A computational study published in Discover Artificial Intelligence by Jun Han demonstrates that coupling hierarchical game theory with probabilistic fuzzy sets improves coordination, convergence speed, and efficiency across simulated industrial networks. By integrating single-leader, multiple-follower Stackelberg game dynamics with probabilistic hesitant fuzzy sets, the framework addresses the dual challenges of physical stochastic disruptions and subjective expert hesitation in automated decision-making. Evaluated across three computational benchmark configurations derived from laboratory-scale industrial testbeds, the bilevel optimization model achieved approximately 45 percent faster empirical convergence toward a common stopping tolerance compared with state-based potential games. Alongside this acceleration, the framework yielded a 34 to 37 percent increase in normalized leader utility, reduced electrical energy consumption by 10 to 14 percent while preserving high production throughput, and decreased utility variance by 52 percent under the highest simulated noise conditions. The findings highlight how combining asymmetric leadership with probability-preserving fuzzy sets enhances uncertainty-aware coordination in self-optimizing manufacturing systems.
The operational motivation for the research stems from the ongoing digital transformation of industrial chains, where cyber-physical production lines and distributed suppliers collaborate under uncertain conditions. In typical manufacturing ecosystems, strategic authority is rarely distributed symmetrically; dominant firms or central production planners hold influential market positions governing system-wide workflows, whereas modular production units and secondary suppliers adjust their operations in response. Traditional multi-agent algorithms frequently assume that all participating entities act simultaneously. This assumption can obscure structural power disparities, causing conflicting local choices, unstable exploration, and intractable tuning of combined objective functions across decentralized subsystems. Furthermore, modern industrial operations inherently generate dual uncertainties: quantifiable random disruptions, such as machine failures or sensor noise, and subjective linguistic ambiguity from imprecise expert judgments. To reconcile these hierarchical and epistemic complexities, the research formulates a sequential decision architecture where a single dominant leader commits to a strategy first, prompting follower entities to respond with coordinated adjustments.
At the core of the framework lies a formalized Stackelberg game structure embedded within a state-based potential-game formulation to coordinate learning across decentralized agents. The industrial chain is modeled as a hierarchical game comprising a single leader entity, a set of m noncooperative follower entities, a continuous state space capturing system conditions, and compact action spaces. System dynamics evolve over discrete time steps governed by transition dynamics and additive Gaussian disturbances. In this bilevel structure, the leader commits to an action first, anticipating follower reactions. Rather than assuming that follower entities centrally maximize aggregate welfare, the model characterizes the joint follower response through a noncooperative Nash equilibrium correspondence, in which each follower maximizes its personal utility conditioned on the leader’s choice and peer actions. The leader then optimizes its strategy relative to this anticipated follower Nash set. By incorporating an exact potential function grounded in social welfare, the formulation provides structured feedback loops guiding distributed learning without ad-hoc exploration.
To represent the dual uncertainty inherent in industrial coordination, the methodology integrates Probabilistic Hesitant Fuzzy Sets into the utility functions of all interacting entities. While traditional stochastic models demand exact probability distributions and standard fuzzy logic relies on fixed membership values, a probabilistic hesitant fuzzy set explicitly accommodates several possible membership degrees along with corresponding probability masses summing to unity. This formulation captures situations where expert evaluations exhibit linguistic hesitation and stochastic variability simultaneously. When combining distinct fuzzy evaluations, standard algebraic operations can suffer from arbitrary pairings or numerical drift. The framework resolves this by evaluating algebraic additions and scalar multiplications across the Cartesian product of the constituent support sets. Each resulting pair receives a probability mass calculated from the component probabilities, after which identical membership degrees are consolidated by summing their associated weights. Normalization subsequently divides grouped probability masses by their total sum, preventing numerical distortion while maintaining unit total probability across policy calculations.
The integration of probabilistic hesitant fuzzy sets directly alters how conflicting multi-objective operational goals are balanced within the hierarchical game. Industrial chain collaboration routinely demands simultaneous trade-offs among production output, electrical power consumption, and system-level operational risks. Rather than eliminating objective scalarization, the proposed architecture retains scalarized multi-objective utility functions but replaces static, crisp weighting coefficients with uncertainty-aware, fuzzy-valued weights. The priorities assigned to competing metrics are treated as discrete hesitant elements paired with subjective probability distributions, reflecting genuine disagreement or ambiguity among domain specialists regarding operational priorities. To aggregate these complex valuations, the framework employs a Probabilistic Weighted Averaging operator, which combines scalar multiplication and Cartesian-product fuzzy addition into a unified utility transformation. As a result, both the leader and follower entities evaluate their strategic choices through fuzzy-transformed utilities that systematically account for linguistic hesitation and stochastic variability, avoiding the fragile parameter hand-tuning that frequently degrades conventional multi-objective optimization schemes in dynamic cyber-physical environments.
From a theoretical standpoint, the study establishes explicit mathematical conditions governing the existence and uniqueness of the resulting PHFS-transformed Stackelberg equilibrium. Equilibrium existence is proved under standard analytical assumptions: action sets must be nonempty, compact, and convex; transformed utilities must remain continuous across the joint action space; follower utilities must be concave in own actions; and the follower Nash correspondence must possess a closed graph. To guarantee equilibrium uniqueness, the framework introduces two structural conditions: the joint follower best-response mapping must operate as a strict contraction with modulus q_F less than one, and the composed leader objective must exhibit strict concavity in the leader’s action. Under these contractive conditions, Banach’s fixed-point theorem ensures a unique follower response for every leader decision, yielding a unique Stackelberg optimum. The author distinguishes theoretical existence from empirical convergence rates, measuring convergence through observed stopping iterations under a tolerance threshold of 10 to the power of negative three.
The experimental validation of the framework was conducted through extensive computational replications across three distinct benchmark configurations derived from previously reported laboratory-scale manufacturing testbeds. The first configuration, the Bulk Goods System, models a bulk-material transportation pipeline focused on balancing overflow prevention, power minimization, and customer demand fulfillment across 10,000 simulated episodes subjected to Gaussian noise standard deviations between 0.05 and 0.20. The second configuration evaluates a Modular Manufacturing testbed simulating distributed self-optimizing machinery using single-step and multi-step deep neural predictors across 15,000 trajectories, with fuzzy probability distributions governed by symmetric Dirichlet priors. The third configuration models a laboratory-scale decentralized manufacturing system known as DS2-SbPG, examining sequential and serial-parallel production lines across 20,000 episodes under explicit disruption-prevention and power-reduction constraints. Across these environments, the hesitation parameter was calibrated to moderate values—specifically five hesitant elements for the bulk and modular setups and seven for the serial-parallel platform—to balance representational resolution against Cartesian-product computational costs.
Empirical benchmarking against four comparative algorithms demonstrated consistent advantages for the integrated PHFS-Stackelberg framework. The tested baselines included standard State-based Potential Games, Distributed Stackelberg Strategies in State-based Potential Games, a Fuzzy-Stackelberg model lacking probabilistic hesitation, and Multi-Agent Deep Deterministic Policy Gradient. Across all three testbeds, the proposed approach achieved convergence in approximately 45 percent fewer iterations toward the stopping tolerance. On the Bulk Goods System, the framework converged in 152 iterations, compared with 278 for standard potential games and 301 for the fuzzy alternative, marking a 45.3 percent improvement. Parallel gains of 45.4 percent and 46.4 percent were recorded on the modular and serial-parallel platforms. Furthermore, the model achieved 90 percent of its asymptotic performance within 6,000 episodes on bulk goods, 9,000 on modular, and 12,000 on serial-parallel benchmarks, outpacing the potential-game baseline by 40 to 50 percent. Crucially, an event-triggered communication mechanism reduced inter-agent message broadcasts by 68 percent relative to time-triggered protocols, causing less than two percent impact on solution quality.
The quantitative outcomes across multi-objective optimization metrics further reinforced the practical effectiveness of the proposed design. Normalized leader utility increased by 33.8 to 37.3 percent over baseline potential games, while electrical energy consumption decreased by 9.8 to 13.5 percent without sacrificing production throughput, which remained between 95 and 99 percent of theoretical capacity. In multi-objective trade-offs mapping production rates against energy use and operational risk, the framework improved the hypervolume indicator across empirical Pareto frontiers by 15 to 22 percent. Ablation experiments isolating architectural components demonstrated that the Stackelberg hierarchy and the PHFS uncertainty mechanism operate synergistically: the unified model delivered a 32 percent improvement in social welfare and a 41 percent reduction in operational risk relative to ablated variants. In robustness evaluations under disturbance noise up to 0.50 on the serial-parallel benchmark, the method achieved a 52 percent reduction in utility variance compared with standard potential games and 38 percent compared with distributed Stackelberg strategies across 20 runs.
Despite these favorable computational outcomes, several analytical limitations define the scope of the reported research. The formulation remains strictly confined to a single-leader hierarchy with fixed roles, meaning that dynamic leader election, automated authority delegation, and multiple-leader competitions remain unaddressed. Furthermore, the mathematical proof of Pareto efficiency applies exclusively to the full-information case with exactly one follower entity; for multi-follower settings, the reported Pareto frontiers represent empirical trade-off comparisons rather than formally guaranteed optimal boundaries. The author also situates the method in relation to recent advances in fractional and orthopair fuzzy decision frameworks that employ complex aggregation operators and sensitivity testing for static multi-criteria ranking, noting that the primary distinction of this study is embedding probabilistic hesitation directly into a dynamic, sequential game-theoretic optimization process. Finally, the published results represent computational replications executed on simulated environments rather than live trials on physical factory hardware, underscoring that practical deployment will require further validation under physical industrial constraints.
Subject of Research: Probabilistic hesitant fuzzy Stackelberg game optimization for industrial chain collaboration
Article Title: Probabilistic hesitant fuzzy decision optimization and Stackelberg games for AI enabled industrial chain collaboration under uncertainty
Article References: Han, J. (2026). Probabilistic hesitant fuzzy decision optimization and Stackelberg games for AI enabled industrial chain collaboration under uncertainty. Discover Artificial Intelligence, 6(1), Article 1827. https://doi.org/10.1007/s44163-026-02601-5
Image Credits: AI Generated
DOI: 10.1007/s44163-026-02601-5
Keywords: Stackelberg games, probabilistic hesitant fuzzy sets, industrial chain collaboration, bilevel decision optimization, hierarchical game theory, event-triggered communication, state-based potential games, follower Nash equilibrium, multi-objective scalarization, probabilistic weighted averaging, bulk goods system, modular manufacturing, DS2-SbPG testbed, energy consumption reduction, utility variance reduction, stochastic disturbance modeling, cyber-physical production systems
News Source: Denise Maddox. (November 12, 2026). Probabilistic Hesitant Fuzzy Stackelberg Framework Accelerates Convergence in Industrial Optimization. Scienmag.
