Unscheduled trips of high-power pulsed converters are a leading source of downtime at large accelerator facilities. At the Spallation Neutron Source (SNS), the High Voltage Converter Modulators (HVCMs) are consistently the second-largest contributor to lost beam time. Each HVCM pulse is recorded across sensor channels spanning currents, voltages, and magnetic fluxes, whose mutual interactions encode the operating state of the system. Fault precursors do not manifest uniformly across these channels: depending on fault type, they may alter the temporal structure of individual signals, change the statistical dependencies among channels, or both. Existing deep-learning approaches typically process multi-channel signals with standard convolutional pipelines that entangle temporal and cross-channel operations from the first layer, giving the model no explicit mechanism to represent channel independence or structured inter-channel interaction. We hypothesise that architectural inductive bias, specifically the ordering of temporal filtering and cross-channel mixing, plays a central role in detection performance on this class of data. To test this, we vary the order in which these two operations are applied, and examine whether per-pulse adaptive channel reweighting further improves sensitivity. Evaluated on the public HVCM dataset across all four SNS subsystems (RFQ, DTL, CCL, SCL), our best variant achieves a pooled AUC-PR of 0.816 and AUC-ROC of 0.934, outperforming the state of the art on most subsystems and five of the six fault families. Ablations identify three dominant input channels and link per-fault-family performance to whether precursors manifest as amplitude shifts in individual channels or as subtler patterns requiring joint channel representations to surface.

Decision-making under uncertainty in energy management is complicated by unknown parameters hindering optimal strategies, particularly in Battery Energy Storage System (BESS) operations. Predict-Then-Optimise (PTO) approaches treat forecasting and optimisation as separate processes, allowing prediction errors to cascade into suboptimal decisions as models minimise forecasting errors rather than optimising downstream tasks. The emerging Decision-Focused Learning (DFL) methods overcome this limitation by integrating prediction and optimisation; however, they are relatively new and have been tested primarily on synthetic datasets with limited evidence of their practical viability. Real-world BESS applications present additional challenges, including greater variability and data scarcity due to collection constraints. Because of these challenges, this work leverages Automated Feature Engineering (AFE) to improve the nascent approach of DFL. This AFE-DFL integration automatically extracts decision-relevant features from limited energy data without requiring domain expertise, while ensuring features directly enhance BESS operational decisions rather than merely improving prediction accuracy metrics. We propose an AFE-DFL framework suitable for small datasets that forecasts electricity prices and demand while optimising BESS operations to minimise costs. We validate the framework's effectiveness on a novel real-world UK property dataset. The evaluation compares DFL methods against PTO, with and without AFE. Results show that DFL yields lower operating costs than PTO, and adding AFE further improves DFL performance by 22.9-56.5 % compared to models without AFE. These findings provide empirical evidence for DFL's practical viability, demonstrating that AFE-DFL integration reduces reliance on domain expertise while achieving superior economic outcomes for BESS optimisation.

The growing size of Convolutional Neural Networks has led to increasingly large and costly models. Knowledge Distillation (KD) addresses this by transferring knowledge from a large network (teacher) to a small one (student), also reducing the training data required. KD is traditionally applied only at the network's final output. However, its behaviour when applied at intermediate network layers has received little attention. This raises the question of whether intermediate block-wise KD, which provides supervision throughout the network, could offer an advantage under specific conditions, such as few instances per class, which is common in fine-grained datasets. This work proposes a student design based on simple, homogeneous blocks mirroring those of the teacher, distilling knowledge between corresponding blocks. Across eleven datasets, we show that on classic datasets, distilling only the last block is sufficient – and often best–, whereas fine-grained, data-scarce settings benefit substantially from intermediate supervision, with even a single additional distillation point narrowing the gap considerably. We further study how this supervision should be guided, exploring configurations of varying granularity and informed by an explainability analysis based on attention maps, Centered Kernel Alignment, and Grad-CAM, alongside the impact of teacher and student fine-tuning strategies. This work shows that intermediate block-wise distillation, guided appropriately, is key to building compact data-efficient models without sacrificing accuracy.

Selecting which instances to label is a key challenge in low-label tabular learning. For recent Tabular Foundation Models such as TabPFN, context selection directly determines predictive performance. Supervised oracle experiments show that carefully chosen labeled context sets can strongly outperform random selection under the same labeling budget. However, the cold-start setting, where instances must be selected before any labels are available, has received little attention in the TFM literature. This problem is fundamentally geometric. In vision and language, foundation models induce embedding spaces where simple geometric selection methods are effective. In contrast, tabular instance selection has so far been performed predominantly in the original tabular space, which lacks a natural metric; heterogeneous types, mixed scales, and nonlinear interactions make raw-space distances unreliable for context construction, and original-space selection falls below random on the majority of datasets as the budget grows. We propose LUCoS (Latent Unsupervised Context Selection), which replaces raw-feature geometry with the latent geometry induced by embeddings from an unsupervised Prior-Fitted Network (PFN) and selects representative medoids as context. Evaluated on 67 OpenML-CC18 datasets across six low-label budgets, LUCoS ranks first under mean AUC, ACC, and F1, with conclusions stable across metrics and dataset-level robustness checks. A gain decomposition reveals a simple mechanism: at the smallest budgets, the main benefit comes from enforcing coverage; as the budget increases, the decisive factor becomes the representation space in which coverage is measured. LUCoS mitigates failures of original feature space selection, showing that reliable unsupervised context selection depends less on selector sophistication than on defining representativeness in a meaningful representation geometry.

Anomaly detection in multivariate time series is a critical task across a wide range of real-world applications, where abnormal behaviour is rare, labels are unavailable, and the cost of a miss is high. The central challenge is learning a characterisation of normality precise enough to flag deviations. Representation self-supervised learning, typically through contrastive approaches, addresses this by embedding temporal patches into a latent space where normality occupies a well-defined region, with anomalies detected by geometric deviation. However, contrastive approaches shape this space indirectly through pair-sampling heuristics, providing no explicit control over the geometric structure that distance-based scoring requires. This means how tightly normal representations are grouped, and whether distances are directionally meaningful. We present VACE (Velocity-Aligned Channel Embeddings), a self-supervised anomaly detection method that represents normality as a compact, directionally coherent region in the embedding space. To this end, VACE trains a channel-aware encoder through a velocity-consistency objective, with no negatives and no synthetic anomalies, so that normal trajectories are locally smooth and aligned. At test time, a Mahalanobis positional score and a velocity-bank directional score are combined multiplicatively, flagging points that are simultaneously off-distribution and dynamically atypical. Despite its simplicity, VACE achieves state-of-the-art performance on TSB-AD-M under rigorous evaluation, significantly outperforming more complex methods trained on substantially larger budgets.

In Artificial Intelligence, there is an increasing demand for adaptive models capable of dealing with a diverse spectrum of learning tasks, surpassing the limitations of systems devised to cope with a single task. The recent emergence of General-Purpose Artificial Intelligence Systems (GPAIS) poses model configuration and adaptability challenges at far greater complexity scales than the optimal design of traditional Machine Learning models. Evolutionary Computation (EC) has been a useful tool for both the design and optimization of Machine Learning models, endowing them with the capability to configure and/or adapt themselves to the task under consideration. Therefore, their application to GPAIS is a natural choice. This paper aims to analyze the role of EC in the field of GPAIS, exploring the use of EC for their design or enrichment. We also match GPAIS properties to Machine Learning areas in which EC has had a notable contribution, highlighting recent milestones of EC for GPAIS. Furthermore, we discuss the challenges of harnessing the benefits of EC for GPAIS, presenting different strategies to both design and improve GPAIS with EC, covering tangential areas, identifying research niches, and outlining potential research directions for EC and GPAIS.

Studying the brain in early life is crucial for understanding how brain development unfolds across the lifespan. Representation learning has been widely applied to EEG, showing potential for learning general-to-specific representations, that is, representations learned by models that capture broadly useful abstract features that can later be specialized. However, infant EEG differs from adult EEG due to the large developmental changes that occur in the first years of life. This developmental variability introduces additional challenges to applying representation learning to developmental infant EEG.In this study, we provide a systematic assessment of representation learning for early-life EEG and highlight simple yet important methodological considerations for its application. In particular, we compare self-supervised and baseline representation learning approaches to analyze their respective strengths and limitations. Evaluation is performed on a real-world dataset of 45 infants, recorded longitudinally through their early growth, using complementary strategies such as downstream prediction of chronological age, unsupervised clustering analyses, and ablation studies across sessions and subjects.

Multi-criteria decision-making (MCDM) techniques are widely used to facilitate systematic and informed decision making. However, real-world problems in diverse fields such as business, cybersecurity, and environmental management often involve significant levels of uncertainty. While progress has been made in developing MCDM methods capable of addressing uncertainty, many challenges remain–particularly in capturing uncertainty, communicating it effectively to stakeholders, and selecting the most suitable MCDM technique for a given context. This paper focuses on interval extensions of the widely used Technique for Order Preference by Similarity to Ideal Solution (TOPSIS), highlighting the need to (1) effectively communicate uncertainty to decision makers and (2) establish a way to compare these extensions correctly. To address these, we propose a Monte Carlo-based approach for interval TOPSIS, offering a new interval extension of the original TOPSIS to handle uncertainty independently. Meanwhile, we leverage a data set from the literature and use the proposed extension as a reference point to compare some of the most recent TOPSIS extensions and analyse their underlying differences in uncertainty handling.

Protein subcellular localization is an important classification task because the location of proteins in a cell is directly linked to their functions. Since a protein can act at two or more locations simultaneously, multi-label classification algorithms are necessary. The currently used algorithms are usually based on supervised learning, which presents some disadvantages such as (i) a need for a large amount of labeled instances for training; (ii) a waste of valuable information that labeled instances can provide; and (iii) a high cost involved in obtaining labeled instances for training. To overcome these disadvantages, semi-supervised learning can be applied, where classifiers exploit both labeled and unlabeled data. Thus, in this paper, we propose a new semi-supervised algorithm for multi-label protein subcellular localization. Our proposal is based on decision tree classifiers induced using predictive clustering trees. We investigate many semi-supervised protein subcellular localization scenarios to test whether unlabeled instances can improve the multi-label classification process. Our results show that the proposal can achieve competitive or better results when compared to the pure supervised version of the predictive clustering trees.

Decision-Focused Learning (DFL) is a paradigm to learn neural network-based predictive models tailored to a specific optimisation problem. A key challenge for DFL methods lies in the non-differentiable nature of most optimisation problems. Recent solutions use a learned, differentiable, model to act as a surrogate loss. To learn the model, the optimisation problem is solved repeatedly using random perturbations of the predictions to calculate a regret value which can be used as a target to learn the surrogate loss. However, this necessitates numerous runs of the, potentially computationally expensive, optimiser. As such, maximising the useful information from each run of the optimizer is paramount. A sample of purely random perturbations may not yield an effective distribution to learning the surrogate from. We propose using a directed perturbation strategy to generate a set of meaningful perturbations for learning a surrogate loss model. We evaluate our approach on a resource allocation problem and real-world case study focused on a critical energy challenge: optimising solar-plus-battery systems. The results show that directed perturbations learn a stronger surrogate loss with fewer runs of the optimiser, enabling a more efficient DFL.

Feature engineering (FE) plays a crucial role in Machine Learning pipelines, yet it remains a time-consuming process requiring heavy domain expertise. While Automated Machine Learning (AutoML) has automated model selection and hyperparameter tuning, it often overlooks FE, which is particularly needed in specialised domains such as Energy Consumption Forecasting (ECF). To address this limitation, we introduce AutoEnergy, a novel, domain-aware FE algorithm tailored for ECF. AutoEnergy automatically generates interpretable features from timestamps and past consumption values through rule-based transformations, integrating them with AutoML for fully automated ECF modelling while reducing human intervention. The performance of AutoEnergy was evaluated using eighteen diverse real-world energy consumption datasets spanning residential, commercial, industrial, and grid power domains. Through extensive benchmarking against baseline AutoML without FE and established FE methods, namely TSFresh (with TSEfficient and TSMinimal configurations) and FeatureTools (FT), AutoEnergy demonstrated significant improvements in both predictive accuracy and computational efficiency. AutoEnergy achieved forecasting error reductions of 19.52% to 84.72% compared to benchmarking methods, with strong performance on smaller datasets and statistical validation via Friedman and Wilcoxon tests. AutoEnergy demonstrated notable computational efficiency by running 1.31 and 4.41 times faster than FT and TSEff, respectively. Although 1.58 times slower than TSMin, AutoEnergy achieved 82.38% lower forecasting errors. Integrating AutoEnergy with the state-of-the-art Tabular Prior Data Fitted Network (TabPFN) resulted in significant forecasting error reductions across test sets. These findings highlight AutoEnergy's potential to improve AutoML performance while reducing reliance on domain expertise for FE, paving the way for fully automated ML pipelines in ECF applications.

Generalising deep-learning models to perform well on unseen data domains with minimal retraining remains a significant challenge in computer vision. Even when the target task—such as quantifying the number of elements in an image—stays the same, data quality, shape, or form variations can deviate from the training conditions, often necessitating manual intervention. As a real-world industry problem, we aim to automate stock level estimation in retail cabinets. As technology advances, new cabinet models with varying shapes emerge alongside new camera types. This evolving scenario poses a substantial obstacle to deploying long-term, scalable solutions. To surmount the challenge of generalising to new cabinet models and cameras with minimal amounts of sample images, this research introduces a new solution. This paper proposes a novel ensemble model that combines DenseNet-201 and Vision Transformer (ViT-B/8) architectures to achieve generalisation in stock-level classification. The novelty aspect of our solution comes from the fact that we combine a transformer with a DenseNet model in order to capture both the local, hierarchical details and the long-range dependencies within the images, improving generalisation accuracy with less data. Key contributions include (i) a novel DenseNet-201 + ViT-B/8 feature-level fusion, (ii) an adaptation workflow that needs only two images per class, (iii) a balanced layer-unfreezing schedule, (iv) a publicly described domain-shift benchmark, and (v) a 47 pp accuracy gain over four standard few-shot baselines. Our approach leverages fine-tuning techniques to adapt two pre-trained models to the new retail cabinets (i.e., standing or horizontal) and camera types using only two images per class. Experimental results demonstrate that our method achieves high accuracy rates of 91% on new cabinets with the same camera and 89% on new cabinets with different cameras, significantly outperforming standard few-shot learning methods.

Procedural level generation in video games has made significant strides, yet achieving high-quality automated level design remains a major challenge. Over the years, techniques have evolved from simple constructive algorithms to advanced Artificial Intelligence (AI) models like Generative Adversarial Networks and Large Language Models. However, the lack of a standardised evaluation framework has hindered direct numerical comparisons and the ability to gauge true progress in the field. To address this gap, we propose an evaluation methodology to benchmark key generation techniques and explore the potential of general-purpose AI models. As a case study, we present a Super Mario Bros level generator powered by ChatGPT, leveraging general-purpose natural language for design tasks. The results show the different strengths and weaknesses of existing models, indicating that traditional algorithms still outperform the most advanced AI methods in this domain, highlighting the need for further innovation to bridge the gap.

Zero-Shot Learning is an important paradigm within General-Purpose Artificial Intelligence Systems, particularly in those that operate in open-world scenarios where systems must adapt to new tasks dynamically. Semantic spaces play a pivotal role as they bridge seen and unseen classes, but whether human-annotated or generated by a machine learning model, they often contain noisy, redundant, or irrelevant attributes that hinder performance. To address this, we introduce a partitioning scheme that simulates unseen conditions in an inductive setting (which is the most challenging), allowing attribute relevance to be assessed without access to semantic information from unseen classes. Within this framework, we study two complementary feature-selection strategies and assess their generalisation. The first adapts embedded feature selection to the particular demands of ZSL, turning model-driven rankings into meaningful semantic pruning; the second leverages evolutionary computation to directly explore the space of attribute subsets more broadly. Experiments on five benchmark datasets (AWA2, CUB, SUN, aPY, FLO) show that both methods consistently improve accuracy on unseen classes by reducing redundancy, but in complementary ways: RFS is efficient and competitive though dependent on critical hyperparameters, whereas GA is more costly yet explores the search space more broadly and avoids such dependence. These results confirm that semantic spaces are inherently redundant and highlight the proposed partitioning scheme as an effective tool to refine them under inductive conditions.

Predict+Optimize frameworks integrate forecasting and optimization to address real-world challenges such as renewable energy scheduling, where variability and uncertainty are critical factors. This paper benchmarks solutions from the IEEE-CIS Technical Challenge on Predict+Optimize for Renewable Energy Scheduling, focusing on forecasting renewable production and demand and optimizing energy cost. The competition attracted 49 participants in total. The top-ranked method employed stochastic optimization using LightGBM ensembles, and achieved at least a 2% reduction in energy costs compared to deterministic approaches, demonstrating that the most accurate point forecast does not necessarily guarantee the best performance in downstream optimization. The published data and problem setting establish a benchmark for further research into integrated forecasting-optimization methods for energy systems, highlighting the importance of considering forecast uncertainty in optimization models to achieve cost-effective and reliable energy management. The novelty of this work lies in its comprehensive evaluation of Predict+Optimize methodologies applied to a real-world renewable energy scheduling problem, providing insights into the scalability, generalizability, and effectiveness of the proposed solutions. Potential applications extend beyond energy systems to any domain requiring integrated forecasting and optimization, such as supply chain management, transportation planning, and financial portfolio optimization.

As energy demands and costs rise, enhancing energy efficiency in Food and Drink Cold Storage (FDCS) rooms is important for reducing expenses and achieving environmental sustainability ambitions. Forecasting electricity use in FDCSs can help optimise operations and minimise energy consumption by enabling door opening frequency, maintenance, and restocking to be better scheduled. Although Machine Learning (ML) has been applied to forecast energy use in various domains such as commercial and residential buildings, its use in addressing the specific challenges of FDCS, which require stringent temperature and humidity control for food safety and quality, has been less explored. This work addresses this gap by proposing a tailored ML pipeline for FDCS settings capable of predicting one-week into the future and is suitable for small dataset sizes. It provides comparative analysis by employing two distinct real-world FDCS datasets for training, validation, and testing of the developed models. Moreover, in contrast to existing studies predominantly concerned with energy consumption prediction, this study includes the forecasting of indoor temperature and humidity, given their essential role in preserving the quality and longevity of stored food items. Ensemble-based methods, particularly Random Forest, excelled and achieved the lowest electricity MAEs of 150.65 and 384.88 for each dataset, respectively.

For efficient operation, solar power operators often require generation forecasts for multiple sites with varying data availability. Many proposed methods for forecasting solar irradiance / solar power production formulate the problem as a time-series, using current observations to generate forecasts. This necessitates a real-time data stream and enough historical observations at every location for these methods to be deployed. In this paper, we propose the use of Global methods to train generalised models. Using data from 20 locations distributed throughout the UK, we show that it is possible to learn models without access to data for all locations, enabling them to generate forecasts for unseen locations. We show a single Global model trained on multiple locations can produce more consistent and accurate results across locations. Furthermore, by leveraging weather observations and measurements from other locations we show it is possible to create models capable of accurately forecasting irradiance at locations without any real-time data. We apply our approaches to both classical and state-of-the-art Machine Learning methods, including a Transformer architecture. We compare models using satellite imagery or point observations (temperature, pressure, etc.) as weather data. These methods could facilitate planning and optimisation for both newly deployed solar farms and domestic installations from the moment they come online.

Background Pulmonary fibrosis results from alveolar injury, leading to extracellular matrix remodelling and impaired lung function. This study aimed to classify patients with pulmonary fibrosis according to blood biomarkers to differentiate distinct disease patterns, known as endotypes. Methods In this cluster analysis, we first classified patients from the PROFILE study, a multicentre, prospective, observational cohort of individuals with incident idiopathic pulmonary fibrosis or non-specific interstitial pneumonia in the UK (Nottingham University Hospitals, Nottingham; and Royal Brompton Hospital, London). 13 blood biomarkers representing extracellular matrix remodelling, epithelial stress, and thrombosis were measured by ELISA in the PROFILE study. We classified patients by unsupervised consensus clustering. To evaluate generalisability, a machine learning classifier trained on biomarker signatures derived from consensus clustering was applied to a replication dataset from the Australian Idiopathic Pulmonary Fibrosis Registry (AIPFR). Biomarker associations with mortality and change in percentage of predicted forced vital capacity (FVC%) were assessed, adjusting for age, gender, baseline FVC%, and antifibrotic treatment and steroid treatment before and after baseline. Mortality risk associated with the clusters in the PROFILE cohort was evaluated with Cox proportional hazards models, and mixed-effects models were used to analyse how clustering was associated with longitudinal FVC% in the PROFILE and AIPFR cohorts. Findings 455 of 580 participants from the PROFILE study (348 [76%] men and 107 [24%] women; mean age 72·4 years [SD 8·3]) were included in the analysis. Within this group, three clusters were identified based on blood biomarkers. A basement membrane collagen (BM) cluster (n=248 [55%]) showed high concentrations of PRO-C4, PRO-C28, C3M, and C6M, whereas an epithelial injury (EI) cluster (n=109 [24%]) showed high concentrations of MMP-7, SP-D, CYFRA211, CA19-9, and CA-125. The third cluster (crosslinked fibrin [XF] cluster; n=98 [22%]) had high concentrations of X-FIB. In the replication dataset (117 of 833 patients from AIPFR; 87 [74%] men and 30 [26%] women; mean age 72·9 years [SD 7·9]), we identified the same three clusters (BM cluster, n=93 [79%]; EI cluster, n=8 [7%]; XF cluster, n=16 [14%]). These clusters showed similarities with clusters in the PROFILE dataset regarding blood biomarkers and phenotypic signatures. In the PROFILE dataset, the EI and XF clusters were associated with increased mortality risk compared with the BM cluster (EI vs BM: adjusted hazard ratio [HR] 1·88 [95% CI 1·42–2·49], p<0·0001; XF vs BM: adjusted HR 1·53 [1·13–2·06], p=0·0058). The EI cluster showed the greatest annual FVC% decline, followed by the BM and XF clusters. A similar FVC% decline pattern was observed in these clusters in the AIPFR replication dataset. Interpretation Blood biomarker clustering in pulmonary fibrosis identified three distinct blood biomarker signatures associated with lung function and prognosis, suggesting unique pulmonary fibrosis biomarker patterns. These findings support the presence of pulmonary fibrosis endotypes with the potential to guide targeted therapy development. Funding None.

Machine learning methods are widely used to predict energy consumption, aiming to enhance efficiency and support environmental goals. However, developing these models is traditionally time-consuming and expert-dependent. While Automated Machine Learning (AutoML) has emerged as a valuable approach to streamlining machine learning pipelines, including appropriate preprocessing and learning algorithms, it may still require human experts. These experts might be needed to generate new, interpretable features that could significantly improve model performance. This is particularly relevant in complex settings such as the energy domain, where deep learning's automatic feature extraction lacks interpretability. To address this challenge, this exploratory work introduces an automated feature engineering method tailored for energy forecasting problems. It involves generating a comprehensive set of features that can be fed into AutoML, thereby reducing the need for domain knowledge in feature engineering. The proposed method has been validated using eleven datasets from various energy domains, including residential buildings, renewable energy, and regional energy consumption, with state-of-the-art AutoML methods, namely H20, TPOT, AutoGluon, and FLAML. The results demonstrate a noticeable reduction in prediction errors across all the examined datasets.