Evolution of Artificial Neural Networks for Othello Board State Evaluation

Abstract

To make informed choices of encodings for Evolutionary Artificial Neural Networks, empirical data types of these is needed.In this thesis, the performance of a direct encoding specifying the weights of a fully connected feed-forward network is compared to the performance of two indirect Weighted Function Graph encodings on the task of playing the board game Othello, using evolution and coevolution.The results show that the indirect encodings perform better on the task in early generations in normal evolution, and slightly better in later generations under coevolution.%