PaperlayerView exact revision on arXiv

Source-linked AI summary

Adaptive E2E Monitoring Path Selection with Deep Reinforcement Learning in Multi-domain Optical Networks

Soham Choudhury, Martin Bojinov, Jason P. Jue, Genya Ishigaki

arXiv:2610.10753v1cs.NI

TL;DR

When optical domains conceal internal topology and metrics, a slice coordinator must select E2E paths to localize link failures under limited visibility. The paper formulates initial and progressive path-selection problems and applies DRL to both. The reported approach outperforms baselines in both problems and achieves near-optimal localization in initial selection, while generalization across topologies and path configurations remains limited.

  • Problem

    A slice coordinator must select E2E monitoring paths to localize failures even though domains withhold internal topology and detailed performance metrics.

    Source: §1 · p. 1

  • Method

    The paper uses DRL to select paths for initial failure coverage and progressive targeting of suspected failure regions.

    Source: §1 · p. 2, §7 · p. 20

  • Results

    The approach outperforms baseline algorithms in both path-selection problems and achieves near-optimal localization for initial selection.

    Source: §1 · p. 2, §7 · p. 20

  • Takeaways & Limitations

    The results suggest the feasibility of AI-based autonomous failure detection and localization in future network architectures.

    Source: §1 · p. 2

    Generalization across network topologies and path configurations remains limited, motivating work on model transferability.

    Source: §6.3 · p. 20

Abstract

from arXiv · show

Network slicing over multi-domain optical networks enables resource isolation for high-bandwidth, virtually-dedicated networks. However, guaranteeing end-to-end (E2E) performance remains challenging when domains withhold internal topology and performance metrics from external entities. Under this limited visibility, a control function (slice coordinator) needs to select E2E monitoring paths to detect and localize link failures. This paper formulates two complementary optimization problems, the initial and progressive monitoring path selection problems, which together realize closed-loop autonomous monitoring at the slice coordinator level. The initial problem maximizes failure coverage without prior information, while the progressive problem concentrates additional paths around suspected failure locations identified from prior monitoring. We propose a deep reinforcement learning (DRL) algorithm that adaptively selects an optimal set of monitoring paths for each phase. Experiments using the GNPy optical network simulator demonstrate that our approach outperforms baseline algorithms for both problems and achieves near-optimal localization in the initial selection, suggesting the feasibility of AI-driven closed-loop failure localization in future multi-domain optical network architectures.

1. Introduction

Multi-domain optical slicing makes E2E performance monitoring difficult because domain operators conceal internal topology and detailed metrics. The paper frames initial and progressive path selection as complementary tasks and uses DRL to select paths for both.

1. Introduction

  • Domains may withhold routing logic, topology, and detailed performance metrics, leaving the coordinator with limited visibility into resources.The paper assumes domains expose abstract connectivity between border nodes while keeping intra-domain information hidden.Source: §1 · pp. 1–2
  • The paper optimizes E2E monitoring-path selection for localizing single-link failures under a constraint on the number of paths.The coordinator selects paths from a high-level perspective rather than relying on fully known probing structures.Source: §1 · p. 2
  • Initial selection seeks broad link coverage without prior information, whereas progressive selection adds paths around suspected failures identified by initial monitoring.The two objectives are complementary parts of closed-loop autonomous monitoring by the slice coordinator.Source: §1 · p. 2
  • The approach outperforms baseline algorithms in both selection problems and achieves near-optimal localization for initial selection.The evaluation uses an optical network simulator and implements the slice coordinator as a software module.Source: §1 · p. 2
  • This paper extends earlier DRL work on initial path selection with a progressive-selection problem and solution.The extension is intended to enhance closed-loop autonomy of the monitoring function.Source: §1 · p. 3

2. Related Works

Related work spans multi-domain control, tomography, and data-driven fault management. The paper distinguishes its approach by selecting feasible E2E monitoring paths under limited visibility and a piggyback constraint.

2.1. Multi-Domain Architectures and Monitoring Frameworks

  • Multi-domain telemetry frameworks commonly use intra-domain or per-hop data, whereas this system consumes E2E path observations from an abstracted broker topology.The surrounding literature also discusses hierarchical control and abstraction across administrative domains.Source: §2.1 · p. 3

2.2. End-to-End Failure Localization and Network Tomography

  • Earlier tomography and probing studies examine failure inference and path selection, while this paper uses RL to optimize monitoring paths under its own piggyback constraint.The authors benchmark against a problem-specific optimum because existing selection criteria do not transfer directly.Source: §2.2 · pp. 3–4
  • Existing probe-selection methods often assume freely formed paths over a fully known topology, while this work restricts monitoring to pre-established lightpaths.The feasible measurement set is therefore a fixed, irregular subset of the paths available in the topology.Source: §2.2 · p. 3
  • Prior RL tomography methods target bandwidth estimation or route selection, rather than binary failure localization through discriminative monitoring-path selection.The paper identifies its task as distinguishing the failed component, not estimating a continuous transmission-quality metric.Source: §2.2 · p. 4

2.3. Machine Learning and Data-Driven Fault Management

  • Optical-network ML has focused on impairment estimation, fault classification, and routing, while this work applies RL to selecting monitoring paths.GNPy simulation is used to model physical-layer signal degradation and inject soft failures under controlled conditions.Source: §2.3 · p. 4, §2.4 · p. 4

3. The E2E Monitoring Path Selection Problem

The section models E2E failure localization over an abstract broker graph and defines two path-selection objectives. Initial selection seeks broad coverage; progressive selection uses initial observations to narrow suspected failure locations.

3. The E2E Monitoring Path Selection Problem

  • The problem distinguishes broad coverage without fault-location assumptions from targeting suspicious regions using initial monitoring results.The two objectives correspond to initial and progressive path selection.Source: §3 · p. 5

3.1. System Model

  • The coordinator sees a broker graph that aggregates domains and exposes border-node connectivity, while internal physical topology remains hidden.A star abstraction represents each domain's border-node connectivity; its edges correspond to physical paths within the domain.Source: §3.1 · p. 5, Figure 1 · p. 6

3.2. Localization Function

  • A monitored path signals degradation when its performance metric falls below threshold t; the localization function identifies candidate edges consistent with those signals.Candidate edges are found by intersecting signaling paths and removing edges appearing in nonsignaling paths; a single-link failure is localized when one candidate remains.Source: §3.2 · pp. 5–6

3.3. Initial Monitoring Path Selection

  • Initial selection evaluates path sets across hypothetical single-link failures using the Localization Narrowing Index (LNI), based on remaining candidate edges.The selection problem minimizes LNI, and brute-force evaluation over all path subsets becomes intractable as the broker graph grows.Source: §3.3 · pp. 6–7
  • A constant budget bounds selected monitoring paths; the paper uses unit cost per path to limit the selection to K paths.The cost constraint can be generalized by assigning each path its actual deployment cost.Source: §3.3 · p. 7

3.4. Progressive Monitoring Path Selection

  • Progressive selection starts from initial path telemetry and adds one path at each step to reduce the number of potentially failing edges.When the initial budget is tight, many candidate edges may remain, so later paths target suspected regions rather than broad coverage.Source: §3.4 · pp. 7–8
  • If no initial path signals degradation, the initial signaling set defaults to all initial paths to avoid eliminating every candidate.The progressive process is defined relative to the initial path set and a fixed number of additional paths.Source: §3.4 · pp. 7–8

4. Reinforcement Learning-based E2E Monitoring Path Selection

The slice coordinator learns monitoring-path selections from a limited broker-graph view. DQN supports two distinct phases: choosing an initial set and progressively adding paths as monitoring results arrive.

4. Reinforcement Learning-based E2E Monitoring Path Selection

  • The learning module is part of the E2E slice coordinator, whose network view is limited to a broker graph.The paper motivates learning by the computational complexity of path selection and changing failure dynamics.Source: §4 · p. 8

4.1. Reinforcement Learning Agent

  • The DQN estimates the expected return of selecting a path in a network state, using a target network to define its temporal-difference update.The Q-function is approximated by a one-hidden-layer, 256-neuron ReLU network; the target value is frozen every 500 episodes.Source: §4.1 · p. 8

4.2. State and Action Spaces

  • Initial selection is reformulated as sequential choices: the agent selects one candidate lightpath per round until the cost constraint is reached, then removes it from the candidate set.Candidate paths are represented as bounded-length sequences of node IDs.Source: §4.2.1 · p. 9
  • Progressive-state observations combine selected paths, their OSNR measurements, and path lengths; actions add one existing lightpath to the monitoring set.Fixed-size path representations use padding, with lengths indicating padding positions.Source: §4.2.2 · p. 9

4.3. Reward Function

  • Initial selection gives zero reward until the cost constraint is reached, then scores the selected path set using its localization number index (LNI).The reward uses s = 3 in the experiments to emphasize differences between better- and poorer-performing sets.Source: §4.3.1 · p. 9
  • Progressive selection rewards shrinking the suspected-edge set (+0.15) but penalizes eliminating it entirely (-0.5), with rewards summed across thresholds.This encourages iterative refinement while avoiding inconsistent selections.Source: §4.3.2 · p. 10

5. Evaluation Experiment

The evaluation uses GNPy to simulate optical-network behavior and a broker-graph view for the slice coordinator. It tests failure scenarios and compares RL path selection with heuristic and theoretical baselines.

5. Evaluation Experiment

  • GNPy simulates the physical topology, while a separate simulator removes intra-domain topology and metrics to construct the broker graph seen by the agent.The agent receives network state and reward from E2E measurements and sends selected paths to the network simulator.Source: §5 · pp. 10–11

5.1. GNPy Setup

  • The evaluation uses three multi-domain topologies of differing sizes and complexity, with 50-km fiber segments and randomly selected ROADM endpoints for 50-Gbit/s requests.Fiber loss is 0.2 dB/km; channel slot width starts at 12.5 GHz and increases in 12.5-GHz increments.Source: §5.1 · p. 11

5.2. GNPy Fiber Failure Simulation

  • For initial-selection tests, one fiber is degraded to 1.0 dB/km to model a hard failure, with the failure threshold set to 5.0.The injected noise is restricted to a single fiber to match the paper’s single-link-failure scope.Source: §5.2.1 · p. 11
  • Table 2 shows Path 4’s E2E OSNR worsening as loss on the affected fiber increases, while the other three paths’ OSNR remains unchanged.The other paths do not traverse the affected fiber.Source: Table 2 · p. 12, §5.2.1 · p. 12
  • Progressive-selection tests model soft failures by raising one fiber’s loss from 0.2 to 0.8 dB/km and evaluate thresholds of 16.5, 18.0, and 20.0.The injected loss persists for an entire episode.Source: §5.2.2 · p. 12

5.3. Baseline Algorithms

  • The comparisons include random, longest-path, and least-monitored heuristics, plus theoretical optima; the largest initial case’s optimum was not found after 72 hours.The progressive theoretical minimum suspect-set size is cheaper to compute than finding paths that achieve it, which remains combinatorial and difficult.Source: §5.3 · pp. 12–13

6. Experiment Results

The RL selector leads the initial localization comparisons and improves progressive localization in non-trivial cases, while topology and monitoring thresholds shape the results. The discussion also identifies limited generalization across topologies and path configurations.

6.1. Initial Monitoring Path Selection Performance

  • On Topology 1's 8-edge broker graph, RL and least-monitored choose shorter paths with lower LNI, whereas the longest-path heuristic performs poorly.Longer paths are more likely to share edges, and the star abstraction offers fewer unique paths.Source: §6.1 · p. 13, Figure 3 · p. 14
  • Across all initial-selection settings, RL outperforms the heuristics in Localization Narrowing Index (LNI).The comparison covers Topologies 1–3.Source: §6.1 · p. 13, Figure 3 · p. 14, Figure 4 · p. 14
  • RL ranks first in Topologies 2 and 3; longest-path typically follows in Topology 2, while Topology 3 ranks least-monitored, longest-path, then random.The broker graphs have 18 edges in Topology 2 and 45 in Topology 3.Source: §6.1 · p. 13, Figure 4 · p. 14, Figure 5 · p. 15
  • RL achieves the lowest optimality gap across all reported statistics for Topology 3 with 8 candidate paths and 20 requests: mean 0.2403 versus 0.5060 for least-monitored, 1.1694 for longest, and 1.3916 for random.RL also has a median gap of 0.2336, standard deviation of 0.1862, and maximum gap of 0.6879.Source: §6.1 · p. 14, Table 3 · p. 15

6.2. Progressive Monitoring Path Selection Performance

  • Progressive additions yield no further improvement in trivial cases already localized perfectly or near-perfectly, but RL consistently performs best in non-trivial cases.The evaluation pairs suspect-set size with hit rate; experiments add four paths to a shared initial set.Source: §6.2 · pp. 15–16
  • At t=16.5, RL reached one suspect on average for one failure scenario while all methods retained the true fiber; at t=20.0, it reached near-minimal one-suspect sets while baselines barely reduced candidates.The thresholds reflect operator policy: lower thresholds are more conservative, while higher thresholds may detect weaker failure signals.Source: Figure 6 · p. 16, §6.2 · pp. 17–19, Figure 7 · p. 17
  • With eight paths, RL reached Topology 3's theoretical optima: two candidate edges and 0.5 efficiency, without losing the true failing fiber.Topology 3's larger graph leaves more suspected edges initially than Topology 2.Source: Figure 8 · p. 18, §6.2 · p. 18, Figure 9 · p. 19

6.3. Discussions

  • Because path intersections make localization set-theoretic and non-additive, adding paths can worsen it; RL evaluates long-term selections using past failure experience.The paper reports that generalization across network topologies and path configurations remains limited.Source: §6.3 · pp. 19–20

7. Conclusion

The paper formulates initial and progressive monitoring-path selection for single-link failure localization in E2E optical slices under limited domain visibility. Across tested topologies, its RL approach outperforms baseline heuristics and achieves near-optimal localization in initial selection.

7. Conclusion

  • Across various network topologies, the RL approach consistently outperforms baseline heuristics and achieves near-optimal localization in initial monitoring path selection.Source: §7 · p. 20
  • The method targets single-link failures through initial monitoring coverage and progressive path selection around a detected failure's suspected topological area.Source: §7 · p. 20
  • Future work would extend the setting to partially visible disaggregated networks and improve cross-topology learning with graph neural network layers.Source: §7 · p. 20
  • Threshold-based signaling paths could be relaxed using a recurrent neural network-based RL agent that learns from raw OSNR sequences to infer failure alarms.Source: §7 · p. 20
  • The single-link failure model is proposed for extension to multiple and correlated failures, whose concurrent-failure suspect sets pose a combinatorial-growth challenge.Source: §7 · pp. 20–21
Loading 2610.10753v1…