BOinEA: A Unified Hybrid Approach Integrating Bayesian Optimization and Evolutionary Algorithms for Neural Network Hyperparameter Tuning

Abstract

Hyperparameter optimization (HPO) is a critical yet computationally intensive task in modern neural network design, particularly as model architectures and search spaces increase in complexity. Classical methods such as grid search and random search are often inefficient in high-dimensional spaces, while Bayesian optimization (BO) and evolutionary algorithms (EAs) each offer complementary strengths but also face limitations related to scalability, convergence, and the exploration-exploitation balance. To address these challenges, we propose BOinEA, a novel hybrid HPO method that tightly integrates probabilistic surrogate modeling from BO with the population-based search dynamics of EAs. Unlike existing hybrid approaches that rely on sequential switching or loosely coupled components, BOinEA maintains continuous coordination between global exploration and local exploitation throughout the optimization process. We evaluate BOinEA on benchmark datasets, including CIFAR-10, Fashion-MNIST, and NSL-KDD, and compare it with established HPO techniques such as BOGP, SMAC, and standalone EAs. Experimental results show that BOinEA achieves the best test accuracy on CIFAR-10 and competitive performance on Fashion-MNIST and NSL-KDD, suggesting a favorable trade-off between accuracy, sample efficiency, and convergence speed. These findings suggest that BOinEA provides a robust and scalable framework for neural network hyperparameter optimization and contributes to ongoing advances in automated machine learning.