Abstract

This study investigates input neuron prediction based on the error minimization outcomes of percentage proportion of hidden neurons using two prominent training algorithms: Levenberg-Marquardt (LM) and Scaled Conjugate Gradient (SCG). The research addresses the critical challenge of uncertainty minimization in neural network predictions by systematically evaluating various training state proportions (80%, 70%, 60%, 50%, 40%, and 30%) and hidden neuron configurations (2, 4, 6, 8, 10, 20, 30, 40, and 50). The validation performance analysis reveals that the Levenberg-Marquardt algorithm achieves optimal results at 80% training proportion with an error value of 8.194×10−7 at epoch 363, significantly outperforming the Scaled Conjugate Gradient approach. Error histogram analysis indicates that 60% training proportion yields the least error distribution for both methods, establishing this as the optimal configuration for error histogram minimization. Furthermore, hidden neuron analysis demonstrates that 8 neurons produce the best validation performance with minimal error of 0.00012831 at epoch 94, while configurations approaching or exceeding 40-50 neurons exhibit underfitting characteristics with dispersed error histograms. The comparative study conclusively establishes the Levenberg-Marquardt algorithm's superiority in input neuron prediction for neural network training, providing valuable guidelines for practitioners in medical diagnostics, industrial automation, and traffic prediction applications.