Scholarlink

<p>The paradigm shift from centralized cloud computing to Multi-Access Edge Computing (MEC) has been necessitated by the exponential growth of the Internet of Things (IoT) and the stringent latency requirements of 5G/6G applications. However, the inherent dynamism of edge environments, characterized by fluctuating channel conditions, heterogeneous device capabilities, and finite energy budgets, renders traditional static optimization techniques insufficient. This survey provides a critical review of Deep Learning (DL) and Deep Reinforcement Learning (DRL) as pivotal enablers for autonomous computation offloading. This work systematically dissects the algorithmic trade-offs between Value-Based (e.g., DQN) and Policy-Gradient (e.g., PPO, SAC, A3C) architectures, evaluating their suitability for continuous control, multi-agent coordination, and resource-constrained inference. We further bridge the gap between theoretical algorithms and practical telecommunications deployment by mapping DRL strategies to European Telecommunications Standards Institute (ETSI) MEC specifications and Open Radio Access Network (O-RAN) architectural standards. The review analyzes the "optimization trade-offs" inherent in offloading, specifically the conflict between service immediacy (latency) and resource conservation (energy/security). Finally, we synthesize key challenges hindering large-scale adoption, including the need for lightweight "extreme edge" models, robustness against non-stationary environments, and privacy-preserving federated learning, thereby providing a rigorous roadmap for the evolution of self-optimizing, intelligent edge ecosystems.</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

Deep learning approaches for computation offloading in edge computing: A critical review

  • 期刊论文

  • 发表日期：

    2026年2月24日

查看全文
  • Sapthagiri Miriyala,
  • Venkata Ramireddy Chirra

摘要

The paradigm shift from centralized cloud computing to Multi-Access Edge Computing (MEC) has been necessitated by the exponential growth of the Internet of Things (IoT) and the stringent latency requirements of 5G/6G applications. However, the inherent dynamism of edge environments, characterized by fluctuating channel conditions, heterogeneous device capabilities, and finite energy budgets, renders traditional static optimization techniques insufficient. This survey provides a critical review of Deep Learning (DL) and Deep Reinforcement Learning (DRL) as pivotal enablers for autonomous computation offloading. This work systematically dissects the algorithmic trade-offs between Value-Based (e.g., DQN) and Policy-Gradient (e.g., PPO, SAC, A3C) architectures, evaluating their suitability for continuous control, multi-agent coordination, and resource-constrained inference. We further bridge the gap between theoretical algorithms and practical telecommunications deployment by mapping DRL strategies to European Telecommunications Standards Institute (ETSI) MEC specifications and Open Radio Access Network (O-RAN) architectural standards. The review analyzes the "optimization trade-offs" inherent in offloading, specifically the conflict between service immediacy (latency) and resource conservation (energy/security). Finally, we synthesize key challenges hindering large-scale adoption, including the need for lightweight "extreme edge" models, robustness against non-stationary environments, and privacy-preserving federated learning, thereby providing a rigorous roadmap for the evolution of self-optimizing, intelligent edge ecosystems.