Improving the Neural Network Testing Performance by Transforming the Normalized Data Nonlinearly for Trip Distribution Modelling

Abstract

Previous studies have suggested that the Artificial Neural Network (NN) trip distribution models were unable to calibrate and generalize work trip numbers with the same level accuracy as the Doubly-Constrained Gravity models (DGCM). This study presents some new NN model forms aimed at overcoming these problems trained by using the Levenberg-Marquardt algorithm. A further modification was applied to the model, namely transforming the input data nonlinearly by using logistic functions (Sigmoid) in order to improve the testing/generalization of NN models. This resulted in better performance of NN models, where the average Root Mean Square Error (RMSE) is statistically lower than the DCGM indicating the NN models could have higher generalization ability than DCGM.