frequency-response method.
If the system designed has poor stability margins, it
is possible that the designed system may become unstable if the mathematical
model involves uncertainties.
7. Note that for nth-order systems, classical design methods (root-locus and
frequency-response methods) yield low-order compensators (first or second order).
Since the observer-based controllers are nth-order Cor (N-m)th-order if the
minimum-order observer is usedD for an nth-order system, the designed system
will become 2nth order Cor (2n-m)th orderD. Since lower-order compensators are
cheaper than higher-order ones, the designer should first apply classical methods
and, if no suitable compensators can be determined, then try the pole-placement-
with-observer design approach presented in this chapter.
108 QUADRATIC OPTIMAL REGULATOR SYSTEMS
An advantage of the quadratic optimal control method over the pole-placement method
is that the former provides a systematic way of computing the state feedback control gain
matrix.
Quadratic Optimal Regulator Problems. We shall now consider the optimal
regulator problem that, given the system equation
#
x = Ax + Bu (10112)
determines the matrix K of the optimal control vector
u(t) = -Kx(t) (10113)
so as to minimize the performance index
q
30
J = (x*Qx + u* Ru) dt (10114)
where Q is a positive-definite (or positive-semidefinite) Hermitian or real symmetric
matrix and R is a positive-definite Hermitian or real symmetric matrix. Note that the
second term on the right-hand side of Equation (10114) accounts for the expenditure
of the energy of the control signals. The matrices Q and R determine the relative
importance of the error and the expenditure of this energy. In this problem, we assume
that the control vector u(t) is unconstrained.
As will be seen later, the linear control law given by Equation (10113) is the optimal
control law. Therefore, if the unknown elements of the matrix K are determined so as
to minimize the performance index, then u(t)=Kx(t) is optimal for any initial state
x(0). The block diagram showing the optimal configuration is shown in Figure 1035.
u . x
x = Ax + Bu
Figure 1035
Optimal regulator K
system.
Section 108 / Quadratic Optimal Regulator Systems 793
Now let us solve the optimization problem. Substituting Equation (10113) into
Equation (10112), we obtain
#
x = Ax - BKx = (A - BK) x
In the following derivations, we assume that the matrix A-BK is stable, or that the
eigenvalues of A-BK have negative real parts.
Substituting Equation (10113) into Equation (10114) yields
q
30
J = (x*Qx + x*K*RKx) dt
30
= x*(Q + K*RK) x dt
Let us set
d
x*(Q + K*RK) x = - (x*Px)
dt
where P is a positive-definite Hermitian or real symmetric matrix. Then we obtain
# #
x*(Q + K* RK) x = -x*Px - x*Px = -x*C (A - BK)* P + P(A - BK)D x
Comparing both sides of this last equation and noting that this equation must hold true
for any x, we require that
(A - BK)* P + P(A - BK) = -(Q + K* RK) (10115)
It can be proved that if A-BK is a stable matrix, there exists a positive-definite ma-
trix P that satisfies Equation (10115). (See Problem A1015.)
Hence our procedure is to determine the elements of P from Equation (10115) and
see if it is positive definite. (Note that more than one matrix P may satisfy this equation.
If the system is stable, there always exists one positive-definite matrix P to satisfy this
[Link] means that, if we solve this equation and find one positive-definite matrix
P, the system is stable. Other P matrices that satisfy this equation are not positive definite
and must be discarded.)
The performance index J can be evaluated as
x*(Q + K* RK)x dt = -x* Px 2
q q
30
J = = -x*(q) Px(q) + x*(0) Px(0)
0
Since all eigenvalues of A-BK are assumed to have negative real parts, we have
x(q) S 0. Therefore, we obtain
J = x*(0) Px(0) (10116)
Thus, the performance index J can be obtained in terms of the initial condition x(0)
and P.
To obtain the solution to the quadratic optimal control problem, we proceed as
follows: Since R has been assumed to be a positive-definite Hermitian or real symmetric
matrix, we can write
R = T* T
794 Chapter 10 / Control Systems Design in State Space
where T is a nonsingular matrix. Then Equation (10115) can be written as
(A* - K* B*) P + P(A - BK) + Q + K* T* TK = 0
which can be rewritten as
A*P + PA + C TK - (T*) - 1B *PD* CTK - (T*) - 1B *PD - PBR - 1B*P + Q = 0
The minimization of J with respect to K requires the minimization of
x*CTK - (T*)-1 B* PD *C TK - (T*)-1 B* PD x
with respect to K. (See Problem A1016.) Since this last expression is nonnegative, the
minimum occurs when it is zero, or when
TK = (T*)-1 B* P
Hence,
K = T -1(T*)-1 B* P = R-1 B* P (10117)
Equation (10117) gives the optimal matrix [Link], the optimal control law to the quad-
ratic optimal control problem when the performance index is given by Equation (10114)
is linear and is given by
u(t) = -Kx(t) = -R-1 B* Px(t)
The matrix P in Equation (10117) must satisfy Equation (10115) or the following
reduced equation:
A* P + PA - PBR-1 B* P + Q = 0 (10118)
Equation (10118) is called the reduced-matrix Riccati equation. The design steps may
be stated as follows:
1. Solve Equation (10118), the reduced-matrix Riccati equation, for the matrix P.
[If a positive-definite matrix P exists (certain systems may not have a positive-
definite matrix P), the system is stable, or matrix A-BK is stable.]
2. Substitute this matrix P into Equation (10117). The resulting matrix K is the
optimal matrix.
A design example based on this approach is given in Example 109. Note that if the
matrix A-BK is stable, the present method always gives the correct result.
Finally, note that if the performance index is given in terms of the output vector
rather than the state vector, that is,
q
30
J = (y* Qy + u* Ru) dt
then the index can be modified by using the output equation
y = Cx
to
q
30
J = (x* C* QCx + u* Ru) dt (10119)
and the design steps presented in this section can be applied to obtain the optimal
matrix K.
Section 108 / Quadratic Optimal Regulator Systems 795
EXAMPLE 109 Consider the system shown in Figure 1036. Assuming the control signal to be
u(t) = -Kx(t)
determine the optimal feedback gain matrix K such that the following performance index is
minimized:
q
30
J = AxT Qx + u2 B dt
where
Q = B R
1 0
(m 0)
0 m
From Figure 1036, we find that the state equation for the plant is
#
x = Ax + Bu
where
A = B R, B = B R
0 1 0
0 0 1
We shall demonstrate the use of the reduced-matrix Riccati equation in the design of the
optimal control system. Let us solve Equation (10118), rewritten as
A* P + PA - PBR-1 B* P + Q = 0
Noting that matrix A is real and matrix Q is real symmetric, we see that matrix P is a real sym-
metric matrix. Hence, this last equation can be written as
B RB R + B RB R
0 0 p11 p12 p11 p12 0 1
1 0 p12 p22 p12 p22 0 0
- B R B R [1][0 1] B R + B R = B R
p11 p12 0 p11 p12 1 0 0 0
p12 p22 1 p12 p22 0 m 0 0
This equation can be simplified to
B R + B R - B 12 R + B R = B R
0 0 0 p11 p2 p12 p22 1 0 0 0
p11 p12 0 p12 p12 p22 p222 0 m 0 0
Plant
u x2 x1
Figure 1036 K
Control system.
796 Chapter 10 / Control Systems Design in State Space
Plant
u x2 x1
Figure 1037
!m + 2
Optimal control of
the plant shown in
Figure 1036.
from which we obtain the following three equations:
1 - p212 = 0
p11 - p12 p22 = 0
m + 2p12 - p222 = 0
Solving these three simultaneous equations for p11 , p12 , and p22 , requiring P to be positive definite,
we obtain
P = B R = B R
p11 p12 1m + 2 1
p12 p22 1 1m + 2
Referring to Equation (10117), the optimal feedback gain matrix K is obtained as
K = R-1 B* P
1] B R
p11 p12
= [1][0
p12 p22
= Cp12 p22 D
= C1 1m + 2D
Thus, the optimal control signal is
u = -Kx = -x1 - 1m + 2 x2 (10120)
Note that the control law given by Equation (10120) yields an optimal result for any initial state
under the given performance index. Figure 1037 is the block diagram for this system.
Since the characteristic equation is
s I - A + BK = s2 + 1m + 2 s + 1 = 0
if m=1, the two closed-loop poles are located at
s = -0.866 + j 0.5, s = -0.866 - j 0.5
These correspond to the desired closed-loop poles when m=1.
Solving Quadratic Optimal Regulator Problems with MATLAB. In MATLAB,
the command
lqr(A,B,Q,R)
Section 108 / Quadratic Optimal Regulator Systems 797
