Tuesday, April 12, 2011

Simulating migration

Today I’ve been dipping in and out of the user environments sessions here at the EGI User Forum.

Speaking in the morning session was Jing-Ya You of the Academica Sinica Grid Computing Centre in Taipei, Taiwan. Her talk gave a great example of how distributed computing solutions such as grids and clouds can be used in social science research via her project SimTaiwan.

Social scientists often use a technique called agent based modeling to understand complex social phenomena. By assigning different attributes such as age, gender, health, or socio-economic status to a large number of individual ‘agents’, we can simulate interactions in a dynamic population and changing environment. (N.B. We spoke to Alex Voss, who leads the performance improvement and collaboration with social scientists of SimTaiwan, about agent-based modelling last year. He gives an excellent explanation.)

Flying to a new home
Migration in Taiwan is an important factor in Taiwanese social development and has been so since the 1600s. SimTaiwan therefore aims to test existing theories and investigate recent developments such as increased migration to China. Attributes for the ‘agents’ have been obtained from census data which give a dataset of around 22 million agents with up to 20 attributes each.

SimTaiwan's model is split into two parts - departure and destination. The probability of departure for a single agent is a weighted sum of different factors such as age, sex, education, marital status, population size, and employment at departure location. Weights are provided by a statistical model based on actual migration patterns measured by the census.

Before running the simulation researchers need to test the performance and scale of model, which is where distributed computing comes in. Jing-Ya and her colleagues used a simplified model with only fertility and mortality which they ran for a population of 250k male and 250k female and 365 ‘days’.

By improving the code the team reduced the running time from around 5 to 2 minutes. They also used the grid resource at Academica Sinica to help debug the model, and run stability and sensitivity tests which were used to validate and verify models.

SimTaiwan began in the EUAsiaGrid project and hopes to collaborate with more social scientists in Taiwan in the near future.

Of course Jing-Ya's wasn't the only talk in the User Environment session today. To find out about the rest you can read the overviews from all of the User Environment Sessions in the book of abstracts either online or in paper form at the EGI booth.

No comments: