Improving the Human-Likeness of Game AI’s Moves by Combining Multiple Prediction Models

Strong game AI’s moves are sometimes strange or difficult for humans to understand. To achieve better human-computer interaction, researchers try to create human-like game AI. For chess and Go, supervised learning with deep neural networks is one of the most effective methods to predict human moves. In this study, we first show that supervised learning is also effective in Shogi (Japanese chess) to predict human moves. We also find that the AlphaZero-based model more accurately predicted moves of players with higher skill. We then investigate two evaluation metrics for measuring human-likeness, where one is move-matching accuracy that is often used in existing works, and the other is likelihood (the geometric mean of human moves’ probabilities predicted by the model). To create game AI that is more human-like, we propose two methods to combine multiple move prediction models. One uses a Classifier to select a suitable prediction model according to different situations, and the other is Blend that mixes probabilities from different prediction models because we observe that each model is good at some situations where other models cannot predict well. We show that the Classifier method increases the move-matching accuracy by 1%-3% but fails to improve the likelihood. The Blend method increases the move-matching accuracy by 3%-4% and the likelihood by 2%-5%.The 15th International Conference on Agents and Artificial Intelligence (ICAART 2023), Lisbon, Portuga