【作者】 王辉；

【导师】 吴国文； 沈士根；

【作者基本信息】 东华大学 ， 电子信息， 2025， 硕士

【摘要】 随着移动通信技术与物联网技术的深度融合,工业物联网(Industrial Internet of Things,IIoT)已成为推动智能制造发展的重要技术范式。在IIoT架构中,感知层智能设备产生的实时生产数据需传输至计算中心进行处理分析,而传统云计算模式在实时性要求和资源利用效率方面面临挑战。移动边缘计算(Mobile Edge Com-puting,MEC)通过在网络边缘部署计算节点,有效缓解了核心网络的计算压力,但边缘服务器的资源约束特性对计算任务的调度优化提出了新的要求。如何构建高效的计算卸载策略以优化时延性能与能耗效率,成为提升IIoT系统整体效能的关键问题。本文针对MEC环境下的两类典型场景展开研究:在单服务器资源分配场景中,提出基于随机博弈的优化方法;在多服务器负载均衡场景中,构建基于平均场博弈的协同机制。通过博弈论与多智能体深度强化学习(Multi-Agent Deep Rein-forcement Learning,MADRL)的融合创新,形成了系统性的计算卸载解决方案。同时为了实现对MEC系统中的资源调度和网络管理,本文中还引入了了软件定义网络(Software Defined Network,SDN)框架,通过逻辑中心化的SDN控制器,完成对网络的控制和管理。针对单MEC服务器场景下的多维资源分配问题,本文创新性地建立了随机博弈建模框架。将用户设备抽象为博弈参与者,其策略空间包含任务卸载决策与资源请求行为,收益函数综合考量时延敏感度与能耗成本。通过严格的数学推导,本文论证了该博弈模型存在纳什均衡解,并提出带有优先经验回放机制的随机博弈资源分配(Stochastic Game-based Resource Allocation Algorithm with Proitized Ex-perience Replays,SGRA-PER)算法实现均衡解的智能求解。该方法采用多智能体协同学习机制,通过引入优先经验回放机制优化算法性能,使智能体能在动态环境中自主调整策略。实验研究表明,相较于其他算法,该算法在任务处理时延、服务质量保障等方面展现出显著优势,同时在用户数量增多的场景下保持稳定的优化效果。针对多MEC服务器集群的负载均衡问题,本文构建了基于平均场近似的新型博弈模型。该模型将海量用户的交互行为转化为群体效应表征,通过场论方法将高维决策空间降维至可处理范围。在算法设计层面,创新性地融合平均场博弈理论与深度强化学习技术,使单个智能体的决策仅需考虑系统平均状态而非所有个体行为,有效解决了维度灾难问题,提出了基于平均场博弈的负载均衡方法(Mean-Field Game-based Load Balancing,MFGLB)。理论分析表明,该方法在保证纳什均衡存在性的同时,将计算复杂度从指数级降低至线性级。仿真验证显示,相较于经典负载均衡算法,该策略在系统收敛速度、资源利用率等核心指标上具有明显提升,为大规模IIoT应用场景提供了可行的解决方案。更多还原

【Abstract】 This study addresses two representative scenarios in MEC environments:a stochas-tic game-based optimization method for single-server resource allocation and an average-field game-inspired collaborative mechanism for multi-server load balancing.Through the innovative integration of game theory and Multi-Agent Deep Reinforcement Learn-ing(MADRL),systematic computation offloading solutions are developed.To address resource scheduling and network management in MEC systems,this study incorporates a Software-Defined Networking(SDN)framework,where logically centralized SDN con-trollers enable coordinated network orchestration and resource governance.This archi-tectural integration establishes a unified control plane for dynamic infrastructure opti-mization across distributed edge computing environments.For multi-dimensional resource allocation in single-MEC server scenarios,a stochas-tic game-theoretic framework is established.User equipment is abstracted as game partic-ipants whose strategy spaces encompass task offloading decisions and resource requests,with utility functions incorporating latency sensitivity and energy costs.Mathematical proofs confirm the existence of Nash equilibrium solutions,leading to the development of the SGRA-PER algorithm for intelligent equilibrium attainment.This approach em-ploys a multi-agent collaborative learning mechanism enhanced by prioritized experience replay,enabling autonomous strategy adaptation in dynamic environments.Comparative studies demonstrate the strategy’s superior performance in system throughput and service quality assurance,particularly maintaining stable optimization under high-concurrency conditions.Addressing load balancing in multi-MEC server clusters,this research overcomes computational complexity limitations of traditional game-theoretic methods by construct-ing an average-field approximated game model MFGLB.The model transforms massive user interactions into population effect representations,reducing high-dimensional de-cision spaces through field-theoretic approaches.Algorithmically,it innovatively inte-grates mean-field game theory with deep reinforcement learning,enabling individual agents to base decisions on systemic average states rather than exhaustive peer analy-sis,effectively resolving the curse of dimensionality.Theoretical analyses confirm the method’s Nash equilibrium existence while reducing computational complexity from ex-ponential to linear scales.Simulation validations reveal significant improvements in con-vergence speed and resource utilization compared to classical load balancing algorithms,providing viable solutions for large-scale IIoT applications.更多还原