MFPJ: Mon.17.DEC.2012 -- The Logic of Thought

Right now we feel as though we are on the verge of a major breakthrough 
in our Mentifex AI programming. We want our AI Minds to become able to 
draw conclusions from the knowledge of two or more facts that lead to 
a third fact. For instance, if we inform the AI that "Birds have wings" 
and "Eagles are birds," we want the AI to figure out the inference that 
"Eagles have wings." But how can our AI software achieve such a trick? 

Over time we have learned that just writing about our software dilemmas 
often leads to solving a dilemma. We can brainstorm right here and now 
into the MindForth Programming Journal (MFPJ) and perhaps come up with 
a solution. 

One possibility is, that the 800=BE verbs will somehow pass the 
attributes of a predicate nominative back to the subject of the 
sentence linking a subject with a predicate nominative. There could 
be a kind of BrainSweep module that, encountering a be-verb sentence 
of the Is-a variety, would traverse recent memory in search of the 
predicate nominative used as a nominative subject and then build 
an apparatus of S-V-O linkages connecting the new subject in the 
Is-a relationship to the inferential verbs and direct objects normally 
true of the class specified by the Is-a predicate nominative. 

In other astounding words, we could have our AI Mind software 
stop for a moment and suddenly create one or more non-auditory 
triples, taking everything recently or ever known about the 
Is-a predicate nominative and creating one new silent triple 
for each salient fact. If the AI knows, "Birds have wings," 
and then it hears either "An eagle is a bird" or "Eagles are birds," 
the AI will take the Psi concept number for "eagle" and silently 
link it to the concepts of "have" and the direct object of "wings". 
Lo and behold, we have created an inference engine. 

These inferential engrams would not be linked immediately to 
instances of auditory memory. Instead, any thought process making 
use of the inferential triple would have to search backwards in 
memory for auditory engrams to express the new, not-previously-known 
thought -- which is easy to do in Mentifex AI Minds. MindForth and 
Wotan and Dushka all have the ability to abandon current auditory 
engrams and replace them with grammatically more suitable and 
correct word-forms. 

As we began this session of brainstorming about MACHINE REASONING, 
we thought that we might have to devise some rather weird and ungainly 
mechanisms, such as adding in an extra conceptual panel-flag to hold 
Is-a relationships. We did not really want to go that route, because 
it would be unwieldy. 

Now for inferential machine reasoning we envision a mind-module 
that kicks in when a be-verb is used to express an Is-a relationship 
between a subject and a predicate-nominative class. It is easy to 
set a flag when a be-verb is encountered in the input stream, and to 
use that flag to call the BrainSweep or InFerence module. We should 
probably call it InFerence instead of BrainSweep, even though it 
will sweep the artificial brain to find the inferred relationships. 
Even having "InFerence" as a module of our AI source code will be 
very impressive and very prestigious. 

Anyway, once the be-verb flag ("is_a"?) calls the InFerence module, 
the InFerence mechanism searches back through experiential memory 
to find class-triples with which to instantiate instance-triples. 
These new instance-triples would be readily visible in the diagnostic 
record of the Psi array, not only in MindForth and Wotan but also in 
the JavaScript AiMind.html and Dushka Russian AI. It will be a bizarre 
thing to witness, too -- inferences showing up as hidden engrams 
after the input of an Is-a factoid. If we code this thing, the IQ of 
each AI Mind should increase by at least one tenfold order of magnitude. 

Mentifex (Arthur)
-- 
http://www.scn.org/~mentifex/AiMind.html 
http://www.scn.org/~mentifex/Dushk.html 
http://www.scn.org/~mentifex/mindforth.txt 
http://www.scn.org/~mentifex/DeKi.txt 


-------------------------------------------
AGI
Archives: https://www.listbox.com/member/archive/303/=now
RSS Feed: https://www.listbox.com/member/archive/rss/303/21088071-f452e424
Modify Your Subscription: 
https://www.listbox.com/member/?member_id=21088071&id_secret=21088071-58d57657
Powered by Listbox: http://www.listbox.com

Reply via email to