One subset of black-box optimization methods is called evolution strategies (ES), and it was inspired by the evolution process. With ES, the most successful individuals have the highest influence on the overall direction of the search. There are many different methods that fall into this class, and in this chapter, we will consider the approach taken by the OpenAI researchers Tim Salimans, Jonathan Ho, and others in their paper, Evolution Strategies as a Scalable Alternative to Reinforcement Learning , published in March 2017.
The underlying idea of ES methods is simple: on every iteration, we perform random perturbation of our current policy parameters and evaluate the resulting policy fitness function. Then, we adjust the policy weights proportionally to the relative fitness function value.
The concrete method used in the paper is called covariance matrix adaptation evolution strategy (CMA-ES), in which the perturbation performed is the random noise...