DNFS-VNE: Deep Neuro-Fuzzy System-Driven Virtual Network Embedding Algorithm

Weiping Ding (ae of Tnnls, Tfs, Tits, Jas, Tetci, Tiv, Tai, Inffus, Ins, Asoc, Neucom, EAAI,et al), Ailing Xiao, Ning Chen, Sheng Wu, Shigen Shen, Weiping Ding, Peiying ZhangarXiv 2023 · DOI · arXiv

Abstract

Virtual Network Embedding (VNE) is a critical challenge in network virtualization, requiring efficient mapping of virtual resources onto physical infrastructure. While Deep Reinforcement Learning (DRL) offers powerful optimization, its "black-box" nature limits its applicability in industrial settings where interpretability is paramount. This paper introduces DNFS-VNE, a Deep Neuro-Fuzzy System that combines the learning capabilities of Convolutional Neural Networks (CNNs) with the linguistic interpretability of fuzzy logic. By utilizing CNNs as fuzzy implication operators, the system achieves superior resource allocation efficiency while providing human-readable explanations for its embedding decisions.

Introduction

As the industry transitions toward 5G/6G and the Industrial Internet of Things (IIoT), network virtualization has become essential. It enables multiple independent virtual networks to operate on a shared physical "substrate" network. The core challenge, known as Virtual Network Embedding (VNE), involves mapping virtual nodes and links onto physical hardware to maximize provider revenue while minimizing operational costs.

While standard Deep Neural Networks (DNNs) excel at handling complex, dynamic environments, they are notoriously opaque. In critical infrastructure, engineers require transparency to trust AI-driven decisions. DNFS-VNE addresses this "trust gap" by creating an interpretable AI framework that explains why specific nodes are selected for embedding.

Methodology

The DNFS-VNE architecture functions as the policy network for a DRL agent, organized into five distinct blocks:

  1. Input Block: Captures substrate network features, including available CPU, bandwidth, and average distance.
  2. Fuzzification Block: Transforms raw numerical data into fuzzy sets (e.g., "Low," "Medium," or "High" CPU) using Gaussian membership functions.
  3. Fuzzy Rule Block: This core component uses CNNs as fuzzy implication operators to learn successful embedding patterns. A parallel interpretation branch translates these patterns into Mamdani-type linguistic rules, such as: "If CPU is Very High and Bandwidth is Very High, then Embedding Probability is Very High."
  4. Defuzzification Block: Converts fuzzy linguistic outputs back into concrete numerical values using the "center of gravity" method.
  5. Output Block: Employs a Softmax function to generate final embedding probabilities for candidate substrate nodes.
The five-block architecture of DNFS-VNE, integrating CNNs with fuzzy logic.

Experimental Setup

The system was evaluated using the GT-ITM tool to simulate a realistic network environment:

  • Substrate Network: 100 nodes and 600 links across 4 domains, with resources randomly assigned between 50 and 100 units.
  • Virtual Network Requests (VNRs): 2,000 requests (1,000 for training, 1,000 for testing), each containing 2–10 virtual nodes.
  • Traffic Model: VNRs arrived following a Poisson distribution to simulate continuous real-world traffic patterns.

Key Insights and Analysis

The primary innovation of DNFS-VNE lies in using CNNs as logic operators. Unlike traditional fuzzy logic, which relies on fixed mathematical operators (like Min/Max), this system learns the optimal logic for a specific environment.

Comparison of Neural Networks and Fuzzy Systems, highlighting their complementary strengths.

The researchers found that DNNs and Fuzzy Systems are naturally complementary: DNNs excel at knowledge acquisition from raw data, while Fuzzy Systems are superior at expressing and applying that knowledge in a way humans can understand. By merging them, DNFS-VNE mitigates the "black-box" problem of pure neural networks and the "slow reasoning" problem of pure fuzzy systems.

Results

DNFS-VNE was benchmarked against NodeRank (heuristic), NRM-VNE (MIP-based), and CDRL (standard RL). Key findings include:

  • Superior Revenue: DNFS-VNE consistently achieved higher long-term average revenue than all baselines.
  • Higher Acceptance Rate: The system successfully embedded more virtual networks without exhausting physical resources compared to heuristic and standard RL methods.
  • Rapid Convergence: The model reached stability within 30-40 iterations.
  • Transparency: DNFS-VNE successfully generated a "Fuzzy Rule Base," providing plain-English explanations for its node selection logic, a feat impossible for the CDRL baseline.

Conclusion

DNFS-VNE represents a significant step toward explainable AI in network management. While currently focused on CPU, bandwidth, and distance, future work will explore more complex attributes like link delay and node degree. The authors aim to scale this interpretable framework to massive, global-scale infrastructure to meet the demands of future telecommunications networks.

Summary updated

BibTeX
@misc{al2023dnfs,
  author = {Weiping Ding (AE of TNNLS, TFS, TITS, JAS, TETCI, TIV, TAI, INFFUS, INS, ASOC, NEUCOM, EAAI,et al) and Ailing Xiao and Ning Chen and Sheng Wu and Shigen Shen and Weiping Ding and Peiying Zhang},
  title = {{DNFS-VNE: Deep Neuro-Fuzzy System-Driven Virtual Network Embedding Algorithm}},
  year = {2023},
  eprint = {2310.09078},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2310.09078},
}
Figure