The International Conference on Security, Fault Tolerance, Intelligence

A NAVIGATION METHOD FOR AUTONOMOUS UNMANNED VEHICLES UNDER INTERFERENCE CONDITIONS

Vladyslav Shatrov — 2026-05-08

Autonomous unmanned aerial vehicles require reliable navigation when satellite measurements are degraded, blocked, or intentionally falsified. The relevance of this research is determined by the growing use of UAVs in autonomous missions where GNSS signals may be affected by jamming, spoofing, multipath propagation, or limited satellite visibility. The objective of the paper is to develop a navigation method that preserves a physically plausible UAV trajectory under interference conditions. The proposed approach combines IMU/INS prediction, controlled GNSS correction, adaptive Kalman filtering, and semantic map matching. GNSS measurements are used only when they are consistent with the predicted inertial state. In the case of jamming, the estimator temporarily relies on INS propagation. In the case of spoofing, suspicious measurements are rejected according to abnormal innovation values. To reduce long-term drift, semantic information from UAV camera data is compared with reference map layers, including roads, shorelines, and buildings. Simulation results show that pure inertial navigation accumulates significant positioning error, whereas the adaptive fused estimate remains bounded during jamming and spoofing intervals. The proposed method improves navigation resilience by combining radio navigation, inertial estimation, and visual map constraints. Further research should focus on real flight validation and real-time implementation.

ARITHMETIC CIRCUIT COMPLEXITY OF LEGACY CRYPTOGRAPHIC STANDARDS AS A BARRIER TO ZERO-KNOWLEDGE CREDENTIAL VERIFICATION

Mykhailo Sokolov — 2026-05-08

The study addresses the issues that prevent the implementation of confidential document verification. The first issue is the outdated cryptographic algorithms used by the EDEBO registry (GOST 34.311-95 and DSTU 4145-2002), for which the number of R1CS constraints in the Groth16 system was calculated for the first time. For a 10 KB document, the hash function alone generates over 19 million constraints (~15 minutes to generate a proof), and signature verification generates over 140 million. For comparison, SHA-256 and ECDSA P-256 together require around 2 million constraints (60–120 seconds). The second problem is that the PDF/XML document format does not support selective field exposure, which is an independent obstacle regardless of the algorithms. A transition to SHA-256/ECDSA or “Kupina” and the structured JSON-LD format is proposed.

A GENERALIZED GL-MODEL OF FAILURE BEHAVIOR OF FAULT-TOLERANT MULTIPROCESSOR SYSTEMS WITH INTER-SUBSYSTEM RESOURCE SHARING

Kostiantyn Morozov — 2026-05-08

This paper addresses the problem of constructing models of failure behavior for complex fault-tolerant multiprocessor systems. In particular, systems consisting of several subsystems are analyzed, where the resources of some subsystems (donors) can be used to maintain the operation of others (recipients). Such models can be used to evaluate system reliability parameters by means of statistical experiments. GL-models, which combine the properties of graphs and Boolean functions, are used for modeling. Existing methods for constructing such models assume that the donor subsystem is always operational, which limits the modeling capabilities. The aim of this work is to develop a method for constructing GL-models that takes into account the use of operational processors of donor subsystems both when they are operational and after their failure. The study employs methods of graph theory and Boolean algebra. The proposed method is based on known approaches to constructing GL-models and involves the use of auxiliary models to determine the number of available processors. The obtained results show that the developed model adequately represents system behavior under various failure scenarios. Experimental studies confirm the correctness of the proposed approach and demonstrate that the model complexity is comparable to that of basic models for systems of the same size. The practical significance of the work lies in the application of the proposed approach to reliability analysis of complex systems. Future research directions include extending the method to systems with more complex subsystem behavior and reducing the computational complexity of the models.

SAFETY MEASURES IN PEER-TO-PEER NETWORKS

Herman Kaidaniuk — 2026-05-08

In today's world, computer networks are one of the primary sources of information and communication. As a result, increasing attention is being paid to security when using computer networks. Among computer networks, peer-to-peer networks offer the highest level of anonymity and security compared to alternatives, as this type of network lacks a centralized server and, consequently, a single point of failure. Although peer-to-peer networks have clear advantages over traditional client-server architecture, they are still subject to a large number of potential risks. The purpose of this article is to examine the main modern types of attacks and risks in peer-to-peer networks and ways to prevent or combat them. The primary research method involved reviewing currently known types of attacks and the main countermeasures against them. The article also examines security condition in currently popular P2P networks. As a result, the main risks associated with peer-to-peer networks, the methods that maximize anonymity and security, and how modern systems address these issues were examined.

Key words: peer-to-peer network, cybersecurity, distributed systems, Sybil attack, eclipse attack, botnet

Fig.: 1. Bibl.: 6.

COMPARISON OF TCP CONGESTION CONTROL ALGORITHMS IN NETWORKS WITH DELAY AND LOSS

Yaroslav Fedoriachenko — 2026-05-09

This paper compares TCP Reno, CUBIC, BBR, and Vegas under controlled delay, jitter, and loss profiles, and separately evaluates BBRv1 and BBRv3 in a virtualized kernel testbed. The measurements use Linux tc netem, iperf3, and ping. The results show that BBR is a strong throughput-oriented baseline in satellite-like profiles, whereas Vegas better controls delay and retransmissions. BBRv3 reduced RTT in selected scenarios but had lower throughput than BBRv1. Additional repeats show that TCP measurements are sensitive to the runtime environment; therefore, kernel-version comparisons should be interpreted within the same controlled environment.

INTERACTIVE MOBILE SYSTEM FOR STIMULATING CUSTOMER ACTIVITY IN RETAIL ESTABLISHMENTS

Yaroslav Berlinskyi — 2026-05-08

The article presents the development of an interactive mobile system for stimulating customer activity in retail establishments. The system is aimed at increasing user engagement in physical shopping locations by combining geolocation, augmented reality, gamification, rewards, and role-based management tools for businesses. The relevance of the topic is driven by the transition of mobile applications in retail from static loyalty cards to behavioral platforms that connect digital incentives with real visits to retail locations.

The proposed solution is implemented as a native iOS application with a backend, a relational database, geospatial processing, interaction with AR objects, a reward system, and analytics. The client side includes a location map, AR session, rewards interface, user progress, achievements, and a business dashboard. The server side verifies visits, controls reward distribution, processes AR object collection events, stores promo codes, and aggregates analytical data.

The practical result is a prototype that supports both customer and business scenarios: browsing locations, entering geozones, collecting AR objects, receiving rewards, managing campaigns, and analyzing location activity. Testing covered authentication, role-based routing, AR interaction, reward logic, gamification, and core domain models.

VISUAL COORDINATION METHOD FOR AN AUTONOMOUS UNMANNED AERIAL VEHICLE FORMATION

Anastasiia Danevych — 2026-05-08

This paper addresses the problem of maintaining stable visual coordination in UAV formations with a leader-follower topology under conditions of total target occlusion and the absence of any data input other than visual observation. Although modern deep learning-based predictors demonstrate high accuracy, their computational complexity often limits their deployment on resource-constrained onboard systems. I propose a lightweight coordination framework that combines YOLO (You Only Look Once)-based detection with an Adaptive Kalman Filter (AKF), integrated into the ROS 2 and PX4 ecosystem. The core of the method lies in a dynamic scaling mechanism for the process noise covariance, which allows the system to transition into an inertial prediction mode during the loss of visual contact.

Experimental results obtained in the Gazebo simulation environment demonstrate that the proposed AKF reduces positioning Root Mean Square Error (RMSE) by 50% during non-linear maneuvers compared to standard estimation methods. The system achieved a 97% success rate in maintaining formation in cluttered environments with occlusions lasting up to 5 seconds. These results confirm that the developed method provides a reliable and computationally efficient solution for autonomous UAV operations in complex urban settings without relying on external positioning systems.

BIOMETRIC ACCESS CONTROL FOR BEAGLEBONE BLACK WITH SECURED SSH AND LOCAL WEB INTERFACE: EVENT-DRIVEN SERVICE MANAGEMENT

Dmytro Voznytsia — 2026-05-09

IoT and embedded Linux devices often keep remote access enabled permanently, i.e., they remain continuously exposed to attacks. This paper proposes a different default: SSH does not exist on the network until the user passes local biometric verification. This paper substantiates an approach to reducing the attack surface. It is based on event-driven control of SSH availability with verification of behavior in idle, session, and teardown modes. The prototype is built on BeagleBone Black with an R307 fingerprint sensor and a minimal Yocto-based Linux image. A C control daemon (finite-state machine) enables sshd.socket and a temporary firewall rule only for an authorized session. The web panel with a secrets vault is bound to 127.0.0.1 and is reachable only via an SSH tunnel. Secrets are stored locally using PIN-based key derivation (PBKDF2-HMAC-SHA256) and Fernet encryption. Experiments confirm TCP/22 invisibility in idle mode, controlled access during the session, and an adequate security level for sensitive information storage.

FAST FOURIER TRANSFORMATION BASED FINITE IMPULSE RESPONSE FILTER

Dmytro Bets — 2026-05-09

This article examines the challenges of hardware implementation of high-order finite impulse response (FIR) digital filters. The relevance of this topic stems from the fact that the direct implementation of FIR filters requires a quadratic number of multiply-accumulate operations, O(N2). This leads to the rapid exhaustion of specialized computational resources on the chip and causes unacceptable delays in hardware implementation. The aim of the study is to solve the problem of computational complexity of traditional FIV filters by moving to the frequency domain using the Fast Fourier Transform algorithm. The research methods are based on the fundamental theorem, which states that convolution in the time domain is equivalent to the pointwise product of spectra in the frequency domain. The use of the FFT allows for a significant reduction in computational complexity to O(Nlog N ).

It has been proven that the use of the FFT effectively solves the problem of computational complexity of FIR filters. The application of overlapping algorithms ensures reliable block-by-block processing of infinite data streams. It has been established that the hardware implementation of such systems on FPGA guarantees architectural flexibility and the possibility of massive parallelism, which traditional DSP microprocessors lack. As a result, FFT-based FIR filters remain the standard for developing high-performance digital signal processing systems.

HARDWARE ARCHITECTURE OF A DISTRIBUTED CONTROL SYSTEM FOR A ROBOTIC PLATFORM IN THE ROS2 ECOSYSTEM

Danylo Lytvynenko — 2026-05-09

Modern mobile robotic platforms require reliable control systems capable of processing sensory data and generating commands in real time without performance loss. The use of centralized computing architectures often leads to resource contention between high-level processes and critical drive control tasks. The objective of the study is the design and hardware implementation of a decentralized computing architecture for a differential drive robotic platform. The proposed system is based on a strict hierarchical distribution of computations across three levels. The Odroid M1S microcomputer (upper level) provides general node coordination within the ROS2 ecosystem. Two ESP32 controllers (middle level) are responsible for generating PWM signals for brushless DC (BLDC) motors and processing IMU and GPS data, integrating into the overall network via a micro-ROS bridge. The Arduino Uno microcontroller (lower level) isolatedly executes hard real-time tasks-continuous odometry collection from Hall sensors via hardware interrupts. Communication between levels is implemented through UDP and UART transport protocols using COBS encoding. The developed prototype confirms that the physical separation of tasks eliminates control signal transmission delays, minimizes the risk of cumulative localization errors, and improves the overall fault tolerance of the robot's control system.

OPTIMIZATION OF THE APPLICATION-LEVEL CHECKPOINTING INTERVAL IN STATEFUL MICROSERVICES

Bohdan Marchuk — 2026-05-09

This paper addresses the complex scientific and applied problem of enhancing the operational readiness of stateful microservice architectures. The foundation of the proposed approach is the fine-grained optimization of the chronological frequency of checkpoint generation directly at the application abstraction level. The study aims to formulate a mathematical model and subsequently empirically verify the critical state fixation interval that guarantees a strict balance between context rehydration kinetics and associated infrastructural overhead. An interdisciplinary methodological framework is employed, integrating systems analysis methods, classical rollback recovery theory, and modern chaos engineering tools deployed within a Kubernetes ecosystem. The destructive impact of a hyperactive caching strategy is proven. An aggressive fixation periodicity of 5 seconds induces anomalous CPU load (5.77%) and provokes the exponential accumulation of objects in the garbage collector memory (5.81 MB), degrading the aggregate recovery time to 0.795 s. Conversely, the experimental basis confirms that a calibrated window of 10 - 15 seconds radically neutralizes the intensity of background perturbations. This balanced mode not only ensures a minimal latent period of application recovery (0.503 - 0.563 s) but also strictly limits the amnesia interval - the time gap between the last successful checkpoint and the moment of failure - restricting the volume of event replay to strictly manageable limits (34 - 47 events). The derived regularities form the theoretical foundation for designing highly resilient systems with deterministic cold start characteristics, completely eliminating the probability of operational degradation under standard load conditions.

IOT INFRASTRUCTURE FOR POWER GRID CONTROL AND FACILITY SAFETY

Serhii Bondarenko — 2026-05-09

This article examines the pressing issue of ensuring industrial and energy safety at both private and commercial real estate properties in the context of an unstable power supply caused by increased load on the power grid. The main methods for reducing the risks of accidents and critical failures through the integration of an IoT system for monitoring key indicators within premises are investigated.

EVALUATION OF THE ROBUSTNESS OF MULTIPLE OBJECT TRACKING METHODS TO PARTIAL OCCLUSIONS IN VIDEO SEQUENCES

Mykhailo Pinskyi — 2026-05-09

The article addresses the problem of robust multiple object tracking in video sequences under partial occlusions, short-term detection loss, and object scale reduction. The relevance of the study is determined by the fact that in real video data, objects are often partially hidden by scene elements, overlaid graphical objects, or other moving objects, which may lead to track loss and identity changes. The objective of the paper is to evaluate the robustness of multiple object tracking methods to partial occlusions using ByteTrack and BoT-SORT as examples. The experimental study is based on the tracking-by-detection approach with YOLOv8 used as the object detector. Tracking quality is evaluated using detection rate, failure rate, number of unique track identifiers, main track length, average track length, and track fragmentation indicators. The obtained results show that BoT-SORT produces fewer unique tracks and a longer main track, while ByteTrack demonstrates a slightly lower number of internal track segments. It is concluded that BoT-SORT is a suitable main option for further research on object tracking under partial occlusions.

A METHOD FOR INTELLIGENT PROCESSING AND RETRIEVAL OF MULTIMEDIA DATA

Illia Kniaziev — 2026-05-08

To overcome the limitations of traditional media archives in natural language search, the study proposes a multimodal Retrieval-Augmented Generation architecture for indexing and retrieving visual data. The proposed method combines dense vector search with lexical Best Matching 25 ranking by decomposing automatically generated image captions into atomic sentences. The entire software complex is implemented within a single PostgreSQL instance. Experiments demonstrate that this hybrid approach significantly outperforms baseline vector search, particularly for complex queries with rare proper nouns and vague spatio-temporal hints.

METHOD FOR STATISTICAL DROPOUT RISK ASSESSMENT IN PSYCHOLOGICAL THERAPY BASED ON DE-IDENTIFIED SESSION DATA

Maksym Zaryshniak — 2026-05-08

The paper proposes a method for statistical assessment of the risk of premature termination of psychological therapy, operating exclusively on de-identified session data without access to clients' personal information. The relevance of the research is driven by the rapid growth of the online therapy market and the strengthening of regulatory requirements for the protection of sensitive medical data, particularly GDPR provisions that limit the application of traditional prediction methods based on clinical information. The developed Dropout Risk Score method integrates three independent behavioral signals – miss rate, duration trend, and inter-session interval z-score – into a weighted composite score with defined classification thresholds across three risk levels. A distinctive feature of the method is the use of personalized baselines instead of population norms, which provides increased sensitivity for clients with non-standard attendance patterns. Additionally, an anomaly detection method based on personalized interval analysis is described, and a system architecture with physical separation of identification and analytical layers is proposed for deploying the method under personal data protection requirements. The method does not require large training datasets and ensures interpretability of results for practitioners.

REDUCING THE SEARCH SPACE IN SYMBOLIC PLANNING USING NEURAL COST ESTIMATION

Anton Yasnov — 2026-05-08

The objective of this work is to reduce the search space in GOAP by integrating a neural cost estimation model. A symbolic GOAP-based planner is implemented and extended with a neural evaluation component that adds an additional cost term based on state and action features. A simulation environment is developed to evaluate the proposed approach.

EVALUATION OF THE ROBUSTNESS OF MULTIPLE OBJECT TRACKING METHODS TO PARTIAL OCCLUSIONS IN VIDEO SEQUENCES

Mykhailo Pinskyi — 2026-05-09

METHOD OF ADAPTIVE CRM PROCESS OPTIMIZATION BASED ON SCORING AND STATUS AUTOMATION

Serhii Korniichuk — 2026-05-08

The article presents a method of adaptive CRM process optimization aimed at improving the efficiency of customer request and lead processing. The relevance of the study is determined by the fact that in many CRM systems, key decisions related to lead priority, status updates, and further actions are still made manually by managers. This may cause processing delays, inconsistent decision quality, loss of potential customers, and limited transparency of the business process. The objective of the paper is to develop a method that combines CRM process formalization, lead scoring assessment, and automated status transition based on predefined rules. In the study, the lead management process is represented as a sequence of states, evaluation criteria, and transitions between statuses. An integrated scoring model is proposed, taking into account data completeness, customer activity, time factor, and potential value of the request. The practical validation of the method was carried out through a CRM software prototype based on a client-server architecture, REST API, and an optimization module. The obtained results indicate that the use of scoring and status automation can reduce the time required for initial lead processing, decrease dependence on subjective decisions, and improve the transparency of CRM-oriented business processes.

METHOD OF ATTENDANCE ANALYSIS TAKING INTO EXTERNAL FACTORS IN TIME ACCOUNTING SYSTEMS

Anatolii Khramchenko — 2026-05-08

The article considers a method of attendance analysis in time accounting systems that takes into account external factors and assesses their impact on the deviation of events in time using statistical methods and neural networks.

ARCHITECTURE OF A SECURE DIGITAL REPOSITORY SYSTEM WITH AUTOMATED METADATA GENERATION AND ROLE-BASED ACCESS CONTROL

Tymofii Antonenko — 2026-05-08

This paper focuses on developing a modern architecture for electronic libraries that effectively combines artificial intelligence tools and a high level of information security. A digital repository system is proposed where smart algorithms autonomously analyze uploaded PDF documents and generate metadata (summaries, keywords, authors) for them. This eliminates the routine work of administrators. Furthermore, to address the issue of copyright protection, the system implements Role-Based Access Control (RBAC) based on JWT tokens. It automatically determines whether a user can download a book, considering its license type and the client's current IP address. The implementation of vector search through the PostgreSQL database significantly improves literature retrieval based on its content.

METHOD OF CLASSIFYING IMAGES OF MOLES FOR DETECTING MELANOMA

Kristina Hrabenko — 2026-05-08

The article presents method for automatic malignant melanoma detection based on dermoscopic images. The subsystem is built on the EfficientNet-B0 convolutional neural network for classifying skin lesions as benign or malignant, while the Grad-CAM method is used to visualize and interpret the model’s decisions.

A METHOD OF IMPROVING LLM INFERENCE EFFICIENCY VIA CASCADE ROUTING BASED ON SEMANTIC ENTROPY EVALUATION

Vasyl Khrapko — 2026-05-08

The article presents a method for improving the inference efficiency of large language models using cascade routing based on semantic entropy evaluation to address the issue of high costs and latencies of modern LLMs, which is highly relevant due to the rapid growth of LLM deployment in real-world production and commercial systems. It reveals the working principle, which involves automatically determining the cognitive complexity of the input query using a combined semantic entropy metric and dynamically redirecting it to the most optimal model of the corresponding efficiency and power level. A comparative analysis with existing routing methods and a verification of this method's functionality were conducted on a test dataset of queries.

COMPARATIVE ANALYSIS OF PERSONALIZED RESPONSE GENERATION METHODS BASED ON MESSENGER CORPUS

Bohdan Utenko — 2026-05-08

This paper presents a comparative study of four approaches to automated personalized text response generation in the communication style of a specific individual: TF-IDF retrieval, Markov chains (bigram model), and two fine-tuned neural language models – DistilGPT-2 (82M parameters) and DialoGPT-medium (345M parameters). The training corpus consists of Ukrainian-language private Telegram chat logs (~200,000 messages), from which 10,000 query-response pairs were extracted in an 80/20 train/test split. Neural models were fine-tuned on the Kaggle platform using an NVIDIA T4 GPU; inference and evaluation were performed on CPU. Evaluation was conducted on 500 test pairs across five metrics: BLEU, ROUGE-L, BERTScore-F1, average inference time, and peak RAM consumption. DistilGPT-2 achieved the highest BLEU (0.0088) and ROUGE-L (0.0535) scores and the best BERTScore F1 (0.687), while DialoGPT-medium underperformed across quality metrics despite consuming significantly more memory (Net RAM: 1343 MB vs. 297 MB). TF-IDF retrieval is viable for minimal hardware, and Markov chains are the fastest method (70 µs per query), but both lag behind neural approaches in semantic coherence. The findings indicate that fine-tuned DistilGPT-2 offers the best quality-to-resource trade-off for style personalization in a local CPU deployment scenario.

IMPROVING PORTRAIT STYLE TRANSFER

Pavlo Kopka — 2026-05-09

The paper presents an approach to improving artistic style transfer for portrait images by explicitly accounting for facial geometry and semantic heterogeneity of face regions. The proposed GeometricBiasAttention and SemanticGatedCrossAttention modules are integrated into the transformer-based architectures StyTr2 and S2WAT. Experiments on FFHQ and WikiArt Faces demonstrate improved identity preservation and more controllable stylization. The modified models also show that a smaller configuration can outperform a larger baseline one across several quality metrics, including FID, SSIM, and LPIPS.

AN INTEGRATED APPROACH TO ANALYZING COURSE STRUCTURE USING LLM AND PREREQUISITE GRAPHS

Yehor Hrybenko — 2026-05-08

Educational materials in university courses exhibit a hierarchical structure, where the understanding of new concepts is based on previously introduced ones. During the creation, revision, or integration of materials from different sources, structural inconsistencies may arise, including incorrect ordering of concept introduction, conflicting definitions, isolated concepts, and cyclic dependencies, which reduce the coherence and consistency of the course. Detecting such inconsistencies requires simultaneous consideration of textual semantics and the structure of relationships between concepts.

This paper proposes an integrated method for analyzing the structure of educational materials, combining the use of large language models for extracting concepts, their definitions, introduction time, and prerequisite relationships with the construction of a temporal knowledge graph and subsequent graph-based algorithmic analysis. The method is implemented as a four-stage pipeline, including material structuring, LLM-based extraction, entity normalization, and graph integration, followed by structural analysis using transitive checking, semantic conflict detection, and importance ranking based on PageRank centrality. Additionally, a precision-oriented mode is employed, incorporating validation of results using a language model.

Experimental evaluation was conducted on MIT OpenCourseWare datasets using controlled injection of structural inconsistencies and three evaluation protocols. The proposed method demonstrates significantly higher detection performance (F1 up to 0.738) compared to baseline approaches, including single-prompt language model and TF-IDF. It is also shown that a substantial portion of formally identified false positives corresponds to real structural issues in the course.

The scientific contribution lies in integrating concept and relation extraction using large language models with temporal knowledge graph construction and algorithmic analysis of structural consistency in educational materials.

CLASSIFICATION OF METRICS FOR PROACTIVE ANOMALY DETECTION IN DISTRIBUTED SYSTEM NODE LOADS

Oleh Dubynskyi — 2026-05-08

The paper considers the problem of node state monitoring in high-load distributed systems based on the Actor Model. The disadvantages of reactive load balancing methods are analyzed. A multi-level classification of metrics (system, platform, application) is proposed, the monitoring of which allows implementing proactive migration of computational entities before a critical failure or performance degradation occurs.

ACTION SPACE REDUCTION IN REINFORCEMENT LEARNING USING THE MAXIMUM WIDTH OF THE OPERATION GRAPH FOR FLEXIBLE SCHEDULING PROBLEMS

Oleksii Pysarchuk — 2026-05-08

The paper addresses approach of action space reduction in reinforcement learning for flexible scheduling. The relevance of the study is determined by the large number of operation-machine alternatives in scheduling environments, where only a limited subset of actions is feasible at each decision step. This creates redundant action evaluation and increases inference computational cost. The objective of the paper is to evaluate whether feasibility-aware action processing can improve the quality and inference time trade-off of reinforcement learning policies for scheduling. Three variants are compared: a three-branch MLP baseline (BR), a branch-aware scheduling policy with candidate action scoring (BASP), and BASP Sparse with feasibility-aware reduced action-space evaluation. The experiments are conducted on Brandimarte and Behnke benchmark instances. Schedule quality is measured by relative makespan reduction with respect to BR, while inference efficiency is evaluated by inference-time ratios. The results show that BASP improves makespan but increases inference time compared with BR. BASP Sparse preserves most of the quality gain while significantly improving inference efficiency. On Brandimarte groups, BASP Sparse achieves up to 5.95% makespan reduction and up to 2.21 speedup relative to BASP. On the larger Behnke group, it achieves 28.21% makespan reduction and 3.18 speedup relative to BASP. The results indicate that proposed action space reduction approach becomes especially beneficial for larger scheduling instances.