MACC'97 Poster session
This paper is concerned with a multi-agent model for general problem solving in computer network characterized by {\it openness}, {\it locality}, and {\it diversity} of available knowledge. We are devising a robust multi-agent system in which, given a generally abstract plan, the agent repeatedly refine it into a more concrete one until executable. In this process, to make up for the lack of knowledge, the agent can ask other agents how to refine the partial plan, according to the social evaluation about who has refined the concerning plan most successfully.
In this general system, this paper focuses on a credit assignment problem (Minskey, 1961): If a task is fulfilled by a sequence of activities of agents, how can credit or blame be accurately assigned to early rules that set the stage for the final result? To this problem, we introduce a logic-based model. Unlike the previous researches, this model is devised to specify which agents logically contribute in fulfilling a task. A multi-agent system based on this model has been tested with a simple simulator through an electronic purchase scenario.
task refinement, distributed problem solving, reinforcement learning
PS
file(+gzip) (in Japanese)