Testing and enhancing spatial transferability of artificial neural networks based travel behavior models

Abstract

ABSTRACT Artificial Neural Networks (ANNs) are emerging classes of AI algorithms, and have seen numerousapplications in travel behavior research recently. However, thetransferability of ANN-based travel behavior models is seldom tested. A fewstudies that test transferability, merely use vanilla Feedforward NeuralNetworks. This paper evaluates the spatial transferability of two ANN-basedmodels: first, a Feedforward ANN-basedmode choice model, and next, a Long Short Term Memory (LSTM)-based activitygeneration and activity-timing model, and enhances their transferability usingtransfer learning (TL). Both the modelswere found to exhibit poor transferability in case of naïve transfer. Transferlearning resulted in significant improvements with the TL-enhanced modelsthat utilizeonly 50% of local data achieving results similar to a locallydeveloped model. Further, ANNs performed poorer when compared with nested logit(NL) models during naïve transfer. However, the TL-enhanced ANN-based modelsshowed significant improvement compared to transfer scaling enhanced NL models.