Abstract
This paper introduces a hybrid optimization framework that integrates a Machine Learning (ML) model with the Non-Dominated Sorting Genetic Algorithm-II (NSGAII) for efficient hyperparameter tuning of Convolutional Neural Networks (CNNs). The application focuses on the MNIST dataset for handwritten digit recognition. The proposed Interactive Evolutionary NSGA-II (IE-NSGA-II) leverages a Decision Tree classifier to predict Pareto front ranks after an initial phase of traditional NSGA-II execution. This approach aims to reduce computational overhead associated with standard NSGA-II procedures without significantly compromising solution quality. Experimental results demonstrate that the IE-NSGA-II achieves competitive performance compared to the traditional NSGA-II. At 20 generations, the IE-NSGA-II obtained a validation accuracy of 99.02 % with a validation loss of 0.0349, while the traditional NSGA-II achieved 99.08 % accuracy and 0.0335 loss. The IE-NSGA-II showed a modest reduction in mean execution time, being approximately 4.63 % faster than traditional NSGA-II at 20 generations. This study highlights the potential of MLassisted evolutionary algorithms for creating scalable and efficient hyperparameter optimization techniques in deep learning.

This publication has 4 references indexed in Scilit: