The introduction of digital technologies into the activities of companies is based on software and hardware systems, which must function reliably and without interruption. The forecasting of the completion of storage area networks (SAN) is an essential tool for ensuring the smooth operation of such systems. The aim of this study is to develop a system of the modelling and simulation of the further loading of SAN on previously observed load measurements. The system is based on machine learning applied to the load prediction problem. Its novelty relates to the method used for forming input attributes to solve the machine learning problem. The proposed method is based on the aggregation of data on observed loading measurements and the formalization of the problem in the form of a regression analysis problem. The artificial dataset, synthesized stochastically according to the given parameter intervals and simulating SAN behavior, allowed for more extensive experimentation. The most effective algorithm is CatBoost (gradient boosting on decision trees), which surpasses other regression analysis algorithms in terms of R2 scores and MAE. The selection of the most significant features allows for the simplification of the prediction model with virtually no loss of accuracy, thereby reducing the number of confessions used. The experiments show that the proposed prediction model is adequate to the situation under consideration and allows for the prediction of the SAN load for the planning period under review with an R2 value greater than 0.9. The model has been validated on a series of real data on SAN.

No abstract

The quality of operation of neural networks in solving application problems is determined by the success of the stage of their training. The task of learning neural networks is a complex optimization task. Traditional learning algorithms have a number of disadvantages, such as «sticking» in local minimums and a low convergence rate. Modern approaches are based on solving the problems of adjusting the weights of neural networks using metaheuristic algorithms. Therefore, the problem of selecting the optimal set of values of algorithm parameters is important for solving application problems with symmetry properties. This paper studies the application of a new metaheuristic optimization algorithm for weights adjustment—the algorithm of the spiders-cycle, developed by the authors of this article. The approbation of the proposed approach is carried out to adjust the weights of recurrent neural networks used to solve the time series forecasting problem on the example of three different datasets. The results are compared with the results of neural networks trained by the algorithm of the reverse propagation of the error, as well as three other metaheuristic algorithms: particle swarm optimization, bats, and differential evolution. As performance criteria for the comparison of algorithms of global optimization, in this work, descriptive statistics for metrics of the estimation of quality of predictive models, as well as the number of calculations of the target function, are used. The values of the MSE and MAE metrics on the studied datasets were obtained by adjusting the weights of the neural networks using the cycling spider algorithm at 1.32, 25.48, 8.34 and 0.38, 2.18, 1.36, respectively. Compared to the inverse error propagation algorithm, the cycling spider algorithm reduced the value of the error metrics. According to the results of the study, it is concluded that the developed algorithm showed high results and, in the assessment of performance, was not inferior to the existing algorithm.

Logical analysis of data (LAD), an approach to data analysis based on Boolean functions, combinatorics, and optimization, can be considered one of the methods of interpretable machine learning. A feature of LAD is that, among many patterns, different types of patterns can be identified, for example, prime, strong, spanned, and maximum. This paper proposes a decision-support approach to recognition by sharing different types of patterns to improve the quality of recognition in terms of accuracy, interpretability, and validity. An algorithm was developed to search for pairs of strong patterns (prime and spanned) with the same coverage as the training sample, having the smallest (for the prime pattern) and the largest (for the spanned pattern) number of conditions. The proposed approach leads to a decrease in the number of unrecognized observations (compared with the use of spanned patterns only) by 1.5–2 times (experimental results), to some reduction in recognition errors (compared with the use of prime patterns only) of approximately 1% (depending on the dataset) and makes it possible to assess in more detail the level of confidence of the recognition result due to a refined decision-making scheme that uses the information about the number and type of patterns covering the observation.