Performance of Reinforcement Learning on Traditional Video Games

Reinforcement learning is a model structure that learns and updates the model based on receiving the reward and feedback of certain environments. In recent researches, reinforcement learning has been shown to be very effective in a well- defined or simulated environment to identify optimal strategies for solving given problems. This work test the performance of several Deep Reinforcement Learning model structures on a classic video game, Flappy bird. This project wants to further analyze the performance of different settings of deep Q networks and policy gradient models, to see how different model structures affect the training speed, accuracy, and efficiency. The author has found that the Double DQN model can achieve the best score among all deep learning models, whereas the Q-table model can achieve a better score with a proper feature extraction method, as the environment is rather simple.