Combining Reinforcement Learning With Genetic Algorithm for Many-To-Many Route Optimization of Autonomous Vehicles

This study introduces an approach for route optimization of many-to-many Demand-Responsive Transport (DRT) services. In contrast to conventional fixed-route transit systems, DRT provides dynamic, flexible, and cost-effective alternatives. We present an algorithm that integrates DRT with the autonomous shuttles at Korea National University of Transportation (KNUT), allowing dynamic route modifications in real-time to accommodate incoming service calls. The algorithm is designed to take into account the shuttle’s current position, the destinations of passengers already on board, the current locations and destinations of individuals who have requested shuttle services, and the remaining capacity of the shuttle. The algorithm has been developed to combine genetic algorithms and reinforcement learning. The performance evaluation was conducted using a simulation model that emulates KNUT’s campus and the adjoining local community area. The simulation results show significant improvements in two key metrics, specifically the ‘Request to Pick-up Time’ and ‘Request to Drop-off Time’ during high-demand periods over the single-shuttle operation. Additional simulation test with random service requests and stochastic passenger walking distances showed the potential adaptability across different settings.