A Practical Guide To The Solution of Real-Life Optimal Control Problems

1
A PRACTICAL GUIDE TO THE SOLUTION OF REAL-LIFE OPTIMAL CONTROL

PROBLEMS
H. J. PESCH
Mathematisches Institut, Technische Universitat Munchen
Arcisstr. 21, 80 290 Munchen, Germany
Abstract. The present paper is an introductory and survey paper of the treatment of realistically modelled optimal control
problems from applications in the aerospace eld. Especially those problems are considered which include dierent types
of constraints. In the tutorial part of the paper, recipes are given for the treatment of optimal control problems for which,
among other constraints, control and/or state variable inequality constraints are to be taken into account. Optimal control
problems having singular subarcs and/or discontinuities are also investigated. The discussion of the necessary conditions aims
to the subsequent application of the multiple shooting method, which is known to be a very precise and ecient method for
the solution of those multipoint boundary-value problems that arise from these necessary conditions. Homotopy techniques
as well as the fusion of direct collocation and multiple shooting techniques are described. Both approaches facilitate the
construction of an appropriate initial trajectory to start the multiple shooting iteration. In the survey part of the paper,
some recently published new guidance methods are described. These methods are based, on the one hand, on the theory
of neighboring extremals and, on the other hand, on the multiple shooting method. They are designed for the real-time
computation of optimal trajectories in aerospace applications. Five challenging optimal control problems from the eld of
aerospace engineering run throughout the whole paper and illustrate the use of the recipes, i.e., the necessary conditions that
must be satised by an optimal solution. Numerical results are given for these problems to demonstrate the performance of
the multiple shooting method for the o-line computation of optimal trajectories as well as the performance of the guidance
schemes for the on-line computation of optimal trajectories.
Keywords: Optimal control problems, necessary conditions, control variable inequality constraints, state variable inequal-
ity constraints, singular subarcs, multipoint boundary-value problems, multiple shooting, homotopy techniques, fusion of
direct and indirect methods, neighboring extremals, feedback controls, real-time computation, on-line computation, re-entry
problems, windshear problems, two-stage-to-orbit ascent.
1. Introduction
Many optimization problems in science and engineering can be described by optimal control problems
such as the control of a spacecraft or aircraft, of a chemical reaction or of an industrial robot. Thereby,
the consideration of dierent constraints is very important when realistic models are to be solved. Such
complex optimal control problems can today be treated by sophisticated numerical methods. If the
accuracy of the solution and the judgement of its optimality holds the spotlight, the multiple shooting
method seems to be superior over other methods. The complexity of optimal control problems that have
been solved by the multiple shooting method until now also contributes to this assessment. Some of these
problems may be cited here: the maximum payload ascent of a new generation of space transportation
systems known as the Sanger II project (partly published, currently still under research), maximum
payload missions by ion-driven spacecrafts to planetoids or outer planets (results just published), the
dangerous landing of a passenger aircraft in a windshear downburst (results just published), optimal
heating and cooling by solar energy (already published), and time or energy optimal control of robots
(partly published, still under research). The present investigations and the results obtained clearly show
the trend that control problems are treated of which the mathematical models become closer and closer
to reality. This causes an increase of complexity of optimal control problems by an increasing number of
constraints, by more complicated dierential equations, which themselves must be even generated by a
computer program and, nally, also by an increasing number of unknowns that are involved in the model.
The present survey paper looks out for two aims. First, the necessary conditions of optimal control
theory are summarized in a short course; see Sections 2{6. Thereby, special emphasis is placed on the
treatment of inequality constraints which must be taken into account when real-life applications are
treated. Each section starts with a summary of the necessary conditions, which can be used as a recipe.
For simplicity, the mathematical background is omitted completely. This may help to get some insight
in what kind of routine work has to be done to solve an optimal control problem that ts into the
rather general pattern of the problems discussed here. The summary is then followed by the discussion
2 H. J. PESCH
of a realistic problem from the eld of aerospace engineering. At the end of each section, the resulting
boundary-value problem is given which provides the basis for the subsequent numerical solution by means
of the multiple shooting method. This method and homotopy techniques as well as a new idea to unify
direct and indirect methods are described in Sections 7 and 8.
After this tutorial material, a survey is given on some new methods for the real-time computation
of optimal trajectories. This is the second intention of the present paper. For, if optimal solutions of
processes which run down very fast, such as the optimal ight of space vehicles or aircrafts, are to be
realized practically, it is not sucient just to prescribe the initial data and leave the process to its fate.
Instead, one needs a fast numerical method to compute the future course of the optimal control variables
during the process so that optimality conditions and prescribed constraints are preserved even in the
presence of disturbances. These requirements can be met by guidance schemes that are based on the
theory of neighboring extremals. Based on this theory, the ight path corrections can be obtained from
the solutions of linear boundary-value problems which can be solved very eciently. A rst method, a
so-called neighboring optimum guidance scheme, needs only a few matrix-times-vector operations which
have to be carried through on the onboard computer for the on-line computation of the corrections of the
optimal control variables. The second method, called repeated correction method, combines this linear
guidance law with a single integration of the equations of motion. By this method, not only the control
corrections, but also the control history as well as the associated trajectory can be computed for the
entire remaining ight time. In addition, the observance of the constraints can be veried before feeding
back the adjusted control history. The theoretical and numerical background of these methods is covered
in Sections 9{11.
In detail, the paper has the following outline. In Section 2, problems with control variable inequal-
ity constraints are discussed. The heating constraint re-entry of an Apollo capsula is investigated as an
example. The constraint of the angle of attack represents the control variable inequality constraint in
this example. Next, the maximum cross-range re-entry problem of a space-shuttle-orbiter-type vehicle
provides the example for the treatment of mixed state-control constraints, where both control and state
variables are involved. This kind of problems are discussed in Section 3. The treatment of state variable
inequality constraints is explained in Section 4. Again, the heating constraint re-entry of an Apollo cap-
sula serves as an example. Now, the altitude after the rst dip into the atmosphere is limited because
of safety requirements. This inequality constraint represents a second-order state constraint. Bang-bang
and singular control problems are investigated in Section 5. The abort landing of a passenger aircraft in a
windshear downburst serves as the example for this type of problems. Finally, problems with discontinu-
ities of the state variables appearing in the performance index and in certain constraints are discussed in
Section 6. The maximum payload ascent of a two-stage-to-orbit space transporter yields the illustrating
example. All these challenging optimal control problems of dierent degree of diculty may serve as a
set of benchmark problems for numerical methods designed for the solution of real-life optimal control
problems.
The multiple shooting method is described in brevity in Section 7 as well as homotopy techniques,
which are very useful to overcome the obstacle of nding an appropriate initial guess for starting the
multiple shooting iteration. Direct methods, such as direct collocation, can also be used to get the required
initial trajectory. Some numerical results for all problems are presented in Section 8.
In Section 9, the theory of neighboring extremals is given in brevity. The description of the guidance
methods is presented in Section 10. Some numerical results concerning the performance of these guidance
schemes for the on-line computation of optimal trajectories are given in Section 11.
Section 12 concludes this introductory and survey paper.
PRACTICAL GUIDE TO THE SOLUTION OF OPTIMAL CONTROL PROBLEMS 3
2. Optimal Control Problems with Control Variable Inequality Constraints
2.1. Summary of Necessary Conditions
At the beginning of the tutorial part of the paper, we concentrate on the treatment of the following
simple class of nonlinear optimal control problems. To be minimized is a functional
Ztf
I [u] := '(x(tf ); tf ) + L(x(t); u(t)) dt (1)
0
over all piecewise continuous control vector functions u : [0; tf ] ! U k where U denotes a non-empty,
IR
convex and closed set. The piecewise continuously dierentiable state vector function x : [0; tf ] ! n is
IR
subject to the following constraints in form of a system of ordinary dierential equations and of two-point
boundary conditions,
x_ (t) = f (x(t); u(t)) ; f : n+k !
IR IR
n ; (2a)
x(0) = x0 ; x0 2 n ; IR (2b)
(x(tf ); tf ) = 0 ; : n IR + ! IR
IR
q
(2c)
where the terminal time tf may be either specied or unspecied. It is assumed that all functions appear-
ing in Eqs. (1) and (2) are suciently often continuously dierentiable with respect to their arguments.
In the following, some results of the theory of optimal control are summarized. The notation of Bryson
and Ho (1987) is used to present the following necessary conditions which have to be satised by an
optimal solution. Dening the Hamiltonian H and the auxiliary function
H (x; u; ) := L(x; u) + >f (x; u) ; (3)
(x; t; ) := '(x; t) + > (x; t) ; (4)
an optimal solution of the problem (1) and (2) has to fulll the necessary conditions
_ > = ?Hx ; (5a)
u = arg min
u2U
H; (5b)
>(tf ) = x j tf : (5c)
If the terminal time tf is unspecied, the following necessary condition has to be included,
(t + H ) j tf = 0 : (5d)
Here : [0; tf ] ! n and 2 q denote the so-called Lagrange multipliers or adjoint variables. The
IR IR
partial derivatives Hx and x are to be understood as row >vectors, e.g., Hx := (@H=@x1 ; : : : ; @H=@xn ),
and the transpose of the column vector (:) is denoted by (:) , e.g., x> := (x1 ; : : : ; xn ).
2.2. Example: Apollo Re-Entry (Version 1: Constrained Angle of Attack)
A rst optimal control problem is investigated now to illustrate the procedure of how to transform an
optimal control problem by means of the above necessary conditions into a boundary-value problem that
is well suited for a subsequent numerical treatment. This rst problem describes the atmospheric re-entry
of an Apollo-type capsula. Since dierent versions of this problem are so often published in the literature,
see, e.g., Breakwell et. al. (1963), Scharmack (1967), and Stoer and Bulirsch (1980) and, more recently,
Pesch (1989b, 1990a), it has become a benchmark problem for optimization methods. Its complexity is
still low enough to serve as a tutorial example.
4 H. J. PESCH
The ight path of the capsula is assumed to take place in a vertical plane. Thus, the equations of
motion can be written as
V_ = ? 2Sm % V 2 CD (u) ? (1
g0 sin ;
+ )2 (6a)
_ = 2Sm % V CL (u) + RV(1cos ? g0 cos ;

+ ) V (1 + )2 (6b)
_ = VR sin ; (6c)
_ = 1 V+ cos : (6d)
Here V denotes the velocity, the ight path angle, the normalized altitude ( := h=R with h and R
denoting the altitude above the Earth's surface and the Earth's radius, respectively), and the distance
on the Earth's surface. The control variable is the angle of attack u. For the lift and the drag coecients,
the following relations are assumed, CD (u) = CD0 + CDL cos u with CD0 = 0:88 and CDL = 0:52
and CL (u) = CL0 sin u with CL0 = ?0:505. The air density is assumed to satisfy % = %0 exp(? R ).
All other quantities not mentioned here are constants.
The total stagnation point convective heating per unit area, i.e.,
Ztf
I [u] = 10 V 3 p% dt (7)
0
is to be minimized.
The vehicle is to be maneuvered into an initial position favorable for the nal splashdown in the
Pacic. Thus, all state variables are prescribed at the moment of entry. At the unspecied terminal time,
all state variables except the ight path angle are prescribed, too. More details, especially the values of
the constants and the boundary conditions, can be found in Pesch (1989b).
In addition, the range U of the control variable is described by the inequality constraint
juj umax (8)
with a given positive constant umax . This completes the description of the model.
We now establish the necessary conditions of optimal control theory. In a rst step, the optimal control
variable is eliminated in terms of the state and the adjoint variables. If the optimal control variable lies in
the interior of the set U , the following equations must be satised because of the minimum principle (5b),
Hu = 0 ; (9a)
Huu > 0 : (9b)
In the general case of more than one control variable, the strong Legendre-Clebsch condition (9b) is to
be understood as Huu must be a positive denite matrix. Because of technical reasons, there must hold
u 2 [?; ] here. The optimal control variable u is then given, according to Eq. (9a), by
tan u = ? CCL0V : (10)
DL V
By this condition, u is not uniquely dened. The quadrant is determined by Eq. (9b). We nd sign u =
sign or, in more detail,
sin u = ?CL0 q ; (11a)
(CDL V V )2 + (CL0 )2
cos u = +CDL q V V : (11b)

(CDL V V )2 + (CL0 )2
The following diagram shows the relation between the quadrant of u and the signs of the adjoint
variables V and ,
u ==2

V < 0 V > 0
> 0 > 0
u= u=0
u = ?

V < 0 V > 0
< 0 < 0
u = ?=2
Note that the trajectory has a so-called corner at times tcorner where there holds (tcorner ) = 0 and
V (tcorner) < 0. This results in a jump of the optimal control variable from ? to + or vice versa
if umax = is chosen for the maximum value of the control variable. In this case, the state variables,
however, remain continuously dierentiable. However, this is not true for umax < as we will see. The
trajectory then has, indeed, a corner since the ight path angle has a jump discontinuity at t = tcorner .
The adjoint variables can be computed via the Eqs. (5a), which are not given in detail here.
On constrained subarcs, where the control variable, if chosen according to Eqs. (11), would violate
the inequality contraint (8), we have either u = umax or u = ?umax . The minimum principle (5b)
yields sign u = sign ufree . Here, ufree denotes the competitive control variable, which is not active along
constrained subarcs and determined by Eqs. (11). Hence, we have
u = umax sign ufree (12)
on constrained subarcs. At junction points between unconstrained and constrained subarcs, the control
variable is continuous. However, there arises a discontinuity in the interior of a constrained subarc if
changes its sign, say at t = tcorner , while simultaneously there holds V (tcorner ) < 0. This is because ufree
jumps at t = tcorner from ? to + or vice versa depending on the sign of _ (tcorner ); see the above
diagram. Note that the Erdmann-Weierstra corner condition yields the continuity of , H , and Hu at
t = tcorner; see, e.g., Bryson and Ho (1987), p. 125.
Finally, the set of necessary conditions is completed with the so-called transversality conditions (5c)
and (5d),
(tf ) = 0 ; (13a)
H j tf = 0 : (13b)
Note that is the only state variable not prescribed at the terminal time tf , i.e., = 0. The adjoint
vector drops out.
2.3. The Multipoint Boundary-Value Problem
In summary, the entire set of necessary conditions yields a multipoint boundary-value problem. To achieve
the so-called standard form of the boundary-value problem, we have to transform the undetermined
interval [0; tf ] onto the interval [0; 1] by introducing a new independent variable via := t=tf . Then, tf
becomes a new dependent variable and, because of (:)0 := dd (:) = tf (:_) := tf ddt (:), the right-hand sides
of Eqs. (6) as well as of Eqs. (5a) are to be multiplied by tf . Hence, we have a system of 9 dierential
equations (the dierential equation for the unknown terminal time is given by t0f = 0) with 9 two-
point boundary conditions (7 boundary conditions are given by the model, 2 conditions follow from the
6 H. J. PESCH
transversality conditions (13)). In addition, each junction point t = tjunction between unconstrained and
constrained subarcs yields an interior boundary condition of type
ufree(tjunction ) umax = 0 : (14)
Corners are also determined by an interior boundary condition,
(tcorner ) = 0 (15)
if, in addition, there holds
V (tcorner) < 0 : (16)
By this procedure, the necessary conditions are transformed into a multipoint boundary-value problem
which can be solved by the multiple shooting method very accurately and eciently; see Section 7. Note
that a solution of the above boundary-value problem is not a candidate for an optimal solution of the
control problem if the inequality (8) and the sign conditions sign u = sign ufree and (16) are not fullled
along the entire trajectory or at all corner points tcorner , respectively.
3. Optimal Control Problems with Mixed State-Control Inequality Constraints
Realistically modelled problems generally include inequality constraints of more complicated type than
Eq. (8). At the rst level of increased compexity, we consider so-called mixed state-control constraints of
type
C (x(t); u(t)) 0 ; C : n+k ! l
IR IR (17)
where C explicitly depends on the control vector u, i.e., Cu 6= 0. For the sake of simplicity, we rstly
restrict the following explanations to the case k = l = 1, thus one control variable and one inequality
constraint. Dening the Hamiltonian by
H (x; u; ; ) := L(x; u) + >f (x; u) + C (x; u) ; (18)
the necessary conditions (5) remain unchanged. Additionally we have a necessary sign condition for the
Lagrange parameter ,
= 0 if C < 0 ,
0 if C = 0 . (19)
Therefore, the right-hand sides of the dierential equations for the adjoint variables (5a) are, in general,
to be modied along constrained subarcs, where
C (x; u) = 0 for all t with tentry t texit and tentry < texit : (20)
The moments t = tentry and t = texit denote entry and exit point, respectively, of a constrained subarc.
If Eq. (20) can be uniquely solved for u, i.e., if Cu 6= 0, the control variable can be represented along a
constrained subarc by a function of the state variables,
u = ubound (x) : (21)
If Cu 6= 0, the multiplier is given by means of Eq. (9a),
= ?Cu?1 ( Lu + >fu ) : (22)
If k > 1 or l > 1, but each component of C depends explicitly on one control vector component
only, this case can be treated by adding terms in the denition (18) of the Hamiltonian. If one control
vector component is constrained by several components of C , this problem can be reduced to the previous
problem by introducing a new surrogate inequality constraint. The single case that has to be excluded
because of its possible non-uniqueness is the case where one component of C depends explicitly on several
control vector components. This case is of minor practical importance only and cannot be treated with
sucient generality.
3.2. Example: Space-Shuttle Re-Entry Under a Heating Constraint
For illustration purposes, we consider a second example from the aerospace eld. The following optimal
control problem describes the maximal cross-range re-entry of a space-shuttle-type vehicle. Increased
range capacity of the vehicle allows more frequent returns from the orbit. The functional to be minimized
is the negative cross-range angle at the unspecied terminal time tf ,
I [u] = ?(tf ) : (23)
Assuming no planet rotation and oblateness, a stationary atmosphere, a point mass vehicle, and a constant
drag polar, the equations of motion with respect to a ight-path-oriented coordinate system may be
written as
2
_V = ? (CD0 + CLn) F%0 exp(?h) V 2 ? g0 R
2m R + h sin ; (24a)
_ = CL F%
2m
0 exp(?h) V sin ? V cos cos tan ;
cos R + h (24b)
R 2 !
_ = CL F%
2m
0 exp(?h) V cos ? g0
V R+h ? V
R + h cos ; (24c)
_ = R V+ h cos sin ; (24d)

h_ = V sin ; (24e)
_ = R V+ h coscos
cos :
(24f)
The state variables are the velocity V , the heading angle , the ight path angle , the cross-range
angle , the altitude h, and the down-range angle . The control variables are the bank angle and the
lift coecient CL . The values of the constants of the drag polar are CD0 = 0:04 and n = 1:86.
Initial values for all state variables are prescribed as well as terminal conditions for V , , and h.
Since the dierential equation for the down-range angle is decoupled, it need not be considered for the
optimization process and can be computed afterward.
Moreover, the following mixed state-control or zeroth order state constraint has to be taken into
account, which limits the skin temperature of the vehicle,
CL ? CLH (V; h) ? CLH 0 (25)
where CLH is a rather complicated function of V and h consisting of 20 terms each of which is the
sum of a polynomial of second degree in V=h plus a polynomial of second degree in h=V multiplied by
a polynomial of rst degree in h. Dierent levels of the limit skin temperature T are indicated by the
parameter CLH where CLH = 0 corresponds to a limit skin temperature of T = 1093 C. For details,
see Kugelmann and Pesch (1990b).
According to Eqs. (9), the optimal control variables are given by
sin = ? =(w cos ) ; cos = ? =w ; (26)
where
h i1=2
w = (= cos )2 + 2 ;
and 8 free
>
< CL := [?w=(V V n)]1=(n?1) ; on unconstrained arcs,
CL = > (27)
: CLbound := CLH (V; h) + CLH ; on constrained arcs.
8 H. J. PESCH
Note that Eq. (9b) is fullled if H > 0 and H HCL CL ? HC
2 > 0. This obviously implies HC C >
L L L
0, too. From these inequalities, which are to be satised on unconstrained subarcs only, we obtain an
additional sign condition that must be fullled along unconstrained subarcs,
V < 0 : (28a)
Note that the above determinant condition is trivially satised for the parameter values of the problem.
For mixed state-control constraints, the right-hand sides of the adjoint variables are to be modied
along constraint subarcs; see Eq. (5a) together with the denition (18). For the sake of brevity, we give
here the equation for h only,
2
_ h = ? V (CD0 + k CLn) F%
2m
0 exp(?h) V 2 + 2 g0 R sin
(R + h)3
V cos cos tan
+ ?CL F%
2m
0 exp(?h) V sin +
cos (R + h)2
2 !
F% 0 g 0 R V
+ ?CL 2m exp(?h) V cos + 2 (R + h)3 V ? (R + h)2 cos
V @C LH

+ ? (R + h)2 cos sin ? @h
where the Lagrange parameter is dened by Eq. (9a),

80 ; on unconstrained arcs,
>
<
= > F%0
: n?1
2m exp(?h) V [w + V V n CL ] ; on constrained arcs.
The necessary sign condition (19) implies
V ? w ( < 0) (28b)
V n CLn?1
which must be fullled on constrained subarcs and therefore completes the sign condition (28a).
Finally, the missing boundary conditions are given by the transversality conditions (5c) and (5d),
(tf ) = 0 ; (tf ) = ?1 ; H j tf = 0 : (29)
In summary, we have again a multipoint boundary-value problem which must be solved by a candidate
for the optimal trajectory. We have 11 dierential equations for V , , , , h, their associate adjoint
variables, and for tf . In contrast to the boundary-value problem of Subsection 2.3, the right-hand sides
are piecewise dened here. In addition, 8 two-point boundary conditions are given by the model, and
3 terminal conditions are given by the above transversality conditions. Each entry or exit point, tentry
or texit , repectively, of the mixed state-control constraint (25) is determined by an interior condition of
the type
h i
CLfree ? CLH (V; h) ? CLH tentry=exit = 0 : (30)
If the down-range angle is to be computed, too, its dierential equation is also included together with
an appropriate initial value.
After the computation of the solution of the boundary-value problem, the solution must be checked
whether the inequalities (25) and (28) are satised.
At the end of this section, it should be mentioned that control variable inequality constraints of
type C (u(t)) < 0, which are treated in Section 2, t also into the above pattern. Because of Cx = 0, the
dierential equations for the adjoint variables need not be modied in this case, and therefore need not
be computed, in order to establish the boundary-value problem. However, the sign condition (19) must
be veried.
4. Optimal Control Problems with State Variable Inequality Constraints

The next degree of complexity is given by optimal control problems with state variable inequality con-
straints,
S (x(t)) 0 ; S : IR
n ! IRl : (31)
In the following, we summarize some results of optimal control theory for problems with state variable
inequality constraints. For the sake of simplicity, let us assume that k = l = 1. We rst consider the case
that the constraint is active on a non-vanishing subinterval [tentry ; texit ],
S (x(t)) 0 for all t 2 [tentry ; texit ] [ 0; tf ] : (32)
By successive dierentiation of (32) with respect to time and substituting (2a), we nd the smallest
non-negative number q such that there holds
S (x) 0 ; S (1) (x) 0 ; : : : ; S (q?1) (x) 0 on [tentry ; texit ] (33a)
and
S (q) (x; u) 0 with Su(q) (x; u) 6= 0 : (33b)
Here, S (i) denotes the i-th total time derivative of S . Then, q is called the order of the state constraint.
Now S q(x; u) plays the role of C (x; u) in (18) so that the Hamiltonian, in this case, is dened by
H (x; u; ; ) = L(x; u) + >f (x; u) + S q(x; u) : (34)
Again, for we have the necessary sign condition
= 0 ; if S < 0 ,
0 ; if S = 0 . (35)
On constrained arcs, we obtain the optimal control variable u from (33b) and from (9a). The right-
hand sides of the dierential equations for the adjoint variables (5a) are, in general, to be modied
on [tentry ; texit ]. In order to guarantee that not only (33b) but also (33a) is satised, we also have to
require that the so-called entry or tangency conditions are fullled,

N >(x(tentry ); tentry ) := S (x(tentry )); S (1) (x(tentry )); : : : ; S (q?1) (x(tentry )) = 0 : (36)
Alternatively, the tagency conditions can also be posted at the exit point. The following explanations
have then to be modied suitably.
Interior point conditions of the type
n ( 0; t q
N (x(tinterior ); tinterior ) = 0 ; N : IR f ) ! IR (37)
10 H. J. PESCH
give rise to additional necessary conditions,
>(t+interior ) = >(t?interior ) ? >Nx j tinterior ; (38a)
H (t+interior) = H (t?interior ) + >Nt j tinterior (38b)
where 2 q denotes another Lagrange multiplier. The Eq. (38b) determines tinterior , and the q compon-
IR
ents of are chosen so that the interior point constraint (37) is satised. The multiplier in uences (37)
via the dierential equations indirectly by the jump condition (38a). The generalization to interior point
conditions at several points tinterior is obvious.
From the tangency conditions (36) holding at the entry point of a constrained subarc, we have that
generally is discontinuous at t = tentry whereas is continuous at the exit point texit . In case the tangency
conditions are placed at the exit point, this result holds vice versa.
Sometimes boundary or touch points occur instead of boundary arcs. If, for example, the order is q = 2,
the following conditions hold at a touch point ttouch ,
S (x(ttouch )) = 0 ; S (1) (x(ttouch )) = 0 : (39)
The rst condition is regarded as an interior point condition of type (37) and yields a possible discontinuity
of . The second condition determines the touch point ttouch .
Minimax or so-called Chebyshev optimal control problems also t into the pattern of state-constrained
optimal control problems; see Section 5.
4.2. Example: Apollo Re-Entry (Version 2: Constrained Altitude)
To illustrate the use of the above extended set of necessary conditions, we again consider the re-entry
problem of Subsection 2.2. Now, we replace the control variable inequality constraint (8) by the state
variable inequality constraint
max (40)
which has, because of safety reasons, to be taken into account after the rst dip into the atmosphere, i.e.,
for all t > tmin for which (tmin ) = 0 for the rst time. Here, max is a given positive constant.
One easily nds by successive dierentiation of S (x(t)) := (t) ? max with respect to time t that the
state constraint is of the second order, thus q = 2. The optimal control variable is given on a constrained
subarc by
2 m g 0 1
sin u = F % C ? : (41a)
L0 V 2 (1 + max)2 R (1 + max)
Note that S (1) 0 and S (2) 0 imply _ 0. The optimal control variable along a constrained subarc is
uniquely determined by the minimum principle (5b) from which follows
sign cos u = sign V : (41b)
By an indirect proof, it can be shown that there holds cos u 6= 0 on ]tentry ; texit [. This directly implies
that S (2) (x; u) = 0 can be solved for u on the entire open interval ]tentry ; texit [. Since V (tentry ) < 0
implies V (t) < 0 on ]tentry ; texit [, which can also be proved indirectly, the control variable is continuous
on ]tentry ; texit [. Because of the continuity of the Hamiltonian, see Eq. (38b), it follows that V_ and _ are
continuous at the junction points tentry and texit unless, at these points, there holds V = 0 and = 0
as well. If this case is disregarded, S (2) and also u are continuous at the junction points, too.
In case of a constrained subarc, the entry conditions (36) yield
(tentry ) = max ; (42a)
(t?entry ) = 0 (42b)
where the minus sign at tentry indicates that the control variable is to be chosen according to Eqs. (11).
Because of the continuity of the control variable at the junction points, it follows
ufree(tentry ) = ubound (tentry ) (42c)
ufree(texit ) = ubound (texit ) (42d)
where ufree is dened by the Eqs. (11) and ubound by the Eqs. (41). The present case is completed by the
jump conditions
(t+entry ) = (t?entry ) ? 1 ; 1 2 ;
IR (43a)
(t+entry ) = (t?entry ) ? 2 ; 2 2 :IR (43b)
In summary, the 4 additional unknowns tentry , texit , 1 , and 2 are determined by the 4 interior
boundary conditions (42). The above jump conditions have to be carried through at the entry point of
the constrained subarc and cause that the Lagrange parameters 1 and 2 have in uence on the solution
of the system of dierential equations. The right-hand sides of the adjoint variables are to be modied
along a constrained subarc; compare Section 3.
It is known from the literature that state constraints of even order may have boundary arcs as well as
touch points if the Hamiltonian is regular, i.e., if H has a unique minimum with respect to u in a small
neighborhood of the optimal trajectory for all t 2 [0; tf ]. Moreover, it is known that state constraints of
the rst order do not become active in form of touch points; only boundary arcs occur. Furthermore, for
state constraints of odd order with q > 3, only touch points occur; boundary arcs are impossible. Proofs
for these results can be found in Jacobson et. al. (1971).
Hence, we here have to take into account the appearance of touch points, too. According to Eqs. (39),
we have two conditions at a touch point ttouch ,
(ttouch ) = max ; (44a)
(ttouch ) = 0 : (44b)
The rst of these conditions is considered as an interior point constraint and gives rise to a jump condition,
(t+touch ) = (t?touch ) ? ; 2 :
IR (45)
The second condition (44b) determines the touch point. The Lagrange parameter must be chosen so
as to satisfy the interior point constraint (44a).
In case that several boundary arcs and several touch points exist, the above interior boundary con-
ditions and their associated jump conditions must be multiplied suitably. If a boundary arc is adjacent
to t = 0 or t = tf , either the Eqs. (42a){(42c) and (43) or Eq. (42d) have to be dropped. In the latter
case, it is advisable to reduce the number of unknowns in the boundary-value problem by placing the
jump conditions at t = tf since they can be neglected then.
For the version of the re-entry problem with the altitude constraint, the terminal ight path angle is
prescribed, so that the terminal condition (13a) must be cancelled. In summary, we have again a system
of 9 dierential equations with a corresponding number of two-point boundary conditions. In addition,
there are 4 interior boundary conditions for each interior boundary arc by which the 4 unknowns, entry
and exit point and the two jump parameters 1 and 2 , are determined. For each touch point, there
are 2 interior boundary conditions by which the touch point and the jump parameter is determined.
Thus, we have now a dierent type of a multipoint boundary-value problem which includes, because of
the Eqs. (43) and (45), jump conditions, too.
In addition, necessary sign conditions concerning both the Lagrange parameter and its time deriv-
atives up to the order q and the jump parameter vector must be fullled by an optimal solution. By
12 H. J. PESCH
these necessary conditions, which are based on Jacobson et. al. (1971) and Maurer (1976), non-optimal
solutions can be singled out; see, e.g., Bulirsch et. al. (1991a) where a summary of those sign conditions
can be found. These necessary conditions can be written for the present re-entry problem as
_ 0 ; 0 ; (46a)
1 0 ; 2 0 ; 0 : (46b)
At the end of this section, it should be mentioned that control variable inequality constraints of
type C (x(t); u(t)) < 0, which are treated in Section 3, t also into the above pattern; they can be
considered as zeroth order state constraints. No discontinuities arise for the adjoint variables if q = 0.
5. Optimal Control Problems with Singular Subarcs

Among the most complicated optimal control problems, we have problems where the control variables
appear linearly in both the functional and the equations of motion. The diculties are caused by the fact
that the control variables may have so-called singular subarcs. The treatment of those singular subarcs
will be explained in the following.
We refer to the general formulation of the optimal control problem in Section 2.1. Let us assume that
at least one control variable appears linearly in Eqs. (1) and (2a) and that the set U mentioned after
Eq. (1) is a compact and convex polyhedron which is indeed the case in most practical problems. For the
sake of simplicity, let k = 1 and U = [ umin; umax ]. Then, the Hamiltonian has the form
H (x; u; ) := A(x; ) + u B (x; ) :
If the term B (x; ) does not vanish identically on a non-vanishing subinterval [tentry ; texit ] of [0; tf ],
the minimum principle (5b) yields
u ; if B < 0 ,
u = umax; if B > 0 . (47)
min
In this case, we have so-called bang-bang subarcs only. Each isolated zero of the switching function S (t) :=
B (x(t); (t)) indicates a switch from u = umax to u = umin or vice versa.
If
B (x; ) 0 for all t 2 [tentry ; texit ] [0; tf ] ; tentry < texit ; (48)
singular subarcs occur. The optimal control variable on these singular subarcs can be computed in a
similar way as for state-constrained problems. By successive dierention of the switching function S with
respect to time t and the substitution of the Eqs. (2a) and (5a), we may nd a smallest positive integer q~
such that
S (~q) (x; u; ) 0 for all t 2 [tentry ; texit ] (49a)
and
Su(~q) (x; u; ) 6= 0 : (49b)
If such a smallest integer q~ exists, it must be even; see McDanell and Powers (1971). Therefore, p with
q~ = 2 p is called the order of the singular subarc. The case of most practical interest is p = 1 which
will be considered here only. The control variables near singular subarcs of higher order generally show a
chattering behavior. According to McDanell and Powers (1971), the control variable u is, for rst-order
singular control problems, either discontinuous or continuously dierentiable at the junction points tentry
and texit . The rst case occurs in general. The analogon to the strong Legendre-Clebsch condition (9b) for
singular subarcs is the so-called strong generalized Legendre-Clebsch condition, see Kelley et. al. (1967),
which reads as
(?1)p B2p (x; ) > 0 for tentry t texit (50)
where B2p (x; ) is the factor in front of u after 2p dierentiations of the identity (48).
The interior boundary conditions to determine bang-bang switching points t = tbang and entry and
exit points of singular subarcs are given either by isolated zeros of the switching function
S (tbang ) = 0 (51)
or by two tangency conditions, either by
S (tentry ) = 0 ; S (1) (tentry ) = 0 (52a)
or by
S (tentry ) = 0 ; S (texit ) = 0 : (52b)
Jump conditions do not occur here. Again, we obtain well-dened multipoint bondary-value problems.
Each bang-bang switching point tbang is determined by an equation of type (51). Each pair of an entry
and an exit point of singular subarcs are determined by equations of type (52a) or (52b).
5.2. Example: Abort Landing of a Boeing 727 in a Windshear Downburst
For illustration purposes, we consider here the problem of abort landing of a passenger aircraft in the
presence of a windshear downburst which is taken from Miele et. al. (1987) and Bulirsch et. al. (1991a).
This is one of the most complicated optimal control problems ever solved. Here, we will discuss the
treatment of the bang-bang and singular subarcs only which indeed appear in the candidate trajectory
for the optimal solution of the problem.
To set up the equations of motion, we assume that the aircraft is a particle of constant mass, the ight
takes place in a vertical plane, and Newton's law is valid in an Earth-xed system. Moreover, the wind
ow eld is assumed to be steady. Under these assumptions, the kinematical and dynamical equations
are
x_ = V cos + Wx ; (53a)
h_ = V sin + Wh ; (53b)
V_ = mT cos( + ) ? D ? g sin ? (W_ cos + W_ sin ) ; (53c)
m x h
T sin( + ) + L ? 1 g cos + 1 (W_ sin ? W_ cos ) ;
_ = mV (53d)
mV V V x h
_ = u : (53e)
The state variables are the horizontal distance x, the altitude h, the relative velocity V , and the relative
path inclination . In the formulation above, the relative angle of attack is regarded as a state variable,
too. In fact, its time derivative is chosen as control variable. The approximations of the aerodynamic
forces, the trust T = T (V; ), the drag D = D(V; ), and the lift L = L(V; ), as well as the prescribed
wind velocity components Wx (x) and Wh(x; h) can be found in Bulirsch et. al. (1991a). The power
setting , normally also a control variable, is specied in advance as a function of time; see, e.g., Bulirsch
et. al. (1991a). All other quantities not mentioned here explicitly are constants and given also in Bulirsch
et. al. (1991a).
The following inequality constraints are imposed on the problem
juj umax ; (54a)
max : (54b)
14 H. J. PESCH
The second constraint is a state constraint of the rst order and can be treated analogously to the re-
entry problem of Section 4.2. For the boundary conditions see, e.g., Bulirsch et. al. (1991a). The terminal
time tf is unspecied here, too.
To avoid crashing on the ground, the ground clearance, or in other words the minimal altitude, has to
be maximized,
max min h(t) :
u2U 0tt f
Here, U is described by the control variable inequality constraint (54a). Instead, we can also minimize
the peak value of the altitude drop, that is, the dierence between a constant reference altitude hR and
the instantaneous altitude,
I [u] := 0max fh ? h(t)g :
tt R
f
(55)
The reference altitude hR has to be chosen so as to satisfy hR h(t) for all t 2 [0; tf ]. In any case, we
have a so-called Chebyshev-type optimal control problem, which can be transformed into the standard
form of a Mayer's functional by introducing a new state variable. For the functional (55) dene
(t) := max
^
fhR ? h(t^)g : (56)
0ttf
The variable is now subject to the additional constraints
_ = 0 ; (57a)
hR ? h(t) ? (t) 0 : (57b)
The functional (55) can then be written in standard form
I[u] = (tf ) : (58)
Note that the transformation brings a new state constraint into the game which can be shown to be of
the third order. The treatment can also follow the lines of Section 4. For details, see also Bulirsch et.
al. (1991a). Note that this abort landing problem presents an example with a non-regular Hamiltonian so
that the theoretical results concerning boundary points and boundary arcs induced by state constraints
do not apply here.
As mentioned above, we restrict the investigations to the computation of the control variable for
state-unconstrained subarcs only to demonstrate the use of the necessary conditions for bang-bang and
singular subarcs. From the minimum principle (5b), we obtain a bang-bang expression for the optimal
control variable,

u = ?uumax ;; for < 0 , (59)
max for > 0 .
The adjoint variable plays the role of the switching function. Its isolated zeros,
(tbang ) = 0 ; (60)
mark the switches of the bang-bang subarcs. If non-isolated zeros of occur, the following relation
holds,
(t) 0 on tentry t texit ; tentry < texit : (61)
Repeating twice the dierentiation of this identity with respect to time and substituting Eqs. (53)
and (5a), we obtain an expression for the optimal control on singular subarcs,
Asing ;
using = ? B (62)
sing
with
h i h _ ih i
Asing (t) := _ V T sin( + ) + D ? V1 _ ? VV T cos( + ) + L
h i h i
+V T_ sin( + ) + DV V_ ? V1 T_ cos( + ) + LV V_ ;
h i h i
Bsing (t) := V T cos( + ) + D ? V1 L ? T sin( + ) :
Hence, the order of the singular control is p = 1, and the strong generalized Legendre-Clebsch condition
is
Bsing < 0 for tentry t texit : (63)
The two junction points are determined by the entry or tangency conditions, either by
(tentry ) = 0 ; _ (tentry ) = 0 (64a)
or by
(tentry ) = 0 ; (texit ) = 0 : (64b)
Again, we see that the necessary conditions lead to a well-dened multipoint boundary-value problem.
Each switching point tbang , tentry , or texit is accompanied by an interior boundary condition. For the
complete description of the boundary-value problem, it is refered to Bulirsch et. al. (1991a).
If realistic models are investigated, characteristic parameters of the system are often given by tabular
data. In this case, possible points of non-smoothness of the approximation of the data must be considered
as interior point constraints of type (37), too. This causes additional switching points where the adjoint
variables may have discontinuities. See Bulirsch et. al. (1991a) for details.
6. Optimal Control Problems with Discontinuities in the State Variables
Problems with discontinuities in the state variables at undetermined interior points where instead of (1)
a functional of the form
Ztf
I[u] := '(x(t?stage ); x(t+stage ); x(tf ); tf ) + L(x(t); u(t)) dt (65)
0
is to be minimized subject to the additional equality constraint
(x(t?stage ); x(t+stage )) = 0 ; : 2n ! n ;
IR IR (66)
lead to the additional necessary conditions
H j t?stage = H j t+stage ; (67a)

>(tstage ) = @ (67b)
@x(tstage )
where
(x? ; x+ ; x; t; ) := '(x? ; x+ ; x; t) + > (x? ; x+ ) :
These equations determine x and at the discontinuity t = tstage , the parameters 2 n and the staging
IR
time tstage . If one or more of the quantities tstage , xi (t+stage ) and xi (t?stage ) are specied, the corresponding
equations in (67a) and (67b) are omitted. A generalization to more discontinuities is obvious.
16 H. J. PESCH
6.2. Example: Ascent of a Two-Stage-To-Orbit Vehicle
The Sanger II project of a European space transportation system is, at present, in the focus of industrial
and scientic discussion and development. In the future, a completely reusable space transportation
system will be necessary to maintain and service cost-eciently the planned international space station.
The two-stage-to-orbit vehicle is designed to launch horizontally and to deliver either a manned or an
unmanned cargo unit into orbit. The rst stage is equipped with wings and airbreathing engines. The
second stage is conventionally rocket propelled. The system is capable of performing cruising ights.
Some ideas for such a space transportation system have already been developped by Eugen Sanger in
the thirties. His investigations were republished in 1962; see Sanger (1962).
The rst mathematical model which included a simultaneous staging and trajectory optimization for
a two-stage rocket propelled predecessor model goes back to Shau (1973). A generalization to a three-
dimensional model was presented by Bulirsch and Chudej (1991, 1992a). Recently the model was upgraded
to include the airbreathing engines of the lower stage; see Bulirsch and Chudej (1992b) and Chudej (to
appear).
The payload of the space transporter is to be maximized,
mpayload = m(tf ) ? II (m(t+stage ) ? m(tf )) (68)
subject to the stage separation condition
m(t+stage ) = m(t?stage ) ? I (m0 ? m(t?stage )) : (69)
Here, tstage denotes the unspecied time of separation of the two stages. The terminal time tf is also
unspecied. The functions I and II describe the structural mass consisting of the engines and the fuel
tank in dependence of the fuel used for the two stages. For details, see Bulirsch and Chudej (1992a) and
Chudej (to appear).
The equations of motion in a ight path oriented coordinate system over a spherical Earth with no
wind in the atmosphere are
V_ = 1 [T (V; h; b) cos ? D(V; h; u)] ? g(h) sin
m (70a)
+ !2 (R + h) cos (sin cos ? cos sin sin ) ;
1 [T (V; h; b) sin + L(V; h; u)] cos
_ = mV (70b)
g(h) V
? V ? R + h cos + 2 ! cos cos
+ !2 R V+ h cos (sin sin sin + cos cos ) ;
_ = mV 1cos [T (V; h; b) sin + L(V; h; u)] sin (70c)
? R V+ h cos cos tan

+ 2 ! (sin cos tan ? sin )
? !2 VR cos
+ h cos sin cos ;

h_ = V sin ; (70d)
_ = R V+ h cos sin ; (70e)
_ = (R + hV) cos cos cos ; (70f)
m_ = ?b : (70g)
The state variables are the velocity V , the path inclination , the azimuth inclination , the altitude h,
the geographical latitude , the geographical longitude , and the mass m. The control variables are the
angle of attack u, the lateral inclination angle , the mass ow b, and the thrust angle . The formulae
for the thrust T , the drag D, the lift L, and the gravitational acceleration g(h) can be found in Bulirsch
and Chudej (1992a). All other quantities are assumed to be constant.
The mass ow for both stages is subject to the constraints
0 b bI; max for 0 t tstage ; (71a)
0 b bII; max for tstage < t tf : (71b)
For the boundary conditions, see Bulirsch and Chudej (1992a) and Chudej (to appear).
We concentrate here on the necessary conditions which are brought into the game by the stage sep-
aration condition and the functional including a discontinuity of the mass. According to Eqs. (67), the
stage separation time tstage is determined by the continuity condition of the Hamiltonian; see Eq. (67a).
In addition, the adjoint variable m has a jump discontinuity at t = tstage ,
0 m (t?stage )
m(tstage ) = ? II (m(tstage ) ? m(tf )) +
+ + : (72)
1 + 0I (m0 ? m(t?stage ))
In constrast to the jump conditions introduced by state constraints, this jump condition describes a jump
of xed size which must be carried through at t = tstage .
Due to the constraint describing the stage separation, the dimension of the boundary-value problem is
increased only by 1 to determine the unknowns stage separation time. The constraint (69) yields a jump
condition for the adjoint variable m which must be taken into account together with the jump condition
for the mass given by that Eq. (69). Both jump conditions do not contain any additional parameters to
be determined as part of the solution of the problem as in Section 4.
7. O-line Computation of Optimal Trajectories by Multiple Shooting

7.1. Description of the Method
As we have seen in the preceding sections, the necessary conditions of optimal control theory lead to a
multipoint boundary-value problem of the following form
8 F (t; z(t)) ; 0 t ,
>
< 0 . 1
z_ (t) = F (t; z(t)) = > . (73a)
: Fs(t; z. (t)) ; s t tf ,
z(k+ ) = k (k ; z(k? )) ; 1 k s ; (73b)

ri (z(0); z(tf )) = 0 ; 1 i n1 ; (73c)
ri (ki ; z(k?i )) = 0 ; n1 + 1 i N : (73d)
The piecewise dened dierential equations (73a) include the equations of motion, the system of dier-
ential equations for the adjoint variables, and some so-called trivial dierential equations of form z_j = 0,
18 H. J. PESCH
e.g., for jump parameters entering the jump conditions (73b). These jump conditions are brought into the
problem by possible discontinuities of the adjoint variables when taking into account state variable in-
equality constraints. Equations (73c) and (73d) contain the prescribed boundary conditions for the state
variables, the natural boundary conditions from the transversality conditions for the adjoint variables
and for the terminal time if unspecied, and interior conditions, e.g., junction conditions for boundary or
singular subarcs. All these conditions together determine not only the state and adjoint variables but also
the switching times k and the additional auxiliary variables, namely, the jump parameters mentioned
above. The index ki in the interior conditions serves as an indicator as to whether switching conditions are
associated with a single switching point or not. The piecewise dened right-hand side of (73a) is based on
an assumption of an optimal switching structure associated with a sequence of switching points and their
corresponding control laws between every pair of adjacent switching or boundary points, respectively.
In addition, there are sign conditions which the components zj have to satisfy. These conditions must
be checked after the solution of the boundary-value problem is computed. Non-optimal candidates can
therewith be singled out.
A large class of optimal control problems t into this pattern when analyzing the necessary conditions
of optimal control theory. This kind of multipoint boundary-value problem is especially well suited for
the well-known multiple shooting method (see, e.g., Bulirsch (1971) and Stoer and Bulirsch (1980)). A
new version was developed by Oberle (1982) and the most recent FORTRAN code, called BNDSCO, is
published in the user manual of Oberle and Grimm (1989).
In the following, we give a brief survey of the multiple shooting method with special emphasis on
applications to multipoint boundary-value problems with jump conditions.
The multiple shooting method requires a xed subdivision of the time interval which has to be chosen
by the user,
0 =: t1 < t2 < : : : < tm := tf ; (74)
with tj 6= k , j = 1; : : : ; m and k = 1; : : : ; s. Initial data for the variables zi at the times tj have to
be guessed as well as >
the switching points k . Let the initial guess be Zj(0) for the vectors z (tj ) and (0)
for := (1 ; : : : ; s ) . These data will be changed iteratively. The superscript for the iteration counter will
be dropped in the following. Incidentally, the choice of the partition of the interval is rather uncritical. If
possible, the grid should be ner in regions of stronger changes of the variables, whereas it can be coarser
in other parts.
The basic idea of multiple shooting is to reduce the boundary-value problem to a series of initial-value
problems: For j = 1; : : : ; m ? 1, nd the numerical solution of the initial-value problems

z_ (t) = F (t;0z(t)) ; tj t tj+1 ; with z(tj ) = Zj := Zj ; (75)
where z(t) := (z (t); )> . During the numerical integration, the jump conditions (73b) have to be carried
out at the switching points k 2 [ tj ; tj +1 ]. Moreover, the integration must be stopped at the switching
points, even if no jumps have to be performed. In general, higher derivatives of some variables have
discontinuities here, which may reduce the order of convergence of the integration method. Note that the
right-hand side F changes ? with k. >
Let z(t; tj ; Zj ) = z (t; tj ; Zj ); denote the solution of the initial-value problem (75) in the inter-
val [ tj ; tj +1 ]. Then, a trajectory z (t) and the associated switching points k are a> solution of the above
multipoint boundary-value problem if and only if the vector Z := (Z1 ; : : : ; Zm?1 ) is a zero of
F (Z ) = 0 : (76)
Here, the components of F include the continuity or matching conditions
Fj (Z1 ; : : : ; Zm?1 ) := z(tj+1; tj ; Zj ) ? Zj+1 ; 1 j m ? 2 ; (77a)
and the boundary and switching conditions
R(Z ; Z )
Fm?1 (Z1 ; : : : ; Zm?1 ) := Q(Z1 ; :1 : : ;mZ?m1?1 ) ;
(77b)
with

R(Z1 ; Zm?1 ) := ri (Z1 ; z(tm ; tm?1 ; Zm?1 )
;
i=1;:::;n1
h i
Q(Z1 ; : : : ; Zm?1 ) := ri (ki ; z(k?i ; ti ; Zi )) i=n1 ;:::;N
:
The index i is dened by ti < ki < ti +1 .
The aforementioned code BNDSCO is basically an implementation of the modied Newton method
to determine the zero of F :
Z (i+1) = Z (i) ? (i) DF (Z (i) )?1 F (Z (i) ) ; i = 0; 1; : : : : (78)
Here (i) denotes the relaxation parameter of the i-th iteration step. This modied Newton method is
characterized by the following features: The Jacobian matrix DF is approximated either via numerical
dierentiation or via an appropriate Broyden update. For details, see Stoer and Bulirsch (1980), pp. 266{
270. The Broyden update is applied only during the iteration phase in which the iteration enters the
nal phase of quadratic convergence. Moreover, the Newton method is based on a relaxation strategy
according to Deu hard (1975) using a sophisticated test function check to increase the domain of con-
vergence. In addition, a so-called rank strategy is incorporated to handle ill-conditioned problems; see
Deu hard (1974). These techniques yield a robust algorithm benetting from the advantage of Newton's
method, its quadratic convergence. The disadvantage is the rather small domain of convergence which,
however, is assuaged by the use of homotopy techniques as explained in the following paragraph. In each
Newton iteration, the solution of the system of linear equations is via Householder transformations taking
into account the sparse structure of the coecient matrix. Details can be found in Section 10.1, too. The
magnitude of the norm of that matrix indicates the sensitivity of the solution at tm with respect to the
initial values at t1 . The norm and condition number of the coecient matrix can be estimated from
the Householder decomposition matrices. This is a very useful information assessing the course of the
iteration process. By the way, the elimination of unknowns in the boundary-value problem, if possible,
may reduce the condition number of the boundary-value problem considerably.
7.2. Homotopy Techniques
For the application of the multiple shooting method, the switching structure has to be guessed and
initial guesses for all variables have to be provided. This seems to be a strong restriction when applying
the method. However, the construction of a so-called homotopy chain is a powerful tool to overcome
these obstacles. By a homotopy chain, we understand the construction of a family of problems, naturally
related to the problem to be solved, so that at least one member of that family can be treated easily.
Let that family be associated with a real parameter, say !. So we have to solve, according to Eq. (76), a
one-parameter family of systems of nonlinear equations
F (Z ; !) = 0 ; ! initial ! ! nal :
The solution of a \simpler" problem, say for ! = !i, can then be used as initial guess for the \next"
problem, say for ! = !i+1 . Here, we make an ad hoc choice of the homotopy stepsize ! := !i+1 ? !i .
The starting point for this homotopy chain is ! = ! initial and it terminates with ! = ! nal . It is very
important that ! is a natural parameter of the problem. Otherwise, this procedure might fail. It must
be mentioned that the family of problems generally includes various classes of subproblems, with each
class consisting of a one-parameter family of subsubproblems in such a way that the \last" problem of
the class j is identical to the \rst" problem of the class j + 1. This means that we may have to switch
from one homotopy parameter to another when changing the class. We call this a homotopy strategy.
To illustrate this procedure, we take again the windshear problem from Section 5.2 and show how the
third-order state constraint (57b) is introduced step by step. Because the functional
Ztf
J [u] := (hR ? h(t))2r dt ; (79)
0
20 H. J. PESCH
approximates the Chebyshev functional (55) for r ! 1; see, e.g., Bulirsch et. al. (1991a), we rst solve
the approximating optimal control problem using that functional and choosing r = 3. Then, we combine
the two functionals by
Ztf
=[u] := (1 ? !) (hR ? h(t))6 dt + ! (tf ) (80)
0
and solve a whole chain of optimal control problems by varying the homotopy parameter ! from ! = 0
to ! = 1. By this procedure, the state constraint (57b) is introduced step by step into the problem.
A solution of each problem in the chain of solutions serves as initial guess for the solution of the next
problem which is to be solved until ! = 1 is reached. For details see Bulirsch et. al. (1991b). Note that
the approach via the functional (80) is successful here since the Chebyshev functional (55) and the Bolza
functional (79) are related in a natural way. The coupling of two problems which are not related to each
other in a similar way as by Eq. (80) will generally fail to succeed.
Actually, this is the procedure to treat all aforementioned problems. First, the unconstrained problem
is solved unless the control problem requires the consideration of the inequality constraints to be well-
dened. Hence, for the abort landing problem the control variable inequality constraint (54a) must be
taken into account from the very beginning of the homotopy. In the subsequent steps of the procedure,
the inequality constraints are tightened. Graphical output devices help to detect, for example, whether a
touch point splits into two touch points or into a boundary arc. This is the common behavior for second-
order state constraints; see the numerical results for the altitude-constrained Apollo re-entry problem in
Section 8.3. The appearance of a singular subarc can be detected when the switching function tends to
oscillate around zero during the course of a homotopy run; see, e.g., also Bulirsch et. al. (1991b) and
Section 8.4.
A useful and easy-to-apply homotopy technique is described in Deu hard et. al. (1976) and the recipe is
given here for the example of the Space-Shuttle re-entry. The limit skin temperature can be in uenced in
a natural way by the parameter CLH ; see Eq. (25). If we introduce the homotopy parameter ! := CLH
as additional unknown into the boundary-value problem with the dierential equation
!_ = 0 ; (81a)
the boundary condition
!(tf ) = !i+1 := CLH new ; (81b)
and the initial guess
!(0) (t) !i := CLHold ; (81c)
the convergence of the modied Newton method (78) with an initial relaxation factor (0) < 1 can be
accelerated considerably. The matching conditions are obviously fullled then. The only defect occurs
in the two-point boundary condition for !. However, this approach may not work in all cases. It can
be shown that, when using this approach, the rst iterate of the Newton method is tangential to the
homotopy path, and the relaxation strategy of the modied Newton method provides a homotopy stepsize
control; see Deu hard et. al. (1976) for details.
7.3. Combination of Direct and Indirect Methods
Nevertheless, the construction of an appropriate initial trajectory even for the unconstrained problems
remains often a dicult and time-consuming problem. Many numerical experiments are sometimes needed
for the solution of dierent test initial-value problems since, in general, no information precise enough
is available about the adjoint variables. Recently, a method was developed which combines a direct
collocation method with the multiple shooting method; see von Stryk and Bulirsch (1992). Based on the
approximate solution of the optimal control problem provided by the direct collocation method, initial
estimates can be also obtained for the Lagrange multipliers from the adjoint variables of the Lagrangian
of the nonlinear programming problem which is obtained via parameterization of the optimal control
problem and by using collocation techniques. First numerical experiences show that the approximation
yielded by the direct collocation method is accurate enough to obtain convergence with the multiple
shooting iteration, if the problem is still simple enough. This means the problem must not include
inequality constraints; see Bulirsch et. al. (to appear). So, homotopy techniques still must be used.
By this fusion of direct and indirect methods, one can benet from the superior accuracy provided
by the multiple shooting method, and the disadvantage caused by the small domain of convergence of
the multiple shooting method is considerably diminished. Moreover, direct collocation is a very ecient
method because no integration of dierential equations is to be carried out; the right-hand sides of the
dierential equations are to be fullled pointwise only. On the other side, the disadvantage of direct
methods, namely that numerical solutions may be obtained that are not optimal, do not play a role if
checked by an indirect method afterward.
8. Numerical Results I
Numerical results for all aforementioned problems can be found in the literature. Therefore, the results
are not repeated here in detail. Instead, some of the diculties occuring during the process of solving the
dierent optimal control problems are discussed. In addition, a few components of the optimal solutions
are given here to have the major results at hand and to know how the optimal solutions look like.
8.1. Minimum Heating Re-Entry of an Apollo Capsula Under a Constraint of the Angle
of Attack
The optimal solutions of the re-entry problem with the control variable inequality constraint presented
in Section 2 show eleven dierent classes of switching structures depending on the tightness umax of
the the angle-of-attack constraint. Here, only two aspects shall be pointed out. The rst diculty that
arises after the construction of a starting trajectory for the control-unconstrained problem is described
by the question how the switching structure will look like if the constraint is tighened slightly only.
Starting from the optimal control history for umax = 180 deg, see Fig. 1, it seems to be manifest to expect
one boundary subarc to appear at the end of the ight interval. Indeed, Fig. 2 represents the optimal
solution due to umax = 160 deg obtained via 3 homotopy steps. The dashed line indicate the competitive
nonactive unconstrained control denoted by ufree and determined by Eqs. (11). Note that must have a
zero at t = tf ; see Eq. (13a). In the course of the next homotopy steps, it turns out that becomes zero
in the interior of the constained subarc, too, while simultaneously V is negative. Thus, a corner appears
in the optimal solution. Figure 3 shows the control history for umax = 118 deg. The homotopy stepsize is
umax = 8 deg when using the modied technique of Deu hard et. al. (1976); compare Eqs. (81).
A further tightening of the constraint leads to the appearance of an additional subarc on the lower
bound of the constraint near t = 150 sec, which then merges in the lower bound subarc ending at the
corner; see Fig. 7 in Pesch (1989b). Next, an upper bound subarc occurs at the beginning of the ight
time interval. This situation is shown in Fig. 4 where umax = 62 deg. The homotopy stepsize decreases
to umax = 6 deg.
During the next homotopy steps, the order of the zeros of V and changes: For umax = 62 deg,
has two zeros, say at t = t1 and t = t5 , with < 0 for t1 < t < t5 , and V has three zeros, say
at t = t2 , t = t3 , and t = t4 , with V < 0 for t2 < t < t3 and t4 < t. All zeros are numbered with respect
to their order. Thus, t1 indicates u = 0, t2 , t3 , and t4 indicate u = ?=2, and t5 is just the corner point;
compare Fig. 4. During the subsequent homotopy steps, the zeros t2 and t3 of V disappear, i.e., V > 0
for t < t4 , and the zero t4 of V moves beyond the zero t5 of . Hence, the discontinuous behavior
of the optimal control disappears all of a sudden at a value umax = umax where the problem becomes
singular, i.e., the denominator in the Eqs. (11) becomes zero. The control history for umax = 23 deg, as
a representative of the new class with a continuous optimal control, is given in Fig. 5.
Further results for umax < 23 deg are given in Pesch (1989b) and Pesch (1990a). For this range of
22 H. J. PESCH
Fig. 1. Control history for the Apollo re-entry; u max = 180 deg.
values, the homotopy stepsize decreases from umax = 2 deg to umax = 0:1 deg. The homotopy ends
for umax 15:3 deg having the switching structure u = ?umax , u = ufree , u = umax with the two switching
points very closely side by side. The control variable inequality constraint reduces the maximal altitude
gained by the re-ascent after the rst dip into the atmosphere; see Figs. 9{11 in Pesch (1989b). Compare
also Figs. 7 and 8.
By the way, the construction of a starting trajectory for a related re-entry problem where no inequality
constraint is taken into account is described in Stoer and Bulirsch (1980). The solution of this problem
provides the starting point for a homotopy towards the control-unconstrained problem of Section 2.2.
8.2. Maximum Crossrange Re-Entry of a Space-Shuttle Orbiter
Under a Constraint of the Skin Temperature
For the numerical results of the space-shuttle re-entry problem of Section 3, it is mainly refered to
Dickmanns and Pesch (1975) and Deu hard et. al. (1976). The most recent results are given in Kugelmann
and Pesch (1990b). The changes of the switching structure can be easily obtained when the control-
unconstrained problem is solved rst. For this case, the control history is given in Fig. 1 of Kugelmann
and Pesch (1990b). By tightening the constraint, i.e., decreasing the value of CLH , up to four constrained
subarcs appear. During the further course of the homotopy, ever the last two subarcs merge together until
one long constrained subarc is nally left over. Compare Fig. 3 of Kugelmann and Pesch (1990b). The
homotopy was stopped at T 924 C , where V becomes zero on the constrained subarc, i.e., the free
control CLfree is not dened beyond that point; see Eq. (27). For the results, see Fig. 4 of Kugelmann
and Pesch (1990b). The heating constraint damps the oscillatory behavior of the trajectory; see Fig. 2
of Kugelmann and Pesch (1990b). Note that the problem is ill-conditioned with a condition number
of about 1020 , if the so-called condensed multiple shooting version is used where the system of linear
equations to be solved in each iteration step of the modied Newton method (78) is reduced to a smaller
system with a coecient matrix as given by the Eqs. (101). The condition number is of magnitude 108
only, if the large system with a coecient matrix equivalent to (100) is solved. Notice that the application
of the modied homotopy technique (81) accelerates the homotopy run considerably; see Deu hard et.
al. (1976). Figure 6 gives an impression of the re-entry trajectory for the control-unconstrained case. The
heating constraint smoothes the trajectory. Oscillations only occur after having passed the boundary arc;
see Fig. 2 of Kugelmann and Pesch (1990b).
24 H. J. PESCH
Fig. 6. Re-entry trajectory for the Space-Shuttle re-entry; control-unconstrained case.
8.3. Minimum Heating Re-Entry of an Apollo Capsula Under a Constraint of the Alti-
tude
The introduction of the second-order state constraint has the same eect as the angle of attack constraint
as described in Section 8.1 except that the nal time does not decrease as much. Figures 7 and 8 show
the histories of the altitude for hmax 50:97 km and hmax 42:05 km. The rst trajectory has one touch
point, the second one boundary arc and one touch point. Between these two values of hmax , there are
optimal solutions with two touch points; the rst touch point splits to a boundary arc when the constraint
is intensied. The changes from one switching structure to another can be detected with the help of
graphical output. Pay attention whether the component for the altitude as obtained from the solution of
the boundary-value problem satises the constraint (40) for all t after the rst dip into the atmosphere.
The sign conditions (46) must be obeyed, too, and may help to single out nonoptimal solutions. Note that
the deviation from the boundary hmax between the boundary subarc and the touch point is approximately
30 m only. This problem was solved completely for the rst time by Hiltmann (1983). The control histories
of this problem can be found in Figs. 12{14 of Pesch (1989b).
Fig. 7. Altitude history for the Apollo re-entry; h max 50 97 km.

:
Fig. 8. Altitude history for the Apollo re-entry; h max 42 05 km.

:
8.4. Abort Landing of a Passenger Aircraft Under Windshear Conditions

Because the numerical results and the way of computing the solution are described in great detail in
Bulirsch et. al. (1991b), we concentrate here on the treatment of singular subarcs only. Figures 9 and 10
show the switching function , see Eq. (59), for two distinct sets of boundary conditions at the very
beginning of the homotopy for the state-unconstrained problem with the Bolza type functional (79)
and r = 3. Figure 10 shows an optimal control history with a singular subarc. The transition from the
completely nonsingular case as described by Fig. 9 to the problem of Fig. 10 with a singular subarc needs
to determine which type of switching structure will be optimal. In this case, a solution with two additional
bang-bang switching points, which also seems to be possible, could not be obtained. The multiple shooting
26 H. J. PESCH
method produces solutions contradicting Eq. (59) if the boundary-value problem is formulated for plane
bang-bang structure.
Fig. 9. History of the switching function with plane bang-bang switching structure for the abort landing problem.
Fig. 10. History of the switching function with one singular subarc for the abort landing problem.
For the sake of completeness, Fig.11 shows the two trajectories in the vertical plane for the performance
indices (55) (solid line) and (79) (dashed line) and the windprole. See Pesch (to appear) for the detailed
switching structure due to the third-order state constraint which is induced by the transformation (56)
of the Chebyshev functional. Further results for dierent windshear intensities up to about 200 ft/sec for
the dierence between maximum tailwind and maximum headwind and for windshear proles that also
include upwind zones can be found in Berkmann and Pesch (to appear).
By the way, Bulirsch et. al. (1991a, 1991b) may serve as a user's guide for solving sophisticated optimal
control problems by multiple shooting. In particular, this windshear problem shows a lot of the features
that make this optimal control problem a tough one to solve. The papers provide a good illustration of how
multiple shooting and homotopy techniques work in connection with optimal control problems involving
multiple subarcs. Techniques are presented to detect, besides bang-bang subarcs, singular subarcs. In
addition, the treatment of state constraints is explained in detail with emphasis on detection techniques for
touch points and boundary subarcs. In particular, the modications of the formulation of the multipoint
boundary conditions are explained when changes of the switching structure appear during the homotopy
runs. For the windshear problem, more than 15 such changes of the switching structure occurred, making
that problem one of the most dicult ones ever solved by a numerical method. The complexity of optimal
control problems can also be seen which at present can be solved by the multiple shooting method.
Fig. 11. Abort landing of a passenger aircraft in a windshear; comparison between Bolza and Chebyshev functional.
8.5. Maximum Payload Ascent of a Two-Stage-To-Orbit Space Transporter System

The family of problems for the homotopy strategy to solve this problem consists of three major classes
of subproblems; compare Section 7.2. First, solve the ascent problem in the plane of a great circle. At
the time of stage separation the fuel of the rst stage is assumed to be totally consumed; see Eqs. (71).
Therefore, the following switching structure seems to be reasonable : b = bI; max for 0 t tstage and
b = bII; max , b = 0, and b = bII; max for tstage < t tf . Second, make the transition to a three-dimensional
coordinate system for a ight over a spherical Earth. Third, include Coriolis and centrifugal forces to
obtain a model for a spherical rotating Earth.
For the rst subproblem, convergence diculties with the multiple shooting method arise because of the
non-dierentiability of the Hamiltonian at t = tf if the boundary condition (5d) is used; see Oberle (1976,
1977) for a similar problem. This obstacle can be circumvent if the boundary condition (5d) is replaced
by the equivalent condition H (0) = 0. Note that the problem is autonomous. Thus, H const from
which H 0 follows because of Eqs. (5d) and (67a).
For the numerical results of a more realistic model, see Bulirsch and Chudej (1992a) and Chudej (to
appear). The trajectory of this upgraded model is taken from these references and given in Fig. 12.
9. Neighboring Extremals
If optimal solutions of processes which run down very fast, such as the optimal ight of space vehicles or
aircrafts, are to be realized practically, one needs fast numerical methods to compensate for disturbances
occurring during the course of this process. This guarantees that optimality conditions and prescribed
constraints are preserved. The required minimal computing time for the computation of an adjusted
optimal control program cannot be met by the multiple shooting method if computers with convential
architectures are used. Despite the inherent parallel structure of the multiple shooting algorithm, it is
today not yet thoroughly investigated whether computing times can be achieved by a parallel multiple
shooting algorithm which are fast enough for on-line applications in the aerospace eld. In this survey,
we therefore go a dierent way.
In the following, we describe two numerical methods which are based, with respect to their theoretical
28 H. J. PESCH
Fig. 12. Launch and stage separation of a Sanger-type space vehicle.
part, on the theory of neighboring extremals and, with respect to their numerical part, on the multiple
shooting method. A linearization of the necessary conditions of the disturbed optimal control problem
around the optimal trajectory of the undisturbed problem leads to a linear multipoint boundary-value
problem for the perturbations of the state and adjoint variables. This linear boundary-value problem can
be solved very eciently if appropriate information about the reference trajectory is pre-computed and
stored in the onboard computer. Both methods allow the real-time computation of neighboring optimum
feedback controls for control problems of a rather general class such as described in the foregoing sections.
The two methods are described in detail in Pesch (1989a, 1989b) and in Kugelmann and Pesch (1990a,
1990b), respectively. So, only the basic ideas are given here.
Optimal control problems are investigated which depend on a vector perturbation parameter. For the
sake of simplicity, the considerations are restricted to problems with a perturbation parameter p entering
the initial conditions
x(0) = (p) (82)
and/or the boundary conditions
(x(tf ); tf ; p) = 0 :
By including additional terms, the following linearization technique can be generalized for problems with
perturbations in all other functions which are involved in the description of the underlying model. The
disturbances will give rise to optimal solutions
x(t; p) ; u(t; p) ; tf (p) ; (t; p) ; : : : (83)
of the perturbed optimal control problem which can be shown to exist in a neighborhood of the optimal
solution of the undisturbed problem, if certain regularity assumptions are satised. Moreover, these
optimal solutions are continuously dierentiable with respect to the perturbation parameter p near p = 0.
For the class of problems, where this can be proven, see Maurer and Pesch (to appear). Because of the
continuous dierentiability with respect to p, we may dene the so-called variations
x(t) := @x
@p (t; 0) p ; (t) := @ (t; 0) p ;
@p (84a)
u(t) := @u
@p (t; 0) p ; (t) := @ (t; 0) p ;
@p (84b)
and the dierentials
dtf := ddtpf (0) p ; d := ddp (0) p ; (84c)
dx(tf0 ) := x(tf0 ) + x_ 0 (tf0 ) dtf ; d(tf0 ) := (tf0 ) + _ 0 (tf0 ) dtf (84d)
where x0 = x(t; 0), u0 = u(t; 0), 0 = (t; 0) and tf0 = tf (0) denote an optimal solution of the undisturbed
problem. These variations have now to be computed to obtain a rst-order optimal solution of the
disturbed problem, e.g., by
x(t; p) =: x0 + x(t) ; (85a)
u(t; p) =: u0 + u(t) ; (85b)
:
tf (p) = tf0 + dtf : (85c)
Therefore, the result obtained in this way is called a neighboring extremal. For additional references, see
Pesch (1989a).
For the sake of simplicity and brevity, we consider optimal control problems with one mixed constraint
(l = 1), where Cu 6= 0, and a scalar control variable (k = 1). The boundary conditions are assumed to be
disturbed,
x0 = x(t0 ) =: x(t0 ; p) ? x0 (t0 ) for t0 2 [0; tf [ ; (86a)
d = (x(tf (p); p); tf (p)) : (86b)
The perturbation vector p is assumed to be given at the time of measurement t0 ,
x
p := d 0 2 IR
n+q : (87)
A linearization of the Eqs. (2a), (5a), (9a), (2b), (2c), (5c), and (5d) with the Hamiltonian (18) around
the optimal solution of the undisturbed problem gives
x_ = fx x + fu u ; (88a)
_ = ?Hxx x ? Hxu u ? fx> ? Cx> ; (88b)
0 = Hux x + Huu u + fu> + Cu ; (88c)
x(t0 ) = x0 ; (88d)
d = ( x dx + t dtf ) j tf0 ; (88e)
> >
d(tf ) = (xx dx + x d + xt dtf ) j tf0 ; (88f)

0 = (xt ? _ >0 ) dx + x_ >0 d + tt dtf + t> d j tf0 : (88g)
All vector- or matrix-valued functions are to be evaluated along the optimal solution of the nominal
problem, e.g.,
@ H > = H (x ; u ; ; ) :
Hxu := @u x xu 0 0 0 0
In the last equation, the vanishing terms Hu du (because of Eq. (9a)) and C d (either C = 0 or d = 0
because of Eq. (19)) are omitted.
30 H. J. PESCH
For the further investigation, we assume that the Hamiltonian is regular and that the strong Legendre-
Clebsch condition (9b) is valid on unconstrained subarcs. As in the previous sections, we have to dis-
tinguish between unconstrained and constrained subarcs. In addition, we must take into account that
the control laws being active at the time of correction might be dierent for the nominal and the actual
trajectory, particularly in the neighborhood of the nominal switching points. Here, we consider the two
simpler cases only where the same behavior of the control variable is present along the undisturbed and
disturbed trajectory. For the other cases, see Pesch (1989a).
1st Case: C (x(t; p); u(t; p)) < 0 and C (x0 (t); u0 (t)) < 0 .
Since (t; p) = 0 and also 0 (t) = 0 hold, it follows (t) = 0. From Eq. (88c) and by applying the
Implicit Function Theorem to Eq. (9a), one obtains
u = ?Huu ?1 ( Hux x + f > ) = ux x + u : (89)
u
Substituting Eq. (89) into Eqs. (88a) and (88b) yields a homogeneous system of dierential equations for
the variations x and ,
x_ A(t) B (t) x
_ = B~ (t) ?A>(t) (90)
where
?1 Hux = fx + fuux = df ;
A(t) = fx ? fuHuu dx
?1 f > = fuu = df ;
B (t) = ?fuHuu u d
?1 Hux = ?Hxx ? Hxuux = dg ;
B~ (t) = ?Hxx + HxuHuu dx
dg df >
B (t) = B >(t) ; B(t) = B (t) ; d = ? dx = ?A>(t) ;
~ ~ >
with
g(x; ) := ?Hx>(x; u(x; ); ; (x; )) :
2nd Case: C (x(t; p); u(t; p)) = 0 and C (x0 (t); u0 (t)) = 0 .
Under the assumption Cu 6= 0, we obtain by linearizing C (x(t; p); u(t; p)) = 0,
u = ?Cu?1Cx x = ux x ; (91)
and Eq. (88c) yields
= ?Cu?1 ( Hux x + Huu u + fu> ) : (92a)
Applying the Implicit Function Theorem on Eq. (9a) with (18), this equation can be written as
= x x + u u + (92b)
where
x = ?Cu?1 Hux ; u = ?Cu?1Huu ; = ?Cu?1fu> :
Substituting Eqs. (91) and (92) into (88a) and (88b) leads to
x_ A(t) 0
x
_ = B~ (t) ?A>(t) (93)
where
df ;
A(t) = fx ? fu Cu?1Cx = fx + fuux = dx
B~ (t) = ?Hxx + HxuCu?1 Cx + Cx>Cu?1 ( Hux ? HuuCu?1 Cx ) ;
dg = g + g u + g ( + u ) :
= dx x u x x u x
In both of these cases, the arguments of the coecient functions are just the nominal extremal x0 , u0 ,
0 and 0 .
Note that there may also exist an explicit nonlinear feedback law on constrained subarcs; see Eq. (21).
The gaps between nominal and actual switching points are covered by the two cases C (x(t; p); u(t; p)) <
0 and C (x0 (t); u0 (t)) = 0, and C (x(t; p); u(t; p)) = 0 and C (x0 (t); u0 (t)) < 0, respectively. For these cases,
the linearization must be done so that the system of linear dierential equations for the variations remains
homogeneous. See Pesch (1989a) for details. This property is of utmost importance for the eciency of
the guidance schemes developed in Section 10.
In a similar way, the interior point constraints and the jump conditions are linearized. In addition, linear
relations are obtained from the switching conditions to approximate the displacements of the switching
points. In summary, all these linearized necessary conditions lead to a linear multipoint boundary-value
problem with homogeneous dierential equations, linear multipoint boundary conditions, and linear jump
conditions, that is especially well-suited for real-time computations. See Pesch (1989a).
10. On-line Computation of Optimal Trajectories

by Neighboring Optimum Feedback Guidance Schemes
10.1. Multiple Shooting for Linear Multipoint Boundary-Value Problems
The linearized necessary conditions of Section 9 lead to a linear multipoint boundary-value problem of
the following form,
y_ (t) = T (t) y(t) for j t < j+1 ; j = 1; : : : ; m ? 1 ; (94a)
mX
?1
A y(1) + Asj y(j? ) + B y(m) ? c(p) = 0 ; (94b)
j =2
y(j+ ) = Rsj y(j? ) for j = 2; : : : ; m ? 1 ; (94c)
0 = 1 < 2 < : : : < m = tf (94d)
N
where T : [0; tf ] ! N;
IR is piecewise continuous, A, Asj , B and Rsj are N by N -matrices, and c is an
N -vector. Here, y := (x> ; > ; d1 ; : : : ; dq~) contains the variations x and of the state vector and
>
the adjoint multiplier, respectively, from which the variation u of the control vector can be computed
by a linear relationship of the form,
u(t) = ux (x0 (t); 0 (t)) x(t) + u (x0 (t); 0 (t)) (t) : (95)
Note that there holds u = 0 on state-constrained subarcs. The variations dk are associated with
the multipliers for the interior point conditions. These variations as well as the dierentials dj of the
nominal switching points j , j = 2; : : : ; m ? 1, are needed to approximate the optimal solution of the
perturbed optimal control problem to the rst order. The actual switching points are obtained analogously
to Eqs. (85) by additional linear relations for the displacements dj ,
x( )
_ ?
dj = ?W (Wx; W ) (j )
1 (96)
j
32 H. J. PESCH
where W (x(j ); (j )) = 0 denotes the switching condition associated with the switching point j . The
subscript 0 for the characterization of the nominal switching points is omitted here. See Pesch (1989a,
1989b) for details. Finally, it should be mentioned that the vector perturbation parameter p enters only
the vector c in the above linear boundary-value problem (94). This also is of utmost importance for the
ecency of the guidance methods developed in the following subsections.
The solution of this boundary-value problem can be reduced to the solution of a series of initial-value
problems as in the multiple shooting method of Section 7. We start again on the basis of a subdivision of
the interval [0; tf ], which can be assumed to coincide with the partition (94d) without loss of generality.
Take simply Rsj = I and Asj = 0 to include additional multiple shooting nodes. The following initial-value
problems are now to be solved,
y_ = Ty for j t < j+1 ; j = 1; : : : ; m ? 1 ; (97a)
y(j ; sj ) = sj with sj 2 IR
N : (97b)
The unique solutions
y(t; sj ) = Y (t; j ) sj for j t < j+1 (98)
can be given in terms of the transition matrices Y (t; j ), j = 1; : : : ; m ? 1, dened by the matrix initial-
value problems
@ Y (t; ) = T Y (t; ) ; (99a)
@t j j
Y (j ; j ) = IN : (99b)
Here, IN denotes the identity matrix of dimension N . The unknown vectors sj , denoting the right-hand-
side values of the variations at discontinuities, are to be determined so that the solutions (98) satisfy the
multipoint boundary conditions (94b) and the jump conditions (94c). This leads to the following system
of N (m ? 1) linear equations,
0 Rs2 Y1 ?I O O ::: O 1 0 s1 1 0 0 1
B
B O Rs3 Y2 ?I O ::: O CC BB s2 CC BB 0 CC
B
B O O Rs4 Y3 ?I O C BB s3 CC BB 0 CC
B
B .. ... ... .. CC BB ... CC = BB ... CC (100)
B . . CC BB C B C
B
B O O C CA B@ sm?3 CCA BB@ 0 CCA
@ O O ::: O Rsm?1 Ym?2 ?I sm?2 0
A + As2 Y1 As3 Y2 ::: : : : Asm?1 Ym?2 B Ym?1 sm?1 c
where
Yj := Y (j+1 ; j ) :
By an appropriate elimination, the number of equations in (100) can be reduced to N ,
E s1 = c (101a)
where
X
m jY
?1
E= Asj Yj?i Rji (101b)
j =1 i=1
and
As1 := A ; Asm := B ; Rs1 := I : (101c)
Moreover, we have, for j = 2; : : : ; m, the recursion
sj = Rsj Yj?1 sj?1 ; (102a)
Rsm := I : (102b)
Because of the special block structure of the matrices T and Asi of the boundary-value problem (94),
the coecient matrix E has a special block structure, too,

E = EI EO :
1 2
For details, see Pesch (1989b). Therefore, the number of equations in (101a) can be reduced once more.
According to the partition of the vector y, one obtains, introducing the subpartition
y = (yA>; yB>)> ; yA = x ; yB = (> ; d1 ; : : : ; dq~)> ; (103a)
c = (c>A ; c>B )> ; s1 = (s>1A ; s>1B )> ; (103b)
the following system of linear equations,
E2 s1B = cB ? E1 cA ; s1A = cA : (104)
Generally, this system has considerably fewer equations than (100); but for very sensitive problems, the
elimination process leading to (101) may change the condition number for the worse, since the elimination
is a Gauss algorithm performed blockwise with xed pivoting. For moderately conditioned problems,
however, Eq. (101) is preferable to (100) for on-line computations because of its lower computation and
storage requirements. Compare also the remarks in Section 8.2.
10.2. The Neighboring Optimum Feedback Guidance Scheme
In summary, we obtain from Eq. (104) a rst guidance method where the main computational eort for
computing the neighboring optimal control vector can be carried through before starting the process,
e.g. , before the take-o of a space vehicle. The main part of the computation to be performed before the
process is started is the approximation of the transition matrices Yj .
If we solve Eq. (104) for s1B and substitute the result in the linear formula (95) for u, we obtain a
continuous neighboring optimum feedback law of the form
u(t0 ) = 1 (t0 ) x0 + 2 (t0 ) d (105)
where the so-called gain matrices 1 and 2 can be precomputed. Here t0 denotes the correction time,
x0 the measured deviation from the reference path at this time, and d the change in the terminal
conditions. For more details, see Kugelmann and Pesch (1990a, 1990b). The amount of computation
during the process, namely the matrix-times-vector operations, is negligible. This feedback scheme can
be described by the diagram in Fig. 13.
If we assess the pros and cons of this method, we see that, in spite of the advantageous low onboard
computations, the method shows some disadvantages. For example, the method will fail if measurement
data are absent for a while. Moreover, a precheck of the constraints and a reliable precomputation of
the switching points, before feeding back the adjusted control, are impossible. Even if we would use the
recurrence (102a), the results would be unsatisfactory since the use of transition matrices if integrated
over longer intervals leads to an exponential growth of the linearization error. These disadvantages can
be avoided by the following surer but also costlier method.
10.3. The Repeated Correction Guidance Scheme
Because of the many technical details of the improved method, which is called repeated correction method,
we describe here the idea of the method only. For the details, it is again refered to Pesch (1989a, 1989b).
34 H. J. PESCH
Fig. 13. Chart ow of the neighboring optimum feedback guidance scheme.
The evaluation of the neighboring optimum feedback scheme (105) or, alternatively, the solution of
the linear system (104) is incorporated into the numerical integration of the equations of motion with the
control variables approximated by spline functions for example. A single integration of the equations of
motion then yields an approximation of the actual trajectory for the entire remaining ight time interval.
This approximation also includes the approximation of all switching points. It is obvious that this full
information can, in addition, be used to check the observance of all constraints imposed on the problem.
By this approach, a veried feedback scheme is available which is not so dependent on a continous ow of
the measurement data as the linear feedback scheme of the previous subsection. Note that the repeated
correction, either at many sample points or continuously, reduces the in uence of the linearization error
introduced by the theory of Section 9. This indeed is just the self-correcting property of Newton's method.
By the linearization and the subsequent application of the multiple shooting discretization to the linear
boundary-value problem, we end up with the same system of linear equations that is obtained in the
multiple shooting iteration when applied to the disturbed problem with the undisturbed solution as initial
estimate. This error damping property also holds for the method of Section 10.2 if applied synchronously
with the ow of data. With this repect, both method are equivalent. This leads to controllability regions
of about the same size for both methods. The controllability region of a guidance method describes the
set of all deviations from the reference path which can be successfully compensated during the course of
the process. Because the linearization step and the discretization step can be interpreted as commutative
operators, the controllability regions of both guidance methods are also equivalent to the domain of
convergence of the undamped Newton iteration, i.e., take (i) = 1 for all i in Eq. (78); see Pesch (1990b).
Therefore, the computation of the domain of convergence of the multiple shooting method with the
standard Newton method provides a quick test for controllability; see Kugelmann and Pesch (1991).
One cannot expect the onboard computing time for this more sophisticated method to be negligible
compared with the remaining ight time. Therefore, we rst integrate the equations of motion in real-
time over a small interval [t0 ? t0 ; t0 ], t0 > 0, by using the measured actual state vector at t0 ? t0
as the initial value. The control vector is chosen either as the actual control computed last or, if not
available, as the nominal control. The actual control history due to this precomputed future deviation
from the nominal trajectory at time t0 is then computed. During the onboard computation, the vehicle
is assumed to y to this so precomputed state at t0 so that the actual control can be started in due time
after completion of the computation. Here t0 is selected as an upper bound for the onboard computing
time needed. This feedback scheme can be described by the diagram of Fig. 14.
Fig. 14. Chart ow of the repeated correction guidance scheme.
Finally, it should be mentioned that the parallel structure of the multiple shooting method can still be
preserved, if the linear system (100) is taken for the repeated correction algorithm which, moreover, has
the better numerical stability properties. Then, the numerical integration and the check of the constraints
can be split up into dierent segments and, therefore, can be carried through on dierent processors. The
linearization error, however, cannot be smoothed as well as by the serial version of the repeated correction
method; for the discussion of this eect, see Kugelmann and Pesch (1990c).
11. Numerical Results II

For the optimal control problems considered in the Sections 2{4 and 6, investigations have been made
concerning the controllability of the optimal trajectories. In Pesch (1979, 1989b, 1990a), the Apollo
re-entry problems are discussed. Investigations for the Space-Shuttle re-entry problem can be found
in Pesch (1980) and in Kugelmann and Pesch (1990b) and for the Sanger ascent in Kugelmann and
Pesch (1991). Here, results are presented only for the Space-Shuttle re-entry problem. Figure 15 shows
some trajectories around the reference ight path for simulated deviations from the nominal altitude,
which have been computed by means of the neighboring optimum guidance scheme of Section 10.2.
The cross section through the controllability tube around the nominal trajectory in the direction of the
altitude abscissa is also indicated.
The numerical results for the cross-range maximization problem of a space-shuttle glider under a
reradiative heating constraint, as obtained and discused in Kugelmann and Pesch (1990b), show that,
even for that extremely sensitive problem, the domain of controllability is large enough for practical
applications. The domain of controllability for terminal perturbations is considerably smaller, since it
may be inherent for all feedback schemes that use some information from a reference trajectory.
Together with the more costly repeated correction method, two variants of a multiple shooting
based guidance method are available which are numerically stable and nearly optimal. These can be
applied either to guidance problems requiring extremely fast corrections|the observance of the con-
straints can only be guaranteed to the rst order in this case|or to problems that allow more expensive
computations|in which case all constraints are checked and even a limited absence of measurement data
is not disasterous.
12. Conclusions and Outlock

The trend in the numerical treatment of optimal control problems in aerospace engineering points towards
problems of increasing complexity in order to approximate reality as closely as possible. These problems
lead rather to an increasing number of constraints than to an increasing number of unknowns involved
in the problems. Thus, engineers and mathematicians who are involved in the design and development
36 H. J. PESCH
Fig. 15. Neighboring optimal trajectories around the optimal reference trajectory for the Space-Shuttle re-entry.
of future aerospace enterprises have a strong demand for reliable and ecient software that can handle
optimal control problems with dierent types of constraints. Despite the fact that a direct method, such as
a direct collocation method, may be easier to apply, since only a little knowledge of optimal control theory
is required, the indirect multiple shooting method has advantages with respect to reliability, precision,
and getting insight into the structure of the solution and its optimality. Moreover, the multiple shooting
method is especially appropriate for an implementation on parallel computers; see Kiehl (1989).
Concerning the o-line computation of optimal trajectories of complex control problems, both theory
and software have reached a high standard. The limit of problems which today can be solved by the
multiple shooting method in the eld of aerospace applications can be seen, for example, from the
investigations of ion-driven gravity-assisted missions to asteroids like Flora and Vesta, see Bulirsch and
Callies (1991a, 1991b), and to outer planets like Neptune, see Callies (to appear). The multiple rendezvous
mission to asteroids, for example, is modelled by an optimal control problem subject to several control
and state variable inequality constraints, several interior point constraints and parameter constraints. The
fully optimized trajectory including a Moon swing-by and spiraling down to and up from low asteroid
orbits exhibits more than 50 switching points. For complex missions like this, the outstanding accuracy
provided by the multiple shooting method is no longer an unnecessary by-product of an over-precise or
overdeveloped method, but of vital and decisive importance for the mission planning. The high accuracy
renders possible the computation of the optimal trajectory at all.
For the optimal control of industrial robots, the equations of motion themselves are very complicated
and, in general, established by means of appropriate software. The adjoint dierential equations can
then be obtained via symbolic dierentiation. Minimum-time and minimum-energy trajectories for an
industrial robot of three degrees of freedom are investigated in Pesch et. al. (to appear) and von Stryk
and Schlemmer (to appear). See also Pesch (to appear). The numerical results have been obtained both
by means of a direct collocation and by means of a multiple shooting method. Because of the high
complexity of the adjoint variables (more than 3,000 FORTRAN statements for the right-hand sides of
the dierential equations), the direct collocation method is preferable to any indirect method for this
kind of problems.
Contrary to the o-line computation of optimal trajectories, both the theoretical and the numerical
basis for the on-line computation of optimal trajectories are not as well-developed when optimal control
problems of a rather general class, as covered in this paper, are considered. The theoretical foundation
which is related to second-order sucient optimality conditions, see Maurer and Pesch (to appear), is
not yet well understood if state inequality constraints are taken into account, although their numerical
treatment turns out to be a routine work today. Despite the missing mathematical justication, the
guidance schemes based on the theory of neighboring extremals show their applicability for the real-time
computation of optimal trajectories for control problems which include control and/or state variable in-
equality constraints; see Pesch (1989a, 1989b) and Kugelmann and Pesch (1990a, 1990b). Open questions
are concerned with problems having singular subarcs, and also the numerical realization of a neighboring
optimum guidance scheme which can compensate disturbances of system parameters, e.g., air density
uctuations, is still pending.
A new approach for problems of the latter typ is via dierential game theory where the unknown
air density uctuations are modelled as the controls of an antogonistic player in a two-person zero-sum
dierential game; see Breitner and Pesch (to appear). By this approach, not only the worst case can be
studied but optimal strategies can be computed for density uctuations within limits that are known
from long time measurements; see Breitner (to appear) and Fig. 16.
In the near future, methods designed for the o-line computation of optimal trajectories will be able
to compete because of the rapidly increasing speed of computation due to parallel computers if they are
available for onboard computers on spacecraft, too. Parallelized versions of indirect or direct multiple
shooting seem to be the most promising techniques because of the inherent parallel structure of the
multiple shooting algorithm.
References
Berkmann, P. and Pesch, H. J.: to appear, Àbort Landing under Dierent Windshear Conditions'.
Breakwell, J. V., Speyer, J. L., and Bryson, A. E.: 1963, Òptimization and Control of Nonlinear Systems Using the Second
Variation', SIAM Journal on Control 1, 193{223.
Breitner, M. H.: to appear, `Construction of the Optimal Feedback Controller for Constrained Optimal Control Problems
with Unknown Disturbances', Proceedings of the 9th IFAC Workshop of Control Applications of Optimization, Edited
by R. Bulirsch and D. Kraft, International Series of Numerical Mathematics, Basel: Birkhauser.
Breitner, M. H. and Pesch, H. J.: to appear, `Re-entry Trajectory Optimization Under Atmospheric Uncertainty as a
Dierential Game', Advances in Dynamic Games and Applications, Edited by T. Basar et al. Annals of the ISDG,
Vol. 1, Basel: Birkhauser.
Bryson, A. E. and Ho, Y. C.: 1987, Applied Optimal Control, Washington, D.C.: Hemisphere.
Bulirsch, R.: 1971, Die Mehrzielmethode zur numerischen Losung von nichtlinearen Randwertproblemen und Aufgaben der
optimalen Steuerung, Report of the Carl-Cranz Gesellschaft, Oberpfaenhofen: Carl-Cranz Gesellschaft.
Bulirsch, R. and Callies, R.: 1991a, Òptimal Trajectories for an Ion Driven Spacecraft from Earth to the Planetoid Vesta',
Proceedings of the AIAA Guidance, Navigation and Control Conference, New Orleans, Louisiana, 1991, AIAA Paper
91-2683.
Bulirsch, R. and Callies, R.: 1991b, Òptimal Trajectories for a Multiple Rendezvous Mission to Asteroids', 42nd International
Astronautical Congress, Montreal, Canada, 1991, IAF-Paper IAF-91-342.
Bulirsch, R. and Chudej, K.: 1991, Àscent Optimization of an Airbreathing Space Vehicle', Proceedings of the AIAA
Guidance, Navigation and Control Conference, New Orleans, Louisiana, 1991, AIAA Paper 91-2656.
Bulirsch, R. and Chudej, K.: 1992a, `Staging and Ascent Optimization of a Dual-Stage Space Transporter', Zeitschrift fur
Flugwissenschaften und Weltraumforschung 16, 143{151.
Bulirsch, R. and Chudej, K.: 1992b, `Guidance and Trajectory Optimization under State Constraints', Preprint of the 12th
IFAC Symposium on Automatic Control in Aerospace | Aerospace Control 1992, Edited by D. B. DeBra and E. Gottzein,
Dusseldorf: VDI/VDE-GMA, 553{538.
Bulirsch, R., Montrone, F., and Pesch, H. J.: 1991a, Àbort Landing in the Presence of a Windshear as a Minimax Optimal
Control Problem, Part 1: Necessary Conditions', Journal of Optimization Theory and Applications 70, 1{23.
Bulirsch, R., Montrone, F., and Pesch, H. J.: 1991b, Àbort Landing in the Presence of a Windshear as a Minimax Optimal
Control Problem, Part 2: Multiple Shooting and Homotopy', Journal of Optimization Theory and Applications 70,
223{254.
Bulirsch, R., Nerz, E., Pesch, H. J., and von Stryk, O.: to appear, `Combining Direct and Indirect Methods in Optimal
Control: Range Maximization of a Hang Glider', Optimal Control, Proceedings of the Conference in Optimal Control
and Variational Calculus, Oberwolfach, 1991, Edited by R. Bulirsch et. al., Basel: Birkhauser.
38 H. J. PESCH
Fig. 16. Re-entry of a shuttle under uncertain air density disturbances: Optimal control solution (thin line) and dierential
game solution (thick line) for the control unconstrained case of Section 3.
Callies, R.: to appear, Òptimal Design of a Mission to Neptune', Optimal Control, Proceedings of the Conference in Optimal
Control and Variational Calculus, Oberwolfach, 1991, Edited by R. Bulirsch et. al., Basel: Birkhauser.
Chudej, K.: to appear, Òptimization of the Stage Separation and the Flight of a Future Launch Vehicle', Proceedings
of the 16th IFIP Conference on System Modelling and Optimization, Compiegne, 1993, Lecture Notes in Control and
Information Science, Berlin: Springer.
Deu hard, P.: 1974, À Modied Newton Method for the Solution of Ill-conditioned Systems of Nonlinear Equations with
Application to Multiple Shooting', Numerische Mathematik 22, 289{315.
Deu hard, P.: 1975, À Relaxation Strategy for the Modied Newton Method', Optimization and Optimal Control, Lecture
Notes in Mathematics 477, Edited by R. Bulirsch et. al., Berlin: Springer, 59{73.
Deu hard, P., Pesch, H. J., and Rentrop, P.: 1976, À Modied Continuation Method for the Numerical Solution of Nonlinear
Two-Point Boundary Value Problems by Shooting Techniques', Numerische Mathematik 26, 327{343.
Dickmanns, E. D. and Pesch, H. J.: 1975, Ìn uence of a Reradiative Heating Constraint on Lifting Entry Trajectories for
Maximum Lateral Range', Proceedings of the 11th International Symposium on Space Technology and Science, Tokyo,
Japan, 241{246.
Hiltmann, P.: 1983, Numerische Behandlung optimaler Steuerprozesse mit Zustandsbeschrankungen mittels der Mehrziel-
methode, Diploma Thesis, Department of Mathematics, Munich University of Technology, Munich.
Jacobson, D. H., Lele, M. M., and Speyer, J. L.: 1971, `New Necessary Conditions of Optimality for Control Problems with
State-Variable Inequality Constraints', Journal of Mathematical Analysis and Application 35, 255{284.
Kelley, H. J., Kopp, R. E., and Moyer, H. G.: 1967, `Singular Extremals', Topics in Optimization, Edited by G. Leitmann,
New York: Academic Press, 63{101.
Kiehl, M: 1989, `Vectorizing the Multiple-Shooting Method for the Solution of Boundary-Value Problems and Optimal
Control Problems', Proceedings of the 2nd International Conference on Vector and Parallel Computing Issues in Applied
Research and Development, Troms, 1988, Edited by J. Dongarra et. al., London: Ellis Horwood, 179{188.
Kugelmann, B. and Pesch, H. J.: 1990a, `New General Guidance Method in Constrained Optimal Control, Part 1: Numerical
Method', Journal of Optimization Theory and Applications 67, 421{435.
Kugelmann, B. and Pesch, H. J.: 1990b, `New General Guidance Method in Constrained Optimal Control, Part 2: Application
to Space Shuttle Guidance', Journal of Optimization Theory and Applications 67, 437{446.
Kugelmann, B. and Pesch, H. J.: 1990c, `Serielle und parallele Algorithmen zur Korrektur optimaler Flugbahnen in
Echtzeit-Rechnung', Jahrestagung der Deutschen Gesellschaft fur Luft- und Raumfahrt, Friedrichshafen, 1990, DGLR-
Jahrbuch 1990 1, 233{241.
Kugelmann, B. and Pesch, H. J.: 1991, `Real-Time Computation of Feedback Controls with Applications in Aerospace
Engineering', Proceedings of the AIAA Guidance, Navigation and Control Conference, New Orleans, Louisiana, 1991,
AIAA Paper 91-2658.
Maurer, H.: 1976, Optimale Steuerprozesse mit Zustandsbeschrankungen, Habilitationsschrift, University of Wurzburg,
Wurzburg.
Maurer, H. and Pesch, H. J.: to appear, `Solution Dierentiability for Nonlinear Parametric Control Problems', SIAM Journal
on Control and Optimization.
Maurer, H. and Pesch, H. J.: to appear, `Solution Dierentiability for Parametric Nonlinear Control Problems with Inequality
Constraints', Proceedings of the 16th IFIP Conference on System Modelling and Optimization, Compiegne, 1993, Lecture
Notes in Control and Information Science, Berlin: Springer.
McDanell, J. P. and Powers, W. F.: 1971, `Necessary Conditions for Joining Optimal Singular and Nonsingular Subarcs',
SIAM Journal on Control 9, 161{173.
Miele, A., Wang, T., and Melvin, W. W.: 1987, Òptimal Abort Landing Trajectories in the Presence of Windshear', Journal
of Optimization Theory and Applications 55, 165{202.
Oberle, H. J.: 1976, Numerical Computation of Minimum-Fuel Space-Travel Problems by Multiple Shooting, Department of
Mathematics, Munich University of Technology, Munich, Report M-7635.
Oberle, H. J.: 1977, Òn the Numerical Computation of Minimum-Fuel, Earth-Mars Transfer', Journal of Optimization
Theory and Applications 22, 447{453.
Oberle, H. J.: 1982, Numerische Berechnung optimaler Steuerungen von Heizung und Kuhlung fur ein realistisches Sonnen-
hausmodell, Habilitationsschrift, Munich University of Technology, Munich.
Oberle, H. J. and Grimm, W.: 1989, BNDSCO|A Program for the Numerical Solution of Optimal Control Problems,
Internal Report No. 515-89/22, Institute for Flight Systems Dynamics, Oberpfaenhofen: German Aerospace Research
Establishment DLR.
Pesch, H. J.: 1979, `Numerical Computation of Neighboring Optimum Feedback Control Schemes in Real-Time', Applied
Mathematics and Optimization 5, pp. 231{252.
Pesch, H. J.: 1980, `Neighboring Optimum Guidance of a Space-Shuttle-Orbiter-Type Vehicle', Journal of Guidance and
Control 3, 386{391.
Pesch, H. J.: 1989a, `Real-time Computation of Feedback Controls for Constrained Optimal Control Problems, Part 1:
Neighbouring Extremals', Optimal Control Applications and Methods 10, 129{145.
Pesch, H. J.: 1989b, `Real-time Computation of Feedback Controls for Constrained Optimal Control Problems, Part 2: A
Correction Method Based on Multiple Shooting', Optimal Control Applications and Methods 10, 147{171.
Pesch, H. J.: 1990a, Òptimal Re-Entry Guidance under Control and State Constraints', Proceedings of the 8th IFAC Work-
shop on Control Applications of Nonlinear Programming and Optimization, Paris, 1989, Edited by P. Bernhard and
H. Bourdache-Siguerdidjane, IFAC Workshop Series, 1990, No. 2, Oxford: Pergamon Press, 33{38.
Pesch, H. J.: 1990b, Òptimal and Nearly Optimal Guidance by Multiple Shooting', Mecanique Spatiale - Space Dynamics,
Proceedings of the International Symposium, Toulouse, 1989, Edited by Centre National dEtudes Spatiales, Toulouse:
Cepadues Editions, 761{771.
Pesch, H. J.: to appear, `Solving Optimal Control and Pursuit-Evasion Game Problems of High Complexity', Proc. of the
9th IFAC Workshop on Control Applications of Optimization, Munich, 1992, Edited by R. Bulirsch and D. Kraft, Basel:
Birkhauser (ISNM).
Pesch, H. J., Schlemmer, M., and von Stryk, O.: to appear, `Minimum-Energy and Minimum-Time Control of Three-Degrees-
of-Freedom Robots, Part 1: Mathematical Model and Necessary Conditions'.
Pesch, H. J., Schlemmer, M., and von Stryk, O.: to appear, `Minimum-Energy and Minimum-Time Control of Three-Degrees-
of-Freedom Robots, Part 2: Numerical Methods and Results for the Manutec r3 Robot.
Sanger, E.: 1962, `Raumfahrt | gestern, heute und morgen', Astronautica Acta VIII 6, 323{343.
Scharmack, D. K.: 1967, Àn Initial Value Method for Trajectory Optimization Problems', Advances in Control Systems 5,
Edited by C. T. Leondes, New York: Academic Press, 1{65.
Shau, G.-C.: 1973, Der Ein u ugmechanischer Parameter auf die Aufstiegsbahn von horizontal startenden Raumtransport-
ern bei gleichzeitiger Bahn- und Stufungsoptimierung, Dissertation, Department of Mechanical and Electrical Engineer-
ing, University of Technology, Braunschweig.
Stoer, J. and Bulirsch, R.: 1980, Introduction to Numerical Analysis, Springer: New York (2nd Edition, 1993).
von Stryk, O. and Schlemmer, M.: Òptimal Control of the Industrial Robot Manutec r3', Proceedings of the 9th IFAC
Workshop of Control Applications of Optimization, Edited by R. Bulirsch and D. Kraft, International Series of Numerical
Mathematics, Basel: Birkhauser.
von Stryk, O. and Bulirsch, R.: 1992, `Direct and Indirect Methods for Trajectory Optimization', Annals of Operations
Research 37, 357{373.

A Practical Guide To The Solution of Real-Life Optimal Control Problems

Uploaded by

Copyright:

Available Formats

A Practical Guide To The Solution of Real-Life Optimal Control Problems

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

A Practical Guide To The Solution of Real-Life Optimal Control Problems

Uploaded by

Copyright:

Available Formats

1

A PRACTICAL GUIDE TO THE SOLUTION OF REAL-LIFE OPTIMAL CONTROL

_ = 2Sm % V CL (u) + RV(1cos ? g0 cos ;

cos u = +CDL q V V : (11b)

_ = R V+ h cos sin  ; (24d)

where the Lagrange parameter  is de ned by Eq. (9a),

4. Optimal Control Problems with State Variable Inequality Constraints

5. Optimal Control Problems with Singular Subarcs

_ = mV 1cos [T (V; h; b) sin  + L(V; h; u)] sin  (70c)

? R V+ h cos cos  tan

7. O -line Computation of Optimal Trajectories by Multiple Shooting

z(k+ ) = k (k ; z(k? )) ; 1  k  s ; (73b)

Fig. 4. Control history for the Apollo re-entry; u max = 62 deg.

Fig. 5. Control history for the Apollo re-entry; u max = 23 deg.

Fig. 6. Re-entry trajectory for the Space-Shuttle re-entry; control-unconstrained case.

Fig. 7. Altitude history for the Apollo re-entry; h max  50 97 km.

Fig. 8. Altitude history for the Apollo re-entry; h max  42 05 km.

8.4. Abort Landing of a Passenger Aircraft Under Windshear Conditions

8.5. Maximum Payload Ascent of a Two-Stage-To-Orbit Space Transporter System

Fig. 12. Launch and stage separation of a Sanger-type space vehicle.

10. On-line Computation of Optimal Trajectories

Fig. 13. Chart ow of the neighboring optimum feedback guidance scheme.

Fig. 14. Chart ow of the repeated correction guidance scheme.

11. Numerical Results II

12. Conclusions and Outlock

You might also like

cos u = +CDL q V V : (11b)

_ = R V+ h cos sin ; (24d)

where the Lagrange parameter is dened by Eq. (9a),

_ = mV 1cos [T (V; h; b) sin + L(V; h; u)] sin (70c)

? R V+ h cos cos tan

7. O-line Computation of Optimal Trajectories by Multiple Shooting

z(k+ ) = k (k ; z(k? )) ; 1 k s ; (73b)

Fig. 7. Altitude history for the Apollo re-entry; h max 50 97 km.

Fig. 8. Altitude history for the Apollo re-entry; h max 42 05 km.

Fig. 12. Launch and stage separation of a Sanger-type space vehicle.