M
Mentifex
Am 9. Juni 2011, Donnerstag Morgen in der Eigerwand.
I am not going to sit here and try to write the great
American Usenet post. I just want to record my
thoughts and feelings.
Yesterday I spent hours and hours coding the
http://www.scn.org/~mentifex/AiMind.html
JavaScript artificial intelligence (JSAI).
The AI was failing to provide the correct
answer to a simple query of "Kids (make)
[what]?" Instead of activating the engram of
"KIDS MAKE ROBOTS", the AI was
activating (and then inhibiting) the engram
of the query itself. I had to test and track
the causation-chain through one mind-
module after another. EnReify was not
properly putting the eligible activation
values up, so I had to completely
rewrite the EnReify module.
Then the AI started choosing the correct
verb-engram, but not the correct direct object.
In SpreadAct I discovered that the "spike"
for the direct object had a useless value of
zero. So I had to follow the chain back
into the VerbAct module and see what was
going wrong there. It was really strange
that the activation-value in the VerbAct
module was at minus thirty-two, so the
"spike" for the direct object was coming
out at zero activation.
It turned out that in the VerbPhrase module
I was doing the minus-thirty-two inhibition of
the verb engram too soon -- before the call
to VerbAct, so that a negative activation
was being passed into VerbAct. I solved
the problem by moving the VerbPhrase
inhibition code down below the call to
VerbAct, where the negative inhibition
values could no longer interfere with the
"spike" going from VerbAct to SpreadAct.
Debugging the JSAI took me hour after hour
of coding. At one point, I had to create
8jun11T.F as a "Test" version of MindForth,
so that I could fix the JavaScript AI in
comparison with the Forth AI.
Mentifex (Arthur)
I am not going to sit here and try to write the great
American Usenet post. I just want to record my
thoughts and feelings.
Yesterday I spent hours and hours coding the
http://www.scn.org/~mentifex/AiMind.html
JavaScript artificial intelligence (JSAI).
The AI was failing to provide the correct
answer to a simple query of "Kids (make)
[what]?" Instead of activating the engram of
"KIDS MAKE ROBOTS", the AI was
activating (and then inhibiting) the engram
of the query itself. I had to test and track
the causation-chain through one mind-
module after another. EnReify was not
properly putting the eligible activation
values up, so I had to completely
rewrite the EnReify module.
Then the AI started choosing the correct
verb-engram, but not the correct direct object.
In SpreadAct I discovered that the "spike"
for the direct object had a useless value of
zero. So I had to follow the chain back
into the VerbAct module and see what was
going wrong there. It was really strange
that the activation-value in the VerbAct
module was at minus thirty-two, so the
"spike" for the direct object was coming
out at zero activation.
It turned out that in the VerbPhrase module
I was doing the minus-thirty-two inhibition of
the verb engram too soon -- before the call
to VerbAct, so that a negative activation
was being passed into VerbAct. I solved
the problem by moving the VerbPhrase
inhibition code down below the call to
VerbAct, where the negative inhibition
values could no longer interfere with the
"spike" going from VerbAct to SpreadAct.
Debugging the JSAI took me hour after hour
of coding. At one point, I had to create
8jun11T.F as a "Test" version of MindForth,
so that I could fix the JavaScript AI in
comparison with the Forth AI.
Mentifex (Arthur)