Controlling and tracking objects more precisely have 1 The formulation of the problem
been a goal unremittingly pursued in many elds including
aeronautics, astronautics, navigation, and manufacturing. Consider system
However, the classical optimal control is over-designed, and x(t) + B1w (t h) + B2u (t)
x (t) = Ax (1)
the H control is designed for the worst-case reference sig-
nal and may lead to insucient disturbance attenuation z (t) = Cxx(t) + D2 u(t) (2)
and robustness properties of the closed-loop system. where x , w , u and z are the state, the exogenous input, the
The development of the infrared and sensor technol- control input, and the regulated signal, respectively, and
ogy renders us access to more or less information in ad- h( 0) is a time delay.
vance, namely, preview information, which can improve The H preview control problem in nite horizon is
the performance eectively when applied appropriately. As stated as follows: for a given > 0, nd the control law
shown in the industrial suspension systems[13] , the closed- x(t), w (s)|s [t h, t]) such that
like u (t) = F (x
loop performance can be improved considerably by using
the preview controller. The extensive applications encour- zz 2L2 [0,tf ] + x (tf )T Ptf x (tf )
aged many works[49] with respect to the preview con- sup < 2 (3)
w w 2L2 [0,t h]
w
trol. So far, the optimal preview theory has been relatively f
thereof. As for the proof of the solvability, we will resort X(t) in (11) will be characterized further.
to the abstract space theory. More details will be found in Substituting (11) into (9), we can nd that X(t) satises
the following sections. the relationship
2.1 The ground to solve
X = XA + AT X + C T C XB2 B2T X, X(tf ) = Ptf (12)
It is necessary to show the solvability of the problem
For a compact arrangement, dierential Riccati equation
max{min(zz 2L2 [0,tf ] + x (tf )T Ptf x (tf )) 2 w
w 2L2 [0,tf h]} (12) omits the time index t. We proceed to study the prop-
w u
erty of X(t). In view of (10), we have
(5)
min zz 2L2 [0,tf ] + x (tf )T Ptf x (tf ) = (13)
Finally, we introduce the projection theorem in Hilbert u
space. T
x(0), X(0)x
x x(0) x
x(tf ), X(tf )x
x(tf ) + x (tf ) Ptf x (tf )
Lemma 1. Let V1 and V2 be the Hilbert spaces with
bounded linear operators J : V2 V2 and S : V1 V2 . for any initial value x (0). If X(tf ) = Ptf , the minimal
Suppose J T = J and S T JS > I for some > 0. Then, value in the above is simplied as
given any v 2 V2 , there exists a unique solution to the
optimization problem min zz 2L2 [0,tf ] + x (tf )T Ptf x (tf ) = x
x(0), X(0)x
x(0) 0 (14)
u
min Svv 1 v 2 2J = min Svv 1 v 2 , J(Svv 1 v 2 ) (6) which together with Lemma 2 shows that X(0) is positive
v 1 V1 v 1 V1
semi-denite. For any t > 0, t [0, tf ], consider the mini-
Denote the optimal value max min J(u u, w ) by gopt . On mization problem minu zz 2L2 [t,tf ] + x (tf )T Ptf x (tf ). Like-
the basis of Lemma 1, we achieve the following conclusion. wise, we will verify that X(t) is positive semi-denite.
Lemma 2. If > opt , the game problem (5) is solvable. However, in most cases, w 0 does not hold. Therefore,
Moreover, gopt is nonnegative for any initial data. more work is needed to solve the H problem.
Proof. We only prove the case Ptf = 0, and the case
Ptf = 0 is similar except that the operators S and J are 2.3 Maximization on w
more complicated. Now the roles of S, J, v 1 , and v 2 in We still resort to the above-mentioned traditional idea
Lemma 1 are taken by the input-output operator of the to handle the outer maximization problem in (5).
system (1) and (2), identity operator, the control input, w (t h)
w
and the output driven by initial data and the disturbance, In order to avoid computing , we need to
ww (t)
respectively. For convenience, we will omit the interval of rewrite the game problem as
the norms in L2 [a, b]. It is clear that the optimal problem
minu z 2L2 2 w 2L2 is solvable. As > opt , there max{min(zz (t)2L2 [0,tf ] + x (tf )T Ptf x (tf ))
w u
exists a > 0 such that 2 w 2L2 z 2L2 < 2 w 2L2 .
With the similar line again, the roles of S, J, v 1 , and v 2 2 w
w (t h)2L2 [h,tf ] } (15)
in Lemma 1 are taken by the input-output operator (after
applying the optimal control input) from the disturbance Choose the Hamilton function H = 12 (zz (t)Tz (t) 2w (t
w to z , diag{I, 2 I}, w , and the output driven by initial h)Tw (th))+pp2 (t)T (Ax
x(t)+B1w (th)+B2u (t)) connected
data, respectively. That the optimal value is nonnegative with (1) and (15). After a straightforward calculation, we
follows from that max min J(u u, w ) min J(uu, 0) 0. note that the optimal w (t h) admits
2.2 Minimization on u
p 2 (t) = ATp 2 (t) C T Cx
x(t) (16)
With the solvability, the following will take a traditional
idea to solve the inner minimization problem in (5). 0 = B1Tp 2 (t) 2
w (t h) (17)
Let H = 12 z (t)Tz (t) + p 1 (t)T (Ax
x(t) + B1w (t h) + It is found that (16) is identical with (7), so we can have
B2 u(t)) be the Hamilton function related with minu zz 2L2 the Hamilton-Jacobi type ODE
and (1), naturally, the optimal u should satisfy
T 1 T
x (t) A B 2 B2 + 2 B1 B1
x (t)
p 1 (t) = ATp 1 (t) C T Cx
x(t) (7) = (18)
p 1 (t) p 1 (t)
C T C AT
u (t) = B2Tp 1 (t) (8)
where p1 (t) is the so-called costate. Moreover, (10) clearly still holds for (18).
Combine system (1) and (2) with (7) and (8), a However, prudential readers will see that (18) holds only
Hamilton-Jacobi type original dierential equation (ODE) in the partial interval [h, tf ], which stems from that w (t
is given as h), t [0, h) is known and cannot be chosen. Meanwhile,
another fact is that
x (t) A B2 B2T x (t) B
= T T + 1 w (t h) (9) zz 2L2 [0,tf ] + x (tf )T Ptf x (tf ) 2 w
w 2L2 [0,tf h] =
p 1 (t) C C A p 1 (t) 0
x (tf )T Ptf x (tf ) + w
w (t h), B1Tp 1 (t)L2 [0,h]
From (1) and (9), for any t0 , t1 [0, tf ], and t0 < t1 , t
x(t), p 1 (t)|0f
x (19)
zz 2L2 [t0 ,t1 ] = x
x, p 1 |tt10 + w
w (t h), B1Tp 1 (t)L2 [t0 ,t1 ] (10)
As h = 0, (18) holds in the whole interval [0, tf ], and
If w 0, the game problem (5) is only a standard linear (19) is reduced to
quadratic regulated (LQR) one. Given that u should be
casual, we assume zz 2L2 [0,tf ] + x (tf )T Ptf x (tf ) 2 w
w 2L2 [0,tf ] =
t
x(t)
p 1 (t) = X(t)x (11) x(t), p 1 (t)|0f + x (tf )T Ptf x (tf )
x (20)
No. 2 WANG Hong-Xia and ZHANG Huan-Shui: Finite Horizon H Preview Control 329
h r
x(0) + (s, t)B1w (s h)ds
x (t) = (t, 0)x (26) 2 w (r h),E2
w w (s h)dsdr(39)
(r, s)B(s)w
0
t
tf 0 0
(s, t)B2 B2T (r, s)T E2Tw (r h)drds According to Lemma 2, for any initial value, gopt 0;
0 s naturally, for w (t h) 0, t [0, h) and any x (0),
where (, ) is the state transition matrix corresponding
A2 . Note that the computation uses the boundary value x(0), X(0)x
gopt = x x(0) +
y (tf ) = p 1 (tf ) Xx x(tf ) = 0. x(0), (h, 0)T Z(h)(h, 0)x
x x(0) 0 (40)
Denote
t while (5) implies that
Gc (t) = (r, 0)B2 B2T (r, 0)dr (27)
0 gopt min zz L2 [0,tf ] + x (tf )T P (tf )x
x(tf ) =
B(t) = B1 Gc (t)E2T (28) x(0)
x (0), X(0)x (41)
h
x(0) +
m (h) = (h, 0)x w (r h)dr (29)
(t, r)B(r)w So Z(h) is positive semi-denite via (40) and (41). With
0 the similar lines of Z(h), we can prove that Z(t) 0, t > 0.
In terms of the above analysis, it is not hard to reach the
Then from (25) and (26), we have the results directly
conclusion.
tf Theorem 1. Consider the system (1) and (2) and the
y (h) = (s, h)E2Tw (s h)ds (30) performance (3). If > opt holds, then the H -like Ric-
h cati equation (37) has a positive semi-denite solution.
x (h) = m (h) Gc (h)yy (h) (31)
2.4 The H preview controller
From (29), we know that m (h) (only contains the initial Similar to the general time-delay systems, the system
data) qualies for the initial value. What is more impor- (1) and (2) can also be reformed as an equivalent abstract
tant, we can thus assume delay-free system[5, 15] . The H preview control (3) for the
m(t)
y (t) = Z(t)m (32) system (1) and (2) can thus be converted into a standard
H control. Denote f (t) = (x x(t), w
t ) as the abstract state,
at h because y (h) is determined completely by the optimal where w w (t + s), s [h, 0]}. Then, the H control
t = {w
w (t), t > 0, which is to be determined. for the abstract system generates the optimal game value
330 ACTA AUTOMATICA SINICA Vol. 36
x(0), w
like (x x(0), w
0 ), X (x 0 ) with positive semi-denite op- with A = 2, B1 = 1, B2 = 1, C = 0, and D = 1, where the
erator kernel X . X satises operator Riccati equation choices of C and D follow the aforementioned orthogonal
assumption.
1
X = X A + AT X + X B1 B1
T
B 2 B2
T
+ C T C (42) Because the nite horizon problem is considered, we will
2 compute the left-side of the performance index (3) instead
of the standard -iteration in our numerical experiment.
where For prescribed = 2.5, the state responses using the stan-
dard H control law and our H preview control law
d
A( , ) = A + B1 (h), () (11) are displayed in Fig. 1, and the control performance,
d
0 namely, the exact ratio of the left-side in (3) is shown in
Fig. 2.
D(A) = ( , ) M2 , () = ()d, L2 [h, 0]
The state x1 in the solid line in Fig. 1 is generated by the
B1w = (0, ())w w , B2u = (B2u , 0) H preview control law, and the state x2 in the dash-dot
line is yielded by the standard H control law. Given their
C( , ) = C , D2u = D2u amplitudes and response-speeds, the former, by contrast, is
more favorable.
Immediately, the -suboptimal H control-law can be
From Fig. 2, it is as expected that the H preview con-
given as
trol law (h = 0) leads to better performance than the stan-
u (t) = B2T X (x
x(t), w
t) (43) dard H control law (h = 0). Another fact is also seen
that the H preview control performance improves as the
As (43) is combined with (38), the following theorem holds. preview length h increases, it keeps nearly close to certain
Theorem 2. Considering the system (1) and (2), if level when the preview length h is larger than 1. In other
given > opt , then a suboptimal H control law satisfy- words, it is not the case that the more information, the bet-
ing (3) is given as ter performance. Such phenomenon is in line with Mirkin s
for 0 t tf h, research[16] and is the so-called saturation of performance.
u (t) = B2T (X(t) + (h, 0)T Z(t + h)(h, 0))x x(t)+
h
T T T
((r, 0) E2 + (h, 0) Z(t + h)(h, r)B(r))
0
w (t h + r)dr)
(44a)
for tf h < t tf ,
tf
T T
u (t) = B2 X(t)x x(t) + (s, t) E2Tw (r h)dr
t
(44b)
where B() is dened as in (28), and X() and Z() are
solutions to (12) and (37), respectively.
It is easy to observe that the controller in (44) takes
advantage of not only the state but also the preview in-
formation w (t + s), s [h, 0] at t instant. Meanwhile, Fig. 1 The state responses
compared with the general control law, the gain matrix in
(44) is unusual and involves the two Riccati equations (12)
and (37).
We proceed to discuss the sucient condition for >
opt , namely, the solvable conditions for the H preview
control problem.
Theorem 3. Considering the system (1) and (2), if
there exist positive semi-denite solutions to (12) and (37),
respectively, then > opt .
Proof. The proof mainly depends on the standard H
control conclusions[15] of the equivalent abstract system
and is omitted.
Theorem 3 together with Theorems 1 and 2 in eect
results in the following theorem.
Fig. 2 The performance improvement
Theorem 4. Considering the system (1) and (2) and
the performance (3), > opt if and only if that there
exist positive semi-denite solutions to (12) and (37), re- 4 Conclusion
spectively. Moreover, if > opt holds, (44) provides a
The paper investigates the H preview control problem
suboptimal H full-information controller.
in the nite horizon, and obtains a sucient and necessary
3 Numerical example condition for the solvability of the H preview problem.
Meanwhile, a simple control law is formulated. It utilizes
We have already proposed the control law for the H the future information of exogenous input so that it is able
preview control problem in the preceding section. In or- to preempt the eect of w when u and w are introduced,
der to illustrate that it is eective, let us consider the H and the performance is thus improved.
preview control problem for the scalar system (1) and (2) As a matter of fact, the H -like Riccati equation is
No. 2 WANG Hong-Xia and ZHANG Huan-Shui: Finite Horizon H Preview Control 331
achieved by the similitude transformation of a delay-free 10 Anderson B D O, Moore J B. Optimal Filtering. New Jersey:
H Riccati equation from of a mathematical viewpoint. Prentice-Hall, 1979
The transformation may change the original stable sub- 11 Tomizuka M. Optimal continuous nite preview problem.
space and also the solution of the Riccati equation, so the IEEE Transactions on Automatic Control, 1975, 20(3):
paper analyzes the rationality of the transformation, and 362365
veries that the solution of the H -like Riccati equation is 12 Blondel V D, Sontag E D, Vidyasagar M, Willems J C.
Open Problems in Mathematical Systems and Control The-
positive semi-denite. ory. London: Springer-Verlag, 1999
References 13 Zhang H S, Xie L H, Duan G R. H control of discrete-time
systems with multiple input delays. IEEE Transactions on
1 Marzbanrada J, Ahmadi G, Zohoorc H, Hojjat Y. Stochastic Automatic Control, 2007, 52(2): 271283
optimal preview control of a vehicle suspension. Journal of
Sound and Vibration, 2004, 275(3-5): 973990 14 Basar T, Bernhard P. H -Optimal Control and Related
Minimax Design Problems: A Dynamic Game Approach.
2 Hat A. Optimal linear preview control of active vehicle sus- Boston: Birkhauser, 1995
pension. In: Proceedings of the 29th Conference on Decision
and Control. Hawaii, USA: IEEE, 1990. 27792784 15 van Keulen B. H Control for Distributed Parameter Sys-
tems: A State Space Approach. Boston: Birkhauser, 1993
3 Hac A, Youn I. Optimal semi-active suspension with preview
based on a quarter car model. In: Proceedings of American 16 Mirkin L, Meinsma G. When does the H xed-lag smooth-
Control Conference. Boston, USA: IEEE, 1991. 433438 ing performance saturate for a nite smoothing lag? IEEE
Transactions on Automatic Control, 2004, 49(1): 131134
4 Kojima A, Ishijima S. Formulas on preview and delayed
H control. IEEE Transactions on Automatic Control, 2006,
51(12): 19201937 WANG Hong-Xia Received her B. Sc. degree in applied
5 Tadmor G, Mirkin L. H control and estimation with pre- mathematics in 2003, and the M. Sc. degree in computational
view part I: matrix ARE solutions in continuous time. mathematics in 2005, both from Harbin Institute of Technology.
IEEE Transactions on Automatic Control, 2005, 50(1): Now, she is a Ph. D. candidate in control science and engineering
1928 at Harbin Institute of Technology. Her research interest covers
6 Tadmor G, Mirkin L. H control and estimation with pre- H preview control, H control, and time delay systems.
view part II: xed-size ARE solutions in discrete time. E-mail: whx1123@126.com
IEEE Transactions on Automatic Control, 2005, 50(1): ZHANG Huan-Shui Received his Ph. D. degree from North-
2940 eastern University in 1997. From 1998 to 2001, he worked as a
7 Halpern M E. Optimal tracking with previewed commands. postdoctoral fellow at Nanyang Technological University, Singa-
IEE Proceedings D: Control Theory and Applications, 1991, pore, and from 2001 to 2003 a research fellow at Hong Kong
138(3): 237241 Polytechnic University. Since 1999, he has been a professor at
8 Bender E K. Optimum linear preview control with applica- Shandong University. His research interest covers optimal esti-
tion to vehicle suspension. ASME Journal of Basic Engineer- mation, robust ltering and control, time delay systems, singular
ing, 1968, 90(2): 213221 systems, wireless communication, and signal processing. Corre-
9 Zhang H S, Xie L H, Soh Y C, Zhang D. H xed-lag sponding author of this paper. E-mail: hszhang@sdu.edu.cn
smoothing for discrete linear time-varying systems. Auto-
matica, 2005, 41(5): 839846