Scholarlink

In complex and dynamic vehicular edge computing (VEC) environments, task offloading is critical but often hindered by high mobility, incomplete information, and frequent topology changes, which result in unpredictable latency and severe load imbalance. To address these issues, we propose DTO-LB, a deep reinforcement learning-based joint optimization framework for task offloading. To ensure effective task offloading under incomplete information, we introduce a Vehicle-Relationship (V-R) model, which integrates signal strength, vehicle quality, and task urgency into an extended Coulomb force model to better capture inter-vehicle interactions. This V-R model guides vehicle clustering and is used to formulate an Imprecise Game Theory (IGT) framework for inter-group task offloading. On this basis, we design a Multi-Agent Deep Deterministic Potential Function Strategy Gradient (MADDPFSG) algorithm, which enables agents to make robust decisions under high dynamics, thereby reducing service latency. To address load imbalance caused by frequent topology changes, we introduce a Dynamic Weight Updating (DWU) mechanism, which employs threshold-based detection to identify and mitigate malicious or overloaded base stations (BSs). This strategy adaptively adjusts task allocations, improving system robustness and efficiency. Experimental results demonstrate that DTO-LB outperforms baselines, improving cumulative reward (0.63%), success rate (2.8%), and completion time (0.65%). It maintains robustness under attacks, with success rates declining from 93% (no attack) to 86% (weak), 72% (moderate), and 60% (strong), demonstrating effective offloading optimization in dynamic adversarial environments.

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

Task Offloading in Dynamic Vehicular Networks Based on Deep Reinforcement Learning

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Yini Pu,
  • Xiaolin Zhang,
  • Jingyu Wang,
  • Lixin Liu,
  • Jinling Yu,
  • Lisha Li,
  • Cuiyun Shi

摘要

In complex and dynamic vehicular edge computing (VEC) environments, task offloading is critical but often hindered by high mobility, incomplete information, and frequent topology changes, which result in unpredictable latency and severe load imbalance. To address these issues, we propose DTO-LB, a deep reinforcement learning-based joint optimization framework for task offloading. To ensure effective task offloading under incomplete information, we introduce a Vehicle-Relationship (V-R) model, which integrates signal strength, vehicle quality, and task urgency into an extended Coulomb force model to better capture inter-vehicle interactions. This V-R model guides vehicle clustering and is used to formulate an Imprecise Game Theory (IGT) framework for inter-group task offloading. On this basis, we design a Multi-Agent Deep Deterministic Potential Function Strategy Gradient (MADDPFSG) algorithm, which enables agents to make robust decisions under high dynamics, thereby reducing service latency. To address load imbalance caused by frequent topology changes, we introduce a Dynamic Weight Updating (DWU) mechanism, which employs threshold-based detection to identify and mitigate malicious or overloaded base stations (BSs). This strategy adaptively adjusts task allocations, improving system robustness and efficiency. Experimental results demonstrate that DTO-LB outperforms baselines, improving cumulative reward (0.63%), success rate (2.8%), and completion time (0.65%). It maintains robustness under attacks, with success rates declining from 93% (no attack) to 86% (weak), 72% (moderate), and 60% (strong), demonstrating effective offloading optimization in dynamic adversarial environments.