Eastern-European Journal of Enterprise Technologies

Improving the efficiency of greenhouse control by using a Markov decision-making process model

Fri, 31 Oct 2025 00:00:00 +0200

The object of this paper is the process of greenhouse control. The study solves the task of rational greenhouse control based on the Markov decision-making process taking into account two-level optimization. A random Markov decision-making process has been defined for the problem of greenhouse operation improvement.

A greenhouse control model was built, which makes it possible to determine rational microclimate parameters to grow agricultural crops. To validate the greenhouse control model, real data from an experiment on growing strawberries in a greenhouse complex were used.

Observations lasted from May 17 to June 8, 2025. Monitoring of microclimate parameters was carried out around the clock with an interval of 1 minute, which ensured high accuracy of the analysis. The experimental scenario included three irrigation circuits, a heating system, LED lighting, ventilation, and CO₂ monitoring.

The proposed approach to greenhouse management based on the Markov decision-making process model demonstrates high practical value, especially in the context of growing sensitive crops such as strawberries. The simulation shows that the implementation of two-level optimization in autonomous greenhouse control systems could provide an increase in yield by 10.15%. At the same time, due to the significant volume of the greenhouse and the high thermal inertia of the structures, the actual values of the microclimate parameters deviate from the rational ones by 10–15%, as a result of which the calculated yield increase for the model built is about 7%.

Development of a multi-modal fully guided attention gate (MM-FGAG) framework for spatio-temporal flood detection

Neni Mulyani, Anjar Wanto, Jhonson Efendi Hutagalung — Fri, 31 Oct 2025 00:00:00 +0200

Floods are one of the most frequent hydrometeorological disasters in Indonesia, causing severe social, economic, and environmental impacts. The object of this research is spatio-temporal flood detection in Simpang Empat, Asahan Regency, North Sumatra, an area that faces annual flooding due to high rainfall, low-lying topography, and land-use changes. Conventional detection approaches based on either spatial or temporal data often fail to capture complex interactions, thereby limiting predictive accuracy. To address this problem, this study developed a multi-modal fully guided attention gate (MM-FGAG) framework that integrates Sentinel-2 multispectral imagery, SRTM elevation, CHIRPS rainfall, and ERA5 atmospheric variables. The model employs CNN-based spatial priors to guide temporal attention in LSTM, ensuring that predictions focus on the most flood-relevant regions and time periods. Experimental results show that MM-FGAG achieved 91.72% accuracy, 92.05% precision, 90.29% recall, and an AUC of 0.945, significantly outperforming CNN, LSTM, and CNN-LSTM baselines. This improvement is explained by explicit spatial-to-temporal guidance, which enhances predictive accuracy while also increasing interpretability through attention maps. Distinctive features of the framework include multimodal integration, guided attention, and the ability to generate flood risk maps with more than 90% agreement with observed data. These findings confirm that MM-FGAG is robust, adaptive, and capable of producing accurate and explainable predictions. The framework shows strong potential for use in flood early warning systems and disaster risk management, providing timely information for evacuation planning and resource allocation in vulnerable regions.

A hybrid multi-scale convolution neural network with attention and texture features for improved image classification

Irpan Adiputra Pardosi, Tengku Henny Febriana Harumy, Syahril Efendi — Fri, 31 Oct 2025 00:00:00 +0200

The object of this study is the classification of low-resolution and multi-class images, represented by the CIFAR-10 benchmark dataset. It is challenging to accurately classify low-resolution and multi-class images because traditional CNNs usually have trouble identifying both global and complex texture patterns. To address this issue, this study employs the CIFAR-10 dataset as a representative benchmark for real-world scenarios where image quality is limited, such as in low-cost medical imaging, remote sensing, and security surveillance systems. The limited discriminability of traditional CNNs in these situations is the primary issue addressed. The proposed method employs three parallel convolutional streams with distinct kernel sizes (3 × 3, 5 × 5, and 7 × 7) to capture hierarchical spatial patterns, followed by the integration of two attention mechanisms – squeeze-and-excitation and convolutional block attention module – that adaptively emphasize the most relevant spatial and channel-wise information. In addition, structural texture descriptors such as Gray-level co-occurrence matrix, local binary pattern, and Gabor filters are computed independently and later fused with the deep representations to enrich the feature space. Experiments were carried out on the CIFAR-10 dataset under varying levels of class complexity: 10, 5, and 3 categories. The results reveal that the hybrid approach significantly improves precision, recall, and F1-score across all scenarios, with the highest accuracy of 90.87% obtained when only three classes are involved. These improvements are explained by the complementary nature of deep and handcrafted features, which together enable the model to learn both global semantics and fine-grained local textures can achieve higher classification accuracy, improved reliability, and reduced misclassification errors, ultimately enhancing the effectiveness of applications ranging from medical decision support to intelligent surveillance.

Evaluating an integrated user‑feedback approach to software quality monitoring: enhancing accuracy and timeliness

Fri, 31 Oct 2025 00:00:00 +0200

The object of the research is a monitoring and analysis system for software quality based on user feedback collected from open-source projects on GitHub. The problem addressed is the lack of effective automated tools that can process large volumes of unstructured user feedback to identify quality issues, prioritize tasks, and detect negative trends in real time. Traditional quality assurance methods, while important, fail to capture the nuance of user sentiment and the contextual details present in natural language feedback, leading to delays in problem detection and resolution. The developed system integrates three key modules: sentiment analysis for assessing user satisfaction, issue categorization for structuring feedback into actionable types, and anomaly detection for identifying sudden changes in sentiment or feedback dynamics. The results show that transformer-based models, particularly fine-tuned BERT, outperform rule-based and traditional machine learning approaches in both accuracy and robustness. This advantage is explained by their ability to capture domain-specific language, sarcasm, and contextual dependencies, enabling more precise interpretation of complex feedback. The anomaly detection component, using LSTM autoencoders and Isolation Forest, demonstrated the ability to identify critical quality regressions up to two days before official issue reporting. These results can be applied in practice for continuous software quality monitoring in agile, open-source, or user-centric development environments where timely, data-driven decision-making is essential. The approach supports real-time insight generation, helping development teams respond proactively to quality risks and improve overall user satisfaction

Development of a multi-agent adaptive recommendation system based on reinforcement learning

Bohdan Romaniuk, Olha Peliushkevych — Fri, 31 Oct 2025 00:00:00 +0200

This study's object is the process that improves efficiency and accuracy in delivering personalized recommendations to users in systems based on reinforcement learning.

The principal task addressed in the study is to improve recommendation adaptation and personalization by assigning a dedicated agent to each user. This approach reduces the influence of other users’ activity and allows for more precise modeling of individual preferences.

The proposed approach employs an Actor–Critic model implemented using the Deep Deterministic Policy Gradient algorithm to achieve more stable training and maximize long-term rewards in sequential decision-making processes. Recommendations are generated using the unique characteristics of items that are based on users’ historical interactions. Neural networks are trained with separate parameter configurations for single-agent and multi-agent models.

Experimental results on the MovieLens dataset demonstrate the superiority of the multi-agent model over the single-agent baseline across key evaluation metrics. For top-5 recommendations, the multi-agent model achieved improvements of + 4% for Precision@5; + 0.32% for Recall@5; and + 2.92% in Normalized Discounted Cumulative Gain NDCG@5. For top-10 recommendations, gains were + 1% for Precision@10; + 0.18% for Recall@10; and + 1.14% for NDCG@10, respectively.

Simulations for individual users showed that the multi-agent model outperformed the single-agent baseline in 66 out of 100 cases in terms of cumulative reward. The proposed system demonstrates effectiveness in capturing user preferences, improving recommendation quality, and adapting to evolving user preferences over time.

The main area of practical application for the results includes dynamic online environments such as e-commerce systems, media platforms, social networks, and news aggregators.

Devising a code-free method for detecting signs of informational-psychological influences in messages

Dmytro Lande, Kostiantyn Yefremov, Artem Soboliev, Ivan Pyshnograiev — Fri, 31 Oct 2025 00:00:00 +0200

This study investigates text messages that potentially contain signs of informational-psychological operations (IPSOs). The task addressed aims to solve the problem of detecting signs of IPSOs in the media space.

An innovative method for detecting such signs has been proposed, based on the construction and analysis of semantic networks and implemented without the use of program code by using large language models (LLMs). This makes it possible to generate formalized analytical queries to LLMs in the form of a code-free system based on the composition of structured prompts.

The method's unique feature is the parallel analysis of data from two sources of knowledge: internal and external. The internal one contains generalized IPSO patterns formed on the basis of a wide corpus of data. The external one includes verified examples of fake messages from social networks, news outlets, and archives of fact-checking organizations.

To improve the accuracy of analysis, semantic normalization of concepts is used, which employs embedded vectors to unify terminology, as well as comparison of causal paths in semantic networks to identify connections. The assessment of the probability of a message belonging to IPSO is formed by aggregating the results using a weighted average, which makes it possible to take into account semantic and structural similarity. An example of applying the method to the analysis of a disinformation message is given, demonstrating the ability to detect key signs of psychological influence: manipulative narratives, emotional loading, and cause-and-effect relationships.

The proposed method is flexible, reproducible, and accessible to researchers without programming skills, which makes it a valuable tool for monitoring information threats and analyzing disinformation in the context of information confrontations

Enhancing retrieval performance in social media with corpus-based query expansion using bidirectional encoder representations from transformers

Roberto Kaban, Poltak Sihombing, Syahril Efendi, Maya Silvi Lydia — Fri, 31 Oct 2025 00:00:00 +0200

This study focuses on a collection of tweets related to government services (e-government), which are preprocessed and transformed into a domain-specific corpus for query expansion. Conventional IR models struggle with unstructured and noisy content containing informal language and abbreviations, which reduces retrieval accuracy. To overcome these issues, this study proposes a hybrid query expansion (QE) model named ROCBERT-QE, which combines corpus-based retrieval (CBR) with bidirectional encoder representations from transformers (BERT). The model applies dual expansion, using corpus-based co-occurrence frequencies to capture lexical relationships and BERT embeddings to preserve semantic context. A domain-specific corpus consisting of 5,017 preprocessed tweets related to Indonesia’s National Health Insurance (BPJS) was constructed, encompassing 6,215 unique terms that represent linguistic variation and informality in public discourse. Experimental results demonstrate that ROCBERT-QE outperforms baseline retrieval methods such as TF-IDF, BM25, and standard BERT. For single-word queries, Recall reached 0.8574 and Precision 0.8807, while for sentence-level queries, Recall was 0.8932 and Precision 0.9175. The synergy of frequency-based and contextual expansion enables effective handling of lexical noise and semantic ambiguity. The results confirm the scientific potential of combining corpus-based and transformer-based approaches in IR tasks involving unstructured content. Practically, ROCBERT-QE can be applied for real-time analysis of citizen discourse in e-government contexts, such as service evaluation, policy feedback, and early detection of public issues. The framework is scalable and adaptable to other domains with informal or multilingual data characteristics

Formalization of text prompts to artificial intelligence systems

Vladyslav Oliinyk, Andrii Biziuk, Zhanna Deineko, Viktor Chelombitko — Fri, 31 Oct 2025 00:00:00 +0200

This study's object is the process that formalizes text prompts to large language models of artificial intelligence for the purpose of automatically generating action cards for hexagonal tabletop wargames. The task relates to the ambiguity (42% of terms are misinterpreted), contextual incompleteness (37%), and syntactic variability (21%) of natural language prompts, resulting in unhelpful and unpredictable responses.

To address this problem, a conceptual-practical model has been proposed, which combines structured prompt templates, a localized glossary of key terms, as well as clear instructions on response format.

Practical verification was carried out by generating a set of action cards for solo wargames on modern large language models by generating over 100 prompts and analyzing more than 300 responses. The experiments demonstrated that the formalized prompts reduced the total error rate by 58%, as well as increased the relevance of responses from 55–65% to 88–92%. The average time for preparing prompts was reduced by 25–40%. The “d6-table” templates ensured the stability of the output format in 90–95% of cases while JSON structures provided stability in 85–90% of cases. The glossary and structure definition integrated into the prompts minimized semantic discrepancies and syntactic errors.

A special feature of the proposed template structures for prompts is adaptability to different subject areas through the use of a description language specific to each of these areas. The research results have practical value for automating game content development processes and could be adapted for other subject areas where accuracy, consistency, and structure of language model responses are important.

The proposed systematic approach facilitates the automation of complex content development with a guaranteed increase in the quality and predictability of responses from large language models.

Determining the effectiveness of GPT-4.1-mini for multiclass text categorization

Yurii Voloshchuk, Oleksandr Mitsa — Fri, 31 Oct 2025 00:00:00 +0200

The object of this study is the process of multiclass automatic categorization of user queries using large language models under the conditions of a language transition from English to Ukrainian.

The scientific task relates to the fact that most modern large language models (LLMs) are optimized for English while their effectiveness for morphologically complex and low-resource languages, particularly Ukrainian, remains insufficiently studied.

In this work, an experimental approach was devised and implemented to evaluate the transferability of the GPT-4.1-mini model from English to Ukrainian in the task to categorize 11,047 user queries spanning nine applied domains. The analysis employed conventional metrics (Recall, Precision, Weighted-F₁, Macro-F₁) alongside a novel indicator, the Uncertainty/Error Rate (U/E), which captures the proportion of model refusals and “hallucinations.”

The findings demonstrate that the highest quality was achieved on the English dataset (Macro-F₁ = 69.78%, U/E = 0.05%). When Ukrainian prompts were applied, Macro-F₁ decreased to 63.73%; however, the U/E equaled 0%, indicating higher reliability of responses. Using English prompts with Ukrainian-language data preserved nearly the same level of accuracy (Macro-F₁ = 69.66%), thereby revealing strong internal translation and generalization mechanisms.

The novelty of this study is attributed to the use of a large multidomain parallel corpus, the systematic comparison of prompts in two languages, the application of the state-of-the-art model GPT-4.1-mini, and the introduction of the U/E metric as a reliability criterion. The proposed approach demonstrates the feasibility of applying GPT-4.1-mini to Ukrainian-language information services without additional training, particularly for automatic query routing in financial, medical, legal, and other domains.

Identifying the impact of forecast errors and flexibility preferences in decision support for optimal day-ahead prosumer operational planning

Oleh Lukianykhin, Vira Shendryk — Fri, 31 Oct 2025 00:00:00 +0200

The object of the study is operational planning in decision support systems (DSSs) for prosumers. The study addresses a lack of explicit flexibility modeling in DSSs and a limited understanding of how forecast quality impacts planning results.

A novel control module for short-term planning of flexible energy demand and battery dispatch in prosumers is presented. The proposed solution improves prosumers’ information support by integrating consumption and generation forecasts, user-defined flexibility preferences, and battery constraints to reduce operational costs and increase profit from energy sales via optimal planning. Unlike methods that obscure decision logic, the module enables explicit flexibility modeling, enhancing transparency and better reflecting individual behaviors. Validation using real-world data across diverse prosumer segments confirms the module’s robustness and effectiveness in achieving cost savings.

The module maintained positive cost improvements under realistic and extreme forecast errors (up to 75%) across most flexibility settings, with performance influenced by forecast accuracy and flexibility configuration. A linear dependency was found between forecast error and cost savings. In rare edge cases – very low flexibility and high forecast error – the control plans led to underperformance. Increasing flexibility relaxes accuracy requirements, highlighting an important trade-off. Higher flexibility led to stronger initial performance but faster degradation as forecast errors increased. Lower flexibility setups declined more slowly but were more prone to underperformance in edge conditions.

These findings offer practical insights into flexibility modeling and forecast error tolerance, enabling improved planning and control design for prosumers