Method and algorithm for calculating the minimum and maximum connectivity probability machine learning

DOI: 10.31673/2412-9070.2025.025563

Abstract

The increasing use of information systems in various fields often necessitates studying their functional stability. For this purpose, various functional stability indicators have been developed to provide quantitative assessments. Despite the clarity and convenience of some of these indicators, most have a significant drawback: the large number of computations required for their use. Consequently, there is a need to develop methods for estimating these indicators that enable sufficiently fast and accurate evaluation. This paper attempts to construct such an estimation method for connectivity probability using machine learning models.
When working with connectivity probability, it is possible to determine its minimum and maximum values for a given information system. Since their direct computation is highly complex, especially for large-scale systems, an alternative approach is to develop estimation techniques. This study attempts to separately estimate the maximum and minimum connectivity probabilities using two machine learning models: polynomial regression and feedforward neural networks. To train each model, a set of randomly selected graphs representing the theoretical topology of the information system was used. For each graph, the minimum and maximum connectivity probabilities, as well as several numerical characteristics, were computed. Using this dataset, the models were trained and tested under different hyperparameter settings.
The experiments showed that using polynomial regression to estimate the minimum and maximum connectivity probabilities based on numerical characteristics is impractical, as achieving adequate accuracy with this approach requires an excessive number of computations. On the other hand, feed-forward neural networks with three hidden layers have the potential to provide a sufficiently accurate estimation of both the maximum and minimum connectivity probabilities. At the same time, the required computational cost remains acceptable.
The main challenge of this approach is the large number of epochs required for training the neural networks, which leads to a significant training time.
The obtained results indicate that with further research, the application of neural networks for estimating the connectivity probability of an information system can achieve sufficient accuracy while maintaining an acceptable computational load. These findings can be used for developing new information system design methods and for analyzing existing systems.

Keywords: information systems, functional stability, machine learning, regression, neural networks, connectivity probability.