> Computer: I don't *think* a nonalgorithmic search is possible.
> Human: We don't have decent specs, why should I presume anything about
> the impossibility of nonalgorithmic searches?
Start from the definition of a search and state that a "nonalgorithmic
search" is an oxymoron of the first class?
> One module I spun off this semester, while the flu was trying to convince
my
> body I was sick (it partially succeeded, it knocked out 800 math for all
> of 2 days), was a bare-bones expert-system inference engine, anything but
> commerical quality.
Expert systems, in general, aren't so much "search algorithms" so much
as "fit algorithms," at least for most I've ever seen. Rather than
searching for a match you take the elements of information you have
and see where they fall in what the ES knows, and generate further
queries based on the most likely things to make a better fit.
> I lack intuition for how an AI search programmed for this module works.
> On combinatoric problems of decent size [execution time hours], with
> "normalized rules" [they must inform the engine when they are useless],
> the AI is usually *faster* than a dedicated search loop with the
> corresponding rules explicitly coded. [C++, of course.]
>
> One of my advising professors, for a brief instant, entertained the idea
> that I had somehow built a quantum effect into the AI. He discarded this
> within 2 minutes.
No need for quantum effects to understand this; dedicated search loops
with explicit rules codings suffer a lack of dynamicism compared to a
system in which rules can flag themselves as "useless" dynamically;
nearing the solution, the ES throws away rules it no longer need
consider for a better fit, reducing the number of `decisions' that
need be made.
Just as I don't need to look for the Hilton Atlanta when I've finally
found my way to the Perimeter and can stop looking for it.
> Unless you're trying to verify that the algorithm does what you said it
> did, and you're trying to rule out a compiler error! [It usually isn't
> worth trying to figure out directly if the ASM listing actually
> corresponds to the source code.]
At this point you're no longer trying to solve the problem, you're
trying to prove you did it right. :) My test is simple, `does it
work? Yes? Then does it matter if it exploits a compiler error? :)'