Playing the Snake Game with Reinforcement Learning

Abstract

When it is necessary to make several decisions to solve a problem, reinforcement learning works well. In this project, we demonstrate the outcomes of playing with snakes using a Deep Q-Network (DQN). Unfortunately, DQN has a tendency to overfit, which results in a deterioration in terms of scores that are subpar after numerous training sessions. To solve this issue, we enhanced the reward setting and provided the snake with more information about its surroundings in order to raise our score and prevent some overfitting issues. Also, we went through the differences between DQN and conventional evolutionary algorithms as well as possible avenues for DQN optimization.