Discrete choice modeling is a fundamental part of travel demand forecasting. To date, this field has been dominated by parametric approaches (e.g., logit models), but non-parametric approaches such as artificial neural networks (ANNs) possess much potential since choice problems can be assimilated to pattern recognition problems. In particular, ANN models are easily applicable with their higher capability to identify nonlinear relationships between inputs and designated outputs to predict choice behaviors. This article investigates the capability of four types of ANN model and compares their prediction performance with a conventional multinomial logit model (MNL) for mode choice problems. The four ANNs are: backpropagation neural networks (BPNNs), radial basis function networks (RBFNs), probabilistic neural networks (PNNs), and clustered probabilistic neural networks (CPNNs). To compare the modeling techniques, we present the algorithmic differences of each ANN technique, and we assess their prediction accuracy with a 10-fold cross-validation method. Furthermore, we assess the contribution of explanatory variables by conducting sensitivity analyses on significant variables. The results show that ANN models outperform MNL, with prediction accuracies around 80% compared with 70% for MNL. Moreover, PNN performs best out of all ANNs, especially to predict underrepresented modes.
MAPPING OF ACCESS MODE CHOICE PROBABILITY BASED ON MULTINOMIAL LOGIT MODEL TO GET THE COMPETITION DESCRIPTION OF AIRPORT ACCESS MODES
