Learning by Exchanging Advice

Abstract

Decentralization and distribution of processes became an issue in Artifi cial Intelligence (AI) as well as in several other areas of Computer Science in the past decades. These needs favored the creation of more autonomous, distributed and intelligent, software tools. Among these new software entities, labelled agents, there are those whose emphasis is in the intelligence component. Intelligence, as we perceive it,is strongly related to the capability of learning from previous experience and using stored knowledge to improve future behaviour. These forms of intelligent, or adaptive, behaviour are becoming a competitive factor in today's software. This new trend, that emphasizes autonomy, distribution and learning, brought new challenges. One of these challenges is to expand the Machine Learning (ML) paradigms from, the old, single-agent, perspective, to this new world where software inhabits an environment that is much more dynamic and in which agents have only a partial, and often noisy, view of the state that surrounds them. The extension of learning to these new environments must overcome the dif culties of this new paradigm, but should also take advantage of its bene ts. The fact that a multitude of agents populates the software environments, and in some cases they are learning to solve similar problems, leads s to the current research issue that can be summarized as: (How) can agents bene t from the exchange of information during the learning process?