Utilization of Support Vector Models and Gene Expression Programming for Soil Strength Modeling

Abstract

The subgrade strength of roads and highways is based on the California bearing ratio (CBR) value. In this investigation, attempts have been made to overcome the limited boundary condition approach by using advanced methods, support vector machine (SVM) and gene expression programming (GEP) for prediction of CBR value. A large and wide range of datasets of different types of soils have been utilized in the analysis. The grain size distribution, Atterberg’s limits and compaction characteristics of soils have been used as the input variables. Best models with different variables were developed by using GEP and the same were used for SVM analysis. The advantage of SVM over others is that it works on the principle of statistical risk minimization. A comparative study of SVM and GEP models indicates that the SVM has better predictability than GEP. Further, it was found that the five-input variable (including gravel content, sand content, plasticity index, maximum dry density and optimum moisture content) model is the best one to predict the CBR value. The detailed statistical analysis including Pearson coefficient correlation (R) and Error analysis have also been carried out. Based upon the statistical analysis, overfitting ratio of SVM was found to be 0.630 against the value of 1.02 in GEP analysis. Further, sensitivity analysis was carried out and it was found that the CBR value is highly dependent on gravel and sand contents. On the other hand, plastic limit plays an insignificant role in determining the CBR value of soils.