Constrained Policy Optimization

TLDR

Reinforcement learning often benefits from specifying reward functions alongside constraints, especially for safety‑critical systems, yet recent high‑dimensional policy search advances have not addressed constrained settings. The authors introduce Constrained Policy Optimization (CPO), a general‑purpose policy search algorithm that guarantees near‑constraint satisfaction at every iteration. CPO trains neural network policies for high‑dimensional control by enforcing near‑constraint satisfaction through a theoretical bound that relates expected returns of two policies to their average divergence. The algorithm is shown to successfully satisfy safety constraints in simulated robot locomotion tasks.

Abstract

For many applications of reinforcement learning it can be more convenient to specify both a reward function and constraints, rather than trying to design behavior through the reward function. For example, systems that physically interact with or around humans should satisfy safety constraints. Recent advances in policy search algorithms (Mnih et al., 2016, Schulman et al., 2015, Lillicrap et al., 2016, Levine et al., 2016) have enabled new capabilities in high-dimensional control, but do not consider the constrained setting. We propose Constrained Policy Optimization (CPO), the first general-purpose policy search algorithm for constrained reinforcement learning with guarantees for near-constraint satisfaction at each iteration. Our method allows us to train neural network policies for high-dimensional control while making guarantees about policy behavior all throughout training. Our guarantees are based on a new theoretical result, which is of independent interest: we prove a bound relating the expected returns of two policies to an average divergence between them. We demonstrate the effectiveness of our approach on simulated robot locomotion tasks where the agent must satisfy constraints motivated by safety.