From: Proceedings of the First International Conference on Multiagent Systems. Copyright © 1995, AAAI (www.aaai.org). All rights reserved. Map Making as a Support for Cooperation Fumihiko Chimura and Mario Tokoro Departmentof ComputerScience, Faculty of Science and Technology,Keio University. 3 - 14 - 1 Hiyoshi, Kohoku-ku,Yokohama 223, Japan. {chimura,mario}Gmt.cs.keio.ac.jp This research deals with search tasks of agents, in particular, a meeting task whereagents try to gather into a single location starting fromdistributed arrangements. The meeting task is ubiquitous in cooperation because agents need to approach one another if they wishto worktogether effectively. For example,in cooperative computation, multiple processes must come near in order to reduce communicationcost. Anefficient solution is one in whichall agents worksimultaneously, and in whichthe meeting location determines adaptively. However,the difficulty with this solution is that the agents, and hence the meetinglocation, will moveunpredictably. The search task of movingtargets requires many moveswhenagents try to react, as soon as possible, to each moveof the targets (Ishida & Koff 1991). The trailblazer search makesa tradeoff betweenreaction time and the number of moves (Chimura & Tokoro 1994). It dynamically maintains a mapof where the agent has already searched through in the problem space. The mapis a routing table that records paths found in the searched region. Oncethe agent finds a mappedpath to the target, it flashes back using those paths. The idea contrasts with that of conducting a lookaheadsearch (Ishida 1992). Assumingthe mapis useful to overcomethe unpredictability issue of the meetingtask, weevaluate a dynamicsolution that uses the trailblazer search as the search strategy of agents. Wecomparethree solutions. ¯ Single Agent Traveling Search (SATS) Without using a map,a single agent tries to reach the locations of all the other agentsthat stay still. ¯ Static Meeting Point Search (SMPS) Without using maps, all agents simultaneously try to reach a predefined meetinglocation. ¯ Dynamic Meeting Point Search (DMPS) Using mapsthat record movesof agents, all agents simultaneouslytry to reach the locations of all the other agents. For evaluation, we use a simple grid-like, rectangular problemspace of size 40 junctions on each side, placing agents and obstacles randomlyon junctions. Obstacles 442 ICMAS-9$ realize unpredictable motionsof agents. The estimated total cost of searchis the sumd. D-I- e ¯ E of the cost D of map maintenance and the number E of moves, for the last agent to reach the meetinglocation, and whered is set to 10-s and e to 1. 35O [] 3OO / 2.5O Aa ID SATS O 8MPS b DMPS ~ 150 2 Agents 100 .... 5O ........4 Agents 0 3 Agents m v ...~.l=..,,-IP- 0 5 10 15 20 253035 Obstacle Occupation (%) Figure 1: Results for the MeetingTask DMPS is efficient due to the benefit of the map.The reason is that, even in the face of unpredictable motions, whenagents try to comenear and meet each other, there is high probability that they run into passed regions of each other. In this case, the map can be highly utilized. Weare further studying the case of agents using locally valid, incompletemaps. References Chimura, F., and Tokoro, M. 1994. The Trailblazer Search: A NewMethod for Searching and Capturing MovingTargets. In Proceedings of the Twelfth National Conferenceon Artificial Intelligence, 13471352. AAAI. Ishida, T., and Koff, R. E. 1991. MovingTarget Search. In Proceedingsof the Twelfth International Joint Conferenceon Artificial Intelligence, 204-210. Ishida, T. 1992. MovingTarget Search with Inteiligence. In Proceedingsof the Tenth National Conference on Artificial Intelligence, 525-532.AAAI.