Empirical study on the performance of Neuro Evolution of Augmenting Topologies (NEAT)

In this paper we provide empirical results on training a neural network with a genetic algorithm. We test various features of the generalized genetic algorithms, namely spieciation and fitness sharing and present the statistical analysis of all three variations. An obstacle avoidance problem was created in which the objective is for vehicles to traverse the course. We present interesting observations about the differences between evolutionary techniques and argue that there is a significant benefit in approaches that aim to diversify the gene pool as a mechanism for avoiding local minima.