This article presents the results of research on the application of the Analytic Hierarchy Process method for the selection of optimal process parameters in abrasive waterjet cutting of 21CrMoV5-7+QT steel. The selected process parameters significantly influence the cutting performance and quality of the machined surface, while parameters with negligible impact on the process can be identified and excluded from further analysis. The AHP method, as a multi-criteria decision-making approach, enables the evaluation and prioritization of technological parameters through the construction of a hierarchical decision structure consisting of objectives, criteria, and decision alternatives. The developed experimental design allowed for the identification of the most significant input variables affecting the abrasive waterjet cutting process, determination of parameter values ensuring the most favorable machining conditions, and assessment of the influence of parameter variation on overall process performance. The proposed approach supports the optimization of AWJ machining by considering multiple technological and quality-related criteria simultaneously, leading to improved efficiency, cutting quality, and process stability.
In Japanese high school mathematics, expected value should be taught not only as a computational procedure but also as a basis for judgment under uncertainty. However, concrete learner-centered instructional models for teaching this topic with digital tools remain limited. This study proposes a curriculum using Minecraft Education (ME) for the expected value component of the Mathematics A unit on counting and probability. The curriculum was designed as a five-lesson sequence consisting of basic operation training in ME, data collection through in-game fishing, data organization, expected value calculation, and student-designed application tasks. To examine the theoretical behavior of the scoring system prior to classroom implementation, a single preliminary 300-trial simulation was conducted under the same conditions assumed in the curriculum. Although the three team conditions were designed to yield the same expected value per trial (6.75 points; expected total score over 300 trials = 2025), the simulated total scores differed across teams (Team A = 2009, Team B = 1885, Team C = 1567). This discrepancy was largely explained by differences in variance across the scoring patterns. The simulation illustrates a key design feature of the curriculum: students can be asked to compare identical theoretical expected values with different observed outcomes in a finite number of trials. Because this study focused on curriculum design and a single preliminary simulation, classroom learning effectiveness was not evaluated. The proposed curriculum is therefore intended to support learners in distinguishing between theoretical expectation and observed outcomes and may serve as a bridge from expected value to broader statistical ideas, including variance and distributional reasoning.
Arabic retrieval poses a well-known difficulty: a single concept may surface in dozens of orthographic and morphological variants, causing keyword-based methods to miss semantically equivalent passages. We investigate this problem through a controlled empirical study of contrastive fine-tuning applied to three Arabic transformer encoders:AraBERTv2, ARBERT, and CAMeLBERT across three structurally distinct benchmarks (FAAQA, TyDiQA-Arabic, ARCD). All models train under identical conditions and are compared against two lexical baselines: plain BM25 and a morphology-aware BM25-Stem variant using ISRI stemming. As a dataset contribution, we release FAAQA-Enhanced: a cleaned, fully annotated, and dialectally augmented extension of the FAAQA medical QA corpus, growing from 1,443 to 4,159 question-answer pairs across 5 Arabic variety labels, 10 medical specialties, 6 question types, and 3 urgency levels, together with 1,304 retrieval triplets with BM25-mined hard negatives. Our retrieval study yields four findings. First, contrastive optimisation consistently delivers MRR improvements of 15–29% across all encoder dataset combinations. Second, pretraining corpus breadth predicts cross-dataset robustness more reliably than model size or architecture: ARBERT leads on formal content (FAAQA, MRR=0.76), while CAMeLBERT generalises better to topically diverse benchmarks (TyDiQA-Arabic, MRR = 0.61; ARCD, MRR=0.66). Third, extending the input sequence limit from 128 to 256 tokens produces a 5.7% MRR gain on TyDiQA-Arabic. Fourth, geometric analysis of the embedding spaces shows that contrastive training sharpens positive/negative separation in proportion to each encoder’s linguistic coverage. These findings provide practical guidance for Arabic retrieval and RAG systems.
Mathematical modeling based on dynamical systems theory is ideal for describing many natural phenomena. The use of dynamical systems allows for a better understanding of various processes and the prediction of subsequent states. However, predicting their behavior is often challenging due to limited observability, where only a single variable can be measured. This study investigates the efficacy of various machine learning frameworks in forecasting chaotic systems. These frameworks range from classical neighbors-based algorithms to reservoir computing. By employing phase-space reconstruction via time-delay embedding, we evaluated the predictive accuracy of models including k-nearest neighbors (k-nn), Random Forests, and Echo State Networks (ESN). The study utilized both the chaotic Thomas system and solar activity datasets. Our findings reveal that, for short-term forecasting, simpler models that leverage local dependencies often outperform more complex or computationally intensive architectures. Specifically, the kNN method consistently yielded the highest precision across both deterministic and real-world datasets. This research suggests that, in data-constrained environments, simple, locally based learning strategies provide a robust and efficient alternative for short-term prediction of nonlinear dynamics.
Music Genre Classification (MGC) has long been a research challenge in the field of Music Information Retrieval (MIR), with applications ranging from recommendation systems to the management of large-scale audio datasets. Although Deep Learning (DL) methods have been widely applied in this field, there is currently a lack of effective and consistent empirical benchmarks for systematically comparing different model architectures. This study conducts a comprehensive benchmarking of five representative DL models using the Free Music Archive (FMA) dataset, including Fully Connected (FC) neural networks, Convolutional Neural Networks (CNNs), Long Short-Term Memory (LSTM) networks, Gated Recurrent Units (GRUs), and Transformer architectures. All models utilized Mel-Frequency Cepstral Coefficients (MFCCs) as input features and were evaluated under a standardized benchmarking environment. To ensure a fair comparison between model architectures and to conduct robust empirical analysis, we employed multiple evaluation metrics, including precision, recall, F1-score, Receiver Operating Characteristic-Area Under the Curve (ROC-AUC), and the Kolmogorov–Smirnov (K-S) statistic, to capture both classification performance and distributional differences. The experimental results show that, across multiple evaluation metrics, the recurrent models outperformed other architectures, with the LSTM performing best (F1-score of 0.599 and K-S statistic of 0.608). These results highlight the importance of temporal modeling in MGC and demonstrate that, under MFCC-based feature conditions, recurrent architectures remain highly competitive compared to Transformer-based models. Overall, this study establishes a reproducible empirical benchmark and demonstrates the relative advantages of DL models for music genre re-classification on large datasets.
