Learning coordination strategies for cooperative multiagent systems

Abstract

This thesis proposes a new technique for learning multiagent coordination strategies that address the issues of convergence, complexity, credit assignment, and utility. Traditionally, strategies to control the behavior of multiple agents have been hand-coded to meet a designer's goals. This task is complex due to the interactions that can occur among agents. Recent work in this area has focused on how strategies can be learned. Yet, these systems suffer from a variety of problems that include lack of convergence or performance guarantees and from complexity concerns.

Following a formalization of the problem, a review of related works and a discussion of unresolved issues, a generic multiagent learning framework is presented. Then the basis of the proposed technique, probabilistic hill-climbing, is discussed and mapped into this framework Implementation details are then described and experimental results on three different domains reported. Finally, an extension to reduce sample complexity is considered.