Abstract
In large-scale logistics, maintaining equipment balance across a network is a critical yet computationally expensive challenge. Traditional optimization methods often struggle with the NP-hard nature of resource substitution and the practical necessity of fairness in decentralized operations. This paper presents FAIR-SUB, a hybrid framework that integrates Machine Learning (ML) for search space pruning with Operations Research (OR) for fair resource allocation. Evaluated on real-world data from a global delivery firm, FAIR-SUB reduces computation time by 90% and model size by 80%, while significantly improving the equitable distribution of operational burdens among schedulers.
1. Introduction
Large-scale logistics networks, such as those for national trucking fleets or global package delivery, frequently face demand imbalances. These imbalances occur when equipment (e.g., trailers) accumulates at certain nodes while others experience deficits. The conventional remedy, "deadheading" or moving empty equipment, is both costly and environmentally detrimental, with heavy-duty trucks contributing significantly to transportation CO2 emissions.
Resource substitution—using compatible alternative equipment to meet demand—offers a more efficient solution. However, implementing this at scale is computationally intensive. Furthermore, centralized optimal plans often fail in practice if they impose an unfair burden on specific regional schedulers. FAIR-SUB addresses these challenges by balancing global efficiency with local fairness through a hybrid OR-ML approach.
2. Methodology
The FAIR-SUB framework operates through a synergistic two-phase process designed to handle massive datasets and complex constraints.
2.1 ML Component: Search Space Reduction
The first phase utilizes a Deep Neural Network (DNN) trained on 13 weeks of historical operational data to predict the likelihood of specific resource-arc assignments. To optimize efficiency, the authors implement a dynamic top-$\kappa$ mechanism. By calculating "betweenness centrality" for every arc, the model identifies high-traffic bottleneck arcs and allocates more candidate resources to them ($\kappa$ is higher), while pruning peripheral arcs more aggressively. This approach reduces the model size by over 80%.
2.2 OR Component: Fair Optimization
The second phase employs Mixed-Integer Programming (MIP) on the reduced search space. The optimization is conducted in two stages:
- Stage 1: Minimizing total network imbalance.
- Stage 2: Minimizing the number of substitutions while incorporating fairness metrics.
The framework utilizes a Minimax objective to minimize the maximum burden on any single scheduler and a Gini-inspired model to ensure equitable work distribution.
3. Results and Discussion
The framework was tested against eight classes of instances, ranging from small networks to massive systems with over 1,500 nodes and 8.3 million arcs.
3.1 Performance and Scalability
FAIR-SUB achieved a 90% reduction in execution time compared to standard optimization methods. The ML-driven pruning successfully captured the true optimal labels in the top-5 predictions 100% of the time, demonstrating that significant model reduction does not compromise solution quality.
3.2 The Price and Benefit of Fairness
A key finding is that the "price of fairness" is remarkably low; substantial reductions in the burden on overworked schedulers were achieved with only marginal increases in total substitutions. In one case study, the burden on the most-impacted scheduler was reduced by 42% without sacrificing global solution quality. Moreover, "Fair" solutions proved more robust during partial implementation, as they distribute high-value changes across the network rather than clustering them, making them more likely to be adopted by decentralized schedulers.
4. Conclusion and Future Work
FAIR-SUB demonstrates the power of combining structural network properties with predictive modeling to solve large-scale logistics problems. While the current model is deterministic, future research will focus on incorporating stochasticity to handle demand uncertainty and transitioning to online learning for real-time adaptation to network disruptions. Additionally, future iterations could incorporate more granular cost structures, such as arc-dependent fuel costs for different equipment types.
Summary updated