Documente Academic
Documente Profesional
Documente Cultură
Question 1
First, we form the Hamiltonean for the LQR problem:
1
H(x, u) = (xT Qx + uT Ru) + pT (Ax + Bu)
2
To obtain the minimising solution we need to solve:
H
= 0 Ru + B T p = 0 u = R1 B T p
u
H
p = = Qx AT p
x
M (T )
p(T ) = = Sx(T )
x
together with the dynamic equations. In matrix form:
! ! !
x A BR1 B T x
= T
p Q A p
for t [0 T ]. Since this must hold for all state-trajectories given any x(0) = x0 , it is necessary
that:
P = AT P + P A P BR1 B T P + Q
Question 2
Obtain a spectral decomposition of the Hamiltonean:
! !
P1 P1
H =
P2 P2
1
where [P1T P2T ]T is a basis of the stable invariant subspace of H (e.g. the columns of [P1T P2T ]T
are the eigenvectors and generalised eigenvectors of H and is the part of the Jordan block
of H corresponding to all stable eigenvalues). Then P = P2 (P1 )1 . This may fail if we cannot
choose an n-dimensional (where n is the dimension of A) stable invariant subspace for H (i.e.
if H has eigenvalues on the imaginary axis) or if P1 is singular. Both conditions are satisfied if
(A, B) stabilisable and (A, C) detectable.
Question 3
(i) Using the suggested substitution,
u(t) 1 u(t)u(t) u2 (t) 1 u(t) 1 u2 (t)
p(t) = p(t) = = +
au(t) a u2 (t) a u(t) a u2 (t)
and
u2 (t) b u(t)
ap2 (t) + bp(t) + c = a 2 2
+c
a u (t) a u(t)
Hence
u2 (t) bu(t) u(t) u2 (t)
+ c = +
au2 (t) au(t) au(t) au2 (t)
Cancelling and rearranging
u(t) bu(t)
+ c = 0 u(t) bu(t) + cau(t) = 0
au(t) au(t)
which is a linear second-order differential equation.
(ii) Again applying suggested substitution gives:
u(t) u(t)u(t) u2 (t) u(t) u(t) u2 (t)
p(t) = (t) (t) = (t) (t) + (t)
u(t) u2 (t) u(t) u(t) u2 (t)
Hence
u(t) u(t) u2 (t) u2 (t)
(t)((t) + p(t)) = p2 (t) (t)((t) (t) (t) + (t) 2 (t)) = 2 (t) 2
u(t) u(t) u u (t)
Simplifying,
u(t) u(t) d
(t) + (t) (t) = 0 (t)u(t) + (t)u(t) (t)u(t) = 0 ((t)u(t)) = (t)u(t)
u(t) u(t) dt
as required.
Question 4
Here A = 0, B = 1, R = 1 and Q = 0, P (1) = S = 1. The ARE is
Z 1 Z 1 1
2 2 dp 1
p(t) = p (t) p(t) = p (t) 2
= dt =1t
t p t p t
1 1 1
= 1 t p(t) =
p(t) p(1) 2t
2
Thus
1
u(t) = R1 B T P (t)x(t) = x(t)
t2
Hence
Z x(t) Z t
1 dx dt x(t) 2t
x(t) = x(t) = [ln(x)]x(t) t
x0 = [ln(2 t)]0 ln = ln
t2 x0 x 0 2t x0 2
and hence
t
x(t) = 1 x0
2
and therefore
1 t 1
u(t) = 1 x0 = x0
t2 2 2
(constant).
Question 5
This is an LQR finite horizon problem with A = 0, B = R = S = I2 , and Q = 0. The
differential Riccati equation thus gives:
or in full form
p1 = p21 + p2 p3 , p1 (3) = 1
p2 = p2 (p1 + p3 ), p2 (3) = 0
p3 = p22 + p23 , p3 (3) = 1
The second equation implies that p2 (t) = 0 for all t. Then the first and third equations give
pi = p2i , i = 1, 3. Thus
Z 3 Z 3 3
dpi 1 1
= dt = 3 t pi (t) =
t p2i t pi t 4t
Thus !
1
4t
0
P (t) = 1
0 4t
1
and hence ui (t) = x (t),
t4 i
i = 1, 2. Integrating gives:
Z xi (t) Z t
1 dt dxi dt
xi (t) = xi (t) = =
4t 4t xi0 x(t) 0 4t
3
Thus
x (t) xi (t) 4t t
[ln xi ]xii = [ln(4 t)]t0 ln i = ln i
xi (t) = x0 1
0 x0 4 4
and hence
1 i t xi
ui (t) = x0 1 = 0
t4 4 4
for i = 1, 2 (constant).
Question 6
The ARE in this case gives:
2ap p2 + m = 0 p2 2ap m = 0 p = a a2 + m
Since (a, 1) is controllable and (a, m) is observable for m > 0 the stabilising solution must
be positive, and hence p = a + a2 + m > 0. The A matrix of the closed-loop system is
ac = ap = a(a+ a2 + m) = a2 + m < 0. Hence x(t) = exp( a2 + mt) 0 at t
and u(t) = (a + a2 + m) exp( a2 + mt), so that maxt0 |u(t)| = |u(0)| = a + a2 + m.
Note that as m increases the state transient response becomes faster and maxt0 |u(t)| increases
(as m asymptotically).
Question 7
The state-space model is
! ! ! !
x1 0 1 x1 0
= + u
x2 1 0 x2 1
The ARE is
AT P + P A P BR1 B T P + Q = 0
or
! ! ! ! ! !
0 1 p1 p2 p1 p2 0 1 p2 1 0
+ 8 p2 p3 +
1 0 p2 p3 p2 p3 1 0 p3 0 0
which gives:
2p2 8p22 + 1 = 0
p3 + p1 8p2 p3 = 0
2p2 8p23 = 0
4
to a solution P which is not positive semi-definite, so it is disregarded). The second equation
now gives
3
p1 = p3 + 8p2 p3 =
4
and !
3 1
4 4
P = 1 1
4 4
Question 8
Here, ! ! !
0 1 C 1 0
B AB = and =
1 0 CA 0 1
and hence (A, B) is controllable and (A, C) is observable. Hence we expect a unique symmetric
(real) positive definite stabilising solution. The Hamiltonean matrix is:
0 1 0 0
0 0 0 1
H= 1 0 0
0
0 0 1 0
Now !T !
1 4 3 1 4 1
P11 = =
4 3 1 1 4 3 3 1
5
and
! !
1 33 34 4 1
P = P2 P11 =
4 3 23 24 3 1
!
1 3 4 (3 + 4 )(3 4 ) (4 3 )(24 + 3 4 + 23
=
4 3 3 4 (4 3 ) (3 + 4 )(3 4
! !
3 4 (3 + 4 ) 23 + 3 4 + 24 (3 + 4 ) 1
= =
3 4 (3 + 4 ) 1 (3 + 4 )
or !
2 1
P =
1 2
since 23 = j, 24 = j and 3 +4 = 2Re(3 ) = 2. Clearly P is positive definite (eigenvalues
at i = 2 1 > 0) as expected since (A, C) is observable. The closed-loop A-matrix is
! ! ! !
0 1 0 2 1 0 1
Ac = A BB T P = 0 1 =
0 0 1 1 2 1 2
This has characteristic equation det(I Ac ) = 0, or 2 + 2 + 1 = 0 and hence closed-loop
eigenvalues are i = 22 j 22 as expected (stable eigenvalues of H).
Question 9
This is a standard LQR problem with R = 1, Q = diag(1, 1). Let the ARE solution be
!
p1 p2
P =
p2 p3
6
is the stabilising solution and hence
! !
3 1 x1
u = R1 B T P x = 0 1 = x1 3x2
1 3 x2
Question 10
The steady state Kalman filter equation gives:
2Pf Pf2 + 1 = 0 Pf = 21
so that
A LC = 1 ( 2 1) = 2