home *** CD-ROM | disk | FTP | other *** search
- Newsgroups: comp.ai.neural-nets
- Path: sparky!uunet!haven.umd.edu!darwin.sura.net!wupost!gumby!destroyer!ubc-cs!unixg.ubc.ca!kakwa.ucs.ualberta.ca!alberta!arms
- From: arms@cs.UAlberta.CA (Bill Armstrong)
- Subject: Re: NN process emulator
- Message-ID: <arms.715880799@spedden>
- Sender: news@cs.UAlberta.CA (News Administrator)
- Nntp-Posting-Host: spedden.cs.ualberta.ca
- Organization: University of Alberta, Edmonton, Canada
- References: <1992Sep7.090318.8040@nuscc.nus.sg>
- Date: Mon, 7 Sep 1992 15:46:39 GMT
- Lines: 64
-
- kimhock%Solomon.Technet.sg (Teo Kim Hock) writes:
-
- >We have been working on using the feed-forward NN to emulate the
- >dynamics of chemical processes. It works well for simple processes but
- >fail to learn the process with non-linear time-delay, for example
-
- > y(t) = u(t+10-int(y(t-1)/.1)*0.2 + y(t-1)*0.8
- >where 0<y(t)<1.0.
-
- Are you missing a parenthesis ^ here? I assume so.
-
- >This process has time-delay that is a function of the output. We have
- >tried using feed-forward net to learn the dynamics of this process
- >without any sucess. Can anyone give us some advise on this. Thank you
- >in advance.
-
- There shouldnt be any problem in learning this function, provided you train
- a feedforward net that has as inputs
-
- y(t-1), u(t+1), ..., u(t+10)
-
- assuming that the u-values can be predicted. If you meant the prior
- u-values, then you would have to have
-
- > y(t) = u(t-10+int(y(t-1)/.1))*0.2 + y(t-1)*0.8
- >where 0<y(t)<1.0
-
- and use
-
- y(t-1), u(t-1), ..., u(t-10).
-
- I hope this helps, and apologize if I've not changed your problem
- correctly.
-
- If the dimensionality of the input is too high for your chosen learning
- system, try ALNs. If the ALNs do the learning OK, then I can tell you
- what additional steps to take to assure safety.
-
- To train the ALN, just run your system and produce 11-tuples, where
- each one is of the form
-
- y(t-1), u(t-1), ..., u(t-10), y(t).
-
- Depending on what the (known) function u is like, you may have to
- start simulations at quite a few values of y(0) to get enough training
- data. Embed these into an lf program with codings like 12:1 for u's
- and 20:3 for the y's, and, say 100 quantization levels each. You
- might find that some of your inputs could be removed if u is simple
- enough.
-
- You can get the software for ALNs from menaik.cs.ualberta.ca
- [129.128.4.241] in pub/atre27.exe for Windows 3.x and pub/atree2.tar.Z for
- Unix. I suggest you look at the Windows version in any case, as it
- has a lot of documentation and examples, and the Unix version is
- unadorned (and hence very portable).
-
- Good luck.
-
- Bill
- --
- ***************************************************
- 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
-