Application of Boosting-Based Ensemble Learning Method for the Prediction of Compression Index

Abstract

Obtaining geotechnical design parameters by conducting in situ or laboratory testing has always been challenging because of difficulty involved with handling, transportation, release of overburden pressure and poor laboratory conditions. Engineers thus depend on empiricism. Over the decades, many empirical correlations have been proposed to relate index properties of soils with geotechnical design parameters. Conventionally, regression-based methods have been applied to build these empirical relationships. These models though simple, often are not flexible enough to capture more complex relationships between the dependent and the independent variables. Also adding the right interaction terms or polynomials can be tricky and time-consuming. In recent years, advances in the field of data mining have produced robust and efficient ensemble machine learning algorithms like boosting methods, which not only can learn nonlinear relationships but also have strong boundaries in terms of generalization performance. However, boosting methods are yet to be used in regression problems in the field of geotechnical engineering. In this paper, a recently developed boosting method called extra gradient boosting method (XGBoost) is applied to predict the compression index of normally consolidated soils. An efficient grid search algorithm in combination with a three-way hold out technique is used to tune the hyperparameters of the XGBoost algorithm. This study suggests that application of XGBoost in combination with the grid search technique leads to an improvement of 8–11% in prediction accuracies compared to published results that used single- and multi-variable regression, artificial neural network for prediction of compression index. Consequently, XGBoost has potential applicability in estimation of soil properties.