Interpretable Policy Specification and Synthesis through Natural Language and RL

Pradyumna Tambwekar, Andrew Silva, Nakul Gopalan, Matthew C. Gombolay

Abstract: Human-AI collaborative policy synthesis is a procedure in which (1) a human initializes an autonomous agent's behavior, (2) Reinforcement Learning improves the human specified behavior, and (3) the agent can explain the final optimized policy to the user. This paradigm leverages human expertise and facilitates a greater insight into the learned behaviors of an agent. Existing approaches to enabling collaborative policy specification involve black box methods which are unintelligible and are not catered towards non-expert end-users. In this paper, we develop a novel collaborative framework to enable humans to initialize and interpret an autonomous agent's behavior, rooted in principles of human-centered design. Through our framework, we enable humans to specify an initial behavior model in the form of unstructured, natural language, which we then convert to lexical decision trees. Next, we are able to leverage these human-specified policies, to warm-start reinforcement learning and further allow the agent to optimize the policies through reinforcement learning. Finally, to close the loop on human-specification, we produce explanations of the final learned policy, in multiple modalities, to provide the user a final depiction about the learned policy of the agent. We validate our approach by showing that our model can produce >80% accuracy, and that human-initialized policies are able to successfully warm-start RL. We then conduct a novel human-subjects study quantifying the relative subjective and objective benefits of varying XAI modalities(e.g., Tree, Language, and Program) for explaining learned policies to end-users, in terms of usability and interpretability and identify the circumstances that influence these measures. Our findings emphasize the need for personalized explainable systems that can facilitate user-centric policy explanations for a variety of end-users.