Self-Play for Training General Fighting Game AI
Yoshina Takano, Hideyasu Inoue, Ruck Thawonmas, Tomohiro Harada · 2019
In this paper, we train a general fighting game AI from self-play games to outperform an unseen opponent AI. It has been reported that an AI using Deep Q Network (DQN) can outperform the training partner. However, according to our experience, the DQN AI is not always superior to a new opponent, unseen before. By learning from self-play, we overcome this drawback while maintaining the DQN AI’s strong points. Our experimental results show that it is more effective to use a variety of AIs with different behaviors as training partners.