Hierarchical Population Generation in Transportation Modelling

Abstract

Synthetic populations are necessary for agent-based travel models. However, conventional approaches based upon reweighting data from an available small sample suffer from low heterogeneity; simulation-based methods are promising, but have not so far taken into account hierarchies in the data. The aim of this thesis is to develop a simulation-based approach for population synthesis that allows to consider hierarchies, as in the case of persons and households, and to apply it in a case study. The current state of the art of hierarchical population generation is first reviewed, with a special attention to the theoretical foundation of each technique. An approach belonging to the family of Markov Chain Monte Carlo methods is chosen for implementation, offering several advantages compared to the popular Iterative Proportional Fitting. MCMC has only recently been applied for this purpose in transportation modelling, even though it has a long history in other fields. It is a two-step procedure based upon fitting conditionals and Gibbs sampling, which exploits these conditionals in a Markov chain to simulate new agents. In this thesis, MCMC is both generalised (iMCMC) and extended to handle hierarchies (hMCMC). A careful analysis of the results proves the validity of the developed approach. Besides, some advanced data visualisation and clustering techniques are implemented to study the available demographic sample. Firstly, Multiple Correspondence Analysis, the categorical version of the well-known Principal Component Analysis, which finds a new basis system highlighting the distribution of the data points, rather than simply using the original variables. Secondly, Self-Organising Map, to produce low-dimensional views of the considered high-dimensional data which allow to easily identify the present clusters. Thirdly, Chow-Liu Tree, which approximates the joint probability distribution underlying the data sample by finding the most evident interdependencies among its variables. These methods allow to distinguish the characteristics of the clusters into which the available data sample is divided. It is also illustrated how they can become further lines of research for the problem of population generation, by fitting a model for the joint probability distribution.