Path: utzoo!utgpu!jarvis.csri.toronto.edu!mailrus!tut.cis.ohio-state.edu!cs.utexas.edu!uunet!portal!cup.portal.com!dan-hankins
From: dan-hankins@cup.portal.com (Daniel B Hankins)
Newsgroups: comp.ai
Subject: Re: Simulation and Understanding (was Re: Simulation verus real
Message-ID: <17191@cup.portal.com>
Date: 15 Apr 89 04:14:49 GMT
References: <827@htsa.uucp> <5790@watdcsu.waterloo.edu> <5106@cs.Buffalo.EDU> <1259@rpi.edu> <5254@cs.Buffalo.EDU>
Organization: The Portal System (TM)
Lines: 48

In article <5254@cs.Buffalo.EDU> lammens@sunybcs.uucp (Jo Lammens) writes:

>[...] In a previous posting I used the analogy of trying to understand how
>an operating system works by modeling (and simulating) the transistors
>that make up the machine on which it runs.  Suppose I know nothing about
>operating systems nor computers, and I want to simulate an o.s. using some
>other technology, say a mechanical construction with gears and pulleys
>etc.

>[...] Even though I model a transistor sufficiently precisely, and throw
>in a lot of them, do you think I will ever get the simulated o.s. to work
>if I don't know how it works or even what it's supposed to do?
>Going back to the original theme of neurons and brain functions, do you
>think that throwing in a lot of simulated neurons (I mean a whole lot)
>will automatically result in brain functions, consciousness or what have
>you, if you don't know what they are or even what they're supposed to do?

     The trick in getting the simulated opsys to work is not in merely
modelling the function and connectivity of the gates, _but also their state
at a particular point in time_.  That is, if you model the transistors of a
given system, and also initialize them to the actual states of the
transistors in the modelled system, as determined by taking some kind of
'snapshot', then I really think you will preserve the opsys as well.

     An alternative to a snapshot is to model the machine without copying
the state as determined instantaneously (something which is very difficult
to do).  Then one loads up the simulated computer with the operating system
just as one would load the real machine.  This is, in fact, the approach
that is used where I work to load up a machine which simulates another
machine.  We often do this because the actual machine is not yet available,
due to hardware problems, incomplete designs of some sections, cost of the
real machine, and so on.

     This is quite analogous to the way one would treat a neural-network
simulation.  There are two approaches;  one can build the simulated network
and try to program it via a snapshot taken of the equivalent real NN, or
one can program the NN the way the real one was originally programmed.

     The first of these approaches will be impractical for quite some time
to come.  The second seems to be practical.  Build a simulated NN which is
equivalent to that of an infant - a blank slate.  Then it will learn by
experience as humans and other animals do.

     Of course, for quite some time it will be more economical to produce
NNs by manual labor, with a 9-month delivery schedule.


Dan Hankins