Robert.Frederking@CMU-CS-CAD.ARPA (06/25/84)
There is a good article in the Winter 83 AI Magazine (4;4) about
non-logical AI (it is a rebuttal to Nils Nilsson's Presidential Address
at AAAI-83). The authors point out that certain problems are intractable if
dealt with symbolically, whereas they are easily solved if one uses
real numbers and ordinary math. I suspect that the human brain uses a
combination of analog and digital/symbolic processing, and that some
cases of intuition might arise from the results of an analog
computation into which introspection is not possible.
As for Ken Laws's comment about switching to a new optimal
strategy at each step (rather than Berliner's smoothing of
transitions), one of the things he is trying to get around is the
"horizon effect", where the existance of a sharp cut-off in the
program's evaluation makes it think that postponing a problem solves it
(since you no longer see the problem if it is pushed back over your
horizon). In other words, perhaps the optimal strategy at each point *is*
a non-linear combination of several discrete strategies.
Also, I think it is a mistake to say that "pattern-matching"
and "reasoning" are different things. After all, one must
pattern-match in order to find appropriate objects to combine with an
inference rule (obvious in OPS5, but also true in PROLOG). The
question at hand is perhaps more whether one is allowed to use logically
unsound inferences (a.k.a. heuristics).