The calibration of complex-system models with OpenMOLE

Mathieu Leclaire, Romain Reuillon
Géographie-cités, CNRS


Large scale experimentations on models


Methods

Design of experiments on a space of parameters

Calibration

Methods

Optimization

Sensitivity analysis

Naturally parallel algorithms => leverage parallelism.

What does OpenMOLE do ?

It implements exploration algorithms.

It transparently delegates computational loads to massively parallel environments.

Upscaling

Prototype small
Experiment large

A naturally parallel formalism to design experiments.

Embed your model as a black box

C
R
C++
Java
Scala
Scilab
Octave
Python
Netlogo
...

A Netlogo Task in OpenMOLE GUI

Assign execution environments to tasks.

Download: http://www.openmole.org

Chromosome structuring

C++
2 days per simulation
1600 simulations
8.5 years / CPU

Junier et al., CTCF-mediated transcriptional regulation through cell type-specific chromosome organization in the β-globin locus, Nucleic Acids Research, 2012.

SimTRAP project

NetLogo
5 minutes per simulation
100000 simulations
1 year / CPU

PhD thesis of J. Figuel, Modélisation et simulation des comportements piétonniers dans les espaces de transport – Application aux échanges quai / train de voyageurs.

Simpop project

Scala
5 minutes per simulation
360 000 000 simulations
22 years / CPU

Reuillon et al., Algorithmes évolutionnaires sur grille de calcul pour le calibrage de modéles géographiques, proceedings of France Grilles 2012.

1.5s by simulation
100 replications by fitness evaluation: 3min
millions of fitness evaluation

=> several years of computation

Gain > 2000: 5.5 year of computation / day

We have established that the modeling hypothesis are sufficient to reproduce the dynamic.

Reusable / model independant / integrated in OpenMOLE... for instance reused for an epidemic model in NetLogo.

Join us on the users@list.openmole.org

and on openmole.org