Advertisement
English transcript
Hello everybody, we will continue our lecture series.
At last lecture, we saw this the concept of calculus of variation extended that to constant
optimization problems, as well as well as we discussed about an example problem, which
motivated that this concept of variation of calculus can be used for optimal control very
well directly rather. So, now, we want to generalize this concept and then talk about
how you formalize this optimal control concepts coming from calculus of variations, ok. So,
we will try to use these equations again, but now it will be in the fully prime work
of optimal control theory again, I mean we do not want to talk in a disguised way and
thing like that anymore, we will define optimal control problem and try to attempt to kind
of solve this problem as it is actually.
Alright, so what is the summary of various solve problem and multiple dimension with
constrains?
So, that we that is what we are we are were looking for, that is what we discussed in
the just in the last class, that we are we were interested in the optimizing this kind
of a cross function along with this this general equality constraints sort of thing, where
x up in to end dimension of the phi up to the till the dimension.
And then, we talked about this Lagrange’s Existence theorem and things like that and
ultimately, let us to this definition of this l star, again this l star happen to be a functional
number also, so this this l star can be defined like this.
And then, we have this EL equation and thus EL equations happen to be something like this
actually, ok. So, we have n equations coming from there and until the equations coming
from there, transversality conditions were like this, we discussed everything in the
in the previous lecture, some of you is who do not recall probably can revise.
That actually and the ultimately lend up with these these kind of e l equations with with
transversality conditions, ok. So, this transversaility condition depending on whether t f is fixed
or t f is free we have these two classes of different things and we discussed about that
in the last class. Ultimately the point is, key point is we have e l equations, lagrange’s
equation and there are transversality condition or boundary conditions like that actually.
Alright, so now how do we use these concepts for these applications of these deals of optimal
control problem rather ah.
So, that is the formal definition of optimal control that also we discussed in the last
class, a little bit towards end of the last class, but all I mean to revise it again.
What you are looking for is to find and admissible control u t would causes the system this x
star equal to f of X U t, where initial condition is fixed to follow an admissible trajectory
that optimizes the performance index on the on the way while satisfying the appropriate
boundary conditions. So, we have the p h of a good optimal control problem definition,
you to you need to have a state equations system dynamics, you need to have a initial
conditions or other final conditions, initial conditions together that means appropriate
boundary conditions. And as a designer you should always try to construct a cross function
that you want to minimize or maximize either way, which will lead to the object basically,
ok. So, there were the three component, that is
always essential for a for a good optimal control problem formulation actually, alright.
And also remember, now if you look at this thing, this this constant equation contain
an x dot term, x x dot t equal to 0, where as this constraint equation what you are looking,
what you are talking here, I mean this cross function, let me go back to cross function
also there, cross function contain a x dot term also here. Now, this cross function,
let me go back to cos function also there, cos function contain an extra term also here,
now this cos function what we are looking for here typically does not contain an x dot
most of the cases, ok.
But x dot is also required, because x dot consists part of this state equation and by
the way it does not mean that you do not really need x dot all the time, I mean you can have
a prolonged x dot is also there, then in especially this l q r clause, linear quadratic regulator
clause when we discuss that point at time, will see that the clause of problem for which
you really need a x dot cross functional, in the cross functional also. But, in general,
most of the time you do not need it, we it is sufficient to define in the form of x and
u, because once x and u is this then x dot is constant anyway like this, you do not have
a freedom of, you do not have a further freedom basically what I mean.
Once you get x and u and x initial condition is defined and then this x dot represent to
be equal to that, so you do not have too much freedom for minimizing of maximizing that
that quantity basically. So, that is the reason why we are not putting it here most of the
time, alright, so this is the problem definition there how do you going do that, ok.
Now, the problem here is this this particular term, I mean if you say this this equation
what we discussed in the calculus of of variation setting it did not contain any term outside
the integral. What suddenly we are talking at term which is outside the integral actually,
this term, this is outside the integral, how do you handle that, alright.
One way to handle that is something like this, let us first consider this integral, we will
consider the integral t naught to t f d by of this quantity and this happens to be like
that, I mean this function evaluated at final value minus evaluated in initial value. That
means, this function what we are looking for, this phi of t f x, f small par f t f x f,
that what we are looking for, this can be interpreted as something this one plus integral
of this quantity, ok But once you evaluate any function with a
number, this particularly this as a number basically, so that means, that is irrelevant
for optimization really, you are talking about this one plus something actually, which is
which is variable. Anything that is has flexibility we can do something for minimizing and maximizing,
anything that is number we can do anything too much on that, so that means, what I mean
is this phi f x naught t naught is a constant value. So, instead of trying to optimize this
j, it is equivalent to optimize j 1 ignoring this fellow.
So, we will ignore that and then tell now this is a integral of that and that is the
original of thing is also integral of that, there is a term inside the integral, there
is a term outside, this is gone anyway, again this is integral term, where combine these
two, this what you are looking for here, now is compatible to what you what you seen before
actually, right; this does not have anything outside. So, this is what you are telling
this is the t that we are following actually in a way.
So, instead of trying to optimize the cross function j directly, which will which is also
possible, I mean, I will talk to you in few slides to down the line actually, that is
on the way of doing things, but we will consider in one to one equivalent of what we already
know and how do we want to put it in that frame work actually, then we we do this algebra
and then look at this, is just a constant value, does not matter. So, we will ignore
that and then construct this j 1 and then take both the terms inside the integral and
then it happens to be very comfortable to what you already know basically. So, now we
can use this calculus of various ideas that we already know.
What is that? Then, obviously we will construct a j bar first and j bar is nothing but whatever
we already have inside the integral plus lambda transpose, this constraint equation, f minus
x dot is equal to 0, we have this constraint equation. So, f minus x dot will be equal
to 0 and that constraint equation should form here, inside the integral itself, ok.
Now, the further algebra talks something like this, whatever, where ever you do not see
any derivative term, you define that something like Hamiltonian, that this definition comes
from continuous instantaneous and all that, that one, that history and all that is not
here, that is not here, that essentially come from his ideas basically. Any way before that
people where all talking about e l equation and going directly there, but this this particular
concept of simplifying things and all that was. We will see some of this other contribution
is constraint optimal control lectures later actually.
Anyway, come right this Hamiltonian is defined as something like this, all the terms without
derivative term. Then I can apply simply j bar, something like this l plus lambda transpose
anything, but h, now I will drop the arrangements for simplifying simplicity of algebra, I know
that which one can be easy function of what, with that assumption I will drop this arrangement
and then tell this is first h and then the next term is d phi by d t and the next term
is this this quantity lambda transverse x dot, ok.
So, this is my quantity, so this is nothing but l star actually, so J bar is nothing but
integration t 0 to t f l star td where l star is nothing but, that now it is very very compatible
to what we know before actually. But what is l star and what are Hamiltonians and all,
level l Hamiltonians is a function of h come l lambda and l star is also function of h
c 1 lambda.
So, we have to apply the EL equations thrice once with respect to x, once with respect
to u, once with respect to lambda actually. Alright, so we will do that and this l star
just to summarize l star is defined like this, where h is defined like that, h is l plus
lambda transverse, where l star is nothing but h plus d phi by d t, means lambda transverse
t x dot. Now, what is the necessary condition? We have to apply this condition thrice.
And again this this expression nowhere contain u dot and lambda dot, again this to quantifier
goes to 0.
But, we are not done yet, this algebra we have to get actually, because what is del
l star by del x, del l star by del x, so that by definition is del by del x l star, an l
star is nothing but this quantity, right. So, we substitute this quantity here, whatever
is l star is fine, here and then we carry out this algebra. So, first thing talks about
del h by del x plus del by del x of this quantity on the where you have here and d phi is nothing
but del phi by del t plus phi is a function of x del of t, so del phi by del t plus del
phi by del x transverse time of x dot actually. So, these two quantities come from this quantity,
really of this quantity does not contain any x or I mean lambda, so the derivative of that,
with that, specially with respect to that x is zero as equal to to left all to be right
there. This is del h by del x plus this quantity del square phi by del x del t plus this quantity
del by del square phi by del x square into x dot, this is double derivative of the thing
actually, alright. Now, what about the other quantity, this is
one quantity, what about the other quantity del l star by del x, this is the first quantity
that we discussed, where is the second quantity? d by d minus one of del l star by del x. So,
that is why d by dt of del l star by del x dot. Now, we have to go to this l star sort
of thing and d by d t of that, d by d t of that is nothing but d by d t of del phi by
del x minus lambda basically, ok. Alright, because very it is easy to see that
from l star definition something like this, so you can think of this del phi by del t
again I have explained that here and we talk about d phi by d t, this d phi by d t is nothing
but del phi by del t plus this term that is why you get an additional x dot t actually,
ok. So, now, we need to talk about del l star by del x dot, now we are actually looking
for this expression, h is not a function of x dot, so any personal derivative with respect
to that is not there, but it will come from these two expressions, what is these two,
one is del phi by del x, another one is lambda, that will come here and d by d t that d t,
so del d by d t of that is nothing but again del by del t of that plus del by del x of
that time x dot. Remember, whenever you talk, there are two
things remain here, one is the total derivative, something like d phi by d t, other one is
the personal derivative del phi by del t. So, whenever there is d phi by d t then is
the first with the personal derivative with respect to this variable and then the account
for change in the derivative as well basically, so that is how the algebra will proceed actually,
sorry. So, this is this is the one what we are looking for here, d by d t of that turns
out to be like this actually, because of that del phi by del by del t of that one plus del
by del x of that into x dot minus this d by d t of lambda is nothing but lambda dot.
Because of this, look at this expression, whenever you have this del l by del x do not
worry about it actually, any way this is the type of expression that I was talking about.
So, do by algebra yourself, then you then the quantity will emerge actually and ultimately
you will end up you in something like this. So, what is the EL equation, EL equation tells
that this term minus this term x is equal to 0, then we talk this term minus this term,
obviously these two will be cancelled of del square phi by del x del t, is in these both
the places will get cancelled out.
And then you will be left out with other term, so it is here anyway, so one term will cancelled
out, the other term will will tell us that del h by del x from here, the other term will
also cancelled out, this two terms, this is how this actually, ok.
So, then we are left out with nothing by del h by del x plus lambda dot is equal to 0,
that means, lambda dot is equal to nothing but minus del h by del x and this particular
equation is called costate equation or adjoint equation really. Then coming to the second
part of the equation del l star by del u equal to 0, that is easy, that t will nowhere else
u appears, u appears only in the definition of Hamiltonian. So, what you are talking is
del h by del u is equal to 0 and then by del l star by del lambda will be nothing but the
boundary I mean boundary, sorry this constraint equation sort of thing.
So, it will give us this thing like that del h by del lambda minus x dot equal to zero;
that means, x dot is nothing but del h by del lambda and when you talk about h will
like that, so what is del h by del lambda, it is nothing but f. So, x dot equal to f
of x u t that is nothing but here state equation, so ultimately what you are looking for is
what you one landing up a, where you landed up is something like costae equation and optimal
control equation now is also called as stationary equation. So, it is just as algebraic equation
sort of thing and then this is a state equation actually, so this equation has to be solved
together basically that way.
So, this what a summary. Summary is you have to define lamination which nothing but l plus
lambda transpose f and then carry out this algebra, well one thing is x dot equal to
f of x u t state equation and then del h by del t is equal to zero, optimal control equation,
once you solve this equation at every point of time u will be a function of x and lambda
alternatively. Hamiltanion is a function of x and lambda, everything, so if you solve
this equation for U then we got that optimal control, for solving that you remember U will
be function of x and lambda, both and for that you need these two equation together
basically, ok. So, one part of the equation is state equation,
other part is costate equation, in regarding that I will talk one or two slides latterly,
now what about boundary conditions, ok.
Boundary conditions are again, go back to that that transverse facility condition and
all, the general transversality happens to be like that, but L star you remember something
like this, this is that when the d phi by d t is nothing but del phi by del t plus del
phi by del x into x dot sort of thing. So, you substitute this L star, whatever you see
L star is something like this and then we get variety of conditions this way and that
way and all, you can derive anything actually. For example, when both t naught and x of t
not are assumed to be fixed, then this delta t not and delta x not, though those were 0
by default, we will, I mean we all that we have is delta x f by delta t f, that is the
variation actually. Alright, now what you do is you substitute
these expression del l star by del x dot something like that, you are going to carry out these
algebra will be more, cancelled out some of these other terms and land up some of these
equations like this actually. So, what I mean is when ever this t naught and x of t naught
are assumed to be fixed, which is many of the terms which is true, then this is a boundary,
I mean transversality conditions, ultimately this will be place of something like that
in the form of hamiltonian actually. And in addition to that many times people
tell a case at sometimes delta x f is 0, that means, x f is fixed, then this only these
one you left out, delta t f is free, that means hamiltonian at the final time is nothing
but minus of del phi by del t. Similarly when t f is fixed, what x f is free, then this
is 0, but this is free, then this cannot be here, lambda f equal to del phi by del x actually.
These is what is tried here, when t of f is fixed x of f is free, you land up with this
equation, this what I just now described, so that means lambda f will turn of to be
del phi by del x f. Actually, when on the other end when t f is free, but x f is fixed,
then h of t f will be minus del phi by del t value to the t f and when both are free
you have both these necessary anyway. If you x f is free and t f is also free then both
these condition will come into the picture and also remember how many condition are there,
those kind of phi is a scalar, x is a vector, del phi by del x happen to be n, so that means
you are getting n boundary condition from here.
So, how many differential equation you have here, it is also again n, H is is a scalar,
x is a vector, so lambda dot is an n dimension is x actually, so that means lambda is also
n dimensional in a way you can give that picture. So, sincerely what is going on here, we have
problem, optimal control problem and we have a once of differential equation and once of
algebric equation, once of boundary conditions and things like that and all of them to be
satisfied for meaning full solutions basically, ok.
Now, what about an alternate approach? We do not have to carry out all these algebra
and things like that, but in a simplistic sense let as see whether we can have the same
solution or not actually.
So that the performance index to be minimized, happens to be like this, the same performance
index that we started with, path constraints nothing but the state equation and again the
boundary condition are there, x of t j, X of 0 is the initial condition which is specified
and t f is fixed and x of t f is free actually.
Anyway so this is what it is, then the necessary condition of of optimal value t what you do,
will not worry about the calculus of various of e l equation and all we will go back little
further and little more before that, then we will concept augmented performance index,
j bar is nothing but that. So, L plus lambda transpose f minus X dot, so f minus anywhere,
f minus x dot f minus f minus x along the path. So, will take it inside the integral,
which is path dependent co structurally, where is something laminoarized here, put is as
j bar or something like this, then directly define lamainarization like that, Hamiltonian
is same Hamiltonian what we discussed L plus lambda transpose f.
And now we are interested net replying el equation, but we will consider first variation
of j bar directly, when we do first variation of j bar is nothing but first variati
…
At last lecture, we saw this the concept of calculus of variation extended that to constant
optimization problems, as well as well as we discussed about an example problem, which
motivated that this concept of variation of calculus can be used for optimal control very
well directly rather. So, now, we want to generalize this concept and then talk about
how you formalize this optimal control concepts coming from calculus of variations, ok. So,
we will try to use these equations again, but now it will be in the fully prime work
of optimal control theory again, I mean we do not want to talk in a disguised way and
thing like that anymore, we will define optimal control problem and try to attempt to kind
of solve this problem as it is actually.
Alright, so what is the summary of various solve problem and multiple dimension with
constrains?
So, that we that is what we are we are were looking for, that is what we discussed in
the just in the last class, that we are we were interested in the optimizing this kind
of a cross function along with this this general equality constraints sort of thing, where
x up in to end dimension of the phi up to the till the dimension.
And then, we talked about this Lagrange’s Existence theorem and things like that and
ultimately, let us to this definition of this l star, again this l star happen to be a functional
number also, so this this l star can be defined like this.
And then, we have this EL equation and thus EL equations happen to be something like this
actually, ok. So, we have n equations coming from there and until the equations coming
from there, transversality conditions were like this, we discussed everything in the
in the previous lecture, some of you is who do not recall probably can revise.
That actually and the ultimately lend up with these these kind of e l equations with with
transversality conditions, ok. So, this transversaility condition depending on whether t f is fixed
or t f is free we have these two classes of different things and we discussed about that
in the last class. Ultimately the point is, key point is we have e l equations, lagrange’s
equation and there are transversality condition or boundary conditions like that actually.
Alright, so now how do we use these concepts for these applications of these deals of optimal
control problem rather ah.
So, that is the formal definition of optimal control that also we discussed in the last
class, a little bit towards end of the last class, but all I mean to revise it again.
What you are looking for is to find and admissible control u t would causes the system this x
star equal to f of X U t, where initial condition is fixed to follow an admissible trajectory
that optimizes the performance index on the on the way while satisfying the appropriate
boundary conditions. So, we have the p h of a good optimal control problem definition,
you to you need to have a state equations system dynamics, you need to have a initial
conditions or other final conditions, initial conditions together that means appropriate
boundary conditions. And as a designer you should always try to construct a cross function
that you want to minimize or maximize either way, which will lead to the object basically,
ok. So, there were the three component, that is
always essential for a for a good optimal control problem formulation actually, alright.
And also remember, now if you look at this thing, this this constant equation contain
an x dot term, x x dot t equal to 0, where as this constraint equation what you are looking,
what you are talking here, I mean this cross function, let me go back to cross function
also there, cross function contain a x dot term also here. Now, this cross function,
let me go back to cos function also there, cos function contain an extra term also here,
now this cos function what we are looking for here typically does not contain an x dot
most of the cases, ok.
But x dot is also required, because x dot consists part of this state equation and by
the way it does not mean that you do not really need x dot all the time, I mean you can have
a prolonged x dot is also there, then in especially this l q r clause, linear quadratic regulator
clause when we discuss that point at time, will see that the clause of problem for which
you really need a x dot cross functional, in the cross functional also. But, in general,
most of the time you do not need it, we it is sufficient to define in the form of x and
u, because once x and u is this then x dot is constant anyway like this, you do not have
a freedom of, you do not have a further freedom basically what I mean.
Once you get x and u and x initial condition is defined and then this x dot represent to
be equal to that, so you do not have too much freedom for minimizing of maximizing that
that quantity basically. So, that is the reason why we are not putting it here most of the
time, alright, so this is the problem definition there how do you going do that, ok.
Now, the problem here is this this particular term, I mean if you say this this equation
what we discussed in the calculus of of variation setting it did not contain any term outside
the integral. What suddenly we are talking at term which is outside the integral actually,
this term, this is outside the integral, how do you handle that, alright.
One way to handle that is something like this, let us first consider this integral, we will
consider the integral t naught to t f d by of this quantity and this happens to be like
that, I mean this function evaluated at final value minus evaluated in initial value. That
means, this function what we are looking for, this phi of t f x, f small par f t f x f,
that what we are looking for, this can be interpreted as something this one plus integral
of this quantity, ok But once you evaluate any function with a
number, this particularly this as a number basically, so that means, that is irrelevant
for optimization really, you are talking about this one plus something actually, which is
which is variable. Anything that is has flexibility we can do something for minimizing and maximizing,
anything that is number we can do anything too much on that, so that means, what I mean
is this phi f x naught t naught is a constant value. So, instead of trying to optimize this
j, it is equivalent to optimize j 1 ignoring this fellow.
So, we will ignore that and then tell now this is a integral of that and that is the
original of thing is also integral of that, there is a term inside the integral, there
is a term outside, this is gone anyway, again this is integral term, where combine these
two, this what you are looking for here, now is compatible to what you what you seen before
actually, right; this does not have anything outside. So, this is what you are telling
this is the t that we are following actually in a way.
So, instead of trying to optimize the cross function j directly, which will which is also
possible, I mean, I will talk to you in few slides to down the line actually, that is
on the way of doing things, but we will consider in one to one equivalent of what we already
know and how do we want to put it in that frame work actually, then we we do this algebra
and then look at this, is just a constant value, does not matter. So, we will ignore
that and then construct this j 1 and then take both the terms inside the integral and
then it happens to be very comfortable to what you already know basically. So, now we
can use this calculus of various ideas that we already know.
What is that? Then, obviously we will construct a j bar first and j bar is nothing but whatever
we already have inside the integral plus lambda transpose, this constraint equation, f minus
x dot is equal to 0, we have this constraint equation. So, f minus x dot will be equal
to 0 and that constraint equation should form here, inside the integral itself, ok.
Now, the further algebra talks something like this, whatever, where ever you do not see
any derivative term, you define that something like Hamiltonian, that this definition comes
from continuous instantaneous and all that, that one, that history and all that is not
here, that is not here, that essentially come from his ideas basically. Any way before that
people where all talking about e l equation and going directly there, but this this particular
concept of simplifying things and all that was. We will see some of this other contribution
is constraint optimal control lectures later actually.
Anyway, come right this Hamiltonian is defined as something like this, all the terms without
derivative term. Then I can apply simply j bar, something like this l plus lambda transpose
anything, but h, now I will drop the arrangements for simplifying simplicity of algebra, I know
that which one can be easy function of what, with that assumption I will drop this arrangement
and then tell this is first h and then the next term is d phi by d t and the next term
is this this quantity lambda transverse x dot, ok.
So, this is my quantity, so this is nothing but l star actually, so J bar is nothing but
integration t 0 to t f l star td where l star is nothing but, that now it is very very compatible
to what we know before actually. But what is l star and what are Hamiltonians and all,
level l Hamiltonians is a function of h come l lambda and l star is also function of h
c 1 lambda.
So, we have to apply the EL equations thrice once with respect to x, once with respect
to u, once with respect to lambda actually. Alright, so we will do that and this l star
just to summarize l star is defined like this, where h is defined like that, h is l plus
lambda transverse, where l star is nothing but h plus d phi by d t, means lambda transverse
t x dot. Now, what is the necessary condition? We have to apply this condition thrice.
And again this this expression nowhere contain u dot and lambda dot, again this to quantifier
goes to 0.
But, we are not done yet, this algebra we have to get actually, because what is del
l star by del x, del l star by del x, so that by definition is del by del x l star, an l
star is nothing but this quantity, right. So, we substitute this quantity here, whatever
is l star is fine, here and then we carry out this algebra. So, first thing talks about
del h by del x plus del by del x of this quantity on the where you have here and d phi is nothing
but del phi by del t plus phi is a function of x del of t, so del phi by del t plus del
phi by del x transverse time of x dot actually. So, these two quantities come from this quantity,
really of this quantity does not contain any x or I mean lambda, so the derivative of that,
with that, specially with respect to that x is zero as equal to to left all to be right
there. This is del h by del x plus this quantity del square phi by del x del t plus this quantity
del by del square phi by del x square into x dot, this is double derivative of the thing
actually, alright. Now, what about the other quantity, this is
one quantity, what about the other quantity del l star by del x, this is the first quantity
that we discussed, where is the second quantity? d by d minus one of del l star by del x. So,
that is why d by dt of del l star by del x dot. Now, we have to go to this l star sort
of thing and d by d t of that, d by d t of that is nothing but d by d t of del phi by
del x minus lambda basically, ok. Alright, because very it is easy to see that
from l star definition something like this, so you can think of this del phi by del t
again I have explained that here and we talk about d phi by d t, this d phi by d t is nothing
but del phi by del t plus this term that is why you get an additional x dot t actually,
ok. So, now, we need to talk about del l star by del x dot, now we are actually looking
for this expression, h is not a function of x dot, so any personal derivative with respect
to that is not there, but it will come from these two expressions, what is these two,
one is del phi by del x, another one is lambda, that will come here and d by d t that d t,
so del d by d t of that is nothing but again del by del t of that plus del by del x of
that time x dot. Remember, whenever you talk, there are two
things remain here, one is the total derivative, something like d phi by d t, other one is
the personal derivative del phi by del t. So, whenever there is d phi by d t then is
the first with the personal derivative with respect to this variable and then the account
for change in the derivative as well basically, so that is how the algebra will proceed actually,
sorry. So, this is this is the one what we are looking for here, d by d t of that turns
out to be like this actually, because of that del phi by del by del t of that one plus del
by del x of that into x dot minus this d by d t of lambda is nothing but lambda dot.
Because of this, look at this expression, whenever you have this del l by del x do not
worry about it actually, any way this is the type of expression that I was talking about.
So, do by algebra yourself, then you then the quantity will emerge actually and ultimately
you will end up you in something like this. So, what is the EL equation, EL equation tells
that this term minus this term x is equal to 0, then we talk this term minus this term,
obviously these two will be cancelled of del square phi by del x del t, is in these both
the places will get cancelled out.
And then you will be left out with other term, so it is here anyway, so one term will cancelled
out, the other term will will tell us that del h by del x from here, the other term will
also cancelled out, this two terms, this is how this actually, ok.
So, then we are left out with nothing by del h by del x plus lambda dot is equal to 0,
that means, lambda dot is equal to nothing but minus del h by del x and this particular
equation is called costate equation or adjoint equation really. Then coming to the second
part of the equation del l star by del u equal to 0, that is easy, that t will nowhere else
u appears, u appears only in the definition of Hamiltonian. So, what you are talking is
del h by del u is equal to 0 and then by del l star by del lambda will be nothing but the
boundary I mean boundary, sorry this constraint equation sort of thing.
So, it will give us this thing like that del h by del lambda minus x dot equal to zero;
that means, x dot is nothing but del h by del lambda and when you talk about h will
like that, so what is del h by del lambda, it is nothing but f. So, x dot equal to f
of x u t that is nothing but here state equation, so ultimately what you are looking for is
what you one landing up a, where you landed up is something like costae equation and optimal
control equation now is also called as stationary equation. So, it is just as algebraic equation
sort of thing and then this is a state equation actually, so this equation has to be solved
together basically that way.
So, this what a summary. Summary is you have to define lamination which nothing but l plus
lambda transpose f and then carry out this algebra, well one thing is x dot equal to
f of x u t state equation and then del h by del t is equal to zero, optimal control equation,
once you solve this equation at every point of time u will be a function of x and lambda
alternatively. Hamiltanion is a function of x and lambda, everything, so if you solve
this equation for U then we got that optimal control, for solving that you remember U will
be function of x and lambda, both and for that you need these two equation together
basically, ok. So, one part of the equation is state equation,
other part is costate equation, in regarding that I will talk one or two slides latterly,
now what about boundary conditions, ok.
Boundary conditions are again, go back to that that transverse facility condition and
all, the general transversality happens to be like that, but L star you remember something
like this, this is that when the d phi by d t is nothing but del phi by del t plus del
phi by del x into x dot sort of thing. So, you substitute this L star, whatever you see
L star is something like this and then we get variety of conditions this way and that
way and all, you can derive anything actually. For example, when both t naught and x of t
not are assumed to be fixed, then this delta t not and delta x not, though those were 0
by default, we will, I mean we all that we have is delta x f by delta t f, that is the
variation actually. Alright, now what you do is you substitute
these expression del l star by del x dot something like that, you are going to carry out these
algebra will be more, cancelled out some of these other terms and land up some of these
equations like this actually. So, what I mean is when ever this t naught and x of t naught
are assumed to be fixed, which is many of the terms which is true, then this is a boundary,
I mean transversality conditions, ultimately this will be place of something like that
in the form of hamiltonian actually. And in addition to that many times people
tell a case at sometimes delta x f is 0, that means, x f is fixed, then this only these
one you left out, delta t f is free, that means hamiltonian at the final time is nothing
but minus of del phi by del t. Similarly when t f is fixed, what x f is free, then this
is 0, but this is free, then this cannot be here, lambda f equal to del phi by del x actually.
These is what is tried here, when t of f is fixed x of f is free, you land up with this
equation, this what I just now described, so that means lambda f will turn of to be
del phi by del x f. Actually, when on the other end when t f is free, but x f is fixed,
then h of t f will be minus del phi by del t value to the t f and when both are free
you have both these necessary anyway. If you x f is free and t f is also free then both
these condition will come into the picture and also remember how many condition are there,
those kind of phi is a scalar, x is a vector, del phi by del x happen to be n, so that means
you are getting n boundary condition from here.
So, how many differential equation you have here, it is also again n, H is is a scalar,
x is a vector, so lambda dot is an n dimension is x actually, so that means lambda is also
n dimensional in a way you can give that picture. So, sincerely what is going on here, we have
problem, optimal control problem and we have a once of differential equation and once of
algebric equation, once of boundary conditions and things like that and all of them to be
satisfied for meaning full solutions basically, ok.
Now, what about an alternate approach? We do not have to carry out all these algebra
and things like that, but in a simplistic sense let as see whether we can have the same
solution or not actually.
So that the performance index to be minimized, happens to be like this, the same performance
index that we started with, path constraints nothing but the state equation and again the
boundary condition are there, x of t j, X of 0 is the initial condition which is specified
and t f is fixed and x of t f is free actually.
Anyway so this is what it is, then the necessary condition of of optimal value t what you do,
will not worry about the calculus of various of e l equation and all we will go back little
further and little more before that, then we will concept augmented performance index,
j bar is nothing but that. So, L plus lambda transpose f minus X dot, so f minus anywhere,
f minus x dot f minus f minus x along the path. So, will take it inside the integral,
which is path dependent co structurally, where is something laminoarized here, put is as
j bar or something like this, then directly define lamainarization like that, Hamiltonian
is same Hamiltonian what we discussed L plus lambda transpose f.
And now we are interested net replying el equation, but we will consider first variation
of j bar directly, when we do first variation of j bar is nothing but first variati
…