Democratic Re nement in a Multi-Agent World

This project is concerned with how a group of agents can work together to correct faults in their own performance. We believe that reenement can be transformed into a DAI problem which agents will cooperate to solve. All agents in the community would have the opportunity to contribute to this process; hence the word democratic in the project title. The ideas in this document are the culmination of a literature review of three distinct research areas: Distributed Artiicial Intelligence (DAI), Machine Learning in DAI systems and Theory Reenement. Initially we looked at the general uses of Machine Learning in distibuted systems. This led to the observation that no research exists to date on reening the knowledge in a DAI system. Therefore we decided to investigate this speciic area. Section 2 is a general overview of the reasons for studying DAI and the research to date in this area. In Section 3 the limited existing research on integrating learning into DAI systems is reviewed. Sections 3 and 4 outline the reasons why reenement is necessary in a DAI system, the current approaches to reenement of a single theory or knowledge base and why a new approach is required in order to reene knowledge in a DAI system. Finally our current research objectives and projected future work are set out in Section 5. At this stage I will introduce an example DAI system which is used to illustrate various points throughout this document. It is a predator-prey scenario of a kind that is often used in the DAI literature. A system of hunters cooperate in order to capture large prey. There are two diierent types of agents: scouts and assassins. Scouts have better perception than assassins but are physically weaker and cannot kill large prey on their own although they may attack prey smaller than themselves. Assassins nd it advantageous to cooperate with scouts in order to locate prey which is then shared among all those who participated in the hunt. The number of hunters needed to ensure a kill will depend on the species of prey. Communication is possible between all hunters. All agents may be peers or there may be a pack leader who directs the other hunters. The following is a reproduction of a slide 1