Testing the Limits of Emergent Behavior in MAS Using Learning of Cooperative Behavior

Abstract. We present a method to test a group of agents for (unwanted) emergent behavior by using techniques from learning of cooperative behavior. The general idea is to mimick users or other systems interacting with the tested agents by a group of tester agents and to evolve the actions of these tester agents. The goal of this evolutionary learning is to get the tested agents to exhibit the (unwanted) emergent behavior. We used our method to test the emergent properties of a team of agents written by students for an assignment in a basic MAS class. Our method produced tester agents that helped the tested agents to perform at their best and another configuration of our system showed how much the tested agents could hold their own against very competitive agents, which revealed breakdowns in the tested agents' cooperation.

• Type: Conference
• Year: 2006
• Where: ECAI
• Authors: Jordan Kidney, Jörg Denzinger
• Keywords: Artificial Intelligence, ECAI 2006, Emergent Behavior, Method Produced Tester, Tester Agents