home *** CD-ROM | disk | FTP | other *** search
- Path: sparky!uunet!stanford.edu!bu.edu!wupost!darwin.sura.net!paladin.american.edu!auvm!DRETOR.DCIEM.DND.CA!JEFF
- From: jeff@DRETOR.DCIEM.DND.CA ((Jeffrey C. Hunter))
- Newsgroups: bit.listserv.csg-l
- Subject: Re: Jeff's Loose control system
- Message-ID: <9211070148.AA26354@dretor.dciem.dnd.ca>
- Date: 7 Nov 92 00:48:34 GMT
- Sender: "Control Systems Group Network (CSGnet)" <CSG-L@UIUCVMD.BITNET>
- Lines: 74
- Comments: Gated by NETNEWS@AUVM.AMERICAN.EDU
- X-Mailer: ELM [version 2.2 PL9]
- In-Reply-To: <9211060115.AA06523@ben.dciem.dnd.ca>; from "William T. Powers"
- at Nov 5, 92 3:45 pm
-
- [From Jeff Hunter (921106-A)]
-
- > Re: Bill Powers (921105.1430)
-
- Great. I was afraid when I posted that you would either agree,
- or disagree with the post. Since you've done both I can be happy :-)
- I'll try and state my points again more carefully. If I do it
- right you should reply "yeah. that's trivially obvious".
-
- First your comments on the "wavers" program. (Which by-the-way
- are pretty perceptive.)
-
- > I've got your program to run, and think I understand it now. By the
- > way, Turbo C doesn't like long integers as indexes to arrays
-
- Ta. That bin stuff didn't need to be "long" anyways.
-
- > The environmental part of the loop is a simple integrator.
-
- Yup. It's just one-dimensional brownian motion. Fortuitously
- Hans Blom is also into the topic.
- The default change-per-tick is +-4 (just to be greater than 2).
-
- > The perceptual function and comparator are in "if (cont_tot > 0)"...
-
- Yup. Reference is zero. Linear input function. No delay.
- The comparator is actually *trinary* since it can produce <0, >0,
- or =0. This point is obscured by the code, and it makes an analysis
- of the beast a little tricky.
-
- > The steady-state loop gain of these control systems is nominally
- > infinite.
- Great. Glad you agree.
-
- > ... The reason is the integration in the environmental part of
- > the loop.
- Not exactly. The reason is that the long term average of the
- disturbances is zero. If I had biassed the random numbers so that
- there were 3 increases for each decrease then the control would
- disappear.
-
- This is the main point of "wavers". The effective loop
- gain of the control can be varied wildly merely by applying
- a different "pattern of disturbances". For example in the
- "bias" case the disturbances were +5,-3 rather than +4,-4.
-
- > ... It is
- > impossible to estimate the effective loop gain except for the biased
- > case, where a constant of 1 was added on each iteration. Over 10,000
- > iterations, this should have added 10,000 to the value of bias_tot
- > (and a lot more to the squared error); instead, bias_tot changed by
- > only 104, or about 1% of the expected value. This yields an effective
- > loop gain of about 100.
-
-
- Forgive me if this point (effective loop gain depends on the
- disturbances) is obvious. However a lot of discussion on the group
- appears to ignore this.
- I think that the one of the points of the "leaning on the
- environment" posts was that a perfectly valid control system in one
- setup could fail spectacularly if the disturbances changed in
- pattern even if they stayed the same in magnitude.
-
-
- > If you intended to portray a loose (low-loop-gain) control system,
- > you didn't succeed.
- I actually wanted to emphasise the difference between
- moment-by-moment gain (in which "cont" is totally swamped by
- almost every disturbance) and the long-term gain (in which "cont"
- has infinite gain).
-
- ... Jeff
- --
- De apibus semper dubitandum est - Winni Ille Pu
-