizontal lines. The pair (cid:104)a, b(cid:105) is ∆θ0,θ1 related to (cid:104)c, d(cid:105) if there is a sequence of matrices as in Figure 2. That is, ∆θ0,θ1 is the transitive closure of M (θ0, θ1), where these matrices are considered to be pairs of θ0-pairs. Of course, one can also define ∆θ1,θ0 = Cgθ1({(cid:104)(cid:104)x, x(cid:105),(cid:104)y, y(cid:105)(cid:105) : (cid:104)x, y(cid:105) ∈ θ0}) now considered to be θ1-pairs. This congruence is the transitive closure of M (θ0, θ1), where these matrices are Both ∆θ0,θ1 and ∆θ1,θ0 are subalgebras of A22 , see Figure 2. In fact, an aspect Figure 2 of the symmetry of the binary commutator is given in the following theorem. Theorem 4.1 (See Chapter 4 of [4]). ∆θ0,θ1 = ∆θ1,θ0 The ∆ relation provides a nice characterization of the binary commutator of θ0, θ1. Theorem 4.2 (Theorem 4.9 of [4]). For x, y ∈ A the following are equivalent: (1) (cid:104)x, y(cid:105) ∈ [x, y] (2) (cid:20) x y (cid:20) x b y y (cid:21) (cid:21) ∈ ∆θ0,θ1 ∈ ∆θ0,θ1 for some b ∈ A (3) y b abcdθ0θ1(θ0,θ1)-matrixabcdabcd∈∆θ0,θ1,∆θ1,θ0≤A22(cid:21) (cid:20) x y c c (4) ∈ ∆θ0,θ1 for some c ∈ A. 7 The first portion of this note is devoted to extending Theorem 4.2 to the ternary commutator. 5. Complexes of Matrices Let A be an algebra and take (θ0, θ1, θ2) ∈ Con(A)3. We begin this section defining what we call a (θ0, θ1, θ2)-matrix complex. Informally, these are blocks that are built from (θ0, θ1, θ2)-matrices. Formally, take some integers n0, n1, n2 ≥ 2, and set Now take Cn0,n1,n2 = An0×n1×n2 (cid:91) C = Cn0,n1,n2 n0,n1,n2≥2 We call C the set of 3-dimensional complexes of the algebra A, and Cn0,n1,n2 the set of complexes with dimensions (n0, n1, n2). For f = (f0, f1, f2) ∈ n0 × n1 × n2 and g = (g0, g1, g2) ∈ 23 = 2 × 2 × 2 let f + g = (f0 + g0, f1 + g1, f2 + g2). Let f = (f0, f1, f2) ∈ n0 × n1 × n2 be such that for each i ∈ 3 we have fi < ni − 1. Now define Matf : Cn0,n1,n2 → A23 to be the map defined by (Matf (a))g = af +g. The image of such a map is called a component matrix. Now define a (θ0, θ1, θ2)-matrix complex with dimensions (n0, n1, n2) to be a complex a ∈ Cn0,n1,n2 with the property that Matf (a) ∈ M (θ0, θ1, θ2) for each map Matf . The collection of all (θ0, θ1, θ2)-complexes with dimensions (n0, n1, n2) is called Cn0,n1,n2 (θ0, θ1, θ2) and the collection of all (θ0, θ1, θ2)-complexes is called C(θ0, θ1, θ2). We can take cross-sections of complexes in the same way as we do for matrices. Consider c ∈ Cn0,n1,n2. Let D be a tuple that lists some coordinates in order and R be a tuple that specifies a value that is assigned to each corresponding coordinate in D. Set crsecR D(c) to be the lower dimensional complex obtained by projecting c onto those factors indexed by functions that extend the partial function specified by D and R. We use the names line and square in the same way as for complexes. A typical complex, one of its component matrices and a cross-section square are shown in Figure 3. We now define a map that allows us to identify the corners of a complex. For integers n0, n1, n2 ≥ 2 and a function g ∈ 23, let γ(g) ∈ n0 × n1 × n2 be the tuple that is 0 or ni − 1 in the i-th coordinate if g(i) is 0 or 1, respectively. Then set to be the map defined by Cornern0,n1,n2 : C → A23 Cornern0,n1,n2 (a)g = aγ(g) 8 ANDREW MOORHEAD Figure 3 We will often refer to this map as Corner when no confusion over the dimensions of a complex is possible. Finally, we define Cornersn0,n1,n2(θ0, θ1, θ2) = {Corner(a) : a ∈ Cn1,n2,n3(θ0, θ1, θ2)} and we denote the set of all (θ0, θ1, θ2)-corners by Corners(θ0, θ1, θ2). Notice that because M (θ0, θ1, θ2) = C2,2,2(θ0, θ1, θ2) = Corners2,2,2(θ0, θ1, θ2), we have that M (θ0, θ1, θ2) ⊆ Corners(θ0, θ1, θ2). 6. Transitive Term Condition We now define a stronger term condition which we call the transitive term condition. Definition 6.1. We say that (θ0, θ1, θ2) is transitively centralized at j modulo δ if the following property holds for all (θ0, θ1, θ2)-corners c: () If every (j)-supporting line of c is a δ-pair, then the (j)-pivot line of c is a δ-pair. We abbreviate this property Ctr(θ0, θ1, θ2; j; δ). Because every element of M (θ0, θ1, θ2) is a corner, we have that Ctr(θ0, θ1, θ2; j; δ) implies C(θ0, θ1, θ2; j; δ). One of the results of [9] is that the ordinary term condition is symmetric, so we may unambiguously write C(θ0, θ1, θ2; δ). As it turns out, the transitive term condition is equivalent to the usual term condition in a modular variety. a0,0,0a0,n1−1,0an0−1,0,0an0−1,n1−1,0a0,0,n2−1a0,n1−1,n2−1an0,0,n2−1an0−1,n1−1,n2−1af0,f1,f2af0+1,f1+1,f2+1θ0θ1θ2a∈Cn0,n1,n2(θ0,θ1,θ2)Matf(a)∈M(θ0,θ1,θ2)a0,0,n2−1a0,n1−1,n2−1an0,0,n2−1an0−1,n1−1,n2−1squaren2−12(a)Lemma 6.2. Let V be a modular variety with Day terms m0, . . . , mn. Take A ∈ V and (θ0, θ1, θ2) ∈ Con(A)3. Choose some coordinates i (cid:54)= j (cid:54)= l ∈ 3. Take z ∈ Cn0,n1,n2 (θ0, θ1, θ2) such that 9 Corners(square0 i ) = Corners(squaremi−1 i ) = (cid:21) (cid:20) c0 (cid:20) cmi−1 a0 d0 b0 ami−1 (cid:21) , and dmi−1 bmi−1 where the matrices above are oriented so that rows are indexed by j and the columns are indexed by l, with the origin at the bottom left. For each Day term me with e ∈ n + 1 there is an Re dimensions mi ≥ ni, mj = max(nj − 1, 2) and ml = nl, such that j,l(z) ∈ C(θ0, θ1, θ2) with (cid:20) me(c0, c0, a0, a0) me(c0, d0, b0, a0) (cid:20) me(cni−1, cni−1, ani−1, ani−1) me(cni−1, dni−1, bni−1, ani−1) (cid:21) Corners(square0 Corners(squaremi−1 i ) = i ) = and c0 (cid:21) c0 cni−1 cni−1 We call Re j,l(z) the eth shift rotation at (j, l) of z. Proof. Without loss, let i = 2, l = 0 and j = 1. Take z ∈ Cn0,n1,n2(θ0, θ1, θ2). Now, z may be decomposed into the sequence of cross-sections squarek 2(z) for k ∈ n2 as shown in Figure 4. Figure 4 For each cross-section squarek 2(z), let Ak, Bk, Ck and Dk be the following (θ0, θ1)- complexes (see Figure 5): (1) Ak takes the value z0,n1−1,k in each factor. (2) Bk is equal to squarek (3) Ck is such that 2(z). 210squarek2cn2−1=z0,n1,n2−1c0=z0,n1−1,0d0=zn0−1,n1−1,0a0=z0,0,0z0,0,kz0,n1−1,kzn0−1,n1−1,kzn0−1,0,kan2−1=z0,0,n2−1bn2−1=zn0,0,n2−1dn2−1=zn0,n1,n2−1b0=zn0−1,0,010 ANDREW MOORHEAD (4) Dk is such that • line0 • linek • linek • linen1−1 1 1(squarek (Ck) = line0 1(Ck) = linen1−1 1 1(Ck) = line1 1(squarek 2(z)) and 2(z)) for all other k ∈ n1. 1(Dk) is constantly z0,n1−1,k for k ∈ n1 − 1 and (Dk) is constantly z0,0,k. Figure 5 Let Ek = me(Ak, Bk, Ck, Dk) and let Fk ∈ Cn0,n1−1(θ0, θ1) be the complex such that (Fk)g = (Ek)g+(0,1) for all g ∈ n0× (n1− 1). So, Fk is obtained by deleting the bottom row of Ek. Applying the identity me(x, y, y, x) ≈ x, we get that Corners(Fk) is the (θ0, θ1)-matrix shown in 6. Figure 6 We will be done if we can show that there is a complex y ∈ Cn0,n1,m2 , for some = En2−1. We will demonstrate m2 ≥ n2, such that square0 this by showing that there are α, β, γ, ε ∈ Cn0,n1,m2 such that 2 = E0 and squarem2−1 2 (1) square0 (2) square0 (3) square0 2 α = A0 and squarem2−1 2 β = B0 and squarem2−1 2 γ = C0 and squarem2−1 2 2 2 α = Am2−1 β = Bm2−1 γ = Cm2−1 z0,n1−1,kz0,n1−1,kz0,n1−1,kz0,n1−1,kz0,n1−1,kz0,n1−1,kz0,n1−1,kz0,n1−1,kz0,n1−1,kAkz0,n1−1,kz0,n1−2,kz0,0,kz1,n1−1,kzn0−1,n1−1,kzn0−1,n1−2,kz1,n1−2,kzn0−1,0,kz1,0,kBkz0,1,kzn0−1,1,kz1,1,kz0,n1−1,kz0,n1−1,kz0,n1−1,kz0,0,kzn0−1,0,kz1,0,kCkz0,1,kzn0−1,1,kz1,1,kz0,0,kzn0−1,0,kz1,0,kz0,1,kzn0−1,1,kz1,1,kz0,n1−1,kz0,n1−1,kz0,n1−1,kz0,n1−1,kz0,n1−1,kz0,n1−1,kDkz0,n1−1,kz0,n1−1,kz0,n1−1,kz0,0,kz0,0,kz0,0,kme(z0,n1−1,k,z0,n1−1,k,z0,0,k,z0,0,k)z0,n1−1,kme(z0,n1−1,k,zn0−1,n1−1,k,zn0−1,0,k,z0,0,k)z0,n1−1,kε = Dm2−1 11 (4) square0 2 ε = D0 and squarem2−1 2 We then set y = me(α, β, γ, δ). So, let α(cid:48), β(cid:48), γ(cid:48) ∈ Cn0,n1,n2 be such that for each k ∈ n2 (1) squarek (2) squarek (3) squarek 2 α(cid:48) = Ak 2 β(cid:48) = Bk 2 γ(cid:48) = Ck. (cid:20) z0,0,k z0,n1−1,k z0,0,k+1 z0,n1−1,k+1 It is easy to see that α, β, γ ∈ Cn0,n1,n2 . This does not work for the Dk, because (cid:21) need not be a (θ2, θ1)-matrix. However, it is an element of ∆θ2,θ1, and there- fore also an element of ∆θ1,θ2 by Theorem 4.1. So, there exists for a tk ∈ N ≥2 some θ1-pairs (cid:104)uk tk−1, vk 0(cid:105) and (cid:104)z0,0,k+1, z0,n1−1,k+1(cid:105) = (cid:104)uk tk−1, vk tk−1(cid:105), with consecutive pairs forming matrices be- longing to M (θ2, θ1). See Figure 7. tk−1(cid:105) such that (cid:104)z0,0,k, z0,n1−1,k(cid:105) = (cid:104)uk 0(cid:105), . . . ,(cid:104)uk 0, vk 0, vk Figure 7 m2 =(cid:80) For each s ∈ tk set U k s and all other rows equal to vk k∈n2−1 tk and let ε ∈ Cn0,n1,m2 be such that squarer s ∈ Cn0,n1 (θ0, θ1) to be a complex with top row equal 1 . Let U be the collection of all such U k to uk s . We order elements of U by the lexicographical order on pairs (k, s). Now, set 2(γ) is the rth element of U, see Figure 8. It is clear that ε is actually an element of Cn0,n1−1,m2 (θ0, θ1, θ2). Now, α(cid:48), β, γ(cid:48) are (θ0, θ1, θ2)-complexes with dimensions (n0, n1, n2). We extend these complexes by m2 − n2 many copies of An2−1, Bn2−1 and Cn2−1, respectively. That is, define α, β, γ ∈ Cn0,n1,n2(θ0, θ1, θ2) so that 2(α(cid:48)) (cid:40) squarek 2(α) = squarek An2−1 if k ∈ n2 otherwise θ2θ1z0,0,kz0,0,k+1z0,n1−1,kz0,n1−1,k+1z0,1,kz0,1,k+1z0,2,kz0,2,k+1z0,n1−3,kz0,n1−3,k+1z0,n1−2,kz0,n1−2,k+1z0,0,kz0,n1−1,kz0,0,k+1z0,n1−1,k+1uk1vk1uk2vk2uktk−2vktk−212 ANDREW MOORHEAD Figure 8 (cid:40) (cid:40) 2(β(cid:48)) squarek Bn2−1 2(γ(cid:48)) squarek Cn2−1 if k ∈ n2 otherwise if k ∈ n2 otherwise squarek 2(β) = squarek 2(γ) = Now, set y = me(α, β, γ, ε) ∈ Cn0,n1,m2(θ0, θ1, θ2). Notice that square0 2(y) = E0 1,0(z) ∈ Cn0,n1−1,m2(θ0, θ1, θ2) to be the complex (y) = F0. Set Re and squarem2−1 defined by 2 Re 1,0(z)f = yf +(0,1,0) for f ∈ n0 × (n1 − 1) × m2 So, x is the complex that is obtained by deleting the bottom of face of y. We have that square0 1,0(z)) = Fn2−1, so we are done. (cid:3) 1,0(z)) = F0 and squarem2−1 2(Re (Re 2 We can say more in case the dimensions of a complex are minimal in two coor- dinates. Lemma 6.3. Let V be a modular variety with Day terms m0, . . . , mn. Take A ∈ V and (θ0, θ1, θ2) ∈ Con(A)3. Fix an e ∈ n + 1 and choose some coordinates i (cid:54)= j (cid:54)= l ∈ 3. Take z ∈ Cn0,n1,n2 (θ0, θ1, θ2) such that mj = ml = 2. Suppose that for each k ∈ ni we have (cid:20) ck ak (cid:21) dk bk squarek i (z) = Then Re j,l(z) ∈ Cn0,n1,n2 (θ0, θ1, θ2), i.e. the dimensions do not change. Moreover, (cid:20) me(ck, ck, ak, ak) me(ck, dk, bk, ak) (cid:21) ck ck squarek i (Re j,l(z)) = for each k ∈ ni. 210D0U0sD1Un2−2sDn2−1∈Cn0,n1,m2Proof. Without loss, assume i = 2, j = 1 and l = 0. The proof is the same as that of Lemma 6.2, except that now z0,0,k z0,n1−1,k z0,0,k+1 z0,n1−1,k+1 ak + 1 ck+1 ∈ M (θ2, θ1) for each k ∈ n2 − 2. So we choose α, β, γ, ε ∈ Cn0,n1,n2 so that (cid:20) ak ck = (cid:21) 13 (cid:20) (1) squarek 2 α(cid:48) = Ak = (2) squarek 2 β(cid:48) = Bk = (3) squarek 2 γ(cid:48) = Ck = (4) squarek 2 ε(cid:48) = Dk = (cid:21) (cid:21) (cid:21) (cid:21) (cid:21) . ck ck dk bk bk bk ak ck ck (cid:20) ck (cid:20) ck (cid:20) ak (cid:20) ak ak ak ck for each k ∈ n2 − 1. me(α, β, γ, ε). Now we see that It is easy to see that α, β, γ, ε ∈ Cn0,n1,n2(θ0, θ1, θ2). As before, set Re j,l(z) = (cid:20) me(ck, ck, ak, ak) me(ck, dk, bk, ak) (cid:21) squarek 2(Re j,l(z)) = as desired. ck ck , (cid:3) Shift rotations have the nice property of preserving supporting lines. Specifically, we mean that Lemma 6.4. Let V be a modular variety with Day terms m0, . . . , mn. Take A ∈ V and (θ0, θ1, θ2) ∈ Con(A)3. Suppose z ∈ C(θ0, θ1, θ2) is such that every (j)- supporting line is a δ-pair, for some j ∈ 3 and δ ∈ Con(A). Then, for each e ∈ n + 1, every (l)-supporting line of Re Proof. This follows from Lemmas 6.2 and 3.2. j,l is a δ-pair. (cid:3) Theorem 6.5. C(θ0, θ1, θ2; δ) implies Ctr(θ0, θ1, θ2; j; δ) for each coordinate j ∈ 3. Proof. Take z ∈ Cn0,n1,n2(θ0, θ1, θ2) such that each (j)-supporting line of z is a δ-pair. We assume that C(θ0, θ1, θ2; δ) holds and will use this assumption to show that the (j)-pivot line of z is also a δ-pair with an induction on the dimensions nj and nl for l (cid:54)= j. The base case is when both of these dimensions are equal to 2. Without loss, assume that j = 1 and l = 0. Then, for each e ∈ n + 1 we have by Lemma 6.4 1,0(z) is a δ-pair. By Lemma 6.3, Re that every (0)-supporting line of Re 1,0(z) ∈ 2(Re C2,2,n2 (θ0, θ1, θ2), that is, squarek 1,0(z)) ∈ M (θ0, θ1) for each k ∈ n2. Moreover, because the (0)-supporting line of each squarek 1,0(z)) is a δ-pair we can use the assumption that C(θ0, θ1, θ2; δ) holds to conclude that the (0)-pivot line of Re 1,0 is a δ-pair. Because this holds for every e ∈ n + 1, we apply Lemma 3.2 to conclude that the (1)-pivot line of z is a δ-pair. This is shown in Figure 9, where δ-pairs are drawn as curved lines and constant pairs are drawn in bold. The result now follows from an induction on Lemma 6.2. 2(Re (cid:3) 14 ANDREW MOORHEAD Figure 9 7. Definition and Properties of ∆ We define ∆(θ0, θ1, θ2) in a manner analogous to the 2-dimensional ∆. We will prove that this definition is equivalent to a transitive closure of a relation that is obtained by looking at the corners of complexes. This perspective is useful for showing properties of ∆(θ0, θ1, θ2) because of the equivalence of the transitive term condition with the term condition. Recall that ∆θi = θi, where we think of θi as being the collection of 1-dimensional θi-matrices. As before, ∆θi,θj = Cg∆θi ({cube2 j (x, y) : (cid:104)x, y(cid:105) ∈ θj}). Let k ∈ 3 such that k (cid:54)= i, j. The algebra (∆θi,θj )2 is the second power of a square, so it is naturally coordinatized by 23. That is, to each (cid:104)a, b(cid:105) ∈ (∆θi,θj )2 we associate a matrix z ∈ A23 k = b. We now set k = a and face1 such that face0 Without loss, we will show that ∆θ0,θ2,θ1 = ∆θ0,θ1,θ2. The result then follows ∆θi,θj ,θk = Cg∆θi,θj ({cube3 from the symmetry of the 2-dimensional ∆. We begin with the following k(x, y) : (cid:104)x, y(cid:105) ∈ θk}) ⊂ A23 Lemma 7.1. Suppose that z0, z1 ∈ ∆θ1,θ2,θ0 are such that square0 2(z0) = , square1 2(z0) = square0 2(z1) = (cid:20) d1 (cid:20) b1 (cid:21) d0 b0 d1 d0 (cid:21) c1 c0 (cid:21) (cid:20) d1 . d0 (cid:21) . c1 c0 square0 2(z1) = (cid:20) d1 d0 Then b1 b0 ∆θ1,θ2,θ0 (cid:21) (cid:20) a1 a1 a0 a0 and Proof. We provide in Figure 10 a picture of the two matrices z0, z1 ∈ ∆θ1,θ2,θ0 . 210z∈C2,2,n2Re1,0(z)∀e∈n+1followsfromC(θ0,θ1,θ2;δ)15 Figure 10 (cid:21) (cid:21) (cid:21) (cid:21) b1 b0 b1 b0 c1 c0 c1 c0 Set (1) a = (2) b = (3) c = (4) d = d0 (cid:20) d1 (cid:20) a1 (cid:20) d1 (cid:20) a1 a0 d0 a0 Now, we want to show that (cid:104)a, b(cid:105) ∈ ∆θ1,θ2,θ0. By Lemma 3.2 it suffices that (cid:104)me(a, a, c, c), me(a, b, c, d)(cid:105) ∈ ∆θ1,θ2,θ0 for all e ∈ n + 1. This amounts to checking that (cid:20) me(d1, d1, d1, d1) me(b1, b1, c1, c1) (cid:21) me(d0, d0, d0, d0) me(b0, b0, c0, c0) ∆θ1,θ2,θ0 (cid:20) me(d1, a1, a1, d1) me(b1, b1, c1, c1) me(d0, a0, a0, d0) me(b0, b0, c0, c0) (cid:21) , Proof. We provide in Figure 11 a picture of the two matrices z0, z1 ∈ ∆θ1,θ2,θ0 . which follows easily from the identity me(x, y, y, x) ≈ x. Lemma 7.2. Suppose that z0, z1 ∈ ∆θ1,θ2,θ0 are such that square0 2(z0) = , square1 2(z0) = square0 2(z1) = (cid:3) (cid:20) u1 u0 (cid:21) , v1 v0 (cid:20) x1 (cid:20) r1 (cid:21) x0 r0 s1 s0 y1 y0 (cid:21) (cid:21) (cid:20) y1 . y0 (cid:21) s1 s0 Then ∆θ1,θ2,θ0 and square0 2(z1) = r1 r0 x0 (cid:20) x1 (cid:20) x1 (cid:20) u1 (cid:20) y1 u0 x0 y0 Set (1) a = (2) b = (3) c = (cid:21) (cid:21) (cid:21) r1 r0 r1 r0 s1 s0 012a1a1a0a0b1c1b0c0d1d1d0d0a1a1a0a0z0z116 ANDREW MOORHEAD (4) d = (1) (2) s1 s0 (cid:21) (cid:21) v0 (cid:20) v1 (cid:20) v1 (cid:20) v1 (cid:18)(cid:20) u1 (cid:18)(cid:20) u1 s1 s0 v1 v0 u0 v0 v0 Figure 11 (cid:21) u0 r1 r0 (cid:20) u1 (cid:20) u1 u1 (cid:20) u1 (cid:21) (cid:20) u1 u1 (cid:21) u0 u0 r1 r0 u0 (cid:21) (cid:21) (cid:20) v1 (cid:20) v1 v0 , , ∆θ1,θ2,θ0 ∆θ1,θ2,θ0 (cid:21) (cid:21) r1 r0 , We want to show that (cid:104)a, b(cid:105) ∈ ∆θ1,θ2,θ0 . By Lemma 3.2 it suffices that (cid:104)me(a, a, c, c), me(a, b, c, d)(cid:105) ∈ ∆θ1,θ2,θ0 for all e ∈ n + 1. An examination of z1 reveals that that Now, because for all e ∈ n + 1 the identity me(x, y, y, x) ≈ x holds, it follows r1 r0 ∆θ1,θ2,θ0 (1) me s1 s0 s1 s0 r1 r0 , , (2) me u0 Therefore, the matrix shown in Figure 12 belongs to ∆θ1,θ2,θ0 for all e ∈ n + 1. u0 u0 u0 v0 v0 v1 v0 , s1 s0 ∆θ1,θ2,θ0 r1 r0 (cid:21) (cid:21) (cid:20) v1 (cid:20) v1 v0 (cid:21)(cid:19) (cid:21)(cid:19) (cid:20) u1 (cid:20) u1 u0 (cid:21) (cid:21) Figure 12 An examination of z0 reveals that (cid:21) (cid:21) (cid:20) x1 u1 (cid:20) y1 x0 u0 v1 v0 y0 (1) (2) ∆θ1,θ2,θ0 ∆θ1,θ2,θ0 (cid:21) (cid:21) (cid:20) x1 u1 (cid:20) y1 x0 u0 v1 v0 y0 012x1y1x0y0u1v1u0v0u1v1u0v0r1s1r0s0z0z1012me(r1,r1,s1,s1)me(r0,r0,s0,s0)me(r1,u1,v1,s1)me(r0,u0,v0,s0)me(u1,u1,v1,v1)me(u0,u0,v0,v0)me(u1,u1,v1,v1)me(u0,u0,v0,v0)We therefore obtain that the matrix depicted in Figure 13 belongs to ∆θ1,θ2,θ0. An application of Lemma 7.1 finishes the proof. 17 Figure 13 (cid:3) We now provide another characterization of ∆θ0,θ1,θ2. For i ∈ 3 set Ri = {(cid:104)square0 i (Corner(z)), square1 i (Corner(z))(cid:105) : z ∈ C2,2,n2(θ0, θ1, θ2) for some n2 ≥ 2} We will often abuse notation and refer to elements of Ri as elements of A23 , . A picture of a complex that corresponds to rather than pairs of elements of A22 the pair (cid:28)(cid:20) c g a e (cid:21)(cid:29) (cid:21) , (cid:20) d h b f ∈ R0 is given in Figure 14. Figure 14 Lemma 7.3. ∆θi,θj ,θl is the transitive closure of Rl. Proof. Without loss, we show that ∆θ1,θ2,θ0 is the transitive closure of R0. Notice that R0 ∈ (∆θ1,θ2)2. It is easy to see that R0 is reflexive, symmetric and compatible. Therefore, (R0)T C ∈ Con(∆θ1,θ2). Notice that M (θ0, θ1, θ2) ⊆ ∆θ1,θ2,θ0, because each generator of M (θ0, θ1, θ2) belongs to ∆θ1,θ2,θ0 . Therefore, R0 ⊆ ∆θ1,θ2,θ0 by Lemma 7.2, so (R0)T C = ∆θ1,θ2,θ0. (cid:3) 012me(x1,x1,y1,y1)me(x0,x0,y0,y0)me(x1,x1,y1,y1)me(x0,x0,y0,y0)me(u1,u1,v1,v1)me(u0,u0,v0,v0)me(u1,u1,v1,v1)me(u0,u0,v0,v0)012abcdefgh18 ANDREW MOORHEAD A similar argument shows that R2 ∈ ∆θ1,θ2,θ0 . An application of Lemma 7.2 then gives that ∆θ0,θ1,θ2 ⊆ ∆θ1,θ2,θ0 . The argument is clearly symmetric, so we have proved the following Theorem 7.4. ∆θi,θj ,θk = ∆θπ(i),θπ(j),θπ(k) for any permutation of coordinates π. We also have that Theorem 7.5. ∆θi,θj ,θl = ∆∆θi,θj ,∆θi,θl Proof. Without loss, we show that ∆θ1,θ2,θ0 = ∆∆θ1 ,θ2 ,∆θ1,θ0 . Both of these relations can be considered as congruences on ∆θ1,θ2. Each one (cid:3) contains the generators of the other, so they are equal. We now use the ∆ relation to characterize the ternary commutator. Theorem 7.6. The following are equivalent: Proof. That (2) implies (3), (4), (5) is trivial. Suppose that (2) is true for x, y ∈ A. We show that (1) holds. Because ∆θ0,θ1,θ2 = (R2)T C, there is a sequence of complexes z0, . . . , zs−1 ∈ Cn0≥2,2,2(θ0, θ1, θ2) such that This situation is depicted in Figure 15, where [θ0, θ1, θ2]-pairs are connected by a solid curved line. We wish to show that the pair connected by a dashed curved line is also a [θ0, θ1, θ2]-pair. By Lemma 3.2, it suffices to see that (cid:104)me(x, x, x, x), me(x, y, x, x)(cid:105) ∈ [θ0, θ1, θ2] for each e ∈ n + 1. Now, for e ∈ n + 1 we have for each zk the eth shift rotation at (0, 1). By Lemma 6.2, these complexes have the following faces: (cid:21) (cid:20) x x (cid:20) me(x, x, x, x) me(x, y, x, x) x x (cid:21) (1) square0 2(Re 0,1(z0)) = (2) square0 2(Re 0,1(zs−1)) = x (3) square1 2(Re 0,1(zk)) = square0 2(zk+1) = for each k ∈ s − 1. (cid:20) me(ck, ck, ak, ak) me(ck, dk, bk, ak) x (cid:21) ck ck (1) (cid:104)x, y(cid:105) ∈ [θ0, θ1, θ2] (2) x x (cid:21) (cid:28)(cid:20) x x (cid:28)(cid:20) a0 a1 (cid:28)(cid:20) b1 (cid:28)(cid:20) c1 a0 a1 b1 b0 b0 , x x (cid:20) x y (cid:21) (cid:20) a2 (cid:21) (cid:20) b2 (cid:20) c1 (cid:21) y , , (3) (4) (5) , c0 y c2 c0 (cid:21)(cid:29) (cid:21)(cid:29) (cid:21)(cid:29) (cid:21)(cid:29) b2 x x c2 y a2 x (cid:21) (cid:20) x x (cid:20) x y x x x x (cid:21) (1) square0 2(z0) = (2) square0 2(zs−1) = (3) square1 2(zk) = square0 2(zk+1) = ∈ ∆θ0,θ1,θ2 ∈ ∆θ0,θ1,θ2 , for some a0, a1, a2 ∈ A ∈ ∆θ0,θ1,θ2 for some b0, b1, b2 ∈ A ∈ ∆θ0,θ1,θ2 (cid:20) ck ak (cid:21) dk bk for each k ∈ s − 1. 19 Figure 15 Now, C(θ0, θ1, θ2; [θ0, θ1, θ2]) holds, which implies that Ctr(θ0, θ1, θ2; [θ0, θ1, θ2]) holds. So, (cid:104)me(x, x, x, x), me(x, y, x, x)(cid:105) ∈ [θ0, θ1, θ2] by induction. This is shown in Figure, where constant pairs are drawn in bold. Figure 16 A similar argument shows that (3), (4), (5) all imply (1). To finish we show that (1) implies (2). For this it suffices to show that the set of pairs (cid:104)x, y(cid:105) ∈ A2 such that (cid:28)(cid:20) x x (cid:21) x x (cid:21)(cid:29) (cid:20) x y x x , is a congruence of A. Reflexivity and compatibility are obvious. The proof of symmetry is shown in Figure 17. The proof of transitivity is similar to that of symmetry. (cid:3) ∈ ∆θ0,θ1,θ2 xxxxxxxy012c0a0d0b0cs−2as−2ds−2bs−2xx012cs−2c0c0cs−2xxxxme(c0,c0,a0,a0)me(c0,d0,b0,a0)me(cs−2,cs−2,as−2,as−2)me(cs−2,ds−2,bs−2,as−2)me(x,x,x,x)me(x,y,x,x)20 ANDREW MOORHEAD Figure 17 8. Greatest Ternary Congruence Lattice Operation A beautiful result in the theory of the binary commutator is the following from [4]. Theorem 8.1. Let V be a modular variety, and suppose that for each A ∈ V there is a binary operation C : Con(A)2 → Con(A) such that, for all A ∈ V and θ0, θ1 ∈ Con(A), (1) C(θ0, θ1) ≤ θ0 ∧ θ1 (2) If f is a surjective homomorphism with kernel π, then C(θ0, θ1) ∨ π = f−1C(f (θ0 ∨ π, f (θ0 ∨ π)). Then C(θ0, θ1) ≤ [θ0, θ1] for all A ∈ V and θ0, θ1 ∈ Con(A). Interestingly, an analogous property holds for the ternary commutator. 012xxxxxyxx∈∆θ0,θ1,θ2xxxy∈∆θ1,θ2yyyx∈∆θ1,θ2∈∆θ0,θ1,θ2yyyyyxyxyx∈θ2xxxxyyyy∈∆θ0,θ1,θ2xxxxxyxx∈∆θ0,θ1,θ2(1)(2)(3)yyyyxyxx∈∆θ0,θ1,θ2yyyyyyyx∈∆θ0,θ1,θ2(4)(1)+(2)(5)(4)+(3)Proposition 8.2. Let V be a congruence modular variety. Suppose that C : Con(A)3 → Con(A) is defined for each A ∈ V such that (1) C(θ0, θ1, θ2) ≤ θ0 ∧ θ1 ∧ θ2 (2) C(θ0, θ1, θ2)∨ π = f−1(C(f (θ0 ∨ π), f (θ1 ∨ π), f (θ2 ∨ π))) for f a surjective homomorphism with kernel π. 21 Then C(θ0, θ1, θ2) ≤ [θ0, θ1, θ2]. Proof. Take θ0, θ1, θ2 ∈ Con(A) for some A ∈ V. Consider the following congruences of ∆θ0,θ1: ∆ = ∆θ0,θ1,θ2 (cid:26)(cid:28)(cid:20) u0 (cid:26)(cid:28)(cid:20) u0 (cid:26)(cid:28)(cid:20) u0 x v0 w0 x v0 w0 x v0 w0 [θ0, θ1, θ2]0 = η0 = η0 = (cid:21) (cid:21) (cid:21) , , , (cid:20) u1 (cid:20) u0 (cid:20) u1 y v1 w1 y v0 w1 y v0 w0 (cid:21)(cid:29) (cid:21)(cid:29) (cid:21)(cid:29) : (cid:104)x, y(cid:105) ∈ [θ0, θ1, θ2] (cid:27) (cid:27) : (cid:104)x, y(cid:105) ∈ θ0 : (cid:104)x, y(cid:105) ∈ θ1 (cid:27) Let f : ∆θ0,θ1 → A be the projection of ∆θ0,θ1 onto the top right vertex and let π be the kernel of f . One can check that f ([θ0, θ1, θ2]0) = [θ0, θ1, θ2], f (∆∨ π) = θ2, f (η0 ∨ π) = θ0 and f (η1 ∨ π) = θ1. We compute C(η0, η1, ∆) ≤ η0 ∧ η1 ∧ ∆ ≤ [θ0, θ1, θ2]0 where the first inequality follows from the definition of C, and the second from Theorem 7.6. We then have that C(θ0, θ1, θ2) = C(f (η0 ∨ π), f (η1 ∨ π), f (∆ ∨ π)) = f (C(η0, η1, ∆) ∨ π) ≤ [θ0, θ1, θ2] (cid:3) An immediate consequence of Proposition 8.2 is the following Theorem 8.3. For θ0, θ1, θ2 ∈ Con(A), we have that [[θ0, θ1], θ2] ≤ [θ0, θ1, θ2]. 9. Ternary Commutator and 3-Dimensional Cube Terms In [11], Oprsal defines what is called a strong cube term for a congruence per- mutable variety. These strong cube terms are definable from a Mal'cev operation. We reproduce this definition for the 3-dimensional case. Let h(x(0,0), x(1,0), x(0,1)) be a term operation satisfying the identities (1) h(x, x, y) ≈ y (2) h(x, y, x) ≈ y. 22 ANDREW MOORHEAD Notice that h satisfies the Mal'cev identities up to a permutation of variables. We now define p(x(0,0,0), x(1,0,0), x(0,1,0), x(1,1,0), x(0,0,1), x(1,0,1), x(0,1,1)) := h(cid:0)h(x(0,0,0), x(1,0,0), x(0,1,0)), x(1,1,0), h(x(0,0,1), x(1,0,1), x(0,1,1))(cid:1) The operation p is called a 3-dimensional strong cube term. It satisfies the identities (1) p(w, w, x, x, y, y, z) ≈ z (2) p(w, x, w, x, y, z, y) ≈ z (3) p(w, x, y, z, w, x, y) ≈ z Now, every modular variety has a weakened Mal'cev operation called a difference term. This is a term operation d(x(0,0), x(1,0), x(0,1)) that satisfies (1) d(x, x, y) ≈ y (2) d(x, y, x) ≡[θ,θ] y, where θ = Cg((cid:104)x, y(cid:105)). Notice the permutation of variables. We will later need the following Proposition 9.1 (Special Case of 5.7 in [4]). Let θ ∈ Con(A), and suppose x ≡θ y ≡θ z. Then d(x, y, z) x y ∈ ∆θ,θ (cid:20) z (cid:21) We use the difference term to define a term operation that is analogous to a strong cube term. Set d(cid:0)d(x(0,0,0), x(1,0,0), x(0,1,0)), x(1,1,0), d(x(0,0,1), x(1,0,1), x(0,1,1))(cid:1) q(x(0,0,0), x(1,0,0), x(0,1,0), x(1,1,0), x(0,0,1), x(1,0,1), x(0,1,1)) := We call q a 3-dimensional wobbly cube term. In fact, q satisfies (1) q(x, x, x, x, y, y, y) ≈ y (2) q(x, x, y, y, x, x, y) ≈ y (3) q(x, y, x, y, x, y, x) ≡[θ,θ,θ] y, where θ = Cg((cid:104)x, y(cid:105)), which we shall now demonstrate. Theorem 9.2. Let θ ∈ Con(A). Let m ∈ A23 be such that face0 0(m), face0 1(m), face0 2(m) ∈ ∆θ,θ. Set (∗) = q(m(0,0,0), m(1,0,0), m(0,1,0), m(1,1,0), m(0,0,1), m(1,0,1), m(0,1,1)) = q(a, b, c, d, e, f, g) Take z ∈ A23 such that zf = mf when f (cid:54)= (1, 1, 1), and z(1,1,1) = (∗). Then z ∈ ∆θ0,θ1,θ2 . Proof. See Figure 18 for a picture of what is to be shown. Let A, B, C, D, E, F, G ∈ ∆θ0,θ1,θ2 be the matrices in Figure 19. As shown, q(A, B, C, D, E, F, G) ∈ ∆θ0,θ1,θ2 also. Suppose now that the conditions (1), (2) and (3) that are shown in Figure 20 hold. Using that ∆θ,θ,θ is a congruence that relates any two opposing faces will finish the proof. So, it remains to check that the conditions (1), (2) and (3) hold. We will show (2). The proof of the others is similar. 23 (cid:21) ∈ ∆θ,θ so (cid:20) d(a, b, c) d(a, b, c) (cid:21) c c ∈ ∆∆θ,θ,∆θ,θ , so we use Proposition 9.1 to conclude that . By Theorem 7.5, ∆θ,θ,θ = . (cid:3) Now, an application of Proposition 9.1 gives that ∆θ,θ,θ ∈ ∆θ,θ, because (cid:21) b a (cid:20) c (cid:20) d d(a, b, c) (cid:20) d (cid:20) d d(a, b, c) ∆θ,θ d(a, b, c) (cid:21) (cid:20) d(a, b, c) (cid:21) (cid:21) c c c c c c ∆θ,θ,θ Moreover, ∆θ,θ. Therefore, The result now follows, see Figure 21. Figure 18 d(a, b, c) (cid:21) b . a (cid:20) c (cid:20) c (cid:21) (cid:20) c (cid:20) d(a, b, c) d a b (cid:21) a b d(a, b, c) ∈ ∆θ,θ. Also, (cid:21) (cid:20) q(a, b, c, d, a, b, c) d(a, b, c) ∆θ,θ c (cid:21) c c 012cdabgef∈∆θ,θ,θcageabefcdab∈∆θ,θq(a,b,c,d,e,f,g),,24 ANDREW MOORHEAD Figure 19 012ccaacaacdabcabaaaaaaaababaabccaageeaaaaeeeababeefabcdegfABCDEFGaecgq(a,b,a,b,a,b,a)q(a,b,c,d,a,b,c)q(a,b,a,b,e,f,e)(∗)25 Figure 20 Figure 21 012ababaab∈∆θ,θ,θq(a,b,a,b,a,b,a)cdabcabq(a,b,c,d,a,b,c)ababeefq(a,b,a,b,e,f,e),,(1)(2)(3)cdabcaq(a,b,c,d,a,b,c)q(a,b,a,b,a,b,a)(1)+(2)∈∆θ,θ,θ(4)(5)(4)+q(A,B,C,D,E,F,G)∈∆θ,θ.θcdabgeq(a,b,a,b,e,f,e)∈∆θ,θ,θ(∗)(6)(3)+(5)cdabgef∈∆θ,θ,θ(∗)acacbbd(a,b,c)d(a,b,c)ccccd(a,b,c)d(a,b,c)dq(a,b,c,d,a,b,c),∈∆θ,θ,θacacbbdq(a,b,c,d,a,b,c)∈∆θ,θ,θ26 ANDREW MOORHEAD Proposition 9.3. Let q be a 3-dimensional wobbly cube term for a modular variety V. Then q satisfies (1) q(x, x, x, x, y, y, y) ≈ y (2) q(x, x, y, y, x, x, y) ≈ y (3) q(x, y, x, y, x, y, x) ≡[θ,θ,θ] y, where θ = Cg((cid:104)x, y(cid:105)) Proof. The first two identities follow from the definition of q. Item (3) follows from ((cid:104)x, y(cid:105)). Then Theorems 9.2 and 7.6. Let A ∈ V, and take x, y ∈ A. Let θ = Cg cube3 A 0(x, y) ∈ ∆θ,θ,θ. Therefore, (cid:28)(cid:20) x y (cid:21) , (cid:20) x q(x, y, x, y, x, y, x) (cid:21)(cid:29) x y and the result follows. ∈ ∆θ,θ,θ, (cid:3) x y Figure 22 c0d0a0b0g0e0f0q(a0,b0,c0,d0,e0,f0,g0)cm−1dm−1am−1bm−1gm−1em−1fm−1q(am−1,bm−1,cm−1,dm−1,em−1,fm−1,gm−1),,∈∆θ,θ,θt(c)t(d)t(a)t(b)t(g)t(e)t(f)T(Q)∈∆θ,θ,θt(c)t(d)t(a)t(b)t(g)t(e)t(f)Q(T)∈∆θ,θ,θ(1)(2)(3)(1)+(2)t(g)t(e)t(f)Q(T)∈∆θ,θ,θt(g)t(e)t(f)T(Q)References 27 [1] E. Aichinger and N. Mudrinski, Some applications of higher commutators in Mal'cev algebras, Algebra Universalis 63 (2010), no. 4, 367 -- 403. MR2734303 [2] A. Bulatov, On the number of finite Mal'tsev algebras, in Contributions to general algebra, 13 (Velk´e Karlovice, 1999/Dresden, 2000), 41 -- 54, Heyn, Klagenfurt. MR1854568 [3] A. Day, A characterization of modularity for congruence lattices of algebras, Canad. Math. Bull. 12 (1969), 167 -- 173. MR0248063 [4] R. Freese and R. McKenzie, Commutator theory for congruence modular varieties, London Mathematical Society Lecture Note Series, 125, Cambridge Univ. Press, Cambridge, 1987. MR0909290 [5] H. P. Gumm, Geometrical methods in congruence modular algebras, Mem. Amer. Math. Soc. 45 (1983), no. 286, viii+79 pp. MR0714648 [6] J. Hagemann and C. Herrmann, A concrete ideal multiplication for algebraic systems and its relation to congruence distributivity, Arch. Math. (Basel) 32 (1979), no. 3, 234 -- 245. MR0541622 [7] K. A. Kearnes and E. W. Kiss, The shape of congruence lattices, Mem. Amer. Math. Soc. 222 (2013), no. 1046, viii+169 pp. MR3076179 [8] K. A. Kearnes and Szendrei, The relationship between two commutators, Internat. J. Algebra Comput. 8 (1998), no. 4, 497 -- 531. MR1663558 [9] A. Moorhead, Higher commutator theory for congruence modular Varieties, Journal of Alge- bra, 513 (2018), 133-158. [10] A. Moorhead, Supernilpotent Taylor algebras are nilpotent, manuscript in preparation. [11] J. Oprsal, A relational description of higher commutators in Mal'cev varieties, Algebra Uni- versalis 76 (2016), no. 3, 367 -- 383. MR3556818 [12] J. D. H. Smith, Mal'cev varieties, Lecture Notes in Mathematics, Vol. 554, Springer, Berlin, 1976. MR0432511 [13] A. Wires, On supernilpotent algebras, ArXiv e-prints, arXiv:1701.08949 [math.RA], jan. 2017 (Andrew Moorhead) Department of Mathematics; Vanderbilt University; Nashville, TN; U.S.A. E-mail address, Andrew Moorhead: [email protected]
1905.12328	1	1905	2019-05-29T11:14:44	Derivations and dimensionally nilpotent derivations in Lie triple algebras	[ "math.RA" ]	In this paper, we first study derivations in non nilpotent Lie triple algebras. We determine the structure of derivation algebra according to whether the algebra admits an idempotent or a pseudo-idempotent. We study the multiplicative structure of non nilpotent dimensionally nilpotent Lie triple algebras. We show that when $n=2p+1$ the adapted basis coincides with the canonical basis of the gametic algebra $G(2p+2,2)$ or this one obviously associated to a pseudo-idempotent and if $n=2p$ then the algebra is either one of the precedent case or a conservative Bernstein algebra.	math.RA	math	Derivations and dimensionally nilpotent derivations in Lie triple algebras Abdoulaye Dembega∗ Université Norbert Zongo Amidou Konkobo†and Moussa Ouattara‡ Université Joseph KI-ZERBO BP 376 Koudougou, Burkina Faso 03 BP 7021 Ouagadougou 03, Burkina Faso May 30, 2019 Abstract In this paper, we first study derivations in non nilpotent Lie triple algebras. We determine the structure of derivation algebra according to whether the algebra admits an idempotent or a pseudo-idempotent. We study the multiplicative structure of non nilpotent dimensionally nilpotent Lie triple algebras. We show that when n = 2p + 1 the adapted basis coincides with the canonical basis of the gametic algebra G(2p + 2, 2) or this one obviously associated to a pseudo-idempotent and if n = 2p then the algebra is either one of the precedent case or a conservative Bernstein algebra. Keyword: Dimensionally nilpotent Lie triple algebra, pseudo-idempotent, Jordan algebra, ascending basis, adapted basis. 2010 Mathematics Subject Classification : Primary 17A30, secondary 17D92, 17B40, 17C10 1 Introduction A n + 1 finite dimensional algebra A is dimensionally nilpotent if there is a derivation d : A −→ A such that dn+1 = 0 and dn 6= 0. This notion has been studied by G.F. Leger and P.L. Manley[8] for Lie algebras, J.M. Osborn [12] for Jordan algebras, Micali and Ouattara[9] for genetic algebras. Recently, V. Eberlin [4] has deepened the work of the authors of [8] in his thesis. Regarding Jordan algebras, Osborn shows that every dimensionally nilpotent Jordan K-algebra is either nilpotent or satisfies A/Rad(A) ≃ K. We study the case of non nilpotent dimensionally Lie triple algebras. In an adapted basis we caracterize the multiplicative structure of these algebras relative to the parity of n. More precisely we show that when n = 2p + 1, the adapted basis coincides with canonical basis of the gametic algebra G(2p + 2, 2) or this one obviously associated to a pseudo-idempotent. If n = 2p then this algebra is either one of the precedent case or a train algebra of rank 3 which is a Jordan algebra [13]. Since Jordan algebras are also Lie triple ones the final corollary describes non nilpotent dimensionally nilpotent Jordan algebras. ∗[email protected] †[email protected] ‡[email protected] 1 2 Preliminaries A Lie triple algebra is a commutative algebra satisfying (1) 2x(x(xy)) + yx3 = 3x(yx2) while a Jordan algebra is a commutative algebra satisfying Every Jordan algebra satisfies identity (1). x2(yx) = (x2y)x. Theorem 2.1 ([5]). Let A be a Lie triple algebra and L the ideal generated by the associators (x2, x, x). Then L2 = 0 and A/L is a Jordan algebra. Definition 2.2. A pseudo-idempotent of A is a non-zero element e such that there is t 6= 0 in L satisfying e2 = e + t and et = 1 2 t. Theorem 2.3 ([2]). Every Lie triple non nilalgebra contains either a non-zero idempotent, or a pseudo-idempotent. Definition 2.4. An ideal I of an algebra A is said to be caracteristic if d(I) ⊆ I for every derivation d of A. An ideal I of an algebra A is said to be d-invariant if d(I) ⊆ I for a given derivation d of A. 3 Caracterization of derivations In this paragraph we study the derivations in Lie triple non nilalgebras . We give a caracterization, distinguishing two cases: with an idempotent or with a pseudo-idempotent. 3.1 Lie triple algebras with idempotent Relative to the non-zero idempotent e, A admits the following Peirce decomposition A = Ae(1) ⊕ Ae( 1 2 ) ⊕ Ae(0). Relations between Peirce components and the products of their elements are ruled by the following lemma: Lemma 3.1 ([2], Lemme 2.2). Let A = Ae(1) ⊕ Ae(1/2) ⊕ Ae(0) be the Peirce decomposition of A relative to a non-zero idempotent. Then (i) Ae(1/2)Ae(1/2) ⊆ Ae(1) + Ae(0), Ae(λ)Ae(λ) ⊆ Ae(λ), Ae(λ)Ae(1/2) ⊆ Ae(1/2), Ae(λ)Ae(1 − λ) = 0, (λ = 0, 1) ; (ii) (x1y1)a1/2 = x1(y1a1/2) + y1(x1a1/2), (x0y0)a1/2 = x0(y0a1/2) + y0(x0a1/2) ; (iii) [x1(x1/2a1/2)]1 = [(x1x1/2)a1/2 + (x1a1/2)x1/2]1, [x0(x1/2a1/2)]0 = [(x0x1/2)a1/2 + (x0a1/2)x1/2]0 ; (iv) [(x1x1/2)y1/2]0 = [(x1y1/2)x1/2]0, [(x0x1/2)y1/2]1 = [(x0y1/2)x1/2]1 ; (v) x0(y1a1/2) = y1(x0a1/2) ; (vi) x1/2(x2 1/2)1 = x1/2(x2 1/2 ; 1/2)0 = 1 2 x3 2 (vii) (x1/2y1/2)0z1/2 + (y1/2z1/2)0x1/2 + (z1/2x1/2)0y1/2 = (x1/2y1/2)1z1/2 + (y1/2z1/2)1x1/2 + (z1/2x1/2)1y1/2. Since A is e-stable, i.e. Ae(λ)Ae(1/2) ⊆ Ae(1/2) and [(xλx1/2)y1/2]1−λ = [(xλy1/2)x1/2]1−λ with λ = 0, 1, calculations on derivations give results similar to [1, Corollary 2], precisely. Theorem 3.2. Every derivation d of A is determined and only defined by a quadruplet (d(e), fd, gd, hd) with fd ∈ EndK(Ae(1/2)), gd ∈ DerK(Ae(0)) and hd ∈ DerK (Ae(1)) satisfying the following con- ditions: (i) d(e) ∈ Ae(1/2) ; (ii) d(x1) = hd(x1) + 2d(e)x1 ; (iii) d(x1/2) = fd(x1/2) + 2(d(e)x1/2)0 − 2(d(e)x1/2)1 ; (iv) d(x0) = gd(x0) − 2d(e)x0 ; (v) hd(x1y1) = hd(x1)y1 + x1hd(y1) ; (vi) gd(x0y0) = gd(x0)y0 + x0gd(y0) ; (vii) hd((x1/2y1/2)1) = [fd(x1/2)y1/2 + x1/2fd(y1/2)]1 ; (viii) gd((x1/2y1/2)0) = [fd(x1/2)y1/2 + x1/2fd(y1/2)]0 ; (ix) fd(x1x1/2) = hd(x1)x1/2 + x1fd(x1/2) ; (x) fd(x0x1/2) = gd(x0)x1/2 + x0fd(x1/2). Proposition 3.3. Let A be a Lie triple algebra and A = Ae(1) ⊕ Ae(1/2) ⊕ Ae(0) the Peirce decomposition of A relative to an idempotent e 6= 0. Subspaces Jλ = {xλ ∈ Ae(λ) xλAe(1/2) = 0} (λ = 0, 1) and J = J0 ⊕ J1 are caracteristic ideals of A and the quotient algebra A/J is a Jordan algebra. Proof. Considering Jλ = ker(Sλ), with Sλ : Ae(λ) → EndK (Ae(1/2)), xλ 7→ Sλ(xλ) and Sλ(xλ) : a1/2 7→ xλa1/2. We know by ([10]) that Jλ is an ideal of Ae(λ) (λ = 0, 1) and since Ae(λ)Ae(1/2) ⊆ Ae(1/2), then J = J1 + J0 is an ideal of A such that A/J is a Jordan algebra ([11], Proposition 6.7). Let's consider d ∈ DerK (A), xλ ∈ Jλ(e) and a1/2 ∈ Ae(1/2). We have 0 = d(xλa1/2) = xλd(a1/2) + d(xλ)a1/2. But d(a1/2) = fd(a1/2) + 2(d(e)a1/2)0 − 2(d(e)a1/2)1, therefore we have xλd(a1/2) = 0 because xλ(d(e)a1/2)λ = [(xλd(e))a1/2 + (xλa1/2)d(e)]λ. Hence d(xλ)a1/2 = 0 with λ = 0, 1. But, on the one hand we have d(x1) = hd(x1) − 2d(e)x1, and 0 = d(x1)a1/2 = hd(x1)a1/2 and then hd(x1) ∈ J1, on the other hand we have d(x0) = gd(x0) − 2d(e)x0, with 0 = d(x0)a1/2 = gd(x0)a1/2 and then gd(x0) ∈ J0. Hence d(Jλ) ⊆ Jλ and we conclude that d(J) ⊆ J. ✷ 3.2 Lie triple algebras with pseudo-idempotent Lemma 3.4 ([2], Proposition 4.3). Let L = Le(1)⊕Le(1/2)⊕Le(0) and A = Ae(1)⊕Ae(1/2)⊕Ae(0) be the respective Peirce decomposition of L and A, relative to the pseudo-idempotent e, satisfying e2 = e + t with t ∈ L1/2 fixed. Then (i) Ae(0)Le(1/2) ⊆ Le(1/2), Ae(1)Le(1/2) ⊆ Le(1/2), Ae(1)Le(1) ⊆ Le(1), Ae(0)Le(0) ⊆ Le(0), Ae(0)Le(1) = Ae(1)Le(0) = 0, Ae(1/2)Le(0) = Ae(1/2)Le(1) = Ae(1/2)Le(1/2) = 0 ; (ii) Ae(1)Ae(0) ⊆ Le(1/2), Ae(0)Ae(1/2) ⊆ Ae(1/2), Ae(1)Ae(1/2) ⊆ Ae(1/2), Ae(0)Ae(0) ⊆ Ae(0) + Le(1/2), Ae(1)Ae(1) ⊆ Ae(1) + Le(1/2), Ae(1/2)Ae(1/2) ⊆ Ae(1) + Ae(0) ; (iii) (x0y0)1/2 = 4(x0t)y0 = 4(y0t)x0 ; (x1y1)1/2 = 4(x1t)y1 = 4(y1t)x1 ; (x0y1)1/2 = 4(x0t)y1 = 4(y1t)x0 ; 3 (iv) (x1y1)a1/2 = x1(y1a1/2) + y1(x1a1/2) ; (v) (x0y0)a1/2 = x0(y0a1/2) + y0(x0a1/2) ; (vi) x0(y1a1/2) = y1(x0a1/2) ; (vii) [x0(x1/2a1/2)]0 = [(x0x1/2)a1/2 + (x0a1/2)x1/2]0 ; (viii) [x1(x1/2a1/2)]1 = [(x1x1/2)a1/2 + (x1a1/2)x1/2]1 ; (ix) [(x0x1/2)y1/2]1 = [(x0y1/2)x1/2]1 ; [(x1x1/2)y1/2]0 = [(x1y1/2)x1/2]0 ; (x) (x1/2y1/2)0z1/2 + (y1/2z1/2)0x1/2 + (z1/2x1/2)0y1/2 = (x1/2y1/2)1z1/2 + (y1/2z1/2)1x1/2 + (z1/2x1/2)1y1/2. Lemma 3.5. Let A be a Lie triple algebra and e a pseudo-idempotent of A: e2 = e + t, et = 1 2 t, t2 = 0 with t ∈ L. For every derivation d of A, we have d(t) = 0 and d(e) ∈ Ae(1/2). Proof. Let's consider d ∈ DerK(A). Since e2 = e + t, we have 2ed(e) = d(e) + d(t). Setting d(e) = [d(e)]1 + [d(e)]1/2 + [d(e)]0, we have d(t) = [d(e)]1 − [d(e)]0. Because of 2et = t, we deduce 2ed(t) + 2d(e)t = d(t). We have 2d(e)t = −[d(e)]1 − [d(e)]0. We know that t ∈ Le(1/2) and Le(1/2) is an ideal of A. It follows that [d(e)]1 = [d(e)]0 = 0. ✷ Theorem 3.6. Every derivation d of A is determined and only defined by a quadruplet (d(e), fd, gd, hd) with fd ∈ EndK (Ae(1/2)), gd ∈ EndK(Ae(0)) and hd ∈ EndK(Ae(1)) satisfying the following con- ditions: (i) d(e) ∈ Ae(1/2) ; (ii) d(x1) = hd(x1) + 2d(e)x1 ; (iii) d(x1/2) = fd(x1/2) + 2(d(e)x1/2)0 − 2(d(e)x1/2)1 ; (iv) d(x0) = gd(x0) − 2d(e)x0 ; (v) hd((x1y1)1) = [hd(x1)y1 + x1hd(y1)]1 ; fd((x1y1)1/2) = [hd(x1)y1 + x1hd(y1)]1/2 = 2hd((x1y1)1)t ; (vi) gd((x0y0)0) = [gd(x0)y0 + x0gd(y0)]0 ; fd((x0y0)1/2) = [gd(x0)y0 + x0gd(y0)]1/2 = 2gd((x0y0)0)t ; (vii) hd((x1/2y1/2)1) = [fd(x1/2)y1/2 + x1/2fd(y1/2)]1 ; gd((x1/2y1/2)0) = [fd(x1/2)y1/2 + x1/2fd(y1/2)]0 ; (viii) fd(x1x0) = hd(x1)x0 + x1gd(x0) ; (ix) fd(x1x1/2) = hd(x1)x1/2 + x1fd(x1/2) ; (x) fd(x0x1/2) = gd(x0)x1/2 + x0fd(x1/2). Proof. Let d be a derivation of A and e a pseudo-idempotent of A. Since d(e) ∈ Ae(1/2), we have (i). Let x1 ∈ Ae(1). We have ex1 = x1, and then d(e)x1 + ed(x1) = d(x1). Let's set d(x1) = a1 + a1/2 + a0. Then d(e)x1 + a1 + 1 2 a1/2 = a1 + a1/2 + a0, and we have a1/2 = 2d(e)x1 and a0 = 0. Hence d(x1) = hd(x1) + 2d(e)x1 with hd an endomorphism of Ae(1) and (ii) is prooved. By similar calculations we have (iii) and (iv). with x1, y1 ∈ Ae(1), we have (x1y1)1/2 = 4x1(y1t) = 4y1(x1t) = 2(x1y1)t d(x1y1) = d[(x1y1)1] + d[(x1y1)1/2] = d[(x1y1)1] + 2d((x1y1)t) = hd[(x1y1)1] + 2d(e)(x1y1) + 2d((x1y1))t + 2(x1y1)d(t) = hd[(x1y1)1] + 2hd(x1y1)t + 2d(e)(x1y1), because d((x1y1))t = hd(x1y1)t. 4 We also have, d(x1y1) = d(x1)y1 + x1d(y1) = x1[hd(y1) + 2d(e)y1] + [hd(x1) + 2d(e)x1]y1 = hd(x1)y1 + x1hd(y1) + 2[d(e)y1]x1 + 2[d(e)x1]y1. = hd(x1)y1 + x1hd(y1) + 2d(e)(x1y1). It follows that hd((x1y1)1) = [hd(x1)y1 + x1hd(y1)]1 et fd((x1y1)1/2) = [hd(x1)y1 + x1hd(y1)]1/2 = 2hd(x1y1)t, and we have (v). We show by similar calculations that: gd((x0y0)0) = [gd(x0)y0 + x0gd(y0)]0 et fd((x0y0)1/2) = [gd(x0)y0 + x0gd(y0)]1/2 = 2gd(x0y0)t, and we have (vi). Let x1/2, y1/2 ∈ Ae(1/2). We have d(x1/2y1/2) = d((x1/2y1/2)1) + d((x1/2y1/2)0) = hd((x1/2y1/2)1) + 2d(e)(x1/2y1/2)1 +gd((x1/2y1/2)0) − 2d(e)(x1/2y1/2)0, But d(x1/2y1/2) = d(x1/2)y1/2 + x1/2d(y1/2) = (fd(x1/2) + 2[d(e)x1/2]0 − 2[d(e)x1/2]1)y1/2 +x1/2(fd(y1/2) + 2[d(e)y1/2]0 − 2[d(e)y1/2]1) = fd(x1/2)y1/2 + 2y1/2[d(e)x1/2]0 − 2y1/2[d(e)x1/2]1 +x1/2fd(y1/2) + 2x1/2[d(e)y1/2]0 − 2x1/2[d(e)y1/2]1 = fd(x1/2)y1/2 + x1/2fd(y1/2) + 2d(e)(x1/2y1/2)1 −2d(e)(x1/2y1/2)0 because of identity (x) of Lemma 3.4. It follows that: d(x1/2y1/2) = fd(x1/2)y1/2 + x1/2fd(y1/2) + 2d(e)(x1/2y1/2)1 − 2d(e)(x1/2y1/2)0 and we have We have hd((x1/2y1/2)1) = [fd(x1/2)y1/2 + x1/2fd(y1/2)]1 et gd((x1/2y1/2)0) = [fd(x1/2)y1/2 + x1/2fd(y1/2)]0, and we have (vii). d(x1x0) = fd((x1x0)1/2) = d(x1)x0 + x1d(x0) = [hd(x1) + 2d(e)x1]x0 + x1[gd(x0) − 2d(e)x0] = hd(x1)x0 + x1gd(x0), 5 fd((x1x0)1/2) = hd(x1)x0 + x1gd(x0), and we have (viii). In a similar way d(x1x1/2) = fd(x1x1/2) + 2[d(e)(x1x1/2)]0 − 2[d(e)(x1x1/2)]1 = d(x1)x1/2 + x1d(x1/2) = [hd(x1) + 2d(e)x1]x1/2 + x1[fd(x1/2) + 2[d(e)x1/2]0 − 2[d(e)x1/2]1] = hd(x1)x1/2 + x1fd(x1/2) + 2[d(e)x1]x1/2 + 2x1[d(e)x1/2]0 − 2x1[d(e)x1/2]1, fd(x1x1/2) = hd(x1)x1/2 + x1fd(x1/2), and we have (ix). So we have d(x0x1/2) = fd(x0x1/2) = d(x0)x1/2 + x0d(x1/2) = [gd(x0) − 2d(e)x0]x1/2 + x0[fd(x1/2) + 2[d(e)x1/2]0 − 2[d(e)x1/2]1] = gd(x0)x1/2 + x0fd(x1/2) − 2[d(e)x0]x1/2 + x0[2[d(e)x1/2]0 −2[d(e)x1/2]1], fd(x0x1/2) = gd(x0)x1/2 + x0fd(x1/2), and finally (x). Conversely, once we have identities (i) to (xi), setting x = x1 + x1/2 + x0 and y = y1 + y1/2 + y0, ✷ we show that d(xy) = d(x)y + xd(y). Example 3.7. Let A be the four dimensional Lie triple K-algebra which multiplication table in the basis {e, t, u, r} is given by : e2 = e + t, u2 = u + r, et = 1 2 r, all other product being zero. The Peirce decomposition of A relative to pseudo-idempotent e gives Ae(1) = K(e + 2t), Ae(1/2) = Kt, Ae(0) =< u, r >. Let d be a derivation of A. Since e and u are pseudo-idempotents, we have d(t) = d(r) = 0, d(e) = αt, d(u) = βr. The derivation algebra is two dimensional. 2 t, ur = 1 Example 3.8. Let's consider the four dimensional Lie triple K-algebra A which multiplication table in the basis {e, t1, t2, v} is given by : e2 = e + t1, et1 = 1 2 t2, ev = v and vt1 = t2, all other products being zero. Then we have Ae(1) =< (e + 2t), v >, Ae(1/2) =< t1, t2 >, Ae(0) = 0. Let d be a derivation of A. We have d(t1) = 0. Set d(e) = α1t1 + β1t2. Thus α1t1 + β1t2 = d(e + 2t1) = hd(e + 2t1) + 2d(e)(e + 2t1) = hd(e + 2t1) + α1t1 + β1t2. It follows that hd(e + 2t1) = 0. 2 t1, et2 = 1 Setting d(v) = α2(e + 2t1) + β2v + γ2t1 + η2t2, relation 0 = d(v2) = 2d(v)v gives α2 = γ2 = 0. It follows that hd(v) = β2v. Furthermore, relation fd((vt1)1/2) = hd(t1) + fd(t1) gives fd(t2) = β2t2. We have d(e) = α1t1+β1t2, hd(e+2t1) = 0, hd(v) = β2v, fd(t1) = 0, fd(t2) = β2t2. The derivation algebra is three dimensional. Proposition 3.9. Let's consider a pseudo-idempotent e 6= 0. Subspace Je(1/2) = {x1/2 ∈ Ae(1/2) x1/2Ae(1/2) = 0} is a caracteristic ideal and A/Je(1/2) is a Lie triple algebra with e as idempotent. Proof. Let x1/2 ∈ Je(1/2), a1/2 ∈ Ae(1/2) and yλ ∈ Ae(λ) (λ = 0, 1). We have [(x1/2yλ)a1/2]λ = [(a1/2yλ)x1/2]λ = 0 and [(x1/2yλ)a1/2]1−λ = [(a1/2yλ)x1/2]1−λ = 0, and then (x1/2yλ)a1/2 = 0. Hence Ae(λ)Je(1/2) ⊆ Je(1/2), and it follows that AJe(1/2) ⊆ Je(1/2). Je(1/2) is an ideal of A. Since t ∈ L1/2 ⊆ Je(1/2), e is an idempotent of quotient algebra A/Je(1/2). Let's consider now d ∈ DerK(A), x1/2 ∈ Je(1/2) and a1/2 ∈ Ae(1/2). We have 0 = d(x1/2a1/2) = x1/2d(a1/2) + d(x1/2)a1/2. But d(x1/2) = fd(x1/2) ∈ A1/2 and d(a1/2) = fd(a1/2) + 2(d(e)a1/2)0 − 2(d(e)a1/2)1, it follows that x1/2d(a1/2) = 0 because x1/2(d(e)a1/2)1 = x1/2(d(e)a1/2)0. So d(x1/2)a1/2 = 0, and d(x1/2) ∈ Je(1/2). We conclude that d(Je(1/2)) ⊆ Je(1/2). ✷ 6 4 Dimensionally nilpotent Lie algebras Definition 4.1. Let A be a n + 1 finite dimensional K-algebra . If there is a nilpotent K-derivation d of A such that dn+1 = 0 and dn 6= 0, d is said to be dimensionally nilpotent, and so is the algebra A, though A is not necessarily nilpotent. If so it is, there is a basis {e0, e1, . . . , en} of A such that d(ei) = ei+1 (i = 0, . . . n − 1) and d(en) = 0 and the basis {e0, e1, . . . , en} is said to be adapted to d. Example 4.2. [9, Exemple 2.5] Let K be a commutative field of characteristic 6= 2 and A = G(n + 1, 2) the gametic diploid algebra with n + 1 alleles. Its multiplication table in the natural basis {a0, . . . , an} is given by aiaj = 1 2 aj. We know that the mapping ω : A → K, ai 7→ 1 is a weight function and if we set ei = a0 − ai (i 6= 0) then {e1, . . . , en} is a basis of the ideal N = ker ω and e0 = a0 is an idempotent of A such that {e0, e1, . . . , en} is the canonical basis of A, so e0ei = 1 2 ei (i = l, . . . , n) and if d is a derivation of A, e0d(ei) = 1 2 d(ei) (i = 1, . . . , n), because d(e0) ∈ N and N is a zero algebra. So we just need to define d : A → A by d(ei) = ei+1 (i = 1, . . . , n − 1), d(e0) = e1 and d(en) = 0. It follows that dn+1 = 0 and dn 6= 0, showing the gametic algebra A = G(n + 1, 2) is dimensionally nilpotent. 2 ai + 1 0 = e0 + en, all other product being zero. If d is a derivation of A, e0d(ei) = 1 Example 4.3. Let K be a commutative field of characteristic 6= 2 and A the n + 1 dimensional commutative K-algebra, which multiplication table in the basis {e0, e1, . . . , en} is given by e0ei = 1 2 ei (i = 1, . . . , n), e2 2 d(ei) (i = 1, . . . , n) because d(e0) ∈ N =< e1, . . . , en > and N is a zero algebra. Here, we just need again to define d : A → A by d(ei) = ei+1 (i = 1, . . . , n − 1), d(e0) = e1 and d(en) = 0. We have dn+1 = 0 and dn 6= 0, that shows the algebra A is dimensionally nilpotent. Since Ken is an ideal, the quotient algebra A/Ken is isomorphic to G(n, 2). 4.1 Basic tools Theorem 4.4 ([12]). Let K be a perfect field of characteristic 6= 2 and 3 and A finite dimensional K-Jordan algebra, dimensionally nilpotent. Then either A is nilpotent or dimK (A/rad(A)) = 1. Remark 4.5. Let A be a dimensionally nilpotent Lie triple non nilalgebra. Because of Theorem 2.3 we consider two cases : 1) A has an idempotent e. Since the ideal J is caracteristic, the quotient algebra A = A/J is a dimensionally nilpotent Jordan algebra. Because of Theorem 4.4 we have dimK(A/rad(A) = 1 and since rad(A) ≃ rad(A)/J, the first isomorphism theorem gives A/rad(A) ≃ A/rad(A) and dimK A/rad(A) = 1. Then we can write A = Ke ⊕ N , with N = rad(A). 2) A has a pseudo-idempotent e. Since the ideal Je(1/2) is caracteristic, the quotient algebra A = A/Je(1/2) is a dimensionally Lie triple algebra with e as idempotent. Because of 1) we can write A = Ke ⊕ N with N = rad(A). So we have A = Ke ⊕ N with N = rad(A). Lemma 4.6. Let x, y ∈ N such that x 6= 0 and α ∈ K. If xy = αy then α = 0 or y = 0. Proof. Since N is nilpotent, there is m ∈ N∗ such that Lm operator by x. Then α = 0 or y = 0. x (y) = αmy = 0, Lx being the multiplicative ✷ From now on, throughout the paper, A is a dimensionally nilpotent Lie triple non nilalgebra of dimension n + 1, with {e0, e1, . . . , en} an adapted basis to the derivation d. We can consider e0 7 2 t and t2 = 0 either, as an idempotent, or a pseudo-idempotent. In the last case, e2 implies d(t) = 0 (Lemma 3.5), that means t = αen with α ∈ K. Since t ∈ Ae(1/2), if α 6= 0, then en ∈ Ae(1/2). 0 = e0 + t, e0t = 1 Lemma 4.7. We have: (i) e0en = λnen (ii) eken = 0 with 1 ≤ k ≤ n i=0 λiei. Deriving k times successively, we have eken = Pn−k Proof. Let's write e0en = Pn i=0 λiei+k. With k = n, it follows that e2 n = λ0en and because of Lemma 4.6 we have λ0 = 0. Set k = n − 1, one has en−1en = λ1en. That implies λ1 = 0. And so on, we have λ0 = λ1 = · · · = λn−1 = 0, e0en = λnen. Deriving successively e0en it follows that eken = 0 with 1 ≤ k ≤ n. ✷ Lemma 4.8. We have : (i) e0ek = λkek + Pn (ii) eiek = Pn (iii) λk ∈ {0, 1 2 , 1} with 1 ≤ k ≤ n. i=k+1 ak,iei with 1 ≤ k ≤ n − 1; j=k+1 γikjej with 1 ≤ i ≤ k ≤ n − 1; 2 e1, e2 0 = e0, e0e1 = 1 Proof. Reason by recurrence on n. With n = 1 the multiplication table of the algebra A is given by e2 1 = 0 and the lemma is satisfied. Assume the lemma is true until an order n. Because of Lemma 4.7 the subspace In+1 = Ken+1 is a d-invariant ideal of A. The quotient algebra A/In+1 is dimensionally nilpotent of dimension n + 1. By the hypothesis, we have e0ek = λkek + Pn j=k+1 γikjej, with 1 ≤ i ≤ k ≤ n. Otherwise e0ek = λkek + Pn j=k+1 γikjej + γik,n+1en+1; and results (i) i=k+1 ak,iei + ak,n+1en+1 and eiek = Pn i=k+1 ak,iei and eiek = Pn and (ii) follow. Now we just need to show (iii). Since 2L3 operator by e0, applying it to ek we have 2λ3 e0 − 3L2 k − 3λ2 e0 + Le0 = 0, with Le0 being the multiplicative k + λk = 0, soit λk ∈ {0, 1 ✷ 2 , 1}. 4.2 Example of low dimensions Here we deal with cases 1 ≤ n ≤ 4. Let A be a dimensionally nilpotent Lie triple algebra, of dimension n + 1 and {e0, e1, . . . , en} be a basis adapted to d. We have ker d = Ken. Since e0 is an idempotent or a pseudo-idempotent, e1 = d(e0) ∈ Ae(1/2), i.e e0e1 = 1 2 e1. Deriving this we have e0e2 + e2 2 e2, that means 1 = 1 (2) λ2 + γ112 = 1 2 et a2,k + γ11k = 0 (3 ≤ k ≤ n). k=2 γ11kek+1 = Pn We also have e2 that means 2γ12k = γ11,k−1 (3 ≤ k ≤ n). Let's derive for the second time e0e1 = 1 e0e3 + 3e1e2 = 1 k=2 γ11kek which derivative is 2e1e2 = Pn−1 1 = Pn 2 e3, that means k=3 γ11,k−1ek, 2 e1. We have (3) λ3 + 3γ123 = 1 2 et a3,k + 3γ1,2,k = 0 (4 ≤ k ≤ n). However we have, d(e0e2) = e0e3 + e1e2 = λ2e3 + Pn 4 a2,k−1ek, that implies (4) λ3 + γ123 = λ2 et a3,k + γ12k = a2,k−1 (4 ≤ k ≤ n). 8 So (4) implies γ123 ∈ {−1, − 1 2 , 0, 1 2 , 1}. So it is necessary to take λ3 = 1 2 in (3). Whence λ3 = λ2 = 1 2 and γ123 = γ112 = 0 if 3 ≤ n. Deriving e0e3 + 3e1e2 = 1 2 e3, one has e0e4 + 4e1e3 + 3e2 2 = 1 2 e4, that means (5) λ4 + 4γ134 + 3γ224 = 1 2 However, d(e0e3) = e0e4 + e1e3 = λ3e4 + Pn k=5 a3,k−1ek, which implies (6) λ4 + γ134 = λ3 et a4,k + γ13k = a3,k−1 (5 ≤ k ≤ n) We also have e1e2 = Pn which implies γ223 = 0 and γ134 + γ224 = 0. k=4 γ12kek because γ123 = 0 and d(e1e2) = e1e3 + e2 2 = Pn k=5 γ12,k−1ek, Case dimK A = 2 i.e n = 1. We obviously have e0e1 = 1 2 e1, e2 1 = 0, e2 0 = e0 or e2 0 = e0 + e1 all other product being zero. e0 e0 e1 1 2 e1 0 e0 e1 e0 e0 + e1 e1 1 2 e1 0 e0 e1 Case dimK A = 3 i.e n = 2. Because of (2) we have λ2 + γ112 = 1 ∗ λ2 = 0 ⇒ γ112 = 1 0 = e0, e0e1 = 1 2 , so e2 2 . Let's discuss the possible values of λ2. 2 e1, e2 1 = 1 2 e2 all other product being zero. e0 e0 e1 1 2 e1 2 e2 1 e2 0 0 0 e0 e1 e2 ∗ λ2 = 1 2 ⇒ γ112 = 0, so e2 0 = e0 or e2 0 = e0 + e2, e0e1 = 1 2 e1, e0e2 = 1 2 e2 all other product being zero. e0 e0 e1 1 2 e1 0 e2 1 2 e2 0 0 e0 e1 e2 e0 e0 + e2 e1 1 2 e1 0 e2 1 2 e2 0 0 e0 e1 e2 ∗ λ2 = 1 ⇒ γ112 = − 1 2, so e2 0 = e0, e0e1 = 1 2 e1, e0e2 = e2, e2 1 = − 1 2 e2 all other product being zero. e0 e0 e1 1 2 e1 − 1 2 e2 e2 e2 0 0 e0 e1 e2 9 Case dimK A = 4 i.e n = 3. Because of the preliminary calculations, λ3 = λ2 = λ1 = 1 2 e3 ⇒ e3 ∈ A 1 So e0e3 = 1 and finally e0e2 = 1 e0e3 = 1 2 e3, e2 0 = e0 or e2 and e2 1 = γ113e3. Since A2 1 2 2 2 , γ112 = γ123 = 0 and a2,3 + γ113 = 0. ⊆ A0 + A1 we have γ113 = 0 implying a2,3 = 0 2 e2. So we have the following multiplication table : e0e1 = 1 2 e1, e0e2 = 1 2 e2, 0 = e0 + e3, all other product being zero. e0 e0 e1 1 2 e1 0 e2 1 2 e2 0 0 e3 1 2 e3 0 0 0 e0 e1 e2 e3 e0 e0 + e3 e1 1 2 e1 0 e2 1 2 e2 0 0 e3 1 2 e3 0 0 0 e0 e1 e2 e3 Case dimK A = 5 i.e n = 4. 2 because of (6). Since e1e3 = γ134e4 = 1 ∗ λ4 = 0 ⇒ γ134 = 1 ⊆ A0 + A1, A 1 1 2 2 , A 1 A1 ⊆ A 1 A2 2 = − 1 (4) and finally γ113 = a2,3 = 0. In the same way e2 Ae(1/2)2 ⊆ Ae(0)+Ae(1) and Ae(0)2 ⊆ Ae(0). But e0e2 = 1 because of (2). Whence the following multiplication table . So e0e3 = 1 A0 ⊆ A 1 2 2 2 2 e4 we have e3 ∈ A 1 because 2 e3 and a3,4 = 0 ⇒ γ124 = 0 because of , because 2 e4 ⇒ e2 ∈ Ae(0) or e2 ∈ A 1 2 e2+a2,4e4 ⇒ e2 ∈ A 1 so a2,4 = 0 = γ114 2 2 2 e0 e0 e1 1 2 e1 0 e2 1 2 e2 0 − 1 2 e4 1 e3 1 2 e3 2 e4 0 0 e4 0 0 0 0 0 e0 e1 e2 e3 e4 ∗ λ4 = 1 2 ⇒ γ134 = γ224 = 0 because of (5) and (6). We have e1e2 = γ1,2,4e4 ∈ A 1 ⇒ 2 e2 + a2,3e3) so γ124 = 0 and then 2 e2 ∈ Ae(0) or e2 ∈ Ae(1) this is a contradiction (because e0e2 = 1 a3,4 = γ113 = 0. Whence the following multiplication table: e0 e0 + e4 e1 1 2 e1 0 e2 1 2 e2 0 0 e3 1 2 e3 0 0 0 e4 1 2 e4 0 0 0 0 e0 e0 e1 1 2 e1 0 e2 1 2 e2 0 0 e3 1 2 e3 0 0 0 e4 1 2 e4 0 0 0 0 e0 e1 e2 e3 e4 ∗ λ4 = 1 ⇒ e4 ∈ Ae(1) ⇒ γ134 = − 1 2 e0 e1 e2 e3 e4 10 We have e1e3 = − 1 2 e4 ⇒ e3 ∈ A 1 2 = 1 same way e2 Whence this table 2 e4 ⇒ e2 ∈ A 1 2 ⇒ e0e3 = 1 (because e2 can not be in Ae(1)), e0e2 = 1 2 e3 ⇒ a3,4 = 0. So γ124 = γ113 = a2,3 = 0. In the 2 e2 ⇒ a2,4 = γ114 = 0. 2 e0 e0 e1 1 2 e1 0 e2 1 2 e2 0 1 2 e4 e3 1 2 e3 − 1 0 0 2 e4 e4 e4 0 0 0 0 e0 e1 e2 e3 e4 4.3 Main results in general case Theorem 4.9 (Main theorem). Let A be dimensionally nilpotent Lie triple non nilalgebra. Let {e0, e1, . . . , en} be an adapted basis of A. Then: 1◦) If n = 2p + 1, the multiplication table of A is one of the two following: (i) e2 0 = e0, e0ei = 1 2 ei (1 ≤ i ≤ 2p + 1), all other product being zero. (i′) e2 0 = e0 + e2p+1, e0ei = 1 2 ei (1 ≤ i ≤ 2p + 1), all other product being zero. 2◦) If n = 2p, the multiplication table of A is one of the four following : (i) e2 0 = e0, e0ei = 1 2 ei (1 ≤ i ≤ 2p), all other product being zero. (i′) e2 0 = e0 + e2p, e0ei = 1 2 ei (1 ≤ i ≤ 2p), all other product being zero. (ii) e2 0 = e0, e0ei = 1 product being zero. 2 ei (1 ≤ i ≤ 2p − 1), e0e2p = 0, eie2p−i = 1 2 (−1)i−1e2p, (1 ≤ i ≤ p), all other (iii) e2 0 = e0, e0ei = 1 product being zero. 2 ei (1 ≤ i ≤ 2p − 1), e0e2p = e2p, eie2p−i = 1 2 (−1)ie2p (1 ≤ i ≤ p), all other Proof. Reason by recurrence on n. Subsection 4.2 shows that the theorem is true when n ≤ 4. Assume it is true until an order n > 4 and let's show it remains true for n + 1. Integer n being either even or odd, we consider two cases : 1) n = 2p is even. The multiplication table of A has the following form e0ek = 1 2 ek +ak,2p+1e2p+1 (1 ≤ k ≤ 2p − 1), e0e2p = λ2pe2p + a2p,2p+1e2p+1, e0e2p+1 = λ2p+1e2p+1 and eie2p−i = εie2p + 2 (−1)i (i = 1, . . . , p) according to λ2p = 1 γi,2p−i,2p+1e2p+1, with εi = 0, εi = 1 2 , λ2p = 0 or λ2p = 1, respectively. We have d(eie2p−i) = eie2p+1−i + ei+1e2p−i = εie2p+1, and then the following system 2 (−1)i−1 or εi = 1 (S) e1e2p + e2e2p−1 = ε1e2p+1, · · · · · · , eie2p+1−i + ei+1e2p−i = εie2p+1, · · · · · · , epep+1 + ep+1ep = 2epep+1 = εpe2p+1.   11 We see that epep+1 = 1 2 εp−ie2p+1, e1e2p = 2p−1 2i+1 e1e2p = (λ2p − λ2p+1)e2p+1, and also λ2p − λ2p+1 = 2p−1 2 εpe2p+1, ep−1ep+2 = (εp−1 − 1 2 εp−1e2p+1, ep−iep+1+i = 2 ε1e2p+1. But d(e0e2p) = e0e2p+1 + e1e2p = λ2pe2p+1, that means 2 εp)e2p+1 = 3 2 ε1. If λ2p = 0, we have ε1 = 1 2 and λ2p+1 = − 2p−1 4 6∈ {0, 1 2 , 1}, impossible. If λ2p = 1, we have ε1 = − 1 2 and λ2p+1 = 1 + 2p−1 4 = 2p+3 4 6∈ {0, 1 2 , 1}, impossible. Hence λ2p = 1 2 , ε1 = 0 and λ2p+1 = 1 2 . Since all the λk are equal to 1 2ak,2p+1λ2p+1(λ2p+1 − 1) = 0. If λ2p+1 = 1 all k. Hence eiej = 0, with 1 ≤ i ≤ j ≤ n. 2 (k 6= 2p + 1), applying 2L3 2 , then we have ak,2p+1 = 0, which means e0ek = 1 e0 + Le0 = 0 to ek, it follows that 2 ek for e0 − 3L2 2) n = 2p − 1 is odd. The multiplication table of A has the following form e0ek = 1 2 ek + ak,2pe2p (k = 1, . . . , 2p − 1), e0e2p = λ2pe2p and eie2p−1−i = γi,2p−1−i,2pe2p (i = 1, . . . , p − 1). Deriving this last relation, we have eie2p−i + ei+1e2p−1−i = 0, and the following system (S′ p) e1e2p−1 + e2e2p−2 = 0, e2e2p−2 + e3e2p−3 = 0, · · · · · · , ep−1ep+1 + e2 p = 0.   So e1e2p−1 = −e2e2p−2 = e3e2p−3 = · · · = (−1)i−1eie2p−i = · · · = (−1)p−1e2 (−1)i−1e1e2p−1. However, since d(e0e2p−1) = e0e2p+e1e2p−1 = 1 Because λ2p ∈ {0, 1 2 , 1}, we consider three situations : 2 e2p, we have e1e2p−1 = ( 1 p, that means eie2p−i = 2 −λ2p)e2p. If λ2p = 0, we have e1e2p−1 = 1 2 e2p, eie2p−i = 1 2 (−1)i−1e2p (i = 1, . . . , p) and e2 0 = e0. If λ2p = 1, we have e1e2p−1 = − 1 2 e2p, eie2p−i = 1 2 (−1)ie2p (i = 1, . . . , p) and e2 0 = e0. If λ2p = 1 2 , we have e1e2p−1 = eie2p−i = 0 (i = 1 . . . p) and 2ai,2pλ2p(λ2p − 1) = 0 shows 0 = e0, or 2 ei (i = 1, . . . , p). Furthermore we have, either e2 that ai,2p−i = 0. Hence e0ei = 1 e2 0 = e0 + e2p. For cases λ2p = 0 and λ2p = 1, We just need to show eiej = 0 for i + j < 2p. The following lemma completes the proof of the theorem. And Note 4.12 shows that all algebras defined in this theorem are Lie triple. ✷ Lemma 4.10. e0ei = 1 2 ei for i = 1, . . . , n − 1. Proof. One has eie2k−i = γi,2k−i,nen for i = 1, . . . , k − 1. Deriving this we have eie2k−i+1 + ei+1e2k−i = 0. By varying i we have the following system (Sk) e1e2k + e2e2k−1 = 0, e2e2k−1 + e3e2k−2 = 0, · · · · · · , ek−1ek+2 + ekek+1 = 0, ekek+1 + ek+1ek = 2ekek+1 = 0.   12 Going up the lines of this system we see that eie2k+1−i = 0 for i = 1, . . . , k in particular e1e2k = 0, so e0e2k+1 + e1e2k = 1 2 e2k+1 and e0e2k+1 = 1 2 e2k+1, k = 0, . . . , p − 1. Now we make a recurrence on n. Assume it is true until an order n. We distinguish two cases : • n + 1 = 2p + 1 is odd. We have e0en+1 = 1 2 en+1, which imposes e0en = 1 2 en. • n + 1 = 2p is even. Since n = 2p − 1 is odd, we have e0en = 1 2 en. ✷ Since every commutative Jordan algebra is a Lie triple algebra, we have the following result: Corollary 4.11. Let A be a commutative Jordan non nilalgebra, dimensionally nilpotent. Let {e0, e1, . . . , en} be an adapted basis of A. Then: 1◦) If n = 2p + 1, the multiplication table of A is: 0 = e0, e0ei = 1 e2 2 ei (1 ≤ i ≤ 2p + 1), all other product being zero. 2◦) If n = 2p, the multiplication table of A is one of the following three tables : (i) e2 0 = e0, e0ei = 1 2 ei (1 ≤ i ≤ 2p), all other product being zero. (ii) e2 0 = e0, e0ei = 1 product being zero. 2 ei (1 ≤ i ≤ 2p − 1), e0e2p = 0, eie2p−i = 1 2 (−1)i−1e2p, (1 ≤ i ≤ p), all other (iii) e2 0 = e0, e0ei = 1 product being zero. 2 ei (1 ≤ i ≤ 2p − 1), e0e2p = e2p, eie2p−i = 1 2 (−1)ie2p (1 ≤ i ≤ p), all other Proof. The Corollary follows from Theorem 4.9 knowing that Jordan algebras do not admit pseudo- idempotent. ✷ Note 4.12. 1) Multiplication tables in Theorem 4.9 1) (i) and in Corollary 4.11 1) when n = 2p + 1 is odd are those of gametic algebras G(2p + 2, 2) (Example 4.2). It is the same for those defined in Theorem 4.9 2) (i) and in Corollary 4.11 2) (i) when n = 2p is even. These are gametic algebras G(2p + 1, 2). They are caracterized as elementary train algebras with equation x2 − ω(x)x = 0, in which ω : A → K, e0 7→ 1, ei 7→ 0 is a homomorphism of algebras. 2) Multiplication tables in Theorem 4.9 2) (ii) and in Corollary 4.11 2) (ii), when n = 2p is even, are those of normal Bernstein algebras of type (2p, 1). Normal Bernstein algebras are defined by equation x2y = ω(x)xy. These are Bernstein-Jordan algebras, caracterized by the train equation x3 − ω(x)x2 = 0 [13, 15]. 3) Multiplication tables in Theorem 4.9 2) (iii) and in Corollary 4.11 2) (iii), when n = 2p is even, are those of the other class of train algebras of rank 3 which are Jordan algebras of type (2p, 1). They are defined by equation x3 − 2ω(x)x2 + ω(x)2x = 0 [13, Theorem 2.1]. 4) Multiplication tables in Theorem 4.9 1) (i′) and 2) (i′) are those of train algebras satisfying x3 − 3 2 ω(x)x2 + 1 2 ω(x)2x = 0. These are Lie triple algebras because of [2, Proposition 5.2, (iii)]. REFERENCES [1] J. Bayara, A. Conseibo, A. Micali and M. Ouattara, in power- Discrete Contin. Dyn. Syst., Ser. S, 4 6, 1359 -- 1370 (2011) Derivations associative algebras, https://doi.org/10.3934/dcdss.2011.4.1359 13 [2] J. Bayara, A. Konkobo and M. Ouattara, Algèbres de Lie triple sans idempotent, Afr. Mat., 25(4): 1063 -- 1075, (2014). doi:10.1007/s13370-013-0172-4 [3] J. Bayara, A. Konkobo and M. Ouattara, qui sont des algèbres doi:10.1016/j.indag.2016.10.004 train, Indagationes Mathematicae, 28(2): Equations des algèbres Lie triple (2017). 390 -- 405, [4] V. Eberlin, Algèbres de Lie dimensionnellement nilpotentes, Comm. Algebra, 28(1), 183 -- 191 (2000). doi:10.1080/00927870008826835 [5] I.R. Hentzel and L.A. Peresi, Almost Jordan rings, Proc. Am. Math. Soc. 104(2), 343- -- 348, (1988). doi:10.2307/2046977 [6] P. Holgate, The interpretation of derivations in genetic algebras, Linear Algebra Appl., 85, 75 -- 79 (1987). doi:10.1016/0024-3795(87)90209-6 [7] N. Jacobson, Lie algebras., Dover publications., Dover, 1979. [8] G.F. Leger and P.L. Manley, Dimensionally nilpotent Lie algebras, Journal of Algebra 117: 162 -- 164 (1988). doi:10.1016/0021-8693(88)90247-5 [9] A. Micali and M. Ouattara Algèbres génétiques dimensionnellement nilpotentes Linear Algebra Appl. 266: 271 -- 290 (1997). doi:10.1016/S0024-3795(97)86524-X [10] J.M. Osborn, Commutative algebras satisfying an identity of degree four, Trans. Amer. Math. Soc. 16 (1965), 1114 -- 1120. doi:10.2307/2035628 [11] J.M. Osborn, Varieties of algebras, Adv. Math. 8 (1972), 163 -- 369. doi:10.1016/0001-8708(72)90003-5 [12] J.M. Osborn, Dimensionally nilpotent Jordan algebras, Proc. Amer. Math. Soc., 116, 949 -- 953 (1992). doi:10.2307/2159472 [13] M. Ouattara, Sur les T-algèbres de Jordan, Linear Algebra Appl. 144 (1991), 11 -- 21. doi:10.1016/0024-3795(91)90056-3 [14] R.D. Schafer, An introduction to nonassociative algebras Academic Press, New York, 1966. [15] A. Wörz-Busekros, Algebras in Genetics, Lecture Notes in Biomathematics 36, Springer- Verlag, Berlin-New York, 1980. 14
1707.05083	1	1707	2017-07-17T10:39:00	Eigenvalues and Wiener index of the Zero Divisor graph $\Gamma[\mathbb {Z}_n]$	[ "math.RA", "math.AC", "math.CO" ]	The Zero divisor Graph of a commutative ring $R$, denoted by $\Gamma[R]$, is a graph whose vertices are non-zero zero divisors of $R$ and two vertices are adjacent if their product is zero. In this paper, we consider the zero divisor graph $\Gamma[\mathbb{Z}_n]$ for $n=p^3$ and $n=p^2q$ with $p$ and $q$ primes. We discuss the adjacency matrix and eigenvalues of the zero divisor graph $\Gamma[\mathbb{Z}_n]$. We also calculate the energy of the graph $\Gamma[\mathbb{Z}_n]$.	math.RA	math	Eigenvalues and Wiener index of the Zero Divisor graph Γ[Zn] B.Surendranath Reddy,Rupali.S.Jain and N.Laxmikanth [email protected],[email protected] and [email protected] Abstract The zero divisor graph of a commutative ring R, denoted by Γ[R], is a graph whose vertices are non-zero zero divisors of R and two vertices are adjacent if their product is zero. In this paper, we consider the zero divisor graph Γ[Zn] for n = p3 and n = p2q with p and q primes. We discuss the adjacency matrix and eigenvalues of the zero divisor graph Γ[Zn]. We also calculate the Wiener index of the graph Γ[Zn]. Keywords: zero divisor graph adjacency matrix eigenvalues energy Wiener index. AMS(2010): 13A99,05C12, 05C50. 1 Introduction The concept of the zero divisor graph of a ring R was first introduced by I.Beck [1] in 1988 where he considered the set of zero divisors including zero and introduced the concepts such as diameter, grith and clique number of a zero divisor graph. Then later on Anderson and Livingston in [2], Akbari and Mohammadian in [3] continued the study of zero divisor graph and they considered only the non-zero zero divisors. The concepts of the energy and Wiener index of a zero divisor graph was introduced by Mohammad Reza and Reza Jahani in [4], for the ring Γ[Zn] for n = p2 and n = pq where p and q are prime numbers. Motivated by the work of Mohammad Reza and Reza Jahani, in this article, we are extending the concepts of adjacency matrix, energy and Wiener index of the zero divisor graph to the rings Zn for n = p3 and p2q, where p,q are primes. In this article, section 2, is about the preliminaries and notations related to zero divisor graph of a commutative ring R, in section 3, we derive the 1 standard form of the adjacency matrix of the zero divisor graph Γ[Zn], in section 4 we discuss the eigenvalues of the graph Γ[Zn] and in section 5, we find the Wiener index of Γ[Zn] for n = p3 and p2q. 2 Preliminaries and Notations Definition 2.1. (Zero divisor Graph)[1] Let R be a commutative ring with unity and Z[R] be the set of its zero divi- sors. Then the zero divisor graph of R denoted by Γ[R], is the graph(undirected) with vertex set Z∗[R] = Z[R] − {0}, the non-zero zero divisors of R, such that two vertices v, w ∈ Z∗[R] are adjacent if vw = 0. Definition 2.2. (Adjacency matrix of Γ[R]) The adjacency matrix of the zero divisor graph Γ[R] is the matrix [vij] with rows and columns labeled by the vertices and is given by vij = (cid:26) 1, vi is adjacent to vj; 0, otherwise. Adjacency matrix of Γ[R] is denoted by M(Γ[R]). Clearly for an undirected graph, the adjacency matrix is symmetric. Definition 2.3. (Energy of Γ[R])[6] The energy of the zero divisor graph Γ[R] is defined as the sum of the ab- solute values of all the eigenvalues of its adjacency matrix M(Γ[R]) i.e., if λ1, λ2,· · · , λn are n eigenvalues of M(Γ[R]), then the energy of Γ[R] is E(Γ[R]) = n Pi=1λi. Definition 2.4. (Weiner index of Γ[R])[7] Let Γ[R] be a zero divisor graph with vertex set V . We denote the length of the shortest path between every pair of vertices x, y ∈ V with d(x,y). Then the Wiener index of Γ[R] is the sum of the distances between all pair of vertices of Γ[R], i.e., W (Γ[R]) = Px,y∈V 3 Adjacency Matrix of Γ[Zn] d(x, y). In this section, we derive the standard form of the adjacency matrix of Γ[Zn] for n = p3 and n = p2q, p and q being primes. To start with we con- sider p = 3 and n = 33. Then the set of non-zero zero divisors of Z27 is {3, 6, 9, 12, 15, 18, 21, 24}. Now by considering these non-zero zero divisors in 2 ascending order, we get the adjacency matrix as M(Γ[Z27]) =     0 0 1 0 0 1 0 0 0 0 1 0 0 1 0 0 1 1 1 1 1 1 1 1 0 0 1 0 0 1 0 0 0 0 1 0 0 1 0 0 1 1 1 1 1 1 1 1 0 0 1 0 0 1 0 0 0 0 1 0 0 1 0 0   0 0 0 0 0 0 1 1 0 0 0 0 0 0 1 1 0 0 0 0 0 0 1 1 0 0 0 0 0 0 1 1 0 0 0 0 0 0 1 1 0 0 0 0 0 0 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1   Instead of this, if we rearrange the non-zero zero divisors such that all the multiples of 3 appear first and then the multiples of 32, we get the adjacency matrix as M(Γ[Z27]) = =(cid:20) O R Rt S (cid:21) where O is the zero matrix of order 6 and R and S are matrices of ones of size 6 × 2 and 2 respectively. After getting the above interesting form of the adjacency matrix, we tried for several values of n = 8, 125, 343, etc and got the adjacency matrix in similar form and this leads us to derive the standard form for n = p3 in the following theorem. Theorem 3.1. Let n = p3 with p prime. Then the adjacency matrix of the graph Γ[Zn] is M(Γ[Zn]) = (cid:20) O(p2−p) Rt (p−1)×(p2−p) R(p2−p)×(p−1) S(p−1) (cid:21) where O is the zero matrix and R and S are the matrices of ones. Proof. Let n = p3. Then the set of non-zero zero divisors of Zn is Z∗[Zn] = {p, 2p, 3p,· · · , (p2 − 1)p} with cardinality p2 − 1. As explained in the example, we rewrite Z∗[Zn] = A ∪ B, where A = {kpk = 1, 2,· · · , p2 − 1 and p ∤ k} B = {lp2l = 1, 2, 3,· · · , p − 1} and the adjacency matrix is labeled with elements of A first then the elements of B. 3 It is clear that if x, y ∈ A then xy 6= 0. Since A = p2 − p, we get a zero matrix O of size p2 − p in the adjacency matrix. And if x ∈ A, y ∈ B or y ∈ A, x ∈ B then xy = 0. Since A = p2 − p and B = p − 1, we get a matrix R of ones of size (p2 − p) × (p − 1). Also if x, y ∈ B, then xy = 0 and hence we get a matrix S of ones of size p − 1 in the adjacency matrix. Hence the adjacency matrix by considering the elements of A first and then the elements of B, we get M(Γ[Zn]) = (cid:20) O(p2−p) Rt (p−1)×(p2−p) R(p2−p)×(p−1) S(p−1) (cid:21). Now we consider the case for n = p2q with p and q different primes and get the standard form of adjacency matrix in the following theorem. Theorem 3.2. Let n = p2q, where p,q are primes. Then the adjacency matrix of the graph Γ[Zn] is O O R(p−1)(q−1)×(p−1) O O Rt O O St O Tp−1 U t O Sp(p−1)×(q−1) U(p−1)×(q−1) O     M(Γ[Zn]) = where R, S, T and U are matrices of ones. Proof. For n = p2q, the non-zero zero divisors are multiples of p, multiples of q, multiples of p2 and multiples of pq. Hence Z∗[Zn] = A ∪ B ∪ C ∪ D, where A = {k1p k1 = 1, 2,· · · , pq − 1 and k1 ∤ p, q} B = {k2q k2 = 1, 2,· · · , p2 − 1 and k2 ∤ p} C = {k3pq k3 = 1, 2,· · · , p − 1 } D = {k4p2 k4 = 1, 2,· · · , q − 1} Then A = (p − 1)(q − 1), B = p(p − 1), C = p − 1 and D = q − 1. Since no element of A is adjacent with an element of A and B, we get the zero matrix of order (p − 1)(p + q − 1). Similarly we get zero matrices corresponding to A and D; B and C; and D and D. And as every element of A is adjacent with every element of C implies we get a matrix R of ones of order (p − 1)(q − 1) × (p − 1). Similarly we get the matrix S of ones corresponding to B and D; the matrix 4 T of ones corresponding to C and C; and the matrix U of ones corresponding to C and D. Hence we get the adjacency matrix by considering the elements of A first, then B, then C and then D as M(Γ[Zn]) = O O R(p−1)(q−1)×(p−1) O O Rt O O St O Tp−1 U t O Sp(p−1)×(q−1) U(p−1)×(q−1) O     4 Eigenvalues of the zero divisor graph Γ[Zn] In this section we discuss the non zero eigenvalues of the zero divisor graph Γ[Zn] for n = p3 and n = p2q. Theorem 4.1. Let n = p3. Then the only non-zero eigenvalues of the zero divisor graph Γ(Zn) are given by (p−1)(1±√1+4p) . 2 Proof. By theorem 3.1, the adjacency matrix of Γ(Zp3)) is given by M(Γ[Zp3]) = (cid:20) O(p2−p) Rt (p−1)×(p2−p) R(p2−p)×(p−1) S(p−1) (cid:21) where O is the zero matrix and R and S are the matrices of ones. Let λ be a non-zero eigenvalue of M(Γ[Zp3]). 0 · · · −λ 0 0 −λ · · · 0 ... ... . . . 0 · · · −λ 1 1 · · · 1 · · · 1 ... ... ... 1 1 · · · A B ... 0 1 1 ... 1 (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) 1 1 ... 1 1 − λ 1 ... 1 1 1 ... 1 1 · · · · · · ... · · · · · · 1 − λ · · · . . . · · · 1 − λ 1 1 ... 1 1 1 ... ... 1 = 0 (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) Then M − λI = C D(cid:12)(cid:12)(cid:12)(cid:12) Let M − λI = (cid:12)(cid:12)(cid:12)(cid:12) (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) Then M − λI = AD − CA−1B = 0. Since A is a scalar matrix of order p2 − p, we get A = (−λ)p2−p. p−p2 λ p−p2 λ ... p−p2 λ And CA−1B = p−p2 λ p−p2 λ ... p−p2 λ p−p2 λ p−p2 λ ... p−p2 λ · · · · · · . . . · · · (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) 5 So D − CA−1B = (1 − λ) + p−p2 1 + p−p2 λ λ ... 1 + p−p2 λ (1 − λ) + p−p2 λ ... (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) · · · · · · . . . · · · 1 + p−p2 1 + p−p2 λ λ ... (1 − λ) + p−p2 λ (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) 1 + p−p2 Using the properties of the determinant, we get 1 + p−p2 λ λ D − CA−1B = (−1)p−1 × λp−3[λ2 − (p − 1)λ − (p2 − p)(p − 1)] Therefore, M − λI = λp2−p(−1)p−1× λp−3[λ2−(p−1)λ−(p2− p)(p−1)] = 0 Since λ 6= 0, we have λ2 − (p − 1)λ − (p2 − p)(p − 1) = 0 Hence, λ = = = (p − 1) ±q(p − 12) + 4(p2 − p)(p − 1) (p − 1) ± (p − 1)√1 + 4p (p − 1)(1 ± √1 + 4p) 2 2 2 Theorem 4.2. If p is a prime number then the Energy of the zero divisor graph Γ(Zp3) is E(Γ(Zp3)) = (p − 1)(√1 + 4p). Proof. Let n = p3. Then by theorem 4.1, the non-zero eigenvalues of Γ(Zn)) are given by λ1 = λ2 = (p − 1)(1 + √1 + 4p) (p − 1)(1 − √1 + 4p) 2 2 , Thus the Energy, E(Γ(Zp3)) = λ1 + λ2 (p − 1)(1 − √1 + 4p) 2 (p − 1)(1 + √1 + 4p) 2 (cid:17)(cid:16)p1 + 4p + 1 +p1 + 4p − 1(cid:17) = (cid:12)(cid:12)(cid:12) = (cid:16) p − 1 = (p − 1)(p1 + 4p) +(cid:12)(cid:12)(cid:12) (cid:12)(cid:12)(cid:12) 2 (cid:12)(cid:12)(cid:12) Theorem 4.3. Let n = p2q with p and q primes. If λ 6= 0 is a nonzero eigenvalue of the zero divisor graph Γ(Zn), then λ4− (p− 1)λ3 − 2p(q− 1)2(p− 1)λ2 + p(p− 1)2(q− 1)λ + p(p− 1)3(q− 1)2 = 0. 6 Proof. By theorem 3.2, the adjacency matrix is given by O O R(p−1)(q−1)×(p−1) O O Rt O O St O Tp−1 U t O Sp(p−1)×(q−1) U(p−1)×(q−1) O   where R, S, T and U are matrices of ones. Let λ be a nonzero eigenvalue of M(Γ[Zp2q]). M(Γ[Zn]) =   Then M − λI = (cid:12)(cid:12)(cid:12)(cid:12) where A = = 0 A B C D (cid:12)(cid:12)(cid:12)(cid:12)   0 0 · · · −λ 0 0 −λ · · · ... ... . . . · · · −λ 0 ... 0 ; B = C t =     ... ... 1 · · · 1 0 · · · 0 ... ... 1 · · · 1 0 · · · 0 0 · · · 0 1 · · · 1 ... ... 0 · · · 0 1 · · · 1 ... ...   and D = 1 − λ 1 ... 1 1 1 ... 1   1 ... 1 1 1 ... · · · 1 − λ · · · . . . · · · 1 − λ · · · · · · 1 1 ... 1 · · · · · · · · · 1 1 1 1 ... ... 1 · · · 1 0 · · · −λ 0 0 −λ · · · ... ... . . . · · · −λ 0 ... 0 0   (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) Clearly A = λpq+p2−p−5 · · · (1−p)(q−1) λ (1−p)(q−1) CA−1B = · · · · · · · · · and hence D − CA−1B = ... λ 0 ... 0 (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) 0 ... 0 p(1−p) λ ... p(1−p) λ (1−p)(q−1) λ ... (1−p)(q−1) λ 0 ... 0 · · · · · · · · · · · · 0 ... 0 p(1−p) λ ... p(1−p) λ 7 1 − λ + (1−p)(q−1) λ ... · · · (1−p)(q−1) λ ... (1−p)(q−1) λ 1 ... 1 · · · 1 − λ + (1−p)(q−1) · · · λ 1 ... 1 · · · (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) 1 ... 1 −λ + p(1−p) λ ... p(1−p) λ · · · · · · · · · 1 ... 1 p(1−p) λ ... · · · −λ + p(1−p) λ (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) = λ4 − (p − 1)λ3 − 2p(p − 1)(q − 1)λ2 + p(p − 1)2(q − 1)λ + p(p − 1)3(q − 1)2. Since λ 6= 0, A = λpq+p2−p−5 and AD − CA−1B = 0, we get λ4 − (p− 1)λ3 − 2p(p− 1)(q − 1)λ2 + p(p− 1)2(q − 1)λ + p(p− 1)3(q − 1)2 = 0. 5 Wiener index of the zero divisor graph Γ(Zn) In this section we calculate the Wiener index of Γ(Zn) for n = p3 and n = p2q. Theorem 5.1. The Weiner index of the zero divisor graph Γ(Zp3) is ( p−1 2 )(2p3 − 3p − 2). Proof. Let Γ(Zp3) be a zero divisor graph where p is a prime number. Then the vertex set of Γ(Zp3) is divided into two sets A = {k1p k1 = 1, 2, 3, ..., p2 − 1 and k1 ∤ p} B = {k2p2 k2 = 1, 2, 3, ..., p − 1} where A = p2 − p = n(say) and B = p − 1 = m(say) Let x, y ∈ V (Γ(Zp3)) Then either x, y ∈ A or x, y ∈ B or x ∈ A, y ∈ B Suppose x, y ∈ A . Since no two elements of A are adjacent and every element of A is adjacent with an element of B, we get d(x, y) = 2 Suppose x, y ∈ B . Since any two elements of B are adjacent to each other,we get d(x, y) = 1 Suppose x ∈ A, y ∈ B. Since every element of A is adjacent with an element of B ,we get d(x, y) = 1 Thus n n Xx,y∈A d(x, y) = d(x1, xi) + Xi=2 = 2(n − 1) + 2(n − 2) + ... + 2 = n(n − 1) d(x2, xi) + ... + d(xn−1, xn) Xi=3 8 Similarly Px,y∈A And d(x, y) = (m − 1) + (m − 2) + ... + 1 = m(m−1) 2 d(x, y) = Xx∈A,y∈B m Xi=1 d(x1, yi) + m Xi=1 d(x2, yi) + ... + m Xi=1 d(xn, yi) = m + m + ... + m(n times) = nm Therefore,the Weiner index is W (Γ(Zp3)) = Xx∈A,y∈B d(x, y) + Xx,y∈A d(x, y) + Xx,y∈B d(x, y) 2 (cid:17) = nm + n(n − 1) +(cid:16) m(m − 1) = (p2 − p)(p − 1) + (p2 − p)(p − 2) + = (cid:16) p − 1 2 (cid:17)(2p3 − 3p − 2) (p − 1)(p − 2) 2 Theorem 5.2. The Wiener index of the zero divisor graph Γ(Zp2q) is ( 1 2 )[p2(2p2 − 4p − 1) + p(4p2q + 2pq2 − 8pq − 2q + 5) + 2] Proof. Let Γ(Zp2q) be a zero divisor graph where p ,q are primes. Then the vertex set of Γ(Zp2q) is divided into four sets A = {k1p k1 = 1, 2, 3, ..., pq − 1 and k1 ∤ p, q} B = {k2q k2 = 1, 2, 3, ..., p2 − 1 and k2 ∤ p} C = {k3pq k3 = 1, 2, 3, ..., p − 1 } D = {k4p2 k4 = 1, 2, 3, ..., q − 1} Then α = A = (p − 1)(q − 1), β = B = p(p − 1) , γ = C = (q − 1) and δ = D = (p − 1) For x, y ∈ V (Γ(Zp2q)), we have the following cases: 1. Suppose x, y ∈ A. Since no two elements of A are adjacent and every element of A is adjacent with an element of D, we get d(x, y) = 2 2. If x, y ∈ B, since no two elements of B are adjacent and every element of B is adjacent with an element of C,we get d(x, y) = 2 9 3. If x, y ∈ C, as no two elements of C are adjacent and every element of C is adjacent with an element of either B or D ,we get d(x, y) = 2 4. If x, y ∈ D, being any two elements of D are adjacent, we get d(x, y) = 1 5. Now suppose x ∈ A, y ∈ B, as no element of A is adjacent with an element of B, we get d(x, y) = 2 With similar arguments we get, 6. If x ∈ A, y ∈ C, d(x, y) = 2 7. If x ∈ A, y ∈ D, d(x, y) = 1 8. If x ∈ B, y ∈ C, d(x, y) = 1 9. If x ∈ B, y ∈ D, d(x, y) = 2 10. If x ∈ C, y ∈ D, d(x, y) = 1 Hence the Weiner index is W (Γ(Zp2q) = Xx,y∈{same sets} d(x, y) + Xx,y∈{different sets} d(x, y) d(x1, xi) +Xi=3 d(x1, yi) +Xi=1 d(x2, xi) + ... + d(xn−1, xn) d(x2, yi) + ... +Xi=1 = Xi=2 +Xi=1 = 2[(α − 1) + (α − 2) + ... + 2 + 1] + 2[(β − 1) + (β − 2) + ... + 2 + 1] + 2[(γ − 1) + (γ − 2) + ... + 2 + 1] + 1[(δ − 1) + (δ − 2) + ... + 2 + 1] + 2[αβ] + 2[αγ] + [αδ] + [βγ] + 2[βδ] + [γδ]. δ(δ − 1) d(xn, yi) = α(α − 1) + β(β − 1) + γ(γ − 1) + + 2αβ + 2αγ + αδ + βγ + 2βδ + γδ 2 By substituting the values of α, β, γ, δ in above and on solving, we get W (Γ(Zp2q) = ( 1 2)[p2(2p2 − 4p − 1) + p(4p2q + 2pq2 − 8pq − 2q + 5) + 2]. References [1] I.Beck, Coloring of Commutative rings, J.Algebra, 116(1), 208-226 (1988). 10 [2] D.F.Anderson and P.S.Livingston, The Zero divisor graph of Commuta- tive ring, J.Algebra, 217(2), 434-447 (1999). [3] S.Akbari, A.Mohammadian, On the Zero divisor graph of a commutative rings, J.Algebra, 274, 847-855 (2004). [4] Mohammad Reza Ahmadi and Reza Jahani-Nezhad, Energy and Weiner Index of Zero divisor graphs, Iranian Journal of Mathematical Chemistry, 2(1), 45-51 (2011). [5] I.Gutman, The Energy of a graph, Ber.Math-Stat.Sekt.Forschungszent Graz, 103, 1-22 (1978). [6] I.Gutman, S.Kaivzar, B.Mohar(Eds), Fifty Years of Weiner Index, 1-259, Match Commun.Math.Comput.Chem (1997). [7] I.Gutman, The Energy of a graph:old and new results, Algebraic Combi- natorics and Applications,,196-211, Springer-verlag.Berlin(2001). 11
1004.3650	2	1004	2010-11-02T01:32:51	Lie bialgebra structures on extended Schrodinger-Virasoro Lie algebra	[ "math.RA", "math.QA" ]	In this paper, Lie bialgebra structures on the extended Schrodinger-Virasoro Lie algebra are classified. It is obtained that all the Lie bialgebra structures on L are triangular coboundary. As a by-product, it is derived that the first cohomology group is trivial.	math.RA	math	Lie bialgebra structures on extended Schrodinger-Virasoro Lie algebra1 Lamei Yuan ∗, Yongping Wu †, Ying Xu ∗ ∗Department of Mathematics, University of Science and Technology of China, Hefei 230026, China †School of Mathematics and Computer Science, Longyan College, Fujian 364111, China E-mail: [email protected], wyp [email protected], [email protected] Abstract. In this paper, Lie bialgebra structures on the extended Schrodinger-Virasoro Lie algebra L are classified. It is obtained that all the Lie bialgebra structures on L are triangular coboundary. As a by-product, it is derived that the first cohomology group H 1(L, L ⊗ L) is trivial. Key words: Lie bialgebras, Yang-Baxter equation, Extended Schrodinger-Virasoro Lie algebras. 0 1 0 2 v o N 2 ] . A R h t a m [ 2 v 0 5 6 3 . 4 0 0 1 : v i X r a Mathematics Subject Classification : 17B05, 17B37, 17B62, 17B68. §1. Introduction The Schrodinger-Virasoro Lie algebras [6] were introduced in the context of non-equilibrium statistical physics during the process of investigating the free Schrodinger equations. They are closely related to the Schrodinger algebra and the Virasoro algebra, both of which play important roles in many areas of mathematics and physics (e.g., statistical physics, integrable system) and have been investigated in a series of papers [7 -- 10, 13, 18, 21, 25]. In order to investigate vertex representations of the Schrodinger-Virasoro Lie algebra, J. Unterberger introducted (see Definition 1.5 in [22]) a class of infinite-dimensional Lie algebras called the extended Schrodinger-Virasoro Lie algebra L, which can be viewed as an extension of the Schrodinger-Virasoro Lie algebra by a conformal current of weight 1 and generated by {Ln, Mn, Nn, Yp n ∈ Z, p ∈ Z + 1/2} with the following Lie brackets: [Lm, Ln] = (n − m)Ln+m, [Ln, Yp ] = (p − n/2)Yp+n, [Mn, Yp ] = [Nm, Nn] = 0, [Lm, Nn] = nNm+n, [Nm, Yp ] = Ym+p, [Mm, Mn] = 0, [Lm, Mn ] = nMn+m, [Nm, Mn] = 2Mm+n, [ Yp, Yq ] = (q − p)Mp+q. (1.1) Note that L is centerless and finitely generated with a generating set {L−2, L−1, L1, L2, N1, Y1/2}. Moreover, it is 1 2 Z -graded by L = Ln∈Z Ln/2 = (cid:0) Ln∈Z Ln(cid:1)L(cid:0) Ln∈Z Ln+1/2(cid:1), where Ln = span{Ln, Mn, Nn} and Ln+1/2 = span{Yn+1/2}, for all n ∈ Z. The derivations, central extensions and automorphisms of L have been studied in [5]. To search for the solutions of the Yang-Baxter quantum equation, Drinfel'd [1] introduced the notion of Lie bialgebras in 1983. Since then, a number of people have studied further Lie bialgebra structures (e.g., [3, 11, 12, 14 -- 16). Witt type Lie bialgebras introduced in [20] were classified in [17]. This work has been generalized in [19, 23]. Lie bialgebra structures on generalized Virasoro-like and 1Supported by NSF grants 10825101 of China China ∗Corresponding author: [email protected] 1 Block Lie algebras were investigated in [11, 24]. Drinfel'd [2] posed the problem that whether or not there exists a general way to quantilize all Lie bialgebras. Etingof and Kazhdan [3] gave a positive answer to this problem, but there does not exist an uniform method to realize quantilizations of all Lie bialgebras. Actually, investigating Lie bialgebras and quantilizations is a complicated problem. The authors in [8] prove that not all Lie bialgebra structures on the Schrodinger-Virasoro Lie algebra are triangular coboundary. For the extended Schrodinger-Virasoro Lie algebra L, this is not the case. Namely, we obtain that all Lie bialgebra structures on L are triangular coboundary. In particular, we derive that the first cohomology group H 1(L, L ⊗ L) is trivial. §2. Preliminaries Throughout the paper, F denotes a filed with characteristic zero. All vector spaces and tensor products are over F. Let Z+ (resp. Z>0) be the set of all nonnegative (resp. positive) integers and Z∗ be the set of all nonzero elements of Z. Let L be a vector space, ξ the cyclic map of L ⊗ L ⊗ L, namely, ξ(x1 ⊗ x2 ⊗ x3) = x2 ⊗ x3 ⊗ x1 for x1, x2, x3 ∈ L, and τ the twist map of L ⊗ L, i.e., τ (x ⊗ y) = y ⊗ x for x, y ∈ L. A Lie algebra is a pair (L, δ), where δ : L ⊗ L → L is a bilinear map with the conditions: Ker( Id − τ ) ⊂ Ker δ, δ · ( Id ⊗ δ) · ( Id + ξ + ξ2) = 0 : L ⊗ L ⊗ L → L, where Id is the identity map. Dually, a Lie coalgebra is a pair (L, ∆) with a linear map ∆ : L → L⊗L satisfying: Im ∆ ⊂ Im( Id − τ ), ( Id + ξ + ξ2) · ( Id ⊗ ∆) · ∆ = 0 : L → L ⊗ L ⊗ L. (2.1) A Lie bialgebra is a triple (L, δ, ∆) such that (L, δ) is a Lie algebra, (L, ∆) is a Lie coalgebra, and the following compatible condition holds: ∆δ(x ⊗ y) = x · ∆y − y · ∆x, ∀ x, y ∈ L. (2.2) where "·" means the diagonal adjoint action, i.e., x · (Piai ⊗ bi) = Pi([x, ai] ⊗ bi + ai ⊗ [x, bi]), and in general, δ(x ⊗ y) = [x, y], for all x, y, ai, bi ∈ L. Denote by U the universal enveloping algebra of L and 1 the identity element of U . For any r = Pi ai ⊗ bi ∈ L ⊗ L, define c(r) ∈ U ⊗ U ⊗ U by c(r) = [r12, r13] + [r12, r23] + [r13, r23], where r12 = Pi ai ⊗ bi ⊗ 1, r13 = Pi ai ⊗ 1 ⊗ bi, r23 = Pi 1 ⊗ ai ⊗ bi. Obviously, c(r) = Pi,j [ai, aj] ⊗ bi ⊗ bj + Pi,j ai ⊗ [bi, aj] ⊗ bj + Pi,j ai ⊗ aj ⊗ [bi, bj]. Definition 2.1 (1) A coboundary Lie bialgebra (L, δ, ∆, r) is a Lie bialgebra such that the cobracket ∆ is an inner derivation, i.e., there exists an element r ∈ L ⊗ L such that ∆(x) = x · r for all x ∈ L. (2.3) 2 ∆ is called a coboundary of r, denoted by ∆r. (2) A coboundary Lie bialgebra (L, δ, ∆, r) is called triangular if it satisfies the following classical Yang-Baxter Equation (CYBE): c(r) = 0. (2.4) (3) An element r ∈ Im( Id − τ ) ⊂ L ⊗ L is said to satisfy the modified Yang-Baxter equation (MYBE) if x · c(r) = 0, ∀ x ∈ L. (2.5) The following famous results are due to Drinfel'd [2], Michaelis [14] and Taft [20], respectively. We combine them into one theorem as follows: Theorem 2.2 (i) For a Lie algebra (L, [·, ·]) and r ∈ Im( Id − τ ) ⊂ L ⊗ L, the triple (L, [·, ·], ∆r) is a Lie bialgebra if and only if r satisfies MYBE. (ii) Let L be a Lie algebra containing two linear independent elements a, b satisfying [a, b] = kb for some nonzero k ∈ F, and set r = a ⊗ b − b ⊗ a. Then r is a solution of CYBE and equips L with a structure of triangular coboundary Lie bialgebra. (iii) Let L be a Lie algebra and r ∈ Im( Id − τ ) ⊂ L ⊗ L. Then for any x ∈ L, ( Id + ξ + ξ2) · (1 + ∆r) · ∆r(x) = x · c(r). §3. Structures of Lie bialgebra of the extended Schrodinger-Virasoro Lie algebra Regard V = L ⊗ L as a L-module under the adjoint diagonal action. A linear map D : L → V is called a derivation if D([x, y]) = x · D(y) − y · D(x) for all x, y ∈ L. (3.1) If there exists some v ∈ V such that D(x) = x · v, then D is called an inner derivation. Denote by vinn the inner derivation determined by v. Let Der(L, V) (resp. Inn(L, V)) be the set of all derivations (resp. inner derivations). Then it is well known that H 1(L, V) ∼= Der(L, V)/Inn(L, V), where H 1(L, V) is the first cohomology group of the Lie algebra L with coefficients in V. A derivation D ∈ Der(L, V) is homogeneous of degree α ∈ 1 2 Z if D(Lp) ⊂ Vα+p for all p ∈ 1 2 Z. Z, define a linear map Dα : L → V as follows: Let Der(L, V)α be the set of all the homogeneous derivations of degree α. For any D ∈ Der(L, V) and α ∈ 1 2 D(µ) = Pp∈ 1 have Z µp with µp ∈ Vp, then we set Dα(µ) = µq+α. Obviously, Dα ∈ Der(L, V)α and we 2 for any µ ∈ Lq with q ∈ 1 2 Z, write D = Pα∈ 1 2 Z Dα, (3.2) which holds in the sense that only finitely many Dα(u) 6= 0 and D(u) = Pα∈ 1 u ∈ L. Actually, for any D ∈ Der(L, V), (3.2) is a finite sum, referring to [8] for details. 2 Z Dα(u) for any 3 Lemma 3.1 H 1(L0, Vn/2) = 0 for all n ∈ Z∗. Proof. For any D ∈ Der(L, V), we have D = Pn∈Z Dn/2. Suppose n 6= 0, then the restriction of Dn/2 to L0 induces a derivation from L0 to the L0-module Vn/2. That is, Dn/2L0 ∈ Der(L0, Vn/2). Conveniently, we denote Dn/2L0 by Dn/2. Let r = 2 n Dn/2(L0) ∈ Vn/2. For any X0 ∈ L0, one has n 2 Dn/2(X0) = L0 · Dn/2(X0) = X0 · Dn/2(L0), since [L0, X0] = 0. It follows Dn/2(X0) = X0 · r, which implies Dn/2 is inner. (cid:3) Lemma 3.2 HomL0(Vm/2, Vn/2) = 0 for all m 6= n. Proof. Let f ∈ HomL0(Vm/2, Vn/2) = 0 with m 6= n. One has f ([X0, Em/2]) = [X0, f (Em/2)] for any X0 ∈ L0 and Em/2 ∈ Vm/2. In particular, f ([L0, Em/2]) = [L0, f (Em/2)]. That is, m 2 f (Em/2) = n 2 f (Em/2). It follows f (Em/2) = 0, since m 6= n. Consequently, f = 0. (cid:3) Taking these two Lemmas above into account, we can immediately derive the following result from Proposition 1.2 in [4]. Proposition 3.3 Der(L, V) = Der0(L, V) + Inn(L, V). Lemma 3.4 Let L⊗n = L ⊗ · · · ⊗ L be the tensor product of n copies of L, and regard L⊗n as an L-module under the adjoint diagonal action. Suppose r ∈ L⊗n satisfying x · r = 0, ∀ x ∈ L. Then r = 0. Proof. It is easy to see that L⊗n is 1 2 Z-graded by L⊗n p = Pp1+p2+···+pn=p Lp1 ⊗ Lp2 ⊗ · · · ⊗ Lpn, ∀ p, pi ∈ 1 2 Z, i = 1, 2, · · · n. Z rp as a finite sum with rp ∈ L⊗n p . By hypothesis, L0 · r = 0, which implies r = r0. Write r = Pp∈ 1 So r = 2 Pr1+r2+···+rn=0 cr1,r2,···rnEr1 ⊗ Er2 ⊗ · · · Ern for some cr1,r2,···rn ∈ F and Eri ∈ Lri with ri ∈ 1 2 Z. Since M0 · r = 0 by the assumption, all the coefficients of the terms containing Nj for j ∈ Z are zero, hence these terms in the sum vanish. Similarly, by N0 · r = 0, one can kill the coefficients of the terms containing Mj and Yj+1/2 with j ∈ Z. Now we can rewrite r by r = Pr1+r2+···+rn=0 cr1,r2,···rnLr1 ⊗ Lr2 ⊗ · · · Lrn for some cr1,r2,···rn ∈ F. But M1 · r = 0 forces all the coefficients cr1,r2,···rn are zero. This proves the lemma. (cid:3) Theorem 3.5 Der(L, V) = Inn(L, V). Proof. It suffices to show Der0(L, V) ⊆ Inn(L, V) by virtue of Proposition 3.3. For any 0 6= D ∈ Der0(L, V), we shall prove that the zero derivation is obtained after a number of steps in each of which D is replaced by D − uinn for some u ∈ V0. This will be done by a little bit complicated calculations. For clarity, we divide them into three claims. Claim 1. D(L0) = 0. 4 In fact, for any Xp ∈ L with p ∈ 1 2 Z, applying D to [L0, Xp] = pXp, one has Xp · D(L0) = 0. Then it follows from Lemma 3.4 that D(L0) = 0. Claim 2. D(L±1) = 0. For any n ∈ Z, one can write D(Ln), D(Mn), D(Nn) and D(Yn−1/2) as follows: D(Ln) = Pi∈Z +a(n) (a(n) 1,i Li⊗Ln−i+a(n) 2,i Li ⊗Mn−i+a(n) 3,i Mi ⊗Ln−i+a(n) 4,i Li⊗Nn−i+a(n) 5,i Ni ⊗Ln−i 6,iMi⊗Mn−i+a(n) 7,i Mi⊗Nn−i+a(n) 8,i Ni ⊗Mn−i+a(n) 9,i Ni⊗Nn−i+a(n) 10,iYi−1/2 ⊗Yn−i+1/2), D(Mn) = Pi∈Z (b(n) 1,i Li⊗Ln−i+b(n) 2,i Li⊗Mn−i+b(n) 3,i Mi ⊗Ln−i+b(n) 4,i Li⊗Nn−i+b(n) 5,i Ni ⊗Ln−i +b(n) 6,iMi ⊗Mn−i+b(n) 7,i Mi ⊗Nn−i+b(n) 8,i Ni⊗Mn−i+b(n) 9,i Ni⊗Nn−i+b(n) 10,iYi−1/2 ⊗Yn−i+1/2), D(Nn) = Pi∈Z (d(n) 1,i Li⊗Ln−i+d(n) 2,i Li⊗Mn−i+d(n) 3,i Mi⊗Ln−i+d(n) 4,i Li⊗Nn−i+d(n) 5,i Ni⊗Ln−i +d(n) 6,iMi⊗Mn−i+d(n) 7,iMi⊗Nn−i+d(n) 8,i Ni⊗Mn−i+d(n) 9,i Ni⊗Nn−i+d(n) 10,iYi−1/2 ⊗Yn−i+1/2), D(Yn−1/2) = Pi∈Z (αn,iLi⊗Yn−1/2−i+α† n,iYi−1/2 ⊗Ln−i+βn,iMi⊗Yn−1/2−i+β† n,iYi−1/2 ⊗Mn−i +γn,iNi ⊗ Yn−1/2−i + γ† n,iYi−1/2 ⊗ Nn−i). Note that all the sums are finite. For any n ∈ Z, one can easily get the following identities by (1.1): L1 · (Nn ⊗ N−n) = nNn+1 ⊗ N−n − nNn ⊗ N1−n, L1 · (Mn ⊗ N−n) = nMn+1 ⊗ N−n − nMn ⊗ N1−n, L1 · (Nn ⊗ M−n) = nNn+1 ⊗ M−n − nNn ⊗ M1−n, L1 · (Mn ⊗ M−n) = nMn+1 ⊗ M−n − nMn ⊗ M1−n, L1 · (Ln ⊗ N−n) = (n − 1)Ln+1 ⊗ N−n − nLn ⊗ N1−n, L1 · (Nn ⊗ L−n) = nNn+1 ⊗ L−n − (1 + n)Nn ⊗ L1−n, L1 · (Ln ⊗ M−n) = (n − 1)Ln+1 ⊗ M−n − nLn ⊗ M1−n, L1 · (Mn ⊗ L−n) = nMn+1 ⊗ L−n − (1 + n)Mn ⊗ L1−n, L1 · (Ln ⊗ L−n) = (n − 1)Ln+1 ⊗ L−n − (1 + n)Ln ⊗ L1−n, L1 · (Yn−1/2 ⊗ Y1/2−n) = (n − 1)Yn+1/2 ⊗ Y1/2−n − nYn−1/2 ⊗ Y3/2−n. Let Qi = max{p a(1) i,p 6= 0} for i = 1,· · · , 10. Applying D − uinn to L1, where u is a proper linear combination of Lp ⊗ L−p, Lp ⊗ M−p, Mp ⊗ L−p, Mp ⊗ M−p, Lp ⊗ N−p, Np ⊗ L−p, Mp ⊗ N−p, Np ⊗ M−p, Np ⊗ N−p and Yp−1/2 ⊗ Y1/2−p with p ∈ Z, and using induction on P10 i=1Qi, one can safely suppose a(1) 1,i = a(1) 2,k = a(1) a(1) 6,j = a(1) 4,k = a(1) 7,j = a(1) 10,k = a(1) 8,j = a(1) 3,n = a(1) 9,j = 0, 5,n = 0, for i 6= −1, 2, j 6= 0, 1, for k 6= 0, 2, n 6= ±1. (3.3) (3.4) 5 Applying D to [ L1, L−1] = −2L0 and using D(L0) = 0, we have (cid:16)(cid:0)(p − 2)a(−1) Pp∈Z + (cid:0)(p − 2)a(−1) + (cid:0)(p − 1)a(−1) + (cid:0)(p − 2)a(−1) + (cid:0)(p − 1)a(−1) + (cid:0)(p − 1)a(−1) + (cid:0)(p − 1)a(−1) + (cid:0)(p − 1)a(−1) + (cid:0)(p − 1)a(−1) + (cid:0)(p − 2)a(−1) 1,p−1 − (p + 2)a(−1) 2,p−1 − (p + 1)a(−1) 3,p−1 − (p + 2)a(−1) 4,p−1 − (p + 1)a(−1) 5,p−1 − (p + 2)a(−1) 6,p−1 − (p + 1)a(−1) 7,p−1 − (p + 1)a(−1) 8,p−1 − (p + 1)a(−1) 9,p−1 − (p + 1)a(−1) 10,p−1 − (p + 1)a(−1) 1,p + (p − 2)a(1) 2,p + (p − 1)a(1) 3,p + (p − 2)a(1) 4,p + (p − 1)a(1) 5,p + (p − 2)a(1) 6,p + (p − 1)a(1) 7,p + (p − 1)a(1) 8,p + (p − 1)a(1) 9,p + (p − 1)a(1) 10,p + (p − 2)a(1) 1,p − (p + 2)a(1) 2,p − (p + 2)a(1) 3,p − (p + 1)a(1) 4,p − (p + 2)a(1) 5,p − (p + 1)a(1) 6,p − (p + 1)a(1) 7,p − (p + 1)a(1) 8,p − (p + 1)a(1) 9,p − (p + 1)a(1) 10,p − (p + 1)a(1) 1,p+1(cid:1)Lp ⊗ L−p 2,p+1(cid:1)Lp ⊗ M−p 3,p+1(cid:1)Mp ⊗ L−p 4,p+1(cid:1)Lp ⊗ N−p 5,p+1(cid:1)Np ⊗ L−p 6,p+1(cid:1)Mp ⊗ M−p 7,p+1(cid:1)Mp ⊗ N−p 8,p+1(cid:1)Np ⊗ M−p 9,p+1(cid:1)Np ⊗ N−p 10,p+1(cid:1)Yp−1/2 ⊗ Y1/2−p(cid:17) = 0. In particular, one has (p − 2)a(−1) 1,p−1 − (p + 2)a(−1) 1,p + (p − 2)a(1) 1,p − (p + 2)a(1) 1,p+1 = 0, ∀ p ∈ Z, which together with the fact that {p ∈ Z a(−1) 1,p 6= 0} is finite and (3.3), forces a(−1) 1,p = 3a(−1) 1,−2 + a(−1) 1,−1 + 3a(1) 1,−1 = a(−1) 1,0 + 3a(−1) 1,1 + 3a(1) 1,2 = a(−1) 1,−1 + a(−1) 1,0 = 0, (3.5) for p ∈ Z\{−2, 0, ±1}. Similarly, comparing the coefficients of Lp ⊗ M−p, Mp ⊗ L−p ,Lp ⊗ N−p, Np ⊗ L−p, Mp ⊗ M−p, Mp ⊗ N−p, Np ⊗ M−p, Np ⊗ N−p and Yp−1/2 ⊗ Y1/2−p and taking (3.3) and (3.4) into account, one has a(−1) 2,0 + 2a(−1) 2,1 = a(−1) 2,0 + 2a(−1) 2,−1 = a(−1) 3,−1 + 2a(−1) 3,0 = a(−1) 3,−1 + 2a(−1) 3,−2 = 0, a(−1) 4,0 + 2a(−1) 4,1 = 2a(−1) 4,−1 + a(−1) 4,0 + a(1) 4,0 = a(−1) 5,−1 + 2a(−1) 5,0 = a(−1) 5,−1 + 2a(−1) 5,−2 = 0, a(−1) i,−1 + a(−1) 2,p = a(1) a(1) i,0 + a(1) 3,p = a(1) i,0 + a(1) i,1 = 2a(−1) 10,−1 +a(−1) 10,0 +2a(1) 10,0 = a(−1) 10,0 + 2a(−1) 10,1 +2a(1) 10,2 = 0, 4,p = a(1) 5,p = a(−1) 2,p2 = a(−1) 3,p3 = a(−1) 4,p4 = a(−1) 5,p5 = a(−1) i,pi = a(−1) 10,p10 = 0, (3.6) (3.7) (3.8) (3.9) for any p ∈ Z, p2 ∈ Z \ {0, ±1}, p3 ∈ Z \ {0, −1, −2}, p4 ∈ Z \ {0, ±1}, p5 ∈ Z \ {0, −1, −2}, pi ∈ Z\{−1, 0} with i = 6, 7, 8, 9 and p10 ∈ Z\{0, ±1}. By (3.3) and (3.4) as well as applying D to [L1, N0] = 0 and [L1, M0] = 0, respectively, we have 6,n = a(1) a(1) 7,n = a(1) 8,n = a(1) 9,n = a(1) 10,n = 0, for all n ∈ Z, d(0) 1,j1 = d(0) 2,j2 = d(0) 3,j3 = d(0) 4,j4 = d(0) 5,j5 1,j1 = b(0) b(0) d(0) 1,0 + 2d(0) 2,j2 = b(0) 1,−1 = d(0) 3,j3 = b(0) 1,0 + 2d(0) 4,j4 = b(0) 1,1 = d(0) b(0) 1,0 + 2b(0) 1,−1 = b(0) 1,0 + 2b(0) 1,1 = b(0) = d(0) 10,p = 0, i,j = d(0) i,j = b(0) l,1 = d(0) l,1 = b(0) 10,p = 0, k,0 + d(0) k,0 + b(0) 5,j5 = b(0) l,0 + d(0) l,0 + b(0) k,−1 = 0, k,−1 = 0, (3.10) (3.11) (3.12) (3.13) (3.14) 6 where j1 ∈ Z\{0, ±1}, j2 ∈ Z\{0, 1}, j3 ∈ Z\{0, −1}, j4 ∈ Z\{0, 1}, j5 ∈ Z\{0, −1}, j ∈ Z\{0}, p ∈ Z\{0, 1}, i = 6, 7, 8, 9, l = 2, 4, 10 and k = 3, 5. Then it follows from (3.3), (3.9) and (3.10) that D(L1) = Pi∈Z(cid:0)a(1) 1,i Li ⊗ L1−i(cid:1) = a(1) 1,−1L−1 ⊗ L2 + a(1) 1,2L2 ⊗ L−1. (3.15) Furthermore, applying D to [L1, Y1/2] = 0, we get D(L1) = 0 from (3.15). Similarly, applying D to [L−1, N0] = 0 and [L−1, M0] = 0, respectively, we obtain from (3.6)-(3.9) and (3.11)-(3.14) that a(−1) 6,i = a(−1) 7,i = a(−1) 8,i = a(−1) 9,i = a(−1) 10,i = b(0) 4,i = b(0) 5,i = d(0) 4,i = d(0) 5,i = 0, d(0) 2,1 − a(−1) 2,0 = d(0) 3,−1 − a(−1) 3,−1 = b(0) 2,1 + a(−1) 4,0 = b(0) 3,−1 + a(−1) 5,0 = 0, (3.16) (3.17) for all i ∈ Z. Set u := L1 ⊗ M−1 − L0 ⊗ M0. Observe that L1 · u = 0. Substitute D + a(−1) the expression of D(L−1), one can safely assume a(−1) 2,1 = 0, since such replacement would not affect the expression of D(L1). Similarly, set u(1) := M−1 ⊗ L1 − M0 ⊗ L0, u(2) := L1 ⊗ N−1 − L0 ⊗ N0, and u(3) := N−1 ⊗ L1 − N0 ⊗ L0, then replace D by D + a(−1) 5,0 u(3) in turn, one can assume a(−1) inn and D + a(−1) inn, D + a(−1) 2,1 uinn into 4,1 u(2) 3,0 u(1) 5,0 = 0. Hence we get 4,1 = a(−1) 3,0 = a(−1) inn D(L−1) = a(−1) 1,−2L−2 ⊗ L1 + a(−1) 1,−1L−1 ⊗ L0 + a(−1) 1,0 L0 ⊗ L−1 + a(−1) 1,1 L−1 ⊗ L2, by (3.5)-(3.7). Finally, using D([L−1, Y−1/2]) = 0 and (3.5), one has D(L−1) = 0. Claim 3. D(L±2) = D(N1) = D(Y1/2) = 0. It follows (3.14) and (3.16) that D(M0) = b(0) 1,−1L−1 ⊗ L1 + b(0) 8,0N0 ⊗ M0 + b(0) +b(0) 1,0L0 ⊗ L0 + b(0) 1,1L1 ⊗ L−1 + b(0) 6,0M0 ⊗ M0 + b(0) 7,0M0 ⊗ N0 9,0N0 ⊗ N0 + b(0) 10,0Y−1/2 ⊗ Y1/2 + b(0) 10,1Y1/2 ⊗ Y−1/2, D(N0) = d(0) 1,−1L−1 ⊗ L1 + d(0) 8,0N0 ⊗ M0 + b(0) 1,0L0 ⊗ L0 + d(0) 9,0N0 ⊗ N0 + d(0) 1,1L1 ⊗ L−1 + d(0) 10,0Y−1/2 ⊗ Y1/2 + d(0) +d(0) 10,1Y1/2 ⊗ Y−1/2. 6,0M0 ⊗ M0 + d(0) 7,0M0 ⊗ N0 (3.18) (3.19) Set v(1) := M0⊗M0, v(2) := M0⊗N0, v(3) := N0⊗M0 and v(4) := Y1/2⊗Y−1/2−Y−1/2⊗Y1/2. Observe 7,0v(2) inn, that L±1 · v(i) = 0, but N0 · v(i) 6= 0 for i = 1, 2, 3, 4. Replacing D by D − 1 inn, D − 1 4 d(0) 6,0 = d(0) 6,0v(1) 7,0 = d(0) 2 d(0) 8,0 = d(0) 10,0 = 0. D− 1 2 d(0) 8,0v(3) inn and D− 1 2 d(0) 10,0v(4) inn in turn in (3.19), one can assume that d(0) By applying D to [N0, N1] = 0 and using (3.13), we have d(0) follows from D([N0, M0]) = 2D(M0) and (3.18)-(3.19)that D(N0) = 0 and 1,−1 = d(0) 1,1 = d(0) 1,−1 = d(0) 10,1 = 0. Then it D(M0) = b(0) 7,0M0 ⊗ N0 + b(0) 8,0N0 ⊗ M0 + b(0) 10,0Y−1/2 ⊗ Y1/2 + b(0) 10,1Y1/2 ⊗ Y−1/2 . (3.20) 7 Now considering D([L±2, M0]) = 0 and D([L±2, N0]) = 0, one has D(L2) = Pia(2) 1,i Li ⊗ L2−i, and D(L−2) = Pia(−2) 1,i Li ⊗ L−2−i. (3.21) As a by-product, we also get b(0) 10,0 = b(0) 10,1 = 0. Replacing D by D + 1 2 b(0) 8,0(N0 ⊗ N0)inn in (3.20), one can assume b(0) 8,0 = 0. Now we get from (3.20) that D(M0) = b(0) 7,0M0 ⊗ N0. (3.22) Applying D to [L1, L−2] = −3L−1 and [L−1, L2] = 3L1, respectively, and using D(L±1) = 0, we have a(2) 1,p = a(−2) 1,q = 2a(2) 1,1 + 3a(2) 1,0 = a(2) 1,0 + 4a(2) 1,−1 = a(2) 1,1 + a(2) 1,2 = a(2) 1,2 + 4a(2) 1,3 = 0, 2a(−2) 1,−1 + 3a(−2) 1,0 = a(−2) 1,0 + 4a(−2) 1,1 = a(−2) 1,−1 + a(−2) 1,−2 = a(−2) 1,−2 + 4a(−2) 1,−3 = 0, (3.23) (3.24) 4 a(2) where p ∈ Z\{0, ±1, 2, 3} and q ∈ Z\{−3, −2, 0, ±1}. Set v := L−1 ⊗ L1 − 2L0 ⊗ L0 + L1 ⊗ L−1 and take D − 1 1,0 = 0. Then it follows (3.21) and (3.23) that D(L2) = 0. Consequently, one can easily get D(L−2) = 0 by applying D to [L−2, L2] = 4L0 and using (3.24). 1,0vinn in place of D in the first equation of (3.21), one can assume a(2) Applying D to [M0, Y1/2] = 0 and [M0, Y−1/2] = 0, respectively, one has γ† 1,i = γ† 0,i = γ0,j = γ0,j = 0, ∀ i ∈ Z, j ∈ Z\{0}. (3.25) Similarly, when D is applied to [N0, Y1/2] = Y1/2 and [N0, Y−1/2] = Y−1/2, respectively, it follows β1,i = β† 1,i = β0,i = β† 0,i = 0, for all i ∈ Z. Using D([L1, Y1/2]) = 0, D([L−1, Y−1/2]) = 0 and D(L±1) = 0, we have α1,i = α† 1,i = α† 0,i = α0,j = 0, ∀ i ∈ Z\{0, 1}, j ∈ Z\{0, −1}, α1,0 + α1,1 = α0,0 + α0,−1 = α† 1,0 + α† 1,1 = α† 0,0 + α† 0,1 = 0. (3.26) (3.27) Applying D to [L1, Y−1/2] = −Y1/2, one has α0,0 = α1,1, α† b = α† 0,0 and c = γ0,0 . Hence, we can rewrite D(Y±1/2) as follows: 0,0 = α† 1,0 and γ0,0 = γ1,0. Let a = α0,0, D(Y1/2) = −aL0 ⊗ Y1/2 + aL1 ⊗ Y−1/2 + bY−1/2 ⊗ L1 − bY1/2 ⊗ L0 + cN0 ⊗ Y1/2, (3.28) D(Y−1/2) = aL0 ⊗ Y−1/2 − aL−1 ⊗ Y1/2 + bY−1/2 ⊗ L0 − bY1/2 ⊗ L−1 + cN0 ⊗ Y−1/2. (3.29) When D is applied to [L2, Y−1/2] = − 3 0, since D(L±2) = 0. It follows c = d(0) D(Y±1/2) = D(M0) = 0 by (3.22), (3.28) and (3.29). 2 Y3/2 and [L−2, Y3/2] = 5 7,0 = 0 by applying D to [Y−1/2, Y1/2] = M0, which proves 2 Y−1/2, respectively, one has a = b = 8 Now it is left to calculate D(N1). Firstly, using D([N0, N1]) = 0 and D(N0) = 0, we have d(1) 2,i = d(1) 3,i = d(1) 6,i = d(1) 7,i = d(1) 8,i = d(1) 10,i = 0, ∀ i ∈ Z. Then applying D to [L−1, N1] = N0 and using D(N0) = 0, we obtain d(1) 1,i1 = d(1) 4,0 + 2d(1) d(1) 4,i2 = d(1) 4,1 = d(1) 5,i3 = d(1) 4,0 + 2d(1) 9,i4 = d(1) 4,−1 = d(1) 1,0 + d(1) 5,1 + 2d(1) 1,1 = d(1) 1,1 + 3d(1) 1,2 = d(1) 5,0 = d(1) 5,1 + 2d(1) 5,2 = d(1) 1,0 + 3d(1) 9,0 + d(1) 1,−1 = 0, 9,1 = 0, (3.30) (3.31) where i1 ∈ Z\{0, ±1, 2}, i2 ∈ Z\{0, ±1}, i3 ∈ Z\{0, 1, 2}, and i4 ∈ Z\{0, 1}. Finally, by applying D to [N1, Y−1/2] = Y1/2 and (3.31) as well as D(Y±1/2) = 0, we get d(1) 1,i = d(1) 4,i = d(1) 5,i = d(1) 9,i = 0, ∀ i ∈ Z, which together with (3.30), yields D(N1) = 0. Hence, the claim is proved, so is the theorem, since L is generated by L±1, L±2, N1 and Y1/2. (cid:3) The following lemma is very useful to the main theorem in the paper. Lemma 3.6 Suppose v ∈ V such that x · v ∈ Im( Id − τ ) for all x ∈ L. Then v ∈ Im( Id − τ ). Proof. First note that L · Im( Id − τ ) ⊂ Im( Id − τ ). We shall show that after several steps in each of which v is replaced by v − u for some u ∈ Im( Id − τ ), the zero element is obtained, which leads us to the result. Write v = Pn∈ 1 Z vn. Obviously, 2 v ∈ Im( Id − τ ) ⇐⇒ vn ∈ Im( Id − τ ), ∀ n ∈ 1 2 Z. (3.32) Z nvn = L0 · v ∈ Im( Id − τ ). By (3.32), nvn ∈ Im( Id − τ ). In particular, vn ∈ Im( Id − τ ) Then Pn∈ 1 if n 6= 0. Thus when replacing v by v − Pn∈ 1 2 2 Z∗ vn, one can suppose v = v0 ∈ V0. Write (aiLi ⊗ L−i + biLi ⊗ M−i + ciMi ⊗ L−i + diMi ⊗ M−i + eiYi−1/2 ⊗ Y1/2−i v = Pi∈Z + fiLi ⊗ N−i + giNi ⊗ L−i + hiNi ⊗ N−i + kiMi ⊗ N−i + riNi ⊗ M−i). Since all the elements of the forms Ei ⊗ F−i − F−i ⊗ Ei and Yi−1/2 ⊗ Y1/2−i − Y1/2−i ⊗ Yi−1/2 are contained in Im( Id − τ ), where {Ei, Fi} ⊂ {Li, Mi, Ni} for all i ∈ Z. Replacing v by v − u, where u is a combination of some of these elements, we can assume ci = gi = ri = 0, ∀ i ∈ Z; ai, di, hi 6= 0 =⇒ i > 0 or i = 0; ei 6= 0 =⇒ i > 0. (3.33) Then v can be rewritten as (aiLi ⊗ L−i + diMi ⊗ M−i + hiNi ⊗ N−i) v = Pi∈Z+ + Pi∈Z (biLi ⊗ M−i + fiLi ⊗ N−i + kiMi ⊗ N−i) + Pi∈Z>0 eiYi−1/2 ⊗ Y1/2−i. (3.34) 9 Assume ap 6= 0 for some p > 0. Choose q > 0 such that q 6= p. Then Lp+q ⊗ L−p appears in Lq · v, but (3.33) implies the term L−p ⊗ Lp+q does not appear in Lq · v, which contradicts the fact that Lq ·v ∈ Im( Id−τ ). Hence we get ai = 0, ∀ i ∈ Z∗. Similarly, one can suppose di = hi = 0, ∀ i ∈ Z∗ and ei = 0, ∀ i ∈ Z. Then (3.34) becomes v = Pi∈Z (biLi ⊗ M−i + fiLi ⊗ N−i + kiMi ⊗ N−i) + a0L0 ⊗ L0 + d0M0 ⊗ M0 + h0N0 ⊗ N0 . (3.35) By Im( Id − τ ) ⊂ Ker( Id + τ ) and our hypothesis L · v ⊂ Im( Id − τ ), we have 0 = ( Id + τ )M0 · v = −4h0(M0 ⊗ N0 + N0 ⊗ M0) − 2Pi∈Z fi(Li ⊗ M−i + M−i ⊗ Li) − 4Pi∈Z ki(Mi ⊗ M−i) . Comparing the coefficients, one gets h0 = 0 and fi = ki = 0, ∀ i ∈ Z. Similarly, ( Id + τ )(N0 · v) = 0 implies d0 = 0 and bi = 0 for all i ∈ Z; and ( Id + τ )(L1 · v) = 0 leads to a0 = 0. Then the lemma follows from (3.35). (cid:3) By now we have enough in hand to classify the Lie bialgebra structures on the extended Schrodinger-Virasoro Lie algebra. The following theorem is the central result of the paper. Theorem 3.7 Let (L, [·, ·]) be the extended Schrodinger-Virasoro Lie algebra. Then each Lie bial- gebra structure on L is triangular coboundary. Proof. Let (L, [·, ·], ∆) be a Lie bialgebra structure on L. Thanks to Theorem 3.5, there exists some r ∈ V such that ∆ = ∆r. By (2.1), Im(∆) ⊂ Im( Id − τ ). Then it follows from Lemma 3.6 that r ∈ Im( Id − τ ). But (2.1), Theorem 2.2(iii) and Lemma 3.4 show that c(r) = 0; as a result, (L, [·, ·], ∆) is a triangular coboundary Lie bialgebra by Definition 2.1. (cid:3) References [1] V.G. Drinfel'd, Constant quasiclassical solutions of the Yang-Baxter quantum equation, Soviet Math. Dokl. 28(3) (1983), 667 -- 671. [2] V.G. Drinfel'd, Quantum groups, in: Proceeding of the International Congress of Mathemati- cians, Vol. 1, 2, Berkeley, Calif. 1986, Amer. Math. Soc., Providence, RI, 1987, pp. 798 -- 820. [3] P. Etingof, D. Kazhdan, Quantization of Lie bialgebras, I, Selecta Math. (New Series) 2 (1996) 1-41. [4] R. Farnsteiner, Derivations and extensions of finitely generated graded Lie algebras, J. Algebra, 118(1) (1988), 34 -- 45. [5] S. Gao, C. Jiang, Y. Pei , Structure of the extended Schrodinger-Virasoro Lie algebra, Alg. Colloq, in press (2008). [6] M. Henkel, Schrodinger invariance and strongly anisotropic critical systems. J. Stat. Phys., 75 (1994), 1023 -- 1029. [7] M. Henkel, Unterberger J. Schrodinger invariance and space-time symmetries. Nucl. Phys. B, 660 (2003), 407 -- 412. 10 [8] J. Han, Y. Su, Lie bialgebra structures on the Schrodinger-Virasoro Lie algebra, J. Math. Phys. 50 (2009), no. 8, 083504, 12 pp. [9] J. Li, Y. Su, Representations of the Schrodinger-Virasoro algebras, J. Math. Phys, 49 (2008), 053512. [10] J. Li, Y. Su, The derivation algebra and automorphism group of the twisted Schrodinger- Virasoro algebra, arXiv:0801.2207v1, (2008). [11] J. Li, Y. Su, B, Xin, Lie bialgebras of a family of Block type, Chinese Annals of Math. (Series.B) 29 (2008), 487 -- 500. [12] J. Li, Y. Su, B. Xin, Lie Bialgebra Structures on the Centerless W -Algebra W (2; 2), Algebra Colloquium, accepted. [13] J. Li, Y. Su, L. Zhu, 2-cocycles of original deformative Schrodinger-Virasoro algebras, Sci- ence in China: Series A 51 (2008), 1989 -- 1999. [14] W. Michaelis, A class of infinite-dimensional Lie bialgebras containing the Virasoro algebras, Adv. Math. 107 (1994), 365 -- 392. [15] W. Michaelis, Lie coalgebras, Adv. Math. 38 (1980), 1 -- 54. [16] W. Michaelis, The dual Poincare-Birkhoff-Witt theorem, Adv.Math. 57 (1985), 93 -- 162. [17] S.H. Ng, E.J. Taft, Classification of the Lie bialgebra structures on the Witt and Virasoro algebras, J. Pure Appl. Alg. 151 (2000), 67 -- 88. [18] C. Roger, J. Unterberger, The Schrodinger-Virasoro Lie group and algebra: representation theory and cohomological study, Ann. Henri Poincar´e, 7 (2006), 1477 -- 1529. [19] G. Song, Y. Su, Lie bialgebras of generalized Witt type, Science in China: Series A 49 (2006), 533 -- 544. [20] E.J. Taft, Witt and Virasoro algebras as Lie bialgebras, J. Pure Appl. Al. 87 (1993), 301 -- 312. [21] S. Tan, X. Zhang, Automorphisms and Verma modules for Generalized Schrodinger-Virasoro algebras, arXiv:0804.1610v2. [22] J. Unterberger, On vertex algebra representations of the Schrodinger-Virasoro algebra. arXiv:cond-mat/0703 214v2, (2007). [23] Y. Wu, G. Song, Y. Su, Lie bialgebras of generalized Witt type. II. Comm. Algebra, 35(6) (2007), 1992-2007. [24] Y. Wu, G. Song, Y. Su, Lie bialgebras of generalized Virasoro-like type, Acta Mathematica Sinica, English Series, 22 (2006), 1915 -- 1922. [25] X. Zhang, S. Tan, Whittaker modules and a class of new modules similar as Whittaker modules for the Schrodinger-Virasoro algebra, arXiv:0812.3245v1. 11
1503.05364	2	1503	2016-06-13T12:15:29	Hochschild products and global non-abelian cohomology for algebras. Applications	[ "math.RA" ]	Let $A$ be a unital associative algebra over a field $k$, $E$ a vector space and $\pi : E \to A$ a surjective linear map with $V = {\rm Ker} (\pi)$. All algebra structures on $E$ such that $\pi : E \to A$ becomes an algebra map are described and classified by an explicitly constructed global cohomological type object ${\mathbb G} {\mathbb H}^{2} \, (A, \, V)$. Any such algebra is isomorphic to a Hochschild product $A \star V$, an algebra introduced as a generalization of a classical construction. We prove that ${\mathbb G} {\mathbb H}^{2} \, (A, \, V)$ is the coproduct of all non-abelian cohomologies ${\mathbb H}^{2} \, \, (A, \, (V, \cdot))$. The key object ${\mathbb G} {\mathbb H}^{2} \, (A, \, k)$ responsible for the classification of all co-flag algebras is computed. All Hochschild products $A \star k$ are also classified and the automorphism groups ${\rm Aut}_{\rm Alg} (A \star k)$ are fully determined as subgroups of a semidirect product $A^* \, \ltimes \bigl(k^* \times {\rm Aut}_{\rm Alg} (A) \bigl)$ of groups. Several examples are given as well as applications to the theory of supersolvable coalgebras or Poisson algebras. In particular, for a given Poisson algebra $P$, all Poisson algebras having a Poisson algebra surjection on $P$ with a $1$-dimensional kernel are described and classified.	math.RA	math	HOCHSCHILD PRODUCTS AND GLOBAL NON-ABELIAN COHOMOLOGY FOR ALGEBRAS. APPLICATIONS A. L. AGORE AND G. MILITARU Abstract. Let A be a unital associative algebra over a field k, E a vector space and π : E → A a surjective linear map with V = Ker(π). All algebra structures on E such that π : E → A becomes an algebra map are described and classified by an explicitly constructed global cohomological type object GH2 (A, V ). Any such algebra is isomorphic to a Hochschild product A ⋆ V , an algebra introduced as a generalization of a classical construction. We prove that GH2 (A, V ) is the coproduct of all non- abelian cohomologies H2 (A, (V, ·)). The key object GH2 (A, k) responsible for the classification of all co-flag algebras is computed. All Hochschild products A ⋆ k are also classified and the automorphism groups AutAlg(A⋆k) are fully determined as subgroups of a semidirect product A∗ ⋉ (cid:0)k∗ × AutAlg(A)(cid:1) of groups. Several examples are given as well as applications to the theory of supersolvable coalgebras or Poisson algebras. In particular, for a given Poisson algebra P , all Poisson algebras having a Poisson algebra surjection on P with a 1-dimensional kernel are described and classified. Introduction Introduced at the level of groups by Holder [28], the extension problem is a famous and still open problem to which a vast literature was devoted (see [1] and the references therein). Fundamental results obtained for groups [1, 15, 37] served as a model for studying the extension problem for several other fields such as Lie/Leibniz algebras [14, 31], super Lie algebras [6], associative algebras [17, 26], Hopf algebras [8], Poisson algebras [24], Lie-Rinehart algebras [12, 25] etc. The extension problem is one of the main tools for classifying 'finite objects' and has been a source of inspiration for developing cohomology theories in all fields mentioned above. We recall the extension problem using the language of category theory. Let C be a category having a zero (i.e. an initial and final) object 0 and for which it is possible to define an exact sequence. Given A, B two fixed objects of C, the extension problem consists of the following question: Describe and classify all extensions of A by B, i.e. all triples (E, i, π) consisting of an object E of C and two morphisms in C that fit into an exact sequence of the form: 0 / B i / E π / A / 0 2010 Mathematics Subject Classification. 16D70, 16S70, 16E40. Key words and phrases. The extension problem, non-abelian cohomology, classification of algebras. A.L. Agore is Postdoctoral Fellow of the Fund for Scientific Research-Flanders (Belgium) (F.W.O. Vlaanderen). This work was supported by a grant of the Romanian National Authority for Scientific Research, CNCS-UEFISCDI, grant no. 88/05.10.2011. 1 / / / / 2 A. L. AGORE AND G. MILITARU Two extensions (E, i, π) and (E′, i′, π′) of A by B are called equivalent if there exists an isomorphism ϕ : E → E′ in C that stabilizes B and co-stabilizes A, i.e. ϕ ◦ i = i′ and π′ ◦ ϕ = π. The answer to the extension problem is given by explicitly computing the set Ext(A, B) of all equivalence classes of extensions of A by B via this equiva- lence relation. The simplest case is that of extensions with an 'abelian' kernel B for which a Schreier type theorem proves that all extensions of A by an abelian object B are classified by the second cohomology group H2(A, B) - the result is valid for groups, Lie/Leibniz/associative/Poisson/Hopf algebras, but the construction of the second co- homology group is different for each of the above categories [8, 14, 15, 26, 37]. The difficult part of the extension problem is the case when B is not abelian: as a general principle, the Schreier type theorems remain valid, but this time the classifying object of all extensions of A by B is not the cohomology group of a given complex anymore, but only a pointed set called the non-abelian cohomology H2 nab(A, B). For its construction in the case of groups we refer to [7], while for Lie algebras to [19] where it was proved that the non-abelian cohomology H2 nab(A, B) is the Deligne groupoid of a suitable differ- ential graded Lie algebra. The difficulty of the problem consists in explicitly computing H2 nab(A, B): lacking an efficient cohomology tool as in the abelian case [35, 38], it needs to be computed 'case by case' using different computational and combinatorial approaches. This paper deals, at the level of associative algebras, with a generalization of the ex- tension problem, called the global extension (GE) problem, introduced recently in [3, 33] for Poisson/Leibniz algebras as a categorical dual of the extending structures problem [2, 4, 5]. The GE-problem can be formulated for any category C using a simple idea: in the classical extension problem we drop the hypothesis 'B is a fixed object in C' and replace it by a weaker one, namely 'B has a fixed dimension'. For example, if C is the category of unital associative algebras over a field k, the GE-problem can be formulated as follows: for a given algebra A, classify all associative algebras E for which there exists a surjective algebra map E → A → 0 whose kernel has a given dimension c as a vector space. Of course, any such algebra has A× V as the underlying vector space, where V is a vector space such that dim(V ) = c. Among several equivalent possibilities for writing down the GE-problem for algebras, we prefer the following: Let A be a unital associative algebra, E a vector space and π : E → A a linear epi- morphism of vector spaces. Describe and classify the set of all unital associative algebra structures that can be defined on E such that π : E → A becomes a morphism of algebras. By classification of two algebra structures ·E and ·′ E on E we mean the classification up to an isomorphism of algebras (E, ·E) ∼= (E, ·′ E) that stabilizes V := Ker(π) and co-stabilizes A: we shall denote by Gext(A, E) the set of equivalence classes of all algebra structures on E such that π : E → A is an algebra map. Let us explain now the significant differences between the GE-problem and the classical extension problem for associative algebras whose study was initiated in [17, 26]. Let (E, ·E) be a unital algebra structure on E such that π : (E, ·E) → A is an algebra map. Then (E, ·E) is an extension of the unital algebra A by the associative algebra V = Ker(π), which is a non-unital subalgebra (in fact a two-sided ideal) of (E, ·E). However, the multiplication on V is not fixed from the input data, as in the case of the classical extension problem: it depends essentially on the algebra structures on E which we are looking for. Thus the HOCHSCHILD PRODUCTS AND NON-ABELIAN COHOMOLOGY FOR ALGEBRAS 3 classical extension problem is in some sense the 'local' version of the GE-problem. The partial answer and the best result obtained so far for the classical extension problem was given in [26, Theorem 6.2]: all algebra structures ·E on E such that V is a two-sided ideal of null square (i.e. x ·E y = 0, for all x, y ∈ V - that is V is an 'abelian' algebra) are classified by the second Hochschild cohomological group H2(A, V ). The result, as remarkable as it is, has its limitations: for example, if A := Mn(k) is the algebra of n × n-matrices then there is no associative algebra E of dimension 1 + n2 for which we have a surjective algebra map A → Mn(k) with a null square kernel - however, there is a large family of such algebras with a projection on Mn(k), one of them being of course the direct product of algebras Mn(k) × k (more details are given in Example 2.13). This example provides enough motivation for studying the GE-problem and the fact that it covers the missing part in the classical Hochschild approach of the extension problem. The paper is organized as follows. Section 1 gives the theoretical answer to the GE- problem in three steps. First of all, in Proposition 1.2 we introduce a new product A ⋆ V = A ⋆(⊳, ⊲, ϑ, ·) V , associated to an algebra A and a vector space V connected by two 'actions' of A on V , a 'cocycle' and an associative multiplication · on V satisfying several axioms. We call the algebra A ⋆ V the Hochschild product since, in the particu- lar case when · is the trivial multiplication on V the above product reduces to the one in this case the axioms involved in the construction introduced by Hochschild in [26]: of A ⋆ V come down to the fact that V is an A-bimodule and ϑ : A × A → V is an usual normalized 2-cocycle. On the other hand, if the cocycle ϑ is the trivial map then the associated Hochschild product A ⋆ V is just the semidirect product A# V of algebras and the corresponding axioms show that (V, ·) is an associative algebra in the monoidal category (AMA, − ⊗A −, A) of A-bimodules. The canonical surjection A ⋆ V → A is an algebra map having the kernel V . Proposition 1.4 proves the converse: any algebra structure ·E which can be defined on the vector space E such that π : (E,·E ) → A is a morphism of algebras is isomorphic to a Hochschild product A ⋆ V . Based on these results and a technical lemma, the theoretical answer to the GE-problem is given in Theorem 1.9: the classifying set Gext(A, E) is parameterized by an explicitly con- structed global cohomological object GH2 (A, V ) and the bijection between the elements of GH2 (A, V ) and Gext(A, E) is given. On the route, Corollary 1.5 proves that any finite dimensional algebra is isomorphic to an iteration of Hochschild products of the form (cid:0)···(cid:0)(S ⋆ V1) ⋆ V2(cid:1)⋆··· ⋆ Vt(cid:1), where S is a finite dimensional simple algebra and V1,··· , Vt are finite dimensional vector spaces. Corollary 1.11 provides a decomposition of GH2 (A, V ) as the coproduct of all non-abelian cohomologies H2(cid:0)A, (V, ·V )(cid:1), which are classifying objects for the extensions of A by all associative algebra structures · on V -- the second Hochschild cohomological group H2 (A, V ) is the most elementary piece among all components of GH2 (A, V ). Computing the classifying object GH2 (A, V ) is a highly nontrivial task: if A is finite dimensional and V := kn this object parameterizes the equivalence classes of all unital associative algebras of dimension n + dim(A) that admit an algebra surjection on A. In Section 2 we shall identify a way of computing this object for a class of algebras called co-flag algebras over A, i.e. algebras E that have a finite π1−→ A0 := A, chain of surjective morphisms of algebras An := E such that dim(Ker(πi)) = 1, for all i = 1,··· , n. All co-flag algebras over A can be πn−→ An−1 ··· π2−→ A1 4 A. L. AGORE AND G. MILITARU completely described and classified by a recursive reasoning whose key step is given in Proposition 2.6 and Corollary 2.7 where GH2 (A, k) is computed and all Hochschild prod- ucts A ⋆ k are described by generators and relations. The less restrictive classification of these algebras, i.e. only up to an agebra isomorphism, is given in Theorem 2.8 where the second classifying object HOC (A, k) is computed: it parameterizes the isomorphism classes of all Hochschild products A ⋆ k, that is, it classifies up to an isomorphism, all algebras B which admit a surjective algebra map B → A with a 1-dimensional kernel. As a bonus of our approach, the automorphism groups AutAlg(A⋆k) are fully determined in Corollary 2.12 as subgroups of a semidirect product A∗ ⋉(cid:0)k∗ × AutAlg(A)(cid:1) of groups. At this point we should mention that determining the structure of the automorphism group of a given algebra is an old problem, intensively studied and very difficult, arising from invariant theory (see [10, 13] and the references therein). Several examples where both classifying objects GH2 (A, k) and HOC (A, k) are explicitly computed for different algebras A are worked out in details. For instance, if A = k[Cn], is the group algebra of the cyclic group Cn of order n, then GH2 (k[Cn], k) ∼= (cid:0)Un(k) × Un(k)(cid:1) ⊔ k∗, where Un(k) is the group of n-th roots of unity in k. The classification of these algebras, by de- scribing the classifying object HOC (k[Cn], k), is also indicated and reduces the question to a challenging number theory problem which depends heavily on the arithmetics of n and the base field k -- it is also related to two intensively studied problems in the theory of group rings, namely the description of all invertible elements and the automorphism group of a group algebra [23, 32, 34]. An intriguing example is A := Tn(k), the algebra of upper triangular (n × n)-matrices. The global cohomological object GH2 (Tn(k), k) is computed in Example 2.19 being described by a very interesting set of matrices of trace 0. In particular, HOC (T2(k), k) = 8, i.e. up to an isomorphism of algebras there exist exactly eight 4-dimensional algebras that have an algebra projection on the Heisenberg algebra T2(k). Applications for coalgebras and Poisson algebras are given in Section 3 based on the same idea: namely, that of rephrasing the concepts and results of this paper for coalgebras (resp. Poisson manifolds) via two different contravariant functors (−)∗ := Homk (−, k) (resp. C ∞(−)) from the category of coalgebras (resp. Poisson manifolds) to the category of algebras (resp. Poisson algebras). Having the theory of supersolvable Lie algebras [9] as a source of inspiration we introduce the concept of a supersolvable coalgebra in such a way that a coalgebra C is supersolvable if and only if the convolution algebra C ∗ is a co-flag algebra. In particular, Corollary 2.18 classifies all 3-dimensional supersolvable coalgebras over a field of characteristic 6= 2. On the other hand, for a given Poisson algebra P , Theorem 3.3 classifies up to an isomorphism all Poisson algebras Q which admit a Poisson surjection Q → P → 0 with a 1-dimensional kernel. The result is the algebraic counterpart of the classification problem of all Poisson manifolds containing a given Poisson manifold M of codimension 1. As an example, we show that there exist exactly six families of 4-dimensional Poisson algebras with a Poisson algebra surjection of the Heisenberg-Poisson algebra H(3, k). For applications and further motivation for studying Poisson algebras we refer to [21, 22, 30]. HOCHSCHILD PRODUCTS AND NON-ABELIAN COHOMOLOGY FOR ALGEBRAS 5 1. The global extension problem Notations and terminology. For two sets X and Y we shall denote by X ⊔ Y their coproduct in the category of sets, i.e. X ⊔ Y is the disjoint union of X and Y . Unless otherwise specified, all vector spaces, linear or bilinear maps are over an arbitrary field k. A map f : V → W between two vector spaces is called the trivial map if f (v) = 0, for all v ∈ V . By an algebra A we always mean a unital associative algebra over k whose unit will be denoted by 1A. All algebra maps preserve units and any left/right A-module is unital. For an algebra A, AutAlg(A) denotes the group of algebra automorphisms of A, Alg (A, k) is the space of all algebra maps A → k while AMA stands for the category of A-bimodules. If (V, ⊲, ⊳) ∈ AMA is an A-bimodule, then the trivial extension of A by V is the algebra A× V , with the multiplication defined for any a, b ∈ A, x, y ∈ V by: (1) Let A be an algebra, E a vector space, π : E → A a linear epimorphism of vector spaces with V := Ker(π) and denote by i : V → E the inclusion map. We say that a linear map ϕ : E → E stabilizes V (resp. co-stabilizes A) if the left square (resp. the right square) of the following diagram (a, x) · (b, y) := (cid:0)ab, a ⊲ y + x ⊳ b(cid:1) V Id V i i E π A ϕ Id / E π / A (2) is commutative. Two unital associative algebra structures · and ·′ on E such that π : E → A is a morphism of algebras are called cohomologous and we denote this by (E,·) ≈ (E,·′), if there exists an algebra map ϕ : (E,·) → (E,·′) which stabilizes V and co- stabilizes A. We can easily prove that any such morphism is bijective and thus, ≈ is an equivalence relation on the set of all algebra structures on E such that π : E → A is an algebra map and we denote by Gext (A, E) the set of all equivalence classes via the equivalence relation ≈. Gext (A, E) is the classifying object for the GE-problem. In what follows we will prove that Gext (A, E) is parameterized by a global cohomological object GH2 (A, V ) which will be explicitly constructed. To start with, we introduce the following: Definition 1.1. Let A be an algebra and V a vector space. A Hochschild data of A by V is a system Θ(A, V ) = (⊲, ⊳, ϑ, ·) consisting of four bilinear maps ⊲ : A × V → V, ⊳ : V × A → V, ϑ : A × A → V, · : V × V → V For a Hochschild data Θ(A, V ) = (⊲, ⊳, ϑ, ·) we denote by A ⋆ V = A ⋆(⊳, ⊲, ϑ, ·) V the vector space A × V with the multiplication defined for any a, b ∈ A and x, y ∈ V by: (a, x) ⋆ (b, y) := (ab, ϑ(a, b) + a ⊲ y + x ⊳ b + x · y) (3) A ⋆ V is called the Hochschild product associated to Θ(A, V ) if it is an associative algebra with the multiplication given by (3) and the unit (1A, 0V ). In this case Θ(A, V ) = (⊲, ⊳, ϑ, ·) is called a Hochschild system of A by V . The multiplication defined by (3) is more general than the one appearing in the proof of [26, Theorem 6.2] -- the latter arises / / / / / / 6 A. L. AGORE AND G. MILITARU as a special case of A ⋆ V for · : V × V → V the trivial map, that is x · y = 0, for all x, y ∈ V . Moreover, the GE-problem is the dual, in the sense of category theory, of the extending structures problem studied for algebras in [5]: hence, from this viewpoint the Hochschild product A ⋆ V can be seen as a categorical dual of the unified product A ⋉ V introduced in [5, Theorem 2.2]. The necessary and sufficient conditions for A ⋆ V to be a Hochschild product are given in the following: Proposition 1.2. Let A be an algebra, V a vector space and Θ(A, V ) = (⊲, ⊳, ϑ, ·) a Hochschild data of A by V . Then A ⋆ V is a Hochschild product if and only if the following compatibility conditions hold for any a, b, c ∈ A and x, y ∈ V : (H0) ϑ(a, 1A) = ϑ(1A, a) = 0, x ⊳ 1A = x, 1A ⊲ x = x (H1) (x · y) ⊳ a = x · (y ⊳ a) (H2) (x ⊳ a) · y = x · (a ⊲ y) (H3) a ⊲ (x · y) = (a ⊲ x) · y (H4) (a ⊲ x) ⊳ b = a ⊲ (x ⊳ b) (H5) ϑ(a, bc) − ϑ(ab, c) = ϑ(a, b) ⊳ c − a ⊲ ϑ(b, c) (H6) (ab) ⊲ x = a ⊲ (b ⊲ x) − ϑ(a, b) · x (H7) x ⊳ (ab) = (x ⊳ a) ⊳ b − x · ϑ(a, b) (H8) The bilinear map · : V × V → V is associative. Before going into the proof, we make some comments on the relations (H0)-(H8). The first relation in (H0) together with (H5) show that ϑ is a normalized Hochschild 2- cocycle. (H6) and (H7) are deformations of the usual left and respectively right A-module conditions: together with (H4) and the last two relations of (H0) they measure how far (V, ⊲, ⊳) is from being an A-bimodule. Finally, axioms (H1)-(H3) are compatibilities between the associative multiplication · on V and the 'actions' (⊲, ⊳) of A on V which are missing in the classical theory [26] since · is the trivial map and thus they are automatically fulfilled. When ϑ is the trivial map, axioms (H1)-(H3) together with (H6)-(H8) imply that (V,·) is an associative algebra in the monoidal category AMA = (AMA, − ⊗A −, A) of A-bimodules (see Example 1.3 below). Proof. To start with, we can easily prove that (1A, 0V ) is the unit for the multiplication defined by (3) if and only if (H0) holds. The rest of the proof relies on a detailed analysis of the associativity condition for the multiplication given by (3). Since in A ⋆ V we have (a, x) = (a, 0) + (0, x), it follows that the associativity condition holds if and only if it holds for all generators of A⋆V , i.e. for the set {(a, 0) a ∈ A}∪{(0, x) x ∈ V }. To save space we will illustrate only a few cases, the rest of the details being left to the reader. For instance, the associativity condition for the multiplication given by (3) holds in {(0, (a, 0)} if and only if (H1) holds. Similarly, the associativity condition holds x), in {(0, x), (a, 0), (0, y)} if and only if (H2) holds while, the associativity condition holds in {(0, x), From now on a Hochschild system of A by V will be viewed as a system of bilinear maps Θ(A, V ) = (⊲, ⊳, ϑ, ·) satisfying the axioms (H0)-(H8) and we denote by HS (A, V ) the set consisting of all Hochschild systems of A by V . (0, z)} if and only if · : V × V → V is associative. (cid:3) (0, y), (0, y), HOCHSCHILD PRODUCTS AND NON-ABELIAN COHOMOLOGY FOR ALGEBRAS 7 Examples 1.3. 1. By applying Proposition 1.2 we obtain that a Hochschild data (⊲, ⊳, ϑ, ·) for which · is the trivial map is a Hochschild system if and only if (V, ⊲, ⊳) is an A-bimodule and ϑ : A × A → V is a normalized 2-cocycle. Furthermore, if ϑ is also the trivial map, then the associated Hochschild product is just the trivial extension of the algebra A by the A-bimodule V as defined by (1). 2. A Hochschild system Θ(A, V ) = (⊲, ⊳, ϑ, ·) for which ϑ is the trivial map is called a semidirect system of A by V . In this case ϑ will be omitted when writing down Θ(A, V ) and axioms in Proposition 1.2 take a simplified form: Θ(A, V ) = (⊲, ⊳, ·) is a semidirect system if and only if (V, ⊲, ⊳) ∈ AMA is an A-bimodule, (V,·) is an associative algebra and a ⊲ (x · y) = (a ⊲ x) · y, (x ⊳ a) · y = x · (a ⊲ y) (x · y) ⊳ a = x · (y ⊳ a), (4) for all a ∈ A, x, y ∈ V . The Hochschild product associated to a semidirect system Θ(A, V ) = (⊲, ⊳, ·) is called a semidirect product of algebras and will be denoted by A#V := A#(⊳, ⊲, ·)V . The terminology will be motivated below in Corollary 1.6: exactly as in the case of groups or Lie algebras, the semidirect product of algebras describes split epimorphisms in the category of algebras. We will rephrase the axioms of a semidirect system Θ(A, V ) = (⊲, ⊳, ·) of A by V using the language of monoidal categories. The first and the second axioms of (4) are equivalent to the fact that the bilinear map · : V × V → V is an A-bimodule map, while the last one is the same as saying that the map is A-balanced. The space of these maps is in one-to-one correspondence with the set of all A-bimodule maps V ⊗A V → V . This fact together with the other two axioms can be rephrased as follows: Θ(A, V ) = (⊲, ⊳, ·) is a semidirect system of A by V if and only if (V,·) is a (not-necessarily unital) associative algebra in the monoidal category AMA = (AMA, − ⊗A −, A) of A-bimodules. The Hochschild product is the tool to answer the GE-problem. Indeed, first we observe that the canonical projection πA : A ⋆ V → A, πA(a, x) := a is a surjective algebra map with kernel {0} × V ∼= V . Hence, the algebra A ⋆ V is an extension of the algebra A by the associative algebra (V,·) via / V (5) 0 iV / / A ⋆ V πA / A / 0 where iV (x) = (0, x). Conversely, we have: Proposition 1.4. Let A be an algebra, E a vector space and π : E → A an epimorphism of vector spaces with V = Ker(π). Then any algebra structure · which can be defined on the vector space E such that π : (E,·) → A becomes a morphism of algebras is isomorphic to a Hochschild product A ⋆ V and moreover, the isomorphism of algebras (E,·) ∼= A ⋆ V can be chosen such that it stabilizes V and co-stabilizes A. Thus, any unital associative algebra structure on E such that π : E → A is an algebra map is cohomologous to an extension of the form (5). Proof. Let · be an algebra structure of E such that π : (E,·) → A is an algebra map. Since k is a field we can pick a k-linear section s : A → E of π, i.e. π ◦ s = IdA and s(1A) = 1E. Then ϕ : A × V → E, ϕ(a, x) := s(a) + x is an isomorphism of vector / / / 8 A. L. AGORE AND G. MILITARU spaces with the inverse ϕ−1(y) = (cid:0)π(y), y − s(π(y))(cid:1), for all y ∈ E. Using the section s we define three bilinear maps given for any a, b ∈ A and x ∈ V by: : V × A → V, x ⊳ a := x · s(a) ⊳ = ⊳s ⊲ = ⊲s : A × V → V, a ⊲ x := s(a) · x ϑ = ϑs : A × A → V, ϑ(a, b) := s(a) · s(b) − s(ab) and let ·V : V × V → V be the restriction of · at V , i.e. x ·V y := x · y, for all x, y ∈ V . We can easily see that they are well-defined maps. The key step is the following: using the system (⊳, ⊲, ϑ, ·V = ·) connecting A and V we can prove that the unique algebra structure ⋆ that can be defined on the direct product of vector spaces A × V such that ϕ : A × V → (E,·) is an isomorphism of algebras is given by: (6) for all a, b ∈ A, x, y ∈ V . Indeed, let ⋆ be such an algebra structure on A × V . Then: (a, x) ⋆ (b, y) := (ab, ϑ(a, b) + a ⊲ y + x ⊳ b + x · y) (a, x) ⋆ (b, y) = ϕ−1(cid:0)ϕ(a, x) · ϕ(b, y)(cid:1)= ϕ−1(cid:0)(s(a) + x) · (s(b) + y)(cid:1) = ϕ−1(s(a) · s(b) + s(a) · y + x · s(b) + x · y) = (cid:0)ab, s(a) · s(b) − s(ab) + s(a) · y + x · s(b) + x · y(cid:1) = (cid:0)ab, ϑ(a, b) + a ⊲ y + x ⊳ b + x · y) as needed. Thus, ϕ : A ⋆ V → (E,·) is an isomorphism of algebras and we can see that it stabilizes V and co-stabilizes A. (cid:3) Using Proposition 1.4 we obtain the following result concerning the structure of finite dimensional algebras which indicates the crucial role played by Hochschild products. We can survey all algebras of a given dimension if we are able to compute various Hochschild systems starting with a simple algebra (whose structure is well-known due to the Wederburn-Artin theorem) and the associated Hochschild products. It is the associative algebra counterpart of a similar result from group theory [36, pages 283-284]. Corollary 1.5. Any finite dimensional algebra is isomorphic to an iteration of Hochschild products of the form (cid:16)···(cid:0)(S ⋆ V1) ⋆ V2(cid:1)⋆··· ⋆ Vt(cid:17), where S is a finite dimensional simple k-algebra, t is a positive integer and V1,··· , Vt are finite dimensional vector spaces. Proof. Let A be an algebra of dimension n. The proof goes by induction on n. If n = 1 then A ∼= k ∼= k ⋆ {0} and k is a simple algebra. Assume now that n > 1. If A is simple there is nothing to prove. On the contrary, if A has a proper two-sided ideal {0} 6= Vt 6= A, let π : A → A1 := A/Vt be the canonical projection. It follows from Proposition 1.4 that A ∼= A1 ⋆ Vt, for some Hochschild system of A1 by Vt. If A1 is simple the proof is finished; if A1 is not simple, we apply induction since dimk(A1) < n. (cid:3) The semidirect products of algebras characterize split epimorphism in this category: Corollary 1.6. An algebra map π : B → A is a split epimorphism in the category of algebras if and only if there exists an isomorphism of algebras B ∼= A#V , where V = Ker(π) and A#V is a semidirect product of algebras. HOCHSCHILD PRODUCTS AND NON-ABELIAN COHOMOLOGY FOR ALGEBRAS 9 Proof. First we note that for a semidirect product A#V , the canonical projection pA : A#V → A, pA(a, x) = a has a section that is an algebra map defined by sA(a) = (a, 0), for all a ∈ A. Conversely, let s : A → B be an algebra map such that π ◦ s = IdA. Then, the bilinear map ϑs constructed in the proof of Proposition 1.4 is the trivial map and hence the corresponding Hochschild product A ⋆ V is a semidirect product A#V . (cid:3) Proposition 1.4 shows that the classification part of the GE-problem reduces to the classification of all Hochschild products associated to all Hochschild systems of A by V . This is what we will do next. Lemma 1.7. Let Θ(A, V ) = (⊲, ⊳, ϑ, ·) and Θ′(A, V ) = (⊲′, ⊳′, ϑ′, ·′) be two Hochschild systems and A⋆V , respectively A⋆′ V , the corresponding Hochschild products. Then there exists a bijection between the set of all morphisms of algebras ψ : A ⋆ V → A ⋆′ V which stabilize V and co-stabilize A and the set of all linear maps r : A → V with r(1A) = 0 satisfying the following compatibilities for all a, b ∈ A, x, y ∈ V : (CH1) x · y = x ·′ y; (CH2) x ⊳ a = x ⊳′ a + x ·′ r(a); (CH3) a ⊲ x = a ⊲′ x + r(a) ·′ x; (CH4) ϑ(a, b) + r(ab) = ϑ′(a, b) + a ⊲′ r(b) + r(a) ⊳′ b + r(a) ·′ r(b) Under the above bijection the morphism of algebras ψ = ψr : A⋆V → A⋆′V corresponding to r : A → V is given by ψ(a, x) = (a, r(a) + x), for all a ∈ A, x ∈ V . Moreover, ψ = ψr is an isomorphism with the inverse given by ψ−1 r = ψ−r. Proof. It is an elementary fact that a linear map ψ : A × V → A × V stabilizes V and co-stabilizes A if and only if there exists a uniquely determined linear map r : A → V such that ψ(a, x) = (a, r(a) + x), for all a ∈ A, x ∈ V . Let ψ = ψr be such a linear map. We will prove that ψ : A ⋆ V → A ⋆′ V is an algebra map if and only if r(1A) = 0 and the compatibility conditions (CH1)-(CH4) hold. To start with it is straightforward to see that ψ preserve the unit (1A, 0) if and only if r(1A) = 0. The proof will be finished if we check that the following compatibility holds for all generators of A × V : ψ(cid:0)(a, x) ⋆ (b, y)(cid:1)= ψ(cid:0)(a, x)(cid:1) ⋆′ ψ(cid:0)(b, y)(cid:1) (7) By a straightforward computation it follows that (7) holds for the pair (a, 0), (b, 0) if and only if (CH4) is fulfilled while (7) holds for the pair (0, x), (a, 0) if and only if (CH2) is satisfied. Finally, (7) holds for the pair (a, 0), (0, x) and respectively (0, x), (0, y) if and only if (CH3) and respectively (CH1) hold. (cid:3) Lemma 1.7 leads to the following: Definition 1.8. Let A be an algebra and V a vector space. Two Hochschild systems Θ(A, V ) = (⊲, ⊳, ϑ, ·) and Θ′(A, V ) = (⊲′, ⊳′, ϑ′, ·′) are called cohomologous, and we denote this by Θ(A, V ) ≈ Θ′(A, V ), if and only if · = ·′ and there exists a linear map 10 A. L. AGORE AND G. MILITARU r : A → V such that r(1A) = 0 and for any a, b ∈ A, x, y ∈ V we have: x ⊳ a = x ⊳′ a + x ·′ r(a) a ⊲ x = a ⊲′ x + r(a) ·′ x ϑ(a, b) = ϑ′(a, b) − r(ab) + a ⊲′ r(b) + r(a) ⊳′ b + r(a) ·′ r(b) (8) (9) (10) As a conclusion of this section, we obtain the theoretical answer to the GE-problem: Theorem 1.9. Let A be an algebra, E a vector space and π : E → A a linear epi- morphism of vector spaces with V = Ker(π). Then ≈ defined in Definition 1.8 is an equivalence relation on the set HS(A, V ) of all Hochschild systems of A by V . If we denote by GH2 (A, V ) := HS(A, V )/ ≈, then the map GH2 (A, V ) → Gext (A, E), (⊲, ⊳, ϑ, ·) 7−→ A ⋆(⊳, ⊲, ϑ, ·) V (11) is bijective, where (⊲, ⊳, ϑ, ·) denotes the equivalence class of (⊲, ⊳, ϑ, ·) via ≈. Proof. Follows from Proposition 1.2, Proposition 1.4 and Lemma 1.7. (cid:3) ′ Computing the classifying object GH2 (A, V ), for a given algebra A and a given vector space V is a very difficult task. The first step in decomposing this object is suggested by the way the equivalence relation ≈ was introduced in Definition 1.8: it shows that two different associative algebra structures · and · on V give two different equivalence classes of the relation ≈ on HS(A, V ). Let us fix ·V an associative multiplication on V and denote by HS·V (A, V ) the set of all triples (cid:0)⊳, ⊲, ϑ(cid:1) such that (cid:0)⊲, ⊳, ϑ, ·V(cid:1)∈ HS(A, V ). Two triples (cid:0)⊳, ⊲, ϑ(cid:1) and (cid:0)⊳′, ⊲′, ϑ′(cid:1)∈ HS·V (A, V ) are ·V -cohomologous and we denote this by (cid:0)⊳, ⊲, ϑ(cid:1)≈·V (cid:0)⊳′, ⊲′, ϑ′(cid:1) if there exists a linear map r : A → V such that r(1A) = 0 and the compatibility conditions (8)-(10) are fulfilled for ·′ = ·V . Then ≈·V is an equivalence relation on HS·V (A, V ) and we denote by H2(cid:0)A, (V, ·V )(cid:1) the quotient set HS·V (A, V )/ ≈·V . The non-abelian cohomology H2(cid:0)A, (V, ·V )(cid:1) classifies all extensions of the unital associative algebra A by a fixed associative algebra (V, ·V ). This can be seen as a Schreier type theorem for associative algebras: we mention that [26, Theorem 6.2] (see also [35, Proposition 3.7]) follows as a special case of Corollary 1.10 if we let ·V to be the trivial map. Corollary 1.10. Let A be a unital associative algebra and (V,·V ) an associative multi- plication on V . Then, the map (⊲, ⊳, ϑ) 7−→ A ⋆(⊳, ⊲, ϑ, ·V ) V H2(cid:0)A, (V, ·V )(cid:1)→ Ext (A, (V, ·V )), (12) is bijective, where Ext (A, (V, ·V )) is the set of equivalence classes of all unital associa- tive algebras that are extensions of the algebra A by (V,·V ) and (⊲, ⊳, ϑ) denotes the equivalence class of (⊲, ⊳, ϑ) via ≈l. The above considerations give also the following decomposition of GH2 (A, V ): Corollary 1.11. Let A be an algebra, E a vector space and π : E → A an epimorphism of vector spaces with V = Ker(π). Then GH2 (A, V ) = ⊔·V H2(cid:0)A, (V, ·V )(cid:1) (13) HOCHSCHILD PRODUCTS AND NON-ABELIAN COHOMOLOGY FOR ALGEBRAS 11 where the coproduct on the right hand side is in the category of sets over all possible associative algebra structures ·V on the vector space V . By looking at formula (13) one can see that computing GH2 (A, V ) is a very laborious task in which the first major barrier is describing all associative multiplications on V . The complexity of the computations involved increases along side with dim(V ). Among all components of the coproduct in (13) the simplest one is that corresponding to the trivial associative algebra structure on V , i.e. x ·V y := 0, for all x, y ∈ V . We shall denote this trivial algebra structure on V by V0 := (V, ·V = 0) and we shall prove that H2(cid:0)A, V0(cid:1) is the coproduct of all classical second cohomological groups. Indeed, let HS0(A, V0) be the set of all triples (cid:0)⊳, ⊲, ϑ(cid:1) such that (cid:0)⊲, ⊳, ϑ, ·V := 0(cid:1)∈ HS(A, V ). Example 1.3 shows that a triple (cid:0)⊳, ⊲, ϑ(cid:1)∈ HS0(A, V0) if and only if (V, ⊲, ⊳) is an A- bimodule and ϑ : A × A → V is a normalized 2-cocycle. Two triples (cid:0)⊳, ⊲, ϑ(cid:1) and (cid:0)⊳′, ⊲′, ϑ′(cid:1)∈ HS0(A, V0) are 0-cohomologous (cid:0)⊳, ⊲, ϑ(cid:1)≈0 (cid:0)⊳′, ⊲′, ϑ′(cid:1) if and only if ⊳ = ⊳′, ⊲ = ⊲′ and there exists a linear map r : A → V such that r(1A) = 0 and ϑ(a, b) = ϑ′(a, b) − r(ab) + a ⊲ r(b) + r(a) ⊳ b (14) for all a, b ∈ A -- these are the conditions remaining from Definition 1.8 applied for the trivial multiplication · := 0. The equalities ⊳ = ⊳′ and ⊲ = ⊲′ show that two different A-bimodule structures over V give different equivalence classes in the classifying object H2(cid:0)A, V0(cid:1). Thus, for computing it we can also fix (V, ⊳, ⊲) an A-bimodule structure over V and consider the set Z2 i.e. bilinear maps ϑ : A × A → V satisfying (H5) and the first condition of (H0). Two normalized 2-cocycles ϑ and ϑ′ are cohomologous ϑ ≈0 ϑ′ if and only if there exists a linear map r : A → V such that r(1A) = 0 and (14) holds. ≈0 is an equivalence relation on the set Z2 (⊳, ⊲) (A, V0)/ ≈0 is just the classical second Hochschild cohomological group which we denote by H2 (⊳, ⊲) (A, V0). All the above considerations prove the following: (⊳, ⊲) (A, V0) of all normalized Hochschild 2-cocycles: (⊳, ⊲) (A, V0) and the quotient set Z2 Corollary 1.12. Let A be an algebra and V a vector space viewed with the trivial asso- ciative algebra structure V0. Then: where the coproduct on the right hand side is in the category of sets over all possible A-bimodule structures (⊳, ⊲) on the vector space V . H2(cid:0)A, V0(cid:1) = ⊔(⊳, ⊲) H2 (⊳, ⊲) (A, V0) (15) 2. Co-flag algebras. Examples. In this section we apply the theoretical results obtained in Section 1 for some concrete examples: more precisely, for a given algebra A we shall classify all unital associative algebras B such that there exists a surjective algebra map π : B → A having a 1- dimensional kernel, which as a vector space will be assumed to be k. First, we shall compute GH2 (A, k): it will classify all these algebras up to an isomorphism which stabilizes k and co-stabilizes A. Then, we will compute the second classifying object, denoted by HOC (A, k), which will provide the classification of these algebras only up 12 A. L. AGORE AND G. MILITARU to an isomorphism. Computing both classifying objects is the key step in a recursive algorithm for describing and classifying the new class of algebras defined as follows: Definition 2.1. Let A be an algebra and E a vector space. A unital associative algebra structure ·E on E is called a co-flag algebra over A if there exists a positive integer n and a finite chain of surjective morphisms of algebras An := (E,·E ) (16) such that dimk(Ker(πi)) = 1, for all i = 1,··· , n. A finite dimensional algebra is called a co-flag algebra if it is a co-flag algebra over the unital algebra k. π1−→ A0 := A πn−→ An−1 πn−1−→ An−2 ··· π2−→ A1 By applying successively Proposition 1.4 we obtain that a co-flag algebra over an algebra A is isomorphic to an iteration of Hochschild products of the form(cid:0)···(cid:0)(A⋆k)⋆k(cid:1)⋆···⋆k(cid:1), where the 1-dimensional vector space k appears n times in the above product. The tools used for describing co-flag algebras are the following: Definition 2.2. Let A be an algebra. A co-flag datum of the first kind of A is a triple (λ, Λ, ϑ) consisting of two algebra maps1 λ, Λ : A → k and a bilinear map ϑ : A× A → k satisfying the following compatibilities for any a, b, c ∈ A: ϑ(a, 1A) = ϑ(1A, a) = 0, ϑ(a, bc) − ϑ(ab, c) = ϑ(a, b)Λ(c) − ϑ(b, c)λ(a) (17) A co-flag datum of the second kind of A is a pair (λ, u) consisting of a linear map λ : A → k such that λ(1A) = 1 and a non-zero scalar u ∈ k∗. We denote by CF1 (A) (resp. CF2 (A)) the set of all co-flag data of the first (resp. second) kind of A and by CF (A) := CF1 (A) ⊔ CF2 (A) their coproduct; the elements of CF (A) are called co-flag data of A. The set of co-flag data CF (A) parameterizes the set of all Hochschild systems of A by a 1-dimensional vector space. The next result also describes the first algebra A1 from the exact sequence (16) in terms depending only on A. Proposition 2.3. Let A be an algebra. Then there exists a bijection HS (A, k) ∼= CF (A) between the set of all Hochschild systems of A by k and the set of all co-flag data of A given such that the Hochschild product A ⋆ k associated to (λ, Λ, ϑ) ∈ CF1 (A) is the algebra denoted by A(λ,Λ,ϑ) with the multiplication given for any a, b ∈ A, x, y ∈ k by: (a, x) ⋆ (b, y) = (cid:0)ab, ϑ(a, b) + λ(a)y + Λ(b)x(cid:1) (18) while the Hochschild product A ⋆ k associated to (λ, u) ∈ CF2 (A) is the algebra denoted by A(λ,u) with the multiplication given for any a, b ∈ A, x, y ∈ k by: (a, x) ⋆ (b, y) = (cid:0)ab, u−1(cid:0)λ(a)λ(b) − λ(ab)(cid:1)+λ(a)y + λ(b)x + u xy(cid:1) (19) Proof. We have to compute the set of all bilinear maps ⊲ : A × k → k, ⊳ : k × A → k, ϑ : A × A → k and · : k × k → k satisfying the compatibility conditions (H0)-(H8) of Proposition 1.2. Since k has dimension 1 there exists a bijection between the set of all Hochschild datums (cid:0)⊲, ⊳, ϑ, ·(cid:1) of A by k and the set of all 4-tuples (Λ, λ, ϑ, u) consisting of two linear maps Λ, λ : A → k, a bilinear map ϑ : A × A → k and a scalar 1Recall that we assume the algebra maps λ : A → k to be unit preserving, i.e. λ(1A) = 1. HOCHSCHILD PRODUCTS AND NON-ABELIAN COHOMOLOGY FOR ALGEBRAS 13 u ∈ k. The bijection is given such that the Hochschild datum (cid:0)⊲, ⊳, ϑ, ·(cid:1) corresponding to (Λ, λ, ϑ, u) is defined as follows: a ⊲ x := λ(a)x, x ⊳ a := Λ(a)x, x · y := u xy for all a ∈ A and x, y ∈ k. Now, axiom (H0) holds if and only if ϑ(a, 1A) = ϑ(1A, a) = 0 and λ(1A) = Λ(1A) = 1. Axioms (H1), (H3), (H4) and (H8) are trivially fulfilled. Axiom (H5) is equivalent to ϑ(a, bc) − ϑ(ab, c) = ϑ(a, b)Λ(c) − ϑ(b, c)λ(a), axiom (H6) is equivalent to (20) while axiom (H7) is equivalent to Λ(ab) = Λ(a)Λ(b) − u ϑ(a, b). Finally, axiom (H2) is equivalent to uΛ(a) = uλ(a), for all a ∈ A. A discussion on u is imposed by the last compatibility condition and the conclusion follows easily: CF1 (A) corresponds to the case when u = 0 and this will give rise to the algebras A ⋆ k = A(λ,Λ,θ). The case CF2 (A) corresponds to u 6= 0; in this case Λ = λ and the cocycle ϑ is implemented by u and λ via the formula ϑ(a, b) := u−1(cid:0)λ(a)λ(b) − λ(ab)(cid:1), for all a, b ∈ A, that arises from (20). Moreover, we can easily check that axiom (17) is trivially fulfilled for ϑ defined as above. The algebra A(λ,u) is just the Hochschild product A ⋆ k associated to this context. (cid:3) λ(ab) = λ(a)λ(b) − u ϑ(a, b) Remarks 2.4. (1) The first family of Hochschild products A(λ,Λ,θ) constructed in Propo- sition 2.3 corresponds to the classical case in which k ∼= 0 × k is a two-sided ideal of null square in the algebra A(λ,Λ,θ). The algebras A(λ,Λ,θ) will be classified up to an isomor- phism in Theorem 2.8 below. For the new families of algebras A(λ,u) the kernel of the canonical projection πA : A(λ,u) → k is equal to k ∼= 0 × k and this is not a null square ideal since (0, 1) ⋆ (0, 1) = (0, u) 6= (0, 0). Let (λ, u) ∈ CF2 (A) be a co-flag datum of the second kind of A. Taking into account the multiplication on A(λ, u) given by (19) we can easily prove that the map: ϕ : A(λ, u) → A × k, ϕ(a, x) := (a, λ(a) + u x) (21) for all a ∈ A and x ∈ k is an isomorphism of algebras (which does not stabilize k, if u 6= 1), where A × k is the usual direct product of algebras. The inverse of ϕ is given by ϕ−1(a, x) = (cid:0)a, u−1(x − λ(a)(cid:1), for all a ∈ A and x ∈ k. We will now describe the algebra A(λ,Λ,ϑ) and A(λ,u) by generators and relations. The elements of A will be seen as elements in A × k via the identification a = (a, 0) and we denote by f := (0A, 1) ∈ A × k. Let {ei i ∈ I} be a basis of A as a vector space over k. Then the algebra A(λ,Λ,ϑ) is the vector space having {f, ei i ∈ I} as a basis and the multiplication ⋆ given for any i ∈ I by: ei ⋆ ej = ei ·A ej + ϑ(ei, ej) f, f 2 = 0, ei ⋆ f = λ(ei) f, f ⋆ ei = Λ(ei) f (22) where ·A denotes the multiplication on A. The algebra A(λ,u) is the vector space having {f, ei i ∈ I} as a basis and the multiplication ⋆ given for any i ∈ I by: ei ⋆ej = ei·A ej +u−1(cid:0)λ(ei)λ(ej)−λ(ei·A ej)(cid:1) f, f 2 = u f, ei ⋆f = f ⋆ei = λ(ei) f (23) Using Proposition 2.3, Proposition 1.4 and the isomorphism A(λ, u) ≃ A × k we obtain: 14 A. L. AGORE AND G. MILITARU Corollary 2.5. Let A be an algebra. A unital associative algebra B has a surjective algebra map B → A → 0 whose kernel is 1-dimensional if and only if B is isomorphic to A × k or A(λ,Λ,ϑ), for some (λ, Λ, ϑ) ∈ CF1 (A). We are now able to compute the classifying object GH2 (A, k). Proposition 2.6. Let A be an algebra. Then, GH2 (A, k) ∼= (cid:16)CF1 (A)/ ≈1 (cid:17) ⊔ k∗ where ≈1 is the following equivalence relation on CF1 (A): (λ, Λ, ϑ) ≈1 (λ′, Λ′, ϑ′) if and only if λ = λ′, Λ = Λ′ and there exists a linear map t : A → k such that for any a, b ∈ A: (24) ϑ(a, b) = ϑ′(a, b) − t(ab) + λ′(a)t(b) + Λ′(b)t(a) Proof. It follows from Theorem 1.9 and Proposition 2.3 that GH2 (A, k) ∼= (cid:16)CF1 (A)/ ≈1 (cid:17) ⊔ (cid:16)CF2 (A)/ ≈2 (cid:17) where the equivalence relation ≈i on CFi (A), for i = 1, 2, is just the equivalence relation ≈ from Definition 1.8 written for the sets CFi (A) via the bijection HS (A, k) ∼= CF (A) given in Proposition 2.3. The equivalence relation ≈ written on the set of all co-flag data of the first kind takes precisely the form from the statement -- we mention that a linear map t satisfying (24) has the property that t(1A) = 0. The equivalence relation ≈ written on CF2 (A) takes the following form: (λ, u) ≈2 (λ′, u′) if and only if u = u′ and there exists a linear map t : A → k such that for any a ∈ A we have: λ(a) = λ′(a) + t(a) u′ (25) Now, if we fix a unit preserving linear map λ0 : A → k we obtain that the set {(λ0, u) u ∈ k∗} is a system of representatives for the equivalence relation ≈2 on CF2 (A) and hence CF2 (A)/ ≈2 ∼= k∗, which finishes the proof. The way ≈1 is defined in Proposition 2.6 indicates the decomposition of (cid:16)CF1 (A)/ ≈1 (cid:17) as follows: for two fixed algebra maps (λ, Λ) ∈ Alg(A, k) we shall denote by Z2 (λ, Λ) (A, k) the set of all normalized (λ, Λ)-cocycles; that is, the set of all bilinear maps ϑ : A×A → k satisfying the following compatibilities for any a, b, c ∈ A: (cid:3) ϑ(a, 1A) = ϑ(1A, a) = 0, ϑ(a, bc) − ϑ(ab, c) = ϑ(a, b)Λ(c) − ϑ(b, c)λ(a) Two (λ, Λ)-cocycles ϑ, ϑ′ : A × A → k are equivalent ϑ ≈(λ, Λ) exists a linear map t : A → k such that 1 ϑ′ if and only if there ϑ(a, b) = ϑ′(a, b) − t(ab) + λ(a)t(b) + Λ(b)t(a) If we denote H2 (λ, Λ) (A, k) := Z2 (λ, Λ) (A, k)/ ≈(λ, Λ) 1 for all a, b ∈ A. following decomposition of GH2 (A, k): (26) we obtain the HOCHSCHILD PRODUCTS AND NON-ABELIAN COHOMOLOGY FOR ALGEBRAS 15 Corollary 2.7. Let A be an algebra. Then, GH2 (A, k) ∼= (cid:16)⊔(λ,Λ) H2 (λ, Λ) (A, k)(cid:17)⊔ k∗ (27) where the coproduct on the right hand side is in the category of sets over all possible algebra maps λ, Λ : A → k. The classifying object GH2 (A, k) computed in Corollary 2.7 classifies all Hochschild products A ⋆ k up to an isomorphism of algebras which stabilizes k and co-stabilizes A. In what follows we will consider a less restrictive classification: we denote by HOC (A, k) the set of algebra isomorphism classes of all Hochschild products A⋆k. Two cohomologous Hochschild products A ⋆ k and A ⋆′ k are of course isomorphic and therefore there exists a canonical projection GH2 (A, k) ։ HOC (A, k) between the two classifying objects. Next we compute HOC (A, k). Theorem 2.8. Let A be an algebra. Then there exists a bijection: HOC (A, k) ∼= (cid:16)CF1 (A)/ ≡ (cid:17)⊔{ A × k } where ≡ is the equivalence relation on CF1 (A) defined by: (λ, Λ, ϑ) ≡ (λ′, Λ′, ϑ′) if and only if there exists a triple (s0, ψ, r) ∈ k∗ × AutAlg(A) × Homk(A, k) consisting of a non-zero scalar s0 ∈ k∗, an algebra automorphism ψ of A and a linear map r : A → k such that for any a, b ∈ A we have: (28) (29) (30) (31) (32) λ = λ′ ◦ ψ, ϑ(a, b) s0 = ϑ′(cid:0)ψ(a), ψ(b)(cid:1)+λ(a)r(b) + Λ(b)r(a) − r(ab) Λ = Λ′ ◦ ψ Proof. Corollary 2.5 shows that any Hochschild product A⋆k is isomorphic to A(λ,Λ,ϑ), for some (λ, Λ, ϑ) ∈ CF1 (A) or to A(λ′,u′), for some (λ′, u′) ∈ CF2 (A). Since A(λ′,u′) ∼= A×k, the proof relies on the following two steps: (1) Let (λ, Λ, ϑ) and (λ′, Λ′, ϑ′) ∈ CF1 (A). Then, there exists a bijection between the set of all algebra isomorphisms ϕ : A(λ,Λ,ϑ) → A(λ′,Λ′,ϑ′) and the set of all triples (s0, ψ, r) ∈ k∗ × AutAlg(A)× Homk(A, k) satisfying the compatibility conditions (29) and (30). The bijection is given such that the algebra isomorphism ϕ = ϕ(s0, ψ, r) associated to (s0, ψ, r) is defined for any a ∈ A and x ∈ k by: ϕ(s0, ψ, r)(a, x) = (cid:0)ψ(a), r(a) + xs0(cid:1) (2) The algebras A(λ,Λ,ϑ) and A(λ′,u′) ∼= A × k are not isomorphic. We start by proving (1); although this is more than we need for proving our theorem, this more general statement will be used later on in computing the automorphism groups for the algebras A(λ,Λ,ϑ). First we note that there exists a bijection between the set of all linear maps ϕ : A × k → A × k and the set of quadruples (s0, β0, ψ, r) ∈ k × A × Homk(A, A)× Homk(A, k) given such that the linear map ϕ = ϕ(s0, β0, ψ, r) associated to (s0, β0, ψ, r) is given for any a ∈ A and x ∈ k by: ϕ(a, x) = (cid:0)ψ(a) + x β0, r(a) + x s0(cid:1) 16 A. L. AGORE AND G. MILITARU We will prove now the following technical fact: a linear map given by (32) is an iso- morphism of algebras from A(λ,Λ,ϑ) to A(λ′,Λ′,ϑ′) if and only if β0 = 0, s0 6= 0, ψ is an algebra automorphism of A and (29)-(30) hold. Taking into account the multiplication on A(λ,Λ,ϑ) given by (18), we can easily obtain that ϕ(cid:0)(0, x) ⋆ (0, y)(cid:1)= ϕ(0, x) ⋆′ ϕ(0, x) if and only if β0 = 0, where by ⋆′ we denote the multiplication of A(λ′,Λ′,ϑ′). Hence, in order for ϕ to be an algebra map it should take the following simplified form for any a ∈ A and x ∈ k: (33) for some triple (s0, ψ, r) ∈ k × Homk(A, A) × Homk(A, k). Next we prove that a linear map given by (33) is an algebra morphism from A(λ,Λ,ϑ) to A(λ′,Λ′,ϑ′) if and only if ψ : A → A is an algebra map and the following compatibilities are fulfilled for any a, b ∈ A: ϕ(a, x) = (cid:0)ψ(a), r(a) + x s0(cid:1) λ(a) s0 = λ′(cid:0)ψ(a)(cid:1) s0, r(ab) + ϑ(a, b) s0 = ϑ′(cid:0)ψ(a), ψ(b)(cid:1)+λ′(ψ(a))r(b) + Λ′(ψ(b))r(a) Λ(a) s0 = Λ′(cid:0)ψ(a)(cid:1) s0 (34) (35) Indeed, ϕ preserves the unit (1A, 0) if and only if ψ(1A) = 1A and r(1A) = 0. On the other hand we can prove that the first (resp. the second) compatibility of (34) is exactly the 0 , ψ−1, −(r◦ψ−1)s−1 condition ϕ(cid:0)(a, 0) ⋆ (0, x)(cid:1)= ϕ(a, 0) ⋆′ ϕ(0, x) (resp. ϕ(cid:0)(0, x) ⋆ (a, 0)(cid:1)= ϕ(0, x) ⋆′ ϕ(a, 0)). Finally, the condition ϕ(cid:0)(a, 0) ⋆ (b, 0)(cid:1)= ϕ(a, 0) ⋆′ ϕ(b, 0) is equivalent to the fact that ψ is an algebra endomorphism of A and (35) holds. Finally, the condition r(1A) = 0 follows by taking a = b = 1A in (35). Step (1) is finished if we prove that an algebra map ϕ = ϕ(s0, ψ, r) given by (33) is bijective if and only if s0 6= 0 and ψ is an automorphism of A. Assume first that s0 6= 0 and ψ is bijective with the inverse ψ−1. Then, we can see that ϕ(s0, ψ, r) is an isomorphism of algebras with the inverse given by ϕ−1 (s0, ψ, r) := 0 ). Conversely, assume that ϕ is bijective. Then its inverse ϕ−1 is ϕ(s−1 an algebra map and thus has the form ϕ−1(a, x) = (ψ′(a), r′(a) + xs′ 0), for some triple 0, r′, ψ′). If we write ϕ−1 ◦ ϕ(0, 1) = (0, 1) we obtain that s0s′ (s′ 0 = 1 i.e. s0 is invertible in k. In the same way ϕ−1 ◦ ϕ(a, 0) = (a, 0) = ϕ ◦ ϕ−1(a, 0) gives that ψ is bijective and ψ′ = ψ−1. We will prove now the assertion from step (2). Assume that ϕ : A(λ,Λ,ϑ) → A(λ′,u′) is an algebra map. Thus, ϕ is given by (32), for some quadruple (s0, β0, ψ, r). Now, we can see that the algebra map condition ϕ(cid:0)(0, x) ⋆ (0, y)(cid:1)= ϕ(0, x) ⋆′ ϕ(0, y) holds if and only if β0 = 0 and s0 = 0, where ⋆′ denotes the multiplication on the algebra A(λ′,u′). Hence, ϕ takes the form ϕ(a, x) = (cid:0)ψ(a), r(a)(cid:1), for all a ∈ A and x ∈ k. Such a map is never an isomorphism of algebras since is not injective and thus two algebras of the form A(λ,Λ,ϑ) and A(λ′,u′) are never isomorphic. The theorem is now completely proved. (cid:3) Remark 2.9. The compatibility condition (30) of Theorem 2.8 highlights the difficulty of classifying co-flag algebras over a given algebra A: it generalizes the classical Kroneker- Williamson equivalence of bilinear forms whose classification was started in [39] and finished in [27] over algebraically closed fields. We recall that two bilinear forms ϑ and ϑ′ on a vector space A are called isometric if there exists a linear automorphism ψ ∈ Autk(A) such that ϑ(x, y) = ϑ′(ψ(x), ψ(y)), for all x, y ∈ A. If the cocycles ϑ and ϑ′ are isometric as bilinear forms on A and ψ is an algebra automorphism of A, then HOCHSCHILD PRODUCTS AND NON-ABELIAN COHOMOLOGY FOR ALGEBRAS 17 (30) holds by taking s0 := 1 and r := 0, the trivial map. For future references to the problem of classifying bilinear forms up to an isometry we refer to [27]. Theorem 2.8 can be applied to classify all semidirect products of algebras of the form A#k. We recall from Example 1.3 that a semidirect product A#k is just a Hochschild product A(λ,Λ,ϑ) = A ⋆ k having a trivial cocycle. The algebra obtained in this way will be denoted by A(λ, Λ). Directly from the proof of Theorem 2.8 we obtain: Corollary 2.10. Let A be an algebra, (λ, Λ) and (λ′, Λ′) two pairs consisting of algebra maps from A to k. Then there exists an isomorphism of algebras A(λ, Λ) ∼= A(λ′, Λ′) if and only if there exists ψ ∈ AutAlg(A) such that λ = λ′ ◦ ψ and Λ = Λ′ ◦ ψ. An interesting special case occurs for the algebras A such that there is no algebra map A → k (e.g. the classical Weyl algebra W1(k) = k < x, y xy − yx = 1 > or the matrix algebra Mn(k), for n ≥ 2). Using Proposition 2.6 and Theorem 2.8 we obtain: Corollary 2.11. Let A be an algebra for which there is no algebra map A → k. Then GH2 (A, k) ∼= k∗, HOC (A, k) = { A × k } In particular, up to an isomorphism, the only algebra B for which there exists a surjective algebra map B → A having a 1-dimensional kernel is the direct product A × k. Determining the automorphism group of a given algebra is an old and very difficult problem, intensively studied in invariant theory (see [13] and their references). As already mentioned, the first step proved in the proof of Theorem 2.8 allows us to compute the automorphism group AutAlg (A(λ,Λ,ϑ)), for any (λ, Λ, ϑ) ∈ CF1 (A). Let k∗ be the units group of k, k∗ × AutAlg(A) the direct product of groups and (A∗, +) the underlying abelian group of the linear dual A∗ = Homk(A, k). Then the map given for any s0 ∈ k∗, ψ ∈ AutAlg(A) and r ∈ A∗ by: ζ : k∗ × AutAlg(A) → AutGr (A∗, +), ζ(s0, ψ) (r) := s−1 0 r ◦ ψ is a morphism of groups. Thus, we can construct the semidirect product of groups A∗ ⋉ζ (cid:0)k∗ × AutAlg(A)(cid:1) associated to ζ. The next result shows that AutAlg(A(λ,Λ,ϑ)) is isomorphic to a certain subgroup of the semidirect product A∗ ⋉ζ (cid:0)k∗ × AutAlg(A)(cid:1). Corollary 2.12. Let A be an algebra, (λ, Λ, ϑ) ∈ CF1 (A) a co-flag datum of the first kind of A and let G(cid:0)A, (λ, Λ, ϑ)(cid:1) be the set of all triples (s0, ψ, r) ∈ k∗ × AutAlg(A)× A∗ such that for any a, b ∈ A: λ = λ ◦ ψ, Λ = Λ ◦ ψ, ϑ(a, b) s0 = ϑ(cid:0)ψ(a), ψ(b)(cid:1)+λ(a)r(b) + Λ(b)r(a) − r(ab) Then, there exists an isomorphism of groups AutAlg(A(λ,Λ,ϑ)) ∼= G(cid:0)A, (λ, Λ, ϑ)(cid:1), where G(cid:0)A, (λ, Λ, ϑ)(cid:1) is a group with respect to the following multiplication: 0, ψ ◦ ψ′, r ◦ ψ′ + s0r′) for all (s0, ψ, r), (s′ (s0, ψ, r) · (s′ 0, ψ′, r′) ∈∈ G(cid:0)A, (λ, Λ, ϑ)(cid:1). Moreover, the canonical map 0, ψ′, r′) := (s0s′ (36) G(cid:0)A, (λ, Λ, ϑ)(cid:1)−→ A∗ ⋉ζ (cid:0)k∗ × AutAlg(A)(cid:1), in an injective morphism of groups. (s0, ψ, r) 7→ (cid:0)s−1 0 r, (s0, ψ)(cid:1) 18 A. L. AGORE AND G. MILITARU Proof. The fact that G(cid:0)A, (λ, Λ, ϑ)(cid:1) is a group with respect to the multiplication (36) follows by a straightforward computation which is left to the reader: the unit is (1, IdA, 0) and the inverse of (s0, ψ, r) is (s−1 0 (r ◦ ψ−1)). The first statement follows from the proof of Theorem 2.8, step (1), since ϕ(s0,ψ,r)◦ ϕ(s′ 0, ψ◦ψ′, r◦ψ′+s0r′), where ϕ(s0,ψ,r) is an automorphism of A(λ,Λ,ϑ) given by (31). The last assertion follows by a routine computation. (cid:3) 0 , ψ−1, −s−1 0,ψ′,r′) = ϕ(s0s′ (n,n) denote the Kroneker symbols. Now we shall provide some explicit examples. The first example shows the limitations of the classical approach for the extension problem: there is no (1 + n2)-dimensional algebra with an algebra projection on the matrix algebra Mn(k) whose kernel is a null square ideal, but there exists a family of (1 + n2)-dimensional algebras which project on the matrix algebra Mn(k). We denote by {eij i, j = 1,··· , n} the canonical basis of Mn(k), i.e. eij is the matrix having 1 in the (i, j)th position and zeros elsewhere while j and δ(i,j) δk Example 2.13. Let n ≥ 2 be a positive integer. Then, GH2 (Mn(k), k) ∼= k∗ and the equivalence classes of all (1 + n2)-dimensional algebras with an algebra projection on Mn(k) are the following algebras denoted by Mn(k)u and defined for any u ∈ k∗ as the vector space having {f, eij i, j = 1,··· , n} as a basis and the multiplication given for any i, j = 1,··· , n by: f 2 := u f, eij ⋆ f = f ⋆ eij := δ(i,j) eij ⋆ ekl := δj (n,n) f, Furthermore, HOC (Mn(k), k) = {Mn(k) × k}. The result follows by applying Proposition 2.6 and Corollary 2.11 since there is no unitary algebra map Mn(k) → k. If we consider λ0 : Mn(k) → k defined by λ0(eij) := δ(i,j) (n,n), for all i, j = 1,··· , n, then {(λ0, u) u ∈ k∗} is a system of representatives for the equivalence relation ≈2. The algebra Mn(k)u associated to the pair (λ0, u) is the vector space having {f, eij i, j = 1,··· , n} as a basis while the multiplication given by (23) comes down to the one in the statement. k eil + u−1(cid:0)δ(i,j) (n,n)δ(k,l) (n,n) − δk j δ(i,l) (n,n)(cid:1) f An interesting example through the subtle arithmetics involved in the classification of the corresponding Hochschild products is the group algebra k[Cn], where for a positive integer n ≥ 2 we denote by Cn the cyclic group of order n generated by d. We introduce the following notation: for any i, j = 1,··· , n − 1 we shall denote by i ∗ j the positive integer given by i ∗ j := (cid:26) i + j if i + j − n if j + i < n j + i ≥ n In what follows Un(k) := {ω ∈ k ωn = 1} denotes the cyclic group of n-th roots of unity in k and A(n, k) := {x ∈ U (k[Cn]) ψ : k[Cn] → k[Cn], ψ(di) = xi, i = 0, 1,··· , n − 1, is an algebra automorphism}. Example 2.14. Let k be a field such that n is invertible in k. Then: GH2 (k[Cn], k) ∼= (cid:0)Un(k) × Un(k)(cid:1) ⊔ k∗ HOCHSCHILD PRODUCTS AND NON-ABELIAN COHOMOLOGY FOR ALGEBRAS 19 and the equivalence classes of (n + 1)-dimensional algebras with an algebra projection on k[Cn] are the families of algebras having {f, di i = 1,··· , n} as a basis over k and the multiplication ⋆ defined for any (α, β) ∈ Un(k) × Un(k), u ∈ k∗ and i, j = 1,··· , n by: k[Cn](α, β) : k[Cn]u : di ⋆ dj = di+j, di ⋆ dj = di+j + u−1(δn f 2 = 0, di ⋆ f = αi f, f ⋆ di = βi f i δn j − δn i+j)f, f 2 = uf, di ⋆ f = f ⋆ di = δn i f Furthermore, there exists a bijection HOC (k[Cn], k) ∼= (cid:0)Un(k) × Un(k)/ ≡ (cid:1)⊔{k[Cn] × k} where ≡ is the following equivalence relation on Un(k)× Un(k): two pairs (α, β), (α′, β′) of n-th roots of unity in k are equivalent (α, β) ≡ (α′, β′) if and only if there exists x0 + x1 d + ··· + xn−1 dn−1 ∈ A(n, k) such that α′ = x0 + x1 α + ··· + xn−1 αn−1, β′ = x0 + x1 β + ··· + xn−1 βn−1 (37) To start with we point out that the algebra maps k[Cn] → k are parameterized by the cyclic group of n-th roots of unity in k. Consider α, β ∈ Un(k) which implement λ and respectively Λ, i.e. λ(d) = α and Λ(d) = β. We are left to compute the set of all (λ, Λ)- cocycles. To this end we denote ϑ(di, d) := ξi, i = 1,··· , n − 1 and we will see that these elements will allow us to completely determine the cocycle ϑ : k[Cn] × k[Cn] → k. Indeed, by writing down (17) for triples of the form (di, dj, d) and using induction we obtain the following general formula: ϑ(di, dj) = j−1 Xk=0 ξi∗k βj−1−k −(cid:0) j−1 Xl=1 ξl βj−1−l(cid:1)αi for all i, j = 1,··· , n − 1, where ξ0 := 0. Furthermore, by writing down (17) for triples of the form (di, dn−i, di) and using the above formula for ϑ we obtain the following compatibility which needs to be fulfilled for any i = 1,··· , n − 1: (αi − βi)(ξn−1 + ξn−2 β + ... + ξ1 βn−2) = 0 Therefore we distinguish two cases, namely: α = β or α 6= β and ξn−1 + ξn−2 β + ... + ξ1 βn−2 = 0. It follows that CF1 (k[Cn]) ∼= (Un(k) × kn−1) ∪ {(α, β, ξ1, ξ2, ..., ξn−2) ∈ Un(k)× Un(k)× kn−2 α 6= β} and the bijection associates to any (α, δ1, δ2, ..., δn−1) ∈ Un(k)×kn−1 the co-flag datum of the first kind (λα, Λα, ϑδ) given for all i, j = 1,··· , n−1 by: λ(d) = Λ(d) := α, ϑδ(di, dj) := j−1 Xk=0 j−1 δi∗k αj−1−k −(cid:0) Xl=1 δl αi+j−1−l(cid:1) where δ0 = 0, and to any (β, γ, ξ1, ξ2, ..., ξn−2) ∈ Un(k) × Un(k) × kn−2, with β 6= γ, associates the co-flag datum of the first kind (λβ, Λγ, ϑξ) given for any i, j = 1,··· , n−1 by: λβ(d) := β, Λγ(d) := γ, ϑξ(di, dj) := j−1 Xk=0 ξi∗k γj−1−k −(cid:0) j−1 Xl=1 ξl γj−1−l(cid:1)βi 20 A. L. AGORE AND G. MILITARU where ξ0 = 0 and ξn−1 = −(cid:0)ξn−2 β + ... + ξ1 βn−2(cid:1). It is now obvious that a (λα, Λα)- cocycle is never equivalent to a (λβ, Λγ)-cocycle for any α, β, γ ∈ Un(k), β 6= γ. Fur- thermore, by a rather long but straightforward computation it can be easily seen that for all α ∈ Un(k), any (λα, Λα)-cocycle, say ϑδ, is equivalent (in the sense of (26)) to the trivial cocycle via the linear map t : k[Cn] → k defined by t(1) := 0 and for any i = 2,··· , n − 1: t(d) := n−1 n−1 Xj=1 αjδn−j, t(di) := n−1iαi−1 n−1 Xj=1 αjδn−j − i−2 Xj=0 αjδi−1−j (λα, Λα) (k[Cn], k) ∼= {(α, α) α ∈ Un(k)}. A similar state- Therefore we have ⊔(λα,Λα) H2 ment holds for the second family of co-flag data of the first kind: for all β, γ ∈ Un(k), with β 6= γ, any (λβ, Λγ)-cocycle, say ϑξ, is equivalent to the trivial cocycle via the linear map t : k[Cn] → k defined by t(1) := t(d) := 0 and t(di) := −Pi−2 j=0 ξi−1−j γj, for all (k[Cn], k) ∼= {(α, β) ∈ Un(k) × Un(k) α 6= β}. i = 2,··· , n − 1. Thus ⊔(λβ ,Λγ ) H2 Therefore, we have proved that (cid:16)CF1 (k[Cn])/ ≈1 (cid:17)∼= Un(k)×Un(k) and the correspond- ing co-flag algebras are those denoted by k[Cn](α, β). For the co-flag data of the second kind of k[Cn] we choose the set of representatives {(λ0, u) u ∈ k∗} for the equivalence relation ≈2, where λ0 : k[Cn] → k is given by λ0(di) = δn i , for all i = 1,··· , n. They give rise to the co-flag algebras denoted by k[Cn]u. The conclusion now follows from Corollary 2.7. Finally, the assertion regarding HOC (k[Cn], k) is an easy consequence of Corollary 2.10. (λβ , Λγ ) Remark 2.15. Example 2.14 shows that any Hochschild product k[Cn] ⋆ k is isomorphic either to the direct product k[Cn] × k, or to a semi-direct product k[Cn](α, β), parame- terized by the group Un(k) × Un(k). The explicit description of the isomorphism classes of the algebras k[Cn](α, β) indicated by the equivalence relation (37) is a difficult number theory problem which relies heavily on the arithmetics of the positive integer n as well as on the base field k. Furthermore, the problem is also related to other two intensively studied problems in the theory of group algebras, namely the description of all invertible elements and the automorphism group of a group algebra [23, 32, 34]. Indeed, the key set A(n, k) which appears in the description of the classifying object HOC (k[Cn], k) parameterizes in fact the automorphism group AutAlg(k[Cn]). Any element of A(n, k) is invertible in k[Cn] and has order n in the group U (k[Cn]). These elements depend es- sentially on n and the base field k. Indeed, let X n − 1 = f1f2 ··· ft be the decomposition of X n − 1 as a product of irreducible polynomials in k[X]. If we denote by εi a root of fi in a fixed algebraic closure of k, we obtain that there exists a canonical isomorphism of k-algebras k[Cn] ∼= k(ε1)×··· × k(εt) that maps the generator d of Cn to (ε1,··· , εt). Thus, AutAlg(k[Cn]) is isomorphic to a direct product between all wreath product of Aut(cid:0)k(εi)(cid:1) and the symmetric groups [34]. algebras with an algebra projection on k[C2] ∼= k × k. Applying Example 2.14 for n = 2, we obtain the classification of all 3-dimensional HOCHSCHILD PRODUCTS AND NON-ABELIAN COHOMOLOGY FOR ALGEBRAS 21 Example 2.16. If k a field of characteristic 6= 2, then: GH2 (k[C2], k) ∼= (cid:0){±1} × {±1}(cid:1) ⊔ k∗ HOC (k[C2], k) ∼= {k3, k[X, Y ]/(X 2 − 1, Y 2, XY − Y ), A21} (38) (39) where A21 is the 3-dimensional non-commutative algebra having {1, d, f} as a basis and the multiplication given by d2 = 1, f 2 = 0, df = −f d = f . Now we highlight the efficiency of our methods in order to classify co-flag algebras of a given dimension. If k is a field of characteristic 6= 2, then, up to an isomorphism, there exists only two co-flag algebras of dimension 2: the algebras k[X]/(X 2) and k[C2] ∼= k×k [5, Corollary 4.5]. If k 6= k2, we mention that the other family of 2-dimensional algebras, namely the quadratic field extension k(√d), for some d ∈ k \ k2 does not contain co-flag algebras since there is no algebra map k(√d) → k. The co-flag algebras over k[C2] are classified by (39) and thus, in order to classify all 3-dimensional co-flag algebras we need to classify the co-flag algebras over k[X]/(X 2). Example 2.17. Let A := k[X]/(X 2). Then GH2 (k[X]/(X 2), k) ∼= k ⊔ k∗ and the equiv- alence classes of 3-dimensional algebras that have an algebra projection on k[X]/(X 2) are two families of algebras defined for any a ∈ k and u ∈ k∗ as follows: Aa := k[X, Y ]/(X 2 − a Y, Y 2, XY ), Au := k[X, Y ]/(X 2, Y 2 − u Y, XY ) Furthermore, HOC (k[X]/(X 2), k) = {A0, A1, A1}, i.e. up to an isomorphism there exist three co-flag algebras of dimension 3 over k[X]/(X 2). Indeed, A is the 2-dimensional algebra having 1 and x as a basis and x2 = 0. Thus A has only one algebra map A → k, namely the one sending x to 0. Hence, there exists a bijection CF1 (A) ∼= k such that the co-flag datum of the first kind (λ, Λ, ϑ) associated to a ∈ k is given by ϑ(x, x) := a, λ(x) = Λ(x) = ϑ(1, x) = ϑ(x, 1) = ϑ(1, 1) := 0 We can easily see that the equivalence relation ≈1 of Proposition 2.6 becomes equality, i.e. a ≈1 a′ if and only if a = a′ and hence CF1 (A)/ ≈1 ∼= k. The families of algebras as- sociated to such a co-flag datum of the first kind as defined by (22) are the 3-dimensional algebras having {f, 1, x} as a basis and the multiplication given by: x ⋆ x = af , f 2 = x⋆f = f ⋆x = 0, which is the algebra Aa. For the last part we apply Proposition 2.6 which proves that CF2 (A)/ ≈2 ∼= k∗: the algebra Au, for all u ∈ k∗, is precisely the alge- bra defined by (23) associated to the co-flag datum of the second king (λ0, u), where λ0 is the linear map given by λ0(x) := 0, λ0(1) := 1. The last statement follows from The- orem 2.8 or it can be proved directly as follows: we observe that, for any u ∈ k∗, there exists and isomorphism of algebras Au ∼= A1 = k[X, Y ]/(X 2, Y 2−Y, XY ). On the other hand, there exists an isomorphism of algebras Aa ∼= A1 = k[X, Y ]/(X 2 − Y, Y 2, XY ), for all a ∈ k∗ and any two algebras A0, A1 and A1 are not isomorphic to each other. To conclude, using Example 2.16 and Example 2.17 we obtain: 22 A. L. AGORE AND G. MILITARU k[X, Y ]/(X 2 − 1, Y 2, XY − Y ), Corollary 2.18. If k is a field of characteristic 6= 2 then, up to an isomorphism, there exist exactly six 3-dimensional co-flag algebras namely: k3, k < x, y x2 = 1, y2 = 0, xy = −yx = y > k[X, Y ]/(X 2, Y 2, XY ), k[X, Y ]/(X 2 − Y, Y 2, XY ), k[X, Y ]/(X 2, Y 2 − Y, XY ) In particular, if k := C the field of complex numbers, Corollary 2.18 shows that only 6 out of the 22 types of algebras of dimension 3 are co-flag algebras. Moreover, we also highlight the efficiency of Theorem 2.8 in classifying co-flag algebras by turning the problem into a purely computational one using a recursive method: if we consider A to be each of the algebras from Corollary 2.18 and using the results of this section we will arrive at the classification of 4-dimensional co-flag algebras. Of course the difficulty of the computations increases along side with the dimension. Very interesting and completely different from Mn(k) is the case when A := Tn(k) is the algebra of upper triangular matrices, i.e. Tn(k) is the subalgebra of Mn(k) having B := {eij i, j = 1, 2, ··· , n, i ≤ j} as the canonical basis over k. In order to write down the classifying object GH2 (Tn(k), k) we introduce the following three sets of matrices of trace zero, defined for any u, v, w = 1, 2,··· , n by: M u := {A = (aij) ∈ Mn(k) aii = 0, for all i = 1,··· , n and aur = 0, for all u < r} n M v,w := {A = (aij) ∈ Mn(k) v,w M := {A = (aij) ∈ Mn(k) Xi=1 Xi=1 n aii = 0, avv = 0, and aws = 0, for all w ≤ s} aii = 0, avv = 0, and aws = 0, for all w ≤ s 6= v} Example 2.19. Let k be a field of characteristic zero. Then: M u(cid:1)⊔(cid:0) GH2 (Tn(k), k) ∼= (cid:0) [u∈{1,2,··· ,n} [ × kn−w × kv−w−1. The equivalence where we denote: U := Sv,w∈{1,2,··· ,n}, v>w M classes of (cid:0) n(n+1) 2 + 1(cid:1)-dimensional algebras that have an algebra projection on Tn(k) are the families of algebras having {f, eij i, j = 1, 2,··· , n, i ≤ j} as a basis over k and the multiplication given below (we only write down the non-zero products): M v,w × kn−w(cid:1)⊔ U ⊔ k∗ v,w∈{1,2,··· ,n}, v<w v,w Tn(k)u A : Tn(k)v,w B,Γ : eit ⋆ ets = eil − αisf, euj ⋆ ejl = eul, euu ⋆ ekl = αklf, eij ⋆ euu = αijf, eij ⋆ ejj = eij − αijf, eiu ⋆ euu = eiu, euu ⋆ f = f ⋆ euu = f, where u ∈ {1, 2,··· , n}, A = (αpq)p,q=1,n ∈ M u, i, j 6= u, t 6= s; eip ⋆ epl = eil − βilf, ewi ⋆ eil = ewl − βlf, eww ⋆ ekl = ewlδw eit ⋆ evv = βitf, eij ⋆ ejj = eij − βij(1 − δv ewj ⋆ ejj = ewj + γj(1 − δw where v, w ∈ {1, 2,··· , n}, v < w, B = (βpq)p,q=1,n ∈ M v,w, Γ = (γr)w<r, i 6= w, p 6= l, i 6= l, t 6= v, s /∈ {v, w}; j )f, ews ⋆ evv = −γsf, j )f, evv ⋆ f = eww ⋆ f = f, k + βklf, HOCHSCHILD PRODUCTS AND NON-ABELIAN COHOMOLOGY FOR ALGEBRAS 23 Tn(k)w,v C,Ψ,Ω : λTn(k) : eip ⋆ epl = eil − δilf, ewi ⋆ eij = ewj + ψjf, ewt ⋆ etv = ewv + ωtf, eww ⋆ ekl = ewlδw k + δklf, eij ⋆ ejj = eij − δijf, ewj ⋆ evv = −ψjf eiv ⋆ evv = eiv, evv ⋆ f = f ⋆ eww = f, where v, w ∈ {1, 2,··· , n}, v > w, C = (δpq)p,q=1,n ∈ M , Ψ = (ψr)w<r, Ω = (ωs)w<s<v, i 6= w, j 6= v, p 6= l, t /∈ {v, w}; eij ⋆ ejl = eil + λ−1δn where λ ∈ k∗. j − 1)f, enn ⋆ f = f ⋆ enn = f, f 2 = λf, v,w i δn l (δn Indeed, we start by discussing the algebra maps λ : Tn(k) → k. Denote λ(eij) = αij ∈ k, for all eij ∈ B. Since e2 ii = eii we have αii ∈ {0, 1}, for all i = 1,··· , n. Moreover, since char(k) = 0 and we assume λ to be unitary it follows that Σn i=1αii = 1. Therefore, αuu = 1, for some u ∈ {1, ··· , n} and αii = 0, for all i 6= u; we denote by λu this algebra map. As eij = eiieij, for all i ≤ j, we obtain that αij = 0, for all i 6= u and i ≤ j. Finally, since eujeuu = 0 and λu(euu) = 1, we obtain that αuj = 0, for any u < j. To conclude, the set of algebra maps Tn(k) → k are in bijection to the set {1,··· , n} and the algebra map corresponding to some j ∈ {1,··· , n} is given by λj(ejj) := 1 and λj(euv) := 0, for all (u, v) 6= (j, j). The next step of the proof is a computational one: namely, for any u, v ∈ {1, 2, ··· , n} we are left to compute the set of all (λv, Λu)-cocycles ϑ. This is achieved by straightforward but lengthy checking of (17) which in this case comes down to the following compatibility condition: ϑ(eij, ersepq) − ϑ(eijers, epq) = ϑ(eij, ers)Λu − ϑ(ers, epq)λv (40) with i ≤ j, r ≤ s and p ≤ q. Rather than including here the cumbersome computations we will just point out the main steps taken; the detailed proof can be provided upon request. First, notice that since (40) is not symmetric with respect to the maps λv and Λu we distinguish three cases, namely: u = v, u < v and respectively u > v. For the case u = v the (λu, Λu)-cocycles obtained are implemented by a family of (n− u) scalars and a matrix of trace zero with zeros on the line u strictly above the diagonal. Then, it can be proved that any such cocycle is equivalent (in the sense of (26)) to a cocycle implemented by a matrix in M u. The corresponding co-flag algebras are those denoted by Tn(k)u A. If u < v then any (λv, Λu)-cocycle is equivalent to a cocycle implemented by (n − v) scalars and a matrix in M u,v. The corresponding co-flag algebras are those denoted by Tn(k)u,v B,Γ. Finally, u > v then any (λv, Λu)-cocycle is equivalent to a cocycle implemented by two families of (n− v) and respectively (u− v − 1) scalars and a matrix C,Ψ,Ω. Finally, in M the last family of co-flag algebras, denoted by λTn(k), corresponds to a co-flag datum of the second kind associated (δn, λ), where δn(i) = δn is the Kronecker symbol and i λ ∈ k∗. Moreover, for n = 2 we can also write down in a transparent way the other classifying object, namely HOC (T2(k), k). By a long but straightforward computation it can easily be seen that HOC (T2(k), k) contains the algebras whose multiplication is depicted below together with the direct product of algebras T2(k) × k: . The corresponding co-flag algebras are those denoted by Tn(k)v,u u,v 24 A. L. AGORE AND G. MILITARU ⋆ e11 e12 e22 f ⋆ e11 e12 e22 f e11 e11 0 0 f e11 e11 0 0 0 e12 e12 0 0 0 e12 e12 0 0 0 e22 f 0 f 0 e12 0 e22 0 0 e22 f f 0 0 e12 e22 0 0 f ⋆ e11 e12 e22 f ⋆ e11 e12 e22 f e11 e11 0 0 0 e11 e11 f 0 0 e12 e12 0 0 0 e22 f 0 0 e12 0 e22 f f 0 ⋆ e11 e12 e22 f e11 e11 − f 0 −f 0 e12 e12 0 0 0 e22 f e12 e22 − f f f 0 0 f 0 e12 e12 − f 0 f 0 e22 0 e12 − f e22 f f f 0 0 0 ⋆ e11 e12 e22 f e11 e11 0 0 f e12 e12 0 0 0 e22 f 0 0 e12 0 e22 f 0 0 ⋆ e11 e12 e22 f e11 e11 0 0 f e12 e12 + f 0 0 0 e22 f 0 0 e12 0 e22 f 0 0 3. Applications to coalgebras and Poisson algebras In this section we shall present two applications of our results to the theory of coalgebras and respectively Poisson algebras, the latter being the algebraic counterpart of Poisson manifolds. Our strategy is to use two different contravariant functors which have both the category of algebras as a codomain, namely the linear dual functor (−)∗ := Homk (−, k) and respectively Fun (−) := C ∞(−) the real smooth functions on a Poisson manifold functor. Supersolvable coalgebras. We recall that a coalgebra C = (C, ∆, ε) is a vector space C equipped with a comultiplication ∆ : C → C ⊗ C and a counit ǫ : C → k such that (∆⊗ Id)◦ ∆ = (Id⊗ ∆)◦ ∆ and (Id⊗ ε)◦ ∆ = (ε⊗ Id)◦ ∆ = Id, where ⊗ = ⊗k and Id is the identity map on C. We use the Σ-notation for comultiplication: ∆(c) = c(1) ⊗ c(2), for all c ∈ C (summation understood). The base field k, with the obvious structures, is the final object in the category of coalgebras. We only provide some basic information of coalgebras, referring the reader to [11] for more detail. If C = (C, ∆, ε) is a colgebra, then the linear dual C ∗ = Homk(C, k) is an associative algebra under the convolution product (f ∗ g)(c) := f (c(1))g(c(2)), for all f , g ∈ C ∗ and c ∈ C having the unit 1C ∗ = ε. Conversely, if A is a finite dimensional algebra with a basis {ei i = 1,··· , n}, then the linear dual A∗ is a coalgebra with the comultiplication and the counit given for any i = 1,··· , n by: n ∆A∗(e∗ i ) := Xj,l=1 i (ej el) e∗ e∗ j ⊗ e∗ l , εA∗ (e∗ i ) := e∗ i (1A) (41) where {e∗ i i = 1,··· , n} is the dual basis of {ei i = 1,··· , n}. The contravariant functor (−)∗ := Homk (−, k) gives a duality between the category of all finite dimensional coalgebras and the category of finite dimensional algebras [11]. Having the supersolvable Lie algebras [9] as a source of inspiration we introduce the following concept: HOCHSCHILD PRODUCTS AND NON-ABELIAN COHOMOLOGY FOR ALGEBRAS 25 Definition 3.1. A coalgebra C is called supersolvable if there exists a positive integer n such that C has a finite chain of subcoalgebras k ∼= C1 ⊂ C2 ⊂ ··· ⊂ Cn := C (42) such that Ci has codimension 1 in Ci+1, for all i = 1,··· , n − 1. Any supersolvable coalgebra is finite dimensional and the definition is equivalent to the fact that dim(Ci) = i, for all i = 1,··· , n. Furthermore, since any supersolvable coalgebra C contains a 1-dimensional subcolagebra, we obtain that G(C) 6= ∅, where G(C) is the space of group-like elements of C. Using the duality given by the functor (−)∗ it follows that a finite dimensional coalgebra C is supersolvable if and only if its convolution algebra C ∗ is a co-flag algebra in the sense of Definition 2.1 and vice- versa: a finite dimensional algebra A is a co-flag algebra if and only if its dual A∗ is a supersolvable coalgebra. Thus, the results obtained in the previous section can be applied for the classification problem of supersolvable coalgebras of a given dimension.2 Using the above facts, Corollary 2.18 classifies in fact all 3-dimensional supersolvable coalgebras: the isomorphism classes are the duals A∗ of the algebras listed in the statement. For example, the dual coalgebra associated to the noncommutative algebra k < x, y x2 = 1, y2 = 0, xy = −yx = y > is the non-cocommutative supersolvable coalgebra having {f1, f2, f3} as a basis and the comultiplication and the counit given by: ∆(f1) := f1 ⊗ f1 + f2 ⊗ f2, ∆(f2) := f1 ⊗ f2 + f2 ⊗ f1, ∆(f3) := f1 ⊗ f3 + f3 ⊗ f1 + f2 ⊗ f3 − f3 ⊗ f2, ε(f3) := 0 ε(f1) := 1, ε(f2) := 0 The coalgebra obtained in this way is indeed supersolvable by choosing C1 := k(f1 − f2) (we observe that f1± f2 is a group-like element) and C2 := kf1 + kf2 as the intermediary coalgebras of dimension 1 respectively 2 in the sequence (42). Another application of our theory of Section 1 is the following: let C be a finite di- mensional coalgebra, n a positive integer and V := kn. Then by taking the convolution algebra A := C ∗ we obtain that the object GH2 (C ∗, kn) classifies, up to an isomorphism which stabilizes C and costabilizes kn, all coalgebras which contain C as a subcoalgebra of codimension n. Moreover, a coalgebra D contains C as a subcoalgebra of codimension n if and only if D is isomorphic to the dual coalgebra (C ∗ ⋆ kn)∗ of a Hochschild product C ∗ ⋆ kn between the convolution algebra C ∗ and the vector space kn. The formula for the comultiplication of any such coalgebra (C ∗ ⋆ kn)∗ can be written down effectively by using (41) and (3). This observation shows that the GE-problem applied for finite dimensional algebras and finite dimensional vector spaces gives the answer at the level of finite dimensional coalgebras to what we have called the extending structures problem studied in [2, 4, 5] for Jacobi, Lie and respectively associative algebras. Applications to Poisson algebras. Commutative Poisson algebras are algebraic coun- terparts of Poisson manifolds from differential geometry: for a given smooth manifold M , there is a one-to-one correspondence between Poisson brackets on the commutative alge- bra P := C ∞(M ) of smooth functions on M and all Poisson structures on M [22]. The 2The classification of solvable Lie algebras [20], over arbitrary fields, was achieved up to dimension 4. 26 A. L. AGORE AND G. MILITARU [pq, r] = [p, r] q + p [q, r] importance of Poisson algebras in several areas of mathematics and physics (Hamilton- ian mechanics, differential geometry, Lie groups, noncommutative algebraic/diferential geometry, (super)integrable systems, quantum field theory, vertex operator algebras, quantum groups) it is also well known -- see [21, 22, 30] and the references therein. In fact, C ∞(−) gives a contravariant functor from the category of Poisson manifolds to the category of Poisson algebras and constitutes the tool through which geometrical prob- lems can be restated and approached at the level of Poisson algebras. In particular, the GE-problem, formulated for Poisson algebras, is just the algebraic counterpart of the following question from differential geometry: Let M be a Poisson manifold. Describe and classify all Poisson manifolds containing M as a manifold of a given codimension. We recall that a Poisson algebra is a triple P = (P, mP , [−, −]), where (P, mP ) is a unital associative algebra, (P, [−, −]) is a Lie algebra such that the Leibniz law holds for any p, q, r ∈ P : (43) Usually, a Poisson algebra P is by definition assumed to be commutative like the algebra P := C ∞(M ) of real smooth functions on a Poisson manifold M which is the typical example of a Poisson algebra. However, following [18, 29] in order to broaden the class of Poisson algebras and to be able to construct relevant examples, throughout this paper we do not impose this restriction. A morphism between two Poisson algebras P and P ′ is a linear map ϕ : P → P ′ that is a morphism of associative algebras as well as of Lie algebras; we will denote by AutPoss(P ) the group of automorphisms of a Poisson algebra P . For basic concepts and unexplained notions on Lie algebras we refer to [16] and to [30] for those concerning Poisson algebras. Let P = (P, mP ) be an algebra and u ∈ k. Then (P, mP , [−, −]u) is a Poisson algebra, where [a, b]u := u(ab − ba), for all a, b ∈ P . In particular, any associative algebra P = (P, mP ) is a Poisson algebra with the abelian Lie bracket, i.e. [a, b] := 0, for all a, b ∈ P . On the other hand, let g = (g, [−, −]g) be a Lie algebra with a linear basis {ei i ∈ I}. Then the symmetric algebra P := S(g) of g (i.e. the polynomial algebra k[ei i ∈ I]) is a Poisson algebra with the bracket defined by [ei, ej] := [ei, ej]g, for all i, j ∈ I and extended to the entire algebra k[ei i ∈ I] via the Leibniz law (43). If P is a Poisson algebra then the direct product P × k is a Poisson algebra with the direct product structures of associative/Lie algebra: that is the multiplication and the bracket is defined for any p, q ∈ P and x, y ∈ k by: (p, x) ⋆ (q, y) := (pq, xy), (44) Furthermore, the canonical projection πP : P × k → P , πP (p, x) := p is a surjective Poisson algebra map having a 1-dimensional kernel. It what follows we shall classify all Poisson algebras Q that admit a surjective Poisson algebra map Q → P → 0 with a 1-dimensional kernel. In order to do this we first recall from [3] the following concept: {(p, x), (q, y)} := (cid:0)[p, q], 0(cid:1) Definition 3.2. A co-flag datum of a Poisson algebra P is a 5-tuple (λ, Λ, ϑ, γ, f ), where λ, Λ, γ : P → k are linear maps, ϑ, f : P × P → k are bilinear maps such that: (CF1) (λ, Λ, ϑ) is a co-flag datum of the first kind of the associative algebra P (CF2) λ([p, q]) = Λ([p, q]) = γ([p, q]) = f (p, p) = 0 HOCHSCHILD PRODUCTS AND NON-ABELIAN COHOMOLOGY FOR ALGEBRAS 27 (CF3) f (p, [q, r]) + f (q, [r, p]) + f (r, [p, q]) + γ(p)f (q, r) + γ(q)f (r, p) + γ(r)f (p, q) = 0 (CF4) f (pq, r) − Λ(q)f (p, r) − λ(p)f (q, r) = γ(r)ϑ(p, q) + ϑ([p, r], q) + ϑ(p, [q, r]) (CF5) γ(pq) = γ(p)Λ(q) + λ(p)γ(q) for all p, q, r ∈ P . We denote by F (P ) the set of all co-flag data of P . The above concept was introduced in [3, Definition 3.2] under the name of 'abelian co- flag datum' of P . We dropped the adjective 'abelian' since the other family of co-flag datum introduced in [3, Definition 3.4] will generate a family of Poisson algebras which are all isomorphic to the direct product P × k (see the proof of Theorem 3.3 below) -- hence, they are irrelevant from the classification view point. Let (λ, Λ, ϑ, γ, f ) ∈ F (P ) be a co-flag datum of a Poisson algebra P . Then we shall denote by P(λ, Λ, ϑ, γ, f ) := P × k the direct product of vector spaces which is a Poisson algebra with the multiplication ⋆ and the bracket {−, −} defined [3, Section 3] for any p, q ∈ P , x, y ∈ k by: (p, x) ⋆ (q, y) := (pq, ϑ(p, q) + λ(p)y + Λ(q)x) {(p, x), (q, y)} := (cid:0)[p, q], f (p, q) + γ(p)y − γ(q)x(cid:1) (45) (46) Now we can prove the following classification result: it is the Poisson version of The- orem 2.8 and improves the classification given in [3, Theorem 3.6] where all Poisson algebras Q having a Poisson surjection Q → P → 0 with a 1-dimensional kernel are classified in a more restrictive fashion: up to an isomorphism which stabilizes k and co-stabilizes P . Theorem 3.3. Let P be a Poisson algebra. Then: (1) A Poisson algebra Q has a surjective Poisson algebra map Q → P → 0 with a 1- dimensional kernel if and only if Q ∼= P×k or Q ∼= P(λ, Λ, ϑ, γ, f ), for some (λ, Λ, ϑ, γ, f ) ∈ F (P ). (2) Two Poisson algebras P(λ, Λ, ϑ, γ, f ) and P(λ′, Λ′, ϑ′, γ ′, f ′) are isomorphic if and only if there exists a triple (s0, ψ, r) ∈ k∗×AutPoss(P )×Homk(P, k) such that for any p, q ∈ P : (47) (48) λ = λ′ ◦ ψ, ϑ(p, q) s0 = ϑ′(cid:0)ψ(p), ψ(q)(cid:1)+λ(p)r(q) + Λ(q)r(p) − r(pq) f (p, q) s0 = f ′(cid:0)ψ(p), ψ(q)(cid:1)+γ(p)r(q) − γ(q)r(p) − r([p, q]) (49) Λ = Λ′ ◦ ψ, γ = γ′ ◦ ψ (3) The Poisson algebras P(λ, Λ, ϑ, γ, f ) and P × k are not isomorphic. Proof. (1) Let Q be a Poisson algebra having a surjective Poisson algebra map Q → P → 0 with 1-dimensional kernel. Then, using [3, Proposition 2.4 and Proposition 3.5] we obtain that Q ∼= P(λ, Λ, ϑ, γ, f ), for some (λ, Λ, ϑ, γ, f ) ∈ F (P ) or Q ∼= P (λ, ϑ, u), where u ∈ k \ {0}, λ : P → k is a unit preserving linear map, ϑ : P × P → k is a bilinear map satisfying the following two compatibilities for any p, q, r ∈ P : λ(pq) = λ(p)λ(q) − u θ(p, q), θ(p, qr) − θ(pq, r) = θ(p, q)λ(r) − θ(q, r)λ(p) (50) 28 A. L. AGORE AND G. MILITARU which we called in [3, Definition 3.4] a non-abelian co-flag datum of P . The Poisson algebra P (λ, ϑ, u) is the vector space P × k having the multiplication and the Poisson bracket defined by: (p, x) ⋆ (q, y) := (pq, ϑ(p, q) + λ(p)y + λ(q)x + u xy) (51) {(p, x), (q, y)} := (cid:0)[p, q], −u−1 λ([p, q])(cid:1) (52) for all p, q ∈ P and x, y ∈ k. Since u 6= 0 we obtain from the first equation of (50) that ϑ is implemented by u and λ by the formula ϑ(p, q) = u−1(cid:16)λ(p)λ(q)− λ(pq)(cid:17) and hence the multiplication of the Poisson algebra P (λ, u) given by (51) takes the form: (p, x) ⋆ (q, y) = (pq, u−1(cid:16)λ(p)λ(q) − λ(pq)(cid:17)+λ(p)y + λ(q)x + u xy) which is precisely (19). The first part is finished once we observe that the map given by the formula (21), namely ϕ : P (λ, u) → P × k, ϕ(p, x) := (p, λ(p) + u x), for all p ∈ P and x ∈ k is an isomorphism of Poisson algebras. Indeed, Remark 2.4 shows that ϕ is an isomorphism of associative algebras; hence we only have to prove that it is also a Lie algebra map, where the bracket on P (λ, u) is given by (52). A straightforward compu- tation shows that ϕ(cid:16){(p, x), (q, y)}(cid:17)= [ϕ(p, x), ϕ(q, y)]P ×k = (cid:0)[p, q], 0(cid:1) and thus any Poisson algebra P (λ, u) is in fact isomorphic to the direct product of Poisson algebras P × k. (2) The first step in proving Theorem 2.8 gives a bijection between all associative algebra isomorphism corresponding to the Poisson algebras P(λ, Λ, ϑ, γ, f ) and P(λ′, Λ′, ϑ′, γ ′, f ′) and the set of all triples (s0, ψ, r) ∈ k∗ × AutAlg(P ) × Homk(P, k) satisfying (48) and the first two compatibilities of (47). The bijection is given such that the associative algebra isomorphism ϕ = ϕ(s0, ψ, r) : P(λ, Λ, ϑ, γ, f ) → P(λ′, Λ′, ϑ′, γ ′, f ′) associated to (s0, ψ, r) is given by the formula (33), that is ϕ(p, x) = (cid:0)ψ(p), r(p) + x s0(cid:1), for all p ∈ P and x ∈ k. The proof will be finished if we show that such a map ϕ = ϕ(s0, ψ, r) is also a morphism of Lie algebras if and only if ψ is an automorphism of the Lie algebra P = (P, [−, −]) and the last equation of (47) and (49) hold. This is an elementary fact: by a straightforward computation we can show that ϕ(cid:16){(p, 0), (q, 0)}(cid:17)= {ϕ(p, 0), ϕ(q, 0)} if and only if ψ : P → P is a Lie algebra map and (49) holds. ϕ(cid:16){(p, 0), (0, x)}(cid:17)= {ϕ(p, 0), ϕ(0, x)} if and only if γ = γ′ ◦ ψ. The rest of the details In a similar fashion are left to the reader. (3) Follows from step (2) of the proof of Theorem 2.8 which proves that the associative algebras P(λ, Λ, ϑ, γ, f ) and P × k are never isomorphic. (cid:3) Corollary 3.4. Let P be a Poisson algebra for which there is no algebra map P → k or P is perfect as a Lie algebra, i.e. P = [P, P ]. Then, up to an isomorphism, the only Poisson algebra Q for which there exists a surjective algebra map Q → P → 0 having a 1-dimensional kernel is the direct product P × k of Poisson algebras. Proof. The proof follows from Theorem 3.3 since in both cases the set F (P ) is empty. Indeed, the first case follows form Corollary 2.11, while if P is perfect as a Lie algebra HOCHSCHILD PRODUCTS AND NON-ABELIAN COHOMOLOGY FOR ALGEBRAS 29 then, using the compatibilities (CF2) we obtain that λ = Λ ≡ 0, the trivial maps, which contradicts the fact that the algebra maps λ and Λ are unit preserving. (cid:3) The geometrical meaning of Corollary 3.4 is the following: if M is a real Poisson manifold such that the algebra C ∞(M ) of all real smooth functions on M is perfect as a Lie algebra, then up to an isomorphism, there is only one Poisson manifold containing P as a sub-manifold of codimension 1. The group of Poisson algebra automorphisms of P(λ, Λ, ϑ, γ, f ) can also be described. Using the proof of Theorem 3.3, the Poisson version of Corollary 2.12 takes the following form: Corollary 3.5. Let P be a Poisson algebra, (λ, Λ, ϑ, γ, f ) ∈ F (P ) a co-flag datum of P and let GP(cid:0)P, (λ, Λ, ϑ, γ, f )(cid:1) be the set of all triples (s0, ψ, r) ∈ k∗ × AutPoss(P ) × P ∗ satisfying the compatibility conditions (47)-(49) written for λ′ = λ, Λ′ = Λ, γ′ = γ, ϑ′ = ϑ and f ′ = f . Then, there exists an isomorphism of groups AutPoss(cid:0)P(λ, Λ, ϑ, γ, f )(cid:1)∼= GP (cid:0)P, (λ, Λ, ϑ, γ, f )(cid:1) where the latter is a group with respect to the following multiplication: 0, ψ ◦ ψ′, r ◦ ψ′ + s0r′) 0, ψ′, r′) := (s0s′ for all (s0, ψ, r), (s′ (s0, ψ, r) · (s′ 0, ψ′, r′) ∈∈ GP (cid:0)P, (λ, Λ, ϑ, γ, f )(cid:1). Moreover, the map (s0, ψ, r) 7→ (cid:0)s−1 GP (cid:0)P, (λ, Λ, ϑ, γ, f )(cid:1)−→ P ∗ ⋉ζ (cid:0)k∗ × AutPoss(P )(cid:1), 0 r, (s0, ψ)(cid:1) is an injective morphism of groups. We end the paper with a relevant example which follows by a long computation based on Theorem 3.3; the detailed proof can be provided upon request. Example 3.6. Let k be a field of characteristic zero and P := H(3, k) the 3-dimensional i.e., H(3, k) is the set of all upper tri- non-commutative Heisenberg-Poisson algebra: angular 2 × 2 matrices with the usual multiplication and the Lie bracket given by [x, y] := xy − yx. Consider {e11, e12, e22} a basis of H(3, k) over k. Then the set of isomorphism types of all 4-dimensional Poisson algebras which admit a surjective Poisson algebra map on H(3, k) are the ones listed below together with the usual direct product H(3, k) × k (we only write down the non-zero products): P1 : P2 : P3 : e11 ⋆ e11 = e11, e11 ⋆ e12 = e12, e12 ⋆ e22 = e12, e22 ⋆ e22 = e22, e11 ⋆ f = f ⋆ e11 = f, {e11, e12} = e12, {e12, e22} = e12. e11 ⋆ e11 = e11, e11 ⋆ e12 = e12, e12 ⋆ e22 = e12, e22 ⋆ e22 = e22, e22 ⋆ f = f ⋆ e22 = f, {e11, e12} = e12, {e12, e22} = e12 e11 ⋆ e11 = e11, e11 ⋆ e12 = e12 − f, e12 ⋆ e11 = f, e12 ⋆ e22 = e12 − f e22 ⋆ e12 = f, e22 ⋆ e22 = e22, e11 ⋆ f = f ⋆ e22 = f, {e11, e12} = e12, {e12, e22} = e12, {e11, f} = f, {e22, f} = −f. 30 A. L. AGORE AND G. MILITARU P ω 4 : P τ 5 : e11 ⋆ e11 = e11, e11 ⋆ e12 = e12, e12 ⋆ e22 = e12, e22 ⋆ e22 = e22, e11 ⋆ f = f ⋆ e22 = f, {e11, e12} = e12, {e12, e22} = e12, {e11, f} = ωf, {e22, f} = −ωf, where ω ∈ k. e11 ⋆ e11 = e11, e11 ⋆ e12 = e12, e12 ⋆ e22 = e12, e22 ⋆ e22 = e22, e22 ⋆ f = f ⋆ e11 = f, {e11, e12} = e12, {e12, e22} = e12, {e11, f} = τ f, {e22, f} = −τ f, where τ ∈ k. We point out that, even if up to an isomorphism there are only 8 associative algebras of dimension 4 with a surjective algebra map on the algebra H(3, k) (indicated at the end of Example 2.19), the set of isomorphism types of Poisson algebras having the same algebra structure can be infinite due to the 1-parameter families P ω 4 and P τ 5 . Acknowledgement Parts of this work were undertaken while the first named author was visiting the Max Planck Institute for Mathematics in Bonn. Their hospitality and financial support is gratefully acknowledged. References [1] Adem, A. and Milgram, R. J. - Cohomology of finite groups, Springer, 2004. [2] Agore, A.L. and Militaru, G. - Jacobi and Poisson algebras, J. Noncommutative Geometry 9 (2015), 1295 -- 1342. [3] Agore, A.L. and Militaru, G. - The global extension problem, crossed products and co-flag non- commutative Poisson algebras, J. Algebra, 426 (2015), 1 -- 31. [4] Agore, A.L. and Militaru, G. - Extending structures for Lie algebras, Monatsh. fur Mathematik, 174 (2014), 169 -- 193. [5] Agore, A.L. and Militaru, G. - Extending structures, Galois groups and supersolvable associative algebras, Monatsh. fur Mathematik, DOI:10.1007/s00605-015-0814-8. [6] Alekseevsky, D., Michor, P. W. and Ruppert, W. - Extensions of super Lie algebras, J. Lie Theory 15 (2005), 125 -- 134. [7] Alperin, J.L. and Bell, R.R. - Groups and representations, Springer-Verlag, New York, 1995. [8] Andruskiewitsch, N. and Devoto, J. - Extensions of Hopf algebras, Algebra i Analiz 7 (1995), 22 -- 61. [9] Barnes, D.W. and Newell, M.L. - Some theorems on saturated homomorphs of soluble Lie algebras, Math. Z., 115 (1970) 179-187. [10] Bavula, V.V. and Jordan, D.A. - Isomorphism problems and groups of automorphisms for generalized Weyl algebras, Trans. Amer. Math. Soc., 353 (2001), 769-794. [11] Brzezi´nski, T. and Wisbauer, R. - Corings and comodules, London Math Soc. Lect. Note Ser., 309, Cambridge University Press, 2003 [12] Castiglioni, J. L., Garca-Martnez, X. and Ladra, M. - Universal central extensions of Lie-Rinehart algebras, arXiv:1403.7159 [13] Ceken, S., Palmieri, J., Wang Y.- H. and Zhang, J.J. - The discriminant controls automorphism groups of noncommutative algebras, Adv. Math., 269 (2015), 551 -- 584. [14] Chevalley, C. and Eilenberg S. - Cohomology theory of Lie groups and Lie algebras, Trans. Amer. Math. Soc., 63(1948), 85 -- 124. [15] Eilenberg, S. and Maclane, S. - Cohomology theory in abstract groups, II, Annals of Mathematics, 48(1947), 326 -- 341. [16] Erdmann, K. and Wildon, M.J. - Introduction to Lie algebras, Springer, 2006. HOCHSCHILD PRODUCTS AND NON-ABELIAN COHOMOLOGY FOR ALGEBRAS 31 [17] Everett, C. J., Jr. - An extension theory for rings, Amer. J. Math., 64 (1942), 363-370 [18] Farkas, D. R. and Letzter, G. - Ring theory from symplectic geometry, J. Pure Appl. Algebra, 125(1998), 155-190. [19] Fr´egier, Y. - Non-abelian cohomology of extensions of Lie algebras as Deligne groupoid, J. Algebra, 398 (2014), 243-257. [20] de Graaf, W.A. - Classification of Solvable Lie Algebras, Experiment. Math., 14 (2005), 15 -- 25. [21] Grabowski, J., Marmo, G. and Perelomov, A. M. - Poisson structures: towards a classification, Modern Phys. Lett. A 8(1993), 1719 -- 1733. [22] Grabowski, J. - Brackets, Int. J. Geom. Methods Mod. Phys., 10(8):1360001, 45, 2013. [23] Janusz, G. J. - Automorphism group of simple algebras and group algebras, Representation Theory of Algebras, Lecture Notes in Pure and Applied Mathematics, Vol. 37 (1976), 381-388. [24] Huebschmann, J. - Poisson cohomology and quantization, J. Reine Angew. Math., 408 (1990), 57-113. [25] Huebschmann, J., Perlmutter, M. and Ratiu, T.S. - Extensions of Lie-Rinehart algebras and cotan- gent bundle reduction, . Proc. London Math. Soc., 107 (2013), 1135 -- 1172. [26] Hochschild, G. - Cohomology and representations of associative algebras, Duke Math. J., 14(1947), 921 -- 948. [27] Horn, R. and Sergeichuk, V. - Canonical matrices of bilinear and sesquilinear forms, Linear Algebra and its Applications, 428 (2008), 193-223. [28] O. Holder, Bildung zusammengesetzter Gruppen, Math. Ann. 46(1895), 321 -- 422. [29] Kubo, F. - Finite-dimensional non-commutative Poisson algebras, J. Pure Applied Algebra, 113 (1996), 307 -- 314. [30] Laurent-Gengoux, C., Pichereau, A. and Vanhaecke, P. - Poisson Structures, Vol. 347, 2013, Springer. [31] Loday, J.-L. and Pirashvili, T. - Universal enveloping algebras of Leibniz algebras and (co)homology, Math. Ann., 296(1993), 139 -- 158. [32] Milies, C. P. and Sehgal S.K. - An introduction to group rings, Kluwer Academic Publishers, 2002. [33] Militaru, G. - The global extension problem, co-flag and metabelian Leibniz algebras, Linear Mul- tilinear Algebra 63 (2015), 601 -- 621. [34] Olivieri, A., del Rio, A. and Sim´on, J. J. - The Group of Automorphisms of the Rational Group Algebra of a Finite Metacyclic Group, Comm. in Algebra, 34 (2006), 3543-3567. [35] Redondo, M. J. - Hochschild cohomology: some methods for computations, IX Algebra Meeting USP/UNICAMP/UNESP (Sao Pedro, 2001), Resenhas IME-USP, 5(2001), 113 -- 137. [36] Rotman, J.J. - Advanced Modern Algebra, New York: Prentice Hall, 2nd Edition, 2003. [37] Schreier, O. - Uber die Erweiterung von Gruppen, I, Monatshefte fur Mathematik und Physik, 34 (1926), 165 -- 180. [38] Weibel, C.A. - An introduction to homological algebra, Cambridge University Press, 1994. [39] Williamson, J. - On the algebraic problem concerning the normal forms of linear dynamical systems, Amer. J. Math., 58 (1936), 141 -- 163. Faculty of Engineering, Vrije Universiteit Brussel, Pleinlaan 2, B-1050 Brussels, Belgium and "Simion Stoilow" Institute of Mathematics of the Romanian Academy, P.O. Box 1-764, 014700 Bucharest, Romania E-mail address: [email protected] and [email protected] Faculty of Mathematics and Computer Science, University of Bucharest, Str. Academiei 14, RO-010014 Bucharest 1, Romania E-mail address: [email protected] and [email protected]

Subsets and Splits

No saved queries yet

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