Evolving Neural Networks for the Capture Game

This paper proposes the use of a genetic algorithm to develop neural networks to play the Capture Game, a subgame of Go. The motivation for this is twofold: to evaluate and possibly improve upon current genetic algorithm variants in order to produce a good player and (more importantly) to use this process to examine the properties and processes that are present in evolutionary systems in an attempt to shed some light on the phenomena that are required for an evolutionary process to produce robust, perpetually improving individuals and avoid local minima without any outside interaction. A brief survey of related work in the area is given, which highlights some of the interesting research questions that remain. This is followed by an outline of a distributed system that has been developed for use in the experimental evaluation of some of the proposed ideas and some of the initial results generated by the system