Scholarlink

<p>The rapid proliferation of Internet of Things (IoT) devices and latency-sensitive applications has amplified the need for efficient task scheduling in hybrid cloud-edge environments. Traditional heuristic and metaheuristic algorithms often fall short in addressing the dynamic nature of workloads and the conflicting objectives of performance, energy efficiency, and cost-effectiveness. To overcome these challenges, this study introduces <b>R</b>einforcement <b>L</b>earning-Based <b>M</b>ulti-<b>O</b>bjective <b>T</b>ask <b>S</b>cheduling (RL-MOTS), a framework leveraging Deep Q-Networks (DQNs) for intelligent and adaptive resource allocation. The proposed model formulates scheduling as a Markov Decision Process, incorporating a priority-aware dynamic queueing mechanism and a multi-objective reward function that balances task latency, energy consumption, and operational costs. Additionally, the framework employs a state-reward tensor to capture trade-offs among objectives, enabling real-time decision-making across heterogeneous cloud and edge nodes. Comprehensive simulations using CloudSim validate the robustness of RL-MOTS under varying workload conditions. Compared to baseline strategies such as FCFS, Min-Min, and multi-objective heuristic models, RL-MOTS achieves up to 28% reduction in energy consumption, 20% improvement in cost efficiency, and significant reductions in makespan and deadline violations, while maintaining strict Quality of Service (QoS) requirements. The framework’s adaptability to preemptive and non-preemptive scheduling further enhances its resilience and scalability. These findings establish RL-MOTS as a forward-looking solution for sustainable, cost-efficient, and performance-oriented computing in next-generation distributed systems. Future research will focus on integrating transfer learning and federated learning to increase scalability and privacy in large, decentralized environments, including those applicable to the medical industry.</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

Reinforcement learning based multi objective task scheduling for energy efficient and cost effective cloud edge computing

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年11月24日

查看全文
  • Wenfan Zhang,
  • Haijiao Ou

摘要

The rapid proliferation of Internet of Things (IoT) devices and latency-sensitive applications has amplified the need for efficient task scheduling in hybrid cloud-edge environments. Traditional heuristic and metaheuristic algorithms often fall short in addressing the dynamic nature of workloads and the conflicting objectives of performance, energy efficiency, and cost-effectiveness. To overcome these challenges, this study introduces Reinforcement Learning-Based Multi-Objective Task Scheduling (RL-MOTS), a framework leveraging Deep Q-Networks (DQNs) for intelligent and adaptive resource allocation. The proposed model formulates scheduling as a Markov Decision Process, incorporating a priority-aware dynamic queueing mechanism and a multi-objective reward function that balances task latency, energy consumption, and operational costs. Additionally, the framework employs a state-reward tensor to capture trade-offs among objectives, enabling real-time decision-making across heterogeneous cloud and edge nodes. Comprehensive simulations using CloudSim validate the robustness of RL-MOTS under varying workload conditions. Compared to baseline strategies such as FCFS, Min-Min, and multi-objective heuristic models, RL-MOTS achieves up to 28% reduction in energy consumption, 20% improvement in cost efficiency, and significant reductions in makespan and deadline violations, while maintaining strict Quality of Service (QoS) requirements. The framework’s adaptability to preemptive and non-preemptive scheduling further enhances its resilience and scalability. These findings establish RL-MOTS as a forward-looking solution for sustainable, cost-efficient, and performance-oriented computing in next-generation distributed systems. Future research will focus on integrating transfer learning and federated learning to increase scalability and privacy in large, decentralized environments, including those applicable to the medical industry.