home *** CD-ROM | disk | FTP | other *** search
- Newsgroups: comp.ai.fuzzy
- Path: sparky!uunet!cs.utexas.edu!usc!elroy.jpl.nasa.gov!ames!ads.com!marcel
- From: marcel@ADS.COM (Marcel Schoppers)
- Subject: Re: What does "AND" mean?
- Message-ID: <1993Jan22.194643.19867@ads.com>
- Summary: normal distributions, etc
- Sender: Marcel Schoppers
- Organization: Advanced Decision Systems, Mtn. View, CA (415)960-7300
- References: <1993Jan21.225423.25301@netcom.com> <1993Jan22.154705.11200@unocal.com>
- Distribution: na
- Date: Fri, 22 Jan 1993 19:46:43 GMT
- Lines: 58
-
-
- Let me propose a non-fuzzy way of thinking about what the fuzzy AND "should"
- do. (Get your wooden stakes and bibles out...)
-
- The human population exhibits normal distribution on every property people
- have examined, including intelligence (I mean IQ), opinions of looks, etc.
- So suppose I decide to calculate person X's rating on the beauty scale as the
- probability that a randomly chosen subject will think X is beautiful. That
- probability IS that person's fuzzy membership of the "beautiful people" set.
- Then I collect such probabilities for a large number of X's, and I'll get a
- normal distribution: probability of perceived beauty is the independent
- variable, number of people having that beauty rating is the dependent variable.
- OK, let's divide the vertical axis (number of people) by the total number of
- people rated, so the vertical axis becomes the probability that a randomly
- chosen person will have a given beauty rating.
-
- I can do the same thing with people's nice-ness, and get a second distribution,
- (presumably) orthogonal to the first, i.e. now I've got a 2-D normal
- distribution with two horizontal axes; the original distributions were seen
- to be "projections" of this 2-D distribution onto one of the two axes (i.e.,
- integration over the other axis).
-
- Now I want to rate someone as a possessor of both beauty and nice-ness. The
- obvious way to do that is to
-
- 1. rate that person on the two characteristics, thus finding their
- point on the 2-D distribution,
-
- 2. construct an "evaluation axis" that is at 45 degrees to the original
- axes, thus weighting beauty and nice-ness equally,
-
- 3. perform a projection/integration of the 2-D distribution onto the
- evaluation axis, thus getting a new 1-D distribution (which happens
- also to be normal)
-
- 4. along the way we will also be projecting the candidate's point onto
- the evaluation axis, and so will get both a rating on that axis
- (horizontal) and a corresponding probability that a random person
- has that rating (vertical).
-
- 5. the evaluation axis, running from [0,0] to [1,1], has length
- sqrt(2), and the mode of the distribution will occur at its mid-
- point, so let's re-calibrate it to run from 0 to 1.
-
- The interesting thing is to see what rating the person gets on the evaluation
- axis. This is in fact trivial to figure out, we get
-
- pos-on-evaluation-axis = (X + Y) / 2
-
- and I would propose this as being one possible "beauty + nice-ness" rating.
- (If you're thinking that YOU wouldn't be interested in someone who had beauty
- = 1 and nice-ness = 0, I'd argue that maybe your personal opinion doesn't
- matter, many men WOULD be so silly, and I'm talking about distributions over
- populations.)
-
- marcel
-
-
-