Cooperative Learning Using Advice Exchange

Abstract

One of the main questions concerning learning in a Multi-Agent System’s environment is: “(How) can agents benefit from mutual interaction during the learning process?” This paper describes a technique that enables a heterogeneous group of Learning Agents (LAs) to improve its learning performance by exchanging advice. This technique uses supervised learning (backpropagation), where the desired response is not given by the environment but is based on advice given by peers with better performance score. The LAs are facing problems with similar structure, in environments where only reinforcement information is available. Each LA applies a different, well known, learning technique. The problem used for the evaluation of LAs performance is a simplified traffic-control simulation. In this paper the reader can find a summarized description of the traffic simulation and Learning Agents (focused on the advice-exchange mechanism), a discussion of the first results obtained and suggested techniques to overcome the problems that have been observed.