On 20/10/2007, Robert Wensman <[EMAIL PROTECTED]> wrote:
> First of all, I do not believe science can allways respect past exact
> definitions of words in order to make progress. How about if Einstein
> refrained from publishing his relativity theory, because it would contradict
> the way people normally think about time/gravity? Beeing conservative in
> this aspect is the same thing as writing a program, without ever going back
> and changing names, revising old code. In the end you would be left with a
> mess.
I think sometimes the code base is so old and opaque, you might need
to start from scratch and develop a new one. Possibly even with a new
project name ;-)
> Secondly, the scientific field of AI has already proved their incompetence
> in producing anything near AGI. It is apparent the field is too much focused
> on narrow AI, so why should we respect their narrow definitions which
> probably are set up to lead us into a hole again? Basically: Screw them,
> they have had their chance... ;-).
Redefine it then.... But realise you need to redefine it, and
publicise that redefinition. I find it easier to call my interest a
different name (selectionist computer systems is one I am musing
about), but if you think you will have luck trying to work against the
status quo, good luck to you.
> Thirdly, the definition you outline on the topic is so broad that actually
> all learning systems that maintain and improve a set of solutions would fit
> into it. "select some set of the population and make a new population" could
> basically mean anything. The designs I have considered involve maintaining a
> set of favoured model components, and combining them to produe new models.
Do not all learning systems evolve in a way ;)
It is quite a common template to analyse a evolutionary algorithm. How
does it vary, select and what size population it has.
Wikipedia has similar pseudo code for a GA, I just generalised to
describe other EA that don't necessarily have crossover or have
different methods of selection.
Wiki's Code -
1. Choose initial population
2. Evaluate the fitness of each individual in the population
3. Repeat
1. Select best-ranking individuals to reproduce
2. Breed new generation through crossover and mutation
(genetic operations) and give birth to offspring
3. Evaluate the individual fitnesses of the offspring
4. Replace worst ranked part of population with offspring
4. Until <terminating condition>
GA and GP are nothing special at the moment, just another form of
search of a fixed space with large populations and using lots of
randomness in the variation and selection. They are subject to the no
free lunch theorems of search, which you should know about if you are
interested in that sort of thing.
And it is not quite true that all learning systems would follow that
definition. I explicitly said it was pseudo code, so even if a system
has variation and selection in it, if it didn't follow those steps in
a centralised code like fashion it would be something different from
standard EA or standard learning.
> So because I use a different definition of evolution, compared to this more
> narrow form of evolution you refer to, your attack on the inneficiency of
> "narrow evolution" is not directly relevant to anything that I said
> previously.
Give us then an example of what you mean by evolutionary systems.That
has been my major point all along, without that people will assume
what you are interested in is like something else they know.
> Also, I do not believe that the human mind creates optimal routes from point
> A to point B.
Who said anything about optimal? I was just talking about finding a
route from A to B in a realistic time frame..
Will Pearson
-----
This list is sponsored by AGIRI: http://www.agiri.org/email
To unsubscribe or change your options, please go to:
http://v2.listbox.com/member/?member_id=8660244&id_secret=56022282-fc21b1
