Automatic adaptation to generated content via car setup optimization in TORCS

The car setup optimization problem as employed for a recent competition is a real-valued, 22 variable gray box problem (some dependencies are known) which challenges optimization algorithms in many ways. Runs must be short, there is a considerable amount of noise, and evaluation times for each solution candidate have to be determined by the algorithm itself. We take this as example problem and ask what happens if a user or a procedural content generator provides new tracks on which the standard cars are almost undriveable. Consequently, we suggest to use an optimization algorithm to adapt the cars to the track in almost real-time (minutes). We investigate how the CMA-ES, a modern evolutionary strategy, fares in this context and suggest some means to adapt it to the requirements. Attempts to improve the results using special noise handling methods unfortunately fail, most likely due to the very hard time constraints. Additionally, we perform a basic test with humans driving the optimized cars and have a short look at the properties of the cars changed for improving their performance.