Autonomous agents in snake game via deep reinforcement learning

Abstract

Since DeepMind pioneered a deep reinforcement learning (DRL) model to play the Atari games, DRL has become a commonly adopted method to enable the agents to learn complex control policies in various video games. However, similar approaches may still need to be improved when applied to more challenging scenarios, where reward signals are sparse and delayed. In this paper, we develop a refined DRL model to enable our autonomous agent to play the classical Snake Game, whose constraint gets stricter as the game progresses. Specifically, we employ a convolutional neural network (CNN) trained with a variant of Q-learning. Moreover, we propose a carefully designed reward mechanism to properly train the network, adopt a training gap strategy to temporarily bypass training after the location of the target changes, and introduce a dual experience replay method to categorize different experiences for better training efficacy. The experimental results show that our agent outperforms the baseline model and surpasses human-level performance in terms of playing the Snake Game.