EE635 - Control System Theory

Jitkomut Songsiri

7. Linear Quadratic Gaussian Control

output feedback
Kalman filter
LQG/LQR

7-1

Output feedback
consider a linear system
x = Ax + Bu,

y = Cx

a state-feedback controller has a form

u(t) = Kx(t)
which requires the availability of the process measurement
when the state variables are not accessible, one can use
u(t) = K x
(t)
where x
(t) is an estimate of x(t) based on the output y
Linear Quadratic Gaussian Control

7-2

Full-order observers
x cannot be fully measured and the goal is to estimate x based on y
our approach is to replicate the process dynamic in x

x
= A
x + Bu
define the state estimation error e = x x
, we can see
e = Ax A
x = Ae
if A is stable, then the error goes to zero asymptotically
if A is unstable, e is unbounded and x
grows further apart from x
to avoid this problem, one can consider a correction term as
x
= A
x + Bu + L(y y),

x
(0) = 0

(feed y back to the estimator)

Linear Quadratic Gaussian Control

7-3

L is a given matrix, called observer gain matrix

the error dynamic now is
e = Ax A
x L(Cx C x
) = (A LC)e,

e(0) = x(0)

the observer error goes to zero if L is chosen such that

A LC is stable
we can make e goes to zero fast if the eigenvalues of A LC can be
arbitrarily assigned
eigenvalues of A LC are same as those of (AT C T LT )
hence, choosing L is the dual of state-feedback design problem for the
pair (AT , C T )
eigenvalues of A LC can be freely reassigned if and only if (A, C)
observable
Linear Quadratic Gaussian Control

7-4

Observer-based controller
closed-loop equations after a state feedback
u(t) = r(t) K x
(t)
where r(t) is reference input, is
x = Ax BK x
+ Br
x
= A
x + L(y y) + Br BK x

or

  
d x
A
=

LC
dt x

Linear Quadratic Gaussian Control

   
BK
x
B
+
r
A LC BK x

7-5

Separation principle
change the coordinate


I
I

  
  
0
x
x
x
=
=
I x

xx

the dynamics in the new coordinate is

  
d x
A BK
=
0
dt e

   
BK
x
I
+
r
A LC e
0

the dynamic of e does not depend on x

closed-loop eigenvalues are
{eig(A BK)} {eig(A LC)}
one can assign eigenvalues of the system and the observer independently
Linear Quadratic Gaussian Control

7-6

Observer of noisy system

in general, the system is corrupted by noise
x = Ax + Bu + w,

y = Cx + v

where w is process noise and v is measurement noise

rewrite the error dynamic of an observer
e = Ax + Bu + w A
x Bu L(Cx + v C x
)
= (A LC)e + w Lv

due to w, v, the estimation will generally not go to zero

one would like the error to remain small by a good choice of L
the optimal choise of L is given by the Kalman gain
Linear Quadratic Gaussian Control

7-7

Kalman filter
assume w, v are uncorrelated zero-mean Gaussian white noise, i.e.,
E w(t)w( ) = W (t ),

E v(t)v( ) = V (t )

E w(t)v( ) = 0

spectral density matrices of w, v are

W  0,

and V 0,

respectively

The optimal observer gain which minimizes Eke(t)k2 is

L = P C V 1
where P is the unique positive-semidefinite solution of the ARE
P A + AP P C V 1CP + W = 0
when using the optimal gain, this system is called the Kalman-Bucy filter
Linear Quadratic Gaussian Control

7-8

A LC is stable as long as the two conditions hold

(A, C) is observable
(A, W ) is controllable
Recall: LQR control with J =
we need to solve ARE

R
0

x(t)Qx(t) + u(t)Ru(t)dt

P A + AP P BR1B P + Q = 0
with conditions
(A, B) controllable to guarantee Jmin <
(A, Q) observable to obtain a unique positive definition P
Duality: Kalmain filter is equivalent to designing an LQR controller on the
dual system (A, C , B , D ) with Q = W , R = V
Linear Quadratic Gaussian Control

7-9

Linear Quadratic Gaussian Control

a combination of optimal state estimation and optimal state feedback
w
u

Plant

B
x

R
A

Linear Quadratic Gaussian Control

L
+
+

y
-

Kalman Filter

LQ regulator

7-10

Example
we design an LQG controller to a spring-mass system

 



0
0 1
u + w, y = 1 0 x + v
x+
x =
1
1 0
where covariances of w are v are I and 0.1 respectively
we use the parameters
Q = 5C C,

R = 1,

W = I,

V = 0.1

(A, C) observable and (A, W ) controllable

MATLAB codes
n
Q
K
W
L

=
=
=
=
=

2; A = [0 1;-1 0]; B = [0 1]; C = [1 0];

5*C*C; R = 1;
lqr(A,B,Q,R);
eye(n); V = 0.1;
lqe(A,eye(n),C,W,V);

Linear Quadratic Gaussian Control

7-11

closed-loop response to unit step in reference

2.5
output
estimation error
input

2
1.5
1
0.5
0
0.5
1
1.5
0

10

15

20

25

30

t
(a standard LQG does not have an integral action)
Linear Quadratic Gaussian Control

7-12

Robustness properties
LQR-controlled systems if R is diagonal, then
the system will have a gain margin equal to and a phase margin of 60
The input u = Kx can have a complex perturbation
ui = kieji ,

i = 1, 2, . . . , m

m = number of inputs

without causing instability providing

1. i = 0 and 0.5 ki ,
2. ki = 1 and |i| 60,

i = 1, 2, . . . , m

i = 1, 2, . . . , m

LQG-controlled systems LQG-controlled system with a combined Kalman

filter and LQR control law, there are NO guaranteed stability margins
Linear Quadratic Gaussian Control

7-13

Counterexample
we will show that the gain margin of LQG becomes very small as feedback
and observer gains become large

consider an unstable SISO system

 

   
 
x 1
1 1 x1
0
1
=
+
u+
w
x 2
0 1 x2
1
1


y = 1 0 x+v
with parameters in controller and observer design given by

1 1
,
Q=
1 1


Linear Quadratic Gaussian Control

R = 1,


1 1
,
W =
1 1


V =1

7-14

ARE for controller design has the positive definite solution




p11 p12
, p12 = p22 = 2 + 4 + , p11 = (p212 )/2
P =
p12 p22
the LQR gain is
K = R



B P = (2 + 4 + ) 1 1

k k

solving the ARE for observer design, the Kalman gain is

  
p
1
l
L = (2 + 2 + )
=
1
l
the equations of the controlled system and the observer are
x
x = Ax BK
+ Fw
x
= A
x + Bu + L(Cx + v C x
)
where
Linear Quadratic Gaussian Control

7-15

 is the actual input matrix for the system (B

contains an uncertainty)
B
B is the input matrix used in control design

were concerned if the LQG has a robust stability (from uncertainty in B)

 
= 0 where 6= 1 (so B
=
assume B
6 B)

the eigenvalues of the dynamic matrix



A
LC

BK
(A LC + BK)

satisfies the characteristic equations of the form

s4 + c 3 s3 + c 2 s2 + c 1 s + c 0 = 0
and from Rouths criterion, all ck s must be positive
Linear Quadratic Gaussian Control

7-16

however, we can show that

c1 = l + k 4 + 2( 1)kl,

c0 = 1 + (1 )kl

and the value of can make these two coefficients to change sign:
if > (1 + 1/kl) then c0 < 0
if < [1 (l + k 4)/2lk] then c1 < 0
so the gain margin is smaller as kl becomes larger
increasing l and k relates to increasing Q and W
robustness is enhanced by low weighting on the state and covariance of
process noise

Linear Quadratic Gaussian Control

7-17

References
Lecture note on
Observer-based controller design, David Banjerdphongchai, Chulalongkorn
University
Chapter 9 in
S. Skogestad and I. Postlethwaite, Multivariable Feedback Control:
Analysis and Design, John Wiley & Sons, 1996
Chapter 6 in
R. F. Stengel, Optimal Control and Estimation, Dover, 1994

Linear Quadratic Gaussian Control

7-18