Research on an Intelligent Single-Path Routing Model Based on OSPF Metrics

Abstract

The work presents and thoroughly investigates an intelligent single-path routing model based on OSPF protocol metrics and deep learning. The pro-posed intelligent model is based on a mathematical formalization of the problem in the form of Boolean programming, ensuring the implementation of strictly single-path routing without flow branching. The model is based on a multilayer perceptron (MLP) architecture. The integration of neural net-work predictive capabilities directly into the path-selection process enables dynamic optimization of interface costs. An experimental study was conducted on a topology with five nodes and six communication links. The work provides a comparative analysis of MLP regression and classification models across bandwidth ranges: 10 Mbit/s – 400 Gbit/s, 1 – 400 Gbit/s, and 1 – 100 Gbit/s. It was found that MLP forecasting accuracy reaches 99–100% when using the “optimal” data set, but decreases significantly with excessive variability in input parameters. An important aspect of the study is the comparison of software implementation environments. It has been found that using Python (TensorFlow and PyTorch) provides 5–10% higher prediction accuracy than MATLAB, which is explained by the specialization of Python libraries for network analytics tasks. The conclusions of the work justify the “accuracy limit” of MLPs due to the neglect of the network graph`s topological structure and point to the promising transition to graph neural networks (GNNs) for large systems. At the same time, it is emphasized that, due to its low computational complexity, MLP remains the optimal choice for local solutions and routers with limited resources, where microsecond-level decision-making speed is critical.