Using Decision-Tree Learning to Assess Liquefaction Potential from CPT and V_s

Damages attributed to earthquake-induced liquefaction have resulted in substantial loss of lives and property around the globe. This paper reports on a study that evaluates the effectiveness of bagged and boosted decision-tree machine learning methods to develop models for assessing liquefaction potential. Major advantages of decision tree learning include their ease of use and their production of easily inspected models with which engineers can better understand the mechanism of the liquefaction phenomenon and verify the trained models. This study demonstrates that the accuracies of the models learned by either bagged or boosted decision-tree learning are competitive with accuracies reported for other learning methods, such as neural network learning. Test-set accuracies of the models learned in this study were 96% on a dataset of 178 cone penetrometer soundings and 89% on a dataset of 225 shear wave velocity soundings, with combined training and testing accuracies of these models on the datasets of 98% and 96%, respectively.