Geodemographic modelling collaboration Alex Voss, Andy Turner Presentation to Academia Sinica Centre for Survey Research

Overview Background MoSeS –Population Reconstruction –Dynamic Simulation Summary of on-going projects/effort Future Work Next Steps Acknowledgements

Background Social science does not traditionally use advanced ICTs but emergence of new analytical methods is driven by: –Increased availability of data about social phenomena –Issues with data management and integration –Challenges to analyse social phenomena at scale –Challenges to inform practical policy and decision making (e.g., evidence-based policy making) National Centre for e-Social Science (NCeSS) in the UK is investigating ways to respond to these challenges. EUAsiaGrid is supporting e-Social Science amongst other application domains

MoSeS MoSeS develops modeling and simulation approaches for social science –First phase research node of NCeSS, now continued through second round node GENeSIS Contemporary demographic modeling of the UK based on UK census data and other datasets Using agent-based simulation to project population forward in time by 25 years Simulate the impact of distinct demographic processes such as mortality, fertility, health status, household formation, migration E.g., to inform policy making – what impact do policy decisions have

Population Reconstruction Generation of an individual level population data for the UK –Based on 2001 census data –Works with ‘public release’ versions of census that are restricted, Census Aggregate Statistics at Output Area Level 1% of population (anonymisation) –Reconstructed data has same attributes as real population and same number of individuals but is still anonymised –Uses a genetic algorithm to select a well fitting set of sample of anonymised records to assign to an output area –Need for attributes in the SAR to be matched with those in the CAS This is often complicated because of different categories –Aggregation to a lowest common categorisation

Population Reconstruction (II) Output Generation Raw Data Population Reconstruction Data Output Simulation Synthetic Population Disclosure Control Validation* 

Population Reconstruction (III) If  was 0, we would have a dataset that happens to match the raw census We are looking for a synthetic population with a ‘reasonable’  How can we check if  is ‘reasonable’?

Input Data Example: Individual-level Samples of Anonymised Records (ISARs) Downloadable for UK academics from Click-through End-User License Disclosure control through: –Selection –Aggregation –Permutation No additional output checking required

Output Variables Demographics Employment Health status Household composition Two types of constraints: –Control constraints (have to be met) –Optimisation constraints (used in fitness function)

Control Constraint

Optimisation constraint

Running Pop. Reconstruction Can use geographic segmentation to compute each output area in parallel Good example of high-throughput computing using master-slave architecture Each output area takes about 5 minutes to compute (on average) UK contains about 223,060 output areas, so a complete run would take about 775 days to complete Different control constraints lead to different runtime behaviour for each output area This can be reduced by applying, e.g., 128 processors, reducing the runtime to 6 days More than likely to have errors in such a large compute job, need robust error-recovery Realistically, it takes 2 weeks to compute a complete output

Application Porting Step 0 Code runs through PBS UK Census Data Local Compute Resource Step 1 Code on NGS UK Census Data Any NGS resource Step 2 Code on NGS UK Census Data Run through p-Grade portal Step 2 Code on EGEE UK Census Data GridPP Resources (Step 2.5) Code on EGEE UK Census Data ASGC Resource Step 3 Code on EGEE Using TW Census Data ASGC Resource Currently porting population reconstruction code to EGEE, investigating TW data assets and exploring other links, e.g., with healthcare research Setting up e-Social Science VO

Demo…

p-Grade Portal

Experiences Integrating existing code into grid environment required some changes to source code –management of input arguments –code scalability –log management –error handling Finding the right input size and parameters for testing to keep execution times low Making sense of execution failures –lack of ways to debug code in distributed environments

Experiences II Step-wise process works well, –ensures we encounter problems piece by piece –allows us to comply with data protection / licensing Population reconstruction is resource intensive –may run up against limits on wall clock time Importance of ‘at elbow’ support –but hindered by data protection/licensing issues Licensing means we need to limit execution to UK resources No e-Social Science VO for EGEE porting (yet) –Needed to get support from other VOs in the meantime

Dynamic Modelling Daily activity modelling –Commuting –Retail modelling –Transportation Population Forecasting –Annual time step –Birth –Death –Migration

Future Work Next steps until code runs in Taiwan with Taiwanese data –Proof of concept execution on Quanta cluster at ASGC –Definition of data outputs from Modularising computation so it can exploit multiple NGS nodes or EGEE CEs Improving data and code staging Moving from population reconstruction to supporting the simulation process Integration into ‘science gateway’ for the social sciences and developing a repository for models

Acknowledgements National Centre for e-Social Science –MoSeS Node: Mark Birkin (PI) –GENeSIS Node: Mike Batty (PI) –NCeSS Hub: Peter Halfpenny and Rob Procter EUAsiaGrid Consortium –Marco Paganoni (Project Director) CPC at Westminster University –Gabor Szmetanko –Gabor Terstyanszky –Tamas Kiss GridPP –Jens Jensen and Jeremy Coles National Grid Service –Jason Lander and Shiv Kaushal (Leeds), Steven Young (Oxford), Mike Jones (Manchester)