Zero-Order Optimization Methods with Applications to Reinforcement Learning

We consider the problem of minimizing a noisy objective function using only function values. We first review the main approaches in the literature and then focus on two strategies that aim at estimating gradients. One is based on Gaussian smoothing, and we illustrate its performance on some reinforcement learning tasks. The second strategy makes use of a noise estimation procedure due to Hamming, and we test it on general black-box optimization problems.