Reinforcement Learning of Intelligent Characters in Fighting Action Games

In this paper, we investigate reinforcement learning (RL) of intelligent characters, based on neural network technology, for fighting action games. RL can be either on-policy or off-policy. We apply both schemes to tabula rasa learning and adaptation. The experimental results show that (1) in tabula rasa leaning, off-policy RL outperforms on-policy RL, but (2) in adaptation, on-policy RL outperforms off-policy RL.