Trust Region Policy Optimization

ABSTRACT

We describe an iterative procedure for optimizing policies, with guaranteed monotonic improvement. By making several approximations to the theoretically-justified procedure, we develop a practical algorithm, called Trust Region Policy Optimization (TRPO). This algorithm is similar to natural policy gradient methods and is effective for optimizing large nonlinear policies such as neural networks. Our experiments demonstrate its robust performance on a wide variety of tasks: learning simulated robotic swimming, hopping, and walking gaits, as well as playing Atari games using images for input. TRPO's key innovation is a KL-divergence constraint that prevents policy updates from changing the network's behavior too drastically in a single step.