How costly is a good compromise: Multi-objective TORCS controller parameter optimization

Abstract

We extend existing work on the offline parameter optimization for The Open Racing Car Simulator (TORCS) controllers and take it to a truly multi-objective level. By means of the (100+1)-SMS-EMOA, we optimize the parameter set for the controller named `Mr. Racer' on three significantly different tracks simultaneously, with a budget of 3 × 6000 function evaluations. In the ten runs performed, the SMS-EMOA reliably finds good compromise solutions, as well as selfish optima that are comparable in quality to the ones previously obtained by means of the CMA-ES for each particular track. We further analyze how to select parameter set(s) for the controller from the results of the evolutionary optimization, for the case that a controller has the chance to further finetune its behavior on an unknown track, as it is done in the warinup phase of the Simulated Car Racing Championship. Experimental results show that one parameter set is not sufficient. To perform well, a controller as Mr. Racer needs at least two different parameter sets from which it can choose in the warinup stage. The best performance is gained by using three parameter sets, which leads to an increase in championship points of 17% compared to the 2013 version of Mr. Racer.