Learning Representations and Robust Exploration for Improved Generalization in Reinforcement Learning

Nasik Muhammad Nafi

Abstract: Deep Reinforcement Learning agents typically aim to learn a task through interacting in a particular environment. However, training on such singleton RL tasks, where the agent interacts with the same environment in every episode, implicitly leads to overfitting. Thus, the agent fails to generalize to minor changes in the environment, especially in image-based observation. Generalization is one of the main contemporary research challenges and recently proposed environments that enable diversified episode generation opens up the possibility to investigate generalization. My initial work towards this objective includes representation learning through the partial decoupling of policy and value networks and hyperbolic discounting in a single-agent setting. Efficient exploration is another crucial aspect of achieving generalization when learning from limited data. My dissertation would focus on proposing and evaluating methods that enable better representation learning and exploration for unseen scenarios. Another key objective is to extend my work to multi-agent generalization which is comparatively less studied.