A Generalized Approach to Population Synthesis

Abstract

Agent-based microsimulation model systems for transportation planning operate at the level of individual agents, and allow more detailed simulation and analysis than traditional four-step models. At the core of any agent-based model is the synthetic population, a disaggregate representation of the agents in the study area. While population synthesis of human individuals has been explored broadly, the generation of household populations for the simulation of entire households has been studied only recently in the field of transport planning. This thesis focuses on methods for generating synthetic populations of households that satisfy exogenous constraints at both household and person levels. A novel method, Hierarchical Iterative Proportional Fitting, and two known methods, Iterative Proportional Updating and Entropy Optimization, are presented in an algorithmic framework. The demonstrated equivalence of Entropy Optimization to Generalized Raking, a method from survey statistics, provides access to strong theoretical results and allows application in a broader scope. In particular, for the first time a fast feasibility check for the exogenous constraints is available. A generic cross-platform open-source implementation is presented. Empirical results suggest that Generalized Raking outperforms existing methods: the synthetic population can be devised with less computational effort and better represents the input data. This method is then applied to generate a synthetic population with activity chains for entire Switzerland for 2030. The activity chains are reweighted to match the assumed frequency of the particular activity types. Generalized Raking allowed obtaining a suitable set of activity chains with only little modeling effort. Weighted random sampling without replacement is another important method for generating synthetic populations. A fairly recent efficient algorithm is described, and its open-source implementation with an empirical validation is presented.