Artificial Intelligence and Soft Computing

Air quality is crucial for public health and the environment, therefore, accurate prediction of particulate matter (PM) concentration is a significant challenge in the field of environmental protection. In this paper, we present and compare three approaches to predicting particulate matter concentration: convolutional-recurrent autoencoder, hierarchical autoencoder and variational autoencoder. We conducted experiments on real air quality datasets, evaluating the performance of the methods in terms of prediction accuracy. The results indicate that the proposed approaches can effectively predict changes in particulate matter concentration, which is of great importance for early warning systems and air quality management strategies.

The conjugate gradient (CG) algorithm is a widely used approach for training neural networks. Its most computationally demanding step is directional minimization. This paper introduces a novel modification of the CG algorithm that accelerates directional minimization, leading to a significant reduction in computation time. The proposed modification was evaluated on selected test cases, and its performance was compared with the classical CG method.

In this paper, we present a novel supervised learning algorithm for neural network training based on QR decomposition with Householder reflections. The core of our study outlines the fundamental mathematical principles underlying this approach. To validate its effectiveness and robustness, we provide a detailed analysis of benchmark experiments, demonstrating the advantages of our method.

Fusarium Head Blight (FHB), among other crop diseases, is responsible for major yield losses in small grains. As such, methodologies to detect these diseases from unmanned aerial vehicle (UAV) drones equipped with hyper-spectral sensors have been developed as a time and cost-effective solution for identifying these diseases. The International Conference on Pattern Recognition (ICPR) 2024 held a Kaggle competition to identify the severity of FHB. The dataset provided included 32 \(\times \) 32 \(\times \) 101 images for classification, emphasizing the challenging nature of feature extraction and the complex relationship between the channels. A SOTA 100% classification accuracy was reported on the test data using a simple Resnet-inspired architecture using less than three averaged input bands instead of the full 101 bands. This suggests that perhaps an overemphasis is being placed on utilizing all the information from the multiple bands in hyper-spectral imaging (HSI). However, not all well-established techniques transferred this HSI dataset, with little purpose in augmentations or transfer learning. The findings suggest that with change point analysis of the radiance of pixels in the hyper-spectral imagery, one could identify the splits necessary for a significant reduction in input channels via channel-wise normalization followed by simple arithmetic mean averaging across the channels. Suggesting that already well-established CNN architectures are well-suited for crop disease detection from HSI.

Fusarium head blight (FHB) is a crop disease that significantly threatens grain production and the global agricultural economy. Recent advancements in remote sensing and image-based methods for plant disease diagnosis, emphasizing the superior spectral-spatial information provided by hyperspectral imaging (HSI), aim to address this issue. Accurate and automated FHB detection is crucial for disease management and crop production. This paper explores the potential of HSI for automated crop disease detection, focusing on FHB in wheat, and provides two deep learning-based approaches to address the challenge of FHB detection. The results show that the modified Resnet18 model achieved 100% evaluation accuracy while the DarkNet19 only managed to achieve 73% evaluation accuracy. The t-distributed stochastic neighbor embedding (t-SNE) visualizations used to visualize the latent space for both models further validate these results and illustrate distinctive separation between classes in their feature space. These findings demonstrate the potential of HSI for rapid, non-destructive, and accurate crop disease diagnosis, contributing to the development of efficient, large-scale monitoring systems for improved agricultural management and food security.

Large language models (LLMs) are becoming increasingly proficient in processing and generating multilingual texts, which allows them to address real-world problems more effectively. However, language understanding is a far more complex issue that goes beyond simple text analysis. It requires familiarity with cultural context, including references to everyday life, historical events, traditions, folklore, literature, and pop culture. A lack of such knowledge can lead to misinterpretations and subtle, hard-to-detect errors. To examine language models’ knowledge of the Polish cultural context, we introduce the Polish linguistic and cultural competency benchmark, consisting of 600 manually crafted questions. The benchmark is divided into six categories: history, geography, culture & tradition, art & entertainment, grammar, and vocabulary. As part of our study, we conduct an extensive evaluation involving over 30 open-weight and commercial LLMs. Our experiments provide a new perspective on Polish competencies in language models, moving past traditional natural language processing tasks and general knowledge assessment.

This paper introduces a simplified version of the BBATDD algorithm (Boosting-Based Algorithm Trained with Drift Detector), called A-BBATDD, which aims to speed up neural network training. Instead of updating the importance of each training example individually, the new version uses an approximation based on the whole mini-batches. This makes the method easier to use with standard machine learning tools such as TensorFlow or PyTorch. The algorithm still focuses on sampling the most difficult examples for the network and includes a drift detection mechanism (CUSUM) to reset the importance values when the network performance changes significantly. Experiments on the MNIST dataset show that the performance of the method is especially sensitive to the mini-batch size.

In this paper, we compare different approaches for prototype classification learning schemes on spherical manifolds and consider them from the perspective of quantum computing. More precisely, these approaches are based on the concept of learning vector quantization, i.e. prototype based models which are known for its robustness and interpretability. We analyze them to what extent these models are applicable on quantum devices and which quantum assumptions are met. Thereby, the relation of learning on spherical manifolds to the learning on Bloch-spheres in quantum approaches is of special interest.

Crop diseases, particularly Fusarium Head Blight (FHB), pose a significant threat to agricultural productivity, causing severe economic losses and contributing to food insecurity. Traditional diagnostic methods, which are based heavily on expert visual inspections, are often time-consuming, error-prone, and insufficient for early intervention. Hyperspectral imaging (HSI) provides a powerful alternative by capturing spectral information across multiple wavelengths, revealing subtle physiological changes caused by diseases imperceptible to the naked eye. This study combines HSI with deep learning, using a hybrid Inception-ResNet architecture for the binary classification of the severity of FHB. The model efficiently extracts localized spectral features from hyperspectral datasets by implementing a sliding-window approach and custom preprocessing techniques. The proposed framework achieved a validation accuracy of \(\sim 98.92\%\) and a training accuracy of \(\sim 99.4\%\), demonstrating its effectiveness in distinguishing between mild and severe FHB. The fusion of HSI and CNNs offers a scalable, noninvasive solution for precision agriculture, enabling timely disease detection and reducing reliance on labour-intensive diagnostic methods. This framework holds promise for broader agricultural applications, supporting sustainable farming practices and contributing to global food security. Future research will explore the extension of this approach to detect other plant diseases and the deployment of the model in real-time monitoring systems for more efficient crop health management.

This paper presents a federated learning (FL) framework designed to counter Byzantine and Sybil-type attacks, which are major security threats in distributed systems. One of the key elements in such systems is business continuity. By combining the Multi-Krum client selection mechanism with the FoolsGold aggregation approach, the framework offers resilience against these two attack types. Byzantine attacks, characterized by clients sending arbitrary or misleading updates, can disrupt model convergence. In contrast, Sybil attacks, in which attackers overwhelm the network with multiple malicious clients forming a majority, can significantly disrupt the aggregation. These attack types may decrease performance, increase communication load, and result in incorrect predictions. By leveraging a statistical method based on Chernoff bounds, the system determines the optimal number of clients selected via the Multi-Krum method, removing the outlier weights from the malicious clients. The FoolsGold aggregation method utilizes cosine similarity and clients’ gradient history and assigns weights to each selected client, reducing the risk of the aggregation process being overwhelmed by Sybil’s attackers. Experiments with attention-augmented CNNs on the MNIST and Covid-19 Radiology datasets confirmed the system’s effectiveness under various adversarial conditions, achieving 97% and 77% accuracy, respectively, comparable to results without attacks. The results indicate the effectiveness and robustness of the proposed strategy as a defense mechanism against FL poisoning attacks and provide a basis for further development of techniques for detecting non-metric-based attacks.

We propose an automated vulnerability detection system that synergizes static analysis and fuzzing target identification through LLMs and AI agents. Building on the Dante system, our solution integrates various reasoning models and leverages a dynamic dataset from student code contributions. The system employs reinforcement learning, prompt crafting, and test-time computing techniques to refine the detection of critical vulnerabilities in C codebases. A multi-step automated workflow performs detailed static code analysis, extracts fuzzing targets, and iteratively compiles and tests code snippets. Log outputs are summarized using reasoning models, highlighting only the most relevant errors.

This article explores the use of quasirecurrent neural networks (QRNNs) in selected natural language processing (NLP) tasks—specifically encoding, classification, regression, and named entity recognition. Our hypothesis suggests that QRNNs, as a simpler recurrent architecture, might offer advantages over long short-term memory (LSTM) networks, such as faster inference times and better utilisation of computational resources. To test our hypothesis, we conducted an empirical investigation that implemented both QRNN and LSTM architectures, and compared their performance in terms of quality and speed in selected NLP tasks. Although QRNNs performed better than LSTM networks, the improvement was less substantial than in established architectures. QRNNs deliver various benefits, such as rapid training and fine-tuning, lower memory and data needs, and the prevention of overfitting on small datasets through proper regularisation techniques. We believe that QRNNs represent a compelling and unique concept that has yet to be fully embraced as a mainstream NLP architecture.

This study introduces a novel Semi-Decentralized Long Short-Term Memory (SD-LSTM) architecture and compares its performance against a traditional LSTM model for stock price prediction, examining both accuracy and training time. All experiments employ canonical settings. Results indicate that SD-LSTM consistently achieves better prediction accuracy—evidenced by significantly lower mean squared error—across stock data from 5 major U.S. companies (Apple, NVIDIA, Amazon, Alphabet, Microsoft). Moreover, SD-LSTM accomplishes these improvements with fewer parameters. In terms of training speed, SD-LSTM is substantially faster than traditional LSTM when handling larger datasets and more complex configurations, highlighting its efficiency in parallel processing. Overall, these findings underscore the potential of this new SD-LSTM architecture for large-scale applications and its viability for integration into both established and emerging hybrid approaches that demand advanced predictive accuracy and computational efficiency.

Viscosity, a key property of liquids, is crucial for quality control in industries like chemical, pharmaceutical, and food processing. Traditional viscosity measurement tools, such as capillary viscometers, are costly and unsuitable for real-time monitoring. This study investigates the impact of convolutional neural network (CNN) architectures on predicting the viscosity of water – PVP solutions. Using droplet images from 14 solutions with varying water-to-PVP ratios, we evaluated three CNN architectures—SimpleModel, DeepModel, and ComplexModel. The findings highlight the influence of model complexity on prediction accuracy.

We employ a custom-designed convolutional-glial network architecture for the classification of large text corpora. The problem selected to demonstrate the high effectiveness of the proposed technique, was defined in collaboration with Poland’s largest railway operator. The task involves categorizing incoming reports (submitted by passengers, station staff, and railway line personnel) into appropriate departments based on their textual content. The input data consists of short texts containing complaints, inquiries, and remarks, which were transformed into vector representations using the BERT language model. A hybrid convolutional architecture was then applied, augmented with a glial-type control module that dynamically regulates the activation levels of CNN filters. This glial control layer, inspired by biological mechanisms, was trained alternately with the convolutional component, enabling better adaptation of information flow within the network. Experimental results confirm that this approach outperforms classic architectures in terms of classification accuracy, while also offering greater flexibility and development potential. The proposed solution can be effectively applied in customer service support systems and automated text analysis for public transportation.

Crime, a universal phenomenon rooted in social dynamics, requires innovative approaches to public security. This paper presents a method that integrates Process Mining, LSTM (Long Short Term Memory), and Markov Chains to predict crime types, supporting resource allocation and preventive strategies. Using 512,657 crime records from Denver, Colorado, the method follows six stages: (i) data preparation, (ii) DFG (Directly Follows Graph) extraction, (iii) transition matrix construction (Markov Chains), (iv) LSTM-based prediction, (v) model evaluation, and (vi) visualization for decision making. The proposed method achieved 94% accuracy, with precision and recall around 95% (Top-1) in the LSTM model, significantly outperforming Markov Chains. This approach provides interpretable visualizations that enhance strategic decision-making and highlights the potential of combining machine learning and statistical modeling for more effective public security policies.

Air pollution remains a critical issue in many countries, particularly in developing regions like India. In November 2019, New Delhi experienced an alarming Air Quality Index (AQI) level of 900, far exceeding the ‘severe’ threshold. Accurate air pollution forecasting is essential for informed decision-making due to its direct impact on public health. Effective forecasting depends on selecting suitable methods and evaluation measures to maximize prediction accuracy. This study investigates the prediction of air quality levels in India, emphasizing techniques to enhance forecasting precision and identify areas for further model improvement. Our research evaluates various Neural Networks, deep learning, and machine learning algorithms for predicting AQI. The findings reveal that among the techniques explored, the Long Short-Term Memory (LSTM) model outperforms others, demonstrating very good accuracy in capturing complex patterns in AQI data. This contribution serves as a valuable benchmark for experts aiming to refine forecasting methods and provides a reference for emerging researchers in the field.

In this contribution we demonstrate the flexibility of learning vector quantization (LVQ) variants for reliable classification learning in a medical domain. More specifically, we consider classification of breast cancer subtypes by means of gene expression profiles using knowledge-informed relevance learning, where the medical knowledge regarding co-expression of genes and resulting pathways and biological processes are obtained from medical databases. This leads to better interpretability and allows partial causal inferences about the influence of those relations to identification of the subtypes from a gene-expression perspective. Further, relevance learning information from LVQ can be used to detect and to mitigate unwanted distortion from the data and, hence, contributes to more reliable classification ability.

This paper explores the automatic classification of English literary texts by epoch using the CoLiE dataset, a curated collection of texts from Project Gutenberg. The authors employ various computational approaches, including traditional bag-of-words models with TF-IDF weighting, enhancements incorporating part-of-speech n-grams, and maxLogit-based sequence classification. In addition, they examine the performance of transformer-based models such as BERT and RoBERTa, as well as the impact of named entity masking on classification accuracy. Furthermore, the study evaluates the capabilities of large language models (LLMs) in epoch detection using different prompting strategies. The results indicate that, while TF-IDF-based models achieve competitive performance, transformer-based classifiers show promise, but are occasionally outperformed. In particular, LLMs exhibit sensitivity to prompt formulation, with classification accuracy varying significantly depending on the input structure. Even in the best-case scenario, LLMs achieve accuracy approximately seven percentage points lower than other methods.

The commonality of high-dimensional data encountered in numerous application areas poses significant challenges and impacts the performance and accuracy of machine learning models, including fuzzy models. The effect known as a concentration phenomenon has been discussed intensively in the literature. However, this detrimental aspect has not been thoroughly studied and quantified in the area of fuzzy models. To narrow down the existing gap and bring some original design insights, in this study, we investigate the impact of data dimensionality on the performance of fuzzy rule-based models, in particular, Takagi-Sugeno models. The effect of increased dimensionality becomes clearly manifested when building fuzzy condition parts (fuzzy sets) realized with the use of Fuzzy C-Means clustering. The Fuzzy C-Means algorithm dwells on the distances between data and prototypes, and these distances are directly impacted by the concentration phenomenon. Furthermore, the study leverages partition matrix-based concentration indices and histograms of membership grades to evaluate clustering results that have a direct impact on the performance of ensuing fuzzy models. Based on the results of the Fuzzy C-Means clustering, we explore the effect of the increasing dimensionality of the feature space on the performance of K-Nearest Neighbor classifiers. By working with high-dimensional data, we experimentally explore the effect of successive random reduction of the number of features (input variables). The experimental results reveal that the performance of the model (both predictors and classifiers) can often be enhanced by reducing the dimensionality of the feature space, leading to models of lower complexity and improved generalization abilities. In the study, we also offer some design guidelines.

CliftonStrengths (formerly StrengthsFinder) is a popular psychometric tool created by Donald Clifton at the Gallup Institute, used to assess the intensity of 34 traits called “talents”, of which 5 are considered dominant. Talent is understood as a natural pattern of thinking, feeling and acting that can be effectively used in various areas of life. According to the assumptions of the theory, people who develop their key talents are more likely to achieve above-average results.

The use of CliftonStrengths in higher education has wide implications. It is commonly perceived that science and technology studies are chosen by people with a greater degree of introversion, disciplined and analytical thinking, while humanities and arts courses attract people who are extroverted, creative and focused on Relationship Building. It is therefore assumed that having certain talents can predispose to fulfilling specific roles in the professional community.

The aim of this study was to determine the most frequently appearing talents depending on the chosen field of study. Statistical analysis of the obtained results showed the existence of certain relationships between the talent profile and educational preferences, which are described in detail in the article. The applied machine learning tools and fuzzy techniques allowed to indicate relationships that had not been considered by specialists so far and shed new light on the issues considered.

Multi-Population-based Algorithms (MPAs) are algorithms for searching the solution space in which the population is divided into subpopulations (islands). One of the MPA methods is the Multi-population-based Nature-Inspired Algorithm (MNIA), in which each subpopulation is processed by a different Population-Based Algorithm (PBA). This integrates various approaches to searching the solution space and increasing the efficiency of exploration and exploitation. The aim of this paper is to investigate the influence of the selection of base algorithms in the MNIA on its efficiency. Knowledge in this area can be important in the context of designing mechanisms for the automatic selection of the MNIA formulas during its operation. The approaches considered in the simulations gave very good results and contributed to the formulation of several interesting conclusions.

We have been experimenting with an idea of extending the information encoded in the ACO pheromone to enhance the efficiency and efficacy of swarm optimization. Up to now this idea has been applied to single-criteria problems; this paper, however, aims at extending it and applying to multi-criteria TSP problems. We focus on presenting the concept of the problem and showing in a statistically significant way that the proposed optimization method works for the selected problems.

This paper presents the application of socio-cognitive mutation operators inspired by the TOPSIS method to the Low Autocorrelation Binary Sequence (LABS) problem. Traditional evolutionary algorithms, while effective, often suffer from premature convergence and poor exploration-exploitation balance. To address these challenges, we introduce socio-cognitive mutation mechanisms that integrate strategies of following the best solutions and avoiding the worst. By guiding search agents to imitate high-performing solutions and avoid poor ones, these operators enhance both solution diversity and convergence efficiency. Experimental results demonstrate that TOPSIS-inspired mutation outperforms the base algorithm in optimizing LABS sequences. The study highlights the potential of socio-cognitive learning principles in evolutionary computation and suggests directions for further refinement.

This paper investigates the critical aspect of migration timing in hybrid island-based metaheuristic algorithms. Migration timing plays a pivotal role in balancing exploration and exploitation, ensuring that the algorithm avoids premature convergence while effectively exploring the search space. We propose and evaluate several migration timing strategies, including periodic migration, fitness-based triggers, and diversity-driven approaches. Our experiments are conducted on a set of benchmark optimization problems, including both discrete (Traveling Salesman Problem) and continuous (Black-box Optimization Benchmarking) tasks. The results demonstrate that adaptive migration strategies, which dynamically adjust based on population diversity and fitness stagnation, outperform static approaches. This study provides insights into the optimal conditions for triggering migration and offers guidelines for designing more effective hybrid metaheuristic frameworks.