Journal of Artificial Intelligence and Metaheuristics

When multiple LLM-assisted editors maintain a shared knowledge base, the conflicts that arise are seldom outright factual contradictions. More often, two users reading the same source extract compatible assertions that differ in specificity, coverage, or editorial emphasis. This paper asks a prior question: when should a conflict be resolved by selecting a winner, and when should both perspectives be preserved? We study this question in a controlled single-source collaborative setting. Five role-conditioned instances of GLM-4-Flash—simulating a senior researcher, a teacher, a student, a clinician, and a conservative editor—independently extract assertions from AI-Wiki-21, a 21-page AI knowledge base. Their outputs yield 4,556 candidate conflict pairs; annotation of 138 selected pairs produces a 100-conflict gold standard. Three findings emerge, each with scope limited to the single-source, single- LLM-family, simulated-role setting studied. First, confirmed conflicts are exclusively editorial—54% Granularity and 46% Coverage—with zero Factual conflicts, suggesting that in this controlled setting the dominant challenge is editorial coordination, not truth arbitration. Second, direct LLM judgment achieves 97.5% agreement with human annotators on this editorial-conflict gold standard, while a DPO-inspired quality heuristic (DIQS) reaches 81.0% without API overhead; collective voting plateaus at 41.8%, echoing social-choice instability. Third, sensitivity analysis of the fiber split threshold across τ ∈ [0.1,0.9] identifies τ∗ = 0.5 as the stable optimum, partitioning 3.1% of conflicts while preserving diverse user perspectives. The results also raise ethical questions about automated knowledge governance.

Point-of-Sale (POS) systems support retail transaction processing, inventory management, and operational decisionmaking. However, many traditional POS systems rely on a single centralized database, creating a critical point of failure during server faults, database errors, or network disruptions. Conventional POS systems may also lack intelligent forecasting capabilities, causing inventory decisions to depend on manual estimation or fixed reorder rules. These limitations can lead to operational interruption, inventory inconsistency, and inefficient stock management. This study proposes a high-availability smart POS system integrated with AI-based inventory forecasting. The framework uses a dual-database architecture consisting of a primary database and a secondary backup database. Application-level failover redirects operations automatically when the primary database becomes unavailable, while an event-driven backup mechanism synchronizes important data after critical transactions. The forecasting component was developed using the Rossmann Store Sales dataset. Several models were evaluated, including Random Forest (RF), Bagging, CatBoost, XGBoost, and KNN Regressor. Optimization algorithms including PSO, MVO, HHO, and DE were also integrated with RF. The results showed that baseline RF achieved the best overall performance, with MSE = 0.080454, RMSE = 0.283643, MAE = 0.184264, R2 = 0.919712, and WI = 0.978542. Among optimized models, PSO + RF achieved the strongest optimizer-based result, with MSE = 0.081480 and RMSE = 0.285446. The implemented system also demonstrated successful database connectivity, failover readiness, and automatic backup execution. The findings confirm that the proposed framework can support continuous retail operation while providing accurate AI-based forecasting and inventory recommendations.

Accurate solar radiation forecasting is essential for improving the reliability of photovoltaic energy generation and supporting effective solar battery management, particularly because solar radiation is highly variable and depends on nonlinear interactions among meteorological and temporal factors. Although conventional prediction methods can provide useful estimates, they often struggle to capture the sequential behavior of solar radiation caused by daily sunlight cycles, atmospheric variation, and changing weather conditions. Therefore, this study aims to develop and evaluate deep learning models for predicting Solar_radiation using meteorological data collected from the HI-SEAS weather station over four months, from September to December 2016, where the main input variables include temperature, humidity, pressure, wind direction, wind speed, and time-related features. Five recurrent deep learning models were implemented and compared, namely Recurrent Neural Network (RNN), Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), Bidirectional Long Short Term Memory (BiLSTM), and Attention-LSTM. Before model training, the dataset was preprocessed by handling missing values, checking temporal consistency, arranging the observations chronologically, and applying Min–Max normalization to ensure stable learning. Model performance was assessed using multiple regression metrics, including MSE, RMSE, MAE, MBE, correlation coefficient, R2, RRMSE, NSE, and WI. The experimental results showed that BiLSTM achieved the best overall forecasting performance, with an MSE of 0.0014, RMSE of 0.0379, MAE of 0.0182, MBE of 0.0039, correlation coefficient of 0.9750, R2 of 0.9494, RRMSE of 0.3645, NSE of 0.9494, and WI of 0.9865. GRU and RNN also produced competitive results, achieving RMSE values of 0.0381 and 0.0382 and R2 values of 0.9489 and 0.9486, respectively, while Attention-LSTM showed comparatively lower performance with an RMSE of 0.0492 and R2 of 0.9149. These findings indicate that recurrent deep learning models are effective for learning nonlinear and temporal patterns in solar radiation data, with BiLSTM providing the most accurate and reliable predictions. The proposed forecasting framework can support photovoltaic energy planning and solar battery decision-making by estimating future solar radiation levels and helping determine whether solar energy utilization will be feasible under expected weather conditions.

The growing availability of educational data has prompted the use of machine learning methods to predict student academic performance and support data-driven decision-making in education. Nevertheless, such models for predicting performance rely heavily on proper data preprocessing, model selection, and optimal hyperparameter settings. This research proposes a hybrid predictive architecture that combines machine learning classifiers with bio-inspired metaheuristic optimization algorithms to improve classification efficiency in educational data mining. It is based on the xAPI-Edu.A dataset of 480 students’ demographic, academic, and behavioral characteristics is used to first analyze a set of baseline machine learning models, including Random Forest, XGBoost, Support Vector Machine, Multilayer Perceptron, K-Nearest Neighbors, and Gaussian Naive Bayes, using standard classification metrics. The initial experimental findings on the baseline layer show that the Random Forest classifier outperforms the other models before optimization, achieving accuracies of 0.8889 and 0.8814, and F-scores of 0.8889 and 0.8814, respectively, indicating strong generalization and equal discrimination among the classes. To further improve the predictive performance, the state-of-the-art metaheuristic algorithms, i.e., the Al-Biruni Earth Radius Optimizer (BER), the Gray Wolf Optimizer (GWO), the Particle Swarm Optimization (PSO), the Genetic Algorithms (GA) and the Whale Optimization Algorithms (WOA) are adopted to optimize the hyperparameters of the Random Forest. It has been experimentally demonstrated that every optimization approach provides a measurable performance increase, but the BER-optimized Random Forest consistently performs better. In particular, the BER-Random Forest model achieves an F-score of 0.9477 and an accuracy of 0.9439, both of which are much higher than the baseline configuration. Full statistical and visual analyses, such as kernel density estimation, Z-score heatmaps, and swarm plots, also support the strength, stability and superiority of the proposed BER-based optimization framework. Such findings demonstrate the effectiveness of metaheuristic-based hyperparameter optimization in educational predictive analytics and provide significant insights into the creation of intelligent, efficient, and data-driven systems of academic assistance.

The ability to accurately predict student academic performance is now an established pillar of contemporary educational data mining, driven by the desire for data-driven interventions and personalized learning. This paper presents a mixed-methodology optimization and machine learning model that combines the Al-Biruni Earth Radius-Stochastic Fractal Search (BER-SFS) algorithm with eXtreme Gradient Boosting (XGBoost) to improve predictive accuracy and model stability in forecasting student performance. The proposed BER-SFS + XGBoost model is an optimized system that systematically optimises the hyperparameter based on dual exploration-exploitation mechanism to reduce the Mean Squared Error (MSE) of the baseline XGBoost model of 0.0226 to 0.00029 and the Root Mean Squared Error (RMSE) of the baseline XGBoost model to 0.1504 to 0.00194 and the coefficient of determination R2 of 0.9019 to Comparison to the other metaheuristics, including theGrey Wolf Optimizer (GWO), Particle Swarm Optimization (PSO) and Whale Optimization Algorithm (WOA) proved the superiority of the suggested hybrid model in all metrics of the evaluation. These findings support the potential to combine the dynamics of geometric exploration provided by BER with those of diffusion-based refinement offered by SFS to achieve high generalization and minimal bias. The implications extend beyond predictive analytics, demonstrating that the hybrid metaheuristic optimization approach can serve as a scalable, explainable method for adaptive educational systems, enabling early detection of at-risk learners and the development of data-driven academic support mechanisms.

Keystroke Dynamics Analysis (KDA) is a prominent behavioral biometric technique for continuous user authentication in digital environments. Yet, keystroke timing prediction remains challenging due to individual typing variability, temporal inconsistencies, and the tendency of machine learning models to overfit in high dimensional spaces when hyperparameters are poorly tuned. This study formulates the task as predicting keystroke timing intervals—dwell times, keydown–keydown latencies, and keyup–keydown latencies—for a fixed password sequence. We introduce a predictive framework that integrates the Dipper Throated Optimizer (DTO) with regression modeling, using a sequential dual optimization strategy: binary DTO (bDTO) first selects informative feature subsets, followed by standard DTO to fine-tune the hyperparameters of a Gradient Boosting Regressor (GBR). This design balances exploration and exploitation to address the complexity of optimization in behavioral biometric data. Experimental validation on the Keystroke Dynamics Benchmark Dataset demonstrates stepwise performance gains: the baseline GBR achieved an MSE of 0.014244, reduced to 0.004768 after bDTO based feature selection (66.5% improvement), and further refined to an MSE of 0.000003 with DTO hyperparameter tuning (99.97% relative improvement), a result interpreted with caution due to potential overfitting risks. The optimized model also attained R2 = 0.9824, Nash–Sutcliffe Efficiency = 0.9786, and Willmott Index = 0.9810, underscoring strong predictive agreement between observed and predicted timing intervals.