Scholarlink

<p>Multi-access edge computing (MEC) enables low-latency services by bringing computation closer to mobile users, but joint task offloading and resource allocation remain challenging due to user mobility, time-varying wireless links, and limited edge capacity. This paper proposes a constrained learning framework for Artificial Intelligence-Defined Wireless Networking (AIDWN)-enabled multi-MEC systems based on a constrained Markov decision process (CMDP). Three key novelties distinguish the proposed approach: (i)&#xa0;a CMDP formulation of the joint offloading and bandwidth allocation problem with explicit capacity and delay constraints enforced via Lagrangian dual-variable updates; (ii)&#xa0;a queue-aware end-to-end delay model that jointly captures transmission, queueing, and execution components, enabling congestion-sensitive policy learning; and (iii)&#xa0;a Constrained Double Deep Q-Network (C-DDQN) augmented with adaptive service-rate estimation and one-step delay prediction, embedded directly in the learning state to support proactive congestion control under non-stationary traffic. Simulation results under congested network conditions with up to <InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(U{=}100\)</EquationSource> </InlineEquation> users demonstrate stable convergence and improved constraint feasibility: compared with an unconstrained DDQN baseline, C-DDQN reduces capacity and delay violation rates by <b>76%</b> and <b>70%</b>, respectively, while achieving a Jain fairness index of <InlineEquation ID="IEq2"> <EquationSource Format="TEX">\(J{=}0.950\)</EquationSource> </InlineEquation> and maintaining competitive throughput. Robustness experiments under bursty Pareto ON/OFF traffic confirm a <b>58%</b> violation reduction over the unconstrained baseline even under traffic model mismatch. The source code is publicly available at: <a href="https://github.com/amzil-abdellah/C-DDQN-MEC">https://github.com/amzil-abdellah/C-DDQN-MEC</a>.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

A constrained deep reinforcement learning framework for joint task offloading and resource allocation in multi-access edge computing systems

  • 期刊论文

  • 发表日期：

    2026年6月18日

查看全文
  • Abdellah Amzil,
  • Mohamed Hanini,
  • Said El Kafhali

摘要

Multi-access edge computing (MEC) enables low-latency services by bringing computation closer to mobile users, but joint task offloading and resource allocation remain challenging due to user mobility, time-varying wireless links, and limited edge capacity. This paper proposes a constrained learning framework for Artificial Intelligence-Defined Wireless Networking (AIDWN)-enabled multi-MEC systems based on a constrained Markov decision process (CMDP). Three key novelties distinguish the proposed approach: (i) a CMDP formulation of the joint offloading and bandwidth allocation problem with explicit capacity and delay constraints enforced via Lagrangian dual-variable updates; (ii) a queue-aware end-to-end delay model that jointly captures transmission, queueing, and execution components, enabling congestion-sensitive policy learning; and (iii) a Constrained Double Deep Q-Network (C-DDQN) augmented with adaptive service-rate estimation and one-step delay prediction, embedded directly in the learning state to support proactive congestion control under non-stationary traffic. Simulation results under congested network conditions with up to \(U{=}100\) users demonstrate stable convergence and improved constraint feasibility: compared with an unconstrained DDQN baseline, C-DDQN reduces capacity and delay violation rates by 76% and 70%, respectively, while achieving a Jain fairness index of \(J{=}0.950\) and maintaining competitive throughput. Robustness experiments under bursty Pareto ON/OFF traffic confirm a 58% violation reduction over the unconstrained baseline even under traffic model mismatch. The source code is publicly available at: https://github.com/amzil-abdellah/C-DDQN-MEC.