Dealing with “Miscreant Agents”

Agent-based systems. An agent is an entity that perceives the world, makes decisions about how to behave, and then takes action. Multiagent systems (MAS) are software or software/hardware systems that are organized around agents. Multiagent systems research is increasingly important in controlling groups of robots and in controlling interacting software agents.

Agents behaving badly. The most common way of thinking about MASs is to treat the agents as benign, and possibly cooperating. However, once self-interested agents are allowed in the mix, there is the very real potential for an agent to decide that its best interests are not in line with the interests of others or with the group as a whole. In this case, the agent may be considered a “miscreant” agent, even if there is no overt goal on the part of the agent to hurt the system.

Predicting and protecting. The CA-SLAM (Context-Aware Society-Level Analysis of Motives) project on systems in which this kind of bad behavior could arise. Research issues include determining when the system risks this kind of behavior, how to detect such agents, and how to design the system so that this kind of behavior is avoided. CA-SLAM began as a PhD project in MaineSAIL; the student is now at UMA.