Another class of black-box methods that has recently become a popular alternative to the value-based and PG methods is genetic algorithms or GA. It is a large family of optimization methods with more than two decades of history behind it and a simple core idea of generating the population of N individuals, each of which is evaluated with the fitness function. Every individual means some combination of model parameters. Then some subset of top performers is used to produce (which is called mutation) the next generation of the population. This process is repeated until we're satisfied with the performance of our population.

There are lots of different methods in the GA family, for example, how to complete the mutation of the individuals for the next generation or how to rank the performers. Here we'll consider the simple GA method with some extensions, published in the paper by Felipe Petroski Such, Vashisht Madhavan, and others, called Deep Neuroevolution: Genetic Algorithms...