home *** CD-ROM | disk | FTP | other *** search
- Newsgroups: comp.ai.neural-nets
- Path: sparky!uunet!zaphod.mps.ohio-state.edu!sdd.hp.com!wupost!gumby!destroyer!ubc-cs!alberta!arms
- From: arms@cs.UAlberta.CA (Bill Armstrong)
- Subject: Re: Rule extraction from NN
- Message-ID: <arms.712257807@spedden>
- Sender: news@cs.UAlberta.CA (News Administrator)
- Nntp-Posting-Host: spedden.cs.ualberta.ca
- Organization: University of Alberta, Edmonton, Canada
- References: <1992Jul27.141356.281701@cs.cmu.edu>
- Date: Mon, 27 Jul 1992 17:23:27 GMT
- Lines: 45
-
- sef@sef-pmax.slisp.cs.cmu.edu writes:
-
-
- >The general problem with rule extraction from neural nets is not solvable.
- >A neural net provides a richer language than can be captured in a small,
- >human-readable set of rules, whether symbolic or fuzzy.
- ...
- >Looked at from the other side, this is a kind of advantage: neural nets
- >allow us to explore the wide universe of mappings for which there is no
- >simple rule-based equivalent.
-
- >-- Scott
-
- I agree with Scott. An NN solution may or may not be easy to
- understand; and a wide universe of mappings is offered.
-
- However, a disadvantage of the NN approach is that the universe of
- possible mappings may be *too wide* in the case of a particular
- problem. If a priori knowledge implies there are some restrictions on
- the solution mapping, such as monotonic behavior or other a priori
- constraints that have to be satisfied (e.g. the value of output B is
- always between output A and output C), it may be hard or impossible to
- force the neural solution to respect those constraints. Since the NN
- solution would certainly be incorrect if the constraints don't hold,
- it might be hard in such cases to have confidence in any explanation
- derived from it.
-
- Do people have any experience enforcing monotonicity constraints? How
- would this be done in a BP system? If you forced all weights in the
- net to be positive, then the function would be increasing in all
- variables. If your variables are chosen as x1 (for monotonic
- increasing), minus x2 (for decreasing),etc. then this would work. You
- still have to worry about functions that change from increasing to
- decreasing or vice-versa, and I have no idea how to do this within
- the usual MLP paradigm.
-
- Needless to say, in some control applications, monotonicity is very
- important. If the pressure in a boiler is above a certain point and
- rising, you certainly don't want to add more fuel!
-
- --
- ***************************************************
- Prof. William W. Armstrong, Computing Science Dept.
- University of Alberta; Edmonton, Alberta, Canada T6G 2H1
- arms@cs.ualberta.ca Tel(403)492 2374 FAX 492 1071
-