UC Berkeley

In this thesis, we discuss various techniques for improving exploration for deep reinforcement learning. We begin with a brief review of reinforcement learning (RL) and the fundamental v.s. exploitation trade-off. Then we review how deep RL has improved upon classical and summarize six categories of the latest exploration methods for deep RL, in the order increasing usage of prior information. We then explore representative works in three categories discuss their strengths and weaknesses. The first category, represented by Soft Q-learning, uses regularization to encourage exploration. The second category, represented by count-based via hashing, maps states to hash codes for counting and assigns higher exploration to less-encountered states. The third category utilizes hierarchy and is represented by modular architecture for RL agents to play StarCraft II. Finally, we conclude that exploration by prior knowledge is a promising research direction and suggest topics of potentially impact.