Learning a Shogi Evaluation Function Using Genetic Algorithms

This paper introduces a novel approach to evolving a shogi evaluation function using genetic algorithms. This study explores an alternative to the commonly used mentor-assisted learning methods for shogi, also known as Japanese chess. Instead of relying on established concepts for learning evaluation functions, such as mentor-assisted learning, we employ a genetic algorithm utilizing the winning player as the sole learning input. Our novel dataset, compiled from 1 million board states scraped from professional games, served as the foundation for training. Our approach yielded a 70 % classification accuracy in determining the winner of shogi games from previously unseen board states when tested on a validation set. The results highlight the effectiveness of using genetic algorithms to evolve a shogi evaluation function and provide a further understanding of enhancing shogi AI using game outcomes as primary training data. Our method's computational efficiency also stands as an advantage over other techniques commonly employed in this domain. This work offers a fresh perspective in the realm of computer shogi with implications for future research and development.