ABSTRACT. Previous work has made use of a parameterized plane curve polynomial representation for mathematical handwriting, with the polynomials represented in a Legendre or Legendre–Sobolev graded basis. This provides a compact geometric representation for the digital ink. Preliminary results have also been shown for Chebyshev and Chebyshev–Sobolev bases. This article explores the trade-offs between basis choice and polynomial degree to achieve accurate modeling with a low computational cost. To do this, we consider the condition number for polynomial evaluation in these bases and bound how the various inner products give norms for the variations between symbols.

ABSTRACT. Stirling numbers were originally introduced in order to relate monomial powers and factorial powers. The notion of factorial powers has been generalized beyond integer powers, and Stirling numbers have been generalized beyond integer arguments. Here we explore the transformations between monomial and factorial powers by identifying special values of the generalized Stirling numbers.

ABSTRACT. The location of real eigenvalues offers critical insights into the stability and resonance of physical systems. This paper introduces a hybrid symbolic-numeric method for certified real eigenvalue localization. Our approach integrates Gershgorin disk analysis and Hermite matrix certification to obtain certified intervals that contain real eigenvalues. The certified intervals can be refined using bisection-like approaches up to a certain precision. This method provides interval certifications while maintaining computational efficiency. We illustrate our framework through a simple, complete computational example and briefly describe how it can be extended to use this method on complex matrices with complex eigenvalues.

ABSTRACT. We present a version of the REDUCE computer algebra system as it was in the early 1970s. We show how this historical version of REDUCE may be built and run in very modest present-day environments and outline some of its capabilities. This shows that a large part of computer algebra function can be delivered by a system that is small by today's standards and can be deployed in resource constrained settings.

*** Note this short paper was used as a software demo proposal for ISSAC 2025. The demo was accepted, but the demo descriptions are not published.

So to be clear, this paper has not been submitted for publication elsewhere. ***

ABSTRACT. We explore new approaches for finding matrix multiplication algorithms in the commutative setting by adapting the flip graph technique: a method previously shown to be effective for discovering fast algorithms in the non-commutative case. While an earlier attempt to apply flip graphs to commutative algorithms saw limited success, we overcome both theoretical and practical obstacles using two strategies: one inspired by Marakov’s algorithm to multiply 3x3 matrices, in which we construct a commutative tensor and approximate its rank using the standard flip graph; and a second that introduces a fully commutative variant of the flip graph defined via a quotient tensor space. We also present a hybrid method that combines the strengths of both. Across all matrix sizes up to 5x5, these methods recover the best known bounds on the number of multiplications and allow for a comparison of their efficiency and efficacy. Although no new improvements are found, our results demonstrate strong potential for these techniques at larger scales.

ABSTRACT. This paper presents a geolocation system to assist in identifying the location in images associated with illicit activities, thereby reducing investigation time for authorities. The system is implemented as a multi-agent system. The geolocalisation is performed based on image analysis, which involves both image classification and regression of the geographical coordinates. The final result is obtained by fusioning both results. The classification model is based on the EfficientNet architecture. The evaluation is performed on a dataset with images from Romania collected from Google Street View. Moreover, the best version of the model succeeds in solving difficult classification tasks, correctly identifying the answer among two or more similar counties in 80% of the cases. Alternatively, from a regression point of view, the predicted coordinates are at most 80 km away from the ground truth location in over 80% of the evaluation cases. By employing various data augmentation techniques, such as noise addition or severe compression, a model robust to potential adversarial interference was achieved.

ABSTRACT. The long-term goal of artificial intelligence (AI) is to develop systems that can generalize across diverse domains and solve complex coordination problems. Logistics scenarios provide an effective benchmark for this objective because they require decentralized decision-making, temporal scheduling, and dynamic cooperation. This paper presents a systematic evaluation of the AbstractSwarm Multi-Agent Logistics Competition, held at GECCO 2023, which employed the AbstractSwarm simulation system as a testbed for cooperative agents. Six heterogeneous scenarios were used to model challenges ranging from patient scheduling and ward coordination to traffic control, warehouse storage, and last-mile delivery. Agents submitted by participants were executed for 100 runs with 20 repetitions per scenario, and performance was measured using normalized waiting times with scenario-specific baselines. The results, summarized in detailed tables, reveal a trade-off between specialization and generalization: some agents excelled in structured scheduling tasks such as Treatments and X-Ray Scheduling, while others demonstrated moderate robustness across all scenarios but struggled in dynamic environments such as Crossroads or sparse-reward tasks like Storage. These findings highlight both the promise and limitations of swarm intelligence for logistics and motivate future research into hybrid methods, reinforcement learning strategies, and scalable coordination mechanisms for real-world applications.

ABSTRACT. An increasing trend in data-driven web and mobile applications has significantly raised the importance of proper data architecture and management. Establishing a robust and agile data platform is paramount to fueling data-driven applications. In response, businesses are leveraging innovative data strategies to cater to their evolving technological needs. Traditional cloud computing has formed the foundation for modernizing such applications, offering scalable computing, storage, networking, and robust data management capabilities. Moreover, integrating machine learning with big data on the cloud has allowed organizations to process and analyze vast data volumes, deriving valuable insights and driving innovation. However, with the growing interest in real-time analytics, the need for edge computing, specifically fog computing, has emerged. The study proposes a reference architecture, which couples the power of fog computing with a cloud-based big data-driven e-commerce platform. This hybrid approach aims to overcome the limitations of traditional big data platforms enabling faster processing and improved latency, which are crucial for real-time applications. The reference architecture provides the building blocks upon which e-commerce enterprises can build their personalized big data platforms. Additionally, potential challenges linked to the implementation of a fog-based architecture are discussed, providing insight into future research direction.

ABSTRACT. Efficient patient scheduling is a critical challenge in hospital management, impacting both patient satisfaction and resource utilization. This paper proposes a patient scheduling recommender system based on a multi-agent architecture. The system integrates real-time hospital data, patient preferences, and medical staff availability to optimize scheduling. The knowledge base incorporates patient health records and hospital resource constraints, allowing dynamic scheduling adjustments. A rule-based inference engine evaluates scheduling conflicts and prioritizes appointments based on urgency and availability. The proposed system enhances hospital efficiency by reducing patient waiting times and ensuring fair resource allocation.

ABSTRACT. Symbolic datatypes are central to abstract reasoning about dynamical systems. Successful examples include BDDs and SAT for finite-state systems, and polyhedra and SMT for discrete dynamical systems over certain infinite state spaces. Neural networks provide a natural symbolic datatype over continuous state spaces, with the added benefit of supporting key operations while being trainable. We advocate using neural networks for multiple purposes in reasoning about continuous state spaces, particularly for representing both deterministic dynamics—such as controllers—and correctness certificates. A correctness certificate is a succinct witness for a desired property of a dynamical system. For example, invariants and barrier functions certify safety, while Lyapunov functions and supermartingales certify progress. Stochastic barrier functions and supermartingales can further account for uncertainty (noise) in system dynamics. Established techniques from machine learning and formal reasoning about neural networks—such as SMT, abstract interpretation, and counterexample-guided refinement—enable the joint synthesis of controllers and correctness certificates. This allows for the synthesis of guaranteed-to-be-correct controllers, where both the controllers and their certificates are represented, learned, and verified as neural networks. This talk includes joint work with Krishnendu Chatterjee, Mathias Lechner, and Djordje Zikelic.

ABSTRACT. Breast cancer represents a major global health challenge, and early detection is crucial for the success of treatment. This research proposes a unified diagnostic pipeline designed to automate preliminary diagnosis based on breast ultrasound images, which are widely accessible and comfortable for patients. The pipeline integrates three distinct stages: detection, classification, and segmentation. The detection model is used to localize tumors, the classification network is used to identify image-level labels (normal, benign, malignant), and the segmentation model aims to delineate tumor regions. During evaluation, particularly when analyzing results based on tumor sizes, we found that incorporating a classification model improves the segmentation performance to distinguish healthy patients, by reducing false positives in normal cases. With all three models integrated, our approach achieves a Dice score comparable to state-of-the-art methods and leads to improved accessibility and reliability of automated diagnostic tools towards earlier breast cancer detection.

ABSTRACT. This paper presents a comparative analysis of various machine learning (ML) architectures in predicting and classifying neurological diseases, specifically epilepsy and multiple sclerosis (MS), using electroencephalogram (EEG) data. Using two distinct datasets—an open-source dataset for epilepsy and a private dataset for MS—our paper evaluates a diverse set of classifiers, and compares their performances. The results demonstrate that while ensemble methods like Random Forest and temporal models like LSTM achieve superior accuracy with large, well-defined datasets, simpler models and non-parametric approaches can perform better on smaller, more tightly structured datasets. These findings underline the critical need for aligning ML model selection with dataset characteristics and highlight potential pathways for future research.

ABSTRACT. Predicting corporate bankruptcy is vital for financial stability and risk management. This paper presents a large-scale, open data study on bankruptcy risk prediction for Romanian small and medium-sized enterprises (SMEs), leveraging a unique dataset of over 4 million company-year financial statements from 2019 to 2023. Our robust data pipeline addresses real-world challenges such as noise, missing values, and the prevalence of “atypical” (inactive, shell, or chronically distressed) companies, through systematic cleaning, outlier removal, and targeted exclusion from the healthy class. We construct and compare two modeling datasets: one using raw accounting indicators and another based on engineered financial ratios, both linked with multi-year company histories. We benchmark classical machine learning models (Random Forest, XGBoost, LightGBM, CatBoost, Decision Tree, Logistic Regression) and deep learning architectures (LSTM, GRU) on these variants. Our results show that tree-based ensemble ML models perform best on raw accounting data with multi-year sequences, achieving AUC-ROC >94.00%, while deep learning models achieve their highest precision and F1-score when trained on engineered financial ratios. Excluding atypical companies substantially improves model precision and minority class recall. These findings highlight the importance of careful dataset design, robust data curation, and multi-year context for effective large-scale bankruptcy prediction. Our work establishes best practices and actionable insights for credit risk assessment in emerging economies.

ABSTRACT. In this work, the Gravitational Search Algorithm (GSA) is introduced to approach Protein Structure Prediction (PSP) on the classical hydrophobic-polar (HP) lattice model. The importance of the PSP on the HP lattice model lies in its ability to simplify and model the complex process of protein folding, enabling researchers to study protein structure and stability using computationally efficient methods. Despite its simplicity, the HP model presents a highly complex and NP-hard optimization problem, making it an attractive target for metaheuristic search strategies. The proposed method utilizes the population-based optimization framework of GSA, where agents are guided by gravitational forces, and integrates it with the energy function of the HP lattice model to evaluate protein conformations based on hydrophobic interactions. GSA algorithm was compared with other approaches for predicting short and long protein sequences. Experimental results showed that DGSA provides a highly accurate, reproducible, and stabile prediction ability for the protein structure prediction on the 2D square and 3D cubic HP model.

ABSTRACT. Deep learning has made remarkable strides in medical imaging, often achieving high accuracy on benchmark datasets. In this study, we train a ResNet50 model on a standardized chest X-ray dataset for binary classification of pneumonia versus normal cases. Our training process achieves over 99\% training accuracy, with a validation accuracy and AUC of 100\%, indicating perfect performance on the validation split. However, when applied to a larger, unseen test set, the model’s performance drops to 90–94\% accuracy. Through detailed error analysis using Grad-CAM visualizations and confusion matrix inspection, we uncover consistent misclassifications attributed to variations in shoulder positioning, rib visibility, and image background artifacts. These anatomical and acquisition inconsistencies mislead the model, highlighting a significant challenge in generalization. This paper argues for the necessity of preprocessing strategies, such as lung-region isolation and semantic segmentation, to mitigate spurious correlations. Our findings emphasize that even models with seemingly perfect metrics require interpretability-driven validation before clinical deployment.

ABSTRACT. This study investigates the integration of wind energy with solar photovoltaic (PV) systems to improve wintertime residential heating, particularly during periods of no solar output such as during the winter night. Using five years of hourly weather data from a real-world station in western Romania, we analyzed wind speed patterns relevant to compact Vertical Axis Wind Turbines (VAWTs) rated at 3kW, 5kW, and 10kW. A simplified power curve model is applied to predict output based on observed wind profiles, and machine learning techniques are used to estimate hybrid system performance. Notably, Sanandrei, despite being geographically close to Timisoara, shows up to 8.9\% higher winter energy yield due to slightly better wind exposure and a higher elevation. The findings highlight the critical importance of local microclimate variations in renewable system planning.

ABSTRACT. Effective glucose management for people with Type 1 Diabetes (T1D) remains a significant clinical challenge, particularly in the context of intensive insulin therapy with advanced devices such as the Medtronic 740 insulin pump and the Guardian 3 continuous glucose monitoring (CGM) system. This study proposes a multivariate prediction approach to forecast future glucose levels and recommend optimal insulin dosing or corrective actions. We leverage rich, synchronized time-series data from the insulin pump, CGM, and user-logged events (e.g., meals, physical activity) to train and evaluate three state-of-the-art algorithms: Long Short-Term Memory (LSTM) networks, Temporal Convolutional Networks (TCN), and Transformer-based models. Each algorithm is assessed for its ability to accurately predict short-term glucose trends and recommend proactive interventions, with the goal of improving glycemic control and reducing the risk of hypoglycemia and hyperglycemia. Our findings provide actionable information for the development of closed-loop and decision support systems tailored for T1D management.

ABSTRACT. Traditional n-gram and image-resize features break down when binary code is packed, masked, or encrypted, because they rely on literal byte values and short-range context. This paper introduces Arithmetic Analysis, a family of symbolic feature grammars that encode additive and subtractive relationships among neighbouring bytes instead of their exact sequences. Five variants are defined: additive, subtractive, mixed, combinatorialmixed, and accumulative-combinatorial-mixed, offering a tunable balance between expressive power and dimensional growth. We evaluate eleven concrete feature sets on one hundred thousand Executable and Linkable Format (ELF) binary files. Experiments cover two regimes: plaintext files and duplicate files encrypted with Advanced Encryption Standard (AES) in Electronic Code Book (ECB) mode. A four-order accumulativecombinatorial- mixed grammar on clean data matches a bigram histogram at half the runtime and an F1 score of 99%. Under fullfile encryption, low-order arithmetic grammars retain meaningful discrimination, achieving an F1 score of around 67%, whereas frequency and resize-based methods collapse to random chance. Arithmetic Analysis delivers compact, interpretable, and obfuscation-resilient representations for static binary detection. Its operator grammar is domain-agnostic, opening future applications from malware triage to automatic text-generation metrics such as Arithmetic BLEU and Arithmetic ROUGE.