Implementing Game Strategies Based on Reinforcement Learning

Abstract

Artificial intelligence (AI) technology such as reinforcement learning is increasingly used in playing game in recent years. A deep reinforcement learning model was used to play the game Flappy Bird. This paper aimed to let the computer play a simple game and get the corresponding data for AI learning. Game image was sequentially scaled, grayed, and adjusted for brightness. Before the current frame entered a state, the multi-dimensional image data of several frames of image superposition and combination was processed. Deep Q Network algorithm realized the best action prediction of the game execution in a specific game state, and successfully converted a game decision problem into the classification and recognition problem of instant multi-dimensional images and solved it with a convolutional neural network. After analysis, computer players controlled by deep neural networks had better results than human players. This experiment was a model combined between a deep neural network model and reinforcement learning, and could be applied in other games.