Performance of Genetic Programming and Multivariate Adaptive Regression Spline Models to Predict Vibration Response of Geocell Reinforced Soil Bed: A Comparative Study

This paper explores the applicability of rapidly growing machine learning techniques (MLTs) for predicting the vibration response of geocell reinforced soil beds. Peak particle velocity (PPV) is used as an indicator to represent the vibration response. Two machine learning techniques namely, Genetic programming (GP), and multivariate adaptive regression splines (MARS) are used for the PPV prediction. Primarily, a series of field vibration tests were conducted over the geocell reinforced beds to obtain the dataset for model development. During the field test, PPV variation was studied by varying the test parameters namely, footing embedment, dynamic load, modulus of infill material, width, and depth of placement of geocell mattress. In total, 240 field measurements were used to formulate the PPV prediction models. The prediction performance of a developed model was examined by determining the different statistical indices. In addition, the ranking of each input parameter was calculated to identify the parameter, which influences the PPV most. According to the outcome of developed models, coefficient of determination (R²) values of (0.9918, 0.9852), and (0.9949, 0.9941), were observed for training and testing data sets of GP and MARS models, respectively. The sensitivity analysis of both the models revealed that the distance from the source to the measurement point indicating the damping properties of the reinforced bed is predominantly affecting PPV. Further, a comparative study has been carried out to examine the efficiency of the developed model in predicting the PPV response at the unknown dynamic excitation. The results of the comparative analysis revealed that the MARS model exhibits a high degree of accuracy in predicting the PPV variation in comparison to GP.