Keywords: Machine/Reinforcement Learning

Keywords: Machine Learning, Biomedical Engineering, Computer Vision
Abstract: Breast cancer remains among the leading causes of death among women worldwide. The integration of advanced technologies such as healthcare cyber-physical systems, computer vision, and deep learning models has enhanced the efficiency of breast cancer diagnosis. This study proposes HybNet, a hybrid deep learning-based backbone feature extraction model comprising convolutional layers and inception modules, for breast tumor classification from ultrasound images. The model is trained and validated on two datasets: one publicly available dataset (Breast Ultrasound Images (BUSI) Dataset) and another private dataset (i.e., Dataset B). Moreover, the robustness of the model is enhanced by increasing the reliability of the datasets via morphological data augmentation. Experimental results demonstrate that HybNet achieves accuracy, precision, sensitivity, F1-score, and specificity values of 95.1%, 96.6%, 93.7%, 94.9%, and 96.3%, respectively, on BUSI, and 96.8%, 97.2%, 98.6%, 97.9%, and 98.8%, respectively, on Dataset B, thereby outperforming ResNet-50 and Inception-V3.

Keywords: Communication Systems, Communication Networks
Abstract: Wireless protocols are undergoing increasingly rapid development. Currently, most wireless networks that provide Internet access use the 802.11ac protocol. This paper evaluates its performance in the 5 GHz band. In the proposed scenario, the access point is configured using MikroTik equipment capable of supporting a channel width of up to 80 MHz. The study also analyzes bandwidth sharing under these conditions. The results show that, with an 80 MHz radio channel, 802.11ac maintains a stable connection and achieves transfer speeds exceeding 10 Mbps for both upload and download, even when multiple clients generate traffic simultaneously to the access point.

Keywords: Neural Networks, Machine Learning, Biomedical Engineering
Abstract: Ovarian cancer is one of the most lethal gynecologic malignancies for women worldwide, with a high mortality rate due to its late diagnosis and limited effective treatment plans. Considering its severity, automated segmentation of ovarian tumors in medical imaging can be important for early detection and treatment planning. Recent advances in deep learning and neural networks have significantly contributed to this field. In this paper, we present a comprehensive study of multiple neural network architectures for ovarian tumor segmentation, with a particular focus on U-Net and its variants. We aim to investigate various network configurations, including different encoder backbones, to optimize segmentation performance. Our experiments utilize the publicly available MMOTU dataset, providing a diverse ovarian ultrasound image collection. The results demonstrate that carefully selected configurations and augmentations can enhance segmentation accuracy, as quantified by metrics such as Intersection over Union and Dice Coefficient. This study contributes to research by offering insights into the design and optimization of deep learning models for ovarian cancer segmentation and highlights the potential for improved diagnostic support through automated image analysis.

Keywords: Intelligent Systems, Neural Networks, Machine Learning
Abstract: The power of generative AI is well recognized nowadays in offering answers to many topics. However, the quality of the response may not be the best when a large input is given, especially when a technical response is requested. This paper presents an innovative approach to the Text-to-SQL problem, leveraging Retrieval-Augmented Generation. The proposed solution addresses the challenge of translating natural language requests into SQL statements by focusing on the extraction of relevant tables and columns from extensive database schemas relevant to a given question. The results were analyzed based on the actual columns and tables that a specific question is pointing to, using several fine-tuned Bidirectional Encoder Representations from Transformer models.

Keywords: Neural Networks, Machine Learning, Biomedical Engineering
Abstract: The research focuses on a deep learning ensemble algorithms approach for classifying ovarian tumor stages, using medical images and specific annotations from the TCGA-OV dataset. Ovarian cancer, a particularly dangerous form of gynecological malignancy, is often diagnosed in advanced stages. After implementing transfer techniques on 5 individual neural networks, including ResNet-50, DenseNet-12, EfficientNet-B3, Vision Transformer (ViT), and MobileNetV2, the best 3 were chosen to create an ensemble model. The ensemble model (Weighted Voting Ensemble) demonstrated an enhanced 97.2% accuracy and an AUC-ROC of 0.985 compared to the best individual neural network performances obtained for Vision Transformer (ViT) -accuracy of 95.4% and an AUC-ROC of 0.971. This approach diminishes classification errors and enhances generalization. The research emphasizes the importance of dataset quality, interpretability, and clinical relevance, proposing future efforts in improving datasets and implementing clinical AI in real-time. The results indicate that AI can greatly enhance the staging of ovarian tumors, lessening diagnostic bias and tailoring personalized treatments.

Keywords: Machine/Reinforcement Learning, Intelligent Systems, Control Applications
Abstract: Low voltage Direct Current (LVDC) systems represent an emerging alternative for efficient usage and local energy distribution. Such systems are able to integrate renewable energy sources, energy storage and DC-native and AC loads using a common DC bus and multiple voltage levels. We present an end-to-end implementation of a prediction framework for LVDC microgrids loads in a residential scenario using a data-driven approach. Given the limited availability of public DC microgrid datasets, a data augmentation pipeline is first proposed and deployed using open source software libraries. Results provide a comparative analysis between the SARIMA and Prophet load forecasting (regression) models using standardised evaluation metrics.

Keywords: Intelligent Systems
Abstract: This article presents methods for extracting data from documents in an automated process. The extraction of data is a critical step in automation, and supporting this technique is the UiPath platform, which includes the Document Understanding package. This package offers several types of extractors capable of retrieving data from various document formats, such as native PDFs or scanned documents. The purpose of data extraction is to identify and isolate relevant information from documents in order to process it further within a fully automated workflow

Keywords: Intelligent Systems, Machine Learning, Software Methods and Tools
Abstract: Intelligent decision making systems are being used more and more in precision agriculture to maximize input use and improve green practices. However, because of the complexity, computational expense, and lack of transparency of many current methods, small-scale farmers' adoption of such systems is still limited. In this paper, we compare a number of lightweight machine learning algorithms for fertilizer recommendation, focusing on their interpretability, simplicity, and suitability for application in low-resource agricultural environments.

By combining characteristics like soil type, crop type, season, and rainfall, a synthetic dataset is created to replicate authentic agricultural situations. Measures such as classification accuracy, computational efficiency, and model transparency are used to evaluate the evaluated models, which include K-Nearest Neighbors (KNN), Naive Bayes, Decision Trees, Singular Value Decomposition (SVD), and Gradient Boosted Trees implemented using the Smile library.

The findings provide an organized comparison of traditional recommendation techniques designed for agricultural applications, showing that Decision Trees strike the best balance between interpretability and predictive performance. These results support larger objectives in precision agriculture and intelligent decision support systems by helping to create easily navigable, low-complexity recommendation modules that can help small-scale farmers make informed fertilizer decisions.

Keywords: Software Methods and Tools, Cloud Computing, Machine Learning
Abstract: This paper presents the design, implementation, and evaluation of a real-time sound source localization system for Microsoft HoloLens 2, which aims to enhance spatial awareness through mixed reality feedback. The architecture follows a client-server model in which audio signals, captured by the built-in microphone array of Microsoft HoloLens 2, are preprocessed locally and sent to a remote Python server for inference using a machine learning model deployed in Google Vertex AI. The predicted sound direction is visualized directly in the user’s field of view through mixed reality cues, enabling intuitive spatial orientation without requiring external hardware. To assess system performance, we conducted a controlled experiment measuring response times under different combinations of server hardware and network configurations. Results demonstrate the feasibility of real-time audio-based localization in mixed reality and suggest that the system can be extended with multimodal feedback for more effective navigation in complex acoustic environments.

Keywords: Software Methods and Tools, Computational Methods, Communication Networks
Abstract: Networked programs and the tools to build them have become subjects of great interest with the rise of distributed computing and web services. A programming language built for such tasks is the Go programming language, which has a focus on high parallelism, a feature needed to serve large amounts of network requests. We created a functional extension to the Go programming language that allows the user to create pure functions that can be called from pre-existing Go programs. Our solution brings the strong points of functional programming such as function purity, partial function application, and higher-order functions while preserving the strengths already present in the Go programming language. The experimental results show that the LambdaGo programs perform as well as their Go counterparts in some instances while incurring some performance overhead in others due to the added abstractions.

Keywords: Quantum Information and Control
Abstract: This study aims to evaluate the capabilities of state-of-the-art quantum computing applications, with a particular emphasis on their effectiveness in solving optimization problems. Specifically, it focuses on Quantum Annealing algorithms, examining their potential advantages and inherent limitations, reviews the capabilities of existing models, replicates proposed solutions for addressing NP-Hard problems, and systematically compares the results obtained with those produced by classical algorithms. Through this analysis, the study seeks to provide a rigorous assessment of the practical effectiveness of quantum annealing in optimization tasks and its viability as an alternative to classical computational methods in the near future.

Keywords: Neural Networks, Machine Learning, Intelligent Systems
Abstract: Thyroid ultrasound is the investigation of choice for detection of thyroid abnormalities , especially when used in conjunction with thyroid function tests. Thyroid ultrasound could be an accessible screening option for the early identification of thyroid cancer. Through thyroid ultrasound, endocrinologists differentiate various aspects of thyroid pathology, such as: autoimmune thyroiditis, nodules, or the normal appearance of the gland. Deep learning neural networks have gained an increasing weight in medicine, having a promising role in assisting computer-based medical decisions. In this paper we decided to use deep neural networks to assist endocrinology physicians to better diagnose the thyroid pathologies. We decided to implement a neural network based on Xception model to help the endocrinologist in making medical decisions by offering a presumptive diagnosis based on the patient's ultrasound images. The simplified Xception network is a powerful option for medical image processing, especially pediatric thyroid ultrasound, but it is important to consider the specific dataset requirements for endocrinology data and to prepare properly to get the best possible results. Similar with the simplified Xception structure, the proposed neural network model was based on a 32-layer convolutional neural network (CNN) (31 convolutional layers and one fully connected layer) and was developed from scratch without pre-trained weights.

Keywords: Robotics, Control Applications, Mechatronics
Abstract: The paper focuses on the design of a control system for a class of grippers with continuum arms. The associated mathematical models for continuum arms are analyzed, insisting on the influence of dead times on driving performance. The model of the force exerted on the contact surface of the load is studied. The grasping function is obtained by coiling the terminal element around the load and exerting a corresponding pressure force on the load surface. The effects generated by dead times are discussed. Methods based the Lambert function are used for the dead time analyze and the stability of the system is investigated using Lyapunov techniques. The theoretical results are verified by numerical simulation.

Keywords: Robotics, Autonomous Systems
Abstract: This paper presents an embedded indoor localization and telemetry unit for mobile robots operating where GNSS signals are absent. Specifically, Ultra-Wideband ranging, inertial sensing, and Bluetooth Low Energy execute on a single microcontroller assembled from off-the-shelf modules. Since all estimation and telemetry run onboard, packets are transmitted at 10 ~Hz to a host PC for monitoring, logging, and fault diagnosis. Moreover, the firmware is configurable at build time, enabling component deactivation and continued operation after partial subsystem loss; concurrently, detectors report BLE drops, delayed packets, and sensor timeouts. During trials with ten autonomous 1/10 scale vehicles in a competition-style environment, the unit delivered >5 h of operation from a 1200 ~mAh cell, maintained BLE latency <100 ms, and kept position error <20 cm. Although the system relies on fixed UWB anchors, it requires neither custom PCBs nor an external localization server.

Keywords: Robotics, Multi - Agents Systems, Optimization and Optimal Control
Abstract: This paper addresses a pick-up and delivery vehicle routing problem for mobile robots motion planning where tasks are subjected to strict time windows and precedence constraints. The current formulation extends the classical problem by integrating obstacles in the environment, while also extending our previous work by incorporating multiple robots. Our solution comprises two main steps: the construction of a visibility graph to represent the environment and the subsequent reduction of this graph, followed by the formulation and solution of a Mixed-Integer Linear Programming problem (MILP) to compute efficient routes. The proposed methodology is supported through illustrative numerical simulations, validating its effectiveness in solving a multi-vehicle routing problem with pick-up and delivery tasks in an environment cluttered with obstacles.

Keywords: Machine/Reinforcement Learning, Control Applications, Robotics
Abstract: With the recent trend of integrating Machine Learning techniques in the robotics field in mind, this paper proposes a multi-objective optimization of PID control laws for a Nao robot. The two optimization techniques implemented, namely the Multi-Objective Genetic Algorithm and the Multi-Objective Bayesian Optimization, are analyzed and compared. For testing and validating the solutions obtained with both methods, a Digital Twin of the Nao robot was utilized. The key benefit of these approaches is the ability to optimize multiple objectives individually and concomitantly, opening further research perspectives in robotics.

Keywords: Robotics, Signal Processing, Embedded Systems
Abstract: Abstract—This paper presents the development of a charge coupled device sensor module used for mobile robots, such as line follower robots. This kind of robots are mostly used in robotics competitions where the goal of the robot is to follow a line on a contrasting background and to be capable to detect discontinuities or any obstacles on the track. The experience of the last competitions leaded to needing an improvement and an update from the typical sensor bar to a charge-coupled device based sensor module. This module, despite his reduced dimensions, comes with features that will provide more efficiency, stability and greater “vision” to the robot. The following sections will describe each step, from the electrical design to the software development and simulations

Keywords: Communication Systems, Communication Networks, Information Systems Applications
Abstract: This paper presents a communication framework for connected robots in a platoon formation, designed as a platform for real-time testing and validation of control algorithms. The framework utilizes ZigBee-based wireless communication, ensuring reliable, low-power, and efficient data exchange between robots. It supports scalable and adaptable communication, allowing efficient integration into various multi-robot applications while maintaining robustness in dynamic environments. The proposed system facilitates real-time monitoring and coordination, providing a robust environment for evaluating different strategies for platoon formation and cooperative decision-making. This framework provides a versatile solution for developing and validating control strategies in connected robotic systems by ensuring structured message handling, efficient network management, and optimized data processing.

Keywords: Machine/Reinforcement Learning, Robotics, Control Systems Design
Abstract: Recent advances in robotics and artificial intelligence have driven the rapid development of sophisticated solutions for automating complex tasks. A key example is mobile robots with manipulator arms, which require efficient training methods to perform intricate operations. The integration of reinforcement learning algorithms has significantly enhanced the ability of these robots to learn optimal actions through environmental interaction. Central to this approach is the reward mechanism, which guides behavior adaptation and task-specific performance improvements.

This paper explores applying reinforcement learning techniques to mobile robots with manipulator arms, focusing on improving their adaptability and operational efficiency in real-world environments. The proposed framework leverages reward-based learning to optimize decision-making, enabling autonomous systems to handle diverse and dynamic tasks more robustly. Experimental results demonstrate the potential of this approach to advance autonomous manipulation and navigation capabilities in complex scenarios.

Keywords: Machine Learning, Computer Vision, Autonomous Systems
Abstract: Accurate identification of road surface types plays a vital role in enabling autonomous vehicles to adapt their driving behavior and maintain safety under varying environmental and road conditions. Current systems struggle with imbalanced datasets and visually similar surfaces. This paper proposes an early fusion approach integrating supplementary channels with RGB data in modified ResNet-50 architectures. The input layer processes 4-channel inputs, combining RGB with either grayscale or gradient information, enabling the network to capture complementary features from the earliest processing stage. Experiments on balanced and imbalanced datasets demonstrate that the grayscale early fusion configuration consistently outperforms gradient fusion and standard RGB-only approaches. The grayscale configuration particularly excels in distinguishing between visually similar road surfaces and shows enhanced robustness to imbalanced class distributions. These results demonstrate that early fusion with grayscale information significantly improves road surface classification accuracy, addressing critical challenges in autonomous vehicle perception systems.

Keywords: Machine Learning, Cyber - Security, Neural Networks
Abstract: Web application firewalls (WAFs) play an essential role in protecting web applications against security threats. However, traditional rule-based WAFs often struggle to adapt to evolving attack patterns. This paper investigates the application of machine learning (ML) models for enhancing WAF performance, with a specific focus on Recurrent Neural Networks (RNN), Gated Recurrent Units (GRU), and Long Short-Term Memory (LSTM) networks. Initially, an RNN model was explored, but subsequent experimentation with GRU and LSTM architectures produced superior results in terms of detection accuracy and robustness. The models were trained on a dataset consisting of web traffic logs, and performance was evaluated using metrics such as accuracy, precision, and F1-score. The obtained results demonstrate that LSTM and GRU outperformed the RNN in handling long-term dependencies, offering better generalization and detection capabilities for complex attack patterns. The findings suggest that advanced deep learning techniques can significantly improve the efficacy of WAFs, reducing false positives and enhancing security measures. Future work will explore further optimizations and the potential integration of more advanced models to strengthen WAF systems.

Keywords: Machine Learning, Web services and applications, Neural Networks
Abstract: The paper aims to develop a web application which analyzes online articles to detect clickbait. Using advanced Natural Language Processing techniques and Artificial Intelligence together with sentiment analysis technique, the proposed method provides for users the ability to verify the authenticity of the news. After selecting an article posted on a news website, the title, the content and the connection between them are analyzed, relevant statistics are generated and used to classify the article. The tests carried out revealed a very good accuracy for the method proposed.

Keywords: Machine Learning, Neural Networks, Computer Vision
Abstract: Lip reading is a promising field in accessibility and assistive technologies. It involves investigating video sequences and tracking the movements of the speaker's lips to decipher the text they speak. This paper explores encoder-decoder architectures tailored for this task, where the encoder extracts visual features that effectively capture both spatial and temporal information, and the decoder, based on Long Short-Term Memory (LSTM) cells, processes the sequence of visual features to generate the corresponding text sequence. Within this framework, we introduce models based on early fusion, which incorporate additional feature maps into the input arrays to support and enhance visual feature extraction. The primary objective is to investigate the impact of this early fusion approach in combination with different types of encoders that perform either spatial or spatio-temporal analysis. The encoder architectures considered include 2D convolutional neural networks (CNNs), 3D CNNs, and Transformers of varying complexity. Furthermore, two variants of the LSTM decoder are explored: one that imposes a fixed output sequence length through the architecture, and another that allows for flexible interpretation of the output sequence via the loss function. The model is trained end-to-end and evaluated on the GRID Corpus dataset.

Keywords: Machine Learning, Neural Networks, Computer Vision
Abstract: The application aims to use deep learning technologies to analyze and interpret visual signals of sign language. Sign language is a communication system that includes movements of the hands, face and body. It is used by people who cannot speak, but also in everyday communication between people, especially when encrypted communication is desired. The aim is to detect and understand sign signals using machine learning algorithms and deep neural networks. These signals are usually transmitted through movements of the limbs and facial expressions. The detection process involves identifying and classifying these movements with the aim of translating their meaning. For this purpose, the paper suggests a variant of SlowFast neural network that combines ResNet50 or AlexNet and LSTM (Long Short-Term Memory) neural networks.

Keywords: Machine Learning, Neural Networks, Fault Diagnosis and Fault Tolerant Control
Abstract: The importance of anomaly detection in time-series data is widely acknowledged in diverse sectors such as industrial processes, healthcare systems, and financial markets. Solutions to this problem are diverse but can be broadly classified as either classical statistical methods or more recent machine learning techniques that leverage data to detect anomalies. The primary goal of this paper was to examine the strengths and weaknesses of a specific type of anomaly detection that relies on data reconstruction using autoencoders. To this end, we sought to evaluate the relative performance of various latent space modelling strategies employing a consistent BiLSTM architecture for the encoding and decoding components. The results indicate that reducing the latent space dimensionality without additional structural modifications leads to degraded performance. This degradation is more pronounced in deterministic methods compared to stochastic methods, which demonstrate better generalisation capabilities even at smaller latent space sizes. Additionally, the study reveals that incorporating a GAN-based methodology into reconstruction methods does not significantly enhance performance but can help when the design is cohesive. Also, when an appropriate threshold is selected, reconstruction methods achieve a high degree of reliability in minimising false negatives, where anomalies are incorrectly classified as normal.

Keywords: Neural Networks, Machine Learning, Computer Vision
Abstract: Person re-identification aims to match individuals across different camera views and lighting conditions, which is crucial for various applications in surveillance and security. However, traditional re-identification methods often rely on training and testing within the same dataset, limiting their generalizability to real-world scenarios characterized by domain shifts, e.g. illumination variations, and different camera viewpoints. In this paper, we investigate the performance of established person re-identification methods when trained on one dataset and evaluated on multiple different datasets, thereby providing a more realistic assessment of their robustness and cross-domain generalization ability. Our experimental setup involves a comprehensive comparison of several state-of-the-art deep neural models, trained on some datasets and tested on different datasets. We employ standard evaluation metrics such as rank-1 accuracy and mean average precision to quantify performance disparities. Our results reveal significant variations in methods effectiveness across domains, which further enabled us to formulate valuable insights related to the factors influencing their robustness. The findings offer valuable directions for future research in person re-identification.

Keywords: Nonlinear Systems
Abstract: Exact feedback linearization serves as a fundamental method for nonlinear control, as it enables the construction of a reversible coordinate transformation which maps the original nonlinear system into a linear one. This thereby allows the application of standard linear control techniques. However, its application relies on certain prerequisites, notably the system's relative degree, i.e., pole-zero excess, being well-defined and constant. A key challenge arises when this condition is not met, leading to so-called singularity cases, which can obstruct the straightforward implementation of the linearization technique. The scope of the current paper is to propose a composite control law, comprised in a state feedback linearization in addition to a control barrier function-type auxiliary command to compensate residual nonlinearities by avoiding such domains from the state-space. We use the classical ball and beam process benchmark from the literature. The development of the composite control law is formulated in terms of linear matrix inequalities, leading to a convex solution. The potential of the method is illustrated through numerical simulation, by emphasizing problematic initial conditions for the ball and beam process.

Keywords: Nonlinear Systems
Abstract: This paper examines a broadly applicable triangular normal form for x-flat control-affine systems with two inputs. First, we show that this triangular form encompasses a wide range of established normal forms. Next, we prove that any x-flat system can be transformed into this triangular structure after a finite number of prolongations of each input. Finally, we introduce a refined algorithm for identifying candidates for x-flat outputs. Through illustrative examples, we demonstrate the usefulness of our results. In particular, we show that the refined algorithm exceeds the capabilities of existing methods for computing flat outputs based on triangular forms.

Keywords: Other Topics
Abstract: This paper presents two pseudo-random number generators (PRNGs) based on chaotic dynamical systems, specifically the Lorenz system and a derived Three-Scroll attractor. Designed with simplicity and modularity, the proposed generators are implemented in Java and evaluated in terms of generation speed and statistical randomness using the National Institute of Standards and Technology (NIST) test suite. A detailed parametrization strategy is employed to optimize bit extraction from floating-point representations. While the proposed methods do not yet meet the criteria for cryptographically secure generators (CSPRNGs), the results indicate strong potential for use in applications where moderate security and high performance are required, such as embedded systems or lightweight cryptographic protocols. The design also enables easy customization for further research and adaptation.

Keywords: Nonlinear Systems, Control Applications, Control Systems Design
Abstract: Armature current control of a DC machine- using a bidirectional DC-DC converter- is proposed in this paper for load torque emulation to test electrical drives used in electric vehicles. A detailed discussion on the validity of the linear ripple condition is provided for the given circuit parameters. The state-space averaging method is applied to determine the steady-state circuit variables in the presence of non-ideal switching devices. A Lyapunov-based control law is then derived using the results obtained from state-space averaging, and its performance is verified through MATLAB Simulink simulations.

Keywords: Nonlinear Systems, Computational Methods, System Identification and Modeling
Abstract: The accessibility of a system is a necessary condition for its controllability. This property is linked to the rank of the Lie algebra generated by the drift and input vector fields. In case of polynomial dynamical systems the Lie subalgebra can be computed using methods of algebraic geometry. An interesting system falling into this class is the rotation of a rigid body, in particular its angular momentum. We attempt to decide the accessibility of this system in dependence of the inertia parameters for simple cases.

Keywords: Nonlinear Systems, Uncertain Systems, Industrial Applications
Abstract: This paper focuses on set-based transient rotor angle stability analysis of power systems, in which a complex power network can be decomposed into smaller subsystems through a bounded decoupling variable. This way, analyses of decomposed power system elements can be conducted individually, eventually incorporating them into a statement about the whole system’s stability. The computational problem is to find lower and upper limits for the disturbance inputs so that an appropriate invariant set exists for each generator. Using the observed properties of barrier phases, an iterative procedure is proposed to find a feasible set of angular constraints. The operation of the method is illustrated on an example taken from the literature.

Keywords: Multi - Agents Systems
Abstract: We introduce the blended dynamics theorems in various forms, which clarify how enforced synchronization against heterogeneity in a networked dynamical system gives rise to emergent behavior. This finding naturally leads to engineering applications such as distributed state estimation and distributed decision making. The blended dynamics theorems also reveal how robustness emerges in multi-agent systems under various sources of uncertainty. Finally, we illustrate how enforced synchronization induces adaptation in multi-agent systems.

Keywords: Data Management Systems, Information Systems Applications, Distributed Systems
Abstract: The Digital Product Passport (DPP) is envisioned as a key enabler for circular economy objectives, aiming to provide standardized and lifecycle-based product information across value chains. While a complete DPP system involves complex data models, identity frameworks, and regulatory integration, this paper focuses on a specific technical aspect: evaluating the feasibility of secure and policy-controlled data transfer using the Eclipse Dataspace Connector (EDC). A working prototype was developed to test the actual exchange of structured product data between economic operators, based on pre-negotiated contracts and HTTP push mechanisms. The implementation demonstrates core data space principles such as data sovereignty, decentralized architecture, and contract-based governance. Although limited in scope, the system validates EDC’s suitability for DPP-related data sharing and sets the stage for future extensions toward full-featured, interoperable DPP systems.

Keywords: Distributed Systems, Cyber - Security, Agent - Based Systems
Abstract: This paper aims to present an efficient data security mechanism through the implementation of Transparent Data Encryption (TDE) in the case of crypto-ransomware variant of attacks. The proposed solution relies on a Man-in-the-middle (MITM) proxy, which enables dynamic encryption and decryption of data both in transit and at rest, using state-of-the-art authenticated encryption algorithms AES-GCM and ChaCha20-Poly135, which are dynamically selected according to the characteristics of the files, minimizing the performance impact on the system. For a secure management of cryptographic keys, a vault service was integrated, enabling the development of a Zero Knowledge architecture. Ransomware detection process adds minimal overhead compared to the traditional analysis methods. The proposed architecture provides improved data privacy and resilience against exfiltration attacks or data leaks. The stored data is fully encrypted and lacks identifiable metadata, enhancing protection against unauthorized access. Preliminary results confirm the effectiveness of the proposed solution and demonstrate the benefit of its integration into GlusterFS and HDFS platforms by substantially enhancing data security and protection within distributed systems as storage agnostic encryption security layer.

Keywords: Computational Methods, Biologically Inspired Systems
Abstract: Genetic algorithms can tackle many problems that are otherwise intractable, yet their most notable drawback is the lengthy time needed to converge. This paper investigates how GPU acceleration can address this challenge and reduce the time genetic algorithms require to produce a solution. Using the Traveling Salesman Problem as a case study, we isolate the two most expensive stages, construction of the distance matrix and per-tour fitness evaluation, and port them to CUDA. We then introduce two kernel-level optimizations for the distance matrix computation. Shared memory tiling caches column coordinates on the chip, cutting global memory traffic and shrinking distance matrix time by 38.7%. Then, multistream execution with pinned host memory overlaps data transfers with computation, trimming end-to-end matrix build time by 28.5%. The fitness function, when ported to the GPU, achieves up to an order of magnitude speedup for medium-sized instances and almost doubles performance for larger cases, compared to the CPU baseline. All experiments were conducted on an Intel Xeon CPU and an NVIDIA Tesla T4 GPU, using randomly generated datasets ranging in size from 100 to 10,000 locations. The proposed techniques generalize to any evolutionary or swarm method whose fitness decomposes into pairwise calcu- lations, offering a practical recipe for scaling population-based optimizers on modern GPUs.

Keywords: Web services and applications, Distributed Systems, Cloud Computing
Abstract: EduPlay is an interactive educational platform that reimagines traditional homework as a personalized, engaging, and secure experience for middle school students. Combining blockchain technology, gamification, and decentralized storage via IPFS (InterPlanetary File System), EduPlay ensures transparency, authorship traceability, and student motivation. Learners complete tasks using an intuitive visual canvas, and in collaborative projects, each user’s contribution is automatically recorded and verifiably linked to their identity on the blockchain.

A central feature of EduPlay is its use NFTs (Non-Fungible Tokens) as immutable certificates of achievement. Blockchain is also used for tracking collectible cards within a trading system. Students are encouraged to build collections, trade cards with peers, and unlock learning incentives, making academic progress tangible and engaging.

Teachers can create and evaluate assignments with customizable templates and receive real-time insights into student performance. Parents gain visibility into their child’s progress through a dedicated dashboard. EduPlay promotes creativity, collaboration, and transparent learning outcomes in a gamified, student-centered environment.

Keywords: Web services and applications, Distributed Systems, Cloud Computing
Abstract: In response to the growing demand for secure, transparent, and adaptive digital education systems, this paper presents a novel decentralized platform that redefines how learning outcomes are assessed, stored, and certified. The proposed architecture integrates blockchain technology, dynamic non-fungible tokens (dNFTs), and artificial intelligence to establish a verifiable and learner-centric educational ecosystem. Decentralized storage is handled via InterPlanetary File System (IPFS), secured through hybrid encryption and controlled access, ensuring that sensitive data such as assessments and performance records remain both confidential and tamper-proof.

A distinctive feature of the platform is the use of dynamic dNFTs to represent and continuously update each learner’s academic progress. Unlike static certificates, these credentials evolve as students complete assessments and advance through course material, offering a transparent and verifiable record of learning achievements anchored on the blockchain. Each dNFT encodes metadata such as course names, skill levels, and timestamps, ensuring long-term integrity and resistance to tampering. Complemented by an AI module for personalized test generation, this model replaces traditional certification with an adaptive, decentralized alternative that enhances trust, autonomy, and credential portability across educational and professional domains.

Keywords: Distributed Systems, Information Systems Applications, Web services and applications
Abstract: In the rapidly evolving digital landscape, ensuring trust, transparency, and security in online collaborations remains a significant challenge, particularly for innovators and experts engaged in knowledge exchange. The proposed SPARK-IT platform leverages blockchain, AI-driven matchmaking, decentralized identity management, and tokenomics to foster a secure innovation ecosystem. By utilizing a permissioned blockchain, smart contracts, and decentralized storage, SPARK-IT ensures intellectual property protection, traceability of contributions, and non-repudiation in mentor-innovator engagements.

This paper presents the technical architecture of the platform, demonstrating how distributed ledger technology and AI-driven methodologies can establish a human-centered, sustainable and trustworthy online innovation ecosystem. By bridging academia, startups, and industry, SPARK-IT redefines digital trust and collaboration in the innovation economy.

Keywords: Control Applications, Linear Systems
Abstract: The buck converter is a very common switched mode DC-DC converter. This converter is designed to reduce an input voltage supply to a lower output voltage. The generation of a voltage higher than the input voltage requires other converter topologies such as the boost converter. However, with a minor modification, the buck converter can simultaneously provide both a lower and a high output voltage. In this paper we derive a model of such a modified buck converter. It is also shown how these voltages can be adjusted largely independently of each other by suitable control of the converter.

Keywords: Distributed Parameter Systems, Control Systems Design, Control Applications
Abstract: An improved model for the 1D mechanical systems with linear motion coordinates is obtained emph{via} the Hamilton variational principle, incorporating the elastic strain at the coupling of the driving motor with the driven system. Consequently it was possible to use a linear dynamic compensator subject to a weak frequency domain inequality and to obtain asymptotic stability from a ``weak'' energy-like Lyapunov functional by applying the Barbashin-Krasovskii LaSalle invariance principle.

Keywords: Control Applications, Stochastic Systems, Distributed Systems
Abstract: In high-performance computing (HPC), many research has been conducted on how to efficiently utilize the idle time of HPC resources (periods when no jobs are submitted to the platform). One promising approach is to exploit this idle time by injecting small, flexible, independent, and interruptible jobs that have no strict time constraints. However, managing the injection of these jobs is challenging due to the stochastic nature of job parameters, such as execution times and resource consumption. Additionally, process noise—resulting from system complexity and the arrival and execution of varying external workloads—can interrupt or terminate these filler jobs. In this paper, we propose and evaluate, using real data, a Stochastic Model Predictive Control (SMPC) approach that addresses system uncertainty and incorporates a feed-forward compensation mechanism for disturbance rejection. The proposed algorithm shows promising results: it ensures a platform usage rate of 98%, significantly improving overall resource efficiency, and reducing the number of early terminated jobs compared to previous work.

Keywords: Control Applications, Adaptive and Robust Control
Abstract: Driven by the Pharma 4.0 vision and the growing shift toward continuous manufacturing, the pharmaceutical industry increasingly requires advanced control strategies that ensure real-time quality, flexibility, and regulatory compliance. The rotary tablet press plays a central role, directly influencing critical quality attributes. Due to its tightly coupled mechanical stages and sensitivity to material properties, it presents significant challenges for process understanding, modeling, and control. We propose a robust control weighting scheme to ensure consistent performance across varying conditions. This approach employs the Glover-McFarlane open-loop shaping synthesis, utilizing a simple tuning method with two degrees of freedom: the decoupling frequency, determined via a matrix spectral alignment algorithm, and the open-loop gain to adjust response aggressiveness. The process model's linear approximation is derived through system identification based on experimental data from a high-fidelity digital twin of the actual process. The performance achieved with the resulting suboptimal regulator is illustrated through numerical simulation.

Keywords: Control Applications, Adaptive and Robust Control, Control Systems Design
Abstract: Computer numerical control (CNC) systems require high precision, together with a fast computation of the command signals, to ensure low overhead, irrespective of the uncertainties inherent to the electromechanical system, its nonlinearities and its load conditions. We consider a robust control design scheme for a two-axis CNC machine which encompasses the parametric uncertainties of each individual axis and their cross-coupling effects, which also minimizes the sensitivity of the resulting regulator to the influence of the sampling rate of the controller and quantization effects caused by the encoder measurements and pulse width modulation command inputs. We impose a fixed structure of a simple cascade of proportional and proportional–integral regulators, as this structure can be universally implemented on most CNC drivers. The proposed regulator design further includes the delay caused by the zero-order hold circuit, while the sensitivity analysis of the numerical effects is quantified by the areas of the closed-loop poles spanned in the complex plane. The performance of the closed-loop system is illustrated through numerical simulation.

Keywords: Robotics, Control Applications, Neural Networks
Abstract: In recent years, medical devices utilizing soft robotic solutions have gained popularity for small, minimally invasive, and complex procedures. Another critical category is represented by prostheses and recovery devices, mainly implemented using small servomotors and complex mechanical assemblies. Soft-actuated prostheses are needed for safer and more elastic prosthetic devices and wearable technologies. In these cases, a classical control solution, using a computational unit with limited speed, becomes rigid and uncomfortable. Furthermore, most procedures require a calibration process, numerical evaluation, and rigorous testing before they can be used. The solution proposed in this research is based on a hardware spiking neural network that can control the bending motion of a soft hybrid actuator. The bending is considered according to a natural finger flexing, commanded by the human neural system through small electrical signals closer to the network behavior. The solution offers a promising response speed, not limited by a digital frequency or numerical determinism. Furthermore, the solution offers out-of-the-box functionality, eliminating the need for additional setup. This research is sustained by experimental results guided by a novel design in controlling the motion of hybrid soft actuators.

Keywords: Robotics, Manufacturing Systems, Computer Vision
Abstract: Advancements in robotics increasingly emphasize close interaction with the environment, often integrating advanced sensing and processing to enhance perception, autonomy, and situational awareness. However, unsupervised operation can introduce unpredictability and risks for human operators. Enhancing environmental awareness is, therefore, essential for both safety and performance. This research proposes an architecture for recurrent scene capture, enabling the integration of environmental features into a Digital Twin. The system repeatedly detects environmental changes and reconstructs the recognized objects for use in both virtual and physical spaces. The proposed architecture addresses the need for a perception approach capable of understanding its environment, which is critical to ensuring flexibility and safety in industrial settings. Validation is performed in two scenarios: avoiding the object detected if it interferes with a robot task, and transferring object awareness and manipulation from virtual space to a real robot. The experimental results demonstrate the method's potential for developing safer, more adaptive and flexible collaborative robotic systems.

Keywords: Computational Methods, Autonomous Systems, Robotics
Abstract: Autonomy is a key feature in robotic systems, especially for tasks that require minimal or no supervision. Reliable operation in such contexts depends on methods that account for environmental complexity, resource limitations, and energy efficiency. For mobile robots and aerial drones, autonomy is closely tied to navigation and spatial awareness. Success often hinges on how efficiently a robot can plan and follow a path through its environment. This paper focuses on trajectory generation, a core aspect of autonomous motion planning. BiSpiral is a general-purpose method designed for complete area coverage in convex environments. The proposed approach constructs a spiral-shaped path by leveraging the geometric properties of convex shapes, enabling uniform coverage while minimizing redundancy and optimizing space usage. The current research presents the method’s conceptual design, algorithmic implementation and results from experiments conducted on various convex layouts compared with the classical spiral trajectory approach and validated using the Crazyflie 2.1 UAV under laboratory conditions.

Keywords: Robotics, Mechatronics, Human - Computer Interaction
Abstract: Even though robots have been developed that can create artistic paintings under the new developments brought by artificial intelligence, human painters, through the "soul" they dedicate to the artistic act, bring a certain uniqueness to the works, something that is currently unmatched by technology. However, the use of robots can help human artists to reduce their working time, which is largely occupied with operations in the preparation stage, such as multiplying certain models or cutting stone for mosaics. This paper studies the possibility of creating a low-cost robot for making templates, useful both for professional painters and for beginners who will be helped to learn certain elementary techniques. In the initial stage, an elementary model was simulated, based on an Arduino Uno development board. Starting from the simulated model, a new model was proposed by adding an SD card module on which the G-code files with the models to be drawn are loaded. An LCD display and buttons for selecting the desired file were also added. This model may be challenging for users without technical skills. Therefore, a new constructive variant was proposed, in which the Arduino board was replaced with an ESP32 development board, with a built-in WiFi module and can be connected directly to a smartphone, on which a dedicated mobile application runs. The robot was tested to multiply some templates for mosaic work. The results showed that the total working time was reduced by approximately 20 hours from 25 hours by using the equipment, compared to the situation of making the entire template only manually. As a future direction of research, the addition of a stone cutting module is considered, this being also an operation that, if done manually, takes up a lot of time.

Keywords: Robotics, Computational Methods, Other Topics
Abstract: In this paper, we revisit the path planning approach based on convex lifting and extend it to accommodate non-convex obstacles. Obstacles are modeled as overlapping unions of polyhedra. To handle complex environments, a preprocessing module is required to reconstruct the obstacles, and convex lifting is used to obtain an interconnected graph. A post-processing module will then is then introduced to remove graph edges that intersect obstacles, thereby extracting only feasible paths. We propose two obstacle representations that are compatible with the convex lifting framework, significantly broadening its applicability. Furthermore, we address the potential loss of connectivity resulting from the removal of obstacle-intersecting edges and propose a solution to maintain connectivity and thus ensure the preservation of collision-free paths.

Keywords: Robotics, Optimization and Optimal Control, Manufacturing Systems
Abstract: This project aims to develop a Delta robot designed for educational purposes, focusing on optimizing the mechanical configuration to achieve an efficient workspace and minimize the required motor torque. The article includes CAD modeling and additive manufacturing of components, followed by a comparative analysis of multiple geometric configurations to identify the optimal variant.

Keywords: Robotics, Autonomous Systems, Other Topics
Abstract: This paper focuses on the coverage problem for UAVs aiming for trapezoidal cells returned by a cell decomposition technique applied for obstacle avoidance. The classical coverage algorithms, ZigZag and Spiral, are implemented and objectively evaluated based on performance metrics including coverage percentage, overlap percentage, path length, and number of turns. In addition, two premises have been analyzed, considering safe and non-safe coverage strategies, specifically for trapezoidal workspaces. The results of this evaluation enable the extraction of comparative conclusions between the two algorithms, which contribute to a more structured and informed coverage problem approach.

Keywords: Neural Networks, Other Topics
Abstract: In the contemporary digital era, the necessity for swift and reliable diagnostic tools in medical imaging has become increasingly paramount. However, brain magnetic resonance imaging analysis poses significant challenges, including high intraclass variability, low interclass contrast, and the need for expert annotation. This study presents the development of an intelligent diagnostic system based on a custom Convolutional Neural Network architecture, specifically designed to assist in the detection and classification of brain tumors. The system processes magnetic resonance images to distinguish between glioma, meningioma, pituitary tumors, and healthy brain tissue. The proposed model was trained and evaluated on a benchmark dataset and compared with several established architectures. Experimental results demonstrate that the custom architecture achieves comparable classification accuracy while showing increased computational efficiency, offering a promising tool to support clinicians with fast, interpretable, and robust diagnostic predictions. Furthermore, the results highlight important considerations for selecting specific types of layers, a process substantiated by mathematical arguments presented throughout the paper.

Keywords: Neural Networks, Other Topics
Abstract: This paper proposes the development and evaluation of advanced machine learning models specifically designed for the analysis of facial emotions extracted from video recordings, with the objective of identifying and classifying potential mental disorders. In the context of the increasing need for early diagnostic methods, the use of convolutional neural networks is proposed to extract and interpret relevant emotional features. The system seeks to correlate expressive patterns with psychiatric conditions such as depression, schizophrenia, or anxiety disorders. The goal is to adapt these neural networks to accurately interpret emotional fluctuations and subtle expressions indicative of changes in mental states. The study employs diverse and relevant datasets to adequately cover a wide range of facial expressions. Furthermore, efforts are directed toward adjusting the neural network architecture and tuning its parameters to enhance model accuracy under various analytical conditions.

Keywords: Machine Learning, Neural Networks, Computer Vision
Abstract: Deploying artificial intelligence (AI) on legacy robotic platforms is increasingly relevant as industries and educational institutions continue to rely on aging, yet functional hardware. This paper presents a novel approach to running an emotion recognition system entirely on the NAO robot using the deprecated C++ SDK, an environment no longer actively supported or maintained. The implementation required reconstructing a full development and cross-compilation toolchain on an end-of-life Linux distribution, rebuilding libraries manually, and resolving numerous compatibility issues. Despite these significant obstacles, a fully local, real-time solution was achieved using classical machine learning methods and OpenCV, avoiding reliance on cloud or external compute resources. The result demonstrates not only technical feasibility but also the enduring value of legacy systems when paired with lightweight, carefully optimized AI models. This work contributes a replicable blueprint for extending the life and capabilities of constrained robotics platforms through persistent engineering effort.

Keywords: Machine/Reinforcement Learning, Computer Vision, Neural Networks
Abstract: The development of intelligent tools capable of analyzing images in real conditions constitutes a major challenge in the field of computer vision. In this paper, we present a hybrid deep learning architecture, named DenseSwinNet, which combines the DenseNet201 convolutional neural network for visual feature extraction with the Swin Transformer V2 for classification. The aim of this approach is to improve the accuracy, robustness and generalizability of automatic detection systems. Our model is based on adaptive attention mechanisms to efficiently process contextual and visual variations. To evaluate the effectiveness of this architecture, we applied it to a case study on the identification of potato foliar diseases. The system is designed to automatically identify and categorize leaves into eight groups, including various diseases such as Viruses, Bacteria, Alternaria, Fungi, Parasites, Nematodes, Phytophthora, as well as healthy leaves. The experiment, conducted on a mixed dataset (controlled and uncontrolled environments), achieved an average accuracy of 99.24% in cross-validation — a result that confirms the effectiveness of our approach. Therefore, this contribution provides a high-performance and flexible IT solution that helps advance smart farming and improve the way crops are managed.

Keywords: Intelligent Systems, Machine Learning, Neural Networks
Abstract: This study introduces a method for magnetic field-based Simultaneous Localization and Mapping (SLAM), utilizing an Unscented Kalman Filter (UKF) in combination with a Physics-Informed Neural Network (PINN) model. The aim is to improve the accuracy and efficiency of magnetic field-based localization in environments with spatially varying magnetic characteristics. The proposed approach leverages an offline training of a PINN to model the magnetic field by embedding physical constraints directly into the learning process, resulting in a compact and computationally efficient representation. The dynamic system states are then estimated using UKF design. Simulation results validate the effectiveness of the proposed approach in accurately estimating the true system states, with considerable improvement on the localization and tracking accuracy compared to the current state of the art approaches.

Keywords: Machine/Reinforcement Learning, Neural Networks, Optimization and Optimal Control
Abstract: Prompt optimization for Large Language Models (LLMs) in black-box settings—where model gradients and likelihoods are inaccessible—is a major challenge in applied natural language generation, especially for educational content like video script generation. In this paper, we present textbf{NovaEduPTuneBO}, a novel likelihood-free Bayesian optimization framework for prompt tuning that integrates Gaussian Processes (GP), Markov Chain Monte Carlo (MCMC) sampling, and a Deep Neural Network (MLP) encoder. The MLP encodes prompts into compact representations and is trained online during the optimization loop to guide sampling more effectively. Experiments conducted on standard NLP benchmarks demonstrate that NovaEduPTuneBO outperforms prior approaches (Hyper-BO, Prompt-BO), confirming its robustness and suitability for educational applications. Our framework sets the foundation for adaptive educational content generation with LLMs, particularly in constrained real-world deployment scenarios.

Keywords: System Identification and Modeling, Machine Learning, Fault Diagnosis and Fault Tolerant Control
Abstract: Voltage and temperature estimation in Li-Ion batteries is crucial for battery health monitoring and performance optimization. This article presents how to use a data-driven solution to better estimate the overall voltage and temperature of Li-Ion batteries. The main focus is on estimating the Open-Circuit Voltage (OCV) and the temperature when discharging with constant current. Also, the paper studies the degradation of the model parameters over time, which can lead to better predictions that keep track of the number of discharging cycles and can act as a safety feature that the Battery Management System (BMS) can use to protect the battery pack. The results are compared with those obtained with the existing state-of-the-art models.

Keywords: Autonomous Systems, Fault Diagnosis and Fault Tolerant Control
Abstract: In the context of the ongoing migration of the population to urban areas and the increasing demand for sustainable solutions for cultivating plants in confined spaces, the smart vertical garden is emerging as a significant alternative. The development of a smart vertical garden that utilizes an advanced automated irrigation system, integrating monitoring and control technologies, is presented. This system is designed to optimize water consumption and enhance efficiency in plant cultivation. The intelligent system has been developed to address the challenges posed by constrained urban environments, offering an innovative solution that promotes resource conservation and efficient use of space. The research encompasses not only technological aspects but also reliability calculations employing Reliability Block Diagrams (RBD) and Markov chains. This enables the assessment of the reliability and durability of system components, with failure intensity and mean time between failures serving as parameters

Keywords: Fault Diagnosis and Fault Tolerant Control, Neural Networks, Machine Learning
Abstract: This paper presents a comparative evaluation of a custom Radial Basis Function (RBF) Network and a classical RBF Network applied to the detection of mechanical faults in wastewater treatment plants (WWTPs). The custom RBF Network introduces adaptive sigmas for each center, providing greater flexibility to adjust to the complexities and variations inherent in the fault data. In contrast, the classical RBF model uses a uniform sigma value across all centers, potentially limiting its ability to identify complex patterns in the data. In order to ascertain which network is best for fault detection in WWTPs, both models are assessed using performance metrics (accuracy, precision, recall and F1-score). According to the results obtained from the simulations, the custom RBF performs better than the conventional RBF one on every metric, demonstrating its improved capacity to identify mechanical faults in WWTPs.

Keywords: System Identification and Modeling, Linear Systems, Signal Processing
Abstract: This paper presents the online identification, based on finite impulse response filter coefficients, of a DC motor. The coefficients are obtained based on least mean-square identification using experimental data. They are used to construct the Hankel matrix based on which the mathematical model is determined. The results are subsequently compared with a standard method in Matlab. The method is applied in real time directly on a low-cost development board, where it successfully replicates the identification process previously tested in simulation. The results obtained online confirm the accuracy and reliability of the approach in a real-time setting. This work bridges the gap between theoretical system identification and practical real-time implementation, enabling motor identification with accessible hardware. The proposed approach supports rapid prototyping, educational use, and cost-effective industrial applications, particularly in scenarios that require real-time system monitoring.

Keywords: Fault Diagnosis and Fault Tolerant Control, Automotive Control Systems, Neural Networks
Abstract: Enhancing driving safety systems is a critical objective in the automotive industry, with advanced driver assistance systems and fallback strategies playing essential roles in accident prevention. This paper uses machine learning techniques to investigate underbrake anomalies, i.e., scenarios where braking performance is unexpectedly diminished. A dataset specifically simulating underbrake scenarios was created through Hardware-in-the-Loop (HiL) simulations, replicating real-world driving conditions, including both hydraulic and non-hydraulic braking systems.A systematic comparison of the performance of several neural network architectures, including Multi-Layer Perceptrons (MLP), Convolutional Neural Networks (CNN), Long Short-Term Memory (LSTM) networks, and hybrid CNN-LSTM models, was performed on this dataset. Results indicate that the LSTM architecture achieves the highest accuracy and F1 scores in detecting anomalous braking behavior, outperforming baseline models. However, further improvements in accuracy, computational efficiency, and generalization are needed for real-world deployment. This study establishes a robust benchmark for underbrake anomaly detection and highlights future directions, including enhanced feature extraction techniques and real-time implementation considerations.

Keywords: System Identification and Modeling, Linear Parameter - Varying Systems, Data - Driven Control
Abstract: Closed-loop system identification is essential for controlling and monitoring high-order mechatronic systems under real-time constraints. This paper presents a comparative study of four identification methods: Closed-Loop Output Error (CLOE), Observer/Kalman Filter Identification combined with the Eigensystem Realization Algorithm (OKID+ERA), Welch's frequency domain method, and a Time-Varying Multi-Model Selection (TVMMS) strategy based on a bank of parallel static models. All methods are evaluated under consistent experimental conditions using a shared dataset from a realistic mechatronic plateform. The comparative analysis focuses on key performance metrics including estimation accuracy, robustness to noise, computational complexity, and suitability for real-time embedded implementation. The study highlights trade-offs between estimation accuracy and implementation efficiency, thereby supporting the selection of identification methods suitable for adaptive control of mecahtronic systems.

Keywords: Machine/Reinforcement Learning
Abstract: Parkinson's Disease (PD) is a neurodegenerative disorder characterized by motor symptoms such as tremors, bradykinesia, and rigidity. Early diagnosis is crucial for effective disease management and improved patient outcomes. In recent years, machine learning techniques have emerged as powerful tools for biomedical data analysis and disease classification. This study investigates the application of Gradient Boosting Machine (GBM) for classifying PD based on accelerometer data.

The primary goal of this paper is to develop a robust classification model capable of accurately distinguishing PD patients from healthy controls using accelerometer readings collected during various activities. Statistical and frequency domain features, including measures of central tendency, variability, and spectral characteristics, are extracted from the data. A GBM model is then trained to identify movement patterns indicative of PD.

Accurate classification of PD patients can facilitate timely interventions, personalized treatment, and improved patient care. Future research may refine feature selection, explore ensemble learning methods, and validate the model on larger, more diverse datasets. This study underscores the potential of machine learning in enhancing the diagnosis and understanding of Parkinson's Disease.

Keywords: Multi - Agents Systems, Agent - Based Systems, Intelligent Systems
Abstract: We propose a simple and effective strategy for maze exploration and evacuation by reactive agents, which operate strictly on local observations with no memory or global knowledge. Such agents only react to nearby local stimuli. Our method introduces a dynamic cell-based weighting mechanism, where each explorable cell maintains directional weights guiding the agents’ movements. We employ a weight decay strategy which helps encourage the exploration of lesser-visited paths and reduce redundant traversal. Additionally, using the same weighting mechanism, the agents are able to locally detect dead-ends and block paths leading to them, further improving maze navigation. We validate our approach through simulations across a range of maze sizes, agent counts and weight decay values. Our experiments consistently show that the cell weighting strategy leads to faster exploration and earlier evacuation compared to purely random movement. The results show that reactive behavior, together with a weight-based local adaptation mechanism, can significantly enhance both exploration coverage and evacuation efficiency in maze environments. We achieve these improvements without requiring global coordination, memory or path planning, which makes our approach suitable for environments with significant resource constraints.

Keywords: Cyber - Security, Machine Learning
Abstract: Cross-Site Scripting remains a prevalent vulnerability in modern web applications, despite the widespread adoption of output encoding libraries, hardened browser environments, and policy-driven security mechanisms such as Content Security Policy. A core challenge lies in the ever-expanding and rapidly evolving nature of the attack surface. As HTML parsing rules and JavaScript semantics continue to evolve, new contexts emerge where malicious scripts can be injected and executed. Simultaneously, attackers continually refine their filter-evasion techniques, further complicating defense efforts. Security testing workflows commonly rely on static payload collections of known attack strings curated over time. While these datasets are crucial, they often fail to keep pace with emerging attack strategies, potentially overlooking novel or subtle injection vectors. This paper highlights the resulting gap in evaluation coverage and explores the potential of automatically generated payloads as a forward-looking research avenue. In particular, we present early-stage experiments in synthesizing such payloads and discuss their dual role: not only can they help identify vulnerabilities introduced by previously unseen attack patterns, but they can also serve to augment training datasets for machine learning-based detection systems, where real-world exploit data is often limited.

Keywords: Cyber - Security, Software Methods and Tools, Computational Methods
Abstract: Modern operating systems manage a varied set of processes, demanding sophisticated tools for supervision, management, and optimization to ensure system stability and performance. This paper describes the structure and theoretical foundations of a complete process management and performance optimization system using the functions of the Windows Application Programming Interface. The system presented here is meant to provide detailed information on process behavior, resource usage (CPU, memory, I/O), and inter-process relationships, with the ability for dynamic parameter-informed enhancement through experiential insight. Key Windows APIs for process enumeration and resource querying (including Performance Data Helper - PDH), prioritization, and exploring affinities such as alteration, memory resource management, and inter-process communication, form a fundamental part of the system architecture. While several existing tools offer resource insights (e.g., Task Manager, Process Hacker), they often lack extensibility, dynamic rule-based controls, and integrated heuristic engines. Our system addresses this gap through a multi-layered, API-centric architecture that emphasizes process behavior modeling and real-time optimization.

Keywords: Cyber - Security, Cloud Computing, Internet of Things
Abstract: This paper presents the implementation of a lightweight cyber threat intelligence (CTI) system that automatically gathers security news and forum posts, then summarizes them using NLP model with an average latency of 8 seconds per summary. The results are indexed in OpenSearch and visualised in Grafana. The entire pipeline runs on a Raspberry Pi 4 (4 GB RAM), providing an energy efficient edge solution independent of cloud resources, with a consumption below 12W and constant CPU temperature below 42 ◦C. Remote redundancy is ensured through both dynamic DNS port forwarding and VPN-based access.

Keywords: Cyber - Security, Communication Networks, Machine Learning
Abstract: Network security has become increasingly important as cyber-attacks continue to evolve and threaten both personal and organizational systems. This paper presents NetShield Lite, a lightweight network security tool that provides vulnerability scanning, real-time DoS attack detection, and automated security hardening recommendations. The system integrates Python-based port scanning with nmap, packet analysis using Scapy, and local AI assistance through Ollama. NetShield Lite features a modular architecture combining command-line functionality with a web dashboard for real-time monitoring. The tool implements a novel timestamp-based approach for tracking DoS sessions and an adaptive hardening system based on user experience levels. Experimental results demonstrate effectiveness in detecting SYN flood, UDP flood, and ICMP flood attacks while maintaining low resource consumption. The system is suitable for both educational purposes and practical network security monitoring.

Keywords: Optimization and Optimal Control, Networked Control, Cyber - Security
Abstract: A smart contract is a program deployed on a blockchain network and becomes immutable once deployed. Reentrancy bugs are among the most significant vulnerabilities in blockchain technology. Numerous approaches have been proposed to detect and analyze them. However, minimizing reentrancy errors in smart contracts remains a challenge. The aim of this paper is to optimize the handling of reentrancy bugs in smart contracts across different versions of Solidity. It also proposes a new method for syntactic and lexical code detection. Our approach is based on static analysis, and the tool can detect different versions of smart contract code. We used regex to identify external calls in the contract and Control Flow Graph (CFG) to identify the components of the code. A detection algorithm was developed to identify reentrancy vulnerabilities. This approach enhances contract reliability and reduces risk prior to deployment. Our method has been evaluated using True Positive Rate (TPR) and False Positive Rate (FPR) metrics. We tested it on the SmartBugs benchmark suite and contracts from Etherscan. The contracts analyzed include two solidity versions: 0.4.x and 0.8.x. Thanks to the integration of multiple tools, our solution supports testing across different solidity versions

Keywords: Industrial Applications, Software Methods and Tools, Internet of Things
Abstract: Fault detection and fast diagnosis in electric power networks are critical concerns for energy systems performance and safety, as untimely electrical events (temporary over-increases of specific parameters, decreases of values below established thresholds, or protection triggering at the power line level) have either minor or major effect over the electric power infrastructure. Regardless of the diversity of the industrial architectures met while on-field situations, a comprehensive overview of recorded data must be available. In this context, designing a distributed data acquisition, monitoring and recording system with a high degree of adaptability, such as providing synchronized event-triggered recordings, is required. This work provides an efficient event synchronization and record triggering method, based on IoT remote interaction between the nodes of the distributed system, in which the relevant processing is exclusively performed at the edge of the afferent network. The method has already proven its utility in a controlled experimental fault testing environment, being suitable for various electric power architectures.

Keywords: Data - Driven Control, Control Applications, Mechatronics
Abstract: The current paper proposes the introduction of fuzzy logic to replace the linear Proportional-Derivative (PD) component of second Active Disturbance Rejection Control, resulting in the so-called Fuzzy Active Disturbance Rejection Control (FADRC), whose parameters are tuned using Fictitious Reference Iterative Tuning (FRIT). The novel FADRC-FRIT algorithm is validated using experiments on the 3D crane laboratory equipment by controlling the x-, y-, and z-axes. The main purpose of the current mix lies in its ability to automatically and optimally tune the parameters of ADRC, which is improved with PD Takagi-Sugeno fuzzy control and tuned in a model-free FRIT manner. The secondary purpose is the time efficiency achieved in identifying the optimal control parameters.

Keywords: Industrial Applications, Control Applications, Control Systems Design
Abstract: This paper presents the integration of a communication system for a flexible PCB manufacturing line intended for the electronics industry. The line supports the production of multiple product models and integrates a communication system based on a Programmable Logic Controller (PLC) and a personal computer (PC) containing the Node-Red program. The system enables real-time comparison between the product data selected for the placement machines— entered via the line PC—and the product data loaded on the reflow oven, both of which can be selected manually by the operator or automatically through a traceability software. The hardware configuration and PLC program are designed with modularity in mind, allowing for future expansion and new modes of interaction within the manufacturing process. The control system incorporates modular programming techniques that enhance the safety and integrity of the products being manufactured. Additionally, a web interface developed using Node-RED provides intuitive monitoring and visualization of system status.

This design ensures high operational efficiency, ease of use for operators, compliance with quality standards and scalability for future production needs, all while minimizing implementation effort.

Keywords: Sensor Networks, Industrial Applications, Control Applications
Abstract: This paper examines the extension of IEEE 802.15.4 for wireless process control in industrial settings, emphasizing secure and efficient data transfer. It identifies the shortcomings of existing protocols like ISA100 Wireless and Wireless HART in supporting wireless control loops. A new approach using advanced microcontrollers and transceivers is proposed to create a robust wireless control network (WCN) with low latency and interference resistance. Key developments include a custom development board, hardware and software comparisons, and the TSCH (Time Slotted Channel Hopping) mode. The solution is validated through simulations and experiments, showing stable control loops in tough industrial environments. Results suggest the network meets stringent industrial standards, promoting wider use of wireless technologies in process control and aiding the creation of cost-effective, reliable systems.

Keywords: Internet of Things, Embedded Systems, Control Applications
Abstract: This paper presents a cost-effective microclimate automation system designed for poultry farms, combining Arduino-based sensor integration with a Node-RED interface. A custom-built plexiglass model replicates poultry housing conditions, integrating sensors for temperature, humidity, air quality, and light, with actuators for ventilation and lighting control. The system enables real-time monitoring, threshold-triggered automation, and customizable settings via a web dashboard. Notifications are sent to farm managers in critical situations such as excessive CO₂ or NH₃ levels. The solution demonstrates a scalable, low-cost prototype adaptable for real farm deployments.

Keywords: Optimization and Optimal Control
Abstract: OSQP is a general purpose solver, based upon the alternating direction method of multipliers, for convex quadratic programs. Within this solver’s algorithm, the idea of the augmented Lagrangian with a penalty parameter- a parameter which captures the relative weight-age on the objective function and the constraints of the problem in-hand- is utilized to develop an algorithm with so-called augmentation parameters. Although, the selection of these parameters is a crucial task, the optimal way to do the selection is not yet known. This work proposes a new method to select these parameters by utilizing the information of the conjugate directions of the coefficient matrix of a linear system of equations present in the algorithm. This selection makes it possible to cache these conjugate directions, instead of computing them at each iteration, resulting in a faster computation of the solution of the linear system, thus reducing the overall computation time. This reduction is demonstrated by a numerical example by comparing the total time it takes for the algorithms to converge sufficiently close to the optimal solution.

Keywords: Petri Nets, Control Applications, Control Systems Design
Abstract: Object Enhanced Time Petri Nets (OETPNs) are best suited to be used throughout all the development stages of distributed control systems, reactive applications, and real-time systems, including discrete time and event-driven systems. This paper focuses on the specifications, design (synthesis), implementation, and verification of event-driven systems. The OETPNs are further extended to process events that occur immediately between ticks. The OETPN executor algorithm is further developed to receive, process, and respond to events by adding a new kind of transitions, “Asynchronous transitions”, those transitions are connected to input channels to receive events and are executed immediately, even between ticks. For verification, a new method for converting OETPN to Stochastic Time Petri Nets (STPNs) is proposed to assess the temporal behavior of the system so that the state class graph can be generated to detect the change of the program structure during execution time. An example of an event-driven system (Urban Vehicle System Traffic Jam Detection and Traffic Light Control Systems) is demonstrated starting from the specifications, design, and implementation with OETPNs and the verifications using the proposed method to transform OETPNs to STPNs.

Keywords: Petri Nets, Control Systems Design, Fuzzy Systems/Logic
Abstract: The problem of a sporadically house heating is approached by Fuzzy Logic Enhanced Time Petri Net (FLETPN) that is used for modeling the system composed of a house, a photo-voltaic system and a heating pump with its controller. The solar radiation and the external temperature are considered as variable disturbances. The influence of the first disturbance is estimated using the electric power generated by the photo-voltaic system. The models include fuzzy logic rule sets driven by the Petri Nets (PNs) and they must be obtained during the identification process. A Genetic Algorithm (GA) uses measured fused data for the controlled process identification. Further, the GA is used for the synthesis of the supervisor fuzzy logic rule sets.

Keywords: Data - Driven Control, Automata/Discrete Event Systems and Petri Nets, Linear Systems
Abstract: Data-driven methods have been recently introduced in the analysis of Boolean control networks (BCNs), with the goal of assessing the solvability of key control problems, even in the absence of a known model, provided that a sufficiently rich data set collected from the network is available. In this paper, we focus on the finite-horizon and infinite-horizon optimal control problems for BCNs and investigate when the collected data allow us to solve the problems and comment on the sub-optimality of the obtained solutions.

Keywords: Linear Systems
Abstract: “Strong stability” is introduced and characterized in control systems literature as a special type of stability allowing a deeper exploration of some system properties that depend on the Euclidean norm of the state-space vector. Our paper proposes a larger point of view for studying norm-dependent properties of system trajectories, which is not limited to the Euclidean norm and addresses such properties as associated with the existence of invariant sets defined by the considered norms. Thus, we develop an analysis framework based on set invariance theory applied to discrete- and continuous-time linear systems, which is able to accommodate the concept of “strong stability relative to a given (but arbitrary) vector norm” as a natural generalization of the previously mentioned approach. We prove that, regardless of the considered vector norm, this concept has an algebraic characterization, by equivalence, in terms of the induced matrix norm (for discrete-time dynamics) and matrix measure (for continuous-time dynamics). The developed framework clearly highlights the difference between the standard stability notions (that refer to norm-independent properties) and the strong-stability notions (that refer to norm-dependent properties and are intimately related to set invariance theory).

Keywords: Control Systems Design, Control Applications, Biomedical Engineering
Abstract: Numerous surgical procedures require general anaesthesia, and a current issue is how the drug dosage for this anaesthesia could be automatically controlled to avoid potential errors of under or overdose, while also providing support to the anesthesiologist for decision making. The Depth of Hypnosis is an essential component of anesthesia, measured by the Bispectral Index using the electroencephalography technique. The present study demonstrates how this index can be controlled in both the induction and maintenance phases of the anaesthesia procedure using fractional order controllers. Using a patient simulator that mimics the patient’s Bispectral index, the proposed algorithm for designing a fractional order PI controller is developed. The imposed control strategy is tested and validated on a group of 23 patients in the simulator, in the presence of the output disturbances mimicking surgical stimuli, both during the induction and maintenance phase.

Keywords: Biomedical Engineering, System Biology
Abstract: Found in all living cells, the DeoxyriboNucleic Acid (DNA) represents the blueprint of life. DNA sequences reassembly is one of the most expensive and computationally intensive tasks in molecular biology. To solve such complex applications specially conceived high computational power hardware architectures are required. The Field Programmable Gate Arrays (FPGA) – based accelerators play a major role in the efforts of shrinking as much possible the processing times related to genes reassembly. This paper focuses on two leading FPGA-based technologies that may be used for DNA reassembly hardware and software implementation purposes. One is the Xilinx EDK (Embedded Development Kit) Platform Studio software toolkit based on the 32bit MicroBlaze soft processor framework. The second one is the Vivado Design Suite toolkit specially developed to program the AMD Zynq 7000 SoC (System-on-a-Chip) family processing units. As well as, two general purpose development boards have been considered for experiments: a Genesys Virtex-5 development board that explores the MicroBlaze soft processor technology, and a ZyboZ7-10 board built upon the Zynq-7000 ARM/FPGA hardware platform. On both of them the same DNA processing algorithm has been implemented, with the same gene reads and reference sequences as inputs. Then, the efficiency of the two different hardware architectures programmed under different software technologies has been analyzed. In this way useful information may be collected regarding the utility, respectively the advantages or disadvantages of the above mentioned technologies. Possessing this knowledge, the user can decide then what technology looks more appropriate for its DNA reassembly purposes.

Keywords: Biomedical Engineering, Control Systems Design, Linear Parameter - Varying Systems
Abstract: The Artificial Pancreas Problem addresses glucose regulation via continuous monitoring and intervention. While existing nonlinear models assume exact patient parameters, incorporating uncertainty yields a more realistic framework. This paper develops and compares two robust control methods for input-affine insulin–glucose models. Both use an inner linearization layer and an outer robust component to handle uncertainties. The first combines exact feedback linearization with a robust linear controller, while the second adapts Koopman linearization into a robust form. Numerical results demonstrate the strengths and limitations of each approach for nonlinear glucose control.

Keywords: Neural Networks, System Identification and Modeling, Biomedical Engineering
Abstract: The control law is as good as the mathematical model of the system. Therefore, there is a clear need for better and better models. In the case of drug concentration diffusion, the dynamics that govern the time and space evolution can be very hard to incorporate into a classical neural network (NN). The present paper puts forward a hypothesis regarding the advantages of a neural network that incorporates knowledge of the system process. This is referred to as a physical informed neural network (PINN), and the case study that is the subject of consideration is the modelling of the dynamics of propofol diffusion used in anaesthesia. The paper proves this statement by showcasing the predictions of both NNs and PINNs on a time horizon that extends beyond the limits of the training set. This results in a marked enhancement in the performance of the PINNs.

Keywords: Industrial Applications, Fault Diagnosis and Fault Tolerant Control, Intelligent Systems
Abstract: Collecting increasingly large image datasets is commonly viewed as essential for enhancing the accuracy of computer vision models in industrial applications. However, simply acquiring more data does not always guarantee substantial improvements in performance. An overlooked factor is the correlation between data samples, particularly when samples originate from the same production batch. Such correlation can significantly influence model accuracy. In this study, we explore the impact of 2D image correlations among samples from identical production batches, specifically examining uniformly colored lunch plates used in an industrial manufacturing scenario. Traditionally, deep learning methodologies suggest that dataset expansion directly correlates with improved detection accuracy. Contrary to this widely-held assumption, our experimental results indicate that indiscriminate data collection eventually reaches a saturation point. Through analysis, we demonstrate that carefully selecting less correlated samples—rather than simply increasing the dataset size can enhance model performance. Specifically, by strategically curating smaller but better diversified datasets, we achieved an improvement exceeding 10% in binary accuracy on our test sets compared to larger datasets, that were not curated. Furthermore, our analysis highlights the significance of dataset quality and diversity, suggesting practical guidelines for efficient dataset construction in industrial defect detection applications. These findings underscore the necessity of integrating even an easy 2D correlation analysis into the data selection process can be more beneficial than accumulating vast volumes of less curated samples.

Keywords: Machine Learning, Machine/Reinforcement Learning, Biomedical Engineering
Abstract: Heart failure (HF) is a leading cause of mortality worldwide, with early detection critical for improving patient outcomes. This study investigates the application of machine learning (ML) models to predict HF using a comprehensive dataset of 918 patients, compiled from five public sources. Eleven clinical features, including chest pain type, blood pressure, cholesterol, and electrocardiogram results, were analyzed for their predictive value. Exploratory data analysis (EDA) revealed significant associations between HF and factors such as asymptomatic chest pain, elevated fasting blood sugar, and abnormal ST segment characteristics. Multiple ML algorithms were evaluated, with a primary focus on maximizing the recall for the No_Disease class of the target feature. While standard and pruned Decision Tree (DT) models exhibited substantial overfitting, ensemble approaches, particularly a pruned Random Forest (RF) model, demonstrated superior generalization, achieving a recall gap of less than 10% for the No_Disease class, and a weighted mean recall (WMR) gap of only 4.96% between the training and testing datasets. Feature importance analysis identified the slope of the ST segment, ST depression (old peak), cholesterol, chest pain type, exercise-induced angina, and maximum heart rate (HR) as the most influential predictors. These findings underscore the potential of carefully tuned ensemble ML models to enable accurate, early identification of HF, supporting timely clinical intervention and improved patient care.

Keywords: Biomedical Engineering, Machine Learning, Software Methods and Tools
Abstract: Non-Alcoholic Fatty Liver Disease (NAFLD) is a prevalent condition increasing risks for cardiovascular diseases, type 2 diabetes, and metabolic disorders. Traditional diagnostic methods, such as liver biopsy, are invasive and may result in classification errors. This work explores the use of deep learning models for automated detection and evaluation of hepatic steatosis from ultrasound images, aiming for a non-invasive, accurate alternative. A segmentation model based on U-Net architecture is employed to identify regions of interest, followed by a deep neural network - EfficientNetV2 - for classification of fatty liver grades. The proposed methodology achieves high accuracy, specificity, and sensitivity, demonstrating its potential as an effective tool for early detection and staging of hepatic steatosis. These findings contribute to advancing non-invasive diagnostic methods, mitigating the risks associated with biopsy, and improving patient outcomes in liver disease assessment.

Keywords: Machine Learning
Abstract: This paper investigates Sound Source Localization (SSL) on the Microsoft HoloLens 2 using machine learning models trained on spatialized audio data from its built-in microphone array. Although prior studies have explored SSL in controlled environments, few have addressed its application to Augmented Reality (AR) devices with limited sensors. To address this gap, we collected a dataset of 2,880 audio samples in increments of 5◦, covering the full 360◦ horizontal plane, and grouped them into 4, 8, 16, and 32 directional classes. Logistic Regression, Random Forest, XGBoost, and Vertex AI models were trained and evaluated using Receiver Operating Characteristic (ROC) curves and mean angular deviation metrics. The results show that while accuracy declines with finer angular resolution, Vertex AI consistently provides the most precise estimations, supporting its potential for real-time SSL in AR-based spatial awareness applications.

Keywords: Machine Learning
Abstract: This paper presents a transformers-based what-if scenario analysis (TBWISA) framework for simulating pricing interventions in e-commerce. The method integrates structural causal modeling, price elasticity estimation, and a transformer forecasting model to generate robust, interpretable demand scenarios. A sliding window algorithm identifies optimal revenue windows under each pricing strategy. TBWISA is evaluated against XGBoost and log-linear models using forecasting accuracy, elasticity realism, and economic plausibility. TBWISA outperforms the benchmarks and produces consistent and economically valid outcomes. The framework supports data-driven decision-making in retail and can be extended to other domains that require scenario analysis under uncertainty.

Keywords: Reinforcement Learning
Abstract: Safety is a critical concern in certain real-world reinforcement learning (RL) applications, where unsafe actions can lead to irreversible failures or high costs. Many approaches rely on manually designed or domain-specific safety rules, which may limit adaptability or require extensive prior knowledge. This paper introduces an environment with action-dependent dynamics and partial observability, designed to test and compare safe RL methods. A classic unconstrained RL algorithm is evaluated alongside two safety-oriented techniques: hard action masking and a novel soft action masking method. The latter estimates empirical risk during training and uses an annealing schedule to gradually reduce the likelihood of unsafe actions, without relying on predefined constraints. Experimental results show that soft action masking provides a flexible and effective compromise between exploration and safety in both deterministic and non-deterministic settings.

Keywords: Machine Learning, Computational Methods, Computer Vision
Abstract: Sky cameras provide valuable visual context that can enhance a model's ability to estimate and forecast photovoltaic power generation. However, accurately estimating solar panel energy production based on sky images is a challenging task that requires advanced modeling techniques. This paper explores different regression approaches to estimate energy production, using both manually extracted features and Deep Learning-based feature extraction. Ten classical regression models, all with their specific strengths and weaknesses, are evaluated on manually extracted features. The results, expressed in terms of mean absolute error and coefficient of determination ((R^2)), identify Gradient Boosting Regression as the most effective method. A Convolutional Neural Network (CNN) trained directly on sky images also demonstrates strong performance. As a result, we propose hybrid approaches that combine CNN-based feature extraction with various regression models, leading to improved prediction accuracy and reduced error. These findings highlight the potential of integrating Deep Learning-based feature extraction with classical regression methods to enhance the accuracy of solar energy estimation.

Keywords: Machine Learning, Biologically Inspired Systems, System Identification and Modeling
Abstract: This paper investigates how a swarm-inspired metaheuristic, namely Grey Wolf Optimizer (GWO) can automatically calibrate tree-ensemble learners to accurately predict the agrometeorological parameters (temperature and relative humidity). Using a rich set of agrometeorological observations from the NARDI Fundulea station in Romania, GWO explores a six-dimensional hyper-parameter space for two competing regressors: Random Forest and Extreme Gradient Boosting. The search converges quickly, producing fully tuned models that are then evaluated on a chronologically held-out year. Comparative analysis shows that the GWO-XGBoost captures seasonal trends and rapid atmospheric variations more faithfully than the corresponding GWO-Random Forest, while the latter maintains slightly tighter median errors. The study demonstrates that coupling metaheuristic optimization with boosted trees offers an interpretable, high precision forecasting of the agrometeorological parameters and establishes the first machine-learning benchmark on the full NARDI Fundulea dataset.

Keywords: Machine Learning, Biologically Inspired Systems, Statistical Learning
Abstract: Reaction-diffusion (RD) systems have emerged as a physically-motivated alternative to classical graph clustering techniques. In this paper we present an implicit-explicit numerical scheme that simulates a multi-species RD process on the k-nearest-neighbor graph of a data set. Cluster membership is read off from the steady-state concentration field. Using the canonical Iris data set, we compare the proposed method against k-means and spectral clustering. The RD simulator achieves an Adjusted Rand Index (ARI) of 0.76 and a Normalized Mutual Information (NMI) of 0.81, which is comparable to the best baseline. To the best of our knowledge this is one of the first evaluations of fully-implicit RD solvers for data clustering in an IEEE article. All code and reproducibility material are publicly released, available at https://github.com/silviup/graph-based-data-clustering.