Optimizing Large-Scale Sugarcane Transportation under Capacity and Queue Constraints Using a Hybrid Evolutionary Framework

Abstract

Large-scale sugarcane transportation requires balancing transportation efficiency with congestion at processing facilities under capacity constraints. This study proposes a queue-aware bi-objective optimization framework integrating Non-dominated Sorting Genetic Algorithm II (NSGA-II), Adaptive Large Neighborhood Search (ALNS), and reinforcement learning (RL) for state-dependent search-mode control. The objectives are to minimize total transportation distance and average queue waiting time. The framework was evaluated using a real-world case comprising 199 subdistricts and four sugar mills in northeastern Thailand. Exhaustive enumeration on a reduced instance was first used to validate the ability of the proposed method to approximate a known reference Pareto front. Large-scale experiments compared the RL-guided framework with standalone NSGA-II, Random ALNS, tuned MOEA/D, and a greedy baseline. The RL-guided method achieved the best mean performance, with a transportation distance of 144,392.49 km and an average waiting time of 1.383 h/trip, together with the best mean HV and IGD values. Friedman and Holm-adjusted Wilcoxon tests confirmed significant improvements over NSGA-II for distance, HV, and IGD and over MOEA/D for all evaluated measures, while differences from Random ALNS were not statistically significant. Relative to the non-optimized baseline, the proposed method reduced transportation distance by 9.93% and fuel consumption and CO₂ emissions by 9.05%. The results demonstrate the potential of queue-aware RL-guided optimization to support efficient and sustainable large-scale agricultural transportation planning.