Performance Comparison Among Classical Metaheuristic Algorithms for Stochastic Last-Mile Delivery Routing Problem

Last-mile delivery routing requires simultaneous control of distance, travel time, service deadlines, vehicle capacity, workload balance, operating cost, emissions, and uncertainty. This study provides a controlled low-budget comparison of ten classical metaheuristics for a LaDe-calibrated stochastic CVRPTW. The GA, DE, PSO, ACO, ABC, SA, GWO, WOA, TLBO, and JAYA used the same random-key representation, capacity-aware split decoder, repair rules, CRN scenarios, and exactly 50 objective evaluations per run. The experiment comprised 1065 fixed design conditions, 30 seeded runs per algorithm-condition cell, and 319,500 metaheuristic runs, with NNS, Clarke–Wright Savings, and OR-Tools Guided Local Search as benchmarks. PSO achieved the lowest grand-mean weighted logistics cost, whereas the WOA achieved the lowest median and best Friedman mean rank. The grand-mean difference between the WOA and PSO was only 125.2 objective units, and the context winner changed across the three cities and five customer-size levels (PSO: 4, WOA: 2, and JAYA: 2). Accordingly, corrected p-values are treated as conditional design diagnostics, while practical interpretation prioritises paired magnitude, stratified consistency, feasibility, and benchmark proximity. A design-stratified sensitivity analysis showed that the preferred method changed across cities and customer sizes. Crucially, every metaheuristic and benchmark recorded 0% full feasibility under the common evaluator; capacity- and lateness-violation profiles are therefore elevated to primary outcomes, and no algorithm is claimed to be operationally superior. The 50-evaluation protocol is interpreted as an early-budget screening regime rather than evidence of asymptotic convergence. The results show that algorithm choice is conditional on the tested data, weights, problem sizes, uncertainty settings, and evaluation budget and that feasibility-preserving decoding is more important for deployment than small differences in penalised objective value.