Title: A hybrid ant colony genetic optimisation framework for automated CNN architecture and hyperparameter tuning in image classification

Authors: Hong-Ren Chen; Mu-Yen Chen; Jia-Lang Xu; Yi-Syuan Wang; Po-Yen Hsu

Addresses: Department of Digital Content and Technology, National Taichung University of Education, No.140, Minsheng Rd., West Dist., Taichung City 403514, Taiwan, ROC ' Department of Engineering Science, National Cheng Kung University, Tainan City, 701, Taiwan, ROC ' Department of Applied Statistics, National Taichung University of Science and Technology, No. 129, Section 3, Sanmin Road, North District, Taichung City 404336, Taiwan, ROC ' Department of Engineering Science, National Cheng Kung University, Tainan City, 701, Taiwan, ROC ' Department of Engineering Science, National Cheng Kung University, Tainan City, 701, Taiwan, ROC

Abstract: Hyperparameter and architectural optimisation remain critical challenges in deep learning, directly affecting model accuracy, generalisation, and efficiency. This study proposes a hybrid ant colony genetic optimisation (ACGO) algorithm for image classification, integrating the pheromone-guided search of ant colony optimisation with the genetic diversity mechanisms of crossover and mutation. A novel encoding scheme for convolutional neural network (CNN) hyperparameters enables automated tuning across datasets of varying complexity. The proposed ACGO is evaluated against five metaheuristic algorithms - genetic algorithm, ant colony optimisation, coral reef optimisation, particle swarm optimisation, and Grey Wolf optimiser - on CIFAR-10 and CIFAR-100. ACGO achieved 81.1% accuracy on CIFAR-10 and 52.7% on CIFAR-100 (an absolute improvement of approximately 4% and 3% against competing methods), demonstrating strong adaptability and computational efficiency. These results highlight ACGO's potential as a scalable, robust optimisation framework for deep learning-based image classification.

Keywords: metaheuristic algorithms; hyperparameter optimisation; image classification; model encoding; convolutional neural networks; CNN.

DOI: 10.1504/IJAHUC.2026.155554

International Journal of Ad Hoc and Ubiquitous Computing, 2026 Vol.52 No.4, pp.243 - 256

Received: 14 Nov 2025
Accepted: 14 Feb 2026
Published online: 05 Aug 2026 *

Full-text access for editors Full-text access for subscribers Purchase this article Comment on this article