Simplicity versus Complexity in Machine Learning Models—Focusing on Soil Resilient Modulus Prediction

The resilient modulus (M_R) of subgrade soil is an important factor in pavement design, which can be determined through repeated load laboratory testing. However, due to its high cost, time-intensive nature, and intricate process, geotechnical engineers and researchers have consistently expressed interest in creating an M_R prediction model using alternative soil test data. With advancements in prediction tools using machine learning (ML) algorithms, during the past decade, much research has been focused on implementing ML algorithms for the development of M_R prediction models. Many recent studies have been focused on using what is referred to as “black box” models, such as artificial neural networks (ANN). While these black box models provide high prediction accuracy, they are not explainable. In other words, there is no closed-form solution in which the relationship between the predicted label and the input features is explicitly stated. Moreover, these black box models are computationally complex and require advanced coding knowledge, making them inaccessible to many design professionals. In this paper, for illustration purposes, a dataset of 279 M_R test results that have been extensively used for ML-based model development of M_R is adopted from the literature. Different regression approaches, from simple regression models to advanced ML algorithms, are applied to this dataset, and the developed models are compared in terms of accuracy, simplicity, and explainability. Additionally, the bias and variance in error are investigated for different algorithms. It is concluded that the improved accuracy of the developed models is not always adequate to justify the effort, time, and complexity of adopting complex advanced algorithms instead of simple regression models. Moreover, it is shown that overly complex algorithms could lead to high variance, causing the algorithm to model random noise in the training data.