Policy Gradient Methods for Reinforcement Learning with Function Approximation

Function approximation is essential to reinforcement learning, but the standard approach of approximating a value function and determining a policy from it has so far proven theoretically intractable. In this paper