Learning to play games has been a topic of interest to researchers since the early days of artificial intelligence. The goal is to create programs that enable computers to play games intelligently. In recent years, we have seen deep learning being used more and more. AlphaZero is one of the deep reinforcement learning algorithms that has achieved superhuman level of play in Chess, Shogi and Go without any domain knowledge. In this paper, we used AlphaZero to learn how to play the game Connect Four, with a focus on using expert knowledge to improve it. Several methods are presented that introduce expert heuristics into the learning phase of the AlphaZero algorithm. Using field and feature heuristics, we analyzed different methods on sets of positions, games with error corrections, and four different opponents, one of which plays optimally. By using the feature heuristic, which encourages connecting game pieces, we were able to slightly improve the results of the position sets as measured by various metrics.
|
