Analyzing non-linear contributions to predictive performance in a neural network based scheduling model

Abstract

This paper aims to investigate whether increasing non-linear opportunities in a neural network-based scheduling model improves its predictive performance. More specifically, this paper experiments on a trip distribution model that is part of an activity-based scheduling model called Skyline-seqNN from the ongoing thesis A neural network scheduling model. The motivation behind that model's proposed structure is to lay the groundwork for a neural network discrete choice model (DCM) that achieves to model travel demand on a detailed level while also being suitable for experimental analysis.Similar to a four-step model framework in the sequential aspect, the model system from the referenced paper utilizes the three sub-models; trip generation, trip distribution, and mode choice using a utility-maximizing micro-simulation approach. The trip generation model first decides whether, at every 10-minute interval between 05:00 am and 11:00 pm, an individual in the next time step should stay and continue the current activity or take an activity-defined trip. The distribution and mode choice models are used whenever a trip is selected. The trip distribution model decides the trip's destination by evaluating travel times and land use descriptions of each zone. The mode choice model learns the probability distribution of modes given each mode's travel time to the selected destination zone.Tests performed in this paper show how successive non-linear opportunities between input features in the trip distribution model increase its predictive performance. The data used for training and evaluation comes from a travel questionnaire from 2015 performed in Stockholm containing 10819 individuals and days.