FightingICE is a challenging fighting game Fighting game that requires the player to decide an action to perform among many actions. We developed several RL-based agents (DQN, Double DQN, Proiritized DQN, Dueling DQN, ACKTR, etc.) to fight against existing opponents. Our ACKTR algorithm achieved 50% win rate after 550 episodes and over 90% win rate after 1000 trainings. We further improved the algorithm by predicting the next action of the opponent through a three-layer network, which only took 500 episodes to achieve a 90 winning rate. Details stated in video.
