Communicating During Learning

Abstract

This work studies the e ects of communicating during learning. It focuses on the exchange of information between teams of agents that are solving similar problems at di erent locations using di fferent learning algorithms. The objectives are: 1) assert the bene ts and costs of diff erent types of information-exchange during learning; 2) compare this process in heterogeneous versus homogeneous environments. By \heterogeneous" we mean that each team of agents uses a di erent learning algorithm. The experiments reported here use the Predator-Prey problem as a test-case. We conclude that: Exchange of information can bene t learning performance, at the expense of communication and o ine processingtime; Heterogeneous systems coupled with role-attribution show only a slight improvement of performance over homogeneous ones in the chosen test-case.