Optimal Neuro-Evolution for Simulated Self-Driving Race Cars

NICHOLAS S. WEINTRAUT

In automobile racing, the objective of racers is to drive a car around a lap faster than all opponents. Such a task is well suited to neuro-evolution training for a neural network-based AI, as the results of the AI at each attempt can be easily graded by the lap time. The neuro-evolution method for neural networks has often been applied for various digital car-racing solutions, but often with naive training systems or network inputs. This paper aims to improve upon several existing methods by using an adversarial “King of the Hill” training solution, as well as a spline-based perfectlocalization method for state input to the neural network. The AI and training system are implemented in the Unity game engine for simulation and later playback of results. Existing naive training methods are tested alongside the adversarial method, and different network input feeds are tested, all in order to evaluate the agent with the quickest times on track laps.