Learning to Play Pac-Xon with Q-Learning and Two Double Q-Learning Variants

Pac-Xon is an arcade video game in which the player tries to fill a level space by conquering blocks while being threatened by enemies. In this paper it is investigated whether a reinforcement learning (RL) agent can successfully learn to play this game. The RL agent consists of a multilayer perceptron (MLP) that uses a feature representation of the game state through input variables and gives Q-values for each possible action as output. For training the agent, the use of Q-learning is compared to two double Q-learning variants, the original algorithm and a novel variant. Furthermore, we have set up an alternative reward function which presents higher rewards towards the end of a level to try to increase the performance of the algorithms. The results show that all algorithms can be used to successfully learn to play Pac-Xon. Furthermore both double Q-learning variants obtain significantly higher performances than Q-learning and the progressive reward function does not yield better results than the regular reward function