A Consistent Neural Network Model for Doubly Constrained Spatial Movement Estimation

抄録

Despite its successful applications in mode choice modelling, the Neural Network (NN) Approach is often seen as a poor technique for trip distribution estimation. However, empirical results from this study show that the NN can also be used as a good trip distribution modelling tool, by using appropriate Training Algorithm (TA), which is the most critical property of NN. This study will describe the results of NN approach in work trip distribution estimation with three different TAs, namely Back Propagation (BP), Variable Learning Rate (VLR), and Lavemberg-Marquardt (LM). The experiments were conducted 30 times for each TA. The results suggest that both BP and VLR models tend to have overestimated total trip numbers, while LM model generates the closest total trip number to the observed one, with less than five per cent difference. Statistical analysis also suggests that LM model is able to generate a consistent performance, while other models have a great fluctuation.