A NEAT Visualisation of Neuroevolution Trajectories

NeuroEvolution of Augmenting Topologies (NEAT) is a system for evolving neural network topologies along with weights that has proven highly effective and adaptable solving challenging reinforcement learning tasks. This paper analyses NEAT through the lens Search Trajectory Networks (STNs), recently proposed visual approach to study dynamics evolutionary algorithms. Our goal improve understanding neuroevolution systems. We present statistical analysis contrasting behaviour NEAT, without using crossover operator, when two benchmark problems outlined in original article: XOR double-pole balancing. Contrary what reported article, our experiments perform significantly better both domains.