Groupe d’études et de recherche en analyse des décisions

Learning and Distributed Learning in Multi-Agent Systems #1

Paolo Caravani University of L'Aquila, Italie

Recent challenges in the governance of networks and networked systems revived interest in game theory as a paradigm to study collective behaviour. Common traits of multi-agent systems are the lack of complete information, the large and often unknown number of players, the incomplete knowledge of other players’ payoffs or rationality. These features hinder the possibility of knowing in advance equilibrium strategies. The informative gap can be partially filled by transmission and reception of selected information among players and in certain conditions individually-optimal strategies can be learned while playing. Game morphology, information structure and transmission protocols greatly influence and condition the degree of success of different learning schemes.

In these two lectures I first survey classical and more recent results on game classification from the viewpoint of their amenability to learning. Then I will survey particular learning schemes that can be adopted in function of the available game structure.

Lecture one
Classification of games into types:

Zero-sum games
Common-interest games
Passive games
Exact potential games
Ordinal potential games
Dominance-solvable games
Max-solvable games

At the end of each lecture there will be a discussion session. Reading material will also be made available.