Answers are grounded in this paper and cite their sources.
Transportation planning with heterogeneous vehicle capacities, integer-trip requirements, limited fleet availability, and restricted routes creates a mixed-integer decision space that grows rapidly with network size. Although a full mixed-integer linear programming (MILP) model can produce feasible optimal plans, many source–destination–vehicle variables remain inactive while still enlarging the search space. This study develops an artificial neural network (ANN)-assisted framework that learns to identify promising route–vehicle combinations before optimization. A full integer-trip MILP generates reference solutions for controlled scenarios derived from the study network. Route-level operational and cost features are then used to train a multilayer perceptron that estimates the probability that each source–destination–vehicle combination will be active. The resulting probabilities are combined with deterministic coverage and low-cost safety rules to form a reduced MILP, while an iterative recovery mechanism restores excluded variables when necessary. The framework was evaluated on 1200 normal scenarios, 60 stress scenarios, five independent training seeds, and networks of increasing size. The principal ANN achieved 98.10% active-route recall, with a mean of 98.86% across the five seeds. On 180 held-out instances, the proposed method reduced the initial variable set by 41.57% while maintaining a mean cost gap of 0.0185% relative to the full MILP. All normal test instances were feasible without recovery, and complete feasibility was restored under stress. The results show that feasibility-preserving neural screening can reduce the transportation search space without transferring final decision authority from the optimization model.
No takes yet. Share an insight, caveat, or question.
Alshabibi et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: