Models of Mechanisms: The Case of the Replicator Dynamics - Mechanism and Causality in Biology and Economics - page 93

The author of this little vignette thus did not solely intend to entertain with the
Gestalt shift, but rather used this shift in order to infer an obviously absurd and
hence satirical conclusion: because the same image represents both a rabbit and
duck, it is suggested, we must conclude that rabbit and duck are indeed most
similar.
Obviously, this inference is absurd for a number of reasons. I want to focus here
on a rather subtle one, namely, that the same image relates to the two objects it
supposedly represents in different ways. When we use the above image as a
representation of a rabbit, we make certain kinds of idealisation. For example, we
idealise the size of the rabbit's mouth and nose, as well as the shape of its ears.
When we use the image as a representation of a duck, however, we make different
idealisations: the back of a duck's head looks different, and it has different
markings on its feathers. Thus, when using the image to represent either the one
or the other, we make different allowances for which part of the image may not be
representationally accurate. The ingenuity of the draughtsman lay in creating one
image that allowed us to make the respective idealisations in such a way that it can
function either as a representation of a rabbit or a duck. By making these different
idealisations, we adapt the image for its respective uses. Although a duck shares
some features with the image, and a rabit also share some features with the image,
these are not the same features. Thus understood, there is little reason to believe in
the similarity of rabbits and ducks because they are representable with the same
image.
The same holds for the RD model. To use a model as a representation, we always
have to make some idealising assumptions. But when interpreting the RD model
biologically, we make idealisations that systematically differ from those we make
when interpreting the RD model socially. Let me list some of these differences.
First, all three learning models require that players in some way identify actions
and strategies - either of their own or possibly of others. If agents could not identify
strategies in this way, they would not be able to link a diagnosis of 'success' with
the choice of a successful strategy. This stands in contrast to the biological model,
where the strategy notion only fulfils a theoretical role: differential reproduction
does not require that the organism identify the strategies.
This additional requirement pushes these learning models beyond a simple
notion of copying. Rather, it involves the ability to attribute goals and intentions.
'Something other than copying is taking place' (Sperber 2000 , p. 171), and this
other factor may have the power to lead the process in directions that mere
copying would not. Yet such factors are idealised away in all of the three learning
models.
Second, unlike the biological model, the learning models make specific
assumptions about the learning rules players employ, at the exclusion of other,
possible rules. In the biological model, if the payoffs are interpreted as fitness, there
is a natural justification for a linear relationship between payoffs and differential
reproduction. Yet in the learning models, specific imitation and reinforcement rules
have to be chosen to arrive at a linear relationship.