Learning by exchanging advice

Abstract

The emergence of Multiagent systems brought new challenges to the field of Machine Learning, as it did to many others. One of the main challenges is to take advantage of the information available when several agents, possibly using different learning techniques, are dealing with similar problems, either in the same location (i.e. acting as a team) or in different ones. This work aims at studying the possible advantages and pitfalls of exchanging information during the learning process, leading to better adaptation. We will discuss the subject of when, how and to whom ask for advice, and present the results obtained in two experimental scenarios: the Pursuit (Predator-Prey) Domain and a Traffic Control simulation. Results show that exchange of information can improve the average performance of learning agents enabling them to escape from local maxima in some cases, although it may reduce the exploration of the space, preventing successful agents from finding better local maxima of the quality function.