A Comparative Assessment of Regularized Regression Techniques for Modeling the Mechanical Properties of Rubberized Concrete

Abstract

Over the last few decades, many researchers have investigated the properties and behavior of concrete mixtures incorporating rubber-based solid wastes as a partial substitution of natural aggregates. Within this context, they have conducted experimental studies and developed numerical models that simulate the nature of rubberized concrete. Some of these mathematical simulations were intended to provide a rapid mixture of proportioning approaches and property estimation methods. Currently, it is believed that regression analysis provides an effective tool to simply construct a mathematical expression that models a set of data. For that reason, multiple linear regression was extensively utilized in predicting rubberized concrete properties in the literature. However, the performances of regularized regression analysis approaches were not evaluated even though they provide better alternatives to traditional regression methods in terms of controlling the overfitting issue. This study aims to assess the performance of Ridge, Lasso, and elastic net regression models in estimating the compressive and tensile strengths, and modulus of elasticity of rubberized concrete. Additionally, it intends to benchmark their capabilities against the traditional multiple linear regression method. Multiple linear regression, Ridge regression, Lasso regression, ElasticNet regression, Bayesian ridge regression, Stochastic gradient descent, Huber regression, and Quantile regression methods were used in the study. In general, the research findings illustrated the superior performance of regression assessment in modeling the mechanical properties of rubberized concrete. Indeed rubberized concrete mechanical properties can be better modeled using regularized regression techniques, such as ElasticNet-based SGD compared to traditional methods, such as MLR.