Evaluation of Alternative Route Choice Models in Large Scale Networks

Evaluation of Alternative Route Choice Models in Large Scale Networks


P Kant, J Morris, E Mein, Omnitrans International, NL; D van Amelsfort, Goudappel Coffeng, NL


The paper evaluates several route choice models on real large scale networks. Further a method is presented to reduce parameter estimation and increase application possibilities for modelling route choice.


In this paper we compare alternative techniques to model route choice within traffic assignment models. We show that the performance of different models is dependent on the size of the choice set and the correlation between routes. We also present a method of setting the scale of utility to reflect differences between link type, quality of information provided to drivers and traffic management measures. This reduces the task of parameter estimation and provides a systematic approach to model the effects of proposed traffic management measures on route choice .

To account explicitly for correlation between routes, a Monte Carlo simulation of the Probit model was used as the reference against which the alternative route choice models were evaluated. Following the approach by Bovy and Bliemer (2008), parameters were estimated for each of the route choice models: Multinomial Logit (Ben-Akiva and Lerman, 1985), Cross-Nested Logit (Vovsha and Bekhor, 1998), Paired Combinatorial Logit (Koppelman and Wen, 2000), Path Size Logit (Ben-Akiva and Ramming, 1998) and C-Logit (Cascetta et al, 1997).

To ensure practical application of the research, the tests were conducted using the national model of the Netherlands ? comprising 4100 zones and over 200.000 links. Routesets of up to 6 routes were generated for a sample of origins and destinations. The results showed that the performance of each model depended on the characteristics of the routeset available between each origin and destination ? in particular, on the size of the choice set and the overlap between alternative routes.

Bovy and Bliemer (2008) showed that unless the same route choice sets were used for calibration and application, the choice probabilities would be biased, making application of route choice models difficult for real-world networks.

We show that introducing a parameter defining the ratio of variance to mean travel cost according to Bovy, Bekhor and Prato (2008), can overcome this problem. Allowing this parameter to vary according to type of road, user or traffic management measures provides the basis for efficient estimation of route choice models accommodating a range of network conditions and traveller behaviour, including the effects of navigation systems and variable message signs.


Association for European Transport