XPost: comp.ai.edu
From: ondra@dynawest.cz
"Dmitry A. Kazakov" writes:
| On Wed, 07 May 2008 10:57:58 GMT, Ondra Zizka wrote:
|
| > Is there some AI theory ( or idea / area of research aiming to create
| > a theory) which would cover most currently known concepts and use
| > them together? What about some fuzzy graph-like database of
| > n-tuples holding all knowledge of an intelligent system, perhaps using
| > neural networks to create the fuzzy relations and to perform
| > tranformations of both short-term knowledge (aka. cogitation) and long-term
| > knowledge (learning, memorizing, creating memories) ?
|
| The topology of the graph in effect induces some distance/similarity
| measure in n-dimensional space of tuples, which in turn determines how
| learning works. This implies that there cannot be any universal structure,
| because for any distance we could construct a problem, for which the least
| distance learning will not work. Now, if the structure is to define the
| distance, then that is not universal. If the distance is determined by
| something else, then the structure is not *all* knowledge.
|
Sure, the structure would not hold *all* knowledge, just the storable
part of it. Having human brain as inspiration, it has specialized
centers created as described in DNA, which could hardly learn do their
task (sound processing, optical object recognition, cause/effect
principle - all these seem to be "hard-wired").
The structure would hold experience (actions done and its effects,
learned techniques), memory (remembered objects, remembered "classes
of objects", social memory), current environmental information (like
"where am I", "what's the time"), current, mid-term and long-term
"goals", etc etc.
| Hence there cannot be such universal thing. There could only be ones,
| suitable for some class of problems. So the question is, which class of
| problems is equivalent to/required by "cognition". So far, nobody knows
| this.
And any estimates?
My personal bet is that it will have something of Prolog's inference
mechanism, only the associations will be fuzzy and self-learned,
stored in the structure I've described above, and the rules will be
also subject of inference and storing - and that's the way the
intelligent system will learn:
1) current sate -> accidental actions done -> their effect ->
associations update
2) current sate -> observated environmental changes -> their effect
-> associations update
3) current sate + desired state -> inference -> assumed actions
needed -> actual effect of actions done -> associations update
Say, our intelligent system is a newborn.
Ad 1) He is hungry. Someone offers feeding. He accepts and is
not hungry. Thus, the intelligent system associates: "I am
hungry" + "Feeding" -> leads to "I am not hungry".
Ad 2) He is alone. He cries. Someone comes and offers feeding.
Associates: "I am alone" + "I cry" -> leads to "Someone comes and
offers feeding".
Ad 3) He is hungry and not alone. He wants to be fed, and by
inference, he finds the "I cry" action, with value say 0.5,
because he is not alone; it's low value, but he has nothing better,
so he tries the action "I cry". And it works! - he is fed. Thus,
updating the (fuzzy) association: "I am hungry" + "I cry" += 0.25,
"I am alone" + "I cry" -= 0.15.
Not that the newborn is powered by fuzzy prolog, but it could work
this way.
The question now is, whether such infering could solve all problems.
As far as I can imagine, it could solve quite complex tasks. What's
your opinion?
Ondra
[ comp.ai is moderated ... your article may take a while to appear. ]
--- SoupGate-Win32 v1.05
* Origin: you cannot sedate... all the things you hate (1:229/2)
|
