Knowledge Base Or Bust

.. drclue at drclue.net
Fri Dec 14 17:43:34 EST 2001


The parent child relationship is not properly biased 
against the positional token/structure.

To actually make this thought work would require an input
feedback loop of some sort so as to derive weights
from actual data over time in a fashion that builds
it's own patterns , which then is mated.




"Arthur T. Murray" wrote:
> 
> KNOWLEDGE BASE OR BUST
> 
> Today after a hiatus of two and a half months we resume coding
> http://mind.sourceforge.net/index.html the AI Mind in JavaScript.
> First we change the "description" meta-tag to read:  "Please find
> room for this AI as mind.html on your own website so that others
> may obtain it."  The idea is to let the AI spread memetically
> across the Web.
> 
> Our next goal is to obtain a working knowledge base (KB).  That
> goal is immensely worthwhile because, once our KB works reliably,
> it will be like the http://www.cyc.com Cyc project, and we will
> be able to pack the KB full of all kinds of mundane or expert
> knowledge.
> 
> The source code left over from 29sep2001 still gives a transcript
> with faulty KB reasoning:
> 
> Human:  cats eat fish
> Robot:  CATS EAT FISH
>         32   28  10
> 
> Human:  fish eat bugs
> Robot:  FISH EAT BUGS
>         20   41  10
> 
> Human:  bugs eat dirt
> Robot:  BUGS EAT DIRT
>         20   41  10
> 
> Human:  cats eat
> Robot:  CATS EAT BUGS
>         10   41  10
> 
> The AI gives the wrong answer in the fourth exchange because the
> concept of "bugs" has retained too high an activation.
> 
> I ran the above four exchanges through only in Transcript mode so
> that I could inspect the logic of the reasoning of the AI. Then I
> went back and used the same inputs in Troubleshoot mode.  I
> obtained the same results, but this time each line of output had
> the concept-activation numbers interspersed among the English
> words of the output.  I look at those three-number sets and I
> think, gee, maybe I should have written the numbers down beneath
> each of the four lines of output that I transcribed above.  Then
> I realize that I can run the same inputs all over again and write
> down the numbers in red.  Here goes.  See above for the results.
> 
> Well, what do you know?  The output activation-levels were the
> same (20, 41, 10) in the middle two exchanges but different in
> the first (32, 28, 10) and in the fourth (10, 41, 10), where I
> entered only a partial sentence of "cats eat" in order to see
> what the AI knew about cats.
> 
> Now, I myself as the programmer and theoretician Arthur T. Murray
> realize how complex the deep mental processing is behind these AI
> Mind outputs, but I do not immediately realize exactly what to do
> in order to correct the problems and to perfect the knowledge
> base (KB).  I only know some general methods of tracking down and
> analysing exactly what is happening in the artificial mind at
> each stage in the process of thinking.  The work is daunting, and
> I think, am I up to this?  Yes, I have to be, because nobody else
> is going to understand the work that has gone before or is going
> to have faith in the future of this project.  When I finally have
> a knowledge base that works reliably, then people will make the
> effort to understand the AI and to develop it further.

-- 
--=<> Dr. Clue (A.K.A. Ian A. Storms) <>=--
--=<[]>=- http://www.drclue.net
--=<[]>=- C++ HTML JavaScript DHTML CGI TCP/IP SQL JAVA VRML NSAPI 
--=<[]>=- http://www.drclue.net/F1.cgi/HTML/HTML.html	(My famous
HTML/CGI guide.)
--=<[]>=- http://www.drclue.net/beta			(My X-BROWSER DHTML library.)




More information about the Neur-sci mailing list