Published Paper Details:

Hyperparameter optimization (HPO) of deep neural networks is a computationally expensive black-box problem that significantly affects model performance. Manual tuning is impractical for modern architectures, motivating the use of automated optimization methods. This paper proposes a hybrid metaheuristic algorithm that integrates the Aquila Optimizer (AO) with Genetic Algorithm (GA) operators for automated HPO of a convolutional neural network (CNN). The proposed method incorporates three mechanisms: opposition-based learning (OBL) for population initialization, a cosine-annealed mutation schedule, and distance-aware archive-guided crossover to preserve population diversity during exploitation. The optimizer searches for four CNN hyperparameters--learning rate, batch size, dropout rate, and SGD momentum--by minimizing classification error on a validation set. Experiments are conducted on four image classification benchmarks: MNIST, Fashion-MNIST, EMNIST, and CIFAR-10, using three independent random seeds and a fixed evaluation budget of 610 fitness evaluations per run. The proposed hybrid achieves accuracies of 98.87%, 89.38%, 99.08%, and 62.38% on the respective datasets, consistently outperforming the standalone AO and achieving performance comparable to the GA, with the largest improvement observed on Fashion-MNIST (+0.37%).