Impact of hyperparameter tuning on hybridised convolutional neural networks for pathloss modelling in mobile communication systems

Abstract

The performance of machine learning models, particularly Convolutional Neural Networks (CNNs), is profoundly influenced by effective hyperparameter tuning. However, a comprehensive understanding of how these hyperparameters affect the predictive accuracy of CNN-based pathloss models has not been adequately carried out. This study explores the role of hyper-parameter tuning in a hybridised CNN architecture that integrates DenseNet121 and ResNet50 to enhance pathloss prediction in mobile network environments. Field measurements were conducted along strategically selected urban and suburban routes in Ilorin, Kwara State, Nigeria. The results revealed the critical influence of key hyperparameters, such ashidden layers, batch size, training epochs, and computational efficiency, on model performance. Initially, with only two (2) hidden layers, the model showed suboptimal predictive accuracy, characterised by an MAE of 25.15, a  MSE of 34.43, and a highly negative R² value of 6.01. However, increasing the hidden layers to seventeen(17) yielded a substantial improvement, with the MAE reducing to 2.08, the MSE decreasing to 7.35, and the R² shifting positively to 0.80. Further analysis of batch sizes revealed that smaller sizes resulted in poor model performance, increasing it to 8 significantly enhanced accuracy. Additionally, an increase in training epochs from 50 to 200 led to a marked reduction in prediction errors, albeit at the expense of extended training time per iteration. These findings underscore the pivotal role of strategic hyperparameter selection in optimising CNN-based pathloss modelling, offering valuable insights for enhancing predictive performance in mobile network systems.