Determination of Suitable Hyperparameters of Artificial Neural Network for the Best Prediction of Geotechnical Properties of Soil

Abstract

Abstract: The artificial neural network is robust in predicting soil properties. The present study aims to determine the suitable hyperparameters such as number of hidden layers, neurons, and backpropagation algorithms for the best prediction of geotechnical properties of soil. The supervised learning category-based multilayer perceptron artificial neural network approach is used, and models are developed in MATLAB R2020a. The ANN models are configured with neurons (5, 10 & 15), hidden layers (one to five), and a backpropagation algorithm (LM, BFG, SCG, GDA, GD & GDA). Fifteen ANN models are developed for each algorithm. The study shows that the LM, BFG, and SCG algorithm-based ANN models require strongly (0.61-0.8) to very strongly (0.81-1) correlated datasets. On the other hand, the GDM, GD, and GDA algorithm-based ANN models require only strongly correlated datasets to achieve a performance of more than 0.9. In most cases, it is also found that the GDM, GD, and GDA algorithm-based ANN models achieve high performance with three hidden layers interconnected with ten neurons. Still, LM algorithm-based ANN model achieves high performance with a single hidden layer interconnected with 5/15 neurons. The present work draws a relationship between the correlation coefficient and the number of hidden layers & neurons. It also helps to study the effect of hidden layers and neurons on the performance of ANN models. Formulas are derived from the performance of ANN models to calculate the required number of hidden layers and neurons for a particular backpropagation algorithm to achieve a testing performance of more than 0.9. Keywords: Artificial Neural Network, Consistency Limits, Compaction Parameters, Hidden Layers, Backpropagation Algorithms