Computer systems and information technologies

AI-ASSISTED PIPELINE FOR SEMI-AUTOMATIC CONSTRUCTION OF HTR CORPORA FROM HISTORICAL DOCUMENTS

Andrii IVASECHKO — 2026-09-30

This paper presents an improved software application for the semi-automatic annotation of handwritten words from historical documents using a CRNN deep learning model. The proposed approach integrates an artificial intelligence module directly into the annotation process, automatically generating a preliminary transcription for each image. It implements the Human-in-the-Loop concept, where the prediction serves as a recommendation that can be accepted, corrected, or replaced by the annotator, preserving expert control over the final annotation while reducing repetitive manual input.

The application provides automated ordering of images according to metadata, structured storage of transcriptions and document information, text normalization, detection of already processed samples, and resumption of interrupted annotation sessions. The AI prediction module is integrated directly into the graphical interface, allowing the annotator to review and edit the generated transcription without switching between separate tools.

The CRNN model was trained on a custom corpus of 2,763 segmented handwritten word images with diverse graphical and linguistic characteristics. Training employed image augmentation, AdamW optimization, CTC loss, Gradient Clipping, adaptive learning-rate reduction, and Early Stopping. The main focus of the study is the evaluation of annotation performance and the effect of AI-assisted predictions on user productivity.

Experimental results showed that the average processing time decreased from 8.4 s per word with manual transcription to 4.7 s when an AI prediction was available. This corresponds to a 1.79-fold increase in processing speed and a 44.0% reduction in processing time. Annotator productivity increased from 7.1 to 12.8 words per minute. These results demonstrate the practical effectiveness of integrating HTR prediction into an expert annotation workflow and its potential to accelerate the construction of training corpora for historical handwritten text recognition systems while maintaining expert verification of the final annotations.

SEMANTIC EMBEDDINGS AS A CANDIDATE-GENERATION CHANNEL FOR SANCTIONS NAME SCREENING

Iegor PAVLENKO — 2026-09-30

Sanctions name screening is an open-set identification problem with asymmetric error costs: a missed designee is more consequential than an extra alert. We evaluate OpenAI text-embedding-3-large and Google gemini-embedding-001 as candidate-generation channels over a fixed OFAC SDN individuals snapshot (7,426 entries; primary firstName + lastName only), using a controlled stress test in which the same 150 designated individuals are rendered in four Slavic-name regimes — direct Latin spelling, ICAO-style romanisation, French-convention transliteration, and a Russian form converted to Ukrainian and officially romanised — at two specificity levels, against easy and deliberately hard look-alikes, at three Matryoshka dimensions (768, 1536, 3072). Gemini led in every non-trivial regime and its margin grew with transliteration distance, reaching 34.6 pp of Recall@1 on two-part Russian-to-Ukrainian forms (entity-cluster bootstrap 95% CI 27.1–42.1 pp; exact McNemar p ≈ 1×10⁻¹⁶), with no systematic loss down to 768 dimensions, whereas OpenAI degraded with both transliteration distance and reduced dimension. Top-1 similarity did not separate transliterated positives from hard look-alikes: for OpenAI on the hardest regime the positives scored below the negatives (AUC 0.318, 95% CI 0.251–0.385). At an equal candidate budget, unioning both models' top-5 lists matched Gemini alone at k = 10 exactly (99.36%), and one designee sat at rank 60 despite a one-character surname difference. Embedding similarity is therefore a useful candidate-generation channel but not a reliable accept/reject rule; on a limited, partly synthetic corpus queried against mutable hosted APIs these are benchmark-specific findings, not a no-miss guarantee. Data, results and figures are released.

PATTERN PAGE OBJECT MODEL FOR MODULAR TEST AUTOMATION ARCHITECTURE

Yuriy PONOCHOVNYI — 2026-09-30

Modern web applications evolve through short development cycles and frequent interface redesigns, which often make UI test suites fragile because scenarios depend directly on page structure, selectors, timing, and navigation logic. The paper examines the Page Object Model (POM) as an architectural foundation for modular test automation rather than a technique for storing locators. The study combines analysis of recent empirical literature on GUI test maintenance with a case illustration based on a Playwright and TypeScript framework. POM is compared with the Screenplay pattern and Component Object approach by separation of concerns, abstraction granularity, readability, reuse, scalability, and implementation complexity. The proposed interpretation distinguishes test scenarios, page interfaces, reusable component objects, fixtures, services, data factories, and utilities as separate architectural layers, formalising change propagation, cohesion, reuse, and maintenance cost through six numbered indicators. Attention is paid to TypeScript contracts, dependency injection, Playwright locators, fixtures, and isolated browser contexts as means of maintaining explicit dependencies and predictable module boundaries. A qualitative maintainability model, consistent with ISO/IEC 25010, explains how centralising interface knowledge reduces change propagation. The illustration shows page objects exposing business operations while component objects encapsulate recurring fragments such as navigation panels, modals, and notifications. POM improves maintainability only when responsibilities are limited to interface interaction and state representation; oversized classes, assertion-heavy methods, deep inheritance, and uncontrolled cross-page dependencies turn POM into a source of coupling. A modular POM implementation should be treated as an adaptable architectural layer whose effectiveness depends on cohesion, explicit contracts, composition, and controlled abstraction.

RISK-SENSITIVE Q-LEARNING IN A TWO-AGENT MARKOV ATTACK–DEFENSE MODEL

Andrii PRYTULA — 2026-09-30

A risk-sensitive method for forming a web application security testing policy based on entropic-risk Q-learning in a two-agent Markov attack–defense model is proposed. The study addresses the need to detect short critical compromise trajectories under partial observability, noisy estimation of defensive attributes, and stochastic changes in the defensive profile of the environment. Unlike risk-neutral approaches, the proposed method employs a dense reward oriented towards the time to compromise (TTC) and a post-episode evaluation of tail losses in terms of the conditional value at risk at the 0.95 confidence level (CVaR0.95), where the risk-sensitivity parameter β of the entropic-risk operator controls the degree of risk aversion. Two modes of the simulation environment are introduced, namely the controlled environment Env1 and the stochastic environment Env2, which are used to analyze policy stability under different levels of information uncertainty. The ToN-IoT and CIC-IDS2017 datasets were used for the statistical parameterization of the environments. Aggregation of semantically related MAL attack steps into web components reduces the dimensionality of the state representation by a factor of 8–12, which makes the training of the risk-sensitive policy computationally manageable. The obtained results show that the proposed method provides the lowest average time to compromise among all the compared approaches. In Env2 the TTC value decreased from 36.9±1.5 to 27.9±1.1 steps for ToN-IoT and from 39.7±1.7 to 31.4±1.3 steps for CIC-IDS2017, the CVaR0.95 of normalized tail losses decreased by approximately 31 %, and the number of episodes required for policy convergence was reduced by approximately 35.4%. The most effective mode is achieved at β ≈ -0.6. The practical significance lies in the possibility of applying the method for automated penetration testing and for evaluating the robustness of web application protection mechanisms.

DESIGN AND IMPLEMENTATION OF A RAG-BASED AI AGENT FOR ROBO-ADVISORS

Serhii SAVCHENKO — 2026-09-30

This study focuses on the design and implementation of an agentic system based on Retrieval-Augmented Generation (RAG) architecture for automated financial advisory services (Robo-Advisors, RA). The relevance of this study lies in the knowledge limitations of large language models (LLMs) in domain-specific areas, such as the financial sector, where factual errors can lead to direct financial losses and regulatory risks. Based on a literature review, the class of tasks for a modern Robo-Advisor platform suitable for RAG-based solutions has been identified. The purpose of this research is to develop a RAG agent architecture for a Robo-Advisor service and verify its functionality on a working prototype. Functional requirements for an RA module were established to track the behaviour of financial assets within a specific investment portfolio. An analysis was conducted on the tools used to implement typical interaction scenarios with a RAG system (data acquisition, text preprocessing, vectorisation, data normalisation, LLM interaction, etc.). A software architecture was designed in which agent functions (obtaining stock prices, analysing financial news, automated reports, user communication) are reflected in separate software modules. The system is divided into an asynchronous indexing circuit and a synchronous user interaction circuit. The agent's memory is implemented as hybrid storage, where a relational database serves as the source of truth, and the vector index is a fully recoverable derivative artifact. A method for re-ranking news search results using a normalised time-of-publication coefficient is proposed. User interaction is implemented as a chatbot for a messenger. Pilot operation confirmed that the system meets the established requirements; identified shortcomings and limitations of the agent are described, along with methods for their resolution. The results of this study can be used to integrate RAG agents into full-scale Robo-Advisor services.

ADAPTIVE PID CONTROL OF UAV POSITION USING VISUAL DISPLACEMENT

Mykola SUSLO — 2026-09-30

Visual-displacement-based control of unmanned aerial vehicles (UAVs) requires the position controller to operate directly on image-space errors while adapting to changes in motion conditions and limited control authority. This study presents an adaptive PID gain-scheduling method for UAV route following using visual displacement estimated from a downward-facing camera. The controller operates in image pixel space and adjusts the PID gains according to the current motion condition, visual-positioning confidence, route progression, cross-track error, and control-signal saturation. Saturation is handled independently for the two image-plane axes, allowing the affected axis to use a saturation-limited gain schedule while the other axis retains the configuration corresponding to the underlying motion condition. The proposed method was implemented and evaluated in a Gazebo/ArduPilot Software-in-the-Loop environment and compared with a fixed-gain PID controller. Preliminary development tests also included triangular and zigzag route geometries, but these runs were excluded from the final statistical validation. The final validation dataset consisted of 100 experimental runs, with 50 runs performed for each controller mode under the same predefined square route and simulation configuration. Performance was evaluated using route-completion rate, trajectory-error measures, cross-track error, route-completion time, and control-signal saturation occupancy. Route-completion proportions were compared using Fisher’s exact test, while continuous run-level metrics were analysed using Mann–Whitney U tests with Holm correction for multiple comparisons. The adaptive controller completed 46 of 50 validation runs, corresponding to a route-completion rate of 92%, whereas the fixed-gain controller completed 35 of 50 runs, corresponding to 70%. The difference in completion proportions was statistically significant. For successfully completed routes, no statistically significant differences were observed after Holm correction in the principal trajectory-error measures. Mean control-signal saturation occupancy decreased from 52.60% for the fixed-gain controller to 48.00% for the adaptive controller, and this difference remained statistically significant after correction for multiple comparisons. The results indicate that, under the investigated simulation conditions, the proposed adaptive PID gain-scheduling method was associated with a higher route-completion rate and reduced operation at the control limits without statistically significant deterioration in observed trajectory-following accuracy. The conclusions are limited to the investigated square-route validation and Software-in-the-Loop environment and require further evaluation using additional route geometries, disturbances, and hardware-based UAV experiments.

COMPARISON OF GPS/IMU INTEGRATION METHODS FOR REAL-TIME VEHICLE POSITIONING

Viktor TKACHENKO — 2026-09-30

This paper addresses the challenge of determining vehicle coordinates in real-time telemetry systems through the integration of Global Positioning System (GPS) and inertial measurement unit (IMU) data. A comprehensive analysis of existing sensor-fusion approaches enabled the formulation of the problem within the framework of embedded, resource-constrained real-time platforms and underscored the necessity of a systematic comparative evaluation of methods with fundamentally different computational and accuracy characteristics. Three sensor-fusion methods are considered: the complementary filter, which combines data through frequency-domain separation; the Extended Kalman Filter, which relies on Jacobian-based linearization of the nonlinear motion model; and the Unscented Kalman Filter, which applies a deterministic sigma-point transform to approximate nonlinear state propagation without linearization. The paper substantiates the choice of methodological foundations underlying each method, including the kinematic vehicle state model comprising geographic coordinates, velocity, heading angle, and sensor bias offsets, as well as the metrics used for their quantitative comparison — positioning RMSE, circular error probable, dead-reckoning drift during GPS outages, and update latency. The growing demand for reliable positioning under GPS-denied conditions, combined with the limited computational resources of embedded telemetry platforms and the nonlinear nature of vehicle kinematics, highlights the need to establish practical, hardware-aware criteria for filter selection. To address this, all three algorithms were implemented on a common hardware and software platform — a Raspberry Pi 5 single-board computer running Python 3.11 — and validated against real driving data collected on a 4.7 km route, with ground truth provided by an RTK GPS receiver. The paper presents experimental results demonstrating the relative effectiveness of the proposed methods in terms of positioning accuracy, robustness to GPS signal loss, and behavior under aggressive maneuvering. Additionally, it outlines practical recommendations for selecting an integration method depending on the computational constraints of the target embedded platform, ranging from microcontroller-based systems to single-board computers and DSP platforms

SENSOR COMPLEX AND HIERARCHICAL DATA PROCESSING ARCHITECTURE FOR SMART RETAIL ENTERPRISES

Ivan TSMOTS — 2026-09-30

The paper addresses the applied scientific problem of designing a sensor complex and a hierarchical data processing system for a smart retail enterprise as the basic component of an adaptive management information technology. It is shown that conventional centralised data acquisition and processing architectures fail to provide the required management responsiveness when a retail enterprise simultaneously generates data streams from radio-frequency identification devices, smart shelves, POS terminals, video analytics systems, microclimate and energy consumption sensors, as well as from corporate ERP, CRM and WMS information systems and digital sales channels. Twelve principles for constructing the sensor complex of a smart retail enterprise are formulated. They cover modularity, peripheral intelligent data processing, interface compatibility and standardisation, multi-source data integration, continuous monitoring, scalability, distributed processing, reliability and fault tolerance, energy efficiency, security and privacy, self-diagnostics and adaptability. Based on these principles, a structure of the sensor complex is developed that combines nine functional groups of data acquisition means together with edge computing nodes and IoT gateways. For the first time, an architecture of a five-level hierarchical data processing system is developed. It comprises the level of sensors and primary data sources, the level of local processing, the level of data integration and consolidation, the level of intelligent analysis and decision making, and the level of decision implementation, communication and monitoring. For each level the functions are defined and the corresponding data transformation operators are formalised, and their composition is presented as a generalised model describing a closed adaptive management loop in which the results of implementing managerial decisions form updated data for the next processing cycle. A mathematical model for evaluating the efficiency of the architecture is developed. It is based on a set of temporal, resource, informational and functional indicators: data processing time and managerial decision formation time, network traffic volume, utilisation of the computing resources of the central platform, the scalability indicator, availability and fault tolerance, the composite quality indicator of integrated data, and the adaptive management efficiency indicator. An integral efficiency criterion and a system of conditions are proposed, under the simultaneous fulfilment of which the developed hierarchical architecture outperforms the traditional centralised one. It is demonstrated that the proposed solutions reduce the time of data processing and managerial decision making, decrease the network load and the load on the central information and computing platform, and increase scalability, integrated data quality, system availability and the efficiency of adaptive management of a smart retail enterprise in real time.

Keywords: sensor complex, smart retail enterprise, hierarchical data processing system, edge computing, Internet of Things, data integration, adaptive management, architecture efficiency evaluation.

SUMMARY OF THE RESULTS OF INFORMATION TECHNOLOGY DEVELOPMENT FOR OPTIMIZING VECTOR GRAPHIC OBJECT PLACEMENT ON A PLANE

Oleksandr TYMCHENKO — 2026-09-30

The article summarizes the results of developing an information technology for optimizing the placement of vector graphic objects within a limited area for automated cutting, contour cutting, and other manufacturing processes associated with efficient material utilization. The problem of placing arbitrarily shaped objects is formalized with consideration of workspace boundaries, allowable rotation angles, technological clearances, non-overlap constraints, and layout quality criteria. The developed technology combines SVG data preparation, discrete contour approximation, genetic optimization, clustering, caching, and parallel processing of placement variants. To reduce computational complexity, each contour together with its buffer zone is converted into a two-dimensional binary matrix, making it possible to replace complex geometric checks with comparisons of matrix masks using bitwise operations. The optimization core is implemented on the basis of a genetic algorithm in which solutions are represented as chromosomes, while new generations are formed using elitist selection, selection, crossover, mutation, and clustering. The technology was implemented as a desktop software tool developed in Kotlin on the JVM platform using Java Swing, which supports SVG file loading, parameter configuration, optimization execution, visualization, and layout saving. Experimental studies were conducted using synthetic and practically oriented sets of vector objects. The relationship between the discretization cell size, layout accuracy, and computation time was established. In the example involving the placement of paper model parts, the number of required A4 sheets was reduced from 10 to 6, corresponding to a 40% reduction. The obtained results confirm the effectiveness of the technology for automated cutting, contour cutting, and other vector object placement tasks.

CRYPTOGRAPHIC HASH-FUNCTIONS BASED ON CELLULAR AUTOMATA AND KECCAK ARCHITECTURE

Serhii YANUSHEVSKYI — 2026-09-30

This research presents a comprehensive study of cryptographic hash functions and the development of novel hashing algorithms based on elementary cellular automata (CA). Modern cybersecurity relies on cryptographic hashing to guarantee reliable authentication and data integrity. To achieve high security alongside computational efficiency, authors propose utilizing non-linear cellular automata as binary pseudorandom sequence generators within a SHA-3 sponge framework. The study examines elementary CA rules from the “Rule 30” equivalence class, specifically Rules 30, 86, 135, and 149, alongside the linear additive Rule 150.

The authors formalized an algebraic state space model over the Galois field F2. Linear diffusion properties are mathematically guaranteed by a symmetric tridiagonal circulant operator matrix M150, while non-linear confusion is introduced by Rule 86. Two distinct architectural approaches were developed and analyzed: a Keccak-like sponge analogue featuring a custom CA-driven round function (Approach A) and a lightweight direct rule interaction model (Approach B). The custom round function combines Rule 86, Rule 150, and cyclic 31-bit rotations across a 1600-bit state register.

Rigorous empirical evaluation was performed using the NIST STS v800-22 suite, confirming high randomness with pass rates P >= 0,96 across subtests. Strict Avalanche Criterion (SAC) analysis demonstrated near-ideal bit mixing. Approach A achieved a mean bit flip ratio of 49.87% and a low absolute SAC error of 0,0013. Performance benchmarking was executed in a Java 17 JVM environment. Approach A attained a processing throughput of 134,62 MB/s, outperforming standard Java SHA-3 by 29,4%. Approach B reached 206,10 MB/s, providing an ultra-fast lightweight alternative.

The findings prove that combining non-linear CA rules with linear matrix operators in sponge constructions significantly enhances diffusion speed, collision resistance, and execution performance, establishing CA as a highly promising foundation for modern cryptographic tools.

ROUTING MULTIMEDIA OBJECTS BASED ON INPUT FEATURE COMPATIBILITY

Dmytro DENYSIUK — 2026-09-30

Identifying the format variant of a multimedia object affects the choice of processing procedure, although the filename, declared MIME type, and initial bytes may imply incompatible interpretations. This study examines the relationship between a model of format representation admissibility, a preliminary candidate identification method, and experimental responses to conflicting input features. The model represents the sources as separate sets of possible format variants and distinguishes missing information, an unrecognized value, an established contradiction, and an approximate byte match. The method selects the candidate solely from the bytes. Metadata constrains whether processing may continue and, for the ISO BMFF family, also determines the target class. A unique route is defined with a fixed target format variant, required level, procedures, and resource limits. Route uniqueness is established for a finite candidate set and well-defined procedures. For identical bytes and target class, the operational part of the route is shown to be invariant under admissible metadata changes. The candidate remains unconfirmed until the subsequent method completes a full read. The S3 test matrix contains 60 scenarios involving a changed suffix, a conflicting declared type, or a concealed prohibited suffix, with 20 scenarios per group. All 60 scenarios resulted in the expected blocking. This result corresponds to specific model constraints and algorithmic steps but does not establish format identification accuracy in a naturally occurring file population. The proposed approach separates the byte-based grounds for selection from permission to continue and preserves candidate status in the route passed between stages. This allows signature-based or statistical identification tools to be integrated while treating their output as a preliminary conclusion that requires confirmation of the object’s structural admissibility.

RIGHT-INVARIANT STEREO VISUAL-INERTIAL ODOMETRY: REJECTION THRESHOLDS AND NOISE SCALE

Anatolii ZHARKOV — 2026-09-30

A mathematical model is presented of a visual-inertial odometry filter built on a right-invariant error on the matrix Lie group with appended inertial measurement unit biases, with landmarks eliminated by projecting the residual onto the left null space of the Jacobian and with measurements rejected by a normalised Mahalanobis distance. For every element of the model the transition matrices, Jacobians and retraction are given in both error conventions, the right-invariant and the classical additive one, which makes it possible to isolate the effect of the convention itself under otherwise identical conditions.

On five EuRoC MH sequences with four front ends arranged in a design (two ways of tracking a point two detectors: the learned XFeat and the classical ORB), the separate and joint contributions of two factors are measured: the error convention and the form of the rejection rule. The widely used constant threshold is shown to correspond to a quantile for a two-component measurement and not to agree with the actual dimension of the projected residual. Replacing it by a quantile threshold reduces the mean position error by 21.1, 23.4 and 20.2 % on three front ends out of four: the best result of the grid is 0.136 m against 0.177 m with the constant threshold and 0.222 m with no rejection at all. The contribution of the error convention is smaller, 2.4–10.0 %, and on ORB with descriptor matching the ratio of the two factors reverses. At the same time the same replacement, with the noise scale left unchanged, causes a loss of solution in one configuration out of twenty (644.9 m), which is removed by matching (0.436 m) rather than by relaxing the rule.

The scientific novelty consists in establishing that the form of the rejection threshold and the adopted measurement-noise scale are subject to joint rather than separate calibration: a threshold defined by a quantile at the actual residual dimension needs no tuned constant, yet transferred without matching it costs the solution. The practical value lies in assigning the threshold without manual tuning and in the quantitatively stated limits of transferability of settings between front ends.

METHOD FOR FORMING INFORMATION SERVICES BASED ON THE FUNCTIONAL COMPONENTS OF AN INFORMATION TECHNOLOGY PLATFORM

Nataliia KUNANETS — 2026-09-30

Modern information technology platforms support complex digital environments by integrating information resources, technological infrastructure, analytical components, information services, users, and intelligent agents. However, existing approaches to information service formation are primarily based on software component composition and insufficiently consider the functional structure of the platform, the relationships among its components, and the development patterns of the information technology ecosystem. The purpose of this study is to develop a method for forming information services based on the functional components of an information technology platform by defining methodological principles, formalizing the component composition process, and constructing a conceptual model for digital service formation. The research methodology is based on systemic, service-oriented, and ecosystem approaches, structural-functional analysis, conceptual modeling, and formalization of information service construction processes. Particular attention is paid to establishing relationships among functional requirements, technological and analytical components, interaction scenarios, and digital services. The results include the development of a method that provides a consistent transition from functional requirements to component composition, information service formation, and their integration into interaction scenarios involving users and intelligent agents. The methodological principles of functional completeness, requirement alignment, scenario orientation, composability, loose coupling, reusability, interoperability, adaptability and personalization, controlled autonomy, and security by design are formulated. The scientific novelty lies in the development of a method that establishes a cause-and-effect relationship between the functional components of an information technology platform, the principles of their composition, and the resulting digital services. The practical significance of the proposed method lies in its applicability to the design of information technology platforms for smart cities, digital twins, decision-support systems, research information systems, and multi-agent environments. Future research will focus on automating service composition, evaluating service quality, and integrating generative artificial intelligence into the information service formation process.

METHOD FOR LARGE INTEGER FACTORIZATION BASED ON THE RESIDUE NUMBER SYSTEM

Stepan IVASIEV — 2026-09-30

The paper investigates the problem of factoring large integers, which remains one of the fundamental challenges in modern number theory and public-key cryptography. An analysis of the classical Fermat factorization method is performed, and its computational bottlenecks associated with repeated square computations and exhaustive iteration are identified. To overcome these limitations, a modified version of the Fermat method based on a recurrent computation of the difference of squares is proposed. The proposed modification eliminates repeated exponentiation operations by replacing them with incremental arithmetic operations, thereby creating the mathematical foundation for the application of residue number systems.

Based on the proposed modification, two novel factorization algorithms are developed. The first algorithm employs a key-folding mechanism in the Residue Number System, where sets of admissible quadratic residues are generated for individual moduli and subsequently combined according to the Chinese Remainder Theorem. This approach provides a multistage algebraic filtering procedure that substantially reduces the number of candidate values requiring verification in the positional number system. The second algorithm introduces an offset-lattice approach, in which the search process is transferred from the value space to the iteration-index space. By constructing periodic sets of admissible offsets for groups of moduli, the algorithm performs lattice-based filtering and evaluates only those iterations satisfying all modular constraints, significantly decreasing the number of expensive perfect-square tests.

The computational complexity of the proposed algorithms is analyzed and compared with that of the classical Fermat method. Theoretical analysis demonstrates that both algorithms reduce the effective search space through modular filtering, while the offset-lattice algorithm additionally decreases the number of examined iterations by exploiting the periodic structure of admissible offsets. Owing to the independence of computations for different moduli, both algorithms naturally support parallel implementation, making them suitable for execution on multicore and distributed computing platforms.

Experimental studies performed on integers of various sizes confirm the effectiveness of the proposed approaches. Compared with the classical Fermat method, the developed algorithms reduce the number of candidate checks and improve computational efficiency while preserving the correctness of factorization. The obtained results demonstrate that residue number systems, combined with key-folding and offset-lattice techniques, provide a promising direction for accelerating the factorization of medium-sized integers and may serve as a basis for further research on high-performance modular factorization algorithms.

COMPARATIVE ANALYSIS OF FOG SYSTEM ARCHITECTURE OPTIMIZATION METHODS: FROM METAHEURISTICS TO HYBRID SOLUTIONS

Oleksandr SBITNEV — 2026-09-30

The rapid proliferation of Internet of Things devices has intensified demands on distributed computing infrastructures, making fog computing — a paradigm that positions computational resources at the network edge — a critical enabler of low-latency real-time applications. Optimally placing services across fog nodes is a proven NP-hard combinatorial problem. Despite a decade of active research, existing studies examine individual algorithm families in isolation, leaving system designers without an evidence-based framework for selecting an optimisation method suited to their deployment context. The aim of this paper is to conduct a systematic comparative analysis of more than ten optimization approaches within a unified analytical framework and to construct a formalised five-dimensional classification model M: X → Y (X = S×W×L×P×T) that maps IoT system characteristics onto the most suitable method class. The study synthesises results from twenty-six primary sources across three method classes — metaheuristic algorithms (GA, PSO, ACO, Firefly), machine learning methods (supervised learning, LSTM, CNN-BiLSTM, GNN, federated learning), and reinforcement learning (single-agent DQN and multi-agent MARL) — evaluated against five metrics: processing latency, energy efficiency, scalability, adaptability to dynamic workloads, and implementation complexity. The analysis yields three principal results. First, PSO and ACO exhibit linear O(n) per-iteration complexity relative to population size, whereas GA and Firefly exhibit quadratic O(n²), making complexity the decisive selection criterion for fog networks exceeding 50 nodes: at that threshold, T_GA ≈ 100·T_PSO. Second, federated learning achieves over 93% classification accuracy in IoT transportation systems without centralising raw data, and is the only method class that satisfies legal data-sovereignty constraints. Third, the fully integrated MARL+GNN+FL approach is theoretically projected to achieve a latency of 65–98 ms and a 22–30% improvement in energy efficiency relative to greedy baselines, based on the principle of architectural orthogonality and on results independently validated for the MARL+GNN and FL sub-architectures; however, none of the 26 reviewed sources validates this combined architecture within a single unified testbed, and empirical co-validation constitutes the primary direction for future work. The formalised model M: X → Y includes explicit numerical thresholds for all five input dimensions, a priority-ordered decision function (Algorithm 1), and six verified output classes; it is falsifiable, reproducible, and extensible, and its application is demonstrated on a smart-city fog deployment scenario with N = 75 nodes.

TOPOLOGICAL OPTIMIZATION AND ARTIFICIAL INTELLIGENCE TECHNOLOGIES FOR DESIGNING DRONE STRUCTURES

Nazarii MULIAK — 2026-09-30

Unmanned aerial vehicles (UAVs) have become an indispensable tool for solving a wide range of complex tasks. To improve and increase the efficiency of unmanned aerial vehicles, it is important to optimize their structural elements. Reducing weight and strategically distributing loads while maintaining structural strength and rigidity directly impacts flight endurance, stability, payload capacity, and overall airframe integrity. Integrating topological optimization with artificial intelligence enables the automated design of components tailored to specific operational requirements.

The application of an integrated approach to automate the design of small-type UAV structures and the combination of classical methods of topological optimization with artificial intelligence and additive manufacturing technologies were considered. A microdrone airframe, fabricated via 3D printing using carbon fiber-reinforced nylon (Nylon PA12 + CF15), was analyzed both before and after generative design optimization. The effectiveness of the optimization and the changes in the structure’s rigidity and stability were assessed. Optimization resulted in an approximately 12% reduction in structural mass relative to the baseline model. Structural modeling and load simulations demonstrated that the maximum displacement of the optimized model does not exceed 3 mm, compared to 9 mm for the baseline structure. The results from virtual load simulations and physical testing of the printed models confirm that structural mass can be significantly reduced while preserving the integrity and core functional performance of the airframe. The study is based on simplified loading conditions and is limited to the number of experimental samples. Furthermore, dynamic and vibrational loads were not accounted for in the current study. Future research will focus on expanding the experimental framework and conducting an in-depth analysis of combined loading conditions that significantly affect the aerodynamic performance and real-world flight envelopes of drones.

SPURIOUS CORRELATIONS IN EXPLAINABLE AI: DETECTION, QUANTIFICATION AND MITIGATION

Oleksandr VERBYTSKYI — 2026-09-30

Modern explainable AI (XAI) methods are widely used to audit deep neural networks, yet a growing body of evidence shows that the very explanations they produce can be hijacked by spurious correlations - statistical shortcuts between non-causal input features (image background, demographic attributes, textual artefacts) and the target label. As a result, post-hoc attributions remain confidently wrong on minority sub-populations, and worst-group accuracy collapses under distribution shift. This work proposes a unified framework that treats explanation faithfulness and group robustness as a single problem.

Objective. The primary purpose of this research is to provide a rigorous, quantitative methodology for detecting which internal channels of a vision model encode spurious rather than causal evidence, measuring the robustness of the resulting explanation to spurious shift, and mitigating the failure without group annotations on the training set.

Methodology. The methodology is centred on a constrained group-loss optimisation framework that searches for a minimal channel subset whose removal eliminates dependence on the spurious attribute while preserving the target-class confidence above a threshold . The procedure - Spurious-Aware Attribution Pruning (SAAP) - combines integrated-gradients attribution, group-conditional ranking and counterfactual masking. Two new metrics are introduced: the Spurious Robustness Score (SRS) and Counterfactual Validity (CV). Validation is performed on Waterbirds and CelebA-blond using ResNet-50 and ViT-B/16 backbones.

Results. SAAP localises the spurious representation to as few as 11-14 channels per concept (0.5-0.7% of the channels in the ResNet-50 last block; for ViT-B/16 block 11). On Waterbirds, worst-group accuracy improves by 28.1 percentage points (pp) on ResNet-50 () and by 24.9 pp on ViT-B/16 (). On CelebA-blond, the gain is 38.5 pp on ResNet-50 () and 35.5 pp on ViT-B/16 (), at the cost of only 3-4 pp in average accuracy. Counterfactual Validity rises from 0.45-0.55 to 0.79-0.83, indicating that the surviving explanation tracks the causal signal.

Original contributions. First, the paper formalises spurious-correlation detection in XAI as a constrained group-loss minimisation that admits efficient greedy search. Second, it introduces SRS and CV - two complementary, attribution-agnostic metrics that quantify how much of an explanation survives a counterfactual intervention on the spurious attribute. Third, it demonstrates an end-to-end pipeline (detect mitigate audit) that does not require group labels at training time and generalises across CNN and transformer backbones.

Practical significance. The method makes black-box vision models auditable on a per-concept basis, which is critical for medical AI, autonomous driving and content moderation, where shortcut-driven errors disproportionately affect minority sub-populations. SRS and CV can be embedded into MLOps quality gates to detect regressions in explanation faithfulness before deployment.

EARLY WARNING OF CRITICAL INFORMATION SYSTEM STATES FROM SHORT SEQUENCES

Svitlana POPERESHNYAK — 2026-09-30

The reliability of modern information systems largely depends on the timely detection of changes that may lead to performance degradation, violation of established response-time requirements, and service unavailability. Conventional monitoring tools primarily register the exceedance of predefined thresholds but do not always make it possible to recognize the preceding development of an adverse trend and distinguish it from a short-term deviation. The objective of this study is to develop an approach to the early warning of critical information system states based on the sequential aggregation of structural estimates obtained from short monitoring-data sequences. The article proposes a procedure for normalizing heterogeneous indicators, forming binary sequences within sliding time windows, and calculating local structural estimates that characterize the frequency, fragmentation, and persistence of adverse observations. To detect developing degradation, the local estimates are sequentially aggregated with consideration of the current deviation level, trend, accumulation, duration, and variability. Based on the resulting aggregated estimate, information-response rules are developed using activation thresholds, confirmation intervals, and hysteresis during system recovery. The proposed approach distinguishes stable operation, an isolated deviation, gradually developing degradation, and a persistent critical state. A simulation-based evaluation was conducted for the corresponding server-infrastructure workload scenarios. The results showed that the sequential aggregation of structural estimates enables persistent adverse changes to be detected before established quality-of-service requirements are violated while limiting responses to isolated short-term fluctuations. Comparison with conventional threshold-based monitoring confirmed the possibility of providing a time interval for preventive actions, including resource scaling, workload redistribution, and restriction of noncritical requests. The proposed procedure can be integrated into information technologies for monitoring server, distributed, and cloud systems and used to support decision-making aimed at preventing critical operating conditions.

SELECTION OF DATA IN SYSTEMS AND TOOLS FOR COMPUTER ATTACK DEFENSE TAKING INTO ACCOUNT THE HISTORICAL ASPECT OF THEIR USE

Vadym PAIUK — 2026-09-30

The article considers the problem of data selection for systems and means of countering computer attacks in corporate networks in the context of constant growth of information volumes, high dynamism of information flows and limited computing resources. It is shown that the effectiveness of the functioning of modern systems for countering computer attacks largely depends not only on the quality of threat detection algorithms, but also on the timely provision of decision-making processes with relevant, relevant and most valuable data. It is substantiated that the use of the entire accumulated array of information leads to an increase in the time of its processing, an increase in information redundancy and a decrease in the efficiency of response to computer attacks, while an excessive reduction in data volumes can lead to the loss of important information necessary to identify complex or recurring threats.

A method of data selection for systems and means of countering computer attacks is proposed, which is based on the complex use of the depth of the analysis horizon, self-learning mechanisms and previous experience of the system functioning. A feature of the method is the integrated evaluation of data according to a set of criteria, which allows taking into account their current relevance, predicted usefulness in different time horizons, the results of previous use, and the ability to adaptively change the selection parameters in accordance with changes in the information environment. In contrast to existing approaches, the proposed method provides an intelligent choice of data, taking into account the relationship between the historical effectiveness of information, its prospective significance and the results of the system's self-learning.

To implement the proposed method, an architecture of the memory subsystem of systems and means of countering computer attacks has been developed, which provides accumulation, classification, optimization, modification and adaptive selection of data in accordance with the needs of means of different levels. The proposed architecture allows maintaining a balance between completeness, duration of storage and relevance of information, reducing information redundancy, reducing access time to critical data and increasing the efficiency of information support of decision-making processes.

The practical significance of the results obtained lies in the possibility of using the proposed method in the development of intelligent memory subsystems of systems and means of countering computer attacks in corporate networks, which provides an increase in the accuracy of data selection, acceleration of information analysis processes, improvement of the adaptability of systems to changes in operating conditions and increase in the efficiency of detection and neutralization of computer attacks.

Prospects for further research are the development of a mathematical model for adaptive adjustment of the weights of the data selection method, the improvement of self-learning mechanisms using modern methods of artificial intelligence, and the experimental assessment of the effectiveness of the proposed method in real corporate networks under various scenarios of computer attacks.

ARCHITECTURAL PATTERNS FOR SOFTWARE DESIGN OF INTELLIGENT SYSTEMS

Oleksandr PANASIUK — 2026-09-30

The design of software for intelligent systems increasingly extends beyond model training to encompass the full lifecycle of a machine learning (ML) component within a production software system: data ingestion, feature preparation, model serving, monitoring, and retraining. Conventional software architecture patterns, developed for deterministic systems, do not directly address the specific engineering challenges posed by ML components, namely non-determinism in outputs, dependence on data distributions that change over time, and coupling between code artifacts and data or model artifacts. This study addresses the problem of formulating a coherent architectural methodology for designing software for intelligent systems that treats the ML model as a first-class, versioned software component rather than as an external black box invoked by conventional application code. Recent research on MLOps, continuous training (CT) pipelines, microservice decomposition of inference services, and data-drift monitoring was analyzed, and the previously insufficiently addressed problem of integrating these separate practices into a single, reproducible architectural pattern was identified. A layered reference architecture for intelligent systems that separates the data layer, the model-training and versioning layer, the inference/serving layer, and the observability layer was proposed and justified. The supporting continuous integration, continuous delivery, and continuous training (CI/CD/CT) pipeline that binds these layers together was described. The main research material presents the architecture in detail and validates it with three reproducible Python experiments: a regression case study that predicts the volume electrical resistivity of carbon-filled polyethylene composites from experimental data used as a concrete intelligent-system component situated in the training and versioning layer; a discrete simulation of the observability layer's Kolmogorov-Smirnov drift monitor, which detected an injected gradual distribution shift with a measured delay of 100 requests; a discrete-event simulation of the serving layer comparing a synchronous and an asynchronous, task-queue-based design under identical offered load, showing a 136-fold reduction in 95th-percentile latency for the asynchronous design. The results indicate that separating concerns across the proposed four layers reduces the coupling between data-science experimentation and production software engineering, improves deployment reproducibility, and provides an explicit, empirically bounded mechanism for continuous training that keeps deployed models aligned with the current data distribution. These advantages, together with the architecture's current limitations, point to prospects for further research, including the extension of the architecture to support multi-model orchestration and automated rollback strategies driven by online quality metrics.

SEMIPARAMETRIC IDENTIFICATION AND UNCERTAINTY QUANTIFICATION OF BENEFICIATION UNIT SEPARATION CHARACTERISTICS

Vitalii TRON — 2026-09-30

Identifying the separation characteristics of iron ore beneficiation units from process data provides the model and uncertainty information required for model predictive control. Parametric models, including the Rosin–Rammler and sigmoid models, have fixed analytical forms that may not describe the characteristics of individual units with sufficient accuracy. Point estimation without a dedicated statistical procedure does not quantify predictive uncertainty, while purely nonparametric estimates may be unstable when data are limited. This study developed a unified semiparametric identification method in which a physically motivated parametric prior mean function was augmented with a correction estimated by Gaussian process regression from empirical deviations of the observations. The same estimation procedure was used for model fitting, cross-validation, and sensitivity analysis. A semiparametric Nadaraya–Watson estimator constructed from deviations from the prior mean function served as a benchmark. Experimental recovery was calculated using a common mass-balance definition. The ratio between product mass flow rates was either obtained from measurements or estimated from the size-by-size balance. For the classifier, this ratio was well constrained by the data, with a 95% bootstrap interval of [1.95, 2.42]. The proposed semiparametric model reduces the within-sample root-mean-square error by approximately 82% relative to the Rosin–Rammler model and outperforms the stand-alone Gaussian process model in leave-one-out cross-validation, with coefficients of determination of 0.970 and 0.945, respectively. The mass-flow-rate ratio was re-estimated in every cross-validation iteration using only the training size classes. For the hydrocyclone, the ratio was estimated with a 95% bootstrap interval of [0.96, 2.56], and the proposed model retains the highest leave-one-out coefficient of determination, approximately 0.81, throughout this interval. Results for magnetic separators with measured mass flow rates confirm that the argument of a separation characteristic must reflect the physical separation mechanism. Negative leave-one-out coefficients of determination obtained when the separation characteristic was modelled as a function of particle size demonstrate that particle size is an unsuitable argument and justify representing the characteristic as a function of iron content, which more directly reflects the magnetic separation mechanism. Monte Carlo simulation under the specified noise model was used to quantify sensitivity; for the classifier, the mean relative width of the uncertainty band is approximately 14%.

HALLUCINATION-AWARE UAV DATASET BOOTSTRAPPING FOR ISR

Yehor TERESHCHENKO — 2026-09-30

This paper studies hallucination-aware annotation for UAV-style imagery under publication-safe conditions and treats the problem as one of measurement, reliability, and deployment rather than as a small prompt-engineering demonstration. The proposed framework combines ontology normalization, overlay masking, grounded summary generation, controlled prompting, and structured metadata capture so that labels, counts, and textual descriptions can be audited jointly. To address the gap between convenient multimodal demonstrations and publishable evidence, the study combines an archival audit of 7 historical military-style frames with a public proxy benchmark of 350 aerial images and frames drawn from AerialMPT, DLR-MVDA, and DLR-ACD. The executed public benchmark contains 274,581 ground-truth objects, moving the empirical basis of the work from a narrow pilot into a hundreds-scale evaluation. The local YOLO-World baseline completed all 350 public samples with mean latency 0.1807 s per image, but it under-counted dense scenes severely, with mean absolute count error of 143.889 on AerialMPT, 297.300 on full DLR-MVDA scenes, and 6848.242 on DLR-ACD. For a matched cloud-versus-local comparison, the paper introduces a dense-tile MVDA benchmark with 10 derived 1024x1024 crops and 315 labeled vehicles. A three-profile Gemini Robotics-ER ablation shows that prompt design materially changes outcome quality: the generic profile completed all tiles but over-counted heavily, the constrained ISR profile reduced error but dropped one request, and the best profile, coarse-to-fine, achieved MAE 5.300, MAPE 10.161%, and 10 / 10 completion on this slice. A supplementary video study on publicly available tactical footage then isolates text control: the freeform condition produced ontology expansion in 21 of 21 non-empty summaries, whereas the matched grounded condition reduced expansion to 0 of 20 while cutting latency from 491.070 s to 272.711 s. The archival audit further reveals unsupported unit, status, or intent language in 4 of 7 historical frames (57.1%) and four raw label variants (mil targets, military target, military targets, vehicle) for the same armored-platform concept. The contribution of the paper is therefore an experimentally grounded framework for hallucination-aware UAV dataset bootstrapping that reports semantic control, counting quality, prompt sensitivity, and execution robustness within one evaluation design.

COMPRESSIVE LOGARITHMIC GENERATING ELEMENTS FOR INERTIALESS FILTERING OF SYMMETRIC NOISE

Pavlo KLOPOTOVSKYI — 2026-09-30

We develop a polynomial inertialess filter for recovering a Gaussian signal from additive non-Gaussian noise within the Kunchenko decomposition framework, taking the generating element to be a compressive, sub-polynomial family: after the mandatory linear element , the odd, monotone elements . Because these elements grow more slowly than any power of , they compress rather than amplify the tails, and the normal system is finite whenever the noise has finite variance alone – no moment above the second is ever required, so the body cannot «diverge» on heavy tails. The cost of this generality is numerical: successive powers of a slowly varying function are strongly collinear, so the body is severely ill-conditioned and is solved under Tikhonov regularisation with a cross-validated ridge. We show that this controlled cost buys a measurable advantage on the noise regime to which the compressive shape is matched. Monte-Carlo experiments with 95% bootstrap confidence intervals, benchmarked against the Wiener linear filter through the Kunchenko coefficient , show the logarithmic basis to be the better filter on platykurtic / multimodal noise, where its slowly growing nonlinearity resolves inter-mode structure that the linear filter does not use ( = 0.851 at =0.5, widening to 0.729 – a 27% reduction in residual variance over the linear filter – at q = 0.25). A controlled experiment on a real CWRU bearing carrier with a platykurtic disturbance reproduces this finding ( = 0.853 on a bimodal disturbance at q = 0.5), whereas raw real noise with negative excess kurtosis alone gives no advantage – i.e. the gain is driven by the disturbance’s structure, not by its kurtosis. Because the elements are odd and sign-preserving, the mechanism is matched to symmetric disturbances as a class: it improves on the linear filter on all three symmetric classes examined, and is only marginal on the asymmetric one, where odd compression captures skew inefficiently. We therefore recommend the logarithmic basis for symmetric non-Gaussian disturbances, most strongly on their platykurtic and multimodal corner, and we note that negative excess kurtosis on its own is not a sufficient condition for an advantage.

COMPUTER MODELING OF WAREHOUSE LOGISTICS PROCESSES: TECHNOLOGY OF RATIONAL SLOT SELECTION FOR PICKER’S ROUTE MINIMIZATION

Tetiana KYSIL — 2026-09-30

The article is devoted to solving scientific and practical problems of increasing the efficiency of logistics systems by minimizing the duration of material handling equipment (MHE) routes using the example of picking orders in warehouses with shared storage. The order picking process is one of the most labor-intensive and expensive operations, as the movement of staff or machinery accounts for over 50 percent of the total picking time. When goods are scattered throughout the warehouse, an additional problem arises regarding the rational selection of slots for picking goods in accordance with the order. To ensure the rational selection of locations, the PIV (Proximity Indexed Value) multi-criteria decision-making method was applied in this study. The locations of a particular item are considered as alternatives, each characterized by different criteria. Three criteria have been selected. To evaluate the alternatives (slots) according to the criterion 'degree of satisfaction of demand for other items in the order', it is proposed to take into account the potential of adjacent locations.

A hypothesis has been formulated regarding the impact of sorting goods within orders for the subsequent application of the PIV method. A heuristic for the preliminary ranking of items within orders has been justified and proposed, based on the scarcity of alternative locations in the warehouse, thereby implementing the principle of 'limited choice'. To verify the hypothesis, a three-dimensional simulation model of the goods storage area of the sorting center with a capacity of 360 unique slots was developed. Computer modeling in the MATLAB environment allowed us to recreate the topological labyrinth of the warehouse, record the coordinates of the aisles between racks and the operational routes, and ensure that virtual equipment moved strictly along the center lines of the aisles via designated access points. As part of experimental research, a comparative analysis of route construction using the nearest neighbor algorithm for two scenarios was performed: classical processing of the consignment note “top-down” and processing according to the proposed principle of “limited choice”. The results of a series of computer experiments for order waves of 5 to 15 items confirmed the validity of the hypothesis and showed a reduction in the total duration of the material handling equipment route by 2–9% when processing an order according to the principle of limited choice.