Comparative Analysis of Hidden Layer Neuron Configurations for Prediction Performance in Neural Networks Trained Using Levenberg–Marquardt and Scaled Conjugate Gradient Algorithms

Authors: *Onah, T.O., Aka, C.C. And Onah, B.C.

DOI Info: http://doi.org/10.5281/zenodo.21046185

ABSTRACT

This study investigates the performance characteristics of neural network training using varying hidden layer neuron configurations with paired complementary values (2 and 50, 4 and 40, 6 and 30, 8 and 20), employing the Levenberg-Marquardt (LM) and Scaled Conjugate Gradient (SCG) algorithms. The research aims to determine optimal neuron configurations for minimizing prediction uncertainty and training errors. Experiments were conducted using MATLAB R2019a's Neural Network Fitting Tool with systematic variation of hidden layer neurons from the default value of 10. Results demonstrate that the Levenberg-Marquardt algorithm consistently achieves superior validation performance with lower mean squared error values compared to the Scaled Conjugate Gradient approach. For a hidden neuron configuration of 8, LM achieved a validation error of 0.00012831 at epoch 94, representing the optimal performance among tested configurations. The error histogram analysis reveals that configurations approaching the default target of 10 neurons (specifically 6, 8, and their complements) exhibit tighter error distributions centered on zero. Conversely, extreme neuron counts (2, 50) demonstrated increased error dispersion indicative of underfitting and overfitting, respectively. These findings provide practical guidelines for neural network architecture design in regression and function approximation tasks, with applications extending to medical diagnostics, industrial automation, and predictive systems.

Download Full Text