An modeling processing method for video games based on deep reinforcement learning

In the reinforcement learning field, the traditional Q-Learning method can make the agent get a good grades on some simple games whose states are limited. Meanwhile, if the model of game is relatively simple, the continuous states can also be transformed into finite states through discretization so that the agent can achieve good results. However, traditional Q-Learning will face many difficulties for video games with image output, such as too complicated image states to store. In this paper, an algorithm of deep reinforcement learning(DRL), called Deep Q-Network(DQN), which is originated from Q-Learning and combining with artificial neural network, is used to model the video games with its image output. Then in the model training process, some image processing optimization methods and neural network structure are put forward for the preparation of test. Finally, from the experimental results, it is concluded that DQN can make the agent get high scores from video game, and the human game data can make the model training faster and better.