Recently state space techniques for control of
Linear Parameter Varying Systems LPV have proved extremely fruitful [BP94],
[AG95], as have LFT based methods (cf Workshop ACC 99).
One of the best references for current information is [P].
We shall work with an input-output analog of these.
Consider the rather standard picture of a controlled
discrete time system with an uncertainty
.
We shall take
to be an uncertainty which is a stable
SISO system whose supremum norm over the unit disk is less
than or equal to
. We wish to design a controller K offline
of course and the ground rules presume that once the system is
running the controller can identify
, that is, we take
K to have the form
.
This
is in the spirit of
LPV control, since
in this approach the architecture in Figure 1,
the uncertain part
would be determined
by a parameter
and so would have the form
.
Then the usual set up has a controller
depending only on the state x and the parameter
.
Notice that once we design our controller K ,
if we are in a LPV set up, then at each instance of running
the closed loop system
is known, and so a sensible controller is
In LFT based control
would be determined as a time varying system
and the controller would change as
changes.
Remark
This contrasts with the time domain Linear Fractional Transform LFT
approach taken by Packard and collaborators
where the plant P and controller K
are both linear fractional in
.
Thus P and K have a much more restricted form
in this LFT approach than we use here.
We could do a frequency domain analog here
and assume a linear fractional dependence of
on
, see Section 4.1.
This is
reasonable, but requiring K to be linear fractional in
seems very restrictive and
probably leads to a very conservative answer.
The advantage of this frequency domain LFT method is that
the setup leads to a complex
synthesis problem (not convex but biconvex).
Our setup on the other hand leads to an LMI, but possibly an infinite
one.