Generalized Multivariate Extreme Value Models for Explicit Route Choice Sets

TL;DR

This paper introduces a unified framework for multivariate extreme value route choice models, identifying new variants and decision rules, and demonstrates their improved performance over traditional models through empirical validation.

Contribution

It generalizes existing GMEV models, introduces new multiplicative variants and decision rules, and empirically validates their superior performance in route choice modeling.

Findings

New multiplicative models outperform additive ones in tests.

Variance of utility should differ from traditional assumptions.

Four new GMEV models based on route differences.

Abstract

This paper analyses a class of route choice models with closed-form probability expressions, namely, Generalized Multivariate Extreme Value (GMEV) models. A large group of these models emerge from different utility formulas that combine systematic utility and random error terms. Twelve models are captured in a single discrete choice framework. The additive utility formula leads to the known logit family, being multinomial, path-size, paired combinatorial and link-nested. For the multiplicative formulation only the multinomial and path-size weibit models have been identified; this study also identifies the paired combinatorial and link-nested variations, and generalizes the path-size variant. Furthermore, a new traveller's decision rule based on the multiplicative utility formula with a reference route is presented. Here the traveller chooses exclusively based on the differences between routes. This leads to four new GMEV models. We assess the models qualitatively based on a generic structure of route utility with random foreseen travel times, for which we empirically identify that the variance of utility should be different from thus far assumed for multinomial probit and logit-kernel models. The expected travellers' behaviour and model-behaviour under simple network changes are analysed. Furthermore, all models are estimated and validated on an illustrative network example with long distance and short distance origin-destination pairs. The new multiplicative models based on differences outperform the additive models in both tests.