A Hybrid GA–PSO Framework for Neural Network Architecture and Parameter Optimization

The main motivation for this study is to develop a predictive framework that provides high accuracy at lower computational and experimental costs, resulting in better decision-making in the chosen application domain. Artificial neural networks (ANNs) are widely used for prediction, classification, and pattern recognition tasks. However, their performance is sensitive to the selection of architectural and learning parameters. Hence, an important research challenge is the effective selection of architectural and learning parameters. Several hybrid GA–PSO approaches have been proposed, but most of the existing studies simultaneously optimize network architecture and trainable parameters or focus on a single application domain. However, there is still a lack of systematic framework that optimizes these components separately and validates its performance on multiple heterogeneous datasets. To fill this gap, this study proposes a novel hybrid optimization algorithm, called GAPSO, which combines the genetic algorithm (GA) and particle swarm optimization (PSO) for efficient tuning of artificial neural network (ANN) parameters. The proposed framework is evaluated on five benchmark datasets, including AirPassengers, Sunspots, Death and Injury, Earthquake, and Insurance. In the proposed approach, PSO is used for determination of optimal network architecture (number of hidden neurons) and GA is used for optimization of connection weights and threshold values. The experimental results demonstrate that for four out of five datasets, the lowest MAPE values were achieved by GAPSO-ANN, and were competitive compared to ANN, GA-ANN, PSO-ANN, LSTM and XGBoost models. Additionally, the Wilcoxon signed-rank test showed statistically significant performance improvements (p = 0.03125).