home *** CD-ROM | disk | FTP | other *** search
- Comments: Gated by NETNEWS@AUVM.AMERICAN.EDU
- Path: sparky!uunet!usc!howland.reston.ans.net!paladin.american.edu!auvm!IMAG.FR!MALER
- Return-Path: <@VMD.CSO.UIUC.EDU:maler@vercors.imag.fr>
- X-Mailer: Mail User's Shell (6.5.6 6/30/89)
- Message-ID: <9301121816.AA23292@pelvoux.imag.fr>
- Newsgroups: bit.listserv.csg-l
- Date: Tue, 12 Jan 1993 19:16:23 MET
- Sender: "Control Systems Group Network (CSGnet)" <CSG-L@UIUCVMD.BITNET>
- From: Maler <maler@IMAG.FR>
- Subject: Re: Koza, Genetic Programming, & LISP
- In-Reply-To: "CZIKO Gary A." <g-cziko@UIUC.EDU> "Koza, Genetic Programming,
- & LISP" (Jan 12, 10:55)
- Lines: 58
-
- [From Oded Maler 930112]
-
- [Gary Cziko 930112.1620 and before and others]
-
- The last thing I want is to cool down the enthusiasm of the first
- encounter with machine learning, inductive inference and automatic
- programming. But.
-
- Few days ago I realized that the subject of my first (but fortunately
- not the last) Ph.D. proposal was in fact about (pseudo-) mind reading.
- It was intended to use machine learning techniques for "student
- modeling". That is, given that a student has a wrong "program" to do
- arithmetical subtraction, I give him some excercises and from his
- answers I try to infer which program he is using. All this was in the
- cognitive framework (funny, but people in my (computer) age learned
- about cognitivism before behaviorism). Anyway when I realized that the
- gap between theory and reality in this case is too big, I moved to the
- more theoretical aspects of inductive inference and what I can tell
- you from my visit there is:
-
- 1) There are many methods besides GA for trying to build expressions/
- functions/programs which are compatible with a given sample.
-
- 2) The dark side of any such methods might be that they require an
- enormous amount of data or generate a huge program (in the worst case
- you might get the look-up table containing all the sample points).
-
- 3) GA is a kind of combination of a random/heuristic search which seems
- to work in some cases (witness Koza's successes). On the other hand
- this field (like many others) includes some anti-scientific jumps
- between the fact (a computer program that searches among syntactic
- objects) and the metaphor (evolution, mutation, cross-over) and other
- stuff that gives a theoretical computer scientist the same feeling
- as Rick gets when he hears "reinforcement".
-
- 4) In your intended application of GA you should distinguish between
- the scientific inductive inference you want to perform "from above",
- e.g., to find some (god forbid) S-R regularity in tracking data, and
- between modeling learning of an individual (re-organization). For the
- former you don't need to care whether it is done by GA, non-parametric
- estimation, search in a space of decision lists, perceptron, backprop
- and any other method. It is just a tool for trying to find regularity
- in some data. For the latter, GA might have some advantage because
- it seems consistent with the postulates or reorganization, but so will
- any procedure that will work on random perturbation on the control
- network.
-
- 5) All this does not come to undetermine the impressive HACKING
- results achieved by Koza and others.
-
- --Oded
-
-
- --
- ===============================================================
- Oded Maler, LGI-IMAG (Campus), B.P. 53x, 38041 Grenoble, France
- Phone: 76635846 Fax: 76446675 e-mail: maler@vercors.imag.fr
- ===============================================================
-