Deep generative models for combined population and job synthesis

Abstract

Population synthesis is an essential step when modelling transport demand. In this paper, we consider a problem of generating population on a very detailed level which includes origins and destinations of the agents among other socio-demographic characteristics. To model this high-dimensional distribution, we propose to use generative approaches based on artificial neural networks from the deep learning area. As a case study based on a large travel survey data from Denmark, we show that the Generative Adversarial Network (GAN) and the Variational AutoEncoder (VAE) are capable of producing synthetic populations which statistical properties are in good agreement with the observed data.