It may be more important to learn from failure than from success, in order to minimize chances of damaging procedures that work quite well. One method to minimize this damage is to use censors and suppressors. "A safer way to deal with this would be to modify M by adding special memory devices called `censors' and `suppressors'...which remember particular circumstances in which M fails and later proceed to suprress M when similar conditions recur. Such censors would not tell you what to do, only what you shouldn't do; still, they prevent your wasting time by repeating old mistakes."
The real world is not mathematically precise. So, "we can't so willfully make up the rules for objects that already exist, so our only course is to begin with imprerfect guesses -- collections of rough and ready rules -- and then proceed to find out where they're wrong."
(NB: All of this brings to mind Gold's theories of learning (which indicate that powerful learning techniques do involve learning from positive and negative information) and how these theories are at odds with empirical studies of language learning. Is Minsky following a similar path here? Specifically, how much of (say) language learning is learning from failure?)
Pearl Street | Society of Mind Home Page | Dawson Home Page |