Extensions to the guaranteed service model for industrial applications of multi-echelon inventory optimization

Victoria G. Achkar, Braulio Brunaud, Hector D. Perez, Rami Musa, Carlos A. Méndez, Ignacio E. Grossmann

TL;DR

This paper introduces a generalized "Guaranteed Service Model" for multi-echelon inventory optimization that accounts for complex real-world constraints like hybrid demand nodes and minimum order quantities. By reformulating the optimization problem into a Quadratically Constrained Program (QCP), the authors achieved massive computational speedups—solving in seconds what standard solvers couldn't handle in hours—while reducing inventory costs by up to 17%.

What Problem Does This Paper Address?

Managing inventory across a complex supply chain (multi-echelon) is a balancing act: you want enough stock to keep customers happy (high service levels) but not so much that you tie up millions in "dead" capital.

The authors identify a major gap between academic models and industrial reality, particularly in the pharmaceutical sector. Most existing models assume demand only happens at the very end of the chain. In reality, large distribution hubs often act as "hybrid nodes," serving both external customers and downstream internal warehouses. Furthermore, standard models often ignore "Minimum Order Quantities" (MOQs) and manufacturing lead times within the network. This paper aims to create a model that captures these "messy" industrial features without becoming too complex to solve.

Methods

The researchers built upon the Guaranteed Service Model (GSM), a classic framework for safety stock placement. Their "extended" version introduces several novel components:

  • Hybrid Node Modeling: They explicitly account for nodes that face both independent (external) and dependent (internal) demand.
  • Fill Rate Approximation: While most models use "Cycle Service Level" (the probability of not stocking out), industry prefers "Fill Rate" (the % of total demand met). The authors used polynomial regression to create a surrogate model that translates fill rate targets into safety stock factors.
  • MOQ Integration: They modeled how large batch sizes (MOQs) affect inventory. Interestingly, while large MOQs increase "cycle stock," they can actually reduce the required safety stock because the large shipments act as a buffer.
  • QCP Reformulation: This is the technical "secret sauce." The original problem is a non-convex Nonlinear Program (NLP), which is notoriously hard to solve. The authors proved that by defining new variables to replace square root terms, the problem can be transformed into a Quadratically Constrained Program (QCP). This allows the use of high-performance solvers like Gurobi or CPLEX.

Datasets

The model was evaluated using three distinct setups:

  1. Illustrative Example: A small, controlled network used for sensitivity analysis (e.g., how changing lead times affects stock placement).
  2. Small-size Industrial Case: A real-world scenario involving 4 SKUs and 31 raw materials.
  3. Medium-size Industrial Case: A larger real-world network with 20 finished goods, 120 raw materials, and 196 distinct safety stock decisions.
  4. Simulation Validation: All results were validated using InventoryManagement.jl, an open-source discrete-time simulation package, to ensure the theoretical stock levels actually met service targets in a "live" environment.

Key Insights and Ideas

  • The "MOQ Buffer" Effect: One of the most interesting findings is that a large MOQ can eliminate the need for safety stock entirely for certain products. If you are forced to order 10 weeks of supply at once, that massive "overshoot" often covers demand variability better than a dedicated safety stock pile would.
  • Mathematical Structure is Everything: The jump from NLP to QCP wasn't just a minor tweak; it was the difference between a model that is a theoretical curiosity and one that can be used in a daily business environment.
  • Risk Pooling Works: The model confirms that holding stock further "upstream" (at the plant) is often cheaper than "downstream" (at retailers) because the plant can "pool" the uncertainty of all its retailers together.

Key Results

  • Massive Speedups: For the medium-sized industrial case, the standard NLP solver (BARON) failed to find a feasible solution within 1000 seconds. The reformulated QCP model solved it to optimality in just 3 seconds.
  • Superior Cost Savings: In a direct comparison with a leading commercial MEIO software, the authors' model achieved a 17% reduction in holding costs compared to current levels, while the commercial tool only managed a 10% reduction.
  • High Accuracy: Simulation results showed that the predicted safety stocks met the target service levels with a maximum error of only 0.02 points, proving the polynomial regression for fill rates was highly reliable.

Limitations and Future Work

The authors acknowledge a few areas for improvement:

  • Normal Distribution Assumption: The model assumes demand follows a normal distribution. In cases of highly "lumpy" or intermittent demand (common for rare medicines), this assumption might break down.
  • Capacity Constraints: The current model assumes warehouses and plants have infinite capacity. In the real world, storage space is a hard limit.
  • Future Directions: The team plans to extend the model to handle non-normal demand and integrate it with tactical planning tools like "rhythm wheels" (production sequencing) for a truly end-to-end supply chain optimization.

Summary updated

Abstract

Highlights•Multi-echelon inventory optimization based on guaranteed service.•Model integrates existing and novel features present in real applications.•Quadratically Constrained formulation yields order magnitude computational savings.•Model tested on real-world application and validated with simulation.
BibTeX
@article{achkar2024extensions,
  author = {Victoria G. Achkar and Braulio Brunaud and Hector D. Perez and Rami Musa and Carlos A. Méndez and Ignacio E. Grossmann},
  title = {{Extensions to the guaranteed service model for industrial applications of multi-echelon inventory optimization}},
  journal = {Eur. J. Oper. Res.},
  year = {2024},
  doi = {10.1016/j.ejor.2023.08.013},
}
Figure