Journal of Problems in Computer Science and Information Technologies

APPLYING SELF-PLAY REINFORCEMENT LEARNING TO KAZAKH NATIONAL BOARD GAMES: A CASE STUDY ON TOGYZKUMALAK

Dinara Zhussupova — Fri, 19 Jun 2026 00:00:00 +0000

The paper presents the results of a study on training agents for competitive board games, using the Kazakh national Togyzkumalak game, a zero--sum game, as a case study. As demonstrated by Alpha Zero, it is possible to train a champion without human expert knowledge or supervised learning. However, each game has unique characteristics that require specific research to develop an excellent player. In this study, two approaches were used to train agents: model-free and model-based. The self-play technique was used effectively during the training. As a result, the agents developed the ability to play Togyzkumalak at a competitive level.

PRE-FILTERING OF GRID MAPS WITH SKELETONIZATION AND MORPHOLOGY METHOD FOR EFFICIENT PATH PLANNING FOR MOBILE ROBOTICS

Abdulla Sapargali, Khansa Arshad, Muhammad Ilyas — Fri, 19 Jun 2026 00:00:00 +0000

SLAM-based occupancy grid maps are widely used for indoor mobile robots, but their noise, jagged obstacle borders, and spurious narrow gaps can mislead the classical grid-based path planners, e.g. A* algorithm, causing unstable routes, edge-hugging behavior, and increased computation. This manuscript proposes a lightweight pipeline to improve global path planning robustness and speed in warehouse-like environments. The approach first converts the SLAM map into a conservative binary representation and applies morphological refinement to remove small artifacts and smooth free-space connectivity. Next, the free space is skeletonized and transformed into a graph that captures the topology of navigable corridors. Global path planning is then performed using a hybrid strategy: a local grid search connects the start and goal regions to the skeleton, while the main route is computed on the skeleton graph. Experiments in two warehouse scenarios show that the proposed method substantially reduces node expansions and path planning time compared to full-grid A* algorithm, while producing more physically plausible paths that avoid false passages in noisy maps. These results indicate that morphological and skeleton-based global path planning can serve as a practical solution to enhance efficiency for robust navigation in realistic SLAM based path planning.

LOW-COST NIGHT VISION CAMERA BASED ON A MODIFIED CONSUMER WEBCAM AND LIGHTWEIGHT REAL-TIME ENHANCEMENT

Fri, 19 Jun 2026 00:00:00 +0000

Low-light perception is a core requirement for robotics, inspection, and wearable situational awareness, yet reliable night- vision solutions often require expensive sensors or computation- ally heavy processing. This manuscript presents (i) a low-cost near- infrared (NIR) night-vision camera system with real-time enhancement and mixed-reality visualization, and (ii) a controlled benchmarking protocol that compares multiple sensor modalities and classical en- hancement pipelines on a Linux embedded platform. To prioritize real-time feasibility, we deliberately avoid neural networks and other AI/ML models, and instead benchmark lightweight classical methods: CLAHE, bilateral filtering + CLAHE, Non-Local Means (NLM) + CLAHE, Retinex SSR, Retinex SSR with percentile normalization, and two proposed pipelines (a CLAHE-based pipeline and a Retinex integrated variant). Experiments were conducted in a fully dark room using identical illumination conditions and multiple observation distances. We report runtime (ms/frame, FPS) and no- reference quality metrics (entropy, edge strength, Laplacian variance, RMS contrast, mean intensity) to accommodate the absence of ground-truth references. A representative run shows that the proposed CLAHE-based pipeline achieves 246.5 FPS at 640×480 while substantially improving objective no-reference metrics relative to raw frames, whereas Retinex with percentile normalization yields higher contrast/entropy but at 12.7 FPS. Hardware benchmarking indicates that thermal imaging provides the clearest visibility but at high cost/power, Kinect requires stronger illumination, and event-based sensing offers extremely low latency but requires temporal modulation to observe static targets.

COMPUTATIONAL MODEL OF KAZAKH VOWEL–CONSONANT HARMONY

Zhansaya Segizbayeva, Marek Miłosz — Fri, 19 Jun 2026 00:00:00 +0000

This study presents a computational model of morphological generation based on the agglutinative nature of the Kazakh language and vowel–consonant harmony. In this work, we model the Kazakh vowel–consonant harmony system, which governs allomorphic selection of suffixes based on both vowel and consonant properties. Using a rule based approach, the model generates verb and noun forms across various grammatical categories (tense, aspect, case, possession). Experimental results demonstrate the model’s high efficiency and full compliance of the generated word forms with vowel harmony rules. Specifically, thousands of combinations were created for derivative words (DTL), possession (POSS1, POSS2), and case (CASE) categories. This research makes a significant contribution to the development of natural language processing (NLP) morphological analysis and automatic translation systems for the Kazakh language.

COMPARATIVE ANALYSIS OF SEQUENCE-BASED AND GRAPH-BASED DEEP LEARNING FOR ANOMALY DETECTION IN MICROSERVICE TRACES

Yesset Zhussupov , Ainur Zhumadillayeva, Shona Shinassylov, Ainur Amirtayeva — Fri, 19 Jun 2026 00:00:00 +0000

The transition from monolithic to microservice architectures has significantly increased the complexity of system observability. Distributed tracing has emerged as a vital tool for monitoring inter-service interactions, yet traditional anomaly detection methods often fail to capture the intricate structural relationships inherent in these systems. This article presents a comprehensive comparative analysis of sequence-based machine learning approaches, specifically Long Short-Term Memory (LSTM) networks, with Graph-based Deep Learning methods such as Graph Neural Networks (GNNs) for anomaly detection in microservice traces. We evaluate these paradigms across three critical dimensions: structural awareness, detection accuracy, and computational efficiency. Our analysis, drawn from recent frameworks including DeepTraLog and TraceVAE, demonstrates that graph-based models provide superior structural fidelity and achieve higher precision-recall metrics while maintaining scalability for high-dimensional telemetry data. The findings indicate that GNN-based approaches achieve F1-scores up to 0.954, significantly outperforming sequence-based alternatives which typically achieve F1-scores around 0.741. These results have important implications for the design of next-generation observability platforms for cloud-native applications.

ENHANCING WIND SPEED FORECASTING WITH LARGE LANGUAGE MODELS: A CASE STUDY OF KAZAKHSTAN

Fri, 19 Jun 2026 00:00:00 +0000

Accurate wind speed forecasting plays an important role in the planning, operation, and control optimization of modern wind energy systems. Kazakhstan, the country with the largest wind energy potential in Central Asia, is facing significant challenges due to the highly variable, nonlinear, and non-stationary nature of wind. This study proposes a wind-speed forecasting framework based on frozen pre-trained Transformer backbones (BERT/BART) with lightweight projection-based adaptation for time-series inputs. The study employs frozen pre-trained Transformer backbones with self-attention and lightweight projection-based adaptation of Bidirectional Encoder Representations from Transformers (BERT) and Bidirectional and Auto-Regressive Transformers (BART) as the core of the model. In this retrospective analysis, hourly wind speed data from five representative cities across Kazakhstan are used to evaluate simulated operational forecast performance at seven time steps: 1, 3, 6, 36, 72, 144, and 432 hours. The LLMs were compared with AutoRegressive Integrated Moving Average (ARIMA), Segmented Recurrent Neural Network (SegRNN), and Patch Time Series Transformer (PatchTST) models, and the results showed superior accuracy and higher stability in long-term forecasting. These results support the potential of pre-trained Transformer backbones as effective sequence models for wind-speed forecasting under diverse climatic conditions.

APPLICATION OF MACHINE LEARNING METHODS FOR PHISHING ATTACK DETECTION: A COMPARATIVE ANALYSIS OF MODELS AND THEIR EFFECTIVENESS

Aerna Aheti, Hadi Mukhtar — Fri, 19 Jun 2026 00:00:00 +0000

Despite significant progress in combating, phishing continues to be one of the top cybersecurity risks, capitalizing on both technological gaps and human behavior. In this work, we propose a machine learning (ML) framework for creating phishing URL detector based on the LegitPhish labeled dataset that consists of 101,219 URLs and 17 engineered features for each URL. There is a particular focus on pre-modelling data diagnostics to resolve any data leakage, shortcuts and highly correlated variables that can drive the model results and cause artificial over-optimism. Six classification models were tested such as Logistic Regression (LR), Support Vector Machine (SVM), Random Forest, XGBoost, Multi-Layer Perceptron (MLP) and CNN-LSTM architecture. Seven informative features were selected for model development: After discarding two leakage-prone features has_ip_address, https_flag and eight redundant features. Experimental results revealed that the overall performance of the two models, Random Forest and XGBoost, were the best with an F1-score of 0.952 and an ROC-AUC value of 0.992 respectively. The precision score for the Random Forest model was 0.928 and the Recall score was 0.978, which provides a good balance between the detection of phishing and control of false positives. Despite the complexity of the CNN-LSTM model, it did not perform better than the ensemble-based models, while the MLP model proved competitive. The results confirm the significance of careful feature diagnostics and leakage prevention and show that well-tuned ensemble classifiers are able to deliver accurate, efficient, and viable results for the phishing URL detection task.

AI-DRIVEN TEXT AND AUDIO SYNTHESIS FOR LOW-RESOURCE LANGUAGE DATASETS

Aidana Karibayeva, Balzhan Abduali — Fri, 19 Jun 2026 00:00:00 +0000

This paper presents an AI-driven methodology for constructing parallel text corpora and synthesizing parallel audio datasets for low-resource Turkic language pairs, namely Kazakh-Kyrgyz and Kazakh-Uzbek. The scarcity of high-quality linguistic and speech resources for these languages poses significant challenges for the development of neural machine translation and automatic speech recognition systems. The paper proposes and validates a methodology for parallel dataset construction driven by artificial intelligence, in which the selection of a translation system is guided by a set of criteria encompassing free accessibility, translation quality, and processing efficiency, assessed through experiments on a 2000-sentence test set. Based on this evaluation, large-scale parallel corpora for the Kazakh–Kyrgyz and Kazakh–Uzbek language pairs were generated using the selected AI system. A subsequent manual error analysis revealed that approximately 3.7% (Kazakh–Kyrgyz) and 4.3% (Kazakh–Uzbek) of the translations contained inaccuracies, indicating the need for post-editing and further model refinement. Audio synthesis experiments using MMS-TTS and TurkicTTS systems demonstrated that synthetic speech of near-natural quality can be generated for both languages, with NISQA scores reaching up to 4.44 for Uzbek. The findings confirm that the proposed methodology provides a practical and scalable foundation for expanding linguistic resources for low-resource Turkic languages and supporting further research in machine translation and speech processing.

PPO-BASED PHYSICAL RESOURCE BLOCK SCHEDULING IN 5G NR: A REINFORCEMENT LEARNING APPROACH WITH 3GPP-COMPLIANT CHANNEL MODELING

Yedil Nurakhov, Abzal Kyzyrkanov, Syrym Aldabergen — Fri, 19 Jun 2026 00:00:00 +0000

This paper proposes a deep reinforcement learning approach to Physical Resource Block (PRB) scheduling in 5G New Radio networks. A Gymnasium-compatible simulation environment is developed on the 3GPP TR 38.901 UMa NLOS channel model with Rayleigh fading, serving as a reproducible experimental platform. A Proximal Policy Optimization (PPO) agent is trained with a composite reward function that jointly optimizes aggregate throughput and Jain's Fairness Index. The agent is evaluated over 100 episodes with five independent random seeds against three classical baselines — Round Robin, Proportional Fair, and Maximum CQI. The PPO agent achieves a mean throughput of 57.09 ± 1.04 Mbps with a Jain's Fairness Index of 0.358, surpassing Round Robin (32.37 Mbps, JFI 0.630) and Proportional Fair (35.31 Mbps, JFI 0.597) in throughput while exceeding Maximum CQI in fairness (JFI 0.100). A throughput-fairness tradeoff analysis confirms that the proposed agent occupies an operating point inaccessible to any single classical method. The composite reward function, which weights both objectives equally, enables learned adaptive behavior that balances spectral efficiency and user equity. The code and evaluation pipeline are fully open-source and reproducible using the Stable-Baselines3 library.

KNOWLEDGE FUSION THROUGH DATALESS BERT PARAMETER MERGING VIA A REGRESSION-BASED AVERAGING

Alisher Kaziz, Richard Harrison — Fri, 19 Jun 2026 00:00:00 +0000

Recent progress in dataless knowledge fusion methods has focused on migrating knowledge from a teacher neural network to a student neural network without utilizing the teacher model's training data. As a result, fine-tuning pre-trained language models (PLM) has emerged as a simple method to improve the performance of Natural Language Processing (NLP) models in specific domains. These refined models are available as open source, but typically their training datasets are not, due to issues related to data privacy or intellectual property. This sets up an obstruction to combining knowledge from separate models to produce an enhanced single model. In this paper, we investigate the issue of combining separate models developed on distinct training datasets to create a unified model that performs on out-of-domain data for the student model. We suggest a dataless knowledge fusion technique that integrates models within their parameter space, directed by weights that reduce prediction discrepancies between the combined model and the separate models. Across a wide range of parameter configurations, our assessment indicates that the suggested approach substantially exceeds the performance of traditional baselines like Fisher-weighted averaging or model ensembling. Additionally, we observe that our approach serves as a viable alternative to multi-task learning, capable of maintaining and enhancing the individual models without requiring access to the training data. Our experiments confirm the dataless merging methods, integrating finely adjusted parameters to achieve robust multitask performance with negligible impact on class ratio degradation, approximately 2.1%.