Simulation-Based Optimization of Autonomous Vehicles Using Genetic Algorithm

Autonomous vehicles are becoming a reality, yet it is only truly autonomous with a strong model. The system can be trained by implementing a genetic algorithm as an optimizer in its simulations. The authors explore existing studies in learning algorithms for AVs training and, through this paper, propose NeuroEvolution of Augmenting Topologies (NEAT), a type of genetic algorithm, to be implemented to train neural network models for AVs in a simulated environment. An optimization method is further done which involves hyperparameter tuning to find the optimal set of hyperparameters for the learning algorithm. The findings suggest that NEAT not only accelerates the development process but also enhances the AV's adaptability across different environments, achieving peak performance with a population size of 60 and a mutation rate of 0.8. This configuration attained the highest fitness scores 84, 98, and 100 on Tracks A, B, and C, respectively, with optimal results observed around the 50th generation. This research focuses on the application of genetic algorithm to AV simulations with optimization, uncovering NEAT's potential. Future work will focus on simulations in more complex and real-world environments.