In fighting games, players defeated repeatedly may lose interest in the games and quit playing. Therefore, it’s essential to develop support AI assisting them at a low intervention rate to maintain their enjoyment. However, prior works have shown controlling the intervention rate is an issue in environments with a large action space like fighting games. In this study, we propose the following methods (1) linear interventional penalty function to suppress excessive interventions, (2) \(\epsilon \) -submission method to encourage exploration of player actions, (3) A special action that represents executing the same action as the player. We also utilize policy ensemble to generate diverse virtual players to enhance the AI’s collaboration with unknown players. To evaluate, we compare the training results of the baseline method and our method in Lunar Lander and FightingICE. Especially in FightingICE, our AI achieves the same score with only 11% intervention rate compared to the baseline’s 45% intervention rate. Furthermore, we conduct an experiment where 15 subjects use our support AI with intervention rates of 20%, 30%, and 60%. The average number of their wins in 10 matches is 4.5 even with our 20% intervention support AI, while it is 0.8 without it. Moreover, subjects rate the 20% and 30% intervention support AIs higher than the 60% one in terms of the appropriateness of the frequency of support and timing, the freedom of operation and the enjoyment. Our model has the potential for application to other games besides fighting games. Therefore, it’s expected to be studied for further applications.
