Resumen de Large-scale simulation of population dynamics for socio-demographic analysis
Cristina Montañola Sales
Computer modelling and complex systems simulation have dominated the scientific debate over the last decade, providing important outcomes in biology, geology and life sciences. In the social sciences, the number of research groups currently developing research programs in this direction is increasing. The results are extremely promising since simulation technologies have the potential to become an essential tool in the field. Agent-based modelling is widely recognised as one of the techniques with more potential to develop useful simulations of social interacting systems. The approach allows to specify complex behavioural and cognitive rules at the individual level; and through aggregation, the output at the macro level can be derived. Increasingly, the output of micro-level simulation models is used as an input to policy models. Policy models not only requires detailed micro-level data, but also significant compute power since the number of agents and interactions can be extremely large in some cases. High performance computing offers a massive supercomputing power which allows us to simulate a large artificial society. In that context, parallel simulation could provide an alternative to speed up the execution of such compute-intensive socio-demographic models. It deals with techniques that allow the use of multiple processors to run a single simulation. Although research in parallel simulation has been around for more than two decades, the number of applications in the social sciences is scarce. In this thesis, we present a methodology for simulating population dynamics at a large-scale. Specifically, we developed a parallel simulation framework to run demographic models. It simulates the interactions of individuals in a society so the population projection can be obtained. Two of the main obstacles hindering the use of agent-based simulation in practice are (a) its scalability when the analysis requires large-scale models, and (b) its ease-of-use, especially for users with no programming experience. Our approach proposes a solution for both challenges. On one hand, we give a solution in to simulate large social systems in a parallel environment. We show its potential by studying the performance of our approach by identifying the factors that affect the simulation execution time. Moreover, we investigate the impact of three well-known configurations of computer architecture. Since the application of parallel simulation in demography is new, it is useful to quantify the effect of these factors on performance. On the other hand, we provide a graphical user interface which allows modellers with no programming background to specify agent-based demographic models and transparently run them in parallel. We believe this will help to remove a major barrier on using simulation although we are aware technical knowledge is necessary to execute scenarios in High Performance Computing facilities. Two cases studies are presented to support the feasibility of the approach for the social sciences. The first case under study carries out an analysis of the evolution of the emigrated population of The Gambia between 2001 and 2011, a relevant period for immigrations in Spain. The second case study simulates the socio-demographic changes of South Korean during one hundred years. The objective is to rate the feasibility of our methodology for forecasting individual demographic processes. Our results show that agent-based modelling can be very useful in the study of demography. Furthermore, the use of a parallel environment enables the use of larger scale demographic models
