Advice-Exchange Amongst Heterogeneous Learning Agents: Experiments in the Pursuit Domain

Abstract

The question that is addressed in this paper is: "(How) can a heterogeneous group of learning-agents, involved in solving similar problems, cooperate by exchanging information in order to improve their own performance?" The approach taken, entitled "Advice-Exchange", consists on requesting advice from agents that show good performance on the current problem and using this knowledge either as a desired response for supervised training or to provide extra reinforcement to the agent about a given action. This is the first step towards a technique that aims at providing added capabilities to heterogeneous groups of learning-agents that are solving similar problems in parallel. Results of several experiments in the Pursuit (predator-prey) domain show that information exchange can improve the performance of the learning algorithms tested. Contrary to initial expectations the use of heterogeneous groups of learners, despite having good results in the easier tasks, does not seem to be critical for the harder problems.