Adaptive diversity for personalized multimodal transport optimization

The shift toward sustainable urban mobility requires transport systems that are efficient, environmentally friendly, and tailored to individual user needs. This paper presents a framework for optimizing multimodal transport routes by simultaneously addressing four conflicting objectives: travel time, monetary cost, carbon emissions, and user comfort. To capture the subjective nature of comfort, we develop a machine learning surrogate model based on a multilayer perceptron, trained on synthetic user preferences generated from explicit behavioral heuristics and integrated into the optimization process. We compare two leading multiobjective evolutionary algorithms, NSGA-II and NSGA-III, for solving this personalized multimodal transport problem. Although NSGA-III is theoretically better suited for many-objective problems, our experimental results reveal an unexpected finding: NSGA-II significantly outperforms NSGA-III in this domain. Through statistical analysis of normalized hypervolumes over 30 independent runs across 150 diverse user profiles, we show that NSGA-II achieves a large effect size (Cohen's $d = 1.37$) and consistently produces superior Pareto fronts. These results suggest that NSGA-II's crowding distance mechanism is more effective at maintaining diversity in the discontinuous and non-convex search space typical of multimodal transport networks than NSGA-III's reference point based approach. This study provides useful insights that may contribute to the design of future mobility as a service platforms.