Purpose: Discussion of Strong AI Minds thinking in English, German or Russian. Privacy policy: Third parties advertising here may place and read cookies on your browser; and may use web beacons to collect information as a result of ads displayed here.

Friday, April 07, 2017


Wrong solution to a bug briefly ruins word-recognition.

[2017-04-06] Let us run the ghost192.pl AI without input and try to fix the first thing that goes wrong with it. After a series of sensible outputs, at t=2562 the AI suddenly says "HELP I" without a subject for the verb. As we investigate, we see that EnNounPhrase is trying to activate a subject at t=2427, but the pronoun "I" is stored at t=2426 with an erroneous recall-vector "rv" of t=2427. The error in auditory storage causes the AI at a later moment not to find the auditory engram.

[2017-04-06] We notice that MindForth sets tult in the AudInput module, while the Perlmind is setting $tult in both the InStantiate module and the AudInput module. However, where $tult is set, does not seem to matter. We eventually notice that some MindForth code ported into AudInput() was letting the $rv recall-vector be set erroneously not only for an alphabetic character, but also for a CR-13 carriage-return or a SPACE-32. When we restricted the $rv setting to alphabetic characters, our current bug was fixed, and the AI no longer said "HELP I".

Letting $rv be set only once per word correctly solves a bug.

[2017-04-07] Yesterday in ghost192.pl our attempt at solving a recall-vector $rv bug made the AI unable to recognize reentrant words. Now in ghost193.pl we would like to isolate $rv so that its value can be set only once in each cycle of recognizing a word. When we do so, we obtain the proper $rv value for the first word stored by the AI, but it remains the same value for all subsequent words being stored. We must determine where to reset $rv to zero. We try resetting $rv to zero at the start of the Speech() module, as MindForth does. Immediately we see fresh values of $rv being stored for each reentrant word. We let the AI run on at length, and it no longer says "HELP I" without a subject for the verb. Then we start the AI with an input of "you know me" and somewhat later the AI remembers the self-referential knowledge and it outputs, "I KNOW YOU". Thus we have made a major improvement to the AI functionality by fixing the $rv bug. There remain grammatical issues, probably based on software bugs.