Compressing the Evaluation Function With Small-Scale Deep Learning on Othello

Abstract

Othello AI has made significant progress in both evaluation and search algorithms over time. However, a major challenge in creating a highly accurate evaluation function is that the number of evaluation parameters can be enormous, typically around 10 to 40 million. In this article, we compress the evaluation function of Othello AI by using deep learning for function approximation. As a result, we demonstrate that it is possible to reduce the data volume to approximately 0.45% of that of the uncompressed evaluation function while maintaining the same performance. Furthermore, we show that compressing the data to 6.4% of its original size can actually improve performance compared to the uncompressed evaluation function. We also compare our method with other lossless and lossy compression techniques and show that our approach outperforms them in terms of compression efficiency. Using this method, we developed an Othello AI that is capable of defeating Edax 4.6, an open-source Othello AI currently regarded as one of the strongest in the world. Additionally, we applied this method to the Othello AI global contest on CodinGame, where our AI secured first place worldwide.