Geoscience Reference
In-Depth Information
included in the calibration. In the second case, even if the simulation results are
compared with the observed data (calibration and evaluation phases), this step is less
tightly controlled than in the case of the statistics field. The reasoning is nearer the
model of a “candidate for explanation”: the set of rules formalized with the MAS
enables the reproduction of a given organization (sufficient conditions), but this does
not exclude that other sets of rules can also achieve it (conditions not necessarily
necessary).
In the statistical approach, the emphasis is therefore placed on the relation
between the attributes of the objects of interest, and the constraints attached to the
statistical methods imply that these objects are independent [MAC 09]. The agent
models, on the contrary, place the interactions in the center of the method. It is
because of the interactions that the “objects” are interpreted as “agents”. After all,
the objective is indeed to explore the forms and properties that emerge from the
interactions
between the agents.
4.1.2.
Different types of explanation: from the percentage of variance explained to
the generation of mechanisms
The points of view of the explanation, which are derived from these two
approaches, differ [BUL 05, MAN 07]. The statistical explanation relies on
highlighting the covariations of different phenomena. Thus, the term “explanatory”
is excessive; the statistical models do not allow us to put forward causal relations.
The “percentage of variance explained” by a model refers to the ability of the model
to reproduce, from the combination of a set of variables, said share of
differentiations between the objects of interest. In the case of high schools in the
Parisian region, for example, (see Chapter 2, section 2.3.4), 65% of the interschool
differences in terms of grades in the “brevet” (French high school exam) are
“explained” by positioning differences on an axis of social composition.
3
However, the statistical model does not allow us to identify the real causes behind
the difference in results between schools. As Bulle puts it: “The real mechanisms
behave as 'black boxes': the models tend to reproduce the 'inputs' and the 'outputs'
of the real sets being studied without revealing their operation secrets” [BUL 05].
The title of one of Epstein's books [EPS 07] about MAS, “Generative Social
Science”, clearly illustrates the explanation that the author associates with this
formalism: to him, to explain a phenomenon means to simulate it (“If you didn't
grow it, you didn't explain it” [EPS 99, EPS 07]). If the interaction rules
implemented at the agent level generate an organization at a macro level of
3 This axis corresponds to the first factor of a principal component analysis (PCA) performed
on the social compositions of the schools.
