Evolutionary Optimization of a Neural Network Controller for Car Racing Simulation

摘要

In this paper a novel method for car racing controller learning is proposed. Car racing simulation is an active research field where new advances in aerodynamics, consumption and engine power are modelled and tested. The proposed approach is based on Neural Networks that learn the driving behaviour of other rule-based bots. Additionally, the resulted neural-networks controllers are evolved in order to adapt and increase their performance to a given racing track using genetic algorithms. The proposed bots are implemented and tested on several tracks of the open racing car simulator (TORCS) providing smoother driving behaviour than the corresponding rule-based bots and increased performance using the evolutionary adaptation.