Работая с сайтом, я даю свое согласие на использование файлов cookie. Это необходимо для нормального функционирования сайта, показа целевой рекламы и анализа трафика. Статистика использования сайта обрабатывается системой Яндекс.Метрика

Принять

Научный журнал Моделирование, оптимизация и информационные технологииThe scientific journal Modeling, Optimization and Information Technology

Online media
issn 2310-6018

submit
a manuscript

metadata of articles for the last 2 years

Modeling and optimization of data collection process for artificial intelligence in medicine

2026. T.14. № 4. id 2232

Ivaschenko A.V. Terekhin M.A. Poretskova G.Y. Zhdanovich G.E. Melnikov D.A. Radaev D.E.

DOI: 10.26102/2310-6018/2026.55.4.020

Development of Artificial Intelligence technologies in medicine requires a systematic approach to collecting and processing structured datasets for training, testing, and validating machine learning models. This paper proposes a solution to this problem through simulation modeling based on queueing theory. This modeling requires estimating the planned throughput of each data collection point, ensuring a sufficient number of patients, the availability and reliability of their medical information, and meeting legal requirements regarding personal data protection and medical ethics. The proposed approach was studied using the analysis of biomedical data collection processes designed to train artificial intelligence models for remote diagnostic methods. The empirical part of the study was conducted at biomedical signal collection points over a six-month period. The total sample size was 574 patients. A simulation model was developed to optimize the data collection process. According to the simulation modeling, the average data collection intensity was 7.28 patients per day with significant variability in the workload. During the optimization process, changes were made to the data collection process through parallelization, which increased productivity by reducing the time spent on questionnaires and temperature measurements and increasing patient throughput. The optimization of the data collection process increased the workload from 4.67 to 12.12 patients per day. The proposed approach allows us to validate the architecture of the organizational and technological process for data collection before scaling and minimizes the risk of exceeding the schedule deadlines for generating medical datasets.

Keywords: medical dataset, simulation modeling, queueing theory, digital twin, throughput, artificial intelligence

Automated decision support system for predicting online shopping behavior of e-commerce users

2026. T.14. № 3. id 2230

Svyatov R.S.

DOI: 10.26102/2310-6018/2026.54.3.010

The relevance of this study is caused by the rapid development of electronic commerce and the growing need for effective tools to predict user behavior in online retail environments. The main problem lies in the fact that existing solutions in this domain are often limited to specific datasets, lack sufficient scalability, and rarely support real-time automation of the forecasting process. The purpose of this study is to develop a decision support system that enables the estimation of the probability of future purchase completion based on the analysis of user behavioral data and provides decision-makers with actionable recommendations for subsequent marketing activities. The methodological framework of the study is based on the use of a web analytics system as a source of information on user activities, data preprocessing and structuring procedures, and the application of gradient boosting as a machine learning algorithm for predicting the probability of purchase. To identify internal and external factors that could have a positive or negative impact on achieving the goal, a SWOT analysis was conducted. Experimental validation of the system was conducted using data from four online stores representing different business domains. The results demonstrate that the overall F-score exceeds 80 % across all experiments. The materials presented in this article have practical relevance for e-commerce professionals, data analysts, and marketing specialists, as well as for decision-makers, since the proposed system enables automated prediction of purchasing behavior, the formation of interpretable user segments, and the application of the obtained results to marketing personalization and optimization of managerial decision-making.

Keywords: machine learning, decision support system, user behavior analysis, e-commerce, consumer behavior prediction, online stores

Architecture of a distributed multimodal medical data analysis system based on variational semantic alignment

2026. T.14. № 5. id 2229

Pozharsky R.V. Ryndin A.A.

DOI: 10.26102/2310-6018/2026.56.5.010

The article presents the architecture of a distributed system for intelligent analysis of multimodal medical data (DICOM images and text reports), combining theoretical methods of variational inference with modern MLOps engineering practices. The key problem addressed is the integration of heterogeneous data (DICOM imaging studies and text clinical reports) under real-world time and computational constraints. The main scientific contribution lies in the formalization and implementation of a new semantic alignment criterion conditioned on unobserved clinically significant latent factors. This criterion, maximized using variational inference (Evidence Lower Bound), ensures deep integration of modalities based on a common pathophysiological basis rather than superficial correlations. On the practical side, a fault-tolerant distributed infrastructure based on Docker, Apache Spark, MinIO, and MLflow has been developed and deployed, providing a complete data lifecycle –from storage and distributed processing to experiment tracking. For adaptive load management, a reinforcement learning-based controller is proposed and implemented, formalizing patient routing between fast (deterministic algorithms) and deep (full ViT+BERT models) pipelines as a partially observable Markov decision process (POMDP). The architectural framework and mathematical model of variational semantic alignment are presented. Experiments on synthetic data confirmed the correctness of the software implementation in the WSL2/Docker environment and the efficient interaction of Spark and MinIO components. The next stage of research will be scaling the system to the full MIMIC-CXR dataset for clinical validation of the proposed hypotheses.

Keywords: multimodal analysis, variational inference, semantic alignment, distributed computing, reinforcement learning, medical data, DICOM, MLOps

A two-level stochastic-adaptive model for managing the operational risks of retail network facilities

2026. T.14. № 5. id 2227

Ustimov M.G. Lvovich I.Y.

DOI: 10.26102/2310-6018/2026.56.5.006

The article presents an innovative two-level stochastic-adaptive operational risk management model designed for large distributed infrastructure networks. The study solves the problem of the inability of traditional deterministic models to adequately assess the "tail" risks in conditions of high uncertainty of energy consumption, equipment failures and logistical failures. The proposed methodology combines strategic planning and tactical online adaptation. At the top level, two-stage stochastic programming is used to generate robust maintenance and capacity redundancy plans that take into account the probabilistic nature of threats. Intelligent clustering of objects using self-organizing Kohonen maps allows you to divide the network into four categories: critical, high-risk, logistically vulnerable and stable. At the lower level, reinforcement learning agents (PPO and DQN algorithms) adapt operational solutions in real time using customized reward functions for each cluster. Experimental results confirm the effectiveness of the approach: for critical facilities, the share of downtime has been reduced to 2 %, and for stable facilities, maximum resource savings have been achieved. The implementation of the model makes it possible to reduce operating costs by 10–15% and increase the reliability of critical infrastructure by 20–30%. The model ensures transparency of management through clear KPIs and contributes to the implementation of a sustainable development strategy.

Keywords: multimodal data analysis, semantic alignment, medical diagnostics, reinforcement learning, distributed computing

Experimental study of cooperative resource allocation in big data processing systems based on Shapley value and machine learning

2026. T.14. № 5. id 2226

Blyakherov M.V. Lvovich I.Y.

DOI: 10.26102/2310-6018/2026.56.5.012

Optimal and equitable allocation of computing resources in dynamic Big Data environments such as Apache Spark remains a challenge. Traditional planners often do not take into account the synergetic effects of cooperation between tasks, which leads to inefficiency and conflicts. The purpose of this work is to experimentally investigate and verify a hybrid approach to cooperative resource allocation based on the formal principles of cooperative game theory and adaptive machine learning capabilities. The paper formalizes a model of a cooperative game, where coalitions of parallel tasks are characterized by a utility function depending on the allocated resources. To ensure stability and fairness, equilibrium conditions (the core of the game) have been introduced, and the distribution is based on the Shapley value, which estimates the marginal contribution of each task. To overcome the analytical complexity of evaluating utility in real conditions, it is proposed to use ML models (gradient boosting, graph neural networks) trained on historical cluster metrics as approximators of the characteristic function of the game. An experimental bench based on Apache Spark with the Prometheus/Grafana monitoring system has been developed and deployed. Experiments have shown that the proposed approach provides a dynamic and balanced allocation of resources (CPU, memory), increases the stability of task coalitions, and improves overall distribution equity (Ginny index) compared to the baseline scenarios. Visualization of key metrics confirmed the achievement of states close to the core of the game. The study demonstrates the practical applicability and effectiveness of combining game-theoretic models and machine learning for intelligent resource management in distributed Big Data systems, paving the way for the creation of self-optimizing and cooperative orchestrators.

Keywords: resource allocation, cooperative game theory, shapley value, machine learning, big data, apache Spark, monitoring

Methodology for managing the emergency response process based on multi-criteria optimization methods

2026. T.14. № 4. id 2224

Barkalov S.A. Bekirova O.N. Vtornikova Y.A.

DOI: 10.26102/2310-6018/2026.55.4.021

In modern conditions, due to the unstable economic and political situation around the world, emergencies of various natures are becoming more frequent and large-scale phenomena. This is caused both by natural factors and man-made reasons, as well as by deliberate actions resulting from conflicts and sabotage, which necessitates the improvement of rapid response methods. Consequently, the relevance of developing automated decision support systems for effectively countering contemporary challenges and threats in the field of emergency consequence management is increasing. This paper describes a methodology for the effective management of a set of works and measures for emergency response, based on multi-criteria optimization methods. The following were chosen as optimization criteria: efficiency or the ability to complete assigned tasks in the shortest possible time, availability or the ability to provide resources for all work being carried out in the required volume, and information content or the implementation of measures to ensure up-to-date and objective information about the current situation. Three models for conducting optimization and obtaining a Pareto-optimal solution are considered: the generalized objective function method, the criterion constraints method, and the method of successive concessions. The article provides the mathematical formulation and description of the models and presents an algorithm for selecting a model for different conditions.

Keywords: emergencies, decision making, threat response, multi-criteria optimization, mathematical modeling

Traffic flow management based on reinforcement learning

2026. T.14. № 4. id 2223

Minakov E.I. Khazov N.I.

DOI: 10.26102/2310-6018/2026.55.4.015

Traffic jams often occur due to inefficient control of traffic lights at intersections, that is, due to the fact that their settings are not sufficiently adapted to specific conditions. Currently, foreign research is actively underway in the field of applying machine learning methods with reinforcement to optimize traffic flows at intersections, which once again shows the urgency of the problem. The prospect of using reinforcement learning lies in the ability to control the dynamics of complex processes without human intervention. To maintain the efficiency and safety of moving cars in urban environments, there are systems that control traffic flows using traffic lights. The paper considers the existing types of traffic flow management systems. The analysis revealed their positive and negative qualities. The article proposes an intelligent control system based on the principles of reinforcement learning, supplemented by an approximator using a neural network. The network architecture is a multi-layered perceptron, with two hidden layers with ReLU activation functions. The process of agent training and the results of control system modeling in the SUMO microscopic modeling environment are presented. The results are presented in the form of a graph of the dynamics of agent training, heat maps of intersections when simulating rush hour traffic and in case of an accident before and after exposure. The proposed system makes it possible to increase the traffic intensity in the intersection network by 40% and 25% during rush hour and traffic accidents, respectively. In addition, the future prospects of its development are reflected.

Keywords: traffic flow, traffic management, reinforcement learning, neural networks, machine learning, adaptive management

Features of evaluating process efficiency in an IT team when applying Agile management methodologies

2026. T.14. № 5. id 2221

Kotova M.R.

DOI: 10.26102/2310-6018/2026.56.5.004

Evaluating the efficiency of processes in IT teams applying agile management methodologies (Agile) is a relevant research problem associated with the need to reduce change delivery lead time while simultaneously ensuring delivery stability (quality) and the economic feasibility of software development. Traditional outcome-based indicators focused on deadlines and the volume of completed work prove to be insufficiently informative in iterative and incremental development contexts, as they fail to reflect flow variability, process losses, and the consequences of quality degradation. This paper proposes a process-oriented model for evaluating IT team efficiency, based on aggregating flow metrics, delivery stability metrics, and economic loss indicators into an integral process efficiency index. The model relies on digital trace data generated throughout the IT product life cycle in issue tracking systems, version control systems, CI/CD pipelines, and monitoring tools. Within the framework of the study, the use of a multiplicative aggregation approach is substantiated, which makes it possible to account for the impact of limiting factors in the development process. The approbation of the model using data from product teams applying Agile approaches confirms that the integral assessment enables early detection of process degradation and the localization of problem areas related to flow management, delivery stability, and the accumulation of technical debt. The results obtained demonstrate the feasibility of using the proposed model as a tool for managerial decision support and continuous monitoring of IT team process efficiency.

Keywords: agile development, lean, devOps, process efficiency, flow metrics, delivery stability, technical debt, integral index, IT team

Hybrid semantic reduction of texts in library information systems

2026. T.14. № 3. id 2220

Rzyankin I.S. Noskov M.V.

DOI: 10.26102/2310-6018/2026.54.3.013

The relevance of the study is determined by the continuous growth of textual information in library information systems and the need to ensure fast and meaningful navigation across electronic collections under constrained computational resources. Existing automatic summarization solutions are primarily oriented toward large-scale language models, which limits their practical deployment within local library infrastructures. In this context, the paper aims to develop a resource-efficient method of semantic text reduction that balances the quality of semantic representation with computational feasibility. The proposed approach is based on a hybrid architecture that sequentially combines lexical reduction using word clouds with neural summarization performed by compact models. In addition, a context-oriented evaluation metric is introduced to assess relevance with regard to semantic coherence, structural characteristics, and domain-specific terms significant for the library environment. An experimental study conducted on a corpus of 1178 documents demonstrates that the hybrid approach improves relevance indicators while simultaneously reducing inference time compared to direct neural summarization of the full text. The obtained results confirm the practical applicability of the proposed method for library information systems operating under limited computational infrastructure and its usefulness for navigation and cataloging tasks.

Keywords: semantic text reduction, automatic summarization, word cloud, library information systems, hybrid text processing methods, neural models, relevance evaluation, library Relevance Score

Application of Markov decision process in developing an information security strategy

2026. T.14. № 4. id 2217

Zubchenko K.V. Sultanov N.Z. Shevtsova G.A.

DOI: 10.26102/2310-6018/2026.55.4.019

The relevance of this study is driven by the growing number and complexity of cyberattacks, in particular the need to continually improve organizations' security levels, as well as the ongoing planning and modeling of security strategies in the face of limited resources. This work aims to develop a model for developing an information security strategy for a given organization, taking into account economic indicators. The primary research methods are modeling, comparative analysis, and synthesis. The paper contains the characteristics of the simulated organizations, the formulas and algorithms used in the prototype, as well as numerical indicators of criteria and parameters. The relationships between the model parameters are presented. As a result, the model's performance on the simulated organizations was demonstrated: optimal strategies were obtained for each of them, correlating with generally accepted approaches to developing strategies in real companies. The resulting graphs of the system states are demonstrated. For all organizations, integrated strategies proved to be the most optimal. In the short term, the use of a Markov decision process allows for the successful optimization of management decisions, regardless of the company's maturity level. Allocating a large budget for information security has a significant impact on efficiency only for companies with a low maturity level. The results of the work are of practical value to information security specialists and managers, providing a tool for developing an optimal information security strategy within a given budget.

Keywords: markov decision process, information security strategy, security strategy modeling, economic costs, strategy optimization

Development of an adaptive resource management system for containerized CAD systems based on reinforcement learning

2026. T.14. № 3. id 2216

Chudinova A.A.

DOI: 10.26102/2310-6018/2026.54.3.016

This project is dedicated to the development of an adaptive resource management system for containerised computer-aided design (CAD) applications using reinforcement learning. Modern CAD workloads are characterised by highly variable computing requirements, which makes traditional threshold-based auto-scaling mechanisms insufficient for maintaining performance and reliability in dynamic conditions. To address this issue, the proposed system compares classic Kubernetes pod scaling based on thresholds (HPA) with a Q-learning-based auto-scaling strategy applied to container clusters. The experimental setup is implemented as a simulation of a distributed containerised cluster and includes customisable workload models representing light, medium, heavy, and peak request patterns. System performance is evaluated using metrics such as response time, throughput, availability, cost-effectiveness, mean time to recovery, and false positive scaling events. A reinforcement learning agent monitors tracked system metrics and learns scaling policies that optimise long-term performance and stability through repeated interactions with the environment. The application interface allows users to control simulation parameters, including the number of policy runs, the number of episodes per run, and the number of steps per episode, as well as cluster configuration parameters such as the number of nodes and cores per node. The workload intensity can be adjusted to analyse system behaviour in different operating scenarios. This configuration allows for systematic evaluation of adaptive auto-scaling strategies and their impact on resource efficiency and fault tolerance in containerised CAD systems. The study represents a methodological innovation thanks to its interactive, experiment-based evaluation interface, which combines modelling and orchestration logic.

Keywords: adaptive resource management, experimental setup, containerized cluster, workloads, kubernetes, classic pod autoscaling, thresholds (HPA), autoscaling strategy, q-learning

Algorithm for constructing fully interpretable segmented linear regressions

2026. T.14. № 3. id 2212

Bazilevskiy M.P.

DOI: 10.26102/2310-6018/2026.54.3.018

This article is devoted to the relevant scientific field – interpretable machine learning. Previously, the author introduced the concept of «fully interpretable linear regression», which is constructed using ordinary least squares for the entire set of statistical data. In this article, this concept is generalized to segmented linear regression, in which data is first divided into segments, and then its own linear regression is constructed on each of them. An algorithm for constructing fully interpretable segmented linear regressions has been developed. Its peculiarity is that, firstly, the division of the predictor space into segments is carried out using logical activation functions for the arguments of binary operations min. Secondly, paired regression is construct in each segment, which completely solves the problem of multicollinearity. Using the developed algorithm, a segmented linear regression of concrete compressive strength was constructed based on a sample of 1030 observations. In all its eight segments, the values of the linear regression determination coefficients do not exceed 0.8, which indicates the presence of unaccounted-for factors, so the constructed model cannot be strictly attributed to fully interpretable ones. However, all other interpretability conditions are met. In addition, the segmented model turned out to be much better in terms of approximation quality than simple linear regression.

Keywords: regression analysis, interpretability, segmented linear regression, ordinary least squares, multicollinearity, significance of estimates

Deep learning architectures for multiphase CT image segmentation

2026. T.14. № 3. id 2211

Samsonenko S.V. Kashirina I.L.

DOI: 10.26102/2310-6018/2026.54.3.012

The article provides a comprehensive systematic analysis of modern deep learning architectures for automatic segmentation of multiphase CT images. The specific features of multiphase data are considered in detail, the main of which are spatial mismatches (offsets) between phases caused by patient movements and the different nature of the accumulation of contrast agent in pathological tissues at different phases. These features make direct adaptation of classical segmentation methods ineffective and require the development of specialized architectures. The article traces the evolution of approaches: from basic convolutional networks (U-Net, 3D U-Net, nnU-Net) and hybrid models (TransUNet, UNETR) combining convolutions and transformers to specialized solutions. Special attention is paid to models with mechanisms of cross-attention between phases, such as PA-ResSeg, M3Net and MULLET, which allow for implicit alignment of features and adaptive merging of information from different phases without explicit registration (alignment) of images. The paper also analyzes the comparative advantages of various data fusion strategies from different phases (early, late, cross-interaction), discusses issues of computational efficiency and availability of open datasets. Key trends and promising areas of development of the field have been identified, including the use of fundamental models (MedSAM, VoxTell) and modal-agnostic learning. It is concluded that further progress in the field of multiphase segmentation of CT images is associated with the creation of computationally efficient architectures capable of integration into the real clinical process to support diagnostic solutions.

Keywords: hybrid architectures, image segmentation, attention mechanisms, multiphase CT, feature fusion, medical imaging, deep learning, computer vision, PA-ResSeg, m3Net

Research on uncertainty in multi-agent road surface monitoring

2026. T.14. № 2. id 2210

Podberezkin A.A.

DOI: 10.26102/2310-6018/2026.53.2.009

The relevance of this study is determined by the fact that, in road-infrastructure monitoring platforms, errors at the stage of detection and interpretation of object conditions can propagate into normative and managerial decision errors, especially under real-world acquisition conditions (shadows, glare, wet/snow-covered pavement, contamination, and ambiguous defect boundaries), where the risk of misclassification and inaccurate localization increases. This is critical for threshold-based normative assessment, since even small inaccuracies may change the condition category and, consequently, lead either to unjustified maintenance assignments or to missing hazardous defects. Therefore, this paper investigates the use of detection uncertainty for road-surface defect monitoring within a multi-agent pipeline, where observation results are transferred between components together with the processing context via the Model Context Protocol as a unified mechanism for exchanging events, metadata, and interpretation parameters. The main approach is to build a computational pipeline that includes video-data preprocessing, defect detection, computation of the uncertainty indicator H(p) from the class-probability distribution, assignment of the status "automatic/validation/refinement" subsequent normative interpretation, and aggregation over road-network segments. To ensure reproducibility, each run is recorded as a unified "experiment context" (scene/frame identifier, model version, threshold parameters, decision status), enabling comparable mode-to-mode evaluation and auditing of discrepancy causes. Verification is based on comparing normative decisions with expert assessment and analyzing how the share of erroneous normative decisions depends on the automatic-decision threshold for H(p), while the risk-oriented logic routes high-uncertainty detections to validation and reduces the probability of errors in borderline cases. The results show that context logging via Model Context Protocol and accounting for H(p) improve experimental reproducibility and the soundness of normative interpretation, decreasing the risk of incorrect maintenance prioritization by separating ambiguous observations and preserving the decision rationale.

Keywords: multi agent system, road surface monitoring, road surface defects, computer vision, detection uncertainty, normative interpretation, context logging

A method for information extraction based on extractive question-answering models and strategies for evaluating and aggregating relevant text fragments

2026. T.14. № 3. id 2207

Martynyuk P.A.

DOI: 10.26102/2310-6018/2026.54.3.008

In the context of accelerated growth of heterogeneous textual data volumes, universal approaches to information extraction that are independent of the specific structure and domain of source texts have become particularly important. Despite the widespread adoption of large generative language models, the problem of accurate and resource-efficient information extraction from textual data remains relevant. While possessing broad capabilities, generative models are often excessive for specialized information retrieval tasks and may demonstrate low interpretability of results. This study is part of research work aimed at developing an alternative method for information extraction from unstructured texts to form a structural model of a text document. The proposed approach focuses on identifying semantically rich text fragments through relevance analysis relative to given thematic aspects of the text. This research presents an information extraction method using an extractive question answering model, based on multi-level answer aggregation combining strategies for assessing text fragment relevance, semantic clustering, and final answer selection for a given question. The proposed approach enables identification of words in the text that are most relevant to the target thematic aspects, which can subsequently be used to extract reliable information from the document. The article presents experimental results confirming the effectiveness of the proposed method in identifying semantically relevant elements of a text document. The obtained results have practical value for developing automated systems of text semantic structure construction and can be applied in document analysis, information retrieval, and intelligent text processing tasks.

Keywords: natural language processing, information extraction, unstructured text, question-answering model, self-attention mechanism

Experimental analysis of the stability of unmanned aerial vehicles multi-system GNSS receivers to deliberate radio frequency impacts

2026. T.14. № 4. id 2205

Stupina A.A. Kukartsev V.V. Kirill K.I. Masyuk M.A.

DOI: 10.26102/2310-6018/2026.55.4.013

In most modern unmanned aerial vehicles (UAVs), global navigation satellite systems (GNSS) are used as the main means of determining spatial position. However, civilian navigation signals have low energy security and are vulnerable to deliberate radio frequency influences at the physical level, such as signal suppression and substitution, which can lead to loss of navigation solutions or the formation of false coordinates. The purpose of this work is an experimental analysis of the stability of UAV navigation receivers to deliberate radio frequency influences and an assessment of the influence of interfering signal parameters on the reliability of receiving GNSS navigation information. As part of the study, the frequency and signal characteristics of GPS, GLONASS, Galileo and BeiDou systems were analyzed, as well as experimental measurements of the signal-to-noise ratio C/N₀ when exposed to barrage interference of various power and geometry of the interference source location. Additionally, the effect of shielding the navigation receiver was investigated and an asynchronous attack using software-defined radio devices was implemented. As a result, it was found that a decrease in C/N₀ below 25–28 dB·Hz leads to a loss of stable navigation reception, regardless of the navigation system used. It is shown that low-power sources of interference can disrupt the navigation support of UAVs at distances up to several hundred meters, and the shielding of the receiver reduces the effectiveness of interference, but does not provide complete protection.

Keywords: unmanned aerial vehicles, global navigation satellite systems, navigation receivers, radio frequency interference, navigation stability

A computational method for image segmentation based on a Dirichlet field and an analysis of the asymptotic accuracy of spatial regularizer discretization

2026. T.14. № 3. id 2204

Shchetinin E.Y. Andreychuk A.A.

DOI: 10.26102/2310-6018/2026.54.3.009

A computational method for semantic image segmentation with distributional uncertainty estimation is proposed based on representing the prediction as a Dirichlet distribution field. Unlike approaches that require multiple stochastic inference runs (MC dropout) or averaging over an ensemble of independent models, the method computes uncertainty maps in closed form based on the Dirichlet field parameters predicted in a single forward pass of the neural network. The method is formulated as the minimization of a composite functional including the expected logarithmic loss function (expected log-loss), KL regularization for controlling the distribution concentration, and spatial smoothing that takes into account local image intensity variations (edge-aware). For fixed smooth fields, the asymptotic discretization accuracy of the spatial regularizers used is established: the discrete Dirichlet energy approximates the corresponding continuous integral with a first-order error over the grid step. Additionally, a formal decomposition of the overall uncertainty into epistemic and data-supported components was introduced, which can be used in further analysis of the method's behavior and the development of extensions. Computational experiments were performed on three medical image datasets (ACDC, Synapse, CHAOS) with 10 independent initializations. In the main comparison with the baseline model trained using cross-entropy, the differences are statistically significant across initializations on all datasets; for ACDC, significance at the patient level was further confirmed. The method improves segmentation quality and improves the calibration of probability estimates with an overhead of approximately 17 %. In the task of detecting pixel-level segmentation errors, the uncertainty map achieves an AUROC of 0.891.

Keywords: image segmentation, neural network methods, dirichlet distribution, uncertainty estimation, calibration, dirichlet energy, edge-aware regularization, asymptotic sampling accuracy

Simulation model for managing laboratory staff workload during a pandemic using the AnyLogic platform

2026. T.14. № 4. id 2203

Donsckaia A.R. Lomakin A.S. Zubkov A.V. Orlov D.V. Nazarov N.O. Kovaleva E.S.

DOI: 10.26102/2310-6018/2026.55.4.003

The sharp increase in the burden on healthcare systems during the COVID-19 pandemic has shown the inefficiency of traditional methods of calculating labor productivity based on mathematical formulas. They do not take into account the dynamics of work processes, problems in the planning of labor resources, equipment and areas. This leads to inefficient load distribution, especially when, using the example of clinical laboratories, it became necessary to process thousands of samples for PCR testing every day. The aim of the research is to develop and analyze a method for workload planning using simulation modeling in AnyLogic, which allows visualizing and optimizing laboratory processes. The tasks include an analysis of existing approaches, a description of the methodology, application using the example of a PCR laboratory, and an assessment of the benefits in a pandemic. The proposed approach includes timekeeping of technological processes, data collection in tabular form, and creation of a digital laboratory model to identify bottlenecks, equipment and personnel downtime. Using the example of a PCR laboratory, the possibility of optimizing resources, calculating maximum productivity, and justifying purchases is demonstrated. The method makes it possible to increase the efficiency of laboratory production in situations of unpredictable demand, minimizing the risks of disruptions and financial losses.

Keywords: simulation modeling, anyLogic, workload planning, laboratory production, COVID-19 pandemic

Optimal control of finite increments of model factors based on sensitivity analysis

2026. T.14. № 2. id 2202

Sysoev A.S.

DOI: 10.26102/2310-6018/2026.53.2.010

The article addresses the topical inverse problem of target-oriented control: determining the necessary finite changes to the system's input factors to achieve a desired target state, as opposed to the classical direct problem of forecasting. To solve it, a new methodological approach is proposed. This approach is based on sensitivity analysis utilizing the Lagrange mean value theorem. This framework allows for moving beyond local linearization to precisely account for nonlinear effects and factor interactions under substantial, practically observed changes. The key scientific result is the development of a universal iterative algorithm, which, for a given mathematical model, determines the vector of finite changes for the controllable factors that ensures the required increment in the output indicator with minimal total cost of the introduced changes and within given constraints. At each iteration step, the model's gradient (sensitivity estimate) is computed at an intermediate point, whose position is sequentially refined, and an auxiliary constrained optimization problem is solved. The practical efficiency and operability of the proposed method are verified using a numerical example with the nonlinear Ishigami model. The algorithm successfully found the optimal control action, ensuring high accuracy in achieving the target.

Keywords: inverse control problem, sensitivity analysis, finite change analysis, lagrange mean value theorem, constrained optimization

Application of artificial intelligence methods to analyze human behavioral biometrics in ensuring the security of complex information systems

2026. T.14. № 2. id 2201

Shelestova O.V. Kochkarov A.A.

DOI: 10.26102/2310-6018/2026.53.2.015

This article examines the application of artificial intelligence methods and technologies to analyzing human behavioral biometrics in the security of complex information systems. The relevance of the study stems from the limitations of traditional authentication mechanisms, which focus primarily on the initial stage of a user session and are ineffective in detecting user impersonation during interaction with the system. An alternative approach is proposed, using user behavioral characteristics to continuously assess trust in the current session. The paper analyzes anonymized text input data on a mobile device, reflecting the temporal and structural features of user interaction with the interface. It is shown that the combination of such characteristics allows for the identification of stable behavioral patterns suitable for user profiling. Using dimensionality reduction and cluster analysis methods, typical behavioral profiles are identified, differing in input style and rhythm, as well as the nature of corrections. Cluster membership is established to be maintained across multiple sessions with acceptable variability in individual characteristics. A risk-based approach to assessing behavioral deviations is proposed, based on comparing current behavioral indicators with a typical cluster profile. The study's results confirm the feasibility of using cluster behavioral profiles in risk-based access control systems and can be used in the design and development of continuous authentication mechanisms in complex information systems.

Keywords: behavioral biometrics, information security, artificial intelligence, machine learning, cluster analysis, continuous authentication, user behavior analysis

Agent-based approach to intelligent search in library systems

2026. T.14. № 2. id 2199

Rzyankin I.S. Baryshev R.A. Guchko A.A.

DOI: 10.26102/2310-6018/2026.53.2.008

The article explores the application of an agent-based Retrieval-Augmented Generation (Agentic RAG) approach to intelligent search tasks in library collections. The object of the study is the Agentic RAG architecture, which integrates information retrieval mechanisms with agent-based planning and self-evaluation of intermediate results. The addressed problem concerns the limitations of classical Retrieval-Augmented Generation in handling complex thematic and contextual queries within semantically rich library data environments. Unlike traditional RAG pipelines, the agent-based architecture enables iterative refinement of search strategies, adaptive decision-making, and reassessment of intermediate outcomes. The research methodology is based on the development of a software prototype implementing Agentic RAG and its experimental comparison with a classical RAG baseline using a real university library corpus comprising bibliographic metadata, annotations, and full-text fragments. The evaluation framework includes standard information retrieval metrics (Precision@k, Recall@k, MRR, nDCG) as well as expert-based assessment of answer relevance. The results demonstrate a consistent superiority of Agentic RAG in terms of retrieval accuracy, recall, and ranking quality, particularly for complex queries. However, the interpretation of findings is constrained by the selected evaluation metrics and the characteristics of the experimental corpus. The practical significance lies in the potential integration of agent-based architectures into library information systems without requiring substantial infrastructural changes.

Keywords: agent-based search, retrieval-Augmented Generation, library information systems, intelligent search, semantic search, neural network technologies, agent architectures

Ontology-based approach to predicting consumer purchasing behavior in e-commerce

2026. T.14. № 2. id 2196

Svyatov R.S.

DOI: 10.26102/2310-6018/2026.53.2.018

The relevance of this study is determined by the need to improve the accuracy and interpretability of models for predicting consumer purchasing behavior in online stores. Existing machine learning methods demonstrate high performance; however, their effectiveness largely depends on the composition and structure of the feature space, which is typically formed empirically and does not reflect the causal relationships between user actions. This study aims to develop a purchasing behavior prediction method based on an ontological analysis of the e-commerce domain. A formalized approach is proposed for describing entities and their interrelations, providing a systematic construction of the feature space and enabling its scalability across various online stores. The gradient boosting algorithm CatBoost was employed as the machine learning tool, trained on data obtained from the Yandex.Metrica web analytics system. The proposed method was tested on five online stores with different thematic focuses. Experimental results demonstrated stable quality metrics, with F-scores ranging from 65 % to 83 %, confirming the applicability and reproducibility of the developed approach. The findings have practical significance for the development of intelligent decision support systems in e-commerce and can be utilized in designing scalable analytical platforms for predicting user activity and purchase conversion.

Keywords: machine learning, ontology analysis, user behavior analysis, e-commerce, consumer behavior prediction, online stores

A comparative study of deep learning architectures for interpretable diagnosis of retinal diseases

2026. T.14. № 2. id 2195

Miroshnichenko V.V. Kashirina I.L.

DOI: 10.26102/2310-6018/2026.53.2.016

Interpretability of deep learning decisions remains a critical requirement for their application in medical diagnostics. This study presents a comparative analysis of three modern neural network architectures—Vision Transformer (ViT), Swin Transformer, and ConvNeXt – for multiclass classification of retinal diseases using optical coherence tomography (OCT) images. The research was conducted on the open OCTDL dataset containing 2.064 images across seven diagnostic categories with pronounced class imbalance. To compensate for this imbalance, a loss function weighting strategy was employed. All three models achieved validation accuracy exceeding 0.91, with ConvNeXt demonstrating the best performance (0.945) and an optimal balance of sensitivity and specificity, particularly for rare pathologies. Model interpretability was evaluated using Grad-CAM, attention weight visualization, and the model-agnostic LIME method. The analysis revealed that ConvNeXt combined with Grad-CAM provides the most reliable localization of clinically significant features, whereas ViT attention maps and Swin Transformer activation maps often appeared blurred or focused on non-informative regions. The results confirm the advantage of ConvNeXt as the most promising architecture for clinical deployment in ophthalmological diagnostics, owing to its combination of high accuracy, interpretability, and moderate computational requirements.

Keywords: deep learning, vision Transformer, swin Transformer, convNeXt, retinal diseases, grad-CAM

Application of an actuator model to improve the performance of an unmanned aerial vehicle lateral g-load control system

2026. T.14. № 3. id 2194

Smirnov V.A. Orlov V.P.

DOI: 10.26102/2310-6018/2026.54.3.004

This article examines the problem of improving the performance and accuracy of g-load control loops for highly maneuverable unmanned aerial vehicles. It is noted that traditional approaches based on a full range of physical sensors and linearized models lead to design complexity and are insufficient to compensate for significant aerodynamic nonlinearities and parameter spreads. A proposed solution is a transition to model-based control, replacing the steering actuator position sensor signal with the output signal of its virtual mathematical model. The study aims to develop the structure of an astatic loop implementing this approach. A three-loop system with an integral angular velocity stabilizer and compensation for the nonlinear torque characteristic is presented, ensuring astatic control without additional integrating links. To implement the approach in practice, the introduction of correcting devices in the high-frequency loop considers the total phase delays is proposed. The effectiveness of the solution is demonstrated using statistical modeling with random variations in the system parameters. It is shown that replacing a real actuator signal with its model does not lead to a statistically significant deterioration in the quality of transient processes, which confirms the possibility of increasing the speed and reliability of the system while simultaneously simplifying its hardware implementation.

Keywords: control system, lateral g-load, astatic loop, stabilization actuator, model-based control

Improving traffic quality of service in hybrid networks with cloud and fog layers

2026. T.14. № 3. id 2193

Glushak E.V. Mikhailova P.D.

DOI: 10.26102/2310-6018/2026.54.3.001

Improving the quality of service (QoS) in hybrid networks with cloud and fog levels is an urgent task of modern development of telecommunication systems. As the volume of data transferred increases, traditional resource management methods become insufficiently effective. Hybrid networks combining cloud and fog computing can significantly improve performance and reduce latency. An urgent task is to ensure a balance between high throughput, minimal delays and low packet loss. Efficient resource allocation helps to reduce energy consumption and operating costs. The article is devoted to optimizing the quality of traffic service in hybrid networks combining cloud and fog computing. A mathematical model based on a system of differential equations is presented that describes the dynamics of load, queues, resource allocation, delays, and packet losses. The model formalizes the task of optimal resource management in order to minimize delays and losses with limited capabilities. Numerical integration methods are used for the solution. The developed algorithm makes it possible to effectively balance the load between cloudy and foggy levels. The proposed approach proves its effectiveness for optimizing modern telecommunication systems, especially for applications with critical response time requirements.

Keywords: hybrid networks, cloud computing, fog computing, quality of service (QoS), traffic optimization, load balancing, data latency, layered architecture, resource allocation, routing

Optimization-simulation modeling for resource allocation management in geographically distributed organizational systems with variable workloads

2026. T.14. № 4. id 2191

Boklashov I.I. Ivanov D.V. Lvovich Y.E.

DOI: 10.26102/2310-6018/2026.55.4.008

This paper addresses the integration of optimization approaches and simulation modeling to manage resource allocation within an organizational system characterized by a geographically distributed operational environment and variable activity volumes. The research methodology employs a systems approach, utilizing structural modeling to represent the organization's functioning and management. By structuring the interaction between the control center and operational units, the study establishes quantitative connection characteristics, which are recorded via the system's digital monitoring. The core component of this optimization-simulation model involves the multi-alternative selection of priority units for integrated resource allocation, subject to balance constraints and a stochastic flow of requests defining work requirements. Variable activity volumes are accounted for through a multi-period distribution of integrated resources. Consequently, the set of candidate units for the subsequent period includes those excluded from the optimized subset in the previous step, alongside a random component determined by the simulation results. The study demonstrates that single-period optimization utilizes real-time data to identify priority units for resource allocation. Furthermore, the multi-period optimization-simulation process generates sufficient synthetic data on resource demand; when combined with retrospective monitoring data, this forms a representative training dataset for machine learning predictive models. Finally, the paper defines management decisions supported by these predictive models for both the operational and developmental stages of the organizational system.

Keywords: organizational system, management, optimization, simulation modeling, machine learning, forecasting

A method for implementing pseudo-realistic movement of non-player characters in open virtual worlds

2026. T.14. № 3. id 2190

Shutov K.I. Lobanov A.A.

DOI: 10.26102/2310-6018/2026.54.3.006

The open-world game market increasingly demands NPC (non-player character) behaviour that feels believable yet remains designer-controllable under tight computational budgets. Common solutions tend to be extreme: either they attempt full simulation and overload the system, or they rely on predictable scripted patterns. This paper proposes a pseudo-realistic NPC movement method that bridges these extremes. The core idea is to verify spawn reachability using a matrix of shortest-path distances between world areas. When the player enters an area, the algorithm selects only those NPCs that could have physically reached it given elapsed time, movement speed and available routes, making an encounter consistent with hidden travel rather than instantaneous spawning. Encounter frequency is controlled via a priority scheme, allowing designers to tune event density and the rarity of specific characters without maintaining a detailed simulation. Candidate selection is further accelerated by reordering an almost-sorted list, reducing the cost of repeated queries under similar conditions. Experiments on synthetic graphs show that the core client-side runtime stays within milliseconds for up to 1000 NPCs. The method delivers believability and control at low computational cost and can be integrated into existing engines to adjust difficulty and balance.

Keywords: game design, game development, video games, pathfinding algorithm, sorting algorithm, NPC, non-player character

Privacy-preserving threat intelligence sharing across government agencies using FEGB-Net

2026. T.14. № 3. id 2189

Arm A. Lyapuntsova E.V.

DOI: 10.26102/2310-6018/2026.54.3.011

Government networks are increasingly targeted by coordinated cyberattacks that exploit similarities in infrastructure and operational practices across agencies. Although early detection at one organization could provide valuable warnings to others, effective threat intelligence sharing is often constrained by data sovereignty and privacy regulations. This paper presents an extension of the federated ensemble graph-based network (FEGB-Net) framework that enables privacy-preserving threat intelligence sharing across government agencies. The proposed approach extracts compact behavioral threat signatures from locally trained federated graph neural network models, protects these signatures using differential privacy, and supports real-time cross-agency threat matching. Experimental evaluation using the CICIDS2017 dataset demonstrates that detection accuracy remains comparable to isolated operation, while coordinated attack detection time is reduced by up to 88.5 %. Privacy analysis confirms that ε-differential privacy with ε = 2.0 limits membership inference attacks to near-random success. The results show that collaborative defense can be achieved without compromising data privacy or sovereignty.

Keywords: federated learning, threat intelligence sharing, graph neural networks, differential privacy, government cybersecurity

A modified algorithm for simulated annealing for the task of diagnosing failures of analog electronic devices

2026. T.14. № 2. id 2188

Uvaysov S.U. Chernoverskaya V.V. Hai N.D. Hai V.T. Pham X.T.

DOI: 10.26102/2310-6018/2026.53.2.013

The article presents the results of a study that developed a new method for automated diagnostics of functional components of electronic devices in order to identify parametric component failures, electrical failures, and short circuit detection. The relevance of the study is due to the ever-increasing complexity of modern electronics, when traditional diagnostic methods do not provide the necessary accuracy and efficiency of diagnostic procedures, which leads to an increase in equipment failures during operation and an increase in the cost of its maintenance and repair. The proposed method is based on a well-known algorithm for simulated annealing, which has been adapted to solve the problems of troubleshooting electronic devices. Objective: to propose a new method for diagnosing failures of electronic equipment based on a modified algorithm for simulated annealing, aimed at increasing the reliability of identification of faults occurring in nodes and modules during the operation of modern electronics, as well as to increase the degree of automation of diagnostic procedures. Physical and model experiments conducted during the study showed that the proposed method based on a modified algorithm effectively detects a number of failures, including complex cases of sequential failures that could not be identified using traditional methods. In addition, the proposed approach requires less time for analysis and makes it possible to increase the reliability of diagnostics of the studied nodes and modules of electronic equipment. The results obtained confirm the promising application of the method in the tasks of technical diagnostics, including its further integration into automated control systems of electronic equipment.

Keywords: electronic device, fault diagnosis, malfunction, annealing simulation algorithm, optimal solution, function extremum, solution generation mechanism, markov chain

Rating model with latent parameters based on the softmax function

2026. T.14. № 3. id 2185

Bratischenko V.V.

DOI: 10.26102/2310-6018/2026.54.3.002

The relevance of the work is due to the widespread use of recommendation systems using rating assessments. Based on the results of the review of recommendation methods, it is concluded that it is possible and expedient to build a probabilistic rating model similar to the Item Response Theory models. It is proposed to use latent interest parameters for each subject, characterizing its tendency to set a certain rating, and latent agreeability parameters for each object, characterizing the frequency of obtaining a certain rating. The probabilities of the estimates are determined by a softmax function with interest and matching parameters. The equations connecting observations and latent parameters are obtained using the maximum likelihood method. An iterative procedure for calculating parameters based on rating estimates has been developed and its convergence has been substantiated. The model was tested using the well-known Nexflix set with movie ratings and statistical characteristics of the ratings predictions were presented. The accuracy of predicting ratings turned out to be comparable with the accuracy of predictions of other models. The advantage of the proposed model is a compact description of the assessment probabilities in the form of sets of latent parameters of subjects and objects, which makes it possible to predict rating estimates. The disadvantages include the computational complexity of estimating the parameters and the need to recalculate the parameters when new data becomes available. The proposed model can be used to study and predict ratings.

Keywords: recommender system, rating assessment, collaborative filtering, probabilistic model with latent parameters, softmax function