YouSub
Untitled

en English

Untitled English subtitles

0 downloads · 611 lines · added Aug 2, 2015

Report a problem
What's wrong?
Advertisement

English transcript

Hello everybody, we will continue with our lecture series for Optimal Control, Guidance
and Estimation. Last lecture we have seen LQR - Linear Quadratic Regulator design development
and as I told in the last class, we will continue several lectures on this topic. This is one
of the very popular optimal control techniques as well as a technique, which is already implemented
in industry and all that; it is so important, so we will give the importance in about three
four lectures actually.
So, this particular lecture, I want to cover some of these, first very brief summary of
this LQR design what we discussed in the last class. Then, we will formally prove the stability
nature of the closed loop system with a LQR control, so that will be through the optimum
theory and all that, so but formal proof we will give actually.
Then as a bi-product, we will also be able to say what is the optimum value of the cost
function in an infinite time problem and that is that is kind of closely link to each other
basically, we will see that. Then, we will have some couple of extensions of this LQR
design and you can see the literature, there are quite a few extensions possible actually,
but we are not going to talk everything, but we will give a few extensions which will kind
of give you confidence that you can given a problem, you can probably read, understand
and probably develop yourself as well actually. So, for those things will come, I mean this
four five examples that will give all extensions that will discuss here, first term is the
cross product term in cost function and that is something like X transpose U sort of thing,
if you if you have that kind of thing, then what you do. And then, there is a rate of
state rate of state minimization, that means you want to minimize X dot, similarly rate
of control minimization, you want to minimize U dot term and if there is also a extension,
which talks about some prescribed degree of stability, how do you execute that? So, this
kind of things we will we will discuss in this particular lecture, we will also follow
it up with the formulation for command tracking as well as all formulation for inhomogeneous
system. That means, if you have X dot equal to X plus V plus C then, this called inhomogeneous
system, because when X and U are 0, X dot is also 0, I mean a sorry when X and U are
0, then X dot is still C basically. So, that kind of things are called inhomogeneous system
and all that, we will see the all this extensions in this class actually.
So, first a brief summary of this LQR design what we discussed in the last class. The performance
index in this LQR formulation is something like this, you have half X f transpose S of
X f, which is the total thing is phi of X f plus integral t naught to t f per half of
this thing and then X transpose Q X plus U transpose R U. Several things we discussed
last class that R is to U positive definite, Q X is to U positive semi definite, S f is
to U positive semi definite like that and there are some guide lines about the selecting
of X of Q and R, all those things we discussed. And this also a fundamental requirement of
A and B pair has to be controllable as well as A and square root of Q that pair has to
be detectable or in general observable actually. So, this is the perform index that we want
to minimize subject to the path constraint, X dot equal to A X plus B U is a system dynamic
linear systems and then followed by boundary conditions, where X of 0 is X 0 which is seem
to be specified, t f is fixed and X of X of t f is free, so those kinds of formulations
we discussed in the last class. So, now went back and then started with this Hamiltonian
definition phi f phi f x f definition like that.
So, phi of X f is very naturally like this and Hamiltonian is l plus lambda transpose
f and l turns out to be this one and f turns out to be this one. So, l plus lambda transpose
A X plus B U, so we are ready now to apply our necessary conditions of optimality, which
is state equation, which is X dot equal to l del X by del lambda which is nothing but,
A X plus B U, f of X U basically. So, X dot equal A X plus B U and then we have this costae
equation, lambda dot equal to minus del H by del X, which is minus Q X plus A transpose
lambda. Then you have optimal control equation, which
is nothing but, del H by del U equal to 0 which leads to U equal to minus R inverse
B transpose lambda, we all derive this actually last class. Then finally, the boundary condition
turns out to be lambda f equal to del f by del X f and phi of X f is this one. So, we
can do this partial derivative and tell lambda f is to be S f X f. Then motivated by this
I mean this expression, where that t equal to t f lambda equal lambda is a linear function
of X f, so and as well as some other argument.
So, we happened I mean we assume that, how about how about lambda of t is something like
linear function of X of t at every point of time, that linear dependency can be different;
that means, P of t can be actually time varying function, time varying matrix rather, but
it is at every point of time, we still have this linear relationship actually. Then we
went back and it applied these conditions, whatever conditions we have, state equation
and costae equation and optimal control equation and we started with differentiating this expression
what we have.
So, lambda dot turns out to be like that, X dot we substituted as system dynamic, then
U be substituted as this expression, where lambda is again substituted as P times X actually.
So, we have then lambda dot was nothing but, this expression minus Q X minus A transpose
lambda, where lambda is nothing but, again P times X. So, we putting all together and
bringing everything to left hand side and assuming that X is not equal to 0 and this
expression is valid for infinite combinations of X and all that, so you can tell that coefficient
has to be 0.
So, this essentially leads to this Riccati equation, which is essentially is a differential
equation, differential matrix equation rather, with a boundary condition popping of from
this boundary condition. So, you have lambda f is P f X f which is equal to S f of X f,
so P f turns out to be something like S f actually. So, what we told is we start with
this boundary condition and we have a differential equation for this Riccati matrix anyway, so
you can back propagate it from t f to t naught and then from P f T for all t belong to t
naught to t f is available. Once P f t is available, then we can calculate lambda of
t and once lambda of t is available, you can calculate this U of T, so U is available for
lambda. So, on the on the way we also observe that
by Kalman’s theorem, when t f goes infinity for constant Q and R matrices, P dot happens
to be 0 for all t; so that means, this equation what you have as a differential equation,
it based on to an algebraic equation and this is what is popularly used all over the place
and by default, if t f is not mentioned then it is assumed that, t f goes to infinity.
So, this is the way to solve, but this we don't need to propagate anything, but still
it is a non-linear matrix equation, which needs to be solved actually, but there are
efficient algorithms available because this equation is heavily studied. So, once you
get a solution either this way or that way at P f T, then we have this control ready
which is U equal to minus R inverse B transpose lambda, where lambda is P times X. So, combining
this P to the side, you can tell that there is a gain matrix R inverse B transpose P,
and hence U equal to minus K X, where K is nothing but, R inverse B transpose P, so this
is the whole summary of the LQR design actually, alright.
So, with this idea or with this knowledge, we will proceed further and first we saw stability
of the closed loop system in LQR actually. Remember, when I am not talking anything about
finite term, infinite term this are all we will do discussions based on infinite time
formulation actually.
Alright, So, we have a closed loop system X dot equal to A X plus B U, which is nothing
but that and Lyapunov function V of X is X transpose P X. So, V dot of X is nothing but,
X transpose P X plus X transpose P times X dot, remember we are talking about infinite
time, so that means P is a constant matrix actually, so that P dot is not involved for.
So, we have X dot equal to A X plus B U again, so substitute that X dot equal to this expression
A minus B K times X. So, I can substitute it here, a minus X dot is nothing but, A times
B K, A minus B K times X whole transpose P X plus X transpose P times X dot again, so
X dot is nothing but, A minus B K X substitute like here.
So, that results to like if I expand this transpose and that is also X transpose A minus
B K transpose times B, so that you can substitute it back here, where K you know that sorry
K is nothing but, this R inverse B transpose P. So, that you can substitute back here,
R inverse B transpose P nothing but the K basically, so this similar thing, this is
also K. So, substitute back and then expand it fully and then do this plus Q minus Q thing
here, once I do this plus Q minus Q, you can observe that one part of the equation is nothing
but, the Riccati equation actually. So, then this being the Riccati equations, obviously
this X, this part of the equation will go to 0 and you will be left out with the rest
of the term actually. So, what is this? Now B dot you can is nothing
but, this expression actually, this is remember, this is already in quadratic form, so all
that we need to so and this is personately there is negative sign in both the expression
actually. So, we take out the negative sign and all that we need to show is, this V dot
needs to be negative definite function, so these are all Lyapunov theory based arguments
sort of thing. So, we start with a positive definite P, that is our Lyapunov function
that means, V of X is a positive definite function and we need to show that V dot is
a negative definite function. So, that is another reason why we need to select positive
definite free, I mean P may be multiple solution in this Riccati equation, but we need to select
a P which is positive definite actually. Anyway, so this being a positive definite, so this
positive definite function and V dot turns out to be this expression with if I take the
minus sign out, it will turns out to be Q plus this expression, PBR inverse B transpose
P actually, now you can already see that Q is a positive semi definite matrix.
So, if Q is a positive semi definite matrix, what about this expression? Remember R is
a positive definite matrix, so R inverse is also positive definite, remember if a matrix
has Eigen values lambda, then it is inverse is 1 by lambda is Eigen values. So, if all
this Eigen values are positive for R, there R inverse also will have positive Eigen values
actually. Now, this expression you can see B times B
transpose, that is a quadratic expression sort of thing and this is positive P already
a positive definite matrix. So, this will lead essentially lead to this entire P B R
inverse B transpose P is a positive definite matrix actually. So, you have a positive definite
matrix and negative that of negative that is negative definite and minus Q is also there,
so that means, that is negative semi definite. So, the total expression become something
like negative definite matrix actually, these all explained here, same thing what I told
for R; this notation what you see here is not just greater than, they are not scalar
quantities R is a matrix, so by symbolically it means R is positive definite matrix.
So, R inverse is also a positive definite matrix, P is also positive definite matrix,
this happens to be a positive definite matrix and Q is a positive semi definite matrix.
So, the combination of that is positive definite matrix and hence negative of that happens
to be negative definite matrix. So, V dot happens to be negative definite and by the
Lyapunov’s theorem, direct theorem actually that if V of X is a positive definite function
and in V dot of X happens to be a negative definite function, then the close loop system
is always asymptotically stable. And this is the this being the linear system and this
is like if you can extend further, these are all radically unbounded system, I mean radically
unbounded function and things like that and so essentially it happens to be globally asymptotically
stable. So, that kind of thing, but also remember
we are all talking about linearize system, linear system or typically linearize system,
so for the linearize system, the linearization is valid in a local neighborhood any way.
So, what you are talking here is, it is a global asymptotically stable for the linear
system, I mean not for the original non-linear system actually, so just keep that in mind
it is a comment sort of thing. As far as linear system is concerned, if you apply LQR you
will have global asymptotic stability guarantee basically and also it turns out that because
of this expression, this nature X dot equal to this is something like a close loop matrix
times X. So, you can talk about X of T is nothing but,
E to the power A minus K into T times X 0, so that is the solution. So, essentially it
turns out to be exponentially stable also basically. So, it is the strongest motion
of stability in non-linear systems theory, it satisfies actually, so it is globally exponentially
stable sort of thing actually. Any way, so this is I mean more on that, if someone is
interested what to see, what not to see and all that actually in the Lyapunov theory,
I encourage my other course which is already available in NPTEL program itself actually,
which is advance control systems theory, where I have discussed about two three lectures
on Lyapunov theory actually. So, you can see that for details somewhat actually, any ways
the summary is we have we have one V of X which his positive definite, for which V dot
of X tends to be negative definite. Hence by lyapunov’s direct theorem or sufficiency
condition and all, it turns out that the close loop system is always asymptotically stable
and hence it is globally asymptotically stable, because this is a radically unbounded function
and all and is also exponentially stable, so there is very strong motion of stability
actually.
Alright. So, there is stability there, so what about the minimum value of the cost function,
is it possible to estimate that? So, now you can go back and see this value of the cost
function is nothing but that, this t naught infinity we are talking about infinite cost
functions here, infinite time problems here. So, J equal to this half of T naught 2 infinity
all this expression, now U you substitute is minus R inverse B transpose B times X,
so that is the expression for U and this is U transpose.
So, you expand that again X transpose Q X from this side and X transpose P is a symmetric
matrix, so P transpose is P and then B, because B transpose is B then R inverse coming from
here. So, essentially you have this expression, I mean if you if you expand this expression,
you have this X transpose and then P transpose, which is P, then B transpose transpose which
is B, then R inverse transpose, so that is R is also symmetric matrix. So, R inverse
transpose is R inverse times R, times this matrix you have actually here. So, it turns
out that, this is identity you can talk about taking it out and hence we left we are left
out with this expression, half of X transpose Q, Q coming from here and then X transpose
here and X is there, so that goes out, and then you have this P B R inverse P B R inverse
B transpose P, so this is expression actually. But, what is it? I mean this expression what
you have here is nothing but our V dot, V dot is negative of that actually; so, minus
V dot happens to be that basically. So, we put it that expression that this is minus
V dot, so integration of T naught to infinity minus V dot times d t, remember this is D
V by d t sort of thing actually. So, this integration and differentiation will cancel
out and left out and we are left out with this V, half of V T naught to infinity and
this is minus half of X transpose P because X transpose P X because V is that that one
we selected that actually. So, put it back and then it is T naught to
infinity, so we evaluate it actually. So, X naught transpose times P times X naught
minus X infinity transpose P times X infinity, but we just proven, just this is a asymptotically
stable, it is always asymptotically stable, so that means, X of infinity is equal to 0
basically. So, with that term analysis and you are left out with that; that means, if
you just know the initial condition and the Riccati matrix solution, you straight away
have a value of the cost function actually, really I mean if you operate based on that,
you are supposed to get this and probably this expression will help you in cross validating
your results or kind of verifying results and all that actually.
So, this is minimum value of the cost function and now some extensions of LQR design, this
is I mean LQR procedure, we know by now and infinite time procedure is also we know, we
are guarantee to have globally a asymptotic stability for linear systems, stability theorem
we proved and minimum value of the cost function also we derived actually.
Now, we see some extensions of LQR theory, LQR design; now let us see one by one actually.
So, the first thing is cross product term in performance index. So, remember for originally,
we had the performance index X transpose Q U plus U transpose R U, we did not have a
cross product on something like this, X transpose W U. So, this W is also like a matrix sort
of thing but, remember X and U are not of same dimension; that means, W need not be
a square matrix actually. So, with this expression let us analyze, how
do you do that and the whole idea here is you can always go back to this Hamiltonian,
try to derive the necessary condition and try from the beginning and proceed further,
as we proceeded before, but that that turns out to be not necessary actually; if we manipulate
this expressions little bit, then with existing solutions available, we can actually solve
this problem as well. Now, how do you do that? Now, just see this
expression whatever you have and then let us start with this expression, if somebody
starts with this big expression and you try to simplify. So, this first term will be X
transpose Q times X, the first term over here is U transpose R times U and some of the terms
and all will cancel out and you will be left out with this term actually and then ultimately,
it turns out that this is nothing but, X transpose Q X and minus see this one it is a scalar,
ultimately all this things are individually scalar quantity.
So, the same expression times its own transpose equal to that actually, any time they were
matrix multiplication is a scalar quantity, you know that scalar transpose is a scalar
value itself. So, we can tell that this expression X transpose W U whole transpose is equal to
X transpose W U. So, whatever we have this expression is a transpose of that, so you
can tell that this is nothing but, same thing of that. So, that turns out to be 2 times
this actually; that means, whatever expression we had here, we can actually substitute by
this long hand expression actually. Now, somebody can ask me, we have in how you
got it and all, so this certificatelly does not happen overnight, but the idea here turns
out to be in the reverse direction, you start with this expression here and this is now
easy because you can talk about 2 terms like that and one term you keep it as it is, the
other term you take a transpose of that. So, coming from third line to second line is rather
very easy actually, but here we do little bit more manipulation, we add and add and
subtract the term, try out I mean a couple of times and then and finally, you land up
with this expression
…

Need subtitles from another video?