Hands-On Neuroevolution with Python

The RL training of the game agent

The RL training in our experiment was implemented using the neuroevolution method. This method is based on a simple genetic algorithm that evolves a population of individuals. The genotype of each individual encodes the vector of the trainable parameters of the controller ANN. By trainable parameters, we mean the connection weights between the network nodes. In every generation, each genotype is evaluated against a test environment by playing Frostbite and produces a specific fitness score. We evaluate each agent (genome) against 20,000 frames of the game. During the evaluation period, the game character can play multiple times, and the final Atari game score is the fitness score, which is a reward signal ...