Discovering Emergent Agent Behaviour with Evolutionary Finite State Machines

Martin Masek, Chiou-Peng Lam, Lyndon Benke, Luke Kelly, Michael Papasimeon

Abstract: In this paper we introduce a novel approach to discovering emergent behaviour in multiagent simulations, using evolutionary finite state machines to model intelligent agents in an adversarial two-player game. Agent behaviour is modelled as a finite set of predetermined states. The logic that leads to transitions between states is evolved to maximise fitness, which is determined through execution in a constructive simulation environment. The resultant evolved finite state machine (E-FSM) is evaluated for two finite state machine implementations, one with states specifically designed to perform a known behaviour and the other with states consisting of generic actions. Our experiments demonstrate that this approach can discover complex emergent behaviours from simple, generic actions, and use these behaviours to achieve a position of tactical superiority in the domain of air combat simulation.