Re: Randomnization of thoughts and (project three)

Hi.  I am trying to read your message carefully to see if I can come up with
any ideas that you might be interested in without getting into an argument
against GA's.

During the past few days I got hit by more 200 emails a day that were caused
or sent by the Swen virus even though my computer was never affected.
However, it seems to have gone back to normal now.  I just thought I would
let you know.

If you feel like telling me more about your ideas I would appreciate it.  I
think that in the sixties Chomsky wrote about the manifestations of
traditional grammatical semantics in explicit, simple abstractions suitable
for computational slots.  Although you didn't seem to like the idea of an
ontology, these abstract semantics of grammatical parts of a sentence always
appealed to me for their abstract semantic simplicity as well as their
extent.  These could be used for a superficial search for the possible
relationships in the surface features of the text.

I will try to have some more to add in a few days, but unfortunately I am
pretty busy right now.  So if you can tell me more about how this program of
yours is going to work, and how you see these active agents working during
the initial stages of your program, I would enjoy reading it.

Jim Bromer

"Patty Cutman" <[EMAIL PROTECTED]> wrote in message
news:[EMAIL PROTECTED]
> Jim Bromer wrote:
>
> > When I used the term bonding areas I meant something along the lines of
what
> > you and what Roger Schank were thinking of, but I can generalize that
idea
> > to apply to all kinds of thoughts.  You mentioned Schank's example of a
> > semantic relation between the expression 'give' and the unspoken idea of
a
> > gift. I think that kind of semantic relationship in which the sentence
needs
> > to be interpreted according to previously learned knowledge is typical
of
> > all conceptualization or thought.  If I see a faucet I know what it
> > represents even if I can't get any water out of it.  This ability is
> > fundamental to survival and its significance in intelligence certainly
> > transcends linguistic expression.  The faucet is a symbol of water.  But
> > since faucets are different and varied the actual sight of a faucet I
have
> > never seen before must be interpreted through some kind of
generalization of
> > what faucets look like and what they 'mean'.
> >
> > I am not sure what you meant by, " But we are really looking for binding
> > sites for active processes."
>
> The bot personas that I envision will have a passive semantic network.
> Such a network, however well designed, would always lay there flat and
> do absolutely nothing.  For behavior to happen we need active agents
> that do something.  These are the agents that we contrive to live or die
> based upon their performance.  But these agents (sometimes I call them
> gremlins) bind to the passive semantic network at sites that are
> particular to the agent (or agent class).  This is very much like a
> protine molecule which binds to a specific site on a cell membrane.
> Once bound, the agent performs its task.
>
> However here is where I would make a slight change to the biological
> model.  Any number of gremlins could bind to the same semantic network
> site.  This way they can compete to perform their task.  The ones that
> fail will be selected against.
>
> > A pronoun has a strong grammatical potential to be bound with other
surface
> > features of a sentence.  In other words, other referents to the (symbol
of
> > the) object or being that the pronoun would refer to might appear in the
> > input in the vicinity of the pronoun.  The object that a semantic
binding
> > would refer to (as in the Schank example) might not appear in the input,
but
> > the program would have to figure it out.  This is the thing.  Even if a
> > primary referent did appear in the text along with a pronoun that
referred
> > to it, the program would have to figure out whether that binding made
sense
> > or not.  A gremlin would have difficulty doing this unless it had access
to
> > a lot of knowledge.
>
> The gremlin has access to the entire semantic network .. it can navigate
> it and search or retireve whatever it needs.  Within itself it should
> have very little knowledge, just its own fixed behavior.  Of course the
> fun part is where gremlins bind to other gremlins as well as to the
> semantic network.
>
> > This is why I am not that impressed by neural networks,
> > agents and genetic algorithms in their simple forms.
>
> Like I said, project three is very speculative.  But I have been
> impressed with what Alife can do on its own - very impressed.
> Incidentally there is nothing preventing a agent from growing its own NN
> or even using traditional logic programming.  There should be no
> prohibition on what paradigm gremlins can use.
>
> > (By the way, I have
> > always said that I am very well aware that if a neural network model was
> > adequately modified, it could have the full potential of any other kind
of
> > program.  I say the same thing about genetic algorithms and
> > quasi-independent agents.)  But back to my question: are you looking for
> > surface active features in text that might be used in initial conceptual
> > binding??
>
> Well initially that is all there is - just the surface features that can
> be found in the syntax and the lexicon of the words in the network.  But
> I suspect that people will want to create gremlins that add their own
> nodes to this network and those added nodes can represent things in
> whatever world.
>
> > One of the problems with AILife is that I don't see how the genetic
strings
> > are capable of using insight about previous learning in solving new
problems
> > unless that insight is purely coincidental (in two meanings of the term
> > coincidental.)  Your own rhetoric (your belief in AI Life) suggests that
you
> > might be dismissive of the concept of 'insight' when it comes to AI even
> > though you probably would take it seriously in an examination of your
own
> > intelligence.  That is, even if you believe that 'insight' is a complex
that
> > can be explained by reference to simpler components, when it comes to
your
> > AI project, aren't you are just ignoring the entire question?  I am not
> > saying you will never make any progress with AI Life, I am just saying
that
> > you won't come up with an adequate explanation of things like 'insight'
> > working only with genetic algorithms (in their current form).
>
> Well I don't even know what insight is anymore, let alone setting out to
> create it artifically.  When I was 16, I knew I had insight.  Now I'm
> not so sure.
>
> > When I spoke of independent/dependent learning I just meant that new
> > learning can take place only on through previously acquired knowledge.
That
> > is, you won't be able to learn something new without figuring some
things on
> > your own.
>
> I want my artificial personas to start out with a fair amount of
> prebuilt abilities - eg POS Tagging of English text, and generating of
> grammatical English text from a semantic network.  What the gremlins
> need to figure out on their own - collectively of course - is how to
> choose a focus from a network that will interest their hosts.
>
> > You were right, I did append the sentence, "The game of life is best
played
> > with other life forms," to the end of my message.   I was trying to give
you
> > sagely wisdom, and I just couldn't throw it away even though I couldn't
fit
> > it into anything that I was saying!  I figured that you might be working
> > with some group.  I didn't even think about the fact that the school
year
> > just started.  Are you working in an university setting?  I wanted to
> > emphasize that the learning that can take place in a group effort is
much
> > more substantial than the learning that can take place in a genetic
> > algorithm.  But even aside from the value of working with a group, just
> > working in the garden makes me think that the real game of life is a
very
> > complicated process.  The belief that insight garnered from scientific
> > reductionism implies that complicated things are actually simple is
naïvely
> > simplistic.  Effective AI is difficult to achieve because it is
complicated.
>
> I totally agree with the thrust of that paragraph.  Project three, for
> me, is 90% social engineering with humans, and 10% programming
> technology.  One of my big complaints with university AI projects is
> that they tend not to have interaction with a broad community - at best
> they just interact with a small group of grad students and the
> professor.  And that is *the* primary reason that I have come to these
> groups to run my mouth.
>
> Note my new email address - the excite email is so severely infested
> with worms and viruses and spam that it is useless - so if anybody wants
> to contact me directly, now you can.
>
> [EMAIL PROTECTED]
> An intelligent human should be able to figure out how to mail me there.
>   Hopefully the spam bots can't.
>
> Patty
>