Video Game Design and Performance Analysis Based on Reinforcement Learning

Abstract

The Reinforcement Learning (RL) implement real-time decision-making behavior in uncertain environment in order to optimize overall target. RL is considered to be the paradigm of future general artificial intelligence (AI) because it is very similar to the way human thinks, which is expected to be widely used in game designing. In order to make RL algorithm perform better in video games, the model-free algorithm is used and convolutional neural network (CNN) is used in the pre-processing. In this paper, the performance of Deep Q Network (DQN) and Double Deep Q Network (DDQN) algorithms were tested and compared in the Arkanoid Game. The experiment shows that the DQN algorithm has a problem of overestimating the Q value and the DDQN algorithm can solve it well. In addition, the two algorithms are similar in terms of loss changes, but the learning effect of DDQN performs better.