In an advance for autonomic network operations, computer systems researchers have introduced a reinforcement learning framework capable of autonomously managing and recovering from diverse network faults in Software-Defined Networking environments. Developed by E. H. Makiyah, M. N. Rasool, and F. A. Rawdhan, the system operates as a data-driven assurance engine within an intent-based networking paradigm, bridging high-level administrative policies and programmatic infrastructure adjustments. Evaluated on an emulated virtual network governed by the open-source Faucet controller, the framework deploys a Proximal Policy Optimisation agent that learns to diagnose operational anomalies and execute discrete remediation measures across both the data plane and the control plane. Experimental benchmarks show that the trained agent achieves autonomous fault recovery within an average of 26 seconds across controlled failure episodes. By executing timely interventions, the system reduced a packet loss proxy metric by 82.6 percent and cut OpenFlow error rates by 83.5 percent compared to unmanaged network conditions, establishing a viable approach for automated self-healing network fabrics.
Software-Defined Networking architecture decouples logically centralised control planes from underlying packet forwarding hardware through standardised southbound interfaces such as the OpenFlow protocol. While this programmable separation allows network operators to dynamically adjust routing paths and balance operational traffic loads, expanding deployments increasingly encounter complex disruptions that jeopardize overall availability. Unmitigated physical link failures, local port malfunctions, flow table entry inconsistencies, and unexpected controller crashes can trigger substantial service interruptions and degrade end-to-end quality of service. Traditional recovery mechanisms have relied on reactive recalculations that incur substantial restoration latency or proactive path pre-provisioning that consumes scarce, expensive Ternary Content-Addressable Memory space within hardware switches. Although prior artificial intelligence approaches applied reinforcement learning to network maintenance, earlier frameworks remained narrowly confined to isolated link rerouting tasks or virtual network function migrations, failing to observe comprehensive operational telemetry or address multi-category failures across the entire software-defined infrastructure.
To overcome these limitations, the research team constructed a reproducible virtual network testbed using the Mininet emulation platform, establishing a tree topology consisting of three OpenFlow switches and four end hosts connected to a remote Faucet controller operating on port 6653. The environment was instrumented with a continuous multi-source monitoring pipeline centered on Prometheus, which gathered telemetry by scraping data from the Faucet controller process, the Gauge switch-monitoring agent, and the underlying Node Exporter at fixed ten-second intervals. Alongside normal operational monitoring, the researchers implemented an automated fault injection and congestion generation suite. This automated harness systematically introduced randomized perturbations, including physical trunk link outages, individual port toggling events, Faucet controller reloads, and high-volume UDP flood attacks designed to replicate real-world transport congestion bottlenecks. The resulting data stream combined physical, logical, and performance indicators into a labeled telemetry log that preserved operational timestamps and explicit failure labels for algorithmic training.
The researchers mathematically formulated the autonomic self-healing problem as a discrete-time Markov Decision Process, defining network interactions through a five-element tuple comprising states, actions, transition dynamics, reward structures, and a discount factor. At each discrete ten-second monitoring timestep, the reinforcement learning agent receives an eight-dimensional continuous state vector scraped directly from the Prometheus monitoring stack. This telemetry vector incorporates specific operational variables: the Faucet controller process central processing unit execution time, resident memory byte allocations, virtual memory byte allocations, cumulative OpenFlow messages transmitted, accumulated OpenFlow error counts, active datapath connection tallies, datapath disconnection events, and the aggregate quantity of network switch ports residing in a down state. By integrating low-level process metrics with switch-level OpenFlow counters, the state space provides the agent with unified visibility into control-plane stability, data-plane forwarding integrity, and transient congestion spikes without relying on synthetic approximations or single-source telemetry feeds.
Decision-making within this Markovian environment is governed by a discrete action space containing four mutually exclusive restorative actions designed for automated deployment. The agent can select an action index of zero to execute a no-operation decision when the network operates smoothly, an index of one to reset all ports on an affected switch, an index of two to trigger a complete switch restart, or an index of three to initiate a bespoke self-healing procedure tailored for complex forwarding irregularities. To optimize policy parameters across this action space, the architecture utilizes the Proximal Policy Optimisation algorithm implemented via stable-baselines3 and TensorFlow, structured around a multilayer perceptron neural network policy. The optimization routine relies on a clipped surrogate objective function with a clipping parameter threshold set to 0.35, which constrains policy parameter updates between successive iterations. This clipping mechanism, paired with value function loss tracking and an exploratory entropy bonus, prevents policy collapse amidst stochastic network transitions.
Central to the agent’s decision quality is a severity-weighted reward function that assigns numerical penalties and bonuses directly proportional to the operational gravity of specific network disruptions. Structural link faults, which disconnect inter-switch communication and sever reachability, incur the heaviest penalty of negative ten and yield a restorative heal bonus of positive twelve. Local port faults receive a negative seven penalty and a positive ten bonus, while application-level flow errors receive a negative five penalty and a positive eight bonus. Minor anomalies, including soft and temporary warnings, are assigned mild penalties and bonuses ranging between negative two and positive two. Critically, restoring the control plane following a Faucet reload triggers a substantial positive reward bonus of twenty. To prevent erratic over-reaction, unnecessary interventions during completely healthy network states incur an explicit negative one penalty, while maintaining a no-operation state yields a positive two reward, establishing an asymmetric incentive structure that rewards decisive, timely intervention while punishing unneeded operational tampering.
The investigation evaluated the progressive contribution of individual reward components by training three distinct Proximal Policy Optimisation agent variants across ten thousand operational timesteps. The baseline variant, designated PPO1, received only infrastructure-level telemetry penalties based on down ports and error counters without awareness of specific fault types. The intermediate variant, PPO_FAUCET_HEAL_1, integrated explicit controller reload restoration bonuses, while the final proposed model, PPO_FAUCET_HEAL_2, incorporated full severity weighting, duration penalties, and fault-specific healing bonuses. Throughout training runs, TensorBoard diagnostic tracking revealed stable convergence, with learning rates adjusting smoothly, Kullback-Leibler divergence remaining tightly bounded, and processing throughput reaching approximately 880 frames per second. The fully enriched agent reached an average episodic reward of 1,550 with extended episode lengths averaging approximately 2,200 timesteps. Systematic sensitivity analysis across halved, base, and doubled reward scales confirmed that the selected base configuration yielded the fastest recovery times and highest cumulative rewards, avoiding value function instability.
Empirical validation across fifty controlled fault injection events on the base three-switch Mininet topology demonstrated that the trained agent substantially outperformed both unmanaged and deterministic management paradigms. When pitted against a no-healing baseline where faults persist until external clearing, and a rule-based reactive baseline that issues static repair scripts upon fault detection, the reinforcement learning agent proved markedly superior. The proposed model attained an average fault recovery time of 2.6 timesteps, corresponding to exactly 26 seconds with a standard deviation of 1.1 timesteps, compared to 4.8 timesteps or 48 seconds for the rule-based reactive controller. Concurrently, the OpenFlow error rate dropped from 2.31 errors per timestep under the no-healing baseline to 0.38 errors per timestep under the autonomous agent, an 83.5 percent improvement.
A granular policy output analysis stratified across evaluation timesteps confirmed that the reinforcement learning model developed true state-conditional recovery behaviors rather than relying on trivially biased action distributions. In completely healthy network states without active faults, the agent selected the no-operation action during 91.2 percent of timesteps, demonstrating robust restraint against unwarranted interventions. When confronted with physical link faults, the agent directed its interventions appropriately, selecting port resets in 47.3 percent of cases and switch restarts in 31.6 percent of instances. Under isolated port failures, the model favored targeted port resets in 62.4 percent of timesteps. Flow table errors prompted bespoke recovery routines in 47.8 percent of cases, whereas Faucet controller reload events triggered switch restarts 61.8 percent of the time to re-establish datapath registrations. For transient soft and temporary warnings, the agent appropriately maintained no-operation states in 73.5 percent and 84.1 percent of timesteps, executing appropriate actions across more than 85 percent of all faulted states.
To examine framework scalability, the researchers extended testing to larger topologies using transfer learning, fine-tuning the base policy for 5,000 steps on an 8-switch linear network with 16 hosts and a 16-switch fat-tree architecture comprising 4 core, 4 aggregation, and 8 edge switches with 64 hosts. In both expanded networks, mean episodic rewards and episode lengths remained within twelve percent of base values, while recovery times scaled modestly but remained below fifty seconds, showing effective knowledge transfer across topologies. Nevertheless, the authors identified critical operational limitations inherent in software emulation. Because Mininet cannot replicate the precise line-rate forwarding throughput, queuing delays, and hardware timing characteristics of physical switching silicon, recorded recovery durations reflect algorithmic decision speeds rather than physical hardware restoration. The authors noted that achieving carrier-grade industry restoration benchmarks of sub-fifty milliseconds will require testing physical OpenFlow devices such as Zodiac FX or Pica8 switches alongside distributed multi-controller deployments to ensure data-plane and control-plane resilience under enterprise-scale stress.
Subject of Research: Autonomous reinforcement learning framework for multi-fault self-healing in software-defined networks
Article Title: A reinforcement learning framework for self-healing fault recovery in intent-based SDNs
Article References: Makiyah, E. H., Rasool, M. N., & Rawdhan, F. A. (2026). A reinforcement learning framework for self-healing fault recovery in intent-based SDNs. Discover Networks, 2(1), Article 14. https://doi.org/10.1007/s44354-026-00028-z
Image Credits: AI Generated
DOI: 10.1007/s44354-026-00028-z
Keywords: software-defined networking, reinforcement learning, proximal policy optimization, fault recovery, self-healing networks, Faucet controller, Prometheus monitoring, intent-based networking, reinforcement, learning, framework, self-healing
News Source: Denise Maddox. (November 11, 2026). Reinforcement Learning Automates Multi-Fault Self-Healing Across Software-Defined Networks. Scienmag.
