Amirkabir University of TechnologyAUT Journal of Modeling and Simulation2588-295357220251201Efficient Arabic Hate Speech Detection via LLaMA-3: A Prompting and Instruction-Tuning Approach127138587510.22060/miscj.2025.24262.5414ENAnas KhudhurAbbassAlborz Campus, University of Tehran, Tehran, Iran0000-0002-9793-5907HeshaamFailiCollege of Engineering, School of Electrical and Computer Engineering, University of Tehran, Tehran, Iran0000-0003-0443-3762Journal Article20250713Cyberspace generates user-generated content daily, but also gives freedom of expression, potentially spreading hate speech and endangering minorities. So, there’s the issue of identifying hate speech as quickly as possible to stop it from being shared. This is particularly challenging for Arabic given its rich morphology and lack of good quality linguistic resources. In this article, we examine zero-shot and few-shot prompting for detecting Arab hate speech using the LLaMA-3-8B language model while also refining performance via supervised fine-tuning utilizing a custom instruction-based dataset. In the zero-shot approach, the outputs from the model are unstructured textual outputs, so we take the unstructured responses and run them through a lightweight TF-IDF + Logistic Regression classifier to classify the responses in one of the predefined hate speech categories. To obtain better classification, we construct the instruction-based training set by creating tweet embeddings from Arabic-BERT and using K-Means clustering to enforce semantic/topical variety. Next, we use the GPT-4o model to generate representative instructions from each cluster and create an instruction-based fine-tuning data set. We then fine-tune LLaMA-3-8B using QLoRA, which also allows the model to be fine-tuned with a lower memory footprint. The experimental results presented in this paper show that zero-shot and few-shot prompting achieved relatively low F1-scores of 42.2% and 45.0%, respectively, and instruction-tuning fine-tuning achieves the overall performance of an F1-score of 90.1%, which exceeds stronger benchmarks like AraBERT. Our results exemplify the potential impact of instruction tuning and QLoRA-based fine-tuning over prompting-based approaches in low-resource contexts like Arabic.https://miscj.aut.ac.ir/article_5875_767d01b4bac1a1e8824c9b9f7cc79a04.pdfAmirkabir University of TechnologyAUT Journal of Modeling and Simulation2588-295357220251201Designing a new robust control method to reduce LFOs in the power system139156592810.22060/miscj.2025.24959.5443ENFarhadAmiriDepartment of Electrical Engineering, Tafresh University, Tafresh 39518-79611, Iran0000-0001-9335-7823SajadSadrDepartment of Electrical Engineering, Tafresh University, Tafresh 39518-79611, Iran0000-0002-6113-8930Journal Article20251022LFOs are among the most significant variables that lead to the loss of power system stability and the reduction of transmission line capacity. The SSSC is one of the parts used to increase the power system's stability. In this paper, in order to improve the stability of the power system and also to reduce the LFOs, the design of the SSSC equipped with a new robust controller is discussed, which is called output feedback based on LMI. The stability criterion of the suggested method is established based on the Lyapunov stability theory. The control objectives in the proposed method are the asymptotic stability of the system under the influence of disruption and parameter uncertainty as well as minimizing the impact of disturbance on the system states. SSSC equipped with output feedback method based on LMI with compensation based on RMPC methods, Fuzzy lead-lag optimized by PSO, Fuzzy lead-lag optimized by MWOA has been compared and the results are that the proposed method has a favorable performance compared to other control methods presented and has been able to withstand the uncertainty of parameters and disturbances. In order to compare the performance of the proposed method, simulations have been performed in different scenarios. The maximum deviations related to the rotor angular velocity are improved by 58% using the proposed method. The settling time related to rotor angular velocity deviations has been improved by 6% using the proposed method.https://miscj.aut.ac.ir/article_5928_8ae1da0fe37c98412768453f82490da2.pdfAmirkabir University of TechnologyAUT Journal of Modeling and Simulation2588-295357220251201Artificial Neural Network-Based Feedback Control Strategy for Epidemiological SIR and SIER Models157172595110.22060/miscj.2026.24611.5428ENOussamaChayoukhDepartment of Mathematics and Computer Science, Ben M'Sick Faculty of Science, Hassan II University of Casablanca, Casablanca, Morocco.0009-0003-9027-4637OmarZakaryDepartment of Mathematics and Computer Science, Ben M'Sick Faculty of Science, Hassan II University of Casablanca, Casablanca, Morocco.0000-0003-0176-8233MariamRedouaneDepartment of Mathematics and Applications, Faculty of Science and Techniques, Abdelmalek Essaadi University, Tangier, Morocco.AadilLahrouzDepartment of Mathematics and Applications, Faculty of Science and Techniques, Abdelmalek Essaadi University, Tangier, Morocco.0000-0002-6050-6586Journal Article20250823We investigate a branched artificial neural network (ANN) feedback controller for mitigating outbreaks in compartmental epidemic models. The architecture couples a shared trunk with two specialized branches and employs Soboleva–modified hyperbolic tangent (SMHT) activations to approximate the shape and boundedness of analytic control laws, yielding smooth, non–bang–bang signals suited to implementable interventions. The network is trained offline in supervised fashion on synthetic SIR trajectories labelled by a control that steers the infected population toward a low terminal target over a fixed horizon. On unseen SIR scenarios, the learned policy lowers peak prevalence and shortens outbreak duration relative to uncontrolled dynamics. When compared against simple baselines; however, the ANN achieves these outcomes with markedly smoother profiles and reduced actuation effort (time integral of the control), a property desirable for practice. Without retraining, the controller transfers to SEIR and retains qualitative benefits consistent with partial observability induced by the latent exposed class. We evaluate our suggested controller against conventional neural network baselines through ablation studies and robustness tests incorporating multiplicative process noise. The results demonstrate that our branched architecture reduces the attack size and peak infection with a comparable control effort. Importantly, the controller exhibits smooth, bounded actuation signals even when subjected to significant uncertainty. We discuss limitations and outline extensions: identification from data, observer design for latent/noisy states, explicit resource and rate constraints, and online adaptation under distribution shift.https://miscj.aut.ac.ir/article_5951_800b03685c22049f049801f6841861a2.pdfAmirkabir University of TechnologyAUT Journal of Modeling and Simulation2588-295357220251201A Hybrid Deep Graph and Kernel Ensemble Approach to Mental Health Prediction in Social Network173190596710.22060/miscj.2026.23944.5403ENNazilaTaghvaeiFaculty of Computer and Information Technology Engineering, Qazvin Branch, Islamic Azad University, Qazvin, IranBehroozMasoumiFaculty of Computer and Information Technology Engineering, Qazvin Branch, Islamic Azad University, Qazvin, IranMohammadRezaKeyvanpourDepartment of Computer Engineering, Alzahra University, Vanak, Tehran, IranOmidSojoodiFaculty of Computer and Information Technology Engineering, Qazvin Branch, Islamic Azad University, Qazvin, IranJournal Article20250223Mental-health forecasting from social-media data is a complex multimodal challenge involving temporal, textual, and relational information. This study presents a hybrid two-stage framework that integrates a Long Short-Term Memory (LSTM)-based graph ensemble with an Ensemble Deep Kernel Learning (EDKL) meta-model to predict depressive risk and emotional trajectories within online social networks. In Stage 1, user-level representations are encoded using an LSTM encoder combined with multiple graph neural backbones, including Graph Convolutional Networks (GCN), Graph Attention Networks (GAT), and Graph Transformer Networks (GTN). Their outputs are stacked and calibrated via logistic regression to produce reliable depression-risk probabilities. In Stage 2, an EDKL meta-learner aggregates predictions from heterogeneous deep models (MLP, CNN, and LSTM) through kernel ridge regression with hybrid kernels optimized by a meta-heuristic search algorithm. This hybrid architecture supports robust, fine-grained forecasting across temporal, behavioral, and relational dimensions. Experiments on publicly available Twitter and MHASN datasets demonstrate substantial improvements over transformer-based and single-stage baselines, achieving up to 99% accuracy with consistently low error variance. The study also addresses ethical considerations related to privacy, bias, and potential misuse, emphasizes reproducibility through transparent experimental protocols, and outlines promising directions for future multimodal extensions, including richer linguistic, visual, and interaction signals for clinically relevant mental-health monitoring.https://miscj.aut.ac.ir/article_5967_8682cc30db9c025ecd3fee433f8ab54c.pdfAmirkabir University of TechnologyAUT Journal of Modeling and Simulation2588-295357220251201Complete Automated Structure Discovery and Parameter Estimation for Piecewise Affine Models with Guaranteed Convergence191214598110.22060/miscj.2026.25021.5447ENMohammadLotfiDepartment of Electrical Engineering
Amirkabir University of Technology
Tehran, IranMohammad BagherMenhajDepartment of Electrical Engineering
Amirkabir University of Technology
Tehran, Iran0000-0002-9470-5532MehdiKarrariDepartment of Electrical Engineering
Amirkabir University of Technology
Tehran, IranMohammadHaeriDepartment of Electrical Engineering
Sharif University of Technology, Tehran, IranJournal Article20251031This paper presents a novel, fully automated framework for the complete structure discovery and parameter estimation of Piecewise Affine (PWA) models. To the best of our knowledge, this is the first approach that simultaneously determines the number of submodels, their orders, parameter vectors, and polyhedral partitions from data, without any prior structural knowledge or the need for tuning parameters. The methodology integrates three key innovations: (1) Automated submodel order selection via Orthogonal Least Squares with an Error-to-Signal Ratio test; (2) A clustering-based algorithm for determining the number of submodels and generating a robust initial labeled dataset; and (3) An iterative algorithm that integrates a novel self-labeling support vector machine (SL-SVM) for estimating polyhedral partitions with a recursive least squares (RLS) scheme for refining submodel parameters, both with guaranteed convergence. Theoretical analysis demonstrates both computational efficiency and convergence properties, with the SL-SVM algorithm significantly reducing complexity compared to standard SVM. Extensive simulations validate the framework's performance across multiple benchmark systems, achieving Best Fit Rates exceeding 98% in scenarios of complete structural uncertainty. The approach consistently outperforms existing methods in accuracy while maintaining computational efficiency. Furthermore, we demonstrate the method's applicability to nonlinear system identification through PWARX approximation, showcasing its versatility for practical engineering applications. The proposed framework represents a significant advancement in automated system identification, providing a comprehensive solution for black-box modeling of hybrid and nonlinear systems.https://miscj.aut.ac.ir/article_5981_abb9d15b3293a96a3ea116867b2b16d5.pdfAmirkabir University of TechnologyAUT Journal of Modeling and Simulation2588-295357220251201Distributional Soft Actor-Critic with Adaptive Entropy Regularization: An Extended Theoretical Analysis215228604710.22060/miscj.2026.23574.5387ENMeysamFoziDepartment of Computer Engineering, Amirkabir University of Technology, Tehran, IranMohammad MehdiEbadzadehDepartment of Computer Engineering, Amirkabir University of Technology, Tehran, Iran0000-0001-6466-5229Journal Article20241006In reinforcement learning, one of the key challenges is the overestimation of the Q-value function, which can negatively impact policy performance. This paper introduces an adaptive extension of the distributional soft actor-critic (DSAC) algorithm, designed specifically for continuous control tasks, with the goal of mitigating Q-value overestimation. In addition, the proposed approach addresses the exploration-exploitation trade-off by taking into account model variance. We develop and evaluate four distinct versions of this adaptive extension, each incorporating different entropy regularization techniques: linearly decaying, exponentially decaying, linear adaptive, and exponentially adaptive regularization. These regularization methods are applied during the training process to balance exploration and exploitation more effectively. Our experimental results, conducted on OpenAI’s MuJoCo humanoid control tasks, demonstrate that the exponentially adaptive entropy regularization version of the DSAC algorithm performs significantly better than both the baseline method and the other proposed extensions. This performance improvement highlights the importance of adaptive entropy regularization strategies in reinforcement learning, particularly for tasks requiring fine-tuned control in continuous environments. The findings suggest that the proposed adaptive DSAC algorithm not only enhances learning stability by reducing overestimation but also offers a more efficient solution to the exploration-exploitation dilemma, providing a promising direction for future research in reinforcement learning for continuous control settings.https://miscj.aut.ac.ir/article_6047_98baeb82b676b662e12a7af8ad9212f6.pdf