Title: Tuning GENIE Earth System Model Components using a Grid Enabled Data Management System
1Tuning GENIE Earth System Model Components using
a Grid Enabled Data Management System
- Andrew Price, Gang Xue, Andrew Yool, Dan Lunt,
Tim Lenton, - Jasmin Wason, Graeme Pound, Simon Cox and the
GENIE team. - http//www.genie.ac.uk/
- UK e-Science All Hands Meeting
- 3rd September 2004
2Outline
- Introduction
- Scientific aims of GENIE
- e-Science tools
- Data Management System
- Geodise Toolboxes
- OPTIONS Design Search and Optimisation
- Results
- Future work
- Conclusions
3Introduction
- The GENIE project is developing a Grid-based
system to - Flexibly couple together state-of-the-art
components to form a unified Earth system model - Execute the resulting model on the Grid
- Share the distributed data produced in
simulations - Provide high-level open access to the system,
creating and supporting virtual organisations of
Earth system modellers
4Scientific Aims
- Orbital parameters affect incident radiation and
climate - Biological and geological processes interact
with, and feedback upon, the climate (via, for
instance, CO2)
5The target GENIE Model
6Initial GENIE experiments
- Initial studies in GENIE performed parameter
sweeps to investigate the properties of the model
7e-Science Tools
- Data Management System (augmented version of the
Geodise Database System) - Matlab scripting environment
- Geodise Toolboxes
- XML Toolbox
- OPTIONS Design Search and Optimisation package
- Template and Example scripts
8Data Management System
Client
Grid
Geodise Database Toolbox
Java Client Code
Jython Functions
SOAP
Apache Axis
Matlab Functions
Metadata Database
CoG
GridFTP
XML Schema
9Grid Computation
National Grid Service (GT2)
Oxford
Leeds
RAL
Manchester
Java CoG
Flocked Condor Pools
10 Geodise Toolboxes
11Scripting a Tuning Study
MATLAB
function RMS_Error cgoldstein(params)
optimum fminsearch( _at_cgoldstein, params, )
GENIE Database
gd_query(results)
Grid Resource
CG binary
gd_putfile(CG binary)
config file
gd_putfile(config file)
gd_jobsubmit(RSL)
results file
gd_getfile(results file)
gd_archive(results)
return RMS_Error
12Matlab Optimisation Toolbox
Specify a starting
point parameters
0.5 Perform the
minimisation optimum
fminsearch( _at_cgoldstein_1D, parameters,
optimisation_parameters )
Specify a starting point
parameters 420
5000000 Perform
the minimisation optimu
m fminsearch( _at_cgoldstein_2D, parameters,
optimisation_parameters )
13OPTIONS
- Matlab interface to the Options design
exploration system - http//www.soton.ac.uk/ajk/options/welcome.html
- State of the art design search and optimisation
algorithms - Design of Experiment methods
- Response Surface Modelling
- Over 30 search methods including
- Adaptive Random Search (ADRANS), Powell's Direct
Search (PDS), - Simplex Method (SIMP), Genetic Algorithm (GA),
- Simulated Annealing (SA), Evolutionary
Programming (EP)
14Grid Computation
OptionsMatlab
optjobparallel.m
GENIE Database
objfun.m
objfun_parse.m
National Grid Service (GT2)
Local Resource (GT2)
Oxford
Leeds
RAL
Manchester
15OptionsMatlab
gtgt OptionsInput createOptionsStructure(4.0)
DNULL -777 OLEVEL 2 MAXJOBS 100
NVRS 12 VNAM 'SCLTAU' 'INVDRAG'
'OCNHORZDF' ... LVARS 1.3000 2.0000
2500 ... UVARS 2.1000 4.8000 5700 ...
VARS 1.7000 3.4000 4100 ...
ONAM 'RMSERROR' OMETHD 4.0000 DIRCTN
-1 NITERS 1000 OPTFUN
'cgoldstein_12D' OPTJOB 'optjobparallel'
GA_NPOP 100 gtgt OptionsOutput
OptionsMatlab(OptionsInput)
Available Optimisation Methods 1.1 for OPTIVAR
routine ADRANS 1.2 for OPTIVAR routine DAVID
1.3 for OPTIVAR routine FLETCH 1.4 for OPTIVAR
routine JO 1.5 for OPTIVAR routine PDS 1.6 for
OPTIVAR routine SEEK 1.7 for OPTIVAR routine
SIMPLX 1.8 for OPTIVAR routine APPROX 1.9 for
OPTIVAR routine RANDOM 2.1 for user specified
routine OPTUM1 2.2 for user specified routine
OPTUM2 2.3 for NAG routine E04UCF 2.4 for bit
climbing 2.5 for dynamic hill climbing 2.6 for
population based incremental learning 2.7 for
numerical recipes routines 2.8 for design of
experiment based routines 3.11 for Schwefel
library Fibonacci search 3.12 for Schwefel
library Golden section search 3.13 for Schwefel
library Lagrange interval search 3.2 for
Schwefel library Hooke and Jeeves search 3.3 for
Schwefel library Rosenbrock search 3.41 for
Schwefel library DSCG search 3.42 for Schwefel
library DSCP search 3.5 for Schwefel library
Powell search 3.6 for Schwefel library DFPS
search 3.7 for Schwefel library Simplexsearch
3.8 for Schwefel library Complexsearch 3.91 for
Schwefel library twomembered evolution strategy
3.92 for Schwefel library multimembered
evolution strategy 4 for genetic algorithm
search 5 for simulated annealing 6 for
evolutionary programming 7 for evolution
strategy
16Twin-Test Experiment
Attempt to recover a known state of the model
using a Genetic Algorithm.
Performed 10 generations of a 100 member
population. Then applied a local Simplex search
of the best candidate.
Population too small to find optimal solution
suitable for finding local minima
17Tuning using Observational Data
- Apply the same method but calculate the RMS
error statistic by comparing the model state with
NCEP observational data.
- The lack of a land surface in the model means
tuning cannot match the observational data.
18IGCM3 Atmosphere Model
- The objective function is a weighted sum of the
RMS differences between the model state and NCEP
data. - Winter and Summer averages for a number model
fields.
19IGCM Results
- 25 reduction in error statistic compared to
default parameters - Similar result to a parallel study performed
using the Ensemble Kalman Filter - Model physics insufficient to perfectly match
observational data.
20e-Science Summary
21Conclusions
- Provided the environmental scientist with a
toolset for tuning GENIE models - Scripting environment
- Database repository
- Computational Grid interface
- Suite of generic optimisation algorithms
- A Global minimum can reliably be found in low
dimensional problem space. - For higher dimensional problems, the tools are
appropriate for locating local minima in the
state space.
22The GENIE Team
- Coordinator
- Tim Lenton CEH Edinburgh
- Principal investigator
- Paul Valdes Bristol
- Research Team and Collaborators
- James Annan FRSGC, Japan
- Chris Brockwell UEA Norwich
- David Cameron CEH Edinburgh
- Peter Cox Hadley Centre (UKMO)
- Neil Edwards Bern, Switzerland
- Murtaza Gulamali London e-Science Centre
- Julia Hargreaves FRSGC, Japan
- Phil Harris CEH Wallingford
- Dan Lunt Bristol
- Bob Marsh SOC
- Andrew Price Southampton e-Science Centre
- Andy Ridgwell UBC, Canada
Management Team Melvin Cannell CEH
Edinburgh Trevor Cooper-Chadwick Southampton
e-Sci. Centre Simon Cox Southampton e-Sci.
Centre John Darlington London e-Science
Centre Richard Harding CEH Wallingford Tony
Payne Bristol John Shepherd
SOC Andrew Watson UEA Norwich Thanks
to Steven Newhouse