Learning to Play Atari Games

Teaching computers to play video games is a complex learning problem that has recently seen increased attention. In this paper, we develop a system that, using constant model and hyperparameter settings, learns to play a variety of Atari games. In order to accomplish this task, we extract object features from the game screen, and provide these features as input into reinforcement learning algorithms. We detail our testing of different hyperparameter settings for two reinforcement learning algorithms, using both linear and neural network function approximators, in order to compare the performance of these approaches. We also consider learning techniques such as the use of replay memory with Q-learning, finding that even on simple MDPs this method significantly improves performance. Finally, we evaluate our model, selected through validation on the game Breakout, on a test set of different video games.