Modeling compaction parameters using support vector and decision tree regression algorithms

Abstract

Shortening the periods of compaction tests can be possible by analyzing the data obtained from previous laboratory tests with regression methods. The regression analysis applied to current data reduces the cost of experiments, saves time, and gives estimated outputs. In this study, the MLS-SVR, KB-SVR, and DTR algorithms were employed for the first time for the estimation of soil compaction parameters. The performances of these regression algorithms in estimating maximum dry unit weight (MDD) and optimum water content (OMC) were compared. Furthermore, the soil properties (fine-grained soil, sand, gravel, specific gravity, liquid limit, and plastic limit) were employed as inputs in the study. The data used for the study were supplied from the experimental soil tests from small dams in Nigde, a province in the southern part of Central Anatolia, Turkey. Polynomial-based KB-SVR yielded the best R-values with 0.93 in the prediction of both OMC and MDD. Moreover, in the multioutput estimation model, polynomial and RBF-based KB-SVR methods were successful with 0.98 and 0.99, respectively. Additionally, while the MSE value was 1.33 in the estimation of OMC, this value was 0.04 in the estimation of MDD. Accordingly, MDD was the most successfully estimated parameter in all processes. It was concluded that through the algorithms used in this study, the prediction of soil compaction parameters could be possible without the need for further laboratory tests.