Empirical determination of optimal configuration for characteristics of a multilayer perceptron neural network in nonlinear regression

Abstract

In this paper, we determine an optimal configuration for characteristics of a multilayer perceptron neural network (MPL) in nonlinear  regression for predicting crop yield. Monte Carlo simulation approach has been used to train several databases generated by varying the internal structure of 3-MLP from simple to complex for 5 different algorithms most commonly used. Results showed that the optimal configuration is obtained with the Levenberg Marquard algorithm, 75% of the number of input variables as number of hidden nodes, learning rate 40%, minimum sample size 150, tangent hyperbolic and exponential functions in the hidden and output layers respectively. This configuration has been illustrated with real life data.

Key words: artificial neural network; machine learning; sample-size effect; nonlinear models; prediction