I went to The Festival of Inappropriate Technology (XCOM2002) on Sunday. It
was rather chaotic (especially with the inaudible PA system for the first
hour or so), but quite interesting. I got to ask a question of Freeman Dyson
and have a quick chat with him at the book signing afterwards. Subsequent
attempts at name-dropping have been frustrated by the fact that none of my
friends or work colleagues have heard of him!
Anyway an interesting question came up that touched an issue raised here by
the Hermit. Can computers be programmed to create machine intelligence, for
example using genetic algorithms? I think there are some problems with this
approach:
* Genetic algorithms allow us to search a space of possible solutions, but
its up to us to define that space. At present I don't think we even know
enough about the problem to describe the problem space when it comes to
intelligence.
* Genetic algorithms are computationally very inefficient - they quickly
choke on large problems and are usually regarded as a method of last resort
for optimisation problems. Evolution took billions of years to evolve
intelligence on earth using a degree of parallelism we may never be able to
approach, and as far as we know this may have been a fluke.
* Genetic algorithms work well (at present) where the problem space is not
too large and we can quickly decide whether a solution is good, neither of
these apply to the problem of creating intelligence.
* Genetic algorithms need a reasonable pool of starting solutions if you
want a good end result. At present I think we have no idea how to even
create a reasonable starting population.
A colleague and I wrote a genetic algorithm based program to lay-out
chemical plant schematics according to various aesthetic criteria (minimum
line crossings etc). The time response was approx n^2. It choked on any
problem much bigger than 20 nodes. I set it the London underground as a
'stress test' - it still had nothing like a good solution 24 hours later.
Andy Brice
This archive was generated by hypermail 2b30 : Sun Sep 22 2002 - 05:06:14 MDT