Advertisement
English transcript
Hello every one, we will continue our lecture series in optimal control guidance in estimation
course. In the last couple of lectures,
we saw this optimal control formulation from calculus of variations and then we gave a
couple of examples which motivated that in general we need numerical methods to solve
all in optimal control problems. So, let us see a couple of such numerical
techniques and then we will follow it up with further things. Now,
let’s see the topics covered today. This lecture is first to very briefly review the
necessary conditions of optimality, then we will talk about 23 methods and then the first
shooting methods followed by gradient method and quasi linearization method.
There are various other techniques as well but it’s not necessary to continue talking
about all the methods. You can, after knowing these techniques, probably see some literature
or book or something to follow the other techniques
as well. As just started, this is the generic optimal control problem that we are interested
in that kind of optimizing or minimize or maximize certain performance index of this
form and along with the performance index we have this associated path constraint as
well as boundary conditions actually.
So, the necessary conditions of optimality that we derived from previous lectures come
out to be like that we have a state equation which is x dot be del H by del lambda which
is the f since this sort of thing is coming from the system dynamics. Typically, when
when followed by costate equation as well which is lambda dot is minus del H by del
X and both of the these referred to be differential equations dynamic equations really and then
the the third equation is optimal control equation which is a static equation. From
this equation we should be able to solve optimal control U H is a function of X and lambda
actually . So, that is how it develops now, for this 2 differential equations, we have
this two boundary conditions as well. So, the problem is well defined really and
we can see that this state equation takes the initial condition whereas costate equation
takes the final condition and that is how it lines up with those 2 point boundary value
problems. And then, this 2 point boundary value problem should account for all this
boundary conditions properly. Also review if a solution does not satisfy any one of
the equations including the boundary conditions that means that the solution is non optimal.
So, we must get to further conditions; almost get to all the conditions with equivalent
process itself. So, that’s the point. So, in this lecture,
before getting started with some of these techniques, we also discussed in last class
that this state costate equations are dynamic equations and it turns out that if 1 is stable
the other turns to be unstable in general. The optional control is a stationary equation
from which you need to solve for control and for which we need to know the lambda. Unless
you know lambda, we cannot solve it from here and boundary conditions are going to be split.
That’s what I told you here, that state equation is initial condition whereas co-state
equation is final condition and typically because of this, the the state the state equation
is double of forward whereas costate equation should be double of backward because you know
the the final condition. Basically, that’s that’s why in a typically costate equation
we can integrated backwards because we know the boundary condition at the final thing
in general. I mean the numerical techniques can circumference over this and then talk
about forward integration above the equations and all that we will just see that in a in
a while. Ok, so, with all this difficulty that the boundary conditions are split and
if one is stable and other one is unstable, that it leads to this great difficulty of,
I mean, this complexity issues and all enhance this this particular features that we are
talking about, is also known as curse of complexity is optimal control.
So, we will end up with nice formulation but it’s a part of necessary conditions we are
leading towards this curse of complexity and traditionally, this two point boundary value
problems require computationally intensive procedures and we are going to see some of
these techniques in a while. Anyway ultimately also remember that when every things are written
down, it ultimately leads to this this open loop control structure actually.
Why so, because ultimately after we get the solution in iterative sense we are getting
a solution for a particular initial condition. If the initial condition is changed, then
the solution is not valid. So, because of that and once the initial condition is fixed
for all features, time t not to t f we will control strategy and it also gets fixed actually.
So, because of that, this is a future of getting getting into this open loop control structure
actually. So, here are some of the comments and difficulties
associated with always. So let’s proceed further and then talk about one particular
method called shooting. This is the traditional method, very iterative obviously, but I utilize
it on problems as well.
So, let us see what is the philosophy first and again just to summarize, you have the
state equation, costate equation, optimal control and boundary conditions. You once
again solve this optimal control equation ok, del is by del equal to 0 then I can end
up with this u equal to psi of X and lambda. So, in general, you may or may not be able
to solve it in explicit from. Ok, if you are not able to solve the explicit form, then
you have to solve using numerical techniques like neutralizing method or something like
that; but suppose many classes of problems, it is possible to solve u as an explicit function
of x and lambda. So, let us assume that it is available either
symbolically or numerically. Either way, its actually ok once we solve it. You can if necessary,
we put it back here in the lambda dot equation and if necessary, you can put it back here
as well. So, then, this u will be eliminated from both
the equations. Then you will end up with Xdot is nothing but X and lambda if it is a function
of X and lambda and lambda dot is a function of t x and lambda. Actually, most of the time,
we will not discuss about this time dependency but if it is, I mean, if it is there, then
it is not a matter of concern. Actually it can be handled directly. Now, let’s solve
this boundary conditions. I told you that this is final boundary condition and this
is initial condition; so, this strategy tells us that let us be little more intelligent
and have a guess for lambda 0 actually. Ok, so, how about starting with a guess and
again, this you need to emphasize here that queuing a lambda 0 value is not a trivial
task. It requires a lot of insight into the problem or essentially it it takes tricks
and techniques how to guess this this lambda 0. One way that I will recommend probably
is that you can guess a iterative control history, ok you you can rest from t not to
t a and once you guess, you then you can probably, I mean integrate, this equation state equation.
Ok, x is not available and your control history is available, so, you can integrate from t
not to t f. If using that particular control, guess control history, it is nowhere close
to optimal in general. But we know that control history becaus it is an iterative control,
anyway, so, if you integrate this state equation from t not to t f and get your x f, and once
you get your x f, then you can get this lambda f. Once you get your lambda f then you can,
I mean using this condition, you can integrate this equation backwards from t f to t not
now, and that is how you get a guess value for lambda 0. So, lambda 0 is not a trivial
task because several several regions like that, and one of the primary region is lambda,
in general does not have any physical value if it does have some some geometric value.
We will see that later but, but when geometrically it will be the only sense that gives some
sort of direction value. It does not talk about original value and lambda is not a physical
variable. We cannot talk about position velocity, current voltage, nothing like that, so, we
cannot really have a good guess for lambda in general. However, controlling the physical
quantity, we can guess the control history from t not to t f. We have agent on stabilizing
controller pid controller or just some heuristic controller from the from the problem of definition
itself ok. So, using that, or even sometimes I do recruitment
using 0 control if the system is stable. Then, simply start with a 0 based history and and
you can integrate the equations; the homogenous equation and then get it x f and then let
it be coaxially. There is also a possibility, and again we have prolong dependent once we
start working on a problem. We will see that in a later lecture on that, but the point
here is, guessing lambda 0 is not trivial, so, give some special emphasis first for guessing
a gold value of lambda 0. But once you guess it, for a while you can ignore this and then
take a look at these 2 differential equations, have a corresponding initial conditions available.
So, I can take these initial conditions and integrate this 2 equations forward state and
costate equation forward, because, you have already eliminated in terms of x and lambda
so these two are functions of x and lambda only I have two differential equations two
sets of differential equations two sets of differential equations and two sets of initial
condition, so you can use this and then integrating further, I will see that okay one side lambda
and t equal t f and that lambda what I am getting here, and this lambda computed this
they not be same and most of the time they will not be same. That is why, we were telling
that this integration process is something like suiting, I mean, you are suiting again
sort of thing towards target and and finally, it ands of somewhere with lambda f, but, the
the goal point is somewhere here. Lambda is del phi by del f what is should be and then
it we sense the error between the 2, and using that error we have to compute and improve
this lambda 0.
How you do that is, this procedure here; so you start with some lambda 0 lambda 0 for
first guess sort of thing and then start integrating forward I not to t f you learned of somewhere;
and then lambda f star to be somewhere. This delta lambda f you have to compute and correct,
this lambda 0 1 to lambda 0 2 and the procedure may go slightly more closer, may be lower
suite next time probably, usually the equations proceeds that way. Ok, so, to summarize, guess
the initial condition for the costate first, compute the control at each grid point because,
once you have the lambda 0 we have x 0 so u 0 is nothing but function of x 0 lambda
0. Once you have u 0, you can get from here; once once you have x one and lambda, one is
well. Once you have x one and lambda one, integrating this equations together,
you can calculate u one; and then, if u one said u one, you can calculate x2 and anything
like that usually. Now, let us so compute the control at each
grid point and then propagate the state and costate equations and then calculate the final
boundary conditions and error in the costate at the final time and then correct the costate
vector at the initial time based on this error at the final time. Then repeat the procedure,
keep repeating the procedure until and unless this lambda f, what you were getting from
integrating this point, is very close to what it should be. Then then you tell okay, everything
is satisfied because I am starting with initial conditions of the state anyway and then I
satisfy whatever I started, I learned of there and this two values are very close and this
is also satisfied and I used this control anywhere so this is satisfied and these two
equations I am integrating as a part of the procedure anyway, so, all the conditions will
be satisfied initially. So little bit breadth now until.
How do you do that actually? Okay, so, this what high recommend that form a meta state
vector x and lambda, that means instead of talking these two separate equations we can
always tell that is a function of x and lambda in a meta state; that we called as a meta
state vector because u is already eliminated from here. So, this implies that the error
in z dz is nothing but dx by d lambda, okay, and the information is used in the linear
equations getting basically okay; will see that in a way.
Okay, one side you can find this meta state vector, what is z dot z dot is nothing but
a and lambda dot is nothing but if I go back to here this is what in terms of to be f of
three. Okay, so and I associated with that z of t not is now available. Because of t
0, I have already guessed structure, so now corresponding to this differential equation,
which is knowing near, we can also obtain a corresponding linearised error dynamics
which is, a dz dot is nothing but dx dot ah d lambda dot now see nothing
but equal to del f by del z in to dz, using this taylors series linearized structure.
So, here is a meta state vector differential equation associated with boundary condition
and associated with these equation and and want to come in initial condition, we have
this corresponding error dynamics as well. All these things, we have to take advantage
of actually, okay? Now, okay, the point here is this: we have
a linearised version of a error dynamics, but remember that del f by del z is not a
constant matrix, okay, so, this matrix is certainly not a constant matrix, that is time
varying matrix. So, we cannot write this close form solution of if the power at n think like
that, we cannot do like that, so, what what is the usual way of time varying linear systems?
We know that it can be written in the form of state transition matrix. Ok, so, we have
this linear equation here, but this time, a varying matrix is involved here
and because of that the solutions terms out be something like that. So, dZ at any point
t j is nothing but phi of t j t I and the corresponding value dZ of t I; that is how
it proceeds this way ok? Now, if you have this form, that means if
you know the error at any point of time t I, you can calculated any other point of time
t j. Now, the corresponding equations of the state transition matrix turns out to be phi
of its t j t I and the dynamics and initial conditions. Okay, now that we have written
in this form, we need a solution, need a number for this matrix actually. So, this is very
standard linear system equations theory and then if you write a state transition matrix
the same, I mean the state transition matrix ingenerate, will satisfy the same differential
equation whatever differential equation we have. But, in the matrixes d z d z is a vector
were phi is a matrix usually but the differential equation turning out to be a I mean, exactly
in the same form basically So, you can take this d f by d z and plug in here and will
satisfy this equation, okay?, and then we have this phi of t 0 t 0 initial condition
is is invariably identity. So, if you see this t 0 t 0, both are same, then then this
has to be identity usually. So, the beauty here is that irrespective of
the problem definition, irrespective of the initial condition of the problem, this particular
phi can be obtained independently; actually form I t not to t f. This is a differential
equation and this is a corresponding, I mean, initial condition. We need the information
of del f by del z also all the way. That’s why, it is coupled in a way, about in which
segment you are talking about, was the value of del f bu del z can can vary depending on
that basically; but those are details Having said this, what happens is, you have a state
transition matrix and the corresponding initial conditions actually. So, this is the system
of equation that can be integrated and you get phi of t z t I. Okay, to proceed further
actually, so what it tells us is, you can now integrate this equation with corresponding,
I mean, corresponding state transition matrix equation as well okay?, and from t not to
t f and then once you once you know this actually, then your solution is almost ready because you you know this
d z. Once you know this d z of t I, then you can, I mean, you can proceed further, and
then you can integrate this equation. Also in a way this this equation and the corresponding
initial conditions are available. Similarly, this equation and corresponding initial conditions
are available, so, this set of equations along with this set of equation can be integrated
forward. Okay, once you integrate forward, then you can compute this thing at any point
of time, ok? Now, let we just use this differential equation, and just using this algebraic equation,
because the solution of this is now available from this actually.
Now, say this is the thing and then what we will get finally, we are getting something
like this actually. At t equal to t f dZ f by definition is d X f and d lambda f and
this is simply when we put it, we will get phi of t f t 0 dZ 0. Okay, now the question
it is we are actually interested in getting a value for dZ 0 because of what you know
is dZ f actually. Ok, ultimately what you know is because lambda f will be embedded
in lambda f I mean dZ f will contain in a d lambda f term and we using that will be
interested to calculate d lambda 0. The idea actually convert to that. dZ f is calculated
this way and hence, in principle this dZ 0 is nothing but this this phi inverse t f t
0 d z f. Also remember state transition matrix are never singular and hence this inverse
always exist actually. ok So, this is always possible to compute, Now,
the question here is, we do not want part of this d x not actually right, d x not is
fixed and we want to preserve that. So, invariably no matter what we get from the solution, that
we don’t want to consider that actually. So, consider this d x is not equal to 0, increased
only in updating lambda 0 okay? So, you can keep on doing that until some
convergence happened. This is fine okay, this is fine. I mean, if we can implement this
way. However, there is a
small computationally to equate the procedure, then we make this faster and this idea, therefore
you can you can partition this state transition matrix this way: Half way phi one and phi
2 exactly the made loop vertically, then d z f tends out. This way you can write it phi
1 times d x not plus phi 2 f tends d lambda not and then you can write this equation.
What you are looking for this equation it can be it can be we written in the form of
phi 1 f d x not plus phi 2 f d lambda not actually. Okay, now we are interested in forcing
this ah this ah lambda dX not is 0 anyway that’s what we discovered in in here, Okay,
so, because of that mm this terms are to be zero and hence we will apply actually okay?
So, what we are looking for is, now you can locate; for convenience, you can think that
H equal to some lambda f be the vector of boundary conditions at t f. That means the
way that you compute the desired lambda f, we can consider that is a function of lambda
f itself okay? in the okay You can generalize this problem a little bit further, I mean,
it can be simply lambda f itself do not be a function directly emitting, okay? So, anyway
so the what I mean is this this condition is available, okay, what we see that this this equation is available, so, using
this now you can think okay and I will consider that, I mean, call that is h or something
that, okay, then you can look at d lambda d I, mean what you’re saying here d z f
is actually nothing but del h by del z into d z f okay Alright, so you can do that and
then that tend turns out be lambda f minus lambda f star, but it should be, and then
you can substitute all that and then tell if I substituted back here. What I what I
am getting here is this d z f is, I mean, substitute by here and then required it do
and then I can
…
course. In the last couple of lectures,
we saw this optimal control formulation from calculus of variations and then we gave a
couple of examples which motivated that in general we need numerical methods to solve
all in optimal control problems. So, let us see a couple of such numerical
techniques and then we will follow it up with further things. Now,
let’s see the topics covered today. This lecture is first to very briefly review the
necessary conditions of optimality, then we will talk about 23 methods and then the first
shooting methods followed by gradient method and quasi linearization method.
There are various other techniques as well but it’s not necessary to continue talking
about all the methods. You can, after knowing these techniques, probably see some literature
or book or something to follow the other techniques
as well. As just started, this is the generic optimal control problem that we are interested
in that kind of optimizing or minimize or maximize certain performance index of this
form and along with the performance index we have this associated path constraint as
well as boundary conditions actually.
So, the necessary conditions of optimality that we derived from previous lectures come
out to be like that we have a state equation which is x dot be del H by del lambda which
is the f since this sort of thing is coming from the system dynamics. Typically, when
when followed by costate equation as well which is lambda dot is minus del H by del
X and both of the these referred to be differential equations dynamic equations really and then
the the third equation is optimal control equation which is a static equation. From
this equation we should be able to solve optimal control U H is a function of X and lambda
actually . So, that is how it develops now, for this 2 differential equations, we have
this two boundary conditions as well. So, the problem is well defined really and
we can see that this state equation takes the initial condition whereas costate equation
takes the final condition and that is how it lines up with those 2 point boundary value
problems. And then, this 2 point boundary value problem should account for all this
boundary conditions properly. Also review if a solution does not satisfy any one of
the equations including the boundary conditions that means that the solution is non optimal.
So, we must get to further conditions; almost get to all the conditions with equivalent
process itself. So, that’s the point. So, in this lecture,
before getting started with some of these techniques, we also discussed in last class
that this state costate equations are dynamic equations and it turns out that if 1 is stable
the other turns to be unstable in general. The optional control is a stationary equation
from which you need to solve for control and for which we need to know the lambda. Unless
you know lambda, we cannot solve it from here and boundary conditions are going to be split.
That’s what I told you here, that state equation is initial condition whereas co-state
equation is final condition and typically because of this, the the state the state equation
is double of forward whereas costate equation should be double of backward because you know
the the final condition. Basically, that’s that’s why in a typically costate equation
we can integrated backwards because we know the boundary condition at the final thing
in general. I mean the numerical techniques can circumference over this and then talk
about forward integration above the equations and all that we will just see that in a in
a while. Ok, so, with all this difficulty that the boundary conditions are split and
if one is stable and other one is unstable, that it leads to this great difficulty of,
I mean, this complexity issues and all enhance this this particular features that we are
talking about, is also known as curse of complexity is optimal control.
So, we will end up with nice formulation but it’s a part of necessary conditions we are
leading towards this curse of complexity and traditionally, this two point boundary value
problems require computationally intensive procedures and we are going to see some of
these techniques in a while. Anyway ultimately also remember that when every things are written
down, it ultimately leads to this this open loop control structure actually.
Why so, because ultimately after we get the solution in iterative sense we are getting
a solution for a particular initial condition. If the initial condition is changed, then
the solution is not valid. So, because of that and once the initial condition is fixed
for all features, time t not to t f we will control strategy and it also gets fixed actually.
So, because of that, this is a future of getting getting into this open loop control structure
actually. So, here are some of the comments and difficulties
associated with always. So let’s proceed further and then talk about one particular
method called shooting. This is the traditional method, very iterative obviously, but I utilize
it on problems as well.
So, let us see what is the philosophy first and again just to summarize, you have the
state equation, costate equation, optimal control and boundary conditions. You once
again solve this optimal control equation ok, del is by del equal to 0 then I can end
up with this u equal to psi of X and lambda. So, in general, you may or may not be able
to solve it in explicit from. Ok, if you are not able to solve the explicit form, then
you have to solve using numerical techniques like neutralizing method or something like
that; but suppose many classes of problems, it is possible to solve u as an explicit function
of x and lambda. So, let us assume that it is available either
symbolically or numerically. Either way, its actually ok once we solve it. You can if necessary,
we put it back here in the lambda dot equation and if necessary, you can put it back here
as well. So, then, this u will be eliminated from both
the equations. Then you will end up with Xdot is nothing but X and lambda if it is a function
of X and lambda and lambda dot is a function of t x and lambda. Actually, most of the time,
we will not discuss about this time dependency but if it is, I mean, if it is there, then
it is not a matter of concern. Actually it can be handled directly. Now, let’s solve
this boundary conditions. I told you that this is final boundary condition and this
is initial condition; so, this strategy tells us that let us be little more intelligent
and have a guess for lambda 0 actually. Ok, so, how about starting with a guess and
again, this you need to emphasize here that queuing a lambda 0 value is not a trivial
task. It requires a lot of insight into the problem or essentially it it takes tricks
and techniques how to guess this this lambda 0. One way that I will recommend probably
is that you can guess a iterative control history, ok you you can rest from t not to
t a and once you guess, you then you can probably, I mean integrate, this equation state equation.
Ok, x is not available and your control history is available, so, you can integrate from t
not to t f. If using that particular control, guess control history, it is nowhere close
to optimal in general. But we know that control history becaus it is an iterative control,
anyway, so, if you integrate this state equation from t not to t f and get your x f, and once
you get your x f, then you can get this lambda f. Once you get your lambda f then you can,
I mean using this condition, you can integrate this equation backwards from t f to t not
now, and that is how you get a guess value for lambda 0. So, lambda 0 is not a trivial
task because several several regions like that, and one of the primary region is lambda,
in general does not have any physical value if it does have some some geometric value.
We will see that later but, but when geometrically it will be the only sense that gives some
sort of direction value. It does not talk about original value and lambda is not a physical
variable. We cannot talk about position velocity, current voltage, nothing like that, so, we
cannot really have a good guess for lambda in general. However, controlling the physical
quantity, we can guess the control history from t not to t f. We have agent on stabilizing
controller pid controller or just some heuristic controller from the from the problem of definition
itself ok. So, using that, or even sometimes I do recruitment
using 0 control if the system is stable. Then, simply start with a 0 based history and and
you can integrate the equations; the homogenous equation and then get it x f and then let
it be coaxially. There is also a possibility, and again we have prolong dependent once we
start working on a problem. We will see that in a later lecture on that, but the point
here is, guessing lambda 0 is not trivial, so, give some special emphasis first for guessing
a gold value of lambda 0. But once you guess it, for a while you can ignore this and then
take a look at these 2 differential equations, have a corresponding initial conditions available.
So, I can take these initial conditions and integrate this 2 equations forward state and
costate equation forward, because, you have already eliminated in terms of x and lambda
so these two are functions of x and lambda only I have two differential equations two
sets of differential equations two sets of differential equations and two sets of initial
condition, so you can use this and then integrating further, I will see that okay one side lambda
and t equal t f and that lambda what I am getting here, and this lambda computed this
they not be same and most of the time they will not be same. That is why, we were telling
that this integration process is something like suiting, I mean, you are suiting again
sort of thing towards target and and finally, it ands of somewhere with lambda f, but, the
the goal point is somewhere here. Lambda is del phi by del f what is should be and then
it we sense the error between the 2, and using that error we have to compute and improve
this lambda 0.
How you do that is, this procedure here; so you start with some lambda 0 lambda 0 for
first guess sort of thing and then start integrating forward I not to t f you learned of somewhere;
and then lambda f star to be somewhere. This delta lambda f you have to compute and correct,
this lambda 0 1 to lambda 0 2 and the procedure may go slightly more closer, may be lower
suite next time probably, usually the equations proceeds that way. Ok, so, to summarize, guess
the initial condition for the costate first, compute the control at each grid point because,
once you have the lambda 0 we have x 0 so u 0 is nothing but function of x 0 lambda
0. Once you have u 0, you can get from here; once once you have x one and lambda, one is
well. Once you have x one and lambda one, integrating this equations together,
you can calculate u one; and then, if u one said u one, you can calculate x2 and anything
like that usually. Now, let us so compute the control at each
grid point and then propagate the state and costate equations and then calculate the final
boundary conditions and error in the costate at the final time and then correct the costate
vector at the initial time based on this error at the final time. Then repeat the procedure,
keep repeating the procedure until and unless this lambda f, what you were getting from
integrating this point, is very close to what it should be. Then then you tell okay, everything
is satisfied because I am starting with initial conditions of the state anyway and then I
satisfy whatever I started, I learned of there and this two values are very close and this
is also satisfied and I used this control anywhere so this is satisfied and these two
equations I am integrating as a part of the procedure anyway, so, all the conditions will
be satisfied initially. So little bit breadth now until.
How do you do that actually? Okay, so, this what high recommend that form a meta state
vector x and lambda, that means instead of talking these two separate equations we can
always tell that is a function of x and lambda in a meta state; that we called as a meta
state vector because u is already eliminated from here. So, this implies that the error
in z dz is nothing but dx by d lambda, okay, and the information is used in the linear
equations getting basically okay; will see that in a way.
Okay, one side you can find this meta state vector, what is z dot z dot is nothing but
a and lambda dot is nothing but if I go back to here this is what in terms of to be f of
three. Okay, so and I associated with that z of t not is now available. Because of t
0, I have already guessed structure, so now corresponding to this differential equation,
which is knowing near, we can also obtain a corresponding linearised error dynamics
which is, a dz dot is nothing but dx dot ah d lambda dot now see nothing
but equal to del f by del z in to dz, using this taylors series linearized structure.
So, here is a meta state vector differential equation associated with boundary condition
and associated with these equation and and want to come in initial condition, we have
this corresponding error dynamics as well. All these things, we have to take advantage
of actually, okay? Now, okay, the point here is this: we have
a linearised version of a error dynamics, but remember that del f by del z is not a
constant matrix, okay, so, this matrix is certainly not a constant matrix, that is time
varying matrix. So, we cannot write this close form solution of if the power at n think like
that, we cannot do like that, so, what what is the usual way of time varying linear systems?
We know that it can be written in the form of state transition matrix. Ok, so, we have
this linear equation here, but this time, a varying matrix is involved here
and because of that the solutions terms out be something like that. So, dZ at any point
t j is nothing but phi of t j t I and the corresponding value dZ of t I; that is how
it proceeds this way ok? Now, if you have this form, that means if
you know the error at any point of time t I, you can calculated any other point of time
t j. Now, the corresponding equations of the state transition matrix turns out to be phi
of its t j t I and the dynamics and initial conditions. Okay, now that we have written
in this form, we need a solution, need a number for this matrix actually. So, this is very
standard linear system equations theory and then if you write a state transition matrix
the same, I mean the state transition matrix ingenerate, will satisfy the same differential
equation whatever differential equation we have. But, in the matrixes d z d z is a vector
were phi is a matrix usually but the differential equation turning out to be a I mean, exactly
in the same form basically So, you can take this d f by d z and plug in here and will
satisfy this equation, okay?, and then we have this phi of t 0 t 0 initial condition
is is invariably identity. So, if you see this t 0 t 0, both are same, then then this
has to be identity usually. So, the beauty here is that irrespective of
the problem definition, irrespective of the initial condition of the problem, this particular
phi can be obtained independently; actually form I t not to t f. This is a differential
equation and this is a corresponding, I mean, initial condition. We need the information
of del f by del z also all the way. That’s why, it is coupled in a way, about in which
segment you are talking about, was the value of del f bu del z can can vary depending on
that basically; but those are details Having said this, what happens is, you have a state
transition matrix and the corresponding initial conditions actually. So, this is the system
of equation that can be integrated and you get phi of t z t I. Okay, to proceed further
actually, so what it tells us is, you can now integrate this equation with corresponding,
I mean, corresponding state transition matrix equation as well okay?, and from t not to
t f and then once you once you know this actually, then your solution is almost ready because you you know this
d z. Once you know this d z of t I, then you can, I mean, you can proceed further, and
then you can integrate this equation. Also in a way this this equation and the corresponding
initial conditions are available. Similarly, this equation and corresponding initial conditions
are available, so, this set of equations along with this set of equation can be integrated
forward. Okay, once you integrate forward, then you can compute this thing at any point
of time, ok? Now, let we just use this differential equation, and just using this algebraic equation,
because the solution of this is now available from this actually.
Now, say this is the thing and then what we will get finally, we are getting something
like this actually. At t equal to t f dZ f by definition is d X f and d lambda f and
this is simply when we put it, we will get phi of t f t 0 dZ 0. Okay, now the question
it is we are actually interested in getting a value for dZ 0 because of what you know
is dZ f actually. Ok, ultimately what you know is because lambda f will be embedded
in lambda f I mean dZ f will contain in a d lambda f term and we using that will be
interested to calculate d lambda 0. The idea actually convert to that. dZ f is calculated
this way and hence, in principle this dZ 0 is nothing but this this phi inverse t f t
0 d z f. Also remember state transition matrix are never singular and hence this inverse
always exist actually. ok So, this is always possible to compute, Now,
the question here is, we do not want part of this d x not actually right, d x not is
fixed and we want to preserve that. So, invariably no matter what we get from the solution, that
we don’t want to consider that actually. So, consider this d x is not equal to 0, increased
only in updating lambda 0 okay? So, you can keep on doing that until some
convergence happened. This is fine okay, this is fine. I mean, if we can implement this
way. However, there is a
small computationally to equate the procedure, then we make this faster and this idea, therefore
you can you can partition this state transition matrix this way: Half way phi one and phi
2 exactly the made loop vertically, then d z f tends out. This way you can write it phi
1 times d x not plus phi 2 f tends d lambda not and then you can write this equation.
What you are looking for this equation it can be it can be we written in the form of
phi 1 f d x not plus phi 2 f d lambda not actually. Okay, now we are interested in forcing
this ah this ah lambda dX not is 0 anyway that’s what we discovered in in here, Okay,
so, because of that mm this terms are to be zero and hence we will apply actually okay?
So, what we are looking for is, now you can locate; for convenience, you can think that
H equal to some lambda f be the vector of boundary conditions at t f. That means the
way that you compute the desired lambda f, we can consider that is a function of lambda
f itself okay? in the okay You can generalize this problem a little bit further, I mean,
it can be simply lambda f itself do not be a function directly emitting, okay? So, anyway
so the what I mean is this this condition is available, okay, what we see that this this equation is available, so, using
this now you can think okay and I will consider that, I mean, call that is h or something
that, okay, then you can look at d lambda d I, mean what you’re saying here d z f
is actually nothing but del h by del z into d z f okay Alright, so you can do that and
then that tend turns out be lambda f minus lambda f star, but it should be, and then
you can substitute all that and then tell if I substituted back here. What I what I
am getting here is this d z f is, I mean, substitute by here and then required it do
and then I can
…