Multilayer feedforward networks for transportation mode choice analysis: An analysis and a comparison with random utility models

Usually,. discrete choice analysis as regards a transportation system is based on random utility theory. Recently, a different approach to choice analysis, based on Artificial Neural Network (actually a multilayer feedforward network-MLFFN) models, has been proposed by several researchers. Such modelling approach can address three main demand simulation issues (trip generation, trip distribution and modal split) and has shown good predictive capability. Most of the papers deal with extra-urban (inter-city or intra-regional) trips and they are calibrated on aggregate data, simulating demand flows, results have been compared with regressive models. An alternative and more stimulating approach can be followed up by using disaggregate data, as only one paper does, to simulate single-user choice. The aim of this paper is twofold, first to describe the main step towards the successful application of MLFFNs to support travel demand analysis, and second to show that they can be fruitfully applied to analyse transportation mode choice. A deep analysis has been carried out to address each of the major issues needed to make operational an MLFFN. The proposed approach relies on disaggregated (revealed preferences survey data-sets) data taken from two different case studies. The two case studies focus on medium distance intercity journeys, and allow to investigate mode choice for two different trip purposes. and two different geographical contexts: journey-to-work of commuters within the Italian region of Veneto and journey-to-study of students towards the rural location of the University of Salerno. MLFFNs performances have been compared with the most effective and advanced closed-form Random Utility Models (RUMs) that can be calibrated on the same survey data-sets. Models validation and comparison have been carried out by using indices commonly used in MLFFNs or RUMs applications, and by introducing many others expressly defined to aid results interpretation. (c) 2005 Elsevier Ltd. All rights reserved.