Scholarlink

<p>Reinforcement Learning (RL) has been widely investigated in Software Defined Networks (SDN) during the last few years, as a promising technique for various network challenges, such as routing optimization. In this context, different RL algorithms have been deployed, which are based on either, policy-optimization or value-iteration methods. According to this latter, only the Q-Learning algorithm has been considered in the literature for routing optimization in SDN. In this paper, we propose a comparative evaluation of the two main algorithms according to value-iteration methods, namely Q-Learning and State-Action-Reward-State-Action (SARSA), to address latency minimization within a multipath flow-routing SDN network. Through various evaluations, we show that SARSA performs better than Q-Learning regarding convergence time and computation time, while achieving comparable performance in terms of average latency. Furthermore, we propose, for the first time in this work, to implement a novel hybrid RL approach in the SDN environment, namely, QL-SARSA, to take advantage of both algorithms. Findings demonstrate that QL-SARSA effectively improves network traffic management by reducing flow latency, convergence time, computation time, and jitter while enhancing throughput. Further performance improvements are achieved by integrating a congestion avoidance mechanism and the Max-Boltzman Exploration (MBE) strategy in the QL-SARSA algorithm.</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

A hybrid reinforcement learning approach for multipath routing optimization in software-defined networks

  • 期刊论文

  • 发表日期：

    2025年7月1日

查看全文
  • Houda Hassen,
  • Soumaya Meherzi,
  • Zouhair Ben Jemaa

摘要

Reinforcement Learning (RL) has been widely investigated in Software Defined Networks (SDN) during the last few years, as a promising technique for various network challenges, such as routing optimization. In this context, different RL algorithms have been deployed, which are based on either, policy-optimization or value-iteration methods. According to this latter, only the Q-Learning algorithm has been considered in the literature for routing optimization in SDN. In this paper, we propose a comparative evaluation of the two main algorithms according to value-iteration methods, namely Q-Learning and State-Action-Reward-State-Action (SARSA), to address latency minimization within a multipath flow-routing SDN network. Through various evaluations, we show that SARSA performs better than Q-Learning regarding convergence time and computation time, while achieving comparable performance in terms of average latency. Furthermore, we propose, for the first time in this work, to implement a novel hybrid RL approach in the SDN environment, namely, QL-SARSA, to take advantage of both algorithms. Findings demonstrate that QL-SARSA effectively improves network traffic management by reducing flow latency, convergence time, computation time, and jitter while enhancing throughput. Further performance improvements are achieved by integrating a congestion avoidance mechanism and the Max-Boltzman Exploration (MBE) strategy in the QL-SARSA algorithm.