Evolving Neural Controllers for Autonomous Driving using NEAT

In this work, we represent the investigations of the neuroevolution approaches for training neural controllers in autonomous driving. We compared some neuroevolution methods in the context of a simulated vehicle training. We train the navigation with tracking using velocity and spatial inputs: classical neuroevolution, NEAT, HyperNEAT, and NEAT with novelty search. In addition, we incorporated adaptive mutation strategies that respond to fitness stagnation. Our results demonstrated that applied neuroevolution approaches can successfully train neural controllers and dynamic mutation strategies. The solution improves the convergence of stability and performance, particularly for complex tasks with evolving network structures. These findings highlight the effectiveness of adaptive neuroevolution and define directions for future refinement in controller design and evolution strategies.