Neural network-based vehicle control in simulated environments using real-coded genetic algorithms

¹ Siberian Federal University, 79, Svobodny Ave., Krasnoyarsk, 660041, Russia
² Reshetnev Siberian State University of Science and Technology, 31, Krasnoyarsky Rabochy Ave, Krasnoyarsk, 660037, Russia

^* Corresponding author: d.s.shabalin@yandex.ru

Abstract

This research explores a method of optimizing neural networks for vehicle control in a simulation environment using a real-coded genetic algorithm (RCGA). The study focuses on applying RCGA in conjunction with multiple genetic operators, including simulated binary crossover (SBX), power mutation (PM), and tournament selection, to evolve neural network weights and biases, enhancing control performance for simulated vehicles. By utilizing RCGA to adjust neural network parameters, the approach enables adaptive and efficient vehicle control. The experiments demonstrate that combining sensor data with neuroevolutionary optimization in a simulation leads to a highly reliable control system, achieving performance metrics comparable to human operators. These findings suggest that RCGA-based optimization methods can be effectively applied to complex dynamic systems in various technical fields.