YouSub
Untitled

en English

Untitled English subtitles

0 downloads · 620 lines · added Aug 2, 2015

Report a problem
What's wrong?
Advertisement

English transcript

We will continue further in our lecture series and many times in our control theory as well,
we will need lot of numerical methods on the way, especially with respect to non-linear
systems theory the things are not available in close form to that and in linear systems.
Also, remember, that we need some operating point first before we talk about deviation
dynamics, which we linearize and then interrupt, that is, linear system. We will talk about
linearization in one of those lectures later, I mean, about the linearization method in
its useful. And then we will see some of these concepts, other concepts also like root finding
and other things are useful in lot of application actually.
So, let us quickly review some of these numerical methods that are frequently, I mean, kind
of required in our analysis and synthesis actually. So, we will, first concept, that
comes to my mind is, first problem is linear equation solution. Suppose, you have A X equal
to b and remember, we do not want, we know for sure, that if A is some rectangular and
things like that, I mean, A, A is square matrix for which determinant of A is not 0. That
means, A is non-singular and b is not equal to 0, then X equal to A inverse b, we know
for sure, but A inverse is divided by determinant of A and that requires a lot of computation.
We do not want to go to that definition all the time, whenever that is required actually.
Now, the question is, why do we require this kind of solution in our control theory? And
one of that is you can see some of these, I mean, some of these motivation part of it.
Suppose, we want to find out something like, like this, this is a typical equation, X dot
equal to AX plus BU and we, let us say we divide this X into controlled state and uncontrolled
state, like X c is controlled state, X n is uncontrolled state and we divide it in such
a way, that it is like dimension of X c is equal to dimension of U.
So, if I just consider, that upper portion of the equation, that means, X c dot is equal
to A 1 X plus B 1 U. Then, I tell this B 1, remember the dimension of X c is equal to
dimension of U; that means, B 1 is a square matrix now. Then, I will probably try to find
out what is equilibrium condition? That means, this X c dot is 0, right. If I take about
the controlled state equilibrium and I want to interrupt that as a forced equilibrium
point, that means, my control is not 0, so that I want to maintain, that particular equilibrium
point. And then, what is my control required for
that? It is also like what is called as trim control in aircraft actually, like if your
aircraft flies level, you really require some sort of a trim movement because remember,
your and C p are not at the same point. That means, there is continuously, there is a gravity
related. I mean, this aerodynamic lift related movement, which will act on the vehicle actually.
If I, if I just turn that in a little small diagram, see something like there and your
is somewhere here and C p is somewhere there. So, that means, this, this particular thing
will give us some sort of a movement about that actually. So, this is turning movement
sort of thing. So, you want to cancel that by using something
like an equivalent movement actually. So, that means, you really require some sort of
a small elevator deflection actually. So, those control trim condition is actually a
post equilibrium point in aircraft plane and those cannot relate to that, we will see that
later. And that is very obvious, it is not too much, you can see that this problem from
other example problems as well.
So, what I am looking at? I am looking at some solution for a, for something like X
c dot is equal to 0. That means, the control states are operating under steady state, actually.
If I really want to find out what is my control necessary for that particular situation, then
I can find a control. That way, if I remember, B is actually, B 1 is actually non-singular,
square and non-singular. So, I can take that way.
If I apply this control equation, control, then I will ensure, that this happens to be
0, so that my, my aircraft continues to fly at level, that particular control is actually
necessary for maintaining X e dot 0, X e is X dot for example, and BU is delta e, elevated
deflection. Then, this elevated deflection I have to compute it that way, so obviously,
we require this tabulation. Now, this is A 1 of X, this is nothing, but
our B, if we are going back to that, this is our B and this A is nothing, but B 1 actually,
in this case. So, we require this kind of solution, I mean, efficient way of getting
solutions and things like that. Remember, this control has to be applied online also,
that means, we cannot afford to have lot of computational time getting wasted for computing
this B 1 inverse. We do not want that actually, we want to have an efficient way of controlling,
I mean, finding out this B 1 inverse, actually.
So, coming back to that, can we... So, this is the equation, that you are asking and then
we observing, that A inverse involves too many computations actually and computational
complexity sense, it requires n square into n factorial computations. That means, if A
is larger and larger and larger, you are getting trapped in this computational actually. So,
you want to do it in an efficient way and one of that efficient way happens to be like
elimination, which we will see in this class, actually. Now, this also, this approach is
not just suffers from the computational headache, I mean, it also suffers from these, these
other problems, like what is one thing is called ill-conditioning, remember A inverse
is nothing, but, but adjoint of A divided by determinant of A; A inverse is equal to
adjoint of A divided by determinant of A. So, determinant of A goes to 0.
That means, if you have a singularity is approaching, you are approaching singularity, then this
is a serious problem, actually. You cannot just talk about division actually. And also
remember, there will be like so many computations and lot of computations, all these computations
will have small round off errors, the too many computations, then round off errors can
be large actually. So, you do not want to do too many computations in any situation
basically. The computers are always like, finite length you can take, I mean, that number,
you know, that finite digits only, then round of errors all the time, actually, so, as you
do not want to do too many computations, actually.
Now, Gaussian elimination is a substitute to that instead of general things. We will
just take it as a small example and try to see what is going on here. So, we have this
kind of an equation, let us say 3 by 3 and then, I mean, skip all these again, again
small mistake here, all these are small x, x 1, this is small x 2, small x 3, these all
are elements, actually. Anyway, so this equation, what we want to do?
We want to reduce this equation, whatever equation you have in this A matrix to be a
upper triangular matrix, actually. And we take advantage of the fact, that you can multiply
by any row with, by any constant and then take addition of of other row, then the equation
will not change actually, equation remains same. So, then, we will take advantage of
that thing and then, tell we want to reduce this matrix to a upper triangular form actually
by taking advantage of this row transform, I mean, row properties, actually, like row
multiplication, addition, things like that actually. So, what you do actually? We will
try to see the, this, these elements, whatever you see here, 1, 1, 0, this has to be all
zeroes actually. Now, 1, 0, already there, so I want to make sure of this 0 also happens
here or this 0 happens here, either way. So, now I see, how do I multiply this particular
thing by minus half, one-half and then add it off, actually. That means, I want to make
sure, that it becomes 0 itself. If I do that, then 2 by 2 is 1, then 1 minus 1 is 0, basically.
So, that will pop out there corresponding to that. Now, you say 2 minus one-half actually,
again 2 minus one-half is 3 by 2 like that actually and 0. We will not change this, that
is, the equation that I will leave it actually. Again, 2 minus one-half is 3 by 2 actually
here. Now, you see, that this is already 0, 0, but, but I have to make sure, that is also
0. So, I will multiply this element, this row by 2 by 3, this time and then subtract
it actually.
If I do that, then it will, I will end up with something like this actually. Now, this
is an upper triangular matrix, so we start looking at the equation from down to top,
actually. So, then it turns out, that minus one-third
is nothing, but minus one-third of x 3 is equal to 3 and hence, x 3 is 9. Once I get
x 3 value, then I put that this particular is nowhere and I mean is a function of x 3,
this particular, remember this let us do that, this, this particular thing has given me x
3 is equal to 9. Now, if I do this equation, rather second, second thing, then I get 3
by 2 x 2 plus 9 because x 3 is 9 is equal to 3 by 2. So, I, I tried to solve this and
if I try to solve, then x 2 happens to be minus 5, actually.
Now, I go back to the first one equation. Now, first this, this equation now and tell,
this is 2x 1 plus x 2, x 2 is already minus 5, so this minus 5 equal to 1, right. So,
this, then 2x 1 turns out to be 6 and x 1 is equal to 3, that is why, you get x 1 equal
to, 3x 1 is equal to 3x 2 equal to minus 5 and x 3 is equal to 9, that is easy now. So,
what is, what are you doing here? You are somehow trying to eliminate these, these are
called pivotal elements and make it 0, 0. And then, once you get an upper triangular
matrix sort of thing here, you try to look at the equations from bottom to top because
then they try to solve, solve this equations straight forward actually. So, that is Gauss
elimination.
And now, Gauss eliminations suddenly make this n square into n factorial operations,
this reduces to just this much operation actually. Remember, this, this is actually something
like a very high computational thing to a very low computational thing. Remember, n
cube is nothing, but simply n, not n factorial here and that will get multiplied with n square,
which is a lot of computationally. So, obviously, it leads to far lesser computation
and then, hence it is kind of referred actually, especially, especially in metal of command
window, if you, some of you use that, then you have a choice of using INV function, INV
into B, which is like X or the same thing you can do that using this slashes, a slash,
that will give you Gauss elimination, actually. So, instead of using the first one, I will
suggest, that you use this one actually, A slash B, alright.
All is, all is well here, but, but you see some problems actually. So, first problem
is, so this, this Gauss elimination method will also encounter problems if the pivotal
elements happen to be zero, actually. I mean, that one of the diagonal element becomes,
suppose this is zero, now, sorry I took it there, suppose if you do this exercise and
this happens to be zero, now no matter how much you multiply and try to add and subtract,
nothing will happen to this one actually, not be able to do that.
So, in this, those situations, one easy way to do that A is just to exchange the rows
actually. Suppose, if it, this one happens to be 0 here, then you take you take the entire
equation, you put in a 3rd equation and this one is substitute for 2nd equation, you exchange
the equations actually, then proceed further. That is the idea there, actually.
So, this Gauss elimination is a very standard practice now and lot of people make use of
that again, a metal of window, I mean, metal of command, metal of environment. I suggest
that you use this, this one. Anyway, now going back to the next concept.
What we saw there is algebraic equation solution for linear systems. That means, we are talking
about, this is a linear sort of equation actually, right. AX is equal to B, what you, what you
have is actually a linear equation. Now, what if you have a non-linear equation? Remember,
numerical methods are more powerful in linear cases and all that, alright.
Now, let us see something like this, F of X equal to 0, and you want to find out what
is X equal to? And again, the similar motivation there, like you can, I can partition this
state’s, whatever X dot equal to f of X, U into two parts, I still maintain that condition,
dimension of X c is equal to dimension of U. Then, in this case you land up with, like
X c dot equal to 0. If I want to entrust that question, then I have this non-linear, this
non-linear equation now. So, f c of X 0, U 0 is equal to 0 that is the condition that
I want to find out. Now, that means, now can I find out some U
0, which will satisfy this equation, this equation, this can be a function of X 0, that
is, that means, if I know a particular number for X, that is, X 0, what is the particular
number for U, which will give me that and that is why this 0, 0 notations are there,
actually. So, I will go back to this equation now, f c is of 0, f c of U equal to 0, and
try to find out what is use for that. That is the motivation why want to address this,
why we want to address this particular problem actually. This is just one of the examples
why you need that. You may need for many different cases also actually.
Now, we will start with the simple solution sort of idea. First, we will see what is a
scalar case? By the way, this has many, many numerical methods available, one of that,
which is popularly known is bisection method. That means, if you, if you talk about the
scalar expression here, for example, and you tell I have two guess values now, then there
is something called bisection method and there are very, very other ideas as well, actually.
So, this is not what I am talking here, is not an exhaustive review of methods available.
So, what I am doing really, doing here, is rather talking about what is called as Newton-Raphson
method, which is, which is very neat, it has its own beauty physically and that is what
is mostly used in practice. That is why we want to review this particular method of getting
the solution actually. So, what you, what you really want to find out in this case?
Remember, non-linear equation solution will require some sort of iterative solutions actually,
that means, in one go we will not be able to find out, but, but we may need some couple
of iterations before arrive at the final solutions, actually. So, what is the idea here?
Now, let us say that we are having some x k value, some value for x k is already available
and k can be 0 also, that means, you can start with some initial guess, then if k is 1, 2,
3, 4 than the already iterated values available. Obviously, that is not the solution; that
is why we want to do iteration actually. That means, f of x k is not really 0. However,
you want to find out delta x k such that if I make x k plus delta x k, then f of x k delta
x k is equal to 0. That is what that mean, that means. If my delta x k is proper, then
x k plus delta x k will be a root of this equation actually and that is the whole idea
there. Now, I will interpret this, f k, f of x k
plus delta x k in Taylor series expansion and tell these are all higher order terms
and all that, I will not be interested in that, I will suppress that and I will just
take the 1st order term, up to 1st order term and this is what my equation actually. Remember,
this entire expression is zero because that is supposed to be a root actually. That is
what you want to find a delta x k, such that f of this x k plus delta x k equal to zero
basically. So, that is your aim actually, that is, how we want to find a delta x k.
That means, this is equal to 0 up to that actually, up to this term, then I can, I can
relate, that delta x, this particular term is zero. Hence, this del f by, I mean, d f
by d x evaluated at x k times delta x k, this term, whatever term is there, is nothing,
but minus f of x and assuming that is f dash of x, remember that, evaluated at x k.
So, assuming that, that is non-zero, I can do that actually. Delta x k is nothing, but
this term, this whatever term you see here, that is nothing, but, but delta x k actually.
This particular term is nothing, but, but delta x k. So, if this is delta x k, then
this is my, what I really wanted is x k plus delta x k, so that is my x k plus 1. So, this
is, this is this particular term. What you see here is x k plus 1; that is why if f of
x k plus 1 is 0 basically right, that is what you want it. So, this is why you will get
x k plus 1 is x k plus delta x k and delta x k is nothing, but that actually.
So, x k, x k plus 1 is equal to x k minus f of x k divided by inverse of x k, that is
the, that is the algorithm and you keep on doing that until you get some, some convergence.
That means, that you do not get too much, I mean, very much of improvement afterwards
actually, that is, up to that you do that and stop.
Now, picture, really, let us see what is going on here. You have sum f of, sum of x, sum
of f x, arbitrarily it is plotted something like that, f of x is equal to 0 and assuming
that, that is the zero-axis of course, of course it starts with 0, 0 probably, and then
obviously, what you want? You want this particular value, this particular value, this is your
root actually, this is what you really want. Ultimately, what you do not know? The value
to start with, actually. So, you start with some sort of a guess value, this is my guess
value, I will start with some guess value, let us say, then f of x is that most, that
is my f of x, then f of x. f of x is nothing, but the slope actually, evaluated at x. So,
that is the slope what is giving me actually. So, if I compute this delta x i sort of thing,
and if you see this is actually a linear equation. So, what you are doing here? x i plus 1 is
x i minus this term actually and if you just carefully look at it, it is kind of a, to
start with this equation line, I mean straight line equation and try to find out the interpretation,
the meaning of that. That means, it all that tells me, that I have a guess value, I go
to that particular point where it intersects this particular function, then evaluate this
slope and with that extension of this slope I will cut this x, I mean, this x-axis somewhere
and that is going to my next iteration value. So, again, I evaluate the function there,
I again take their local derivative, extend that and wherever it cuts, it is my i plus
2. Then, again I put, go there and I will proceed further itself like that and ultimately
I will converge there actually. So, that is the interpretation of that.
Multi variable case, this is similar extension we can do. Now, if you want this f of x is
equal to 0, you tell, while I will do the same exercise. That means, I will take multi
variable Taylor series in expansion and then I will tell, I mean, higher order terms I
will neglect and I found I want to find output delta X k in such a way, that, that X k plus
1, this is my X k plus 1, anyway, this is my X k plus 1, that is going to be a root.
That means f of X k plus 1 is going to be zero, actually.
So, this entire thing, what I see here I interpret, that I will find delta X k in such a way,
that this is equal to 0. Now, if that happens, then I define this fellow as A k plus 1, I
define this del F by del X evaluated at k is nothing, but, but A k is something like
that and then, the equation tells me, that A k times delta x k is nothing, but, but negative
of this particular thing, F of X k. And hence, delta X k is nothing, but A inverse, A k inverse
of F k. And remember, when you see, whenever you see this kind of a thing, you can use,
you can use Gaussian elimination. You can use Gauss elimination actually to compute
that in a faster way, then you can update and then you get X k plus 1 is X k plus delta
X k. So, very similar to what you have here, instead
of divided by, if
…

Need subtitles from another video?