Evolving Fuzzy Rules for Reactive Agents in Dynamic Environments

Fuzzy logic controllers have been applied to a wide range of control problems, but are very difficult to build for situations where the environment changes quickly and there is a lot of uncertainty. This work investigates a new method of creating fuzzy controllers, in the form of reactive agents, for such environments. The framework for this investigation is the RoboCup soccer simulation environment, where the agents are in the form of simulated soccer players evolved to exhibit competent dribble-and-score behaviours. The method proposed uses a messy genetic algorithm to evolve a set of behaviour producing fuzzy rules which define the agents. The results presented indicate that the messy genetic algorithm is well suited to this task, enabling a performance improvement over traditional evolutionary methods by reducing complexity, and that the agents produced perform well in their environment. The best agent evolved is consistently and reliably able to locate the ball, dribble it to the goal and score.