Lecture 2: Fuzzy Control II

Jens Kober Robert Babuška

Knowledge-Based Control Systems (SC42050)

Cognitive Robotics
3mE, Delft University of Technology, The Netherlands

07-04-2018
Outline

1 Singleton and Takagi–Sugeno fuzzy system.

2 Knowledge based fuzzy modeling.

3 Data-driven construction.

4 Direct fuzzy control.

5 Supervisory fuzzy control.

2 / 81
Singleton Fuzzy Model

If x is Ai then y = bi

Inference/defuzzification:
PK
—Ai (x)bi
y = Pi=1
K
i=1 —Ai (x)

• well-understood approximation properties

• straightforward parameter estimation

3 / 81
Piece-wise Linear Approximation
y y = f(x)
b2

b4

b3
b1
x
A1 A2 A3 A4
1

4 / 81
Linear Mapping with a Singleton Model
p
X
y = kT x + q = kj xj + q
j=1

• Triangular partition:
A1 A2 A3 A4 A5

a1 a2 a3 a4 a5

• Consequent singletons are equal to:

p
X
bi = kj ai;j + q
j=1

5 / 81
Takagi–Sugeno (TS) Fuzzy Model

If x is Ai then yi = ai x + bi

K
X K
X
—Ai (x)yi —Ai (x)(ai x + bi )
i=1 i=1
y = K
= K
X X
—Ai (x) —Ai (x)
i=1 i=1

6 / 81
Input-Output Mapping of the TS Model

y
y=
a2

b1
x+

x+
b2
a1
y=
y = a3 x + b3

x
Small Medium Large
µ

x
Consequents are approximate local linear models of the system.

7 / 81
TS Model is a Quasi-Linear System

K
X K
X
—Ai (x)yi —Ai (x)(aiT x + bi )
i=1 i=1
y = K
= K
X X
—Aj (x) —Aj (x)
j=1 j=1

8 / 81
TS Model is a Quasi-Linear System

K
X K
X
—Ai (x)yi —Ai (x)(aiT x + bi )
i=1 i=1
y = K
= K
X X
—Aj (x) —Aj (x)
j=1 j=1

K K
!
X X
y = ‚i (x)aiT x+ ‚i (x)bi
i=1 i=1
| {z } | {z }
a(x)T b(x)

linear in parameters ai and bi , pseudo-linear in x (LPV)

8 / 81
TS Model is a Polytopic System

Parameter space
Rules
Antecedent space
Big
Medium

Polytope
a2
Small

x2 a1
Small Medium Big
Parameters of a consequent
function: y = a1x1 + a2x2
x1

9 / 81
Construction of Fuzzy Models

10 / 81
Modeling Paradigms

• Mechanistic (white-box, physical)

• Qualitative (naive physics, knowledge-based)
• Data-driven (black-box, inductive)
Often combination of different approaches semi-mechanistic, gray-box
modeling.

11 / 81
Parameterization of nonlinear models

• polynomials, splines
• look-up tables
• fuzzy systems
• neural networks
• (neuro-)fuzzy systems
• radial basis function networks
• wavelet networks
• ...

12 / 81
Modeling of Complex Systems

Data Model
- static
- dynamic

13 / 81
Modeling of Complex Systems

Data Model
- static
- dynamic

Prior knowledge
- partial models
- stability
- gain, nonlinearity
- settling time

13 / 81
Modeling of Complex Systems

Data Model User

- static - accuracy
- dynamic - interpretation
- mathematical form
- reliability (extrapolation)
Prior knowledge
- partial models
- stability
- gain, nonlinearity
- settling time

13 / 81
Building Fuzzy Models

Knowledge-based approach:
• expert knowledge → rules & membership functions
• fuzzy model of human operator
• linguistic interpretation

14 / 81
Building Fuzzy Models

Knowledge-based approach:
• expert knowledge → rules & membership functions
• fuzzy model of human operator
• linguistic interpretation

Data-driven approach:
• nonlinear mapping, universal approximation
• extract rules & membership functions from data

14 / 81
Knowledge-Based Modeling

• Problems where little or no data are available.

• Similar to expert systems.
• Presence of both quantitative and qualitative variables or
parameters.
Typical applications: fuzzy control and decision support, but also
modeling of poorly understood processes

15 / 81
Wear Prediction for a Trencher

Trencher T-850 (Vermeer) Chain Detail

Goal: Given the terrain properties, predict bit wear and production rate
of trencher.

16 / 81
Why Knowledge-Based Modeling?

• Interaction between tool and environment is complex, dynamic and

highly nonlinear, rigorous mathematical models are not available.
• Little data (15 data points) to develop statistical regression models.
• Input data are a mixture of numerical measurements (rock
strength, joint spacing, trench dimensions) and qualitative
information (joint orientation).
• Precise numerical output not needed, qualitative assessment is
sufficient.

17 / 81
Dimensionality Problem: Hierarchical Structure

Assume 5 membership functions for each input

625 rules in a flat rule base vs. 75 rules in a hierarchical one
x1
rule base
x2 A
y1
rule base z
C
x3 y2
rule base
x4 B

18 / 81
Trencher: Fuzzy Rule Bases
Orientation

Rock strength

Rule base 3 Feed Rule base 4 Production rate

(feed) (production) (linguistic output)
Rock dim.

Defuzzification
Joint set 1
Rule base 1 Block size Production rate
Joint set 2 (numerical output)
(block size)
Joint set 3

If TRENCH-DIM is SMALL and STRENGTH is LOW Then FEED is VERY-HIGH;

If TRENCH-DIM is SMALL and STRENGTH is MEDIUM Then FEED is HIGH;
....
If JOINT-SP is EXTRA-LARGE and FEED is VERY-HIGH Then PROD is VERY-HIGH

19 / 81
Example of Membership Functions
VS SM ME LA VL Low Medium High
1 1
Membership grade

Membership grade
0.8 0.8
0.6 0.6

0.4 0.4

0.2 0.2

0 0
0 0.2 0.4 0.6 0.8 1 0 50 100 150 200

VL LO ME HI VH
1
Membership grade

0.8

0.6

0.4

0.2

0
0 1 2 3 4 5
Bit consumption [bit/m^3]

20 / 81
Output: Prediction of Production Rate

data no. measured value predicted linguistic value(s)

--------------------------------------------------------

1 2.07 VERY-LOW 1.00

2 5.56 HIGH 1.00
3 23.60 VERY-HIGH 0.50
4 11.90 HIGH 0.40 VERY-HIGH 0.60
5 7.71 MEDIUM 1.00
6 7.17 LOW 0.72
7 8.05 MEDIUM 0.80
8 7.39 LOW 1.00
9 4.58 LOW 0.50
10 8.74 MEDIUM 1.00
11 134.84 EXTREMELY-HIGH 1.00

21 / 81
Data-Driven Construction

22 / 81
Structure and Parameters

Structure:
• Input and output variables. For dynamic systems also the
representation of the dynamics.
• Number of membership functions per variable, type of membership
functions, number of rules.

Parameters:
• Consequent parameters (least squares).
• Antecedent membership functions (various methods).

23 / 81
Least-Squares Estimation of Singletons

Ri : If x is Ai then y = bi

• Given Ai and a set of input–output data:

{(xk ; yk ) | k = 1; 2; : : : ; N}
• Estimate optimal consequent parameters bi .

24 / 81
Least-Squares Estimation of Singletons

1 Compute the membership degrees —Ai (xk )

2 Normalize
K
X
‚ki = —Ai (xk )= —Aj (xk )
j=1
PK
(Output: yk = i=1 ‚ki bi , in a matrix form: y = Γb)
h i−1
3 Least-squares estimate: b = ΓT Γ ΓT y

25 / 81
Least-Square Estimation of TS Consequents

2 3 2 3 2 3
x1T y1 ‚i1 0 · · · 0
6 T7
6x2 7 y 6 0 ‚i2 · · · 0
6 7 6 7
6 27 7
X=6
6 .. 7 ;
7 y =6 .
6 . 7
7 ; Γi = 6 .
6 . .. .. .. 7
4 . 5 4 . 5 4 . . . .
7
5
xNT yN 0 0 ··· ‚iN
h iT
„i = aiT bi ; Xe = [X 1]

26 / 81
Least-Square Estimation of TS Consequents

h i−1
• Global LS: „ 0 = (X 0 )T X 0 (X 0 )T y

with X 0 = [Γ1 Xe Γ2 Xe ::: Γc Xe ]

h iT
and „ 0 = „1T „2T ::: „cT

27 / 81
Least-Square Estimation of TS Consequents

h i−1
• Global LS: „ 0 = (X 0 )T X 0 (X 0 )T y

with X 0 = [Γ1 Xe Γ2 Xe ::: Γc Xe ]

h iT
and „ 0 = „1T „2T ::: „cT
h i−1
• Local LS: „i = XeT Γi Xe XeT Γi y

27 / 81
Antecedent Membership Functions

• templates (grid partitioning),

• nonlinear optimization (neuro-fuzzy methods),
• tree-construction
• product space fuzzy clustering

28 / 81
Fuzzy Clustering: Data

29 / 81
Fuzzy Clustering: Prototypes

v2

Cluster centers (means):

v1 [ ]
V= v1 v2

29 / 81
Fuzzy Clustering: Distance

v2

Cluster centers (means):

v1 d(zk v, 1) zk
V= [ v1 v2 ]

29 / 81
Fuzzy Clustering: Distance

v2

)
,2
v
zk
d(
Cluster centers (means):
v1 d(zk v, 1) zk
V= [ v1 v2 ]

29 / 81
Fuzzy Clustering: Partition Matrix

v2

)
,2
v
zk
d(
Cluster centers (means):
v1 d(zk v, 1) zk
V= [ v1 v2 ]

Fuzzy partition matrix:

U=
[ ]
29 / 81
Fuzzy Clustering: Shapes

v2

)
,2
v
zk
d(
Cluster centers (means):
v1 d(zk v, 1) zk
V= [ v1 v2 ]

Fuzzy partition matrix:

U=
[ ]
29 / 81
Fuzzy Clustering Problem

Given the data:

zk = [z1k ; z2k ; : : : ; znk ]T ∈ Rn ; k = 1; : : : ; N

Find:
the fuzzy partition matrix:

—11 : : : —1k : : : —1N

2 3
6 .. .. .. 7
U=4 . ::: . ::: . 5
—c1 : : : —ck : : : —cN

and the cluster centers:

V = {v1 ; v2 ; : : : ; vc }; vi ∈ Rn

30 / 81
Fuzzy Clustering: an Optimization Approach

Objective function (least-squares criterion):

c X
X N
J(Z; V ; U; A) = —m 2
i;j dAi (zj ; vi )
i=1 j=1

subject to constraints:

0 ≤ —i;j ≤ 1; i = 1; : : : ; c; j = 1; : : : ; N membership degree

N
X
0< —i;j < N; i = 1; : : : ; c no cluster empty
j=1
c
X
—i;j = 1; j = 1; : : : ; N total membership
i=1

31 / 81
Fuzzy c-Means Algorithm

Repeat:
PN m
— z
k=1 i;k k
1 Compute cluster prototypes (means): vi = P N m
k=1
—i;k

32 / 81
Fuzzy c-Means Algorithm

Repeat:
PN m
— z
k=1 i;k k
1 Compute cluster prototypes (means): vi = P N m
k=1
—i;k

2 Calculate distances: dik = (zk − vi )T (zk − vi )

32 / 81
Fuzzy c-Means Algorithm

Repeat:
PN m
— z
k=1 i;k k
1 Compute cluster prototypes (means): vi = P N m
k=1
—i;k

2 Calculate distances: dik = (zk − vi )T (zk − vi )

1
3 Update partition matrix: —ik = Pc (dik =djk )1=(m−1)
j=1

until k∆Uk < ›

32 / 81
Distance Measures

• Euclidean norm:

d 2 (zj ; vi ) = (zj − vi )T (zj − vi )

• Inner-product norm:

dA2 i (zj ; vi ) = (zj − vi )T Ai (zj − vi )

• Many other possibilities . . .

33 / 81
Fuzzy Clustering – Demo

1 Fuzzy c-means

34 / 81
Extraction of Rules by Fuzzy Clustering

y
10

2 4 6 8 x

35 / 81
Extraction of Rules by Fuzzy Clustering

y Cluster 4
10 x

r3
8

ste
6

Clu
x
4

r2
te
2
Cluster 1 x

us
Cl
x
2 4 6 8 x

35 / 81
Extraction of Rules by Fuzzy Clustering

y Cluster 4
10 x

r3
8

ste

projection
6

Clu
x
4

r2
te
2
Cluster 1 x

us
Cl
x
Takagi-Sugeno model 2 4 6 8 x
1
m(x)

Rule-based description: 0.8

If x is A1 then y = a1x + b1
0.6
A1 A2 A3 A4
0.4
If x is A2 then y = a2x + b2
0.2

etc... 0
2 4 6 8
x

35 / 81
Extraction of Rules by Fuzzy Clustering

y
B4 y Cluster 4

10
10 x
projection

r3
8
B3

ste
6
6

Clu
4
x
4

2
B1 B2

r
te
2

2
Cluster 1 x

us
Cl
x
1

0
0.8

0.6

0.4

0.2

2 4 6 8 x
m(y)

Rule-based description:
If y is B1 then x = a1y + b1 Inverse Takagi-Sugeno model
If y is B2 then x = a2y + b2
etc...

35 / 81
Rule Extraction – Demo

• Extraction of Takagi–Sugeno rules

36 / 81
Fuzzy Control

37 / 81
Fuzzy Control: Background

• controller designed by using If–Then rules instead of mathematical

formulas (knowledge-based control),
• early motivation: mimic experienced operators,
• fuzzy reasoning: interpolation between discrete outputs,
• currently: also controllers designed on the basis of a fuzzy model
(model-based fuzzy control),
• a fuzzy controller represents a nonlinear mapping
(but completely deterministic!).

38 / 81
Parameterization of Nonlinear Controllers

Controller Process
1

0.5
theta

−0.5
1

0.5
−1
−1 0
−0.5
0 −0.5
0.5
1 −1 x
phi

Parameterizations

PL PS Z

PS Z NS

Z NS NL

Sigmoidal Fuzzy Systems Radial Basis Wavelets Splines

Neural Networks Functions

39 / 81
Fuzzy System is a Nonlinear Mapping

40 / 81
Basic Fuzzy Control Schemes

• Direct (low-level, Mamdani) fuzzy control

• Fuzzy supervisory (high-level, Takagi–Sugeno) control

• Fuzzy model-based control

41 / 81
Process Controlled by Operators
Goal (reference) Disturbances

Human u y
operator Process

42 / 81
Knowledge Acquisition
Goal (reference) Disturbances

Human u y
operator Process

Knowledge
acquisition
Fuzzy if-then
rules

42 / 81
Direct Fuzzy Control
Goal (reference) Disturbances

Fuzzy u y
controller Process

Implementation
Fuzzy if-then
rules

42 / 81
Example of Operator Knowledge

Case Condition Action to be taken Reason

11 BZ OK a. Decrease fuel rate slightly To raise percentage of oxygen
OX low
BE OK
12 BZ OK a. Reduce fuel rate To increase percentage of oxygen
OX low for action b
BE high b. Reduce fan speed To lower back-end temperature and
maintain burning zone temperature
13 BZ OK a. Increase fan speed To raise back-end temperature
OX OK b. Increase fuel rate To maintain burning zone
BE low temperature

Extract from Peray’s textbook for kiln operators (Oestergaard, 1999)

43 / 81
FLC Analogue to PID Control

r e u y
PID Process
-

44 / 81
PID Control: Internal View
e(fi )dfi + D de(t)
Rt
u(t) = P e(t) + I 0 dt

P
e u
I +

d
dt D

45 / 81
PID Control: Internal View
e(fi )dfi + D de(t)
Rt
u(t) = P e(t) + I 0 dt

P
e u
I +

d
dt D

dynamic filter static mapping

45 / 81
Fuzzy PID Control
“ ”
e(fi )dfi; de(t)
Rt
u(t) = f e(t); 0 dt

knowledge-based
system
e u

Fuzzy
d
dt
dynamic filter

46 / 81
Fuzzy PID Control
r e Fuzzy u y
Controller Process
-

e e Du Dynamic u
Dynamic
Pre-filter òe Static map
Post-filter
De

Knowledge base
Fuzzifier Defuzzifier
Inference

47 / 81
Example: Proportional Control

r e u y
P Process
-

48 / 81
Controller’s Input–Output Mapping
u
u=Pe

49 / 81
Fuzzy Proportional Control: Rules

If error is Negative Big then control input is Negative Big

If error is Positive Big then control input is Positive Big

If error is Zero then control input is Zero

50 / 81
Controller’s Input–Output Mapping
u
PB
ZE

e
NB

NB ZE PB

51 / 81
Example: Friction Compensation

1 DC motor with static friction.

2 Fuzzy rules to represent “normal” proportional control.
3 Additional rules to prevent undesirable states.

52 / 81
DC Motor: Model

Armature Load
Friction
K(s) 1 1
1 1
L.s+R J.s+b s
voltage angle

53 / 81
Proportional Controller

Control u

P Motor
Mux

Angle

54 / 81
Linear Control
0.15
shaft angle [rad] 0.1

0.05

−0.05

−0.1

0 5 10 15 20 25 30
time [s]

1.5

1
control input [V]

0.5

−0.5

−1

−1.5
0 5 10 15 20 25 30
time [s]

55 / 81
Fuzzy Control Rule Base

If error is Positive Big then control input is Positive Big;

If error is Negative Big then control input is Negative Big;

If error is Zero then control input is Zero;

56 / 81
Membership Functions for Error

Negative Big Zero Positive Big

1

0
-0.2 -0.15 -0.1 -0.05 0 0.05 0.1 0.15 0.2

57 / 81
Additional Rules

If error is Positive Big then control input is Positive Big;

If error is Negative Big then control input is Negative Big;

If error is Zero then control input is Zero;

If error is Negative Small then control input is not Negative Small;

If error is Positive Small then control input is not Positive Small;

58 / 81
Membership Functions for Error

Negative Big Zero Positive Big

1 Negative Positive
Small Small

0
-0.2 -0.15 -0.1 -0.05 0 0.05 0.1 0.15 0.2

59 / 81
Fuzzy Control
0.15
shaft angle [rad] 0.1

0.05

−0.05

−0.1

0 5 10 15 20 25 30
time [s]

1.5

1
control input [V]

0.5

−0.5

−1

−1.5
0 5 10 15 20 25 30
time [s]

60 / 81
Input–Output Mapping of the Controller
u
1
local nonlinearity

0.5

-0.15 -0.1 -0.05 .0 0.05 0.1 0.15

e
-0.5

-1

61 / 81
Another Solution: Sliding Mode Control
0.15
shaft angle [rad] 0.1

0.05

−0.05

−0.1

0 5 10 15 20 25 30
time [s]

1.5

1
control input [V]

0.5

−0.5

−1

−1.5
0 5 10 15 20 25 30
time [s]

62 / 81
Experimental Results
0.2

0.1
position [rad]

−0.1
Proportional Control
−0.2

0 2 4 6 8 10 12 14 16 18 20
time [s]

10
control input [V]

−5

−10
0 2 4 6 8 10 12 14 16 18 20
time [s]

63 / 81
Experimental Results
0.2

0.1
position [rad]

−0.1
Proportional Control
Fuzzy Control
−0.2

0 2 4 6 8 10 12 14 16 18 20
time [s]

10
control input [V]

−5

−10
0 2 4 6 8 10 12 14 16 18 20
time [s]

63 / 81
Membership Functions
1

membership degree 0.8

0.6

0.4

0.2

0
−1 −0.9 −0.8 −0.7 −0.6 −0.5 −0.4 −0.3 −0.2 −0.1 0
Control error

0.8
membership degree

0.6

0.4

0.2

0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Control error

64 / 81
Fuzzy PD Controller: Rule Table

error rate

NB ZE PB
NB NB NB ZE
error
ZE NB ZE PB
PB ZE PB PB

R12 : If error is NB and error rate is ZE then control is NB

65 / 81
Fuzzy PD Controller – cont’d

0.5 0.5

0 0
control

control
−0.5 −0.5

−1 −1
1 1
0.5 1 0.5 1
0 0.5 0 0.5
0 0
−0.5 −0.5
−0.5 −0.5
error rate −1 −1 error rate −1 −1
error error

0.5

0
control

−0.5

−1
1
0.5 1
0 0.5
0
−0.5
−0.5
error rate −1 −1
error

66 / 81
Supervisory Fuzzy Control

Classical u y
Process
controller

67 / 81
Supervisory Fuzzy Control
external signals

Fuzzy Supervisor

Classical u y
Process
controller

67 / 81
Supervisory Control Rules: Example

If process output is High

then reduce proportional gain Slightly and
increase derivative gain Moderately.
(Supervised PD controller)

68 / 81
Example: Inverted Pendulum

d
69 / 81
Cascade Control Scheme

Reference Position
controller Angle Inverted
controller pendulum

70 / 81
Experimental Results
0.5

cart position [m]

0
Conventional PD controller

−0.5
0 10 20 30 40 50 60
time [s]
0.2

0.1
angle [rad]

−0.1

−0.2
0 10 20 30 40 50 60
time [s]
1
control input [−]

0.5

−0.5

−1
0 10 20 30 40 50 60
time [s]

71 / 81
Experimental Results
0.5

cart position [m]

0
Conventional PD controller
Fuzzy Supervised PD controller
−0.5
0 10 20 30 40 50 60
time [s]
0.2

0.1
angle [rad]

−0.1

−0.2
0 10 20 30 40 50 60
time [s]
1
control input [−]

0.5

−0.5

−1
0 10 20 30 40 50 60
time [s]

71 / 81
Takagi–Sugeno Control
Takagi–Sugeno PD controller:

R1 : If r is Low then uL = PL e + DL ė
R2 : If r is High then uH = PH e + DH ė

—L (r ) uL + —H (r ) uH
u = = ‚L (r ) uL + ‚H (r ) uH
—L (r ) + —H (r )

72 / 81
Takagi–Sugeno Control
Takagi–Sugeno PD controller:

R1 : If r is Low then uL = PL e + DL ė
R2 : If r is High then uH = PH e + DH ė

—L (r ) uL + —H (r ) uH
u = = ‚L (r ) uL + ‚H (r ) uH
—L (r ) + —H (r )

= {‚L (r )PL + ‚H (r )PH } e + {‚L (r )DL + ‚H (r )DH } ė

72 / 81
Takagi–Sugeno Control
Takagi–Sugeno PD controller:

R1 : If r is Low then uL = PL e + DL ė
R2 : If r is High then uH = PH e + DH ė

—L (r ) uL + —H (r ) uH
u = = ‚L (r ) uL + ‚H (r ) uH
—L (r ) + —H (r )

= {‚L (r )PL + ‚H (r )PH } e + {‚L (r )DL + ‚H (r )DH } ė

= P (r )e + D(r )ė;

72 / 81
Takagi–Sugeno Control
Takagi–Sugeno PD controller:

R1 : If r is Low then uL = PL e + DL ė
R2 : If r is High then uH = PH e + DH ė

—L (r ) uL + —H (r ) uH
u = = ‚L (r ) uL + ‚H (r ) uH
—L (r ) + —H (r )

= {‚L (r )PL + ‚H (r )PH } e + {‚L (r )DL + ‚H (r )DH } ė

= P (r )e + D(r )ė; P (r ) ∈ conv(PL ; PH ); : : :

72 / 81
Takagi–Sugeno Control is Gain Scheduling

fuzzy scheduling

Controller K
Inputs Outputs
Controller 2

Controller 1

73 / 81
TS Control: Input–Output Mapping

2 2

1.5 1.5

1 1

0.5 0.5

0 0
1 1

0.5 1 0.5 1
0.5 0.5
0 0
0 0
−0.5 −0.5
−0.5 −0.5
−1 −1 −1 −1

1.5

0.5

0
1

0.5 1
0.5
0
0
−0.5
−0.5
−1 −1

74 / 81
TS Control: Example

1 Strongly nonlinear process (output-dependent gain).

2 Fuzzy supervisor to adjust the gain of a proportional controller.
3 Comparison with linear (fixed-gain) proportional control.

75 / 81
TS Control: Example

Nonlinear process:

d 3 y (t) d 2 y (t) dy (t)

+ + = y 2 (t)u(t)
dt 3 dt 2 dt
Problems with linear control:
• stability and performance depend on process output
• re-tuning the controller does not help
• nonlinear control is the only solution

76 / 81
TS Control: Example

Goal: Design a controller to stabilize the process for a wide range of

operating points (y > 0):

TS (proportional) control rules:

If y is Small then u(k) = PSmall · e(k)

If y is Medium then u(k) = PMedium · e(k)
If y is Large then u(k) = PLarge · e(k)

77 / 81
Comparison of Performance
2

1.9

1.8

1.7

1.6

1.5

1.4

1.3

1.2

1.1

1
0 5 10 15 20 25 30

5.5

4.5

3.5

2.5

2
0 5 10 15 20 25 30

78 / 81
Typical Applications

• Tune parameters of low-level controllers (auto-tuning).

• Improve performance of classical control (response-assisted PID).
• Adaptation, gain scheduling (aircraft control).

79 / 81
Typical Applications

• Tune parameters of low-level controllers (auto-tuning).

• Improve performance of classical control (response-assisted PID).
• Adaptation, gain scheduling (aircraft control).

+ Enhancement of classical controllers.

+ Interface between low-level and high-level control.
− Ad hoc approach, difficult analysis.

79 / 81
Fuzzy Control: Design Steps

control engineering approaches + heuristic knowledge

1 Determine inputs and outputs.

2 Define membership functions.

3 Design rule base.

4 Test (completeness, stability, performance).

5 Fine-tune the controller.

80 / 81
Parameters in a Fuzzy Controller

Fuzzification module Defuzzification module

Inference
Scaling Fuzzifier engine Defuzzifier Scaling

Scaling Membership Membership Scaling

factors functions Rule base functions factors

Data base Data base

Knowledge base

81 / 81