A quantum fuzzy logic-enhanced evolutionary framework for energy-latency optimization in 6G Edge-IioT Authors: Siavash Siavashian Rashidi, Ali Broumandnia, Abbas Mirzaei, Ramin Karimi Published: 2026 ABSTRACT The transition toward 6G-enabled Industrial Internet of Things (IIoT) requires strict Ultra-Reliable Low-Latency Communication (URLLC), creating a fundamental trade-off with energy consumption. Orchestrating resources in dense, hyper-dynamic edge environments is an NP-hard problem where traditional algorithms often exhibit brittleness and premature convergence. In this paper, we propose a novel multi-objective orchestration framework, QFLN-AMRO, designed to robustly resolve the energy-latency conflict. The architecture integrates a Fractional-Order NSGA-II evolutionary engine with a cognitive Quantum Fuzzy Logic Network (QFLN) to dynamically adjust evolutionary parameters, thereby establishing a superior deterministic Pareto front. To prevent network saturation and queue drift under extreme URLLC stress, we further incorporate a Lyapunov-guided resilience mechanism. The scientific validity and robustness of the proposed framework are rigorously evaluated using a tripartite real-world data setup comprising EUA spatial topologies, Edge-IIoT payloads, and Google Cluster Traces. Empirical results demonstrate that QFLN-AMRO significantly outperforms state-of-the-art deep reinforcement learning models (PPO, GAT) and metaheuristics (GA, PSO, GWO) in minimizing energy expenditure while guaranteeing deterministic latency thresholds. The superior performance of our approach is validated by rigorous statistical analysis, including Omnibus ANOVA (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:p\approx\:4.66\times\:{10}^{-14}$$\end{document}) and Tukey HSD tests, confirming its reliability for next-generation edge orchestration. Introduction The transition from the Fourth Industrial Revolution to the Industry 5.0 paradigm has profoundly reshaped global manufacturing and service landscapes1,2. Fueled by the rapid expansion of the Industrial Internet of Things (IIoT), billions of interconnected, heterogeneous devices now generate vast quantities of critical data3,4. However, centralized cloud architectures, despite their computational power, struggle to meet the stringent low-latency, high-bandwidth, and data privacy requirements of modern industrial applications. Consequently, edge computing has emerged as a key distributed architectural paradigm5,6. By migrating computational resources to the network periphery—in close proximity to data sources—edge architectures significantly reduce transmission latency, alleviate core network congestion, and facilitate real-time, autonomous decision-making in densely connected industrial ecosystems7–9. A central mechanism in this edge-enabled ecosystem is task offloading. Executing complex, AI-driven, or data-intensive workloads on end-tier IIoT devices is often infeasible due to their inherent limitations, such as constrained battery life and limited processing capacity10. Task offloading addresses this challenge by migrating computationally intensive tasks to more powerful edge servers, thereby extending device longevity and enhancing the overall Quality of Service (QoS)11–13. Nevertheless, orchestrating this offloading process in dense, dynamic environments presents a multifaceted, NP-hard optimization problem5,14. The complexity of this resource orchestration involves a critical trade-off between algorithmic latency and infrastructure energy consumption. While edge servers reduce the computational strain on IIoT nodes, aggressive workload aggregation at the edge can create thermal hotspots and increase total systemic energy consumption11,15,16. According to the International Energy Agency (IEA), data centers and transmission networks accounted for approximately 1.4%–1.7% of global electricity use in 2022. This substantial energy demand increases operational expenditures (OpEx) for industrial entities and conflicts with global mandates for carbon footprint reduction16,17. Conversely, prioritizing energy conservation may require local queueing or delayed execution, which fails to meet the Ultra-Reliable Low-Latency Communication (URLLC) constraints essential for critical operations like industrial robotics and real-time process control18–20. Although meta-heuristic algorithms, such as Genetic Algorithms (GA) and Particle Swarm Optimization (PSO), are widely used to address this Multi-Objective Optimization Problem (MOOP)5,21,22, existing methods have notable vulnerabilities. Most conventional frameworks are deterministic or single-objective, failing to account for the stochastic nature of modern IIoT workloads, including fluctuating data volumes, dynamic bandwidth, and heterogeneous task lifecycles. Moreover, standard evolutionary algorithms often experience premature convergence when applied to large, highly constrained search spaces, which limits their ability to dynamically balance latency and energy consumption under peak workload stress. To address this research gap, this paper introduces a novel architecture: the Quantum Fuzzy Logic-Enhanced Adaptive Multi-Objective Resource Orchestrator (QFLN-AMRO). This framework advances beyond static heuristics by integrating a Fractional-Order Non-dominated Sorting Genetic Algorithm II (NSGA-II) with an advanced Quantum Fuzzy Logic cognitive layer. This synergy mitigates premature convergence and redefines the Pareto-optimal frontier, enabling a dynamic balance between URLLC latency requirements and green-computing energy objectives. The primary contributions of this research are as follows: Formulate a Multi-Objective Orchestration Model: We develop a highly constrained multi-objective optimization model for task offloading in ultra-dense Edge-IIoT environments, designed to minimize the trade-off between total execution latency and systemic energy expenditure.Design the QFLN-AMRO Framework: We introduce a predictive and adaptive orchestration matrix that embeds quantum-fuzzy mechanics into the NSGA-II engine. This design establishes a more deterministic Pareto-optimal front for intelligent and dynamic QoS compliance.Utilize Real-World Data Integration: In contrast to studies that rely on synthetic probability distributions, our proposed architecture is evaluated using a high-fidelity tripartite empirical setup: the real-world Edge Cloud Ecosystem (EUA) dataset for spatial topologies, Edge-IIoT for realistic payload profiling, and Google Cluster Traces for temporal workload dynamics.Conduct Extensive Stress-Testing and Statistical Validation: We rigorously validate the framework’s algorithmic performance and resilience under extreme load scenarios, involving up to 1000 concurrent URLLC tasks. The statistical superiority of QFLN-AMRO is confirmed through comprehensive omnibus ANOVA and Tukey HSD post-hoc analyses. Formulate a Multi-Objective Orchestration Model: We develop a highly constrained multi-objective optimization model for task offloading in ultra-dense Edge-IIoT environments, designed to minimize the trade-off between total execution latency and systemic energy expenditure. Formulate a Multi-Objective Orchestration Model: We develop a highly constrained multi-objective optimization model for task offloading in ultra-dense Edge-IIoT environments, designed to minimize the trade-off between total execution latency and systemic energy expenditure. Design the QFLN-AMRO Framework: We introduce a predictive and adaptive orchestration matrix that embeds quantum-fuzzy mechanics into the NSGA-II engine. This design establishes a more deterministic Pareto-optimal front for intelligent and dynamic QoS compliance. Design the QFLN-AMRO Framework: We introduce a predictive and adaptive orchestration matrix that embeds quantum-fuzzy mechanics into the NSGA-II engine. This design establishes a more deterministic Pareto-optimal front for intelligent and dynamic QoS compliance. Utilize Real-World Data Integration: In contrast to studies that rely on synthetic probability distributions, our proposed architecture is evaluated using a high-fidelity tripartite empirical setup: the real-world Edge Cloud Ecosystem (EUA) dataset for spatial topologies, Edge-IIoT for realistic payload profiling, and Google Cluster Traces for temporal workload dynamics. Utilize Real-World Data Integration: In contrast to studies that rely on synthetic probability distributions, our proposed architecture is evaluated using a high-fidelity tripartite empirical setup: the real-world Edge Cloud Ecosystem (EUA) dataset for spatial topologies, Edge-IIoT for realistic payload profiling, and Google Cluster Traces for temporal workload dynamics. Conduct Extensive Stress-Testing and Statistical Validation: We rigorously validate the framework’s algorithmic performance and resilience under extreme load scenarios, involving up to 1000 concurrent URLLC tasks. The statistical superiority of QFLN-AMRO is confirmed through comprehensive omnibus ANOVA and Tukey HSD post-hoc analyses. Conduct Extensive Stress-Testing and Statistical Validation: We rigorously validate the framework’s algorithmic performance and resilience under extreme load scenarios, involving up to 1000 concurrent URLLC tasks. The statistical superiority of QFLN-AMRO is confirmed through comprehensive omnibus ANOVA and Tukey HSD post-hoc analyses. To address this, the QFLN integrates a cognitive layer into the NSGA-II framework, dynamically self-tuning the genetic operators—namely, the crossover and mutation probabilities. This is achieved through a set of fuzzy inference rules that monitor the algorithm’s state, such as population diversity and convergence progress, in real-time. Based on these metrics, the fuzzy membership functions adjust the operators’ rates to strike an optimal balance between exploration and exploitation. For instance, if population diversity decreases, the mutation rate is increased to introduce new genetic material, whereas in later generations, the crossover rate might be prioritized to refine high-quality solutions. Furthermore, the “quantum” aspect of QFLN introduces a significant advantage over classical fuzzy logic by leveraging the principles of quantum mechanics in its cognitive layer. Unlike traditional fuzzy systems that use real-numbered membership values between [0, 1], our QFLN employs qubits (quantum bits). A qubit, existing in a superposition of states, can represent a fuzzy membership value not as a single point, but as a probabilistic distribution. This allows the system to model uncertainty and ambiguity with much higher fidelity. Consequently, the QFLN’s decision-making process for tuning NSGA-II parameters becomes more robust and nuanced, enabling a more intelligent and adaptive search of the complex solution space for energy-latency trade-offs. Contributions To address the limitations of conventional heuristics, such as premature convergence, and deep reinforcement learning architectures, which can be hindered by the curse of dimensionality and substantial training overhead in deterministic 6G edge environments, this paper introduces the Quantum Fuzzy Logic-Enhanced Adaptive Multi-Objective Resource Orchestrator (QFLN-AMRO). The principal contributions of this research are: A Novel Architectural Synergy: We introduce a multi-objective framework that integrates a Quantum Fuzzy Logic Network (QFLN) with a modified Non-dominated Sorting Genetic Algorithm II (NSGA-II) core. The QFLN leverages fuzzy membership functions to dynamically recalibrate mutation and crossover rates, which mitigates the static scalarization problem and prevents the premature convergence common in traditional swarm intelligence algorithms (e.g., PSO, GWO).Deterministic URLLC Resilience via Lyapunov Stability: QFLN-AMRO incorporates a rigorous admission control and selective task-dropping mechanism based on Lyapunov drift optimization. This design ensures stability under extreme network saturation. We empirically demonstrate that with hyper-dynamic payloads (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:N\ge\:1000$$\end{document} concurrent tasks), the framework maintains a near-optimal system stability ratio of 0.0105, effectively shielding mission-critical URLLC traffic.A Tripartite Data-Driven Evaluation: To ensure methodological rigor and avoid biases from purely synthetic simulations, the framework’s evaluation is based on a tripartite data-driven approach. We synthesize spatial constraints from the EUA dataset ( M=20 servers), stochastic temporal patterns from the Google Cluster Trace, and deterministic payloads from the Edge-IIoT dataset (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:N=100\to\:500$$\end{document}). This method ensures that the evaluated latency-energy Pareto fronts reflect realistic Industry 5.0 operational conditions.Demonstrated Statistical and Operational Superiority: We provide conclusive quantitative evidence of the framework’s superior performance. QFLN-AMRO outperforms state-of-the-art Deep Reinforcement Learning (PPO, GAT) and metaheuristic (PSO, GWO) baselines, achieving a 25% reduction in overall energy consumption while preserving URLLC latency thresholds. A rigorous statistical validation using ANOVA and Tukey HSD post-hoc tests yields a p-value of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:4.66\times\:{10}^{-14}$$\end{document}, confirming that the observed performance gains are statistically significant and not a result of stochastic variance. A Novel Architectural Synergy: We introduce a multi-objective framework that integrates a Quantum Fuzzy Logic Network (QFLN) with a modified Non-dominated Sorting Genetic Algorithm II (NSGA-II) core. The QFLN leverages fuzzy membership functions to dynamically recalibrate mutation and crossover rates, which mitigates the static scalarization problem and prevents the premature convergence common in traditional swarm intelligence algorithms (e.g., PSO, GWO). A Novel Architectural Synergy: We introduce a multi-objective framework that integrates a Quantum Fuzzy Logic Network (QFLN) with a modified Non-dominated Sorting Genetic Algorithm II (NSGA-II) core. The QFLN leverages fuzzy membership functions to dynamically recalibrate mutation and crossover rates, which mitigates the static scalarization problem and prevents the premature convergence common in traditional swarm intelligence algorithms (e.g., PSO, GWO). Deterministic URLLC Resilience via Lyapunov Stability: QFLN-AMRO incorporates a rigorous admission control and selective task-dropping mechanism based on Lyapunov drift optimization. This design ensures stability under extreme network saturation. We empirically demonstrate that with hyper-dynamic payloads (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:N\ge\:1000$$\end{document} concurrent tasks), the framework maintains a near-optimal system stability ratio of 0.0105, effectively shielding mission-critical URLLC traffic. Deterministic URLLC Resilience via Lyapunov Stability: QFLN-AMRO incorporates a rigorous admission control and selective task-dropping mechanism based on Lyapunov drift optimization. This design ensures stability under extreme network saturation. We empirically demonstrate that with hyper-dynamic payloads (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:N\ge\:1000$$\end{document} concurrent tasks), the framework maintains a near-optimal system stability ratio of 0.0105, effectively shielding mission-critical URLLC traffic. A Tripartite Data-Driven Evaluation: To ensure methodological rigor and avoid biases from purely synthetic simulations, the framework’s evaluation is based on a tripartite data-driven approach. We synthesize spatial constraints from the EUA dataset ( M=20 servers), stochastic temporal patterns from the Google Cluster Trace, and deterministic payloads from the Edge-IIoT dataset (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:N=100\to\:500$$\end{document}). This method ensures that the evaluated latency-energy Pareto fronts reflect realistic Industry 5.0 operational conditions. A Tripartite Data-Driven Evaluation: To ensure methodological rigor and avoid biases from purely synthetic simulations, the framework’s evaluation is based on a tripartite data-driven approach. We synthesize spatial constraints from the EUA dataset ( M=20 servers), stochastic temporal patterns from the Google Cluster Trace, and deterministic payloads from the Edge-IIoT dataset (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:N=100\to\:500$$\end{document}). This method ensures that the evaluated latency-energy Pareto fronts reflect realistic Industry 5.0 operational conditions. Demonstrated Statistical and Operational Superiority: We provide conclusive quantitative evidence of the framework’s superior performance. QFLN-AMRO outperforms state-of-the-art Deep Reinforcement Learning (PPO, GAT) and metaheuristic (PSO, GWO) baselines, achieving a 25% reduction in overall energy consumption while preserving URLLC latency thresholds. A rigorous statistical validation using ANOVA and Tukey HSD post-hoc tests yields a p-value of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:4.66\times\:{10}^{-14}$$\end{document}, confirming that the observed performance gains are statistically significant and not a result of stochastic variance. Demonstrated Statistical and Operational Superiority: We provide conclusive quantitative evidence of the framework’s superior performance. QFLN-AMRO outperforms state-of-the-art Deep Reinforcement Learning (PPO, GAT) and metaheuristic (PSO, GWO) baselines, achieving a 25% reduction in overall energy consumption while preserving URLLC latency thresholds. A rigorous statistical validation using ANOVA and Tukey HSD post-hoc tests yields a p-value of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:4.66\times\:{10}^{-14}$$\end{document}, confirming that the observed performance gains are statistically significant and not a result of stochastic variance. The remainder of this paper is organized as follows. Section 2 reviews the relevant literature and identifies gaps in existing task offloading models for dense Edge-IIoT environments. Section 3 describes the proposed methodology, including the dataset synthesis, the mathematical MOOP formulation, and the core QFLN-AMRO architecture. Section 4 presents a comprehensive performance evaluation, covering resilience, benchmarking, and large-scale workload stress testing. Section 5 discusses the results, focusing on algorithmic convergence and statistical robustness. Finally, Sect. 6 concludes the paper and outlines future research directions. Related work and critical literature review The proliferation of the Industrial Internet of Things (IIoT) and Edge Computing has made task offloading a critical mechanism for mitigating the computational constraints of end-devices while optimizing energy and latency. Due to the NP-hard nature of resource orchestration in dense IIoT environments, the literature from 2021 to 2025 has primarily focused on meta-heuristic and learning-based multi-objective optimization strategies. This section provides a critical synthesis of these methodologies, delineating their contributions and identifying the structural limitations that motivate the development of the proposed QFLN-AMRO architecture. Limitations of conventional meta-heuristic approaches Nature-inspired meta-heuristics are widely applied to address the Multi-Objective Optimization Problem (MOOP) of balancing energy and latency in IIoT networks. For instance, You and Tang5 employed Particle Swarm Optimization (PSO) to achieve a more effective energy-latency balance compared to baseline Genetic Algorithms (GA) and Simulated Annealing (SA). Similarly, Wang and Lu14 introduced a hybrid Grey Wolf Optimizer-GA (GWO-GA) that reduced QoS-aware service discovery costs by 30%. Other research has focused on expanding exploration capabilities; Bali et al.22 hybridized the Artificial Bee Colony and Whale Optimization Algorithms (HAWO), reporting significant systemic gains including a 50% reduction in energy and a 25% decrease in latency. Furthermore, Mir et al.21 integrated chaos theory and fuzzy logic into a Grasshopper Optimization Algorithm (FLGOA) to extend network lifespan by 16%. Critical Insight: While effective for baseline optimization, these frameworks are susceptible to premature convergence (stochastic stagnation) when applied to the vast and highly constrained search spaces characteristic of ultra-dense 6G/IIoT environments. Their reliance on standard evolutionary mechanics limits their ability to adapt to high-frequency workload volatility, exposing a critical need for more advanced exploration techniques, such as quantum parallelism, to avoid local optima. Critical Insight: While effective for baseline optimization, these frameworks are susceptible to premature convergence (stochastic stagnation) when applied to the vast and highly constrained search spaces characteristic of ultra-dense 6G/IIoT environments. Their reliance on standard evolutionary mechanics limits their ability to adapt to high-frequency workload volatility, exposing a critical need for more advanced exploration techniques, such as quantum parallelism, to avoid local optima. Critical Insight: While effective for baseline optimization, these frameworks are susceptible to premature convergence (stochastic stagnation) when applied to the vast and highly constrained search spaces characteristic of ultra-dense 6G/IIoT environments. Their reliance on standard evolutionary mechanics limits their ability to adapt to high-frequency workload volatility, exposing a critical need for more advanced exploration techniques, such as quantum parallelism, to avoid local optima. The computational bottleneck of learning-based paradigms To overcome the static nature of conventional meta-heuristics, recent research has incorporated Deep Reinforcement Learning (DRL) for adaptive offloading. Chen et al.23 modeled the offloading problem as a Markov Decision Process (MDP) and utilized a DDTO algorithm to balance task completion rates with energy consumption. Moghaddasi et al.24 leveraged a Rainbow Deep Q-Network (DQN) in a tiered D2D-Edge-Cloud architecture, improving energy efficiency by 29.8%. For highly dynamic topologies, Li et al.18 combined Graph Neural Networks (GNN) with the Advantage Actor-Critic (A2C) algorithm in aerial Edge-IoT systems. In a similar vein, Deng et al.12 employed Q-learning and Deep Deterministic Policy Gradient (DDPG) to reduce latency by up to 30%. Notably, Lin et al.7 pioneered an RL-based algorithm enhanced with quantum parallelism for virtual reality viewport rendering, demonstrating the potential of quantum-inspired approaches. Critical Insight: Despite their adaptive capabilities, DRL-based architectures are frequently constrained by the “curse of dimensionality” and significant training overheads. In mission-critical industrial settings where Ultra-Reliable Low-Latency Communication (URLLC) is essential, the non-deterministic exploration phases inherent to pure RL models can lead to unacceptable transient latency spikes and dropped tasks. Critical Insight: Despite their adaptive capabilities, DRL-based architectures are frequently constrained by the “curse of dimensionality” and significant training overheads. In mission-critical industrial settings where Ultra-Reliable Low-Latency Communication (URLLC) is essential, the non-deterministic exploration phases inherent to pure RL models can lead to unacceptable transient latency spikes and dropped tasks. Critical Insight: Despite their adaptive capabilities, DRL-based architectures are frequently constrained by the “curse of dimensionality” and significant training overheads. In mission-critical industrial settings where Ultra-Reliable Low-Latency Communication (URLLC) is essential, the non-deterministic exploration phases inherent to pure RL models can lead to unacceptable transient latency spikes and dropped tasks. Environmental sustainability and QoS Trade-offs The “Green Edge Computing” paradigm has catalyzed architectural designs focused on sustainable IIoT. Alharbi and Aldossary16 showed that edge-fog architectures could reduce carbon emissions by 43%, while Song et al.25 and Chai et al.26 extended orchestration to satellite-based MEC systems for global energy optimization. To ensure strict QoS compliance, particularly in hazardous industrial environments, Lv et al.19 developed a parallel optimization framework that achieved a task success rate above 90%. Other studies have scrutinized resource allocation models; Jiang et al.6 formulated a mixed-integer nonlinear programming problem solved with a binary Fireworks Algorithm, and Hou et al.27 created incentive-based homogeneous allocation models. Critical Insight: Although these studies address important aspects of architectural sustainability and QoS, they often treat the energy-latency trade-off as a static scalarization problem. Consequently, they typically fail to produce a mathematically rigorous, dynamically adapting Pareto-optimal front that can maintain URLLC standards under extreme network stress, such as massive concurrent task injections. Critical Insight: Although these studies address important aspects of architectural sustainability and QoS, they often treat the energy-latency trade-off as a static scalarization problem. Consequently, they typically fail to produce a mathematically rigorous, dynamically adapting Pareto-optimal front that can maintain URLLC standards under extreme network stress, such as massive concurrent task injections. Critical Insight: Although these studies address important aspects of architectural sustainability and QoS, they often treat the energy-latency trade-off as a static scalarization problem. Consequently, they typically fail to produce a mathematically rigorous, dynamically adapting Pareto-optimal front that can maintain URLLC standards under extreme network stress, such as massive concurrent task injections. Identified research gaps and proposed contribution This review of the state-of-the-art reveals three fundamental deficiencies in existing task offloading frameworks: Algorithmic Myopia: Current multi-objective algorithms often lack a “historical memory” mechanism, causing them to converge on local optima within the complex topologies of ultra-dense networks3.Inadequate Constraint Handling: Most models do not dynamically penalize constraint violations (e.g., latency threshold breaches) during the optimization process, leading to deterministic failures under heavy workloads.Absence of Hybrid Cognitive Architectures: The literature is notably deficient in architectures that fuse the rapid convergence of evolutionary algorithms with the deterministic precision of advanced logic systems2,4. Algorithmic Myopia: Current multi-objective algorithms often lack a “historical memory” mechanism, causing them to converge on local optima within the complex topologies of ultra-dense networks3. Algorithmic Myopia: Current multi-objective algorithms often lack a “historical memory” mechanism, causing them to converge on local optima within the complex topologies of ultra-dense networks3. Inadequate Constraint Handling: Most models do not dynamically penalize constraint violations (e.g., latency threshold breaches) during the optimization process, leading to deterministic failures under heavy workloads. Inadequate Constraint Handling: Most models do not dynamically penalize constraint violations (e.g., latency threshold breaches) during the optimization process, leading to deterministic failures under heavy workloads. Absence of Hybrid Cognitive Architectures: The literature is notably deficient in architectures that fuse the rapid convergence of evolutionary algorithms with the deterministic precision of advanced logic systems2,4. Absence of Hybrid Cognitive Architectures: The literature is notably deficient in architectures that fuse the rapid convergence of evolutionary algorithms with the deterministic precision of advanced logic systems2,4. To systematically address these limitations, this paper introduces the QFLN-AMRO architecture. Our approach integrates a Fractional-Order NSGA-II engine, which provides a mathematical memory for deeper convergence, with Quantum Computing mechanics to ensure robust search space exploration without stagnation. This core is governed by a Fuzzy Logic controller that dynamically adapts mutation rates and penalty functions. This tripartite cognitive synergy bridges the critical gap between environmental sustainability and deterministic URLLC QoS compliance in massive-scale IIoT ecosystems See Table 1. Table 1Comparison of algorithms and energy optimization results in IIoT task offloading.RowReferenceAlgorithm UsedDataset UsedObjectiveLimitationsResults Obtained1You & Tang5Particle Swarm Optimization (PSO)SimulationEnergy-efficient and low-latency task offloading in IIoT; minimize latency, energy consumption, and execution cost.Not specifiedReduced latency, balanced energy consumption, and more effective resource allocation compared to GA and SA.2Song et al.25Sequential Fractional Programming, Dual Lagrange Decomposition, E-CORA AlgorithmSimulationMinimize weighted total energy consumption of IoT mobile devices by offloading computations to LEO satellites.Not specifiedE-CORA outperformed full offloading/local computation. A specific amount of offloaded bits minimized energy.3Zhu et al.11Novel scheduling optimization algorithm for AI tasks in a heterogeneous smart edge computing frameworkSmall-scale testbed and extensive simulationReduce energy consumption in AI task processing in IIoT by offloading tasks from servers.Not specifiedProposed online scheduling consumed less than 80% of static scheduling energy and 70% of FIFO energy.4Li et al.18Graph Neural Network-based Reinforcement Learning (GNN-A2C: Advantage Actor-Critic with Graph Neural Networks)Implementation with Google Tensorflow, performance analysis (implies simulation)Maximize tasks offloaded to a drone with joint optimization of drone navigation and task offloading in aerial EdgeIoT.Not specified (real-time network state for drones unknown, posing a challenge)GNN-A2C achieved fast convergence and significantly reduced task drop rates.5Zhu et al.15Novel online method for task scheduling optimization in edge and cloudSmart edge computing testbed and extensive simulationOptimize AIoT energy efficiency with smart edge, focusing on edge device and cloud energy.Not specifiedProposed method outperformed other strategies with lower energy consumption.6Chen et al.23Dynamic Task Offloading Algorithm based on Deep Reinforcement Learning (DDTO) (MDP-based)ExperimentsBalance maximizing task completion rate with tolerable latency and minimizing IoT device power consumption.Not specifiedDDTO converged quickly and effectively balanced completion rate and power consumption.7Lin et al.7Online RL-based algorithm with quantum parallelism (MDP-based problem transformation)SimulationsReduce VR equipment energy consumption and ensure smooth VR experience in IIoT with joint offloading, computation, and spectrum allocation optimization.Not specifiedEfficiency in reducing energy consumption and stall rate, improved learning efficiency.8Lv et al.19Industrial gas tracking algorithm based on parallel optimization; two-stage gas tracking algorithm based on state transition modelSimulation and evaluationInter-layer digital twin optimization in IIoT for hazardous gas boundary tracking.Not specifiedEffective gas boundary tracking, reduced network energy. Parallel optimization task success rate > 0.9 in 5 s. State transition algorithm energy consumption 2.11 joules (N = 600), minimum tracking error 0.31.9Hou et al.27Incentive-based online task allocation algorithmNumerical results (simulation)Stimulate collaborative computing between EC servers and IIoT devices to accelerate processing and reduce latency. Maximize system utility.Not specifiedSuperior performance and efficiency; satisfied incentive compatibility, individual rationality, computational efficiency, and feasibility.10Deng et al.12Q-learning (discrete partial offloading), Deep Deterministic Policy Gradient (DDPG) (continuous offloading), both RL-basedSimulation resultsProvide automated partial offloading with minimal latency for better QoS in multi-user IIoT MEC systems.Not specifiedQ-learning reduced latency by 23%, DDPG by 30%.11Wang Dayong et al.3(Review)Review paperAnalysis of existing papersProvide comprehensive understanding of IoT task offloading in MEC; identify problems, techniques, methods, parameters, shortcomings, and future trends.Limited to reviewed papersProvided a perspective on IoT task offloading in MEC to guide researchers.12Gali & Mahamkali28Distributed Deep Meta-Learning Task Offloading (DDMTO) (DNN with backpropagation)Not specified in abstract (likely simulation)Address poor portability of DLTs for task offloading in dynamic IoT environments; ensure effective/efficient offloading in heterogeneous edge-cloud systems.Not specifiedAimed for better performance and reduced computational complexity with heterogeneous cloud collaboration (quantitative results not specified in abstract).13Moghaddasi et al.24Rainbow Deep Q-Network (Rainbow DQN) - Advanced DRLRigorous simulationsOptimize task offloading in D2D-edge-cloud for enhanced operational efficiency in IoT systems.Not specified29.8% increase in energy efficiency, 27.5% latency reduction, 43.1% utility increase compared to state-of-the-art methods.14Firouzi et al.8RL-based offloading techniqueCase study: ECG-based arrhythmia detectionAddress edge-fog-cloud IoT interaction for healthcare applications; effective load distribution.Not specifiedAimed to demonstrate and evaluate model efficiency (quantitative RL results not specified in abstract).15Jiang et al.6Customized Binary Fireworks Algorithm (enhanced with Gaussian mutation and elite selection)Simulation resultsJoint task offloading and resource allocation optimization in a new cloud-edge network architecture for IoT to ensure QoS.Not specifiedConfirmed efficiency; significant performance gains compared to newer offloading strategies.16Huang et al.29Drone-assisted task offloading optimization algorithm (decomposing non-convex problem into two subproblems)Simulation resultsMinimize total network overhead in drone-assisted fog-based IoT with joint optimization of drone path, power, and offloading ratios.Finding local optima for non-convex problemAchieved superior network overhead performance compared to existing algorithms.17Huang et al.30Distributed Offloading Algorithm (DOLA) based on non-cooperative game theoryReal-world dataset-based simulations (Melbourne CBD, Australia)Minimize task completion latency with offloading to multiple BSs in overlapping MEC regions using CoMP.Not specifiedProved Nash equilibrium existence. DOLA showed efficiency and superiority over existing schemes.18Wei et al.31Local Differential Privacy-based Deep Reinforcement Learning (LDP-DRL) with prioritized experience replayExtensive experimentsProvide personalized privacy-aware task offloading in edge-cloud IIoT for automated production.Not specifiedFormal privacy/convergence analysis. Experiments evaluated efficiency, privacy protection, discount factor impact, and cost efficiency (implying positive results).19Dai et al.32Learning-based co-offloading algorithmSimulationsMinimize system cost (latency and migration cost) in D2D-assisted MEC for IIoT with co-offloading to MEC servers or nearby devices.Not specifiedAlgorithm enabled optimal edge node selection without complete information; validated by simulations.20Bradbury et al.10Trust-based system architecture; Beta Reputation System (BRS) as trust model exampleImplementation on real hardware (Zolertia RE-Motes)Facilitate selection of suitable edge nodes for task offloading from resource-constrained IoT devices using trust and considering threats.Risk analysis not conducted, threat modeling excludes trust model itselfDemonstrated architecture feasibility, threat management efficiency. Showed trust value evolution. Identified future work.21Chai et al.26Attention mechanism and collaborative Proximal Policy Optimization (A-PPO) algorithmSimulation resultsImprove offloading efficiency in multi-task MEC for satellite IoT with joint dependent task offloading and resource allocation.Not specifiedA-PPO converged in 25 steps; reduced cost by at least 8.87% compared to baseline algorithms.22Gong4(Review)Review paperAnalysis of 18 papersSystematic review of metaheuristic/nature-inspired algorithms for reliable IoT networks (reliability, security, fault tolerance, sustainability).Limited to reviewed papersMechanisms improved reliability, fault detection, network efficiency, lifetime, and energy optimization. Current studies focus on edge capabilities.23Rana et al.33(Classification/Framework)Proposed framework and classification (framework with predictive models for metaheuristic selection/application)Case study on smart manufacturing (for fitness metric generation)Address issues, principles, and significance of metaheuristic routing for energy-efficient IoT transmission; provide classification and framework.Theoretical framework, not empirically testedProposed theoretical framework and case study; aimed to provide a benchmark solution.24Mir et al.21Fuzzy Chaotic Grasshopper Optimization Algorithm (FLGOA) (GOA enhanced with Lorenz chaos theory and fuzzy logic)Evaluation in two different scenarios (implies simulation)Energy-aware routing in IoT with optimized node sleep/wake scheduling.Not specifiedFLGOA showed 9–16% better network lifetime, 16–22% better residual energy, and 12–16% better coverage rate than FGOA, GOA, GSO.25Wang & Lu14Hybrid multi-objective metaheuristic: Grey Wolf Optimizer + Genetic Algorithm (GWO-GA)Experimental (implies simulation)QoS-aware service discovery and selection in edge-cloud IoT to reduce time, cost, and improve energy efficiency/security.Not specifiedGWO-GA performed 30% better in reducing cost factor compared to other algorithms.26Agarwal & Alam34Maximum Score Pooling (MSP) algorithm (hybrid metaheuristics) for feature selection; MSP-optimized Decision Tree (MSP-DT) classifierExperiments (implies lifelogging dataset)Personalized, edge-optimized lifelogging framework, robust to sensor failure, reducing network traffic and latency.Not specifiedDetected 12 activities with 97.67% average accuracy; reduced transferred data by 47.66%.27Singh et al.35Jaya Algorithm (JA)Benchmark-generated instances for various fog scenariosOptimize fog device deployment (FDD) for maximum network connectivity (FDC) and edge device coverage (EDC).Not specifiedSignificant improvement in EDC and FDC; JA outperformed APSO, LDWPSO, MPA in convergence.28Al-Hashimi et al.36Hybrid metaheuristics (GWO + GA); exploration algorithms (SVM, Naïve Bayes)Cardiac data collected from sensorsDevelop optimized e-health monitoring framework using IoT for lower costs, higher efficiency, accurate data analysis, and better patient care.Not specifiedHybrid GWO + GA outperformed standalone versions. Exploration algorithms with hybrid metaheuristic were more accurate.29Lakhan et al.13Evolutionary Metaheuristic Offloading Scheduling Algorithm (EMOS) frameworkSimulation resultsEfficient, low-latency ICPS for IoT applications addressing task offloading and scheduling to reduce various latencies (processing, offloading, execution, failure, security).Not specifiedReduced local processing latency by 39%, offloading latency by 41%, execution/propagation latency by 40%, failure latency by 43%, security/migration latency by 50% compared to existing ICPS metaheuristics.30Vaiyapuri et al.37FLIDS-BSAFSC: Federated Learning with Bird Swarm Algorithm (BSA) for feature selection and Social Group Optimization with Kernel Extreme Learning Machine for classificationBenchmark datasetsIntrusion Detection System (IDS) based on FL in IoT for classification, detection, and defense against attacks with improved security and privacy.Not specifiedFLIDS-BSAFSC outperformed recent approaches.31Satouf et al.20Adapted Grey Wolf Optimizer (GWO)Simulation for task schedulingSemi-dynamic real-time task scheduling for latency-sensitive IoT in cloud-fog, minimizing energy, cost, and makespan.Not specifiedOutperformed GA, PSO, ABC in makespan, total execution time, cost, and energy consumption.32Tu et al.38Collaborative filtering based on Artificial Bee Colony (ABC) and genetic operatorsComputational tests (implies agricultural service data)Improve accuracy of agricultural service recommendations and reduce MAE in fog-based IoT systems.Not specifiedImproved ranking score (5.8%), accuracy (5%), recall (5.7%), intra-similarity (13%), Hamming distance (4.8%) compared to other methods.33Bali et al.22Hybrid Artificial Bee Colony and Whale Optimization Algorithm (HAWO)Simulation results in MatlabOptimize energy and latency for task allocation to edge processing devices in IoT.Not specifiedHAWO showed promising results compared to ABC and WOA. Compared to benchmarks: 50% energy reduction, 25% latency reduction, 60% best cost improvement.34Nematollahi et al.39Moth Flame Optimization with Opposition-Based Learning (OBLMFO); stacked caching approach; blockchainEvaluation with distinct tasks in Python 3.9 (implies simulation)Efficient task offloading and resource allocation in IoT using fog computing, ensuring accuracy with blockchain.Not specifiedOBLMFO reduced latency factor by 12.18% and energy consumption by 6.22%.35Haseeb et al.9Intelligent and Secure Edge-based Computing Model (ISEC) (deep learning for routing, distributed hashing with chaining for security, symmetric cryptography, DHT)Experimental results (implies simulation for green IoT in sustainable cities)Manage energy and information security for data transmission in sustainable cities using green IoT.Not specifiedImproved energy consumption (21%), throughput (15%), end-to-end latency (12%), path disruption (36%), and network overhead (52%) compared to other solutions.36Rane1(Discussion)Review/position paper (discussion on AI, IoT, big data, blockchain)ConceptualExplore challenges and future directions of AI, IoT, and big data integration for smart/sustainable AEC industry.Analytical and forward-lookingProvided a roadmap for stakeholders; highlighted importance of AI, blockchain, SDGs (no new experimental results).37Alharbi & Aldossary16Mixed Integer Linear Programming (MILP); heuristic algorithmData from thousands of agricultural sensors (implies simulation)Enhance energy efficiency of smart agriculture systems and reduce carbon emissions using edge-fog-cloud architecture.Not specifiedReduced total energy consumption by 36%, carbon emissions by 43%, network traffic by up to 86% compared to traditional cloud. Heuristic comparable to MILP.38Fraga-Lamas et al.2(Case Study/Guideline)Fog computing architecture with AI-equipped IoT nodes (case study)Industry 5.0 smart workshop use case (for energy consumption analysis)Examine aspects of edge AI G-IoT systems; provide guidelines for sustainable development considering edge AI energy consumption.Focus on analysis and guidelinesEvaluated energy consumption, analyzed carbon footprint impact. Provided guidelines for edge AI G-IoT.39Alsamhi et al.17(Review)Review paperReview of existing papersOverview of techniques/strategies for green IoT using UAVs in B5G networks (applications, energy efficiency, sustainability).Limited to reviewed papers and UAV scope in green IoTDiscussed use cases, challenges, and opportunities for UAVs in green IoT (no new experimental results). Table 1Comparison of algorithms and energy optimization results in IIoT task offloading.RowReferenceAlgorithm UsedDataset UsedObjectiveLimitationsResults Obtained1You & Tang5Particle Swarm Optimization (PSO)SimulationEnergy-efficient and low-latency task offloading in IIoT; minimize latency, energy consumption, and execution cost.Not specifiedReduced latency, balanced energy consumption, and more effective resource allocation compared to GA and SA.2Song et al.25Sequential Fractional Programming, Dual Lagrange Decomposition, E-CORA AlgorithmSimulationMinimize weighted total energy consumption of IoT mobile devices by offloading computations to LEO satellites.Not specifiedE-CORA outperformed full offloading/local computation. A specific amount of offloaded bits minimized energy.3Zhu et al.11Novel scheduling optimization algorithm for AI tasks in a heterogeneous smart edge computing frameworkSmall-scale testbed and extensive simulationReduce energy consumption in AI task processing in IIoT by offloading tasks from servers.Not specifiedProposed online scheduling consumed less than 80% of static scheduling energy and 70% of FIFO energy.4Li et al.18Graph Neural Network-based Reinforcement Learning (GNN-A2C: Advantage Actor-Critic with Graph Neural Networks)Implementation with Google Tensorflow, performance analysis (implies simulation)Maximize tasks offloaded to a drone with joint optimization of drone navigation and task offloading in aerial EdgeIoT.Not specified (real-time network state for drones unknown, posing a challenge)GNN-A2C achieved fast convergence and significantly reduced task drop rates.5Zhu et al.15Novel online method for task scheduling optimization in edge and cloudSmart edge computing testbed and extensive simulationOptimize AIoT energy efficiency with smart edge, focusing on edge device and cloud energy.Not specifiedProposed method outperformed other strategies with lower energy consumption.6Chen et al.23Dynamic Task Offloading Algorithm based on Deep Reinforcement Learning (DDTO) (MDP-based)ExperimentsBalance maximizing task completion rate with tolerable latency and minimizing IoT device power consumption.Not specifiedDDTO converged quickly and effectively balanced completion rate and power consumption.7Lin et al.7Online RL-based algorithm with quantum parallelism (MDP-based problem transformation)SimulationsReduce VR equipment energy consumption and ensure smooth VR experience in IIoT with joint offloading, computation, and spectrum allocation optimization.Not specifiedEfficiency in reducing energy consumption and stall rate, improved learning efficiency.8Lv et al.19Industrial gas tracking algorithm based on parallel optimization; two-stage gas tracking algorithm based on state transition modelSimulation and evaluationInter-layer digital twin optimization in IIoT for hazardous gas boundary tracking.Not specifiedEffective gas boundary tracking, reduced network energy. Parallel optimization task success rate > 0.9 in 5 s. State transition algorithm energy consumption 2.11 joules (N = 600), minimum tracking error 0.31.9Hou et al.27Incentive-based online task allocation algorithmNumerical results (simulation)Stimulate collaborative computing between EC servers and IIoT devices to accelerate processing and reduce latency. Maximize system utility.Not specifiedSuperior performance and efficiency; satisfied incentive compatibility, individual rationality, computational efficiency, and feasibility.10Deng et al.12Q-learning (discrete partial offloading), Deep Deterministic Policy Gradient (DDPG) (continuous offloading), both RL-basedSimulation resultsProvide automated partial offloading with minimal latency for better QoS in multi-user IIoT MEC systems.Not specifiedQ-learning reduced latency by 23%, DDPG by 30%.11Wang Dayong et al.3(Review)Review paperAnalysis of existing papersProvide comprehensive understanding of IoT task offloading in MEC; identify problems, techniques, methods, parameters, shortcomings, and future trends.Limited to reviewed papersProvided a perspective on IoT task offloading in MEC to guide researchers.12Gali & Mahamkali28Distributed Deep Meta-Learning Task Offloading (DDMTO) (DNN with backpropagation)Not specified in abstract (likely simulation)Address poor portability of DLTs for task offloading in dynamic IoT environments; ensure effective/efficient offloading in heterogeneous edge-cloud systems.Not specifiedAimed for better performance and reduced computational complexity with heterogeneous cloud collaboration (quantitative results not specified in abstract).13Moghaddasi et al.24Rainbow Deep Q-Network (Rainbow DQN) - Advanced DRLRigorous simulationsOptimize task offloading in D2D-edge-cloud for enhanced operational efficiency in IoT systems.Not specified29.8% increase in energy efficiency, 27.5% latency reduction, 43.1% utility increase compared to state-of-the-art methods.14Firouzi et al.8RL-based offloading techniqueCase study: ECG-based arrhythmia detectionAddress edge-fog-cloud IoT interaction for healthcare applications; effective load distribution.Not specifiedAimed to demonstrate and evaluate model efficiency (quantitative RL results not specified in abstract).15Jiang et al.6Customized Binary Fireworks Algorithm (enhanced with Gaussian mutation and elite selection)Simulation resultsJoint task offloading and resource allocation optimization in a new cloud-edge network architecture for IoT to ensure QoS.Not specifiedConfirmed efficiency; significant performance gains compared to newer offloading strategies.16Huang et al.29Drone-assisted task offloading optimization algorithm (decomposing non-convex problem into two subproblems)Simulation resultsMinimize total network overhead in drone-assisted fog-based IoT with joint optimization of drone path, power, and offloading ratios.Finding local optima for non-convex problemAchieved superior network overhead performance compared to existing algorithms.17Huang et al.30Distributed Offloading Algorithm (DOLA) based on non-cooperative game theoryReal-world dataset-based simulations (Melbourne CBD, Australia)Minimize task completion latency with offloading to multiple BSs in overlapping MEC regions using CoMP.Not specifiedProved Nash equilibrium existence. DOLA showed efficiency and superiority over existing schemes.18Wei et al.31Local Differential Privacy-based Deep Reinforcement Learning (LDP-DRL) with prioritized experience replayExtensive experimentsProvide personalized privacy-aware task offloading in edge-cloud IIoT for automated production.Not specifiedFormal privacy/convergence analysis. Experiments evaluated efficiency, privacy protection, discount factor impact, and cost efficiency (implying positive results).19Dai et al.32Learning-based co-offloading algorithmSimulationsMinimize system cost (latency and migration cost) in D2D-assisted MEC for IIoT with co-offloading to MEC servers or nearby devices.Not specifiedAlgorithm enabled optimal edge node selection without complete information; validated by simulations.20Bradbury et al.10Trust-based system architecture; Beta Reputation System (BRS) as trust model exampleImplementation on real hardware (Zolertia RE-Motes)Facilitate selection of suitable edge nodes for task offloading from resource-constrained IoT devices using trust and considering threats.Risk analysis not conducted, threat modeling excludes trust model itselfDemonstrated architecture feasibility, threat management efficiency. Showed trust value evolution. Identified future work.21Chai et al.26Attention mechanism and collaborative Proximal Policy Optimization (A-PPO) algorithmSimulation resultsImprove offloading efficiency in multi-task MEC for satellite IoT with joint dependent task offloading and resource allocation.Not specifiedA-PPO converged in 25 steps; reduced cost by at least 8.87% compared to baseline algorithms.22Gong4(Review)Review paperAnalysis of 18 papersSystematic review of metaheuristic/nature-inspired algorithms for reliable IoT networks (reliability, security, fault tolerance, sustainability).Limited to reviewed papersMechanisms improved reliability, fault detection, network efficiency, lifetime, and energy optimization. Current studies focus on edge capabilities.23Rana et al.33(Classification/Framework)Proposed framework and classification (framework with predictive models for metaheuristic selection/application)Case study on smart manufacturing (for fitness metric generation)Address issues, principles, and significance of metaheuristic routing for energy-efficient IoT transmission; provide classification and framework.Theoretical framework, not empirically testedProposed theoretical framework and case study; aimed to provide a benchmark solution.24Mir et al.21Fuzzy Chaotic Grasshopper Optimization Algorithm (FLGOA) (GOA enhanced with Lorenz chaos theory and fuzzy logic)Evaluation in two different scenarios (implies simulation)Energy-aware routing in IoT with optimized node sleep/wake scheduling.Not specifiedFLGOA showed 9–16% better network lifetime, 16–22% better residual energy, and 12–16% better coverage rate than FGOA, GOA, GSO.25Wang & Lu14Hybrid multi-objective metaheuristic: Grey Wolf Optimizer + Genetic Algorithm (GWO-GA)Experimental (implies simulation)QoS-aware service discovery and selection in edge-cloud IoT to reduce time, cost, and improve energy efficiency/security.Not specifiedGWO-GA performed 30% better in reducing cost factor compared to other algorithms.26Agarwal & Alam34Maximum Score Pooling (MSP) algorithm (hybrid metaheuristics) for feature selection; MSP-optimized Decision Tree (MSP-DT) classifierExperiments (implies lifelogging dataset)Personalized, edge-optimized lifelogging framework, robust to sensor failure, reducing network traffic and latency.Not specifiedDetected 12 activities with 97.67% average accuracy; reduced transferred data by 47.66%.27Singh et al.35Jaya Algorithm (JA)Benchmark-generated instances for various fog scenariosOptimize fog device deployment (FDD) for maximum network connectivity (FDC) and edge device coverage (EDC).Not specifiedSignificant improvement in EDC and FDC; JA outperformed APSO, LDWPSO, MPA in convergence.28Al-Hashimi et al.36Hybrid metaheuristics (GWO + GA); exploration algorithms (SVM, Naïve Bayes)Cardiac data collected from sensorsDevelop optimized e-health monitoring framework using IoT for lower costs, higher efficiency, accurate data analysis, and better patient care.Not specifiedHybrid GWO + GA outperformed standalone versions. Exploration algorithms with hybrid metaheuristic were more accurate.29Lakhan et al.13Evolutionary Metaheuristic Offloading Scheduling Algorithm (EMOS) frameworkSimulation resultsEfficient, low-latency ICPS for IoT applications addressing task offloading and scheduling to reduce various latencies (processing, offloading, execution, failure, security).Not specifiedReduced local processing latency by 39%, offloading latency by 41%, execution/propagation latency by 40%, failure latency by 43%, security/migration latency by 50% compared to existing ICPS metaheuristics.30Vaiyapuri et al.37FLIDS-BSAFSC: Federated Learning with Bird Swarm Algorithm (BSA) for feature selection and Social Group Optimization with Kernel Extreme Learning Machine for classificationBenchmark datasetsIntrusion Detection System (IDS) based on FL in IoT for classification, detection, and defense against attacks with improved security and privacy.Not specifiedFLIDS-BSAFSC outperformed recent approaches.31Satouf et al.20Adapted Grey Wolf Optimizer (GWO)Simulation for task schedulingSemi-dynamic real-time task scheduling for latency-sensitive IoT in cloud-fog, minimizing energy, cost, and makespan.Not specifiedOutperformed GA, PSO, ABC in makespan, total execution time, cost, and energy consumption.32Tu et al.38Collaborative filtering based on Artificial Bee Colony (ABC) and genetic operatorsComputational tests (implies agricultural service data)Improve accuracy of agricultural service recommendations and reduce MAE in fog-based IoT systems.Not specifiedImproved ranking score (5.8%), accuracy (5%), recall (5.7%), intra-similarity (13%), Hamming distance (4.8%) compared to other methods.33Bali et al.22Hybrid Artificial Bee Colony and Whale Optimization Algorithm (HAWO)Simulation results in MatlabOptimize energy and latency for task allocation to edge processing devices in IoT.Not specifiedHAWO showed promising results compared to ABC and WOA. Compared to benchmarks: 50% energy reduction, 25% latency reduction, 60% best cost improvement.34Nematollahi et al.39Moth Flame Optimization with Opposition-Based Learning (OBLMFO); stacked caching approach; blockchainEvaluation with distinct tasks in Python 3.9 (implies simulation)Efficient task offloading and resource allocation in IoT using fog computing, ensuring accuracy with blockchain.Not specifiedOBLMFO reduced latency factor by 12.18% and energy consumption by 6.22%.35Haseeb et al.9Intelligent and Secure Edge-based Computing Model (ISEC) (deep learning for routing, distributed hashing with chaining for security, symmetric cryptography, DHT)Experimental results (implies simulation for green IoT in sustainable cities)Manage energy and information security for data transmission in sustainable cities using green IoT.Not specifiedImproved energy consumption (21%), throughput (15%), end-to-end latency (12%), path disruption (36%), and network overhead (52%) compared to other solutions.36Rane1(Discussion)Review/position paper (discussion on AI, IoT, big data, blockchain)ConceptualExplore challenges and future directions of AI, IoT, and big data integration for smart/sustainable AEC industry.Analytical and forward-lookingProvided a roadmap for stakeholders; highlighted importance of AI, blockchain, SDGs (no new experimental results).37Alharbi & Aldossary16Mixed Integer Linear Programming (MILP); heuristic algorithmData from thousands of agricultural sensors (implies simulation)Enhance energy efficiency of smart agriculture systems and reduce carbon emissions using edge-fog-cloud architecture.Not specifiedReduced total energy consumption by 36%, carbon emissions by 43%, network traffic by up to 86% compared to traditional cloud. Heuristic comparable to MILP.38Fraga-Lamas et al.2(Case Study/Guideline)Fog computing architecture with AI-equipped IoT nodes (case study)Industry 5.0 smart workshop use case (for energy consumption analysis)Examine aspects of edge AI G-IoT systems; provide guidelines for sustainable development considering edge AI energy consumption.Focus on analysis and guidelinesEvaluated energy consumption, analyzed carbon footprint impact. Provided guidelines for edge AI G-IoT.39Alsamhi et al.17(Review)Review paperReview of existing papersOverview of techniques/strategies for green IoT using UAVs in B5G networks (applications, energy efficiency, sustainability).Limited to reviewed papers and UAV scope in green IoTDiscussed use cases, challenges, and opportunities for UAVs in green IoT (no new experimental results). Comparison of algorithms and energy optimization results in IIoT task offloading. Recent advancements in Industrial IoT (IIoT) and edge computing environments have increasingly focused on the critical intersection of intelligent resource optimization, trust management, and cybersecurity. On the optimization front, recent studies have introduced advanced meta-learning frameworks for hyperparameter optimization in resource-constrained IoT devices40 and secure, energy-aware high-order clustering algorithms for efficient resource allocation41. Additionally, closed-loop digital twin architectures have been proposed to optimize energy-efficient scheduling in manufacturing systems42. Simultaneously, the inherent vulnerabilities of decentralized Edge-IIoT networks have driven the rapid development of robust security and trust evaluation mechanisms. Notable state-of-the-art solutions include dual-modal physical fingerprints for adaptive device authentication43, dynamic trust evaluation utilizing heterogeneous graph convolutional networks44, and distributed trust-driven intelligence for secure edge-level predictions45. Furthermore, hybrid deep learning architectures—such as CNN-BiLSTM-Lite + + and enhanced XGBoost models—have demonstrated significant potential in multi-class cyber threat and intrusion detection within wireless sensor networks and IIoT environments46,47. While these contemporary approaches substantially advance specific domains of IIoT security and resource management, they often address these challenges in isolation. Our proposed Quantum-Fuzzy NSGA-II framework distinctively bridges this gap by simultaneously tackling multi-objective resource allocation and deterministic system stability under extreme stress conditions, reinforced by Lyapunov optimization. Synthesis of literature and problem statement A critical analysis of the current literature (2021–2025) reveals three persistent and fundamental gaps in existing IIoT task offloading architectures, which this research aims to systematically address: Algorithmic Stagnation and Local Optima Entrapment: Conventional meta-heuristics (e.g., PSO, GA) demonstrate a pronounced vulnerability to premature convergence. When deployed in the complex, high-dimensional search spaces of ultra-dense IIoT networks, their stochastic exploration mechanisms often fail, leading to entrapment in local optima and sub-optimal resource allocation.Oversimplification of the Multi-Objective Landscape: The predominant use of static scalarization techniques, particularly weighted-sum methods, represents a critical methodological flaw. By collapsing distinct and often competing objectives (e.g., energy, latency) into a single scalar metric, these approaches fail to produce a true Pareto-optimal front. This severely limits an orchestrator’s ability to navigate the complex, non-linear trade-offs inherent in dynamic operational environments.Lack of Validation under Real-World Stress Conditions: The empirical validation of existing frameworks is overwhelmingly confined to simulations under benign conditions, often using synthetic datasets with uniform probability distributions. There is a significant void in the literature concerning architectures that can prove their resilience and maintain QoS stability when subjected to extreme, real-world stress vectors, such as massive, concurrent task injection bursts (e.g., > 500 simultaneous requests) within high-fidelity spatial and temporal topologies. Algorithmic Stagnation and Local Optima Entrapment: Conventional meta-heuristics (e.g., PSO, GA) demonstrate a pronounced vulnerability to premature convergence. When deployed in the complex, high-dimensional search spaces of ultra-dense IIoT networks, their stochastic exploration mechanisms often fail, leading to entrapment in local optima and sub-optimal resource allocation. Algorithmic Stagnation and Local Optima Entrapment: Conventional meta-heuristics (e.g., PSO, GA) demonstrate a pronounced vulnerability to premature convergence. When deployed in the complex, high-dimensional search spaces of ultra-dense IIoT networks, their stochastic exploration mechanisms often fail, leading to entrapment in local optima and sub-optimal resource allocation. Oversimplification of the Multi-Objective Landscape: The predominant use of static scalarization techniques, particularly weighted-sum methods, represents a critical methodological flaw. By collapsing distinct and often competing objectives (e.g., energy, latency) into a single scalar metric, these approaches fail to produce a true Pareto-optimal front. This severely limits an orchestrator’s ability to navigate the complex, non-linear trade-offs inherent in dynamic operational environments. Oversimplification of the Multi-Objective Landscape: The predominant use of static scalarization techniques, particularly weighted-sum methods, represents a critical methodological flaw. By collapsing distinct and often competing objectives (e.g., energy, latency) into a single scalar metric, these approaches fail to produce a true Pareto-optimal front. This severely limits an orchestrator’s ability to navigate the complex, non-linear trade-offs inherent in dynamic operational environments. Lack of Validation under Real-World Stress Conditions: The empirical validation of existing frameworks is overwhelmingly confined to simulations under benign conditions, often using synthetic datasets with uniform probability distributions. There is a significant void in the literature concerning architectures that can prove their resilience and maintain QoS stability when subjected to extreme, real-world stress vectors, such as massive, concurrent task injection bursts (e.g., > 500 simultaneous requests) within high-fidelity spatial and temporal topologies. Lack of Validation under Real-World Stress Conditions: The empirical validation of existing frameworks is overwhelmingly confined to simulations under benign conditions, often using synthetic datasets with uniform probability distributions. There is a significant void in the literature concerning architectures that can prove their resilience and maintain QoS stability when subjected to extreme, real-world stress vectors, such as massive, concurrent task injection bursts (e.g., > 500 simultaneous requests) within high-fidelity spatial and temporal topologies. The proposed QFLN-AMRO architecture is engineered to overcome these deficiencies. It replaces rudimentary scalarized models with a true Pareto-dominant evolutionary engine (Fractional-Order NSGA-II) to navigate the multi-objective landscape effectively. To counteract algorithmic stagnation, it integrates Quantum-Fuzzy adaptive mechanics, ensuring robust population diversity and enabling escape from local optima. This synergistic design establishes a resilient orchestration framework capable of maintaining deterministic performance under extreme workloads, thus bridging the critical gap between theoretical optimization and practical, mission-critical deployment in next-generation IIoT ecosystems. Methodology This section details the comprehensive scientific methodology underpinning the design and validation of the proposed QFLN-AMRO framework. Our primary objective is to ensure a transparent, rigorous, and reproducible research protocol. The methodology is structured as follows. First, we define the high-fidelity simulation environment, including the real-world datasets and spatial topologies that form the foundation of our empirical analysis. This is followed by an exposition of the data preprocessing and feature engineering pipeline, where we formalize the mathematical models for geospatial distance computation and data normalization, which are critical for algorithmic stability and accuracy. The core of this section is a detailed architectural exposition of the Quantum-Fuzzy-Lyapunov-driven NSGA-II for Adaptive Multi-objective Resource Orchestration (QFLN-AMRO). We dissect its innovative multi-layered design, beginning with the Fractional-Order NSGA-II engine that drives the Pareto-based optimization. We then elaborate on the quantum-fuzzy cognitive layer responsible for adaptively tuning evolutionary parameters and the Lyapunov-based stability mechanism that guarantees QoS constraint adherence. To establish a robust performance benchmark, the framework is evaluated against a comprehensive suite of baseline algorithms, including canonical meta-heuristics (GA, PSO) and a state-of-the-art Deep Reinforcement Learning (DRL) model. Finally, we formally define the quantitative evaluation metrics used to conduct a rigorous comparative analysis, focusing on system-level energy efficiency, task execution latency, network stability, and overall QoS compliance. The comprehensive hierarchical network topology, detailing the multi-tier system model and the communication architecture, is explicitly illustrated in Fig. 1. Fig. 1Hierarchical Network Topology and System Model. The architecture is structured across three primary tiers: IIoT End Devices (bottom), Edge/MEC Tier (middle), and Global Cloud (top). Communication relies on Non-Orthogonal Multiple Access (NOMA) coupled with Ultra-Reliable Low-Latency Communication (URLLC) to ensure strict latency bounds and spectral efficiency under heavy workloads. Hierarchical Network Topology and System Model. The architecture is structured across three primary tiers: IIoT End Devices (bottom), Edge/MEC Tier (middle), and Global Cloud (top). Communication relies on Non-Orthogonal Multiple Access (NOMA) coupled with Ultra-Reliable Low-Latency Communication (URLLC) to ensure strict latency bounds and spectral efficiency under heavy workloads. Datasets and simulation environment The empirical validation of the QFLN-AMRO framework was conducted within a high-fidelity simulation testbed, meticulously designed to mirror the complex dynamics of a real-world IIoT ecosystem. This environment is architected by integrating three specialized, publicly available datasets, which collectively model the physical network topology and the computational workload. A detailed specification of these datasets is provided in Table 2. Physical Infrastructure Modeling: The geospatial distribution of the edge infrastructure is modeled using the EUA (Edge Cloud Ecosystem) Dataset. This dataset provides the precise latitude and longitude coordinates for 20 heterogeneous edge servers. The utilization of real-world geographic data is a critical methodological choice, enabling the accurate calculation of great-circle distances between task origins and compute nodes. This, in turn, facilitates a realistic estimation of network propagation delays, moving beyond the oversimplified, and often inaccurate, assumptions common in prior studies. Computational Workload Synthesis: The computational workload is synthesized from two complementary sources to capture both the specific nature and the operational scale of IIoT demands. The Edge-IIoT Dataset provides the foundational task characteristics, offering granular parameters such as data payload size (input data size) and required computational cycles (task length). These parameters are essential for accurately modeling processing delays and energy expenditure on the edge servers. To rigorously evaluate the framework’s performance under stress and at scale, these IIoT task profiles are augmented with arrival patterns and intensity levels derived from the Google Cluster Trace. This synthesis results in a demanding and dynamic workload profile, featuring the concurrent injection of 500 heterogeneous computational tasks. This methodology ensures a robust stress test of the proposed orchestration algorithm’s scalability, stability, and resource management efficacy under high-throughput conditions. Table 2Detailed characteristics and features of the integrated datasets.DatasetRole in StudyKey Features/AttributesData GranularityGoogle Cluster TracesPredictive Resilience Trainingreq_cpus, req_memory, priority, avg_cpus, avg_memory, max_cpus, max_memoryTask-level (Event traces)EUA DatasetInfrastructure TopologyServer_ID, Latitude, Longitude, User_ID, Location_CoordinatesGeospatial points (Melbourne CBD)Edge-IIoTIoT Traffic ModelingFlow_Duration, Packet_Size, Arrival_Rate, Traffic_Type (Labels y)Network Packet-levelDATA7Dynamic Workload ProfilingTimestamp, Workload_Intensity, CPU_Usage_HistoryTemporal (Time-series)Feature Pre-processing: All continuous features (e.g., \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:avg\_cpus$$\end{document}) are normalized using the Standard Scaler to achieve zero mean and unit variance:. Table 2Detailed characteristics and features of the integrated datasets.DatasetRole in StudyKey Features/AttributesData GranularityGoogle Cluster TracesPredictive Resilience Trainingreq_cpus, req_memory, priority, avg_cpus, avg_memory, max_cpus, max_memoryTask-level (Event traces)EUA DatasetInfrastructure TopologyServer_ID, Latitude, Longitude, User_ID, Location_CoordinatesGeospatial points (Melbourne CBD)Edge-IIoTIoT Traffic ModelingFlow_Duration, Packet_Size, Arrival_Rate, Traffic_Type (Labels y)Network Packet-levelDATA7Dynamic Workload ProfilingTimestamp, Workload_Intensity, CPU_Usage_HistoryTemporal (Time-series)Feature Pre-processing: All continuous features (e.g., \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:avg\_cpus$$\end{document}) are normalized using the Standard Scaler to achieve zero mean and unit variance:. Detailed characteristics and features of the integrated datasets. Feature Pre-processing: All continuous features (e.g., \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:avg\_cpus$$\end{document}) are normalized using the Standard Scaler to achieve zero mean and unit variance:. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:z=\frac{x-\mu\:}{\sigma\:}$$\end{document} Target Variable: For the resilience phase, the binary target failed is derived from Google Cluster traces to train the Ensemble Voting Classifier (RF, XGBoost, LightGBM) Table 3. Table 3Dataset specifications.Dataset NameRole in StudyKey Characteristics UtilizedQuantityEUA DatasetEdge Infrastructure ModelingGeographic Coordinates (Latitude, Longitude)20 ServersEdge-IIoT DatasetPrimary Workload GenerationTask Payload, Computational RequirementsUp to 500 TasksGoogle Cluster TraceWorkload Pattern InspirationTask arrival patterns and resource demandsN/A (Pattern Study) Table 3Dataset specifications.Dataset NameRole in StudyKey Characteristics UtilizedQuantityEUA DatasetEdge Infrastructure ModelingGeographic Coordinates (Latitude, Longitude)20 ServersEdge-IIoT DatasetPrimary Workload GenerationTask Payload, Computational RequirementsUp to 500 TasksGoogle Cluster TraceWorkload Pattern InspirationTask arrival patterns and resource demandsN/A (Pattern Study) Dataset specifications. Data preprocessing and feature engineering Raw data extracted from the aforementioned sources are not directly amenable to the optimization model and require preprocessing to construct a coherent and mathematically sound problem space. This involves two critical steps: geospatial distance calculation to inform the delay cost matrix, and data normalization to ensure numerical stability. The primary feature engineered for this study is the physical distance between a task’s origin and a potential edge server destination. A naive Euclidean distance calculation is mathematically inappropriate as it fails to account for the Earth’s curvature, leading to significant inaccuracies over larger distances. To rectify this, we employ the Haversine formula, which calculates the shortest distance between two points on the surface of a sphere. The distance dbetween two points with geographical coordinates \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\left({\phi\:}_{1},{\lambda\:}_{1}\right)$$\end{document}and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\left({\phi\:}_{2},{\lambda\:}_{2}\right)$$\end{document}is mathematically formulated as:1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$a = \sin ^{2} \left( {\frac{{\Delta \phi }}{2}} \right) + \cos \phi _{1} \cos \phi _{2} \sin ^{2} \left( {\frac{{\Delta \lambda }}{2}} \right)$$\end{document}2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$c = 2 \cdot a\tan 2\left( {\sqrt a ,\sqrt {1 - a} } \right)$$\end{document} 3d = R · cHere, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\Delta\:}\phi\:$$\end{document}is the difference in latitude, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\Delta\:}\lambda\:$$\end{document}is the difference in longitude, Ris the Earth’s mean radius (approximately 6,371 km), and atan2 is the two-argument arctangent function. The resulting distance matrix, d_mat, provides a realistic foundation for modeling the propagation delay component of our first objective function. Furthermore, to mitigate any potential bias arising from disparate scales between the objective functions (e.g., delay measured in milliseconds versus energy in Joules), we apply Min-Max normalization. This standard procedure rescales features to a fixed range of [0, 1]. Although NSGA-II’s dominance-based selection is inherently robust to objective scaling, normalization is a good practice for maintaining numerical stability and facilitating comparative analysis. The transformation for a value xis given by:4\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$x_{{norm}} = \frac{{x - x_{{\min }} }}{{x_{{\max }} - x_{{\min }} }}$$\end{document} Proposed framework: The QFLN-AMRO architecture To address the inherent limitations of conventional task offloading methods in heterogeneous edge-cloud environments, this paper introduces the QFLN-AMRO framework. Traditional optimization approaches often rely on static scalarization—a technique that aggregates multiple, often conflicting, objectives like latency and energy into a single cost function using predefined weights. This approach suffers from a fundamental limitation: it artificially collapses the multi-dimensional decision space. Consequently, it often yields a single, potentially biased solution that fails to capture the complex, non-linear trade-offs inherent in dynamic network operations, leading to sub-optimal performance. In stark contrast, the QFLN-AMRO is architected as a native Multi-Objective Optimization Problem (MOOP). It forgoes scalarization entirely, instead focusing on computing the complete Pareto-optimal front. This front comprises a set of non-dominated solutions, where each solution represents an optimal and distinct trade-off between the objectives. The availability of this entire solution set provides a critical strategic advantage: it grants system orchestrators the flexibility to select the most suitable operating point based on dynamic network conditions or shifting Service-Level Agreements (SLAs)—for instance, prioritizing ultra-low latency for critical industrial tasks or emphasizing energy conservation during off-peak hours—without the prohibitive computational cost of re-running the entire optimization. The framework’s architecture is methodically organized into three sequential layers, ensuring a structured and robust optimization process: Mathematical Formulation Layer: This foundational layer formally defines the optimization problem. It establishes the precise mathematical models, objective functions for minimizing end-to-end latency and total energy consumption, and the operational constraints of the IIoT ecosystem (e.g., computational capacity of nodes, network bandwidth).Multi-Objective Evolutionary Engine: At the core of the framework lies the Non-dominated Sorting Genetic Algorithm II (NSGA-II). We employ this state-of-the-art evolutionary algorithm as the primary optimization engine. NSGA-II is renowned for its efficiency in exploring complex, high-dimensional search spaces and its robust convergence to the true Pareto front, making it exceptionally well-suited for this problem domain.Deterministic Decision Support Layer: Once the Pareto front is generated by the evolutionary engine, this layer provides an autonomous mechanism for selecting the final, optimal solution for deployment. It analyzes the geometric properties of the solution set, employing techniques such as Hypervolume analysis and Knee Point identification to algorithmically pinpoint the most balanced and effective orchestration strategy from the set of optimal trade-offs. Mathematical Formulation Layer: This foundational layer formally defines the optimization problem. It establishes the precise mathematical models, objective functions for minimizing end-to-end latency and total energy consumption, and the operational constraints of the IIoT ecosystem (e.g., computational capacity of nodes, network bandwidth). Mathematical Formulation Layer: This foundational layer formally defines the optimization problem. It establishes the precise mathematical models, objective functions for minimizing end-to-end latency and total energy consumption, and the operational constraints of the IIoT ecosystem (e.g., computational capacity of nodes, network bandwidth). Multi-Objective Evolutionary Engine: At the core of the framework lies the Non-dominated Sorting Genetic Algorithm II (NSGA-II). We employ this state-of-the-art evolutionary algorithm as the primary optimization engine. NSGA-II is renowned for its efficiency in exploring complex, high-dimensional search spaces and its robust convergence to the true Pareto front, making it exceptionally well-suited for this problem domain. Multi-Objective Evolutionary Engine: At the core of the framework lies the Non-dominated Sorting Genetic Algorithm II (NSGA-II). We employ this state-of-the-art evolutionary algorithm as the primary optimization engine. NSGA-II is renowned for its efficiency in exploring complex, high-dimensional search spaces and its robust convergence to the true Pareto front, making it exceptionally well-suited for this problem domain. Deterministic Decision Support Layer: Once the Pareto front is generated by the evolutionary engine, this layer provides an autonomous mechanism for selecting the final, optimal solution for deployment. It analyzes the geometric properties of the solution set, employing techniques such as Hypervolume analysis and Knee Point identification to algorithmically pinpoint the most balanced and effective orchestration strategy from the set of optimal trade-offs. Deterministic Decision Support Layer: Once the Pareto front is generated by the evolutionary engine, this layer provides an autonomous mechanism for selecting the final, optimal solution for deployment. It analyzes the geometric properties of the solution set, employing techniques such as Hypervolume analysis and Knee Point identification to algorithmically pinpoint the most balanced and effective orchestration strategy from the set of optimal trade-offs. The internal workflow of the proposed logic, illustrating the dynamic interaction between the quantum-fuzzy learning components and the multi-objective routing optimization, is meticulously depicted in Fig. 2. Fig. 2Internal flow of the proposed QFLN-AMRO logic. Raw workload data and network states are ingested and passed into the Quantum-Fuzzy Learning Network (QFLN), where quantum superposition states \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\left| {q_{i} } \right\rangle$$\end{document} generate highly adaptive weights through fuzzy rule logic. These weights dynamically guide the Adaptive Multi-Objective Routing Optimization (AMRO) module, strictly driven by the NSGA-II evolutionary algorithm ensuring mathematically proven Lyapunov stability. Internal flow of the proposed QFLN-AMRO logic. Raw workload data and network states are ingested and passed into the Quantum-Fuzzy Learning Network (QFLN), where quantum superposition states \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\left| {q_{i} } \right\rangle$$\end{document} generate highly adaptive weights through fuzzy rule logic. These weights dynamically guide the Adaptive Multi-Objective Routing Optimization (AMRO) module, strictly driven by the NSGA-II evolutionary algorithm ensuring mathematically proven Lyapunov stability. Layer 1: System modeling and problem formulation The foundational phase constructs a deterministic mathematical model of the edge ecosystem. Let the set of independent computational tasks be denoted as \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:T=\{{t}_{1},{t}_{2},\dots\:,{t}_{n}\}$$\end{document}and the array of heterogeneous edge servers as \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:S=\{{s}_{1},{s}_{2},\dots\:,{s}_{m}\}$$\end{document}. The orchestration objective is to map tasks to servers optimally, mathematically defined by the binary decision matrix \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:X\in\:\{\mathrm{0,1}{\}}^{n\times\:m}$$\end{document}:5\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$x_{{ij}} = \left\{ {\begin{array}{*{20}c} {1,} & {if~task~t_{i} ~is~assigned~to~server~s_{j} } \\ {0,} & {otherwise} \\ \end{array} } \right.$$\end{document} To guarantee execution integrity and prevent task duplication or loss, a strict mapping constraint is enforced, ensuring each task is offloaded to exactly one physical node:6\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathop \sum \limits_{{j = 1}}^{m} x_{{ij}} = 1,\forall i \in \left\{ {1, \ldots ,n} \right\}$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\sum\:_{j=1}^{m}{x}_{ij}=1,\forall\:i\in\:\{1,\dots\:,n\}\left(6\right)$$\end{document}The total latency experienced by task \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{t}_{i}$$\end{document}on server \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{s}_{j}$$\end{document}, denoted as \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{D}_{ij}$$\end{document}, is the aggregate of its propagation delay (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{D}_{ij}^{\mathrm{prop}}$$\end{document}) and processing delay (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{D}_{ij}^{\mathrm{proc}}$$\end{document}). Propagation latency is governed by the geospatial distance \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{d}_{ij}$$\end{document}, signal propagation speed \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{v}_{\mathrm{net}}$$\end{document}, and network bandwidth Bfor a task size \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\delta\:}_{i}$$\end{document}:7\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$D_{{ij}}^{{prop}} = \frac{{d_{{ij}} }}{{v_{{net}} }} + \frac{{\delta _{i} }}{B}$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{D}_{ij}^{\mathrm{prop}}=\frac{{d}_{ij}}{{v}_{\mathrm{net}}}+\frac{{\delta\:}_{i}}{B}\left(7\right)$$\end{document}Processing latency is strictly a function of the task’s computational footprint \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{c}_{i}$$\end{document}(e.g., Million Instructions) relative to the deterministic processing capacity of the assigned server \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{p}_{j}$$\end{document}(e.g., MIPS): 8\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$D_{{ij}}^{{proc}} = \frac{{c_{i} }}{{p_{j} }}$$\end{document}Consequently, the first orthogonal objective, Total System Delay (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{f}_{\mathrm{delay}}$$\end{document}), is formulated to minimize the cumulative latency footprint across the entire topology:9\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$Minimize~f_{{delay}} \left( X \right) = \mathop \sum \limits_{{i = 1}}^{n} \mathop \sum \limits_{{j = 1}}^{m} x_{{ij}} \left( {D_{{ij}}^{{prop}} + D_{{ij}}^{{proc}} } \right)$$\end{document} Conversely, the energy expenditure \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{E}_{ij}$$\end{document}incurred by server \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{s}_{j}$$\end{document}is modeled using its specific power consumption coefficient \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\gamma\:}_{j}$$\end{document}(Joules/sec):10\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$E_{{ij}} = \gamma _{j} \cdot D_{{ij}}^{{proc}} = \gamma _{j} \left( {\frac{{c_{i} }}{{p_{j} }}} \right)$$\end{document} This derives the second orthogonal objective, Total Energy Consumption (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{f}_{\mathrm{energy}}$$\end{document}), aimed at maximizing environmental sustainability:11\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$Minimize~f_{{energy}} \left( X \right) = \mathop \sum \limits_{{i = 1}}^{n} \mathop \sum \limits_{{j = 1}}^{m} x_{{ij}} E_{{ij}}$$\end{document} The holistic orchestration challenge is thereby synthesized into a rigorous bi-objective vector optimization problem, constrained by Eq. (6):12\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$Minimize~F\left( X \right) = \left[ {f_{{delay}} \left( X \right),f_{{energy}} \left( X \right)} \right]$$\end{document} Layer 2: The evolutionary optimization engine To resolve the MOOP delineated above, QFLN-AMRO deploys a mathematically rigorous NSGA-II core. The operational sequence relies on four fundamental principles to guarantee convergence without collapsing the objective dimensions: 1. Pareto Dominance: A solution \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{X}_{1}$$\end{document}is mathematically proven to strictly dominate \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{X}_{2}$$\end{document}(denoted as \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{X}_{1}\prec\:{X}_{2}$$\end{document}) if it equals or outperforms \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{X}_{2}$$\end{document}across all dimensions and strictly improves upon it in at least one:13\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$X_{1} \prec X_{2} \left| \Leftrightarrow \right|\forall k \in \left\{ {1,2} \right\},f_{k} \left( {X_{1} } \right) \le f_{k} \left( {X_{2} } \right) \wedge \exists k^{'} \in \left\{ {1,2} \right\},f_{{k^{'} }} \left( {X_{1} } \right) < f_{{k^{,} }} \left( {X_{2} } \right)$$\end{document} 2. Non-dominated Sorting: The population is iteratively partitioned into hierarchical fronts (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{F}_{1},{F}_{2},\dots\:$$\end{document}). The absolute optimal boundary, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{F}_{1}$$\end{document}, contains exclusively non-dominated solutions.14\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$F_{1} = \left\{ {X \in P{\mid }\neg \exists X^{\prime} \in P,X^{\prime} \prec X} \right\}$$\end{document} 3. Crowding Distance: To prevent evolutionary stagnation and ensure widespread coverage of the Pareto front, a spatial diversity metric \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{d}_{\mathrm{crowd}}$$\end{document}is enforced. For solution iin front F.15\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$d_{{crowd}} \left( i \right) = \mathop \sum \limits_{{k = 1}}^{2} \frac{{f_{k} \left( {i + 1} \right) - f_{k} \left( {i - 1} \right)}}{{f_{k}^{{\max }} - f_{k}^{{\min }} }}$$\end{document} 4. Genetic Operators: Exploration is driven by Simulated Binary Crossover (SBX) and Polynomial Mutation. For parents \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{P}_{1},{P}_{2}$$\end{document}, offspring \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{O}_{1},{O}_{2}$$\end{document}are generated via a spread factor\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\beta\:$$\end{document}.16\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$O_{1} = 0.5 \cdot \left[ {\left( {1 + \beta } \right)P_{1} + \left( {1 - \beta } \right)P_{2} } \right]$$\end{document}17\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$O_{2} = 0.5 \cdot \left[ {\left( {1 - \beta } \right)P_{1} + \left( {1 + \beta } \right)P_{2} } \right]$$\end{document} Mutation applies a controlled perturbation \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\delta\:$$\end{document}to gene g:18\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$g^{\prime} = g + \left( {g_{{upper}} - g_{{lower}} } \right) \cdot \delta$$\end{document} As detailed in Algorithm 1, the core evolutionary engine leverages the Non-dominated Sorting Genetic Algorithm II (NSGA-II) framework to identify the Pareto-optimal front for the proposed multi-objective orchestration problem. The procedure is conceptually and mathematically structured into two fundamental phases: Initialization Phase: The algorithm commences by generating an initial population of candidate solutions. Each individual is evaluated against the predefined multi-objective fitness functions—specifically, energy consumption and system latency. Subsequently, an initial non-dominated sorting mechanism is applied to stratify the population into distinct Pareto fronts based on their dominance hierarchy.Evolutionary Loop and Elitism: In the iterative evolutionary phase, parent selection is executed utilizing a binary tournament selection operator, which prioritizes solutions based on a dual-criterion: lower Pareto rank and higher crowding distance. Offspring are then systematically generated by applying Simulated Binary Crossover (SBX) and Polynomial Mutation operators. Upon evaluating the objective functions of the newly generated offspring, a combined population comprising both parents and offspring is constructed. To enforce strict elitism while preserving topological diversity along the Pareto front, the algorithm employs a crowding distance assignment mechanism. This process truncates the combined population, propagating only the most highly fit individuals to the subsequent generation. Initialization Phase: The algorithm commences by generating an initial population of candidate solutions. Each individual is evaluated against the predefined multi-objective fitness functions—specifically, energy consumption and system latency. Subsequently, an initial non-dominated sorting mechanism is applied to stratify the population into distinct Pareto fronts based on their dominance hierarchy. Initialization Phase: The algorithm commences by generating an initial population of candidate solutions. Each individual is evaluated against the predefined multi-objective fitness functions—specifically, energy consumption and system latency. Subsequently, an initial non-dominated sorting mechanism is applied to stratify the population into distinct Pareto fronts based on their dominance hierarchy. Evolutionary Loop and Elitism: In the iterative evolutionary phase, parent selection is executed utilizing a binary tournament selection operator, which prioritizes solutions based on a dual-criterion: lower Pareto rank and higher crowding distance. Offspring are then systematically generated by applying Simulated Binary Crossover (SBX) and Polynomial Mutation operators. Upon evaluating the objective functions of the newly generated offspring, a combined population comprising both parents and offspring is constructed. To enforce strict elitism while preserving topological diversity along the Pareto front, the algorithm employs a crowding distance assignment mechanism. This process truncates the combined population, propagating only the most highly fit individuals to the subsequent generation. Evolutionary Loop and Elitism: In the iterative evolutionary phase, parent selection is executed utilizing a binary tournament selection operator, which prioritizes solutions based on a dual-criterion: lower Pareto rank and higher crowding distance. Offspring are then systematically generated by applying Simulated Binary Crossover (SBX) and Polynomial Mutation operators. Upon evaluating the objective functions of the newly generated offspring, a combined population comprising both parents and offspring is constructed. To enforce strict elitism while preserving topological diversity along the Pareto front, the algorithm employs a crowding distance assignment mechanism. This process truncates the combined population, propagating only the most highly fit individuals to the subsequent generation. This cyclical evolutionary process terminates upon satisfying the maximum generation limit (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{n}_{gen}$$\end{document}), ultimately extracting the first non-dominated front (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{F}_{1}$$\end{document}). This optimal set provides the foundational decision space for the orchestration environment. Algorithm 1Core evolutionary engine (NSGA-II). Core evolutionary engine (NSGA-II). Layer 3: Pareto front analysis and decision support Post-convergence, the framework outputs the final Pareto front (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:P{F}_{\mathrm{final}}$$\end{document}). The absolute quality and convergence depth of this frontier are mathematically validated using the Hypervolume (HV) indicator, which integrates the objective space volume dominated by the front relative to a worst-case reference point R:19\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$HV\left( {PF_{{final}} ,R} \right) = Volume\left( {\bigcup\limits_{{X \in PF_{{final}} }} {\left[ {f_{1} \left( X \right),r_{1} } \right] \times \left[ {f_{2} \left( X \right),r_{2} } \right]} } \right)$$\end{document} To automate deterministic decision-making for single-deployment architectures, the framework geometrically extracts the Knee Point. This point represents the most aggressive trade-off optimum—maximizing gains in one dimension prior to triggering exponential degradation in the other:20\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$X_{{knee}} = \arg \mathop {\max }\limits_{{X \in PF_{{final}} }} \left( {Dis\tan ce\left( {X,Line\left( {X_{{delay}}^{{\min }} ,X_{{energy}}^{{\min }} } \right)} \right)} \right)$$\end{document} Integration of Lyapunov-Guided resilience mechanism The QFLN-AMRO framework employs a two-tier optimization strategy to ensure both long-term efficiency and real-time stability, particularly for stringent URLLC requirements. While the QFLN-tuned NSGA-II engine serves as the primary, proactive optimization tool for discovering the Pareto-optimal trade-offs between energy and latency, the Lyapunov-guided mechanism acts as a complementary, reactive control layer. Its integration is crucial for maintaining network stability under unforeseen traffic surges and preventing violations of latency guarantees. The synergy is realized by incorporating the Lyapunov drift-plus-penalty function directly into the fitness evaluation process of the NSGA-II algorithm. Specifically, we define a Lyapunov function based on the virtual queue lengths of the network nodes. The objective of the Lyapunov optimization is to minimize the “drift” of this function, which corresponds to stabilizing the queues and preventing data backlogs. This drift metric is introduced as a dynamic penalty term into the NSGA-II’s multi-objective fitness function. Consequently, any candidate solution (i.e., a specific resource allocation configuration) that leads to an increase in queue length, thereby threatening URLLC latency bounds, receives a higher penalty score. This forces the evolutionary search to discard unstable solutions and prioritize those that not only reside on the energy-latency Pareto front but also ensure queue stability. In essence, the Lyapunov mechanism acts as a real-time stability-aware constraint, guiding the QFLN-NSGA-II synergy toward robust solutions that guarantee URLLC performance even in high-concurrency scenarios. Baseline algorithms and hyperparameter configuration To rigorously validate the computational dominance of the proposed architecture, it is subjected to intensive comparative analysis against seven state-of-the-art algorithms, explicitly spanning two distinct scientific paradigms. The first category comprises elite Metaheuristics (PSO, GWO, MFO, RHO), representing peak single-objective analytical capacity. The second encompasses advanced Deep Reinforcement Learning (DRL) agents (PPO, GAT, DT-TrPPO-ELS), representing the vanguard of sequential, neural-driven decision-making. The comprehensive architectural hyperparameter matrices are strictly defined in Tables 4 and 5 to ensure absolute experimental reproducibility and algorithmic fairness. Table 4Core hyperparameter matrix for the proposed engine.ParameterValueAnalytical Description & Justificationpop_size 50 Defines the spatial population density per generation; precisely calibrated to balance evolutionary diversity against computational overhead.n_gen 100 The deterministic termination threshold; mathematically proven sufficient for deep front convergence without triggering overfitting.seed 42 Cryptographic seed for the pseudo-random number generator, ensuring \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:100\mathrm{\%}$$\end{document}structural reproducibility of stochastic behaviors.selectionTournamentDeploys a highly aggressive parent selection matrix based concurrently on Pareto rank and crowding distance (preventing local optima).crossoverSBXSimulated Binary Crossover mechanism explicitly tailored for discrete integer matrices (task mapping).eta_crossover 30 The spatial distribution index enforcing tight genetic fidelity between elite parent matrices and generated offspring.mutationPolynomialIntroduces micro-perturbations across the topological gene pool to strictly prevent premature evolutionary stagnation.eta_mutation 20 Controls the mutational magnitude, safeguarding against chaotic degradation of high-performing genetic sequences. Table 4Core hyperparameter matrix for the proposed engine.ParameterValueAnalytical Description & Justificationpop_size 50 Defines the spatial population density per generation; precisely calibrated to balance evolutionary diversity against computational overhead.n_gen 100 The deterministic termination threshold; mathematically proven sufficient for deep front convergence without triggering overfitting.seed 42 Cryptographic seed for the pseudo-random number generator, ensuring \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:100\mathrm{\%}$$\end{document}structural reproducibility of stochastic behaviors.selectionTournamentDeploys a highly aggressive parent selection matrix based concurrently on Pareto rank and crowding distance (preventing local optima).crossoverSBXSimulated Binary Crossover mechanism explicitly tailored for discrete integer matrices (task mapping).eta_crossover 30 The spatial distribution index enforcing tight genetic fidelity between elite parent matrices and generated offspring.mutationPolynomialIntroduces micro-perturbations across the topological gene pool to strictly prevent premature evolutionary stagnation.eta_mutation 20 Controls the mutational magnitude, safeguarding against chaotic degradation of high-performing genetic sequences. Core hyperparameter matrix for the proposed engine. Table 5Structural hyperparameter configurations for baseline benchmarks.Algorithmic ParadigmBaseline AlgorithmArchitectural ParametersConfiguration ValuesProposed Framework QFLN-AMRO (NSGA-II Core) pop_size/n_gen \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:50/\:100$$\end{document} Metaheuristic VanguardPSO, GWO, MFO, RHOpop_size/epochs \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:30/\:50$$\end{document} DRL/Neural AgentsPPOepochs/learning_rate/\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\gamma\:$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:200/\:0.001/\:0.99$$\end{document} GATepochs/learning_rate \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:200/\:0.005$$\end{document} DT-TrPPO-ELSepochs/batch_size \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:200/\:64$$\end{document} Table 5Structural hyperparameter configurations for baseline benchmarks.Algorithmic ParadigmBaseline AlgorithmArchitectural ParametersConfiguration ValuesProposed Framework QFLN-AMRO (NSGA-II Core) pop_size/n_gen \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:50/\:100$$\end{document} Metaheuristic VanguardPSO, GWO, MFO, RHOpop_size/epochs \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:30/\:50$$\end{document} DRL/Neural AgentsPPOepochs/learning_rate/\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\gamma\:$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:200/\:0.001/\:0.99$$\end{document} GATepochs/learning_rate \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:200/\:0.005$$\end{document} DT-TrPPO-ELSepochs/batch_size \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:200/\:64$$\end{document} Structural hyperparameter configurations for baseline benchmarks. Evaluation metrics To evaluate the efficacy of the proposed AMRO framework against baseline metaheuristics (PSO, GWO, MFO, RHO) and DRL agents (PPO, GAT, DT-TrPPO-ELS), we define a multi-dimensional evaluation suite. 1. Total System Cost: A weighted objective function combining time and energy efficiency: 21\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathrm{Cos} t = \omega _{1} \cdot \mathop \sum \limits_{{i = 1}}^{N} Delay_{i} + \omega _{2} \cdot \mathop \sum \limits_{{i = 1}}^{N} Energy_{i}$$\end{document} Where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\omega\:}_{1}={\omega\:}_{2}=0.5$$\end{document}as per the empirical setup in the source code. 2. Task Success Rate (SR) : Determines the percentage of tasks meeting the strict latency threshold (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{T}_{th}=0.2s$$\end{document}) 22\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$SR\left( \% \right) = \left( {\frac{{\mathop \sum \nolimits_{{i = 1}}^{N} {\mathbb{I}}(d_{i} < 0.2)}}{N}} \right) \times 100$$\end{document} Where II(⋅)is the indicator function. 3. Carbon Footprint (CF): Estimates the environmental impact of offloading decisions:23\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$CF\left( {kgCO_{2} } \right) = E_{{total}} \times 0.475$$\end{document} Based on the average carbon intensity per kWh equivalent. 4. Resource Utilization (RU): Measures the efficiency of server allocation and load distribution: 24\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$RU\left( \% \right) = \frac{{{\mid \mathcal{S}}_{{active}} {\mid }}}{{{\mid \mathcal{S}}_{{total}} {\mid }}} \times 100$$\end{document} Where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mathcal{S}}_{active}$$\end{document}is the set of unique servers selected in the solution vector. 5. Jain’s Fairness Index (J): Ensures that the latency experienced by tasks is distributed fairly: 25\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\mathcal{J}} = \frac{{\left( {\mathop \sum \nolimits_{{i = 1}}^{N} d_{i} } \right)^{2} }}{{N \cdot \mathop \sum \nolimits_{{i = 1}}^{N} d_{i}^{2} }}$$\end{document} 6. Predictive Metrics (ML Phase): To assess the resilience module, standard classification metrics are employed:26\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$F1 - Score: \cdot 2\,\frac{{\Pr ecision \cdot \mathrm{Re} call}}{{\Pr ecision + \mathrm{Re} call}}$$\end{document}AUC-ROC: Area under the Receiver Operating Characteristic curve for failure prediction. 26\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$F1 - Score: \cdot 2\,\frac{{\Pr ecision \cdot \mathrm{Re} call}}{{\Pr ecision + \mathrm{Re} call}}$$\end{document} 26\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$F1 - Score: \cdot 2\,\frac{{\Pr ecision \cdot \mathrm{Re} call}}{{\Pr ecision + \mathrm{Re} call}}$$\end{document} AUC-ROC: Area under the Receiver Operating Characteristic curve for failure prediction. AUC-ROC: Area under the Receiver Operating Characteristic curve for failure prediction. To guarantee empirical validity and systematically eliminate the methodological biases frequently observed in contemporary simulation-based literature, the environmental configurations and structural hyperparameters governing the experimental phase are strictly formalized in Table 6. Unlike conventional frameworks evaluated under static, synthetic workloads, our evaluation matrix is explicitly designed to stress-test the architecture under real-world constraints. Consequently, the spatial topology is anchored to the EUA (Edge User Allocation) Dataset, deploying exactly M=50geographically distributed edge servers. Furthermore, to validate the framework’s deterministic resilience under hyper-burst traffic, the concurrent task payload ( N) is dynamically scaled from 100to 500independent operations. Table 6Algorithmic Configuration and Hyperparameters.CategoryParameterValueSimulationNumber of Tasks ( N)100–500 (Dynamic)Number of Servers ( M)50 (EUA Dataset)OptimizationPopulation Size/Epochs30/50Efficiency FactorsPSO/GWO/MFO0.85/0.82/0.79DRL (PPO/GAT)0.72/0.68ML ResilienceEnsemble VotingRF + XGBoost + LightGBM Table 6Algorithmic Configuration and Hyperparameters.CategoryParameterValueSimulationNumber of Tasks ( N)100–500 (Dynamic)Number of Servers ( M)50 (EUA Dataset)OptimizationPopulation Size/Epochs30/50Efficiency FactorsPSO/GWO/MFO0.85/0.82/0.79DRL (PPO/GAT)0.72/0.68ML ResilienceEnsemble VotingRF + XGBoost + LightGBM Algorithmic Configuration and Hyperparameters. Data-driven environmental modeling and workload formalization A critical limitation in contemporary edge orchestration literature is the reliance on purely synthetic, randomly generated environmental variables, which inherently fail to capture the stochastic and heterogeneous nature of real-world Industrial IoT (IIoT) deployments. To categorically eliminate this methodological bias and guarantee empirical fidelity, our experimental framework is engineered upon a tripartite data-driven foundation. As formalized in Table 3, we synthesize three distinct industry-standard datasets to construct a highly realistic evaluation matrix, covering spatial infrastructure, payload characteristics, and temporal arrival patterns. The integration of these specific datasets is not arbitrary; it establishes a strictly constrained, realistic proving ground designed to stress-test the QFLN-AMRO architecture: Spatial Heterogeneity via EUA Dataset: By anchoring the physical topology to the real-world geographic coordinates of 20edge servers from the EUA dataset, we enforce non-uniform server distribution. This categorically invalidates the “perfect grid” assumptions utilized in rudimentary studies. Consequently, the orchestrator is forced to navigate realistic propagation delays and localized congestion, proving its viability in actual urban IIoT deployments.Payload Authenticity via Edge-IIoT Dataset: Instead of executing arbitrary mathematical functions, the framework scales up to 500concurrent tasks drawn directly from the Edge-IIoT dataset. This ensures that the computational requirements and strict latency thresholds perfectly mirror modern Industry 5.0 environments, validating the framework’s energy-latency Pareto optimization under genuine heavy-duty payloads.Temporal Burstiness via Google Cluster Trace: To prevent the predictive models from merely memorizing steady-state traffic, the task arrival rates and resource demand fluctuations are modeled after the Google Cluster Trace. Injecting this real-world temporal burstiness and stochasticity rigorously tests the framework’s URLLC (Ultra-Reliable Low-Latency Communication) resilience and validates the stability of the system’s admission control under hyper-dynamic conditions. Spatial Heterogeneity via EUA Dataset: By anchoring the physical topology to the real-world geographic coordinates of 20edge servers from the EUA dataset, we enforce non-uniform server distribution. This categorically invalidates the “perfect grid” assumptions utilized in rudimentary studies. Consequently, the orchestrator is forced to navigate realistic propagation delays and localized congestion, proving its viability in actual urban IIoT deployments. Spatial Heterogeneity via EUA Dataset: By anchoring the physical topology to the real-world geographic coordinates of 20edge servers from the EUA dataset, we enforce non-uniform server distribution. This categorically invalidates the “perfect grid” assumptions utilized in rudimentary studies. Consequently, the orchestrator is forced to navigate realistic propagation delays and localized congestion, proving its viability in actual urban IIoT deployments. Payload Authenticity via Edge-IIoT Dataset: Instead of executing arbitrary mathematical functions, the framework scales up to 500concurrent tasks drawn directly from the Edge-IIoT dataset. This ensures that the computational requirements and strict latency thresholds perfectly mirror modern Industry 5.0 environments, validating the framework’s energy-latency Pareto optimization under genuine heavy-duty payloads. Payload Authenticity via Edge-IIoT Dataset: Instead of executing arbitrary mathematical functions, the framework scales up to 500concurrent tasks drawn directly from the Edge-IIoT dataset. This ensures that the computational requirements and strict latency thresholds perfectly mirror modern Industry 5.0 environments, validating the framework’s energy-latency Pareto optimization under genuine heavy-duty payloads. Temporal Burstiness via Google Cluster Trace: To prevent the predictive models from merely memorizing steady-state traffic, the task arrival rates and resource demand fluctuations are modeled after the Google Cluster Trace. Injecting this real-world temporal burstiness and stochasticity rigorously tests the framework’s URLLC (Ultra-Reliable Low-Latency Communication) resilience and validates the stability of the system’s admission control under hyper-dynamic conditions. Temporal Burstiness via Google Cluster Trace: To prevent the predictive models from merely memorizing steady-state traffic, the task arrival rates and resource demand fluctuations are modeled after the Google Cluster Trace. Injecting this real-world temporal burstiness and stochasticity rigorously tests the framework’s URLLC (Ultra-Reliable Low-Latency Communication) resilience and validates the stability of the system’s admission control under hyper-dynamic conditions. Performance evaluation and results In this section, we execute a rigorous, multi-stage empirical evaluation of the proposed architectures. To systematically validate the framework’s efficacy under realistic Edge-IIoT constraints, the analysis is fundamentally partitioned into three core phases: (i) Predictive resilience and Pareto front optimality analysis, (ii) Comprehensive comparative benchmarking of the baseline AMRO algorithm against state-of-the-art models, and (iii) Extreme stress-testing coupled with definitive statistical validation of the advanced QFLN-AMRO architecture. Phase I: predictive resilience and multi-objective strategy The integrity of the resource allocation process depends on the precision of the resilience model and the optimality of the decision-making front. Evaluation of the predictive resilience model The proposed resilience model, tasked with forecasting edge node availability and failure patterns, was evaluated using standard classification metrics. As shown in Table 7, the model achieves an Accuracy of 95.72%. The evaluation of the predictive resilience model establishes a highly reliable forecasting baseline, achieving an Accuracy of 0.9572and an exceptionally high Precision of 0.9699, alongside a Recall of 0.8384and an F1-Score of 0.8994(Table 7). The near-perfect precision dictates a drastically minimized false-positive rate in failure detection, definitively preventing the premature and unnecessary reallocation of computational capacity. Furthermore, this statistical equilibrium between sensitivity and specificity guarantees that the framework accurately forecasts edge node availability, empowering the network controller to execute task migrations strictly when node failure is imminent, thereby stabilizing overall resource management in dynamic edge environments. Table 7Evaluation Metrics of the predictive resilience model.MetricValueAccuracy 0.9572 Precision 0.9699 Recall 0.8384 F1-Score 0.8994 Table 7Evaluation Metrics of the predictive resilience model.MetricValueAccuracy 0.9572 Precision 0.9699 Recall 0.8384 F1-Score 0.8994 Evaluation Metrics of the predictive resilience model. The high Precision ( 0.9699) indicates a very low false-positive rate in failure detection, ensuring that system resources are not prematurely reallocated. The F1-Score of 0.8994 confirms that the model maintains a robust balance between sensitivity and specificity in dynamic edge environments. The predictive evaluation of the proposed ensemble model reveals a highly accurate classification capability, capturing \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathrm{15,540}$$\end{document} true positives and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathrm{62,161}$$\end{document} true negatives while strictly limiting false positives to a mere 482 instances (Fig. 3a), alongside an exceptional Area Under the Curve (AUC) of 0.9916 (Fig. 3b). The extreme suppression of false positives combined with such a near-perfect AUC indicates that the mechanism possesses highly deterministic discrimination between healthy and failing nodes. Consequently, this level of precision guarantees that the framework avoids triggering unnecessary, preemptive task migrations, thereby preserving critical computational bandwidth, preventing the premature reallocation of edge resources, and ensuring uninterrupted service continuity within the NOMA network. Fig. 3Predictive performance evaluation of the proposed ensemble model for task failure detection. (a) Confusion matrix illustrating the classification accuracy across successful (0) and failed (1) task instances, highlighting the model’s capability to minimize false positives and false negatives. (b) Receiver Operating Characteristic (ROC) curve demonstrating the robustness of the failure prediction mechanism, achieving an outstanding Area Under the Curve (AUC) of 0.9916. Predictive performance evaluation of the proposed ensemble model for task failure detection. (a) Confusion matrix illustrating the classification accuracy across successful (0) and failed (1) task instances, highlighting the model’s capability to minimize false positives and false negatives. (b) Receiver Operating Characteristic (ROC) curve demonstrating the robustness of the failure prediction mechanism, achieving an outstanding Area Under the Curve (AUC) of 0.9916. Pareto front analysis (Edge Strategies) Our framework treats delay and energy as conflicting objectives. Table 8 presents a sample of the Pareto Front, representing non-dominated strategies where any further reduction in energy consumption would necessitate an increase in processing latency. An extraction of the top edge strategies from the Pareto front demonstrates a strict inverse correlation between execution latency and energy expenditure, where the total system delay scales sequentially from \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathrm{442,166.80}$$\end{document}up to \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathrm{442,959.31}$$\end{document}seconds as the total energy consumption monotonically drops from 0.5337down to 0.5257Joules (Table 8). This quantified gradient confirms that the aggressive reduction of computational latency fundamentally necessitates higher energy injections into the edge nodes. Technically, this deterministic trade-off matrix empowers the network orchestrator to dynamically toggle between discrete offloading policies, seamlessly transitioning from an energy-conserving profile during battery-critical states to a high-speed, high-power execution mode tailored specifically for latency-sensitive applications. Table 8Top edge strategies (Pareto Front Sample).IndexTotal Delay (s)Total Energy (J)0 442166.804955 0.533718 1 442285.720492 0.529830 2 442338.009142 0.526248 3 442550.713327 0.526036 4 442959.317014 0.525719 Table 8Top edge strategies (Pareto Front Sample).IndexTotal Delay (s)Total Energy (J)0 442166.804955 0.533718 1 442285.720492 0.529830 2 442338.009142 0.526248 3 442550.713327 0.526036 4 442959.317014 0.525719 Top edge strategies (Pareto Front Sample). The optimization of task offloading under real-world data constraints yields a distinct set of non-dominated solutions, clearly defining a continuous inverse correlation between total system delay (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{f}_{1}$$\end{document}) and energy consumption (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{f}_{2}$$\end{document}) (Fig. 4). The emergence of this well-distributed Pareto frontier confirms that minimizing computational delay fundamentally necessitates higher energy expenditure, and vice versa. From a network management perspective, this explicitly quantified trade-off empowers the centralized controller to dynamically select an optimal operational point—shifting towards energy-conservative offloading when edge battery levels are critical, or prioritizing ultra-low latency configurations for time-sensitive IIoT applications within the NOMA architecture. Fig. 4Pareto optimal solutions for task offloading based on real-world data, illustrating the trade-off between total system delay (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{f}_{1}$$\end{document}) and energy consumption (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{f}_{2}$$\end{document}). Pareto optimal solutions for task offloading based on real-world data, illustrating the trade-off between total system delay (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{f}_{1}$$\end{document}) and energy consumption (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{f}_{2}$$\end{document}). Phase II: Comparative benchmarking (AMRO vs. Baselines) In this phase, the AMRO architecture is benchmarked against seven established metaheuristic and Deep Reinforcement Learning (DRL) algorithms under a simulated workload of 500 computational tasks. The complete performance metrics for this comparative analysis are presented in Table 9. The results highlight a fundamental trade-off between two dominant classes of offloading strategies. Conventional metaheuristic algorithms, including PSO, GWO, MFO, and RHO, demonstrate rapid decision-making with execution times averaging approximately 0.0007 s. However, their performance deteriorates under congested network conditions, leading to low task success rates that range from 33.50% to 39.04%. Conversely, DRL-based models such as PPO, GAT, and DT-TrPPO-ELS achieve 100% task completion and resource utilization, ensuring high reliability. This reliability, however, is coupled with significant computational overhead. For instance, the DT-TrPPO-ELS and PPO models record execution times of 14.4357s and 12.5023s, respectively, and exhibit higher energy consumption (> 1.28 J) and carbon footprints (> 0.61 kg), which may limit their suitability for energy-constrained and ultra-low-latency edge applications. The proposed AMRO framework effectively bridges this performance gap. It achieves a task success rate of 98.50% and a resource utilization of 95.00%, comparable to DRL models, but with a significantly lower execution time of 2.6784s. This represents an up to 5.3-fold reduction in computational latency compared to the DRL benchmarks. Furthermore, AMRO demonstrates greater energy efficiency, with a total consumption of 0.9202 J and a carbon footprint of 0.4371 kg. These results indicate that the framework successfully integrates the reliability of DRL-based approaches with the computational efficiency required for real-time, dynamic edge computing environments. Table 9Comprehensive performance metrics (Tasks = 500).AlgorithmTotal CostTotal Delay (s)Total Energy (J)Success Rate (%)Carbon Footprint (kg)Resource Util (%)Avg Delay (s)Max Delay (s)Execution Time (s)PSO 0.5661 0.3900 0.7421 33.6388 0.3525 33.6388 0.0008 0.0020 0.0007 GWO 0.6565 0.4539 0.8591 39.0467 0.4081 39.0467 0.0009 0.0023 0.0007 MFO 0.5618 0.3908 0.7328 33.5020 0.3481 33.5020 0.0008 0.0020 0.0006 RHO 0.6020 0.4134 0.7906 35.7300 0.3755 35.7300 0.0008 0.0019 0.0007 PPO 1.2130 0.8363 1.5897 100.00 0.7551 100.00 0.0017 0.0042 12.5023 GAT 1.1495 0.7873 1.5116 100.00 0.7180 100.00 0.0016 0.0040 5.4470 DT-TrPPO-ELS 0.9973 0.7053 1.2894 100.00 0.6124 100.00 0.0014 0.0035 14.4357 Proposed (AMRO) 0.6936 0.4669 0.9202 98.5000 0.4371 95.0000 0.0012 0.0031 2.6784 Table 9Comprehensive performance metrics (Tasks = 500).AlgorithmTotal CostTotal Delay (s)Total Energy (J)Success Rate (%)Carbon Footprint (kg)Resource Util (%)Avg Delay (s)Max Delay (s)Execution Time (s)PSO 0.5661 0.3900 0.7421 33.6388 0.3525 33.6388 0.0008 0.0020 0.0007 GWO 0.6565 0.4539 0.8591 39.0467 0.4081 39.0467 0.0009 0.0023 0.0007 MFO 0.5618 0.3908 0.7328 33.5020 0.3481 33.5020 0.0008 0.0020 0.0006 RHO 0.6020 0.4134 0.7906 35.7300 0.3755 35.7300 0.0008 0.0019 0.0007 PPO 1.2130 0.8363 1.5897 100.00 0.7551 100.00 0.0017 0.0042 12.5023 GAT 1.1495 0.7873 1.5116 100.00 0.7180 100.00 0.0016 0.0040 5.4470 DT-TrPPO-ELS 0.9973 0.7053 1.2894 100.00 0.6124 100.00 0.0014 0.0035 14.4357 Proposed (AMRO) 0.6936 0.4669 0.9202 98.5000 0.4371 95.0000 0.0012 0.0031 2.6784 Comprehensive performance metrics (Tasks = 500). An analysis of the data presented in Table 9 highlights a fundamental trade-off in existing offloading strategies. While traditional metaheuristics such as PSO and GWO offer rapid execution, they exhibit low task success rates, ranging from 33% to 39%. Conversely, sophisticated Deep Reinforcement Learning (DRL) models achieve near-perfect success rates but at the cost of significant computational overhead, with execution times reaching up to 14.43 s. The proposed AMRO framework effectively resolves this dichotomy by achieving a high success rate of 98.5% with a comparatively low execution time of 2.67 s, demonstrating its suitability for efficient, real-time task offloading. This performance is further detailed in the multi-dimensional evaluation shown in Fig. 5. The framework demonstrates superior scalability, as the growth in algorithmic execution time remains minimal with increasing workloads, in contrast to the steep rise observed in DRL architectures (Fig. 5a and b). Under high-stress scenarios (1000 tasks), AMRO maintains operational robustness by effectively minimizing both energy dissipation (Fig. 5c) and aggregate delay (Fig. 5d) without compromising reliability, as evidenced by a consistently high task completion rate (Fig. 5e). Furthermore, the framework ensures systemic stability through rapid algorithmic convergence (Fig. 5g) and equitable resource utilization across servers (Fig. 5h), which translates to a reduced variance in task-level delays and thus minimizes network jitter (Fig. 5i). Finally, the projection of the framework’s offloading decisions into the multi-objective Pareto space (Fig. 5j) illustrates its capacity to identify optimal solutions within the conflicting objectives of energy conservation and execution latency. The generated frontier of non-dominated solutions confirms that AMRO consistently derives resource allocation configurations that are optimally adapted to the dynamic conditions of the edge network. Fig. 5Comprehensive performance evaluation and comparative analysis of the proposed AMRO algorithm against baseline models. Comprehensive performance evaluation and comparative analysis of the proposed AMRO algorithm against baseline models. Phase III: Stress testing under resource constraints (QFLN-AMRO) In this phase, the QFLN-AMRO architecture was evaluated under a high-density stress test scenario, comprising 800 tasks allocated across 30 edge servers. The comparative performance against several state-of-the-art metaheuristics is detailed in Table 10. The results indicate that the proposed framework effectively maintains network stability, achieving the lowest Lyapunov Penalty (260.48) and a significantly reduced Queue Drift (159.20). This contrasts with conventional algorithms such as SCA and AOA, which exhibited substantially higher queue drifts of 493.51 and 461.67, respectively, indicating a greater risk of buffer bloat under heavy load. This stable queue management enables QFLN-AMRO to deliver a low total delay of 0.51s and reduce total energy expenditure to 0.75 J, approximately half the energy consumed by baseline models like HHO and WOA. Furthermore, the framework optimizes bandwidth utilization, achieving a spectral efficiency of 926.93 bps/Hz, while ensuring equitable resource distribution, as evidenced by a Jain’s Fairness Index of 0.99. A key architectural feature of QFLN-AMRO is its strategic handling of the Task Drop Rate. While baseline metaheuristics attempt to process the entire overwhelming workload—leading to elevated execution times (e.g., 11.83s for AOA)—QFLN-AMRO exhibits a Task Drop Rate of 54.93%. This is not a systemic failure but rather a deliberate, self-preserving mechanism guided by the Lyapunov optimization constraints. By selectively terminating non-critical or highly resource-intensive tasks during extreme congestion, the framework prevents cascading node failures, conserves energy, and achieves rapid algorithmic convergence (4.46s). This demonstrates that the QFLN-AMRO architecture dynamically prioritizes the overall survivability and responsiveness of the edge network over exhaustive processing when operating under peak saturation. Table 10Resource constraint analysis (Edge Servers = 30, Tasks = 800).AlgorithmLyapunov PenaltyTotal Delay (s)Total Energy (J)Queue DriftSpectral Eff (bps/Hz)Jains FairnessTask Drop Rate (%)Execution Time (s)AOA 537.2031 1.0497 1.5106 461.6744 862.2054 0.99 18.2503 11.8324 SCA 576.4765 1.1215 1.6592 493.5141 840.0895 0.99 19.5089 7.0066 WOA 521.3674 0.9949 1.5123 445.7546 862.0214 0.99 17.6209 5.7464 HHO 484.1141 0.9300 1.4040 413.9150 875.0115 0.99 16.3623 10.6236 SMA 445.8452 0.8523 1.2754 382.0754 890.2199 0.99 15.1037 7.2961 INFO 409.7249 0.7915 1.1898 350.2358 867.0706 0.99 13.8450 9.2112 COA 390.8805 0.7631 1.1313 334.3160 885.1102 0.99 13.2157 7.6569 QFLN-AMRO 260.4818 0.5069 0.7521 159.1981 926.9272 0.99 54.9319 4.4601 Table 10Resource constraint analysis (Edge Servers = 30, Tasks = 800).AlgorithmLyapunov PenaltyTotal Delay (s)Total Energy (J)Queue DriftSpectral Eff (bps/Hz)Jains FairnessTask Drop Rate (%)Execution Time (s)AOA 537.2031 1.0497 1.5106 461.6744 862.2054 0.99 18.2503 11.8324 SCA 576.4765 1.1215 1.6592 493.5141 840.0895 0.99 19.5089 7.0066 WOA 521.3674 0.9949 1.5123 445.7546 862.0214 0.99 17.6209 5.7464 HHO 484.1141 0.9300 1.4040 413.9150 875.0115 0.99 16.3623 10.6236 SMA 445.8452 0.8523 1.2754 382.0754 890.2199 0.99 15.1037 7.2961 INFO 409.7249 0.7915 1.1898 350.2358 867.0706 0.99 13.8450 9.2112 COA 390.8805 0.7631 1.1313 334.3160 885.1102 0.99 13.2157 7.6569 QFLN-AMRO 260.4818 0.5069 0.7521 159.1981 926.9272 0.99 54.9319 4.4601 Resource constraint analysis (Edge Servers = 30, Tasks = 800). The Lyapunov-guided stability mechanism is central to the framework’s ability to manage Quality of Service (QoS) trade-offs under network stress, as detailed in Fig. 6. The proposed QFLN-AMRO framework achieves the lowest Lyapunov Penalty (260.48) and Queue Drift (159.19), indicating a superior capability for maintaining long-term system stability. This stability is achieved through a controlled trade-off: the framework strategically increases its Task Drop Rate (54.93%) to preserve essential QoS metrics such as ultra-low latency (0.50s) and high spectral efficiency (926.92 bps/Hz), particularly under severe congestion. This behavior aligns with the drift-plus-penalty optimization objective (Fig. 6.b), where the system converges to a minimized penalty state by regulating queue backlogs (Fig. 6.a). As the infrastructure scales, the framework’s adaptive control of the task rejection rate (Fig. 6.g) functions as a self-preserving load-shedding mechanism, preventing network collapse while maintaining a high degree of fairness in resource distribution, as quantified by Jain’s Fairness Index (Fig. 6.h). At the communication layer, the integration of Non-Orthogonal Multiple Access (NOMA) enhances data transmission efficiency and resilience. Under a moderately constrained 40-server configuration, the framework maximizes spectral efficiency compared to orthogonal baselines by effectively multiplexing user signals within the power domain (Fig. 6.c). The effectiveness of the framework’s multi-user interference mitigation is demonstrated by the tightly compressed interference distribution shown in the violin plot (Fig. 6.d). This profile indicates successful decoding of superimposed signals and effective isolation of crosstalk, ensuring high-fidelity offloading transmissions. Computationally, the framework demonstrates both rapid convergence and predictable performance. The integration of fractional-order calculus contributes to an accelerated cost-objective convergence path, as observed in the 30-server scenario (Fig. 6.e). This algorithmic enhancement is complemented by consistent execution times. The statistical distribution of the quantum-behaved execution time, visualized in the box plot (Fig. 6.f), exhibits low variance. This suggests that the computational overhead of QFLN-AMRO remains deterministic and resilient to stochastic workload fluctuations. The architectural robustness of QFLN-AMRO is most evident under severe resource starvation, simulated with only 10 active edge servers. In this challenging scenario, QFLN-AMRO outperforms conventional algorithms by effectively capping total energy expenditure (Fig. 6.i) and minimizing aggregate system latency (Fig. 6.j). This extreme-case analysis confirms the architecture’s ability to prevent catastrophic node saturation and queuing failures, ensuring a highly reliable edge orchestration policy even during significant infrastructure degradation. Fig. 6Comprehensive stress testing, statistical robustness, and NOMA network evaluation of the proposed QFLN-AMRO framework under varying resource constraints compared to state-of-the-art algorithms. Comprehensive stress testing, statistical robustness, and NOMA network evaluation of the proposed QFLN-AMRO framework under varying resource constraints compared to state-of-the-art algorithms. Phase IV: Statistical robustness and significance analysis To empirically validate the reliability and statistical significance of the QFLN-AMRO framework, a comprehensive analysis was conducted based on 30 independent execution runs. The results were first evaluated using a one-way Analysis of Variance (ANOVA), which revealed a statistically significant difference among the mean performances of the compared algorithms, yielding an F-value of 12.9659 and a p-value of 4.6617e-14. These findings were further corroborated by non-parametric Wilcoxon tests, confirming the superiority of QFLN-AMRO over key competitors like COA (p-value = 3.2391e-06). A detailed examination of performance consistency, with metrics articulated in Table 11, highlights the framework’s superior stability. QFLN-AMRO achieved a Standard Deviation (STD) of only 6.4584, which is substantially lower than that of other metaheuristics. For example, algorithms such as AOA and WOA exhibited high performance variance, with STDs of 44.2765 and 40.6662, respectively. This low variance in QFLN-AMRO translates to a leading stability index (defined as STD/Mean) of 0.0105, suggesting that its internal mechanics—integrating quantum-behaved and fractional-order concepts—effectively mitigate the stochastic behavior inherent in traditional optimization methods and promote a more deterministic convergence path. The robustness of the framework is further demonstrated through an analysis of its performance boundaries (best versus worst-case outcomes). While some baseline algorithms, such as AOA and INFO, occasionally found marginally better best-case solutions (563.5622 and 578.1780), they were prone to poor performance in their worst-case scenarios (754.0404 and 720.9711). In contrast, QFLN-AMRO maintained a tightly constrained performance range, with its objective cost varying only from 605.0937 to 629.7398. Crucially, the worst-case performance of QFLN-AMRO is better than the average performance of every competing algorithm, including COA (644.0639) and SCA (687.4242). This combination of a competitive mean objective cost (616.9085) and a narrow performance spread confirms its reliability, ensuring consistent and predictable Quality of Service (QoS) for mission-critical edge network orchestration. Table 11Statistical results over 30 independent runs (Objective Cost).AlgorithmMeanSTDBest (Min)Worst (Max)Stability (STD/Mean)QFLN-AMRO 616.9085 6.4584 605.0937 629.7398 0.0105 COA 644.0639 28.6481 605.8564 707.8558 0.0445 INFO 656.1623 35.9437 578.1780 720.9711 0.0548 HHO 657.4490 31.6436 601.8372 715.9205 0.0481 SMA 663.4986 25.5840 604.6122 708.3440 0.0386 WOA 669.3104 40.6662 611.5699 766.4969 0.0608 AOA 677.3353 44.2765 563.5622 754.0404 0.0654 SCA 687.4242 35.1862 612.4259 755.8509 0.0512 Table 11Statistical results over 30 independent runs (Objective Cost).AlgorithmMeanSTDBest (Min)Worst (Max)Stability (STD/Mean)QFLN-AMRO 616.9085 6.4584 605.0937 629.7398 0.0105 COA 644.0639 28.6481 605.8564 707.8558 0.0445 INFO 656.1623 35.9437 578.1780 720.9711 0.0548 HHO 657.4490 31.6436 601.8372 715.9205 0.0481 SMA 663.4986 25.5840 604.6122 708.3440 0.0386 WOA 669.3104 40.6662 611.5699 766.4969 0.0608 AOA 677.3353 44.2765 563.5622 754.0404 0.0654 SCA 687.4242 35.1862 612.4259 755.8509 0.0512 Statistical results over 30 independent runs (Objective Cost). A comprehensive analysis was performed to evaluate the statistical robustness, convergence reliability, and operational stability of the QFLN-AMRO framework over 30 independent executions, with the detailed results presented in Fig. 7. This analysis confirms that the framework’s structural design provides significant immunity to stochastic variance, a critical attribute for deterministic performance in dynamic environments. Convergence and cost distribution analysis The consistency of the framework’s convergence is demonstrated through its objective cost distribution. The boxplot analysis (Fig. 7a) reveals a tightly compressed interquartile range and the absence of significant outliers, indicating consistent convergence and mitigating the stochastic deviations common in other metaheuristics. This finding is further substantiated by the Kernel Density Estimation (Fig. 7e), which shows a sharp, unimodal peak at the lower end of the cost spectrum. Concurrently, the Empirical Cumulative Distribution Function (Fig. 7f) exhibits a steep ascent, confirming a high probability of achieving near-optimal solutions in any given run. Computational overhead and resource management From a computational perspective, the violin plot of execution times (Fig. 7b) shows a narrow and dense distribution, confirming that the framework’s overhead is both predictable and stable. This determinism extends to resource management, where QFLN-AMRO consistently minimizes both the mean and standard deviation for total energy consumption (Fig. 7c) and overall system delay (Fig. 7d). The compact error bars (Mean ± STD) for these metrics affirm the framework’s capacity to uphold strict Quality of Service (QoS) guarantees under variable network loads. Resilience and algorithmic stability The framework’s resilience is particularly evident in the worst-case versus best-case performance analysis (Fig. 7g). The tight clustering of results near the identity line (y = x) indicates that even the worst-case outcomes remain close to the optimal, effectively bounding the performance degradation under unfavorable conditions. This stability is formally quantified by the Coefficient of Variation (STD/Mean) in Fig. 7h. The markedly low value for QFLN-AMRO underscores its robustness against variations stemming from random initializations, a key differentiator from conventional algorithms. Holistic multi-objective performance Collectively, these attributes are synthesized in the multi-metric radar chart (Fig. 7i). The chart visually demonstrates the framework’s holistic superiority, with its performance polygon significantly contracted towards the origin across all key metrics—Mean Cost, Execution Time, Delay, Energy, and STD—relative to baseline models. This multi-dimensional view confirms that QFLN-AMRO achieves a balanced, Pareto-optimal solution, delivering high operational efficiency and statistical reliability suitable for dynamic edge computing environments. Fig. 7Statistical robustness, distribution analysis, and stability validation of the proposed QFLN-AMRO framework over multiple independent executions. Statistical robustness, distribution analysis, and stability validation of the proposed QFLN-AMRO framework over multiple independent executions. A granular statistical analysis was conducted to validate the performance superiority of the QFLN-AMRO framework against other metaheuristic baselines, the results of which are detailed in Table 12. Following the initial omnibus ANOVA, a Tukey Honestly Significant Difference (HSD) post-hoc test was employed for rigorous pairwise comparisons. This analysis confirms that the enhanced performance of QFLN-AMRO is statistically significant in every comparison, leading to a conclusive rejection of the null hypothesis. The adjusted p-values (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{p}_{adj}$$\end{document}) remained decisively below the critical significance level of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\alpha\:=0.05$$\end{document}. Specifically, against algorithms such as AOA, SCA, SMA, and WOA, the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{p}_{adj}$$\end{document} value was effectively zero (0.0000), while even against highly competitive baselines like COA (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{p}_{adj}=0.0335$$\end{document}), HHO (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{p}_{adj}=0.0001$$\end{document}), and INFO (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{p}_{adj}=0.0002$$\end{document}), the statistical distinction remains unambiguous. This outcome provides strong evidence that the framework’s superiority is not a result of stochastic chance but stems from its inherent architectural advantages. Beyond establishing statistical significance, the Tukey HSD test also quantifies the magnitude of the performance gap through the “Mean Difference” metric. The results demonstrate a substantial margin of outperformance, with the mean difference ranging from 27.1554 against a competitive algorithm like COA to more extreme differentials of 70.5157 against SCA and 60.4268 against AOA. This considerable performance delta is a direct consequence of the framework’s advanced optimization engine. The synergy between quantum-behaved mechanics and fractional-order calculus enables a more efficient exploration of the search space, allowing QFLN-AMRO to systematically bypass the local optima that typically constrain conventional metaheuristics. In summary, this post-hoc statistical evaluation provides robust, quantitative validation of the framework’s reliability and effectiveness, confirming its suitability as a deterministic and superior orchestration policy for demanding Edge-IIoT environments. Table 12Tukey HSD Post-Hoc Test (QFLN-AMRO vs. Baselines).Group 1Group 2Mean Diffp-adjRejectAOAQFLN-AMRO -60.4268 0.0000 TrueCOAQFLN-AMRO -27.1554 0.0335 TrueHHOQFLN-AMRO -40.5405 0.0001 TrueINFOQFLN-AMRO -39.2538 0.0002 TrueQFLN-AMRO (Proposed)SCA 70.5157 0.0000 TrueQFLN-AMRO (Proposed)SMA 46.5901 0.0000 TrueQFLN-AMRO (Proposed)WOA 52.4019 0.0000 True Table 12Tukey HSD Post-Hoc Test (QFLN-AMRO vs. Baselines).Group 1Group 2Mean Diffp-adjRejectAOAQFLN-AMRO -60.4268 0.0000 TrueCOAQFLN-AMRO -27.1554 0.0335 TrueHHOQFLN-AMRO -40.5405 0.0001 TrueINFOQFLN-AMRO -39.2538 0.0002 TrueQFLN-AMRO (Proposed)SCA 70.5157 0.0000 TrueQFLN-AMRO (Proposed)SMA 46.5901 0.0000 TrueQFLN-AMRO (Proposed)WOA 52.4019 0.0000 True Tukey HSD Post-Hoc Test (QFLN-AMRO vs. Baselines). Discussion and interpretation The comprehensive performance evaluation presented in Sect. 4 provides strong empirical evidence for the superiority of the proposed QFLN-AMRO framework. This section moves beyond the empirical results to offer a theoretical interpretation of these findings. We analyze the underlying mechanisms driving the framework’s performance, dissect the dynamics of multi-objective trade-offs, and examine the architectural foundations of its resilience in dense Edge-IIoT environments. Algorithmic synergy and convergence dynamics The convergence analysis highlights a fundamental difference between conventional metaheuristics (e.g., GA, PSO) and the QFLN-AMRO architecture. Standard algorithms are susceptible to premature convergence and stochastic stagnation, particularly when navigating the high-dimensional, combinatorial search space typical of large-scale task offloading scenarios (> 500 tasks). While the baseline AMRO employs NSGA-II to preserve population diversity, it lacks a mechanism for adaptive control. The principal innovation of QFLN-AMRO lies in the synergy between its Fractional-Order Calculus core and its Quantum Fuzzy Logic Network (QFLN) controller. The fractional-order component introduces a long-range memory property, enabling the optimization process to utilize historical gradient information and effectively avoid local optima. Simultaneously, the QFLN functions as an intelligent cognitive controller. It dynamically maps Pareto front topological states, such as hypervolume growth and solution spread, to fuzzy membership functions. These functions, in turn, leverage quantum probability amplitudes to precisely adjust mutation and crossover rates. This synergy facilitates the rapid and steep descent in objective cost observed within the initial 20 generations, ensuring the orchestrator achieves near-optimal solutions with the fast convergence required for real-time 5G/6G Ultra-Reliable Low-Latency Communication (URLLC) services. Resolution of multi-objective trade-offs A critical finding of this research is the effective resolution of the inherent conflict between latency (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{f}_{1}$$\end{document}) and energy consumption (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{f}_{2}$$\end{document}). Simple heuristic schedulers like Greedy or Round-Robin may achieve low latency for specific task subsets, but our results show this comes at the cost of significant energy inefficiencies and thermal imbalances (“hotspots”) across the edge infrastructure. In contrast, QFLN-AMRO synthesizes a mathematically robust, non-dominated Pareto front, characterized by a superior Hypervolume (HV) and optimal solution distribution (Spread). This provides network operators with a diverse set of high-quality “knee-point” solutions. The framework integrates spatial awareness, using Haversine distance for propagation modeling, and spectral efficiency awareness through NOMA-inspired protocols. This allows for intelligent load balancing, leading to assignment matrices that reduce total energy consumption by up to 25% compared to GA, while limiting the associated latency increase to a marginal 5%. This demonstrates the framework’s ability to achieve true multi-objective harmony rather than collapsing into a single-objective solution. Infrastructural resilience via lyapunov-guided admission control The stress-testing scenarios, which involved scaling the network to 1000 IIoT tasks with only 10 active servers, revealed critical scalability limitations in baseline algorithms. Under such high-density conditions, conventional schedulers suffer from uncontrolled queue growth and cascading latency failures. QFLN-AMRO circumvents these bottlenecks through its integrated Lyapunov-guided optimization strategy. The framework incorporates a proactive admission control mechanism that continuously monitors the system’s drift-plus-penalty trajectory (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\Delta\:}Q\left(t\right)$$\end{document}). Instead of allowing the network to reach saturation, QFLN-AMRO strategically rejects tasks that would violate hard delay constraints under heavy congestion. This load-shedding is not a system failure but a strategic self-preservation mechanism. It guarantees that the core communication backbone remains uncongested, thereby maintaining high spectral efficiency (> 926 bps/Hz) and ensuring equitable resource distribution (confirmed by Jain’s Fairness Index) for all admitted mission-critical payloads. Statistical Validation and Robustness To ensure that the observed performance gains were not an artifact of stochastic variation, a rigorous statistical validation was performed. The omnibus ANOVA yielded a highly significant result (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:F(d{f}_{between},d{f}_{within})=12.97,p<.001$$\end{document}), confirming a structural difference between the tested algorithms. This result rejects the null hypothesis that the observed means are equal. Furthermore, a post-hoc Tukey Honestly Significant Difference (HSD) test was conducted to perform pairwise comparisons. The results placed QFLN-AMRO in a distinct performance category, with adjusted p-values approaching zero ( p<.0001) when compared against algorithms such as AOA, SCA, and WOA. This statistical outcome underscores the deterministic reliability introduced by the quantum-fuzzy control mechanism. This reliability is further evidenced by the framework’s exceptionally low Coefficient of Variation (0.0105) across 30 independent runs. For Industry 4.0 systems where operational predictability is paramount, QFLN-AMRO provides a mathematically validated and robust performance envelope. Holistic architectural advantages In summary, the theoretical and empirical findings confirm the multi-dimensional architectural evolution embodied by QFLN-AMRO. As systematically outlined in Table 13, the framework represents a significant advancement over both rigid heuristics and unstable metaheuristics. While conventional methods are limited by static logic or blind stochastic search, QFLN-AMRO’s cognitive layer elevates the orchestration paradigm. Despite its enhanced intelligence, the framework maintains a scalable computational complexity of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$O\left( {G \cdot N\log N} \right) + \in$$\end{document}, where ∈represents the negligible inference overhead of the fuzzy logic engine. This holistic design effectively mitigates node saturation and stochastic instability, delivering a highly resilient, intelligent, and statistically superior orchestration policy tailored for the severe operational constraints of dynamic Edge-IIoT environments. Table 13Comprehensive comparison of orchestration architectures.Performance MetricRound-Robin/GreedyGenetic Algorithm (GA)Particle Swarm (PSO)AMRO (Base)QFLN-AMRO (Proposed)Optimization GoalSingle-ObjectiveMulti-ObjectiveMulti-ObjectiveMulti-ObjectiveDynamic Multi-ObjectiveConvergence SpeedImmediate (Heuristic)ModerateFast (Local)Fast (Global)Ultra-Fast (Quantum)Pareto Front QualityPoor (N/A)AverageLow DiversityHigh HVSuperior HV & SpreadPeak Load ResilienceVery LowModerateModerateHighVery High (Lyapunov-Guided)Energy EfficiencyLowModerateHighVery HighMaximum OptimizationResource BalancingStatic/RandomStochasticStochasticAdaptiveIntelligent/ProactiveStatistical StabilityHigh (Deterministic)Low (Stochastic)ModerateHighVery High (Validated)Decision IntelligenceNoneFixed RulesFixed VelocityAdaptive WeightsQuantum-Fuzzy LogicComplexity \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:O\left(N\right)$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:O(G\cdot\:{N}^{2})$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:O(G\cdot\:N)$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:O(G\cdot\:N\mathrm{l}\mathrm{o}\mathrm{g}N)$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$O\left( {G \cdot N\log N} \right) + \in$$\end{document} Note: Grepresents the number of generations and Nrepresents the number of tasks. ∈represents the negligible computational overhead of the QFLN inference engine. Table 13Comprehensive comparison of orchestration architectures.Performance MetricRound-Robin/GreedyGenetic Algorithm (GA)Particle Swarm (PSO)AMRO (Base)QFLN-AMRO (Proposed)Optimization GoalSingle-ObjectiveMulti-ObjectiveMulti-ObjectiveMulti-ObjectiveDynamic Multi-ObjectiveConvergence SpeedImmediate (Heuristic)ModerateFast (Local)Fast (Global)Ultra-Fast (Quantum)Pareto Front QualityPoor (N/A)AverageLow DiversityHigh HVSuperior HV & SpreadPeak Load ResilienceVery LowModerateModerateHighVery High (Lyapunov-Guided)Energy EfficiencyLowModerateHighVery HighMaximum OptimizationResource BalancingStatic/RandomStochasticStochasticAdaptiveIntelligent/ProactiveStatistical StabilityHigh (Deterministic)Low (Stochastic)ModerateHighVery High (Validated)Decision IntelligenceNoneFixed RulesFixed VelocityAdaptive WeightsQuantum-Fuzzy LogicComplexity \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:O\left(N\right)$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:O(G\cdot\:{N}^{2})$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:O(G\cdot\:N)$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:O(G\cdot\:N\mathrm{l}\mathrm{o}\mathrm{g}N)$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$O\left( {G \cdot N\log N} \right) + \in$$\end{document} Note: Grepresents the number of generations and Nrepresents the number of tasks. ∈represents the negligible computational overhead of the QFLN inference engine. Comprehensive comparison of orchestration architectures. Note: Grepresents the number of generations and Nrepresents the number of tasks. ∈represents the negligible computational overhead of the QFLN inference engine. Quantitative validation and empirical dominance While Table 13 establishes the theoretical and architectural distinctions between the evaluated frameworks, Table 14 provides the rigorous, empirical substantiation of these claims under a dense, hyper-dynamic payload scenario ( N=1000 concurrent tasks). The quantitative evaluation strictly focuses on validating the latency-energy Pareto optimality and the algorithmic resilience against peak load destabilization. Observation 1 (Energy-Latency Synergy): As delineated in Table 14, traditional heuristics fundamentally fail to meet the strict URLLC threshold (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:<5\mathrm{\:ms}$$\end{document}). While PSO and GA achieve average latencies of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:9.80\mathrm{\:ms}$$\end{document}and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:12.10\mathrm{\:ms}$$\end{document}respectively, they simultaneously suffer from high energy expenditure due to sub-optimal resource trapping. In stark contrast, QFLN-AMRO secures a mean latency of strictly \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:4.15\mathrm{\:ms}$$\end{document}while radically suppressing energy consumption to \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:55.65\mathrm{\:J}$$\end{document}.Technical Implication: This explicitly confirms the efficacy of the Quantum-Fuzzy logic integration. The \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\sim\:25\mathrm{\%}$$\end{document}reduction in energy (compared to the ablated AMRO base) is not a product of stochastic variance, but a direct consequence of the QFLN layer dynamically steering the evolutionary operators away from local optima, resulting in a significantly superior Hypervolume (HV = 0.958).Observation 2 (Systematic Resilience via Stability Ratio): The most critical divergence is observed in the peak load resilience metrics. Under the extreme \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:1000\mathrm{-task}$$\end{document}stress test, the Round-Robin baseline completely destabilizes, and swarm intelligence models (PSO, GA) drop nearly \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:40\mathrm{\%}\mathrm{-}50\mathrm{\%}$$\end{document}of the critical payload. Conversely, QFLN-AMRO maintains a near-perfect Task Admission Rate of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:98.7\mathrm{\%}$$\end{document}alongside an unprecedented Lyapunov Stability Ratio of 0.0105.Technical Implication: This numerical evidence fundamentally validates the theoretical claim of “Very High (Lyapunov-Guided)” resilience presented in Table 13. The ultra-low stability ratio quantitatively proves that the proposed orchestrator possesses the deterministic capacity to absorb stochastic temporal burstiness (inspired by the Google Cluster Trace) without cascading queue overflow or violating URLLC latency bounds See Table 15. Observation 1 (Energy-Latency Synergy): As delineated in Table 14, traditional heuristics fundamentally fail to meet the strict URLLC threshold (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:<5\mathrm{\:ms}$$\end{document}). While PSO and GA achieve average latencies of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:9.80\mathrm{\:ms}$$\end{document}and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:12.10\mathrm{\:ms}$$\end{document}respectively, they simultaneously suffer from high energy expenditure due to sub-optimal resource trapping. In stark contrast, QFLN-AMRO secures a mean latency of strictly \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:4.15\mathrm{\:ms}$$\end{document}while radically suppressing energy consumption to \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:55.65\mathrm{\:J}$$\end{document}. Observation 1 (Energy-Latency Synergy): As delineated in Table 14, traditional heuristics fundamentally fail to meet the strict URLLC threshold (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:<5\mathrm{\:ms}$$\end{document}). While PSO and GA achieve average latencies of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:9.80\mathrm{\:ms}$$\end{document}and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:12.10\mathrm{\:ms}$$\end{document}respectively, they simultaneously suffer from high energy expenditure due to sub-optimal resource trapping. In stark contrast, QFLN-AMRO secures a mean latency of strictly \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:4.15\mathrm{\:ms}$$\end{document}while radically suppressing energy consumption to \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:55.65\mathrm{\:J}$$\end{document}. Technical Implication: This explicitly confirms the efficacy of the Quantum-Fuzzy logic integration. The \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\sim\:25\mathrm{\%}$$\end{document}reduction in energy (compared to the ablated AMRO base) is not a product of stochastic variance, but a direct consequence of the QFLN layer dynamically steering the evolutionary operators away from local optima, resulting in a significantly superior Hypervolume (HV = 0.958). Technical Implication: This explicitly confirms the efficacy of the Quantum-Fuzzy logic integration. The \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\sim\:25\mathrm{\%}$$\end{document}reduction in energy (compared to the ablated AMRO base) is not a product of stochastic variance, but a direct consequence of the QFLN layer dynamically steering the evolutionary operators away from local optima, resulting in a significantly superior Hypervolume (HV = 0.958). Observation 2 (Systematic Resilience via Stability Ratio): The most critical divergence is observed in the peak load resilience metrics. Under the extreme \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:1000\mathrm{-task}$$\end{document}stress test, the Round-Robin baseline completely destabilizes, and swarm intelligence models (PSO, GA) drop nearly \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:40\mathrm{\%}\mathrm{-}50\mathrm{\%}$$\end{document}of the critical payload. Conversely, QFLN-AMRO maintains a near-perfect Task Admission Rate of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:98.7\mathrm{\%}$$\end{document}alongside an unprecedented Lyapunov Stability Ratio of 0.0105. Observation 2 (Systematic Resilience via Stability Ratio): The most critical divergence is observed in the peak load resilience metrics. Under the extreme \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:1000\mathrm{-task}$$\end{document}stress test, the Round-Robin baseline completely destabilizes, and swarm intelligence models (PSO, GA) drop nearly \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:40\mathrm{\%}\mathrm{-}50\mathrm{\%}$$\end{document}of the critical payload. Conversely, QFLN-AMRO maintains a near-perfect Task Admission Rate of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:98.7\mathrm{\%}$$\end{document}alongside an unprecedented Lyapunov Stability Ratio of 0.0105. Technical Implication: This numerical evidence fundamentally validates the theoretical claim of “Very High (Lyapunov-Guided)” resilience presented in Table 13. The ultra-low stability ratio quantitatively proves that the proposed orchestrator possesses the deterministic capacity to absorb stochastic temporal burstiness (inspired by the Google Cluster Trace) without cascading queue overflow or violating URLLC latency bounds See Table 15. Technical Implication: This numerical evidence fundamentally validates the theoretical claim of “Very High (Lyapunov-Guided)” resilience presented in Table 13. The ultra-low stability ratio quantitatively proves that the proposed orchestrator possesses the deterministic capacity to absorb stochastic temporal burstiness (inspired by the Google Cluster Trace) without cascading queue overflow or violating URLLC latency bounds See Table 15. Table 14Quantitative performance evaluation of orchestration architectures (Under Dense Workload \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathrm{N}=1000$$\end{document}).Performance MetricRound-Robin (Baseline)Genetic Algorithm (GA)Particle Swarm (PSO)AMRO (Ablated Base)QFLN-AMRO (Proposed)Average Energy Cons. (J) \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:124.50\pm\:8.2$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:98.30\pm\:5.1$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:86.70\pm\:4.5$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:74.20\pm\:2.8$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:55.65\pm\:1.2$$\end{document} Average URLLC Latency (ms) \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:18.45\pm\:3.1$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:12.10\pm\:1.8$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:9.80\pm\:1.5$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:6.20\pm\:0.7$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:4.15\pm\:0.2$$\end{document} Hypervolume Indicator (HV)N/A (Single Obj.) 0.684 0.742 0.865 0.958 Task Admission Rate (%) \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:32.4\mathrm{\%}$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:51.2\mathrm{\%}$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:63.8\mathrm{\%}$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:82.5\mathrm{\%}$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:98.7\mathbf{\%}$$\end{document} Lyapunov Stability RatioDestabilized 0.485 0.392 0.084 0.0105 Convergence GenerationN/AGen 120Gen 85Gen 60Gen 22 Table 14Quantitative performance evaluation of orchestration architectures (Under Dense Workload \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathrm{N}=1000$$\end{document}).Performance MetricRound-Robin (Baseline)Genetic Algorithm (GA)Particle Swarm (PSO)AMRO (Ablated Base)QFLN-AMRO (Proposed)Average Energy Cons. (J) \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:124.50\pm\:8.2$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:98.30\pm\:5.1$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:86.70\pm\:4.5$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:74.20\pm\:2.8$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:55.65\pm\:1.2$$\end{document} Average URLLC Latency (ms) \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:18.45\pm\:3.1$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:12.10\pm\:1.8$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:9.80\pm\:1.5$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:6.20\pm\:0.7$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:4.15\pm\:0.2$$\end{document} Hypervolume Indicator (HV)N/A (Single Obj.) 0.684 0.742 0.865 0.958 Task Admission Rate (%) \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:32.4\mathrm{\%}$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:51.2\mathrm{\%}$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:63.8\mathrm{\%}$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:82.5\mathrm{\%}$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:98.7\mathbf{\%}$$\end{document} Lyapunov Stability RatioDestabilized 0.485 0.392 0.084 0.0105 Convergence GenerationN/AGen 120Gen 85Gen 60Gen 22 Quantitative performance evaluation of orchestration architectures (Under Dense Workload \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathrm{N}=1000$$\end{document}). Table 15List of Abbreviations.(Acronym)(Full Form)AMROAdaptive Multi-objective Resource OrchestratorDRLDeep Reinforcement LearningEUAEdge User AllocationGAGenetic AlgorithmHILHardware-in-the-LoopHSDHonestly Significant Difference (Tukey’s HSD)HVHypervolumeIIoTIndustrial Internet of ThingsLEOLow Earth OrbitMADRLMulti-Agent Deep Reinforcement LearningMECMulti-Access Edge ComputingNISQNoisy Intermediate-Scale QuantumNOMANon-Orthogonal Multiple AccessNP-hardNon-deterministic Polynomial-time hardNSGA-IINon-dominated Sorting Genetic Algorithm IIPSOParticle Swarm OptimizationQFLNQuantum Fuzzy Logic NetworkQoSQuality of ServiceURLLCUltra-Reliable Low-Latency Communication Table 15List of Abbreviations.(Acronym)(Full Form)AMROAdaptive Multi-objective Resource OrchestratorDRLDeep Reinforcement LearningEUAEdge User AllocationGAGenetic AlgorithmHILHardware-in-the-LoopHSDHonestly Significant Difference (Tukey’s HSD)HVHypervolumeIIoTIndustrial Internet of ThingsLEOLow Earth OrbitMADRLMulti-Agent Deep Reinforcement LearningMECMulti-Access Edge ComputingNISQNoisy Intermediate-Scale QuantumNOMANon-Orthogonal Multiple AccessNP-hardNon-deterministic Polynomial-time hardNSGA-IINon-dominated Sorting Genetic Algorithm IIPSOParticle Swarm OptimizationQFLNQuantum Fuzzy Logic NetworkQoSQuality of ServiceURLLCUltra-Reliable Low-Latency Communication List of Abbreviations. Conclusion and future work This paper introduced, formulated, and empirically validated the Quantum Fuzzy Logic-Enhanced Adaptive Multi-Objective Resource Orchestrator (QFLN-AMRO), a novel framework designed to address the NP-hard problem of dynamic task offloading in heterogeneous Edge-IIoT environments. By treating resource allocation as a native Multi-Objective Optimization Problem (MOOP), our approach circumvents the inherent limitations of traditional methods that rely on static scalarization, thereby avoiding premature convergence to sub-optimal solutions. The core of the QFLN-AMRO framework is a synergistic integration of a Quantum Fuzzy Logic Network, which adaptively tunes the operators of a Fractional-Order NSGA-II engine. This architecture is further enhanced by NOMA-aware transmission protocols and a Lyapunov-guided admission control mechanism, which collectively ensure both spectral efficiency and systemic stability. Rigorous evaluations were conducted within a high-fidelity simulation environment, leveraging a composite of real-world and synthetic datasets (EUA, Edge-IIoT, and Google Cluster Traces) to ensure the clinical relevance of our findings. The results demonstrate that QFLN-AMRO consistently outperforms conventional metaheuristic (e.g., GA, PSO) and heuristic (e.g., Greedy, Round-Robin) baselines. Statistically, the framework’s superiority was confirmed by a highly significant omnibus ANOVA result (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:p\ll\:0.001$$\end{document}), showcasing its ability to redefine the achievable Pareto-optimal frontier. Furthermore, QFLN-AMRO exhibited exceptional stability, evidenced by a minimal coefficient of variation ( 0.0105) across 30 independent runs. Under extreme load conditions, its integrated resilience mechanism effectively prevented network saturation, ensuring robust structural survivability and SLA compliance. In summary, QFLN-AMRO provides a deterministic and computationally efficient solution that simultaneously achieves ultra-low latency and promotes energy-conscious green computing. It establishes a robust and scalable architectural foundation poised to meet the demanding operational requirements of sustainable Industry 4.0 applications. Future work Building upon this foundational research, we identify four strategic trajectories for future investigation to align the QFLN-AMRO architecture with the emergent demands of 6G networks and Industry 5.0: Orchestration in Non-Terrestrial and Dynamic Topologies: The current framework assumes a static spatial topology. A critical next step is to extend the orchestration model to highly mobile and three-dimensional environments, including Unmanned Aerial Vehicle (UAV) swarms, High-Altitude Platform Stations (HAPS), and Low Earth Orbit (LEO) satellite edge nodes. This will necessitate reformulating the propagation models into dynamic, time-varying graphs.Decentralized Control via Multi-Agent Reinforcement Learning: To further reduce decision-making latency into the sub-millisecond range and enhance data privacy, we propose transitioning from the current centralized evolutionary engine to a decentralized Multi-Agent Deep Reinforcement Learning (MADRL) framework. Augmenting this with Federated Learning (FL) would enable distributed, on-device intelligence without exposing raw data.Hardware-in-the-Loop and Digital Twin Validation: To bridge the gap between simulation and real-world deployment, future work will focus on validating the framework on physical testbeds. This involves implementing QFLN-AMRO in Kubernetes-based containerized environments (e.g., KubeEdge) and integrating it with a Digital Twin (DT). This Hardware-in-the-Loop (HIL) approach is essential for quantifying performance against practical challenges such as virtualization overhead and thermal throttling.Migration to Native Quantum Hardware: The success of the quantum-inspired cognitive engine motivates a long-term research goal: mapping the simulated quantum states and superposition mechanics onto physical qubit-based hardware. Executing the orchestration logic on Noisy Intermediate-Scale Quantum (NISQ) devices could unlock unprecedented computational advantages, potentially achieving true quantum supremacy for real-time edge resource management. Orchestration in Non-Terrestrial and Dynamic Topologies: The current framework assumes a static spatial topology. A critical next step is to extend the orchestration model to highly mobile and three-dimensional environments, including Unmanned Aerial Vehicle (UAV) swarms, High-Altitude Platform Stations (HAPS), and Low Earth Orbit (LEO) satellite edge nodes. This will necessitate reformulating the propagation models into dynamic, time-varying graphs. Orchestration in Non-Terrestrial and Dynamic Topologies: The current framework assumes a static spatial topology. A critical next step is to extend the orchestration model to highly mobile and three-dimensional environments, including Unmanned Aerial Vehicle (UAV) swarms, High-Altitude Platform Stations (HAPS), and Low Earth Orbit (LEO) satellite edge nodes. This will necessitate reformulating the propagation models into dynamic, time-varying graphs. Decentralized Control via Multi-Agent Reinforcement Learning: To further reduce decision-making latency into the sub-millisecond range and enhance data privacy, we propose transitioning from the current centralized evolutionary engine to a decentralized Multi-Agent Deep Reinforcement Learning (MADRL) framework. Augmenting this with Federated Learning (FL) would enable distributed, on-device intelligence without exposing raw data. Decentralized Control via Multi-Agent Reinforcement Learning: To further reduce decision-making latency into the sub-millisecond range and enhance data privacy, we propose transitioning from the current centralized evolutionary engine to a decentralized Multi-Agent Deep Reinforcement Learning (MADRL) framework. Augmenting this with Federated Learning (FL) would enable distributed, on-device intelligence without exposing raw data. Hardware-in-the-Loop and Digital Twin Validation: To bridge the gap between simulation and real-world deployment, future work will focus on validating the framework on physical testbeds. This involves implementing QFLN-AMRO in Kubernetes-based containerized environments (e.g., KubeEdge) and integrating it with a Digital Twin (DT). This Hardware-in-the-Loop (HIL) approach is essential for quantifying performance against practical challenges such as virtualization overhead and thermal throttling. Hardware-in-the-Loop and Digital Twin Validation: To bridge the gap between simulation and real-world deployment, future work will focus on validating the framework on physical testbeds. This involves implementing QFLN-AMRO in Kubernetes-based containerized environments (e.g., KubeEdge) and integrating it with a Digital Twin (DT). This Hardware-in-the-Loop (HIL) approach is essential for quantifying performance against practical challenges such as virtualization overhead and thermal throttling. Migration to Native Quantum Hardware: The success of the quantum-inspired cognitive engine motivates a long-term research goal: mapping the simulated quantum states and superposition mechanics onto physical qubit-based hardware. Executing the orchestration logic on Noisy Intermediate-Scale Quantum (NISQ) devices could unlock unprecedented computational advantages, potentially achieving true quantum supremacy for real-time edge resource management. Migration to Native Quantum Hardware: The success of the quantum-inspired cognitive engine motivates a long-term research goal: mapping the simulated quantum states and superposition mechanics onto physical qubit-based hardware. Executing the orchestration logic on Noisy Intermediate-Scale Quantum (NISQ) devices could unlock unprecedented computational advantages, potentially achieving true quantum supremacy for real-time edge resource management.