Using Generative Adversarial Networks to Assist Synthetic Population Creation for Simulations

Synthetic populations are heavily used in agent-based simulations and microsimulations to create realistic representations of real-world populations. Many existing techniques rely on duplicating or selecting a sample of disaggregated records captured via surveys to generate the entire synthetic population. The challenge here is the potential bias present in the sample of disaggregated records. This paper posits that such disaggregated records can be improved or replaced by training a generative adversarial network (GAN). We present a case study of a 1.1 million population using iterative proportional fitting (IPF). We illustrate that IPF makes a better fit using GAN-based disaggregated records rather than original census-based disaggregated records. Our results show a promising use of GANs for synthetic population generation.