Playing Games with Deep Reinforcement Learning

Recently, Google Deepmind showcased how Deep learning can be used in conjunction with existing Reinforcement Learning (RL) techniques to play Atari games[10], beat a world-class player [13] in the game of Go and solve complicated riddles [3]. Deep learning has been shown to be successful in extracting useful, nonlinear features from high-dimensional media such as images, text, video and audio [11]. For control tasks like playing games, researchers have traditionally used handcrafted-features which require a lot of human effort. Convolutional neural networks (CNN) can be used to extract useful visual features from the highdimensional input, such as a video game screen, to learn near-optimal value function in such control tasks. We chose the Arcade Learning environment as our testbed to examine how deep learning can be applied to a reinforcement learning setting and analyze performance of various modern algorithms. The learning is performed end-to-end with no game knowledge included in the architecture or during training. We also show how the architecture of the deep network used, can be modified to achieve superior performances and compare performances on the task of playing a complicated game Seaquest.