Scholarlink

<p>This paper presents a dynamic resource orchestration framework for edge computing environments, utilizing multi-agent reinforcement learning (MARL) to enhance resource allocation and task scheduling. The proposed system consists of edge nodes (E), a centralized resource manager (CRM), and communication infrastructure (CI). Edge nodes execute computational tasks at the network’s periphery, while the CRM oversees resource distribution and task assignment using a global system perspective. The CI supports efficient communication among these components. The MARL framework enables collaborative learning among agents, where each agent selects optimal actions—such as resource allocation, task scheduling, and migration—based on system states that include resource availability, task queue lengths, network conditions, and task priorities. A deep Q-network (DQN)-based training approach is employed, allowing agents to maximize cumulative rewards by balancing task completion efficiency, resource utilization, and latency minimization. The proposed framework is evaluated through comprehensive simulations against traditional heuristic-based and static resource allocation methods. Results demonstrate that our MARL-based approach reduces average task completion latency by 12.3% and improves resource utilization by 8.7% compared to heuristic baselines. Additionally, the framework dynamically adapts to variations in network conditions and workload distribution, ensuring consistent quality of service (QoS) under dynamic edge computing scenarios.</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

Dynamic resource orchestration in edge computing environments using multi-agent reinforcement learning

  • 期刊论文

  • 发表日期：

    2025年6月23日

查看全文
  • Qi Liu,
  • Jianzheng Yang,
  • Zhixian Yan

摘要

This paper presents a dynamic resource orchestration framework for edge computing environments, utilizing multi-agent reinforcement learning (MARL) to enhance resource allocation and task scheduling. The proposed system consists of edge nodes (E), a centralized resource manager (CRM), and communication infrastructure (CI). Edge nodes execute computational tasks at the network’s periphery, while the CRM oversees resource distribution and task assignment using a global system perspective. The CI supports efficient communication among these components. The MARL framework enables collaborative learning among agents, where each agent selects optimal actions—such as resource allocation, task scheduling, and migration—based on system states that include resource availability, task queue lengths, network conditions, and task priorities. A deep Q-network (DQN)-based training approach is employed, allowing agents to maximize cumulative rewards by balancing task completion efficiency, resource utilization, and latency minimization. The proposed framework is evaluated through comprehensive simulations against traditional heuristic-based and static resource allocation methods. Results demonstrate that our MARL-based approach reduces average task completion latency by 12.3% and improves resource utilization by 8.7% compared to heuristic baselines. Additionally, the framework dynamically adapts to variations in network conditions and workload distribution, ensuring consistent quality of service (QoS) under dynamic edge computing scenarios.