Analysis, Combination and Integration of Neuroevolution and Backpropagation Algorithms for Gaming Environment

Abstract

This paper provides a method of combining Neu-roevolution with Backpropagation to achieve lower training times than Neuroevolution when training agents in a video game environment. The combination of these algorithms is reproduced by an alteration of the step of creating a new generation from the most capable agents with the creation of a new generation through the Backpropagation method using the preventively saved data of the most capable agent from the environment. Thus, for the new generation, a Neural Network trained with backpropagation is assigned instead of the best-performing Neural Network from the previous generation. As a result, the Neuroevolution with the Backpropagation method shows better performance when increasing the target of the environmental performance of the agent.