Automatic Optimization of Hyperparameters for Deep Convolutional Neural Networks: Grid Search Enhanced with Coordinate Ascent

Convolutional neural network models of identical depth can yield significantly divergent results depending on the hyperparameter combination employed. Model performance can be improved by adjusting hyperparameter configurations; however, hyperparameter optimization typically incurs substantial computational resources and time. Thus, improving the efficiency of hyperparameter optimization is critical. This study adopts the coordinate ascent method, which provides only initial candidate values for each hyperparameter; at each iteration, only the single hyperparameter with the greatest influence on the model is altered, progressively expanding the search grid until accuracy converges. This method enables efficient and automated discovery of hyperparameter combinations that improve model accuracy. Experimental results show that, using the MWD dataset, the hyperparameter-optimized model achieves 95.71% accuracy on the validation set, and this hyperparameter combination can be regarded as an approximate global optimum. Furthermore, performance within the neighborhood of this combination proves stable, corroborating the robustness of the proposed hyperparameter optimization strategy.