O que é Policy gradient, actor–critic method?

A policy gradient method adjusts the parameters $\theta$ of a stochastic policy in the direction $\E[G_t\nabla_\theta\log\pi_\theta(u_t\mid x_t)]$, the gradient of expected cumulative reward (REINFORCE; Williams, 1992); a baseline subtracted from $G_t$ lowers the variance without biasing it. An actor–critic method replaces $G_t$ by the temporal-difference error of a learned value function (the critic), so that the policy (the actor) learns at every step.