Optimization of Gear Changing using Simulated Annealing

Abstract

The current paper puts forward a study conducted on a racing simulator, with the purpose of finding appropriate revolutions per minute (RPM) values for changing the gears of a car. The intricacy of the problem is given by many factors: it has 12 dimensions, 6 values for increasing and 6 for decreasing the gears, a circuit may contain straight and twisty parts, so distinct RPM values may be optimal for different sections, the move to another circuit makes the previous parameters suboptimal, while adding other competitors increases the amount of noise. Overall, we probably have all the ingredients of a real-world problem and we tackle it with two approaches that are economical as concerns the number of simulations needed and conduct to some interesting findings. We consider as study cases 2 different circuits, and racing alone and against 2 opponents.
This paper is a second approach to optimize gear changing values used in the game TORCS. Previously we used Hill Climbing algorithm to give proof that gear changing learning does reduce lap times. Now by using Simulated Annealing we improved even more the lap times and got better correlated sets of gear values.