This study investigates parameter estimation in a linear three-compartment pharmacokinetic model with absorption. Peripheral-compartment permutation symmetry is explicitly separated from local practical conditioning. Singular-value analysis is supplemented by noise-weighted local design comparisons. XGBoost and LightGBM are trained on 10,000 independently sampled parameter vectors, with clean and noise-augmented training evaluated separately. A bounded, scaled, three-start nonlinear least-squares (NLS) baseline is evaluated on the same held-out observations. On 1000 noise-free test vectors, mean relative parameter errors are 4.2460% and 4.3232% for clean-trained XGBoost and LightGBM, compared with 0.0158% for NLS. Noise experiments use 100 independent parameter vectors and 20 realizations per vector at each nonzero noise level. Bias, conditional standard deviation and RMSE are reported per parameter. The machine learning models exploit the parameter distribution within the ±20% training domain, whereas the bounded NLS baseline includes no explicit penalty toward the domain center. Therefore, their lower errors under noisy observations may partly reflect this prior information rather than an advantage specific to gradient boosting. The results distinguish accurate curve fitting from stable parameter recovery and do not establish universal superiority of gradient boosting over optimization. Conclusions are restricted to the stated synthetic parameter domain.
Knowledge graphs have emerged as a promising approach for representing educational content and supporting personalized learning. However, many existing knowledge graph-based e-learning systems rely on manually constructed domain models that require substantial expert effort and provide limited support for adaptive learning experiences. At the same time, gamification has been shown to improve learner engagement, but educational games often lack structured semantic representations of learner knowledge. This paper presents an adaptive knowledge graph game that integrates Large Language Models (LLMs), knowledge graph-based learner modeling, and game-based learning within a unified educational framework. The proposed system uses an LLM to generate a domain knowledge graph representing concepts and prerequisite relationships within a learning domain. Learner interactions are used to construct a learner knowledge graph that captures concept mastery and learning progress. By comparing the learner model with the domain knowledge graph, the system identifies potential knowledge gaps and generates adaptive learning quests targeting concepts that require further development. The framework incorporates game elements including quests, non-player characters, progression mechanics, and knowledge-based challenges designed to support engagement throughout the learning process. A prototype implementation was developed to demonstrate the integration of knowledge graph generation, learner modeling, gap analysis, adaptive quest generation, and gamified learning mechanics. This work contributes a practical framework for combining LLM-assisted knowledge graph construction with personalized educational gameplay and provides a foundation for future research on intelligent and adaptive learning environments. Unlike conventional graph-based educational systems that store learner information in proprietary graph databases or application-specific structures, AKGG employs Semantic Web technologies (RDF, OWL, and SPARQL) to provide transparent, explainable, and interoperable learner modeling. Adaptive learning decisions can be traced directly to explicit semantic relationships, mastery scores, and SPARQL query results, improving the explainability and auditability of educational recommendations. We further report a selection-policy ablation showing which parts of the adaptation mechanism actually determine what a learner is asked to do: the multi-term priority score changes the selected concept in a minority of decisions relative to a one-term mastery baseline, one of its five terms is provably inert under the current thresholds, and simulated session length is shown analytically to be invariant to the selection policy under the simulator’s assumptions, so the simulation cannot distinguish the policies on that outcome. The classroom deployment reported here is a feasibility pilot; no learning-outcome advantage is claimed.
This study investigates the structural optimisation of a rover wheel-support component using a bio-inspired bone-remodelling algorithm combined with the finite element method (FEM), the radial point interpolation method (RPIM), and the natural neighbour radial point interpolation method (NNRPIM). This study is divided into two stages. First, the influence of the discretisation technique, nodal density and remodelling parameters on the resulting material distributions is examined using a two-dimensional cantilever-beam benchmark. The bone-remodelling algorithm produced admissible load-path configurations for every numerical formulation considered, and the reconstructed benchmark models showed stiffness differences below 1% when cross-analysed with FEM, RPIM and NNRPIM. In the second stage, the algorithm is applied to a two-dimensional design domain representing the component connecting the wheel to the suspension system of the Perseverance rover. Three optimised topologies are independently generated with FEM, RPIM and NNRPIM, reconstructed, and subsequently cross-analysed with the three structural solvers. In two dimensions, the original model presents a mass-normalised stiffness between 205.86 and 210.74 N/mm/kg, whereas the bone-remodelled designs provide values between 351.01 and 446.51 N/mm/kg. For the extruded three-dimensional models, this indicator increases from 208.93–221.53 N/mm/kg to 369.85–476.63 N/mm/kg. The corresponding von Mises stress fields remain well below the aluminium yield strength, and the three numerical formulations provide consistent structural trends. The results show that the adopted bone-remodelling strategy can be transferred from biomechanical analysis to engineering structural optimisation and that RPIM and NNRPIM provide suitable alternatives to FEM for this iterative remodelling framework.
Assessing scholarly contribution across disciplines is challenging because bibliometric indicators capture different aspects of research activity and citation practices vary by field. This study evaluates a weighted cubic-mean composite (Hr) and a career-adaptive extension (DHr) across civil engineering, computer science, mathematics, and neuroscience. The analysis considers 4479 researchers and uses recognition by professional societies as an observable benchmark, while recognizing that awards are imperfect proxies for research quality. reduce selection optimism, we performed nested cross-validation with researcher-level grouping, selecting candidate indicators and metric pairs within inner folds and evaluating the fixed DHr reference formulation on held-out outer folds. On out-of-fold predictions, Hr AUCs were 0.556, 0.692, 0.768, and 0.713, while DHr AUCs were 0.559, 0.690, 0.767, and 0.712 across the four domains, respectively. Paired bootstrap confidence intervals for the Hr–DHr AUC differences included zero in every domain; therefore, these results do not establish that career-adaptive weighting improves discrimination over static aggregation. screening and metric-pair selection were conducted separately in each discipline, these AUCs reflect within-domain validation across four datasets and do not demonstrate transfer to an unseen discipline. The findings support further investigation of composite bibliometric indicators, but independent validation, transparent benchmarking, and fairness assessment are needed before consequential use.
Image classification is a fundamental task in computer vision. Although classical convolutional neural networks are particularly effective at extracting local spatial features, they are inherently limited in their ability to capture global frequency-domain information. To address this limitation, this paper introduces a hybrid model that couples a classical convolutional feature extraction branch with a quantum feature extraction branch built upon the quantum Fourier transform. The quantum branch maps the input into a low-dimensional representation and processes it through a sequence of quantum encoding, the quantum Fourier transform, a variational quantum circuit, and the inverse quantum Fourier transform; the resulting quantum features are concatenated with the classical features and jointly used for the final classification. Ablation studies quantify the individual contribution of each quantum component, and systematic experiments on the MNIST, Fashion-MNIST, and CIFAR-10 benchmark datasets, conducted over multiple independent random seeds on a strictly held-out test set, show that the complete architecture consistently outperforms its purely classical counterparts. All results and comparative analyses are presented in the experimental section.
