Table of Contents
HPC and Geographic Research: An overview of the Human System Modelling Consortium
Contents
Faster Computers are stimulating new ways of doing science!
But
Most geographers and social scientists seemingly do not currently understand what HPC can deliver and few make use of it!
This is a GREAT PITY!
If that PC on your desk had access to a HPC that was 5,000 times faster and had 1,000 times more memory what would you do with it?
?
PPT Slide
PPT Slide
This is one of the things we wanted to change!
Geographers have a KEY role to play here in developing a computational approach to social science
PPT Slide
Typical Data Set Sizes are rapidly increasing
The World about us is becoming increasingly DATA RICH but theory poor
we need NEW TECHNOLOGIES so we can start to cope!
We NEED much FASTER and BIGGER supercomputers to help us cope!
BUT
PPT Slide
PPT Slide
PPT Slide
WE have to be able to demonstrate that if we perform 100,000 or several million times more computation that the benefits are worthwhile!
So what are the Problems?
Also..
PPT Slide
PPT Slide
The challenge is to:
Objectives
Problems Encountered
A new philosophy
PPT Slide
PPT Slide
GeoComputation catches on!
What has the HSM Consortium done?
Codes Ported
Preserving the Investment
Why it is worth the effort
So WHAT are these OPPORTUNITIES?
PPT Slide
Improve legacy models and statistical methods
Apply new approaches based on Computational Intelligent Methods
Investigate novel computational technologies
Seven Brief Case Studies that illustrate:
(1) Parallel Spatial Interaction Models
Spatial Interaction
Examples of Spatial Interaction Models (SIMS)
Origin Constrained Model
Why Parallelise it?
Calibration of the SIM
Porting the Spatial Interaction Model
Performance
New Results and Better Science
PPT Slide
(2) Better ways of Parameter Estimation for Spatial Interaction Models
But there are PROBLEMS with conventional non-linear optimisation methods
PPT Slide
PPT Slide
Argh!!!!!!!!!!
If you use more parameters in your simple model then.. the problem becomes even WORSE!
PPT Slide
Most sophisticated statistical modelling methods PROBABLY suffer from similar risks it is just that few realize it!!
Various Solutions
GA has many advantages
The principal disadvantage is that the GA takes about 10,000 times more compute time than a more conventional nonlinear optimiser
BUT its explicitly parallel
HPC to the rescue!
(3) Spatial Location Optimisation Modelling
Examples
A complex combinatorial optimisation problem
PPT Slide
Results
PPT Slide
PPT Slide
(4) New Computationally Intelligent Methods can be used to build better performing models
If you can AFFORD the computation you can dramatically improve model performances
New Spatial Interaction Models
Often a built-in prejudice against computationally derived models BUT not all are Black Boxes
PPT Slide
HPC challenges many established wisdom's
(5) Engineering Geographical Zoning Systems
Zones are commonly used as:
PPT Slide
PPT Slide
Historically there has been little zone design technology
PPT Slide
Zone Design is a subject of current and recent significance but it is only recently that it has become possible to EXPLICITLY DESIGN zoning systems due to lack of digital map data and fast enough HPC
Zone Design can be viewed as a special type of combinatorial optimization problem
PPT Slide
HARD!
Basic algorithms were first produced 20 years ago
New Algorithms
BUT
BUT
Simulated Annealing takes a lot of compute time!
so.. the big question
PPT Slide
is there?
a Cray T3D 512 processor solution to this problem???
?????????????????????????????? ??????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????????
YES
But the problem is that Simulated Annealing is a highly serial method of optimization
AND.. we did not ‘merely’ want a better result that took longer than previously but a good result that could be produced faster than before
Openshaw-Schmidt hybrid genetic simulated annealer
Advantages
Example using 1991 census data for Leeds - Bradford region
Unemployment Leeds and Bradford Wards
Unemployment Leeds and Bradford EDs
Unemployment Equal Population
(6) Searching for Geographical Patterns in Large Databases
There is a VAST and RAPIDLY growing GEOCYBERSPACE of information
Mark 1 Geographical Analysis Machine
Geographical Analysis Machine
GAM Algorithm
GAM needed HPC because..
Monte Carlo Multiple Testing Outer Loop needed
10 years ago GAM was a mixed blessing!
BUT
Results of a recent evaluation exercise published in Alexander and Boyle (1996)
Overall Performance when Detecting Clustering on 50 synthetic data sets
Estimated Positive Sensitivities in Finding CLUSTER locations
Applying GAM to Long Term Limiting Illness data for Northern England
Ward Level LLTI
Bootstrap, Regional
Bootstrap, Regional Teeside
Bootstrap, Regional Tyneside
Random Data
PPT Slide
(7)Building a Geographical Explanation Machine (GEM)
GEM can be run in 4 modes
Insufficient time to describe how GEM works instead we present some results using as pseudo coverages
GEM is computationally intensive
GEM mode=2
GEM Mode =3
Other Applications
PPT Slide
PPT Slide
PPT Slide
|