No account yet ?

In reinforcement learning, a policy gradient method is an algorithm that learns a policy directly by being directly interested in it.

The policy gradient method allows the optimization of the parameterized policy with respect to the expected performance with the gradient descent method.

At the end of a certain number of iterations, the objective is to obtain a maximization of the performance of the policy for a studied model.

Policy gradient methods are therefore opposed to value-based methods which optimize values in order to then define the optimal policy for these values.