Utilizing Multiple Agents for Decision Making in a Fighting Game

In this paper, we propose a method for implementing a competent game AI using two expert agents that are separately trained for different purposes. Deep reinforcement learning (DRL) has been successfully applied to achieve human-level results for many types of games. However, DRL has not yet been successfully applied to fighting games in terms of performance. We, thus, examine if a competent fighting game AI can be implemented by DRL. The proposed method uses two expert agents which are trained using separate Deep Q-Networks (DQNs) for a fighting game AI. For performance evaluation, we use FightingICE which has been used as a game platform in a game AI competition at a game-AI international conference since 2014. Our experimental results show that proposed AI is superior to an AI using a single DQN.