Path: utzoo!attcan!uunet!lll-winken!arisia!quintus!ok
From: ok@quintus.uucp (Richard A. O'Keefe)
Newsgroups: comp.ai
Subject: Re: The Ignorant assumption
Message-ID: <372@quintus.UUCP>
Date: 9 Sep 88 09:33:18 GMT
References: <1369@garth.UUCP> <2346@uhccux.uhcc.hawaii.edu> <1383@garth.UUCP>
Sender: news@quintus.UUCP
Reply-To: ok@quintus.UUCP (Richard A. O'Keefe)
Organization: Quintus Computer Systems, Inc.
Lines: 31

In article <1383@garth.UUCP> smryan@garth.UUCP (Steven Ryan) writes:
>Seems I made I mistake. When I referred to scientific method I was referring
>to a philosophy:
[lots of stuff deleted]
>Science is philosophy on how the universe can be understood. If some aspect
>if the universe cannot be understood in this way, then science is incomplete.
>Christianity asserts this is true: that there exists transcendental forces
>which science cannot explain. That is an assumption.

This seems to be saying that "science" _ought_ to be able to explain
everything.  If this is so, then I think we just have to put up with
science being "incomplete".

 - in logic, we keep finding unprovable/undecidable things
 - in computing, we keep finding things which are infeasible in principle
 - in biology, neo-Darwinism distinguishes between "selection" (which can
   be explained by reference to phsyical/chemical/mathematic laws) and
   "mutation" (which "just happens").  We may try to explain how a mutant
   survives, but should not look for an explanation of why _that_ mutation.
 - in physics, the Copenhagen interpretation says that events "just happen"
   and rejects as wrong-headed "hidden variables" schemes.

There is a tenuous connection with AI here.  I suggested in an earlier
message that it may be rational for an agent _NOT_ to test its beliefs,
if the expected risk is high enough.  Now we find that some explanations
may not exist, or may not be computationally tractable if they do.

So, when designing a learning system, how do we deal with this?
How should it "decide" when to look for an explanation of something,
and when to change the subject?

Is there a connection with the "noisy data" problem?