Emerging Technologies and Intelligent Systems: Theory, Applications and Innovations
Self-Driving and Self-Healing Networks Using AI: Where Autonomy Is Claimed, Where It Is Evaluated, and What Runs in Production
Authors: Raghuram Malpe Pai, Krishna Suman Dara, Somesh Rahul, Amarnath G S, Shivaprasad Gali and Mohammed Kamran Syed
Publishing Date: 15-11-2026
ISBN: 978-81-975670-4-9
Abstract
A large literature now reports network systems that detect faults, localise them, propose fixes and act. What it does not report in one place is the relationship between where a system was evaluated and how much of the repair loop that evaluation closed. We screen 205 works and audit 101 against two axes applied jointly: the substrate a system was demonstrated on, from analytical models to a live production network, and the deepest stage of the fault lifecycle its evaluation reaches, from detection to execution with verification. Crossing them gives a result neither shows alone. Of the 22 works evaluated on production networks, every one stops at detection, diagnosis or advisory recommendation and none executes a remediation, while all 26 works that evaluate unsupervised execution or execution with verification run on analytical models, simulators, emulators, testbeds or replayed traces. The boundary is categorical rather than graded, and the stage between, execution behind a human approval gate, holds five works. This is a claim about published evidence and not about what operators run unpublished, and the corpus shows why the distinction bites: the only mechanism in it that repairs a production network is a rule-based auto-isolator described in three sentences of background in a paper about something else. Reading 20 of the load-bearing works in full shows part of it obscured by reporting practice: seven describe a deeper stage than they evaluate, and 11 describe a production setting while 5 evaluate on one. Gray degradation is addressed by 11 works and acted on by none, and no work in the corpus evaluates an automated defence against a route hijack. Reported success rates are not comparable, because benchmarks disagree on whether a repair that breaks something else counts as success; where that is scored, reported success falls from 57.8% to 25.5%. We close with a six-field reporting checklist that would make the substrate and loop closure of a self-healing claim legible from its abstract.
Keywords
Self-healing networks, network fault management, autonomous networks, LLM agents, evaluation methodology, systematic review.
Cite as
Raghuram Malpe Pai, Krishna Suman Dara, Somesh Rahul, Amarnath G S, Shivaprasad Gali and Mohammed Kamran Syed, "Self-Driving and Self-Healing Networks Using AI: Where Autonomy Is Claimed, Where It Is Evaluated, and What Runs in Production", In: Harish Sharma and Mukesh Saraswat (eds), Emerging Technologies and Intelligent Systems: Theory, Applications and Innovations, SCRS, India, 2026, pp. 25-50. https://doi.org/10.56155/978-81-975670-4-9-2