An artificial neural network based method to uncover the value-of-travel-time distribution
- PDF / 3,290,782 Bytes
- 39 Pages / 439.37 x 666.142 pts Page_size
- 15 Downloads / 169 Views
An artificial neural network based method to uncover the value‑of‑travel‑time distribution Sander van Cranenburgh1 · Marco Kouwenhoven1
© The Author(s) 2020
Abstract This study proposes a novel Artificial Neural Network (ANN) based method to derive the Value-of-Travel-Time (VTT) distribution. The strength of this method is that it is possible to uncover the VTT distribution (and its moments) without making assumptions about the shape of the distribution or the error terms, while being able to incorporate covariates and taking the panel nature of stated choice data into account. To assess how well the proposed ANN-based method works in terms of being able to recover the VTT distribution, we first conduct a series of Monte Carlo experiments. After having demonstrated that the method works on Monte Carlo data, we apply the method to data from the 2009 Norwegian VTT study. Finally, we extensively cross-validate our method by comparing it with a series of state-of-the-art discrete choice models and nonparametric methods. Based on the promising results we have obtained, we believe that there is a place for ANN-based methods in future VTT studies. Keywords Artificial neural network · Value of travel time · Random valuation · Nonparametric methods · Discrete choice modelling
Introduction The Value-of-Travel Time (VTT) plays a decisive role in the Cost–Benefit Analyses (CBAs) of transport policies and infrastructure projects as well as in travel demand modelling. The VTT expresses travel time changes in monetary values (Small 2012). Due to its importance for transport policies and appraisal, the VTT is one of the most researched notions in transport economics (Abrantes and Wardman 2011). Most Western societies conduct studies to determine the VTT on a regular basis. The focus of such VTT studies is typically not to obtain a single (mean) VTT for all trips, but rather to obtain tables of VTTs which show how the VTT depends on trip characteristics, such as travel purpose and mode. Despite decades of experience with data collection and VTT inference, the best way to obtain the VTT is still under debate. Early studies predominantly used Revealed Preference (RP) data in combination with Multinomial Logit (MNL) models (Wardman et al. * Sander van Cranenburgh [email protected] 1
Transport and Logistics Group, Delft University of Technology, Delft, The Netherlands
13
Vol.:(0123456789)
Transportation
2016). However, despite the well-known advantages of RP data over Stated Choice (SC) data (Train 2003), nowadays RP data are seldom used in VTT studies. The main reason is that while the travellers’ choices are observable, their actual trade-offs across alternatives are not—which hampers estimation of the VTT using RP data. More recent VTT studies therefore favour using SC data in combination with discrete choice models that account for (some of the) potential artefacts of the SC experiment (notably size and sign effects) (Fosgerau et al. 2007; Ramjerdi et al. 2010; Börjesson and Eliasson 2014; Kouwenhoven et al.
Data Loading...
