Optimizing the loading of autoclaves is a critical 2D packing problem in aerospace manufacturing. This paper investigates using deep reinforcement learning (DRL), specifically the proximal policy optimization (PPO) algorithm, to learn effective packing strategies. We developed a custom OpenAI Gym environment that simulates the process, with an observation space that details three matrices representing the initial and final states of the board, the shape of the item and the quantities of the remaining items, a multidiscrete action space to select and place the items, and a reward function that incorporates item profit and invalid action penalties. We evaluated our PPO-based agent against a genetic algorithm (GA) across scenarios involving regular items, mixed regular/irregular items, and item prioritization based on size. In the simplest case with only regular items, PPO achieved a high fill rate (99.37
Developing information retrieval (IR) systems that enable access across multiple languages is crucial in multilingual contexts. In Timor-Leste, where Tetun, Portuguese, English, and Indonesian are official and working languages, no cross-lingual information retrieval (CLIR) solutions currently exist to support information access across these languages. This study addresses that gap by investigating CLIR approaches tailored to the linguistic landscape of Timor-Leste. Leveraging an existing monolingual Tetun document collection and ad-hoc text retrieval baselines, we explore the feasibility of CLIR for Tetun. Queries were manually translated into Portuguese, English, and Indonesian to create a multilingual query set. These were then automatically translated back into Tetun using Google Translate and several large language models, and used to retrieve documents in Tetun. Results show that Google Translate is the most reliable tool for Tetun CLIR overall, and the Hiemstra LM consistently outperforms BM25 and DFR BM25 in cross-lingual retrieval performance. However, overall effectiveness remains up to 26.95
Manual selection of real estate properties can pose considerable challenges for agents since it needs a careful balance of various factors to satisfy client requirements while also manoeuvring through the complexities of the market. Although automated valuation models are widely used to estimate property market values, they are not designed to support property recommendation tasks. To address this gap, filtering-based recommendation methods have been explored, including collaborative and content-based approaches. However, these methods face several limitations in the real estate domain. This paper proposes a recommendation methodology designed to identify houses that closely resemble a given property, allowing agents to select the best matches based on geographical and physical characteristics. To assess the performance of the proposed methodology, we employ a range of evaluation metrics that measure different aspects of the model’s effectiveness in ranking and recommending relevant items. The findings suggest that, while geographic features may slightly influence ranking behaviour, the model is capable of producing diverse and relevant recommendations consistently.
This article addresses the challenge of climate change in agriculture, focusing on the optimization of lettuce production through bio-inspired algorithms supported by an Internet of Things (IoT) architecture. The study underscores the sensitivity of lettuce to thermal stress and proposes a decision-making system that recommends optimal temperature ranges to maximize yield. It highlights the need to maintain optimal environmental temperatures for efficient growth. This research reveals a marked negative correlation between lettuce yield and variables such as evapotranspiration and soil temperature, through data analysis from two regions in Colombia. As a solution, the use of bio-inspired algorithms is proposed to create agricultural recommendation systems capable of identifying temperature ranges that maximize yield. The methodology employs Internet of Things technologies for data collection, and the proposed models include statistical techniques and bio-inspired algorithms, experimentally tested to refine crop management strategies in the face of adverse climatic conditions. The study emphasizes the urgency of implementing adaptive strategies to sustain agricultural productivity and ensure food security in the face of climate volatility.
Log analysis is fundamental to modern software observability systems, playing a key role in improving system reliability. Recently, there has been a growing adoption of Large Language Models (LLMs) for log anomaly detection, due to their ability to learn complex patterns. In this work, we propose a model-agnostic framework that allows seamless plug-and-play integration of different LLMs, making it easy to experiment with and select the model that fits specific needs. These models are first fine-tuned on normal log data, learning their patterns. During inference, the model predicts the most probable next tokens based on the preceding context in each sequence. Anomaly detection is performed using Top-K predictions, where sequences are flagged as anomalous if the actual log entry does not appear among the K most probable next tokens, with K determined using the validation dataset. The proposed framework is evaluated on three widely-used benchmark datasets—HDFS, BGL, and Thunderbird—where it consistently achieves competitive results, outperforming state-of-the-art methods in multiple scenarios. These results highlight the effectiveness of LLM-based log analysis and the importance of flexibility when selecting models for specific operational contexts.
Large language models (LLMs) have shown remarkable capabilities in natural language processing but often exhibit factual inconsistencies when applied to knowledge-intensive tasks, with hallucination rates as high as 30
In the field of artificial neural networks, the Kolmogorov–Arnold Network (KAN), which is inspired by the Kolmogorov–Arnold representation Theorem (KAT), has demonstrated outstanding performance in function fitting. Complex-Valued Kolmogorov–Arnold Network (CVKAN) transfers KAN into the complex domain, providing superior capabilities in complex-valued function fitting. In this paper, we formulate a complex-valued KAT that provides theoretical support for this transformation. Also, given the high suitability of modulus-based activation functions in complex-valued Neural Networks, we propose a ModELU-based CVKAN, which replaces the ℂ SiLU residual function in CVKAN with the ModELU function. Experiments demonstrate that our method outperforms CVKAN in function fitting in terms of accuracy and stability. Furthermore, we adopt RBFs with learnable shape parameters in ModELU-based CVKAN, replacing the previous fixed ones. This replacement enhances the model’s performance in function fitting.
People counting has numerous applications across various domains, from urban planning to crowd management. With advances in computer vision, the use of cameras for real-time video analysis has increased significantly. While camera-based people counting is an effective and versatile approach, particularly in outdoor environments, it raises significant privacy concerns. MobileNet SSD is a widely used computer vision algorithm for detecting and classifying people in images. However, its real-time functionality relies on processing clear pictures of individuals, which can compromise privacy. This study investigates whether applying a blurring technique to video streams can enhance privacy while maintaining accurate people counting. To evaluate this, a video recording of individuals walking was processed with different blur levels and the accuracy of people counting was analyzed across all versions. Notably, one blurred video maintained an accuracy of 76
This paper explores relation classification as a step toward extracting structured information from unstructured Portuguese text. We evaluate prompt-based approaches using generative large language models and compare them with fine-tuned BERT and DeBERTa models, which serve as the baselines in this study. Experiments are conducted on two Portuguese-language datasets: RelEx-PT, a custom sentence-level dataset created by aligning Wikidata triples with Wikipedia sentences, and a second consisting of multi-sentence texts annotated with multiple triples. Results show that, while prompt-based methods offer flexibility and solid performance, they are still not sufficient to overtake fine-tuning, as the latter yields significantly better results in identifying relation types, even widening the gap in more complex task settings.
Coastal upwelling systems are vital to ocean productivity and regional climate regulation, yet quantifying the influence of wind forcing on the persistence of upwelling remains a significant challenge. This study presents a supervised learning framework to predict Upwelling Stability Periods (USPs) by linking daily wind to Sea Surface Temperature (SST)-derived labels in the Canary Upwelling System. Building on the Core-Shell clustering approach, which identifies USPs via unsupervised SST segmentation, we reformulate the problem as a multiclass classification task. A fuzzification–defuzzification scheme is employed to generate USP labels in a manner that reconciles the temporal resolution mismatch between SST-based clustering and daily wind fields. We construct three progressively enriched feature sets from wind and SST data. Decision Trees (DTs) and Random Forests (RFs) are used to model the relationship between wind and USPs. Over a 16-year period, SST-enhanced features yield substantial gains in classification accuracy (AUC > 0.94). DTs achieve strong predictive performance while maintaining interpretability through rule-based structures. RFs consistently deliver robust performance reinforcing their role as a reliable benchmark for classification.
Urban digitalization requires scalable and interoperable methods for accurate building modelling. Traditional approaches to facade extraction and 3D reconstruction often lack seamless generalisation and adaptability, as well as integration with geospatial systems. This work presents an automated pipeline that transforms street-level imagery into georeferenced 3D building models by combining deep learning with texture-based analysis. The pipeline uses window annotations for visual structure detection and integrates geographic information system (GIS) building footprints from OpenStreetMap (OSM) to ensure spatial alignment. Unlike prior approaches that rely on multi-element semantic annotations, our method reduces labeling costs while preserving geometric accuracy. The window detection model achieves a mean Average Precision of 0.94 at 50
In this work, we present a complete pipeline for the detection and three-dimensional reconstruction of archaeological artifacts in underwater environments. Our approach leverages the advantages of Large Multimodal Models (LMMs) that allow for integrating historical, geographical, and contextual data with captured images, a necessary step in the identification and interpretation of objects of possible archaeological interest, otherwise unattainable through unimodal models. In addition to multimodal integration, our pipeline suggests the use of a 3D visualization method (Gaussian Splatting) that has not yet been applied to underwater archaeology but is, in many respects, a natural candidate to replace photogrammetry techniques, which encounter many problems in underwater environments. To demonstrate the effectiveness of the methodology, we have concretely implemented a version of this pipeline and applied it to the study of the wreck of the SS Main steamship (1892), located in Porto Pim Bay, Faial Island, Azores.
This study evaluates the ability of large language models (LLMs) to recognise and reason about discrimination within the legal framework of the European Convention on Human Rights (ECHR). Going beyond conventional bias detection, we assess whether LLMs can apply, interpret, and explain legal concepts in line with judicial reasoning. We introduce a formalised definition of discrimination derived from ECHR case law and apply it in a structured empirical test suite. Our findings reveal systematic limitations in current models’ ability to adhere to legal standards, offering practical insights into enhancing fairness-aware AI in the legal domain.
This work presents LINCS-Dams, a cost-effective prototype for dynamic monitoring of embankment dams. The experimental setup uses affordable sensors, including micro-electromechanical system (MEMS) in-place inclinometers (IPIs), water-level gauges, and vibration accelerometers. Sensor outputs are managed by a finite state machine (FSM) that defines accident driven alert and alarm levels while dynamically adjusting each sensor’s data sampling frequency, optimizing energy consumption and ensuring timely responses. The main goal of this work is implementing and validating the proposed system and assessing its value to small embankment dams, which often lack regular monitoring. Our cost-effective AIoT approach combines sensor networks with intelligent monitoring for early detection and adaptive response, particularly valuable for embankment dams facing increased climate-driven risks. Experimental results confirm that the prototype delivers reliable response and effective dynamic event detection.
Given the current scenario of uncertainty in port operations, the berth allocation problem remains a crucial issue in port operations. This work presents a method that combines the Deep Q-Network (DQN) algorithm with the Long Short-Term Memory (LSTM) neural network architecture within a reinforcement learning framework applied to berth allocation under inventory constraints. The results are promising and indicate that the approach can respect inventory limits while producing quality solutions comparable to those obtained by a commercial solver, thereby offering support for informed decision-making.
This paper presents CrioleSet, a new parallel corpus designed to facilitate neural machine translation (NMT) for Cape Verdean Creole (CVC), a low-resource language spoken by the majority of Cape Verdeans. Comprising over 6,000 translation pairs in English, Portuguese, French, and CVC, the dataset addresses the scarcity of annotated resources for CVC, which is further challenged by dialectal variation across the archipelago. We trained and evaluated three neural network architectures LSTM, GRU with gated attention (GAtt), and Transformer—base to perform English-CVC translation tasks. Experimental results demonstrate that the Transformer-base models significantly outperforms the others, achieving the highest BLEU and METEOR scores and the lowest TER, reflecting better translation quality and robustness. This confirms that attention-based architectures can effectively handle low-resource translation, even with relatively modest datasets. The work underscores the potential of focused data curation and deep learning to advance NLP resources for underrepresented languages. Future work includes expanding the dataset to encompass more dialects and refining models for broader NLP tasks in CVC, thereby contributing to linguistic inclusivity and cross-cultural communication.
The optimisation of massive data obtained from 3D acquisition methodologies through AI represents an innovative research frontier in 3D data management. It arises from the ever-increasing instruments’ capacity to acquire enormous amounts of geometric and radiometric information with a substantial increase in processing times, a demand for computing capacities, and the request to subsample ultra-dense point clouds at the end of the process. On the contrary, a priori intervention on the raw data can mitigate the role of data dimension, reducing processing times while preserving the valuable information to analyse and interpret the artefacts. The research presents a new methodological approach based on integrating photogrammetry and AI. Through AI algorithms, it was possible to optimise the weight of the images, automatically cluster and segment image areas, and assign different resolutions according to the image content. This experimental pipeline significantly reduced calculation times, extracted point clouds with variable resolution according to the elements represented, and preserved the architectural artefacts’ geometry.
Muography provides a non-invasive method for exploring the internal structures of cultural heritage artifacts using cosmic-ray-derived muons. In this work, we present an integrated pipeline combining synthetic data generation via Monte Carlo simulations with advanced deep learning techniques, aiming to overcome traditional limitations in muographic imaging. Utilizing the Geant4 toolkit, muon interactions were simulated within materials such as concrete, limestone, and wood, including concealed metallic elements to replicate realistic structural scenarios. To enhance image quality without requiring prolonged exposure times and to reduce detector costs, our approach employs two neural networks sequentially: an event augmentation network based on a U-Net architecture enriched with residual dense blocks, and a resolution augmentation network designed to improve spatial detail.
Gastrointestinal diseases have a growing impact on public health, often requiring timely and accurate diagnosis to prevent complications and improve patient outcomes. In this context, artificial intelligence (AI) has emerged as a promising tool to support clinicians in image-based diagnosis. This study presents the design and evaluation of an explainable deep learning system for the automatic detection and classification of gastrointestinal anomalies in colonoscopy images. Using transfer learning and convolutional neural networks, the proposed architecture incorporates a fine-tuned ResNet18 model alongside explainable AI (XAI) methods to ensure both high diagnostic performance and model transparency. The system was trained and validated using the Kvasir dataset, a clinically annotated collection of endoscopic images covering multiple gastrointestinal conditions. Experimental results show that the use of transfer learning significantly improved classification outcomes, with F1-scores exceeding 0.90 for several key categories. A web-based interface was also developed to facilitate clinical adoption, providing visual explanation tools such as heatmaps. These allow healthcare professionals to understand the basis of each prediction, promoting trust and supporting informed decision-making. Overall, the system contributes to more accurate, interpretable, and efficient diagnostic processes in the field of gastrointestinal healthcare.
Generative artificial intelligence (GenAI) is redefining higher education by enabling adaptive and creative learning environments. This study presents a framework applied in a pilot course on Algorithms and Programming, which combines automated skill diagnostics, team formation through classification algorithms (K Nearest Neighbors), content personalization, and continuous feedback based on large language models (LLMs), implemented through the MAKE platform. The results show a significant improvement in learning efficiency and quality: GenAI-assisted teams reduced delivery times by up to 50
