Integrating different data sources to generate synthetic agent populations

The initial population of an agent based simulation (ABS) represents the current state of the underlying real system. Thus obtaining a reasonably detailed population is important for developing a reliable simulation. Population synthesis research explores generating such agent populations. The work presented here falls under the category of sample-free population synthesis methods. Such methods, however, are usually application-specific due to the inclusion of heuristics in the population generation algorithm. The work presented in this thesis has the vision of developing application independent sample-free population synthesis methodologies. This thesis presents two sample-free population synthesis methods. The first generates a population with complex multi-family household structures using the Australian census data by incorporating known population heuristics. The second approach consists of two parts: a population agnostic framework to input heuristics and an algorithm to generate populations based on the proposed framework-constructs. It decouples heuristics from the algorithm, which allows generating different populations by only changing input heuristics, without altering the algorithm logic. The results show that the generated populations are highly consistent with the known distributions of the actual populations. The thesis also proposes a framework for synchronising agent populations in different ABSs, in the context of developing new ABSs by integrating existing ones. The work proposes that obtaining the initial agent population of an integrated ABS is analogous to population synthesis, thus can be achieved using the proposed population agnostic algorithm. Further, we present an evaluation of a set of widely used statistical tests under the specific conditions of population synthesis research. The results suggest using a combination of the Freeman-Tukey's goodness of fit test [Freeman and Tukey 1950] and the standardised absolute error [Voas and Williamson 2001] for analytical evaluations of generated populations.