Interpretable, Verifiable, and Robust Reinforcement Learning Via Program Synthesis

Reinforcement learning is a promising strategy for automatically training policies for challenging control tasks. However, state-of-the-art deep reinforcement learning algorithms focus on training deep neural network (DNN) policies, which are black box models that are hard to interpret and reason about. In this chapter, we describe recent progress towards learning policies in the form of programs. Compared to DNNs, such programmatic policies are significantly more interpretable, easier to formally verify, and more robust. We give an overview of algorithms designed to learn programmatic policies, and describe several case studies demonstrating their various advantages.