Raw Run Records and Frozen Analysis Artifacts for a Hybrid Matheuristic (AMO-LNES) on a Three-Objective Meal-Delivery Routing Problem with Time Windows

This dataset contains the complete raw output of 2,250 formal algorithm runs and 180 ablation runs for a benchmark study of AMO-LNES, a hybrid matheuristic combining NSGA-II-style evolution with Large Neighborhood Search (LNS), on a synthetic Solomon-style three-objective meal-delivery routing problem with time windows. The three maximized objectives are total demand fulfilled (Z1), time-window satisfaction rate (Z2), and routing efficiency (Z3).In the formal comparison, AMO-LNES-FROZEN-B20 is benchmarked against four baselines (NSGA-II, MOEA/D, ALNS-classic, MO-ALNS) across 3 instance families (R1 dispersed, C1 clustered, RC1 mixed) × 3 scales (n = 20, 60, 100) × 5 instances per cell (instance seeds 101–105) × 10 algorithm seeds, under shared wall-clock budgets of 30/120/300 s.

The ablation study isolates three AMO-LNES components (guided initialization, efficiency-aware operator weighting, LNS stage) in 180 runs at n = 60.Each run record stores the full Pareto archive with explicit routes, a convergence history, and SHA-256 hashes binding the run to frozen code and configurations. Hypervolume and IGD+ are computed on budget-capped archives with per-instance pooled-union-cloud normalization; statistical analysis uses Friedman tests with Holm-corrected Wilcoxon signed-rank post-hoc comparisons over 45 blocks.