In this project Deep Q-Learning was explored for playing Atari Breakout. In particular Deep Q-Learning (DQN) (Mnih et al. 2015), Double Deep Q-Learning (DDQN) (Van Hasselt et al. 2016), Duelling Deep Q-Learning (Wang et al. 2015) and Double Deep Q-Learning with Prioritised Replay Memory (DDQN with PRM) (Schaul et al. 2016) were implemented. The project formed the final coursework for a Reinforcement Learning course. Madge Kelly, Joe Rosenberg and Joe Molyneux-Saunders (myself) made equal contributions to the project. A detailed explaination of the implementations and discussion of the results can be found in the report PDF.

Double Deep Q-Learning Agent Playing Breakout.

Learning curves for each of the 4 deep reinforcement learning methods implemented showing the moving average reward (100 episodes) per training epoch (100 episodes).