10h30 - 12h10
Forecasting Demand in Networks Using Path Choice Data
In many applications, for example, revenue management and traffic simulation, it is important to forecast demand. In this talk we focus on models that allow to forecast users’ path choices in networks. Discrete choice models are often used for this purpose and they specify the probability that a given individual chooses an option among a set of alternatives as a function of attributes. The parameters of the models can be estimated by maximum likelihood using data on observed path choices. An important challenge associated with predicting path choices in networks is the large number path alternatives connecting each node pair.
In this talk we first provide an introduction to discrete choice modeling and maximum likelihood estimation. Second, we present how the path choice problem can be conveniently modeled as a parametric Markov decision process where the transition probabilities are given by a discrete choice model. The resulting model is based on arcs and does not require any sampling of path alternatives. In some cases, the choice probabilities can be computed by solving linear systems which makes the model computationally attractive. We provide illustrations using several different real data sets.
This talk is based on joint work with Fabian Bastin, Mogens Fosgerau, Anders Karlström, Tien Mai and Maëlle Zimmermann.