RGS | London

August 2016

Back to the Future

of Multimodelling

Sébastien Rey-Coyrehourcq | IDEES, Rouen

Romain Reuillon | ISCPIF, Paris

Clémentine Cottineau | UCL, London

slides @ http://goo.gl/gpJTtd

Why multimodelling ?

Validation Dragon !

[ Rey 2015, p11 - 184 ]

1960's - 2016

back to origin ?

Equifinality & self-organisation

[Bertalanffy 1950]

[Ashby 1947]

[Prigogine 1969]

[Foerster 1960]

An Observational Dilemma 

The Nature of Simulation

[Batty 1976]

[Morgan 2005]

[Phan 2008, 2010]

[Winsberg 2009]

[Guala 2008]

[Varenne 2001, 2013]

[Bulle 2005]

Early vision of MM

"Cluster of models" example applied to "ecological community"

Alternative (linked, nested, completary, concurrent, etc. ) models

Multiple Goals of Modelling, Validation != "True"

(see also Hermann 1967 !!)  

generality vs realism vs precision

The strategy of model building in biology, Lewins R., 1966

Interdisciplinarity

Limited ressources

CDC 6000 "super"computers

Washington University IBM 650
& Northwestern's Vogelback computing center CDC 3400/ 6400

ask Swedish & American Simulation pioneers in 60's ... 

[ex Hager III and Hager IV Fortran program]

3 MFlops ...

Openshaw 80's vision

“From data crunching to model crunching: the dawn of a new era”, 1983

The dream is of some kind of model-crunching machine which could be persuaded to search for interesting model specifications in the universe of all possible models relevant to a particular purpose

Openshaw 80's vision

“Building an automated modelling system (AMS) to explore a universe of spatial interaction models”, 1988

7 variables
5 parameterization
5 operators
15 functions

Blueprint to generate spatial interaction models

( up to 1 to 5 linked modules)

into module

1

1

1

2

combine

of

alternatives

Openshaw 80's vision

Models generated hard to interpret

“Building an automated modelling system (AMS) to explore a universe of spatial interaction models”, 1988

Openshaw 80's vision

Cray 1 vector supercomputer 1976

- 160 MFlops -

Random

vs

meta-heuristics

High Performance Computing (HPC)

“Building an automated modelling system (AMS) to explore a universe of spatial interaction models”, 1988

MM method

need

- fastest with less ressources

- find optimal / suboptimal

- explore diversity, uncertainty

better

Back to ecology

Robustness analysis: Deconstructing computational models for ecological theory and applications, Grimm et al. 2016

Complementory with Richard Levins' idea !

Know where and when

parameters and mechanisms break

base model

base model

choices

give different path (1, 2) to explain same pattern

(   )

(   )

Look back and challenge choices

Share your results

and challenge others

(   )

Cumulative knowledge

Back to the future?

Ecology

Geography

2016

2036

1996

1976

1956

Levins

Hermann

Grimm al.

Openshaw

GeoDiver

City

Massive parrallelism era

?

Batty, Torrens

Theory and conceptual framework

Application

Vector processor era

HPC

Operational MM

"This project aims at tackling equifinality in systems dynamics by confronting different mechanisms with similar evaluation criteria [...] and by scanning automatically the parameter space along with the space of model structures (as combinations of modelled dynamics)."

Application

Application

What makes cities grow differentially?

Demographics

Local Resources

Economic Specialization

Politics

a. size effects

b. gen. trend

a. coal and ore

b. oil and gas

c. access to sea

a. State level

b. region level

Example:

USSR 1959-1989

USSR 1989-2010

Experiment Design

Generate models

Confront models (structure + parameters) to evaluation target

Step 1

Step 2

Functional programming paradigm

and mixin methods (Scala)

How to Implement MultiModels?

trait T11

parameter 1 for trait T11

parameters for trait T22

ex: ressource extraction

ex: oil & gas multiplier

ex: degree of rural migration

trait T22

ex: urban transition

Model id

= composition

of mechanisms

What comes out of the calibration of MultiModels?

Model id

= composition

of mechanisms

Parameters

= calibrated values

given composition

Fitness

(for multiple

evaluation targets)

Best 3,200 instances

of 36,000,000 evaluated by GA

for each of the two simulation periods

1

2

...

3200

HPC Needed !!!

~ 1 week of computation

on a grid of 3000 cores

How to Analyse MultiModels ?

Interactive Visualisations & Explorations

shiny.parisgeo.cnrs.fr/VARIUS

A few answers

Are all mechanisms necessary?

Model id

= composition

of mechanisms

Fitness

(for multiple

evaluation targets)

~

A few answers

What are the parameter values of the 'best' parsimonious model?

Given

Fitness

(for multiple

evaluation targets)

~

Parameters

= calibrated values

given composition

Model id

= composition

of mechanisms

1959-1989

1989-2010

A few answers

1959-1989

1989-2010

What is the 'best' model leaving out?

Significant attribute

Non-significant attribute

Large City

Non-Capital

Western

Profile of Cities w

Underestimated

Growth

Western City

Large City

Profile of Cities w

Underestimated

Growth

Perspectives

Build a library of mechanisms

Build a library of mathematical and logical functions

"There is always room for doubt as to whether a result depends on the essentials of a model or on the details of the simplifying assumptions." Levins 1966: 423

Combine and accumulate urban models at different scales

to facilitate the use of tested models for different purposes

"Ultimately, we might assemble libraries of tested submodels representing specific behaviours or processes." Thiele & Grimm 2015

The future has arrived – it’s just not evenly distributed yet."

 

William Gibson

@reyman64

@ClementineCttn

@romainreuillon

GitHub

https://github.com/openmole/family

slides @ http://goo.gl/gpJTtd

bibliography @ http://goo.gl/gRfGvG