Re: [-empyre-] self-modification, emergence

[Paul Brown <paul@paul-brown.com>]

> My favourite example of self-modification and emergence is the work 
> on evolved circuit designs by Adrian Thompson of the COGS lab at 
> Sussex (http://www.cogs.susx.ac.uk/users/adrianth/ade.html). 
> Basically he used a programmable chip (a FGPA, field gate 
> programmable array) to "breed" and automatically test thousands of 
> electronic circuits, on a task such as distinguishing between two 
> different frequencies at their inputs. Many many generations later, 
> circuits were evolved that indeed fulfilled the task. But when the 
> evolved circuits were analysed, they were found to operate according 
> to no known principles of circuit design... for one thing they were 
> tiny and incredibly efficient (using only a small portion of the 
> array) but also they used no internal clock, and instead seemed to 
> use the dynamics of interlinked feedback loops to analyse the input. 
> Best of all, the circuits didn't work so well when the same design 
> was transferred to another chip: the evolved design made use of the 
> specific physical characteristics of its substrate.

In 2000 I was artist-in-residence at the Centre for Computational
Neuroscience and Robotics at the University of Sussex where Adrian
is based.  I had a number of conversations with him about this work
but it's maybe 18 months since we last met.  Like Mitchell I find
this work of particular interest.

Adrian's original experimental rig only worked in one specific location
if he moved it a few centimetres along the bench it stopped.  He
discovered that there was a power cable embedded in the wall at the
original location and that the evolved circuit was "stealing" the mains
impedance frequency as an input.  This example convinced me of the
importance of embodiment in AI and alife.  It also undermines the
authority and potential of GOFAI - good old fashioned top down or
"classical" AI which focuses on disembodied intelligences that exist
in code/dataspace (eg. Moravec)..  It also reinforces Ihnatowicz'
(the Senster) 1960's insight that intelligence is the result of
interaction with the environment which he based on his reading of Piaget.

The other interesting thing about Adrian's work is that although he now
has stable circuits that can be migrated about the environment, about
the FGPA and work at different temperatures etc... he still doesn't know
how they work.  As a classically trained electrical engineer he set
himself the task of evolving a circuit to undertake a task but using
less components than the classical theory demanded.  For the past
few years he has been trying to work out how they are are overcoming this
limitation.  He has looked at quantum tunnelling and more recently
quantum entanglement (and a lot of other physical phenomenon as well).

An important aspect here is that the techniques that alife researchers
are using often lead to solutions that the researchers don't comprehend.
A lot of the research going on in the field is an attempt to analyse
these problems to develop a theoretical underpinning for the discipline.

I think this is also a repost to Christophe's request that we "deal
fully with the question of language (and not only computing language)".

It's my opinion (and I know it's not a popular one!) that language
comes after the event (and cannot predict or pregender new
knowledge except about language itself).  Critics say I'm still
stuck in the Tractatus - a criticism I'm more than happy to accept!

Put more simply - we do the work and then language is developed to
first analyse and then codify that event.  Put another way - theory
comes after the event and cannot come before it (it can only come
before repeats of previously existing events).  This is not to say
that theory cannot create "spaces" (concepts?) for event to take
place - but it can't create events.

So I think that Christophe's suggestion is premature - that language
that will describe alife will not and cannot exist for some time now.
We need to do the work that will create the body of experience so that
language can deconstruct and codify it and turn it into knowledge. As
I have suggested elsewhere when the language to describe an event
evolves then it also kills the event and turns it into history.

LeWitt's instructions did not predict the work.  The idea (concept)
predicted the instructions.  A difference here between language as a
description and language as a set of instructions (like computer
language).  Alife attempts to go beyond this inherent limitation of
the language by using evolutionary and emergent methods.  The up
side is that we can get results.  The down side is that we often
can't comprehend them (because we have gone beyond the limits of
language and have to wait for it to catch up).

Maybe it's important to qualify these assertions by stating that I
am not a theorist - I am a make of things (not necessarily objects!)

And just as a complete aside I can't help (mis?)quoting Minsky that
"language is a system for building ideas in other peoples heads"
and, of course, Burrough's wonderful "language is a virus".

Sorry BTW that I was offline yesterday.  We had storms and my alife
making machines don't like unstable electricity!
--
===============================================================
Paul Brown          PO Box 413, Cotton Tree QLD 4558, Australia
mailto:paul@paul-brown.com            http://www.paul-brown.com
mob 0419 72 74 85                           fax +1 309 216 9900
===============================================================
Visiting Fellow - Birkbeck    http://www.bbk.ac.uk/hafvm/cache/
===============================================================