Driving Cars by Means of Genetic Algorithms
Proceedings of the 10th international conference on Parallel Problem Solving from Nature: PPSN X
The simulated car racing competition @ CIG-2008: events reports
ACM SIGEVOlution
Evolving competitive car controllers for racing games with neuroevolution
Proceedings of the 11th Annual conference on Genetic and evolutionary computation
Robust player imitation using multiobjective evolution
CEC'09 Proceedings of the Eleventh conference on Congress on Evolutionary Computation
On-line neuroevolution applied to the open racing car simulator
CEC'09 Proceedings of the Eleventh conference on Congress on Evolutionary Computation
Learning drivers for TORCS through imitation using supervised methods
CIG'09 Proceedings of the 5th international conference on Computational Intelligence and Games
A modular parametric architecture for the TORCS racing engine
CIG'09 Proceedings of the 5th international conference on Computational Intelligence and Games
Evolving a fuzzy controller for a car racing competition
CIG'09 Proceedings of the 5th international conference on Computational Intelligence and Games
Controller for TORCS created by imitation
CIG'09 Proceedings of the 5th international conference on Computational Intelligence and Games
Evolving driving controllers using genetic programming
CIG'09 Proceedings of the 5th international conference on Computational Intelligence and Games
Optimized sensory-motor couplings plus strategy extensions for the TORCS car racing challenge
CIG'09 Proceedings of the 5th international conference on Computational Intelligence and Games
Evolutionary computation and games
IEEE Computational Intelligence Magazine
Multiple linear regression modeling for compositional data
Neurocomputing
| Hi-index | 0.01 |
Designing controllers for simulated cars is a challenging task because there are numerous sensor inputs and a lot of actuators to be controlled. Although it is possible to use domain-expert knowledge on the car racing, it is not trivial to represent the knowledge into controllers and tune the parameters for them. Therefore, it is natural to adopt machine learning approach into the knowledge-oriented controllers to enhance their performance and minimize tedious parameter tunings. In this paper, we try to enhance our own heuristic controllers using machine learning models which decide the appropriate parameters of the heuristic controllers given the current sensory inputs. At first, we predict the desired speed only using the equations derived by the linear regression analysis for the both curved and straight track segments. Because the decision on reducing speed before the corner is more complex than the one in the straight line and the corners, it is necessary to use a non-linear model such as artificial neural networks. Secondly, the linear regression and artificial neural networks are specialized to predict desired speed in different situations. Experimental results on TORCS-based car racing simulations show that the combination of the two machine learning algorithms with the heuristic outperforms other alternatives (heuristic only, heuristic+linear regression, and heuristic+artificial neural networks).