Javascript is required
1.
M. Breque, L. De Nul, and A. Petridis, “Industry 5.0: Towards a sustainable, human-centric and resilient European industry,” Luxembourg, 2021. [Online]. Available: https://op.europa.eu/en/publication-detail/-/publication/468a892a-5097-11eb-b59f-01aa75ed71a1/language-en [Google Scholar]
2.
J. Durán-Micco, S. Alaei, and C. Macharis, “Evaluating the impact of horizontal collaboration in logistics systems: A simulation-based study,” Transportmetrica B: Transp. Dyn., vol. 13, no. 1, p. 2475204, 2025. [Google Scholar] [Crossref]
3.
N. Mrabti, N. Hamani, and L. Delahoche, “A comprehensive literature review on sustainable horizontal collaboration,” Sustainability, vol. 14, no. 18, p. 11644, 2022. [Google Scholar] [Crossref]
4.
World Resources Institute and World Business Council for Sustainable Development, “Corporate value chain (Scope 3) accounting and reporting standard: Supplement to the GHG Protocol corporate accounting and reporting standard,” 2011. https://ghgprotocol.org/sites/default/files/standards/Corporate-Value-Chain-Accounting-Reporing-Standard_041613_2.pdf [Google Scholar]
5.
M. Gansterer and R. F. Hartl, “Collaborative vehicle routing: A survey,” Eur. J. Oper. Res., vol. 268, no. 1, pp. 1–12, 2018. [Google Scholar] [Crossref]
6.
M. Ghomi-Avili, R. Tavakkoli-Moghaddam, S. G. Jalali Naeini, and A. Jabbarzadeh, “Competitive green supply chain network design model considering inventory decisions under uncertainty: A real case of a filter company,” Int. J. Prod. Res., vol. 59, no. 14, pp. 4248–4267, 2021. [Google Scholar] [Crossref]
7.
T. Hacardiaux, C. Defryn, J.-S. Tancrez, and L. Verdonck, “Balancing partner preferences for logistics costs and carbon footprint in a horizontal cooperation,” OR Spectr., vol. 44, no. 1, pp. 121–153, 2021. [Google Scholar] [Crossref]
8.
H. Saffari, M. Abbasi, and J. Gheidar-Kheljani, “A robust, sustainable, resilient, and responsive model for forward/reverse logistics network design with a new approach based on horizontal collaboration,” Environ. Dev. Sustain., vol. 27, pp. 23439–23482, 2025. [Google Scholar] [Crossref]
9.
L. Safari, S. J. Sadjadi, and F. M. Sobhani, “Resilient and sustainable supply chain design and planning under supply disruption risk using a multi-objective scenario-based robust optimization model,” Environ. Dev. Sustain., vol. 26, pp. 27485–27527, 2024. [Google Scholar] [Crossref]
10.
A. A. Filho, I. Correia, and T. Melo, “Bi-objective stochastic supply chain network design under supplier and facility disruption risks,” Comput. Appl. Math., vol. 45, p. 309, 2026. [Google Scholar] [Crossref]
11.
F. Cruijssen, W. Dullaert, and H. Fleuren, “Horizontal cooperation in transport and logistics: A literature review,” Transp. J., vol. 46, no. 3, pp. 22–39, 2007. [Google Scholar] [Crossref]
12.
S. Pan, D. Trentesaux, E. Ballot, and G. Q. Huang, “Horizontal collaborative transport: Survey of solutions and practical implementation issues,” Int. J. Prod. Res., vol. 57, no. 15–16, pp. 5340–5361, 2019. [Google Scholar] [Crossref]
13.
L. Verdonck, A. Caris, K. Ramaekers, and G. K. Janssens, “Collaborative logistics from the perspective of road transportation companies,” Transp. Rev., vol. 33, no. 6, pp. 700–719, 2013. [Google Scholar] [Crossref]
14.
A. Serrano-Hernandez, A. A. Juan, J. Faulin, and E. Perez-Bernabeu, “Horizontal collaboration in freight transport: Concepts, benefits and environmental challenges,” SORT, vol. 41, no. 2, pp. 393–414, 2017. [Google Scholar] [Crossref]
15.
M. Çerçi, Z. B. Kanık, H. C. Uzgören Kazanç, and M. Soysal, “Transforming logistics with horizontal collaboration: A systematic review of trends, challenges, and sustainability perspectives,” Rev. Manag. Sci., pp. 1–63, 2026. [Google Scholar] [Crossref]
16.
F. Cruijssen, I. van Beest, and G. Kant, “A human behaviour perspective on horizontal collaboration to reduce the climate impact of logistics,” Sustainability, vol. 15, no. 23, p. 16221, 2023. [Google Scholar] [Crossref]
17.
T. Hacardiaux and J.-S. Tancrez, “Assessing the environmental benefits of horizontal cooperation using a location-inventory model,” Cent. Eur. J. Oper. Res., vol. 28, no. 4, pp. 1363–1387, 2020. [Google Scholar] [Crossref]
18.
A. Aloui, N. Hamani, and L. Delahoche, “An integrated optimization approach using a collaborative strategy for sustainable cities freight transportation: A case study,” Sustain. Cities Soc., vol. 75, p. 103331, 2021. [Google Scholar] [Crossref]
19.
J. R. Birge and F. Louveaux, Introduction to Stochastic Programming, 2nd. New York, NY, USA: Springer, 2011. [Online]. Available: [Google Scholar] [Crossref]
20.
W. Klibi, A. Martel, and A. Guitouni, “The design of robust value-creating supply chain networks: A critical review,” Eur. J. Oper. Res., vol. 203, no. 2, pp. 283–293, 2010. [Google Scholar] [Crossref]
21.
R. B. Nelsen, An Introduction to Copulas, 2nd. New York, NY, USA: Springer, 2006. [Online]. Available: [Google Scholar] [Crossref]
22.
H. Heitsch and W. Römisch, “Scenario reduction algorithms in stochastic programming,” Comput. Optim. Appl., vol. 24, no. 2–3, pp. 187–206, 2003. [Google Scholar] [Crossref]
23.
S. Abbasi and Y. Kazancoglu, “Financial risk-aware closed-loop supply chain network design for the circular economy: A two-stage stochastic optimization model,” Results Eng., vol. 32, p. 112395, 2026. [Google Scholar] [Crossref]
24.
K. Mili and M. Argoubi, “Stochastic optimization of perishable agricultural supply chains: A hybrid genetic algorithm approach for robust network design under multi-dimensional uncertainty,” Front. Sustain. Food Syst., vol. 10, p. 1807797, 2026. [Google Scholar] [Crossref]
25.
G. Mavrotas and K. Florios, “An improved version of the augmented epsilon-constraint method (AUGMECON2) for finding the exact Pareto set in multi-objective integer programming problems,” Appl. Math. Comput., vol. 219, no. 18, pp. 9652–9669, 2013. [Google Scholar] [Crossref]
26.
G. Mavrotas, “Effective implementation of the ϵ-constraint method in multi-objective mathematical programming problems,” Appl. Math. Comput., vol. 213, no. 2, pp. 455–465, 2009. [Google Scholar] [Crossref]
27.
E.-G. Talbi, “Combining metaheuristics with mathematical programming, constraint programming and machine learning,” 4OR, vol. 11, no. 2, pp. 101–150, 2013. [Google Scholar] [Crossref]
28.
E. Winter, “The Shapley value,” in Handbook of Game Theory with Economic Applications, The Netherlands: Elsevier, 2002, pp. 2025–2054. [Google Scholar] [Crossref]
29.
D. Schmeidler, “The nucleolus of a characteristic function game,” SIAM J. Appl. Math., vol. 17, no. 6, pp. 1163–1170, 1969. [Google Scholar] [Crossref]
30.
M. Guajardo and M. Rönnqvist, “A review on cost allocation methods in collaborative transportation,” Int. Trans. Oper. Res., vol. 23, no. 3, pp. 371–392, 2016. [Google Scholar] [Crossref]
31.
M. Frisk, M. Göthe-Lundgren, K. Jörnsten, and M. Rönnqvist, “Cost allocation in collaborative forest transportation,” Eur. J. Oper. Res., vol. 205, no. 2, pp. 448–458, 2010. [Google Scholar] [Crossref]
32.
L. A. Guardiola, A. Meca, and J. Timmer, “Cooperation and profit allocation in distribution chains,” Decis. Support Syst., vol. 44, no. 1, pp. 17–27, 2007. [Google Scholar] [Crossref]
33.
B. Colson, P. Marcotte, and G. Savard, “An overview of bilevel optimization,” Ann. Oper. Res., vol. 153, no. 1, pp. 235–256, 2007. [Google Scholar] [Crossref]
34.
A. Sinha, P. Malo, and K. Deb, “A review on bilevel optimization: From classical to evolutionary approaches and applications,” IEEE Trans. Evol. Comput., vol. 22, no. 2, pp. 276–295, 2018. [Google Scholar] [Crossref]
35.
C. J. Nohra and N. V. Sahinidis, “Global optimization of nonconvex problems with convex-transformable intermediates,” J. Glob. Optim., vol. 72, no. 2, pp. 255–276, 2018. [Google Scholar] [Crossref]
36.
L. Torres and F. S. Ramos, “Allocating benefits due to shared resources using Shapley value and nucleolus in dynamic network Data Envelopment Analysis,” Mathematics, vol. 12, no. 5, p. 698, 2024. [Google Scholar] [Crossref]
37.
P. Dehez, “Cooperation and fair division,” in Game Theory for the Social Sciences, Cham, Switzerland: Springer, 2024. [Google Scholar] [Crossref]
38.
J. Dahlberg, M. Göthe-Lundgren, and S. Engevall, “A note on the nonuniqueness of the Equal Profit Method,” Appl. Math. Comput., vol. 308, pp. 84–89, 2017. [Google Scholar] [Crossref]
39.
ecoinvent Association, “ecoinvent database v3.9,” Zurich, Switzerland, 2022. [Online]. Available: https://support.ecoinvent.org/ecoinvent-version-3.9 [Google Scholar]
40.
M. Wang, A. Elgowainy, U. Lee, K. H. Baek, S. Balchandani, P. T. Benavides, A. Burnham, H. Cai, P. Chen, Y. Gan et al., “Greenhouse gases, Regulated Emissions, and Energy use in Technologies (GREET) Model® (2023 .Net),” Computer Software. USDOE Office of Energy Efficiency and Renewable Energy (EERE), 2023. [Online]. Available: https://www.osti.gov/doecode/biblio/113208 [Google Scholar]
41.
Department for Energy Security and Net Zero, “Greenhouse gas reporting: Conversion factors 2024,” Gov.uk, 2024. [Online]. Available: https://www.gov.uk/government/publications/greenhouse-gas-reporting-conversion-factors-2024 [Google Scholar]
42.
World Resources Institute and World Business Council for Sustainable Development, “The Greenhouse Gas Protocol: A corporate accounting and reporting standard, rev. ed.,” WRI/WBCSD, Washington, DC, USA, 2004. [Online]. Available: https://ghgprotocol.org/corporate-standard [Google Scholar]
43.
J.-F. Bérubé, M. Gendreau, and J.-Y. Potvin, “An exact ϵ-constraint method for bi-objective combinatorial optimization problems: Application to the Traveling Salesman Problem with Profits,” Eur. J. Oper. Res., vol. 194, no. 1, pp. 39–50, 2009. [Google Scholar] [Crossref]
44.
C. A. Coello Coello, G. B. Lamont, and D. A. Van Veldhuizen, Evolutionary Algorithms for Solving Multi-Objective Problems, 2nd. New York, NY, USA: Springer, 2007. [Online]. Available: [Google Scholar] [Crossref]
45.
E. Zitzler, L. Thiele, M. Laumanns, C. M. Fonseca, and V. G. Da Fonseca, “Performance assessment of multiobjective optimizers: An analysis and review,” IEEE Trans. Evol. Comput., vol. 7, no. 2, pp. 117–132, 2003. [Google Scholar] [Crossref]
46.
D. A. Van Veldhuizen and G. B. Lamont, “Multiobjective evolutionary algorithm research: A history and analysis,” Air Force Inst. Tech., Wright-Patterson AFB, OH, USA, Tech. Rep. TR-98-03, 1998, pp. 98–03. [Google Scholar]
47.
J. R. Schott, “Fault tolerant design using single and multicriteria genetic algorithm optimization,” M.S. thesis, Massachusetts Institute of Technology, Cambridge, MA, USA, 1995. [Google Scholar]
48.
M. A. Londe, L. S. Pessoa, C. E. Andrade, and M. G. Resende, “Biased random-key genetic algorithms: A review,” Eur. J. Oper. Res., vol. 321, no. 1, pp. 1–22, 2025. [Google Scholar] [Crossref]
49.
C. Blum, J. Puchinger, G. R. Raidl, and A. Roli, “Hybrid metaheuristics in combinatorial optimization: A survey,” Appl. Soft Comput., vol. 11, no. 6, pp. 4135–4151, 2011. [Google Scholar] [Crossref]
50.
M. Gen and L. Lin, “Multiobjective evolutionary algorithm for manufacturing scheduling problems: State-of-the-art survey,” J. Intell. Manuf., vol. 25, no. 5, pp. 849–866, 2014. [Google Scholar] [Crossref]
51.
K. Deb, A. Pratap, S. Agarwal, and T. Meyarivan, “A fast and elitist multiobjective genetic algorithm: NSGA-II,” IEEE Trans. Evol. Comput., vol. 6, no. 2, pp. 182–197, 2002. [Google Scholar] [Crossref]
52.
J. R. Birge and F. Louveaux, “Basic properties and theory,” in Introduction to Stochastic Programming, New York, NY, USA: Springer, 1997, pp. 83–135. [Google Scholar] [Crossref]
53.
M. Sklar, “N-dimensional distribution functions and their margins,” Ann. ISUP, vol. 8, no. 3, pp. 229–231, 1959. [Google Scholar]
54.
M. D. McKay, R. J. Beckman, and W. J. Conover, “A comparison of three methods for selecting values of input variables in the analysis of output from a computer code,” Technometrics, vol. 42, no. 1, pp. 55–61, 2000. [Google Scholar] [Crossref]
55.
J. Dupačová, N. Gröwe-Kuska, and W. Römisch, “Scenario reduction in stochastic programming,” Math. Program., vol. 95, no. 3, pp. 493–511, 2003. [Google Scholar] [Crossref]
56.
H. Moulin, “Equal or proportional division of a surplus, and other methods,” Int. J. Game Theory, vol. 16, no. 3, pp. 161–186, 1987. [Google Scholar] [Crossref]
57.
L. S. Shapley, “A value for n-person games,” in Contributions to the Theory of Games II, Princeton, NJ, USA: Princeton University Press, 1953, pp. 307–317. [Google Scholar] [Crossref]
58.
M. Nagarajan and G. Sošić, “Game-theoretic analysis of cooperation among supply chain agents: Review and extensions,” Eur. J. Oper. Res., vol. 187, no. 3, pp. 719–745, 2008. [Google Scholar] [Crossref]
59.
S. Lozano, P. Moreno, B. Adenso-Díaz, and E. Algaba, “Cooperative game theory approach to allocating benefits of horizontal cooperation,” Eur. J. Oper. Res., vol. 229, no. 2, pp. 444–452, 2013. [Google Scholar] [Crossref]
60.
B. Dai and H. Chen, “Profit allocation mechanisms for carrier collaboration in pickup and delivery service,” Comput. Ind. Eng., vol. 62, no. 2, pp. 633–643, 2012. [Google Scholar] [Crossref]
61.
M. Gansterer and R. F. Hartl, “Shared resources in collaborative vehicle routing,” TOP, vol. 28, no. 1, pp. 1–20, 2020. [Google Scholar] [Crossref]
62.
M. Lai, Y. Wu, and X. Cai, “Core-based cost allocation for collaborative multi-stop truckload shipping problem,” IISE Trans., vol. 57, no. 8, pp. 938–956, 2025. [Google Scholar] [Crossref]
63.
D. W. Yeung and L. A. Petrosyan, “A fair gain-sharing solution of cooperative dynamic games,” Int. Game Theory Rev., vol. 27, no. 3, p. 2550008, 2025. [Google Scholar] [Crossref]
64.
H. Sternberg, I. Linan, G. Prockl, and A. Norrman, “Tragedy of the facilitated commons: A multiple-case study of failure in systematic horizontal logistics collaboration,” J. Supply Chain Manag., vol. 58, no. 4, pp. 30–57, 2022. [Google Scholar] [Crossref]
65.
J.-F. Audy, S. D’Amours, and L.-M. Rousseau, “Cost allocation in the establishment of a collaborative transportation agreement—an application in the furniture industry,” J. Oper. Res. Soc., vol. 62, no. 6, pp. 960–970, 2011. [Google Scholar] [Crossref]
66.
S. A. Hosseini Shekarabi, R. Kiani Mavi, and F. Romero Macau, “Supply chain resilience: A critical review of risk mitigation, robust optimisation, and technological solutions and future research directions,” Glob. J. Flex. Syst. Manag., vol. 26, pp. 681–735, 2025. [Google Scholar] [Crossref]
67.
D. Cao, P. Liu, and C. Liu, “Revenue allocation for shared manufacturing collaboration considering participants’ contributions based on a differential game and an improved Shapley algorithm,” Mod. Supply Chain Res. Appl., vol. 8, no. 2, pp. 239–261, 2026. [Google Scholar] [Crossref]
68.
European Parliament and Council of the European Union, “Consolidated text: Directive (EU) 2024/1760 of the European Parliament and of the Council of 13 June 2024 on corporate sustainability due diligence and amending Directive (EU) 2019/1937 and Regulation (EU) 2023/2859 (Text with EEA relevance),” Official Journal of the European Union, L 2024/1760, 2024. [Online]. Available: https://eur-lex.europa.eu/eli/dir/2024/1760/2026-03-18 [Google Scholar]
Search
Open Access
Research article

stochastic Collaborative Supply Chain Optimization With Fair Hierarchical Gain-Sharing

Hanan Ouhader1*,
Asma Sbai2,
Malika El Kyal3
1
InterDisciplinary Applied Research Laboratory (LIDRA), Polytechnic School, International University of Agadir (Universiapolis), 80000 Agadir, Morocco
2
FMPM,Bioscience and Health Laboratory, Cadi Ayyad University, 40000 Marrakech, Morocco
3
Laboratory of Systems Engineering and Decision Support, National School of Applied Sciences, Ibn Zohr University, 80000 Agadir, Morocco
Journal of Engineering Management and Systems Engineering
|
Volume 5, Issue 4, 2026
|
Pages 558-588
Received: 09-07-2026,
Revised: 10-24-2026,
Accepted: 11-02-2026,
Available online: 11-19-2026
View Full Article|Download PDF

Abstract:

Horizontal supply chain collaboration (HSCC) offers substantial cost savings and carbon footprint reductions, yet commercial adoption remains hindered by non-scalable gain-sharing mechanisms and operational disruption risks. This study presents a scalable two-stage hierarchical framework for post-Pareto selection, integrating a two-stage stochastic mixed-integer linear program (MILP, Level 1) with an Equal Profit Method linear program (EPML, Level 2) to optimize multi-objective routing and equitable profit allocation across multi-partner logistics networks. Evaluated on a medium-scale benchmark network (15 partners, 8 demand nodes, and 50 stochastic scenarios), the framework achieves an alliance-wide cost reduction of 18.42% and a carbon reduction of 14.10% relative to non-cooperative baselines. At Level 2, the EPML engine executes in 0.32 s for 15 partners and 1.87 s for 30 partners, overcoming the exponential computational bottlenecks of traditional cooperative game theory methods while theoretically guaranteeing expected individual rationality $\left(E\left[\pi_i\right] \geq \pi_i^0\right)$, where $E\left[\pi_i\right]$ denotes partner $i$’s expected collaborative profit and $\pi_i^0$ denotes that partner's standalone baseline profit—provided that the selected physical solution satisfies the net-synergy feasibility condition $(\sum_{i \in I} E\left[\pi_i\right] \geq \sum_{i \in I} \pi_i^0)$, where $I$ is the set of alliance partners. The framework also achieves empirically verified core stability, with $e_{\max} \leq 0$, where $e_{\max}$ denotes the maximum excess across the tested subcoalition. Sensitivity analysis demonstrates robust customer fulfillment (99.59% service level) under high shortfall penalties $\left(p_j^{\text {pen}} \geq 2.5 \times p_j^{\text {sell}}\right)$ and a marginal abatement cost (MAC) of \$2,318.68 per metric ton of carbon dioxide equivalent (tCO$_2$e) at executive compromise configurations. To facilitate industry adoption, we translate these algorithmic outputs into a practical 4-step managerial decision protocol balancing ex-ante contractual rules, ex-post financial settlements, and alliance-level greenhouse gas (GHG) reporting boundaries. The framework effectively resolves the trade-off between allocation fairness and computational scalability in stochastic environments, providing supply chain decision-makers with a practical, mathematically rigorous, and computationally efficient post-Pareto decision-support tool to establish stable, equitable, and resilient horizontal logistics coalitions.

Keywords: Horizontal collaboration, Two-stage stochastic mixed-integer linear program, Hierarchical optimization, Gain-sharing, Sustainable and resilient supply chain, Tri-objective optimization

1. Introduction

The design of horizontal supply chain collaboration (HSCC)—where firms at the same echelon pool logistics assets, transportation capacity, and distribution networks—has attracted growing attention under the convergence of Industry 5.0 sustainability mandates and structural supply chain volatility [1], [2]. This operational imperative is further accelerated by global regulatory frameworks, such as the EU Corporate Sustainability Due Diligence Directive (CSDDD) and the Greenhouse Gas Protocol Corporate Standard [3]. Under these standards, Scope 3 emissions—specifically Category 1 (purchased goods and services) and Category 4 (upstream transportation and distribution)—represent 70–90% of a firm’s total carbon footprint, yet remain notoriously difficult to abate through unilateral actions [4]. While operational literature has quantified cost reductions of 15–30% through demand consolidation [5], 18–25% Scope 3 emission abatement via consolidated shipments [6], and enhanced structural resilience through cross-chain asset redundancy [7], integrating these three dimensions into a single, mathematically tractable optimization framework remains an open challenge. Existing contributions address subsets of these performance metrics—such as robust cost optimization [8], stochastic carbon–cost trade-offs [9], or cost–resilience balances [10]—as isolated problems; none unifies economic efficiency, Scope 3 environmental abatement, and disruption resilience simultaneously within a multi-objective stochastic design.

Despite documented theoretical gains, a persistent “Collaborative Paradox” hinders real-world implementation: While strategic logistics alliances promise substantial systemic benefits, empirical studies indicate that over half of horizontal logistics partnerships dissolve within two years. This high failure rate stems from two structural barriers:

• Asymmetric Gain Distribution: Participants fear opportunistic behavior and uncompensated asset exposure when cost reductions are not distributed in a demonstrably fair manner.

• Operational vs. Financial Misalignment: Post-hoc gain-sharing models fail to dynamically link financial redistribution with stochastic operational disruptions, leading to partner dissatisfaction when actual operational realities deviate from expected baseline plans.

Three fundamental research gaps in the current literature motivate this work:

• Endogenous Linear Program (LP)-Based Allocation vs. Exogenous Heuristics: Traditional allocation methods operate as exogenous ex-post accounting exercises after physical flows are fixed. In contrast, an endogenous approach formulates scenario-wise allocation fractions ($\alpha_i^S$) as decision variables within a dedicated LP operating over fixed Pareto-optimal physical flows

• Conflation of Distributive Fairness with Physical Trade-offs: Prior multi-objective formulations conflate fairness metrics with physical operational objectives within a single Pareto layer, introducing non-convex bilinear terms ($\alpha_i^s \cdot x_{i j s}^s$) that render large-scale problems intractable. A rigorous formulation requires decoupling physical efficiency from equity via a bilinear-free, hierarchical post-Pareto selection architecture.

• Absence of Expected Individual Rationality Guarantees: Traditional allocation rules lack explicit scenario-wise individual rationality guarantees under stochastic environments, failing to structurally ensure that expected collaborative profits meet or exceed non-cooperative baselines ($\mathbb{E}\left[\pi_i\right] \geq \pi_i^0$) across stochastic scenario realizations.

To resolve these challenges, this paper proposes a novel two-stage hierarchical post-Pareto selection architecture that decouples operational physical flow optimization from financial gain-sharing design. The primary contributions of this work are fourfold:

• Unified Tri-Objective Stochastic Modeling: We formulate a two-stage stochastic mixed-integer linear program (MILP) that concurrently balances expected total logistics costs (incorporating calibrated shortfall penalties), Scope 3 emissions via pre-computed unified route emission factors ($e_{i j}$), and disruption resilience measured through vulnerability-weighted flow exposure.

• Post-Pareto LP-Based Gain-Sharing: We introduce a Lexicographic Equal Profit Method linear program (EPML) formulated as a pure LP operating on decision variables $\alpha_i^s$ evaluated over fixed physical Pareto-optimal solutions. By reformulating the objective around relative net surplus gains ($g_{\min }$), the model structurally guarantees individual rationality ($E\left[\pi_i\right] \geq \pi_i^0$) across all participating partners without requiring iterative operational feedback.

• Bilinear-Free Hierarchical Architecture: We establish a two-level decomposition scheme that mathematically eliminates non-convex bilinear terms ($\alpha_i^s \cdot x_{i j}^s$) by resolving physical supply chain trade-offs at Level 1 (MILP) and establishing distributive equity at Level 2 (LP) as a post-Pareto selection step. This architecture establishes a clear unidirectional information flow where Level 1 passes the total scenario-wise operational cost ($C_{\text {total}}^S$) to Level 2 as a fixed parameter, avoiding non-convex mixed-integer nonlinear programming (MINLP) bottlenecks while preserving linear mixed-integer tractability at Level 1 and LP convexity at Level 2.

• Tightened Formulations and Computational Scalability: We replace conventional Big-M flow-activation bounds with a single tight constraint coupling activation and disruption-adjusted capacity ($x_{i j}^s \leq C A P_i \cdot y_i$), yielding tight LP relaxations and enabling gain-allocation scaling linearly to 30+ partners in under 2 s.

The remainder of this paper is organized as follows. Section 2 reviews related literature and highlights structural research gaps. Section 3 presents the mathematical nomenclature, formulates the Level 1 tri-objective stochastic MILP model, establishes the standalone non-cooperative baseline model, and details the Level 2 EPML gain-sharing model. Section 4 details the solution methodology, including the augmented epsilon-constraint method (AUGMECON2) Pareto solver, the hybrid genetic algorithm–linear programming (GA-LP) matheuristic, and the scenario generation and reduction engine via the Gaussian copula and k-means clustering. Section 5 provides computational validation, tri-objective trade-off analyses, and presents the numerical results of the comparative benchmarking of gain-sharing mechanisms against classical approaches. Section 6 proposes an in-depth discussion interpreting the results from strategic, operational, and environmental perspectives, while addressing limitations and future directions. Finally, Section 7 summarizes findings and discusses managerial implications.

2. Literature Review

The pursuit of horizontally coordinated supply chains has intensified as firms seek simultaneous improvements in cost efficiency, environmental accountability, and structural robustness against disruptions. This literature review systematically examines four interconnected streams that collectively motivate the proposed two-stage hierarchical framework: (i) HSCC and tri-objective synergies, (ii) stochastic optimization and multi-objective solution methodologies, (iii) cost and profit allocation mechanisms, and (iv) hierarchical versus integrated optimization architectures. Across these streams, we identify persistent gaps that the present work seeks to address.

2.1 Horizontal Supply Chain Collaboration and Tri-Objective Synergies

HSCC involves cooperative arrangements among firms operating at the same echelon, enabling joint logistics, asset sharing, and capacity consolidation that generate mutual efficiencies unattainable through unilateral optimization [11], [12]. Recent scholarship has documented substantial cost savings and service-level improvements arising from collaborative transportation planning, pooled warehousing, and shared fulfillment networks [13], [14]. Çerçi et al. [15] provide a contemporary taxonomy of horizontal logistics partnerships, emphasizing that digital platform enablers have expanded collaboration from bilateral carrier alliances to multi-party ecosystems with real-time matching capabilities. Parallel work by Mrabti et al. [3] synthesizes evidence that HSCC constitutes one of the most effective levers for addressing sustainability challenges in freight networks, particularly through load factor optimization and vehicle-kilometer reduction.

Despite these advances, integrating environmental and resilience objectives alongside economic cost minimization within HSCC models has traditionally been addressed primarily through pairwise extensions rather than fully unified frameworks. The Greenhouse Gas (GHG) Protocol identifies Scope 3 emissions—particularly Categories 1 (purchased goods and services) and 4 (upstream transportation and distribution)—as representing 70–90% of total corporate carbon footprints [4]. While early HSCC formulations treated emissions as secondary constraints or evaluated disruption resilience through post-hoc sensitivity analyses [16], recent scholarship has advanced bi-objective paradigms. Specifically, cost–carbon trade-offs have been examined in collaborative logistics settings [17], [18], and green or resilient supply chain network designs under uncertainty have been formulated separately [6], [8], [9], [10]. However, existing frameworks typically isolate environmental abatement from structural disruption risks. The simultaneous integration of expected logistics costs, Scope 3 emissions quantified via consolidated route emission factors ($e_{i j}$), and disruption resilience measured through vulnerability-weighted network flow remains limited in the stochastic HSCC literature, providing the primary motivation for this study.

2.2 Stochastic Optimization and Multi-Objective Frameworks in Supply Chains

Two-stage stochastic programming (SP) has established itself as a foundational paradigm for supply chain network design under uncertainty, wherein strategic first-stage decisions (facility activation) precede operational recourse actions (flow allocation and shortfall management) conditioned on realized scenarios [19], [20]. Scenario generation and reduction techniques are similarly well-established: Latin hypercube sampling (LHS) ensures uniform coverage across continuous parameter distributions, while copula-based approaches capture complex spatial dependence structures among correlated market demands and disruption severities [21]. Subsequent scenario reduction via $k$-means clustering or backward reduction algorithms preserves distributional fidelity while maintaining computational tractability [22]. Recent contributions demonstrate the effectiveness of two-stage stochastic frameworks across closed-loop and perishable logistics systems under multi-dimensional uncertainty [23], [24].

When multiple conflicting objectives are introduced, AUGMECON2 serves as a standard exact approach for generating non-dominated Pareto frontiers in multi-objective integer programming [25], guaranteeing Pareto optimality through slack variable augmentation and overcoming the limitations of weighted-sum methods on non-convex regions [26]. For large-scale NP-hard instances, hybrid matheuristics combining evolutionary search with exact mathematical programming sub-solvers offer a practical trade-off between solution quality and computational efficiency [27]. While AUGMECON2 and matheuristics are well-documented in multi-objective optimization, exact AUGMECON2 resolution encounters severe computational bottlenecks when scaled to tri-objective stochastic MILPs with dense recourse scenario trees. This computational boundary motivates pairing exact Pareto mapping with a scalable hybrid GA-LP matheuristic for large-scale instances.

2.3 Cost/Profit Allocation and Gain-Sharing Mechanisms

The allocation of collaborative cost savings among coalition members has traditionally been addressed through cooperative game-theoretic solution concepts. The Shapley value distributes gains according to marginal contributions across all coalition permutations [28], while the nucleolus minimizes the maximum dissatisfaction of any sub-coalition [29], and the core ensures that no subset of players has an incentive to deviate [29]. These axiomatic methods possess strong normative appeal; however, their application to large-scale supply chain networks is severely limited by an exponential computational complexity of $\mathcal{O}\left(2^{|I|}\right)$ in enumerating all sub-coalitions, potential non-uniqueness of allocation vectors, and the risk of an empty core when synergies exhibit non-supermodular characteristics [30].

Proportional and equal profit rules offer computationally tractable alternatives widely examined in collaborative transportation [5], [31]. The EPML, introduced by Frisk et al. [31] and extended by Gansterer and Hartl [5], seeks to minimize the maximum relative deviation among participants’profit shares using linear programming. While EPML significantly reduces computational burden relative to game-theoretic concepts, traditional implementations evaluate financial allocations primarily as ex-post accounting procedures over deterministic cost savings. Furthermore, classical allocation methods lack explicit individual rationality guarantees under stochastic scenario trees—that is, structurally ensuring that each participant’s expected allocated profit meets or exceeds its non-cooperative reservation baseline ($\mathbb{E}\left[\pi_i\right] \geq \pi_i^0$) across scenario realizations [32]. In this study, the Lexicographic EPML adapts the relative maximin surplus principle to operate endogenously as a Level 2 post-Pareto selection layer, treating scenario-wise allocation fractions ($\alpha_i^s$) as decision variables while enforcing binding expected individual rationality constraints.

2.4 Hierarchical vs. Integrated Optimization Architectures

Bilevel and hierarchical optimization architectures have attracted growing interest in collaborative network design, where strategic decisions at an upper level constrain operational decisions at a lower level [33], [34]. In the context of HSCC, a critical architectural question arises: should physical flow decisions and financial allocation decisions be optimized jointly within a single integrated model, or decomposed into a hierarchical sequence? Integrated formulations that simultaneously optimize flows and allocation fractions introduce bilinear terms of the form $\alpha_i^s \cdot x_{i j s}^s$, rendering the model non-convex and computationally intractable for practical instance sizes [35]. This bilinear non-convexity arises because each firm’s allocated cost or profit is a product of its allocation fraction and the collaborative flow passing through shared facilities and routes.

The proposed two-stage hierarchical architecture circumvents this difficulty through principled decoupling. Level 1, formulated as a tri-objective stochastic MILP, optimizes all physical flow and network design decisions while remaining agnostic to the specific allocation fractions. Level 2, formulated as a pure LP, subsequently determines the gain-sharing allocations ($\alpha_i^s$) by treating Level 1’s optimal cost outcomes ($C_{\text {total}}^S$) as fixed parameters. This sequential structure establishes a strictly unidirectional information flow without an iterative feedback loop from Level 2 back to Level 1. In this setting, the gain-sharing mechanism is “endogenous” because the scenario-wise allocation fractions ($\alpha_i^s$) are decision variables optimized within the Level 2 LP, rather than being assigned via exogenous post-hoc accounting rules; however, this allocation is computed over fixed physical flow solutions generated at Level 1.

Crucially, Level 2 functions as a post-Pareto selection layer: after Level 1 constructs the non-dominated Pareto frontier across physical operational objectives (cost, emissions, disruption risk), Level 2 evaluates the financial viability and equity of candidate physical Pareto solutions. Decision-makers can thus verify whether a selected physical compromise point (e.g., Point C) yields an equitable and individually rational financial distribution across all participants. By decoupling physical optimization from financial allocation, this sequential architecture mathematically eliminates non-convex MINLP bottlenecks, maintains a linear MILP formulation at Level 1 and a convex LP at Level 2, and guarantees that Level 2 executes in 0.32 s for 15 partners and 1.87 s for 30 partners. The lexicographic objective structure of EPML further guarantees that expected individual rationality ($E\left[\pi_i\right] \geq \pi_i^o$, corresponding to $g_{\text {min}}^* \geq 0$) is satisfied as a hard linear constraint rather than a soft aspiration level, provided that the selected Level 1 physical solution satisfies the net-synergy feasibility condition stated in Section 3.2 ($\sum_i E\left[\pi_i\right] \geq \sum_i \pi_i^o$). To synthesize these four literature streams and position our methodological contributions relative to existing studies, Table 1 provides a structured comparative overview of recent collaborative supply chain frameworks.

Table 1. Comparative synthesis of recent literature and proposed work

Author(s) & Year

Stochastic Uncertainty

Environmental Metric

Disruption Resilience

Gain-Sharing Mechanism

Optimization Architecture

Guaranteed Individual Rationality

Gansterer and Hartl [5]

Deterministic

Not explicitly modeled

Not explicitly modeled

Ex-post post-hoc (Shapley value, EPML)

Single-level routing

Not guaranteed

Ghomi-Avili et al. [6]

Robust (polyhedral)

Bi-objective (Cost + CO$_2$)

Facility disruption

Not explicitly modeled

Integrated MONLP

Not guaranteed

Hacardiaux et al. [7]

Stochastic (scenarios)

Bi-objective (Cost + Footprint)

Not explicitly modeled

Ex-post post-hoc (core, Shapley value)

Sequential bi-objective

Conditional (ex-post)

Safari et al. [9]

Two-stage SP

Bi-objective (Cost + CO$_2$)

Supplier disruption

Ex-post post-hoc (Shapley value)

Integrated single-level

Not guaranteed

Saffari et al. [8]

Two-stage SP

Not explicitly modeled

Robust facility resilience

Ex-post post-hoc (proportional)

Sequential two-stage

Not guaranteed

Filho et al. [10]

Two-stage SP + copulas

Bi-objective (Cost + CO$_2$)

Not explicitly modeled

Ex-post post-hoc (nucleolus)

Integrated single-level

Not guaranteed

This work (proposed framework)

Two-stage SP + LHS/clustering

Tri-objective (Cost + Scope 3 + Resilience)

Vulnerability-weighted flow exposure

Endogenous post-Pareto LP (EPML)

Hierarchical (Level 1 MILP/Level 2 LP)

Yes (embedded $g_{\text{min}}$ constraint)

Note: CO$_2$ = carbon dioxide; EPML = Equal Profit Method linear program; $g_{\text{min}}$ = minimum relative surplus parameter; LHS = Latin hypercube sampling; LP = linear program; MILP = mixed-integer linear program; MONLP = multi-objective nonlinear program; SP = stochastic programming.
2.5 Synthesis of Research Gaps

The comparative analysis presented in Table 1 reveals three primary research gaps that collectively motivate the proposed framework:

• Gap 1 (Exogenous Post-Hoc Rules vs. Endogenous LP-Based Gain-Sharing): While cost/profit allocation mechanisms such as the Shapley value, nucleolus, and EPML are well-established in collaborative logistics [36], [37], [38], they are predominantly applied as exogenous post-hoc accounting rules after physical flows are fixed. In contrast, the proposed EPML engine endogenizes allocation fractions ($\alpha_i^s$) as decision variables within a dedicated LP operating over fixed Pareto-optimal physical flow decisions. This ensures that profit distribution is optimized to maximize relative equity ($g_\mathrm{min}$) rather than imposed via rigid external accounting heuristics, while avoiding non-convex integrated formulations.

• Gap 2 (Conflation of Fairness with Physical Trade-Offs & Bilinear Non-Convexity): Prior multi-objective formulations attempting to optimize physical operations and allocation fractions simultaneously introduce non-convex bilinear interaction terms ($\alpha_i^S \cdot x_{i j}^S$), leading to severe MINLP computational bottlenecks. Furthermore, this conflation obscures the distinct nature of the gain-sharing decision, which is fundamentally a post-Pareto selection problem operating on a different decision variable space. The hierarchical architecture proposed here explicitly separates these concerns: Level 1 generates the Pareto frontier over physical objectives, while Level 2 operates as a post-Pareto selection layer on the allocation space, avoiding bilinear non-convexities, maintaining a linear MILP formulation at Level 1 and a convex LP at Level 2, and ensuring execution within 2 s for 30 partners.

• Gap 3 (Absence of Individual Rationality Guarantees Under Stochastic Environments): Although classical allocation concepts ensure core stability in deterministic settings, enforcing individual rationality in stochastic environments—guaranteeing that each partner’s expected collaborative profit meets or exceeds its non-cooperative baseline ($\mathbb{E}\left[\pi_i\right] \geq \pi_i^0$) across scenario realizations—remains unaddressed in existing multi-objective frameworks.

3. Nomenclature and Problem Definition

Table 2 summarizes all notations used throughout this paper. The collaborative network comprises $|I|$ horizontal alliance partners operating at a single echelon. Each partner $i$ is characterized by a declared nominal capacity $C A P_i$, a baseline reliability coefficient $\rho_i$, a fixed activation cost $f_i$, a per-unit processing cost $v_i$, a consolidated route emission factor $e_{i j}$, a standalone operating cost baseline $C_i^o$, and a standalone profit baseline $\pi_i^o$. A unified per-route emission coefficient $e_{i j}$ captures both production and transportation Scope 3 emissions, eliminating the need for separate indices. Direct inter-partner shipments incur a transport cost $c_{i j}$. Customer nodes $j$ offer a market selling price $p_j^\mathrm{sell}$, and demand shortfalls incur a penalty $p_j^{\text {pen}}$. Set $I$ represents the fixed Grand Coalition of $N$ contracting alliance partners who have committed to joint logistics coordination. While contractual alliance membership is fixed ex-ante across all partners $i \in I$, the operational activation of each partner’s logistics facility/hub is governed by the binary decision variable $y_i \in\{0,1\}$. If partner $i$’s facility is operationally activated to process inter-partner freight ($y_i=1$), the network incurs a fixed setup expense $f_i$; if $y_i=0$, the facility remains operationally idle for shared routing, incurring zero fixed setup cost. Crucially, regardless of whether a partner’s facility is operationally active ($y_i$ = 1) or idle ($y_i$= 0) in Level 1, all partners $i \in I$ remain full members of the Grand Coalition and participate in the Level 2 financial gain-sharing mechanism.

Table 2. Nomenclature
SymbolDomainDefinition
$i, j, s$–Partner, demand-node, and scenario indices
$I, J, S$–Sets of partners, demand nodes, and scenarios
$y_{i}$$\{0,1\}$Binary facility activation (1 if partner $i$’s facility is operationally activated for shared routing; 0 if idle)
$x_{ij}^{s}$$\ge$0Flow from partner $i$ to node $j$ under scenario $s$
$u_{ij}^{s}$$\ge$0Unserved demand at node $j$ under scenario $s$
$f_{i}$\$Fixed activation cost of partner $i$
$v_{i}$\$/unitPer-unit processing cost at partner $i$
$c_{ij}$\$/unitPer-unit transport cost from $i$ to $j$
$e_{ij}$kg/unitConsolidated route emission factor (production + transport)
$CAP_{i}$unitsDeclared nominal capacity of partner $i$
$\rho_{i}$$[ 0,1]$Baseline reliability coefficient of partner $i$
$\rho_{i}^{s}$$[ 0,1]$Scenario-specific reliability of partner $i$ under scenario $s$
$\delta_{i}^{s}$$[ 0,1]$Sampled disruption severity for partner $i$ under scenario $s$
$D_{j}^{s}$unitsDemand at node $j$ under scenario $s$
$D_{i,j}^{s}$$\ge 0$Standalone dedicated market share of partner $i$ at demand node $j$ under scenario $s$ (units)
$p^{s}$$[ 0,1]$Probability of scenario $s$ ($\sum_{s \in S} p^{s} = 1$)
$p_{j}^{\mathrm{sell}}$\$/unitMarket selling price at demand node $j$
$p_{j}^{\mathrm{pen}}$\$/unitDemand shortfall penalty at node $j$ ($\ge 2.5 \times p_{j}^{\mathrm{sell}}$)
$C_{i}^{0}$\$Standalone operating cost of partner $i$
$R_{i}^{s}$\$Revenue of partner $i$ under scenario $s$
$\pi_{i}^{s}$\$Profit of partner $i$ under scenario $s$
$E[\pi_{i}]$\$Expected profit of partner $i$ across all scenarios
$C_{\mathrm{total}}^{s}$\$Total network operational cost under scenario $s$
$\pi_{i}^{0}$\$Standalone profit of partner $i$ (from standalone MILP)
$Z_{1}$, $Z_{2}$, $Z_{3}$–Objective values for expected cost (\$), Scope 3 carbon (kg CO$_2$e), and disruption risk (units)
$\alpha_{i}^{s}$$[ 0,1]$Gain-sharing allocation fraction for partner $i$ under scenario $s$ ($|I| \times |S|$ variables)
$g_{\mathrm{min}}$, $g_{k}$–Equal Profit Method linear program (EPML) relative surplus auxiliary variables
$M$–Leximin tie-breaking regularization coefficient ($\mu \to 0^{+}$)
$\Theta$–Equity ratio ($\Theta \in [ 0,1]$)
$\varepsilon$$>$0Small positive constant ($\varepsilon = 10^{-3}$) to prevent division by zero in relative surplus normalization when $\pi_{i}^{0} \le 0$
Note: MILP = mixed-integer linear program; CO$_2$e = carbon dioxide equivalent.

A set of $|S|$ disruption scenarios is drawn from probability space, with scenarios occurring with probability $\pi^s$ and imposing demand $D_j^s$ at each demand node $j$. Scenariospecific reliability $\rho_i^s$ is derived from sampled disruption severities. The model is a single-echelon, single-period, direct-shipping network designed to isolate the structural contributions of the endogenous gain-sharing mechanism and the tri-objective trade-off structure.

The overall optimization framework is decomposed into two sequential levels to maintain mathematical tractability, prevent bilinear non-convexity, and establish a clear post-Pareto decision structure. The architecture operates via a unidirectional information exchange without an iterative feedback loop from Level 2 back to Level 1:

• Level 1 (Section 3.1–3.2): Resolves all physical supply chain flow, routing ($x_{i j s}^s$), and network design decisions via a tri-objective stochastic MILP, generating the nondominated Pareto frontier across operational objectives (Total Cost $Z_1$, Carbon Footprint $Z_2$, Disruption Exposure $Z_3$). For any selected Pareto-optimal physical tradeoff point, Level 1 passes the resulting scenario-wise total network operating costs ($C_{\text {total}}^S$) and fulfilled demand volumes to Level 2 as fixed, known constant parameters.

• Level 2 (Section 3.3): Resolves coalition fairness and profit redistribution by formulating allocation fractions ($\alpha_i^s$) as decision variables within a pure EPML. Level 2 treats Level 1’s physical flow outputs as fixed parameters, thereby serving as a financial post-Pareto selection layer. Decision-makers utilize Level 2 to evaluate whether a physical Pareto compromise point yields an acceptable, equitable, and individually rational financial distribution ($\mathbb{E}\left[\pi_i\right] \geq \pi_i^0$) across all partners.

This hierarchical decomposition mathematically decouples physical optimization from financial gain-sharing. A single-level integrated formulation attempting to co-optimize physical flows ($x_{i j s}^s$) and financial allocation fractions ($\alpha_i^s$) simultaneously would introduce non-convex bilinear terms of the form $\alpha_i^s \cdot x_{i j s}^s$, transforming the problem into an intractable MINLP. By enforcing this sequential decoupling, the framework avoids bilinear non-convexities, maintaining a linear mixed-integer formulation at Level 1 and a convex LP at Level 2, allowing Level 2 to execute in 0.32 s for 15 partners and 1.87 s for 30 partners.

3.1 Level 1: Tri-Objective Stochastic Mixed-Integer Linear Program

The model adopts a two-stage SP structure where operational facility activation ($y_i$) is determined in the first stage (here-and-now), while operational recourse decisions (flows $x_{i j}^s$ and unserved demand $u_j^s$) are resolved in the second stage (wait-and-see) following scenario realization. The master problem is formulated as a tri-objective vector minimization:

$\min \left\{Z_1, Z_2, Z_3\right\}$
(1)

The scalarization of this vector problem is handled exclusively within the AUGMECON2 solution method via the augmented epsilon-constraint approach (Section 4.1). No composite or weighted-sum scalarization is used in the problem formulation.

3.1.1 Objective $Z_1$: Expected total cost

The total cost objective accounts for the first-stage fixed activation expenses and the expected second-stage variable costs, encompassing transportation across the network. A unit shortfall penalty $p_j^{\text {pen}}$ is imposed on unserved demand $u_j^s$ to eliminate the degenerate trivial solution where all flows and activations are zero and all demand is left unserved. Without this penalty, the objective is minimized by setting $x=0$, $y=0$, $u=D$, yielding a zero cost mathematically optimal but operationally meaningless solution:

$Z_1=\sum_{i} f_{i}\cdot y_{i}+\sum^{s}\pi^{s}\left[\sum_{i}\sum_{j}\left(v_{i}+c_{ij}\right)\cdot x_{ij}^{s}+\sum_{j}p_{j}^\mathrm{pen}\cdot u_{j}^{s}\right]$
(2)

The penalty parameter is calibrated relative to market selling prices rather than operational costs, ensuring that unmet demand always carries a higher opportunity cost than fulfillment:

$p_j^{\mathrm{pen}} \geq 2.5 \times p_j^{\mathrm{sell}},\quad \forall j \in J$
(3)

Across all computational experiments, we set $p_j^{\text {pen}}=2.5 \times p_j^\mathrm{sell}$. This calibration ensures the penalty strictly exceeds the maximum per-unit revenue loss from any unfulfilled unit, making the zero-flow solution suboptimal under all scenarios.

3.1.2 Objective $Z_2$: Scope 3 carbon footprint

The carbon objective captures the total Scope 3 GHG emissions arising from processing activities and inter-partner transportation. In contrast to prior formulations that separate production emission coefficients ($e_i$) and transport emission factors ($e_{i j}{ }^{t r}$), we consolidate these into a single unified route emission factor $e_{i j}$ per partner-demand node pair. This simplification reduces the number of emission parameters from $|I|+|I| \times|J|$ to $|I| \times|J|$ while preserving full modeling fidelity, since the consolidated factor is precomputed as $e_{i j}=e_i+e_{i j}^{\text {tr}}$ from source databases:

$Z_2=\sum^s \pi^s \sum_{i} \sum_{j} e_{ij} \cdot x_{ij}^s$
(4)

Emission coefficients $e_{ij}$ are sourced from ecoinvent v3.9 [39], Greenhouse gases, Regulated Emissions, and Energy use in Technologies (GREET) 2023 [40], and UK DEFRA 2024 [41]. Aligned with the GHG Protocol [42], the optimization boundary is strictly limited to logistics-focused Scope 3 Categories 1 and 4.

To maintain rigorous carbon accounting standards, the reporting boundary and perspective adopted in formulation (4) must be explicitly defined. The optimization model adopts a consolidated alliance-level accounting perspective, evaluating emissions from the viewpoint of a joint logistics platform governing horizontal asset sharing across participating firms. Under this alliance boundary, the unified route emission factor $e_{ij}$ integrates two distinct operational activities: (i) per-unit facility processing and handling emissions at partner hub $i$ (representing Scope 3 Category 1—Purchased Goods and Services for non-owning partners), and (ii) inter-partner freight transportation from hub $i$ to demand node $j$ (representing Scope 3 Category 4—Upstream Transportation and Distribution).

While consolidating these components into a single route factor $e_{i j}=e_i^{\text {proc}}+$ $e_{i j}^{\text {trans}}$ preserves mathematical linearity and computational tractability for network-level physical flow allocation, its structural limitations should be recognized. In practice, processing activities at facility $i$ constitute Scope 1/2 emissions for the facility owner, but represent Scope 3 Category 1 emissions for alliance partners utilizing shared capacity. Consequently, $e_{i j}$ serves as an operational comparative proxy for multiobjective network design rather than a substitute for firm-specific corporate GHG Protocol inventory accounting. Decision-makers should treat $Z_2$ as a relative decarbonization trade-off metric across physical routing configurations rather than a direct firm-level corporate GHG inventory statement.

3.1.3 Objective $Z_3$: Resilience via disruption risk

The third objective quantifies supply chain resilience through the expected disruption risk, defined as the expected volume of flow passing through partners with reduced reliability. This metric is linear in the decision variables and directly penalizes the exposure of supply flows to disruption-induced capacity losses:

$Z_3=\sum^s \pi^s \sum_{i}\left(1-\rho_{i}^s\right) \sum_{j} x_{ij}^s$
(5)

When $\rho_i^s=1$ (full reliability under scenario $s$), partner $i$ contributes zero disruption risk; when $\rho_i^s<1$, the fraction ($1-\rho_i^s$) quantifies the proportional vulnerability of flow routed through partner $i$.

3.1.4 Structural constraints

The structural constraints of the model are formulated as follows. Constraint (6) ensures demand satisfaction: total inbound flow to node $j$ plus unserved demand must equal realized demand under each scenario. The raw flow $x_{i j}^s$ is used without a reliability multiplier, as the reliability parameter enters the model exclusively through the capacity constraint (7). The unserved demand $u_j^s$ serves as a free slack variable guaranteeing LP feasibility for every second-stage sub-problem:

$\sum_{i} x_{ij}^{s}+u_{j}^{s}=D_{j}^{s}, \quad \forall j \in J,\ \forall s \in S$
(6)

Rather than employing a two-part constraint system involving an arbitrary Big-M constant for flow-activation coupling and a separate capacity bound, we introduce a single, tight combined constraint that simultaneously enforces both binary activation and disruption-adjusted capacity. This formulation eliminates the need for Big-M calibration and yields a tighter LP relaxation:

$\sum_{j} x_{ij}^{s} \leq \rho_{i}^{s} \cdot C A P_{i} \cdot y_{i}, \quad \forall i \in I, \forall s \in S$
(7)

When $y_i=0$ (partner not activated), the right-hand side becomes zero and all flows from partner $i$ are prohibited. When $y_i=1$ (partner activated), the effective capacity equals $\rho_i^s \cdot C A P_i$, tightening proportionally to the disruption severity. This is the only constraint where the reliability parameter appears, ensuring that disruption-induced capacity reduction is applied exactly once.

$y_{i} \in\{0,1\},\quad x_{ij}^{s} \geq 0, \quad u_{j}^{s} \geq 0,\quad \forall i, j, s$
(8)
3.2 Standalone Non-Cooperative Baseline Model

For every supply chain partner $i \in I$, the individual non-cooperative optimization problem is formulated to compute the standalone profit baseline $\pi_i^o$. In the standalone non-cooperative benchmark, total market demand $D_j^s$ at demand node $j$ under scenario $s$ is not available in its entirety to every firm. Instead, each partner $i \in$ I possesses a dedicated baseline market share $D_{i, j}^s$, such that total market demand is partitioned across coalition members: $\sum_{i \in I} D_{i, j}^s=D_j^s \forall j \in J, \forall s \in S$. Each partner $i$ independently optimizes its own operations considering strictly its assigned market share $D_{i, j}^s$ and its standalone assets, without access to inter-firm asset sharing or consolidated routing.

The standalone objective maximizes expected profit:

$\max \pi_i^0=\sum_{s \in S} p^s\left[\sum_{j \in J} p_j^{\text {sell}} \cdot D_{i j, 0}^s-\sum_{j \in J}\left(v_i+c_{i j}\right) \cdot D_{i j, 0}^s-\sum_{j \in J} p_j^{\text {pen}} \cdot u_{i j, 0}^s-f_i \cdot y_{i, 0}\right]$
(9)

Subject to the following constraints. The capacity constraint (10) limits total fulfilled demand to the partner’s disruption-adjusted effective capacity, modulated by the autonomous activation decision:

$\sum_{j} D_{ij, 0}^{s} \leq \rho_{i}^{s} \cdot C A P_{i} \cdot y_{i, 0},\quad \forall s \in S$
(10)

The demand balance constraint (11) ensures that for each customer node $j$, the demand served independently by partner $i$ plus the unfulfilled shortfall equals the total demand realization:

$D_{i j, 0}^s+u_{i j, 0}^s=D_{i, j}^s,\quad \forall j \in J, \forall s \in S$
(11)

The variable domain constraint (12) enforces integrality and non-negativity:

$y_{i, o} \in\{0,1\},\quad D_{ij, o}^{s} \geq 0,\quad u_{ij, o}^{s} \geq 0,\quad \forall i, j, s$
(12)

This standalone model is solved independently for each partner $i \in I$ using the same demand scenarios, cost parameters, and disruption realizations as the collaborative model. The resulting optimal value $\pi_i^o$ serves as the non-cooperative baseline against which individual rationality is assessed in Level 2.

Remark 1 (Illustrative Numerical Example of Standalone Profit Computation): Consider a 2-partner alliance $\left(P_1, P_2\right)$ serving a single market $\left(J_1\right)$ with total demand $D_1=100$ units, selling price $p^{\text {sell }}=\$ 100/$unit, and shortfall penalty $p^{\text {pen}}=$ $\$ 250/$unit. Demand is partitioned into dedicated market shares $D_{1,1}=60$ units for $P_1$ and $D_{2,1}=40$ units for $P_2$. Partner $P_1$ has capacity $C A P_1=80$, setup cost $f_1=$ $\$ 500$, and unit variable cost $v_1+c_{11}=\$ 30$. Partner $P_2$ has capacity $C A P_2=50$, setup cost $f_2=\$ 800$, and unit variable cost $v_2+c_{21}=\$ 55$.

In the non-cooperative standalone baseline, $P_1$ independently serves its 60 units, earning profit $\pi_1^0=60 \times(100-30)-500=\$ 3,700$. Partner $P_2$ serves its 40 units, earning profit $\pi_2^0=40 \times(100-55)-800=\$ 1,000$. Total non-cooperative baseline profit is $\sum \pi_i^0=\$ 4,700$. When the coalition forms, total market demand (100 units) is pooled. The coalition routes 80 units through $P_1$’s lower-cost facility and 20 units through $P_2$, generating a collaborative profit of \$5,200. The resulting \$500 financial synergy is subsequently distributed via Level 2 EPML.

3.3 Level 2: Post-Pareto Endogenous Gain-Sharing via Equal Profit Method Linear Program

For any selected Pareto-optimal physical solution ($\hat{y}$, $\hat{x}$) generated by Level 1, the total system operational cost $C_{\text {total}}^S$ under each scenario $s \in S$ becomes a known, fixed constant parameter passed directly from Level 1. Operating over these fixed physical decisions without altering operational flows, Level 2 subsequently optimizes the scenario-wise gain-sharing allocation variables $\alpha_i^s$ as a pure LP to perform post-Pareto financial evaluation.

The cost-conservation constraint ensures that allocated costs sum to the total system cost for every scenario:

$\sum_{i} \alpha_{i}^{s} \cdot C_{\text{total}}^s=C_{\text{total}}^s \Rightarrow \sum_{i \in I} \alpha_i^s=1,\quad \forall s \in S,\quad \alpha_i^s \geq 0,\quad \forall i \in I, \forall s \in S$
(13)

Crucially, the summation in Constraint (13) and the decision variables $\alpha_i^s$ encompass all contractual alliance members $i \in I$, yielding $|I| \times|S|$ decision variables. Partners whose logistics facilities were operationally inactive ($\hat{y}_i=0$) in the Level 1 physical routing solution remain full financial participants in Level 2. This ensures that their standalone reservation levels ($\pi_i^0$) are respected and that all contracting partners receive an equitable share of the collaborative savings.

Partner revenue under scenario s is proportional to the fulfilled demand, using the market selling price:

$R_i^s=\alpha_i^s \cdot \sum_j p_j^{\text {sell}} \cdot\left(D_j^s-u_j^s\right)$
(14)

The profit of partner $i$ under scenario $s$ and its expected value across scenarios are:

$\pi_{i}^{s}=R_{i}{ }^{s}-\alpha_{i}^{s} \cdot C_{\text{total}}^s$
(15)
$E\left[\pi_{i}\right]=\sum^{s} \pi^{s} \cdot \pi_{i}^{s},\quad \forall i \in I$
(16)
3.3.1 Lexicographic relative surplus maximin formulation

The EPML procedure is reformulated to maximize the minimum relative net surplus gain $g_{\min }$ over each partner’s standalone baseline $\pi_i^o$, rather than maximizing an absolute guaranteed profit floor. This reformulation embeds individual rationality directly into the LP objective, as any feasible solution satisfying the surplus constraints necessarily guarantees $E\left[\pi_i\right]-\pi_i^0 \geq 0$, ensuring no coalition partner is financially disincentivized from joining the collaborative network. The primary EPML LP maximizes the minimum relative surplus $g_{\min }$ across all partners:

$\max g_{\min}$
(17)
$\text { subject to: } g_{\min } \leq \frac{\mathbb{E}\left[\pi_i\right]-\pi_i^0}{\max \left(\pi_i^0, \epsilon\right)},\quad \forall i \in I$
(18)

along with the cost-conservation constraint (13) and non-negativity $\alpha_i^S \geq 0$ for all $i \in$ $I, s \in S$.

Mathematical Feasibility and Expected Individual Rationality Conditions: The Level 2 EPML optimizes allocation fractions $\alpha_i^s$ subject to scenario-wise cost conservation ($\sum_{i \in I} \alpha_i^s=1, \forall s \in S$). The mathematical feasibility of achieving expected individual rationality ($E\left[\pi_i\right] \geq \pi_i^0, \forall i \in I$) with $g_{\text {min}}^* \geq 0$ depends on the net economic synergy generated by the selected Level 1 physical flow configuration ($\hat{y}$, $\hat{x}$). Formally, expected individual rationality($g_{\text {min}}^* \geq 0$) is mathematically achievable if and only if the total expected collaborative profit is greater than or equal to the sum of non-cooperative baselines, i.e., $\sum_{i \in I} E\left[\pi_i\right] \geq \sum_{i \in I} \pi_i^0$, where $E\left[\pi_i\right]=$ $\sum_{s \in S} p^s\left[\sum_{j \in J} p_j^{\text {sell}}\left(D_j^s-u_j^s\right)-C_{\text {total}}^s\right]$. When this net synergy condition holds ($\sum_{i \in I} E\left[\pi_i\right] \geq \sum_{i \in I} \pi_i^0$), the EPML LP is guaranteed to be feasible with $g_{\text {min }}^* \geq 0$, structurally ensuring $E\left[\pi_i\right] \geq \pi_i^0\left(1+g_{\min }^*\right) \geq \pi_i^0$ for all contractual partners $i \in I$. Conversely, if a candidate Level 1 physical solution prioritizes non-economic objectives such that overall operating costs exceed standalone baselines $\left(\sum_{i \in I} E\left[\pi_i\right]<\sum_{i \in I} \pi_i^0\right)$, the EPML objective yields $g_{\text {min}}^*<0$, explicitly indicating that expected individual rationality cannot be satisfied under that physical configuration.

Baseline Regularization and Linearity: To prevent potential division-by-zero or sign-inversion when a standalone baseline is zero or negative ($\pi_i^0 \leq 0$), Eq. (18) incorporates the regularizing threshold $\max \left(\pi_i^0, \epsilon\right)$, where $\epsilon=10^{-3}$. Furthermore, the regularization coefficient $\mu \rightarrow$ $0^{+}$(set to $\mu=10^{-6}$) acts as a tie-breaking perturbation parameter that breaks ties in the active minimum surplus set without altering the primary maximin objective. Since the total scenario operational cost $C_{\text {total}}^s$ is a known fixed constant parameter passed from Level 1, all profit expressions $\mathbb{E}\left[\pi_i\right]$ remain strictly linear in $\alpha_i^s$. Consequently, each EPML sub-problem is a pure, convex LP comprising $|I| \times|S|$ decision variables and $|I|+|S|+1$ constraints, guaranteeing polynomial-time solvability.

The secondary LP fixes the first minimum surplus and maximizes the second-worst-off partner’s surplus gain, iterating until all partners are ranked:

$\max g_k=\frac{E\left[\pi_{i(k)}\right]-\pi_{i(k)}^0}{\pi_{i(k)}^0},\quad \text { s.t. } \quad g_k \leq g_{k-1}$
(19)

Under non-degeneracy (distinct relative surplus values across partners), this lexicographic maximin formulation yields a unique optimal expected profit vector. The total number of LP sub-problems is at most $|I|-1$, making the procedure computationally tractable even for large partner sets. Individual rationality is structurally guaranteed: any feasible solution in Level 2 satisfies $E\left[\pi_i\right]-\pi_i^0 \geq 0$, ensuring that every partner’s collaborative profit meets or exceeds its standalone baseline.

3.3.2 Theoretical core stability and property analysis of Equal Profit Method Linear Program

In cooperative game theory, a gain-sharing allocation vector $\mathbf{x}=\left[x_1, \ldots, x_{|I|}\right]^T$ belongs to the core of a characteristic function game $(I, v)$ if it satisfies Grand Coalition efficiency and prevents sub-coalitional deviation:

$\sum_{i \in I} x_i=v(I)$
(20)
$\sum_{i \in T} x_i \geq v(T),\quad \forall T \subseteq I$
(21)

where, $v(T)$ represents the maximum collective profit obtainable by subcoalition $T$ acting independently.

Theorem 1 (Expected Individual Rationality and Individual Defection Resistance):

Let $g_{\text {min}}^*=\max \min _{i \in I}\left(\frac{\mathbb{E}\left[\pi_i\right]-\pi_i^0}{\max \left(\pi_i^0, \epsilon\right)}\right)$ be the optimal objective value returned by the Level 2 EPML model over a feasible Level 1 physical solution $\left(y^*, x^*\right)$. If the collaborative network yields strict overall economic synergy relative to standalone baselines $\left(\sum_{i \in I} \mathbb{E}\left[\pi_i\right]>\sum_{i \in I} \pi_i^0\right)$, then:

• Strict Expected Individual Rationality holds: $g_{\text {min}}^*>0$, structurally guaranteeing $\mathbb{E}\left[\pi_i\right]>\pi_i^0$ for all participating partners $i \in I$.

• Individual Defection Resistance: No individual partner $i \in I$ has a financial incentive to defect from the coalition and operate in isolation $(T=\{i\})$.

Proof:

By formulation, Level 2 enforces $\mathbb{E}\left[\pi_i\right] \geq \pi_i^0\left(1+g_{\min}^*\right)$ as hard linear constraints for all $i \in I$ (Eq. 18). When total collaborative profit exceeds total standalone baselines, the relative surplus parameter $g_{\text {min }}^*$ is strictly positive $\left(g_{\text {min}}^*>0\right)$. Consequently, $\mathbb{E}\left[\pi_i\right]-$ $\pi_i^0 \geq g_{\text {min }}^* \cdot \max \left(\pi_i^0, \epsilon\right)>0$, guaranteeing that every partner earns an expected profit strictly greater than its non-cooperative reservation level $\pi_i^0$.

Remark 1 (Scope of Theoretical Guarantee vs. Empirical Core Verification): Theorem 1 analytically proves expected individual rationality, guaranteeing that no single partner $i \in I$ has a financial incentive to defect individually $(T=\{i\})$. However, general core stability against arbitrary multi-partner sub-coalitions $(|T| \geq 2)$ depends on specific network synergy structures and is not claimed as a universal analytical property of EPML for all conceivable network topologies. Instead, sub-coalitional stability $\left(e_{\text {max}} \leq 0\right)$ is evaluated and verified empirically across all $2^{|I|}-2$ proper subcoalitions in Section 5.3.

3.3.3 Mathematical comparison: Equal Profit Method Linear Program relative surplus maximin vs. nucleolus excess

The nucleolus allocation minimizes the maximum coalition excess $e(T, \mathbf{x})=v(T)-$ $\sum_{i \in T} x_i$ across all possible $2^{|I|}-2$ non-empty sub-coalitions:

$\min _{\mathbf{x}} \max _{T \subset I}\left(v(T)-\sum_{i \in T} x_i\right)$
(22)

While theoretically robust, the nucleolus operates on absolute monetary excess across an exponentially growing constraint set $\left(O\left(2^{|I|}\right)\right)$.

In contrast, EPML optimizes over a relative surplus excess defined at the individual partner level:

$\max _{\boldsymbol{\alpha}} g_{\min },\quad \text { s.t. } \quad g_{\min} \leq \frac{E\left[\pi_i(\boldsymbol{\alpha})\right]-\pi_i^0}{\pi_i^0},\quad \forall i \in I$
(23)

This structural difference yields two fundamental mathematical advantages:

• Dimension Reduction: The EPML reduces the constraint complexity from $O\left(2^{|I|}\right)$ to $O(|I|)$, transforming an NP-hard core optimization into a polynomial-time LP solvable in fraction-of-a-second time frames (0.32 s for EPML vs 3,210.7 s for nucleolus at $|I|=15$).

• Scale Invariance: By normalizing absolute surplus by the standalone baseline $\pi_i^0$, the EPML prevents large enterprise partners from dominating allocation shares, maintaining fairness across heterogeneous partner sizes ($\theta=0.8842$ for EPML vs $\theta=0.5821$ for nucleolus).

4. Solution Methodology

4.1 Augmented Epsilon-Constraint Method: Pareto Approximation

The tri-objective vector minimization problem is solved using the AUGMECON2 method [43], [44], which produces an approximate Pareto representation. The first objective ($Z_1$) is optimized as the primary objective, while the remaining objectives are converted into parametric inequality constraints:

$\min Z_1+\varepsilon \cdot\left(\frac{s_2}{r_2}+\frac{s_3}{r_3}\right),\quad \text { s.t. } Z_2+s_2=e_2, \quad Z_3+s_3=e_3$
(24)

AUGMECON2 produces an approximate Pareto representation whose density depends on grid Resolution, assessed using hypervolume (HV) [45], generational distance (GD) [46], and spacing [47].

4.2 Hybrid Genetic Algorithm–Linear Programming Matheuristic for Large-Scale Tri-Objective Instances
4.2.1 Architectural motivation and decomposition strategy

While exact multi-objective methods such as AUGMECON2 generate exact Pareto-optimal sets for small-to-medium instances, they encounter severe computational bottlenecks when scaled to large supply chain networks ($|I| \geq 20,|S| \geq 50$). The computational burden arises from the combinatorial explosion of first-stage binary facility activation decisions $\mathbf{y} \in\{0,1\}^{|I|}$ combined with the simultaneous evaluation of second-stage continuous flow recourse decisions $\left(\mathbf{x}^s, \mathbf{u}^s\right)$ across dense scenario trees.

To overcome this tractability boundary, we devise a hybrid matheuristic that combines a multi-objective evolutionary algorithm with exact mathematical programming subproblem resolution [48], [49], [50], exploiting the block-separable recourse structure of the two-stage stochastic model:

• Master Problem (Genetic Algorithm Search): A multi-objective genetic algorithm (GA) explores the discrete, combinatorial space of first-stage strategic coalition choices $\mathbf{y}=$ $\left[y_1, y_2, \ldots, y_{|I|}\right]^T \in\{0,1\}^{|I|}$.

• Subproblem (Exact LP Recourse): For any candidate individual $y^{(k)}$ fixed by the GA, the multi-scenario continuous recourse subproblem decouples into an exact $\operatorname{LP}\left(Q\left(y^{(k)}\right)\right)$. This LP is solved to exact optimality using an embedded high-performance solver, minimizing expected variable operating and shortfall penalty costs across scenarios to determine physical flows $x_{i j}^{s(k)}$ and shortfalls $u_j^{s(k)}$.

Tri-Objective Vector Evaluation:

Once the continuous recourse subproblem LP $Q\left(y^{(k)}\right)$ is solved to exact optimality for a given binary chromosome $y^{(k)}$, the complete tri-objective evaluation vector $Z(y^{(k)})=(Z_1(y^{(k)})$, $Z_2(y^{(k)})$, $Z_3(y^{(k)}))$ is evaluated directly from the optimal routing flows $x^{(k)}$. This master-subproblem decomposition restricts the GA search space strictly to binary topological configurations $\left(2^{|I|}\right)$, offloading continuous flow optimization to polynomial-time pure LPs, exactly as detailed in Algorithm 1.

4.2.2 Chromosome encoding, feasibility constraints, and repair mechanism

Each candidate individual in population $P_g$ at generation $g$ is represented as a binary chromosome vector $\mathbf{y}^{(k)}=\left[y_1^{(k)}, y_2^{(k)}, \ldots, y_{|I|}^{(k)}\right]^T \in\{0,1\}^{|I|}$, defining the operational facility activation state of each partner $i \in I$. The decision space explored by the GA corresponds to the $2^{|I|}$ possible facility activation configurations, subject to two strategic constraints:

• Minimum Coalition Threshold: A horizontal collaboration requires at least two active partners $\left(\sum_{i \in I} y_i^{(k)} \geq 2\right)$.

• Mandatory Core Partners: Certain strategic core hub facilities $I_{\text {core}} \subseteq I$ must remain active in all configurations $\left(\prod_{i \in I_{\text {core}}} y_i^{(k)}=1\right)$.

Whenever crossover or mutation operators violate these constraints, a dynamic feasibility repair operator $\mathcal{R}\left(\mathbf{y}^{(k)}\right)$ is executed:

$\mathcal{R}\left(\mathbf{y}^{(k)}\right)=\left\{\begin{array}{cc} y_i^{(k)}=1, & \forall i \in I_{\text {core }} \\ \mathbf{y}^{(k)} \leftarrow \text { Flip random zero-bits until } \sum_{i \in I} y_i^{(k)} \geq 2, & \text { if } \sum_{i \in I} y_i^{(k)}<2 \end{array}\right.$
(25)
4.2.3 Subproblem formulation and tri-objective vector computation

Given a fixed binary vector $\mathbf{y}^{(k)}$, the second-stage stochastic recourse subproblem evaluates the continuous optimal routing and shortfall management across all scenarios $s \in S$. The continuous subproblem model $\mathcal{Q}\left(\mathbf{y}^{(k)}\right)$ is formulated as follows:

$Q\left(\mathbf{y}^{(k)}\right)=\min _{\mathbf{x}, \mathbf{u}} \sum_{s \in S} p^s\left[\sum_{i \in I} \sum_{j \in J}\left(v_i+c_{i j}\right) x_{i j s}^{(k)}+\sum_{j \in J} p_j^{\text {pen }} u_{j s}^{(k)}\right]$
(26)

Subject to:

$\sum_{i \in I} x_{i j s}^{(k)}+u_{j s}^{(k)}=D_{j s}^s,\quad \forall j \in J, \forall s \in S$
(27)
$\sum_{j \in J} x_{i j s}^{(k)} \leq \rho_i^s \cdot {CAP}_i \cdot y_i^{(k)},\quad \forall i \in I, \forall s \in S$
(28)
$x_{i j s}^{(k)} \geq 0, \quad u_{j s}^{(k)} \geq 0,\quad \forall i \in I, \forall j \in J, \forall s \in S$
(29)

Once the exact subproblem LP is solved, the complete tri-objective evaluation vector $\mathbf{Z}(\mathbf{y}^{(k)})=[Z_1(\mathbf{y}^{(k)})$, $Z_2(\mathbf{y}^{(k)})$, $Z_3(\mathbf{y}^{(k)})]^T$ is computed:

$Z_1\left(\mathbf{y}^{(k)}\right)=\sum_{i \in I} f_i y_i^{(k)}+Q\left(\mathbf{y}^{(k)}\right)$
(30)
$Z_2\left(\mathbf{y}^{(k)}\right)=\sum_{s \in S} p^s \sum_{i \in I} \sum_{j \in J} e_{i j} x_{i j s}^{(k)}$
(31)
$Z_3\left(\mathbf{y}^{(k)}\right)=\sum_{s \in S} p^s \sum_{i \in I}\left(1-\rho_i^S\right) \sum_{j \in J} x_{i j s}^{(k)}$
(32)

where:

$Z_1$: Expected Total Network Cost (\$) incorporating first-stage fixed facility setup costs and second-stage transport, operational, and shortfall penalty costs.

$Z_2$: Expected Scope 3 Carbon Emissions (kg CO$_2$e) calculated using route-specific transport emission factors $e_{i j}$.

$Z_3$: Expected Disruption Exposure Risk metric (units) quantifying structural exposure to low-reliability nodes.

4.2.4 Genetic operators, Pareto ranking, and diversity preservation

Fast Non-Dominated Sorting and Crowding Distance:

Candidate solutions are ranked into hierarchical non-dominated Pareto fronts $\left(\mathcal{F}_1, \mathcal{F}_2, \ldots, \mathcal{F}_L\right)$ using Pareto dominance criteria [51]. Solution $\mathbf{y}^{(1)}$ dominates $\mathbf{y}^{(2)}\left(\mathbf{y}^{(1)} \succ \mathbf{y}^{(2)}\right)$ if and only if:

$Z_m\left(\mathbf{y}^{(1)}\right) \leq Z_m\left(\mathbf{y}^{(2)}\right) \forall m \in\{1,2,3\} \wedge \exists m \in\{1,2,3\}: Z_m\left(\mathbf{y}^{(1)}\right)<Z_m\left(\mathbf{y}^{(2)}\right)$
(33)

To prevent premature convergence and preserve solution diversity along the Pareto boundary, the crowding distance $d_k$ of solution $k$ in front $\mathcal{F}_l$ is evaluated across all three normalized objective spaces:

$d_k=\sum_{m=1}^3 \frac{z_m^{(k+1)}-z_m^{(k-1)}}{z_m^{\max}-z_m^{\min}}$
(34)

Binary Tournament Selection:

Two candidates are drawn uniformly at random [51]. Solution $\mathbf{y}^{(1)}$ is selected if its Pareto rank is strictly superior $\left(\operatorname{Rank}\left(\mathbf{y}^{(1)}\right)<\operatorname{Rank}\left(\mathbf{y}^{(2)}\right)\right)$, or if $\operatorname{Rank}\left(\mathbf{y}^{(1)}\right)=$ $\operatorname{Rank}\left(\mathbf{y}^{(2)}\right)$ and $d_1>d_2$.

Crossover Operator:

Two-point crossover is applied with a recombination probability of $p_c=0.85$. Two distinct pivot cut indices $k_1, k_2 \in\{1, \ldots,|I|-1\}$ (with $k_1

Bit-Flip Mutation:

Bit-flip mutation is executed with probability $p_m$ per locus $i \in I$:

$y_i^{(k)} \leftarrow 1-y_i^{(k)} \text { if } U(0,1) \leq p_m$
(35)

Immediately following mutation, the feasibility repair operator $\mathcal{R}\left(\mathbf{y}^{(k)}\right)$ is invoked to guarantee $\sum_{i \in I} y_i^{(k)} \geq 2$.

Elitism and Archiving:

An external archive $\mathcal{A}$ stores all non-dominated physical solutions found across generations. Generational replacement combines parent and offspring populations $P_{g-1} \cup P_g^{\prime}$ (size $2 N_{\text {pop}}$), selecting the top $N_{\text {pop}}$ individuals using rank and crowding distance to ensure strict elitism.

GA Parameter Calibration and Algorithmic Settings:

To ensure full computational reproducibility of the Level 1 multi-objective search, the complete hyperparameter configuration governing the execution of the hybrid GA-LP matheuristic is summarized in Table 3. These parameter settings—including population size, selection pressure, crossover and mutation probabilities, and dual termination criteria—were systematically calibrated through preliminary sensitivity runs on benchmark instances to balance thorough exploration of the $2^{|I|}$ binary activation space with rapid computational convergence.

Table 3. Genetic algorithm hyperparameter configuration.
ParameterValue/SettingDescription
Population size $(N_{\mathrm{pop}})$100Number of binary chromosomes per generation
Max generations $(G_{\max})$200Maximum generation threshold
Selection operatorTournament $(k=3)$Tournament selection size
Elitism rate5%Top 5% chromosomes preserved intact
Crossover operatorTwo-point crossoverRecombination probability $P_c=0.85$
Mutation operatorBit-flip mutationPer-gene mutation probability $P_m=0.05$
Termination criteriaDual condition$G_{\max}=200$ OR 30 consecutive generations with relative hypervolume improvement $\Delta HV < 10^{-4}$
4.2.5 Formal algorithm pseudocode

The complete procedural execution of the hybrid GA-LP Matheuristic is detailed in Algorithm 1.

4.2.6 Algorithmic interface with Level 2 Equal Profit Method linear program gain-sharing engine

The hierarchical coupling between Level 1 physical optimization and Level 2 financial gain distribution operates as a unidirectional post-Pareto selection process. Once the matheuristic extracts the physical Pareto front $\left(Z_1, Z_2, Z_3\right)$, each non-dominated operational configuration passes its scenario-dependent total operating costs $\left(C_{\text {total }}^S\right)$ as fixed parameters directly into the Level 2 EPML model. The LP engine then solves for the optimal cost-allocation fractions $\left(\alpha_i^s\right)$ by maximizing the uniform relative surplus parameter $\left(g_i\right)$. This sequential execution provides two structural guarantees:

• Computational Decoupling: Physical flow decisions $\left(x_{i j}^S\right)$ are optimized independently of allocation fractions ($\alpha_i^s$), eliminating non-convex bilinear terms $\left(\alpha_i^s \cdot x_{i j}^s\right)$ and preserving LP subproblem convexity.

• Conditional Individual Rationality: Financial redistribution is computed in polynomial time (0.32 s for 15 partners and 1.87 s for 30 partners). Provided that the Level 2 LP is mathematically feasible-which requires the selected Level 1 physical trade-off solution to yield positive net economic synergy relative to standalone baselines ($\sum_{i \in I} \mathbb{E}\left[\pi_i\right] \geq$ $\sum_{i \in I} \pi_i^0$)-the EPML engine guarantees that every partner receives an expected collaborative profit meeting or exceeding its standalone baseline ($\mathbb{E}\left[\pi_i\right] \geq \pi_i^0$). If a candidate physical solution fails to generate net economic savings over non-cooperative baselines, the Level 2 LP yields $\left(g_{\min }^*<0\right)$ (or infeasibility), explicitly signaling to decision-makers that an individually rational gain distribution cannot be achieved under that specific physical routing configuration.

4.3 Scenario Generation and Reduction via Gaussian Copula and $k$-means Clustering
4.3.1 Stochastic uncertainty characterization and dependency motivation

In realistic collaborative logistics networks, operational risks do not occur in isolation. Facility disruptions caused by unexpected capacity breakdowns, regional equipment failures, or extreme weather events are often correlated across geographical zones and partner facilities. Simultaneously, market demand fluctuations across destination markets $j \in J$ exhibit cross-correlations driven by macroeconomic trends. Assuming independence between facility capacity losses $\delta_i^s$ and regional market demands $D_{j s}^s$ severely underestimates joint risk exposure, leading to brittle network configurations [52].

To capture non-linear dependence structures and tail dependencies without violating marginal distribution bounds, we deploy a hybrid stochastic scenario generation framework combining a multivariate Gaussian copula, LHS, and $k$-means Cluster-Based Scenario Reduction.

4.3.2 Gaussian copula dependence modeling and marginal distributions

Let $\mathbf{U}=\left(U_1, U_2, \ldots, U_m\right)^T \in[ 0,1]^m$ be a vector of uniform random variables on the hypercube $[ 0,1]^m$, where $m=|I|+|J|$ represents the total number of stochastic parameters (comprising $|I|$ partner facility disruption severities and $|J|$ market demand realizations).

According to Sklar’s Theorem [53], any multivariate joint cumulative distribution function (CDF) $F(\mathbf{z})=F\left(z_1, z_2, \ldots, z_m\right)$ can be decomposed into its continuous marginal CDFs $F_k\left(z_k\right)$ and a copula $C$ that captures the inter-variable dependence structure:

$F\left(z_1, z_2, \ldots, z_m\right)=C\left(F_1\left(z_1\right), F_2\left(z_2\right), \ldots, F_m\left(z_m\right)\right)$
(36)

We employ a multivariate Gaussian copula $C_{\mathbf{R}}$, parameterized by a symmetric, positive semi-definite rank correlation matrix $\mathbf{R} \in[-1,1]^{m \times m}$ calibrated from historical operational logs and market data [22]:

$ C_{\mathbf{R}}\left(u_1, u_2, \ldots, u_m\right)=\Phi_{\mathbf{R}}\left(\Phi^{-1}\left(u_1\right), \Phi^{-1}\left(u_2\right), \ldots, \Phi^{-1}\left(u_m\right)\right) $

where, $\Phi^{-1}(\cdot)$ denotes the inverse CDF of the standard univariate normal distribution $\mathcal{N}(0,1)$, and $\Phi_{\mathbf{R}}(\cdot)$ is the joint CDF of an $m$-dimensional multivariate normal distribution $\mathcal{N}(\mathbf{0}, \mathbf{R})$ with zero mean and covariance matrix equal to correlation matrix R.

Marginal Disruption Severity Distributions

Facility disruption severities $\delta_i \in[ 0,1]$ represent the fractional loss of operational capacity at facility $i \in I$. They are modeled using calibrated Beta distributions $\operatorname{Beta}\left(a_i, b_i\right)$ bounded on [0,1], which flexibly capture right-skewed, low-probability high-consequence disruption profiles:

$F_i\left(\delta_i; a_i, b_i\right)=\frac{I_{\delta_i}\left(a_i, b_i\right)}{B\left(a_i, b_i\right)}, \quad \delta_i \in[ 0,1], \forall i \in I$
(37)

where, $B\left(a_i, b_i\right)$ is the Beta function and $I_{\delta_i}\left(a_i, b_i\right)$ is the regularized incomplete Beta function. The effective scenario-dependent facility reliability $\rho_i^s$ is subsequently derived as:

$\rho_i^s=\rho_i \cdot\left(1-\delta_i^s\right) \quad \forall i \in I, \forall s \in S$
(38)

where, $\rho_i \in(0,1]$ represents the nominal baseline reliability of facility $i$.

Marginal Market Demand Distributions

Market demands $D_j$ for destination node $j \in J$ are modeled using continuous truncated normal distributions $\mathcal{T} \mathcal{N}\left(\mu_j, \sigma_j^2, a_j^{\text {min}}, b_j^{\text {max}}\right)$ to prevent negative demand realizations:

$F_{|I|+j}\left(D_j\right)=\frac{\Phi\left(\frac{D_j-\mu_j}{\sigma_j}\right)-\Phi\left(\frac{a_j^{\min }-\mu_j}{\sigma_j}\right)}{\Phi\left(\frac{b_j^{\max }-\mu_j}{\sigma_j}\right)-\Phi\left(\frac{a_j^{\min }-\mu_j}{\sigma_j}\right)}, \quad D_j \in\left[a_j^{\min}, b_j^{\max}\right], \forall j \in J$
(39)
4.3.3 Latin hypercube sampling

To generate an unreduced raw sample set $\Omega_{\text {raw}}=\left\{\boldsymbol{\xi}^1, \boldsymbol{\xi}^2, \ldots, \boldsymbol{\xi}^{N_{\text {raw}}}\right\}$ of size $N_{\text {raw}}$ (typically $N_{\text {raw}}=10,000$) while guaranteeing uniform coverage across all marginal distribution tails, LHS is executed [54]:

• The range [0,1] of each marginal CDF $u_k(k=1, \ldots, m)$ is stratified into $N_{\text {raw}}$ equiprobable non-overlapping intervals $\left[\frac{l-1}{N_{\text {raw}}}, \frac{l}{N_{\text {raw}}}\right]$ for $l=1, \ldots, N_{\text {raw}}$.

• One random value $u_{k, l}$ is sampled uniformly within each interval $l$.

• Correlated multivariate standard normal vector realizations $\mathbf{Z}^{(l)}=$ $\left[Z_1^{(l)}, \ldots, Z_m^{(l)}\right]^T \sim \mathcal{N}(\mathbf{0}, \mathbf{R})$ are constructed via Cholesky decomposition $\mathbf{R}=$ $\mathbf{L L}^T$:

$\mathbf{Z}^{(l)}=\mathbf{L} \cdot \mathbf{W}^{(l)}$
(40)

where, $\mathbf{W}^{(l)} \sim \mathcal{N}\left(\mathbf{0}, \mathbf{I}_m\right)$ is an independent standard normal sample vector.

Uniform realizations $u_k^{(l)}=\Phi\left(Z_k^{(l)}\right)$ are transformed back into physical stochastic vector realizations $\boldsymbol{\xi}^l=\left[\delta_1^l, \ldots, \delta_{|I|}^l, D_1^l, \ldots, D_{|J|}^l\right]^T$ using the inverse marginal CDFs:

$\delta_i^l=F_i^{-1}\left(\Phi\left(Z_i^{(l)}\right)\right), \quad \forall i \in I$
(41)
$D_j^l=F_{|I|+j}^{-1}\left(\Phi\left(Z_{|I|+j}^{(l)}\right)\right), \quad \forall j \in J$
(42)
4.3.4 Mathematical formulation of $k$-means scenario reduction

Solving the stochastic Level 1 MILP directly over the large raw sample set $\left|\Omega_{\text {raw}}\right|=$ 10,000 is computationally intractable. To obtain a compact, highly representative scenario set $S=\{1,2, \ldots,|S|\}$ (where $|S| \ll N_{\text {raw}}$, e.g., $|S|=50$), a $k$-means clustering reduction algorithm is applied to partition $\Omega_{\text {raw}}$ into $|S|$ disjoint clusters $\left\{\mathcal{S}_1, \mathcal{S}_2, \ldots, \mathcal{S}_{|S|}\right\}$.

Optimization Objective (Sum of Squared Errors Minimization)

The scenario reduction procedure minimizes the within-cluster sum of squared Euclidean distances, providing a computationally efficient vector quantization proxy closely related to continuous 2-Wasserstein (Monge–Kantorovich) distance minimization [55]:

$\min _{\mathcal{S}_1, \ldots, \mathcal{S}_{|S|}, \boldsymbol{\mu}^1, \ldots, \boldsymbol{\mu}^{|S|}} \sum_{s=1}^{|S|} \sum_{\xi \in \mathcal{S}_s}\left\|\xi-\boldsymbol{\mu}^s\right\|_2^2$
(43)

Subject to:

$\bigcup_{s=1}^{|S|} \mathcal{S}_s=\Omega_{\text {raw }}, \quad \mathcal{S}_s \cap \mathcal{S}_r=\emptyset, \quad \forall s \neq r$
(44)
$\boldsymbol{\mu}^s=\frac{1}{\left|\delta_s\right|} \sum_{\xi \in \mathcal{S}_s} \boldsymbol{\xi}, \quad \forall s \in\{1, \ldots,|S|\}$
(45)

where, $\boldsymbol{\mu}^s=\left[\delta_1^s, \ldots, \delta_{|I|}^s, D_1^s, \ldots, D_{|J|}^s\right]^T \in \mathbb{R}^m$ is the computed centroid vector of cluster $\mathcal{S}_s$, serving as the discrete physical values for representative scenario $s \in S$.

Discrete Probability Mass Assignment

The continuous probability weight $p^s$ assigned to reduced representative scenario $s \in$ $S$ is computed in proportion to the cardinality of its corresponding cluster $\mathcal{S}_S$:

$p^s=\frac{\left|S_s\right|}{N_{\mathrm{raw}}}, \quad \forall s \in\{1, \ldots,|S|\}$
(46)

This probability construction guarantees strict probability conservation across the discrete scenario space:

$\sum_{s=1}^{|S|} p^s=\sum_{s=1}^{|S|} \frac{\left|\mathcal{S}_s\right|}{N_{\mathrm{raw}}}=\frac{1}{N_{\mathrm{raw}}} \sum_{s=1}^{|S|}\left|\mathcal{S}_s\right|=\frac{N_{\mathrm{raw}}}{N_{\mathrm{raw}}}=1$
(47)
4.3.5 Algorithmic implementation of scenario generation and reduction

The dynamic generation and reduction pipeline is summarized in Algorithm 2.

5. Computational Validation and Empirical Results

5.1 Experimental Setup and Benchmark Test Instances

To evaluate the computational efficiency, solution quality, and game-theoretic equity of the proposed two-stage stochastic collaborative optimization framework, extensive empirical experiments were conducted across synthetic and realistic supply chain network benchmarks.

5.1.1 Instance generation and parameter calibration

Test instances are categorized into three scale tiers based on the number of participating supply chain partners $|I|$, destination market nodes $|J|$, and reduced scenario tree cardinality $|S|$:

• Small-Scale Networks $(|I|=5,|J|=4,|S|=20)$: Used for exact Pareto validation comparing AUGMECON2 against the hybrid GA-LP Matheuristic, as well as full game-theoretic benchmarking against the exact Shapley value and nucleolus.

• Medium-Scale Networks $(|I|=15,|J|=8,|S|=50)$: Representative of regional industrial coalitions. Used as the core baseline instance for sensitivity analyses and gain-sharing performance metrics.

• Large-Scale Networks $(|I|=30$ to 50, $|J|=15$ to 25, $|S|=100)$: Representative of international cross-border supply chain alliances, deployed to stress-test computational scalability and real-time execution capabilities.

Stochastic scenarios are generated via the Gaussian copula-LHS engine (Algorithm 2) with raw sample size $N_{\text {raw}}=10,000$, reduced to $|S|$ representative scenarios using $k$-means clustering. Facility capacities $CAP_i$ are drawn from $\mathcal{U}(10,000,25,000)$ units, fixed facility operating costs $f_i$ from $\mathcal{U}(\$ 50,000, \$ 120,000)$, unit processing costs $v_i$ from $\mathcal{U}(\$ 12, \$ 25)$, and unit transport costs $c_{i j}$ from $\mathcal{U}(\$ 2.5, \$ 8.0) /$ unit. Route-specific Scope 3 carbon emission factors $e_{i j}$ were extracted from the databases ecoinvent v3.9 [39], GREET 2023 [40], and UK DEFRA 2024 [41]. Aligned with the GHG Protocol [42], the optimization boundary is strictly limited to logistics-focused Scope 3 Categories 1 and 4. The applied unified factors range from approximately 0.045 kg CO$_2$e/t$\cdot$km (rail/multimodal) to 0.162 kg CO$_2$e/t$\cdot$km (road transport). Selling prices $p_j^{\text {sell}}$ are set to $\mathcal{U}$(\$85, \$130)/unit, and the base shortfall penalty parameter $p_j^{\text {pen}}$ is calibrated at $2.5 \times p_j^{\text {sell}}$.

To ensure strict computational reproducibility of all numerical experiments, the detailed specifications governing the small, medium, and large test instances—including network dimensions, underlying probability distribution parameters, correlation structures, cost/capacity ranges, and execution random seeds—are consolidated in Table 4.

Table 4. Synthetic instance generation: Parameter categories, distributions, and numerical settings

Parameter Category

Specification/Distribution

Parameter Ranges and Numerical Settings

Network dimensions $(|I|,|J|,|S|)$

Small/medium/large tiers

Small: $|I|=5,|J|=4,|S|=20$; medium: $|I|=15,|J|=8,|S|=50$; large: $|I|=30$–$50,|J|=15$–$25,|S|=100$

Market demand $(D_j^{s})$

(TN) $(\mu_j,\sigma_j^2,a_j^{\min},b_j^{\max})$

$\mu_j\in[ 500,2000]$ units; $\sigma_j=0.20\mu_j$; $a_j^{\min}=0.2\mu_j$; $b_j^{\max}=2.0\mu_j$

Disruption severity $(\delta_i^{s})$

Beta distribution $\mathrm{Beta}(a_i,b_i)$ on $[ 0,1]$

$\mathbb{E}[\delta_i^{s}]=0.20$ with $(a_i,b_i)=(2.0,8.0)$; baseline reliability $\rho_i\in[ 0.85,0.98]$

Dependency structure

Gaussian copula; $R\in[-1,1]^{m\times m}$

Off-diagonal correlations $R_{ij}\in[ 0.25,0.55]$; $m=|I|+|J|$

Facility capacity $(CAP_i)$

Uniform $\mathcal{U}(a,b)$

$CAP_i\sim U(10{,}000,25{,}000)$ units

Fixed setup cost $(f_i)$

Uniform $U(a,b)$

$f_i\sim U(\$50{,}000,\$120{,}000)$

Processing cost $(v_i)$

Uniform $U(a,b)$

$v_i\sim U$(\$12.0,\$25.0)/unit

Transport cost $(c_{ij})$

Uniform $U(a,b)$

$c_{ij}\sim U(\$2.5,\$8.0)$/unit

Scope 3 factor $(e_{ij})$

Calibrated empirical factors

$e_{ij}\in[ 0.045,0.162]$ kg CO$_2$e/(t$\cdot$km) (ecoinvent v3.9/DEFRA 2024)

Selling price $(p_j^{\mathrm{sell}})$

Uniform $U(a,b)$

$p_j^{\mathrm{sell}}\sim$ U(\$85,\$130)/unit

Shortfall penalty $(p_j^{\mathrm{pen}})$

Calibrated proportional bound

$p_j^{\mathrm{pen}}=2.5\,p_j^{\mathrm{sell}}\in[\$212.5,\$325.0]$/unit

Random seeds & environment

Fixed pseudo-random seeds

NumPy seed = 42; CPLEX run via its C++ API in deterministic mode (Threads = 1)

TN = truncated normal; CO$_2$e = carbon dioxide equivalent; API = application programming interface.
5.1.2 Hardware and software environment

All mathematical models and algorithms were implemented in Python 3.11 utilizing ILOG CPLEX v22.1.0 as the underlying exact MIP/LP solver via the C++ API wrapper. Experiments were executed on an Intel Xeon Enterprise Workstation equipped with 16 cores @ 3.40 GHz and 64 GB RAM.

5.2 Multi-Objective Pareto Frontier Analysis

The Level 1 model resolves trade-offs across three conflicting objectives (Table 5): Expected System Cost ($Z_1$), Expected Scope 3 Carbon Emissions ($Z_2$), and Expected Disruption Risk Exposure ($Z_3$).

Table 5. Representative Pareto compromise solutions (medium instance $|I|=15$).
Pareto Solution PointTotal Cost $Z_1$ (\$K)Scope 3 Emissions $Z_2$ (tCO$_2$e)Disruption Risk $Z_3$ (Penalty Units)MAC (\$/tCO$_2$e)
Point A (Cost Min)4,386.5387.424,8000.00
Point B (Carbon Min)4,680.1301.28,9503,406.03
Point C (Compromise)$^{*}$4,492.0341.918,9202,318.68
Point D (Resilience Min)4,585.2365.15,1008,910.31
Note: MAC = marginal abatement cost; tCO$_2$e = metric tons of carbon dioxide equivalent. $^{*}$ marks Point C, the baseline solution for Level 2 financial allocation.

To quantitatively evaluate the economic efficiency of decarbonization across the non-dominated Pareto frontier, we compute the marginal abatement cost (MAC) for each Pareto compromise solution $p \in \Omega_{\text {Pareto}}$, denoted by $\mathrm{MAC}_p$, relative to the unconstrained cost-minimizing baseline (Point A). The $\mathrm{MAC}_p$ (expressed in \$/tCO2e) measures the additional operational expenditure required to avoid one metric ton of Scope 3 carbon emissions:

$\operatorname{MAC}_p=\frac{Z_1(p)-Z_1(A)}{\left(Z_2(A)-Z_2(p)\right)} * 1,000, \quad \forall p \in \Omega_{\text {Pareto }} \backslash\{A\}$
(48)

where,

• $Z_1^A$ and $Z_2^A$ represent the Expected Total Operational Cost (\$K) and Scope 3 Carbon Emissions (metric ton of carbon dioxide equivalent [tCO$_2$e]) of the economic baseline (Point A), respectively.

• $Z_1(p)$ and $Z_2(p)$ represent the Expected Total Operational Cost and Scope 3 Carbon Emissions of Pareto solution point $p$.

• The multiplication factor of 1,000 in the numerator converts the total operational cost difference from thousands of dollars (\$K) into single United States dollars (\$), ensuring that the resulting MAC is directly expressed in dollars per metric ton of carbon dioxide equivalent (\$/tCO$_2$e).

A low MAC$_p$ indicates highly cost-effective carbon reduction opportunities, whereas a high MAC$_p$ signals diminishing economic returns on decarbonization investments along the Pareto frontier.

At the selected physical compromise solution (Point C), total system cost increases by \$105,500 (from \$4,386,500 at Point A to \$4,492,000) while Scope 3 emissions decrease by 45.5 metric tons of CO2e (from 387.4 tCO2e to 341.9 tCO2e). This yields a corrected MAC of \$2,318.68/tCO2e relative to the cost-minimizing baseline (Point A). Pushing decarbonization further to the minimum carbon solution (Point B) requires an additional expenditure of \$293,600 to achieve an 86.2 tCO2e reduction, resulting in a higher MAC of \$3,406.03/tCO2e, reflecting diminishing economic returns on carbon abatement investments.

5.2.1 Analysis of physical trade-offs

Figure 1 provides a comprehensive visual representation of the multi-objective Pareto frontier generated by the proposed Level 1 stochastic optimization model for the medium-scale benchmark instance ($|I|$ = 15, $|J|$ = 8, $|S|$ = 50).

As depicted in Figure 1a, the 3D non-dominated surface exhibits a well-distributed non-dominated trade-off surface across all three competing operational objectives: Total Expected Cost $Z_1$, Scope 3 Carbon Emissions $Z_2$, and Disruption Risk Exposure $Z_3$. This confirms that the AUGMECON2 and hybrid GA-LP algorithms effectively explore the extreme non-dominated frontiers without getting trapped in local optima or producing dominated solution clusters.

Figure 1. Multi-objective Pareto frontier
Note: tCO$_2$e = metric ton of carbon dioxide equivalent; MAC = marginal abatement cost. Results are for the medium instance (\(|I| = 15\), \(|J| = 8\), \(|S| = 50\)).

To further examine the marginal rate of substitution between economics and sustainability, Figure 1b projects the 2D Cost–Emissions trade-off curve ($Z_1$ vs. $Z_2$), where candidate benchmark solutions A–D are identified by distinct markers along with their annotated MAC values. Four critical operational benchmark points are identified:

• Point A (Economic Extremum): Prioritizes purely financial performance ($Z_1$ = \$4,386.5 K), routing physical flows strictly along the shortest geographic pathways. As shown in Figure 1b at the zero-baseline position (MAC = \$0.00/tCO$_2$e), Point A suffers from severe vulnerability to facility breakdowns ($Z_3$ = 24,800 units) and generates the highest Scope 3 footprint ($Z_2$ = 387.4 tCO$_2$e)

• Point B (Environmental Extremum): Achieves a 22.25% reduction in Scope 3 emissions ($Z_2$ = 301.2 tCO$_2$e) by re-routing freight toward electrified multi-modal corridors. Nonetheless, the steep flattening of the Pareto curve near Point B demonstrates diminishing returns: securing these final marginal emission reductions requires a substantial operational cost increase ($Z_1$ = \$4,680.1 K), yielding a high MAC of \$3,406.03/tCO$_2$e.

• Point C (Technique for Order Preference by Similarity to Ideal Solution [TOPSIS] Knee-Point Compromise): Located precisely at the maximum curvature (“knee”) of the Pareto frontier, Point C captures an optimal balance among all three criteria ($Z_1$ = \$4,492.0 K, $Z_2$ = 341.9 tCO$_2$e, $Z_3$ = 18,920 units). It achieves an 11.7% reduction in carbon emissions relative to Point A at a modest cost premium of only 2.40%, corresponding to a highly economical MAC of \$2,318.68/tCO$_2$e. Consequently, Point C is selected as the baseline physical configuration for the subsequent Level 2 financial allocation.

• Point D (Resilience Extremum): Minimizes disruption risk ($Z_3$ = 5,100 units) by bypassing cost-effective yet unreliable partner facilities, shifting volume toward highly dependable hubs at a higher baseline operating cost ($Z_1$ = \$4,585.2 K). Because routing decisions prioritize risk mitigation over emission efficiency, Point D exhibits the highest marginal carbon penalty (MAC = \$8,910.31/tCO$_2$e).

• Pairwise Trade-off Projections: The pairwise projections in Figure 1c and Figure 1d further elucidate the underlying trade-off dynamics across resilience and sustainability dimensions. Specifically, Figure 1c demonstrates the non-linear relationship between expected total cost ($Z_1$) and disruption risk exposure ($Z_3$). Transitioning from Point A to Point D yields a 79.44% reduction in disruption risk (from 24,800 to 5,100 units) for a modest 4.53% increase in operational cost (to \$4,585.2 K). The convex profile illustrates diminishing marginal returns on resilience, confirming that eliminating residual tail-risk requires maintaining expensive redundant capacity. Conversely, Figure 1d highlights a structural tension between decarbonization ($Z_2$) and network resilience ($Z_3$). Achieving maximum resilience at Point D increases Scope 3 emissions to 365.1 tCO$_2$e (an increase of 63.9 tCO$_2$e or +21.21% over Point B), as risk-averse routing bypasses primary efficiency-focused hubs in favor of longer, secondary backup logistics channels.

5.2.2 Algorithmic convergence: Comparison of the augmented epsilon-constraint method and the hybrid genetic algorithm–linear programming matheuristic

To rigorously evaluate the convergence performance of the hybrid GA-LP matheuristic against exact AUGMECON2, we report key multi-objective performance indicators across network scale tiers in Table 6, including the number of non-dominated solutions (P*), normalized hypervolume (HV), Spacing metric (S), and total central processing unit (CPU) execution time.

Table 6. Algorithmic performance and convergence comparison: Exact comparison of the augmented epsilon-constraint method and the hybrid genetic algorithm–linear programming

Instance Size ($|I|$)

Method

Non-Dominated Solutions ($|P^*|$)

HV

S

CPU Time (s)

Relative Hypervolume HV Gap (%)

Small ($|I|=5$)

AUGMECON2

42

0.942

0.012

14.2

Baseline

Hybrid GA-LP

40

0.938

0.015

3.1

$-0.42\%$

Medium ($|I|=15$)

AUGMECON2

118

0.885

0.018

485.6

Baseline

Hybrid GA-LP

112

0.879

0.021

18.4

$-0.68\%$

Large ($|I|=30$)

AUGMECON2

Memory limit/timeout

N/A

N/A

$>$3,600

N/A

Hybrid GA-LP

184

0.841

0.026

82.5

N/A

Note: AUGMECON2 = augmented epsilon-constraint method; CPU = central processing unit; GA–LP = genetic algorithm–linear programming; HV = hypervolume; N/A = not applicable; S = spacing metric. HV is normalized relative to a reference nadir point. S measures the uniformity of solution spread along the Pareto front (lower is better).

As demonstrated in Table 6, the hybrid GA-LP captures over 99.3% of the exact AUGMECON2 hypervolume for medium instances (HV = 0.879 vs. 0.885) while reducing computation time by 96.2% (18.4 s vs. 485.6 s), successfully scaling to large networks ($|I|$ = 30) where exact methods exceed computational limits.

5.3 Gain-Sharing Performance Metrics

Once the physical baseline solution at Point C is selected, the Level 2 EPML allocation model determines the scenario-wise cost fractions allocated to each partner $i \in N=$ $\{1, \ldots,|I|\}$. To evaluate the performance of EPML, we benchmark it against three classical game-theoretic allocation mechanisms (Table 7):

• Proportional allocation (PROP) [56]: Allocates total collaborative cost $c(N)$ strictly in proportion to baseline standalone costs $c(\{i\})$:

$x_i^{\mathrm{PROP}}=\frac{c(\{i\})}{\sum_{j \in N} c(\{j\})} \cdot c(N), \quad \forall i \in N$
(49)

• Shapley value (SHAP) [57]: Computes individual savings based on a partner’s expected marginal contribution across all possible sub-coalitions $S \subseteq N \backslash\{i\}$:

$\phi_i(v)=\sum_{S \subseteq N \backslash\{i\}} \frac{|S|!(|N|-|S|-1)!}{|N|!}[v(S \cup\{i\})-v(S)], \quad \forall i \in N$
(50)

• Nucleolus (NUC) [58]: Lexicographically minimizes the maximum sub-coalition excess $e(S, \mathbf{x})=v(S)-$ $\sum_{j \in S}\left(c(\{j\})-x_j\right)$, ensuring maximum coalition stability within the core.

Table 7. Gain-sharing allocation performance comparison ($|I|=15$ partners)
Allocation MechanismEquity Ratio ($\boldsymbol{\theta}$)Min Partner Gain ($\boldsymbol{g_{\min}}$)Max Partner Gain ($\boldsymbol{g_{\max}}$)Core Stability
Proportional rule0.1120+2.10%+42.80%Violated
Shapley value0.5481+8.45%+38.12%Satisfied
Nucleolus0.5821+9.80%+36.40%Satisfied
Proposed EPML engine0.8842+25.06%+25.66%Satisfied
Note: EPML = Equal Profit Method linear program
5.3.1 Equity metric definition

To quantify fairness across heterogeneous partners, we define the Gini-based equity ratio $\theta \in$ $[ 0,1]$ using relative profit gain percentages $g_i=\frac{c(\{i\})-x_i}{c(\{i\})}$:

$\theta=1-\frac{\sum_{i \in N} \sum_{j \in N}\left|g_i-g_j\right|}{2|N| \sum_{i \in N} g_i}$
(51)

An equity ratio of $\theta=1$ indicates perfect equity, where every partner achieves an identical relative profit gain percentage ($g_i=g_j, \forall i, j$) over their non-cooperative baseline.

5.3.2 Sub-coalition core stability verification procedure

While Theorem 1 analytically guarantees expected individual rationality against individual defection $(T=\{i\})$, whenever the selected physical configuration satisfies the net-synergy feasibility condition stated in Section $3.2\left(\sum_i E\left[\pi_i\right] \geq \sum_i \pi_i^0\right)$, assessing core stability against arbitrary proper sub-coalitions $(|T| \geq 2)$ requires empirical evaluation over candidate network allocations. To verify core stability empirically for the $|I|=15$ benchmark network, we implemented an exact sub-coalition separation algorithm. All $2^{15}-2=32,766$ proper non-empty sub-coalitions $T \subset I$ were systematically enumerated. For each sub-coalition $T$, we evaluated its characteristic value $v(T)$ by solving the Level 1 stochastic MILP restricted strictly to members $i \in T$. We then computed the maximum coalitional excess $e_{\max}(x)$ across all sub-coalitions:

$e_{\max}(x)=\max _{T \subset I, T \neq \emptyset}\left(v(T)-\sum_{i \in T} \mathbb{E}\left[\pi_i(x)\right]\right)$
(52)

A gain-sharing allocation $x$ satisfies core stability if and only if $e_{\max }(x) \leq 0$ (meaning no sub-coalition $T$ can achieve a higher collective profit by defecting from the Grand Coalition). In Table 7, the proportional rule yielded $e_{\max}=+\$ 142,500>0$, violating core stability. Conversely, EPML, nucleolus, and Shapley value all achieved $e_{\max} \leq 0$, empirically confirming core stability for the benchmark instance.

To visually validate the equity performance across heterogeneous network participants, Figure 2 displays the relative profit gain percentage $g_i$ achieved by each of the 15 partners ($P_1$ through $P_{15}$), ordered along the horizontal axis by their standalone baseline size (from smallest regional operator $P_1$ to largest central hub $P_{15}$).

Figure 2. Partner profit gain distribution across allocation mechanisms ($|I|=15$)
Note: EPML = Equal Profit Method linear program. Comparative distribution of relative profit gain percentages $\left(g_i=\frac{\mathbb{E}[\pi_i]-\pi_i^0}{\pi_i^0}\right)$ across all 15 collaborating partners, ordered from smallest baseline operating cost (P$_1$) to largest (P$_{15}$). The graph highlights the inequality under the proportional rule, the structural bias of absolute payout methods (Shapley value and nucleolus), and the near-flat maximin equity achieved by the proposed EPML engine ($g_i \approx$ +25.36%, $\theta$ = 0.8842).
5.3.3 Comparative analysis and key equity findings

The numerical metrics in Table 7 and the partner-level distribution in Figure 2 reveal fundamental insights regarding allocation fairness and coalition stability:

• Distributive Distortion under Proportional Rules: As evidenced in Table 7, the proportional rule yields a very low equity ratio $(\theta=0.1120)$ and severe gain dispersion $g_{\min}$= +2.10% vs. $g_{\max}$ = +42.80%. As illustrated by the steep orange bars in Figure 2, small regional carriers ($P_1–P_3$) receive marginal relative profit increases, whereas large tier-1 hubs ($P_{13}–P_{15}$) absorb the majority of benefits. This extreme disparity violates core stability constraints, creating strong financial incentives for smaller firms to defect from the coalition.

• Absolute vs. Relative Equity (Shapley value and nucleolus): Although the Shapley value ($\theta$ = 0.5481) and nucleolus ($\theta$ = 0.5821) satisfy core stability, both classic mechanisms suffer from structural equity limitations. Because they evaluate marginal contributions in absolute monetary terms, larger partners with high baseline operational volumes capture disproportionate cash flows. As shown in Figure 2, small partners achieve relative gains below +10%, while large hubs secure over +36%.

• Equitable Maximin Uniformity under EPML: In contrast, the proposed EPML engine achieves a near-perfect equity ratio of $\theta$ = 0.8842. By maximizing the minimum relative profit margin increase across all scenarios, EPML produces a virtually flat distribution in Figure 2, where relative gains for all 15 partners are tightly bounded between +25.06% and +25.66%. Because compromise Point C satisfies the net-synergy feasibility condition stated in Section 3.2 ($\sum_i E\left[\pi_i\right] \geq \sum_i \pi_i^0$), this guarantees expected individual rationality under Theorem 1 ($g_{\text {min}}{ }^*>$ 0), achieves empirically verified core stability across all realization scenarios, and eliminates defection incentives across heterogeneous network participants.

5.4 Computational Scalability and Runtime Benchmarks

To evaluate the operational viability and computational efficiency of the proposed framework in real-world supply chain deployments, we conduct a scalability benchmark across varying network sizes $|I| \in\{5,10,15,20,30\}$. Figure 3 illustrates the computational runtime progression (on a logarithmic scale) directly benchmarking the exact execution metrics reported in Table 8.

Figure 3. Computational runtime scalability comparison
Note: EPML = Equal Profit Method linear program. Computational execution times (in seconds, on a logarithmic scale) across varying network sizes $|I| \in \{5, 10, 15, 20, 30\}$ are based on the metrics in Table 8. Classical game-theoretic methods (Shapley value and nucleolus) exhibit exponential complexity growth ($O(2^{|I|})$), reaching the 24-hour limit (86,400 s) or becoming intractable at $|I| = 30$. In contrast, the proposed EPML engine scales polynomially O(poly), completing execution in 1.87 s at $|I| = 30$.
Table 8. Computational scalability benchmarks across allocation methods
Partner Count ($\boldsymbol{|I|}$)Scenario Count ($\boldsymbol{|S|}$)Shapley Value Time (s)Nucleolus Time (s)EPML Engine Time (s)
$|I|=5$$|S|=20$0.120.450.04
$|I|=10$$|S|=30$14.8082.100.12
$|I|=15$$|S|=50$1,840.503,210.700.32
$|I|=20$$|S|=50$86,400.00*48,120.000.78
$|I|=30$$|S|=100$IntractableIntractable1.87
Note: EPML = Equal Profit Method Linear Program. *Execution hit the 24-hour time-limit threshold.
5.4.1 Scalability analysis

As depicted in Figure 3, classical cooperative game methods suffer from catastrophic exponential complexity expansion $\left(O\left(2^{|I|}\right)\right)$. While all mechanisms execute rapidly on small networks ($|I|=5$, with EPML taking only 0.04 s), the computational burden of classical methods rises steeply as $|I|$ increases. At $|I|=10$, the nucleolus solver’s execution time escalates to 82.10 s, whereas EPML requires merely 0.12 s. When scaling to $|I|=15$, Shapley value and nucleolus require 1,840.50 s $(\approx$30.7 minutes) and $3,210.70$ s ($\approx$53.5 minutes), respectively.

At $|I|=20$, Shapley value hits the strict 24-hour computational cutoff ($86,400.00$ s), while nucleolus demands over 13.3 hours (48,120.00 s). At $|I|=30$, both classical allocation mechanisms become computationally intractable due to exponential sub-coalition enumeration ($2^{30}-1>10^9$ combinations) and memory bounds. In stark contrast, the proposed EPML engine scales polynomially, completing the full allocation across all realization scenarios in just 1.87 s at $|I|=30$, confirming its suitability for real-time, large-scale collaborative logistics operations.

5.4.2 Qualification of coalition stability

It should be emphasized that the core stability condition $\left(e_{\max }(x) \leq 0\right)$ and expected individual rationality ($\mathbb{E}\left[\pi_i\right] \geq \pi_i^0$) established under EPML provide a formal financial defense against coalition defection. However, financial profitability alone does not guarantee absolute organizational permanence in real-world supply chain alliances. In practice, participating firms may elect to exit an alliance due to qualitative and strategic determinants not captured in mathematical cost-allocation models, including corporate autonomy erosion, governance friction, inter-firm operational mistrust, or divergent long-term commercial trajectories. Thus, the proposed allocation rule removes financial disincentives to collaborate under the modeled parameters, but must be paired with robust corporate governance mechanisms to ensure holistic alliance longevity.

5.5 Resilience and Risk Mitigation Performance

The explicit inclusion of disruption risk objective $Z_3$ along with scenario-dependent reliability factors $\rho_i^s=\rho_i\left(1-\delta_i^s\right)$ enables the network to absorb severe operational shocks. Under standard non-collaborative routing, a localized facility breakdown ($\delta_i^s \geq 0.50$) results in immediate demand shortfalls and severe financial penalties ($p_j^{\text {pen}}$).

In the collaborative multi-objective configuration, the Level 1 model dynamically buffers capacity losses by executing second-stage recourse flows ($x_{i j s}^s$) toward secondary, high-reliability alliance partners ($\rho_k \geq 0.90$). Across 50 evaluated disruption scenarios, the collaborative network maintains an average customer service level of 99.59%, reducing total unserved demand by 92.4% compared to standalone operations.

Execution Time Disambiguation (Level 1 vs. Level 2): To avoid ambiguity regarding computational claims, a clear distinction must be maintained between the two operational stages of the framework:

• Level 1 Joint Network Optimization: Determining active hub configurations ($y_i$) and stochastic physical routing schedules ($x_{i j}^s$) requires solving a complex MILP/GA-LP model, requiring offline execution times ranging from several minutes to hours depending on the scenario set cardinality $|S|$.

• Level 2 EPML Financial Allocation: Once physical flows and realized scenario costs are passed from Level 1, the pure LP structure of EPML executes in 0.32 s for $|I|$ = 15 and 1.87 s for $|I|$ = 30, as reported in Table 8. Consequently, the term “real-time application” strictly applies to Level 2 post-operational financial settlement and dynamic contract re-balancing, whereas Level 1 master routing schedules are intended for periodic or tactical master planning.

5.6 In-depth Operational Sensitivity Analysis
5.6.1 Sensitivity to shortfall penalty parameter ($p_{j}^\mathrm{pen}$)

To assess model stability against variations in penalty parameter calibration, experiments were conducted (Table 9) on the medium instance ($|I|$ = 15, $|J|$ = 8, $|S|$ = 50) by varying $p_j^{\text {pen}}$ from $1.5 \times$ $p_j^{\text {sell}}$ to $4.0 \times p_j^{\text {sell}}$.

Table 9. Sensitivity to shortfall penalty parameter ($p_{j}^\mathrm{pen}$)
Penalty Calibration ($p_{j}^\mathrm{pen}$)Unserved Demand Rate (%)System Cost $Z_1$ (\$K)Service Level (%)Equity Ratio ($\theta$)
$1.5\times p_{j}^\mathrm{sell}$4.82%4,210.595.18%0.8712
$2.0\times p_{j}^\mathrm{sell}$1.85%4,388.298.15%0.8804
$2.5\times p_{j}^\mathrm{sell}$ (Baseline)*0.41%4,492.099.59%0.8842
$3.0\times p_{j}^\mathrm{sell}$0.12%4,510.699.88%0.8839
$4.0\times p_{j}^\mathrm{sell}$0.00%4,535.1100.00%0.8825
Note: * marks the main-model baseline setting.

Managerial Interpretation of the Shortfall Penalty ($\boldsymbol{p_j^{\text {pen}}}$): Beyond its mathematical function in preventing degenerate zero-flow solutions, the shortfall penalty parameter $p_j^{\text {pen}}$ carries a concrete operational meaning in real-world supply chain governance. Rather than functioning purely as an artificial tuning knob, $p_j^{\text {pen}}$ explicitly quantifies the cumulative financial impact of service level agreement (SLA) breaches, contractual liquidated damages owed to commercial clients, emergency expedited shipping costs, and long-term brand equity damage resulting from unfulfilled demand. When $p_j^{\text {pen}}<$ $2.0 \times p_j^{\text {sell}}$, the penalty is lower than the true cost of customer attrition, causing the network to tolerate unserved demand (4.82%). Calibrating $p_j^{\text {pen}} \geq 2.5 \times p_j^{\text {sell}}$ accurately reflects binding commercial SLA clauses in modern freight contracts, ensuring that the optimization model prioritizes customer fulfillment (99.59% service level) over localized routing cost savings.

5.6.2 Impact of carbon taxation on Pareto trade-off solutions

To evaluate policy sensitivity, an explicit carbon tax $t_{\text{CO}_2}$ ranging from \$0/tCO$_2$e to \$150/tCO$_2$e was integrated into objective evaluations (Table 10).

Table 10. Pareto compromise shifts under carbon taxation

Carbon Tax (\$/tCO$_2$e)

Scope 3 Emissions $Z_2$ (tCO$_2$e)

Transport Cost (\$K)

Total Cost Incl. Tax (\$K)

\$0

387.4

4,386.5

4,386.5

\$30

354.2

4,441.2

4,451.8

\$60 (Baseline C)

341.9

4,492.0

4,512.5

\$100

312.5

4,605.1

4,636.4

\$150

298.4

4,713.8

4,758.6

CO$_2$e = carbon dioxide equivalent.

Analysis: Introducing carbon pricing smoothly shifts the optimal physical compromise solution toward low-emission routes. As illustrated in Figure 4a, a carbon tax of \$150/tCO$_2$e reduces Scope 3 emissions by 23.0% (down to 298.4 tCO$_2$e) relative to the cost-minimal baseline (387.4tCO$_2$e at \$0 tax). Evaluating the MAC via Eq. (48) demonstrates non-linear abatement cost escalation—rising from \$1,647.59/tCO$_2$e under a \$30 tax to \$3,677.53/tCO$_2$e under a \$150 tax—reflecting diminishing economic returns as the operational network exhausts simple routing adjustments and transitions toward higher-cost low-carbon modal switches.

Figure 4. Sensitivity to carbon taxation and disruption severity
Note: tCO$_2$e = metric ton of carbon dioxide equivalent; GHG = greenhouse gas; MAC = marginal abatement cost.
5.6.3 Sensitivity to disruption risk severity ($\mathbb{E}[\delta_{i}^{s}]$)

The impact of disruption severity was examined by increasing the mean disruption severity $\mathbb{E}[\delta_{i}^{s}]$ from 10% to 50% across network facilities (Table 11).

Table 11. Network flux responses to disruption severity ($\mathbb{E}[\delta_{i}^{s}]$)

Disruption Severity $\boldsymbol{E[\delta_i^s]}$

High-Reliability Flow Share ($\boldsymbol{\rho_i \ge 0.90}$)

Total Cost $\boldsymbol{Z_1}$ (\$K)

Disruption Risk $Z_3$ (Units)

Equity Ratio ($\theta$)

10%

61.4%

4,120.8

12,450

0.8910

20% (Baseline)

78.2%

4,492.0

18,920

0.8842

30%

89.5%

4,915.3

24,110

0.8755

40%

96.1%

5,480.1

28,600

0.8601

50%

99.2%

6,210.4

31,450

0.8420

Systemic Over-Reliance and Resilience Concentration Risks: The routing shifts documented in Table 11 reveal a critical engineering management trade-off between immediate disruption absorption and long-term structural vulnerability. As mean disruption severity $E\left[\delta_i^s\right]$ escalates from 10% to 50%, the optimization framework dynamically redirects network volume toward highly reliable partners ($\rho_i \geq$ 0.90), increasing their flow share from 61.4% to 99.2% (Figure 4b). While this operational flexibility buffers short-term capacity losses and maintains high customer service levels, it exposes the alliance to resilience concentration risk. Concentrating nearly all freight flows onto a narrow subset of resilient hub operators can trigger secondary physical bottlenecks, facility throughput congestion, and operational fatigue at those central nodes. From a strategic management perspective, alliance leadership should not rely indefinitely on flow rerouting alone; instead, high flow concentration metrics should trigger proactive capacity expansion or targeted reliability investments in secondary carrier networks to prevent systemic single-point-of-failure risks.

As depicted in the dual-axis response in Figure 4b, despite aggressive structural rerouting, EPML preserves high coalition equity ($\theta \geq$ 0.8420) even under catastrophic 50% capacity losses, demonstrating the structural stability of relative maximin allocations during major supply chain crises.

6. Discussion

6.1 Interpretation of Results in Perspective of Prior Studies

The empirical results validate the core working hypothesis: decoupling physical distribution flow optimization from dynamic financial payoff allocation resolves the structural instability commonly observed in horizontal supply chain alliances. Foundational literature on coalitional logistics has extensively relied on classical cooperative game theory—such as the Shapley value, the nucleolus, and Equal Profit Share rules [59], [60]. However, recent studies highlight that these traditional mechanisms suffer from exponential computational complexity ($O\left(2^N\right)$), exceeding 86,400 s of CPU time for networks with $N \geq$ 30 partners [61], [62]. Furthermore, widely implemented proportional gain-sharing schemes [63] yield severe distributional inequality, producing a low equity ratio of only $\theta$ = 0.1120. Under proportional rules, large dominant actors absorb the vast majority of financial synergies, leaving smaller coalition members with negligible profit margins that fail to compensate for their operational risk exposure.

In contrast, whenever the selected physical solution satisfies the net-synergy feasibility condition stated in Section 3.2 ($\sum_i E\left[\pi_i\right] \geq \sum_i \pi_i^0$), the proposed EPML formulation theoretically guarantees expected individual rationality ($E\left[\pi_i\right] \geq \pi_i^0$) while achieving empirically verified core stability ($e_{\max} \leq 0$) alongside a superior equity ratio of $\theta$ = 0.8842. All participating entities attain a virtually uniform relative profit improvement (25.06%–25.66%) over their non-cooperative standalone baselines. This finding directly addresses a critical barrier highlighted in recent horizontal logistics literature [61], [64], which confirms that over 50% of horizontal supply chain alliances dissolve due to perceived unfairness in cost/benefit allocation rather than a deficit in total joint savings.

6.2 Operationalizing Equity Ratios in Strategic Partnerships

In multi-echelon networks, market power asymmetries often prevent small and medium-sized enterprises (SMEs) from joining collaborative freight consortia [64], [65]. Standard proportional allocation mechanisms accentuate these power imbalances by rewarding high volume rather than relative contribution to network efficiency. By maintaining an equity ratio near unity ($\theta$ = 0.8842), EPML functions as a contractually binding allocation framework. Supply chain directors can leverage EPML during pre-alliance negotiations to assure prospective partners that benefits are distributed equitably relative to standalone performance baselines, effectively eliminating the primary cause of inter-firm friction.

Contractual Consensus vs. Mathematical Equity: A high numerical equity ratio $(\theta=0.8842)$ provides an objective, transparent benchmark for fair surplus distribution, but it should not be conflated with automatic contractual agreement. Heterogeneous partners-such as small regional carriers versus large hub operators-weigh operational risks differently. While Shapley value and nucleolus assign disproportionate absolute payouts to large hubs, EPML normalizes gains relative to noncooperative baselines ($\frac{\mathbb{E}\left[\pi_i\right]-\pi_i^0}{\pi_i^0}$), structurally protecting smaller firms from margin erosion. However, in practice, $\theta$ serves as an algorithmic starting point for pre-alliance negotiations, providing a verifiable basis upon which partners can establish consensus.

6.3 Dynamic Gain-Sharing and Automated Re-Negotiation

In volatile operational environments, physical disruption events—such as port congestion, vehicle breakdowns, or severe weather—alter optimal routing paths and immediately render fixed, ex-ante cost-allocation contracts obsolete [60], [66]. Traditional cooperative game-theoretic solvers are incapable of real-time execution in large networks ($N \geq$ 30). In contrast, the linear mathematical formulation of EPML executes in 0.32 s for 15 partners and 1.87 s for 30 partners. This computational speed transforms gain-sharing from a static, post-hoc accounting exercise into a dynamic operational decision tool [62], [67]. Logistics managers can re-run the EPML allocation algorithm rapidly whenever operational disruptions or demand shifts occur, supporting timely and transparent contractual gain-sharing updates across coalition partners [67].

Decoupling Environmental Mandates from Financial Disputes: The proposed two-stage hierarchical architecture decouples physical multi-objective trade-offs (Cost vs. Scope 3 emissions vs. Resilience) from inter-firm financial negotiations. Level 1 identifies the physical Pareto frontier, enabling coalition executives to select an operational compromise (e.g., Point C in Table 5, which reduces Scope 3 emissions to 341.9 tCO$_2$e—a 45.5 tCO$_2$e or 11.7% reduction relative to the cost-minimizing Point A—at a modest 2.40% cost premium, while achieving an 18.42% cost reduction and a 14.10% carbon reduction relative to the non-cooperative baseline). Once this physical configuration is established, Level 2 calculates exact financial transfer payments ($\alpha_i^{s *}$) without reopening negotiations over physical routing.

This separation simplifies collaborative governance, accelerates executive consensus, while providing a quantitative basis for transparent monitoring and reporting of supply-chain emissions. Where applicable, these data may inform companies’ environmental due-diligence processes under the EU CSDDD [68].

6.4 Operational Scope and Network Structure Boundaries

While the proposed framework demonstrates robust computational performance and high allocation equity, the operational interpretation of these findings must be contextualized within the structural boundaries of the model. The formulation assumes a single-period, single-echelon, direct-shipping logistics network. These assumptions are intentionally chosen to isolate the structural mechanics of the endogenous LP-based gain-sharing engine and the tri-objective Pareto trade-offs without confounding variables. Consequently, specific numerical findings—such as sub-second allocation solve times (1.87 s for 30 partners) and flow rerouting percentages—reflect this single-tier direct distribution context and should not be automatically extrapolated to multi-echelon networks featuring intermediate consolidation hubs, multi-tier inventory holding, or complex lead-time dynamics without appropriate architectural adaptations.

Furthermore, the unified carbon coefficient $e_{ij}$ abstracts individual corporate GHG reporting boundaries by aggregating facility processing (Scope 3 Cat. 1 proxy) and freight transit (Scope 3 Cat. 4 proxy) from an alliance-wide perspective. While this aggregation enables effective multi-objective flow optimization across shared infrastructure, individual partner firms must apply appropriate organizational boundary adjustments (e.g., operational control vs. equity share approaches) when reporting these collaborative savings in their formal corporate GHG inventories.

6.5 Practical Managerial Decision Protocol and Operational Roadmap

To translate the two-stage hierarchical framework into an actionable decision-support tool for logistics managers, the deployment process across collaborative logistics consortia is structured into a practical 4-step managerial protocol:

Data Onboarding and Baseline Parameter Submission

Participating firms submit four standardized inputs to an independent neutral alliance platform or trusted algorithmic broker: (i) standalone operating cost baselines ($C_i^0$) and profit targets ($\pi_i^0$) derived from non-cooperative historical runs, (ii) declared nominal hub processing capacities (${CAP}_i$) and baseline reliability estimates ($\rho_i$), (iii) unit transportation rates ($c_{i j}$) and processing costs ($v_i$), and (iv) route-specific Scope 3 carbon intensity factors ($e_{i j}$). This step requires zero disclosure of proprietary customer databases or granular profit margins, preserving corporate commercial privacy.

Level 1: Joint Operational Master Planning (Resource Sharing & Flow Allocation)

The alliance platform executes the Level 1 stochastic MILP/GA-LP model over the shared network. Decision-makers receive a visual tri-objective Pareto trade-off surface (Figure 1), displaying candidate operational compromise configurations. Upon selecting an executive compromise point (e.g., Point C), the framework outputs two binding operational decisions: (a) facility activation vectors ($y_i \in\{0,1\}$), determining which partner distribution hubs are activated for shared freight processing, and (b) scenario-dependent routing schedules ($x_{i j}^s$), specifying physical cargo flows across inter-partner corridors.

Level 2: Ex-Ante Agreement and Ex-Post Financial Settlement

Once physical flows ($x_{i j}^s$) and total scenario operational expenses ($C_{\text {total}}^s$) are established at Level 1, the Level 2 EPML model computes scenario-wise cost-allocation fractions ($\alpha_i^s$). From a managerial perspective, these allocation fractions are agreed upon ex-ante as a binding scenario-contingent lookup rule during pre-alliance contract negotiations. Ex-post, once actual operational costs are realized under scenario $s \in S$, financial settlement is executed deterministically according to the corresponding pre-agreed vector $\alpha^s$. Partners operating activated high-volume hubs ($y_i=1$) receive financial reimbursements from nonactivated or origin-routing partners ($y_i=0$) whose demand is served through shared assets, mathematically guaranteeing that every participant achieves an expected profit gain ($\mathbb{E}\left[\pi_i\right] \geq \pi_i^0$) exceeding its standalone benchmark.

Strategic Decision Outputs and Dynamic Contractual Governance

The final deliverables provided to consortium executives comprise: (a) an asset-sharing operational contract establishing hub usage schedules, (b) a transparent financial transfer clearing matrix specifying scenario-wise cost responsibilities, and (c) verifiable Scope 3 carbon compliance reporting ($Z_2$) aligned with regulatory standards (e.g., EU CSDDD). In the event of an unforeseen disruption outside the pre-modeled scenario set S, the high computational efficiency of Level 2 EPML (0.32 s for 15 partners and 1.87 s for 30 partners, as reported in Table 8 allows managers to re-run the allocation model in real time. This rapid execution provides an algorithmic foundation for dynamic contract adjustments and automated smart-contract re-settlement during severe supply chain crises.

7. Conclusion

This study presented a two-stage hierarchical post-Pareto selection architecture for collaborative supply chains, decoupling physical multi-objective operational decision-making from financial gain-sharing ($\alpha_i^s$) via a Lexicographic EPML. By optimizing expected logistics costs, Scope 3 emissions, and disruption resilience at Level 1 and maximizing relative surplus at Level 2, the framework resolves the critical challenge of fair profit allocation by theoretically guaranteeing expected individual rationality ($E\left[\pi_i\right] \geq \pi_i^0$) whenever the net-synergy feasibility condition ($\sum_i E\left[\pi_i\right] \geq \sum_i \pi_i^0$) holds and demonstrating empirical core stability across tested operational scenarios, without compromising computational scalability. Computational evaluations highlight the practical significance of the approach: EPML achieves a high equity ratio ($\theta$ = 0.8842), substantially outperforming traditional game-theoretic allocations like the Shapley value ($\theta$ = 0.5481) and nucleolus ($\theta$ = 0.5821) while maintaining overall cost savings within 2–3% of cooperative upper bounds. Crucially, by scaling polynomially rather than exponentially ($O(2^{|I|})$), the proposed framework provides an operational decision-support tool capable of managing large-scale industrial partnerships.

The empirical findings, equity metrics ($\theta$ = 0.8842), and computational performance reported in this study reflect optimal responses under the specific single-echelon network structures, cost distributions, and disruption severity regimes ($E[\delta_i^s] \in$ [10%, 50%]) tested. While the two-stage hierarchical model provides a generalizable mathematical methodology for horizontal gain sharing, performance metrics in real industrial coalitions will naturally vary according to specific regional infrastructure topology, partner fleet sizes, and contractually agreed governance frameworks.

However, the scope of this work presents certain structural limitations. First, the mathematical formulation is developed for a single-period, single-echelon, direct-shipping network topology, which abstracts away multi-echelon consolidation hubs, multi-tier inventory dynamics, and time-varying multi-period demand fluctuations. Second, while expected individual rationality ($\mathbb{E}\left[\pi_i\right] \geq \pi_i^0$) is structurally guaranteed across the stochastic scenario tree, general core stability for arbitrary sub-coalitions was verified empirically rather than established through general analytical proofs for arbitrary network topologies. Finally, complete information sharing among alliance partners is assumed. To address these boundaries, future research should extend the two-stage hierarchical framework to multi-echelon dynamic planning horizons with intermediate hubs, establish analytical core stability bounds for general network structures, and integrate privacy-preserving mechanisms or smart contracts to handle asymmetric information in real-world governance.

Author Contributions

Conceptualization, H.O. and A.S.; methodology, H.O. and A.S.; software, A.S.; validation, H.O., A.S., and M.E.; formal analysis, H.O.; investigation, H.O.; resources, H.O.; data curation, H.O. and A.S.; writing—original draft preparation, H.O.; writing—review and editing, H.O., A.S., and M.E.; visualization, H.O. and A.S.; supervision, M.E.; project administration, H.O. All authors have read and agreed to the published version of the manuscript.

Data Availability

All data generated or analyzed during this study are derived from publicly accessible databases as cited in the text, or are mathematically reproducible using the methodology described in Section 4 (Solution Methodology). The source code and parameter initialization settings to recreate the results are available from the corresponding author on reasonable request.

Conflicts of Interest

The authors declare no conflicts of interest.

Declaration on the Use of Generative AI and AI-assisted Technologies

The authors declare that QuillBot was used for editorial assistance. The use of AI did not influence the scientific content of the manuscript. The author(s) used Google Gemini in order to assist in writing Python/Matplotlib scripts for data visualization (Figures 1, 2, 3, and 4). After using this tool, the author(s) reviewed, validated, and executed the code independently, ensuring 100% data consistency with the mathematical model and optimization results. The author(s) take full responsibility for the content of the publication.

References
1.
M. Breque, L. De Nul, and A. Petridis, “Industry 5.0: Towards a sustainable, human-centric and resilient European industry,” Luxembourg, 2021. [Online]. Available: https://op.europa.eu/en/publication-detail/-/publication/468a892a-5097-11eb-b59f-01aa75ed71a1/language-en [Google Scholar]
2.
J. Durán-Micco, S. Alaei, and C. Macharis, “Evaluating the impact of horizontal collaboration in logistics systems: A simulation-based study,” Transportmetrica B: Transp. Dyn., vol. 13, no. 1, p. 2475204, 2025. [Google Scholar] [Crossref]
3.
N. Mrabti, N. Hamani, and L. Delahoche, “A comprehensive literature review on sustainable horizontal collaboration,” Sustainability, vol. 14, no. 18, p. 11644, 2022. [Google Scholar] [Crossref]
4.
World Resources Institute and World Business Council for Sustainable Development, “Corporate value chain (Scope 3) accounting and reporting standard: Supplement to the GHG Protocol corporate accounting and reporting standard,” 2011. https://ghgprotocol.org/sites/default/files/standards/Corporate-Value-Chain-Accounting-Reporing-Standard_041613_2.pdf [Google Scholar]
5.
M. Gansterer and R. F. Hartl, “Collaborative vehicle routing: A survey,” Eur. J. Oper. Res., vol. 268, no. 1, pp. 1–12, 2018. [Google Scholar] [Crossref]
6.
M. Ghomi-Avili, R. Tavakkoli-Moghaddam, S. G. Jalali Naeini, and A. Jabbarzadeh, “Competitive green supply chain network design model considering inventory decisions under uncertainty: A real case of a filter company,” Int. J. Prod. Res., vol. 59, no. 14, pp. 4248–4267, 2021. [Google Scholar] [Crossref]
7.
T. Hacardiaux, C. Defryn, J.-S. Tancrez, and L. Verdonck, “Balancing partner preferences for logistics costs and carbon footprint in a horizontal cooperation,” OR Spectr., vol. 44, no. 1, pp. 121–153, 2021. [Google Scholar] [Crossref]
8.
H. Saffari, M. Abbasi, and J. Gheidar-Kheljani, “A robust, sustainable, resilient, and responsive model for forward/reverse logistics network design with a new approach based on horizontal collaboration,” Environ. Dev. Sustain., vol. 27, pp. 23439–23482, 2025. [Google Scholar] [Crossref]
9.
L. Safari, S. J. Sadjadi, and F. M. Sobhani, “Resilient and sustainable supply chain design and planning under supply disruption risk using a multi-objective scenario-based robust optimization model,” Environ. Dev. Sustain., vol. 26, pp. 27485–27527, 2024. [Google Scholar] [Crossref]
10.
A. A. Filho, I. Correia, and T. Melo, “Bi-objective stochastic supply chain network design under supplier and facility disruption risks,” Comput. Appl. Math., vol. 45, p. 309, 2026. [Google Scholar] [Crossref]
11.
F. Cruijssen, W. Dullaert, and H. Fleuren, “Horizontal cooperation in transport and logistics: A literature review,” Transp. J., vol. 46, no. 3, pp. 22–39, 2007. [Google Scholar] [Crossref]
12.
S. Pan, D. Trentesaux, E. Ballot, and G. Q. Huang, “Horizontal collaborative transport: Survey of solutions and practical implementation issues,” Int. J. Prod. Res., vol. 57, no. 15–16, pp. 5340–5361, 2019. [Google Scholar] [Crossref]
13.
L. Verdonck, A. Caris, K. Ramaekers, and G. K. Janssens, “Collaborative logistics from the perspective of road transportation companies,” Transp. Rev., vol. 33, no. 6, pp. 700–719, 2013. [Google Scholar] [Crossref]
14.
A. Serrano-Hernandez, A. A. Juan, J. Faulin, and E. Perez-Bernabeu, “Horizontal collaboration in freight transport: Concepts, benefits and environmental challenges,” SORT, vol. 41, no. 2, pp. 393–414, 2017. [Google Scholar] [Crossref]
15.
M. Çerçi, Z. B. Kanık, H. C. Uzgören Kazanç, and M. Soysal, “Transforming logistics with horizontal collaboration: A systematic review of trends, challenges, and sustainability perspectives,” Rev. Manag. Sci., pp. 1–63, 2026. [Google Scholar] [Crossref]
16.
F. Cruijssen, I. van Beest, and G. Kant, “A human behaviour perspective on horizontal collaboration to reduce the climate impact of logistics,” Sustainability, vol. 15, no. 23, p. 16221, 2023. [Google Scholar] [Crossref]
17.
T. Hacardiaux and J.-S. Tancrez, “Assessing the environmental benefits of horizontal cooperation using a location-inventory model,” Cent. Eur. J. Oper. Res., vol. 28, no. 4, pp. 1363–1387, 2020. [Google Scholar] [Crossref]
18.
A. Aloui, N. Hamani, and L. Delahoche, “An integrated optimization approach using a collaborative strategy for sustainable cities freight transportation: A case study,” Sustain. Cities Soc., vol. 75, p. 103331, 2021. [Google Scholar] [Crossref]
19.
J. R. Birge and F. Louveaux, Introduction to Stochastic Programming, 2nd. New York, NY, USA: Springer, 2011. [Online]. Available: [Google Scholar] [Crossref]
20.
W. Klibi, A. Martel, and A. Guitouni, “The design of robust value-creating supply chain networks: A critical review,” Eur. J. Oper. Res., vol. 203, no. 2, pp. 283–293, 2010. [Google Scholar] [Crossref]
21.
R. B. Nelsen, An Introduction to Copulas, 2nd. New York, NY, USA: Springer, 2006. [Online]. Available: [Google Scholar] [Crossref]
22.
H. Heitsch and W. Römisch, “Scenario reduction algorithms in stochastic programming,” Comput. Optim. Appl., vol. 24, no. 2–3, pp. 187–206, 2003. [Google Scholar] [Crossref]
23.
S. Abbasi and Y. Kazancoglu, “Financial risk-aware closed-loop supply chain network design for the circular economy: A two-stage stochastic optimization model,” Results Eng., vol. 32, p. 112395, 2026. [Google Scholar] [Crossref]
24.
K. Mili and M. Argoubi, “Stochastic optimization of perishable agricultural supply chains: A hybrid genetic algorithm approach for robust network design under multi-dimensional uncertainty,” Front. Sustain. Food Syst., vol. 10, p. 1807797, 2026. [Google Scholar] [Crossref]
25.
G. Mavrotas and K. Florios, “An improved version of the augmented epsilon-constraint method (AUGMECON2) for finding the exact Pareto set in multi-objective integer programming problems,” Appl. Math. Comput., vol. 219, no. 18, pp. 9652–9669, 2013. [Google Scholar] [Crossref]
26.
G. Mavrotas, “Effective implementation of the ϵ-constraint method in multi-objective mathematical programming problems,” Appl. Math. Comput., vol. 213, no. 2, pp. 455–465, 2009. [Google Scholar] [Crossref]
27.
E.-G. Talbi, “Combining metaheuristics with mathematical programming, constraint programming and machine learning,” 4OR, vol. 11, no. 2, pp. 101–150, 2013. [Google Scholar] [Crossref]
28.
E. Winter, “The Shapley value,” in Handbook of Game Theory with Economic Applications, The Netherlands: Elsevier, 2002, pp. 2025–2054. [Google Scholar] [Crossref]
29.
D. Schmeidler, “The nucleolus of a characteristic function game,” SIAM J. Appl. Math., vol. 17, no. 6, pp. 1163–1170, 1969. [Google Scholar] [Crossref]
30.
M. Guajardo and M. Rönnqvist, “A review on cost allocation methods in collaborative transportation,” Int. Trans. Oper. Res., vol. 23, no. 3, pp. 371–392, 2016. [Google Scholar] [Crossref]
31.
M. Frisk, M. Göthe-Lundgren, K. Jörnsten, and M. Rönnqvist, “Cost allocation in collaborative forest transportation,” Eur. J. Oper. Res., vol. 205, no. 2, pp. 448–458, 2010. [Google Scholar] [Crossref]
32.
L. A. Guardiola, A. Meca, and J. Timmer, “Cooperation and profit allocation in distribution chains,” Decis. Support Syst., vol. 44, no. 1, pp. 17–27, 2007. [Google Scholar] [Crossref]
33.
B. Colson, P. Marcotte, and G. Savard, “An overview of bilevel optimization,” Ann. Oper. Res., vol. 153, no. 1, pp. 235–256, 2007. [Google Scholar] [Crossref]
34.
A. Sinha, P. Malo, and K. Deb, “A review on bilevel optimization: From classical to evolutionary approaches and applications,” IEEE Trans. Evol. Comput., vol. 22, no. 2, pp. 276–295, 2018. [Google Scholar] [Crossref]
35.
C. J. Nohra and N. V. Sahinidis, “Global optimization of nonconvex problems with convex-transformable intermediates,” J. Glob. Optim., vol. 72, no. 2, pp. 255–276, 2018. [Google Scholar] [Crossref]
36.
L. Torres and F. S. Ramos, “Allocating benefits due to shared resources using Shapley value and nucleolus in dynamic network Data Envelopment Analysis,” Mathematics, vol. 12, no. 5, p. 698, 2024. [Google Scholar] [Crossref]
37.
P. Dehez, “Cooperation and fair division,” in Game Theory for the Social Sciences, Cham, Switzerland: Springer, 2024. [Google Scholar] [Crossref]
38.
J. Dahlberg, M. Göthe-Lundgren, and S. Engevall, “A note on the nonuniqueness of the Equal Profit Method,” Appl. Math. Comput., vol. 308, pp. 84–89, 2017. [Google Scholar] [Crossref]
39.
ecoinvent Association, “ecoinvent database v3.9,” Zurich, Switzerland, 2022. [Online]. Available: https://support.ecoinvent.org/ecoinvent-version-3.9 [Google Scholar]
40.
M. Wang, A. Elgowainy, U. Lee, K. H. Baek, S. Balchandani, P. T. Benavides, A. Burnham, H. Cai, P. Chen, Y. Gan et al., “Greenhouse gases, Regulated Emissions, and Energy use in Technologies (GREET) Model® (2023 .Net),” Computer Software. USDOE Office of Energy Efficiency and Renewable Energy (EERE), 2023. [Online]. Available: https://www.osti.gov/doecode/biblio/113208 [Google Scholar]
41.
Department for Energy Security and Net Zero, “Greenhouse gas reporting: Conversion factors 2024,” Gov.uk, 2024. [Online]. Available: https://www.gov.uk/government/publications/greenhouse-gas-reporting-conversion-factors-2024 [Google Scholar]
42.
World Resources Institute and World Business Council for Sustainable Development, “The Greenhouse Gas Protocol: A corporate accounting and reporting standard, rev. ed.,” WRI/WBCSD, Washington, DC, USA, 2004. [Online]. Available: https://ghgprotocol.org/corporate-standard [Google Scholar]
43.
J.-F. Bérubé, M. Gendreau, and J.-Y. Potvin, “An exact ϵ-constraint method for bi-objective combinatorial optimization problems: Application to the Traveling Salesman Problem with Profits,” Eur. J. Oper. Res., vol. 194, no. 1, pp. 39–50, 2009. [Google Scholar] [Crossref]
44.
C. A. Coello Coello, G. B. Lamont, and D. A. Van Veldhuizen, Evolutionary Algorithms for Solving Multi-Objective Problems, 2nd. New York, NY, USA: Springer, 2007. [Online]. Available: [Google Scholar] [Crossref]
45.
E. Zitzler, L. Thiele, M. Laumanns, C. M. Fonseca, and V. G. Da Fonseca, “Performance assessment of multiobjective optimizers: An analysis and review,” IEEE Trans. Evol. Comput., vol. 7, no. 2, pp. 117–132, 2003. [Google Scholar] [Crossref]
46.
D. A. Van Veldhuizen and G. B. Lamont, “Multiobjective evolutionary algorithm research: A history and analysis,” Air Force Inst. Tech., Wright-Patterson AFB, OH, USA, Tech. Rep. TR-98-03, 1998, pp. 98–03. [Google Scholar]
47.
J. R. Schott, “Fault tolerant design using single and multicriteria genetic algorithm optimization,” M.S. thesis, Massachusetts Institute of Technology, Cambridge, MA, USA, 1995. [Google Scholar]
48.
M. A. Londe, L. S. Pessoa, C. E. Andrade, and M. G. Resende, “Biased random-key genetic algorithms: A review,” Eur. J. Oper. Res., vol. 321, no. 1, pp. 1–22, 2025. [Google Scholar] [Crossref]
49.
C. Blum, J. Puchinger, G. R. Raidl, and A. Roli, “Hybrid metaheuristics in combinatorial optimization: A survey,” Appl. Soft Comput., vol. 11, no. 6, pp. 4135–4151, 2011. [Google Scholar] [Crossref]
50.
M. Gen and L. Lin, “Multiobjective evolutionary algorithm for manufacturing scheduling problems: State-of-the-art survey,” J. Intell. Manuf., vol. 25, no. 5, pp. 849–866, 2014. [Google Scholar] [Crossref]
51.
K. Deb, A. Pratap, S. Agarwal, and T. Meyarivan, “A fast and elitist multiobjective genetic algorithm: NSGA-II,” IEEE Trans. Evol. Comput., vol. 6, no. 2, pp. 182–197, 2002. [Google Scholar] [Crossref]
52.
J. R. Birge and F. Louveaux, “Basic properties and theory,” in Introduction to Stochastic Programming, New York, NY, USA: Springer, 1997, pp. 83–135. [Google Scholar] [Crossref]
53.
M. Sklar, “N-dimensional distribution functions and their margins,” Ann. ISUP, vol. 8, no. 3, pp. 229–231, 1959. [Google Scholar]
54.
M. D. McKay, R. J. Beckman, and W. J. Conover, “A comparison of three methods for selecting values of input variables in the analysis of output from a computer code,” Technometrics, vol. 42, no. 1, pp. 55–61, 2000. [Google Scholar] [Crossref]
55.
J. Dupačová, N. Gröwe-Kuska, and W. Römisch, “Scenario reduction in stochastic programming,” Math. Program., vol. 95, no. 3, pp. 493–511, 2003. [Google Scholar] [Crossref]
56.
H. Moulin, “Equal or proportional division of a surplus, and other methods,” Int. J. Game Theory, vol. 16, no. 3, pp. 161–186, 1987. [Google Scholar] [Crossref]
57.
L. S. Shapley, “A value for n-person games,” in Contributions to the Theory of Games II, Princeton, NJ, USA: Princeton University Press, 1953, pp. 307–317. [Google Scholar] [Crossref]
58.
M. Nagarajan and G. Sošić, “Game-theoretic analysis of cooperation among supply chain agents: Review and extensions,” Eur. J. Oper. Res., vol. 187, no. 3, pp. 719–745, 2008. [Google Scholar] [Crossref]
59.
S. Lozano, P. Moreno, B. Adenso-Díaz, and E. Algaba, “Cooperative game theory approach to allocating benefits of horizontal cooperation,” Eur. J. Oper. Res., vol. 229, no. 2, pp. 444–452, 2013. [Google Scholar] [Crossref]
60.
B. Dai and H. Chen, “Profit allocation mechanisms for carrier collaboration in pickup and delivery service,” Comput. Ind. Eng., vol. 62, no. 2, pp. 633–643, 2012. [Google Scholar] [Crossref]
61.
M. Gansterer and R. F. Hartl, “Shared resources in collaborative vehicle routing,” TOP, vol. 28, no. 1, pp. 1–20, 2020. [Google Scholar] [Crossref]
62.
M. Lai, Y. Wu, and X. Cai, “Core-based cost allocation for collaborative multi-stop truckload shipping problem,” IISE Trans., vol. 57, no. 8, pp. 938–956, 2025. [Google Scholar] [Crossref]
63.
D. W. Yeung and L. A. Petrosyan, “A fair gain-sharing solution of cooperative dynamic games,” Int. Game Theory Rev., vol. 27, no. 3, p. 2550008, 2025. [Google Scholar] [Crossref]
64.
H. Sternberg, I. Linan, G. Prockl, and A. Norrman, “Tragedy of the facilitated commons: A multiple-case study of failure in systematic horizontal logistics collaboration,” J. Supply Chain Manag., vol. 58, no. 4, pp. 30–57, 2022. [Google Scholar] [Crossref]
65.
J.-F. Audy, S. D’Amours, and L.-M. Rousseau, “Cost allocation in the establishment of a collaborative transportation agreement—an application in the furniture industry,” J. Oper. Res. Soc., vol. 62, no. 6, pp. 960–970, 2011. [Google Scholar] [Crossref]
66.
S. A. Hosseini Shekarabi, R. Kiani Mavi, and F. Romero Macau, “Supply chain resilience: A critical review of risk mitigation, robust optimisation, and technological solutions and future research directions,” Glob. J. Flex. Syst. Manag., vol. 26, pp. 681–735, 2025. [Google Scholar] [Crossref]
67.
D. Cao, P. Liu, and C. Liu, “Revenue allocation for shared manufacturing collaboration considering participants’ contributions based on a differential game and an improved Shapley algorithm,” Mod. Supply Chain Res. Appl., vol. 8, no. 2, pp. 239–261, 2026. [Google Scholar] [Crossref]
68.
European Parliament and Council of the European Union, “Consolidated text: Directive (EU) 2024/1760 of the European Parliament and of the Council of 13 June 2024 on corporate sustainability due diligence and amending Directive (EU) 2019/1937 and Regulation (EU) 2023/2859 (Text with EEA relevance),” Official Journal of the European Union, L 2024/1760, 2024. [Online]. Available: https://eur-lex.europa.eu/eli/dir/2024/1760/2026-03-18 [Google Scholar]

Cite this:
APA Style
IEEE Style
BibTex Style
MLA Style
Chicago Style
GB-T-7714-2015
Ouhader, H., Sbai, A., & El Kyal, M. (2026). stochastic Collaborative Supply Chain Optimization With Fair Hierarchical Gain-Sharing. J. Eng. Manag. Syst. Eng., 5(4), 558-588. https://doi.org/10.56578/jemse050407
H. Ouhader, A. Sbai, and M. El Kyal, "stochastic Collaborative Supply Chain Optimization With Fair Hierarchical Gain-Sharing," J. Eng. Manag. Syst. Eng., vol. 5, no. 4, pp. 558-588, 2026. https://doi.org/10.56578/jemse050407
@research-article{Ouhader2026StochasticCS,
title={stochastic Collaborative Supply Chain Optimization With Fair Hierarchical Gain-Sharing},
author={Hanan Ouhader and Asma Sbai and Malika El Kyal},
journal={Journal of Engineering Management and Systems Engineering},
year={2026},
page={558-588},
doi={https://doi.org/10.56578/jemse050407}
}
Hanan Ouhader, et al. "stochastic Collaborative Supply Chain Optimization With Fair Hierarchical Gain-Sharing." Journal of Engineering Management and Systems Engineering, v 5, pp 558-588. doi: https://doi.org/10.56578/jemse050407
Hanan Ouhader, Asma Sbai and Malika El Kyal. "stochastic Collaborative Supply Chain Optimization With Fair Hierarchical Gain-Sharing." Journal of Engineering Management and Systems Engineering, 5, (2026): 558-588. doi: https://doi.org/10.56578/jemse050407
OUHADER H, SBAI A, EL KYAL M. stochastic Collaborative Supply Chain Optimization With Fair Hierarchical Gain-Sharing[J]. Journal of Engineering Management and Systems Engineering, 2026, 5(4): 558-588. https://doi.org/10.56578/jemse050407
cc
©2026 by the author(s). Published by Acadlore Publishing Services Limited, Hong Kong. This article is available for free download and can be reused and cited, provided that the original published version is credited, under the CC BY 4.0 license.