Re: [Swarm Modelling] Re: The "Art" of Modeling

[Jason Alexander]

Hi Chris -

> Jason; could you clarify the following bit for me?  You write...
>
> > The weather scholar seems to be advocating an instrumentalist view 
> > over
> > a realist view. If the ultimate goal is saving lives, then one will 
> > use
> > any "black box" model which offers the best prediction of the future,
> > regardless of whether it helps us "understand" tornados (where
> > "understanding" a tornado means that we have an accurate description 
> > of
> > the general laws, mechanisms, and processes which serve to produce
> > tornados).
> >
> > But we then face the standard problem of instrumentalism: it seems 
> > that
> > the only justification we can give for why a "black box" model should
> > offer accurate predictions is that it employs, in some way, a
> > description of the general laws, mechanisms, and processes which are
> > really at work in the world. If so, then the best way in which to cook
> > up the "black box" model is to just go out and try to identify the 
> > general laws, mechanisms, and processes that really exist, i.e., we 
> > get
> > a call for realism.
>
> Why is that automatically "the best way"?  It would seem that this, 
> too, requires
> an argument.

If a "black box" model can be used to make successful predictions, it 
must base those predictions on a causal relationship holding between 
variables - otherwise there's no way to account for the model's success 
at making predictions.

The examples you cite from Scientific American don't so much refute the 
claim as illustrate the importance of distinguishing between the task 
of (1) identifying general causal laws, often holding at some 
macro-level of description, and the task of (2) identifying general 
causal mechanisms, operating at some micro-level, which serve to 
account for the macro-level laws.  Redox chemical equations (which are 
basically laws of a sort) are great examples of this.  The statement of 
the equation, which can be found by careful experimentation, doesn't 
itself provide any clue on what the appropriate underlying mechanism is 
which accounts for the chemical equation (law).

Take the example of the GP project having created a circuit that 
outpredicts anything designed by humans, but "we don't understand it."  
Since it is a digital circuit, it takes a set of input variables, 
massages the information contained within them somehow, and then 
produces an output.  If the circuit is better than chance at predicting 
future states of affairs, it has found some causal relationships that 
hold between the variables and used this to construct a law.  Now, when 
we say we "don't understand it" there are (at least) two things we 
might not understand:

(1) The actual law encoded by the circuit, relating input variables to 
output variables,

or

(2) The underlying causal mechanism which accounts for the actual law  
encoded by circuit.

If we "don't understand" in the sense of (2), that's no big deal.  When 
Kepler produced his laws enabling him to predict the motions of the 
planets, people didn't understand them in the sense of (2) because they 
didn't have gravitational theory yet.  They were still general laws 
holding for a class of phenomena.

If we don't understand in the sense of (1), then we just haven't 
translated the general law found by the circuit into a form we can 
comprehend.  We can presumably do this -- let Mathematica grind on the 
logic of the circuit and backtrack it into mathematical notation.  
We'll get something we can then look at and comprehend.  But what if 
the expression runs for hundreds of pages of ugly mathematical 
notation?  You might say at this point that we still "don't understand 
it."  But it seems to me that we don't understand it in the sense of 
(2), not (1).

Either way, this method is compatible with realism: we are searching 
for general laws relating variables.  The GP method just searches for 
laws by a different technique than what we are most familiar with.

> Moreover, there are refutations available; for instance, the current
> issue of Scientific American reports on a GP project that has created 
> at least one
> circuit that outperforms (outpredicts) anything designed by 
> humans--and that the
> humans as yet don't understand it.  Another article talks about using 
> data mining
> to target pharmaceutical research in directions that are most likely 
> to yield
> high-return medications.  The results have been tremendous cost 
> savings and the
> improved targeting of research--but the researchers don't necessarily 
> understand why
> their new targets are "better" than their old ones.  Aren't either one 
> of those
> applications better black boxes than we could presently build using 
> the realist approach?
--
Dr. J. McKenzie Alexander
Department of Philosophy, Logic and Scientific Method
London School of Economics and Political Science
Houghton Street, London WC2A 2AE