CNN Hyperparameter Optimization for Image Classification Using Genetic Algorithm: A Simulation Study

Abstract

Manual determination of Convolutional Neural Network (CNN) hyperparameters is a time-consuming, error-prone, and non-scalable process for high-dimensional search spaces. Genetic Algorithms (GAs) offer a promising global search approach to automate this process, yet a systematic comparison of GAs with other methods—grid search, random search, and Bayesian optimization in the context of image classification remains limited. This study aims to analyze and compare the performance of four CNN hyperparameter optimization methods through a literature-based simulation study covering 2022–2026. The aspects compared include final classification accuracy, convergence patterns, and computational efficiency. The synthesis shows that GAs consistently achieve higher accuracy than grid search and random search; a hybrid GA-PSO approach reaches 94.47% accuracy on CIFAR-10, surpassing manually designed architectures and single evolutionary methods. Bayesian optimization offers competitive convergence speed but is susceptible in high-dimensional hyperparameter spaces. GAs excel in global search-space exploration and can avoid premature convergence through selection, crossover, and mutation mechanisms. This study concludes that GAs are an effective and flexible method for CNN hyperparameter optimization, especially when the search space is complex and fitness evaluation is computationally expensive.