Journal of Computational and Applied Mathematics 120 (2000) 85–108

www.elsevier.nl/locate/cam

SQP-methods for solving optimal control problems with control

and state constraints: adjoint variables, sensitivity analysis and
real-time control
Christof Buskens a; ∗ , Helmut Maurer b
a
Universitat Bayreuth, Lehrstuhl fur Ingenieurmathematik, FAN, D-95440 Bayreuth, Germany
b
Westfalische Wilhelms-Universitat Munster, Institut fur Numerische und instrumentelle Mathematik,
Einsteinstrasse 62, D-48149 Munster, Germany
Received 14 December 1998; received in revised form 8 September 1999

Abstract
Parametric nonlinear optimal control problems subject to control and state constraints are studied. Two discretiza-
tion methods are discussed that transcribe optimal control problems into nonlinear programming problems for which
SQP-methods provide ecient solution methods. It is shown that SQP-methods can be used also for a check of second-order
sucient conditions and for a postoptimal calculation of adjoint variables. In addition, SQP-methods lead to a robust
computation of sensitivity di erentials of optimal solutions with respect to perturbation parameters. Numerical sensitiv-
ity analysis is the basis for real-time control approximations of perturbed solutions which are obtained by evaluating a
rst-order Taylor expansion with respect to the parameter. The proposed numerical methods are illustrated by the optimal
control of a low-thrust satellite transfer to geosynchronous orbit and a complex control problem from aquanautics. The
examples illustrate the robustness, accuracy and eciency of the proposed numerical algorithms. c 2000 Published by
Elsevier Science B.V. All rights reserved.

Keywords: Optimal control problems; Control-state constraints; Nonlinear programming methods; Adjoint variables;
Second-order sucient conditions; Sensitivity analysis; Real-time control

1. Introduction

This paper is concerned with the numerical solution of optimal control problems that are governed
by ordinary di erential equations and are subject to control and state constraints. Such control prob-
lems arise in many applications, e.g., in economics, chemical engineering, robotics, aeronautics and
∗
Corresponding author.

0377-0427/00/$ - see front matter c 2000 Published by Elsevier Science B.V. All rights reserved.
PII: S 0 3 7 7 - 0 4 2 7 ( 0 0 ) 0 0 3 0 5 - 8
86 C. Buskens, H. Maurer / Journal of Computational and Applied Mathematics 120 (2000) 85–108

astronautics. In recent years, the theoretical and numerical treatment of optimal control problems has
also stimulated a growing interest in optimization techniques.
The theory of necessary conditions for optimal control problems with control and state constraints
is well developed and characterizes optimal solutions as solutions of multipoint boundary value
problems (MBVP). For such MBVPs, shooting techniques have been established as ecient and
reliable solution methods providing highly accurate solutions. However, the numerical experience
with shooting methods gained in the last 30 years has revealed a severe drawback: these methods
need a rather precise initial guess of the optimal state, control and adjoint variables and require a
detailed a priori knowledge of the structure of the optimal solution, i.e., of the number of active
time intervals for the constraints. In practice, it is usually rather dicult to determine the structure
of the optimal control and to nd appropriate estimates for the adjoint variables.
Numerical methods that avoid these drawbacks can be found in so-called direct optimization
methods which have been studied extensively in the last 20 years. Direct optimization methods
have proved to be powerful tools for solving optimal control problems (cf., e.g., [1– 6,10,33,34]).
The basic idea of direct optimization methods is to discretize the control problem and to apply
nonlinear programming (NLP) techniques to the resulting nite-dimensional optimization problem.
These methods use only control and state variables as optimization variables and dispense completely
with adjoint variables. Adjoint variables can be obtained by a postoptimal calculation using the
Lagrange multipliers of the resulting nonlinear optimization problem.
Numerical algorithms are usually developed on the basis of rst order necessary optimality con-
ditions. Besides necessary conditions, second-order sucient conditions (SSC) have to be checked
to ensure the optimality of solutions. SSC for continuous control problems are usually dicult to
verify numerically. As an alternative, SSC for the discretized control problem can be tested more
easily using well-known linear algebra techniques for optimization problems.
Another important aspect of SSC appears in sensitivity analysis of optimal control problems. In
practice, control problems are often subject to disturbances or perturbations in the system data. In
mathematical terms, perturbations can be described by parameters in the dynamics, boundary con-
ditions or in control and state constraints. Stability and sensitivity analysis is concerned with the
behavior of optimal solution with respect to parameters. Malanowski, et al. [21,22,24,25] have stud-
ied di erential properties of optimal solutions. The theoretical framework in these papers rests on
indirect methods using boundary value methods. This approach allows to compute the sensitivity dif-
ferentials of optimal solutions with respect to perturbation parameters. Sensitivity di erentials can be
used eciently for real-time control approximations of perturbed solutions (cf. [19,20,24,25,30,31]).
However, instead of working with boundary value methods we shall develop optimization techniques
to compute sensitivity di erentials via the discretized control problem.
The organization of the paper is as follows. Parametric optimal control problems with control
and state constraints are introduced in Section 2. We summarize necessary optimality conditions and
give a short overview on basic results for sensitivity analysis and real-time control approximations.
Section 3 discusses two discretization methods for optimal control problems and the form of the
resulting NLP-problems. Section 4 presents methods for estimating the adjoint variables on the basis
of the two optimization approaches in Section 3. A numerical check of SSC for nite-dimensional
NLP-problems is proposed in Section 5. A formula for the sensitivity di erentials of perturbed so-
lutions with respect to parameters is included. In Section 6, this sensitivity formula is established
as a basis for computing sensitivity di erentials of optimal control, state and adjoint variables.
 C. Buskens, H. Maurer / Journal of Computational and Applied Mathematics 120 (2000) 85–108 87

Section 7 presents two numerical examples for optimal control problems with constraints. The nu-
merical methods from Sections 3– 6 lead to a complete numerical solution and illustrate also the
quality of real-time approximations in the presence of disturbances.

2. Parametric optimal control problems with constraints

2.1. Necessary optimality conditions for optimal control problems

We consider parametric nonlinear control problems subject to control and state constraints. Let
x(t) ∈ Rn denote the state of a system and u(t) ∈ Rm the control in a given time interval [t0 ; tf ].
Data perturbations in the system are modeled by a parameter p ∈ P := Rq .
Parametric control problem OC ( p):
Z tf
Minimize F(x; u; p) = g(x(tf ); p) + f0 (x(t); u(t); p) dt
t0

subject to ẋ(t) = f(x(t); u(t); p); (1)

x(t0 ) = ’(p); (x(tf ); p) = 0;
C(x(t); u(t); p)60; t ∈ [t0 ; tf ]:
The functions g : Rn × P → R; f0 : Rn+m × P → R; f : Rn+m × P → Rn ; ’ : P → Rn ; : Rn × P →
Rr ; 06r6n; and C : Rn+m × P → Rk are assumed to be suciently smooth on appropriate open sets.
The admissible class of control functions is that of piecewise continuous controls. The nal time tf is
either xed or free. Note that the formulation of mixed state-control constraints C(x(t); u(t); p)60 in
(1) includes pure control constraints C(u(t); p)60 as well as pure state constraints C(x(t); p)60 .
Nonautonomous control problems can be reduced to the autonomous problem OC(p) in a standard
way. In the sequel, we assume for simplicity that problem (1) is in Mayer form with f0 (x; u; p) ≡
0. This form can be achieved by introducing an additional state variable x n+1 de ned by the ODE
ẋn+1 = f0 (x; u; p); x n+1 (t0 ) = 0.
The theory of necessary conditions (minimum principle) for control problems of form (1) is
well developed (cf. [32,29,17]). The following form of the necessary conditions requires additional
regularity conditions for the control and state constraints (cf., e.g., [21–23]) which are not discussed
here in detail but which are tacitly assumed to be ful lled. Moreover, for the sake of simplicity we
suppose that the inequality constraints are not active at the initial and nal state:
C(x(t0 ); u(t0 ); p) ¡ 0 and C(x(tf ); u(tf ); p) ¡ 0: (2)
The augmented Hamiltonian function for problem (1) is
H (x; ; ; u; p) := T f(x; u; p) + T C(x; u; p); (3)
where T denotes the transpose,  ∈ Rn is the adjoint variable and  ∈ Rk is a multiplier associated to
the inequality constraints. Let the pair (x∗ ; u∗ ) be an optimal solution for OC(p). Then the necessary
conditions for (x∗ ; u∗ ) (cf. [17,29]) imply that there exists a piecewise continuous and piecewise con-
tinuously di erentiable adjoint function  : [t0 ; tf ] → Rn , a piecewise continuous multiplier function
88 C. Buskens, H. Maurer / Journal of Computational and Applied Mathematics 120 (2000) 85–108

 : [t0 ; tf ] → Rk ; (t)¿0 for t ∈ [t0 ; tf ], and a multiplier  ∈ Rr satisfying

˙ T = −Hx (x∗ (t); (t); (t); u∗ (t); p)
(t)

= −(t)T fx (x∗ (t); u∗ (t); p) − (t)T Cx (x∗ (t); u∗ (t); p); (4)

(tf )T = gx (x∗ (tf ); p) + T x (x∗ (tf ); p); (5)

0 = Hu (x∗ (t); (t); (t); u∗ (t); p)

= (t)T fu (x∗ (t); u∗ (t); p) + (t)T Cu (x∗ (t); u∗ (t); p): (6)
Herein, the subscript x or u denotes the partial derivative with respect to x or u. Eq. (5) is
called the transversality condition for the adjoint variable. In addition, for pure state constraints
C(x(t))60; t ∈ [t0 ; tf ], there exist multipliers (tj )¿0, (tj ) ∈ Rk in each junction or contact point
tj , such that the junction conditions
(tj+ )T = (tj− )T − (tj )T Cx (x(tj )) (7)
are satis ed; the notation + and − in (7) indicates the limit from the left, resp. from the right. The
precise de nition of junction or contact points may be found in [17,22]. Note that the multiplier
(t) can be computed from (4) – (6) as a function of x(t); u(t), and (t).

2.2. Sensitivity analysis and real-time control approximation

Sensitivity analysis is concerned with the behavior of optimal solutions to problem OC(p) with
respect to the parameter p. Of particular practical importance is the property of solution di erentia-
bility of optimal solutions. Malanowski, et al. [21,22,24,25] have derived conditions which ensure
that, roughly speaking, solutions to Eqs. (4) – (6) become di erentiable functions of the parameter p.
To sketch these sensitivity results, let us x a reference or nominal parameter p0 and consider
problem OC(p0 ) as the unperturbed or nominal problem. We assume that there exists a local solution
(x0 ; u0 ) of the reference problem OC(p0 ) such that u0 ∈ C(t0 ; tf ; Rm ) is continuous. Let 0 (t) and 0 (t)
be the associated multipliers satisfying the necessary conditions (4) – (6) for the nominal parameter
p0 . The following result summarizes the main ingredients from Malanowski and Maurer [21,22] and
Maurer and Pesch [24,25].
Solution differentiability for control problems:
Suppose that along the reference solution (x0 ; u0 ) the following assumptions are satis ed:
(a) constraint quali cations hold;
(b) second-order sucient conditions are satis ed;
(c) there are nitely many non-tangential junction or contact points;
(d) strict complementarity holds for the multiplier 0 (t).
Then the unperturbed solution x0 (t); 0 (t); u0 (t) can be embedded into a family of optimal solu-
tions x(t; p); (t; p); u(t; p) to the perturbed problem OC(p). This family is piecewise of class C 1
for all p in a neighborhood of the reference parameter p0 .

We wish to emphasize that this result applies only to special classes of control problems where
the control u0 is supposed to be continuous and where the Legendre–Clebsch condition holds
 C. Buskens, H. Maurer / Journal of Computational and Applied Mathematics 120 (2000) 85–108 89

(cf. [21–23]). These assumptions exclude problems where the control appears linearly leading to
bang–bang or singular controls. However, this statement applies only to sensitivity analysis and
does not prevent the SQP-methods developed in Section 3 from being successful for bang–bang and
singular controls.
The property of solution di erentiability is crucial for designing real-time control approximations
of perturbed solutions. Real-time control algorithms have been developed in the literature since the
early 1970s. Early approaches apply only to special cases, e.g., to problems without control and state
constraints. The repeated correction method of Kugelmann and Pesch [19,20], Pesch [30,31] was
the rst method taking inequality constraints into account. Using the notations of the above solution
di erentiability result, the basic idea of real-time control approximations consists in approximating
the perturbed optimal solution y(t; p) := (x(t; p); (t; p); u(t; p)) by its rst-order Taylor expansion
@y
y(t; p) ≈ y0 (t) + (t; p0 )(p − p0 ): (8)
@p
The nominal functions y0 (t) as well as the sensitivity di erentials (@y=@p)(t; p0 ) are computed
o -line. In case that an actual deviation of the parameter p from the nominal parameter p0 is
detected, Eq. (8) provides a very fast on-line approximation for the perturbed solution since it
requires only matrix-multiplications.
The work in [19 –22,24,25,30,31] employs the indirect method to show that the sensitivity dif-
ferentials (@y=@p)(t; p0 ) satisfy a linear boundary value problem including the nominal state and
adjoint variables. We shall demonstrate in the following sections that sensitivity di erentials can
also be computed eciently via NLP-methods that dispense with adjoint variables and do not need
a precise knowledge of the structure of the optimal solution.

3. Numerical solution of optimal control problems by nonlinear programming techniques

Direct optimization methods for solving the optimal control problem OC(p) are based on a suitable
discretization of problem (1). In this section we describe two optimization approaches where for
simplicity the discussion will be restricted to Euler’s method.

3.1. The full discretization approach

Choose a natural number N and let i ∈ [t0 ; tf ], i = 1; : : : ; N , be mesh or grid points with
t0 = 1 ¡ · · · ¡ N −1 ¡ N = tf : (9)
For notational simplicity we assume that the discretization in (9) is equidistant:
tf − t0
h := ; i = t0 + (i − 1) · h; i = 1; : : : ; N: (10)
N −1
Let the vectors xi ∈ Rn resp. ui ∈ Rm ; i = 1; : : : ; N; be approximations of the state variable x(i ),
resp. control variables u(i ) at the grid points. Then Euler’s approximation applied to the di erential
equation in (1) yields
xi+1 = xi + h · f(xi ; ui ; p); i = 1; : : : ; N − 1: (11)
90 C. Buskens, H. Maurer / Journal of Computational and Applied Mathematics 120 (2000) 85–108

Choosing the optimization variable

z := (x1 ; : : : ; xN ; u1 ; : : : ; uN ) ∈ RNz ; Nz := (n + m)N; (12)
the optimal control problem OC(p) in Mayer form with f0 ≡ 0 is replaced by the following
discretized control problem DOC1 (p) in the form of a nonlinear programming problem with equality
and inequality constraints:
NLP-problem DOC 1 ( p):
Minimize F(z; p) = g(xN ; p) (13)

subject to − xi+1 + xi + h · f(xi ; ui ; p) = 0; i = 1; : : : ; N − 1; (14)

’(p) − x1 = 0; (15)

(xN ; p) = 0; (16)

C(xi ; ui ; p)60; i = 1; : : : ; N: (17)

A free nal time tf can be handled as an additional optimization variable in (12). The nonlinear
optimization problem DOC1 (p) has a large number of optimization variables Nz = (n + m)N and
constraints M =(n+k)N +r. Nevertheless, it can be solved in an ecient way by taking into account
the sparse structure in the Jacobian of the constraints and the Hessian of the associated Lagrangian
(24) (cf. [1–3]).
Let Ne = nN + r be the number of equality constraints and Nc = Ne + kN the total number of
constraints in (14) – (17). Furthermore, let G(z)=(G1 (z); : : : ; GNc (z)) denote the collection of functions
de ning the equality and inequality constraints (14) – (17). Then the nonlinear programming problem
DOC1 (p) has the form:
NLP-problem NLP( p):
Minimize F(z; p)
subject to Gi (z; p) = 0; i = 1; : : : ; Ne ; (18)
Gi (z; p)60; i = Ne + 1; : : : ; Nc :

3.2. A recursive approach

For our purposes, in particular for sensitivity analysis, it is more convenient to reduce the di-
mension of DOC1 (p). This is achieved by treating the control variables as the only optimization
variables while the state variables are obtained recursively from the state equation (11) or (14).
Therefore, we consider the optimization variable
z := (u1 ; u2 ; : : : ; uN −1 ; uN ) ∈ RNz ; Nz := m · N; (19)
and compute recursively the state variables from (11),
xi = xi (z; p) := xi (u1 ; : : : ; ui−1 ; p); (20)
 C. Buskens, H. Maurer / Journal of Computational and Applied Mathematics 120 (2000) 85–108 91

as functions of the control variables with initial condition x1 = ’(p) given in (1). This leads to the
following
NLP-problem DOC 2 ( p):
Minimize F(z; p) := g(xN (z; p); p) (21)
subject to (xN (z; p); p) = 0; (22)
C(xi (z; p); ui ; p)60; i = 1; : : : ; N: (23)
In this case, the reduced number of optimization variables Nz = mN and constraints M = kN + r
results in a dense structure in the Hessian of the Lagrangian (29), whereas about 50% of the
elements in the Jacobian of the constraints are zero. Again, a free nal tf is handled as an additional
optimization variable in (19). Setting the dimensions Ne = r; Nc = Ne + kN , the discretized control
problem DOC2 (p) de nes again a NLP-problem of form (18).
Problems of form (18) can be solved eciently using sequential quadratic programming (SQP)
methods; compare, e.g., the code E04UCF in the NAG-library and the codes mentioned in the sur-
vey [1]. Instead of Euler’s method incorporated into relations (11), (14) or (20) – (23) one can use
higher-order integration methods combined with a higher-order control approximations. All calcula-
tions described in the sequel were performed by the code NUDOCCCS of Buskens [5,6] which has
implemented various higher-order approximations for state and control. Recently, the convergence
of solutions discretized via Euler’s method to solutions of the continuous control problem has been
proved in [23]. Convergence of higher-order Ritz discretization schemes has been established in
[12,13].

4. Estimating adjoint variables

Buskens [6], Enright and Conway [11], Grimm [16] and Stryk [33] demonstrate that direct ap-
proaches for optimal control problems with constraints can be used for a postoptimal calculation of
approximations to adjoint variables. The general idea of these methods is to employ necessary condi-
tions and to use Lagrange multipliers from the equality and inequality constraints of the discretized
control problems for estimating the adjoint variables. In this section we describe two numerical
methods for a postoptimal computation of adjoint variables which refer to the full discretization
approach in Section 3.1 or the recursive approach in Section 3.2.

4.1. Estimating adjoint variables using the full discretization approach

Let us introduce the Lagrangian function for the nonlinear optimization problem DOC1 (p) in
(13) – (17):
L : RNz × RnN × Rr × RkN × P → R;
N
X −1
L(z; ; ; ; p) := g(xN ; p) + (i+1 )T (xi + hf(xi ; ui ; p) − xi+1 )
i=1
N
X
+ (1 )T (’(p) − x1 ) + T (xN ; p) + (i )T C(xi ; ui ; p) (24)
i=1
92 C. Buskens, H. Maurer / Journal of Computational and Applied Mathematics 120 (2000) 85–108

with multipliers i ∈ Rn ;  = (1 ; : : : ; N )T ∈ RnN ;  ∈ Rr ; i ∈ Rk ;  = (1 ; : : : ; N )T ∈ RkN

+ . Assump-
tion (2) yields 0 = N = 0. The rst-order optimality conditions of Karush–Kuhn–Tucker for the
NLP-problem (13) – (17) are
Lz (z; ; ; ; p) = 0; (i )T C(xi ; ui ; p) = 0; (i = 2; : : : ; N − 1); (25)
where the subscript z denotes the partial derivative with respect to z. By evaluating (25) we obtain
an estimate of the continuous adjoint variables (t) in (4) and (5) from the multipliers of the
discretized problem. Omitting the argument in L we get the relations
0 = LxN = gxN (xN ; p) + T xN (x
N
; p) − (N )T ; (26)

0 = Lx i = (i+1 )T + h(i+1 )T fx i (x i ; ui ; p) − (i )T

+ (i )T Cx i (x i ; ui ; p); i = 1; : : : ; N − 1; (27)

0 = Lu i = h(i+1 )T fu i (x i ; ui ; p) + (i )T Cu i (x i ; ui ; p); i = 1; : : : ; N − 1: (28)

Eq. (26) represents the discrete version of the transversality condition (5) while Eq. (27) is the
discrete version of the adjoint variable (t) in (4) using the identi cation (i ) ≈ i =h, i = 1; : : : ; N .
In case of a pure state constraint C(x; p)60 we have to modify this identi cation as follows. Let
l be an approximation for the junction or contact point tj and let l be the corresponding multiplier
in (24). Then we obtain the splitting l ≈ j + h · (tj ) where j ≈ (tj ) is the jump in (7). Finally,
Eq. (28) re ects the optimality condition (6) for the control variable.

4.2. Estimating adjoint variables with the recursive approach

A further approach for estimating adjoint variables is based on the NLP-problem DOC2 (p) in
(21) – (23). In contrast to the method described in the last section, we can eliminate the state vari-
ables x i as optimization variables and thus do not need Lagrange multipliers for the dynamic equa-
tions (14). However, Buskens [6] has shown that an approximation of the adjoint variables (i )
for the optimal control problem OC(p) can be calculated a posteriori from the Lagrange function
corresponding to DOC2 (p). Because of the lengthy proof we omit details and sketch only the main
ideas. The Lagrangian function for the nonlinear optimization problem (21) – (23) is

L : RNz × Rr × RkN × P → R;
N
X
L(z; ; ; p) := g(xN (z; p); p) + T (xN (z; p); p) + (i )T C(x i (z; p); ui ; p): (29)
i=1

Note that Nz = mN now represents the reduced number of optimization variables z = (u1 ; u2 ; : : : ; uN )
for DOC2 (p) in (19) and (20). A rst intuitive approach for computing multipliers i ≈ (i ) is as
follows. The SQP-method provides the multiplier  ∈ Rr in (29) which allows to de ne the vector
N through the transversality condition (26). Then we use Eq. (27) to de ne recursively vectors
i for i = N − 1; : : : ; 1. A careful analysis of the optimality conditions Lz = 0 then reveals that the
necessary condition (28) holds.
A second alternative method for approximating the adjoint variables (i ) is given by the expres-
sions
(i ) ≈ i := Lx i (z; ; ; p); i = 1; : : : ; N: (30)
 C. Buskens, H. Maurer / Journal of Computational and Applied Mathematics 120 (2000) 85–108 93

Again, a careful study of the Karush–Kuhn–Tucker conditions for the NLP-problem (21) – (23) shows
that the vectors i as de ned by (30) satisfy the discrete adjoint equation (27) and the minimum
condition (28) (cf. [6]). Especially, the second proposed method delivers a very reliable and accurate
approximation even for complicated and highly nonlinear problems and problems including pure
state constraints. Further diculties occurring in all direct methods for control systems with mixed
constraints or nonlinear state constraints are discussed in [6].

5. Second-order sucient conditions and sensitivity analysis for NLP-problems

5.1. A numerical check of second-order sucient conditions

Second-order sucient conditions (SSC) for continuous optimal control problems (1) are often
dicult to verify. In this section we describe a numerical check for SSC for the discretized control
problems discussed in Section 3. We have seen that both discretized control problems DOC1 (p) and
DOC2 (p) lead to the parametric NLP-problem NLP(p) in (18) with di erent dimensions Ne and Nc .
The Lagrangian function for problem NLP(p) is
L(z; ; p) = F(z; p) + T G(z; p);  ∈ R Nc : (31)
Let z be an optimal solution for NLP(p) with associated Lagrange multiplier  satisfying necessary
optimality conditions. Consider the set of active indices de ned by Ia (p) := {i ∈ {1; : : : ; Nc } | Gi (z;
 p)
= 0} and let ma := #Ia (p).
Denote the active constraints by G a := (Gi )i∈Ia (p) and the multiplier corresponding to active con-
straints by  a ∈Rma . Let Gza be the Jacobian with dimension ma ×Nz . The following strong second-order
sucient conditions are well known (cf. [14]).

Strong second-order sucient conditions for NLP(p):

Assume that
(a) F and G are twice continuously di erentiable with respect to z and p;
(b) the gradients in Gza are linearly independent; i.e.; rank(Gza (z;
 p)) = ma ;
a
(c) strict complementarity  ¿ 0 of the Lagrange multipliers holds;
(d) the Hessian of the Lagrangian is positive de nite on Ker(Gza (z;  p));
vT Lzz (z;
 ;
 p)v ¿ 0; ∀ v ∈ Ker(Gza (z;
 p)); v 6= 0: (32)
Then z is a local minimum for NLP(p).

A numerical check of the SSC (32) consists in evaluating the projected Hessian on Ker(Gza ) and
in verifying that its eigenvalues are positive. In the sequel, we shall omit all arguments of functions.
The projected Hessian is de ned as the (Nz − ma ) × (Nz − ma ) matrix
Lzz := H T Lzz H; (33)
where the Nz × (Nz − ma ) matrix H contains an orthogonal basis of Ker(Gza ), i.e., Gza H = 0 holds.
The positive de niteness of the projected Hessian Lzz is equivalent to the positive de niteness of
the Hessian Lzz on Ker(Gza ). RQ-factorization is an ecient method to nd an orthogonal basis for
94 C. Buskens, H. Maurer / Journal of Computational and Applied Mathematics 120 (2000) 85–108

the null space of Gza (cf. [15]):

 
Q1
Gza = RQ with R = [R1 | 0] and Q= : (34)
Q2
In this decomposition, R is a ma ×Nz matrix and R1 is a ma ×ma lower triangular matrix with positive
diagonal elements, while Q is a Nz × Nz matrix, Q1 is a ma × Nz matrix and Q2 is a (Nz − ma ) × Nz
matrix. Then H := Q2T forms an orthogonal basis for Ker(Gza ) which follows from Gza H = 0. The
T
matrices Q1 and R1 are unique while Q2 is not. If Q̃2 is another orthogonal basis for the null space
of Gza , then the corresponding projected Hessian is a matrix which is similar to Lzz . Thus, the
eigenvalues of the projected Hessian are independent of the orthogonal basis for Ker(Gza ).

5.2. Sensitivity analysis for parametric nonlinear optimization problems

A thorough treatment of sensitivity analysis can be found in [14].

Solution di erentiability for NLP(p):

Suppose that the optimal solution (z0 ; 0 ) satis es the strong second-order sucient condi-
tions for the nominal problem NLP(p0 ). Then for p near to p0 the unperturbed solution (z0 ; 0 )
can be embedded into a C 1 -family of perturbed optimal solutions (z(p); (p)) for NLP(p) with
(z(p0 ); (p0 ))=(z0 ; 0 ). The active sets Ia (p) coincide with Ia (p0 ) and hence it follows that i (p)=0
for i 6∈ Ia (p). The sensitivity di erentials of the optimal solutions and Lagrange multipliers are
given by the formula
 
dz
!−1 !
 dp (p0 )  Lzz (z0 ; 0 ; p0 ) Gza (z0 ; p0 )T Lzp (z0 ; 0 ; p0 )
 
 =− : (35)
 da  Gza (z0 ; p0 ) 0 Gpa (z0 ; p0 )
(p0 )
dp
Moreover; the sensitivity of the objective function is obtained from
dF
(z(p0 ); p0 ) = Lp (z0 ; 0 ; p0 ): (36)
dp
Note that the so-called Kuhn–Tucker matrix on the right-hand side of (35) is regular since second-order
sucient conditions (32) are assumed to hold. One possibility of evaluating the Kuhn–Tucker matrix
and the incorporated Hessian Lzz consists in using SQP-updates which are obtained in the process
of computing the nominal solution. However, these SQP-updates usually give rather inaccurate ap-
proximations. Hence, in our approach the Hessian Lzz and the Kuhn–Tucker matrix will be evaluated
after computing the optimal solution (z0 ; 0 ), which yields much more accurate results.
Formula (36) can be used in a rst-order Taylor expansion of the objective functional,
dF
F(z(p); p) ≈ F(z; p0 ) + [p0 ](p − p0 ); (37)
dp
where the notation [p0 ] stands for all nominal arguments. Under additional smoothness assumptions
the above solution di erentiability for NLP-problems can be extended to a second-order sensitivity
 C. Buskens, H. Maurer / Journal of Computational and Applied Mathematics 120 (2000) 85–108 95

derivative for the objective function,

 T  T
d2 F dz dz dz
[p0 ] = (p0 ) Lzz [p0 ] (p0 ) + 2 Lpz [p0 ] (p0 ) + Lpp [p0 ]; (38)
dp2 dp dp dp
with (d z=dp)(p0 ) substituted from Eq. (35). This leads to a second-order Taylor-expansion for the
objective function according to
dF d2 F
F(z(p); p) ≈ F(z; p0 ) + [p0 ](p − p0 ) + 12 (p − p0 )T 2 [p0 ](p − p0 ): (39)
dp dp
Among others, Eq. (39) can be used to estimate the error of the rst-order real-time approxi-
mation (8).

6. Sensitivity analysis for perturbed optimal control problems with constraints

The sensitivity analysis in [21,22] is based on the indirect method where boundary value problems
of type (4) – (6) involving state and adjoint variables are resolved. Although this approach provides
a theoretical framework, the numerical examples in [20–22,31] show that a numerical sensitivity
analysis becomes rather tedious using indirect methods. It is somewhat surprising that to our knowl-
edge results from optimization theory have not yet been used in numerical sensitivity analysis of
control problems. The purpose of this section is to develop a robust NLP-method for the computation
of sensitivity derivatives of optimal solutions with respect to parameters. The method is based on
formulas (35) – (38) applied to the discretized control problem DOC2 (p0 ) and allows to compute
sensitivity di erentials for state, control and adjoint variables.
The computation of (d z=dp)(p0 ) via (35) yields an approximation for the sensitivity di erentials
of the perturbed optimal control solutions at the mesh points, i.e., for the quantities
@u dui
(i ; p0 ) ≈ (p0 ); i = 1; : : : ; N: (40)
@p dp
Since a free terminal time tf is handled as an additional optimization variable in DOC2 (p0 ), its
sensitivity di erential (dtf =dp)(p0 ) can likewise be calculated from Eq. (35).
Eq. (40) grants also the sensitivity di erentials of the state with respect to the parameter, if we
take into account the recursive formulation (20) together with relation (11). The state sensitivities
(@x=@p)(i ; p0 ) are obtained from
@x dxi @x i dz @x i
(i ; p0 ) ≈ [p0 ] = (z0 ; p0 ) (p0 ) + (z0 ; p0 ): (41)
@p dp @z dp @p
Moreover, recalling the relation (i ; p0 ) ≈ Lx i (z0 ; 0 ; 0 ; p0 ) in (30) the sensitivity di erentials of
the adjoint variables for the unperturbed optimal control problem OC(p0 ) can be recovered from
relation (35) as (@=@p)(i ; p0 ) ≈ (di =dp)[p0 ], i = 1; : : : ; N; with
 
dz
 i  (p0 ) 
 dp
i i i
d @ @  @
[p0 ] = (z0 ; 0 ; p0 ); (z0 ; 0 ; p0 )  + (z0 ; 0 ; p0 ): (42)
dp @z @  d  @p
(p0 )
dp
96 C. Buskens, H. Maurer / Journal of Computational and Applied Mathematics 120 (2000) 85–108

Considering special initial value perturbations in OC(p) of the type

’(p) = x1 + p; (43)
formula (36) and the multiplier 1 from (30) yield
dF
[p0 ] = 1 : (44)
dp
This formula constitutes the well-known marginal interpretation of the adjoint variable (cf. [7]). If
the perturbation cannot be identi ed with a model parameter, one can measure the deviation of the
perturbed trajectory x(t; p) from the nominal solution x0 (t). This type of problem can be retraced to
problems with initial value perturbations described in Eq. (43).

7. Examples

We shall discuss two numerical examples to illustrate the numerical performance of the algorithms
presented in Sections 3– 6. These algorithms have already been applied to solve problems, e.g., from
electrical engineering, aeronautics, robotics [7–9], production planning [26] and other benchmark
problems. We do not want to dwell on these problems again and instead report on two new examples.
All numerical computations were performed on a 200 MHz Sun Ultra Sparc.

7.1. Time optimal low-thrust planar transfer to a geosynchronous orbit

We consider a time optimal low-thrust trajectory problem (see Fig. 1), where a satellite is trans-
ferred from a circular low-Earth orbit to a geosynchronous orbit (cf. [18]). It is not the purpose of
this section to describe a real-life application with precise parameters for the system. We study the
following simpli ed control problem for a two-dimensional point mass equations of motion in an
inverse square gravity eld:

Fig. 1. A time optimal transfer problem.

 C. Buskens, H. Maurer / Journal of Computational and Applied Mathematics 120 (2000) 85–108 97

Table 1
Convergence behavior of the satellite problem

N CPU F(z; p0 ) = tf ERR

(s)

25 0.17 17:12761030 1:09 × 10−4

26 0.06 17:12761030 1:09 × 10−4
30 0.06 17:12595432 1:25 × 10−5
49 0.68 17:12577446 2:03 × 10−6
63 0.78 17:12574846 5:15 × 10−7
103 1.69 17:12574164 1:85 × 10−7
143 4.20 17:12573995 1:86 × 10−8
208 9.46 17:12573973 5:68 × 10−9

Find a thrust direction control u(t), 06t6tf , that minimizes the nal time F(x; u) = tf subject to

ẋ1 = x2 ; x1 (0) = 6:0; x1 (tf ) = 6:6;

x32 r
ẋ2 = − + a(t; p) sin(u); x2 (0) = 0:0; x2 (tf ) = 0:0;
x1 x12
s s
x2 x3 r r (45)
ẋ3 = − + a(t; p) cos(u); x3 (0) = ; x3 (tf ) = ;
x1 x1 (t0 ) x1 (tf )
x3
ẋ4 = ; x4 (0) = 0:0;
x1
where x1 (t) represents the radial position, x2 (t) the radial velocity, x3 (t) the circumferential velocity
and x4 (t) the polar angle, with r = 62:5 representing the gravitational parameter for the earth. In
the original model, a(t; p) is a function of time describing the thrust acceleration of the spacecraft
and depending on the reduced mass. For simplicity we consider the function a(t; p) = 0:01 + p with
perturbation parameter p to illustrate sensitivity analysis for this problem. The nominal parameter is
p0 = 0. The free nal time tf is treated as an additional optimization variable in problem DOC2 (p0 ).
Table 1 shows the convergence behavior of the objective value with a fourth-order Runge–Kutta
approximation for the state and a linear interpolation of the control variables using the initial esti-
mates u(i ; p0 ) = 0, i = 1; : : : ; N . Note that the routine NUDOCCCS in [6] works with an automatic
grid-re nement algorithm.
In Table 1, N is the number of gridpoints, CPU is the computing time in seconds, and ERR
denotes the relative error of the solution. The number of optimization variables is Nz = N + 1 due to
the free nal time. A numerical check of the SSC described in Section 5.1 shows that the projected
Hessian in (33) is positive de nite with min = 0:8h as the smallest eigenvalue; the chosen stepsize
1
is h = 1=(N − 1) = 200 .
Eqs. (35) and (36) o er a possibility to calculate the sensitivity di erential for the objective func-
tion since the nal time tf is an additional optimization variable. We obtain the value (dF=dp)[p0 ] =
(dtf =dp)(p0 ) = −1990:06 which allows for a rst-order Taylor expansion according to Eq. (37). For
the small perturbation p = −1:0 × 10−3 this gives the approximation F(z(p); p) ≈ 19:1157 with a
relative error of 1:05 × 10−2 compared to the exact solution F(z(p); p) = 18:9173. Better results are
98 C. Buskens, H. Maurer / Journal of Computational and Applied Mathematics 120 (2000) 85–108

Fig. 2. (xi (t; p0 ); (@xi =@p)(t; p0 )), i = 1; : : : ; 4, for the satellite problem.

obtained from the second-order Taylor expansion (39) which yields (d 2 F=dp2 )[p0 ] = −597445:94
and F(z(p); p) ≈ 18:8171 with a relative error of 5:30 × 10−3 . This value is used for the time
interval of the approximated solution in Fig. 3.
The nominal optimal states xi (t; p0 ), i = 1; : : : ; 4, combined with the corresponding sensitivity
di erentials (@x=@p)(t; p0 ) as calculated from Eq. (41) are shown in Fig. 2 on the normalized time
interval [0; 1]. The extreme sensitivity of this problem is demonstrated by the radial velocity x2 (t; p).
Fig. 3 displays the unperturbed solution, the exact perturbed solution and the approximation of
the perturbed trajectory evaluated from Eqs. (8) and (41). The radial velocity exhibits a completely
di erent structure, which is predicted by the rst Taylor approximation. In particular, the unperturbed
optimal control u(t) and its sensitivity di erential (calculated from equation (35)) in Fig. 4 show
a highly nonlinear behavior with large values. This is responsible for the large perturbations of the
radial velocity x2 (t; p).
 C. Buskens, H. Maurer / Journal of Computational and Applied Mathematics 120 (2000) 85–108 99

Fig. 3. Sensitivity of the satellite problem: Radial velocity x2 (t; p).

Fig. 4. (u(t; p0 ); (@u=@p)(t; p0 )) for the satellite problem.

Finally, Fig. 5 depicts the adjoint variables i (t; p0 ), i = 1; : : : ; 4, which are evaluated from Eq.
(30), and the sensitivity di erentials (@=@p)(t; p0 ) as computed from Eq. (42).

7.2. An underwater vehicle problem

As a second example a demanding problem from aquanautics is discussed to indicate that direct
methods are capable of solving highdimensional and highly nonlinear optimal control problems. A
model for the control of an underwater vehicle was recently given by Michalska and Rehman [27].
It is based on a formulation of Nakamuira and Savant [28] and involves six state variables and four
control functions. This model will be modi ed to include an underwater ditch causing an additional
cold water current in horizontal x1 , respectively, in vertical x3 direction (depth) (cf. Fig. 6). The
new control system comprises 10 state and four control variables:
ẋ1 = cos(x6 ) cos(x5 )x7 + rx ;
ẋ2 = sin(x6 ) cos(x5 )x7 ;
100 C. Buskens, H. Maurer / Journal of Computational and Applied Mathematics 120 (2000) 85–108

Fig. 5. (i (t; p0 ); (@i =@p)(t; p0 )); i = 1; : : : ; 4, for the satellite problem.

ẋ3 = −sin(x5 )x7 + rz ;

ẋ4 = x8 + sin(x4 ) tan(x5 )x9 + cos(x4 ) tan(x5 )x10 ;
ẋ5 = cos(x4 )x9 − sin(x4 )x10 ;
sin(x4 ) cos(x4 )
ẋ6 = x9 + x10 ;
cos(x5 ) cos(x5 )
ẋ7 = u1 ;
ẋ8 = u2 ;
ẋ9 = u3 ;
ẋ10 = u4 : (46)
 C. Buskens, H. Maurer / Journal of Computational and Applied Mathematics 120 (2000) 85–108 101

Fig. 6. The underwater vehicle problem.

Table 2
Constants for the underwater vehicle problem

Constants Meaning

uxmax = 2 Maximal horizontal current

uzmax = 1 Maximal vertical current
cx = 0:5 Center of current (center of underwater ditch)
cz = 0:1 Depth of zero current
rx = 0:1 Factor for the expansion of the current (ditch)

The variables x1 –x3 specify the position of the center of the mass, while x4 –x6 describe the orientation
of the vehicle by Euler angles. When the angles are small, x4 corresponds to the roll motion, while
x5 and x6 correspond to the pitch and yaw motions, respectively. It is assumed that the vehicle is
moving with velocity x7 and angular velocities x8 –x10 . Hence, the control function u1 represents the
acceleration of the underwater vehicle, while u2 –u4 describe the angular accelerations. The nonlinear
current is modeled by
 
−((x1 −cx )=rx )2 x3 − c z 2
rx = −uxmax e (x1 − cx ) ;
cz
 2
−((x1 −cx )=rx )2 x3 − cz
rz = −(uzmax + p)e (47)
cz
with constants given in Table 2 and a parameter p representing perturbations in the vertical current.
The nominal solution will be calculated for the parameter p0 = 0.
We wish to determine the minimum energy control u(t); t ∈ [0; 1], that minimizes the energy
functional
Z 1 4
X
F(x; u) = ui (t)2 dt (48)
0 i=1
102 C. Buskens, H. Maurer / Journal of Computational and Applied Mathematics 120 (2000) 85–108

Fig. 7. The optimal trajectory over ground.

subject to the dynamic equations (46) and the 20 initial and terminal conditions:

x1 (0) = 0; x2 (0) = 0; x3 (0) = 0:2; x4 (0) = ; x5 (0) = 0:1;
 2
x6 (0) = − ; x7 (0) = 1; x8 (0) = 0; x9 (0) = 0:5; x10 (0) = 0:1;
4 (49)

x1 (1) = 1; x2 (1) = 0:5; x3 (1) = 0; x4 (1) = ; x5 (1) = 0;
2
x6 (1) = 0; x7 (1) = 0; x8 (1) = 0; x9 (1) = 0; x10 (1) = 0;
and eight additional control constraints
− 156ui (t)615; t ∈ [0; 1]; i = 1; : : : ; 4; (50)
and two further state constraints of order 2 for the velocity of the underwater vehicle:

− 0:026x4 (t) − 60:02; t ∈ [0; 1]: (51)
2
For the de nition of the order of state constraints, compare Hartl et al. [18]. The optimal solution
of this complex problem was calculated from the NLP-problem DOC2 (p) in (21) – (23) using a
fourth-order Runge–Kutta approximation for the state and a linear interpolation of the controls. The
computations are performed with the subroutine NUDOCCCS in [5,6] starting from the initial esti-
mates uj (i ; p0 ) = 0; i = 1; : : : ; N; j = 1; : : : ; 4, in combination with an algorithm for the computation of
admissible initial estimates. Using N =201 gridpoints we obtain a problem with Nz =804 optimization
variables and M = 2020 constraints. The objective function was calculated as F(x0 ; u0 ) = 226:8086.
Fig. 7 shows the unperturbed optimal trajectory (x1 (t); x2 (t); x3 (t)) as a three-dimensional plot.
A numerical check of SSC yields the positive de niteness of the projected Hessian in (33) with
1
stepsize h = 1=(N − 1) = 200 and min = 0:48h as the smallest eigenvalue.
C. Buskens, H. Maurer / Journal of Computational and Applied Mathematics 120 (2000) 85–108 103

Fig. 8. (xi (t; p0 ); (@x=@p)(t; p0 )); i = 1; : : : ; 10, for the underwater vehicle problem.
104 C. Buskens, H. Maurer / Journal of Computational and Applied Mathematics 120 (2000) 85–108

Fig. 9. (ui (t; p0 ); (@u=@p)(t; p0 )); i = 1; : : : ; 10, for the underwater vehicle problem.

The nominal optimal states xi (t; p0 ), i = 1; : : : ; 10, combined with the corresponding sensitivity
di erentials (@x=@p)(t; p0 ) calculated from Eq. (41) are pictured in Fig. 8. The interesting structure
of the problem results from the existence of the additional current. Note that the underwater vehicle
is driving backward for a short interval; compare the velocity x7 (t).
Fig. 9 displays the unperturbed controls ui (t; p0 ), i = 1; : : : ; 4, and the corresponding sensitivity
di erentials (@u=@p)(t; p0 ) which are calculated from Eq. (35). We nd one boundary arc for u1
and three boundary arcs for control u3 while the pertaining sensitivity di erentials are equal to zero
on these arcs. The peaks occurring in the sensitivities result from the linear interpolation and the
non-smoothness of the controls.
Finally, Fig. 10 shows the adjoint variables i (t; p0 ); i = 1; : : : ; 10, as evaluated from Eq. (30) and
the sensitivity di erentials (@=@p)(t; p0 ) which are obtained from Eq. (42). The state constraints (51)
 C. Buskens, H. Maurer / Journal of Computational and Applied Mathematics 120 (2000) 85–108
105

Fig. 10. (i (t; p0 ); (@=@p)(t; p0 )); i = 1; : : : ; 10, for the underwater vehicle problem.
106 C. Buskens, H. Maurer / Journal of Computational and Applied Mathematics 120 (2000) 85–108

Table 3
CPU-time and error for di erent control approximations

p = 0:1 p = 0:4

ERR CPU (s) ERR CPU (s)

ROPT 2:8 × 10−12 39.0 2:6 × 10−12 42.0

SOPT 3:2 × 10−4 1:07 × 10−5 6:0 × 10−3 1:07 × 10−5
WOPT 1:3 × 10−2 — 1:1 × 10−1 —

become active at a contact point and cause a jump in the adjoint variable 4 (t); compare equation
(7). Since the adjoint variables depend on the Lagrange multipliers of the resulting NLP-problem, a
jump occurs also in its sensitivity di erentials (42).
As a nal aspect we study the real-time approach in Section 2.2. Eq. (8) is used to get a rst-order
Taylor expansion for the perturbed controls. Table 3 compares the computing time and the maximal
error in the terminal conditions for di erent control approximations and allows to judge the quality
of the real-time approximation (8).
The notations in Table 3 have the following meaning: ERR denotes the maximal errors in the
terminal states; CPU is the CPU-time in seconds; ROPT means a reoptimization of the perturbed
problem starting from the nominal solution; SOPT is the real-time approximation (8) for the controls
ui (j ; p); i = 1; : : : ; 4; j = 1; : : : ; 201, while the error in the terminal state results from the integration
of the perturbed system with these approximations; WOPT is obtained from an integration of the
perturbed system with the nominal controls.
If the time during the motion of the underwater vehicle is used for computing the needed approx-
imations, the computing time for SOPT can be reduced drastically by an additional factor of 201
(number of gridpoints) to 5:32 × 10−8 s. The two di erent parameters represent a perturbation of
10%, respectively 40% in relation to the maximal vertical current uzmax and cause an error of more
than 11% in x1 (tf ; p) for p = 0:4. Using the real-time approximation this error can be reduced to
0:6%. Note that during our investigations we found several local minima with di erent objective
values. Hence, the speci ed solutions may not represent the global minima.

References

[1] A. Barclay, Ph.E. Gill, J.B. Rosen, SQP methods and their application to numerical optimal control, Report NA
97-3, Department of Mathematics, University of California, San Diego, USA, 1997.
[2] J.T. Betts, W.P. Hu mann, Path constrained trajectory optimization using sparse sequential quadratic programming,
J. Guidance Control Dynamics 16 (1993) 59–68.
[3] J.T. Betts, Issues in the direct transcription of optimal control problems to sparse nonlinear programs, in: R. Bulirsch,
D. Kraft (Eds.), International Series of Numerical Mathematics, ISNM, Vol. 115, Birkhauser, Basel, 1994, pp. 3–18.
[4] H.G. Bock, K.J. Plitt, A multiple shooting algorithm for direct solution of optimal control problems, IFAC nineth
World Congress, Budapest, Hungary, 1984.
[5] Ch. Buskens, Direkte Optimierungsmethoden zur numerischen Berechnung optimaler Steuerungen, Diploma Thesis,
Institut fur Numerische Mathematik, Universitat Munster, Munster, Germany, 1993.
 C. Buskens, H. Maurer / Journal of Computational and Applied Mathematics 120 (2000) 85–108 107

[6] Ch. Buskens, Optimierungsmethoden und Sensitivitatsanalyse fur optimale Steuerprozesse mit Steuer- und
Zustands-Beschrankungen, Dissertation, Institut fur Numerische Mathematik, Universitat Munster, Munster, Germany,
1998.
[7] Ch. Buskens, H. Maurer, Real-time control of robots with initial value perturbations via nonlinear programming
methods, Special Issue of OPTIMIZATION, in press.
[8] Ch. Buskens, H. Maurer, Sensitivity analysis and real-time control of nonlinear optimal control systems via nonlinear
programming methods, in: International Series of Numerical Mathematics, Vol. 124, Birkhauser, Basel, 1998.
[9] Ch. Buskens, H. Maurer, Open-loop real-time control of an industrial robot using nonlinear programming methods,
Automatica (1999), to appear.
[10] Yu.G. Evtushenko, Numerical Optimization Techniques, Translation Series in Mathematics and Engineering,
Optimisation Software Inc. (Publications Division), New York, 1985.
[11] P.J. Enright, B.A. Conway, Discrete approximations to optimal trajectories using direct transcription and nonlinear
programming, AIAA Paper 90-2963-CP, 1990.
[12] U. Felgenhauer, Diskretisierung von Steuerungsproblemen unter stabilen Optimalitatsbedingungen, Habilitations-
schrift, Technische Universitat Cottbus, Institut fur Mathematik, Cottbus, Germany, 1998.
[13] U. Felgenhauer, On higher order methods for control problems with mixed inequality constraints, Preprint M-01=1998,
Technische Universitat Cottbus, Institut fur Mathematik, Cottbus, Germany, 1998.
[14] A.V. Fiacco, Introduction to Sensitivity and Stability Analysis in Nonlinear Programming, Mathematics in Science
and Engineering, Vol. 165, Academic Press, New York, 1983.
[15] P.E. Gill, W. Murray, M.A. Saunders, M.H. Wright, Model building and practical aspects of nonlinear programming,
in: K. Schittkowski (Ed.), Computational Mathematical Programming, Springer, Berlin, 1985, pp. 209–247.
[16] W. Grimm, Convergence relations between optimal control and optimal para-metric control, Schwerpunktprogramm
der Deutschen Forschungsgemeinschaft: Anwendungsbezogene Optimierung und Steuerung, Report No. 420, Institut
fur Flugmechanik und Flugregelung, Universitat Stuttgart, Stuttgart, Germany, 1993.
[17] R.F. Hartl, S.P. Sethi, R.G. Vickson, A survey of the maximum principles for optimal control problems with state
constraints, SIAM Rev. 37 (1995) 181–218.
[18] C.A. Kluever, Optimal feedback guidance for low-thrust orbit insertion, Opt. Control Appl. Methods 16 (1995)
155–173.
[19] B. Kugelmann, H.J. Pesch, A new general guidance method in constrained optimal control, Part 1: Numerical method,
J. Optim. Theory Appl. 67 (1990) 421–435.
[20] B. Kugelmann, H.J. Pesch, A new general guidance method in constrained optimal control, Part 2: application to
space shuttle guidance, J. Optim. Theory Appl. 67 (1990) 437–446.
[21] K. Malanowski, H. Maurer, Sensitivity analysis for parametric control problems with control-state constraints,
Comput. Optim. Appl. 5 (1996) 253–283.
[22] K. Malanowski, H. Maurer, Sensitivity analysis for state constrained optimal control problems, Discrete Continuous
Dynamical Systems 4 (1998) 241–272.
[23] K. Malanowski, Ch. Buskens, H. Maurer, Convergence of approximations to nonlinear optimal control problems,
in: A.V. Fiacco (Ed.), Mathematical Programming with Data Perturbations, Lecture Notes in Pure and Applied
Mathematics, Vol. 195, Marcel Dekker, New York, 1997, pp. 253–284.
[24] H. Maurer, H.J. Pesch, Solution di erentiability for parametric nonlinear control problems, SIAM J. Control Optim.
32 (1994) 1542–1554.
[25] H. Maurer, H.J. Pesch, Solution di erentiability for parametric nonlinear control problems with control-state
constraints, J. Optim. Theory Appl. 23 (1995) 285–309.
[26] H. Maurer, Ch. Buskens, G. Feichtinger, Solution techniques for periodic control problems: a case study in production
planning, Opt. Control Appl. Methods 19 (1998) 185–203.
[27] H. Michalska, F. Rehman, Guiding functions and discontinuous control: the underwater vehicle example, Internat. J.
Control 69 (1998) 1–30.
[28] Y. Nakamuira, S. Savant, Nonlinear tracking control of autonomous underwater vehicles, Proceedings of the 1992
IEEE International Conference on Robotics and Automation, Nice, France, 1992, pp. A4-A9.
[29] L.W. Neustadt, Optimization: A Theory of Necessary Conditions, Princeton University Press, Princeton, NJ, 1976.
[30] H.J. Pesch, Real-time computation of feedback controls for constrained optimal control problems, Part 1: neighbouring
extremals, Opt. Control Appl. Methods 10 (1989) 129–145.
108 C. Buskens, H. Maurer / Journal of Computational and Applied Mathematics 120 (2000) 85–108

[31] H.J. Pesch, Real-time computation of feedback controls for constrained optimal control problems, Part 2: a correction
method based on multiple shooting, Opt. Control Appl. Methods 10 (1989) 147–171.
[32] L.S. Pontryagin, Mathematische Theorie optimaler Prozesse, Oldenbourg Verlag, Munchen-Wien, 1967.
[33] O. von Stryk, Numerische Losung optimaler Steuerungsprobleme: Diskretisierung, Parameteroptimierung und
Berechnung der adjungierten Variablen, Fortschritt-Berichte VDI, Reihe 8, Nr. 441, VDI Verlag, Germany, 1995.
[34] K.L. Teo, C.J. Goh, K.H. Wong, A Uni ed Computational Approach to Optimal Control Problems, Longman
Scienti c and Technical, New York, 1991.