Evolution Strategies

Evolution Strategies is a paper released in 2017 by Open AI that is an exploration of applying neuroevolution at scale.

The ideas presented here have been used in evolutionary algorithms for decades.

The Algorithm

Create a parameter vector.

Generate a population of size n by tweaking that parameter vector with gaussian noise.

Evaluate each member of the population by running the policy network in an environment.

Update the parameter vector by computing the weighted sum of the rewards of the n members of the population.