On Social Reasoning in Multi-Agent Systems

Abstract

"This work presents the core notions of a social reasoning mechanism, based on dependence theory. This modelenables an agent to reason about the others, in particular to calculate his dependence relations and dependencesituations. An agent is said to be dependent on another if the latter can help/prevent him to achieve one of his goals.We consider our social reasoning mechanism as an essential building block for the design of really autonomousartificial agents, which are immersed in an open multi-agent world. By open, we mean that agents may enter orleave the society at any moment. In such systems, as the organization of the agents can not be conceived at designtime, the cooperative problem solving paradigm is based on dynamic coalition formation. In this context, agentsmust be able to adapt themselves to dynamically changing conditions, by evaluating at execution time if theirgoals are achievable and if their plans are feasible. As we do not suppose that agents are benevolent, our modelproposes a criterion to evaluate which partners are more susceptible to accept a proposition of coalition. Finally,as in these kind of systems agents usually do not have a complete and correct representation of each other, ourmodel helps them to detect an agency level inconsistency and to choose a context to be maintained"