Feasible Policy Iteration (2304.08845v2)

Abstract: Safe reinforcement learning (RL) aims to find the optimal policy and its feasible region in a constrained optimal control problem (OCP). Ensuring feasibility and optimality simultaneously has been a major challenge. Existing methods either attempt to solve OCPs directly with constrained optimization algorithms, leading to unstable training processes and unsatisfactory feasibility, or restrict policies in overly small feasible regions, resulting in excessive conservativeness with sacrificed optimality. To address this challenge, we propose an indirect safe RL framework called feasible policy iteration, which guarantees that the feasible region monotonically expands and converges to the maximum one, and the state-value function monotonically improves and converges to the optimal one. We achieve this by designing a policy update principle called region-wise policy improvement, which maximizes the state-value function under the constraint of the constraint decay function (CDF) inside the feasible region and minimizes the CDF outside the feasible region simultaneously. This update scheme ensures that the state-value function monotonically increases state-wise in the feasible region and the CDF monotonically decreases state-wise in the entire state space. We prove that the CDF converges to the solution of the risky BeLLMan equation while the state-value function converges to the solution of the feasible BeLLMan equation. The former represents the maximum feasible region and the latter manifests the optimal state-value function. Experiments show that our algorithm learns strictly safe and near-optimal policies with accurate feasible regions on classic control tasks. It also achieves fewer constraint violations with performance better than (or comparable to) baselines on Safety Gym.