Deep Q networks for visual fighting game AI

Recently, the introduction of vision-based deep Q learning demonstrated successful results in Atari, and Visual Doom AI platform. Unlike the previous study, the fighting game assumes two players with a relatively large number of actions. In this study, we propose to use deep Q Networks (DQN) for the visual fighting game AI competitions. The number of actions was reduced to 11 and the sensitivity of several control parameters was tested using the visual fighting platform. The experimental results show the potential of the DQN approach for the two- player real-time fighting game.