Population Synthesis Meets Deep Generative Modelling

Agent-based transport models depend to a high degree on the formation of the underlying population. Models and methods for generating such populations, possibly under constraints to reflect future margins, are commonly referred to as “population synthesis” models. Historically, different approaches have been proposed, ranging from deterministic approaches, such as the Iterative Proportional Fitting (IPF) algorithm (Deming and Stephan, 1940; Rich and Mulalic, 2012), to simulation based approaches of the Markov Chain Monte Carlo (MCMC) type (Farooq et al., 2013). Often, matrix fitting methods such as the IPF has been used in combination with postsimulation based methods in order to translate prototypical individuals into true micro agents. While the existing models are capable of producing acceptable results for agents with relative few socioeconomic and spatial characteristics, these methods do not scale well when the dimensionality of the underlying distribution becomes large. As a result, in many cases, these methods are not able to accommodate the increasing need for more dimensions that result from, e.g. smaller zones, the combination of household-based and individual-based synthesis and more detailed variables in general. In this paper, we propose a different approach to population synthesis based on generative models from the deep learning framework. Contrarily to existing methods, these new methods are scalable and can handle a very large number of both numerical and categorical attributes at the same time.