A decomposition method for estimating recursive logit based route choice models

Fosgerau et al. (2013) recently proposed the recursive logit (RL) model for route choice problems, that can be consistently estimated and easily used for prediction without any sampling of choice sets. Its estimation however requires solving many large-scale systems of linear equations, which can be...

Full description

Saved in:
Bibliographic Details
Main Authors: MAI, Tien, BASTIN, Fabian, FREJINGER, Emma
Format: text
Language:English
Published: Institutional Knowledge at Singapore Management University 2016
Subjects:
Online Access:https://ink.library.smu.edu.sg/sis_research/5290
https://ink.library.smu.edu.sg/context/sis_research/article/6293/viewcontent/cirrelt_2015_66.pdf
Tags: Add Tag
No Tags, Be the first to tag this record!
Institution: Singapore Management University
Language: English
Description
Summary:Fosgerau et al. (2013) recently proposed the recursive logit (RL) model for route choice problems, that can be consistently estimated and easily used for prediction without any sampling of choice sets. Its estimation however requires solving many large-scale systems of linear equations, which can be computationally costly for real data sets. We design a decomposition (DeC) method in order to reduce the number of linear systems to be solved, opening the possibility to estimate more complex RL based models, for instance mixed RL models. We test the performance of the DeC method by estimating the RL model on two networks of more than 7000 and 40,000 links, and we show that the DeC method significantly reduces the estimation time. We also use the DeC method to estimate two mixed RL specifications, one using random coefficients and one incorporating error components associated with subnetworks (Frejinger and Bierlaire 2007). The models are estimated on a real network and a cross-validation study is performed. The results suggest that the mixed RL models can be estimated in a reasonable time with the DeC method. These models yield sensible parameter estimates and the in-sample and out-of sample fits are significantly better than the RL model.