Towards Improved Robustness of Public Transport by a Machine-Learned Oracle
The design and optimization of public transport systems is a highly complex and challenging process. Here, we focus on the trade-off between two criteria which shall make the transport system attractive for passengers: their travel time and the robustness of the system. The latter is time-consuming to evaluate. A passenger-based evaluation of robustness requires a performance simulation with respect to a large number of possible delay scenarios, making this step computationally very expensive. For optimizing the robustness, we hence apply a machine-learned oracle from previous work which approximates the robustness of a public transport system. We apply this oracle to bi-criteria optimization of integrated public transport planning (timetabling and vehicle scheduling) in two ways: First, w e explore a local search based framework studying several variants of neighborhoods. Second, we evaluate a genetic algorithm. Computational experiments with artificial and close to real-word benchmark datasets yield promising results. In all cases, an existing pool of solutions (i.e., public transport plans) can be significantly improved by finding a number of new non-dominated solutions, providing better and different trade-offs between robustness and travel time.