A Survey Analyzing Generalization in Deep Reinforcement Learning

Reinforcement learning research obtained significant success and attentionwith the utilization of deep neural networks to solve problems in highdimensional state or action spaces. While deep reinforcement learning policiesare currently being deployed in many different fields from medical applicationsto self driving vehicles, there are still ongoing questions the field is tryingto answer on the generalization capabilities of deep reinforcement learningpolicies. In this paper, we will outline the fundamental reasons why deepreinforcement learning policies encounter overfitting problems that limit theirrobustness and generalization capabilities. Furthermore, we will formalize andunify the diverse solution approaches to increase generalization, and overcomeoverfitting in state-action value functions. We believe our study can provide acompact systematic unified analysis for the current advancements in deepreinforcement learning, and help to construct robust deep neural policies withimproved generalization abilities.