Scholarlink

<p>This paper presents a reinforcement learning-based self-healing framework for Software-Defined Networking (SDN) that autonomously manages diverse network faults in a realistically emulated environment. A Mininet testbed controlled by the Faucet SDN controller is instrumented with Prometheus to collect multi-source telemetry, while an automated fault injector and congestion generator produce link, port, flow and controller events alongside UDP-induced bottlenecks to create rich training data. Network features–including controller CPU and memory usage, OpenFlow statistics, port status and explicit fault labels–are periodically scraped and aggregated into a structured dataset that forms the state space of a custom Gym-compatible environment. A Proximal Policy Optimisation (PPO) agent with a multilayer perceptron policy learns discrete self-healing actions such as no-op, port resets, switch restarts and bespoke recovery procedures, guided by a reward function that penalises persistent faults and unnecessary interventions while rewarding timely and appropriate recovery. Experimental evaluation over multiple PPO training runs shows stable optimisation behaviour and high episodic rewards with long episode lengths. Policy output analysis stratified by fault state confirms that the agent has learned state-conditional recovery behaviour, selecting fault-type-appropriate actions in over 85% of fault-state timesteps, thereby providing direct evidence that the agent successfully distinguishes healthy from faulty conditions and among different fault types at the level of individual recovery decisions. Compared with existing RL-based approaches that focus primarily on link failure recovery or service function chain reconfiguration, the proposed framework handles a broader spectrum of SDN fault types and integrates control-plane, data-plane and congestion indicators, thereby offering a more general and robust self-healing capability for operational SDN environments.</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 reinforcement learning framework for self-healing fault recovery in intent-based SDNs

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年5月17日

查看全文
  • E. H. Makiyah,
  • M. N. Rasool,
  • F. A. Rawdhan

摘要

This paper presents a reinforcement learning-based self-healing framework for Software-Defined Networking (SDN) that autonomously manages diverse network faults in a realistically emulated environment. A Mininet testbed controlled by the Faucet SDN controller is instrumented with Prometheus to collect multi-source telemetry, while an automated fault injector and congestion generator produce link, port, flow and controller events alongside UDP-induced bottlenecks to create rich training data. Network features–including controller CPU and memory usage, OpenFlow statistics, port status and explicit fault labels–are periodically scraped and aggregated into a structured dataset that forms the state space of a custom Gym-compatible environment. A Proximal Policy Optimisation (PPO) agent with a multilayer perceptron policy learns discrete self-healing actions such as no-op, port resets, switch restarts and bespoke recovery procedures, guided by a reward function that penalises persistent faults and unnecessary interventions while rewarding timely and appropriate recovery. Experimental evaluation over multiple PPO training runs shows stable optimisation behaviour and high episodic rewards with long episode lengths. Policy output analysis stratified by fault state confirms that the agent has learned state-conditional recovery behaviour, selecting fault-type-appropriate actions in over 85% of fault-state timesteps, thereby providing direct evidence that the agent successfully distinguishes healthy from faulty conditions and among different fault types at the level of individual recovery decisions. Compared with existing RL-based approaches that focus primarily on link failure recovery or service function chain reconfiguration, the proposed framework handles a broader spectrum of SDN fault types and integrates control-plane, data-plane and congestion indicators, thereby offering a more general and robust self-healing capability for operational SDN environments.