(Translation into English by
R.Girvin)
This
optimization task was easily manageable either numerically (see Etude 2) or analytically
(see Etude 7). In those, we learned how to maximise the volume of a bucket
given a chosen manufacturing technology: cutting a sector from a metal disk and
folding the remainder into a cone. But that way the bucket lost not only
optimality, but also the status of bucket: it turned into a fire bowl! It was
also impossible to put on a floor, and would be useless for any purpose except
as a Vietnamese hat. In our rush for Number (the volume of the bucket) we have
lost Essence (the function of a bucket, its capacity for conveniently carrying
liquid).
The
author recalls a cartoon film based on a Norwegian fable about a goat that
learned to count, and decided to count all the farm animals: "I am one;
the hen is two; the pig is three, etc." But as all the animals were moving
around, the process soon turned into uproar. "Oh, have you counted
me?!" "Keep still, or else..!" But everything ended happily,
both for the goat and the fable.
In
this story, as with any other fairy tale, there's a deep wisdom. It's necessary
for us to count things, but as we interact with them we can run into a
conflict. Not only does Nature dislike sharp corners, but also enumeration,
which in a number of cases simply kills her. This can be observed not only in
biology and physics, where the tools of knowledge frequently unrecognizably
spoil the object of research, but also in computer science. This applies not
only when applying computers to natural sciences applications, but also when
applying computers to computers.
To
establish a diagnosis, a doctor has no need to know the exact numerical value
for a patient's body temperature (36.6C, 38.9C, etc). It's enough to express
the thermometer indications by ranges, on which 'medics' have agreed
beforehand: "low", "normal", "raised", "mild
fever", "high fever". The borders of these estimates, though
precisely fixed, are nevertheless imprecise 'fuzzy'. This comes not merely
from modern representations in terms of fuzzy set theory (FST) but also in
terms of practical reality: thermometer error, measurement technique, and so
on. A graduate of a medical school can tell you, without even pausing to think,
where the division between high and very high temperature lies. The skilled
doctor can decide this without consciously thinking of the number, while
diagnosing better than the beginner.
Even
the parameters of patients that are expressed not in real numbers, but in a
kind of binary Wasserman test positive/negative, Koch bacilli present/not
present, HIV positive/negative, etc. have 'fuzzy' borders. Laboratory
analysts know this well. If you glance in any therapeutic reference book, where
the symptoms of illnesses are described, normally you won't see hard numbers
for body temperature, arterial pressure, haemoglobin count, etc. You'll simply see
a broad categorisation: "is raised", "is lowered", and so
on. Perhaps this is why programs that output a diagnosis, based on patient
parameters entered into a computer, have not received wide practical
application. One of difficulties in this area is translation of the parameter
(a number) into a symptom (a category).
There's
a generally accepted identification of three recent revolutions in programming:
structural, object-oriented, and visual. But this revolutionary zeal focused more
on the programs (the art for its own sake) and virtually neglected the point of
programming: the models of the real world, the properties, and the events
programs are simulating. Moreover, in retrospective we can see certain
counter-revolutions: for example, the decline of analogue computers in favour
of digital engineering. Recently, though, there has been a renaissance through
the revival of principles of analogue modelling on modern digital computers
(for example, the MathConnex environment included with Mathcad 7 and 8 Pro: see
Appendix 7). It can similarly be seen in the technology of visual programming,
where the former analogue control elements adders, integrators etc. are
reconstructed.
But
the virtual nature of these neo-analogue machines means also their strict
determinism, and this brings not only positive, but also negative consequences.
The principles of FST have been in programmers' hands for a long time
literally. The computer mouse reacts to two events: a click, and a double click.
What, actually, is the difference one double click and two single? The duration
of the pause between clicks. Expressed in non-technical language "very
short", "short", "long", etc.) this is a typical
example of sets with 'fuzzy' borders. How much time is needed to increase a
short pause to turn it into long, so that the double click breaks up into two
single? How many handfuls of grain make a heap?
At
the peak of the structural programming revolution, when in all the programming
'temples' the keyword GOTO was anathema, you frequently heard such statements
as: "It's practically impossible to teach good programming to students
originally taught BASIC; as potential programmers they're crippled
intellectually, without hope of a cure." There were also broader warnings
such as "Caution! Employment in programming is a dead end career. Don't
think that having learned to program, you'll achieve anything in life".
It's
as if traditional programming forces the programmer to look at the Technicolor
world through monochrome glasses: a binary variable can take only two values
(yes/no), and a real variable, values anywhere in a stipulated range strictly
determined by the length of the mantissa. The truth lies somewhere between. The
extreme points of view aren't useless they are like book-ends that stop it
sliding beyond defined extremes. The intermediate truths are termed 'fuzzy'. So
one might say, "If you want to learn about the world (which is fuzzy, and
unquantifiable) and to deal with it, beware of the traditional programming
languages and mathematical programs with their strict determinism."
But
we shall return to our problem of the fire bucket, and try to solve it using
FST techniques and the opinions of people (who thank goodness! don't need
such a rigmarole to use a simple device to escape being roasted).
Let's
carry out an original poll and learn as much as we can about the parameters of
an optimum fire bucket: its most convenient geometry (radius of the base of the
cone to the height) and its optimum volume (the weight of the filled bucket).
These can then be expressed as FST rules. How much water can you add to a
bucket before it turns from light to heavy? How much can you increase or reduce
the radius, or the height, of a bucket before it stops being convenient? These
statements are typical definers of fuzzy sets. In the Mathcad environment, as
well as in other popular packages, there are no variable types for storing such
objects. But nevertheless we'll try to solve the given problem (see figures 6.41, 6.42, 6.43,
6.44 and 6.45 a solution
developed with B. Usyenko). Let's break it down into steps.
Fig. 6.41. Optimum radius of fire bucket
Step
1 (fig. 6.41). The matrix μr stores people's views about
the optimum (most convenient) base radius r of a conical fire bucket, expressed
in millimeters. This data could be gathered by making buckets of various
geometries, giving them to people to try out, and then asking for estimates on
a scale:
·
convenient (1);
·
more convenient than inconvenient (0.67);
·
more inconvenient than convenient (0.34);
·
inconvenient (0).
It
would be possible to have more options within the range 0-1. In step 1 we have
a limited familty of points, but these also could be increased; there are as
many opinions as people. Readers can ask all their friends, and add new columns
to the matrix [μr]
Step
2. The survey data is processed by the least squares method (see Etude 4). We
can see that the data approximately fits a normal distribution curve (see figs. 6.41 and 6.42). The idea of
a 'membership function' μr for the radius of the bucket is one of the basic concepts of FST. In
normal mathematics it would be considered that a certain size either belongs,
or does not belong, to a particular set; in FST it's permissible to say that
the size belongs to the set to some extent
(so many percent).
Step
3. The statistical processing is completed and plotted.
Fig. 6.42. Optimum height of a fire bucket
Steps
4-6 (fig. 6.42) repeat steps 1-3, but for a second
parameter of the bucket, its height.
Fig. 6.43. Optimum volume of a fire bucket
Steps
7-9 repeat steps 1-3 and 1-6 for the third important parameter of the bucket,
its volume (or weight they're proportional). This is based on human
estimates:
·
bucket is light (1);
·
bucket is more light than heavy (0.67);
·
bucket is more heavy than light (0.34):
·
bucket is heavy (0).
The
survey data is processed as before, but using a "one-sided"
cumulative distribution curve (see item 9 in a fig 6.43).
(When designing technical systems, such parameters wouldn't be based on a
survey but on figures provided by experts to the decision-makers).
Fig. 6.44. the "Rotated" optimum fire bucket
Step
10 in fig. 6.44 is the nucleus of the solution of our:
in this, a two-parameter membership function is generated by the merging (akin
to multiplication) of two single-parameter membership functions.
In
FST there are no traditional concepts of addition, subtraction, multiplication
etc.(those represented by the mathematical operators "+",
"-", "×" etc. in the Mathcad environment). In FST,
multiplication (crossing of sets logical AND) is replaced with an operation
of searching for a minimum; and addition (merging of sets logical OR) by
searching for a maximum. The mathematics of precise sets is a special case (a
subset of the mathematics of fuzzy sets) where these operators/functions are
genuinely equivalent. This means that in the Mathcad environment, where there are no built-
in AND and OR operators, we can create equivalents using the min and max functions we described in
Etude 3.
In
our task the membership function mu_rh is generated by the fuzzy addition (min) of the functions mu_r,
mu_h and mu_v. That is, the fuzzy set "the convenient bucket " is the
intersection of three other fuzzy sets: "convenient radius of bucket"
(step 1), "convenient height of bucket" (step 6) and "a not
heavy bucket" (step 9).
Step
11. The top of the 'mountain' the surface plot of function mu_rh is the
point where the parameters of most convenient fire bucket lie.
Fig. 6.45. Designing the optimum fire bucket
Step
12 (fig. 6.45). Searching for the maximum of the
function mu_rh in the Mathcad environment can be done in various ways (see
Etudes 2 and 3). We'll proceed this way: we'll imagine metal disks of various
radius R (from 10mm-500 mm with a step of 1 mm) and cut each size into 2 to 10
identical buckets. This gives a large variety of bucket sizes and geometries,
and we'll consider the optimum (the most convenient) bucket to be the one for
which the membership function mu_rh is maximised.
Even
the reader who hasn't been paying much attention will notice some discrepancies
and simpiflications in our analysis of this task. Three of them are:
·
Unlike our theoretical bucket, a real one is
never filled up right to the brim ("what counts as full?" defines
another fuzzy set).
·
The author has only loosely defined, and often
confused, such concepts as "volume", "weight", and
"weight of bucket".
·
We haven't taken into account the weight of an
empty bucket, and also the material from which it is made.
However it is merely necessary to glance again at the diagrams illustrating fuzzy sets in figs. 6.41, 6.42 and 6.43 to understand a major virtue of applying FST to decision-making tasks. Our solution works in isolation, so it's possible to express the essence of a task disregarding various minor variables: the density of water, weight of the empty bucket, degree of filling, etc. This feature is now realized, for example, in automatic control systems, where regulators based on FST rules are more 'attentive' to the basic signal and less susceptible to noise. It turns out, though it seems paradoxical, that the traditional 'precise' control algorithms qualitatively lose out to 'fuzzy' ones, or are their special cases. In the field of automatic control theory, a certain stagnation could be seen until recently, as any new algorithms couldn't be compared to the older Proportional Integral Derivative (PID) control algorithm. The principles of PID control can be seen in the procedure for a bank's credit check on a client applying for a loan. The banker, in assessing the decision, considers:
1) The sum of money in the client's account (this is the proportional
component: the richer the client, the larger the loan that can be offered);
2) The average sum in the account over, say, the last five years (the
integrated component; checking this ensures that the client didn't borrow a
million pounds the day before, to create the illusion of solvency); and
3) The rate of change of the account (the differential component, showing
whether the client's business is on the rise or decline);
It's
possible to take into account other factors, but three is a nice number!
The
PID-control algorithm has imperceptibly become something of a fetish. Fuzzy
management cames as a fresh breeze in the theory of automatic control, whose
basic rules are now open to review. To be sure, there are other opinions. Some
scientists believe that the use of FST in automatic control, and cybernetics in
general, is just replacing one uncertainty with another ("trying to stitch
soapsuds", the Russian expression goes) and that's all there is to it.
Sceptics explain away the observed improvement in control quality to our
devoting more attention to regulator technology (as if the ritual of attention
could improve performance). Besides some researchers believe that FST, since it
dates back 30 years (it was devised by the Iranian-American Dr Lotfi Zadeh at
UC/Berkeley in the 1960s) is old and best forgotten.
Actually,
the skip from a task about a tripartite duel to one about an optimum bucket wasn't
entirely casual. Traditionally, precise sets are illustrated by circles with
sharply delineated borders. Fuzzy sets are
drawn as circles formed of separate dots, with a high density at the centre and
thinning to zero (as if evaporating) towards the edges.
Such 'fuzzy set' images can be seen on a firing range wall where targets
are hung. The bullet traces form a probability distribution, whose mathematics
is well-known. It appears that the theory for
working on fuzzy sets, as probabilistic distributions, has already existed for
a long time...
We
keep talking about fuzzy sets. But are they mathematically actually sets?
To be consistent, it's necessary to ascertain that the fuzzy set has elements
(fuzzy subsets, fuzzy sub-subsets, etc). We'll return to a classical example: a
heap of grain. An element of this fuzzy set will be, say, a million grains. But
a million grains is not a precise element: it's a fuzzy subset. If you count
out grains, whether manually or automatically), it's no wonder that you might
mistake, say, 999 997 grains for a million. In FST terminology, you could say
that the element 999 997 has a 'membership value' 0.999997 for the fuzzy set {a
million grains}. Besides, even "a grain" is not a precise element,
and is another fuzzy subset: it might be a high-grade grain, but there are also
underdeveloped grains, grain fragments, and bits of husk. Depending which way
you decide, you might count one grain as two, or vice versa.
The
fuzzy set isn't very easily accommodated in the digital computer with classical
data constructs: elements of a file (a vector) should be new files of files
(composed of vectors and matrices, in the case of Mathcad). The classical
mathematics of precise sets (number theory, arithmetic, etc) is rather like a
hook, with the help of which we anchor ourselves on the slippery and fuzzy
environmental world. And a hook is a rough implement, quite often damaging what
it's embedded in. The terms describing fuzzy set membership values (there are
plenty in this book and others "much", "some", "a
few", and so on) are difficult to make into programming statements because
they're contextually dependent. It's one business to say, "Give me a few
sunflower seeds" to someone with a cupful, and another to say the same to
a driver hauling a lorryload.
Whether
it is possible to see a certain crisis in theory and practice of programming
connected with contradict between precise structure of the programs (data) and
fuzzy world? Is it necessary to develop fuzzy programming languages for implementation
fuzzy algorithms that accommodate fuzzy data? Opinions vary. In the author's
view, programmers (and they have the last word) have learned poorly how to cram
the fuzzy world into the strictly determined computer. Figs. 6.41,
6.42, 6.43, 6.44
and 6.45 are an example.
We can imagine A and B
sitting down to pass the time of day in our case, not on a simple pipe, but
on a fire bucket. We'll now explain the 'universal' meaning of these
parametrical factors A and B (more correctly A_r, B_r, A_h, B_h, A_v and B_v)
included in the approximation expressions of stages 2, 5 and 8 (fig. 6.41-43).
Viewing
the fuzzy set as a statistical distribution, A represents the 'spread' of the
distribution, and B the parameter value giving the peak of the distribution.
Imagine we carry out a global statistical experiment. We measure for all adult
humans some parameter weight, height, intelligence, etc whether represented
numerically or estimated by some criterion (genius, talented, very clever,
simply clever... down to idiot). The data points we shall transform into a curve
(see sections 1-3 in figures 6.41, 6.42
and 6.43) where the X-axis is the measured human
parameter, and the Y-axis the percentage of people with a given value of that
parameter. We carry out this statistical exercise separately for men (M) and
women (F). What do we get?
Case 1: AM = AF, but BM > BF:
Curves
turn out like this for parameters such as height, weight, strength of
muscles, etc pertaining particularly to men rather than women. This is related to evolution: sexual dimorphism is
related to the general trend in species size. If the male body is larger than
the female, the next generation tends to be larger. With spiders, for example,
the male is much smaller than the female, and so thank God spiders are
smaller now than the geological past.
Case 2: AM < AF, but BM = BF:
The
curve looks like this for a parameter intelligence whose importance applies
equally to both sexes, and has not varied for millennia. Many consider it quite
proven, that if modern humans were suddenly stripped of education and culture,
we would be no more intelligent (maybe even less so) than the Ancient Greeks.
It's not known why, but there is a greater spread for men: there are more men
of genius (Nobel prizewinners, inventors, writers and artists at the right
edge of the diagram) only because there are also more men of extremely
subnormal intelligence (the left edge). The average woman is cleverer the
average man: the centre of the "female" distribution is raised at the
expense of a smaller spread. The areas of both curves is, however, identical:
no offence intended to anyone, but we believe that the Lord God or Nature
(whichever you believe in) has shown consideration in distributing intelligence
equally to both halves of humanity.
We
can try to apply the theory of fuzzy sets, as described in Etude 3, to the
problem of computer piracy. The law divides all people into two precise sets:
{legal users of programs} and {illegal users pirates}. In real life, it's
much more complex: computer users fall into two fuzzy sets, with a great
variety of motives. At one extreme are people trading 'black' disks at
Gorbushka (the left-hand axis of the diagram above); at the other are those who
work only with legal copies and have never broken any license agreement (the
right-hand axis). Again, it's possible to construct a humpbacked statistical
curve, this time above an axis "sinner ... saint" describing the
condition of the computer market in a particular country. Where does the
maximum of this curve lie, and how is it shifting with time?