A context-based approach to detecting miscreant behavior and collusion in open multiagent systems

Most multiagent systems (MAS) either assume cooperation on the part of the agents or assume that the agents are completely self-interested, for example, in the case of bidding and other market-based approaches. However, an interesting class of MAS is one that is fundamentally cooperative, yet open, and in which one or more of the agents may be self-interested. In such systems, there is the potential for agents to misbehave, i.e., to be miscreants. Detecting this is tricky and context-dependent. Even more difficult is the problem of detecting collusion between agents. In this paper, we report on a project that is beginning to address this problem using a context-based approach. Features of the MAS' situation are used by a subset of the agents to identify it as an instance of one or more known contexts. Knowledge the agent(s) have about those contexts can then be used to directly detect miscreant behavior or collusion or to select the appropriate technique for the context with which to do so. The work is based on context-mediated behavior (CoMB), and it develops a new form of collusion detection called society-level analysis of motives (SLAM).