An Introduction To FUzzy Control Systems

Fuzzy Logic

Introduction To Fuzzy Logic
Fuzzy Membership Functions
A Simple Fuzzy Control System

Why Fuzzy Logic?

Fuzzy logic is an interesting subject to contemplate.

The name implies that you have to be a bit fuzzy headed to use fuzzy logic!
One main claim to fame for fuzzy logic is that it is useful in designing control systems - in spite of the fact that design techniques for control systems abound.
For any number of reasons, fuzzy logic and fuzzy control system design methods have become a controversial topic.

Why Fuzzy Logic? - What the Experts Have to Say!

Fuzzy theory is wrong, wrong, and pernicious. What we need is more logical thinking, not less. The danger of fuzzy logic is that it will encourage the sort of imprecise thinking that has brought us so much trouble. Fuzzy logic is the cocaine of science.

Professor William Kahan
University of California at Berkeley

"Fuzzification" is a kind of scientific permissiveness. It tends to result in socially appealing slogans unaccompanied by the discipline of hard scientific work and patient observation.

Professor Rudolf Kalman
University of Florida at Gainesville

As complexity increases, precise statements lose meaning and meaningful statements lose precision.

Professor Lofti Zadeh
University of California at Berkeley
quoted in McNeill & Freiberger, p 43

So far as the laws of mathematics refer to reality, they are not certain. And so far as they are certain, they do not refer to reality.

Albert Einstein
in Geometry & Experience

Where are the Fuzzy Systems? - What are They Doing Now?

Shifting gears in automatic transmissions in cars
Focussing your camera and camcorder
Running the cruise controls on many cars
Running subways in Japan
Controlling dishwashers and washing machines
and many, many more!

Fuzzy Logic - Background- Your Goals

Goals of this lesson are:

Given a system to be controlled,

Design a simple fuzzy controller using:

Membership functions for the controlled variable,

A Simple Rule Base,

The Centroid Method for Computation (Defuzzification) of the Control Effort

Introduction/Background

Study of Fuzzy logic is a study of a kind of logic. We are all familiar with some of the principles of logic. Fuzzy logic builds on traditional logic and extends traditional logic so that fuzzy logic can solve some long standing problems in traditional logic.

Like many other things, it all started with the ancient Greeks, who first formalized logic. Aristotle may not have been the first to assert the one statement at the foundations of traditional logic, but he certainly is the earliest on record to have done so. In one form or another, Aristotle is credited with claiming:

A thing either is or it is not.

The essence of the claim is that there is no other possibility, and in particular, there is nothing between the two possibilities, and it has come to be known as the law of the excluded middle. Let's look at some examples.

Here are some examples of the law of the excluded middle.

A statement is either true or it is not. In other words a statement is either TRUE or FALSE.
A thing is either living or dead.
A being is either intelligent or not.

Let's examine those examples of the law of the excluded middle.

A statement is either true or it is not. In other words a statement is either TRUE or FALSE. That seems pretty much the case.
A thing is either living or dead. Now we have viruses that lie somewhere between the world of inanimate molecules and living things.
A being is either intelligent or not. But there are serious problems with chimpanzees that can communicate with sign language.

And that leaves us with only the first statement that we can be comfortable with.

Now consider this rather horrible little statement. If all statements are either TRUE or FALSE, which is the case for this statement?

This statement is FALSE.

Think carefully about this statement. Is it TRUE or is it FALSE? Consider this then.

If the statement above is TRUE, then it must be FALSE because that's what the statement says.
If the statement above is FALSE, then it must be TRUE.

Clearly we have a problem here, and it is not trivial. This is a statement that cannot be either TRUE or FALSE. But what else can it be?

Fuzzy logic and fuzzy set theory eliminate paradoxes like the statement above by
assigning a continuum of truth values to statements.

Instead of assuming a statement must be true or false fuzzy logic permits degrees of truth.
For example, consider the statement "Pat is tall". Suppose Pat is right around six feet tall. Suppose, also, that we take tall to mean six feet or taller.
If Pat is .0001 inch under six feet, then Pat is not tall.
In other words, the statement "Pat is tall" would not be true.
If we were to increase Pat's height by .0002 inches, then we would consider Pat tall, and the statement "Pat is tall" would be true.
You would be right if you thought that was a pretty fine distinction to make.
If you didn't think it was a pretty fine line to draw consider what happens if Pat is above six feet in the morning (After sleeping anyone is a little taller.) and just under six feet in the evening. (After standing a day, anyone is a little shorter.)
Do you think that Pat is tall in the morning and not tall in the evening?

There's an interesting situation here.

The problem doesn't lie in the precision of the height measurement.
Actually, the more precisely we know the height, the more difficult it is for us to accept the sharp, arbitrary classification.
It's probably that our concept of "tall" is inherently fuzzy.

In life we sometimes must accept shades of truth, or grades of truth.

In formal logic and in electronics we deal with things that have only two truth values, zero (0 or FALSE) and one (1 or TRUE).
When we design logic circuits we use zeroes and ones for FALSE and TRUE. Let's keep the idea of zero for FALSE and one for TRUE.
However, let's consider the possibility of partial truth. If Pat is six feet tall, then the statement "Pat is tall" might have a truth value of .9.
Another way of stating this might be to say "Pat is a member of the set of tall people with a membership grade of 0.9."

The concept of grades of truth or grades of membership in sets is something that, at this writing, does not seem to come naturally to many even though it is something that many of us actually use in making decisions in our life.

We choose a place to live or a career to follow not because of yes or no factors.
Rather, we tend to think in terms of the truth value of statements like "This place is a good place to live." or "This career is one for which I am well suited".
We use fuzzy truth measures in our lives, often because we have no choice.
However, engineers are now choosing to incorporate fuzzy logic rules in control of trains, washing machines and cameras.

If the underlying reasoning seems a little obscure, then it is time to begin to look at how ideas about fuzzy logic get to be incorporated into products that work better. That's a pretty good distance to go, so let's get started.

We need to formalize the intuitive ideas about partial truth and partial membership that we uncovered earlier.

Fuzzy Membership Functions

An appealing way to represent some ideas about truth of statements and membership functions is to draw a graph of a membership function.

Shown below is a graph of a membership function for a traditional logic case where the graph is a graph of the truth value of the statement:

X is Tall, assuming

Tall means more than 6 feet.

Next is a similar graph. It represents a membership function for a fuzzy case.

Shown below is a graph of a membership function for a fuzzy logic case where the graph is a graph of the truth value of the statement:

X is Tall, assuming

Someone more than 6 feet is tall.

In this fuzzy membership function you can give varying interpretations to what it means. Here are some examples.

For a person exactly six (6) feet tall, the value of the membership function is 0.5. That means:

That person has a membership value of 0.5 in the set of tall people.

In the usual case, that isn't possible. You're either in (value of membership = 1) or you are out (value of membership = 0).

The truth of the statement "This person is tall" is 0.5.

In the usual case, the statement is either TRUE (value of the truth of the statement = 1) or it is FALSE (value of the truth of the statement = 0).

This plot lets you compare the fuzzy membership function with the crisp membership function.

Membership functions are a useful and integral part of fuzzy control systems. However, in fuzzy control, it is often more convenient to use membership functions that are not smooth - like the one above. Instead, it is often the case that membership functions are composed of straight lines - because the computation of the membership value is much simpler in that case.

Here's a membership function of the type we want to examine. This particular membership
function can be used to determine the truth value of the statement:

The room is very hot.

Problem

P1 Assume that you measure the temperature of the root, and you get 80 degrees. Determine the truth value of the statement:

"The room is very hot."

Here's another kind of membership function. This membership function gives truth value for the statement:

The room is very cold.

Problem

P2 Assume that you measure the temperature of the root, and you get 60 degrees. Determine the truth value of the statement:

"The room is very cold."

Finally, here's a triangular membership function that gives truth values for the statement:

The temperature of the room is OK.

Problem

P3 Assume that you measure the temperature of the root, and you get 60 degrees. Determine the truth value of the statement:

"The temperature of the room is OK."

Let's review all of these membership functions. Here they are together. There's one interesting observation to make.

Two different membership functions can both have non-zero truth values.
Assuming a measurement of 80 degrees, the truth value of the statement

"The temperature of the room is OK." is 0.67. (The black function)

While, the truth value of the statement

"The temperature of the room is very hot." is 0.33. (The red function)

Fuzzy Control

Membership functions are a useful and integral part of fuzzy control systems. Let us examine a control problem that we can solve with a fuzzy approach. Let's take a look at a very simple fuzzy control system. We'll use the membership functions from the last section.

Here are the membership functions. They can be used to measure the truth value of the three statements:

The room is very cold.
The temperature of the room is OK.
The room is very hot.

And, for any temperature, one or more of these functions can have a non-zero value.

Now, to build/design a control system, we will need something else - something we haven't encountered before - a rule base. A rule base is exactly what the phrase implies. It is a set of rules for doing the control. The rule base tells you how to change the control effort. Here's an example rule base for a simple fuzzy control system.

If the room is very cold, then the control effort is high.
If the temperature of the room is OK, then the control effort is medium.
If the room is very hot, then the control effort is low.

Different colors are used here to show the correlation between the statements in the rule base and the function. The rule base consists of statements with two parts. Here's the rule base again.

If the room is very cold, then the control effort is high.
If the temperature of the room is OK, then the control effort is medium.
If the room is very hot, then the control effort is low.

The statements in bold text are the antecedents, and the regular text parts are the consequents in the rule base. In any event, this rule base - when implemented - will give us a fuzzy control system.

Now, let's examine the control algorithm for fuzzy control - proportional and others. Here's the algorithm.

Measure the variable to be controlled. Temperature would be an example of a variable to be controlled.
From the measured value, determine the truth value of all antecedents in the rule base.

NOTE: These truth values are usually referred to as the firing levels for the rules in the rule base.

From the firing levels of the rules in the rule base, defuzzify to obtain the level(s) of control effort.

We'll take these different steps up separately. The first step in the algorithm is:

Measure the variable to be controlled. Temperature would be an example of a variable to be controlled.

We'll assume that you have a way of measuring temperature and getting the resultant value into a computer program, say one written in C or Visual Basic.

The second step in the algorithm is:

From the measured value, determine the truth value of all antecedents in the rule base.

NOTE: These truth values are usually referred to as the firing levels for the rules in the rule base.

Problem

P4 Assume that you measure the temperature of the root, and you get 60 degrees. Determine the firing level of the antecedent statment:

"If the room is very cold, then the control effort is high."

Use this set of graphs for your calculation.

Problem

P5 Assume that you measure the temperature of the root, and you get 60 degrees. Determine the firing level of the antecedent statment:

"If the temperature of the room is OK, then the control effort is medium."

Use this set of graphs for your calculation.

Problem

P6 Assume that you measure the temperature of the root, and you get 60 degrees. Determine the firing level of the antecedent statment:

"If the room is very hot, then the control effort is low."

Use this set of graphs for your calculation.

Click on the box to enter your answer.

We need to go a little slower when it comes to computing the control effort.

We will assume that the control effort is represented with membership functions also. In particular, we need to have membership functions for the three different antecedents:

Control effort is high.
Control effort is medium.
Control effort is low.

Here are some membership functions we can use. These are not unique, but they will work for us.

L is for "Control effort is low" (red),
M is for "Control effort is medium" (blue).
H is for "Control effor is high" (black).

Now, we need an algorithm for defuzzification (and isn't that a great word?).

The defuzzification algorithm should use the firing levels calculated in the first steps of the control algorithm.
One way - commonly used - is to use the firing levels to weight the control effort membership functions.
In other words, we will first multiply the appropriate control effort membership functions by the firing levels.
Next, we add up the resulting functions.

Here are the membership functions for the antecedents again.

Go back to the problems above where you got the firing levels for each of the membership functions. The firing levels you should have found are as follows.

If the room is very cold - 0.333
If the temperature of the room is OK - 0.667
If the room is very hot - 0

Then, we are going to multiply these firing levels by L(C), M(C) and H(C) - the membership functions for the control effort. If we multiply the control effort membership functions by the firing levels, we will get the graphs shown below - which actually contains three graphs!

0.333 (firing level for very cold) * H(C) (high control effort)

That's the smaller triangular function at the right in the plot.

0.667 (firing level for OK) * M(C) (M(C) is for medium control effort.)

That's the larger blue triangle in the middle of the plot.

0.00 (firing level for very hot) * L(C) (L(C) is for low control effort.)

That's the black triangle of ZERO height. It's not really visible.

Here is the sum of those three functions - shown in orange.

We still have one unanswered question - and it is important.

How do we get from this composite function to a value for the control effort?
The commonest way to get a single number for the control effort is this:

Compute the center of gravity (centroid) of the composite function.

Good grief, now we have to remember how to compute a centroid! Well, here's the definition of the centroid of a function.

Centroid = Moment of Function/Area of Function

For the composite function, the area is the integral of the function over the entire function.
For the composite function, the moment of the function is the integral of the function multiplied by the variable.
In practice, the integral will only be taken over a finite interval since the function is non-zero only on a finite interval.

Now, if we can express the centroid as a function of the parameters of the membership functions, we have a chance to get a simple expression. Let's work on that.

The composite function (shown in orange in the plot above) is a weighted sum of the individual membership functions for the control effort (Low, Medium and High in our example).

Call the membership functions Cold(T) (Cold), OK(T) (OK) and Hot(T) (Hot).

The weights are the firing levels of the associated antecedent functions in the rule base.

Call the firing levels a_C, a_O, and a_H.

Now, we multiply the firing levels my the control effort functions.

Call the control effort functions L(C) (Low), M(C) (Medium) and H(C) (High).

Here's the expression for the composite function.

Composite function = a_CH(C) + a_OM(C) + a_H.L(C)

Next, note that each of the control effort membership functions is a triangle.
For each triangle, the height is 1, and we will call the width, W.

Then the area of each triangle is W/2.
Note, for our example functions, the width is 100 (percent of max control effort).

Now, we can compute the area of the composite function in terms of these areas.

Now, the area of the composite function can be computed. Note that the areas (integrals) of H(C), M(C) and L(C) all equal W/2, where W is the total width of the triangle.

Next examine the moment of a triangle function. Here's the derivation for the moment.

Get an expression for the function, Tri(C).

For C in the interval from L (Left/Low point) to M (Mid point), Tri(C) can be written as:

Tri(C) = (C - L)/(M - L)

For C in the interval from M to R (Rightmost point),Tri(C) can be written as:

Tri(C) = 1 - (C - M)/(R - M)

Now, integrate to get the moment.

The integration should be broken into two parts - two intervals - to simplify.

For C in the interval from L (Left/Low point) to M (Mid point), Tri(C) can be written as:

Tri(C) = (C - L)/(M - L)

For C in the interval from M to R (Rightmost point),Tri(C) can be written as:

Tri(C) = 1 - (C - M)/(R - M)

Now, integrate from L to M:

The result is: Moment_Left = (1/3)M² - (1/6)LM - (1/6)L²

Similarly, integrate the function over the right half of the interval:

The result is: Moment_Right = (-1/3)M² + (1/6)MR + (1/6)R²

Add the two to get the moment:

Moment = Moment_Left + Moment_Right
Moment = (1/3)M² - (1/6)LM - (1/6)L² + (-1/3)M² + (1/6)MR + (1/6)R²
Moment = (1/6)MR + (1/6)R² - (1/6)LM - (1/6)L²
Moment = (1/6)[R(M + R) - L(M + L)]

Example

E1 In the graph below, compute the moment of the Medium Control Effort Membership function (shown in blue as M(C)). Note, for this function:

L = 0,
M = 50,
R = 100

Putting the values for L, M and R into the expression above, we have:

Moment = (1/6)[R(M + R) - L(M + L)]
Moment = (1/6)[100(50 + 100) - 0(50 + 0)]
Moment = (1/6)[15,000] = 2500

Problem

P7 In the graph below, compute the moment of the High Control Effort Membership function (shown in black as H(C)). Note, for this function:

L = 50,
M = 100,
R = 150

Now, let's compute the centroid of the medium function, M(C). In the example above we found:

Moment = 2500

We can also compute the area as half of the product of the height (1) and the width (from 0 to 100):

Area = (1/2)[100 - 0] = 50

Then, we know that the centroid is given by:

Centroid = Moment/Area
Centroid = 2500/50 = 50

And, that should make sense to you.

Problem

P8 In the graph below, compute the centroid of the High Control Effort Membership function (shown in black as H(C)). Note, for this function:

L = 50,
M = 100,
R = 150

Now, let's go back to the control computation we were looking at earlier. We now have all the numerical results we need. Here is the composite function discussed earlier. The control effort is the centroid of the composite function.

Previously, we computed the firing levels of the three membership functions for Very Cold, OK (Medium) and Very Hot. Here are those firing levels.

0.333 (firing level for very cold)
0.667 (firing level for OK)
0.00 (firing level for very hot)

From these firing levels, we constructed a composite membership function for the Control Effort. That composite function is shown in the graph above. It is constructed as follows.

0.333 (firing level for very cold) * H(C) (high control effort)

That's the smaller triangular function at the right in the plot.

0.667 (firing level for OK) * M(C) (M(C) is for medium control effort.)

That's the larger blue triangle in the middle of the plot.

0.00 (firing level for very hot) * L(C) (L(C) is for low control effort.)

That's the black triangle of ZERO height. It's not really visible.

The sum of these three functions (Firing level x Control Effort Membership Function) is the orange function shown in the figure. The "orange" function - the composite function is given by:

Composite function(C) = 0.333*H(C) + 0.667*M(C) + 0.0*L(C)

The centroid for the composite function is the control effort level that we will use in this situation. That centroid is given by:

Centroid of Composite function(C) =

0.333*Centroid of H(C) +
0.667*Centroid of M(C) +
0.0*Centroid of L(C)

Here is the function again.

The numerical values for the centroids are:

Centroids of Control Effort Membership Functions:

Centroid of H(C) = 100
Centroid of M(C) = 50
Centroid of L(C) = 0

The computed control effort is then:

Control Effort = 0.333x100 + 0.667x50 + 0x0
Control Effort = 0.667

So, the control law (the computation for the actual control effort) reduces to:

Determine the firing levels for each of the membership functions in the antecedents.

This is usually referred to as measuring the "firing levels".

Multiply the firing levels by the corresponding Control Effort Membership Functions to obtain a composite function.

The correspondence is determined by the rule base.

Determine the centroid of the composite function to determine the control effort to apply.

This last step is referred to as "defuzzification".

Simple as this approach seems, it does often work, and it is a reasonable way to implement a control system. It has some pros and cons.

You can easily add more rules to the rule base, and they can be compound rules:

Example: If A is true and B is false, then do C.

You can't easily predict performance.

Marks of Acceptance of Fuzzy Logic

At this point, Fuzzy Logic, Fuzzy Set Theory and Fuzzy control are all accepted and widely used. As illustrations:

Visit one of many fuzzy web pages.

http://www.quadralay.com/www/Fuzzy/tutorial.html

Note Zadeh's "Profile" in IEEE Spectrum, June 1995, pp32-35 and Bart Kosko's "Profile" in IEEE Spectrum, February 1996, pp58-62.
Read any copy of IEEE Transactions on Fuzzy Systems.
Note that MatLab sells two toolboxes

The Fuzzy Systems Toolbox
The Fuzzy Logic Toolbox.

Buy one of Bart Kosko's books, especially "Fuzzy Thinking" to get yourself started.