Programmatic Policy

Programmatic policy research focuses on representing and learning policies as executable programs, aiming for improved interpretability, controllability, and generalization compared to traditional black-box methods like deep reinforcement learning. Current research explores various approaches, including tree-based policies, state machine integrations, and the use of large language models for policy synthesis and evaluation, often within the context of interactive agents and complex tasks. This work holds significant promise for enhancing the reliability and trustworthiness of AI systems across diverse applications, from robotics and game playing to access control and human-computer interaction, by providing more understandable and modifiable decision-making processes.