Resource Substitution Problem Overview

• Resource Substitution Problem (RSP) is a framework for optimal reassignments in directed networks that minimizes resource imbalances using mixed-integer programming and ML-driven candidate pruning.
• It employs fairness metrics including minimax and Gini-based approaches to balance efficiency with equitable distribution of substitution burdens.
• The integrated OR–ML method achieves significant model-size and computation-time reductions in large-scale logistics networks while preserving optimality and fairness.

The Resource Substitution Problem (RSP) concerns determining optimal reassignments among a set of interchangeable and potentially composite resource types, so as to minimize persistent imbalances induced by asymmetrical flows in large-scale, directed networks. Originating in domains such as logistics, the RSP is to find a new assignment—subject to task requirements, resource compatibility, and operational fairness—that minimizes the number of resource substitutions needed to balance the overall network. Practical settings commonly include cost-sensitive substitution and decentralized scheduling, further motivating solutions that integrate formal optimization (operations research, OR) and data-driven guidance (machine learning, ML) (Mohan et al., 23 Nov 2025).

1. Formal Problem Specification

The RSP is defined on a directed graph $\mathcal{G}=(\mathcal{N},\mathcal{A})$ where $\mathcal{N}$ is the set of nodes and $\mathcal{A}$ the set of arcs, each representing a required task assigned to one resource from a predefined set $\mathcal{R}$ (resources may represent equipment, vehicles, or composite arrangements). Each arc $a$ admits a subset $\mathcal{R}_a\subseteq\mathcal{R}$ of feasible resources, e.g., two short trailers may substitute for a long trailer if both endpoints permit. $\Phi_0:\mathcal{A}\rightarrow\mathcal{R}$ encodes the initial resource assignment.

For any assignment $\Phi$, the node-level imbalance of resource $r$ at node $n$ is defined as:

$\mathcal{N}$0

with $\mathcal{N}$1 and $\mathcal{N}$2 the sets of arcs into/out of $\mathcal{N}$3, respectively. The total network imbalance is:

$\mathcal{N}$4

Binary variables $\mathcal{N}$5 indicate assignment $\mathcal{N}$6 iff resource $\mathcal{N}$7 is selected for arc $\mathcal{N}$8. Additional variables—$\mathcal{N}$9 (linearization of imbalance), $\mathcal{A}$0 (max per-scheduler substitutions), and $\mathcal{A}$1 (pairwise difference for Gini-based fairness)—are used in Stage 2 modeling.

2. Mixed-Integer Programming Formulations

The RSP solution process is structured in two mathematical programming stages:

Stage 1: Minimize Network Imbalance

Formulation:

Minimize network imbalance:

$\mathcal{A}$2

subject to:

$\mathcal{A}$3

$\mathcal{A}$4

$\mathcal{A}$5

Stage 2: Minimize Substitutions Under Fairness

Let $\mathcal{A}$6 (total substitutions), $\mathcal{A}$7 the set of schedulers, $\mathcal{A}$8 arcs owned by scheduler $\mathcal{A}$9, and $\mathcal{R}$0 scheduled substitutions.

• Pure efficiency: $\mathcal{R}$1 s.t. $\mathcal{R}$2 (Stage 1 constraints)
• Minimax fairness: $\mathcal{R}$3 s.t. $\mathcal{R}$4 $\mathcal{R}$5 and $\mathcal{R}$6
• Weighted trade-off: $\mathcal{R}$7 $\mathcal{R}$8
• Gini-based fairness: $\mathcal{R}$9 (with $a$0 for all $a$1)

These formulations enable multiple fairness paradigms: minimizing peak burden (minimax), balancing overall changes (weighted), or reducing inequality (Gini).

3. Fairness Considerations in Resource Substitution

Multiple fairness metrics operationalize diverse organizational policies and scheduler equity:

• Minimax fairness: minimizes maximum substitutions incurred by any single scheduler, i.e., $a$2.
• Weighted objectives: convex combinations, $a$3, interpolate between total-change and peak-burden criteria.
• Gini-style fairness: minimizes total pairwise difference $a$4, reflecting a Gini index–like inequality measure.

Weighted trade-off and Gini approaches support generating Pareto fronts to facilitate scheduler review and selection of context-specific trade-offs between efficiency and equity.

4. ML-Enhanced Candidate Pruning and Decision Support

To address scale and complexity, ML augments the OR formulations in two principal ways:

Learning Arc-Resource Preference Scores

A deep neural network (DNN) is trained on historical labeled (arc, resource) assignments $a$5 to estimate $a$6. Feature vectors $a$7 include node IDs, volume, scheduler identities, temporal indicators, distance, equipment size, and other operational variables. The DNN outputs a softmax vector $a$8 trained via cross-entropy.

Dynamic Top-$a$9 Resource Selection

For each arc $\mathcal{R}_a\subseteq\mathcal{R}$0, betweenness centrality $\mathcal{R}_a\subseteq\mathcal{R}$1 partitions arcs into Low, Medium, or High centrality. Thresholds $\mathcal{R}_a\subseteq\mathcal{R}$2 determine $\mathcal{R}_a\subseteq\mathcal{R}$3. For each arc, only the top $\mathcal{R}_a\subseteq\mathcal{R}$4 resources (ranked by $\mathcal{R}_a\subseteq\mathcal{R}$5) are retained; other assignment variables are fixed to zero. This procedure sharply reduces problem dimensionality: $\Phi$3 This dynamic pruning executes prior to Stage 1 optimization, resulting in an arc-resource candidate set tailored for both operational centrality and empirical preference.

5. Integrated OR–ML Workflow and Computational Acceleration

The solution architecture integrates ML-driven candidate reduction with two-stage mixed-integer programming, as follows:

1. Offline DNN training on historical $\mathcal{R}_a\subseteq\mathcal{R}$6 data.
2. For new instances: predict $\mathcal{R}_a\subseteq\mathcal{R}$7, compute $\mathcal{R}_a\subseteq\mathcal{R}$8, and dynamically select reduced resource candidate sets $\mathcal{R}_a\subseteq\mathcal{R}$9 for each arc.
3. Solve Stage 1 MIP on the pruned assignment model for the minimum imbalance $\Phi_0:\mathcal{A}\rightarrow\mathcal{R}$0.
4. Solve Stage 2 for diverse fairness objectives, yielding a portfolio of solutions using the same reduced model.
5. Output the Pareto frontier of efficiency–fairness trade-offs for scheduler selection.

Additional acceleration strategies include warm-starting via iterated local search around the initial assignment $\Phi_0:\mathcal{A}\rightarrow\mathcal{R}$1, Dantzig–Wolfe decomposition or column generation for $\Phi_0:\mathcal{A}\rightarrow\mathcal{R}$2, and targeted MIP solver tuning (e.g., cut aggressiveness, presolve configuration).

6. Empirical Performance in Large-Scale Logistics Networks

Applied to a package delivery network (2–40 schedulers, up to 1,588 nodes, up to $\Phi_0:\mathcal{A}\rightarrow\mathcal{R}$3 arcs, 13–23 equipment types), the hybrid OR–ML framework yields:

• Model-size reduction: 80–93% via dynamic top-$\Phi_0:\mathcal{A}\rightarrow\mathcal{R}$4 pruning.
• Execution-time reduction: 90–99% (MIP times cut from hundreds to 1.0–26.5 s in Stage 1).
• Imbalance reduction: Final imbalance $\Phi_0:\mathcal{A}\rightarrow\mathcal{R}$5 reduces 16–22%.
• Optimality preservation: Reduced models retain global optima for $\Phi_0:\mathcal{A}\rightarrow\mathcal{R}$6 or with dynamic pruning.
• Fairness trade-offs: As fairness weight $\Phi_0:\mathcal{A}\rightarrow\mathcal{R}$7 (in weighted trade-offs) increases $\Phi_0:\mathcal{A}\rightarrow\mathcal{R}$8, total substitutions increase up to 89% in the largest instances, but maximum per-scheduler burden drops by 17–18%.
• Portfolio insights: Weighted objectives generate a Pareto front for schedulers, and partially implemented fair solutions (applying only top-$\Phi_0:\mathcal{A}\rightarrow\mathcal{R}$9 changes) improve over truncations of pure-efficiency plans. Increased collaboration in arc ownership further reduces both imbalance and fairness cost.

For a representative case (50 schedulers, 8,000 nodes, 24 million arcs): static $\Phi$0 yields a 9.7% optimality gap in 133 s; $\Phi$1 achieves optimality in 2,376 s; dynamic $\Phi$2 averaging 2.2 attains the optimum in 132 s with 86% arc reduction. These results demonstrate preservation of optimality and substantial computational gains using the ML-augmented, candidate-pruned approach.

7. Summary of Benefits and Implications

The hybrid OR–ML approach for the Resource Substitution Problem addressed in (Mohan et al., 23 Nov 2025) demonstrates significant scalability and fairness advances:

• Up to 93% reduction in model size.
• Up to 99% reduction in solution times.
• 16–22% reduction in network imbalance.
• Up to 18% reduction in maximum per-scheduler burden for fairer solutions.

A plausible implication is the feasibility of deploying near-optimal, fairness-aware substitution plans in decentralized, real-world logistics contexts at scales unattainable by classical methods. Further, portfolio-based Pareto-efficient trade-offs facilitate more informed, preference-aligned choices by human schedulers within large organizations.