Estimating Route Choice Models from Stochastically Generated Choice Sets on Large-Scale Networks

Abstract

Route choice is one of the most complex decision-making contexts to represent mathematically, and the most frequently used approach to model route choice consists of generating alternative routes and modeling the preferences of utility-maximizing travelers. The main drawback of this approach is the dependency of the parameter estimates from the choice set generation technique. Bias introduced in model estimation has been corrected only for the random walk algorithm, which has problematic applicability to large-scale networks. This study proposes a correction term for the sampling probability of routes extracted with stochastic route generation. The term is easily applicable to large-scale networks and various environments, given its dependence only on a random number generator and the Dijkstra shortest path algorithm. The implementation for revealed preferences data, which consist of actual route choices collected in Cagliari, Italy, shows the feasibility of generating routes stochastically in a high-resolution network and calculating the correction factor. The model estimation with and without correction illustrates how the correction not only improves the goodness of fit but also turns illogical signs for parameter estimates to logical signs.

Using path decomposition enumeration to enhance route choice models

“…It is argued that the Path Size attribute needs to be evaluated over the set of all paths, which is infeasible, and the paper determines a minimum set required to obtain unbiased results. The technique was applied to a synthetic network and for real networks it is reported by Rieser-Schüssler et al [8] and Vacca et al [9] to be slow. Rieser-Schüssler et al [8] also reports it to be very sensitive to selected input parameters that severely affect the distribution for the lengths ratio actual/shortest of the generated routes.…”

Timetable-based simulation method for choice set generation in large-scale public transport networks

“…Recent literature has tackled the issue of consistent estimates after instances of importance sampling (e.g., Frejinger et al 2009;Flötteröd and Bierlaire 2013;Ben-Akiva 2013a, 2013b), but limitations in the implementation to large-scale networks emerge when considering the following: (i) the random walk (Frejinger et al 2009) has convergence problems when tested on large networks with two-way links and the original application was on a small network with one-way links obviating loop formation; (ii) the Metropolis-Hastings algorithm (Flötteröd and Bierlaire 2013) has computational requirements as shown by its application that also lacks comparison with observed routes; (iii) the sampling and the importance sampling correction proposed by Ben-Akiva (2013a, 2013b) performed excellently with Monte Carlo simulation, but was tested for a real data set with an extremely low number of observations and consequently rather large standard errors that facilitated a positive comparison for very large samples. Although a couple of recent studies have succeeded in estimating models while correcting for importance sampling in medium-size networks (Mai et al 2015a;Vacca et al 2015), recent literature has obviated the problem by introducing recursive models that can be consistently estimated for logit-type choice probabilities (Fosgerau et al 2013;Mai et al 2015b). These models rely on a dynamic specification of link choices and hence avoid a priori choice set generation and are consistently estimated or used for prediction in a computationally efficient way.…”