home *** CD-ROM | disk | FTP | other *** search
- Newsgroups: comp.ai
- Path: sparky!uunet!til!til!tanaka
- From: tanaka@til.til.com (Yokichi Tanaka)
- Subject: Re: Fuzzy Control
- Message-ID: <1992Jul22.014706.24354@til.til.com>
- Keywords: fuzzy
- Sender: tanaka@til (Yokichi Tanaka)
- Nntp-Posting-Host: til
- Organization: Togai InfraLogic, Inc.
- Date: Wed, 22 Jul 1992 01:47:06 GMT
- Lines: 153
-
-
- In article <1992Jul9.141506.9440@tamsun.tamu.edu>, glass@eemips.tamu.edu
- (Emily Glass) writes:
-
- |> 1) Appropriateness of the fuzzy OR defined as MAX:
-
- |> variable, speed. Suppose in our fuzzy control rules we have need for the
- |> following membership sets: VERY_FAST, FAST, MEDIUM, and SLOW.
- |>
- |> ___________ ___________ ___________ ___________________
- |> | \ / \ / \ /
- |> | SLOW \/ MEDIUM \/ FAST \/ VERY_FAST
- |> | /\ /\ /\
- |> __|___________/__\___________/__\___________/__\___________________
- |>
- |> Now also consider that one rule will be testing if the speed is
- |> `fast or very fast.' We could design a membership function called
- |> FAST_OR_VERY_FAST that looks like:
- |> __________________________________
- |> /
- |> /
- |> / FAST_OR_VERY_FAST
- |> ____________________________/_____________________________________
- |>
- |> But since we already have FAST and VERY_FAST, one would think that we
- |> could just use these with an `OR' in the rule:
-
- Potentially dangerous assumption.
-
- |> if FAST OR VERY_FAST, then ...
- |> as opposed to
- |> if FAST_OR_VERY_FAST, then ...
- |>
- |> Conceptually these should be the same thing.
-
- Maybe, maybe not ... depends on your system.
-
- |> FAST OR VERY_FAST
- |> ___________ ___________________
- |> / \ /
- |> / \/
- |> /
- |> ____________________________/_____________________________________
- |>
- |> I realize that one way to get around this would be to redefine
- |> the overlap between FAST and VERY_FAST as follows:
- |>
- |> FAST VERY_FAST
- |> __________________________________
- |> / /\
- |> / / \
- |> / / \
- |> ____________________________/_____________/______\________________
- |>
- |> But then, this removes some of much flexibility for adjusting
- |> the FAST and VERY_FAST membership functions to improve control.
-
- Right. The better way (assuming your FAST and VERY_FAST are working well)
- would be your original idea of defining a fifth membership function
- FAST_OR_VERY_FAST, IMO.
-
- |> Has anyone addressed this issue? What is physically happening in the
- |> fuzzy-OR case?
-
- You came up with a very specific case. In your specific case, it would
- seem like FAST "OR" VERY_FAST should yield max(1, FAST + VERY_FAST).
- (... which does result in a "clean" MBF with no dips.)
- However, in many other cases, you want the fuzzy-OR to behave as a max
- function and not an additive function. One such case might be if you
- substituted FAST with "AROUND_6_FEET_TALL" and VERY_FAST with
- "AROUND_7_FEET_TALL". In such a case a 6-footer will have degree of
- membership of 1 for "AROUND_6_FEET_TALL" OR "AROUND_7_FEET_TALL", as
- would a 7-footer. But do you want to assign someone 6'6" a degree of
- membership of 1? I think 0.5 would be more appropriate.
- I don't think one specific case should necessitate a change in how
- fuzzy logic works.
-
- In English, we tend to view FAST and VERY_FAST as quite similar,
- and as being continuous. If your example was modified to only have 2 input
- MBFs (membership functions), SLOW and FAST, as drawn below:
-
- SLOW FAST
- ___ ___
- \ /
- \/
- /\
- ___/__\___
-
- with the cross-over at 40 mph, then you might agree "SLOW" OR "FAST" should
- yield 0.5 and not 1.0 for 40 mph. (But 1.0 for 0mph or 100mph.)
- Or, better yet, replace SLOW with FAST_DESCENT and FAST with FAST_ASCENT.
-
- |> 2) Splitting compound rules into multiple rules.
- |>
- |> Now consider the Fuzzy System which includes the two rules
- |>
- |> If A and C, then D.
- |> If B and C, then D.
- |>
- |> Can we simplify this fam system by making the rule:
- |>
- |> If (A or B) and C, then D?
- |>
- |> It turns out that the answer is `No' when one uses the additive
- |> combination technique and `Yes' when one uses the maximum membership
- |> combination technique.
-
- Right.
-
- |> In some applications this accumulation
- |> of weights with multiple rules is desirable (for example, in the case of
- |> medical fuzzy experts), because one wants more weight for each
- |> conditional that supports a decision.
-
- Right. I think in almost all cases accumulation is more desirable.
-
- |> Kosko showed that a FAM system using the maximum combination technique
- |> loses sensitivity as the number of FAM rules increases (NEURAL NETWORKS
- |> AND FUZZY SYSTEMS, Prentice Hall, 1992). For this reason, in many cases,
- |> one would want to use the additive combination technique. However, if one
- |> also wanted a true fuzzy OR between rules, one is in a dilemma.
-
- Fuzzy-OR'ing rule outputs doesn't make sense to me. It might be just a
- matter of lingo, but I tend to use the term "fuzzy-OR" for premises (or
- antecedents or left-hand-side) only. Outputs are "accumulated" or
- "max'ed/union'ed"; the latter being the fuzzy-OR equivalent, I suppose.
- In any case, I have yet to see an application in which "max'ing" is
- preferred over "accumulating".
-
- |> One solution (although a complicated one) would be to use the maximum
- |> combination technique between rules with identical control membership
- |> functions, and then use additive combination technique to combine these
- |> results with the rest.
-
- I don't know if this has much advantage over a straight "accumulation".
- If output MBF (membership function) D fires 5 times (with values 0.5,
- 0.8, 0.8, 0.6, 0.7) and output MBF E fires once with value 0.8. I
- would like my output to be much closer to the center of D than with E.
- By max'ing for each output (or control) MBF, you would end up with
- D 0.8 and E 0.8. Assuming D and E are weighted equally, the output will
- end up exactly between the centers of D and E.
-
- As a side note, I am really curious to find out if there is enough interest
- to merit a newsgroup "comp.ai.fuzzy". If you want, send me e-mail so I
- can get a feel for the level of interest.
-
- Yoke
-
- ------------------------------------------------------------------------------
- Yokichi (Yoke) Tanaka Togai InfraLogic, Inc. "fuzzy"
- tanaka@til.com Irvine, California, USA
- uunet!til!tanaka 714-975-8522
- "The more you drive, the less intelligent you are."
-