Item

Abstract

A policy gradient method is a reinforcement learning approach that directly optimizes a parametrized control policy by a variant of gradient descent. These methods belong to the class of policy search techniques that maximize the expected return of a policy in a fixed policy class, in contrast with traditional value function approximation approaches that derive policies from a value function. Policy gradient approaches have various advantages: they enable the straightforward incorporation of domain knowledge in policy parametrization and often an optimal policy is more compactly represented than the corresponding value function; many such methods guarantee to convergence to at least a locally optimal policy; the methods naturally handle continuous states and actions and often even imperfect state information. The counterveiling drawbacks include difficulties in off-policy settings, the potential for very slow convergence and high sample complexity, as well as identifying local optima that are not globally optimal.