Coordination in Large-Scale Multi-Agent Systems

There exists no official, universally accepted definition of agency within the artificial intelligence community, only perceptions of what an agent ‘should be’ [1]. The most widely cited description of agency is that which Woolridge and Jennings call ‘a weak notion of agency’: a system with properties of autonomy, social ability, reactivity and pro-activity [2]. A multi-agent system is simply one in which many agents run simultaneously, communicating amongst themselves and working toward either their own goal or a common objective. RoboCup is an international competition which takes place each year in a different city around the world; for example, this year the competition took place in Atlanta, USA [3]. It consists of four main competitions, each containing several leagues; those competitions are RoboCupSoccer, RoboCupRescue, RoboCupJunior and RoboCup@Home. The league amongst these which I feel would best suit the proposed research is the simulation league of RoboCupRescue, which involves writing agents which aim to best cope with a disaster scenario. RoboCupRescue is a good example of a multi-agent system, as it involves many agents running simultaneously and toward goals. The idea and main scenario of RoboCupRescue came about after a major earthquake in Kobe, Japan, which required emergency services to co-ordinate amongst themselves in a situation where information was scarce and so needed to be shared in the most efficient way [4]. The competition has three types of agent, each with their own goal: police, ambulance and fire, which aspire to clear road blockades, rescue civilians and extinguish fires, respectively. This situation shows a need for each agent to co-ordinate with the other members of its team, not only to share information but also to plan the best course of action to take. This kind of co-ordination is difficult to implement efficiently at the current time as the agent situation is under a large amount of uncertainty, which means a good reasoning algorithm needs to be discovered. I do not intend for this algorithm to be specific only to the RoboCupRescue domain; I will merely use RoboCupRescue as a platform to test potential algorithms. I also intend to make use of the ECSKernel tool, which I helped to develop in 2006 as an aid to those researching areas of RoboCupRescue. The tool allows the platform to be controlled and finely tuned so that it can be used for particular research purposes. I propose to start my research by investigating existing reasoning algorithms such as solutions to MDP’s (Markov Decision Processes) and POM-DP’s (Par-