Mind.Forth Programming Journal (MFPJ) Thurs.27.DEC.2007
http://tech.groups.yahoo.com/group/win32forth/message/13076
In Mind.Forth artificial intelligence for robots,
as we try to make the AI Mind balk at thinking a
thought for which it has insufficent knowledge,
we need to coordinate a drastic reduction of
post-thought activation of concepts in the
psiDamp mind-module with a testing for too little
activation on candidate-verbs in the verbPhrase
module. The psiDamp module must make sure that
conceptual activations are low -- but not too low
for a meandering chain of thought to snake its way
through the knowledge base (KB) blanketing the mindgrid.
In the verbPhrase module, the "detour" variable
must abort and force into detour any incipient
thought for which all candidate verbs have too low
an activation for a KB-compatible thought to arise.
Human input of a new word properly used as a direct
object should not permit the AI to use the same word
as the subject of a nonsense statement connecting
the new word with some inappropriate verb that
chances to have a modicum of conceptual activation.
Instead, verbPhrase should reject low-activation
verbs and detour the AI Mind into asking a question,
or making a self-referential EGO-module statement,
or finding something else to think about in a
thought-engendering traversal of the AI knowledge base.
In the verbPhrase module we try using the following code.
act @ 13 < IF \ Detour away from selecting a low-activation verb.
1 detour ! \ 27dec2007 Set the detour flag to 1-as-true.
fyi @ 2 = IF \ 27dec2007 In Tutorial mode show any detour taken.
CR ." verbPhr: detouring when verb-activation is only " act @ .
THEN \ 27dec2007 End of test for Tutorial mode
LEAVE \ 27dec2007 Stop low-activation verb before SPEECH.
THEN \ 27dec2007 End of test for verb with activation too low.
There seems to be a problem with too high an activation
being set when the human user enters a new word as input.
For example, after "FISH EAT BUGS" is said by the robot
looping through a chain of thoughts, the human user types
in "fish eat eggs" and waits for a response. After a line
of diagnostic code reveals "verbPhr: act = 24" the AI Mind
incorrectly utters the thought "EGGS EAT EGGS" in response,
because at twenty-four (24) the activation on the EAT verb
was too high to fail the "detour" test that would have
aborted the incipient nonsense-thought of "EGGS EAT EGGS".
Luckily, however, we know that human user input has a
cognitive privilege of leaving relatively high activation
on words contained in the input stream, so that the AI Mind
will pay attention to user input and generate a response.
There may be a problem in the Audition mind-module, where
the following code reveals a high setting on "actset" --
the variable used to decrement input activation in such a way
that subject, verb and object have descending activations.
\ 31 actset ! \ 1apr2007 From JSAI CR() module.
27 actset ! \ 4apr2007 Aiming for a wider S-V-O gap.
1 lastword ! \ 3apr2007 To reset "seq" tag
The above code may explain why there was so high an
activation as twenty-four (24) on the verb "EAT" in our
Mind.Forth Programming Journal (MFPJ) entry of 19.DEC.2007.
We have an opportunity now to lower the initial actset value
at the end of the enBoot module and the "actset" used in the
Audition module. By lowering the "actset" values gradually,
we may observe any drastic changes in AI mind functionality.
Meanwhile the AI Mind is continuing its "shakedown cruise"
as it becomes better and better at maintaining meandering
chains of thought. Values and settings which interfere with
chains of thought are gradually being adjusted into relative
harmony in a search for ideal rather than merely adequate
settings.
In verbPhrase we have changed important code as follows.
\ motjuste @ 0 = IF 7 EMIT THEN \ 3apr2007 A test.
\ motjuste @ 0 = IF 7 EMIT THEN \ 27dec2007 Commenting out above.
motjuste @ 0 = IF \ 27dec2007
1 detour ! \ 27dec2007 Set the detour flag to 1-as-true.
fyi @ 2 = IF \ 27dec2007 In Tutorial mode show any detour taken.
CR ." verbPhr: detouring when no candidate-verb is found.
THEN \ 27dec2007 End of test for Tutorial mode
LEAVE \ 27dec2007 Go back up to any calling module. e.g., SVO.
THEN \ 27dec2007 End of test for no candidate verb found.
It is time to upload our code for release on the Web,
because we have implemented the detour-trigger in the
verbPhrase module, as indicated in the MindForth changelog.
\ 17dec07A.F introduces "detour" as abort-flag for insufficient knowledge.
\ 19dec07A.F flushes out verb-activation-too-high and BIRDS-as-IS bugs.
\ 23dec07B.F solves aboriginal audRecog false-positive recognition bug.
\ 27dec07A.F causes verbPhrase to LEAVE when a detour-condition is met.
Although we have not coded SVO to "detour" into questions, EGO,
or KB-traversal, we have set the stage for implementing code
dependent upon the "detour" status in the SVO module.
Soon we may be able to resume invoking the Ask module or
the wtAuxSDo module, after disabling the asking of questions
while we concentrated upon the generation of chains of thought.
It is significant that our upcoming work on responding to the
"detour" flag in the Subject-Verb-Object SVO module will be
not a bug-fix but simply an elaboration of the features of
the AI Mind. Bugs will still occur and still be fixed, but
we may now have a basically functioning AI that simply needs
further development and refinement.
It will be interesting (and gratifying) to see wtAuxSDo ask
a question not by default because an entered noun is new,
but because the AI tries to use the new noun as the subject
of a sentence and can not find a verb in the knowledge base
that is known to state a factual attribute of the noun.
http://mentifex.virtualentity.com/mind4th.html
-----
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=79547636-236a02
