Toward sustainable and customer-centric reverse logistics: Machine learning-enhanced multi-objective optimization

The rise of e-commerce, coupled with increasing demands for sustainability and customer-oriented logistics, has intensified interest in reverse logistics and time-sensitive delivery planning. This study addresses a green and customer-focused variant of reverse logistics by formulating the Periodic Vehicle Routing Problem with Simultaneous Pickups and Deliveries and Time Windows (PVRPSPDTW). The objective is to jointly minimize logistics costs, CO 2 emissions, and time-window violations. We propose a mixed-integer linear programming model with soft time windows, where small instances are solved using a weighting method in CPLEX. For large-scale problems, we employ a hybrid heuristic (XGP-ALNS), which integrates XGBoost machine learning technique with Adaptive Large Neighborhood Search (ALNS) algorithm. In this framework, XGBoost guides ruin-and-repair operators and serves as a surrogate model for scalarized objectives through the weighting method, improving convergence and solution quality. The performance of XGP-ALNS is benchmarked against Non-Dominated Sorting Genetic Algorithm 2 (NSGA-II) and multi-objective memetic algorithm (MOMA) using five multi-objective evaluation metrics. The results reveal important trade-offs: while stricter emissions limits reduce CO 2 , they also lead to higher costs and more service violations. In urban networks, CO 2 -focused optimization cuts emissions by 23% but increases costs by 29% and time window violations by 16%. Rural networks achieve larger reductions ( ≈ 29%) with minimal cost increases ( ≈ 1–1.5%), although service violations continue to increase. Limiting time window deviations by 20% and 40% increases emissions by 13% and 27%, respectively. Deploying electric vehicles eliminates direct emissions but introduces moderate cost and service trade-offs. In general, the findings underscore the need to tailor sustainability strategies to specific network contexts rather than applying uniform policies. The results highlight that stricter emissions limits lower CO 2 emissions but raise costs and service violations.