Multitask Learning Deep Neural Network to Combine Revealed and Stated Preference Data
It is an enduring question how to combine revealed preference (RP) and stated preference (SP) data to analyze travel behavior. This study presents a new approach of using multitask learning deep neural network (MTLDNN) to combine RP and SP data and incorporate the traditional nest logit approach as a special case. Based on a combined RP and SP survey in Singapore to examine the demand for autonomous vehicles (AV), we designed, estimated and compared one hundred MTLDNN architectures with three major findings. First, the traditional nested logit approach of combining RP and SP can be regarded as a special case of MTLDNN and is only one of a large number of possible MTLDNN architectures, and the nested logit approach imposes the proportional parameter constraint under the MTLDNN framework. Second, out of the 100 MTLDNN models tested, the best one has one shared layer and five domain-specific layers with weak regularization, but the nested logit approach with proportional parameter constraint rivals the best model. Third, the proportional parameter constraint works well in the nested logit model, but is too restrictive for deeper architectures. Overall, this study introduces the MTLDNN model to combine RP and SP data, relates the nested logit approach to the hyperparameter space of MTLDNN, and explores hyperparameter training and architecture design for the joint demand analysis.
READ FULL TEXT