Evolving the autosteering of a car featuring a realistically simulated steering response

We consider the area of intelligent road vehicles, especially, the topic of automated vehicles. Focusing on the importance of the automated steering, we address the challenge of automated keeping of a car in the middle of the driving lane. Our objective is to investigate the feasibility of employing genetic programming (GP) to evolve the automated steering of a car. The latter is implemented in the Open Source Racing Car Simulator (TORCS) with a realistically modeled steering featuring both a delay of response and a rate limit. We propose two approaches aimed at improving the efficiency of evolution via GP. In the first approach we implement an incremental evolution of the steering function by commencing the evolution with an ideal car and gradually increasing the degree of its realism (i.e., the amount of steering delay) in due course of evolution. The second approach is based on incorporating expert knowledge about the (expected) structure of the steering function according to the servo control model of steering. The experimental results verify that the proposed approaches yield an improved efficiency of evolution in that the obtained solutions are both of a better quality and could be obtained faster than those of the canonical GP.