Biomedical Engineering Reference
In-Depth Information
Sensitivities can be retrieved by solving this set of equations for i D 1;2;:::;KC1.
In particular, for our working example we have
LJ LJ LJ LJ Ǜ .j/
i
LJ LJ LJ LJ u .j/
Z
.j/
i
.j/
D
J R
i
@
J R
@ u
@
J R
i
D
C
D
. u d/ i C 1 Ǜ i ; iD1;2;:::;K
LJ LJ LJ LJ u .j/
LJ LJ LJ LJ Ǜ .j/
K
Z
.j/
K C 1
.j/
D
J R
D
@
J R
@ u
@
J R
i
D
C
D
. u d/. K C 1 / C 2 . ref /:
C
1
Notice that from the state equations ( 6.30 ), ( 6.31 ), we have for i D 1;2;:::;KC 1
LJ LJ LJ LJ u .j/ . i / D i C b r i C i C 3. u .j/ / 2 i ;
@
@ u
and
LJ LJ LJ LJ
.j/
F
i
@
Df i ;iD 1;2;:::;K
LJ LJ LJ LJ
.j/
@
@
D u .j/ :
Then, the sensitivities equations read
8
<
i C b r i C i C 3. u .j/ / 2 i D f i
in
K C 1 C b r K C 1 C K C 1 C 3. u .j/ / 2 K C 1 D u .j/
(6.35)
in
:
i D 0; i D 1;2;:::;KC 1
on :
Notice that these equations are linear in the sensitivities. Finally, we have
LJ LJ LJ LJ
LJ LJ LJ LJ
Z
Z
.j/
.j/
D
J R
i
D
J R
D
. u .j/
d/ i C 1 Ǜ .j i ;
. u .j/
d/ K C 1 C 2 .j/ :
D
D
Gradients of the functional with respect to the control variables following this
approach requires therefore the solution of the K C 1 sensitivity equations.
Gradient Computation Through Adjoint Equations
In the following, we omit the iteration index j for simplicity. In the previous section
we computed the operator @
LJ LJ LJ LJ u
@ u
applied to the sensitivities i . Let us consider the
 
Search WWH ::




Custom Search