A Reinforcement Learning–Guided Hybrid NSGA-II + ALNS Framework for Large-Scale Capacitated Transportation Optimization in Thai Sugarcane Logistics

Efficient planning of large-scale agricultural transportation requires balancing travel distance, fleet utilization, and factory capacity constraints. While mixed-integer linear programming (MILP) becomes computationally intractable for large-scale instances and conventional metaheuristics rely on static operator-selection mechanisms, adaptive learning-guided approaches for multi-objective capacitated transportation remain limited. This study proposes a reinforcement learning–guided hybrid NSGA-II + ALNS framework to minimize total transportation distance and truck trips in sugarcane logistics. A real-world case involving 199 subdistricts and four processing plants (796 origin–destination pairs) in northeastern Thailand is examined. Compared with a greedy nearest-assignment baseline, the proposed method reduces total transportation distance from 123,313.52 km to 109,245.22 km (by 11.41%), fuel consumption from 28,131.83 L to 25,108.96 L (by 10.75%), and CO₂ emissions from 75,955.94 kg to 67,794.20 kg (by 10.75%), resulting in an estimated fuel cost saving of approximately 96,550 Thai Baht per cycle. Statistical validation using ANOVA and Tukey’s HSD confirms that performance differences are significant at the 95% confidence level. The results demonstrate that reinforcement learning–guided operator adaptation improves convergence stability, Pareto-front quality, and environmental performance in large-scale bi-objective agricultural transportation systems.