Reuillon R.,Schmitt C., De Aldama R., Mouret J.-B., 2015, « A New Method to Evaluate Simulation Models: The Calibration Profile (CP) Algorithm », JASSS : Journal of Artificial Societies and Social Simulation, Vol. 18, Issue 1, http://jasss.soc.surrey.ac.uk/18/1/12.html
Abstract. Models of social systems generally contain free parameters that cannot be evaluated directly from data. A calibration phase is therefore necessary to assess the capacity of the model to produce the expected dynamics. However, despite the high computational cost of this calibration it doesn’t produce a global picture of the relationship between the parameter space and the behaviour space of the model. The Calibration Profile (CP) algorithm is an innovative method extending the concept of automated calibration processes. It computes a profile that depicts the effect of each single parameter on the model behaviour, independently from the others. A 2-dimensional graph is thus produced exposing the impact of the parameter under study on the capacity of the model to produce expected dynamics. The first part of this paper is devoted to the formal description of the CP algorithm. In the second part,we apply it to an agent based geographical model (SimpopLocal). The analysis of the results brings to light novel insights on the model.
A strong regularity in urban systems has long been identified : the hierarchical distribution of city sizes. Moreover, a closer observation of the evolution of this distribution shows that in the majority of city systems, there is a trend towards a more and more unequal distribution of city sizes. Why does the majority of urban systems show those strong regularities? What are the common growth processes involved? Several dynamic growth models have been proposed but no consensus has yet been reached because of the under-determination of models by those empirical laws. In this presentation we describe a new method of agent-based parsimonious modeling that we think can contribute to the identification of the common urban growth processes. This modeling method is based on intensive model exploration for quantitative evaluation of implemented mechanisms. The exploration tools were first developed for the evaluation of SimpopLocal, a model of the organization of urban systems when cities first emerged. The use of those exploration tools was then generalized into a modeling method that was applied for the first time with the construction of the MARIUS family of models which aims at reproducing the evolution of Soviet urbanisation between 1959 and 1989. Those two examples show how this new modeling method can help the construction of urban theories by helping the evaluation of assumptions made on urban processes.
Clara Schmitt,Sebastien Rey-Coyrehourcq,Romain Reuillon, “Algorithmes évolutionnaires pour le calibrage de modèles géographiques”, October 1st 2012, Journées scientifiques mésocentres et France Grilles. watch online (in French)
The aim of this series of models 1 is to study growth regimes of systems of cities, which are defined by the nature of the interaction between the cities. Three different stages of urbanisation are considered as resource accessibility, which plays a major role in growth dynamics, takes different forms. For each stage, a set of stylized facts characterizing the state of the system and urban growth dynamics is proposed and defines the structure of an ABM model.
The SimpopLocal model tries to characterise a stylised dynamics of urban emergence. This first urban regime is defined by the role of local environmental constraints on growth. Central to this model is the notion of landscape carrying capacity. Settlement sizes and their growth are controlled by the amount of resources locally available. But innovations and their diffusion in the system thanks to interactions between settlements help to overcome those limitations, eventually producing the proto-structure of urban systems.
The SimpopNet model characterise the progressive networking of urban economies. Innovations in communication and transport networks allow the trade and long distance diffusion of goods and techniques which enables cities to overcome, by importing what was lacking, the local constraints and climatic hazards that limited their growth. In this regime, the resource accessibility of each city is defined by its situation in the network. The SimpopNet model simulates the co-evolution of urban systems and transportation networks.
The SimpopClim model will represent a dynamic regime that takes into account the impact of global environmental constraints on urban dynamics.
This PhD is made possible thanks to interdisciplinary work with computer researchers participating in the GeoDiverCity programme, hired after previous collaboration with ISC-PIF (Romain Reuillon and Mathieu Leclaire, http://www.iscpif.fr/) and a PhD in geomatics conducted by Sébastien Rey Coyrehourcq. The SimpopLocal and SimpopNet models are used as case studies for the development of grid exploration procedures and protocols with OpenMole (http://www.openmole.org/). Those explorations and validating tools are designed to meet the validation and reproducibility requirements and to be generic and adapted to the exploration of spatial simulation models.
PhD project of Clara Schmitt, under the supervision of Denise Pumain (founded by ADEME, The French Agency for Environment and Energy) ↩
It’s what it takes to calibrate the SimpopLocal model, that simulates the dynamical hierarchical and spatial organization of settlements at the time when cities emerged, a few thousand years after the emergence of agriculture. Even if this model has been built using a few simple mechanisms, 7 parameters have no known empirical value. To find suitable values for those parameters, an automated calibration algorithm has been designed. In doing so, three quantitative goals have been defined in order to measure the quality of the output of a simulation, hence the quality of a set of parameters. One goal targets the shape of the distribution of the size of settlements, another one the size of the biggest settlement of the distribution and a last one the length of the simulation, or number of iterations, required to achieve those goals. These criteria are evaluated by computing 30 replications (independent execution) of the model (due to its stochasticity). Using the OpenMOLE framework (www.openmole.org) a genetic algorithm (a global optimization algorithm) has been distributed on the European grid EGI, federating computing power all over the world. After running about 10 million model executions, which would take more than 100 years of computation on a bleeding edge computer, the algorithm has finally converged after one week of computation and found suitable sets of parameters for the model calibration. The modelers have validated them and are now taking benefit from the calibrated model to better understand the implications of the mechanisms chosen to simulate a stylized emergence of urbanism.
Romain Reuillon, Sébastien Rey-Coyrehourcqand Clara Schmitt
OpenMOLE has originally been developed in a generic way (and in particular with the cooperation of geographer modelers) to be scientific field independent. That is why dealing with geographical models within the Geodivercity program is straightforward. More specific features for geography will be added during the ERC depending on the needs of the modelers.
A first attempt is to build the Simprocess platform for the multi-level exploration of agent-based models of the Simpop type (PhD by Sébastien Rey-Coyrehourcq).
The next versions will include:
new environments (remote servers through ssh, PBS clusters, cloud),
a standardised serialisation format for workflows,
an integration of cutting edge scientific method for model exploration (optimization, calibration, fitness landscape analysis, sensitivity analysis…),
OpenMOLE (version 0.5) is a piece of software for intensive scientific computing. OpenMOLE is the result of 4 years of daily work with model exploration issues in diverse fields (Human Sciences, Physics, Geography, Food-processing,…). It is 100% FOSS (Free and Open Source Software), 100% written in Scala.
It targets modelers who explore their models at wide scales. It makes it possible to generate automatically wide design of experiments (full-factorial, LHS, Saltelli…) and to take advantage of the computing power of massively parallel execution environments such as computing grids and clusters. It exposes a workflow formalism to design parallel processings in a natural way.
Strengths of OpenMOLE:
a zero-deployment approach: the programs (C, C++, Python, Java, Scala, NetLogo, etc) are embedded at runtime and do not require installing software on execution machines,
a small number of base concepts to handle (4):
tasks (the executable components containing the model for instance);
prototypes (types variables) which are transferred from one task to another;
samplings (how to explore my model?);
environments (where my jobs are executed?);
a modular development for extending the platform with plug-ins in short time and based on OSGi,
an optimized and effortless access to grid resources (automatic resource discovery, eager submission, failure handling, management of data transfer …),
a specific task to embed NetLogo models,
a formal workflow validation prior to the execution based on a strongly typed dataflow,
its scalability: it manages up to hundreds of thousands of task executions from a laptop;
a scripted interface as well as an ergonomic graphical user interface.