Although multi-agent simulations are an intuitive way of conceptualizing systems that consist of autonomous actors, a major problem is the actual design of the agent behavior. In this contribution, we examine the potential of using agent-based learning for implementing the agent behavior. We enhanced SeSAm, a platform for agent-based simulation, by replacing the usual rule-based agent architecture by XCS, a well-known learning classifier system (LCS). The resulting model is tested using a simple evacuation scenario. The results show that on the one hand side plausible agent behavior could be learned. On the other hand side, though, the results are quite brittle concerning the frame of environmental feedback, perception and action modeling.
展开▼