CAMREC: GAs and NNs are just as bad as optimisation!

From: the Campaign for Real Economics (camrec@mmu.ac.uk)
Date: Fri Feb 26 1999 - 18:08:47 GMT


Date: Fri, 26 Feb 1999 18:08:47 +0000
Message-Id: <199902261808.SAA19892@alpheratz.cpm.aca.mmu.ac.uk>
To: bogus
From: camrec@mmu.ac.uk (the Campaign for Real Economics)
Subject: CAMREC: GAs and NNs are just as bad as optimisation!


Message-Id: <199902261808.SAA19892@alpheratz.cpm.aca.mmu.ac.uk>
To: CAMREC list members <camrec@mmu.ac.uk>
From: the Campaign for Real Economics <camrec@mmu.ac.uk>
Date: Fri, 26 Feb 1999 18:08:47 +0000
Subject: CAMREC: GAs and NNs are just as bad as optimisation!

Dear Gianluca Baldassarre,

Thank you for your supportive post and information about your thesis.
Any steps towards greater grounding in actual studies of behaviour (even
if anecdotal) represents progress.

However, in my view, shifting from constrained optimization of utility
to Genetic Algorithms (GA) or neural networks (NN) does not, of itself,
represent any improvement in modelling real economic processes. The
basic problem is when the technique takes precedence over the evidence.
(Obviously, I do not know whether this is true in your case.)

There are big problems with the descriptive credibility of GAs or NNs as
representations of what real agents do. Simply put, there is
(typically) no reason to suppose that they descriptively correspond to
how the agents actually make their decisions. GAs and NNs are just
another formalism (a JAF?). Shifting from one JAF to another will (in
my view) make little difference to economics. It is rather like
inventing a screwdriver and _then_ looking around for something to do
with it!

What is important about a model is that it corresponds descriptively to
the process being modelled, limited only by *justified* assumptions and
the limitations of our modelling technologies. Then it also needs
validating against the data from the object process.

Why start with a GA? Why not start with a model of decision making or
learning that is well validated itself (e.g. from cognitive science or
marketting practitioners)?

The only advantage to introducing GAs and NNs is that it marginally
increases the variety of the stock of JAFs that we have to hand to help
us in our modelling. In this respect the greater the variety the
better, preferably no formal tools or methods of representation should
be disallowed on a priori grounds.

How do these thoughts square with the content of you thesis?

Regards.

PS. If you want some critiques of GAs as descriptions of bounded
(ed.), Compuational Techniques for Modelling Learning in Economics,
Kluwer.

--------------------------------------------------
Bruce Edmonds,
Centre for Policy Modelling,
Manchester Metropolitan University, Aytoun Bldg.,
Aytoun St., Manchester, M1 3GH. UK.
http://www.cpm.mmu.ac.uk/~bruce



This archive was generated by hypermail 2b29 : Thu Jan 06 2000 - 17:00:29 GMT