Training an Agent using Deep Reinforcement Learning: Snake Game

Abstract

Deep Reinforcement Learning has become a commonly adopted method to enable agents to hunt out complex control policies in various video games. Deep-Mind used this technique to play Atari games. However, similar approaches should get to be improved when applied to tougher scenarios, where reward signals are sparse and delayed. This paper illustrates a refined Deep Reinforcement Learning model to enable an autonomous agent to play the classical Snake Game, whose constraints get stricter as the game progresses further. Specifically, to train this model we have used Deep Neural Network (DNN) with a variant of Q-learning where agent will learn from its past experiences. Moreover, we have proposed a designed reward mechanism to properly train the network, adopt a training gap strategy to temporarily bypass training after the situation of the target changes, and also introduces dual experience replay method through which different experiences for better training can be categorized. The final results show that our agent in an environment outperforms the baseline model and surpasses the human-level performance in terms of playing the Snake Game.