Predicting Strength of High-Performance Concrete Using Gradient Boosting Machine Learning: A Comparative Analysis Between Manual and Grid Search Cross-Validation Hyperparameter Tuning
Ryan Tyler, Masengo Ilunga, Bolanle Ikotun, Omphemetse Zimbili
This study evaluates the effectiveness of a gradient boosting regression model in forecasting concrete strength by comparing three hyperparameter configurations: default settings, manual tuning, and automated Grid Search CV. A publicly available dataset of 1030 concrete mixes, featuring cement, slag, fly ash, water, superplasticiser, coarse and fine aggregates, and concrete age, was divided into an 80-20 train-test split. Model performance was assessed using mean absolute error (MAE), mean squared error (MSE), root mean squared error (RMSE), and the coefficient of determination (R2). The default model achieved R2 = 0.89, while targeted adjustments to the parameters, such as the number of estimators, raised R2 to 0.93. Manual fine-tuning of all hyperparameters simultaneously produced the best results of R2 = 0.94, marginally outperforming Grid Search CV 3- and 5-fold of R2 = 0.93. The number of estimators was identified as the most influential parameter. Although exhaustive grid search offers systematic optimisation with high runtimes, manual finetuning can yield superior accuracy within a constrained parameter space. Full Text
|
