Deep Learning Essentials

Deep Q-network (DQN)

Q-learning has been a major backbone to a large number of RL algorithms. However, this does not scale well to a high dimensional environment, such as building a RL system to play a game by Atari. DQN uses a Convolutional Neural Network (CNN) to map this high dimensional state-action space into stable Q-value function outputs. Figure Deep Q-network (DQN): high-level architecture, illustrates this interaction at a high level. The core idea is that CNNs are very useful in learning correlations in structured data. Hence, they take the raw image pixel values from the game screen and learn its correlation with an optimal action value (that is, joystick position) and output a corresponding Q-value. This allows us to approximate ...