Heuristics for car setup optimisation in TORCS

A TORCS-based (The Open Racing Car Simulator) car setup optimisation problem requires a search for the best parameter settings of a race car that improves its performance across different types of race tracks. This problem often exhibits a noisy environment due to the properties of the race track as well as the components of the car. Selection hyper-heuristics are methodologies that control and mix different predefined set of heuristics during the search process for solving computationally hard problems. In this study, we represent the car setup problem as a real valued optimisation problem and investigate the performance of different approaches including a set of heuristics and their combination controlled by a selection hyper-heuristic framework. The results show that selection hyper-heuristics and a tuned heuristic perform well and are promising approaches even in a dynamically changing, noisy environment.

I. Introduction

Many heuristic optimisation algorithms require parameter tuning to perform well in different problems, even in different instances of the same problem. So parameter tuning becomes a very crucial factor in achieving a good optimisation behavior. In the traditional way of parameter tuning, a set of parameters are experimented with before the real run and a good setting of these parameters are obtained [1]. However, this is very time consuming. Since an exhaustive search of all the possible parameter settings is not possible, generally a sub-optimal setting is achieved. Another issue is the fact that a good set of parameters may not be constant over the whole run. Based on the properties of the landscape, optimal parameter settings may also vary over time. There are many studies that deal with parameter tuning techniques. In [2], a detailed history of parameter setting and tuning techniques is provided, while a good survey is provided in [3] and more recently in [4].