Path: utzoo!utgpu!news-server.csri.toronto.edu!cs.utexas.edu!sdd.hp.com!uakari.primate.wisc.edu!uflorida!gatech!mcnc!ncsuvx!news
From: fostel@eos.ncsu.edu (Gary Fostel)
Newsgroups: comp.ai.philosophy
Subject: RE: Emergent Properties
Keywords: chaos, science, prediction
Message-ID: <1990Oct12.214636.7945@ncsuvx.ncsu.edu>
Date: 12 Oct 90 21:46:36 GMT
Sender: news@ncsuvx.ncsu.edu (USENET News System)
Reply-To: fostel@eos.ncsu.edu (Gary Fostel)
Organization: North Carolina State University
Lines: 75


    In spite of the flood of suggestions for what "emergent properties"
    might be, I remain uncertain and quite skeptical of the value of the
    term.   My intent is not to procure a new flood of attempts at
    defining or justifying this term -- the last umpteen are plenty.
    An unsympathetic definition might be that an emergent property is
    one that is not or was not predicted from localized properties 
    of the elements which, when combined, produce the new property.
    My favorite unsympathetic example: from the perspective of
    a Martian, an emergent property of a collection of parts and tools
    and an engineer, might be the production of a useful machine.  Of
    course the martian does not know that the engineer is a particualrly
    important part.  We know that and using "emergent property" seems
    absurd -- from OUR perspective.  

    I suspect that what is and is not going to fall under many peoples
    notion of "emergent" is going to depend on the level of understanding
    of the people and the moment. The more easily a property can be
    predicted using availble methods, the less likely it is to be an
    "emergent property".  For example, many people seem willing to use
    this term for functional properties of neural systems, but I wonder
    if as many would be comfortable with the statement that a tabular
    printout is an emergent property of a particular set of Cobol
    statements.  After all, the table is not at all readily predicted
    from any one of the cobol statements. 

    A more interesting (to me) issue is whether there might be some
    properties that really are VERY hard to predict or model based on
    the constituents.  For example, non-linear dynamical systems are
    often essentially impossible to predict, not due to lack of theory
    but for intrinsic reasons -- so called chaos theory.  In the domain
    of artificial intelligence (or perhaps right outside it :-) are 
    some people who argue that human intelligence can not be duplicated
    or modeled because of the subtle but undeniable infusion of EVERY
    detail of life into the decisions and thoughts of a moment. If
    memory serves me, Penfield is a recent example of this group. 

    Now, it is not so difficult to produce a model that does things
    like the things a chaotic system does -- it is only hard to 
    make a model that does the same thing a chaotic system would do.
    An interesting property, "P", that systems might have, would be
    that they produce behaviors that are drawn from a well defined 
    set of behaviors, even though the direct prediction of which
    behavior is intrinsically intractable.
   
    A cobol program does not have property P, since it's behavior is
    quite predictable; a set of neurons (esp real neurons) might not
    be so easy to predict.  Neural "programming" is more a question
    selecting alternative neural nets from a set of possible nets until
    the bahavior happens to be the one desired.  Such a system might 
    well have property P if it could be shown that the selection 
    strategy was really the only way to get the desired behavior
    with probability 1.  

    The selection stratgy for a neural net may bother some folks,
    who feel they "design" nets. In the case of artificial nets, it
    is probably true that the net can be apriori "designed" and then
    built.  I would say that those systems do not have property P.
    Natural nets and some synthetic nets, are often "trained"
    which really means that a sequence of nets are produced, with the
    sequence terminating when a net with the desired behavior is 
    found.  If it were shown that specific behavior of one of these
    nets was not predictable from any feasible set of measurements of
    properties of the net then the selection scheme would be required
    and the final result would have property P. 

    Perhaps my property P is what others are calling "emergence" and
    I am just befuddled, or perhaps P is something else and I'm still
    befuddled anyway.   If you would like to spend some time sharing
    my befuddlement, consider whther there is a relationship between
    systems with property P and problems which are NP complete.  In
    each case it seems that there is not a way to "get inside" the
    problem, and search  may be the only way to go.
 
----GaryFostel----