Improving artificial neural network performance in calibrating doubly-constrained work trip distribution by using a simple data normalization and linear activation function

The applications of artificial neural networks (NN) in estimating work trip distribution is a unique case study as its performance is not only measured in term of the error level in the estimated trips, but also by its ability to satisfy the production and attraction constraints. Previous research indicated that NN models were unable to fulfil those constraints and had rather poor generalization ability. However, this study has indicated that a NN with simple data normalization and a linear activation function (Purelin) in the output layer could accomplish the two constraints, with average correlation coefficients (r) of 0.958 and 0.997 for production and attraction respectively. The test results have also provided evidence that a validated NN could provide a similar goodness of fit as a doubly-constrained gravity model. However, the error level is still higher than the gravity model as indicated by the average root mean square error (RMSE), where the RMSE for the NN and gravity model are 181 and 174 respectively. Finally, the study suggests that the NN can be used to calibrate doubly-constrained trip distribution matrices; although, further study and refinement is required to improve the model's performance.