💻 computer science

Viable Supply Chain Network Design: Machine Learning-Derived Chance-Constrained Programming

This paper proposes a machine learning-enhanced chance-constrained programming approach to design a viable, resilient, and environmentally sustainable two-echelon supply chain network that accounts for cross-echelon disruptions, demonstrating that the resulting implicit formulation and associated heuristics provide computationally efficient, high-quality solutions for large-scale instances compared to traditional scenario-based methods.

Original authors: Mohammad Rohaninejad, Behdin Vahedi-Nouri, Elham Jelodari Mamaghani, Mehdi Foumani, Olga Battaia

Published 2026-06-10

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Mohammad Rohaninejad, Behdin Vahedi-Nouri, Elham Jelodari Mamaghani, Mehdi Foumani, Olga Battaia

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are the manager of a massive, two-story delivery system. On the bottom floor, you have local shops (the first echelon) that serve customers. On the top floor, you have big warehouses (the second echelon) that supply the shops. Your goal is to decide where to build these shops and warehouses, how to connect them, and how to keep everything running smoothly even when things go wrong.

This paper tackles a very specific problem: How do you design this network so it doesn't just survive a disaster, but actually thrives? The authors call this "viability." They argue that being "resilient" (having backups) isn't enough. You also need "agility" (the ability to move fast) and "sustainability" (keeping pollution low).

Here is a breakdown of their solution using simple analogies:

1. The Three Pillars of a "Viable" Network

The authors say a good network needs three superpowers:

Resilience (The Backup Plan): If a shop closes because of a power outage, customers shouldn't go hungry. They need a backup shop to switch to.
Agility (The Emergency Squad): Sometimes, a backup shop is too far away. In those cases, the network needs "mobile facilities"—think of them as emergency food trucks that can be parked right where they are needed, instantly.
Sustainability (The Green Cap): The network must not pollute too much. The authors put a hard "speed limit" on the total amount of carbon emissions the trucks can produce.

2. The Problem: Too Many "What-Ifs"

Designing this network is like trying to solve a puzzle where the pieces keep changing shape.

The Scenario-Based Approach (The "List" Method): One way to solve this is to write down every single possible disaster (e.g., "Shop A fails," "Shop B fails," "Both fail," "Shop A fails and it rains"). Then, you calculate the cost for every single list.
- The Catch: If you have 100 shops, the number of possible disaster lists is so huge (like the number of grains of sand on a beach) that even the fastest supercomputers get stuck trying to read the list. It takes too long.
The Implicit Approach (The "Rule" Method): The authors tried a smarter way. Instead of listing every disaster, they built a set of rules (a "backup hierarchy"). "If Shop A fails, go to Shop B. If Shop B fails, go to Shop C." This is much faster to calculate.

3. The Secret Weapon: Machine Learning as a "Crystal Ball"

The tricky part is the mobile facilities (the emergency trucks). You can't just say, "We will have 5 trucks." You have to ask, "What is the chance that we will need more than 5 trucks at the exact same time?"

Calculating this probability is mathematically impossible to do perfectly in a reasonable time. It's like trying to predict exactly how many people will show up to a party on a random Tuesday.

The Solution: The authors used Machine Learning to act as a "crystal ball."

They ran thousands of simulations to see what happens when different numbers of shops fail.
They taught a computer (using algorithms like Logistic Regression) to recognize patterns: "Oh, when 3 shops fail in this specific pattern, we always need more than 5 trucks. That's a bad pattern."
They turned this learning into a simple math rule (a "linear cut").
Now, instead of doing complex probability math every time, the computer just checks the rule: "Does this plan look like the 'bad patterns' the computer learned? If yes, reject it. If no, accept it."

This allowed them to solve the problem 95% of the time (the confidence level they wanted) without getting stuck in the math.

4. The "Fix-and-Relax" Strategy

Even with the machine learning trick, the problem is still huge for big cities with hundreds of shops. Solving it all at once is like trying to eat a whole elephant in one bite.

The authors developed a "Fix-and-Relax" strategy:

Step 1: Solve the problem for just the first few shops (the "first bite").
Step 2: Once you decide where those shops go, "freeze" that decision.
Step 3: Relax the rules for the rest of the network (pretend the other shops are flexible) and solve for the next batch.
Step 4: Repeat until the whole network is designed.

They tested two versions of this: one that builds the network layer by layer (Echelon-based) and one that builds it backup-level by backup-level (Backup-based). Both were incredibly fast and found high-quality solutions.

5. The Big Discovery: Work Together, Don't Work Alone

The paper compared two ways of planning:

Hierarchical: The warehouse boss decides where the warehouses go, locks that decision, and then tells the shop boss, "Okay, now you decide where the shops go."
Integrated: Everyone sits in the same room and decides the whole network at once.

The Result: The Integrated approach was significantly better (saving about 12% in costs on average). It proved that if the different levels of the supply chain don't talk to each other and plan together, the whole system becomes inefficient and more expensive.

Summary

The paper presents a new way to design supply chains that are tough, fast, and green. They combined mathematical optimization with machine learning to handle the uncertainty of disasters. They found that by using smart "rules" instead of endless lists, and by having all parts of the network plan together rather than separately, companies can build supply chains that survive disruptions without breaking the bank or the planet.

Technical Summary: Viable Supply Chain Network Design: Machine Learning-Derived Chance-Constrained Programming

Problem Definition
This study addresses the design of a Viable Two-Echelon Supply Chain Network Design Problem (VTE-SCNDP) under conditions of facility disruption. Unlike traditional models that treat supply chain echelons in isolation or focus solely on resilience, this research integrates three core pillars of viability:

Resilience: Achieved through backup reassignment mechanisms where clients are redirected to alternative facilities if their primary assignment fails.
Agility: Modeled via the deployment of mobile facilities that can be rapidly established to serve affected clients when regular facilities are unavailable or when response time thresholds are exceeded.
Sustainability: Enforced through a hard constraint on total expected transportation emissions.

The network consists of two echelons where the first echelon serves as both a service provider to end-users and a client to the second echelon. Disruptions are assumed to occur independently at both levels. The objective is to minimize the expected total cost, comprising fixed establishment costs, transportation costs, and penalties for unmet demand, while ensuring demand satisfaction across both echelons under uncertainty.

Methodology
The paper proposes two distinct Mixed-Integer Programming (MIP) formulations and a suite of solution algorithms:

Scenario-Based Formulation (SBF):
- Explicitly enumerates all possible disruption scenarios.
- Minimizes expected costs across these scenarios.
- While accurate, it suffers from exponential growth in problem size, rendering it computationally intractable for large-scale instances.
Implicit Formulation (IF):
- Avoids explicit scenario enumeration by modeling backup relationships. Clients are pre-assigned to sequences of backup facilities based on priority and proximity.
- Machine Learning Integration: A key innovation is the handling of the chance-constrained programming requirement for mobile facility capacity. The constraint that the probability of exceeding mobile facility capacity must be less than 5% (i.e., 95% service confidence) is computationally intractable due to the lack of a closed-form Cumulative Distribution Function (CDF) for the sum of independent Bernoulli failures.
- Solution: The authors replace this intractable constraint with learned linear cuts. They generate a dataset of client allocation patterns, label them based on whether they satisfy the 95% capacity threshold, and train linear classifiers (Logistic Regression, L1-regularized LR, SGD, Perceptron, and LR with $C=0.1$ ). The best-performing classifier is embedded into the MIP as a linear constraint to enforce the service level.
Solution Algorithms:
- Fix-and-Relax Heuristics: Two heuristics are developed to solve the IF for large-scale problems:
  - Backup-Based Fix-and-Relax (BBFR): Decomposes the problem by backup levels ( $R_1 + R_2$ iterations), optimizing one level at a time while fixing previous levels and relaxing future ones.
  - Echelon-Based Fix-and-Relax (EBFR): Decomposes the problem by echelon, solving the first echelon fully before fixing its decisions to solve the second echelon.
- Sample Average Approximation (SAA): Applied to the SBF to approximate the expected value for medium-sized instances by sampling a subset of scenarios.
- Post-Optimality Adjustment: A correction algorithm is applied to the IF solutions to account for the slight underestimation of costs caused by the probabilistic nature of the chance constraints (violations in the <5% tail are penalized post-hoc).

Key Contributions

Conceptual Integration: The paper operationalizes the concept of "viability" by simultaneously modeling resilience, agility (via mobile facilities), and environmental sustainability within a multi-echelon network design framework.
Cross-Echelon Modeling: Unlike prior studies restricted to single echelons or single-echelon disruptions, this model captures cross-echelon disruptions, acknowledging that failures in one echelon cascade to affect the other.
Machine Learning for Optimization: The study pioneers the use of learned linear cuts derived from classification algorithms to replace intractable chance constraints in supply chain network design, bridging the gap between probabilistic service requirements and computational tractability.
Algorithmic Development: The introduction of tailored Fix-and-Relax heuristics specifically for viable multi-echelon problems and the application of SAA to viable scenarios.

Results
Computational experiments were conducted on 12 randomly generated datasets ranging from small to large scales (up to 375 nodes).

Formulation Comparison: The Implicit Formulation (IF) consistently outperformed the Scenario-Based Formulation (SBF) in terms of computational time and scalability. While SBF failed to solve instances larger than 80 nodes within the time limit, IF solved instances up to 300 nodes. The IF solutions were found to be within 0.7% to 3.69% of the best-known solutions (after post-optimality correction).
Heuristic Performance: Both BBFR and EBFR heuristics proved highly effective. EBFR generally provided the highest quality solutions (often 0.00% RPD1) with reasonable computational times, outperforming SAA on large-scale instances.
Machine Learning Efficacy: The trained classifiers (particularly Logistic Regression with $C=0.1$ ) achieved high accuracy (up to 95.1%) in distinguishing feasible assignment patterns from infeasible ones, validating the substitution of the chance constraint with learned linear cuts.
Integrated vs. Hierarchical Optimization: A comparative analysis revealed that integrated optimization (simultaneous decision-making across echelons) yields significantly better results than hierarchical optimization (optimizing echelons sequentially). The integrated approach reduced costs by an average of 12.26% compared to the hierarchical approach.
Sensitivity: The proposed methods demonstrated robustness under varying failure probabilities, though performance degradation was observed in SAA as failure probabilities increased beyond 0.7.

Significance and Claims
The authors claim that this study represents the first application of fix-and-relax heuristics and machine learning-based constraint learning to viable multi-echelon supply chain network design under disruptions. The significance of the work lies in:

Scalability: Providing a computationally efficient alternative (IF + Heuristics) to exact scenario-based methods for large-scale, complex supply chains.
Practical Viability: Demonstrating that integrating resilience, agility, and sustainability is not only theoretically possible but computationally tractable, offering a pathway for long-term network survival.
Strategic Insight: Highlighting the critical importance of integrated decision-making across supply chain echelons. The results suggest that siloed optimization leads to suboptimal network designs, whereas coordinated planning significantly enhances system performance and cost-efficiency.
Methodological Advancement: Successfully narrowing the theory-practice gap by replacing complex probabilistic constraints with data-driven linear approximations, making advanced optimization models applicable to real-world supply chain challenges.