Evolving Neural Network Controllers for a Team of Self-Organizing Robots

Abstract

Self-organizing systems obtain a global system behavior via typically simple local interactions among a number of components or agents, respectively. The emergent service often displays properties like adaptability, robustness, and scalability, which makes the self-organizing paradigm interesting for technical applications like cooperative autonomous robots. The behavior for the local interactions is usually simple, but it is often difficult to define the right set of interaction rules in order to achieve a desired global behavior. In this paper, we describe a novel design approach using an evolutionary algorithm and artificial neural networks to automatize the part of the design process that requires most of the effort. A simulated robot soccer game was implemented to test and evaluate the proposed method. A new approach in evolving competitive behavior is also introduced using Swiss System instead of the full tournament to cut down the number of necessary simulations.