METAHEURISTIC HYPERPARAMETER OPTIMIZATION FOR DEEP LEARNING: A COMPARATIVE STUDY

Abstract

One of the primary challenges in developing effective deep learning models lies in identifying optimal hyperparameter configurations, particularly within high-dimensional and complex search spaces. Traditional tuning strategies, including grid search and random search, are widely known to be computationally inefficient and frequently yield suboptimal outcomes. To address this limitation, this study presents a comparative analysis of two metaheuristic algorithms — Particle Swarm Optimization (PSO) and Dwarf Mongoose Optimization (DMO) — as advanced alternatives for hyperparameter tuning in deep learning models trained on the CIFAR-10 dataset. Both algorithms were rigorously assessed using a comprehensive set of performance metrics, namely accuracy, precision, recall, and F1-score, supplemented by confusion matrix analysis to capture class-level behavior. Experimental findings confirm that both approaches yield substantial improvements in model performance. Notably, PSO demonstrated superior results, achieving a validation accuracy of 89.15% and a test accuracy of 87.82%, while DMO reached a final accuracy of 86.38%. In terms of optimization behavior, PSO exhibited greater convergence stability and more uniform class-level performance, whereas DMO proved more effective in broadly exploring the search space. Overall, this study reinforces the potential of metaheuristic-based optimization as a robust framework for hyperparameter tuning and underscores the critical role of optimization stability in achieving reliable model generalization.