Advertisement
English transcript
Hello everybody welcome to this lecture four of this course optimal control guidance and
estimation. So far, we have seen some overview of this course, followed by some I mean some
overview of topics related to system theory as well as matrix algebra and in numerical
methods as well; that is relevant in this course actually. So, in the another thing
that is very much relevant in this course is the concept of static optimization; that
means, we are not talking about time varying variables and all, but these are constraint
numbers associated with that. But in general they have some objective, and
all that, how do you optimize that actually. So, that kind of things are call static optimization
or sometimes it is popularly known as a parameter optimization as well, because parameters are
something that typically remains constraint actually. So, anyways those are the topics
of discussion of this particular lecture, and also remember as I told you in beginning
lecture the very fast lecture that one way of solving this, this optimal control problem
is to go heavily to this static optimization ideas and convert the problem entirely to
a large dimensional static optimization problem, and solve it; that is called transcription
method, and all we will see that as we the course develops.
Another motivation of that is even if you formulate this optimal control problem from
calculus variation point of view and all that; the essentially land up with some equation
something call Hamiltonian, and all that; and that happens to be like every point it
happens to be kind of a constant quantity in its variables state and control, and that
you need to be minimize at every point of time in the frame work of static optimization
actually is very relevant concept for optimal control in general. So, it obviously, make
sense to have overview of static optimization before we go to the dynamic optimization;
that means, optimal control in this is now actually.
So, this I will take it in two parts and the first lecture will talk about some overview
or some standard results of static optimization along with some examples and all that way,
then further topics on it will see in the next lecture, and before proceeding to the
calculus of variation ideas and all in that way.
So, the topics of this lecture something like this, we will fast discuss about unconstrained
optimization, then we will go to this constrained optimization with equality constraints primarily
any quality constraints, and all will see in the next lecture actually. Then we will
some numerical examples as well. So, first is unconstraint optimization and
the problem definition turns out to be very clear, that you want to optimization some
vary some function of variable, but the variables are combination of some non time varying quantities.
Actually that is they are not constraint by any constraint equation neither actually before
proceeding with those with then algebra, and then analysis and think that; let us see very
simple straight forward idea first actually.
Now, forget it for a second this j 2 dot line, and all this flat was this flat was x verses
some functions j 1 it is a, and j 1 turns out be this function this solid function actually.
And then very quickly we can see that this 0, 1, 2, 3, 4, 5; they mean something to us
in a in just by looking at the flat actually, and very easily it can be argued that this
0.1 happens to be a local maximum, and point 2 is well we cannot say anything about that
it is a flat thing, but certainly a candidate for getting worried about, then is called
point of inflexion, and all that. And then point 3 turns out to be a local minimum point
4 turns out to be local maximum just by looking at the flat, we can very easily see this is
what is happening actually. And then it also we can see that point three
turns out to be global minimum, and point four turns out to be global maximum because
nothing else is bigger that in this domain of course, I mean x a 2 x a. And all in that
way, but when we talk about global maximum and all typically this x a and x b are also
not there; in other words we are talk about minus infinity plus infinity that domain sort
of thing actually. But anyway within this constraint, and then
you can think that 4 happens to be the global maximum and 3 happens to be the global minimum
actually, and one property that we observe simply observe is that all these points 1
2 3 4, this relationship is true; that means, the gradient of j with respect to x is 0,
if I just take a tangent here that is flat this is any way flat if I take it engine flat
here is a flat. And then it is if I take it engine at the fourth point then they also
parallel to the x axis actually. So, all the points satisfy that.
Now, that it gives is the clue that when we talk about something about optimization then;
obviously, the first derivative turns should happen to be 0 that is how to do actually,
and does not really happen. So, let us the talk about that that way and also remember
that instead of j 1, suppose you have j 2 then really do not have to worry about all
these we just have to worry about one minimum; that is what all that will have is one minimum,
and that one minimum happens to be the global minimum as well actually.
So, most of the time will take advantage of this observation, and value constructs a function
we will want to construct it that way. So, that we really do not want to bother about
global things, and all is not relevant also most of the time it satisfy our need also
basically is quadratic function and all that we will talk later this is how it turns out
to be now coming back to that whether this relationship satisfied for all the points,
and all and analyze it little bit mathematical vigorous sense basically.
So, now again we go back to this a this Taylor series, and all which will repeatedly require
anyway now remember what is a local maximum local maximum tells you that if point one
is local maximum, then very next two it whether go left side or right side either way that
will give a function value which is lesser than the function value at point 1.
Similarly, if it is point 3 if it is minimum local minimum, then whether you go left side
or right side either direction you are going to get higher values, then what you get it
point three actually. So, that kind of constraint is there.
So, let us analytic a point x star and then talk about a value of that function around
x star, and remember delta x can be either positive or negative actually here, and also
remember we have simply talking about a scalar quantity; that means, x is just single I mean
single component j cont j is a function of x, but x does not contain x 1 x 2 components
and all that x is a scalar variable sort of thing.
Then what happens x star plus delta x minus j of x star that is the term that will come
from Taylor series like this, actually now for any hope to x. What we are looking for
a looking for this quantity this j of x star plus delta x should be greater than j of x
star if j of x star happens to be minimum point irrespective of the direction of delta
x at the sign of delta x really actually. So, that has to be happen; that means, this
left side quantity has to be strictly positive for a for this quantity to be on a local minimum
or precisely be strictly negative for the other case actually local maximum case.
So, what we are looking at here is a sign independent property really and also remember
this Taylor series has a this uniform convergence property, and think like that. So that means,
if you right this expression to be something like that if you take a this entire quantity
what we what turns out be like this, the infinite terms what we are talking here is a less dominate
compare to the first term actually. So, that is how it will turn out be… So,
in other words the sign of it is largely dictated by the first term itself actually not by these
terms actually that way. So, if you have any real hope that the this become sign in sensitive,
then it turns out that the first term must be going to 0, then only we have some hope
actually that is how the necessary condition turns out to be and remember delta x is not
0, but this quantity has to be 0. That means, the coefficient has to be 0 basically that
is how it turns out that del j by del x; that x equal to x star is 0 is 0, then x star happens
to be a candidate point I mean you can you still cannot it whether it is minimum maximum
actually. So far that we will remain we will go to the
next one, and then tell if that that is already there; that means, at this point this is 0
anyway, and then this quantity turns out be that one.
And that one means, I can consider the second order term, and then there is a still higher
order terms a from third order fourth order can neglect actually. If this still has be
positive; that means, this has to be greater than this irrespective of the sign of delta
x remember this square term now. So, this has to be strictly greater than 0 provided
this is greater than 0 this is greater than 0 anyway.
So, this quantity is greater than 0 provided this quantity happens to be greater than 0;
that means, second order term has to be greater than 0 basically, and similarly the second
order term happens to be less than 0 it least to the conclusion that this term is less than
0, and other words this term happens to be maximum actually x star happens to be maximum.
So, this is how the how the analysis two to summarize the things what we have doing here
is given any function j of x, we will we will put this del j by del x is equal to 0; and
then solve this and find out candidate points x start did not be 1 point it can be several
point actually and but each of the points
We will try to evaluate this second order derivative and the second order derivative
happens to be 0 then it is a minimum point the second order derivative happens to be
less than 0 it is local minimum that is the sufficient condition actually. Now, we are
carrying out further things, and all let us see this some examples sort of thing, and
before even before the example. Now the question is what if this both derivatives of the 0,
but third of derivative is not 0. So that means, this is true whatever we are
talked about only if the second order derivative is non zero, and the non second order derivative
also happens to be 0, then we have no choice what we go to the third derivative; and the
third derivative happens to be non zero. Then we really cannot tell whether it is strictly
positive or strictly negative, and hence it turns out to be point of influence actually,
and if it is I mean if you still have a hope of getting it something like a positive or
negative; it we must have the third derivative also 0. And then go to the four derivative
and think like that actually right, if it is second derivative is 0 then this lands
up with the third derivative case, and then there is necessary condition of optimality
is if we still any hope then third derivative has to 0, then you land of with four derivative.
And if the fourth derivative happens to be non zero; that means, positive, and then it
is least to be minimum and maximum and otherwise it is maximum actually.
So, the analysis continues it actually toggles between this hard power, and even power sort
of thing actually. So, as a necessary condition all hard power has to be 0 and sufficiency
condition will come even powers actually.
So, this is how the things proceed, and suppose we want to get example here then turns out
to be let us take a very simple example j equal to x fourth, and you can see that well
the x fourth plot is very, very close to these this kind of very similar to x sort of thing.
So; obviously, we know that this is point of minimum; that means, x is equal to 0 is
the answer we know that actually whether we get it or not let see that using this result.
So, if you have j something like this, then del j by del x is four x is equal to 0 then
solution of that is three quantity is, but all happens to be 0 that x star equal to 0
0 0. And then let us go to the second derivative, and second derivative turns out to be like
the taking this term it is twelve x square and x star equal to 0 means twelve x square
is also 0. So, we really have to go third derivative
now, and the third derivative personality happens to be 24 x from here twelve into twenty
four into x star next star is 0; anyway the third derivative is also 0. So, that means
that is still our hope is alive actually, then you go to fourth derivative, and four
derivative it happens to be twenty four which is greater than 0. So, if the fourth derivative
is greater than 0, then obviously, it tells you that my some of my all of my hard powers
also 0, and the first even power that I uncounted is actually greater than 0.
So; obviously, x star is a minimum point actually; that is how we include that the result is
compatible to what we know actually come from graphical.
Since, now example to what if x is j is just x q and remember x q, what is the type of
I mean thing plot x q is something like this, if it is positive is positive, if it is negative
is negative; something like this. So, here you have a point of inflexion really; that
means, at 0 what we are talking here is one side, the second derivative will turn out
to be I mean one side the function turns out to be increasing other side the function turns
out to be decrease actually. So, this will this we know that it is a point of inflexion,
now whether our analysis tells of the tells do not let see.
So, if you have j equal x q then del j by del x is nothing but three x square is equal
to 0, x star turns out to be 0 0, now we have but two candidate solution, but both of happens
to be anyway. Now, the second derivative happens to be six x square is also 0. So, we cannot
conclude anything this point of time this is three x square. So, it will be six x. So,
six x evaluated x star that is six x star x star is 0. So, it is still 0 actually.
Now, we go to third derivative and third derivative del q by del q j by del x q from coming from
here is nothing but six, but the important point note is not whether six is positive
or negative, but six is not equal to 0 remember this third hard derivative actually hard power
derivative, but this hard power derivative has to be non equal to I mean not equal to
0 that is more important, actually if it is not equal to 0 your hope is not alive it conclusion
as a already in there and the conclusion is the x star is point of inflexion.
Because if I take this series this series this terms will start dominating and delta
x is positive defending on whatever sign is there it will lead one sign, and defending
on whatever sign is here if delta x is negative to the exactly opposite say conclusion here.
So, is just one side is a either increasing or decreasing, but other side is just the
positive actually. So, will turn out be point of inflexion. So, that is how it is there.
So, if j equal to x q then x star equal to 0 turns out to be point of inflexion point
actually a point of inflexion really. So, this all comfortable to what we know basically
now, but this is not kind of story this is just very beginning it is talks about as scalar
variable functions sort of thing actually, but we are not interested in scalar variable
function.
In general because we delay system theory, where state can be n dimensional control can
be m dimensional like that. So, we takes it is a mixer of this components
of the state variable really. So, that is what we want analyze then leads to this vector
case, and the problem is turns out to be like that minimize j of x where x is n-dimensional
vector, and remember the objective function is always is a scalar quantity always otherwise
is talk about multi objective of optimization and all that and we are not talking about
that at all in this course actually. So, when somebody talks about standard of optimization
problem then it turns out to be that the optimization function is a scalar quantity, where is the
variables can be n-dimensional there can be several variables playing is playing out different
contributes, and all we want minimize this scalar quantity. Finally, and also remember
this is a this is a non constraint problem; that means, this component of x are not constraint
by any constraint actually you can they can take any values what so ever actually that
way. So, you want analyze that problem and see
what is going on actually and also remember phi definition what is saw in this matrix
theory is reviews, and all that also that by definition this is del j by del x equal
to like this by this is nothing but the gradient vector and del square j by del x square hessian
matrix some sort to be like that by definition actually, and this things we will need. Now
again going by the similar concept this optimization ideas, and all now we are asking a question
that we want a value x star where from x star, you do go whatever direction you want no matter
whatever direction left right top bottom whatever is actually which ever direction we go you
will get some value, which is strictly greater than value of x star or you will get a value
which is all of which strictly less than value of j x star. Basically then only you can talk
about whether is a minimum or maximum quantities sort of thing actually, and picture really
again if really wants picture then they two d sense, you can think about this kind of
a picture this is not this kind of picture actually.
So, if you have minimum points some where no matter which direction, you will look it
you can think of it something like a increasing function in all direction sort of thing. So,
no matter which direction you go in you will land of with higher values compare to the
minimum point actually. So, that kind of questions we are talking about actually.
So, again going back to this ideas Taylor series, we know there is this vector variable
also we can use this Taylor series expansion. So, we talk about j of x is nothing but j
of x star which is x star happens to be candidate for analyze for maximum or minimum, and then
j of x star plus delta x can be expanded that way remember this is the first term j of x
star plus the linear term evaluate del j del x evaluated x star into delta x plus this
second order term which is 1 by 2 factor delta x transpose into this session matrix times
delta x and then the order term. Actually and this case will not very adventurous go
to this, third order fourth derivatives and all that still possible to write.
But mathematical they are quite complex actually to try to in longer algebra sort of thing
actually. So, let us see the idea first in this context most of the time this is efficient
for us also basically. So far minimization problem what we are looking at is this quantity
this a j of x star plus delta x minus j of x star that quantity has to be strictly positive;
that means, no matter which direction you go the delta x can be any direction component
by component, they can take any positive negative any quantity what is our actually? Then it
turns out that if my value turns out to be have in that then it is a minimum point actually.
So, obviously if you want to satisfy, this and then going by this idea for direction
sensitive information, and all that we can always tell that my hope is analyze; that
means, I really want control having a conclusion something like that very free precisely then
it turns out my this term what is fasted out terms has to be 0; otherwise if I have a something
non zero, here the coefficients then depending on this quantity delta x which can be positive
negative whatever component. Since I will land of with a idea with the this quantity
either positive or negative. So, it becomes direct sensitive actually and which we do
not need. So, it turns out that a if you really want
to have a minimum point then this fasted are Jacobian matrix Jacobia
…
estimation. So far, we have seen some overview of this course, followed by some I mean some
overview of topics related to system theory as well as matrix algebra and in numerical
methods as well; that is relevant in this course actually. So, in the another thing
that is very much relevant in this course is the concept of static optimization; that
means, we are not talking about time varying variables and all, but these are constraint
numbers associated with that. But in general they have some objective, and
all that, how do you optimize that actually. So, that kind of things are call static optimization
or sometimes it is popularly known as a parameter optimization as well, because parameters are
something that typically remains constraint actually. So, anyways those are the topics
of discussion of this particular lecture, and also remember as I told you in beginning
lecture the very fast lecture that one way of solving this, this optimal control problem
is to go heavily to this static optimization ideas and convert the problem entirely to
a large dimensional static optimization problem, and solve it; that is called transcription
method, and all we will see that as we the course develops.
Another motivation of that is even if you formulate this optimal control problem from
calculus variation point of view and all that; the essentially land up with some equation
something call Hamiltonian, and all that; and that happens to be like every point it
happens to be kind of a constant quantity in its variables state and control, and that
you need to be minimize at every point of time in the frame work of static optimization
actually is very relevant concept for optimal control in general. So, it obviously, make
sense to have overview of static optimization before we go to the dynamic optimization;
that means, optimal control in this is now actually.
So, this I will take it in two parts and the first lecture will talk about some overview
or some standard results of static optimization along with some examples and all that way,
then further topics on it will see in the next lecture, and before proceeding to the
calculus of variation ideas and all in that way.
So, the topics of this lecture something like this, we will fast discuss about unconstrained
optimization, then we will go to this constrained optimization with equality constraints primarily
any quality constraints, and all will see in the next lecture actually. Then we will
some numerical examples as well. So, first is unconstraint optimization and
the problem definition turns out to be very clear, that you want to optimization some
vary some function of variable, but the variables are combination of some non time varying quantities.
Actually that is they are not constraint by any constraint equation neither actually before
proceeding with those with then algebra, and then analysis and think that; let us see very
simple straight forward idea first actually.
Now, forget it for a second this j 2 dot line, and all this flat was this flat was x verses
some functions j 1 it is a, and j 1 turns out be this function this solid function actually.
And then very quickly we can see that this 0, 1, 2, 3, 4, 5; they mean something to us
in a in just by looking at the flat actually, and very easily it can be argued that this
0.1 happens to be a local maximum, and point 2 is well we cannot say anything about that
it is a flat thing, but certainly a candidate for getting worried about, then is called
point of inflexion, and all that. And then point 3 turns out to be a local minimum point
4 turns out to be local maximum just by looking at the flat, we can very easily see this is
what is happening actually. And then it also we can see that point three
turns out to be global minimum, and point four turns out to be global maximum because
nothing else is bigger that in this domain of course, I mean x a 2 x a. And all in that
way, but when we talk about global maximum and all typically this x a and x b are also
not there; in other words we are talk about minus infinity plus infinity that domain sort
of thing actually. But anyway within this constraint, and then
you can think that 4 happens to be the global maximum and 3 happens to be the global minimum
actually, and one property that we observe simply observe is that all these points 1
2 3 4, this relationship is true; that means, the gradient of j with respect to x is 0,
if I just take a tangent here that is flat this is any way flat if I take it engine flat
here is a flat. And then it is if I take it engine at the fourth point then they also
parallel to the x axis actually. So, all the points satisfy that.
Now, that it gives is the clue that when we talk about something about optimization then;
obviously, the first derivative turns should happen to be 0 that is how to do actually,
and does not really happen. So, let us the talk about that that way and also remember
that instead of j 1, suppose you have j 2 then really do not have to worry about all
these we just have to worry about one minimum; that is what all that will have is one minimum,
and that one minimum happens to be the global minimum as well actually.
So, most of the time will take advantage of this observation, and value constructs a function
we will want to construct it that way. So, that we really do not want to bother about
global things, and all is not relevant also most of the time it satisfy our need also
basically is quadratic function and all that we will talk later this is how it turns out
to be now coming back to that whether this relationship satisfied for all the points,
and all and analyze it little bit mathematical vigorous sense basically.
So, now again we go back to this a this Taylor series, and all which will repeatedly require
anyway now remember what is a local maximum local maximum tells you that if point one
is local maximum, then very next two it whether go left side or right side either way that
will give a function value which is lesser than the function value at point 1.
Similarly, if it is point 3 if it is minimum local minimum, then whether you go left side
or right side either direction you are going to get higher values, then what you get it
point three actually. So, that kind of constraint is there.
So, let us analytic a point x star and then talk about a value of that function around
x star, and remember delta x can be either positive or negative actually here, and also
remember we have simply talking about a scalar quantity; that means, x is just single I mean
single component j cont j is a function of x, but x does not contain x 1 x 2 components
and all that x is a scalar variable sort of thing.
Then what happens x star plus delta x minus j of x star that is the term that will come
from Taylor series like this, actually now for any hope to x. What we are looking for
a looking for this quantity this j of x star plus delta x should be greater than j of x
star if j of x star happens to be minimum point irrespective of the direction of delta
x at the sign of delta x really actually. So, that has to be happen; that means, this
left side quantity has to be strictly positive for a for this quantity to be on a local minimum
or precisely be strictly negative for the other case actually local maximum case.
So, what we are looking at here is a sign independent property really and also remember
this Taylor series has a this uniform convergence property, and think like that. So that means,
if you right this expression to be something like that if you take a this entire quantity
what we what turns out be like this, the infinite terms what we are talking here is a less dominate
compare to the first term actually. So, that is how it will turn out be… So,
in other words the sign of it is largely dictated by the first term itself actually not by these
terms actually that way. So, if you have any real hope that the this become sign in sensitive,
then it turns out that the first term must be going to 0, then only we have some hope
actually that is how the necessary condition turns out to be and remember delta x is not
0, but this quantity has to be 0. That means, the coefficient has to be 0 basically that
is how it turns out that del j by del x; that x equal to x star is 0 is 0, then x star happens
to be a candidate point I mean you can you still cannot it whether it is minimum maximum
actually. So far that we will remain we will go to the
next one, and then tell if that that is already there; that means, at this point this is 0
anyway, and then this quantity turns out be that one.
And that one means, I can consider the second order term, and then there is a still higher
order terms a from third order fourth order can neglect actually. If this still has be
positive; that means, this has to be greater than this irrespective of the sign of delta
x remember this square term now. So, this has to be strictly greater than 0 provided
this is greater than 0 this is greater than 0 anyway.
So, this quantity is greater than 0 provided this quantity happens to be greater than 0;
that means, second order term has to be greater than 0 basically, and similarly the second
order term happens to be less than 0 it least to the conclusion that this term is less than
0, and other words this term happens to be maximum actually x star happens to be maximum.
So, this is how the how the analysis two to summarize the things what we have doing here
is given any function j of x, we will we will put this del j by del x is equal to 0; and
then solve this and find out candidate points x start did not be 1 point it can be several
point actually and but each of the points
We will try to evaluate this second order derivative and the second order derivative
happens to be 0 then it is a minimum point the second order derivative happens to be
less than 0 it is local minimum that is the sufficient condition actually. Now, we are
carrying out further things, and all let us see this some examples sort of thing, and
before even before the example. Now the question is what if this both derivatives of the 0,
but third of derivative is not 0. So that means, this is true whatever we are
talked about only if the second order derivative is non zero, and the non second order derivative
also happens to be 0, then we have no choice what we go to the third derivative; and the
third derivative happens to be non zero. Then we really cannot tell whether it is strictly
positive or strictly negative, and hence it turns out to be point of influence actually,
and if it is I mean if you still have a hope of getting it something like a positive or
negative; it we must have the third derivative also 0. And then go to the four derivative
and think like that actually right, if it is second derivative is 0 then this lands
up with the third derivative case, and then there is necessary condition of optimality
is if we still any hope then third derivative has to 0, then you land of with four derivative.
And if the fourth derivative happens to be non zero; that means, positive, and then it
is least to be minimum and maximum and otherwise it is maximum actually.
So, the analysis continues it actually toggles between this hard power, and even power sort
of thing actually. So, as a necessary condition all hard power has to be 0 and sufficiency
condition will come even powers actually.
So, this is how the things proceed, and suppose we want to get example here then turns out
to be let us take a very simple example j equal to x fourth, and you can see that well
the x fourth plot is very, very close to these this kind of very similar to x sort of thing.
So; obviously, we know that this is point of minimum; that means, x is equal to 0 is
the answer we know that actually whether we get it or not let see that using this result.
So, if you have j something like this, then del j by del x is four x is equal to 0 then
solution of that is three quantity is, but all happens to be 0 that x star equal to 0
0 0. And then let us go to the second derivative, and second derivative turns out to be like
the taking this term it is twelve x square and x star equal to 0 means twelve x square
is also 0. So, we really have to go third derivative
now, and the third derivative personality happens to be 24 x from here twelve into twenty
four into x star next star is 0; anyway the third derivative is also 0. So, that means
that is still our hope is alive actually, then you go to fourth derivative, and four
derivative it happens to be twenty four which is greater than 0. So, if the fourth derivative
is greater than 0, then obviously, it tells you that my some of my all of my hard powers
also 0, and the first even power that I uncounted is actually greater than 0.
So; obviously, x star is a minimum point actually; that is how we include that the result is
compatible to what we know actually come from graphical.
Since, now example to what if x is j is just x q and remember x q, what is the type of
I mean thing plot x q is something like this, if it is positive is positive, if it is negative
is negative; something like this. So, here you have a point of inflexion really; that
means, at 0 what we are talking here is one side, the second derivative will turn out
to be I mean one side the function turns out to be increasing other side the function turns
out to be decrease actually. So, this will this we know that it is a point of inflexion,
now whether our analysis tells of the tells do not let see.
So, if you have j equal x q then del j by del x is nothing but three x square is equal
to 0, x star turns out to be 0 0, now we have but two candidate solution, but both of happens
to be anyway. Now, the second derivative happens to be six x square is also 0. So, we cannot
conclude anything this point of time this is three x square. So, it will be six x. So,
six x evaluated x star that is six x star x star is 0. So, it is still 0 actually.
Now, we go to third derivative and third derivative del q by del q j by del x q from coming from
here is nothing but six, but the important point note is not whether six is positive
or negative, but six is not equal to 0 remember this third hard derivative actually hard power
derivative, but this hard power derivative has to be non equal to I mean not equal to
0 that is more important, actually if it is not equal to 0 your hope is not alive it conclusion
as a already in there and the conclusion is the x star is point of inflexion.
Because if I take this series this series this terms will start dominating and delta
x is positive defending on whatever sign is there it will lead one sign, and defending
on whatever sign is here if delta x is negative to the exactly opposite say conclusion here.
So, is just one side is a either increasing or decreasing, but other side is just the
positive actually. So, will turn out be point of inflexion. So, that is how it is there.
So, if j equal to x q then x star equal to 0 turns out to be point of inflexion point
actually a point of inflexion really. So, this all comfortable to what we know basically
now, but this is not kind of story this is just very beginning it is talks about as scalar
variable functions sort of thing actually, but we are not interested in scalar variable
function.
In general because we delay system theory, where state can be n dimensional control can
be m dimensional like that. So, we takes it is a mixer of this components
of the state variable really. So, that is what we want analyze then leads to this vector
case, and the problem is turns out to be like that minimize j of x where x is n-dimensional
vector, and remember the objective function is always is a scalar quantity always otherwise
is talk about multi objective of optimization and all that and we are not talking about
that at all in this course actually. So, when somebody talks about standard of optimization
problem then it turns out to be that the optimization function is a scalar quantity, where is the
variables can be n-dimensional there can be several variables playing is playing out different
contributes, and all we want minimize this scalar quantity. Finally, and also remember
this is a this is a non constraint problem; that means, this component of x are not constraint
by any constraint actually you can they can take any values what so ever actually that
way. So, you want analyze that problem and see
what is going on actually and also remember phi definition what is saw in this matrix
theory is reviews, and all that also that by definition this is del j by del x equal
to like this by this is nothing but the gradient vector and del square j by del x square hessian
matrix some sort to be like that by definition actually, and this things we will need. Now
again going by the similar concept this optimization ideas, and all now we are asking a question
that we want a value x star where from x star, you do go whatever direction you want no matter
whatever direction left right top bottom whatever is actually which ever direction we go you
will get some value, which is strictly greater than value of x star or you will get a value
which is all of which strictly less than value of j x star. Basically then only you can talk
about whether is a minimum or maximum quantities sort of thing actually, and picture really
again if really wants picture then they two d sense, you can think about this kind of
a picture this is not this kind of picture actually.
So, if you have minimum points some where no matter which direction, you will look it
you can think of it something like a increasing function in all direction sort of thing. So,
no matter which direction you go in you will land of with higher values compare to the
minimum point actually. So, that kind of questions we are talking about actually.
So, again going back to this ideas Taylor series, we know there is this vector variable
also we can use this Taylor series expansion. So, we talk about j of x is nothing but j
of x star which is x star happens to be candidate for analyze for maximum or minimum, and then
j of x star plus delta x can be expanded that way remember this is the first term j of x
star plus the linear term evaluate del j del x evaluated x star into delta x plus this
second order term which is 1 by 2 factor delta x transpose into this session matrix times
delta x and then the order term. Actually and this case will not very adventurous go
to this, third order fourth derivatives and all that still possible to write.
But mathematical they are quite complex actually to try to in longer algebra sort of thing
actually. So, let us see the idea first in this context most of the time this is efficient
for us also basically. So far minimization problem what we are looking at is this quantity
this a j of x star plus delta x minus j of x star that quantity has to be strictly positive;
that means, no matter which direction you go the delta x can be any direction component
by component, they can take any positive negative any quantity what is our actually? Then it
turns out that if my value turns out to be have in that then it is a minimum point actually.
So, obviously if you want to satisfy, this and then going by this idea for direction
sensitive information, and all that we can always tell that my hope is analyze; that
means, I really want control having a conclusion something like that very free precisely then
it turns out my this term what is fasted out terms has to be 0; otherwise if I have a something
non zero, here the coefficients then depending on this quantity delta x which can be positive
negative whatever component. Since I will land of with a idea with the this quantity
either positive or negative. So, it becomes direct sensitive actually and which we do
not need. So, it turns out that a if you really want
to have a minimum point then this fasted are Jacobian matrix Jacobia
…