Progressive neural network training for the Open Racing Car Simulator

Abstract

In this paper a novel methodology for training neural networks as car racing controllers is proposed. Our effort is focused on finding a new fast and effective way to train neural networks that will avoid stacking in local minima and can learn from advanced bot-teachers to handle the basic tasks of steering and acceleration in The Open Racing Car Simulator (TORCS). The proposed approach is based on Neural Networks that learn progressively the driving behaviour of other bots. Starting with a simple rule-based decision driver, our scope is to handle its decisions with NN and increase its performance as much as possible. In order to do so, we propose a sequence of Neural networks that are gradually trained from more dexterous drivers, as well as, from the simplest to the most skillful controller. Our method is actually, an effective initialization method for Neural Networks that leads to increasingly better driving behavior. We have tested the method in several tracks of increasing difficulty. In all cases the proposed method resulted in improved bot decisions.