Evaluation of MLP-ANN Training Algorithms for Modeling Soil Pore-Water Pressure Responses to Rainfall

Abstract

Knowledge of pore-water pressure responses to rainfall is vital in slope failure and slope hydrological studies. The performance of four artificial neural network (ANN) training algorithms was evaluated to identify the training algorithm appropriate for modeling the dynamics of soil pore-water pressure responses to rainfall patterns using multilayer perceptron (MLP) ANN. The ANN model comprised eight neurons in the input layer, four neurons in the hidden layer, and a single neuron in the output layer representing an 8-4-1 ANN architecture. The training algorithms evaluated include the gradient descent, gradient descent with momentum, scaled conjugate gradient, and Levenberg-Marquardt (LM). The performance of the training algorithms was evaluated using standard performance evaluation measures—root mean square error, coefficient of efficiency, and the time and number of epochs required to reach a predefined accuracy. It was found that all the training algorithms could be used in the prediction of pore-water pressures. However, the LM algorithm required the least time and epochs for training the network and gave the minimum error during both training and testing. The LM training algorithm is therefore proposed as an ideal and fast training algorithm for modeling the dynamics of soil pore-water pressure changes in response to rainfall patterns.