The changing roles of parsimony: Understanding, interpreting and explaining geosimulations via Massively Computer-Aided Modeling-Process

Résumé En

Recent works dedicated to tackling the difficulties of agent-based models (ABM) calibration andassessment have more and more engaged in what could be called a Massively Computer-AidedModeling-Process (MaCAMP): (Schmitt et al., 2015), (Reuillon et al., 2015), (Cottineau et al., 2015).This is a computer-aided modeling process which - thanks to platforms like OPENMole - massivelyuses computations operated via grids to “allow a global exploration of the capabilities” of a givenABM in geosimulation (Schmitt et al., 2015). Thanks to this process, almost every free parameterscombination of values in the different mechanisms hypothesized in the model is tested in its outcomeagainst the intended output. The result is that this massive, computer controlled and systematiccalibration process is no more exposed to the risk of not being a real optimum by being a local oneonly. Indeed, this risk is high when free parameters are numerous, not easily interpretable and whentheir estimation relies on incomplete because partially human controlled trial-and-error processes(Schmitt et al., 2015).In front of this trend towards calibration through massive computations, a question arises: what’sthe epistemic role of the models’ parsimony if there remains any? Why not get rid of this apparentlyout-of-age limitation? The surprising fact is that the research works developing this approach stillinvoke the extreme importance of simplicity, parsimony and controlled complexification for theirmodel building, even if their models finally are complex. Surprisingly enough, parsimony still hasan epistemic value in the context of MaCAMP. But what is it? In this talk, I will defend three claims:1. The use of parsimony is still there but it is not exactly the same as it was in non massivelycomputer aided modeling processes. As a consequence, it appears that the epistemic values ofparsimony in geosimulation are diverse and changing; 2. These different values of parsimony canbe related to its different use at different levels - or for different aspects - of each of the differentmechanisms represented in the model. Sometimes parsimony is sought for assuring genericity,sometimes for improving understanding, sometimes for assuring the strict incremental nature ofmodel complexification, sometimes for enabling interpretation and sometimes for establishing theexplaining power of some mechanism of the model. 3. Nevertheless, one can discern a generaltrend: in the case of MaCAMP, parsimony is less sought for global understanding of the model orvia the model that it is for the interpretation of mechanisms or for the establishing of some partialexplanation of the target system’s behavior via mechanisms.In a first section dedicated to some definitions, I will assume the distinctive meanings of“understanding” and “explanation” that are most frequently used: “By ‘explanation’, I mean theintelligible representation (i.e. by concepts) of a system of interactions or a mechanism (elements+ actions) that are assumed to be the cause of a phenomenon […] By ‘comprehension’ or‘understanding’, I mean a unifying conceptual representation that can be mobilized by an unassistedhuman mind. We understand a phenomenon that is composed of a variety of sub-phenomena whenwe can, by means of a single mental (mathematical or logical) operation, reconstruct the gist of thestructure of that variety” (Varenne, 2018, p. 165). By “interpretation” of a model or of some partof it, I mean the opinion of what it means or of what it refers to. Interpretation seems necessaryfor explanation and understanding. But the reverse does not hold. In the second section, relyingon an analysis of some seminal passages of the System of Logic by Stuart Mill, I will recall thetraditional reasons why parsimony is authorized and sought for not only in natural sciences but alsoin theoretical and quantitative social sciences. From this viewpoint, the relevance of a parsimonioustheory was related to our desire to both understand and explain. In the third section, taking theexamples of Hägerstrand’s theory and models of diffusion (Hägerstrand, 1967) and of Pumain’sevolutionary theory of cities (Pumain, 1997), I will emphasize the difference between theoriesconceived as sets of principles and theories conceived as sets of hypotheses and mechanismssuch as the ones implemented by geosimulation. The latter may gather different mechanisms thataffect different entities as much as different aspects of the same entities. When tackling thesetheories via MaCAMP, the search for parsimony could be reduced to the search for a “minimal set ofmechanisms”. But, in fact, as the last section will show through an analysis of the quoted papers,many trade-offs between different - and sometimes contradictory - needs of parsimony appear tobe necessary if one wants to assure interpretability and/or explainability at different levels of themodels and during its conception. Most of the time, an overall understanding has to be sacrificedbut to the benefit of a distributed explanation.