home *** CD-ROM | disk | FTP | other *** search
- Newsgroups: comp.ai.neural-nets
- Path: sparky!uunet!gatech!destroyer!ubc-cs!alberta!arms
- From: arms@cs.UAlberta.CA (Bill Armstrong)
- Subject: Re: Advice re suitabilty of ANN's for this application
- Message-ID: <arms.713898156@spedden>
- Sender: news@cs.UAlberta.CA (News Administrator)
- Nntp-Posting-Host: spedden.cs.ualberta.ca
- Organization: University of Alberta, Edmonton, Canada
- References: <14AUG92.09433373@wl.aecl.ca> <1992Aug15.051739.428@afterlife.ncsc.mil>
- Date: Sat, 15 Aug 1992 17:02:36 GMT
- Lines: 101
-
- hcbarth@afterlife.ncsc.mil (Bart Bartholomew) writes:
-
- >In article <14AUG92.09433373@wl.aecl.ca> harrisp@wl.aecl.ca writes:
- >>
-
- Details of nuclear reactor problem deleted...
-
-
- >>
- >> a) Problem / No Problem (minimal information needed)
- >>
- >> This would be what we would consider successful, but not
- >> very satisfactory solution to the problem
- >>
- >> b) Problem Identification (ie there is a problem due to the combined
- >> effects of high dissolved oxygen and
- >> low pH {this type of information to
- >> have been obtained from the training
- >> sets} )
- >>
- >> This would be both successful and satisfactory if achieved.
- >>
- >> c) Problem Identification and/or Solution Proposal
- >>
- >> (ie. the network identifies the problem and proposes
- >> a method, based on past experience, for resolution
- >> of the problem)
- >>
- >> This would be the ideal solution to the problem.
- >>
- >>The main desire in this program is provide an automated method for dealing
- >>with the data to minimize the need of the reactor operators to concern
- >>themselves with the chemistry of the heat transport loop, and focus even
- >>more of there attention of reactor operating conditions and safety (which is
- >>already there primary responsibility, they would prefer not to have to worry
- >>to much about the chemistry end of the heat transport)
-
- > Even as I type this, without having read the remaining
- >articles, I just *know* that Prof Armstrong is similarly entering
- >a reply. The thing is, I think he will be right. This type of
- >problem looks like a natural for ALNs. The CoDomain feature
- >of the ALN is the best approach I know of for the answer to
- >the questions
- > 1: Is there a problem?
- > 2: If so, what is it?
- >The answer to 2 would presumably lead easily to the required
- >corrective action.
- > I'm also sure that Prof Armstrong would emphasize
- >that the network output should be construed *only* as advisory,
- >perhaps spotting a problem earlier than otherwise, perhaps
- >confirming a decision that the operators would make anyway.
- > BackProp could probably answer #1, but I don't see how
- >to get it to answer #2, unless you have a series of small nets
- >feeding a final decision, and then you see which of the smaller
- >nets set things off.
- > Prof Armstrong, it looks like this might be the problem
- >you have dreamed of.
- > Bart
- >--
-
- Hi Bart,
-
- Here's the reply I wrote. You were perfectly right about its
- existence (but it didn't get posted due to a technical problem), and
- now we can all check to see if you have mastered my internal
- state-transition and output functions too.
-
- I think your reply raises excellent points that I agree with.
-
- Bill
-
- PS Thanks for the nice words about ALNs.
-
- *******************
-
- A priori, it seems like if you were going to use a neural net, you
- would have to generate additional data to
-
- 1) cover all possible bad events,
-
- 2) complete the data from the events you do have where some really bad
- situations don't show up because the operators prevented it from
- happening.
-
- I believe you will have to use all sorts of a priori knowledge,
- perhaps indicating the need for a model. Because the costs of a
- mistake are probably very high, I would also advise against using any
- kind of trained neural network that could produce any wild value that
- differs significantly from values one would expect based on the sample
- of points tested. This includes networks trained with BP and ALNs that
- use random walks for encoding reals.
-
- Here is a case where careful analysis is called for, where the
- interactions of cause and effect are critical, and where blindly
- applying a neural net to your data could result in disaster.
-
- --
- ***************************************************
- 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
-