<p>Efficient and privacy-aware task offloading remains a critical challenge for 6&#xa0;G-enabled Internet of Vehicles (IoV) due to highly dynamic vehicular mobility, heterogeneous edge resources, and stringent latency and energy constraints. To address these challenges, an intelligent task offloading framework based on federated learning-assisted multi-agent deep deterministic policy gradient Internet of Vehicles (FL-MADDPG-IoV) is developed for a multi-tier edge computing architecture comprising roadside units, unmanned aerial vehicle-based edge servers, and macro base stations. The framework integrates federated reinforcement learning with hierarchical multi-agent coordination, enabling distributed edge agents to collaboratively optimize offloading policies while preserving data privacy by eliminating the exchange of raw data. The learning process is evaluated on a large-scale dataset comprising 184,482 samples with six statistical features, demonstrating adaptability across varying vehicular densities and network conditions. Simulation results indicate substantial performance improvements over existing deep reinforcement learning and federated learning approaches. Specifically, task latency is reduced by <InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(62\%\)</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mn>62</mn> <mo>%</mo> </mrow> </math></EquationSource> </InlineEquation> (mean&#xa0;±&#xa0;2.1% across 5 independent runs), energy consumption by <InlineEquation ID="IEq2"> <EquationSource Format="TEX">\(65\%\)</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mn>65</mn> <mo>%</mo> </mrow> </math></EquationSource> </InlineEquation> (mean&#xa0;±&#xa0;0.8%), and system throughput is improved by <InlineEquation ID="IEq3"> <EquationSource Format="TEX">\(90\%\)</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mn>90</mn> <mo>%</mo> </mrow> </math></EquationSource> </InlineEquation>, all under the defined simulation conditions. In addition, the federated model achieves <InlineEquation ID="IEq4"> <EquationSource Format="TEX">\(14\%\)</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mn>14</mn> <mo>%</mo> </mrow> </math></EquationSource> </InlineEquation> to <InlineEquation ID="IEq5"> <EquationSource Format="TEX">\(16\%\)</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mn>16</mn> <mo>%</mo> </mrow> </math></EquationSource> </InlineEquation> improvement in accuracy over benchmark methods. Pareto-optimal analysis further reveals an effective trade-off between latency (0.23 s) and energy consumption (0.025 J), establishing the framework’s real-time decision-making ability, with actor inference for a single step taking less than 5&#xa0;ms—a fraction of the time window of 40-100&#xa0;s during which vehicles are connected to the UAV at the simulated speeds of 10-25&#xa0;m/s. The results confirm the success of the proposed FL-MADDPG-IoV framework as a scalable, autonomous and energy-efficient approach.</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

FL-MADDPG-IoV: federated multi-agent deep reinforcement learning for task offloading in 6G IoV using multi-tier UAV-enabled edge computing

  • Sofia Shafiq,
  • Muhammad Waseem Iqbal,
  • Ahsan Humayun,
  • Sohail Jabbar,
  • Muhammad Asif Habib

摘要

Efficient and privacy-aware task offloading remains a critical challenge for 6 G-enabled Internet of Vehicles (IoV) due to highly dynamic vehicular mobility, heterogeneous edge resources, and stringent latency and energy constraints. To address these challenges, an intelligent task offloading framework based on federated learning-assisted multi-agent deep deterministic policy gradient Internet of Vehicles (FL-MADDPG-IoV) is developed for a multi-tier edge computing architecture comprising roadside units, unmanned aerial vehicle-based edge servers, and macro base stations. The framework integrates federated reinforcement learning with hierarchical multi-agent coordination, enabling distributed edge agents to collaboratively optimize offloading policies while preserving data privacy by eliminating the exchange of raw data. The learning process is evaluated on a large-scale dataset comprising 184,482 samples with six statistical features, demonstrating adaptability across varying vehicular densities and network conditions. Simulation results indicate substantial performance improvements over existing deep reinforcement learning and federated learning approaches. Specifically, task latency is reduced by \(62\%\) 62 % (mean ± 2.1% across 5 independent runs), energy consumption by \(65\%\) 65 % (mean ± 0.8%), and system throughput is improved by \(90\%\) 90 % , all under the defined simulation conditions. In addition, the federated model achieves \(14\%\) 14 % to \(16\%\) 16 % improvement in accuracy over benchmark methods. Pareto-optimal analysis further reveals an effective trade-off between latency (0.23 s) and energy consumption (0.025 J), establishing the framework’s real-time decision-making ability, with actor inference for a single step taking less than 5 ms—a fraction of the time window of 40-100 s during which vehicles are connected to the UAV at the simulated speeds of 10-25 m/s. The results confirm the success of the proposed FL-MADDPG-IoV framework as a scalable, autonomous and energy-efficient approach.