A Joint Analysis of Route Choice Set Generation Methods and Route Choice Models on the Basis of Real Data



A Joint Analysis of Route Choice Set Generation Methods and Route Choice Models on the Basis of Real Data

Authors

F Simonelli, V Marzano, A Papola, Università di Napoli ?Federico II?, IT

Description

This paper proposes the results of a research project focused on route choice simulation, both comparing the performances of route choice set generation methods and estimating different route choice models.

Abstract

This paper proposes the results of a research project focused on route choice simulation, both comparing the performances of route choice set generation methods and estimating different route choice models, on the basis of real data collected within the metropolitan area of Napoli (Italy).
In more detail, the context can be classified as urban, but with a significant number of motorway connections between the downtown and the suburbs. The network has been implemented on the basis of a TeleAtlas network database. Network characteristics have been also compared with data already available, therefore providing reliability on directly measured link characteristics. Data are being collected through a computer-aided interview process, based on interactive maps and on an automatic procedure for data collections and storage.
With reference to route choice set generation methods, the most widely adopted procedures follow a deterministic approach, i.e. they define exogenously a route choice set on the basis of reasonable heuristic rules. Their reliability is normally measured through some indicators based on the calculation of the percentage of the observed paths classifiable as ?reproduced? by the generated paths, according to a certain threshold: the most relevant are the percentage coverage and the consistency index indicators (Prato et al. (2005)). Such indicators do not take into account how many additional links and paths not belonging to the set of the observed paths are generated to reach a satisfactory coverage. Since this kind of bias can lead to significant errors in traffic assignment and model estimation, the same coverage indicators above defined have been considered in this paper also with reference to the set of generated paths towards the set of observed paths, i.e. checking directly whether a generated path is significantly different (e.g. longer) with respect to the observed paths it is trying to reproduce. Moreover, the two types of indicators have also been combined together in order to provide aggregate performance measures. The main result is that exploring a significant part of the network (i.e. generating not much overlapped paths) normally leads to a higher coverage, but a higher number of links not belonging to any observed path are introduced at the same time. Therefore, the quality of a route choice set generation method should be measured on the basis of its robustness with respect both to coverage and bias: from this standpoint, the best results have been provided by the randomization method, which is also the less computational effort demanding. Analogous analyses have also been carried out with reference to a disaggregation by o-d pair. The results are also commented and reviewed in the light of recent analogous works (Prato et al. (2007) and Frejinger and Bierlaire (2007)).
With reference to the route choice model estimation, two different aspects have been investigated and addressed. Firstly, further evidence and/or differences with respect to the results provided by Frejinger and Bierlaire (2006) and Prato et al. (2007) have been analyzed. In more detail, some aspects have been investigated: (a) the influence of choice set generation on route choice estimation, i.e. which choice set generation methods offer higher reliability and stability in model estimation; (b) whether MNL models with path-size and commonality factor terms provide for a better goodness of fit with respect to more complex model structures; (c) whether a positive sign for the path-size coefficient is obtained also in this dataset, providing evidence for the presence of a ?diversion effect? behaviour. Secondly, in order to take into account the previously mentioned bias in the route generation phase, a different specification of the likelihood function to be minimized is proposed for the route choice estimation context. Consistently, theoretical aspects are discussed and the corresponding estimation results compared with those provided by the standard likelihood function.

Publisher

Association for European Transport