[Paper] Simultaneous Genetic Evolution of Neural Networks for Optimal SFC Embedding

Source: arXiv - 2512.09318v1

Overview

The paper introduces GENESIS, a novel genetic‑algorithm framework that simultaneously tackles the three intertwined sub‑problems of Service Function Chain (SFC) embedding: chain composition, virtual network function (VNF) placement, and link routing. By evolving three sine‑function‑activated neural networks whose outputs feed a Gaussian‑based selector and an A* pathfinder, GENESIS achieves optimal embeddings across a wide range of data‑center scenarios—outperforming existing GA and greedy baselines both in solution quality and runtime.

Key Contributions

• Unified optimisation of all three SFC sub‑problems (composition, VNF placement, link routing) within a single evolutionary loop.
• Neuro‑evolution design: three sine‑activation neural networks are co‑evolved, enabling rich, non‑linear decision surfaces for complex placement decisions.
• Hybrid decision pipeline: network outputs are transformed into a Gaussian distribution for probabilistic selection, then refined by an A* algorithm for shortest‑path link embedding.
• Extensive empirical evaluation on 48 realistic data‑center topologies, showing 100 % optimality and a 2× speed‑up over the strongest competing GA.
• Open‑source implementation (emulator‑based) that can be plugged into existing NFV orchestration stacks for reproducible research.

Methodology

1. Encoding – Each individual in the GA population encodes three neural networks (one per sub‑problem). The networks take topology, resource, and traffic descriptors as inputs and output scores for candidate VNFs or links.
2. Sine activation – Instead of conventional ReLU or sigmoid units, the authors use sine‑based activation functions, which provide periodicity and smoother gradients that help explore the combinatorial search space.
3. Fitness evaluation – The raw network scores are first mapped onto a Gaussian probability distribution; higher‑probability candidates are fed to an A* shortest‑path routine that finalises the link embedding. The overall fitness combines three metrics: (i) chain feasibility, (ii) resource utilisation, and (iii) end‑to‑end latency.
4. Genetic operators – Standard crossover and mutation are applied at the weight‑matrix level of the three networks, preserving co‑adapted solutions across sub‑problems.
5. Termination – The algorithm stops after a fixed number of generations or when the fitness plateaus, returning the best individual as the SFC embedding plan.

Results & Findings

• Solution quality: GENESIS found a feasible, cost‑optimal embedding in every test case, whereas the best competing GA succeeded in only 71 % of cases.
• Speed: Despite the added neural‑network evaluation, GENESIS was roughly twice as fast as the second‑best GA, thanks to the Gaussian‑guided selection that reduces the search space early on.
• Scalability: Performance held steady across varying numbers of VNFs (10–50) and link densities, indicating robustness to realistic data‑center growth.

Practical Implications

• NFV Orchestrators can plug GENESIS into their placement engines to obtain higher‑quality SFC deployments without incurring prohibitive compute costs.
• Reduced operational expenditure: Optimal VNF placement translates directly into lower bandwidth and compute waste, which is critical for large‑scale cloud providers and telcos.
• Faster provisioning: The sub‑15‑minute runtime makes GENESIS suitable for near‑real‑time scaling decisions (e.g., auto‑scaling of firewalls or load balancers during traffic spikes).
• Extensibility: Because the approach is model‑agnostic, developers can replace the sine‑activated networks with other differentiable modules (e.g., graph neural networks) to incorporate additional constraints like energy consumption or security policies.

Limitations & Future Work

• Emulation‑only validation: Results are based on a simulated data‑center environment; real‑world deployment may expose latency overheads not captured in the emulator.
• Fixed‑size networks: The current design uses a static three‑network architecture; scaling to extremely large topologies might require hierarchical or modular network structures.
• Parameter sensitivity: The Gaussian‑selection bandwidth and mutation rates were tuned empirically; automated hyper‑parameter search could further improve robustness.
• Future directions proposed by the authors include (i) integrating reinforcement‑learning fine‑tuning after GA convergence, (ii) extending the framework to multi‑domain (cross‑data‑center) SFC embedding, and (iii) open‑sourcing a production‑grade plugin for popular NFV MANO platforms (e.g., OpenStack‑Tacker, ONAP).

Authors

• Theviyanthan Krishnamohan
• Lauritz Thamsen
• Paul Harvey

Paper Information

• arXiv ID: 2512.09318v1
• Categories: cs.NE, cs.AI
• Published: December 10, 2025
• PDF: Download PDF