AI-Based Route Optimization in Urban Public Transport Networks

Abstract

Urban public transport systems are critical arteries of modern cities, yet they frequently grapple with inefficiencies such as prolonged journey times, uneven vehicle workloads, and suboptimal route overlaps. To address these challenges, this study presents an AI-based route optimization framework that synergizes a multi-objective genetic algorithm (GA) with a high-fidelity traffic simulation environment (MATSim). Leveraging real-world datasets from a mid-sized city—comprising one month of GPS bus traces, detailed timetable records, and passenger boarding/alighting profiles—the GA evolves candidate route sets to minimize a composite fitness function of average in-vehicle travel time, passenger waiting time, and fleet operational cost. Through iterative selection, crossover, and mutation over 200 generations, the algorithm identifies route assignments that balance efficiency and service quality.

Post-optimization, each GA-derived configuration undergoes extensive validation via 500 Monte Carlo runs in MATSim, simulating 100,000 agents during a four-hour morning peak. Comparative analysis against baseline schedules employs paired-sample t-tests to ascertain statistical significance. Results indicate a 16.8% reduction in average trip duration (t = 6.12, p < .001), a 33.6% drop in waiting times (t = 5.47, p < .001), and an 18.3% decrease in cost per kilometer (t = 7.35, p < .001). Beyond these aggregate gains, simulation outputs reveal improved load balancing—reducing peak-vehicle overcrowding by 18%—and enhanced on-time performance, with 92% of trips meeting headway targets versus 75% under the original network.

Key contributions include (1) a scalable, multi-objective GA design tailored for urban bus networks; (2) a robust integration methodology linking optimization outputs to agent-based traffic simulation for realistic validation; and (3) empirical evidence of substantial operational savings and passenger experience improvements. The framework’s adaptability to real-time data streams suggests potential for dynamic re-optimization under fluctuating demand or disruptions. Limitations involve the current focus on morning peak periods and computational overhead—approximately two hours per optimization cycle on standard hardware. Future research will explore reinforcement learning hybrids for faster convergence, incorporation of emission and equity objectives, and deployment in diverse urban contexts. By demonstrating a practical pathway from AI optimization to validated simulation insights, this work offers transit planners an evidence-based decision-support tool to enhance service reliability, efficiency, and sustainability.