Professional Documents
Culture Documents
Stochastic
Optimization
Methods
Applications in Engineering and
Operations Research
Third Edition
Stochastic Optimization Methods
Kurt Marti
Stochastic Optimization
Methods
Applications in Engineering
and Operations Research
Third edition
123
Kurt Marti
Institute for Mathematics and Computer
Science
Federal Armed Forces University Munich
Neubiberg/Munich
Germany
v
vi Preface
of the book into the Springer-program. I would like to thank especially the Senior
Editor for Business/Economics of Springer-Verlag, Christian Rauscher, for his very
long patience until the completion of this book.
ix
x Preface to the First Edition
The major change in the second edition of this book is the inclusion of a new,
extended version of Chap. 7 on the approximate computation of probabilities of
survival and failure of technical, economic structures and systems. Based on
a representation of the state function of the structure/system by the minimum
value of a certain optimization problem, approximations of the probability of
survival/failure are obtained by means of polyhedral approximation of the so-
called survival domain, certain probability inequalities and discretizations of the
underlying probability distribution of the random model parameters. All the other
chapters have been updated by including some more recent material and by
correcting some typing errors.
The author thanks again Ms. Elisabeth Lößl for the excellent LaTeX-typesetting
of all revisions and completions of the text. Moreover, the author thanks Dr. Müller,
Vice President Business/Economics & Statistics, Springer-Verlag, for the possibility
to publish a second edition of “Stochastic Optimization Methods.”
xiii
Outline of the 3rd Edition
xv
xvi Outline of the 3rd Edition
optimal open-loop controls are found. For approximate solutions of the stochastic
two-point boundary problem, cf. [112].
Stochastic optimal regulators or optimal feedback control under stochastic
uncertainty can be obtained in general by approximative methods only. In Chap. 3,
stochastic optimal feedback controls are determined by means of the fundamental
stochastic open-loop feedback method. This very efficient approximation method is
also the basis of model predictive control procedures. Using the methods described
in Chap. 2, stochastic optimal open-loop feedback controls are constructed by
computing next to stochastic optimal open-loop controls on the remaining time
intervals tb t tf with t0 tb tf . Having a stochastic optimal open-loop
feedback control on each remaining time interval tb t tf with t0 tb tf ,
a stochastic optimal open-loop feedback control law follows then immediately
evaluating each of the stochastic optimal open-loop controls on tb t tf at
the corresponding initial time point t D tb . The efficiency of this method has
been proved already by applications to the stochastic optimization of regulators for
robots.
Adaptive Optimal Stochastic Trajectory Planning and Control .AOSTPC/ are
considered in Chap. 4: in optimal control of dynamic systems, the standard proce-
dure is to determine first off-line an optimal open-loop control, using some nominal
or estimated values of the model parameters, and to correct then the resulting
deviation of the actual trajectory or system performance from the prescribed
trajectory (prescribed system performance) by on-line measurement and control
actions. However, on-line measurement and control actions are very expensive and
time-consuming. By adaptive optimal stochastic trajectory planning and control
(AOSTPC), based on stochastic optimization methods, the available a priori and
statistical information about the unknown model parameters is incorporating into
the optimal control design. Consequently, the mean absolute deviation between the
actual and prescribed trajectory can be reduced considerably, and robust controls are
obtained. Using only some necessary stability conditions, by means of stochastic
optimization methods also sufficient stability properties of the corresponding feed-
forward, feedback (PD-, PID-)controls, resp., are obtained. Moreover, analytical
estimates are given for the reduction of the tracking error, hence, for the reduction
of the on-line correction expenses by applying (AOSTPC).
Special methods for the optimal design of regulators are described in Chap. 5:
the optimal design of regulators is often based on given, fixed nominal values of
initial conditions, external loads, and other model parameters. However, due to the
variations of material properties, measurement and observational errors, modeling
errors, uncertainty of the working environment, the task to be executed, etc., the
true external loadings, initial conditions, and model parameters are not known
exactly in practice. Hence, a predetermined optimal regulator should be “robust,”
i.e., the regulator should guarantee satisfying results also in case of observational
errors and variations of the different parameters occurring in the model. Robust
regulators have been considered mainly for uncertainty models based on given fixed
sets of parameters, especially certain multiple intervals containing the unknown,
true parameter. However, since in many cases observational errors and parameter
Outline of the 3rd Edition xvii
problems are discussed, e.g., (i) the convexity of the problems and (ii) the “dual
block angular” structure of the resulting (large-scale) deterministic linear program
in case of discrete probability distributions. In this case, several special purpose LP-
solvers are available. In case of a continuous parameter distribution, the resulting
substitute problems can be solved approximatively by means of a discretization of
the underlying probability distribution and using then the SLP-method as described
above.
Structural Analysis and Optimal Structural Design under Stochastic Uncertainty
using Quadratic Cost Functions are treated in Chap. 7: problems from plastic
analysis and optimal plastic design are based on the convex, linear, or linearized
yield/strength condition and the linear equilibrium equation for the stress (state)
vector. In practice, one has to take into account stochastic variations of the vector
a D a.!/ of model parameters (e.g., yield stresses, plastic capacities, external
load factors, cost factors, etc.). Hence, in order to get robust optimal load factors
x, robust optimal designs x, resp., the basic plastic analysis or optimal plastic
design problem with random parameters has to be replaced by an appropriate
deterministic substitute problem. As a basic tool in the analysis and optimal design
of mechanical structures under uncertainty, a state function s D s .a; x/ of the
underlying structure is introduced. The survival of the structure can be described
then by the condition s 0. Interpreting the state function s as a cost function,
several relations s to other cost functions, especially quadratic cost functions, are
derived. Bounds for the probability of survival ps are obtained then by means
of the Tschebyscheff inequality. In order to obtain robust optimal decisions x ,
i.e., maximum load factors, optimal designs insensitive with respect to variations
of the model parameters a D a.!/, a direct approach is presented then based
on the primary costs (weight, volume, costs of construction, costs for missing
carrying capacity, etc.) and the recourse costs (e.g., costs for repair, compensation
for weakness within the structure, damage, failure, etc.), where the abovementioned
quadratic cost criterion is used. The minimum recourse costs can be determined
then by solving an optimization problem having a quadratic objective function
and linear constraints. For each vector a D a.!/ of model parameters and each
design vector x, one obtains then an explicit representation of the “best” internal
load distribution F . Moreover, also the expected recourse costs can be determined
explicitly. The expected recourse function may be represented by means of a
“generalized stiffness matrix.” Hence, corresponding to an elastic approach, the
expected recourse function can be interpreted here as a “generalized expected
compliance function”, which depends on a generalized “stiffness matrix.” Based
on the minimization of the expected primary costs subject to constraints for the
expected recourse costs (“generalized compliance”) or the minimization of the
expected total primary and recourse costs, explicit finite dimensional parameter
optimization problems are achieved for finding robust optimal design x or a
maximal load factor x . The analytical properties of the resulting programming
problem are discussed, and applications, such as limit load/shakedown analysis,
are considered. Furthermore, based on the expected “compliance function,” explicit
Outline of the 3rd Edition xix
upper and lower bounds for the probability ps of survival can be derived using
linearization methods.
Finally, in Chap. 8 the inference and decision strategies applied in stochastic
optimization methods are considered in more detail:
A large number of optimization problems arising in engineering, control, and
economics can be described by the minimization of a certain (cost) function
v D v.a; x/ depending on a random parameter vector a D a.!/ and a decision
vector x 2 D lying in a given set D of feasible decision, design, or control
variables. Hence, in order to get robust optimal decisions, i.e., optimal decisions
being most insensitive with respect to variations of the random parameter vector
a D a.!/, the original optimization problem is replaced by the deterministic
substitute problem which consists in the minimization of the expected objective
function Ev D Ev.a.!/; x/ subject to x 2 D. Since the true probability distribution
of a D a.!/ is not exactly known in practice, one has to replace by a certain
estimate or guess ˇ. Consequently, one has the following inference and decision
problem:
• inference/estimation step
Determine an estimation ˇ of the true probability distribution of a D a.!/,
• decision step R
determine an optimal solution x of min v.a.!/; x/ˇ.d!/ s.t. x 2 D.
Computing approximation, estimation ˇ of , the criterion for judging an approxi-
mation ˇ of should be based on its utility for the decision-making process, i.e., one
should weight the approximation error according to its influence on decision errors,
and the decision errors should be weighted in turn according to the loss caused by
an incorrect decision.
Based on inferential ideas developed among others by Kerridge, Kullback, in
this chapter generalized decision-oriented inaccuracy and divergence functions for
probability distributions , ˇ are developed, taking into account that the outcome ˇ
of the inferential stage is used in a subsequent (ultimate) decision-making problem
modeled by the above-mentioned stochastic optimization problem. In addition,
stability properties of the inference and decision process
! ˇ ! x 2 D .ˇ/;
are studied, where D .ˇ/ denotes the set of -optimal decisions with respect to
probability distribution Pa./ D ˇ of the random parameter vector a D a.!/.
Contents
xxi
xxii Contents
References .. .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 357
Index . . . . . . . . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 365
Chapter 1
Stochastic Optimization Methods
1.1 Introduction
Here, the objective (goal) function f0 D f0 .a; x/ and the constraint functions
fi D fi .a; x/; i D 1; : : : ; mf and gi D gi .a; x/; i D 1; : : : ; mg , defined
on a joint subset of R Rr , depend on a decision, design, control or input
vector x D .x1 ; x2 ; : : : ; xr /T and a vector a D .a1 ; a2 ; : : : ; a /T of model
parameters. Typical model parameters in technical applications, operations research
and economics are material parameters, external load parameters, cost factors,
technological parameters in input–output operators, demand factors. Furthermore,
manufacturing and modeling errors, disturbances or noise factors, etc., may occur.
Frequent decision, control or input variables are material, topological, geometrical
and cross-sectional design variables in structural optimization [77], forces and
moments in optimal control of dynamic systems and factors of production in
operations research and economic design.
The objective function (1.1a) to be optimized describes the aim, the goal of
the modeled optimal decision/design problem or the performance of a technical,
economic system or process to be controlled optimally. Furthermore, the constraints
xL x xU (1.1d’)
Due to the deviation of the actual parameter vector a from the nominal vector
a0 of model parameters, deviations of the actual state, trajectory or performance
of the system from the prescribed state, trajectory, goal values occur.
1.2 Deterministic Substitute Problems: Basic Formulation 3
a D a.!/; ! 2 ˝; (1.3)
of a random -vector a.!/ on a certain probability space .˝; A0 ; P/, where the
probability distribution Pa./ of a.!/ is known, or it is known that Pa./ lies within
a given range W of probability measures on R . Using a Bayesian approach, the
probability distribution Pa./ of a.!/ may also describe the subjective or personal
probability of the decision maker, the designer.
Hence, in order to take into account the stochastic variations of the parameter
vector a, to incorporate the a priori and/or sample information about the unknown
vector a, resp., the standard approach “insert a certain nominal parameter vector
a0 , and correct then the resulting error” must be replaced by a more appropriate
deterministic substitute problem for the basic optimization problem (1.1a–d) under
stochastic uncertainty.
a 2 R ; x 2 Rr .x 2 D0 /;
(1.4a)
4 1 Stochastic Optimization Methods
i W E ! R; i D 0; 1; : : : ; m; (1.4b)
with
an integer
m 0. For the processing of the numerical outcomes
i e.a; x/ ; i D 0; 1; : : : ; m, there are two basic concepts:
where the (conditional) expectation “E” is taken with respect to the time history
A D At ; .Aj / A0 up to a certain time point t or stage j . A short definition of
expectations is given in Sect. 2.1, for more details, see e.g. [13, 47, 135].
Having different expected cost or performance functions F0 ; F1 ; : : : ; Fm to be
minimized or bounded, as a basic deterministic substitute problem for (1.1a–d) with
a random parameter vector a D a.!/ we may consider the multi-objective expected
cost minimization problem
s.t.
x 2 D0 : (1.6b)
Fi .x/ Fi .x 0 /; i D 0; 1; : : : ; m (1.7a)
1.2 Deterministic Substitute Problems: Basic Formulation 5
and
s.t.
X
m
FQ .x/ WD ci Fi .x/ D Ef a.!/; x ; (1.10a)
i D0
where
X
m
f .a; x/ WD ci i e.a; x/ : (1.10b)
i D0
6 1 Stochastic Optimization Methods
X
m
min ci Fi .x/ (1.11a)
i D0
s.t.
x 2 D0 : (1.11b)
s.t.
x 2 D0 : (1.13b)
Instead of taking expectations, we may consider the worst case with respect to
the cost variations caused by the random parameter vector a D a.!/. Hence, the
random cost function
! ! i e a.!/; x (1.14a)
1.3 Optimal Decision/Design Problems with Random Parameters 7
is evaluated by means of
sup
Fi .x/ WD ess sup i e a.!/; x ; i D 0; 1; : : : ; m: (1.14b)
Here, ess sup .: : :/ denotes the (conditional) essential supremum with respect to the
random vector a D a.!/, given information A, i.e., the infimum of the supremum
of (1.14a) on sets A 2 A0 of (conditional) probability one, see e.g. [135].
Consequently, the vector function F D Fsup .x/ is then defined by
0 1
0 1
ess sup 0 e a.!/; x
F0 .x/ B C
B F1 .x/ C B C
B C B ess sup 1 e a.!/; x C
F .x/ D B : C WD B
sup B C: (1.15)
@ :: A :: C
B : C
@ A
Fm .x/ ess sup m e a.!/; x
Working with the vector function F D Fsup .x/, we have then the vector minimiza-
tion problem
s.t.
x 2 D0 : (1.16b)
Important examples are the total weight or volume of a mechanical structure, the
costs of construction, design of a certain technical or economic structure/system,
or the negative utility or reward in a general decision situation. Basic examples in
optimal control, cf. Chap. 2, are the total run time, the total energy consumption of
the process or a weighted mean of these two cost functions.
For the representation of the structural/system safety or failure, for the represen-
tation of the admissibility of the state, or for the formulation of the basic operating
conditions of the underlying plant, structure/system certain state, performance or
response functions
are chosen. In structural design these functions are also called “limit state functions”
or “safety margins”. Frequent examples are some displacement, stress, load (force
and moment) components in structural design, or more general system output
functions in engineering design. Furthermore, production functions and several
cost functions are possible performance functions in production planning problems,
optimal mix problems, transportation problems, allocation problems and other
problems of economic decision.
In (1.17a, b), the design or input vector x denotes the r-vector of design or
input variables, x1 ; x2 ; : : : ; xr , as e.g. structural dimensions, sizing variables, such
as cross-sectional areas, thickness in structural design, or factors of production,
actions in economic decision problems. For the decision, design or input vector x
one has mostly some basic deterministic constraints, e.g. nonnegativity constraints,
box constraints, represented by
x 2 D; (1.17c)
is composed of the following two subvectors: R is the m-vector of the acting exter-
nal loads or structural/system inputs, e.g. wave, wind loads, payload, etc. Moreover,
p denotes the . m/-vector of the further model parameters, as e.g. material
parameters, like strength parameters, yield/allowable stresses, elastic moduli, plastic
capacities, etc., of the members of a mechanical structure, parameters of an electric
circuit, such as resistances, inductances, capacitances, the manufacturing tolerances
and weight or more general cost coefficients.
In linear programming, as e.g. in production planning problems,
a D .A; b; c/ (1.17e)
1.3 Optimal Decision/Design Problems with Random Parameters 9
y.a; x/ 2 B; (1.17g)
where B is a certain subset of Rmy I B D B.a/ may depend also on some model
parameters.
In production planning problems, typical operating conditions are given, cf.
(1.17e), by
yi .a; x/ 0 (1.18c)
s.t.
y.a; x/ 2 B (1.19b)
x 2 D: (1.19c)
s.t.
a D a.!/; ! 2 ˝; (1.21a)
taken with probabilities ˛l ; l D 1; : : : ; l0 . In the second case, the probability that the
realization a.!/ D a lies in a certain (measurable) subset B R is described by
the multiple integral
Z
P a.!/ 2 B D '.a/ da (1.21c)
B
Z
with a certain probability density function ' D '.a/ 0; a 2 R ;
'.a/da D 1.
The properties of the probability distribution Pa./ may be described—fully or
in part—by certain numerical characteristics, called parameters of Pa./ . These
distribution parameters D h are obtained by considering expectations
h WD Eh a.!/ (1.22a)
h W R ! Rsh ; sh 1: (1.22c)
8
ˆ X
l0
ˆ
ˆ
ˆ
< h al ˛l ; in the discrete case (1.21b)
Eh a.!/ D ZlD1 (1.22d)
ˆ
ˆ
ˆ h.a/'.a/ da; in the continuous case (1.21c).
:̂
R
of certain “h-moments” h1 ; : : : ; hs of the type (1.22a). Important examples of the
type (1.22a), (1.23), resp., are the expectation
s.t.
x 2 D: (1.25b)
Here,
pf D pf .x/ WD P y a.!/; x T B (1.25c)
is the probability of failure or the probability that a safe function of the structure,
the system is not guaranteed. Furthermore, cG is a certain weight factor, and cf > 0
describes the failure or recourse costs. In the present definition of expected failure
costs, constant costs for each realization a D a.!/ of a./ are assumed. Obviously,
it is
The objective function (1.25a) may be interpreted as the Lagrangian (with given
cost multiplier cf ) of the following reliability based optimization (RBO) problem,
cf. [7, 92, 132, 147, 163]:
min EG0 a.!/; x (1.26a)
s.t.
where ˛ max > 0 is a prescribed maximum failure probability, e.g. ˛ max D 0:001, cf.
(1.19a–c).
The “dual” version of (1.26a–c) reads
s.t.
EG0 a.!/; x G max (1.27b)
x 2D (1.27c)
W Rm y ! Rm (1.28c)
is a scalar or vector valued cost or loss function (Fig. 1.1). In case B D .0; C1/my
or B D Œ0; C1/my it is often assumed that D .y/ is a non increasing function,
hence,
with a constant 0 > 0. Then for the probability of failure we find the following
upper bound
1
pf .x/ D P y a.!/; x TB E y a.!/; x ; (1.29b)
0
1.4 Deterministic Substitute Problems in Optimal Decision/Design 15
where the right hand side of (1.29b) is obviously an expected cost function of type
(1.28a–c). Hence, the condition (1.26b) can be guaranteed by the expected cost
constraint
E y a.!/; x 0 ˛ max : (1.29c)
Example 1.3 If the loss function .y/ is defined by a vector of individual loss
functions i for each state function yi D yi .a; x/; i D 1; : : : ; my , hence,
T
.y/ D 1 .y1 /; : : : ; my .ymy / ; (1.30a)
then
T
.x/ D 1 .x/; : : : ; my .x/ ; i .x/ WD Ei yi a.!/; x ; 1 i my ;
(1.30b)
s.t.
x2D (1.31b)
II) Expected primary cost minimization under expected failure or recourse cost
constraints
min EG0 a.!/; x (1.32a)
s.t.
III) Expected failure or recourse cost minimization under expected primary cost
constraints
s.t.
EG0 a.!/; x G max (1.33b)
x 2 D: (1.33c)
Here, cG ; cf are (vectorial) weight coefficients, max is the vector of upper loss
bounds, and “min” indicates again that .x/ may be a vector valued function.
or more general
.x/ D Eg a.!/; x ; x 2 D0 : (1.34b)
If x ! y a.!/; x is convex a.s., and is a convex, monotoneous nondecreasing
function, then D .x/ is convex.
It is well known [85] that a convex function is continuous on each open subset of
its domain. A general sufficient condition for the continuity of is given next.
Lemma 1.2 (Continuity) Suppose that Eg a.!/; x exists and is finite for each
x 2 D0 , and assume that x ! g a.!/; x is continuous at x0 2 D0 a.s. If there is
a function D a.!/ having finite expectation such that
ˇ ˇ
ˇ ˇ
ˇg a.!/; x ˇ a.!/ a.s. for all x 2 U.x0 / \ D0 ; (1.35)
Proof Considering the difference quotients ; k D 1; : : : ; r, of at a fixed
xk
point x0 2 D, the assertion follows by means of the mean value theorem, inequality
(1.36a) and Lebesgue’s dominated convergence theorem, cf. [70, 71, 90].
Example 1.4 In case (1.34a), under obvious differentiability
assumptions
concern-
ing and y we have rx g.a; x/ D rx y.a; x/ r y.a; x/ , where rx y.a; x/
T
18 1 Stochastic Optimization Methods
The main problem in solving the deterministic substitute problems defined above
is that the arising probability and expected cost functions pf D pf .x/; D
.x/; x 2 Rr , are defined by means of multiple integrals over a -dimensional
space.
Thus, the substitute problems may be solved, in practice, only by some approx-
imative analytical and numerical methods [44, 70, 90, 100]. In the following we
consider possible approximations for substitute problems based on general expected
recourse cost functions D .x/ according to (1.34a) having a real valued convex
loss function .z/. Note that the probability of failure function pf D pf .x/ may
be approximated from above, see (1.29a, b), by expected cost functions D .x/
having a nonnegative function D .z/ being bounded from below on the failure
domain B c . In the following several basic approximation methods are presented.
which holds for any convex function . Using the mean value theorem, we have
(1.38d)
a) Bounded gradient
If the gradient r is bounded on the range of y D y a.!/; x ; x 2 D, i.e.,
if
r y a.!/; x # max a.s. for each x 2 D; (1.39a)
p
Since t ! t; t 0, is a concave function, we get
p
0 .x/ y.x/ # max q.x/ ; (1.39c)
where
2
q.x/ WD E y a.!/; x y.x/ D trQ.x/ (1.39d)
1
.y/ .y/ D r.y/T .y y/ C .y y/T r 2 .y/.y
Q y/: (1.40a)
2
20 1 Stochastic Optimization Methods
s.t.
y.x/ C c0 q.x/ max (1.41b)
x 2 D; (1.41c)
s.t.
EG0 a.!/; x G max (1.42b)
x 2 D: (1.42c)
The numerical solution is simplified considerably if one can work with one single
state function y D y.a; x/. Formally, this is possible by defining the function
Indeed, according to (1.18b, c) the failure of the structure, the system can be
represented by the condition
Thus, the weakness or failure of the technical or economic device can be evaluated
numerically by the function
.x/ WD E y min a.!/; x (1.43c)
Further approximations of y min .a; x/ and its moments can be found by using the
representation
1
min.a; b/ D a C b ja bj
2
22 1 Stochastic Optimization Methods
1 ˇ ˇ
D min yi .a; x/ C yi C1 .a; x/ ˇyi .a; x/ yi C1 .a; x/ˇ :
i D1;3;:::;my 1 2
for all .a; x/ under consideration with a positive constant A > 0. Thus, defining
and therefore
we have
Hence, the survival/failure of the system or structure can be studied also by means
of the positive function yQ min D yQ min .a; x/. Using now the theory of power or Hölder
means [25], the minimum yQ min .a; x/ of positive functions can be represented also
by the limit
!1=
1 X
my
yQ min
.a; x/ D lim yQi .a; x/
!1 my i D1
!1=
1 X
my
Q WD
of the decreasing family of power means M Œ .y/ yQ ; < 0.
my i D1 i
Consequently, for each fixed p > 0 we also have
!p=
1 X
my
yQ min .a; x/p D lim yQi .a; x/ :
!1 my i D1
1.6 Approximations of Deterministic Substitute Problems in Optimal. . . 23
p
Assuming that the expectation EM Œ yQ a.!/ ; x exists for an exponent D
0 < 0, by means of Lebesgue’s bounded convergence theorem we get the moment
representation
!p=
p 1 X
my
E yQ min a.!/; x D lim E yQi a.!/; x :
!1 my i D1
Since t ! t p= ; t > 0, is convex for each fixed p > 0 and < 0, by Jensen’s
inequality we have the lower moment bound
!p=
p 1 X
my
E yQ min
a.!/; x lim E yQi a.!/; x :
!1 my i D1
Hence, for the pth order moment of yQ min a./; x we get the approximations
!p= !p=
1 X 1 X
my my
E yQi a.!/; x E yQi a.!/; x
my i D1 my i D1
.a; x/ ! yi .a; x/
are bilinear functions. Thus, in this case y min D y min .a; x/ is a piecewise linear
function with respect to a. Fitting a linear or quadratic Response Surface Model
[22, 23, 115, 116]
.j /
with “design” points da 2 R ; j D 1; : : : ; p, the unknown coefficients c D
c.x/; q D q.x/ and Q D Q.x/ are obtained by minimizing the mean square error
X
p
2
.c; q; Q/ WD Q .j / ; x/ y min .a.j / ; x/
y.a (1.43h)
j D1
24 1 Stochastic Optimization Methods
with respect to .c; q; Q/ (Fig. 1.2). Since the model (1.43f) depends linearly on the
function parameters .c; q; Q/, explicit formulas for the optimal coefficients
As can be seen above, cf. (1.34a, b), in the objective and/or in the constraints
of substitute problems for optimization problems with random data mean value
functions of the type
1.6 Approximations of Deterministic Substitute Problems in Optimal. . . 25
.x/ WD Eg a.!/; x
Suppose that on its domain the function g D g.a; x/ has partial derivatives
ral g.a; x/; l D 0; 1; : : : ; lg C 1, up to order lg C 1. Note that the gradient ra g.a; x/
@g
contains the so–called sensitivities .a; x/; j D 1; : : : ; , of g with respect to the
@aj
parameter vector a at .a; x/. In the same way, the higher order partial derivatives
ral g.a; x/; l > 1, represent the higher order sensitivities of g with respect to a at
.a; x/. Taylor expansion of g D g.a; x/ with respect to a at a WD Ea.!/ yields
Xlg
1 l 1
g.a; x/ D N x/ .a a/
r g.a; N lC N lg C1
O x/ .a a/
r lgC1 g.a;
lŠ a .lg C 1/Š a
lD0
(1.44a)
where aO WD aN C #.a a/;
N 0 < # < 1, and .a a/
N l denotes the system of l-th order
products
Y
.aj aN j /lj
j D1
see (1.34a), then the partial derivatives ral g of g up to the second order read:
T
ra g.a; x/ D ra y.a; x/ r y.a; x/ (1.44b)
T
ra2 g.a; x/ D ra y.a; x/ r 2 y.a; x/ ra y.a; x/ (1.44c)
Cr y.a; x/ ra2 y.a; x/;
26 1 Stochastic Optimization Methods
where
2
T @ y
.r / ra2 y WD .r / (1.44d)
@ak @al k;lD1;:::;
X
lg l
Q .x/ WD g.a;
N x/ C ral g.a;
N x/ E a.!/ aN ; (1.45a)
lD2
l
where E a.!/ aN denotes the system of mixed lth central moments of the
T
random vector a.!/ D a1 .!/; : : : ; a .!/ . Assuming that the domain of g D
g.a; x/ is convex with respect to a, we get the error estimate
ˇ ˇ
ˇ ˇ 1 l C1
ˇ .x/ Q .x/ˇ E sup rag g aN C # a.!/ aN ; x
.lg C 1/Š 0#1
lg C1
a.!/ aN : (1.45b)
In many practical cases the random parameter -vector a D a.!/ has a convex,
l C1
bounded support, and rag g is continuous. Then the L1 –norm
1 l C1
r.x/ WD ess sup rag g a.!/; x (1.45c)
.lg C 1/Š
is finite for all x under consideration, and (1.45b, c) yield the error bound
ˇ ˇ lg C1
ˇ ˇ
ˇ .x/ Q .x/ˇ r.x/E a.!/ aN : (1.45d)
hence, the loss function D .a; y/ depends also explicitly on the parameter vector
a. This may occur e.g. in case of randomly varying cost factors.
Linearizing now the vector function y D y.a; x/ with respect to a at a,
N thus,
This approximation is very advantageous in case that the cost function D .a; y/
is a quadratic function in y. In case of a cost function D .a; y/ being linear in
the vector y, also quadratic expansions of y D y.a; x/ with respect to a many be
taken into account.
Corresponding to (1.37), (1.39e), define
y .1/ .x/ WD Ey.1/ a.!/; x D y.a; x/ (1.46d)
Q.1/ .x/ WD cov y.1/ a./; x D ra y.a; x/ cov a./ ra y.a; x/T : (1.46e)
or
0 1 0 1
\
N \
N
P@ Sj A WD P @a.!/ 2 Sj A (1.47b)
j D1 j D1
where
!
X \
k
sk;N WD P Vil ; (1.48b)
1i1 <i2 <<ik N lD1
1.7 Approximation of Probabilities: Probability Inequalities 29
Besides (1.48c, d), a large amount of related bounds of different complexity are
available, cf. [48, 166]. Important bounds of first and second degree are given below:
0 1
[
N
max qj P @ Vj A Q 1 (1.49a)
1j N
j D1
0 1
[
N X
Q1 Q2 P @ Vj A Q1 max qil (1.49b)
1lN
j D1 i 6Dl
0 1
Q12 [
N
P@ Vj A Q 1 : (1.49c)
Q1 C 2Q2 j D1
X
N
Q1 WD qj with qj WD P.Vj / (1.49d)
j D1
j 1
X
N X
Q2 WD qij with qij WD P.Vi \ Vj /: (1.49e)
j D2 i D1
Moreover, defining
we have
0 1
[
N
P@ Vj A q T Q q; (1.49g)
j D1
\
m
In many cases the survival or feasible domain (event) S D Si , is represented by
i D1
a certain number m of inequality constraints of the type
as e.g. operating conditions, behavioral constraints. Hence, for a fixed input, design
or control vector x, the event S D S.x/ is given by
Here,
are certain functions, e.g. response, output or performance functions of the structure,
system, defined on (a subset of) R Rr .
Moreover, yli < yui ; i D 1; : : : ; m, are lower and upper bounds for the variables
yi ; i D 1; : : : ; m. In case of one-sided constraints some bounds yli ; yui are infinite.
Two-Sided Constraints
Consequently, for the probability P.S / of the event S , defined by (1.50b), we have
ˇ ˇ
ˇ ˇ
P.S / D P ˇyi a.!/; x yic ˇ < ./i ; i D 1; : : : ; m : (1.50f)
i / '.y/ 0; y 2 Rm (1.51d)
ii/ '.y/ '0 > 0; if y TB; (1.51e)
with a positive constant '0 (Fig. 1.3), we find the following result:
Theorem 1.1 For any (measurable) function ' W Rm ! R fulfilling conditions
(1.51d, e), the following Tschebyscheff-type inequality holds:
P yli < ./yi a.!/; x < ./yui ; i D 1; : : : ; m
1
1 E' yQ a.!/; x ; (1.52)
'0
Z Z
E' yQ a.!/; x D '.y/Py.a./;x/
Q .dy/ C '.y/Py.a./;x/
Q .dy/
y2B y2B c
Z Z
'.y/Py.a./;x/
Q .dy/ '0 Py.a./;x/
Q .dy/
y2B c y2B c
D '0 P yQ a.!/; x TB D '0 1 P yQ a.!/; x 2 B ;
Thus, the lower bound '0 follows by considering the convex optimization
Q
problems arising in the right hand side of (1.53a). Moreover, the expectation E'.y/
needed in (1.52) is given, see (1.51a), by
E'.y/Q D E yQ T C yQ D EtrC yQ yQ T ;
T
D trC.diag /1 cov y a./; x C y.x/ yc y.x/ yc .diag /1 ;
(1.53b)
reads
T
Q mkC k k.diag /1 k2 kcov y a./; x C y.x/ yc y.x/ yc k:
E'.y/
(1.53c)
Example 1.7 Assuming in the above Example 1.6 that C D diag .ci i / is a diagonal
matrix with positive elements ci i > 0; i D 1; : : : ; m, then
Q is given by
and E'.y/
2
X
m E yi a.!/; x yic
Q D
E'.y/ ci i
i D1
i2
2
X
m y2i a./; x C y i .x/ yic
D ci i : (1.53e)
i D1
i2
One-Sided Inequalities
Suppose that exactly one of the two bounds yli < yui is infinite for each
i D 1; : : : ; m. Multiplying the corresponding constraints in (1.50a) by 1, the
admissible domain S D S.x/, cf. (1.50b), can be represented always by
where yQi WD
yi yui, if yli D 1, and yQi WD yli yi , if yui D C1. If we set
Q x/ WD yQi .a; x/ and
y.a;
n o
BQ WD y 2 Rm W yi < ./ 0; i D 1; : : : ; m ; (1.54b)
then
P S.x/ D P yQ a.!/; x 2 BQ : (1.54c)
Consider also in this case, cf. (1.51d, e), a function ' W Rm ! R such that
i / '.y/ 0; y 2 Rm (1.55a)
Q
ii/'.y/ '0 > 0; if yTB: (1.55b)
34 1 Stochastic Optimization Methods
and
X
m
E' yQ a.!/; x D wi Ee ˛i yQi a.!/;x ; (1.57b)
i D1
N 1 2 2
Ee ˛i yQi D e ˛i yQi .x/ e 2 ˛i yi .a./;x/ : (1.57d)
Example 1.9 Consider '.y/ WD .y b/T C.y b/, where, cf. Example 1.6, C is
a positive definite m m matrix and b < 0 a fixed m-vector. In this case we again
have
and
T
Q D E yQ a.!/; x b C yQ a.!/; x b
E'.y/
T
D trCE yQ a.!/; x b yQ a.!/; x b :
Note that
yi .a; x/ .yui C bi /; if yli D 1
yQi .a; x/ bi D
yli bi yi .a; x/; if yui D C1;
Here, ! is the basic random element taking values in a probability space .˝; A; P /,
and describing the present random variations of model parameters or the influence
of noise terms. The probability space .˝; A; P / consists of the sample space
or set of elementary events ˝, the -algebra A of events and the probability
measure P . The plant state vector z D z.t; !/ is an m-vector involving direct
or indirect measurable/observable quantities like displacements, stresses, voltage,
current, pressure, concentrations, etc., and their time derivatives (velocities), z0 .!/
is the random initial state. The plant control or control input u.t/ is a deterministic
or stochastic n-vector denoting system inputs like external forces or moments,
voltages, field current, thrust program, fuel consumption, production rate, etc.
Furthermore, zP denotes the derivative with respect to the time t. We assume that
an input u D u.t/ is chosen such that u./ 2 U , where U is a suitable linear
space of input functions u./ W Œt0 ; tf ! Rn on the time interval Œt0 ; tf . Examples
for U are subspaces of the space PCn0 Œt0 ; tf of piecewise continuous functions
zP.t/ D g t; z.t/; u.t/ ; t0 t tf ; (2.1c)
z.t0 / D z0 : (2.1d)
Hence, in this case (2.1c, d) represents again an ordinary system of first order
differential equations for the %m unknown functions
zij D zi .t; !j /; i D 1; : : : ; m; j D 1; : : : ; %:
Results on the existence and uniqueness of the systems (2.1a, b) and (2.1c, d) and
their dependence on the inputs can be found in [37].
Also in the general case we consider a solution in the point-wise sense. This
means that for each random element ! 2 ˝, (2.1a, b) is interpreted as a system
of ordinary first order differential equations with the initial values z0 D z0 .!/ and
control input u D u.t/. Hence, we assume that to each deterministic control u./ 2
U and each random element ! 2 ˝ there exists a unique solution
z.; !/ D S !; u./ D S !; u./ ; (2.2a)
Zt
z.t/ D z0 .!/ C g s; !; z.s/; u.s/ ds; t0 t tf ; (2.2b)
t0
40 2 Optimal Control Under Stochastic Uncertainty
such that .t; !/ ! S !; u./ .t/ is measurable. This solution is also denoted by
Obviously, the integral equation (2.2b) is the integral version of the initial value
problem (2.1a, b): Indeed, if, for given ! 2 ˝, z D z.t; !/ is a solution of (2.1a,
b), i.e., z.; !/ is absolutely continuous, satisfies (2.1a) for almost all t 2 Œt0 ; tf
and fulfills (2.1b), then z D z.t; !/ is also a solution of (2.2b). Conversely, if, for
given ! 2 ˝, z D z.t; !/ is a solution of (2.2b), such that the integral on the
right hand side exists in the Lebesgue-sense for each t 2 Œt0 ; tf , then this integral
as a function of the upper bound t and therefore also the function z D z.t; !/ is
absolutely continuous. Hence, by taking t D t0 and by differentiation of (2.2b) with
respect to t, cf. [118], we have that z D z.t; !/ is also a solution of (2.1a, b).
In the following we want to justify the above assumption that the initial value
problem (2.1a, b), the equivalent integral equation (2.2b), resp., has a unique
solution z D z.t; !/. For this purpose, let D .t; !/ be an r-dimensional
stochastic process, as e.g. time-varying disturbances, random parameters, etc., of
the system, such that the sample functions .; !/ are continuous with probability
one. Furthermore, let
gQ W Œt0 ; tf Rr Rm Rn ! Rm
Q Dz g;
be a continuous function having continuous Jacobians D q; Q Du gQ with respect
to ; z; u. Now consider the case that the function g of the process differential
equation (2.1a, b) is given by
g.t; !; z; u/ WD gQ t; .t; !/; z; u ;
.t; !; z; u/ 2 Œt0 ; tf ˝ Rm Rn . The spaces U; Z of possible inputs, trajectories,
resp., of the plant are chosen as follows: U WD Reg.Œt0 ; tf I Rn / is the Banach space
of all regulated functions u./ W Œt0 ; tf ! Rn , normed by the supremum norm kk1 .
Furthermore, we set Z WD C0m Œt0 ; tf and
WD C0r Œt0 ; tf . Here, for an integer
, C0 Œt0 ; tf denotes the Banach space of all continuous functions of Œt0 ; tf into R
normed by the supremum norm k k1 . By our assumption we have .; !/ 2
a.s.
(almost sure). Define
D Rm U I
space of inputs
0 1
z0
WD @ ./ A
u./
into the plant, consisting of the random initial state z0 , the random input function
D .t; !/ and the control function u D u.t/. Let now the mapping W Z ! Z
related to the plant equation (2.1a, b) or (2.2b) be given by
0 1
Zt
; z./ .t/ D z.t/ @z0 C gQ s; .s/; z.s/; u.s/ dsA ; t0 t tf : (2.2d)
t0
Note that for each input vector 2 and function z./ 2 Z the integrand in
(2.2d) is piecewise continuous, bounded or at least essentially bounded on Œt0 ; tf .
Hence, the integral in (2.2d) as a function of its upper limit t yields again a
continuous function on the interval Œt0 ; tf , and therefore an element of Z. This
shows that maps Z into Z.
Obviously, the initial value problem (2.1a, b) or its integral form (2.2b) can be
represented by the operator equation
Operators of the type (2.2d) are well studied, see e.g. [37, 85]: It is known that
is continuously Fréchet .F /-differentiable [37, 85]. Note that the F-differential
is a generalization of the derivatives (Jacobians) of mappings between finite-
dimensional spaces to mappings between arbitrary
normed spaces. Thus, the
F -derivative D of at a certain point ; z./ is given by
N zN./ ; z./ .t/
D ;
Zt
D z.t/ z0 C Dz gQ s; N .s/; zN.s/; uN .s/ z.s/ds
t0
Zt
C D gQ s; N .s/; zN.s/; uN .s/ .s/ds
t0
1
Zt
C Du gQ s; N .s/; zN.s/; uN .s/ u.s/dsA ; t0 t tf ; (2.2f)
t0
42 2 Optimal Control Under Stochastic Uncertainty
where N D zN0 ; N ./; uN ./ and D z0 ; ./; u./ . Especially, for the derivative of
with respect to z./ we find
! Zt
N zN./ z./ .t/ D z.t/
Dz ; Dz gQ s; N .s/; zN.s/; uN .s/ z.s/ds; t0 t tf :
t0
(2.2g)
The related equation
N zN./ z./ D y./; y./ 2 Z;
Dz ; (2.2h)
is a linear vectorial Volterra integral equation. By our assumptions this equation has
a unique solution z./ 2 Z. Note that the corresponding result for scalar Volterra
equations,
see
e.g. [151], can be transferred to the present vectorial case. Therefore,
N
Dz ; zN./ is a linear, continuous one-to-one map from Z onto Z. Hence, its
!
1
inverse Dz ; N zN./ exists. Using the implicit function theorem [37, 85], we
Zt
z.t/ D z0 C gQ s; .s/; z.s/; u.s/ ds; t0 t tf ; (2.3c)
t0
N and it holds
where D z0 ; ./; u./ ; c) S is continuously differentiable on V ./,
1
Du S./ D Dz ; S./ N
Du ; S./ ; 2 V ./: (2.3d)
2.1 Stochastic Control Systems 43
Zt
.t/ Dz gQ s; .s/; S./.s/; u.s/ .s/ds
t0
Zt
D Du gQ s; .s/; S./.s/; u.s/ h.s/ds; (2.3e)
t0
where t0 t tf and D z0 ; ./; u./ .
Remark 2.1 Taking the time derivative of Eq. (2.3e) shows that this integral equa-
tion is equivalent to the so-called perturbation equation, see e.g. [73].
For an arbitrary h./ 2 U the mappings
.t; / ! S./.t/; .t; / ! Du S./h./ .t/; .t; / 2 Œt0 ; tf V ./; (2.3f)
nonlinear dynamics of the plant, random varying model parameters such as material,
load or cost parameters, etc.
Here, the “noise” term D .t; !/ is a certain stochastic process, as e.g. the
Brownian motion, having continuous paths, and gQ D g.t; Q z; u/; hQ D h.t;
Q z; u/ are
given, sufficiently smooth vector/matrix functions.
Corresponding to the integral equation (2.2a, b), the above stochastic differential
equation is replaced by the stochastic integral equation
Zt Zt
z.t; !/ D z0 .!/ C gQ s; z.s; !/; u.s/ ds C hQ s; z.s; !/; u.s/ d.s; !/:
t0 t0
(2.4b)
The meaning of this equation depends essentially on the definition (interpretation)
of the “stochastic integral”
Zt
I .; /; z.; / .t/ WD hQ s; z.s; !/; u.s/ d.s; !/: (2.4c)
t0
Note that in case of closed loop and open-loop feedback controls, see the next
Sect. 2.2, the control function u D u.s; !/ is random. If
Q z; u/ D hQ s; u.s/
h.s; (2.5a)
with a deterministic
control u D u.t/ and a matrix function hQ D h.s;Q u/, such that
s ! hQ s; u.s/ is differentiable, by partial integration we get, cf. [121],
I ; z./ .t/ D I ./; u./ .t/ D hQ t; u.t/ .t/ hQ t0 ; u.t0 / .t0 /
Zt
d Q
.s/ h s; u.s/ ds; (2.5b)
ds
t0
2.1 Stochastic Control Systems 45
Zt
; z./ .t/ WD z.t/ z0 C gQ s; z.s/; u.s/ ds C hQ t; u.t/ .t/
t0
Zt !
d
hQ t0 ; u.t0 / .t0 / .s/ hQ s; u.s/ ds : (2.5c)
ds
t0
The aim is to obtain optimal controls being robust, i.e., most insensitive with
respect to stochastic variations of the model/environmental parameters and initial
values of the process. Hence, incorporating stochastic parameter variations into the
optimization process, for a deterministic
control function u D u.t/; t0 t tf ,
the objective function F D F u./ of the controlled process z D z.t; !; u.// is
defined, cf. [100], by the conditional expectation of the total costs arising along the
whole control process:
F u./ WD Ef !; S !; u./ ; u./ : (2.6a)
Zf t
f !; z./; u./ WD L t; !; z.t/; u.t/ dt C G tf ; !; z.tf / ; (2.6b)
t0
L W Œt0 ; tf ˝ Rm Rn ! R;
G W Œt0 ; tf ˝ Rm !R
46 2 Optimal Control Under Stochastic Uncertainty
are given measurable cost functions. We suppose that L.t; !; ; / and G.t; !; / are
convex functions for each .t; !/ 2 Œt0 ; tf ˝, having continuous partial derivatives
rz L.; !; ; /; ru L.; !; ; /; rz G.; !; /. Note that in this case
Ztf
z./; u./ ! L t; !; z.t/; u.t/ dt C G tf ; !; z.tf / (2.6c)
t0
is a convex function
on Z U for each ! 2 ˝. Moreover, assume that the
expectation F u./ exists and is finite for each admissible control u./ 2 D.
In the case of random inputs u D u.t; !/; t0 t tf ; ! 2 ˝, with definition
(2.6b), the objective function F D F u.; / reads
!
F u.; / WD Ef !; S !; u.; !/ ; u.; !/ : (2.6d)
Example 2.1 (Tracking Problems) If a trajectory zf D zf .t; !/, e.g., the trajectory
of a moving target, known up to a certain stochastic uncertainty, must be followed
or reached during the control process, then the cost function L along the trajectory
can be defined by
2
L t; !; z.t/; u WD z z.t/ zf .t; !/ C '.u/: (2.6e)
In (2.6e) z is a weight matrix, and ' D '.u/ denotes the control costs, as e.g.
Here, Q D Q.t; !/ and R D R.t; !/, resp., are certain positive (semi)definite
m m; n n matrices which may depend also on .t; !/. Moreover, the terminal
cost function G depends then on .!; z/. For example, in case of endpoint control,
the cost function G is given by
with a certain desired, possibly random terminal point zf D zf .!/ and a positive
(semi)definite, possibly random weight matrix Gf D Gf .!/.
For finding optimal controls being robust with respect to stochastic parameter
variations u? ./; u? .; /, resp., in this chapter we are presenting now several methods
for approximation of the following minimum expected total cost problem:
min F u./ s.t. u./ 2 D; (2.7)
min F u.; / s.t. u.; !/ 2 D a.s. (almost sure) ;
u.t; / At -measurable, f
.2:7/
Problem (2.7) is of course equivalent E D E.jAt0 / to the optimal control
problem under stochastic uncertainty:
0 tf 1
Z ˇ
ˇ
min E @ L t; !; z.t/; u.t/ dt C G tf ; !; z.tf / ˇAt0 A (2.9a)
t0
s.t.
zP .t/ D g t; !; z.t/; u.t/ ; t0 t tf ; a.s. (2.9b)
Here, hI D hI .t; !; z/; hII D hII .t; !; z/ are given vector functions of .t; !; z/. By
means of (penalty) cost functions, the random condition (2.9e) can be incorporated
into the objective function (2.9a). As explained in Sect. 2.8, the expectations arising
in the mean value constraints (2.9f) and in the objective function (2.9a) can be
computed approximatively by means of Taylor expansion with respect to the vector
# D #.!/ WD .z0 .!/; .!/ of random initial values and model parameters at
.t0 /
the conditional mean # D # WD E #.!/jAt0 . This yields then ordinary
deterministic constraints for the extended deterministic trajectory (nominal state and
sensitivity)
t ! z.t; #/; D# z.t; #/ ; t t0 :
2.2 Control Laws 49
Control or guidance usually refers [5, 73, 85] to direct influence on a dynamic
system to achieve desired performance. In optimal control of dynamic systems
mostly the following types of control laws or control policies are considered:
I) Open-Loop Control (OL)
Here, the control function u D u.t/ is a deterministic function depending
only on the (a priori) information It0 about the system, the model parameters,
resp., available at the starting time point t0 . Hence, for the optimal selection
of optimal (OL) controls
u.t/ D u tI t0 ; It0 ; t t0 ; (2.10a)
depending on time t and the total information It about the system available up
to time t. Especially It may contain information about the state z.t/ D z.t; !/
up to time t. Optimal (CL) or feedback controls are obtained by solving
f
problems of type (2:7).
Remark 2.4 Information Set At at Time t: Often the information It available
up to time t is described by the information set or algebra At A
of events A occurred up to time t. In the important case At D A.y.t; //,
where y D y.t; !/ denotes the m-vector
N function of state-measurements or
-observations at time t, then an At measurable control u D u.t; !/, see
f has the representation, cf. [12],
problem (2:7),
control function for the remaining time interval t s tf , see Fig. 2.1, is
computed, hence,
uŒt;tf .s/ D u sI t; It ; s t: (2.10d)
cf. (2.10c), are determined first on the time interval Œt; t C Tp with a certain
so-called prediction time horizon Tp > 0. In sampled-data MPC, cf. [45],
optimal open-loop controls u D uŒti ;ti CTp , are determined at certain sampling
instants ti ; i D 0; 1; : : :, using the information Ati about the control process
and its neighborhood up to time ti ; i D 0; 1; : : :, see also [98]. The optimal
52 2 Optimal Control Under Stochastic Uncertainty
until the next sampling instant ti C1 . This method is closely related to the
Adaptive Optimal Stochastic Trajectory Planning and Control (AOSTPC)
procedure described in [95, 98].
Corresponding to the extension of (OLF) control to (SOLF) control,
(NMPC) can be extended to Stochastic Nonlinear Model Predictive Control
(SNMPC). For control policies of this type, robust (NMPC) with respect to
stochastic variations of model parameters and initial values are determined in
the following way:
• Use the a posteriori distribution P .d!jAt / of the basic random element
! 2 ˝, given the process information At up to time t, and
• apply stochastic optimization methods to incorporate random parameter
variations into the optimal (NMPC) control design.
Let v./ 2 D be an arbitrary, but fixed admissible initial or reference control and
assume, see Lemma 2.1, for the input–output map z./ D S.!; u.//:
Assumption 2.1 S.!; / is F -differentiable at v./ for each ! 2 ˝.
Denote by DS !; v./ the F -derivative of S.!; / at v./, and replace now the cost
function F D F u./ by
Fv./ u./ WD Ef !; S !; v./ C DS !; v./ u./ v./ ; u./ (2.11)
where u./ 2 U . Assume that Fv./ u./ 2 R for all pairs u./; v./ 2 D D.
Then, replace the optimization problem (2.7), see [89], by
min Fv./ u./ s:t: u./ 2 D: .2:7/v./
A solution uN ./ 2 D of the convex problem .2.7/v./ is then characterized cf. [94],
by
0
Fv./C uN ./I u./ uN ./ 0 for all u./ 2 D: (2.13)
Definition 2.4 For each v./ 2 D, let M v./ be the set of solutions of problem
.2.7/v./ , i.e., M v./
n o
0
WD u0 ./ 2 D W Fv./C u0 ./I u./ u0 ./ 0; u./ 2 D .
A first relation between our original problem (2.7) and the family of its
approximates .2.7/v./ ; v./ 2 D, is shown in the following.
Theorem 2.1 Suppose that the directional derivative FC0 D FC0 v./I h./ exists
and
FC0 v./I h./ D Fv./C
0
v./I h./ (2.14)
1
D E lim f !; S !; v./ C "h./ ; v./ C "h./
"#0 "
f !; S !; v./ ; v./ : (2.15)
Note. Because of the properties of the operator S D S.!; u.//, the above equations
holds also for arbitrary convex functions f such that the expectations under
consideration exist, see [90].
Due to the definition (2.6a) of the objective function F D F .u.// for condition
(2.14) the following criterion holds:
Lemma 2.4 a) Condition (2.14) in Theorem 2.1 holds if and only if the expectation
operator “E” and the limit process “lim” in (2.15) may be interchanged. b) This
n"#0 o
interchangeability holds e.g. if sup DS !; v./ C "h./ W 0 " 1 is
bounded with probability one, and the convex function f .!; ; / satisfies a Lipschitz
condition
ˇ ˇ
ˇ ˇ
ˇf !; z./; u./ f !; zN./; uN ./ ˇ
.!/ z./; u./ zN./; uN ./
ZU
on a set Q Z U containing all vectors S !; v./ C "h./ ; v./ C "h./ ; 0
" 1, where E.!/ < C1, and kkZU denotes the norm on Z U .
Proof The first assertion a) follows immediately from (2.15) and the definition of
the objective function F . Assertion b) can be obtained by means of the generalized
mean value theorem for vector functions and Lebesgue’s bounded convergence
theorem.
Remark 2.7 Further conditions are given in [90].
A second relation between our original problem (2.7) and the family of its
approximates .2.7/v./ ; v./ 2 D, is shown next.
Lemma 2.5 a) If uN ./ … M uN ./ for a control uN ./ 2 D, then
FuN ./ u./ < FuN ./ uN ./ D F uN ./ for each u./ 2 M uN ./
56 2 Optimal Control Under Stochastic Uncertainty
u 2 D, put j D 1
1
Algorithm 2.1 I) Choose
II) If uj ./ 2 M uj ./ , then uj ./ is stationary and the algorithm stops;
otherwise find a control v j ./ 2 M uj ./
III) Set uj C1 ./ D v j ./ and go to ii), putting j ! j C 1.
u ./ 2 D, put j D 1
1
Algorithm 2.2 I) Choose
II) If uj ./ 2 M uj ./ , then uj ./ is stationary and the algorithm stops;
otherwise find a v .j / ./ 2 M uj ./ , define hj ./ WD v j ./ uj ./
2.4 Computation of Directional Derivatives 57
III) Calculate uN ./ 2 m uj ./; hj ./ , set uj C1./ WD uN ./ and go to ii), putting
j ! j C 1.
Here, based on line search, m u./; h./ is defined by
(
m u./; h./ D u./ C " h./ W F u./ C " h./
)
D min F u./ C "h./ for " 2 Œ0; 1 ; u./; h./ 2 U:
0"1
Suppose here again that U WD Ln1 .Œt0 ; tf ; B 1 ; 1 / is the Banach space of all
essentially bounded measurable functions u./ W Œt0 ; tf ! Rn , normed by the
essential supremum norm. According to Definitions 2.4, 2.5 of a stationary control
and characterization (2.13) of an optimal solution of (2.7)v./, we first have to
58 2 Optimal Control Under Stochastic Uncertainty
0
determined the directional derivative Fv./C . Based on the justification in Sect. 2.1,
we assume again that the solution z.t; !/ D S !; u./ .t/ of (2.2b) is measurable
in .t; !/ 2 Œt0 ; tf ˝ for each u./ 2 D, and u./ ! S !; u./ is continuously
differentiable on D for each ! 2 ˝. Furthermore,
we suppose that the F -
differential .t/ D .t; !/ D Du S !; u./ h./ .t/; h./ 2 U , is measurable
and essentially bounded in .t; !/, and is given according to (2.3a–f) by the linear
integral equation
Zt Zt
.t/ A t; !; u./ .s/ds D B t; !; u./ h.s/ds; (2.16a)
t0 t0
t0 t tf ;
.t0 / D 0: (2.16f)
2.4 Computation of Directional Derivatives 59
The solution D .t; !/ of (2.16a), (2.16e, f), resp., is also denoted, cf. (2.3f), by
!
.t; !/ D u;h .t; !/ WD Du S !; u./ h./ .t/; h./ 2 U: (2.16g)
This means that the approximate (2.7)v./ of (2.7) has the following explicit form:
Ztf
min E L t; !; zv .t; !/ C .t; !/; u.t/ dt (2.17a)
t0
!
C G tf ; !; zv .tf ; !/ C .tf ; !/
s.t.
P !/ D A t; !; v./ .t; !/ C B t; !; v./ u.t/ v.t/ a.s. (2.17b)
.t;
With the convexity assumptions in Sect. 2.1.2, Lemma 2.3 yields that (2.17a–d) is a
convex stochastic control problem, with a linear plant differential equation.
For the subsequent analysis of the stochasticcontrol problem
we need now a
0
representation of the directional derivative Fv./C u./I h./ by a scalar product
Zf
t
0
Fv./C u./I h./ D q.t/T h.t/dt
t0
Ztf T
0
Fv./C u./I h./ D E rz L t; !; zv .t; !/ C v;uv .t; !/; u.t/ v;h .t; !/
t0
!
T
C ru L t; !; zv .t; !/ C v;uv .t; !/; u.t/ h.t/
!
T
C rz G tf ; !; zv .tf ; !/ C v;uv .tf ; !/ v;h .tf ; !/ : (2.18)
60 2 Optimal Control Under Stochastic Uncertainty
0
measurable with respect to .t; !/, the directional derivative Fv./C can be repre-
sented by
Rtf T
0
Fv./C u./I h./ D E a t; !; v./; u./ v;h .t; !/
t0
! !
T T
Cb t; !; v./; u./ h.t/ dt C c tf ; !; v./; u./ v;h .tf ; !/ : (2.20a)
Ztf !
v;h .tf ; !/ D A t; !; v./ v;h .t; !/ C B t; !; v./ h.t/ dt: (2.20b)
t0
Ztf T
0
Fv./C u./I h./ D E aQ t; !; v./; u./ v;h .t; !/dt
t0
Ztf !
T
C bQ t; !; v./; u./ h.t/dt ; (2.20c)
t0
where
T
aQ t; !; v./; u./ WD a t; !; v./; u./ C A t; !; v./ c tf ; !; v./; u./
(2.20d)
T
bQ t; !; v./; u./ WD b t; !; v./; u./ C B t; !; v./ c tf ; !; v./; u./ :
(2.20e)
2.4 Computation of Directional Derivatives 61
Remark 2.9 According to Remark 2.8 also the functions .t; !/ ! aQ t; !; v./;
u./ , .t; !/ ! bQ t; !; v./; u./ are measurable on Œt0 ; tf ˝.
In order to transform the first integral in (2.20c) into the form of the second
integral in (2.20c), we introduce the m-vector function
D v;u .t; !/
T Z f
t
.t/ A t; !; v./ .s/ds D aQ t; !; v./; u./ : (2.21)
t
Under the present assumptions, this Volterra integral equation has [90] a unique
measurable solution .t; !/ ! v;u .t; !/, see also Remark 2.8. By means of (2.21)
we obtain
Ztf T
aQ t; !; v./; u./ v;h .t; !/dt
t0
Ztf !T
T Z f
t
D .t/ A t; !; v./ .s/ds v;h .t; !/dt
t0 t
Ztf
Ztf Ztf
dt dsJ.s; t/.s/T A t; !; v./ v;h .t; !/
t0 t0
Ztf
Ztf Ztf
D .s/ v;h .s; !/ds
T
ds dtJ.s; t/.s/T A t; !; v./ v;h .t; !/
t0 t0 t0
Ztf Zs !
D .s/ T
v;h .s; !/ A t; !; v./ v;h .t; !/dt ds; (2.22a)
t0 t0
62 2 Optimal Control Under Stochastic Uncertainty
Using now again the perturbation equation (2.16a, b), from (2.22a) we get
Ztf Ztf Zs !
T
aQ t; !; v./; u./ v;h .t; !/dt D .s/T B t; !; v./ h.t/dt ds
t0 t0 t0
Ztf Ztf
D ds.s/ T
dtJ.s; t/B t; !; v./ h.t/
t0 t0
Ztf Ztf
D dt dsJ.s; t/.s/T B t; !; v./ h.t/
t0 t0
Ztf !T
Zf Ztf T Z f
t t
T
D .s/ds B t; !; v./ h.t/dt D B t; !; v./ .s/ds h.t/dt:
t0 t t0 t
(2.22b)
Ztf T Z f
t
0
Fv./C u./I h./ D E B t; !; v./ .s/ds
t0 t
! !
T
CbQ t; !; v./; u./ h.t/dt : (2.23)
!
Zf
t
T
.t/ A t; !; v./ c tf ; !; v./; u./ C .s/ds D a t; !; v./; u./ :
t
(2.24)
According to (2.24), defining the m-vector function
Ztf
y D yv;u .t; !/ WD c tf ; !; v./; u./ C v;u .s; !/ds; (2.25a)
t
2.4 Computation of Directional Derivatives 63
we get
T
.t/ A t; !; v./ yv;u .t; !/ D a t; !; v./; u./ : (2.25b)
Replacing in (2.25b) the variable t by s and integrating then the equation (2.25b)
over the time interval Œt; tf , yields
Ztf Ztf !
T
.s/ds D A s; !; v./ yv;u .s; !/ C a s; !; v./; u./ ds: (2.25c)
t t
Ztf !
T
C A s; !; v./ yv;u .s; !/ C a s; !; v./; u./ ds: (2.25d)
t
Obviously, the differential form of the Volterra integral equation (2.25d) for y D
yv;u .t; !/ reads:
T
P
y.t/ D A t; !; v./ y.t/ a t; !; v./; u./ ; t0 t tf ; (2.26a)
y.tf / D c tf ; !; v./; u./ : (2.26b)
System (2.25d), (2.26a, b), resp., is called the adjoint integral, differential
equation related to the perturbation equation (2.16a, b).
By means of (2.25a), from (2.20e) and (2.23) we obtain now
Ztf T Z f
t
0
Fv./C u./I h./ D E B t; !; v./ .s/ds C b t; !; v./; u./
t0 t
!
T T
CB t; !; v./ c tf ; !; v./; u./ h.t/dt
Ztf !
T T
DE B t; !; v./ yv;u .t; !/ C b t; !; v./; u./ h.t/dt:
t0
64 2 Optimal Control Under Stochastic Uncertainty
Zf
t
T
0
Fv./C u./I h./ D E B t; !; v./ yv;u .t; !/
t0
!
T
Cb t; !; v./; u./ h.t/dt: (2.27)
0
Note that Fv./C u./I is also the Gâteaux-differential of Fv./ at u./.
0
For a further discussion of formula (2.27) for Fv./C u./I h./ , in generalization
of the Hamiltonian of a deterministic control problem, see e.g. [73], we introduce
now the stochastic Hamiltonian related to the partly linearized control problem
(2.17a–d) based on a reference control v./:
Hv./ .t; !; ; y; u/ WD L t; !; zv .t; !/ C ; u
!
Cy A t; !; v./ C B t; !; v./ u v.t/ ;
T
(2.28a)
0
.t; !; z; y; u/ 2 Œt0 ; tf ˝ Rm Rm Rn . Using Hv./ , for Fv./C we find the
representation
Zf
t
0
Fv./C u./I h./ D Eru Hv./ t; !; v;uv .t; !/;
t0
!T
yv;u .t; !/; u.t/ h.t/dt; (2.28b)
where v;uv D v;uv .t; !/ is the solution of the perturbation differential, integral
equation (2.17b, c), (2.20b), resp., and yv;u D yv;u .t; !/ denotes the solution of the
adjoint differential, integral equation (2.26a, b), (2.25d).
Let u0 ./ 2 U denote a given initial control. By means of (2.28a, b), the
necessary
and sufficient condition for a control u1 ./ to be an element of the set M u0 ./ , i.e.
2.4 Computation of Directional Derivatives 65
a solution of the approximate convex problem (2.7)u0./ , reads, see Definition 2.4
and (2.13):
Ztf
Eru Hu0 ./ t; !; u0 ;u1 u0 .t; !/; yu0 ;u1 .t; !/;
t0
T
u1 .t/ u.t/ u1 .t/ dt 0; u./ 2 D: (2.29)
Introducing, for given controls u0 ./; u1 ./, the convex mean value function
H u0 ./;u1 ./ D H u0 ./;u1 ./ .u.// defined by
Ztf
H u0 ./;u1 ./ u./ WD EH u0 ./ t; !; u0 ;u1 u0 .t; !/;
t0
yu0 ;u1 .t; !/; u.t/ dt; (2.30a)
0
H u0 ./;u1 ./C u1 .//I u./ u1 ./ 0; u./ 2 D; (2.30b)
Due to the equivalence of the conditions (2.29) and (2.30b), for a control u ./ 2
1
Theorem 2.4 Let u0 ./ 2 D be a given initial control. A control u.1/ ./ 2 U is a
solution of (2.7)u0./ , i.e., u1 ./ 2 M u0 ./ , if and only if u1 ./ is an optimal solution
of the convex stochastic optimization problem
min H u0 ./;u1 ./ u./ s.t. u./ 2 D; (2.31a)
Zf
t
H u0 ./;.;/;y.;/ u./ D EH u0 ./ t; !; .t; !/; y.t; !/; u.t/ dt: (2.31b)
t0
66 2 Optimal Control Under Stochastic Uncertainty
(2.33b)
then
ue u0 ./ ; .; /; y.; / 2 argmin H u0 ./;.;/;y.;/ u./ : (2.33c)
u./2D
Because of Theorem 2.4, problems .P /tu0 ./;;y and (2.33a–c), we introduce, cf.
[73], the following definition:
Definition 2.6 Let u0 ./ 2 D be a given initial control. For measurable functions
D .t; !/; y D y.t; !/ on .˝; A; P /, let denote
ue D ue u0 ./ t; .t; /; y.t; / ; t0 t tf ;
a solution of .P /tu0 ./;;y ; t0 t tf . The function ue D ue u0 ./ t; .t; /; y.t; /
is called a Hu0 ./ -minimal control of (2.17a–d). The stochastic Hamiltonian Hu0 ./
is called regular, strictly regular, resp., if a Hu0 ./ -minimal control exists, and is
determined uniquely.
1
Remark 2.10 Obviously, by means of Definition 2.6, for an optimal solution u ./
of (2.7)u0./ we have then the “model control law” ue D ue u0 ./ t; .t; /; y.t; /
depending on still unknown state, costate functions .t; /; y./, respectively.
2.5 Canonical (Hamiltonian) System 67
t0 t tf ; (2.34a)
.t0 ; !/ D 0 a.s. (2.34b)
T
P !/ D A t; !; u0 ./ y.t; !/
y.t;
!
rz L t; !; zu0 .t; !/ C .t; !/; u e t; .t; /; y.t; / ;
u0 ./
t0 t tf ; (2.34c)
y.tf ; !/ D rz G tf ; !; zu0 .tf ; !/ C .tf ; !/ : (2.34d)
Remark 2.11 Note that the (deterministic) control law ue D ue u0 ./ t; .t; /; y.t; /
depends on the whole random variable .t; !/; y.t; !/ ; ! 2 ˝, or the occurring
moments. In case of a discrete parameter distribution ˝ D f!1 ; : : : ; !% g, see
Sect. 2.1, the control law ue u0 ./ depends,
ue u0 ./ D ue u0 ./ t; .t; !1 /; : : : ; .t; !% /; y.t; !1 /; : : : ; y.t; !% / ;
Assuming that
u1 ./ 2 U; (2.36a)
u1 ./ 2 D: (2.36b)
Suppose here that condition (2.14) holds for all controls v./ under consideration,
cf. Lemma 2.4. Having amethod for the construction of improved approximative
controls u ./ 2 M u ./ related to an initial control u0 ./ 2 D, we consider now
1 0
Ztf !
T T
D E B t; !; v./ yv .t; !/ C b t; !; v./; v./ h.t/dt; (2.38a)
t0
where
related to the basic control problem (2.9a–d), from (2.38a, b), (2.39a–g) we get the
representation, cf. (2.28a, b),
Zf
t
T
FC0 v./I h./ D Eru H t; !; zv .t; !/; yv .t; !/; v.t/ h.t/dt: (2.40b)
t0
Thus, for stationary controls uN ./ 2 D of problem (2.7) we have the characterization
Ztf T
Eru H t; !; zuN .t; !/; yuN .t; !/; uN .t/ u.t/ uN .t/ dt 0; u./ 2 D: (2.41)
t0
Comparing (2.29) and (2.41), corresponding to (2.30a, b), for given w./ 2 D we
introduce here the function
Zf
t
H w u./ WD EH t; !; zw .t; !/; yw .t; !/; u.t/ dt; (2.42a)
t0
Remark 2.12 Because of the assumptions in Sect. 2.1.2, problem (2.42b) is (strictly)
convex, provided that the process differential equation (2.1a) is affine-linear with
respect to u, hence,
O !; z/; BO D B.t;
with a given vector-, matrix-valued function gO D g.t; O !; z/.
2.7 Canonical (Hamiltonian) System of Differential 71
cf. (2.30a, b), (2.31a, b). Corresponding to Theorem 2.4, here we have this result:
Theorem 2.6 (Optimality Condition for Stationary Controls) Suppose that a
control uN ./ 2 D fulfills (2.44). Then uN ./ is a stationary control of (2.7).
Assume now again that the feasible domain D is given by (2.32). In order to solve
the optimization problem (2.42a, b), corresponding to .P /tu0 ./;; , here we consider
the finite-dimensional optimization problem
min EH t; !; .!/; .!/; u s.t. u 2 Dt .P /t;
!
T
P !/ D Dz g t; !; z.t; !/; u t; z.t; /; y.t; /
y.t; y.t; !/
!
rz L t; !; z.t; !/; u t; z.t; /; y.t; / ; t0 t tf (2.45c)
y.tf ; !/ D rz G tf ; !; z.tf ; !/ : (2.45d)
Thus, (2.45a–d) is then an ordinary two-point boundary value problem for the 2%
unknown functions z D z.t; !j /; y D y.t; !j /; j D 1; : : : ; %.
N D zN.t; !/; y.t;
Let denote .Nz; y/ N !/ ; t0 t tf ; ! 2 .˝; A; P /, the unique
measurable solution of (2.45a–d) and define:
uN .t/ WD ue t; zN.t; /; y.t;
N / ; t0 t tf : (2.46)
hence,
uN .t/ D ue t; zuN .t; /; yuN .t; / ; t0 t tf : (2.47c)
Assuming that
Proof
According
N uN /, the control uN ./ 2 D minimizes
to the construction of .Nz; y;
H uN u./ on D. Hence,
uN ./ 2 argmin H uN u./ :
u./2D
g.t; w; z; u/ D g.t;
Q ; z; u/ (2.49a)
z0 .!/ D zQ0 ./ (2.49b)
Q ; z; u/
L.t; !; z; u/ D L.t; (2.49c)
Q ; z/:
G.t; !; z/ D G.t; (2.49d)
Here,
Since the approximate problem (2.17a–d), obtained by the above described inner
linearization, has the same basic structure as the original problem (2.9a–d), it is suf-
ficient to describe the procedure for problem (2.9a–d). Again, for simplification, the
74 2 Optimal Control Under Stochastic Uncertainty
conditional expectation E.: : : jAt0 / given the information At0 up to the considered
starting time t0 is denoted by “E”. Thus, let denote
t0
D WD E.!/ D E .!/jAt0 (2.50a)
the conditional expectation of the random vector .!/ given the information At0 at
time point t0 . Taking into account the properties of the solution
z D z.t; / D S z0 ./; ; u./ .t/; t t0 ; (2.50b)
of the dynamic equation (2.3a–d), see Lemma 2.1, the expectations arising in the
objective function (2.9a) can be computed approximatively by means of Taylor
expansion with respect to at .
Considering first the costs L along the trajectory we obtain, cf. [100],
L t; ; z.t; /; u.t/ D L t; ; z.t; /; u.t/
!
T
C r L t; ; z.t; /; u.t/ C D z.t; / rz L t; ; z.t; /; u.t/
T
. /
1
C . /T QL t; ; z.t; /; D z.t; /; u.t/ . / C : : : : (2.51a)
2
1 T
C E .!/ QL t; ; z.t; /; D z.t; /; u.t/ .!/ C : : :
2
1
D L t; ; z.t; /; u.t/ C trQL t; ; z.t; /; D z.t; /; u.t/ cov ./ C : : : :
2
(2.52)
1
C . /T QG tf ; ; z.tf ; /; D z.tf ; / . / C : : : ; (2.53a)
2
where QG is defined in the same way as QL , see (2.51a). Taking expectations with
respect to .!/, we get
!
EG tf ; .!/; z tf ; .!/ D G tf ; ; z.tf ; /
1
C trQG tf ; ; z.tf ; /; D z.tf ; / cov ./ C : : : : (2.53b)
2
Note. Corresponding to Definition 2.1 and (2.50a), for the mean and covariance
matrix of the random parameter vector D .!/ we have
.t0 /
D WD E .!/jAt0
!
T ˇ
.t 0 / .t 0 / ˇ
cov ./ D cov.t0 / ./ WD E .!/ .!/ ˇAt0 :
Taking expectations in (2.54a), for the expected cost function we get the
approximation
EL t; ; z.t; /; u.t/
EL t; .!/; z.t; / C D z.t; /..!/ /; u.t/ : (2.54b)
In many important cases, as e.g. for cost functions L being quadratic with respect to
the state variable z, the above expectation can be computed analytically. Moreover,
if the cost function L is convex with respect to z, then the expected cost function
EL is convex with respect to both, the state vector z.t; / and the Jacobian matrix
of sensitivities D z.t; / evaluated at the mean parameter vector .
Having the approximate representations (2.52), (2.53b), (2.54b), resp., of the
expectations occurring in the objective function (2.9a), we still have to compute
the trajectory t ! z.t; /; t t0 , related to the mean parameter vector D and
@z
the sensitivities t ! .t; /; i D 1; : : : ; r; t t0 , of the state z D z.t; / with
@i
respect to the parameters i ; i D 1; : : : ; r, at D . According to (2.3a, b) or (2.9b,
c), for z D z.t; / we have the system of differential equations
zP.t; / D g t; ; z.t; /; u.t/ ; t t0 ; (2.55a)
d
D z.t; / D Dz g t; ; z.t; /; u.t/ D z.t; /
dt
CD g t; ; z.t; /; u.t/ ; t t0 ; (2.56a)
d @z @g @z
.t; / D t; ; z.t; /; u.t/ .t; /
dt @j @j @j
@g
C t; ; z.t; /; u.t/ ; t t0 ; j D 1; : : : ; r: (2.56c)
@j
Denoting by
LQ D L
Q t; ; z.t; /; D z.t; /; u.t/ ; (2.57a)
GQ D GQ tf ; ; z.tf ; /; D z.tf ; / ; (2.57b)
Ztf
min E LQ t; .!/; z.t; /; D z.t; /; u.t/ dt
t0
CE GQ tf ; .!/; z.tf ; ; D z.t; / (2.58a)
subject to
zP .t; / D g t; ; z.t; /; u.t/ ; t t0 ; (2.58b)
z.t0 ; / D z0 . / (2.58c)
d
D z.t; / D Dz g t; ; z.t; /; u.t/ D z.t; /
dt
CD g t; ; z.t; /; u.t/ ; t t0 ; (2.58d)
can be evaluated as in (2.52) and (2.53b). This yields then deterministic constraints
for the unknown functions t ! z.t; / and t ! D z.t; /; t t0 .
Remark 2.16 The expectations EH uı ./ ; EH arising in .P /tuı ./;; ; .P /t; , resp., can
be determined approximatively as described above.
Chapter 3
Stochastic Optimal Open-Loop Feedback
Control
feedback control law by minimization of the expected total costs, hence, a stochastic
optimal control law.
As mentioned above, in active control of dynamic structures, cf. [18, 117, 154–
157, 171], the behavior of the m-vector q D q.t/ of displacements with respect to
time t is described by a system of second order linear differential equations for q.t/
having a right hand side being the sum of the stochastic applied load process and
the control force depending on a control n-vector function u.t/:
M qR C D qP C Kq.t/ D f t; !; u.t/ ; t0 t tf : (3.1a)
Hence, the force vector f D f t; !; u.t/ on the right hand side of the dynamic
equation (3.1a) is given by the sum
fa D u u (3.1c)
with
0 I 0
A WD ; B WD ; (3.2b)
M K M 1 D
1
M 1 u
0
b.t; !/ WD (3.2c)
M 1 f0 .t; !/:
3.1 Dynamic Structural Systems Under Stochastic Uncertainty 81
In the optimal design of regulators for dynamic systems, see also Chap. 4, the
(m-)vector q D q.t/ of tracking errors is described by a system of 2nd order linear
differential equations:
Here, M.t/; D.t/; K.t/; Y .t/ denote certain time-dependent Jacobians arising
from the linearization of the dynamic equation around the stochastic optimal
reference trajectory and the conditional expectation pD of the vector of dynamic
parameters pD .!/. The deviation between the vector of dynamic parameters pD .!/
and its conditional expectation pD is denoted by
p D .!/ WD pD .!/ pD .
Furthermore,
u.t/ denotes the correction of the feedforward control u0 D u0 .t/.
By introducing appropriate matrices, system (2.1a, b) can be represented by the
1st order system of linear differential equations:
zP D A.t/z.t; !/ C B
u C b.t; !/ (3.4a)
with
0 I 0
A.t/ WD ; B WD ; (3.4b)
M.t/1 K M.t/1 D.t/ M.t/1
0
b.t; !/ WD : (3.4c)
M.t/1 Y .t/
pD .!/
According to the description in Sect. 2.2, a feedback control is defined, cf. (2.10b),
by
where It denotes again the total information about the control system up to time t
and '.; / designates the feedback control law. If the state zt WD z.t/ is available at
each time point t, the control input n-vector function u D u.t/,
u D
u.t/, resp.,
can be generated by means of a PD-controller, hence,
u.t/.
u.t// WD ' t; z.t/ ; t t0 ; (3.5b)
with a feedback control law ' D '.t; q; q/ P D '.t; z.t//. Efficient approximate
feedback control laws are constructed here by using the concept of open-loop
feedback control. Open-loop feedback control is the main tool in model predictive
control, cf. [1, 101, 136], which is very often used to solve optimal control problems
in practice. The idea of open-loop feedback control is to construct a feedback control
law quasi argument-wise, see cf. [2, 80].
A major issue in optimal control is the robustness, cf. [43], i.e. the insensitivity
of the optimal control with respect to parameter variations. In case of random
parameter variations, robust optimal controls can be obtained by means of stochastic
optimization methods, cf. [100]. Thus, we introduce the following concept of an
stochastic optimal (open-loop) feedback control.
Definition 3.1 In case of stochastic parameter variations, robust, hence, parameter-
insensitive optimal (open-loop) feedback controls obtained by stochastic optimiza-
tion methods are also called stochastic optimal (open-loop) feedback controls.
t tf , is determined first on the remaining time interval Œtb ; tf , see Fig. 3.1, by
stochastic optimization methods, cf. [100] .
Having a stochastic optimal open-loop control u D u tI tb ; Itb ; tb t
tf , on each remaining time interval Œtb ; tf with an arbitrary starting time tb , t0
tb tf , a stochastic optimal open-loop feedback control law is then defined, see
Definition 2.3, as follows:
Definition 3.2
' D ' tb ; Itb WD u .tb / D u tb I tb ; Itb ; t0 tb tf : (3.5c)
Hence, at
time t D tb just the “first” control value
u .tb / D u tb I tb ; Itb of
u I tb ; Itb is used only. For each other argument t; It the same construction is
applied.
For finding stochastic optimal open-loop controls, based on the methods devel-
oped in Chap. 2, on the remaining time intervals tb t tf with t0 tb tf , the
stochastic Hamilton function of the control problem is introduced. Then, the class
of H minimal controls, cf. Definitions 2.6 and 2.7, can be determined in case
of stochastic uncertainty by solving a finite-dimensional stochastic optimization
problem for minimizing the conditional expectation of the stochastic Hamiltonian
subject to the remaining deterministic control constraints at each time point t.
Having a H minimal control, the related two-point boundary value problem
with random parameters can be formulated for the computation of a stochastic
optimal state- and costate-trajectory. Due to the linear-quadratic structure of the
underlying control problem, the state and costate trajectory can be determined
analytically to a large extent. Inserting then these trajectories into the H-minimal
control, stochastic optimal open-loop controls are found on an arbitrary remaining
time interval. According to Definition 3.2, these controls yield then immediately a
stochastic optimal open-loop feedback control law. Moreover, the obtained controls
can be realized in real-time, which is already shown for applications in optimal
control of industrial robots, cf. [139].
Summarizing, we get optimal (open-loop) feedback controls under stochastic
uncertainty minimizing the effects of external influences on system behavior,
subject to the constraints of not having a complete representation of the system,
cf. [43]. Hence, robust or stochastic optimal active controls are obtained by new
techniques from Stochastic Optimization, see [100]. Of course, the construction can
be applied also to PD- and PID-controllers.
84 3 Stochastic Optimal Open-Loop Feedback Control
The performance function F for active structural control systems is defined , cf.
[95, 98, 101], by the conditional expectation of the total costs being the sum of costs
L along the trajectory, arising from the displacements z D z.t; !/ and the control
input u D u.t; !/ , and possible terminal costs G arising at the final state zf . Hence,
on the remaining time interval tb t tf we have the following conditional
expectation of the total cost function with respect to the information Atb available
up to time tb :
0 1
Ztf ˇ
F WD E @ L t; !; z.t; !/; u.t; !/ dt C G.tf ; !; z.tf ; !//ˇ Atb A : (3.6a)
tb
1 T 1
L.t; !; z; u/ WD z Q.t; !/z C uT R.t; !/u (3.6b)
2 2
with positive (semi) definite 2m 2m; n n, resp., matrix functions Q D
Q.t; !/; R D R.t; !/. In the simplest case the weight matrices Q; R are fixed.
A special selection for Q reads
Qq 0
QD (3.6c)
0 QqP
P Furthermore,
with positive (semi) definite weight matrices Qq ; QqP , resp., for q; q.
G D G.tf ; !; z.tf ; !// describes possible terminal costs. In case of endpoint
control G is defined by
1
G.tf ; !; z.tf ; !// WD .z.tf ; !/ zf .!//T Gf .z.tf ; !/ zf .!//; (3.6d)
2
According to (2.28a), (2.40a), see also [101], the stochastic Hamiltonian H related
to the stochastic optimal control problem (3.7a–d) reads:
with
ˇ
h.t/ D h tI tb ; Itb WD E B.!/T y.t; !/ˇAtb ; t tb :
(3.8c)
defined, see Definitions 2.6 and 2.7 and cf. also [101], for tb t tf as a solution
of the following convex stochastic optimization problem, cf. [100]:
ˇ
min E H.t; !; z.t; !/; y.t; !/; u/ˇAtb (3.9a)
s.t.
u 2 Dt ; (3.9b)
is defined by
T
ue .t; h I tb ; Itb / WD argmin C.u/ C h tI tb ; Itb u for t tb : (3.10b)
u2Dt
of the stochastic optimal control problem (3.7a–d), can be obtained, see also [101],
by solving the following stochastic two-point boundary value problem related to
(3.7a–d):
Theorem 3.1 If z D z .t; !/; y D y .t; !/; t0 t tf , is a solution of
.b/
zP.t; !/ D Az.t; !/ C BrC 1 B.!/T y.t; !/ C b.t; !/; tb t tf
(3.13a)
z.tb ; !/ D zb .b/ (3.13b)
88 3 Stochastic Optimal Open-Loop Feedback Control
Having (3.14a, b), for the state trajectory z D z.t; !/ we get, see (3.13a, b), the
following system of ordinary differential equations
.b/
zP.t; !/ D Az.t; !/ C BrC 1 B T e A .tf t / rz G.tf ; !; z.tf ; !//
T
Zt
A.t tb /
z.t; !/ D e zb .b/
C e A.t s/ b.s; !/
tb
!
.b/
1
B T e A .tf s/ rz G.tf ; !; z.tf ; !//
T
CBrC ds;
tb t tf : (3.16)
3.6 Minimum Energy Control 89
Ztf
A.tf tb /
z.tf ; !/ D e zb .b/
C e A.tf s/ b.s; !/
tb
!
.b/
CBrC 1 B T e AT .t f s/
rz G.tf ; !; z.tf ; !// ds : (3.17)
In the case of endpoint control, the terminal cost function is given by the following
definition (3.18a), where zf D zf .!/ denotes the desired—possible random—final
state:
1
G.tf ; !; z.tf ; !// WD kz.tf ; !/ zf .!/k2 : (3.18a)
2
Hence,
and therefore
.b/ .b/
rG.tf ; !; z.tf ; !// D z.tf ; !/ zf .b/ (3.18c)
ˇ ˇ
D E z.tf ; !/ˇAtb E zf ˇAtb :
Thus
Ztf
A.tf tb /
z.tf ; !/ D e zb .b/
C e A.tf s/ b.s; !/
tb
!
.b/
CBrC 1 B T e AT .t f s/
z.tf ; !/ zf .b/ ds : (3.19a)
90 3 Stochastic Optimal Open-Loop Feedback Control
Taking expectations E.: : : jAtb / in (3.19a), we get the following condition for
.b/
z.tf ; !/ :
Ztf
.b/ A.tf tb / .b/
z.tf ; !/ De zb .b/
C e A.tf s/ b.s; !/ ds
tb
Z tf
.b/
e A.tf s/ BrC 1 B T e A .tf s/ z.tf ; !/ zf .b/
T
C ds :
tb
(3.19b)
1 T
C.u/ D u Ru ; (3.20a)
2
hence,
rC D Ru (3.20b)
and therefore
Ztf
.b/ .b/
z.tf ; !/ D e A.tf tb / zb .b/ C e A.tf s/ b.s; !/ ds
tb
Ztf
.b/
e A.tf s/ BR1 B T e A s/
T .t
f
dsz.tf ; !/
tb
Ztf
.b/
e A.tf s/ BR1 B T e A s/
T .t
C f
dszf .!/ : (3.21)
tb
3.6 Minimum Energy Control 91
Define now
Ztf
e A.tf s/ BR1 B T e A s/
T .t
U WD f
ds : (3.22)
tb
.I C U / w D 0:
U w D I w D w D .1/w;
Ztb
.b/ A.tf tb / .b/
.I C U / z.tf ; !/ De zb .b/
C e A.tf s/ b.s; !/ ds C U zf .b/ ;
tb
(3.23a)
hence,
Ztb
.b/ 1 A.tf tb / 1 .b/
z.tf ; !/ D .I C U / e zb C .I C U / e A.tf s/ b.s; !/ ds
tb
1
C .I C U / U zf .b/
: (3.23b)
.b/ .b/
rz G.tf ; !; z.tf ; !// D z.tf ; !/ zf D z.tf ; !/ zf .b/
D .I C U /1 e A.tf tb / zb .b/
Ztf
1 .b/
C .I C U / e A.tf s/ b.s; !/ ds
tb
C .I C U /1 U I zf .b/ : (3.24)
92 3 Stochastic Optimal Open-Loop Feedback Control
Ztf
1 .b/
C .I C U / e A.tf s/ b.s; !/ ds
tb
!
1
C .I C U / U I zf .b/
; tb t tf :
(3.25)
Finally, the stochastic optimal open-loop feedback control law ' D '.t; It / is then
given by
D R1 B T e A tb /
.I C U /1 e A.tf tb / zb .b/
T .t
f
Ztf
1 T AT .tf tb / 1 .b/
R B e .I C U / e A.tf s/ b.s; !/ ds
tb
R1 B T e AT .t f tb /
.I C U /1 U I zf .b/ (3.26)
.b/ .b/
with Itb WD zb .b/ WD z.tb / ; b.; !/ ; zf .b/ .
Replacing tb ! t, we find this result:
Theorem 3.2 The stochastic optimal open-loop feedback control law ' D '.t; It /
is given by
.t /
'.t; It / D R1 B T e A t /
.I C U /1 e A.tf t / z.t/
T .t
f
„ ƒ‚ …
0 .t /
Ztf
1 T AT .tf t / 1 .t /
R B e .I C U / e A.tf s/ b.s; !/ ds
t
„ ƒ‚ …
.t /
1 .t;b.;!/ /
R1 B T e AT .t f t /
.I C U /1 U I zf .t / ; (3.27a)
„ ƒ‚ …
2 .t /
3.6 Minimum Energy Control 93
hence,
.t / .t /
'.t; It / D 0 .t/z.t/ C 1 .t; b.; !/ / C 2 .t/zf .t / ;
.t / .t /
It WD z.t/ ; b.; !/ ; zf .t / : (3.27b)
.t /
Remark 3.2 Note that the stochastic optimal open-loop feedback law z.t/ 7!
'.t; It / is not linear in general, but affine-linear.
In this section we consider more general terminal cost functions G. Hence, suppose
Consequently,
ue .t; h.t// D v .h.t// D rC 1 B T e A .tf t / r.z.tf ; !/ zf .!//
T .b/
(3.29a)
Ztf
z.tf ; !/ D e A.tf tb / zb C e A.tf s/ b.s; !/ ds
tb
Ztf
.b/
e A.tf s/ BrC 1 B T e A .tf s/ r.z.tf ; !/ zf .!//
T
C ds ;
tb
tb t tf : (3.29b)
Special Case:
Now a special terminal cost function is considered in more detail:
X
2m
.z zf / W D .zi zf i /4 (3.30a)
i D1
T
r.z zf / D 4 .z1 zf 1 /3 ; : : : ; .z2m zf 2m /3 : (3.30b)
94 3 Stochastic Optimal Open-Loop Feedback Control
Here,
.b/ T
r.z zf / D 4 E .z1 zf 1 /3 jAtb ; : : : ; E .z2m zf 2m /3 jAtb
T
.b/ .b/
D 4 m3 .z1 .tf ; /I zf 1 .//; : : : ; m3 .z2m .tf ; /I zf 2m .//
.b/
DW 4m3 .z.tf ; /I zf .// : (3.31)
Thus,
Ztf
A.tf tb /
z.tf ; !/ D e zb C e A.tf s/ b.s; !/ ds
tb
Ztf
e A.tf s/ BrC 1 B T e A .tf s/ 4m3 .z.tf ; /I zf .// ds :
T .b/
C
tb
„ ƒ‚ …
.b/
J m3 .z.tf ;/Izf .//
(3.32)
Remark 3.3
ˇ
3 ˇ ˇˇ
ˇ
E z.tf ; !/ zf .!/ Atb ˇ
ˇ
cbyc
3
D E.b/ z.tf ; !/ z.b/ .tf / C z.b/ .tf / zf .!/
3.6 Minimum Energy Control 95
3 2
D E.b/ z.tf ; !/ z.b/ .tf / C 3 z.tf ; !/ z.b/ .tf / z.b/ .tf / zf .!/
!
2 3
C3 z.tf ; !/ z.b/ .tf / z.b/ .tf / zf .!/ C z.b/ .tf / zf .!/ : (3.33c)
.b/ 3 .b/
D m3 .z.tf ; // C 3 2.b/ .z.tf ; //.z.b/ .tf / zf .b/ / C z.b/ .tf / zf .!/ ;
(3.33d)
where 2.b/ .z.tf ; / denotes the conditional variance of the state reached at the final
time point tf , given the information at time tb .
1
.!/ z.tf ; !/ zf .!/ 2 :
G.tf ; !; z.tf ; !// WD (3.34a)
2
This yields
hence,
T .t
t /
.b/
y .b/ .t/ D e A f
.!/T .!/.z.tf ; !/ .!/zf .!//
.b/ .b/
D e A .tf t / .!/T .!/z.tf ; !/ .!/T .!/zf .!/
T
: (3.35b)
96 3 Stochastic Optimal Open-Loop Feedback Control
Zt
A.t tb /
z.t; !/ D e zb C e A.t s/ b.s; !/
tb
.b/
C BrC 1 B T e A s/
T .t
f
.!/T .!/z.tf ; !/
!!
.b/
.!/T .!/zf .!/ ds: (3.37a)
Assume that the control costs and its gradient are given by
1 T
C.u/ D u Ru; rC.u/ D Ru : (3.37b)
2
Here, (3.37a) yields
Ztf
A.tf tb /
z.tf ; !/ D e zb C e A.tf s/ b.s; !/
tb
!
.b/ .b/
BR1 B T e A .tf s/ .!/T .!/z.tf ; !/ .!/T .!/zf .!/
T
ds :
(3.37c)
3.6 Minimum Energy Control 97
Multiplying with .!/T .!/ and taking expectations, from (3.37c) we get
Ztf
.b/ .b/ A.t t / .b/
T z.tf ; !/ D T ef b
zb .b/
C T e A.tf s/ b.s; !/ ds
tb
Ztf
.b/
e A.tf s/ BR1 B T e A s/
T .t
T f
ds
tb
.b/ .b/
.!/T .!/z.tf ; !/ .!/T .!/zf .!/ : (3.38a)
Ztf
e A.tf s/ BR1 B T e A s/
T .t
U D f
ds :
tb
Ztf
.b/ A.t t / .b/
D T ef b
zb .b/ C T e A.tf s/ b.s; !/ ds
tb
.b/ .b/
C T U .!/T .!/zf .!/ : (3.38b)
.b/
Lemma 3.3 I C T U is regular.
.b/
Proof First notice that not only U , but also T is positive semidefinite:
.b/
Then their product T U is positive semidefinite as well. This follows immedi-
ately from [122] as .!/T .!/ is symmetric.
98 3 Stochastic Optimal Open-Loop Feedback Control
.b/
Since the matrix I C T U is regular, we get cf. (3.23a, b),
.b/
.b/
1 .b/
T z.tf ; !/ D I C T U T e A.tf tb / zb .b/
1 Z f
t
.b/ .b/
C I C T U T e A.tf s/ b.s; !/ ds
tb
.b/
1 .b/ .b/
C I C T U T U zf .!/ : (3.38c)
Putting (3.38c) into (3.36), corresponding to (3.25) we get the stochastic optimal
open-loop control
.b/
u .tI tb ; Itb / D R1 B T e A t /
T .t
.!/T .!/z.tf ; !/
f
.b/
.!/T .!/zf .!/
D : : : ; tb t tf ; (3.39)
which yields then the related stochastic optima open-loop feedback control ' D
'.t; It / law corresponding to Theorem 3.2.
P !/ D AT y.t; !/ Qz.t; !/
y.t;
y.tf ; !/ DrG.tf ; !; z.tf ; !// ;
which has the following solution for given z.t; !/ and rG.tf ; !; z.tf ; !//:
Ztf
T .st / T .t t /
y.t; !/ D eA Qz.s; !/ ds C e A f
rG.tf ; !; z.tf ; !//: (3.40)
t
Indeed, we get
Ztf
T .st / T .t t /
AT e A Qz.s; !/ ds AT e A f
rG.tf ; !; z.tf ; !//
t
T 0
D eA Qz.t; !/
0 tf 1
Z
AT @ e A .st / Qz.s; !/ ds C e A .tf t / rG.tf ; !; z.tf ; !//A
T T
D A y.t; !/ Qz.t; !/ :
T
.b/
The unknown function z.t/ , and the vector z.tf ; !/ in this equation are both
.b/
given, based on y.t/ , by the initial value problem, see (3.13a, b),
zP.t; !/ DAz.t; !/ C BrC 1 B T y .b/ .t/ C b.t; !/ (3.42a)
Taking expectations, considering the state vector at the final time point tf , resp.,
yields the expressions:
Zt
A.t tb / .b/
z .t/ De
.b/
zb .b/
C e A.t s/ b.s/ C BrC 1 B T y .b/ .s/ ds;
tb
(3.43a)
Ztf
A.tf tb /
z.tf ; !/ De zb C e A.tf s/ b.s; !/ C BrC 1 B T y .b/ .s/ ds:
tb
(3.43b)
100 3 Stochastic Optimal Open-Loop Feedback Control
According to (3.12) and (3.11c), in the present case the stochastic optimal open-
loop control is given by
ˇ
u .tI tb ; Itb / D ue .t; h.t// D R1 E B T y.t; !/ˇAtb D R1 B T y.t/ :
.b/
(3.44a)
.b/
Hence, we need the function y .b/ D y.t/ . From (3.41) and (3.18a, b) we have
Zf T
t
.b/
.b/ .b/
AT .tf t /
y.t/ De z.tf / zf .b/
C e A .st / Qz.s/ ds : (3.44b)
t
.b/ t /
e A.tf tb / zb .b/ zf .b/
T .t
y.t/ De A f
Ztf !
.b/ .b/
A.tf s/ 1
C e b.s/ BR B y.s/ T
ds
tb
Ztf
T .st /
C eA Q e A.stb / zb .b/
t
Zs !
.b/ .b/
C e A.s / b./ BR1 B T y./ d ds : (3.44c)
tb
1
cB < r : (3.45)
.tf t0 /cQ
cA cR1 .tf t0 / 1 C 2
3.7 Nonzero Costs for Displacements 101
Here,
t /
e A.tf tb / zb .b/ zf .b/
T .t
.T f /.t/ De A f
Ztf !
.b/
A.tf s/ 1
C e b.s/ BR B f .s/ dsT
tb
Ztf
T .st /
C eA Q e A.stb / zb .b/
t
Zs !
.b/
1 T
C e A.s /
b./ BR B f ./ d ds : (3.47)
tb
kT f T gk
(
Ztf
AT .t t /
D sup e f e A.tf s/ BR1 B T .g.s/ f .s// ds
tb t tf
tb
9
Ztf Zs =
C e AT .st /
Q e A.s /
BR B .g./ f .// d ds
1 T
;: (3.48a)
t tb 2
102 3 Stochastic Optimal Open-Loop Feedback Control
Note that the Frobenius norm is submultiplicative and compatible with the Euclidian
norm. Using these properties, we get
kT f T gk
( Ztf
sup cA cA cB cR1 cB kf .s/ g.s/k2 ds
tb t tf
tb
Ztf Zs )
C cA cQ cA cB cR1 cB kf ./ g./k2 d ds
t tb
( Ztf
sup cA cA cB cR1 cB sup kf .s/ g.s/k2 ds
tb t tf tb t tf
tb
Ztf Z s )
C cA cQ cA cB cR1 cB sup kf ./ g./k2 d ds
tb t tf
t tb
n cQ o
Dkf gkcA2 cB2 cR1sup .tf tb / C .tf tb /2 .t tb /2
tb t tf 2
cQ
kf gkcA2 cB2 cR1 .tf tb /.1 C .tf tb // : (3.48b)
2
Thus, T is a contraction if
1
cB2 < cQ (3.48c)
cA2 cR1 .tf tb / 1 C .t
2 f
tb /
and therefore
1
cB < q : (3.48d)
cQ
cA cR1 .tf tb / 1 C .t
2 f
tb /
1
cB < r : (3.48e)
.tf t0 /cQ
cA cR1 .tf t0 / 1 C 2
3.8 Stochastic Weight Matrix Q D Q.t; !/ 103
Remark 3.4 Condition (3.48e) holds if the matrix in (3.1c) has a sufficiently
small Frobenius norm. Indeed, according to (3.2b) we have
0
BD
M 1
and therefore
cB D kBkF D kM 1 kF kM 1 kF k kF :
.b/
u .tI tb ; Itb / D R1 B T y.t/ : (3.49a)
Moreover,
In the following we consider the case that, cf.(2.6h, i), the weight matrix for the
evaluation of the displacements z D z.t; !/ is stochastic and may depend also on
time t; t0 t tf . In order to take into account especially the size of the additive
disturbance term b D b.t; !/, cf.(3.7b), in the following we consider the stochastic
weight matrix
1
G.tf ; !; z.tf ; !// WD ˇ.tf ; !/kz.tf ; !/ zf .!/k2 ; (3.50c)
2
for the adjoint variable y D y.t; !/, we have the boundary value problem
Ztf
T .st /
y.t; !/ D eA Qˇ.s; !/z.s; !/ ds
t
T .t t /
Ce A f
ˇ.tf ; !/z.tf ; !/ ˇ.tf ; !/zf .!/ ; tb t tf
(3.52a)
Ztf
T .st / .b/
C eA Qˇ.s; /z.s; / ds; t tb (3.52b)
t
Since the matrices A; B are assumed to be fixed, see (3.7b), from (3.8c) and (3.52b)
we get
h.t/ DE B T y.t; !/jAtb D B T y .b/ .t/; t tb (3.53a)
Consequently, corresponding to (3.11c) and (3.12), with (3.53a) the optimal open-
loop control u D u .t/ is given then by
.b/
t ! ˇ.t; /z.t; / D E ˇ.t; !/z.t; !/jAtb ; t tb ; (3.54a)
we multiply (3.13a, b) by ˇ.tb ; !/. Thus, the trajectory t ! ˇ.tb ; !/z.t; !/, t tb ,
is the solution of the initial value problem
d
ˇ.tb ; !/z.t; !/ DAˇ.tb ; !/z.t; !/ BR1 Bˇ.tb ; !/y .b/ .t/ C ˇ.tb ; !/b.t; !/
dt
(3.55a)
ˇ.tb ; !/z.tb ; !/ Dˇ.tb ; !/zb : (3.55b)
Taking conditional expectations of (3.55a, b) with respect to Atb , for the approx-
imate weighted conditional mean trajectory (3.54b) we obtain the initial value
problem
.b/
where ˇ .t/ WD E ˇ.t; !/jAtb , t tb . Consequently, the approximate weighted
conditional mean trajectory (3.54b) can be represented, cf. (3.43a, b), by
.b/ .b/
ˇ.tb ; /z.t; / De A.t tb / ˇ.tb ; /zb
Zt
.b/ .b/
C e A.t s/ ˇ.tb ; /b.s; / BR1 B T ˇ .tb /y .b/ .s/ ds;
tb
tb t tf : (3.57)
.b/
t ! ˇ.tf ; /z.tf ; /
Inserting now (3.57) into (3.52b), corresponding to (3.44c) we find the following
approximate fixed point condition for the conditional mean adjoint trajectory t 7!
y .b/ .t/; tb t tf , needed in the representation (3.53b) of the stochastic optimal
open-loop control u D u .t/; tb t tf :
.b/ .b/
T .t t /
y .b/ .t/ De A f
ˇ.tf ; /z.tf ; / ˇ.tf ; /zf ./
Ztf
T .st / .b/
C eA Qˇ.s; /z.s; / ds
tb
t / .b/ .b/
e A.tf tb / ˇ.tb ; /zb ./
T .t
e A f
ˇ.tf ; /zf ./
Ztf .b/
C e A.tf s/ ˇ.tb ; /b.s; /
tb
!
BR1 B T .b/ .tb /y .b/ .s/ ds
Ztf
T .st / .b/
C eA Q e A.stb / ˇ.tb ; /zb ./
t
Zs .b/
C e A.s / ˇ.tb ; /b.; /
tb
!
.b/
1
BR B ˇ .tb /y T .b/
./ d ds (3.58)
1
cB < r : (3.59)
.tf t0 /cQ
cA cR1 .b/ .tb /.tf t0 / 1 C 2
3.9 Uniformly Bounded Sets of Controls Dt ; t0 t tf 107
Here again,
Moreover,
The above shown Theorem 3.4 guaranteeing the existence of a solution of the fixed
point condition (3.58) can be generalized considerably if we suppose that the sets
Dt of feasible controls u D u.t/ are uniformly bounded with respect to the time
t; t0 t tf . Hence, in the following we suppose again:
• time-independent and deterministic matrices of coefficients, hence
1 T 1 T
C.u/ D u Ru Q.z/ D z Qz
2 2
1
z.tf ; !/ zf .!/2
G.tf ; !; z.tf ; !/ D 2
(3.61b)
2
• uniformly bounded sets of feasible controls, hence, we assume that there exists a
constant CD 0 such that
[
kuk2 cD for all u 2 Dt : (3.61c)
t 2T
108 3 Stochastic Optimal Open-Loop Feedback Control
Thus, the integral form of the related 2-point boundary value problem reads:
Zt
Az.!; s/ C b.s; !/ C B ue .s; y.s/
.b/
z.!; t/ Dzb C ds (3.63a)
tb
Ztf
T
y.!; t/ D z.tf ; !/ zf .!/ C A y.!; s/ C Qz.s; !/ ds: (3.63b)
t
Zt
Az.s/ C b.s/ C B ue .s; y.s/
.b/ .b/ .b/ .b/
z.t/ Dzb .b/
C ds (3.64a)
tb
Ztf
.b/ .b/ .b/ .b/
y.t/ D z.tf / zf .b/
C AT y.s/ C Qz.s/ ds: (3.64b)
t
Zt
e A.t s/ b.s/ C B ue .s; y.s/
.b/ A.t tb / .b/ .b/
z.t/ De zb .b/
C ds (3.65a)
tb
Zf T
t
.b/
.b/ .b/
AT .t f t /
y.t/ De z.tf / zf .b/
C e A .st / Qz.s/ ds: (3.65b)
t
.b/
Putting (3.65a) into (3.65b), for y.t/ we get then the following fixed point
condition
.b/
.b/
y.t/ D e A .tf t / z.tf / zf .b/
T
0 1
Ztf Zs
e A.s / b./ C B ue .; y./
.b/ .b/
C e AT .st /
Q @e A.stb / zb .b/ C d A ds
t tb
(3.66)
3.9 Uniformly Bounded Sets of Controls Dt ; t0 t tf 109
For the consideration of the existence of a solution of the above fixed point
equation (3.66) we need several auxiliary tools. According to the assumption
(3.61c), next we have the following lemma:
Lemma 3.4 There exist constants cz ; cG > 0 such that
.b/ .b/
kz.f ./; t/ k cz and krz G.z.f ./; t/ k cG (3.67a)
Zt
e A.t s/ b.s/ C B ue .f .s/; s/ ds
.b/ A.t tb / .b/
z.f ./; t/ WD e zb .b/
C (3.67b)
tb
.b/
Proof With cA WD e kAkF .tf tb / ; cB WD kBkF and cb .b/ WD kb./ k1 the following
inequalities hold:
.b/
kz.f ./; t/ k2 cA kzb .b/ k2 C cb .b/ C cB cD .tf tb / cz (3.68a)
.b/ .b/
krz G.z.f ./; t/ k2 D kz.f ./; tf / zf .b/ k cz C kzf .b/ k2 cG ; (3.68b)
• TQ .X / is bounded:
T
A .tf t / .b/
e z.tf / zf .b/
0
Ztf Zs .b/
C e AT .st /
Q @e A.stb / zb .b/ C e A.s / b./
t tb
1
C B ue .; f .// d A ds
!
tf2 tb2
cA cG C cA cQ cA kzb .b/ k2 .tf tb / C .cA cb .b/ C cB cD / (3.70)
2
• TQ .X / is equicontinuous:
We have to show that for each > 0 there exists a ı > 0 such that, independent
of the mapped function f , the following inequality holds:
Zt
A.t tb / .b/
%.t/ D e zb C e A.t / b./ d; (3.71)
tb
Ztf
T . t /
C eA Q e A. tb / zb
t
1
Z
e A. / b./ C B ue .; f .// dA d
.b/
C
tb
3.9 Uniformly Bounded Sets of Controls Dt ; t0 t tf 111
Ztf
T . s/
eA Q e A. tb / zb
s
1
Z
b./ C B ue .; f .// dA d
.b/
C e A. /
(3.72a)
tb
0 1
Ztf Z
C e AT . t /
Q @%./ C B ue .; f .// dA d
t tb
0 1
Ztf Z
e
e AT . s/
Q @%./ C B u .; f .// dA d
s tb 2
T
e A .tf t / e A .tf s/ cG
T
2
0 1
Ztf Z
AT .t /
C e A Q @%./ C B ue .; f .// dA d
T T
e e A .s/
t tb
0 1
Z t Z
e A .s/ e A Q @%./ C B ue .; f .// dA d
T T
s tb 2
T
A .tf t / AT .tf s/
e e cG
2
T
A .t /
e A .s/ e kAkF tf cQ .c% C cB cD .tf tb //.tf tb /
T
C e
Obviously, the final expression in (3.72b) is independent of f ./. Hence, due to the
continuity of the matrix exponential function and the function %./, the assertion
follows. t
u
112 3 Stochastic Optimal Open-Loop Feedback Control
where
!
tf2 tb2
C WD cA cG C cA cQ cA kzb k2 .tf tb / C .cA cb .b/ C cB cD / : (3.73b)
2
T W M ! M; f 7! TQ f: (3.74)
Obviously, the operator T is continuous and, according to Lemma 3.5, the image of
T is relative compact. Moreover, the set M is closed and convex. Hence, according
to the fixed point theorem of Schauder, T has a fixed point in M. t
u
According to the previous sections, the remaining problem is then to solve the fixed
point equation (3.44c) or (3.58). In the first case, the corresponding equation reads:
t /
Gf e A.tf tb / zb .b/ zf .b/
T .t
y .b/ .t/ De A f
Ztf !
.b/
C e A.tf s/ b.s/ BR1 B T y .b/ .s/ ds
tb
Ztf
T .st /
C eA Q e A.stb / zb .b/
t
Zs !
.b/
1
C e A.s /
b./ BR B y T .b/
./ d ds : (3.75)
tb
3.10 Approximate Solution of the Two-Point Boundary Value Problem (BVP) 113
Based on the given stochastic open-loop feedback control approach, the fixed
point equation (3.75) can be solved by the following approximation method:
Step I
According to Eqs. (3.12) and (3.44a) of the stochastic optimal OLF C , for each
remaining time interval Œtb ; tf the value of the stochastic optimal open-loop control
u D u .tI tb ; Itb /; t tb , is needed at the left time point tb only.
Thus, putting first t D tb in (3.75), we get:
tb /
Gf e A.tf tb / zb .b/ e A tb /
T .t T .t
y .b/ .tb / De A f f
Gf zf .b/
Ztf
AT .tf tb / .b/
Ce Gf e A.tf s/ b.s/ ds
tb
Ztf
AT .tf tb /
e Gf e A.tf s/ BR1 B T y .b/ .s/ ds
tb
Ztf
T .st
C eA b/
Qe A.stb / dszb .b/
tb
Ztf Zs
AT .stb / .b/
C e Q e A.s / b./ d ds
tb tb
Ztf Zs
e AT .stb /
Q e A.s / BR1 B T y .b/ ./ d ds : (3.76a)
tb tb
Step II
Due to the representation (3.44a) of the stochastic optimal open-loop control u and
the stochastic OLF construction principle (2.10d, e), the value of the conditional
mean adjoint variable y .b/ .t/ is needed at the left boundary point t D tb only.
Consequently, y b D y .b/ .s/ is approximated on Œtb ; tf by the constant function
This approach is justified especially if one works with a receding time horizon or
moving time horizon
tf WD tb C
tb /
Gf e A.tf tb / zb .b/ e A tb /
T .t T .t
y .b/ .tb / e A f f
Gf zf .b/
Ztf
AT .tf tb / .b/
Ce Gf e A.tf s/ b.s/ ds
tb
AT .tf tb /
.tf tb /e Gf e A.tf tb / BR1 B T y .b/ .tb /
Ztf
T .st
C eA b/
Qe A.stb / dszb .b/
tb
Ztf Zs
AT .stb / .b/
C e Q e A.s / b./ d ds
tb tb
Ztf
Qe A.stb / ds BR1 B T y .b/ .tb /:
T .st
.s tb /e A b/
(3.76c)
tb
Step III
Rearranging terms, (3.76c) yields a system of linear equations for y .b/ .tb /:
T .t tb /
y .b/ .tb / A0 ..tb ; tf ; Gf ; Q/zb .b/ e A f
Gf zf .b/
.b/
C A1 .tb ; tf ; Gf ; Q/ b Œtb ;tf ./
.b/
where the matrices, linear operator and function, resp., A0 ; A1 ; A23 ; b Œtb ;tf can be
easily read from relation (3.76c). Consequently, (3.76d) yields
I C A23 .tb ; tf ; Gf ; Q/BR1 B T y .b/ .tb / A0 ..tb ; tf ; Gf ; Q/zb .b/
T .t tb / .b/
e A f
Gf zf .b/ C A1 .tb ; tf ; Gf ; Q/ b Œtb ;tf ./: (3.76e)
3.11 Example 115
tb /
Gf e A.tf tb /
T .t
A23 .tb ; tf ; Gf ; Q/ D .tf tb /e A f
Ztf
T .st
C .s tb /e A b/
Qe A.stb / ds :
tb
(3.76f)
According to (3.75) and (2.10d, e), the stochastic optimal open-loop feedback
control ' D ' .tb ; Itb / at t D tb is obtained as follows:
Theorem 3.7 With the approximative solution y .b/ .tb / of the fixed point condition
(3.75) at t D tb , represented by (3.76f), the stochastic optimal open-loop feedback
control law at t D tb is given by
3.11 Example
We consider the structure according to Fig. 3.2, see [18], where we want to control
the supplementary active system while minimizing the expected total costs for the
control and the terminal costs.
116 3 Stochastic Optimal Open-Loop Feedback Control
with
m0 0
M D mass matrix (3.78a)
0 mz
d0 C dz dz
DD damping matrix (3.78b)
dz dz
k0 C kz kz
KD stiffness matrix (3.78c)
kz kz
1
fa .t/ D u.t/ actuator force (3.78d)
C1
f01 .t; !/
f0 .t; !/ D applied load : (3.78e)
0
Here we have n D 1, i.e. u./ 2 C.T; R/, and the weight matrix R becomes a
positive real number.
3.11 Example 117
Zt
s:t: z.t; !/ D zb C Az.s; !/ C Bu.s/ C b.s; !/ ds (3.80b)
tb
1 .b/
zP.t; !/ DAz.t; !/ BB T y.t; / C b.!; t/ (3.81a)
R
P !/ D AT y.t; !/
y.t; (3.81b)
z.tb ; !/ Dzb (3.81c)
y.tf ; !/ DGz.tf ; !/: (3.81d)
118 3 Stochastic Optimal Open-Loop Feedback Control
Hence, the solution of (3.81a)–(3.81d), i.e. the optimal trajectories, reads, cf.
(3.14a), (3.37a),
T .t t /
y.t; !/ De A f
Gz.tf ; !/ (3.82a)
Zt
A.t tb /
z.t; !/ De zb C e A.t s/ b.s; !/
tb
!
1 .b/
BB T e A .tf s/ Gz.tf ; !/
T
ds: (3.82b)
R
R
tb
(3.83)
with
Ztf
1
e A.tf s/ BB T e A s/
T .t
U D f
ds: (3.84)
R
tb
Chapter 4
Adaptive Optimal Stochastic Trajectory
Planning and Control (AOSTPC)
4.1 Introduction
Fig. 4.1 Control of dynamic systems (robots). Here, z.t / WD P / , z.0/ .t /
q.t /; q.t WD
q .0/ .t /; qP .0/ .t /
of model parameters is not a given fixed quantity. The vector p must be represented
therefore by a random vector
on a certain probability space .˝; A; P /, see [8, 58, 123, 148, 169].
Having to control a robotic or more general dynamical system, the control law
u D u.t/, is represented usually by the sum
.0/
optimal geometric path qe .s/; s0 s sf , in configuration space is determined
by using a stochastic optimization approach [93, 109–111, 130]. By means of
.0/
ˇ .0/ .s/ and qe .s/; s0 s sf , we then find a more reliable, robust reference
.0/
trajectory q .0/ .t/; t0 t tf , in configuration space. Applying now the so-called
“inverse dynamics approach” [4, 11, 61], more reliable, robust open-loop controls
.0/
u.0/ .t/; t0 t tf , are obtained. Moreover, by linearization of the dynamic
equation of the robot in a neighborhood of u.0/ .t/; q .0/ .t/; E pM .!/jAt0 ; t
t0 , where At0 denotes the -algebra of informations up to the initial time point t0 ,
a control correction
u.0/ .t/; t t0 , is obtained which is related to the so-called
feedback linearization of a system [11, 61, 131, 150].
At later moments (main correction time points) tj ,
further information on the parameters of the control system and its environment
are available, e.g., by process observation, identification, calibration procedures
etc. Improvements q .j / .t/; u.j / .t/;
u.j /.t/; t tj ; j D 1; 2; : : :, of the preced-
ing reference trajectory q .j 1/ .t/, open loop control u.j 1/ .t/, and local control
correction (feedback control)
u.j 1/.t/ can be determined by replanning, i.e.,
by optimal stochastic trajectory planning (OSTP) for the remaining time interval
t tj ; j D 1; 2; : : :, and by using the information Atj on the robot and its working
environment available up to the time point tj > t0 ; j D 1; 2; : : :, see [62, 141, 142].
According to [11, 126, 148], the dynamic equation for a robot is given by the
following system of second order differential equations
R C h pD ; q.t/; q.t/
M pD ; q.t/ q.t/ P D u.t/; t t0 ; (4.4a)
P WD C.pD ; q; q/
h.pD ; q; q/ P qP C FR .pD ; q; q/
P C G.pD ; q/; (4.4b)
P D C.pD ; q/q,
where C.pD ; q; q/ P and C.pD ; q/ D Cijk .pD ; q/ is the
1i;j;kn
tensor of Coriolis and centrifugal terms, FR D FR .pD ; q; q/ P denotes the vector of
frictional forces and G D G.pD ; q/ is the vector of gravitational forces. Moreover,
u D u.t/ is the vector of controls, i.e., the vector of torques/forces in the joints of
122 4 Adaptive Optimal Stochastic Trajectory Planning and Control (AOSTPC)
the robot. Standard representations of the friction term FR are given [11, 68, 148]
by
P WD Rv .pD ; q/q;
FR .pD ; q; q/ P (4.4c)
P WD R.pD ; q/sgn.q/;
FR .pD ; q; q/ P (4.4d)
T
where sgn.q/ P WD sgn.qP1 /; : : : ; sgn.qPn / . In the first case (4.4c), Rv D
Rv .pD ; q/ is the viscous friction
matrix, and in the Coulomb approach (4.4d),
R D R.pD ; q/ D Ri .p; q/ıij is a diagonal matrix.
Remark 4.1 (Inverse Dynamics) Reading the dynamic equation (4.4a) from the left
to the right hand side, hence, by inverse dynamics [4, 11, 61], the control function
u D u.t/ may be described in terms of the trajectory q D q.t/ in configuration
space.
The relationship between the so-called configuration space fqg of robot coordi-
nates q D .q1 ; : : : ; qn /0 and the work space fxg of world coordinates (position and
orientation of the end-effector) x D .x1 ; : : : ; xn /0 is represented by the kinematic
equation
Ztf
J u./ WD P
L t; pJ ; q.t/; q.t/; u.t/ dt C tf ; pJ ; q.tf /; q.t
P f/ (4.6)
t0
Note that by means of the kinematic equation (4.5), the initial state
.q0 .!/; qP0 .!//in configuration space can be represented by the initial state
x0 .!/; xP0 .!/ in work space.
ii) The terminal conditions
P f / D 0;
tf ; p; q.tf /; q.t (4.7b)
e.g.
Again, by means of (4.5), .qf ; qPf / may be described in terms of the final state
.xf ; xP f / in work space.
Note that more general boundary conditions of this type may occur at some
intermediate time points t0 < 1 < 2 < : : : < r < tf .
iii) Control constraints
Using the kinematic equation (4.5), different types of obstacles in the work
space can be described by (time-invariant) state constraints of the type (4.9a, b).
In robotics [127] often the following state constraints are used:
A special constraint of the type (4.9b) occurs if the trajectory in work space
x.t/ WD T pK ; q.t/ (4.10)
Since the terminal time tf may be given explicitly or implicitly, the trajectory q./
in configuration space may have a varying domain Œt0 ; tf . Hence, in order to work
with a given fixed domain of the unknown functions, the reference trajectory q D
q.t/; t t0 , in configuration space is represented, cf. [68], by
q.t/ WD qe s.t/ ; t t0 : (4.12a)
Here,
s D s.t/; t0 t tf ; (4.12b)
qe D qe .s/; s0 s sf ; (4.12c)
2
ds
ˇ.s/ WD sP t.s/ D
2
t.s/ ; (4.13)
dt
Zs
d
t D t.s/ WD t0 C p : (4.14b)
ˇ./
s0
Z1
d
t.s/ D t0 C .s s0 / r ; s s0 : (4.15a)
0 ˇ s0 C .s s0 /
X
ak
tQ.s/ WD t0 C .s s0 / q ; s s0 : (4.15b)
kD0 ˇ s0 C " 0 C .s s0 2" /
0
k
As long as the basic mechanical equations, the cost and constraint functions do not
depend explicitly on time t, the transformation of the robot control problem from
126 4 Adaptive Optimal Stochastic Trajectory Planning and Control (AOSTPC)
the time onto the s-parameter domain causes no difficulties. In the more general case
one has to use the time representation (4.14b), (4.15a) or its approximates (4.15b, c).
Obviously, the terminal time tf is given, cf. (4.14b), (4.15a), by
Zsf
d
tf D t.sf / D t0 C p (4.16)
ˇ./
s0
Z1
d
D t0 C .sf s0 / r :
0 ˇ s0 C .sf s0 /
1 0
sR D ˇ .s/: (4.17d)
2
Hence, (4.17a–d) yields the following representation
p
P
q.t/ D qe0 .s/ ˇ.s/ (4.18a)
1
R
q.t/ D qe0 .s/ ˇ 0 .s/ C qe00 .s/ˇ.s/ (4.18b)
2
P
of q.t/; R in terms of the new unknown functions qe ./; ˇ./.
q.t/
Inserting now (4.18a, b) into the dynamic equation (4.4a), we find the equivalent
relation
The initial and terminal conditions (4.7a–c) are transformed, see (4.12a, b)
and (4.18a), as follows
p
qe .s0 / D q0 .!/; qe0 .s0 / ˇ.so / D qP 0 .!/ (4.20a)
q
t.sf /; p; qe .sf /; qe0 .sf / ˇ.sf / D 0 (4.20b)
or
q
qe .sf / D qf .!/; qe0 .sf / ˇ.sf / D qP f .!/: (4.20c)
Remark 4.5 In most cases we have the robot resting at time t D t0 and t D tf , i.e.
P 0 / D q.t
q.t P f / D 0, hence,
Using (4.12a, b), the control constraints (4.8a–c) read in s-form as follows:
umin t.s/; pC ue pD ; sI qe ./; ˇ./ umax t.s/; pC ; s0 s sf
p (4.21a)
gI t.s/; pC ; qe .s/; qe0 .s/ ˇ.s/; ue pD ; sI qe ./; ˇ./ 0; s0 s sf
p (4.21b)
gII t.s/; pC ; qe .s/; qe0 .s/ ˇ.s/; ue pD ; sI qe ./; ˇ./ D 0; s0 s sf ;
(4.21c)
where t D t.s/ D t sI ˇ./ or its approximation t D tQ.s/ D tQ sI ˇ./ is defined
by (4.14b), (4.15a–c).
Remark 4.6 I) In the important case
umin .t; pC / WD umin pC ; q.t/; q.t/
P ; umax .t; pC / WD umax pC ; q.t/; q.t/
P
(4.22a)
128 4 Adaptive Optimal Stochastic Trajectory Planning and Control (AOSTPC)
that the bounds for u D u.t/ depend on the system state P
q.t/; q.t/ in
configuration space, condition (4.21a) is reduced to
p
umin pC ; qe .s/; qe0 .s/ ˇ.s/ ue pD ; sI qe ./; ˇ./ (4.22b)
p
umax pC ; qe .s/; qe0 .s/ ˇ.s/ ; s0 s sf :
II) If the bounds for u.t/ in (4.22a) do not depend on the velocity q.t/ P in
configuration space, and the geometric path qe .s/ D qe .s/; s0 s sf , in
configuration space is known in advance, then the bounds
umin pC ; qe .s/ D uQ min .pC ; s/ (4.22c)
umax pC ; qe .s/ D uQ max .pC ; s/; s0 s sf ;
Applying the transformations (4.12a, b), (4.17a) and (4.14b) to the state con-
straints (4.9a, b), we find the following s-form of the state constraints:
p
SI t.s/; pC ; qe .s/; qe0 .s/ ˇ.s/ 0; s0 s sf (4.23a)
p
SII t.s/; pC ; qe .s/; qe0 .s/ ˇ.s/ D 0; s0 s sf : (4.23b)
In the case that the end-effector of the robot has to follow a given path (4.11) in
work space, Eq. (4.23b) reads
T pK ; qe .s/ xe .px ; s/ D 0; s0 s sf ; (4.23e)
with the parameter vector px describing possible uncertainties in the selection of the
path to be followed by the roboter in work space.
4.4 OSTP: Optimal Stochastic Trajectory Planning 129
ds
Applying the integral transformation t D t.s/; dt D p to the integral in the
ˇ.s/
representation (4.6) of the objective function J D J u./ , and transforming also
the terminal costs, we find the following s-form of the objective function:
Zf
s
p ds
J u./ D L .t.s/; pJ ; qe .s/; qe0 .s/ ˇ.s/; ue pD ; sI qe ./; ˇ./ p
ˇ.s/
s0
q
0
C t.sf /; pJ ; qe .sf /; qe .sf / ˇ.sf / : (4.24a)
Zf t Zf s
ds
J u./ WD tf t0 D dt D p : (4.24b)
ˇ.s/
t0 s0
In the following we suppose that the initial and terminal conditions (4.20d) hold, i.e.
Zsf
min LJ pJ ; qe .s/; qe0 .s/; qe00 .s/; ˇ.s/; ˇ 0 .s/ ds C J pJ ; qe .sf /
s0
(4.25a)
s.t.
fIu p; qe .s/; qe0 .s/; qe00 .s/; ˇ.s/; ˇ 0 .s/ 0; s0 s sf (4.25b)
fIIu p; qe .s/; qe0 .s/; qe00 .s/; ˇ.s/; ˇ 0 .s/ D 0; s0 s sf (4.25c)
fIS p; qe .s/; qe0 .s/; ˇ.s/ 0; s0 s sf (4.25d)
fIIS p; qe .s/; qe0 .s/; ˇ.s/ D 0; s0 s sf (4.25e)
ˇ.s/ 0; s0 s sf (4.25f)
p
qe .s0 / D q0 .!/; qe0 .s0 / ˇ.s0 / D qP 0 .!/ (4.25g)
qe .sf / D qf .!/; ˇ.sf / D ˇf : (4.25h)
Here,
LJ D LJ pJ ; qe ; qe0 ; qe00 ; ˇ; ˇ 0 ; J D J .pJ ; qe / (4.26a)
fIu D fIu .p; qe ; qe0 ; qe00 ; ˇ; ˇ 0 /; fIIu D fIIu .p; qe ; qe0 ; qe00 ; ˇ; ˇ 0 / (4.26b)
fIS D fIS .p; qe ; qe0 ; ˇ/; fIIS D fIIS .p; qe ; qe0 ; ˇ/ (4.26c)
are the functions representing the s-form of the objective function (4.24a),
the constraint functions in the control constraints (4.21a–c), and in the state
constraints (4.23a–e), respectively. Define then f u and f S by
fIu fIS
f u WD ; f S WD : (4.26d)
fIIu fIIS
4.4 OSTP: Optimal Stochastic Trajectory Planning 131
In order to get a reliable optimal geometric path qe D qe .s/ in configuration
space and a reliable optimal velocity profile ˇ D ˇ .s/; s0 s sf ,
being robust with respect to random parameter variations of p D p.!/, the
variational problem (4.25a–h) under stochastic uncertainty must be replaced by
an appropriate deterministic substitute problem which is defined according to the
following principles [71, 76, 84, 90, 111], cf. also [70, 71, 90, 107, 108].
Assume first that the a priori information about the robot and its environment up
to time t0 is described by means of a -algebra At0 , and let then
.0/
Pp./ D Pp./ jAt0 (4.27)
denote the a priori distribution of the random vector p D p.!/ given At0 .
Depending on the decision theoretical point of view, different approaches are
possible, e.g. reliability-based substitute problems, belonging essentially to one of
the following two basic classes of substitute problems:
I) Risk(recourse)-constrained minimum expected cost problems
II) Expected total cost-minimum problems.
Substitute problems are constructed by selecting certain scalar or vectorial loss or
cost functions
evaluating the violation of the random constraints (4.25b, c), (4.25d, e), (4.25h’),
respectively.
132 4 Adaptive Optimal Stochastic Trajectory Planning and Control (AOSTPC)
WD ı : (4.28c)
Zsf
min E LJ pJ ; qe .s/; qe0 .s/; qe00 .s/; ˇ.s/; ˇ 0 .s/ jAt0 ds (4.29a)
s0
CE J pJ ; qe .sf / jAt0
s.t.
E fu p; qe .s/; qe0 .s/; qe00 .s/; ˇ.s/; ˇ 0 .s/ jAt0 u ;
s0 s sf (4.29b)
E fS p; qe .s/; qe0 .s/; ˇ.s/ jAt0 S ; s0 s sf (4.29c)
ˇ.s/ 0; s0 s sf (4.29d)
p
qe .s0 / D q 0 ; qe0 .s0 / ˇ.s0 / D qP 0 (4.29e)
qe .sf / D q f .if J D 0/; ˇ.sf / D ˇf ; (4.29f)
Here,
denote scalar or vectorial upper risk bounds which may depend on the path
parameter s 2 Œs0 ; sf . Furthermore, the initial, terminal values q 0 ; qP 0 ; q f
in (4.29e, f) are determined according to one of the following relations:
a)
q 0 WD q.t OP 0 /; q f WD q.t
O 0 /; qP 0 WD q.t O f /; (4.29h)
O
where q.t/; OP
q.t/ denotes an estimate, observation, etc., of the state in
configuration space at time t;
b)
q 0 WD E q0 .!/jAt0 ; qP 0 WD E qP 0 .!/jAt0 ;
q f D qf .0/ WD E qf .!/jAt0 ; (4.29i)
where q0 .!/; qP0 .!/ is a random initial position, and qf .!/ is a random
terminal position.
Having corresponding information about initial and terminal values x0 ; xP 0 , xf
in work space, related equations for q0 ; qP0 ; qf may be obtained by means of the
kinematic equation (4.5).
Remark 4.7 Average Constraints
Taking the average of the pointwise constraints (4.29c, c) with respect to the
path parameter s; s0 s sf , we get the simplified integrated constraints
Zsf
E fu p; qe ; .s/; qe0 .s/; qe00 .s/; ˇ.s/; ˇ 0 .s/ jAt0 ds Q u
s0
(4.29b’)
Zsf
E fS p; qe .s/; qe0 .s/; ˇ.s/ jAt0 ds Q S (4.29c’)
s0
including information about the magnitude ˇ.s/; ˇ 0 .s/ of the motion as well
as information about the direction qe .s/; qe0 .s/; qe00 .s/ of the motion.
s0 s sf ; (4.30a)
min
P q .pC / qe .s/ q max .pC /jAt0 ˛q ; s0 s sf ; (4.30b)
p
P qP min .pC / qe0 .s/ ˇ.s/ qP max .pC /jAt0 ˛qP ; s0 s sf : (4.30c)
1 ˛u ; s0 s sf ; (4.30e)
where ue D ue pD ; sI qe ./; ˇ./ and u .pC /d denotes the diagonal matrix
containing the elements of u .pC / on its diagonal. Moreover, B denotes a
positive definite matrix such that zT Bz 1 for all vectors z such that kzk1 1.
Taking e.g. B D I , (4.30e) reads
E ku .pC /1
d u e p D ; sI qe ./; ˇ./ u c
.p C / k2
jAt 0 1 ˛u ;
s0 s sf : (4.30f)
.0/
initial parameters 0 W t0 ; s0 ; q 0 ; qP 0 ; Pp./ or 0 (4.31a)
and
.0/
the a priori distribution Pp./ may be replaced by a certain vector
!
Y
r
0 WD E pl .!/jAt0 (4.31d)
.l1 ;:::;lr /2
kD1
Zsf
min E LJ p; qe .s/; qe0 .s/; qe00 .s/; ˇ.s/; ˇ 0 .s/ jAt0 ds
s0
CE J p; qe .sf / jAt0 (4.32a)
136 4 Adaptive Optimal Stochastic Trajectory Planning and Control (AOSTPC)
s.t.
ˇ.s/ 0; s0 s sf (4.32b)
p
qe .s0 / D q 0 ; qe0 .s0 / ˇ.s0 / D qP 0 (4.32c)
qe .sf / D q f . if J D 0/; ˇ.sf / D ˇf ; (4.32d)
and Iu ; IIu ; IS ; IIS ; 0 are certain nonnegative (vectorial) scale factors
which may depend on the path parameter s.
We observe that also in this case the initial/terminal parameters characterizing
the second class of substitute problems (4.32a–f) are given again by (4.31a, b).
In the time-varying case the present substitute problems of class II reads:
Zsf
min E LJ s; t0 ; s0 ; ˇ./; pJ ; qe .s/; qe0 .s; qe00 .s/; ˇ.s/; ˇ 0 .s/ jAt0 ds
s0
CE J sf ; t0 ; s0 ; ˇ./; pJ ; qe .sf / jAt0 (4.33a)
s.t.
ˇ.s/ 0; s0 s sf (4.33b)
p
qe .s0 / D q0 ; qe0 .s0 / ˇ.s0 / D qP 0 (4.33c)
qe .sf / D qf . if J D 0/; ˇ.sf / D ˇf : (4.33d)
The following techniques are available for solving substitutes problems of type (I),
(II):
a) Reduction to a finite dimensional
parameter
optimization problem
Here, the unknown functions qe ./; ˇ./ or ˇ./ are approximated by a linear
combination
X
lq
q
qe .s/ WD qOl Bl .s/; s0 s sf (4.34a)
lD1
X̌
l
ˇOl Bl .s/; s0 s sf ;
ˇ
ˇ.s/ WD (4.34b)
lD1
q q ˇ ˇ
where Bl D Bl .s/; Bl D Bl .s/; s0 s sf ; l D 1; : : : ; lq .lˇ /, are given
basis functions, e.g. B-splines, and qO l ; ˇOl ; l D 1; : : : ; lq .lˇ /, are vectorial, scalar
coefficients. Putting (4.34a, b) into (4.29a–f, f’), (4.32a–f), resp., a semiinfinite
optimization problem is obtained. If the inequalities involving explicitly the path
parameter s; s0 s sf , are required for a finite number N of parameter values
s1 ; s2 ; : : : ; sN only, then this problem is reduced finally to a finite dimensional
parameter optimization problem which can be solved now numerically by
standard mathematical programming routines or search techniques. Of course, a
major problem is the approximative computation of the conditional expectations
which is done essentially by means of Taylor expansion with respect to the
parameter vector p at p .0/ . Consequently, several conditional moments have to
be determined (on-line, for stage j 1). For details, see [109–111, 130] and the
program package “OSTP” [8].
b) Variational techniques
Using methods from calculus of variations, necessary and—in some cases—also
sufficient conditions in terms ofcertain differential equations may be derived for
.0/
the optimal solutions qe ; ˇ .0/ ; ˇ .0/ , resp., of the variational problems (4.29a–
f, f’), (4.32a–f). For more details, see [130].
c) Linearization methods
Here,
we assume that we already have an approximative optimal solution
q e .s/; ˇ.s/ ; s0 s sf , of the substitute problem (4.29a–f, f’) or
(4.32a–f)
under
consideration. For example, an approximative optimal solution
q e ./; ˇ./ can be obtained by starting from the deterministic substitute
problem obtained by replacing the random
parameter vector p.!/ just by its
conditional mean p WD E p.!/jAt0 .
.0/
138 4 Adaptive Optimal Stochastic Trajectory Planning and Control (AOSTPC)
Given an approximate optimal solution q e ./; ˇ./ of substitute problem
.0/
(I) or (II), the unknown optimal solution qe ./; ˇ .0/ ./ to be determined is
represented by
and
Zf
s
.0/
F qe ./; ˇ./ WD g .0/ qe .s/; qe0 .s/; qe00 .s/; ˇ.s/; ˇ 0 .s/ ds: (4.36c)
s0
.0/ .0/
CD g;q .s/
ˇ.s/ C E g;q .s/
ˇ 0 .s/ (4.37a)
e ;ˇ e ;ˇ
and
.0/
.0/
.0/
qe .sf / q e .sf / C a ;q e .sf /T
qe .sf /; (4.37b)
4.4 OSTP: Optimal Stochastic Trajectory Planning 139
.0/
0
Ag;q .s/ WD E rq g p.!/; q e .s/; q 0e .s/; q 00e .s/; ˇ.s/; ˇ .s/ jAt0 (4.37c)
e ;ˇ
.0/
0
B g;q ;ˇ .s/ WD E rq 0 g p.!/; q e .s/; q 0e .s/; q 00e .s/; ˇ.s/; ˇ .s/ jAt0 (4.37d)
e
.0/
0
C g;q ;ˇ .s/ WD E rq 00 g p.!/; q e .s/; q 0e .s/; q 00e .s/; ˇ.s/; ˇ .s/ jAt0 (4.37e)
e
.0/ @g 0 00 0
D g;q ;ˇ .s/ WD E p.!/; q e .s/; q e .s/; q e .s/; ˇ.s/; ˇ .s/ jAt0 (4.37f)
e @ˇ
.0/ @g 0 00 0
E g;q ;ˇ .s/ WD E p.!/; q e .s/; q e .s/; q e .s/; ˇ.s/; ˇ .s/ jAt0 (4.37g)
e @ˇ 0
.0/
a;q e .sf / WD E rq p.!/; q e .s/ jAt0 ; s0 s sf : (4.37h)
.0/
As mentioned before, cf. (4.31c, d), the expected values g .0/ ; in (4.37a, b)
and the expected sensitivities defined by (4.37c–h) can be computed approxima-
tively by means of Taylor expansion with respect to p at p D E p.!/jAt0 .
.0/
.0/
According to (4.37a), and using partial integration, for the total costs F
along the path we get the following approximation:
Zsf
.0/ 0
F g .0/ q e .s/; q 0e .s/; q 00e .s/; ˇ.s/; ˇ .s/ ds (4.38a)
s0
Zsf Zsf
.0/ .0/
C G g;q ;ˇ .s/T
qe .s/ ds C H g;q .s/
ˇ.s/ ds
e e ;ˇ
s0 s0
.0/ .0/0
T
C B g;q .sf / C g;q .sf /
qe .sf /
e ;ˇ e ;ˇ
.0/ .0/0
T
C B g;q .s0 / C C g;q .s0 /
qe .s0 /
e ;ˇ e ;ˇ
.0/ .0/
C C g;q .sf /T
qe0 .sf / C g;q .s0 /T
qe0 .s0 /
e ;ˇ e ;ˇ
.0/ .0/
C E g;q .sf /
ˇ.sf / E g;q .s0 /
ˇ.s0 /;
e ;ˇ e ;ˇ
140 4 Adaptive Optimal Stochastic Trajectory Planning and Control (AOSTPC)
where
.0/ .0/ .0/0 .0/00
G g;q .s/ WD Ag;q .s/ B g;q .s/ C C g;q .s/ (4.38b)
e ;ˇ e ;ˇ e ;ˇ e ;ˇ
Conditions (4.29e–f, f’), (4.32c, d), resp., yield the following initial and terminal
conditions for the changes
qe .s/;
ˇ.s/:
ˇ.s0 / D 0;
qe .s0 / D q 0 q e .s0 / (4.38d)
ˇ.sf / D 0;
qe .sf / D q f q e .sf /; if J D 0: (4.38e)
n o
Having a certain collection qe; ./ W 2 of admissible paths in
configuration space, a variant of the above procedure (i), (ii) is to determine—
in an inner optimization loop—the optimal velocity profile ˇ ./ with respect
to a given path qe; ./, and to optimize then the parameter in an outer
optimization loop, see [68].
where
.0/
p K WD E .pK .!/jAt0 / : (4.39d)
Based on the inverse dynamics approach, see Remark 4.1, the optimal reference
trajectory q .0/ D q .0/ .t/; t t0 , is inserted now into the left hand side of the
dynamic equation (4.4a). This yields next to the random optimal control function
v .0/ t; pD .!/ WD M pD .!/; q .0/ .t/ qR .0/ .t/
C h pD .!/; q .0/ .t/; qP .0/ .t/ ; t t0 : (4.40)
142 4 Adaptive Optimal Stochastic Trajectory Planning and Control (AOSTPC)
Starting at the initial state .q 0 ; qP 0 / WD q .0/ .t0 /; qP .0/ .t0 / , this control obviously
keeps the robot exactly on the optimal trajectory q .0/ .t/; t t0 , provided that pD .!/
is the true vector of dynamic parameters.
An optimal feedforward control law u.0/ D u.0/ .t/; t t0 , related to the optimal
reference trajectory q .0/ D q .0/ .t/; t t0 , can be
obtained then by applying a certain
averaging or estimating operator D jAt0 to (4.40), hence,
u.0/ WD v .0/ t; pD ./ jAt0 ; t t0 : (4.41)
.0/
where p D denotes the conditional mean of pD .!/ defined by (4.31c), and the
second equation in formula (4.42a) holds since the dynamic equation of a robot
depends linearly on the parameter vector pD , see Remark 4.2.
.0/
Inserting into the dynamic equation (4.4a), instead of the conditional mean p D
.0/
of pD .!/ given At0 , another estimator pD of the true parameter vector pD or a
.0/
certain realization pD of pD .!/ at the time instant t0 , then we obtain the optimal
feedforward control law
.0/ .0/
u.0/ .t/ WD M pD ; q .0/ .t/ qR .0/ .t/ C h pD ; q .0/ .t/; qP .0/ .t/ ; t t0 : (4.42b)
At . A/; t t0 ; (4.43a)
the -algebra of all information about the random parameter vector p D p.!/ up
to time t. Hence, .At / is an increasing family of -algebras. Note that the flow of
4.5 AOSTP: Adaptive Optimal Stochastic Trajectory Planning 143
information in this control process can be described also by means of the stochastic
process
pt .!/ WD E p.!/jAt ; t t0 ; (4.43b)
see [12].
By parameter identification [65, 143] or robot calibration techniques [15, 145]
we may then determine the conditional distribution
.t /
Pp./ D Pp./jAt (4.43c)
.0/
The optimal functions qe .s/; ˇ .0/ .s/; s0 s sf , based on the a priori
information At0 , loose in course of time more or less their qualification to provide a
satisfactory pair of guiding functions q .0/ .t/; u.0/ .t/ ; t t0 .
However, having at the main correction time points tj ; j D 1; 2; : : :, the updated
.tj /
information -algebras Atj and then the a posteriori probability distributions Pp./
or the updated conditional
.0/ moments
.tj / of p.!/; j D 1; 2; : : :, the pair of
guiding functions q .t/; u .t/ ; t t0 , is replaced by a sequence of renewed
.0/
pairs q .j / .t/; u.j / .t/ ; t tj ; j D 1; 2; : : :, of guiding functions determined
.j /
by replanning, i.e. by repeated (OSTP) for the remaining time intervals Œtj ; tf
and by using the new information given by Atj . Since replanning at a later main
correction time point tj ; j 1, hence on-line, may be very time consuming, in
order to maintain the real-time capability of the method, we may start the replanning
144 4 Adaptive Optimal Stochastic Trajectory Planning and Control (AOSTPC)
procedure for an update of the guiding functions at time tj already at some earlier
time tQj with tj 1 < tQj < tj ; j 1. Of course, in this case
is defined to contain only the information about the control process up to time tQj in
which replanning for time tj starts (Fig. 4.2).
The resulting substitute problem at a stage j 1 follows from the corresponding
.j 1/
substitute problem for the previous stage j 1 just by updating j 1 ! j ; f !
.j /
f , the initial and terminal parameters, see (4.31a, b). The renewed
.j /
initial parameters j W tj ; sj ; q j ; qP j ; Pp./ or j (4.45a)
P j /)
(observation or estimate of q.tj /; q.t
.j / .t /
Pp./ WD Pp./
j
D Pp./jAtj (4.45e)
j WD .tj / : (4.45f)
The renewed
.j / .j / .j /
terminal parameters f W tf ; sf ; q f ; ˇf (4.46a)
sf (given) (4.46b)
.j / .j /
q f WD q.t
O f / or q f WD E qf .!/jAtj (4.46c)
ˇf D 0 (4.46d)
.j /
s .j / tf D sf : (4.46e)
4.5 AOSTP: Adaptive Optimal Stochastic Trajectory Planning 145
As already mentioned above, the (OSTP) for the j -th stage, j 1, is obtained
from the substitute problems (4.29a–f, f’), (4.32a–f), (4.33a–d), resp., formulated
for the 0-th stage, j D 0, just by substituting
.j /
0 ! j and f ! f : (4.47)
the corresponding pair of optimal solutions of the resulting substitute problem for
the j -th stage, j 1.
The pair of guiding functions q .j /.t/; u.j / .t/ ; t tj , for the j -th stage,
j 1, is then defined as described in Sect. 4.4.2 for the 0-th stage. Hence, for
the j -th stage, the reference trajectory in configuration space q .j / .t/; t tj , reads,
cf. (4.39a),
q .j /.t/ WD qe.j / s .j / .t/ ; t tj ; (4.49a)
h i
.j /
where the transformation s .j / W tj ; tf ! Œsj ; sf is defined by the initial value
problem
q
sP .t/ D ˇ .j / .s/; t tj ; s.tj / D sj : (4.49b)
.j /
The terminal time tf for the j -th stage is defined by the equation
.j /
s .j / tf D sf : (4.49c)
where
v .j / .t; pD / WD M pD ; q .j / .t/ qR .j / .t/ C h pD ; q .j / .t/; qP .j / .t/ ; t tj :
(4.50b)
Using the conditional expectation .jAtj / WD E.jAtj /, we find the feedforward
control
.j / .j /
u.j / .t/ WD M p D ; q .j / .t/ qR .j /.t/ C h p D ; q .j / ; qP .j / .t/ ; t tj ; (4.50c)
146 4 Adaptive Optimal Stochastic Trajectory Planning and Control (AOSTPC)
(4.51a)
where
.j /
p K WD E pK .!/jAtj : (4.51b)
4.5.1 (OSTP)-Transformation
Theorem 4.5.1 Let ŒQs0 ; sQf ; sQ0 < sQf WD sf , be a given, fixed reference s-interval,
and consider for a certain stage j; j D 0; 1; : : :, the transformation
sQf sQ0
sQ D sQ .s/ WD sQ0 C .s sj /; sj s sf ; (4.52a)
sf sj
sf sj
s D s.Qs / D sj C .Qs sQ0 /; sQ0 sQ sQf : (4.52b)
sQf sQ0
Represent then the geometric path in work space qe D qe .s/ and the velocity
profile ˇ D ˇ.s/; sj s sf , for the j -th stage by
qe .s/ WD qQe sQ.s/ ; sj s sf (4.53a)
ˇ.s/ WD ˇQ sQ .s/ ; sj s sf (4.53b)
.j /
.; f / D .j ; f / 2 Z Zf (4.54)
and having the fixed reference s-interval ŒQs0 ; sQf . Moreover, the optimal solution
.j / .j /
qe ; ˇ .j / D qe .s/; ˇ .j / .s/ ; sj s sf , may be represented by the optimal
adaptive law
qe.j / .s/ D qQe sQ .s/I j ; f ; sj s sf ;
.j /
(4.55a)
ˇ .j / .s/ D ˇQ sQ .s/I j ; f ; sj s sf ;
.j /
(4.55b)
where
denotes the optimal solution of the above mentioned reference variational problem.
Proof According to (4.53a, b) and (4.52a, b), the derivatives of the functions
qe .s/; ˇ.s/; sj s sf , are given by
sQ sQ
qe0 .s/ D qQ e0 sQ .s/
f 0
; sj s sf ; (4.56a)
sf sj
sQ sQ 2
00 00 f 0
qe .s/ D qQ e sQ .s/ ; sj s sf ; (4.56b)
sf sj
sQ sQ
ˇ 0 .s/ D ˇQ 0 sQ .s/
f 0
; sj s sf : (4.56c)
sf sj
ZsQf
sQf sQ0 00 sQf sQ0 2 Q
min E L J
pJ ; qQe .Qs /; qQe0 .Qs / ; qQ .Qs / ; ˇ.Qs /;
sf sj e sf sj
sQ0
sQf sQ0 sf sj
ˇQ 0 .Qs / jAtj d sQ C E J pJ ; qQ e .sQf jAtj
sf sj sQf sQ0
(4.57a)
s.t.
!
sQf sQ0 00 sQf sQ0 2 Q sQf sQ0
0
E f p; qQe .Qs /; qQe .Qs / ; qQe .Qs / ; ˇ.Qs /; ˇQ 0 .Qs / jAtj
sf sj sf sj sf sj
f ; (4.57b)
sQ0 sQ sQf (4.57c)
Q s / 0; sQ0 sQ sQf
ˇ.Q (4.57d)
q
sQf sQ0 Q s0 / D qP
qQe .Qs0 / D q j ; qQe0 .Qs0 / ˇ.Q j (4.57e)
sf sj
.j / Q sf / D 0
qQe .Qsf / D q f .if J D 0/; ˇ.Q (4.57f)
Q sf / D 0; E
ˇ.Q p; qQ e .Qsf / jAtj ; (4.57e’)
ZsQf
sQf sQ0 00 sQf sQ0 2 Q
min E LJ p; qQ e .Qs /; qQe0 .Qs / ; qQ e .Qs / ; ˇ.Qs /;
sf sj sf sj
sQ0
sQf sQ0 sf sj
ˇQ 0 .Qs / jAtj d sQ C E J p; qQ e .Qsf / jAtj (4.58a)
sf sj sQf sQ0
4.5 AOSTP: Adaptive Optimal Stochastic Trajectory Planning 149
s.t.
Q s / 0; sQ0 sQ sQf
ˇ.Q (4.58b)
q
sQf sQ0 Q s0 / D qP
qQ e .Qs0 / D q j ; qQ e0 .Qs0 / ˇ.Q j (4.58c)
sf sj
.j / Q sf / D 0:
qQe .Qsf / D q f .if J D 0/; ˇ.Q (4.58d)
For the consideration of the time-variant case we note first that by using the
transformation (4.52a, b) and (4.53b) the time t tj can be represented, cf. (4.14a,
b) and (4.15a), also by
Z sQ
Q sf sj d Q
t D t sQ; tj ; sj ; ˇ./ WD tj C q : (4.59a)
sQf sQ0 Q Q /
sQ0 ˇ.
Hence, if the variational problems (4.57a–f) and (4.58a–d) for the j -th stage
depend explicitly on time t tj , then, corresponding to Sect. 4.4, item B), for the
constituting functions LJ ; J ; LJ ; J of the variational problems we have that
Q pJ ; qe ; q 0 ; q 00 ; ˇ; ˇ 0 ; sQ0 sQ sQf
LJ D LJ sQ ; tj ; sj ; ˇ./; (4.59b)
e e
Q pJ ; qe
J D J sQf ; tj ; sj ; ˇ./; (4.59c)
Q p; qe ; q 0 ; q 00 ; ˇ; ˇ 0 ; sQ0 sQ sQf
f D f sQ; tj ; sj ; ˇ./; (4.59d)
e e
Q p; qe ; q 00 ; ˇ; ˇ 0 ; sQ0 sQ sQf
LJ D LJ sQ ; tj ; sj ; ˇ./; (4.59e)
e
Q p; qe :
J D J sQf ; tj ; sj ; ˇ./; (4.59f)
Suppose here that J 6D 0; J 6D 0, resp., and the terminal state condition (4.57f,
e’), (4.58d), resp., is reduced to
Q sf / D 0:
ˇ.Q (4.60a)
Q s / WD ˇj ˇQa .Qs /;
ˇ.Q sQ0 sQ sQf (4.60b)
qQe .Qs / WD q jd qQea .Qs /; sQ0 sQ sQf ; (4.60c)
150 4 Adaptive Optimal Stochastic Trajectory Planning and Control (AOSTPC)
where q jd denotes the diagonal matrix with the components of q j on its main
diagonal, then in terms of the new unknowns ˇQa ./; qQ ea ./ on ŒQs0 ; sQf we have
the nonnegativity and fixed initial/terminal conditions
With the exception of field robots (e.g. Mars rover) and service robots [62], becom-
ing increasingly important, the standard industrial robots move very fast. Hence, for
.j /
industrial robots the optimal solution qe .s/; ˇ .j / .s/ ; ˇ .j / .s/, resp., sj s
sf , generating the renewed pair of guiding functions q .j / .t/; u.j / .t/ ; t tj , on
each stagej D 1; 2;: : : should be provided in real-time. This means that the optimal
.j /
solutions qe ; ˇ .j / ; ˇ .j / , resp., must be prepared off-line as far as possible such
that only relatively simple numerical operations are left on-line.
Numerical methods capable to generate approximate optimal solutions in real-
time are based mostly on discretization techniques, neural network (NN) approxi-
mation [8, 111, 114, 133], linearization techniques (sensitivity analysis) [159].
Discretization Techniques
[
l0
Z Zf D Z l Zfl ; (4.62a)
lD1
in each subdomain Z l Zfl , the optimal adaptive law (4.55c) can be approximated,
cf. [150], by
)
qQO e .Qs I ; f / WD qQ .Qs I l ; fl /; sQ0 sQ sQf
for .; f / 2 Z l Zfl : (4.62c)
ˇOQ .Qs I ; f / WD ˇQ .Qs I l ; fl /; sQ0 sQ sQf
NN-Approximation
For the determination of the optimal adaptive law (4.55a–c) in real-time, according
to (4.34a, b), the reference variational problem (4.57a–f) or (4.58a–d) is reduced
first to a finite dimensional parameter optimization problem by
i) representing the unknown functions qQe D qQe .Qs /; ˇQ D ˇ.Q
Q s /; sQ0 sQ sQf , as
linear combinations
X
lq
q
qQe .Qs / WD qOe Bl .Qs /; sQ0 sQ sQf ; (4.63a)
lD1
l
X̌
Q s / WD
ˇ.Q ˇOl Bl .Qs /; sQ0 sQ sQf ;
ˇ
(4.63b)
lD1
q q ˇ ˇ
of certain basis functions, e.g. cubic B-splines, Bl D Bl .Qs /; Bl D Bl .Qs /; sQ0
sQ sQf ; l D 1; : : : ; lq .lˇ /, with unknown vectorial (scalar) coefficients
qOl ; ˇOl ; l D 1; : : : ; lq .lˇ /, and
ii) demanding the inequalities in (4.57b, c), (4.58b), resp., only for a finite set of
sQ-parameter values sQ0 < sQ1 < : : : < sQk < sQkC1 < : : : < sQ D sQf .
By means of the above described procedure (i), (ii), the optimal coefficients
in an off-line training procedure [8, 111, 133]. Here, the model (4.64a, b) is fitted in
the LSQ-sense to given data
! !
; f ; f
; ; D 1; : : : ; 0 ; (4.64d)
qO l ; l D 1; : : : ; lq ˇOl ; l D 1; : : : ; lˇ
where
. ; f /; D 1; : : : ; 0 ; (4.64e)
are the related optimal coefficients in (4.63a, b) which are determined off-line by an
appropriate parameter optimization procedure.
Having the vectors wq ; wˇ of optimal NN-weights, by means of (4.64a–c), for
.j /
given actual initial/terminal parameters .; f / D .j ; f / at stage j 0, the NN
yields then the optimal parameters
Linearization Methods
@qQe
qQe .Qs ; ; f / D qQ e .Qs I 0 ; f / C
.0/ .0/ .0/
.Qs I 0 ; f /. 0 ; f f / C : : :
@O
(4.65a)
@ˇQ
ˇQ .Qs I ; f / D ˇQ .Qs I 0 ; f / C
.0/ .0/ .0/
.Qs I 0 ; f /. 0 ; f f / C : : : ;
@O
(4.65b)
where the optimal starting functions qQe .Qs I 0 ; f /; ˇQ .Qs I 0 ; f / and the deriva-
.0/ .0/
@qQ .0/ @ˇ
Q .0/
tives e .Qs I 0 ; f /; .Qs I 0 ; f /; : : : can be determined—on a certain grid
@O @O
of ŒQs0 ; sQf —off-line by using sensitivity analysis [159]. The actual values of
qQe ; ˇQ at later stages can be obtained then very rapidly by means of simple
matrix operations. If necessary, the derivatives can be updated later on by a
numerical procedure running in parallel to the control process.
II) Sequential linearization of the (AOSTP) process
.j / .j /
Given the optimal guiding functions qe D qe .s/; ˇ .j / D ˇ .j / .s/; sj s sf
for the j -th stage (Fig. 4.3), corresponding to the representation (4.35a, b), the
.j C1/
optimal guiding functions qe .s/; ˇ .j C1/ .s/; sj C1 s sf , are represented
by
qe .s/;
ˇ.s/; sj C1 s sf , if the following replacements are made in the
formulas (4.35a, b), (4.36a–c), (4.37a–h) and (4.38a–f):
9
s0 s sf ! sj C1 s sf >
>
>
>
qe ; ˇ
.j /
! qe ; ˇ .j / >
>
>
>
At0 ! Atj C1 >
>
>
>
.0/ .0/ .j C1/ .j C1/ >
>
g .0/ ; ; F ! g .j C1/; ;F >
=
.0/ .0/ .j C1/ .j C1/
Ag;q ;ˇ ; : : : ; H g;q ! Ag;q .j / ;ˇ.j / ; : : : ; H g;q .j / ;ˇ.j /
e e ;ˇ e e >
>
ˇ.s0 / D 0 !
ˇ.sj C1 / D ˇj C1 ˇ .j / .sj C1 / >
>
>
>
.j / >
>
qe .s0 / D q0 q e .s0 / !
qe .sj C1 / D q j C1 qe .sj C1 / >
>
>
>
ˇ.sf / D 0 !
ˇ.sf / D 0 >
>
.0/ .j C1/ .j / >
;
qe .sf / D qf q e .sf / !
qe .sf / D q f q f ; if D 0;
J
(4.66c)
.j C1/
where X denotes the conditional expectation of a random variable X with
respect to Atj C1 , cf. (4.50d), (4.51b). Furthermore, (4.37f) yields
0
q
qe.j / .sj C1 / C
qe0 .sj C1 / ˇj C1 D qP j C1 (4.66d)
g WD LJ ; WD J :
Zsf Zsf
.j C1/ .j C1/
min G g;q .j / ;ˇ.j / .s/
qe .s/ds C
T
H g;q .j / ;ˇ.j / .s/
.s/ds (4.67a)
e e
sj C1 sj C1
CRjT
qe .sf / C SjT
qe0 .sf / C Tj
ˇ.sj C1 /
s.t.
qe .sj C1 / D q j C1 qe.j / .sj C1 / (4.67b)
.j C1/ .j /
qe .sf / D q f q f ; if J D 0 (4.67c)
ˇ.sf / D 0
ˇ.s/ ˇ .j / .s/; sj C1 s sf ; (4.67d)
where
.j C1/ .j C1/ .j C1/
Rj WD B g;q .j / ;ˇ.j / .sf / C g;q .j / ;ˇ.j / .sf / C a .j / .sf / (4.67e)
e e ;qe
.j C1/
Sj WD C g;q .j / ;ˇ.j / .sf / (4.67f)
e
.j /0
1 .j C1/ qe .sj C1 / .j C1/
Tj WD C g;q .j / ;ˇ.j / .sj C1 /T .j / E g;q .j / ;ˇ.j / .sj C1 /: (4.67g)
2 e ˇ .sj C1 / e
The linear optimization problem (4.67a–g) can be solved now by the methods
developed e.g. in [52], where for the correction terms
qe .s/;
ˇ.s/; sj C1 s
sf , some box constraints or norm bounds have to be added to (4.67a–g). In case of
ˇ .j / .s/
ˇ.s/
ˇ max ; sj C1 s sf ; (4.67d’)
Obviously, the methods described briefly in Sect. 4.5.3 can be combined in the
following way:
First, by means of discretization (Finite Element Methods), an approximate
optimal control law .qQe ; ˇ / is searched in a class of finitely generated functions
of the type (4.63a, b). Corresponding to (4.65a, b), by means of Taylor expansion
here the optimal coefficients qOl ; ˇOl ; l D 1; : : : ; lq .lˇ /, in the corresponding linear
combination of type (4.63a, b) are represented, cf. (4.63c, d), by
@qO
qOl .; f / D qOl 0 ; f C l .0 ; f / 0 ; f f C : : : ;
.0/ .0/ .0/
(4.68a)
@O
@ˇO
ˇOl .; f / D ˇOl 0 ; f C l 0 ; f
.0/ .0/ .0/
0 ; f f C : : : ;
@O
(4.68b)
@qOl .0/
@ˇO
.0/
0 ; f ; l 0 ; f ; : : : ; l D 1; : : : ; lq .lˇ /; (4.68c)
@O @O
We now consider the control of the robot at the j -th stage, i.e., for time t tj ,
see [4, 6, 11, 20, 58, 61]. In practice we have random variations of the vector p of
the model parameters of the robot and its environment,
moreover,
there are possible
deviations of the true initial state .qj ; qPj / WD q.tj /; q.t
P j / in configuration space
p
.j /0
from the corresponding initial values .q j ; qP j / D q j ; qe .sj / ˇj of the (OSTP)
at stage j . Thus, the actual trajectory
q.t/ D q t; pD ; qj ; qP j ; u./ ; t tj (4.69a)
in configuration space of the robot will deviate more or less from the optimal
reference trajectory
.j /
q .j / .t/ D qe.j / s .j / .t/ D q t; p D ; q j ; qP j ; u.j /./ ; (4.69b)
see (4.44a,
b), (4.45a–f),
(4.49a, b) and (4.50c). In the following we assume that
P
the state q.t/; q.t/ in configuration space may be observed for t > tj . Now, in
order to define an appropriate control correction (feedback control law), see (4.2)
and Fig. 4.1,
u.j /.t/ D u.t/ u.j / .t/ WD ' .j / t;
z.j / .t/ ; t tj ; (4.70a)
the trajectories q.t/ and q .j / .t/; t tj , are embedded into a one-parameter family
of trajectories q D q.t; /; t tj ; 0 1, in configuration space which are
defined as follows:
Consider first the following initial data for stage j :
qj ./ WD q j C
qj ;
qj WD qj q j (4.71a)
P j ;
q
qP j ./ WD qP j C
q P j WD qPj qP j (4.71b)
.j / .j /
pD ./ WD p D C
pD ;
pD WD pD p D ; 0 1: (4.71c)
158 4 Adaptive Optimal Stochastic Trajectory Planning and Control (AOSTPC)
the solution of the following initial value problem consisting of the dynamic
equation (4.4a) with the initial values, the vector of dynamic parameters and the
total control input u.t/ given by (4.71a–d):
P /; q.t;
F .pD ./; q.t; /; q.t; R // D u.t; /; 0 1; t tj ; (4.73a)
where
P q/
and F D F .pD ; q; q; R is defined, cf. (2.6a), by
P q/
F .pD ; q; q; R WD M.pD ; q/qR C h.pD ; q; q/:
P (4.73d)
In the following we suppose [76] that the initial value problem (4.73a–d) has
a unique solution q D q.t; /; t tj , for each parameter value ; 0 1
(Fig. 4.4).
Because of condition (4.70c) of the feedback control law ' .j / to be determined, and
due to the unique solvability assumption of the initial value problem (4.73a–d) at
the j -th stage, for D 0 we have that
Fig. 4.4 Adaptive optimal stochastic trajectore planning and control (AOSTPC)
D 1 WD 1
is the actual trajectory in configuration space under the total control input u.t/ D
u.j /.t/ C
u.j /.t/; t tj , given by (4.71d).
160 4 Adaptive Optimal Stochastic Trajectory Planning and Control (AOSTPC)
Let
D 1 0 D 1, and suppose that the following property known from
parameter-dependent differential equations, cf. [76], holds:
Assumption 4.1 The solution q D q.t; /; t tj ; 0 1, of the initial value
problem (4.71a–d) has continuous derivatives with respect to up to order > 1
for all tj t tj C
tj ; 0 1, with a certain
tj > 0.
Note that .t; / ! q.t; /; t tj ; 0 1, can be interpreted as a homotopy
from the reference trajectory q .j / .t/ to the actual trajectory q.t/; t tj , cf. [134].
Based on the above assumption and (4.74a, b), by Taylor expansion with respect
to at D 0 D 0, the actual trajectory of the robot can be represented by
q.t/ D q t; pD ; qj ; qPj ; u./ D q.t; 1/ D q.t; 0 C
/
D q.t; 0/ C
q.t/ D q .j / .t/ C
q.t/; (4.75a)
X
1
1 l 1 @ q
q.t/ D d q.t/.
/l C .t; #/.
/
lŠ Š @
lD1
X
1
1 l 1 @ q
D d q.t/ C .t; #/; t tj : (4.75b)
lŠ Š @
lD1
@l q
d l q.t/ WD .t; 0/; t tj ; l D 1; 2; : : : ; (4.75c)
@ l
P q/
K.pD ; q; q; R WD Fq .pD ; q; q;
P q/
R D Dq F .pD ; q; q;
P q/
R (4.76b)
P WD FqP .pD ; q; q;
D.pD ; q; q/ P q/
R D hqP .pD ; q; q/:
P (4.76c)
Moreover, it is
and due to the linear parameterization property of robots, see Remark 4.2, F may
be represented by
P q/
F .pD ; q; q; R D Y .q; q;
P q/p
R D (4.76e)
@q d @q P j:
.tj ; / D
qj ; .tj ; / D
q (4.77b)
@ dt @
162 4 Adaptive Optimal Stochastic Trajectory Planning and Control (AOSTPC)
Y .j / .t/
pD C K .j /.t/dq.t/ C D .j / .t/dP q.t/ C M .j /.t/dR q.t/
D du.t/ D 'z.j / .t; 0/dz.t/ D 'q.j /.t; 0/dq.t/ C 'qP .t; 0/dP q.t/; t tj ;
.j /
(4.78a)
with the initial values
dq.tj / D
qj ; dP q.tj / D
q
P j: (4.78b)
Here,
@u
du.t/ WD .t; 0/; (4.78c)
@
dq.t/ d d2
dz.t/ WD P ; dP q WD dq; dR q WD 2 dq; (4.78d)
d q.t/; dt dt
and the matrices Y .j / .t/; K .j /.t/; D .j / .t/ and M .j /.t/ are defined, cf. (4.76b–e), by
Y .j / .t/ WD Y q .j / .t/; qP .j / .t/; qR .j / .t/ (4.78e)
.j /
K .j / .t/ WD K p D ; q .j / .t/; qP .j / .t/; qR .j / .t/ (4.78f)
.j / .j /
D .j / WD D p D ; q .j / .t/; qP .j / .t/ ; M .j /.t/ WD M p D ; q .j / .t/ : (4.78g)
where Kp D .pk ık /; Kd D .dk ık / are positive definite diagonal matrices with
positive diagonal elements pk ; dk > 0; k D 1; : : : ; n.
Inserting (4.79) into (4.78a), and assuming that M .j / D M .j /.t/ is regular [11]
for t tj , we find the following linear system of 2nd order differential equations
for dq D dq.t/:
R
dq.t/ C Kd dP q.t/ C Kp dq.t/ D M .j /.t/1 Y .j / .t/
pD ; t tj ; (4.80a)
dq.tj / D
qj ; dP q.tj / D
q
P j: (4.80b)
4.6 Online Control Corrections: PD-Controller 163
Y .j / .t/
pD
0
1 0
D M
pD ; qe s .t/ .j / .j /
qe.j / s .j / .t/ ˇ .j / s .j / .t/
2
00
Cqe.j / s .j / .t/ ˇ .j / s .j / .t/
!
r
.j /0
Ch
pD ; qe s .t/ ; qe
.j / .j / .j /
s .t/ ˇ .j / s .j / .t/ : (4.81b)
c)
u.j / .t/ C Y .j / .t/
pD D ue pD ; s .j / .t/I qe.j / ./; ˇ .j / ./ ; t tj : (4.82c)
Proof The first equation follows from (4.81b) and (4.19b). Equations (4.50c),
(4.18a, b) and (4.19b) yield (4.82b). Finally, (4.82c) follows from (4.82a, b) and
the linear parameterization of robots, cf. Remark 4.2.
Remark 4.12 Note that according to the transformation (4.19a) of the dynamic
equation onto the s-domain, for the control input u.t/ we have the representation
with s D s.t/.
164 4 Adaptive Optimal Stochastic Trajectory Planning and Control (AOSTPC)
Using (4.78d), it is easy to see that (4.80a, b) can be described also by the 1st
order initial value problem
0
dP z.t/ D Adz.t/ C .j;1/ ; t tj (4.83a)
.t/
!
qj q j
dz.tj / D
zj D ; (4.83b)
qPj qP j
.j;1/
and .t/ is defined, cf. (4.82a), by
.j;1/
.t/ WD M .j /.t/1 Y .j / .t/
pD
D M .j /.t/1 ue
pD ; s .j / .t/I qe.j / ./; ˇ .j / ./ : (4.83d)
.j /
Consequently, for the first order expansion
term dz.t/ of the deviation
z .t/
q.t/
between the actual state z.t/ D and the prescribed state z.j / .t/ D
P
q.t/
.j /
q .t/
; t tj , we have the representation [32, 76]
qP .j / .t/
Zt
A.t tj / 0
dz.t/ D dz .t/ D e
.j /
zj C e A.t / .j;1/ d : (4.84a)
./
tj
Because of E
pD .!/jAtj D 0, we have that
E .t/jAtj D 0;
.j;1/
(4.84b)
E dz.t/jAtj D e A.t tj / E
zj jAtj ; t tj ; (4.84c)
.j /
where 'z .t; 0/ is defined by (4.79). Then, the following relations hold:
a) Asymptotic local stability in the mean:
E dz.t/jAtj ! 0; t ! 1I (4.86b)
Zt q
C a0 e 0 .t / E k .j;1/ ./k2 jA
tj d ; t tj ; (4.86c)
tj
where
E k .j;1/
.t/k2 jAtj kM .j / .t/1 k2 u.je /2 s .j / .t/ ; (4.86d)
2
u.je / s .j / .t/ kY .j / .t/k2 var pD ./jAtj (4.86e)
166 4 Adaptive Optimal Stochastic Trajectory Planning and Control (AOSTPC)
with
r
.j /
u.je / .s/ WD var ue pD ./; sI qe ./; ˇ .j / ./ jAtj ; sj s sf : (4.86f)
Proof
p Follows p from (4.84a), (4.82a–d) and the fact that by Jensen’s inequality
E X.!/ EX.!/ for a nonnegative random variable X D X.!/.
.j /2
Note that ue s .j / .t/ can be interpreted as the risk of the feedforward control
u.j /.t/; t tj . Using (4.69b), (4.78g), (4.86d, e) and then changing variables !
s in the integral in (4.86c), we obtain the following result:
Theorem 4.6.2 Let denote t .j / D t .j / .s/; s sj , the inverse of the parameter
transformation s .j / D s .j / .t/; t tj . Under the assumptions of Theorem 4.6.1, the
.j /
following inequality holds for tj t tf :
r
0 .t tj / .j /2
E kdz.t/kjAtj a0 e z.tj / C kE z.tj /jAtj zj k2
0 t t .j / .s/
1
R
.j / .j /
s .j / .t / a0 e kM p D ;qe .s/ k
C p .j /
ue .s/ ds: (4.87a)
sj ˇ .j / .s/
The minimality or boundedness of the right hand side of (4.87a), hence, the
robustness [43] of the present control scheme, is shown next:
Corollary 4.6.1 The meaning of the above inequality (4.87a) follows from the
following important minimality/boundedness properties depending on the chosen
substitute problem in (OSTP) for the trajectory planning problem under stochastic
uncertainty:
i) The
error contribution
of the initial value zj takes a minimum for zj WD
E z.tj /jAtj , cf. (4.44a, b).
ii) The factor 0 can be increased by an appropriate selection of the matrices
Kp ; Kd ;
iii)
1
.j /
c M kM pD ; qe.j /.s/ k c M ; sj s sf ; (4.87b)
with positive constants c M ; c M > 0. This follows from the fact that the mass
matrix is always positive definite [11].
4.6 Online Control Corrections: PD-Controller 167
iv)
sZ
.j / .t /
Zsf
ds ds .j /
p p D tf tj ; (4.87c)
ˇ .j / .s/ .j /
ˇ .s/
sj sj
.j / .j /
e .s/ WD ue p D ; sI qe ./; ˇ ./ . Hence, with (4.87d) we have then the
where u.j / .j /
condition
2
u.je / .s/ .1 ˛u / min u2e ; sj s sf ; (4.87d’)
1kn
cf. (4.87a). Under special distribution assumptions for pD .!/ more exact explicit
deterministic conditions for (4.30a) may be derived, see Remark 4.2.
If minimum force and moment should be achieved along the trajectory, hence, if
D 0 and L D ku.t/k2 , see (4.6), then, according to substitute problem (4.29a–f,
f’) we have the following minimality property:
Zsf Z f
t
ds
u.je / .s/ C u.j p
2 2
/
e .s/ D min E ku.t/k2 dtjAtj :
ˇ .j / .s/ qe ./;ˇ./
sj tj
(4.87e)
Mean/variance condition for ue : Condition (4.29c) in substitute problem (4.29a–
f, f’) may read in case of fixed bounds umin ; umax for the control u.t/ as follows:
.j /
umin ue p D ; sI qe ./; ˇ./ umax ; sj s sf (4.87f)
According to the above Theorem 4.6.2, further stability results, especially the
convergence
E kdz.t/kjAtj ! 0 for j ! 1; t ! 1 (4.88a)
of the mean absolute first order tracking error can be obtained if, by using a suitable
update law [4, 6, 11, 32] for the parameter estimates, hence, for the a posteriori
distribution P .jAtj /, we have that, see (4.86f),
var pD ./jAtj D var pD ./jAtj ! 0 for j ! 1: (4.88b)
@z @qR @u @z
FpD
pD C Fz C FqR D D 'z.j / ; (4.89a)
@ @ @ @
R z D qqP , is given by (4.73d), see also (4.76e), and therefore
where F D F .pD ; z; q/;
@2 q d @2 q @2 qP
2
.tj ; / D 0; 2
.tj ; / D 2 .tj ; / D 0 (4.90d)
@ dt @ @
2
@2 z @2 z @ q @2 qP
for 2 D 2 .t; / D .t; /; .t; / ; t tj ; 0 1:
@ @ @ 2 @ 2
Putting now D 0, from (4.90a) we obtain the following differential equation
2
for the 2nd order differential d 2 q.t/ D @@q2 .t; 0/ of q D q.t; /:
d d2
K .j /.t/d 2 q.t/ C D .j / .t/ d 2 q.t/ C M .j /.t/ 2 d 2 q.t/
dt dt
.j / .j /
C Fzz p D ; q .t/; qP .t/; qR .t/ 'zz .t; 0/ dz.t/; dz.t/
.j / .j / .j /
'z.j /.t; 0/d 2 z.t/ D 2Fp.jD/z .t/
pD ; dz.t/
C 2FpD qR .t/
pD ; dR q.t/ C 2FzqR .t/ dz.t/; dR q.t/ :
.j / .j /
(4.91a)
Here, we set
d 2 q.t/
d z.t/ WD
2
d 2 ; (4.91b)
dt
d q.t/
.j / .j / .j /
and the vectorial Hessians FpD z .t/; FpD qR .t/; FzqR .t/ follow from (4.90b) by insert-
.j /
ing there the argument .pD ; q; q;P q/
R WD p D ; q .j / .t/; qP .j / .t/; qR .j / .t/ . Further-
more, (4.90d) yields the following initial condition for d 2 q.t/
d 2
d 2 q.tj / D 0; d q.tj / D 0: (4.91c)
dt
According to (4.91a) we define now, cf. (4.79), the second order derivative of
' .j / with respect to z at
z D 0 by
.j /
'zz.j / .t; 0/ WD Fzz p D ; q .j / .t/; qP .j / .t/; qR .j / .t/ ; t tj : (4.92)
170 4 Adaptive Optimal Stochastic Trajectory Planning and Control (AOSTPC)
d2 2 d
2
d q.t/ C Kd d 2 q.t/ C Kp d 2 q.t/ (4.93a)
dt dt
e 2
D M .j /.t/1 D2 F .j / .t/
pD ; dz.t/; dR q.t/ ; t tj ;
d 2
d 2 q.tj / D 0; d q.tj / D 0; (4.93b)
dt
where the sub-Hessian
e e
.j /
D2 F .j / .t/ WD D2 F p D ; q .j / .t/; qP .j / .t/; qR .j / .t/ ; (4.93c)
of F results from the Hessian of F by replacing the diagonal block Fzz by zero. Of
course, we have that, cf. (4.91a),
e
R
r 2 F .j /
pD ; dz.t/; dq.t/
2
D 2 Fp.jD/z .t/
pD ; dz.t/
CFpD qR .t/
pD ; dR q.t/ C FzqR .t/ dR q.t/; dz.t/ :
.j / .j /
(4.93d)
Comparing now the initial value problems (4.80a, b) and (4.93a, b) for the first
and second order differential of q D q.t; /, we recognize that the linear 2nd order
differential equations have—up to the right hand side—exactly the same form.
According to (4.83a–d) and (4.84a) we know that the 1st order expansion terms
dz.t/; dR q.t/ D dq.t/; dP q.t/; dR q.t/ ; t tj ;
.j;2/ e 2
.t/ WD M .j /.t/1 D2 F .j / .t/
pD ; dz.t/; dR q.t/ ; t tj ; (4.95)
4.6 Online Control Corrections: PD-Controller 171
of the error differential equation (4.93a) for d 2 q.t/ is quadratic in the error term
.j / ./.
According to (4.93a–c), the second order expansion term
d
d 2 z.t/ D d 2 q.t/; d 2 q.t/ ; t tj ;
dt
of the Taylor expansion of the tracking error can be represented again by the
solution of the system of linear differential equations (4.83a–c), where now .j;1/ .t/
is replaced by .j;2/ .t/ defined by (4.95), and the initial values are given by
d 2 z.tj / D 0. Thus, applying again solution formula (4.84a), we find
Zt
0
d 2 z.t/ D e A.t / .j;2/ d : (4.96)
./
tj
From (4.94)–(4.96) and Lemma 4.2 we get now the following result:
Theorem
4.6.3 Thesecond
order tracking error2 expansion
terms
2
d d d
d 2 z.t/; 2 d 2 q.t/ D d 2 q.t/; d 2 q.t/; 2 d 2 q.t/ ; t tj , depend
dt dt dt
i) quadratically on the 1st order error terms
pD ; dz.t/; dR q.t/ and
ii) quadratically on the error term
.j / ./ corresponding to the variations/
disturbances of the initial values and dynamics parameters.
Because of (4.96), the stability properties of the second order tracking error
expansion term d 2 q.t/; t tj , are determined again by the matrix exponential
function ˚.t; / D e A.t / and the remainder .j;2/ .t/ given by (4.95).
According to Theorem 4.6.2 and Corollary 4.6.1 we know that the disturbance
term .j;1/ .t/ of (4.80a), (4.83a) is reduced directly by control-constraints
(for ue ) present in (OSTP). Concerning the next disturbance term .j;2/ .t/
of
(4.93a), by (4.95) we note first that a reduction of the 1st order error terms
R
pD ; dz./; d q./ yields a reduction of .j;2/ and by (4.96) also a reduction of
d2 2
d 2 z.t/; d q.t/; t tj . Comparing then definitions (4.83d) and (4.95) of the
dt 2
disturbances .j;1/ ; .j;2/ , we observe that, corresponding to .j;1/ , certain terms
in .j;2/ depend only on the reference trajectory q .j / .t/; t tj , of stage j . Hence,
this observation yields the following result.
172 4 Adaptive Optimal Stochastic Trajectory Planning and Control (AOSTPC)
.j;2/
Theorem 4.6.4 The disturbance of (4.93a), and consequently also the second
d d2
order tracking error expansion terms d 2 q.t/; d 2 q.t/; 2 d 2 q.t/; t tj , can be
dt dt
diminished by
R
i) reducing the 1st order error terms
pD ; dz./; q./ , and by
ii) taking into (OSTP) additional conditions for the unknown functions
qe .s/; ˇ.s/; sj s sf , guaranteeing that (the norm of) the sub-Hessian
e
D2 F .j / .t/; t tj , fulfills a certain minimality or boundedness condition.
Proof Follows from definition (4.95) of .j;2/ and representation (4.96) of the
second order tracking error expansion term d 2 z.
d l 0
d z.t/ D Ad l z.t/ C .j;l/ ; t tj ; l D 1; 2; : : : ; (4.98a)
dt .t/
with the same system matrix A and the disturbance terms .j;l/
.t/; t tj , given
by
.j;l/
e
.t/ D M .j /.t/1 D F .j / .t/; 2 lI
pD ; d j z.t/;
d2 k
2
d q.t/; 1 j; k l 1 ; l 2; (4.98b)
dt
4.7 Online Control Corrections: PID Controllers 173
d2 k
where is a polynomial in the variables
pD and d j z.t/; d q.t/; 1 j; k
e dt 2
l 1, having coefficients from the sub-operators DF .j / .t/ of D F .j / .t/ containing
mixed partial derivatives
of F with respect to
pD; z; qR of order D 2; 3; : : : ; l 1
P q/
at .pD ; q; q; R D pD ; q .t/; qP .j / .t/; qR .j / .t/ such that the disturbance .j;l/
.j / .j /
e
which guarantee a minimality or boundedness condition for (the norm of) the sub-
operators of mixed partial derivatives D2 F .j / .t/; t tj ; D 2; 3; : : : ; l0 .
where
z.j / .t/ WD z.t/ z.j / .t/; t tj ;
174 4 Adaptive Optimal Stochastic Trajectory Planning and Control (AOSTPC)
Furthermore,
' .j / D ' .j / t;
z.t/
D ' .j / t;
z1 .t/;
z2 .t/;
z3 .t/ ; t tj ; (4.100d)
Of course, we have
0 1
q.t/ q .j / .t/
B Rt C
B q.s/ q .j / .s/ ds C
z.t/ D
z.j /.t/ D B C ; t tj : (4.101)
@ tj A
P qP .j / .t/
q.t/
qj ."/ WD q j C "
qj ;
qj WD qj q j (4.102a)
P j ;
q
qPj ."/ WD qP j C "
q P j WD qP j qP j (4.102b)
.j / .j /
pD ."/ WD p D C "
pD ;
pD WD pD p D ; (4.102c)
4.7 Online Control Corrections: PID Controllers 175
the solution of following initial value problem based on the dynamic equation (4.4a)
having the initial values and total control input u.t/ given by (4.102a–d):
P "/; q.t;
F pD ."/; q.t; "/; q.t; R "/ D u.t; "/; t tj ; 0 " 1; (4.104a)
where
P q/
F .pD ; q; q; R D M.pD ; q/qR C h.pD ; q; q/;
P (4.104d)
cf. (4.104a).
In the following we assume that problem (4.104a–d) has a unique solution q D
q.t; "/; t tj , for each parameter "; 0 " 1.
176 4 Adaptive Optimal Stochastic Trajectory Planning and Control (AOSTPC)
where
q.tj ; 0/ D q j ; q.t
P j ; 0/ D qP j : (4.105b)
and
.j /
F p D ; q .j / .t/; qP .j / .t/; qR .j / .t/ D u.j /.t/; t tj ;
t t1 (4.107a)
with
q.tj ; 1/ D qj ; q.t
P j ; 1/ D qP j : (4.107b)
However, since the control input is defined by (4.102d), again due to the unique
solvability property of (4.104a–d), (4.107a, b) yields
Zt
qI .t; "/ WD q.s; "/ds; (4.109)
tj
problem (4.104a–d) can be represented by the equivalent 2nd order initial value
problem:
P "/; q.t;
F pD ."/; q.t; "/; q.t; R "/
.j /
D u.j / .t/ C ' .j / t; q.t; "/ q .j / .t/; qI .t; "/ qI .t/;
P "/ qP .j / .t/
q.t; (4.110a)
qPI .t; "/ WD q.t; "/ (4.110b)
with
and
Zt
.j /
qI .t/ WD q .j / .s/ds: (4.111)
tj
with "0 D 0;
" D 1. Moreover, the expansion on the tracking error
q D
q.t/; t tj , is given by
X
1
1 l 1 @
q.t/ D d q.t/.
"/l C q.t; #/.
"/
lŠ Š @"
lD1
X
1
1 l 1 @ q
D d q.t/ C .t; #/; t tj : (4.112b)
lŠ Š @"
lD1
@l q
d l q.t/ WD .t; 0/; t tj ; l D 1; 2; : : : ; (4.112c)
@"l
denote the l-th order differentials of q D q.t; "/ with respect to " at " D "0 D 0.
Differential equations for the differentials d l q.t/; l D 1; 2; : : :, may be obtained by
successive differentiation of the initial value problem (4.104a–d) with respect to "
at " D 0.
P q/
K.pD ; q; q; R WD Fq .pD ; q; q;
P q/
R (4.113a)
P WD FqP .pD ; q; q;
D.pD ; q; q/ P q/
R D hqP .pD ; q; q/
P (4.113b)
P q/
Y .q; q; R WD FpD .pD ; q; q;
P q/
R (4.113c)
M.pD ; q/ WD FqR .pD ; q; q;
P q/
R (4.113d)
P q/
F .pD ; q; q; R D Y .q; q;
P q/p
R D: (4.113e)
4.7 Online Control Corrections: PID Controllers 179
Taking the partial derivative with respect to ", from (4.110a–e) we obtain the
following equations, cf. (4.77a, b),
@q
Y q.t; "/; q.t; R "/
pD C K pD ."/; q.t;
P "/; q.t; P "/; q.t;
R "/ .t; "/
@"
d @q d 2 @q
C D pD ."/; q.t; "/; q.t; P "/; q.t;R "/ .t; "/ C M pD ."/; q.t; "/ .t; "/
dt @" dt 2 @"
.j / @q
D 'q.j / t; q.t; "/ q .j /.t/; qI .t; "/ qI .t/; q.t; P "/ qP .j / .t/ .t; "/
@"
.j / @qI
C 'q.jI / t; q.t; "/ q .j /.t/; qI .t; "/ qI .t/; q.t; P "/ qP .j / .t/ .t; "/
@"
.j / .j / d @q
C 'qP t; q.t; "/ q .j /.t/; qI .t; "/ qI .t/; q.t; P "/ qP .j / .t/ .t; "/:
dt @"
(4.114a)
d @qI @q
.t; "/ D .t; "/ (4.114b)
dt @" @"
@q
.tj ; "/ D
qj (4.114c)
@"
d @q P j
.tj ; "/ D
q (4.114d)
dt @"
@qI
.t; "/ D 0: (4.114e)
@"
Putting now " D "0 D 0, due to (4.106), (4.109), 4.112c) we obtain
Y .j /.t/
pD C K .j /.t/dq.t/ C D .j / .t/dP q.t/ C M .j /.t/dR q.t/
Zt
D 'q.j / .t; 0; 0; 0/dq.t/ C 'q.jI / .t; 0; 0; 0/
.j / P
dq.s/ds C 'qP .t; 0; 0; 0/dq.t/; t tj
tj
(4.115a)
dq.tj / D
qj (4.115b)
dP q.tj / D
q
P j; (4.115c)
where
2
d R WD d dq:
dP q WD dq; dq (4.116a)
dt dt 2
180 4 Adaptive Optimal Stochastic Trajectory Planning and Control (AOSTPC)
Zt
@qI
.t; 0/ D dq.s/ds; (4.116b)
@"
tj
C M .j /.t/1 K .j /.t/ 'q.j /.t; 0; 0; 0/ dq.t/
Zt
M .j /
.t/1 'q.jI /.t; 0; 0; 0/ dq.s/ds D M .j /.t/1 Y .j /
pD ; t tj
tj
(4.118)
M .j /.t/1 D .j / .t/ 'qP .t; 0; 0; 0/ D Kd
.j /
(4.119a)
M .j /.t/1 K .j /.t/ 'q.j /.t; 0; 0; 0/ D Kp (4.119b)
M .j /.t/1 'q.jI /.t; 0; 0; 0/ D Ki : (4.119c)
Thus, we have
Zt
dR q.t/ C Kd dP q.t/ C Kp dq.t/ C Ki dq.s/ds D .j;1/
.t/; t tj ; (4.121a)
tj
with
dq.tj / D
qj (4.121b)
dP q.tj / D
q
P j; (4.121c)
.j;i1
where .t/ is given, cf. (4.83d), by
.j;1/
.t/ WD M .j /.t/1 Y .j / .t/
pD : (4.121d)
for the 1st order tracking error term dq D dq.t/; t D tj . Moreover, the initial
conditions read, see (4.121b, c),
dq.tj / D
qj (4.122b)
dP q.tj / D
q
P j (4.122c)
dq.t R j WD qRj qR .j / ;
R j / D
q (4.122d)
where
0 1
0 I 0
A WD @ 0 0 I A; (4.123b)
Ki Kp Kd
0 1
qj
dz.tj / WD @
q
P j A D
zj (4.123c)
R j
q
and
0 1
dq.t/
dz.t/ WD @ dP q.t/ A : (4.123d)
R
dq.t/
Obviously, the properties of the 1st order error terms dz.t/; E.dz.t/jAtj /, resp.,
t tj , or the stability properties of the 1st order system (4.123a–d) depend on
the eigenvalues of the matrix A.
4.7 Online Control Corrections: PID Controllers 183
Diagonalmatrices Kp ; Kd ; Ki
with diagonal elements pk ; dk ; ik , resp., k D 1; : : : ; n, system (4.122a) is divided
into the separated ordinary differential equations
d«qk C dk dq
R k C pk dP qk C ik dqk D 0; (4.126b)
for arbitrary initial values (4.122b–d). It is well known that property (4.127a) holds
if
where Re.kl / denotes the real part of the zeros k1 ; k2 ; k3 of the characteristic
equation (4.126c). According to the Hurwitz criterion, a necessary and sufficient
condition for (4.127b) is the set of inequalities
Note that
0 1
dk 1 0
H3k WD @ ik pk dk A (4.128d)
0 0 ik
where
E
zj jAtj D E z.tj /jAtj zNj (4.130c)
Zt
a0 e 0 .t tj /
k
zj k C a0 e 0 .t / k P .j;1/ ./kd : (4.132)
tj
Zt ˇ
ˇ
Ca0 e 0 .t / E k P .j;1/ ./kˇAtj d : (4.133)
tj
where
ˇ ˇ
var pD ./ˇAtj WD E k
pD .!/k2 ˇAtj : (4.134b)
In the following we study the inhomogeneous term P .j;1/ .t/ of the 3rd order
linear differential equation (4.122a) in more detail. According to (4.121d) we have
P .j;1/ .t/ D d M .j /.t/1 Y .j / .t/
pD
dt
d .j / 1
D M .t/ Y .j / .t/
pD
dt
d .j /
M .j /.t/1 Y .t/
pD : (4.135a)
dt
and therefore
d .j / 1
k P .j;1/ .t/k
dt M .t/ kY .j / .t/
pD k
C kM .j /.t/1 k YP .j / .t/
pD : (4.135b)
u.j / .tI
pD /; uP .j / .tI
pD /
are linear with respect to
pD D pD E pD .!/jAtj .
4.7 Online Control Corrections: PID Controllers 187
(4.137b)
where
.j / ˇˇ
var ue pD ./; s .j /.t/I q0 ./; ˇ .j / ./ ˇAtj
.j / ˇ
2 ˇ
WD E ue
pD .!/; s .j / .t/I qe ./; ˇ .j / ./ ˇAtj ; (4.137c)
ˇ !
r ˇ
.j / ˇ
var @u e
p D ./; s .j /
.t/I qe ./; ˇ .j /
./ ˇ .j / s .j / .t/ ˇAtj
@s ˇ
2 ˇ
e .j / ˇ
D E @u
@s
p D .!/; s .j /
.t/I q e ./; ˇ .j /
./ ˇA t j ˇ .j / .j /
s .t/ :
(4.137d)
188 4 Adaptive Optimal Stochastic Trajectory Planning and Control (AOSTPC)
According to the representation (4.124a, b) of the first order tracking error form
dz D dz.t/; t tj , the behavior of dz.t/ is determined mainly by the system matrix
A and the “inhomogeneous term” P .j;1/ .t/; t tj . Obviously, this term plays the
same role as the expression .j;1/ .t/; t tj , in the representation (4.84a) for the
first order error term in case of PD-controllers.
However, in the present case of PID-control the error estimates (4.137a–d) show
that for a satisfactory behavior of the first order error term dz D dz.t/; t tj ,
besides the control constraints (4.8a–c), (4.21a–c), (4.30a–f) for
u.t/ D ue pD ; sI qe .:/; ˇ.:/ with s D s.t/; (4.138a)
here, also corresponding constraints for the input rate, i.e. the time derivative of the
control
@ue p
uP .t/ D pD ; sI qe .:/; ˇ.:/ ˇ.s/ with s D s.t/ (4.138b)
@s
are needed!
The above results can be summarized by the following theorem:
Theorem 4.7.2 Suppose that the matrices Kp ; Ki ; Kd are selected such that the
fundamental matrix ˚.t; / D e A.t / ; t , is exponentially stable (cf. Theo-
rem 4.7.1). Then, based on the definition (4.120a–c) of the linear approximation of
the PID-controller, the following properties hold:
a) Asymptotic local stability in the mean
ˇ
E dz.t/ˇAtj ! 0; t ! 1 (4.139a)
Zt q
.j /
C a0 e 0 .t / M .j /./1 @ue s .j / ./ ˇ .j / s .j / ./ d ; t tj
@s
tj
(4.139b)
4.7 Online Control Corrections: PID Controllers 189
with
r
.j / ˇˇ
u.je / .s/ WD var ue pD .:/; sI qe ./; ˇ .j / ./ ˇAtj ; (4.139c)
r
.j /
@u
e .j / ˇˇ
@ue .s/ WD var pD ./; sI qe ./; ˇ .j / ./ ˇAtj ; (4.139d)
@s @s
s sj . Moreover,
ˇ ˇ
E kdz.t/kˇAtj a0 e 0 .t tj / E k
zj kˇAtj
!
Zt d M .j /.t/1 Y .j / .t/
C a0 e 0 .t /
./ d .j / ; t tj ;
pD
dt
tj
(4.139e)
where
q ˇ
p.jD/ WD var pD .:/ˇAtj : (4.139f)
Zt q
.j /
e 0 .t / M .j /./1 @ue s .j / ./ ˇ .j / s .j / ./ d
@s
tj
sZ
.j / .t /
.j / .j / 1 .j /
D e 0 t t .j / .s/
M t .s/ @ue .s/ds; (4.140b)
@s
sj
Several terms in the above estimates of the 1st order tracking error dz D dz.t/; t
tj , are influenced by means of (OSTP) as shown in the following:
i) Optimal velocity profile ˇ .j /
For minimum-time and related substitute problems the total runtime
sZ
.j / .t /
Zsf
1 1 .j /
p ds p ds D tf tj (4.141)
ˇ .j / .s/ ˇ .j / .s/
sj sj
is minimized by (OSTP).
ii) Properties of the coefficient 0
and 4.7.2 the matrices Kp ; Ki ; Kd can be selected
According to Theorem 4.7.1
such that real parts Re kl ; k D 1; : : : ; n; l D 1; 2; 3, of the eigenvalues
kl ; k D 1; : : : ; n; l D 1; 2; 3, of the matrix A, cf. (4.123b), are negative, see
(4.127b), (4.128a–d). Then, the decisive coefficient 0 > 0 in the norm estimate
(4.131) of the fundamental matrix ˚.t; /; t , can be selected such that
In the following we suppose that the bounds umin ; umax are given deterministic
vectors. After the transformation
s D s .j / .t/; t tj ;
.j /
from the time domain Œtj ; tf to the s-domain Œsj ; sf , see (4.49b), due to
(4.21a) we have the stochastic constraint for qe ./; ˇ./ :
umin ue pD .!/; sI qe ./; ˇ./ umax ; sj s sf :
Demanding that the above constraint holds at least with the probability ˛u , we
get, cf. 4.30a, the probabilistic constraint
ˇ
ˇ
P umin ue pD .!/; sI qe ./; ˇ./ umax ˇAtj ˛u ; sj s sf :
(4.143a)
4.7 Online Control Corrections: PID Controllers 191
where
ˇˇ
e WDE ue pD .!/; sI qe ./; ˇ ./ ˇAtj
u.j / .j / .j /
.j /
Due p D ; sI qe.j / ./; ˇ .j / ./ : (4.144b)
Hence, the sufficient condition (4.144a) for the reliability constraint (4.143a)
yields the variance constraint
!
.j /
ˇ Ztf
ˇ 2 ˇˇ
E J u./ ˇAtj D E
u.t / dt ˇAtj
tj
Zsf 2 ˇˇ ds
D E ue pD .!/; sI qe ./; ˇ./ ˇAtj p
ˇ.s/
sj
Zsf
.j / 2 ˇˇ
D E ue pD .!/; sI qe ./; ˇ./ ue p D ; sI qe ./; ˇ./ ˇAtj
sj
192 4 Adaptive Optimal Stochastic Trajectory Planning and Control (AOSTPC)
.j / 2 ds
C ue p D ; sI qe ./; ˇ./ p
ˇ.s/
Zsf
.j / 2 ds
D u2e .s/ C ue p D ; sI qe ./; ˇ./ p ; (4.145a)
ˇ.s/
sj
where
2 ˇˇ
W D E ue pD .!/; sI qe ./; ˇ./ ue p D ; sI qe ./; ˇ./ ˇˇAtj
.j /
u2e .s/
2 ˇˇ
D E ue
pD .!/; sI qe ./; ˇ./ ˇˇAtj : (4.145b)
!
.j /
Zsf Ztf
/ 2 ˇ
u.je / .s/2 C u.j p ds D ku.t/k dt ˇAtj ;
2
e .s/ min E
ˇ .j / .s/ qe ./;ˇ./
sj s:t: (4.8a–c); tj
(4.9a–d)
(4.146)
where u.j /
e .s/ is defined by (4.144b).
v) Decreasing stochastic uncertainty
According to (4.132), (4.133) and (4.134a, b) we have
ˇ
ˇ
E kdz.t/kAtj ˛0 e 0 .t tj / E k
zj kˇAtj
!
Zt d M .j /.t/1 Y .j / .t/ q
0 .t /
ˇ
C ao e ./
d var pD .:/ˇAtj :
dt
tj
(4.147)
Thus, the mean absolute 1st order tracking error can be decreased further
by removing step by step the uncertainty about the vector pD D pD .!/
of dynamic parameter. This is done in practice by a parameter identification
procedure running parallel to control process of the robot.
vi) Chance constraints for the input rate
According to the representation (4.138b) of the input or control rate we have
@ue p
uP .t/ D pD ; sI qe ./; ˇ./ ˇ.s/; s D s.t/:
@s
4.7 Online Control Corrections: PID Controllers 193
with given, fixed vector bounds uP min uP max , for the input rate @ue
@s
with respect
to the path parameter s sj we obtain the constraint
If we require that the input rate condition (4.148a) holds at least with probability
˛uP , then corresponding to (4.143a) we get the chance constraint
ˇ !
uP min @ue uP max ˇˇ
P p pD .!/; sI qe ./; ˇ./ p ˇAtj ˛uP : (4.149)
ˇ.s/ @s ˇ.s/ ˇ
In the same way as in (iii), condition (4.149) can be guaranteed, cf. (4.139d),
by
.j / 2 1
@ue uP c
@u .s/2 C 1˛uP
.j /
.s/ p min 2 ; sj s sf ;
.j / .s/ 1kn uP k
@s @s .j
ˇ .s/ / ˇ
(4.150a)
where
Ztf !
ˇ 2 ˇˇ
ˇ
E J uP ./ Atj WD E
uP .t/ dt ˇAtj : (4.151a)
tj
194 4 Adaptive Optimal Stochastic Trajectory Planning and Control (AOSTPC)
Ztf 2 ˇ !
ˇ @ue p ds ˇˇ
E J uP ./ˇAtj D E
@s pD .!/; sI qe ./; ˇ./ ˇ.s/ p ˇAtj
ˇ.s/ ˇ
tj
Zsf ˇ
@ue
2 ˇˇ p
D E pD .!/; sI qe ./; ˇ./ ˇAtj ˇ.s/ds
@s
sj
Zsf
@ue ˇˇ
D var pD .!/; sI qe ./; ˇ./ ˇAtj
@s
sj
@u
e .j /
2 p
C p D ; sI qe ./; ˇ./ ˇ.s/ds: (4.151b)
@s
If we consider
Ztf !
ˇ ˇˇ
ˇ
E JQ uP ./ ˇAtj WD E uP .t/ dt ˇAtj ; (4.152a)
tj
then we get
ˇ Z f @u
s ˇ
ˇ e
ˇˇ
Q
E J uP ./ ˇAtj D E pD .!/; sI qe ./; ˇ./ ˇAtj ds: (4.152b)
@s
sj
Chapter 5
Optimal Design of Regulators
The optimal design of regulators is often based on the use of given, fixed nominal
values of initial conditions, pay loads and other model parameters. However, due
to
• variations of the material properties,
• measurement errors (e.g. in case of parameter identification)
• modeling errors (complexity of real systems)
• uncertainty on the working environment, the task to be executed, etc.,
the true payload, initial conditions, and model parameters, like sizing parameters,
mass values, gravity centers, moments of inertia, friction, tolerances, adjustment set-
ting error, etc., are not known exactly in practice. Hence, a predetermined (optimal)
regulator should be “robust”, i.e., the controller should guarantee satisfying results
also in case of variations of the initial conditions, load parameters and further model
parameters.
Robust controls have been considered up to now mainly for uncertainty models
based on given fixed sets of parameters, e.g. a certain multiple intervals, assumed
to contain the unknown, true parameter. In this case one requires then often that the
controlled system fulfills certain properties, as e.g. certain stability properties for
all parameter vectors in the given parameter domain. If the required property can
be described by a scalar criterion, then the controller design is based on a minimax
criterion, such as the H 1 -criterion.
Since in many cases parameter uncertainty can be modeled more adequately
by means of stochastic parameter models, in the following we suppose that the
parameters involved in the regulator design problem are realizations of a random
vector having a known or at least partly known joint probability distribution.
The determination of an optimal controller under uncertainty with respect to
model parameters, working neighborhood, modeling assumptions, etc., is a decision
theoretical problem. Criteria of the “holds for all parameter vectors in a given
set” and the minmax-criterion are very pessimistic and often too strong. Indeed,
in many cases the available a priori and empirical information on the dynamic
system and its working neighborhood allows a more adequate, flexible description
of the uncertainty situation by means of stochastic approaches. Thus, it is often
more appropriate to model unknown and varying initial values, external loads and
other model parameters as well as modeling errors, e.g. incomplete representation
of the dynamic system, by means of realizations of a random vector, a random
function with a given or at least partly known probability distribution. Consequently,
the optimal design of robust regulators is based on an optimization problem under
stochastic uncertainty.
According to Sects. 4.6 and 4.7, at each stage j D 0; 1; 2; : : : the control input u.t/
is defined by
where
u.j /.t/ WD ' .j / t;
z.j / .t/ (5.1b)
D ' .j / t; z.t/ z.j / .t/
D ' .j / t; q.t/ q .j / .t/; q.t/
P qP .j / .t/
Inserting the total control input u.t/; t tj , into the dynamic equation (4.4a), then
the effective further course of the trajectory q.t/; t tj , will be described by the
system of ordinary differential equations (ODE):
P
F .pD ; q.t/; q.t/; R
q.t// D u.j /.t/ C
u.j /.t/ (5.2a)
D u .t/ C '
.j / .j /
t; q.t/ q .t/; q.t/
.j /
P qP .t/ ; t tj
.j /
q.tj / D qj ; q.t
P j / D qP j ; (5.2b)
q.t/ D
q .j /.t/ WD q.t/ q .j / .t/; t tj (5.3a)
198 5 Optimal Design of Regulators
P
q.t/ P .j /.t/ WD q.t/
D
q P qP .j / .t/; t tj (5.3b)
R
q.t/ R .j /.t/ WD q.t/
D
q R qR .j / .t/; t tj (5.3c)
and
qj WD q.tj / q .j / .tj / D qj qj (5.3d)
P j WD q.t
q P j / qP .j / .tj / D qPj qPj (5.3e)
.j /
pD WD pD p D ; (5.3f)
system (5.2a, b) can be rewritten by the system of differential equations for the
tracking error:
.j /
F p D C
pD ; q .j / .t/ C
q.t/; qP .j / .t/
P
C
q.t/; R
qR .j / .t/ C
q.t/
P
D u.j /.t/ C ' .j / t;
q.t/;
q.t/ ; t tj (5.4a)
P j / D
q
q.tj / D
qj ;
q.t P j: (5.4b)
depending on the whole feedback function '.t; ; / and/or on the individual control
inputs
u.t/.
For given model parameters and initial conditions, the quality or performance of
a feedback-function ' .j / D ' .j / .t;
z/ can be evaluated by the functional
J D J .j / ' .j /; pD ; zj (5.6a)
.j /
Ztf
D P
c.t;
q.t/;
q.t/ / C .t; '.t; ; /;
u.t// dt
„ ƒ‚ … „ ƒ‚ …
tj D
z.t / regulator costs
„ ƒ‚ …
costs of the
tracking error
P
c.t;
z/ D
q.t/T Cq
q.t/ C
q.t/C P
qP
q.t/ (5.6c)
involving positive (semi-) definite weight matrices Cq D Cq .t/; CqP D CqP .t/:
Moreover, the regulator costs can be defined by
subject to
P
F pD .!/; q .j / .t/ C
q.t/; qP .j / .t/ C
q.t/; R
qR .j / .t/ C
q.t/
P
D u.j /.t/ C ' t;
q.t/;
q.t/ ; t tj ; a.s. (almost sure) (5.7b)
P j / D
q
q.tj / D
qj .!/;
q.t P j .!/; a.s.; (5.7c)
where
u.t/ is defined by (5.1b–d).
200 5 Optimal Design of Regulators
subject to
P D f0 pD .!/; z.j / .t/ C
z.t/ zP.j / .t/
z (5.8b)
CB pD .!/; z.j / .t/ C
z.t/ u.j / .t/ C '.t;
z.t// ; t tj ; a.s.
z.tj / D
zj .!/; a.s.: (5.8c)
Remark 5.1 Note that state and control constraints can also be considered within
the stochastic optimal regulator optimization problem (5.7a–c).
Considering the m-th order Taylor expansion of the feedback function ' D '.t;
z
with respect to
z, we obtain the representation
X
m X
m
'.t;
z/ WD Gl .t/ .
z/ D
l
Gl .t/ .
z;
z; : : : ;
z/ : (5.9b)
„ ƒ‚ … „ƒ‚… „ ƒ‚ …
n comp. lD1 lD1
l
multilin. form
l-th order
5.2 Parametric Regulator Models 201
'.t;
z/ D G.t/h.
z/ (5.9c)
X
2n
z D
zk ek ; (5.10a)
kD1
where
zk D
qk ; k D 1; : : : ; n; (5.10b)
P l;
znCl D
q l D 1; : : : ; n: (5.10c)
X
2n
D Gl .t/ .ek1 ; : : : ; ekl /
zk1 : : :
zkl (5.11a)
k1 ;k2 ;:::;kl D1
X
2n
Gl .t/ .
z/l D gk1 :::kl .t/
zk1 : : :
zkl : (5.11c)
k1 ;k2 ;:::;kl D1
202 5 Optimal Design of Regulators
Obviously, (5.11c) depends linearly on the functions gk1 :::kl .t/. According to (5.9c)
and (5.11c), the matrix function G D G.t/ has the following columns:
G.t/ D .g1 .t/; : : : ; g2n .t/; g11 .t/; : : : ; g2n2n .t/; : : : ; g1:::1 .t/; : : : ; g2n:::2n .t//:
„ ƒ‚ …
linear regulator
„ ƒ‚ …
quadratic regulator
„ ƒ‚ …
polynomial regulator m-th order
(5.12)
'.t;
z/ D
(5.9a)
Gp .t/
q C P
Gd .t/
q
„ ƒ‚ …
n-vector
P
n P
n
D Gp .t/
qk ek C Gd .t/ P l el
q
kD1 lD1
n
P P
n
D Gp .t/ek
qk C P l
.Gd .t/el /
q
kD1 lD1
Pn P
n
(5.11b)
D gpk .t/
qk C P l
gdl .t/
q
kD1 lD1
(5.12)
q.t/
D G.t/ P
q.t/
with G.t/ WD gp1 .t/; : : : ; gpn .t/; gd1 .t/; : : : ; gdn .t/ .
'.t;
z/ D G.t/ h.
z/:
„ƒ‚… „ ƒ‚ …
unknown given
subject to
P
F pD .!/; q .j / .t/ C
q.t/; qP .j / .t/ C
q.t/; R
qR .j / .t/ C
q.t/ (5.14b)
P
D u.j / .t/ C G.t/h.
q;
q.t//; t tj ; a.s. (almost sure)
P j / D
q
q.tj / D
qj .!/;
q.t P j .!/; f.s. (5.14c)
Working with the expected total costs, in this section we have to determine now the
conditional expectation of the functional (5.6a):
.j /
Ztf
E J .j / jAtj D P
E c.t;
q.t/;
q.t// C .t; '.t; ; /;
u.t// jAtj dt:
tj
(5.15)
P j /;
q.t/ D
q.t;
pD ;
qj ;
q
.j /
tj t tf ; (5.16a)
P
q.t/ P
D
q.t; P j /;
pD ;
qj ;
q
.j /
tj t tf ; (5.16b)
where
pD D
pD .!/ (5.16c)
qj D
qj .!/ (5.16d)
q P j .!/:
P j D
q (5.16e)
204 5 Optimal Design of Regulators
P j .!/ ;
q
E c t;
q t;
pD .!/;
qj .!/;
q P t;
pD .!/;
qj .!/;
!
ˇ
P j .!/ ˇˇ Atj
q
and
!
ˇ
ˇ
E t; '.t; ; /;
u.t; !/ ˇ Atj (5.17a)
with
u.t; !/ D '.t;
z.t;
pD .!/;
zj .!/// (5.17b)
q.t/
z.t/ D P (5.17c)
q.t/
qj
zj D P j : (5.17d)
q
E.c.t;
z.t; a.!///jAtj / D E.c.t;
Q a.!//jAtj / (5.19a)
where
Q a/ WD c.t;
z.t; a//:
c.t; (5.19b)
1
Q a/ D c.t;
c.t; Q a/ C ra c.t;
Q a/T .a a/ C .a a/T ra2 c.t;
Q a/.a a/
2
1
C ra3 c.t;
Q a/ .a a/3 C : : : ; (5.19c)
3Š
and therefore
1
Q a.!//jAtj D c.t;
E c.t; Q a/ C E .a a/T ra2 c.t;
Q a/.a a/jAtj
2
1
C E ra3 c.t;
Q a/ .a.!/ a/3 jAtj C : : : : (5.19d)
3Š
Here, (5.19d) is obtained from (5.19c) due to the following equation:
E ra c.t;
Q a/T .a.!/ a/jAtj D ra c.t;
Q a/T E a.!/jAtj „ƒ‚…
a D 0:
„ ƒ‚ …
Da.j /
Da.j /
„ ƒ‚ …
D0
Q a/
Computation of ra c.t;
Q a/ D c.t;
z.t; a//:
c.t;
206 5 Optimal Design of Regulators
@cQ X @c
2n
@
zk
.t; a/ D .t;
z.t; a// .t; a/; j D 1; : : : ; : (5.20)
@aj @
zk @aj
kD1
hence,
T
ra c.t;
Q a/ D ra
z.t; a/ r
z c t;
z.t; a/ : (5.21)
P WD
q T Cq
q C
q
c.t;
q;
q/ P t tj ;
P T CqP
q;
(ii) Suppose that the cost function for the evaluation of the regulator expenses is
given by (5.6d), hence,
Under the cost assumptions .i /; .ii/ in Sect. 5.3.2 and because of (5.1b), thus
P
u WD '.t;
q.t/;
q.t//;
from (5.7a–c) we obtain now (at stage j ) the following optimal regulator problem
under stochastic uncertainty:
.j /
Ztf
min E P
q.t/T Cq
q.t/ C
q.t/C P
qP
q.t/
tj
ˇ !
T
ˇ
P
C' t;
q.t/;
q.t/ P
Cu ' t;
q.t/;
q.t/ dt ˇˇAtj (5.22a)
s.t.
.j / P
F p D C
pD .!/; q .j / .t/ C
q.t/; qP .j / .t/ C
q.t/; R
qR .j / .t/ C
q.t/
D u.j /.t/ C ' t;
q.t/;
q.t/P f.s. ; t tj ; (5.22b)
P j / D
q
q.tj / D
qj .!/;
q.t P j .!/ f.s. (5.22c)
As already mentioned in Sect. 5.3, (5.16a, b), for the tracking error we obtain
from (5.22b, c) the representation:
P j ; t tj
q.t/ D
q t;
pD ;
qj ;
q
P
q.t/ P t;
pD ;
qj ;
q
D
q P j ; t tj ;
208 5 Optimal Design of Regulators
where
pD D
pD .!/;
qj D
qj .!/;
q P j D
q P j .!/. Under appropriate
regularity assumptions on the functions F; ' as well as on q .j / and u.j / , the solution
q P j/
q.t;
pD ;
qj ;
q
z D P D P P j / ; t tj ;
q
q.t;
pD ;
qj ;
q
@
q @
q @
q P j
q.t/ D
q.t; 0/ C .t; 0/
pD C .t; 0/
qj C .t; 0/
q
@
pD @
qj P j
@
q
C::: (5.23a)
P
@
q P
@
q P
@
q
P
q.t/ P
D
q.t; 0/ C .t; 0
pD C .t; 0/
qj C P j
.t; 0/
q
@
pD @
qj P
@
qj
C::: (5.23b)
0 1
0
with 0 WD @ 0 A and the Jacobians
0
0 1
@
q1 @
q1 @
q1
B @
pD @
pD @
pD C
B 1 2 C
@
q P j / WD B
.t;
pD ;
qj ;
q B
:: :: :: C P j /;
C .t;
pD ;
qj ;
q
@
pD B : : : C
@ @
qn @
qn @
qn A
@
pD1 @
pD2 @
pD
'.t; 0; 0/ D 0;
P j/
q.t/ D
q.t;
pD ;
qj ;
q
@
q @
q @
q P j C :::
D .t; 0/
pD C .t; 0/
qj C .t; 0/
q (5.25a)
@
pD @
qj P j
@
q
P
q.t/ P
D
q.t; P j/
pD ;
qj ;
q
P
@
q P
@
q P
@
q
D .t; 0/
pD C .t; 0/
qj C P j C :::
.t; 0/
q (5.25b)
@
pD @
qj P j
@
q
@
q
The Jacobians @
p ; @
q ; : : : arising in (5.25a, b) can be obtained by partial
D @
qj
differentiation of the system of differential equations (5.22b, c) with respect to
pD ;
qj ;
qP j ; cf. Sect. 4.6.2.
For the computation of the expectation in (5.26) also the feedback function ' D
P is approximated by means of Taylor expansion at
q D 0;
q
'.t;
q;
q/ P D 0.
Because of (5.1d), hence, '.t; 0; 0/ D 0; we get
@' @'
.t; 0; 0/; .t; 0; 0/; : : : (5.27b)
@
q P
@
q
where
E fQ.j / jAj
@'
P
WD E
q.t/T Cq
q.t/ C
q.t/ T P
CqP
q.t/ C .t; 0; 0/
q.t/
@
q
!
T @' ˇ
@' P @' P ˇ
C .t; 0; 0/
q.t/ Cu .t; 0; 0/
q.t/ C .t; 0; 0/
q.t/ ˇAtj
P
@
q @
q P
@
q
@'
P
D E
q.t/T Cq
q.t/ C
q.t/ T P
CqP
q.t/ C
q.t/T .t; 0; 0/T Cu
@
q
!
@' ˇ
P @' @' P ˇ
C
q.t/ T
.t; 0; 0/T Cu .t; 0; 0/
q.t/ C .t; 0; 0/
q.t/ ˇAtj :
P
@
q @
q P
@
q
(5.28b)
This yields
E fQ.j /jAj WD E
q.t/T Cq
q.t/ C
q.t/
P T P
CqP
q.t/
@' @'
C
q.t/T .t; 0; 0/T Cu .t; 0; 0/
q.t/
@
q @
q
P @' @' P
C
q.t/ T
.t; 0; 0/T Cu .t; 0; 0/
q.t/
P
@
q P
@
q
@' @' P
C
q.t/T .t; 0; 0/T Cu .t; 0; 0/
q.t/
@
q P
@
q
!
ˇ
P T @' @' ˇ
C
q.t/ .t; 0; 0/T Cu .t; 0; 0/
q.t/ˇAtj
P
@
q @
q
(5.28c)
5.4 Approximation of the Stochastic Regulator Optimization Problem 211
and therefore
@' @'
E fQ.j / jAj D E
q.t/T Cq C .t; 0; 0/T Cu .t; 0; 0/
q.t/
@
q @
q
@' @'
P
C
q.t/ T
CqP C .t; 0; 0/T Cu P
.t; 0; 0/
q.t/
@
qP P
@
q
!
T @' @' ˇ
C2
q.t/ T
.t; 0; 0/ Cu P
.t; 0; 0/
q.t/ ˇAt : (5.28d)
@
q P
@
q
j
Putting
@' @'
Qq D Qq .t; '/ WD Cq C .t; 0; 0/T Cu .t; 0; 0/ (5.29a)
@
q @
q
@' @'
QqP D QqP .t; '/ WD CqP C .t; 0; 0/T Cu .t; 0; 0/ (5.29b)
P
@
q P
@
q
@' @'
Qu D Qu .t; '/ WD .t; 0; 0/T Cu .t; 0; 0/; (5.29c)
@
q P
@
q
we obtain
@' @'
fQ.j / WD
q.t/T Cq C .t; 0; 0/T Cu .t; 0; 0/
q.t/
@
q @
q
@' @'
P
C
q.t/ T
CqP C .t; 0; 0/T Cu P
.t; 0; 0/
q.t/
P
@
q P
@
q
@' @' P
C 2
q.t/T .t; 0; 0/T Cu .t; 0; 0/
q.t/
@
q P
@
q
Qq Qu
q.t/
P
D
q.t/T ;
q.t/ T
QuT QqP P
q.t/
D
z.t/T Q.t; '/
z.t/: (5.30a)
Here,
q.t/
z.t/ D P ; (5.30b)
q.t/
Qq Qu
Q.t; '/ WD
QuT QqP
@' @' @' !
Cq C @
q
.t; 0; 0/T Cu @
q .t; 0; 0/ @
q
.t; 0; 0/T Cu @@'
P .t; 0; 0/
D @' @'
q
: (5.30c)
T
P .t; 0; 0/ Cu @
q .t; 0; 0/
@
q
CqP C @@' T @'
P .t; 0; 0/ Cu @
q
q P .t; 0; 0/
212 5 Optimal Design of Regulators
Obviously, Qq ; QqP and also Q D Q.t; '/ are symmetric matrices. Since
Cq ; CqP ; Cu are positive (semi-)definite matrices, according to the definition (5.28b)
of the approximate total cost function we get
fQ.j /.t;
z/ > ./ 0; for all
z ¤ 0:
Thus, we have the following result on the approximate total cost function fQ.j / D
fQ.j / .t;
z/:
Lemma 5.1 The quadratic approximation fQ.j / =fQ.j /.t;
z/ of f .j / =f .j /.t;
z/ is
q
a positive semidefinite quadratic form in
z D P . If Cq D Cq .t/; CqP D CqP .t/
q
are positive definite weight matrices, then also the matrix Q D Q.t; '/ is positive
definite, and fQ.j / is then a positive definite quadratic form in
z.
Proof The first part of the assertion is clear. Suppose now that Cq ; CqP are positive
definite weight matrices. Then, according to the definition of fQ.j / D fQ.j / .t;
z/
by (5.28b, c) and since Cu is positive (semi-)definite, we have fQ.j / .t;
z/ D 0 if
and only if
z/ D 0.
For the approximate computation of the expectation E fQ.j /jAtj we use now
only the linear terms of (5.25a, b), hence
q.t/
z.t/ D P
q.t/
! ! 0 @
q
1
@
q @
q .t; 0/
.t; 0/ .t; 0/ P j
Š @
pD
P
pD C
@
qj
P
qj C @
@
q
P
P j
A
q
@
q @
q @
q
@
pD
.t; 0/ @
qj .t; 0/ P j
@
q
.t; 0/
@
z @
z @
z P j:
D .t; 0/
pD C .t; 0/
qj C .t; 0/
q (5.31)
@
pD @
qj P j
@
q
fQ.j / D
zT Q.t; '/
z
T T T !
@
z @
z P j @
z
D
pD T
.t; 0/ C
qj T
.t; 0/ C
q T
.t; 0/
@
pD @
qj P j
@
q
!
@
z @
z @
z P
Q .t; 0/
pD C .t; 0/
qj C .t; 0/
qj
@
pD @
qj @
qP j
T
@
z @
z
D
pDT
.t; 0/ Q .t; 0/
pD
@
pD @
pD
5.4 Approximation of the Stochastic Regulator Optimization Problem 213
T
@
z @
z
C
qjT .t; 0/ Q .t; 0/
qj
@
qj @
qj
T
C P jT @
z .t; 0/ Q @
z .t; 0/
q
q P j
P j
@
q P j
@
q
T
@
z @
z
C 2
qjT .t; 0/ Q .t; 0/
pD
@
qj @
pD
T
C P jT @
z .t; 0/ Q @
z .t; 0/
pD
2
q
P j
@
q @
pD
T
C P jT @
z .t; 0/ Q @
z .t; 0/
qj :
2
q (5.32)
P j
@
q @
qj
pD .!/;
qj .!/ as well as
P j .!/
pD .!/;
q
(5.32) yields
ˇ ˇ
ˇ ˇ
E fQ.j / ˇAtj D E
pD .!/T P
pD
pD .!/ˇAtj
ˇ ˇ
ˇ P j .!/ˇˇAtj
P j .!/T P P
q
C E
qj .!/T P
qj
qj .!/ˇAtj C E
q
qj
ˇ
P j .!/T P P ˇ
C 2E
q
qj ;
qj
qj .!/ˇAtj (5.33a)
T
@
z @
z
P j ;
qj WD
P
q .t; 0/ Q.t; '/ .t; 0/ : (5.33e)
P j
@
q @
qj
ˇ X
ˇ
ˇ ˇ
E
pD .!/T P
pD
pD .!/ˇAtj D E pkl
pDk .!/
pDl .!/ˇAtj
k;lD1
X
ˇ
ˇ
D pkl E
pDk .!/
pDl .!/ˇAtj
k;lD1
X
D pkl cov.j / pDk ; pDl : (5.34a)
k;lD1
Here, the last equation follows, cf. (5.16c), from the following equations:
ˇˇ
cov.j / pDk ; pDl WD E pDk .!/ pDk .j / pDl .!/ pDl .j / ˇAtj
ˇ
ˇ
D E
pDk .!/
pDl .!/ˇAtj ; k; l D 1; : : : ; : (5.34b)
ˇ X
ˇ
E
pD .!/T P
pD
pD .!/ˇAtj D pkl cov.j / pDk ; pDl
k;lD1
X
D pkl cov.j / pDl ; pDk
k;lD1
X
X
D pkl cov.j / pDl ; pDk
kD1 lD1
5.4 Approximation of the Stochastic Regulator Optimization Problem 215
X
D k-th row of P k-th column of cov.j / pD ./
kD1
D tr P
pD cov.j / pD ./ : (5.35a)
ˇ X
ˇ
P j .!/T P P ˇ P j k .!/
qj l .!/ˇˇAtj
E
q
qj ;
qj
qj .!/ ˇAt j D E p kl
q
k;lD1
X
D P j k ./;
qj l ./
pkl cov.j /
q
k;lD1
X
D P j k ./
pkl cov.j /
qj l ./;
q
k;lD1
X
X
D P j k ./
pkl cov.j /
qj l ./;
q
kD1 lD1
X
D P j ./ ;
k-th row of P k-th column of cov.j /
qj ./;
q (5.35d)
kD1
where
.j / ˇˇ
P j ./ WD E
qj .!/
qj .j /
q
cov.j /
qj ./;
q P j .!/
q
P j ˇAtj
(5.35e)
As can be seen from (5.33a–e) and (5.36), for the approximate computation of the
conditional expectation of the objective function (5.22a) we need the Jacobians or
“sensitivities”:
@
q P
@
q @
q P
@
q @
q P
@
q
.t; 0/; .t; 0/; .t; 0/; .t; 0/; .t; 0/; .t; 0/;
@
pD @
pD @
qj @
qj P j
@
q P j
@
q
of the functions
P
q D
q.t;
pD ;
q;
q/; t tj
P D
q.t;
q P P
pD ;
q;
q/; t tj :
Because of
P
@
q @ @
q @ @
q
D D (5.38d)
@
pD @
pD @t @t @
pD
P
@
q @ @
q @ @
q
D D (5.38e)
@
qj @
qj @t @t @
qj
!
P
@
q @ @
q @ @
q
D D ; (5.38f)
P j
@
q P j
@
q @t @t @
q P j
@
qP P P
the derivatives @
p ; @
q ; @
q
D @
qj @
q P j follow, as shown below, together with derivatives
defined by (5.38a–c).
@
q P j /; t tj ; i D 1; : : : ; ;
t 7! .t;
pD ;
qj ;
q
@
pDi
218 5 Optimal Design of Regulators
@F @
pD @F @
q @F @
qP @F @
qR
C C C
@pD @
pDi @q @
pDi @qP @
pDi @qR @
pDi
@' @
q P
@' @
q
D C ; i D 1; : : : ; ; (5.39a)
@
q @
pDi P @
pDi
@
q
@
pD
where @
p Di
D ei (= i-th unit vector). Differentiation of the initial conditions (5.37b,
c) with respect to pDi yields
@
q P j / D 0; i D 1; : : : ; ;
.tj ;
pD ;
qj ;
q (5.39b)
@
pDi
P
@
q P j / D 0; i D 1; : : : ; ;
.tj ;
pD ;
qj ;
q (5.39c)
@
pDi
pD , respectively.
Then, the initial value problems for the sensitivities
@
q P
@
q d @
q
t 7! .t; 0/; t 7! .t; 0/ D .t; 0/
@
pDi @
pDi dt @
pDi
1 0
pD
follow by insertion of
a D @
qj A D 0 into (5.39a–c). Defining still the
P j
q
expressions
@F ˇˇ @F .j / .j /
ˇ WD p D ; q .t/; qP .j / .t/; qR .j / .t/ ; t tj ; (5.40a)
@pD .j / @pD
@F ˇˇ @F .j / .j /
ˇ WD p D ; q .t/; qP .j / .t/; qR .j / .t/ ; t tj ; (5.40b)
@q .j / @q
@F ˇˇ @F .j / .j /
ˇ WD p D ; q .t/; qP .j / .t/; qR .j / .t/ ; t tj ; (5.40c)
@qP .j / @qP
@F ˇˇ @F .j / .j /
ˇ WD p D ; q .t/; qP .j / .t/; qR .j / .t/ ; t tj ; (5.40d)
@qR .j / @qR
5.5 Computation of the Derivatives of the Tracking Error 219
@F ˇˇ @F ˇˇ @
q @F ˇˇ P
@
q @F ˇˇ R
@
q
ˇ C ˇ .t; 0/ C ˇ .t; 0/ C ˇ .t; 0/
@pDi .j / @q .j / @
pDi @qP .j / @
pDi @qR .j / @
pDi
@' @
q @' P
@
q
D .t; 0/ .t; 0/ C .t; 0/ .t; 0/; t tj (5.41a)
@
q @
pDi P
@
q @
pDi
@
q
.tj ; 0/ D 0 (5.41b)
@
pDi
P
@
q
.tj ; 0/ D 0: (5.41c)
@
pDi
@F ˇˇ @F ˇˇ @
q @F ˇˇ @
q P @F ˇˇ @
q R
ˇ C ˇ .t; 0/ C ˇ .t; 0/ C ˇ .t; 0/;
@pD .j / @q .j / @
pD @qP .j / @
pD @qR .j / @
pD
@' @
q @' P
@
q
D .t; 0/ .t; 0/ C .t; 0/ .t; 0/; t tj (5.42a)
@
q @
pD P
@
q @
pD
@
q
.tj ; 0/ D 0 (5.42b)
@
pD
P
@
q
.tj ; 0/ D 0: (5.42c)
@
pD
@
q
.t/ WD .t; 0/; t tj ; (5.43)
@
pDi
@F ˇˇ @F .j / .j /
M .j /.t/ WD ˇ D p D ; q .t/; qP .j / .t/; qR .j / .t/ ; t tj ; (5.45a)
@qR .j / @qR
@F ˇˇ @F .j / .j /
D .j / .t/ WD ˇ D p D ; q .t/; qP .j / .t/; qR .j / .t/ ; t tj ; (5.45b)
@qP .j / @qP
ˇ
@F ˇ @F .j / .j /
K .j /.t/ WD ˇ D p D ; q .t/; qP .j / .t/; qR .j / .t/ ; t tj ; (5.45c)
@q .j / @q
@F ˇ ˇ @F .j / .j /
Y .j /.t/ WD ˇ D p D ; q .t/; qP .j / .t/; qR .j / .t/ ; t tj : (5.45d)
@pD .j / @pD
Putting
.j / .j /
Y .j / .t/ WD y1 ; y2 ; : : : ; y.j / (5.45e)
.j / @F ˇˇ
yi .t/ WD ˇ D i-th column of Y .j / , (5.45f)
@pDi .j /
In the present case the matrix M .j /.t/ is regular. Thus, the above system can be
represented also by the following initial value problem of 2nd order for D .t/:
R C .M .j /.t//1 D .j / .t/ @@'
.t/ P
q
.t; 0/ P
.t/
C.M .j / .t//1 K .j /.t/ @
q
@'
.t; 0/ .t/
.tj / D 0 (5.47b)
P j / D 0:
.t (5.47c)
for each i D 1; : : : ; .
Differentiation of the 2nd order system (5.37a–c) with respect to the initial values
@F @
q @F @
qP @F @
qR @' @
q
C C D
@q @
qj k @qP @
qj k @qR @
qj k @
q @
qj k
P
@' @
q
C ; k D 1; : : : ; n; (5.48a)
P @
qj k
@
q
@
q
.tj / D ek .D k-th unit vector/ (5.48b)
@
qj k
P
@
q
.tj / D 0: (5.48c)
@
qj k
0 1
pD
Inserting again
a D @
qj A D 0 in (5.48a–c), for
P j
q
@
q
.t/ WD .t; 0/; t tj ; (5.49)
@
qj k
with a fixed index k; 1 k n, cf. (5.43), we get the following linear initial value
problem of 2nd order:
@' @' P
K .t/
.j /
.t; 0/ .t/ C D .t/
.j /
.t; 0/ .t/
@
q @
qP
R D 0;
CM .j /.t/.t/ t tj (5.50a)
.tj / D ek (5.50b)
P j / D 0:
.t (5.50c)
@
q
.t/ WD .t; 0/; t tj ; (5.51)
P jk
@
q
222 5 Optimal Design of Regulators
P j k ; k D 1; : : : ; n, at
of
q with respect to the deviation of the initial velocities
q
stage j we have the linear 2nd order initial value problem
@' @' P
K .t/
.j /
.t; 0/ .t/ C D .j / .t/ .t; 0/ .t/
@
q P
@
q
R D 0;
CM .j / .t/.t/ t tj (5.52a)
.tj / D 0 (5.52b)
P j / D ek ;
.t (5.52c)
.tj / D ek (5.53b)
P j / D 0;
.t (5.53c)
R C .M .j /.t//1 D .j / .t/ @' .t; 0/ .t/
.t/ P
P
@
q
C.M .j /.t//1 K .j /.t/ @
q
@'
.t; 0/ .t/ D 0; t tj (5.54a)
.tj / D 0 (5.54b)
P j / D ek
.t (5.54c)
@'
For the functional matrices @
q ; @@'
P to be determined, the approach (5.55a–c)
q
yields the following representation:
@'
.t; 0/ D K .j /.t/ M .j /.t/Kp ; (5.56a)
@
q
@'
.t; 0/ D D .j / .t/ M .j /.t/Kd : (5.56b)
P
@
q
@'
Using only the linear terms @
q
.t; 0/; @@'
P .t; 0/ of the Taylor expansion (5.27a, b)
q
of the feedback function ' D '.t;
q;
q/, P and using the definition (5.55a–c), then
the optimal regulator problem (5.22a–c) is reduced to the optimal selection of the
regulator parameters in the matrices Kp ; Kd .
@'
Defining the Jacobians @
q .t; 0/; @@'
P .t; 0/ by (5.56a, b), the above described 2nd
q
order linear initial value problems read as follows:
@
q
With an index i D 1 : : : , for .t/ D @
q Di
.t; 0/; t tj , according to (5.47a–c)
we have
@
q
For .t/ D @
qj k .t; 0/; t tj , with an index k D 1; : : : ; n we get, cf. (5.53a–c),
R C Kd .t/
.t/ P C Kp .t/ D 0; t tj (5.58a)
.tj / D ek (5.58b)
P j / D 0:
.t (5.58c)
@
q
Finally, for .t/ WD P j k .t; 0/;
@
q
t tj , with an index k D 1; : : : ; n we have,
see (5.54a–c),
R C Kd .t/
.t/ P C Kp .t/ D 0; t tj (5.59a)
.tj / D 0 (5.59b)
P j / D ek :
.t (5.59c)
224 5 Optimal Design of Regulators
and
8
< 0; for (5.57a–c)
dqj WD ek ; for (5.58a–c) (5.60b)
:
0; for (5.59a–c);
8
< 0; for (5.57a–c)
dP qj WD 0; for (5.58a–c) (5.60c)
:
ek ; for (5.59a–c):
Then the 2nd order linear initial value problems (5.57a–c), (5.58a–c), (5.59a–c),
resp., for the computation of the sensitivities
8 9
< .t/; for (5.57a–c) =
dq.t/ WD .t/; for (5.58a–c) ; t tj (5.60d)
: ;
.t/; for (5.59a–c)
dP q.tj / D dP qj : (5.61c)
Defining now
dq.t/
dz.t/ WD ; t tj ; (5.62)
dP q.t/
the 2nd order linear initial value problem (5.61a–c) can be transformed into an
equivalent linear 1st order initial value problem, see (1.2a, b),(1.3):
0
dP z.t/ D A dz.t/ C .j / ; t tj (5.63a)
.t/
dz.tj / D dzj : (5.63b)
5.5 Computation of the Derivatives of the Tracking Error 225
Here, the matrix A and the vector dzj of initial values are defined as follows:
0 I
A WD (5.63c)
Kp Kd
dqj
dzj WD P : (5.63d)
d qj
For the associated homogeneous linear 1st order system of differential equations
dz.t/ ! 0; t ! C1 (5.65)
@
q
lim .t; 0/ D 0: (5.66c)
P j
t !C1 @
q
@
q @
q
Proof Since the systems of differential equations for @
qj
.t; 0/ and P j
@
q
.t; 0/ are
homogeneous, the assertion follows immediately from Lemma 5.3.
226 5 Optimal Design of Regulators
Zt
A.t tj / 0
dz.t/ D e dzj C e A.t / .j / d ; t tj : (5.67)
./
tj
Zt !
@
q 0
.t; 0/ D e A.t / d ; t tj ; (5.68a)
.M .j /.//1 yi ./
.j /
@
pDi
tj
@
q ek
.t; 0/ D e A.t tj / ; t tj ; (5.68b)
@
qj k 0
@
q 0
.t; 0/ D e A.t tj / ; t tj : (5.68c)
P jk
@
q ek
An explicit solution of (5.61a–c) and therefore also of the 2nd order linear systems
of differential equations (5.57a–c),(5.58a–c) and (5.59a–c) as well as an explicit
solution of the 1st order system (5.63a–d) can be obtained under the assumption
that Kp and Kd are diagonal matrices:
dP qk .tj / D dP qj k (5.70c)
As shown in the following, the roots k1;2 ; k D 1; : : : ; n, are just the eigenvalues
of the system matrix A, provided that Kd ; Kp are defined by (5.69):
u
If is an eigenvalue of A and w D an associated eigenvector, then (5.63c)
v
u 0 I u v
D D : (5.73)
v Kp Kd v Kp u Kd v
and therefore
w¤0 , u ¤ 0: (5.75b)
with degree 2n. If the matrices Kp and Kd have the diagonal form (5.69), then also
2 I C Kd C Kp is a diagonal matrix. Thus, we have
Y
n
q./ D .2 C dk C pk /: (5.76c)
kD1
Case 1:
2
dk > 4pk (5.78a)
we have
q
1
k1 D dk C dk
2
4pk < 0; (5.78b)
2
q
1
k2 D dk dk
2
4pk < 0: (5.78c)
2
In this case
dk
k1;2 D (5.79b)
2
is (at least) a double negative eigenvalue of A.
Case 3:
2
dk < 4pk (5.80a)
5.5 Computation of the Derivatives of the Tracking Error 229
Here, we have
q
1
k1;2 D dk ˙ .1/.4pk dk2
/
2
q
1
D dk ˙ i 4pk dk
2
2
1
D .dk ˙ i !k / (5.80b)
2
with
q
!k WD 4pk dk
2
: (5.80c)
In the present case k1;2 are two different complex eigenvalues of A with negative
real part 12 dk .
Consequently, we have this result:
Lemma 5.4 If Kp ; Kd are diagonal matrices having positive diagonal elements
pk ; dk ; k D 1; : : : ; n;, then all eigenvalues k1;2 ; k D 1; : : : ; n; of A have negative
real parts.
According to the Routh–Hurwitz criterion, under the assumption (5.77) of
positive diagonal elements, system of differential equations (5.64a, b) is asymptotic
stable.
With respect to the eigenvalues k1;2 described in Cases 1–3 we have the follow-
ing contributions to the fundamental system of the differential equation (5.70a, b):
Case 1:
Consequently, the solution of system (5.61a–c), (5.63a, b), resp., can be given
explicitly—eventually up to certain integrals occurring in the computation of a
particular solution of the inhomogeneous linear system of differential equations.
For the optimal regulator problem under stochastic uncertainty (5.22a–c), we have
to determine approximatively the conditional expectation
.j /
t
ˇ Zf ˇ
.j / ˇ ˇ
E J ˇAtj D E f .j / .t/ˇAtj dt
tj
.j /
Ztf ˇ
ˇ
E fQ.j / .t/ˇAtj dt; (5.82a)
tj
@'
.t; 0/ D K .j /.t/ M .j /.t/Kp (5.85a)
@
q
@'
.t; 0/ D D .j / .t/ M .j /.t/Kd : (5.85b)
P
@
q
@
q
!
@
z @
pD
.t; 0/
.t; 0/ D P
@
q
@
pD @
pD .t; 0/
@
q @
q @
q !
@
pD1
.t; 0/ @
pD2
.t; 0/ ::: @
pD
.t; 0/
D P
@
q P
@
q P
@
q ; (5.86a)
@
pD1
.t; 0/ @
pD2
.t; 0/ ::: @
pD
.t; 0/
@
q !
@
z @
qj
.t; 0/
.t; 0/ D P
@
q
@
qj @
qj
.t; 0/
@
q @
q @
q !
@
qj1
.t; 0/ @
qj2
.t; 0/ : : : @
qjn
.t; 0/
D P
@
q P
@
q P
@
q ; (5.86b)
@
qj1
.t; 0/ @
qj2
.t; 0/ : : : @
qjn
.t; 0/
0 1
@
q
@
z P j .t; 0/
.t; 0/ D @ A
@
q
P
P j
@
q @
q
.t; 0/
P j
@
q
0 @
q @
q @
q
1
P j .t; 0/ P j .t; 0/ : : : P jn .t; 0/
D@ A:
@
q 1 @
q 2 @
q
P
@
q P
@
q P
@
q (5.86c)
P j
@
q
.t; 0/ P j
@
q
.t; 0/ : : : P jn
@
q
.t; 0/
1 2
Using (5.85a, b), according to (5.42a–c) and (5.45a–d), the functional matrix
@
q
t! .t; 0/; t tj ;
@
pD
232 5 Optimal Design of Regulators
@
qR P
@
q @
q
.t; 0/ C Kd .t; 0/ C Kp .t; 0/
@
pD @
pD @
pD
1 .j /
D M .j /.t/ Y .t/; t tj (5.87a)
@
q
.tj ; 0/ D 0 (5.87b)
@
pD
P
@
q
.tj ; 0/ D 0: (5.87c)
@
pD
Moreover, using (5.85a, b), by (5.48a–c), (5.49) and (5.50a–c) the functional
matrix
@
q
t! .t; 0/; t tj ;
@
qj
satisfies the
R
@
q P
@
q @
q
.t; 0/ C Kd .t; 0/ C Kp .t; 0/ D 0; t tj (5.88a)
@
qj @
qj @
qj
@
q
.tj ; 0/ D I (5.88b)
@
qj
P
@
q
.tj ; 0/ D 0: (5.88c)
@
qj
Finally, using (5.85a, b), according to (5.51) and (5.52a–c) the functional matrix
@
q
t! .t; 0/; t tj ;
P j
@
q
R
@
q P
@
q @
q
.t; 0/ C Kd .t; 0/ C Kp .t; 0/ D 0; t tj (5.89a)
P
@
qj P
@
qj P j
@
q
@
q
.t ; 0/ D 0 (5.89b)
P j j
@
q
P
@
q
.t ; 0/ D I: (5.89c)
P j j
@
q
5.7 Optimal PID-Regulator 233
For simplification of the notation, in the following we consider only the 0-th stage
j WD 0. Moreover, we suppose again that the state z.t/ D .q.t/; q.t// P is available
exactly. Corresponding to Sect. 5.1, the tracking error, i.e., the deviation of the actual
position q D q.t/ and velocity qP D q.t//
P from the reference position q .0/ D q .0/ .t/
and reference velocity qP D qP .t//; t0 t tf is defined by
.0/ .0/
q.t/ WD q.t/ q R .t/; t0 t tf ; (5.90a)
P WD q.t/
q.t/ P qP R .t/; t0 t tf : (5.90b)
Hence, for the difference between the actual state z D z.t/ and the reference state
z.0/ D z.0/ .t/ we have
q.t/ q .0/ .t/
z.t/ WD z.t/ z.0/ .t/ D ; t0 t tf : (5.91)
P qP .0/ .t/
q.t/
Here u.0/ D u.0/ .t/; t0 t tf , denotes the feedforward control based on the
reference trajectory z.0/ .t/ D z.0/ .t/; t0 t tf , obtained e.g. by inverse dynamics,
cf. [4, 11, 61], and
u D
u.t/; t0 t tf ; is the optimal control correction to be
determined.
Generalizing the PD-regulator, depending on the state z.t/ D .q.t/; q.t//,P for the
PID-regulator, cf. Sect. 4.7, also the integrated position
Zt
qI .t/ WD q./ d ; t0 t tf ; (5.93)
t0
is taken into account. Hence, the PID-regulator depends also on the integrative
position error
Zt Zt
qI .t/ WD
q./ d D q./ q R ./ d
t0 t0
Zt Zt
D q./ d q R ./ d D qI .t/ qIR .t/: (5.94)
t0 t0
234 5 Optimal Design of Regulators
u.t/ WD '.t;
q.t/;
qI .t/;
q.t//;
P t0 t tf : (5.95)
Without further global information on the stochastic parameters also in the present
case for the feedback law ' we suppose
'.t; 0; 0; 0/ D 0; t0 t tf : (5.96)
Ztf ˇ !
ˇˇ
min E P / C t; '.t;
q.t /;
qI .t /;
q.t
c t;
q.t /;
qI .t /;
q.t P // dtˇAt0
ˇ
t0
(5.97a)
subject to
P /; q.t
F .pD .!/; q.t /; q.t R //
D uR .t / C '.t;
q.t /;
qI .t /;
q.t
P //; t0 t tf ; a:s: (5.97b)
q.t0 ; !/ D q0 .!/; (5.97c)
P 0 ; !/ D qP 0 .!/;
q.t (5.97d)
'.t; 0/ D 0: (5.97e)
Here,
P
F .pD .!/; q.t/; q.t/; R
q.t// WD M pD ; q.t/ q.t/
R C h pD ; q.t/; q.t/
P (5.98)
pD .!/ WD pD .!/ pD (5.99)
E pD .!/jAt0 denotes the (conditional) expectation of the random vector pD D
pD .!/.
Having the stochastic optimal feedback control ' D ' .t;
q;
qI ;
q/, P the
effective trajectory q D q.t/; t t0 , of the stochastic optimally controlled dynamic
system is then given by the solution of the initial value problem (5.97b–d). Using
the definitions, cf. (4.45c, d), (5.3d, e),
q0 WD q.t0 / q R .t0 / D q0 E q0 jAt0 ; (5.100a)
qP0 WD q.t
P 0 / qP R .t0 / D qP 0 E qP0 jAt0 ; (5.100b)
F .pD C
pD .!/; q R .t/ C
q.t/; qP R .t/ C
q.t/;
P qR R .t/ C
q.t//
R
D uR .t/ C '.t;
q.t/;
qI .t/;
q.t//;
P t0 t tf ; a:s; (5.101a)
q.t0 ; !/ D
q0 .!/; (5.101b)
P 0 ; !/ D
qP0 .!/;
q.t (5.101c)
'.t; 0/ D 0: (5.101d)
P
Remark 5.2 For given control law ' D '.t;
q.t/;
qI .t/;
q.t//, the solution of
the initial value problem (5.101a–d) yields position error function
q WD
q.t; a.!//; t0 t tf ; (5.102a)
Corresponding to Sect. 5.3.2, here we use again quadratic cost functions. Hence,
with symmetric, positive (semi-)definite matrices Cq ; CqI ; CqP the costs c arising
from the tracking error are defined by
P
c t;
q.t/;
qI .t/;
q.t/
WD
q.t/T Cq
q.t/ C
qI .t/T CqI
qI .t/ C
q.t/
P T CqP
q.t/:
P (5.103a)
236 5 Optimal Design of Regulators
Moreover, with a positive (semi-)definite matrix Cu the regulator costs are defined
by
P
t; '.t;
q.t/;
qI .t/;
q.t/
T
P
WD ' t;
q.t/;
qI .t/;
q.t/ P
Cu ' t;
q.t/;
qI .t/;
q.t/ : (5.103b)
f .t/ WD
q.t/T Cq
q.t/ C
qI .t/T CqI
qI .t/
C
q.t/
P T CqP
q.t/
P C
u.t/T Cu
u.t/: (5.104)
Thus, for the objective function of the regulator optimization problem (5.97a–e) we
obtain
Ztf ˇ ! Ztf
ˇ ˇ
ˇ ˇ
E f .t/ dtˇAt0 D E f .t/ˇAt0 dt
ˇ
t0 t0
Ztf
T T T
D E
q.t/ Cq
q.t/ C
qI .t/ CqI
qI .t/ C
q.t/
P P
CqP
q.t/
t0
T ˇˇ
P
C ' t;
q.t/;
qI .t/;
q.t/ P
Cu ' t;
q.t/;
qI .t/;
q.t/ ˇAt0 dt:
(5.105)
Corresponding to Sect. 5.4, here we use also the Taylor expansion method to
compute approximatively the arising conditional expectations. Thus, by Taylor
expansion of the feedback function ' at
q D 0;
qI D 0;
qP D 0, we get
P
' t;
q.t/;
qI .t/;
q.t/ D ' t; 0 C D
q '.t; 0/
q.t/
CD
qI '.t; 0/
qI .t/ C D
qP '.t; 0/
q.t/
P C : (5.106)
Here,
@'
D
q ' D D
q '.t; 0/ D .t; 0/
@
q
denotes the Jacobian, hence, the matrix of partial derivatives of ' with respect to
q at .t; 0/.
5.7 Optimal PID-Regulator 237
Taking into account only the linear terms in (5.106), because of '.t; 0/ D 0 we
get the approximation
ˇ ˇ
ˇ ˇ
E f ˇAt0 E fQˇAt0 ; (5.107)
where
E fQjA0 WD E
q.t/T Cq
q.t/ C
qI .t/T CqI
qI .t/ C
q.t/
P T CqP
q.t/
P
T
C D
q '.t; 0/
q.t/ C D
qI '.t; 0/
qI .t/ C D
qP '.t; 0/
q.t/
P
ˇ !
T ˇ
ˇ
Cu D
q '.t; 0/
q.t/ C D
qI '.t; 0/
qI .t/ C D
qP '.t; 0/
q.t/
P ˇAt0
ˇ
DE
q.t/T Cq
q.t/ C
qI .t/T CqI
qI .t/ C
q.t/
P T CqP
q.t/
P
C
q.t/T D
q '.t; 0/T Cu C
qI .t/T D
qI '.t; 0/T Cu
C
q.t/
P T D
qP '.t; 0/T Cu
ˇ !
ˇ
ˇ
D
q '.t; 0/
q.t/ C D
qI '.t; 0/
qI .t/ C D
qP '.t; 0/
q.t/
P ˇAt0
ˇ
DE
q.t/T Cq
q.t/ C
qI .t/T CqI
qI .t/ C
q.t/
P T CqP
q.t/
P
C
q.t/T D
q '.t; 0/T Cu D
q '.t; 0/
q.t/
C
q.t/T D
q '.t; 0/T Cu D
qI '.t; 0/
qI .t/
P
C
q.t/T D
q '.t; 0/T Cu D
qP '.t; 0/
q.t/
C
qI .t/T D
qI '.t; 0/T Cu D
q '.t; 0/
q.t/
C
qI .t/T D
qI '.t; 0/T Cu D
qI '.t; 0/
qI .t/
P
C
qI .t/T D
qI '.t; 0/T Cu D
qP '.t; 0/
q.t/
C
q.t/
P T D
qP '.t; 0/T Cu D
q '.t; 0/
q.t/
C
q.t/
P T D
qP '.t; 0/T Cu D
qI '.t; 0/
qI .t/
ˇ !
ˇ
ˇ
C
q.t/ P ˇAt0
P T D
qP '.t; 0/T Cu D
qP '.t; 0/
q.t/
ˇ
238 5 Optimal Design of Regulators
DE
q.t/T Cq C D
q '.t; 0/T Cu D
q '.t; 0/
q.t/
C
qI .t/T CqI C D
qI '.t; 0/T Cu D
qI '.t; 0/
qI .t/
C
q.t/
P T CqP C D
qP '.t; 0/T Cu D
qP '.t; 0/
q.t/
P
C 2
q.t/T D
q '.t; 0/T Cu D
qI '.t; 0/
qI .t/
C 2
q.t/T D
q '.t; 0/T Cu D
qP '.t; 0/
q.t/
P
ˇ !
ˇ
ˇ
P ˇAt0 :
C 2
qI .t/T D
qI '.t; 0/T Cu D
qP '.t; 0/
q.t/ (5.108)
ˇ
Qq WDCq C D
q '.t; 0/T Cu D
q '.t; 0/; (5.109a)
QqI WDCqI C D
qI '.t; 0/T Cu D
qI '.t; 0/; (5.109b)
QqP WDCqP C D
qP '.t; 0/ Cu D
qP '.t; 0/;
T
(5.109c)
QqqI WDD
q '.t; 0/ Cu D
qI '.t; 0/;
T
(5.109d)
Qq qP WDD
q '.t; 0/T Cu D
qP '.t; 0/; (5.109e)
QqI qP WDD
qI '.t; 0/T Cu D
qP '.t; 0/: (5.109f)
C 2
q.t/T D
q '.t; 0/T Cu D
qI '.t; 0/
qI .t/
P
C 2
q.t/T D
q '.t; 0/T Cu D
qP '.t; 0/
q.t/
C 2
qI .t/T D
qI '.t; 0/T Cu D
qP '.t; 0/
q.t/
P
0 10 1
Qq QqqI Qq qP
q.t/
B T Q Q C@
P T / @Qqq
D.
q.t/T ;
qI .t/T ;
q.t/ I qI qI qP A
qI .t/A
T T
Qq qP QqI qP QqP P
q.t/
D
zI .t/T Q.t; '/
zI .t/; (5.110)
5.7 Optimal PID-Regulator 239
a D .
pD .!/T ;
q0 .!/T ;
qP0 .!/T /T D 0: (5.112)
Hence, we get
q.t/ D
q.t; 0/ C D
pD
q.t; 0/
pD .!/ C D
q0
q.t; 0/
q0 .!/
C D
qP0
q.t; 0/
qP0 .!/ C : : : ; (5.113a)
P D
q.t;
q.t/ P 0/ C D
pD
q.t;
P 0/
pD .!/
C D
q0
q.t;
P 0/
q0 .!/ C D
qP0
q.t;
P 0/
qP0 .!/ C : : : ; (5.113b)
as well as
Zt Zt
qI .t/ D
q./d D
q.; 0/ C D
pD
q.; 0/
pD .!/
t0 t0
CD
q0
q.; 0/
q0 .!/ C D
qP0
q.; 0/
qP0 .!/ C : : : d : (5.113c)
Using only the linear terms, and taking into account the conditions (5.101a–d)
q.t; 0/ D 0; t t0 ; (5.114a)
P 0/ D 0; t t0 ;
q.t; (5.114b)
q.t/ D
pD
q.t; 0/
pD .!/ C D
q0
q.t; 0/
q0 .!/ C D
qP0
q.t; 0/
qP0 .!/;
(5.115a)
240 5 Optimal Design of Regulators
P D
pD
q.t;
q.t/ P 0/
pD .!/ C D
q0
q.t;
P 0/
q0 .!/ C D
qP0
q.t;
P 0/
qP0 .!/:
(5.115b)
Zt
qI .t/ D
q./ d
t0
Zt
D
pD
q.; 0/
pD .!/
t0
C D
q0
q.; 0/
q0 .!/ C D
qP0
q.; 0/
qP0 .!/ d
Zt Zt
D D
pD
q.; 0/
pD .!/ d C D
q0
q.; 0/
q0 .!/ d
t0 t0
Zt
C D
qP0
q.; 0/
qP0 .!/ d
t0
Zt !ˇ Zt !ˇ
ˇ ˇ
@ ˇ @ ˇ
D
q.; p/ d ˇ
pD .!/ C
q.; p/ d ˇ
q0 .!/
@
pD ˇ @
q0 ˇ
t0 pD0 t0 pD0
„ ƒ‚ … „ ƒ‚ …
D
qI .t;p/ D
qI .t;p/
Zt !ˇ
ˇ
@ ˇ
C
q.; p/ d ˇ
qP0 .!/
@
qP0 ˇ
t0 pD0
„ ƒ‚ …
D
qI .t;p/
DD
pD
qI .t; 0/
pD .!/ C D
q0
qI .t; 0/
q0 .!/ C D
qP0
qI .t; 0/
qP0 .!/:
(5.115c)
Consequently, we find
0 1
q.t/
zI .t/ D @
qI .t/A
P
q.t/
0 1
D
pD
q.t; 0/
pD .!/ C D
q0
q.t; 0/
q0 .!/ C D
Pq0
q.t; 0/
qP 0 .!/
@D
pD
qI .t; 0/
pD .!/ C D
q0
qI .t; 0/
q0 .!/ C D
Pq0
qI .t; 0/
qP 0 .!/A
P 0/
pD .!/ C D
q0
q.t;
D
pD
q.t; P 0/
q0 .!/ C D
Pq0
q.t;P 0/
qP 0 .!/
5.7 Optimal PID-Regulator 241
0 1 0 1
D
pD
q.t; 0/ D
q0
q.t; 0/
D @D
pD
qI .t; 0/A
pD .!/ C @D
q0
qI .t; 0/A
q0 .!/
P 0/
D
pD
q.t; P 0/
D
q0
q.t;
0 1
D
Pq0
q.t; 0/
C @D
Pq0
qI .t; 0/A
qP 0 .!/
P 0/
D
Pq0
q.t;
DD
pD
zI .t; 0/
pD .!/ C D
q0
zI .t; 0/
q0 .!/ C D
Pq0
zI .t; 0/
qP 0 .!/:
(5.116)
fQ D
zI .t/T Q
zI .t/
T T
pD .!/T D
pD
zI .t; 0/ C
q0 .!/T D
q0
zI .t; 0/
T
C
qP0 .!/ D
qP0
zI .t; 0/
T
Q D
pD
zI .t; 0/
pD .!/ C D
q0
zI .t; 0/
q0 .!/
C D
qP0
zI .t; 0/
qP0 .!/
T
D
pD .!/T D
pD
zI .t; 0/ Q D
pD
zI .t; 0/
pD .!/
T
C
q0 .!/T D
q0
zI .t; 0/ Q D
q0
zI .t; 0/
q0 .!/
T
C
qP0 .!/T D
qP0
zI .t; 0/ Q D
qP0
zI .t; 0/
qP0 .!/
T
C 2
q0 .!/T D
qP0
zI .t; 0/ Q D
qP0
zI .t; 0/
pD .!/
T
C 2
qP0 .!/T D
qP0
zI .t; 0/ Q D
q0
zI .t; 0/
q0 .!/
T
C 2
pD .!/T D
qP0
zI .t; 0/ Q D
qP0
zI .t; 0/
qP0 .!/: (5.117)
q0 .!/;
pD .!/ as well as
pD .!/;
qP0 .!/ are stochastically independent.
Since
E
pD .!/jAt0 D
pD D 0; (5.118)
242 5 Optimal Design of Regulators
where
ˇ ˇ ˇ
ˇ ˇ ˇ
E fOˇAt0 WD E
pD .!/T P
pD
pD .!/ˇAt0 C E
q0 .!/T P
q0
q0 .!/ˇAt0
ˇ ˇ
ˇ ˇ
C E
qP0 .!/T P
qP0
qP0 .!/ˇAt0 C E
qP0 .!/T P
qP0 ;
q0
q0 .!/ˇAt0 :
(5.121b)
Thus, we find
ˇ X
ˇ
ˇ ˇ
E
pD .!/ P
pD
pD .!/ˇAt0 D E
T
ŒP
pD kl
pDk .!/
pDl .!/ˇAt0
k;lD1
X
ˇ
ˇ
D ŒP
pD kl E
pDk .!/
pDl .!/ˇAt0
k;lD1
X
ˇˇ
D ŒP
pD kl E
pDk .!/
pDk
pDl .!/
pDl ˇAt0
„ƒ‚… „ƒ‚…
k;lD1
D0 D0
5.7 Optimal PID-Regulator 243
X
ˇˇ
D ŒP
pD kl E pDk .!/ pDk pDl .!/ pDl ˇAt0
k;lD1
X
D ŒP
pD kl cov pDk ; pDl : (5.122a)
k;lD1
ˇ X
ˇ
E
pD .!/ P
pD
pD .!/ˇAt0 D
T
ŒP
pD kl cov pDk ; pDl
k;lD1
X
D ŒP
pD kl cov pDl ; pDk
k;lD1
X
X
D ŒP
pD kl cov pDl ; pDk
kD1 lD1
X
D k-th row of P
pD k-th column of cov pD ./
kD1
D tr P
pD cov pD ./ : (5.123a)
ˇ X
ˇ
ˇ ˇ
E
qP0 .!/T P
qP0 ;
q0
q0 .!/ˇAt0 D E ŒP
qP0 ;
q0 kl
qP0k .!/
q0l .!/ˇAt0
k;lD1
X
D ŒP
qP0 ;
q0 kl cov
qP0k ./;
q0l ./
k;lD1
X
D ŒP
qP0 ;
q0 kl cov
q0l ./;
qP0k ./
k;lD1
X
X
D ŒP
qP0 ;
q0 kl cov
q0l ./;
qP0k ./
kD1 lD1
X
D k-th row of P
qP0 ;
q0
kD1
k-th column of cov
q0 ./;
qP0 ./
D tr P
qP0 ;
q0 cov
q0 ./;
qP0 ./ : (5.125)
Inserting now (5.123a–c) and (5.125) into (5.121b), for the objective function
(5.97a) we find the approximation
ˇ
ˇ
E fQˇAt0 tr P
pD cov pD ./ C tr P
q0 cov
q0 ./
C tr P
qP0 cov
qP0 ./
C tr P
qP0 ;
q0 cov
q0 ./;
qP0 ./ : (5.126)
Ztf
min tr P
pD cov pD ./ C tr P
q0 cov
q0 ./ C tr P
qP0 cov
qP0 ./
t0
C tr P
qP0 ;
q0 cov
q0 ./;
qP0 ./ dtF .p D C
pD .!/; q R .t/ C
q.t/; qP R .t/
C
q.t/;
P qR R .t/ C
q.t//
R
P
D F .pD ; q.t/; q.t/; R
q.t// D uR .t/ C '.t;
q.t/;
qI .t/;
q.t//;
P a:s:
(5.128a)
q.t0 ; !/ D
q0 .!/; (5.128b)
P 0 ; !/ D
qP0 .!/;
q.t (5.128c)
'.t0 ; 0/ D 0: (5.128d)
F D F .p D C
pD ; q R .t/ C
q.t/; qP R .t/ C
q.t/;
P qR R .t/ C
q.t//:
R
@F @
pD @F @
q.t; a/ P a/
@F @
q.t; R a/
@F @
q.t; @ R
C C C D u .t/
@pD @
pD @q @
pD @qP @
pD @qR @
pD @pD
„ ƒ‚ …
D0
@' @
q.t; a/ @' @
qI .t; a/
C .t;
zI / C .t;
zI /
@
q @
pD @
qI @
pD
@' P a/
@
q.t;
C .t;
zI / : (5.130a)
@
qP @
pD
Moreover, by differentiation of Eqs. (5.128b, c) for the initial values with respect
to
pD we still find
@
q
.t0 ; a/ D 0; (5.130b)
@
pD
@
qP
.t0 ; a/ D 0: (5.130c)
@
pD
Here, we have
Zt Zt
@
qI @ @
q
.t; a/ D
q.; a/ d D .; a/ d : (5.131)
@
pD @
pD @
pD
t0 t0
@
q.t; 0/ P 0/
@
q.t; R 0/
@
q.t;
Y R .t/ C K R .t/ C D R .t/ CM R .t/
@
pD @
pD @
pD
@' @
q.t; 0/ @' @
qI .t; 0/
D .t; 0/ C .t; 0/
@
q @
pD @
qI @
pD
@' P 0/
@
q.t;
C .t; 0/ ; (5.132a)
@
qP @
pD
@
q
.t0 ; 0/ D 0; (5.132b)
@
pD
@
qP
.t0 ; 0/ D 0: (5.132c)
@
pD
@
q
.t/ WD .t; 0/; t0 t tf ; (5.133)
@
pD
5.7 Optimal PID-Regulator 247
Since the matrix M D M.t/ is regular, see e.g. [11], Eq. (5.134a) can be
transformed into
@' @'
R C M R .t/1 D R .t/
.t/ P C M R .t/1 K R .t/
.t; 0/ .t/ .t; 0/ .t/
@
qP @
q
Zt
1
@'
CM .t/
R
.t; 0/ .s/ ds D M R .t/1 Y .t/: (5.135)
@
qI
t0
Hence, for the unknown Jacobians of ' we use the following representation:
Definition 5.4 With given fixed, positive definite matrices Kp ; Ki ; Kd , define
@'
Kd DWM R .t/1 D R .t/ .t; 0/ ; (5.136a)
@
qP
@'
Kp DWM R .t/1 K R .t/ .t; 0/ ; (5.136b)
@
q
@'
Ki DWM R .t/1 .t; 0/ : (5.136c)
@
qI
Zt
R C Kd .t/
.t/ P C Kp .t/ C Ki ./ d D M R .t/1 Y R .t/; t t0 ;
t0
(5.137a)
.t0 / D 0; (5.137b)
P 0 / D 0;
.t (5.137c)
248 5 Optimal Design of Regulators
Following the above procedure, we can also (i) determine the partial derivative of
@
pD @
q.t; 0 / P 0/
@
q.t; R 0/
@
q.t;
Y R .t/ C K R .t/ C D R .t/ C M R .t/
@
q0 @
q0 @
q0 @
q0
@' @
q.t; 0/ @' @
qI .t; 0/
D .t; 0/ C .t; 0/
@
q @
q0 @
qI @
q0
@' P 0/
@
q.t;
C .t; 0/ (5.138a)
@
qP @
q0
@
q
.t0 ; 0/ D I; (5.138b)
@
q0
@
qP
.t0 ; 0/ D 0; (5.138c)
@
q0
@
q
.t/ WD .t; 0/; t t0 ; (5.139)
@
q0
@
pD
and having @
q0
D 0, the differential equations (5.138a) reads
R C M.t/1 D.t/ @' .t; 0/ .t/
.t/ P C M.t/1 K.t/ @' .t; 0/ .t/
@
qP @
q
Zt
1
@'
CM.t/ .t; 0/ ./ d D 0: (5.140)
@
qI
t0
Using again definitions (5.136a–c), for D .t/ we get the initial value problem:
Zt
R C Kd .t/
.t/ P C Kp .t/ C Ki ./ d D 0; t t0 ; (5.141a)
t0
.t0 / D I; (5.141b)
P 0 / D 0:
.t (5.141c)
5.7 Optimal PID-Regulator 249
@
pD @
q.t; 0/ P 0/
@
q.t; R 0/
@
q.t;
Y .t/ C K.t/ C D.t/ C M.t/
@
qP0 @
qP0 @
qP0 @
qP0
@' @
q.t; 0/ @' @
qI .t; 0/
D .t; 0/ C .t; 0/
@
q @
qP0 @
qI @
qP0
@' P 0/
@
q.t;
C .t; 0/ ; (5.142a)
@
qP @
qP0
@
q
.t0 ; 0/ D 0; (5.142b)
@
qP0
@
qP
.t0 ; 0/ D I: (5.142c)
@
qP0
@
pD
Also here I denotes the unit matrix and it holds @
qP0 D 0.
Defining here
@
q
.t/ WD .t; 0/; t t0 ; (5.143)
@
qP0
R C M R .t/1 D R .t/ @' .t; 0/ .t/
.t/ P C M R .t/1 K R .t/ @' .t; 0/ .t/
@
qP @
q
Zt
@'
CM R .t/1 .t; 0/ ./ d D 0: (5.144)
@
qI
t0
Taking into account (5.136a–c), one finally gets the initial value problem
Zt
R C Kd .t/
.t/ P C Kp .t/ C Ki ./ d D 0; t t0 ; (5.145a)
t0
.t0 / D 0; (5.145b)
P 0 / D I:
.t (5.145c)
250 5 Optimal Design of Regulators
Ztf
min tr P
pD cov pD ./ C tr P
q0 cov
q0 ./ C tr P
qP0 cov
qP0 ./
t0
!
C tr P
qP0 ;
q0 cov
q0 ./;
qP0 ./ dt (5.146a)
Zt
R C Kd .t/
.t/ P C Kp .t/ C Ki ./ d D M.t/1 Y .t/; t t0 ; (5.146b)
t0
.t0 / D 0; (5.146c)
P 0 / D 0;
.t (5.146d)
Zt
R C Kd .t/
.t/ P C Kp .t/ C Ki ./ d D 0; t t0 ; (5.146e)
t0
.t0 / D I; (5.146f)
P 0 / D 0:
.t (5.146g)
Zt
R C Kd .t/
.t/ P C Kp .t/ C Ki ./ d D 0; t t0 ; (5.146h)
t0
.t0 / D 0; (5.146i)
P 0 / D I:
.t (5.146j)
Introducing appropriate vector and vector functions, the above three systems of
differential equations (5.146b–j) can be represented uniformly. Hence, we set:
8 9
< M R .t/1 Y R .t/; t t0 ; for (5.146b–d) =
.t/ WD 0; for (5.146e–g) ; t t0 ; (5.147a)
: ;
0; for (5.146h–j)
5.7 Optimal PID-Regulator 251
and
8
< 0; for (5.146b–d)
dq0 WD I; for (5.146e–g) (5.147b)
:
0; for (5.146h–j);
8
< 0; for (5.146b–d)
d qP0 WD 0; for (5.146e–g) (5.147c)
:
I; for (5.146h–j):
Zt
R C Kd d q.t/
d q.t/ P C Kp dq.t/ C Ki dq./ d D .t/; t t0 (5.148a)
t0
Defining
0 1
dq.t/
B Rt C
dzI .t/ WD B C
@ dq./ d A ; t t0 ; (5.149)
t0
P
d q.t/
Here, the matrix A and the vector of initial conditions dzI0 is defined by:
0 1
0 0 I
A WD @ I 0 0 A; (5.150c)
Kp Ki Kd
0 1
dq0
dzI0 WD @ 0 A : (5.150d)
d qP0
Chapter 6
Expected Total Cost Minimum Design of Plane
Frames
6.1 Introduction
Yield stresses, allowable stresses, moment capacities (plastic moments with respect
to compression, tension and rotation), applied loadings, cost factors, manufacturing
errors, etc., are not given fixed quantities in structural analysis and optimal design
problems, but must be modeled as random variables with a certain joint probability
distribution. Problems from plastic analysis and optimal plastic design are based on
the convex yield (feasibility) criterion and the linear equilibrium equation for the
stress (state) vector.
After the formulation of the basic mechanical conditions including the relevant
material strength parameters and load components as well as the involved design
variables (as e.g. sizing variables) for plane frames, several approximations are
considered: (i) approximation of the convex yield (feasible) domain by means of
convex polyhedrons (piecewise linearization of the yield domain); (ii) treatment
of symmetric and non symmetric yield stresses with respect to compression
and tension; (iii) approximation of the yield condition by using given reference
capacities.
As a result, for the survival of a plane frame a certain system of necessary and/or
sufficient linear equalities and inequalities is obtained. Evaluating the recourse
costs, i.e., the costs for violations of the survival condition by means of piecewise
linear convex cost functions, a linear program is obtained for the minimization of
the total costs including weight-, volume- or more general initial (construction)
costs. Appropriate cost factors are derived. Considering then the expected total
costs from construction as well as from possible structural damages or failures, a
stochastic linear optimization problem (SLP) is obtained. Finally, discretization of
the probability distribution of the random parameter vector yields a (large scale)
linear program (LP) having a special structure. For LP’s of this type numerous, very
efficient LP-solvers are available—also for the solution of very large scale problems.
Assuming that the material behaves as an elastic-perfectly plastic material [59, 79,
119] a conservative estimate of the collapse load factor T is based [30, 31, 36, 39,
54, 55, 77, 120, 164] on the following linear program:
maximize (6.1a)
s.t.
FL F FU (6.1b)
CF D R0 : (6.1c)
where Aj is the (given) cross-sectional area, and yjL ; yjU , respectively, denotes
the yield stress in compression (negative values) and tension (positive values)
of the j -th member of the truss. In case of a plane frame, F is composed of
subvectors [158],
0 .k/
1 0 1
F1 tk
B C
F .k/ D @ F2.k/ A D @ mC k
A; (6.3a)
.k/
F3 mk
Mkpl D yk
U
Wkpl ; (6.3c)
and Wkpl D Wkpl .Ak / is the plastic section modulus of the cross-section of the k-th
member (beam) with respect to the local z-axis.
For a spatial frame [83, 158], corresponding to the k-th member (beam), F
contains the subvector
C
where tk is the normal (axial) force, mkT the twisting moment, and m k yN ; mkNz ,
mk yN ; mkNz denote four bending moments with respect to the local y; N z-axis at the
positive, negative end of the beam, respectively. Finally, the bounds F L ; F U for F
are given by
L pN yN zN yN z 0
F .k/L D yk Ak ; Mkpl ; Mkpl ; Mkpl ; Mkpl ; Mkpl (6.4b)
U pN yN zN yN
Nz 0
F .k/U D yk Ak ; Mkpl ; Mkpl ; Mkpl ; Mkpl ; Mkpl ; (6.4c)
are the plastic moments of the cross-section of the k-th element with respect to the
pN pN
N zN-axis, respectively. In (6.4d), Wkpl D Wkpl .x/ and
local twisting axis, the local y-,
yN yN zN Nz
Wkpl D Wkpl .X /; Wkpl D Wkpl .X /, respectively, denote the polar, axial modulus
of the cross-sectional area of the k-th beam and yk denotes the yield stress with
respect to torsion; we suppose that yk D ykU
.
Remark 6.1 Possible plastic hinges [59, 82, 119] are taken into account by inserting
appropriate eccentricities ekl > 0; ekr > 0; k D 1; : : : ; n0 , with ekl ; ekr
Lk , where
Lk is the length of the k-th beam.
Remark 6.2 Working with more general yield polygons, see e.g. [164], the stress
condition (6.1b) is replaced by the more general system of inequalities
Here, .H; h/ is a given .n C 1/ matrix, and FdU WD .FjU ıij / denotes the n n
diagonal matrix of principal axial and bending plastic capacities
j
FjU WD yk
U
j
Akj ; FjU WD yk
U
j
Wkj pl ; (6.5b)
where kj; j are indices as arising in (6.3b)–(6.4d). The more general case (6.5a)
can be treated by similar methods as the case (6.1b) which is considered here.
In the plastic design of trusses and frames [77, 87, 92, 97, 128, 152] having n0
members, the n-vectors F L ; F U of lower and upper bounds
y1 ; : : : ; yn 0
, and the r-vector
x D .x1 ; x2 ; : : : ; xr /T (6.7)
of design variables of the structure. In case of trusses we have that, cf. (6.2),
F L D yd
L
A.x/ D A.x/d yL ;
F U D yd
U
A.x/ D A.x/d yU ; (6.8)
where n D n0 , and yd L U
; yd denote the n n diagonal matrices having the diagonal
elements yj ; yj , respectively, j D 1; : : : ; n, moreover,
L U
h iT
A.x/ D A1 .x/; : : : ; An .x/ (6.9)
CF D Ru ; (6.10)
where Ru describes [77] the ultimate load [representing constant external loads or
self-weight expressed in linear terms of A.x/].
6.2 Stochastic Linear Programming Techniques 257
The plastic design of structures can be represented then, cf. [7], by the optimiza-
tion problem
s.t.
F L .yL ; yU ; x/ F F U .yL ; yU ; x/ (6.11b)
CF D Ru ; (6.11c)
where G D G.x/ is a certain objective function, e.g. the volume or weight of the
structure.
Remark 6.3 As mentioned in Remark 6.2, working with more general yield poly-
gons, (6.11b) is replaced by the condition
For the elastic design we must replace the yield stresses yL ; yU by the allowable
stresses aL ; aU and instead of ultimate loads we consider service loads Rs . Hence,
instead of (6.11a–d) we have the related program
s.t.
F L .aL ; aU ; x/ F F U .aL ; aU ; x/; (6.12b)
CF D Rs ; (6.12c)
x xx ;
L U
(6.12d)
Fi WD .ti ; m C C T
i / ; Fi WD .ti ; mi / ;
T
(6.13)
respectively.
258 6 Expected Total Cost Minimum Design of Plane Frames
Furthermore, for each bar/beam with rigid joints we have several plastic capac-
L
ities: The plastic capacity Nipl of the bar with respect to axial compression, hence,
the maximum axial force under compression is given by
L
Nipl D jyiL j Ai ; (6.14a)
where yiL < 0 denotes the (negative) yield stress with respect to compression and Ai
is the cross sectional area of the i th element. Correspondingly, the plastic capacity
with respect to (axial) tension reads:
U
Nipl D yiU Ai ; (6.14b)
where yiU > 0 is the yield stress with respect to tension. Besides the plastic
capacities with respect to the normal force, we have the moment capacity
!T !T
ti mC ti mC
FiLC WD ; i
L M
; FiU C WD ; i
U M
(6.15c,d)
Nipl ipl Nipl ipl
The limit between the elastic and plastic state of the elements is described by the
feasibility or yield condition: At the negative end we have the condition
M/Mpl
1 N/Npl
where x D N
Npl and y D M
Mpl , (see Fig. 6.1).
260 6 Expected Total Cost Minimum Design of Plane Frames
In case (6.18) and supposing symmetric yield stresses, the yield condition
(17a–d) reads
ˇ ˇ
2
ti ˇm ˇ
C ˇˇ i ˇˇ 1; (6.19a)
Nipl Mipl
ˇ ˇ
ti 2 ˇˇ mC i ˇ
ˇ
Cˇ ˇ 1: (6.19b)
Nipl ˇ Mipl ˇ
Remark 6.3 Because of the connection between the normal force ti and the bending
moments, (6.19a, b) are also called “M –N -interaction”.
If the M –N -interaction is not taken into account, K0;sym is approximated from
outside, see Fig. 6.1, by
u
K0;sym WD f.x; y/T W jxj; jyj 1g: (6.20)
Hence, (4.17a–d) are replaced, cf. (6.19a, b), by the simpler conditions
Since the symmetry condition (6.16a) does not hold in general, some modifica-
tions of the basic conditions (6.19a, b) are needed. In the non symmetric case K0;xm
must be replaced by the intersection
of two convex sets K0U and K0L . For a rectangular cross-sectional area we have
p
K0U D f.x; y/T W x 1 jyj; jyj 1g (6.23a)
M/M
pl
1 N/Npl
U
K0
L
K0
v ˇ ˇ
u ˇ mC ˇ
ti u ˇ i ˇ
t
1ˇ ˇ (6.24b)
U
Nipl ˇ Mipl ˇ
ˇ ˇ
ˇ mi ˇ
ˇ ˇ
ˇM ˇ 1 (6.24c)
ipl
ˇ ˇ
ˇ mC ˇ
ˇ i ˇ
ˇ ˇ 1: (6.24d)
ˇ Mipl ˇ
For compression, with (4.17a, c) and (6.23b) we get the feasibility condition
s ˇ ˇ
ti ˇm ˇ
L 1 ˇˇ i ˇˇ (6.24e)
Nipl Mipl
v ˇ ˇ
u ˇ mC ˇ
ti u ˇ ˇ
t1 ˇ i
ˇ (6.24f)
L
Nipl ˇ Mipl ˇ
262 6 Expected Total Cost Minimum Design of Plane Frames
ˇ ˇ
ˇ mi ˇ
ˇ ˇ
ˇM ˇ 1 (6.24g)
ipl
ˇ ˇ
ˇ mC ˇ
ˇ i ˇ
ˇ ˇ 1: (6.24h)
ˇ Mipl ˇ
jm
i j Mipl (6.25c)
jmC
i j Mipl : (6.25d)
For computational purposes, piecewise linearizations are applied [99] to the nonlin-
ear conditions (6.25a, b). A basic approximation of K0L and K0U is given by
and
Nipl
L
ti Nipl
U
(6.27a)
jm
i j Mipl (6.27b)
jmC
i j Mipl : (6.27c)
6.2 Stochastic Linear Programming Techniques 263
M/Mpl
1 N/Npl
K0
approximation KUu
0
approximation KLu
0
Then,
L
Nipl WD jyiL jAi D yiU Ai DW Nipl
U
; (6.28b)
hence,
Nipl WD Nipl
L
D Nipl
U
D yi Ai : (6.28c)
Moreover,
where y ic denotes the arithmetic mean of the centroids of the two half areas of the
cross-sectional area Ai of bar i .
Depending on the geometric from of the cross-sectional areal (rectangle, circle,
etc.), for the element load vectors
0 1
ti
Fi D @ mCi
A ; i D 1; : : : ; B; (6.29)
mi
Here, ˛ > 1 is a constant depending on the type of the cross-sectional area of the
i th bar. Defining the convex set
( ! )
x
K0˛ WD 2 R W jxj C jyj 1 ;
2 ˛
(6.31)
y
ti
!
Nipl
mi
2 K0˛ (negative end) (6.32a)
Mipl
0 1
ti
@ Nipl
A 2 K0˛ (positive end): (6.32b)
mCi
Mipl
x ! x; y ! y;
We have
u1 ˛u1˛1
rf D (6.34a)
u2 1
u1 ˛..u1 //˛1 ˛u1˛1
rf D D (6.34b)
u2 1 1
u1˛..u1 //˛1 ˛u1˛1
rf D D (6.34c)
u2 1 1
˛1
u1 ˛u1
rf D ; (6.34d)
u2 1
where
u1 u1
rf D rf (6.35a)
u2 u2
u1 u1
rf D rf : (6.35b)
u2 u2
we define
1 1
rf WD ; (6.36c)
0 0
1 1
rf WD : (6.36d)
0 0
u1
Using a boundary point of K0˛ with u1 ; u2 > 0, the feasible domain K0˛
u2
can be approximated from outside by the convex polyhedron defined as follows:
From the gradients (6.36a–d) we obtain next to the already known conditions (no
M –N -interaction):
T T
0 x 0 0 x
rf D 0
1 y 1 1 y 1
T T
0 x 0 0 x
rf D 0
1 y 1 1 y C1
T T
1 x 1 1 x1
rf D 0
0 y 0 0 y
T T
1 x 1 1 xC1
rf D 0:
0 y 0 0 y
This yields
hence,
jxj 1 (6.37a)
jyj 1: (6.37b)
6.2 Stochastic Linear Programming Techniques 267
Moreover, with the gradients (6.34a–d), cf. (6.35a, b), we get the additional
conditions
T
u1 x u1
rf
u2 y u2
˛1 T
˛u1 x u1
D 0 (1st quadrant) (6.38a)
1 y u2
T
u1 x u1
rf
u2 y u2
˛1 T
˛u1 x C u1
D 0 (3rd quadrant) (6.38b)
1 y C u2
T
u1 x u1
rf
u2 y u2
T
˛u1˛1 x C u1
D 0 (2nd quadrant) (6.38c)
1 y u2
T
u1 x u1
rf
u2 y u2
T
˛u1˛1 x u1
D 0 (4th quadrant) (6.38d)
1 y C u2
This means
˛u1˛1 x ˛u˛1 C y u2 D ˛u1˛1 x C y ˛u˛1 C u2 0 (6.39a)
˛u1˛1 x C ˛u˛1 C y C u2 D ˛u1˛1 x C y C ˛u˛1 C u2 0 (6.39b)
˛u1˛1 x ˛u˛1 C y u2 D ˛ u1˛1 x C y ˛u˛1 C u2 0 (6.39c)
˛u1˛1 x ˛u˛1 y u2 D ˛ u1˛1 x y ˛u˛1 C u2 0: (6.39d)
With
T
u1 u1
˛u˛1 C u2 D rf DW ˇ.u1 ; u2 / (6.40)
u2 u2
˛u1˛1 x C y ˇ.u1 ; u2 / 0
˛u1˛1 x C y C ˇ.u1 ; u2 / 0
268 6 Expected Total Cost Minimum Design of Plane Frames
˛u1˛1 x C y ˇ.u1 ; u2 / 0
˛u1˛1 x y ˇ.u1 ; u2 / 0:
1 x 1 (6.42a)
1 y 1 (6.42b)
ˇ.u1 ; u2 / ˛u1˛1 x C y ˇ.u1 ; u2 / (6.42c)
ˇ.u1 ; u2 / ˛u1˛1 x y ˇ.u1 ; u2 /: (6.42d)
Obviously, each further point uO 2 @K0˛ with uO 1 > 0; uO 2 > 0 yields additional
inequalities of the type (6.42c, d).
Condition (6.42a–d) can be represented in the following vectorial form:
1 x 1
I (6.43a)
1 y 1
1 x 1
ˇ.u1 ; u2 / H.u1 ; u2 / ˇ.u1 ; u2 / ; (6.43b)
1 y 1
Choosing a further boundary point uO of K0˛ with uO 1 > 0; uO 2 > 0, we get additional
conditions of the type (6.43b).
Using (6.42a–d), for the original yield condition (6.32a, b) we get then the
approximative feasibility condition:
i) Negative end of the bar
ti m
ˇ.u1 ; u2 / ˛u1˛1 C i ˇ.u1 ; u2 / (6.45c)
Nipl Mipl
ti m
ˇ.u1 ; u2 / ˛u1˛1 i ˇ.u1 ; u2 /: (6.45d)
Nipl Mipl
Defining
0 1 1 0 1
Nipl 0 0 ti
B C
.i / a.!/; x WD @ 0 M1ipl 0 A ; Fi D @ mCi
A; (6.46)
1
0 0 Mipl mi
Nipl yi Ai yi Ai 1
D D D ; (6.48)
Mipl yi Wipl yi Ai yNic yNic
m
ˇ.u1 ; u2 /Nipl ˛u1˛1 ti C i
ˇ.u1 ; u2 /Nipl (6.49a)
yNic
270 6 Expected Total Cost Minimum Design of Plane Frames
m
ˇ.u1 ; u2 /Nipl ˛u1˛1 ti i
ˇ.u1 ; u2 /Nipl (6.49b)
yNic
mC
ˇ.u1 ; u2 /Nipl ˛u1˛1 ti C i
ˇ.u1 ; u2 /Nipl (6.49c)
yNic
mC
ˇ.u1 ; u2 /Nipl ˛u1˛1 ti i
ˇ.u1 ; u2 /Nipl : (6.49d)
yNic
t
!
Npl x
m 2 K0˛ D 2 R2 W jxj˛ C jyj 1
Mpl y
Ni 0 > 0; Mi 0 > 0; i D 1; : : : ; B;
Putting
Nipl Mipl
i D i .a.!/; x/ WD min ; ; (6.50)
Ni 0 Mi 0
we have
i i
Nipl
1; Mipl
1
Ni 0 Mi 0
and therefore
ˇ ˇ˛ ˇ ˙ ˇ ˇ ˇ ˇ ˇ˛ ˇ ˇ ˇˇ ˇ ˇ ˇ ˇ ˇ
ˇ ti ˇ ˇ mi ˇ ˇ ti ˇ˛ ˇˇ i ˇˇ ˇ m˙ ˇ ˇ i ˇˇ ˇ ti ˇ˛ ˇ m˙ ˇ
ˇ ˇ Cˇ ˇDˇ ˇ ˇ ˇ C ˇ i ˇˇ ˇ ˇ ˇ Cˇ i ˇ:
ˇN ˇ ˇ M ˇ ˇ N ˇ ˇ Nipl ˇ ˇ M ˇ ˇ Mipl ˇ ˇ N ˇ ˇ M ˇ
ipl ipl i i0 i i0 i i0 i i0
Ni 0 Mi 0
(6.51)
6.2 Stochastic Linear Programming Techniques 271
or
ti !
i Ni 0
m˙ 2 K0˛ : (6.52)
i
i Mi 0
ti m˙
i
Remark 6.4 Multiplying with i we get quotients Ni 0 ; Mi 0 with fixed denominators.
Hence, multiplying (6.53d–g) with i , we get the equivalent system
ti mC
ˇ.u1 ; u2 /i ˛u1˛1 i ˇ.u1 ; u2 /i : (6.54g)
Ni 0 Mi 0
ˇ ˇ
ˇ mC ˇ
ˇ ˛1 ti i ˇ Nipl
ˇ˛u1 ˇ ˇ.u1 ; u2 / (6.56j)
ˇ Ni 0 Mi 0 ˇ Ni 0
ˇ ˇ
ˇ mC ˇˇ
ˇ ˛1 ti Mipl
ˇ˛u1 i ˇ ˇ.u1 ; u2 / (6.56k)
ˇ Ni 0 Mi 0 ˇ Mi 0
ti ; mC
i ; mi ; Ai or x
Symmetric yield stresses with respect to compression and tension, cf. (6.28a), do
not hold for all materials. Thus, we still have to consider the case
L
In case (6.57) we have different plastic capacities Nipl ¤ Nipl
U
with respect
to compression and tension. Thus must be taken into account in the yield condi-
tion (6.30a, b), hence,
ˇ ˇ˛ ˇ ˙ ˇ
ˇ ti ˇ ˇ ˇ
ˇ ˇ C ˇ mi ˇ 1:
ˇN ˇ ˇM ˇ
ipl ipl
hence,
x1 (6.59a)
1 y 1 (6.59b)
˛u1˛1 x C y ˇ.u1 ; u2 / (6.59c)
˛u1˛1 x y ˇ.u1 ; u2 / (6.59d)
Because of
Obviously, the piecewise linearization (6.59a–d) of K0˛U effects only the case
x > 0.
Corresponding to (6.59a–e) the set K0˛L is piecewise linearized by means of the
conditions
1 x (6.63a)
1 y 1 (6.63b)
˛u1˛1 x C y ˇ.u1 ; u2 / (6.63c)
˛u1˛1 x y ˇ.u1 ; u2 /: (6.63d)
Hence, the piecewise linearization (6.63a–d) of K0˛L effects the case x < 0 only.
yiL ¤ yiU :
U
Since the plastic capacities Nipl D yiU Ai and Nipl
L
D jyiL jAi with respect to
tension and compression are different, we have to consider the cases ti > 0 (tension)
and ti < 0 (compression) differently.
Because of K0˛ D K0˛L \ K0˛U , we obtain the following conditions: From
(6.59a–d), hence, for K0˛U we get
ti
U
1 (6.66a)
Nipl
ˇ ˙ˇ
ˇ mi ˇ
ˇ ˇ
ˇM ˇ 1 (6.66b)
ipl
ti m˙
˛u1˛1 U
C i
ˇ.u1 ; u2 / (6.66c)
Nipl Mipl
ti m˙
˛u1˛1 U
i ˇ.u1 ; u2 /: (6.66d)
Nipl Mipl
1 ti
L
Nipl
(6.67a)
ˇ ˇ
ˇ m˙ ˇ
ˇ i ˇ1 (6.67b)
ˇ Mipl ˇ
m˙
˛u1˛1 NtiL C i
Mipl
ˇ.u1 ; u2 / (6.67c)
ipl
m˙
˛u1˛1 NtiL i
Mipl ˇ.u1 ; u2 /: (6.67d)
ipl
276 6 Expected Total Cost Minimum Design of Plane Frames
Nipl
L
ti Nipl
U
(6.68a)
Mipl m
i Mipl (6.68b)
Mipl mC
i Mipl (6.68c)
ti mC
˛u1˛1 U
C i
ˇ.u1 ; u2 / (6.68d)
Nipl Mipl
ti mC
˛u1˛1 U
i
ˇ.u1 ; u2 / (6.68e)
Nipl Mipl
ti mC
˛u1˛1 L
C i ˇ.u1 ; u2 / (6.68f)
Nipl Mipl
ti mC
˛u1˛1 L
i
ˇ.u1 ; u2 / (6.68g)
Nipl Mipl
ti m
˛u1˛1 U
C i
ˇ.u1 ; u2 / (6.68h)
Nipl Mipl
ti m
˛u1˛1 U
i
ˇ.u1 ; u2 / (6.68i)
Nipl Mipl
ti m
˛u1˛1 L
C i
ˇ.u1 ; u2 / (6.68j)
Nipl Mipl
ti m
˛u1˛1 L
i
ˇ.u1 ; u2 /: (6.68k)
Nipl Mipl
U
Corresponding to (6.49a–d), the inequalities (6.68d–k) may be multiplied by Nipl ,
L
Nipl resp., in order to get denominators independent of Ai ; x, respectively.
,
Two-sided constraints are obtained by introducing lower bounds in (6.68d–k)
ˇ.u1 ; u2 / (6.69)
L U
for the plastic capacities Nipl ; Nipl ; Mipl . Consider then the relative capacities iL ; iU ,
cf. (6.50),
( )
L
Nipl Mipl
iL WD min L M
; ; (6.70a)
Ni 0 i0
( )
U
Nipl Mipl
iU WD min ; : (6.70b)
NiU0 Mi 0
Based again on (6.52) and the representation K0˛ D K0˛L \ K0˛U , corresponding
to (6.53d–g) we get
ti m˙
˛u1˛1 C i
ˇ.u1 ; u2 / (6.71a)
iU NiU0 U
i Mi 0
ti m˙
˛u1˛1 i
ˇ.u1 ; u2 / (6.71b)
iU NiU0 iU Mi 0
ti m˙
˛u1˛1 C i
ˇ.u1 ; u2 / (6.71c)
iL NiL0 iL Mi 0
ti m˙
˛u1˛1 i
ˇ.u1 ; u2 /: (6.71d)
iL NiL0 L
i Mi 0
Nipl
L
ti Nipl
U
(6.72a)
Mipl m
i Mipl (6.72b)
Mipl mC
i Mipl (6.72c)
ti mC
˛u˛1
1 C i
iU ˇ.u1 ; u2 / (6.72d)
NiU0 Mi 0
ti mC
˛u1˛1 i
iU ˇ.u1 ; u2 / (6.72e)
NiU0 Mi 0
ti mC
˛u1˛1 C i
iL ˇ.u1 ; u2 / (6.72f)
NiL0 Mi 0
ti mC
˛u1˛1 L
i iL ˇ.u1 ; u2 / (6.72g)
Ni 0 Mi 0
ti m
˛u˛1
1 C i
iU ˇ.u1 ; u2 / (6.72h)
NiU0 Mi 0
278 6 Expected Total Cost Minimum Design of Plane Frames
ti m
˛u1˛1 U
i iU ˇ.u1 ; u2 / (6.72i)
Ni 0 Mi 0
ti m
˛u1˛1 C i
iL ˇ.u1 ; u2 / (6.72j)
NiL0 Mi 0
ti m
˛u1˛1 L
i iL ˇ.u1 ; u2 /: (6.72k)
Ni 0 Mi 0
Stochastic Optimization
CF D R (6.74b)
Thus, since in some cases the vector R of external loads depend also on the
design r-vector x, we get
R D R .a.!/; x/ : (6.75b)
Of course, the plastic capacities depend also on the vectors x and a.!/, hence,
Nipl D Nipl
.a.!/; x/ ; D L; U (6.75c)
Mipl D Mipl .a.!/; x/ : (6.75d)
6.2 Stochastic Linear Programming Techniques 279
We assume that the probability distribution and/or the needed moments of the
random parameter vector a D a.!/ are known [7, 100].
The remaining deterministic constraints for the design vector x are represented
by
x2D (6.76)
H .i / Fi C zi D h.i / ; i D 1; : : : ; B: (6.78)
If
then (6.77) holds, and the yield condition is then fulfilled too, or holds with a
prescribed accuracy.
However, in case
the survival condition is violated at some points of the structure. Hence, structural
failures may occur. The resulting costs Q of failure, damage and reconstructure of
the frame is a function of the vectors zi ; i D 1; : : : ; B, defined by (6.78). Thus, we
have
where
Due to the survival condition (6.79a), we may consider cost functions Q such that
Q.z/ D 0; if z 0; (6.81)
X
B
Q.z/ D Qi .zi /; (6.82)
i D1
the member cost functions Qi D Qi .zi / are often defined [70, 100] by considering
the linear function
with certain vectors qiC ; qi of cost coefficients for the evaluation of the condition
zi 0; zi 6 0, respectively.
The cost function Qi D Qi .zi / is then defined by the minimization problem
X
min .qil yil C qilC yilC / (6.85a)
lD1
6.2 Stochastic Linear Programming Techniques 281
X
Qi .zi / D Qil .zil /
lD1
X
D maxfqilC zil ; qil zil g; (6.89a)
lD1
X
B
Q.z/ D Qi .zi /
i D1
X
B X
D maxfqilC zil ; qil zil g: (6.89b)
i D1 lD1
282 6 Expected Total Cost Minimum Design of Plane Frames
L
L D
; (6.90a)
E
where E denotes the modulus of elasticity.
Assuming that the neutral fiber is equal to the axis of symmetry of the element,
for the total stress
in the upper (“+”) lower (“”) fibre of the boundary we have
t
m
D ˙ ; (6.90b)
A W
where W denotes the axial modulus of the cross-sectional area of the element
(beam).
Representing the change
V of volume of an element by
V D A
L; (6.91a)
then
L L
t
m
V D A
L D A
D A ˙
E E A W
L L A L L 1 L L 1
D
t ˙
m D
t ˙
m D
t ˙
m; (6.91b)
E EW E W
E A E E yNc
L
K WD ; (6.92a)
E
and an appropriate cost factor for the evaluation of violations of moment constraints
reads
L 1
M D : (6.92b)
E yNc
6.2 Stochastic Linear Programming Techniques 283
Total Costs
Denoting by
G0 D G0 a.!/; x (6.93a)
the primary costs, such as material costs, costs of construction, the total costs
including failure or recourse costs are given by
G D G0 a.!/; x C Q z a.!/; x; F .!/ : (6.93b)
Hence, the total costs G D G a.!/; x; F .!/ depend on the vector x D
T
.x1 ; : : : ; xr /T of design variables, the random vector a.!/ D a1 .!/; : : : ; a .!/
of model parameters and the random vector F D F .!/ of all internal loadings.
Minimizing the expected total costs, we get the following stochastic optimization
problem of recourse type [90]
min E G0 a.!/; x C Q z .a.!/; x; F .!// (6.94a)
s.t. H .i / a.!/; x Fi .!/ C zi .!/ D h.i / a.!/; x a.s.;
i D 1; : : : ; B (6.94b)
CF.!/ D R a.!/; x a.s. (6.94c)
x 2 D: (6.94d)
i D 1; : : : ; B (6.95b)
CF.!/ D R .a.!/; x/ a.s. (6.95c)
x 2 D; yiC .!/; yi .!/ 0 a.s.; i D 1; : : : ; B: (6.95d)
Remark 6.5 Stochastic optimization problems of the type (6.95a–d) are called
“two-stage stochastic programs” or “stochastic problems with recourse”.
284 6 Expected Total Cost Minimum Design of Plane Frames
In many cases the primary cost function G0 D G0 a.!/; x represents the
volume or weight of the structural, hence,
X
B
G0 a.!/; x WD i .!/Vi .x/
i D1
X
B
D i .!/Li Ai .x/; (6.96)
i D1
Ai .x/ WD ui xi (6.97)
X
B X
B
G0 .a.!/; x/ D i .!/Li ui .x/ D i .!/Li hi xi
i D1 i D1
D c .a.!//T x; (6.98a)
where
Discretization Methods
X
s
Pa./ WD ˛k a.k/ (6.99a)
kD1
with
X
s
˛k 0; k D 1; : : : ; s; ˛k D 1: (6.99b)
kD1
6.2 Stochastic Linear Programming Techniques 285
X
B
.k/ C.k/
min G 0 .a.k/ ; x/ C ˛k qiT yi C qiCT yi (6.101a)
i D1
.k/ C.k/ .k/
s.t. H .i / .a.k/ ; x/Fi C yi yi D h.i / .a.k/ ; x/;
i D 1; : : : ; B; k D 1; : : : ; s (6.101b)
CF.k/ D R.a.k/ ; x/; k D 1; : : : ; s (6.101c)
C.k/ .k/
x 2 D; yi ; yi 0; k D 1; : : : ; s; i D 1; : : : ; B: (6.101d)
Complete Recourse
According to Sect. 6.2.6, the evaluation of the violation of the yield condition (6.77)
is based on Eqs. (6.78), hence
H .i / Fi C zi D h.i / ; i D 1; : : : ; B:
zi D h.i / H .i / Fi
286 6 Expected Total Cost Minimum Design of Plane Frames
min q .i /T y .i / (6.102a)
bzgl. M y .i / .i /
D zi (6.102b)
y .i / 0: (6.102c)
Here, q .i / is a given cost vector and M .i / denotes the so-called recourse matrix
[69, 70].
We assume that the linear equation (6.102b) has a solution y .i / 0 for each
vector zi . This property is called ‘‘complete recourse”.
In the present case the stochastic optimization problem (6.94a–c) reads
!
XB
min E G0 a.!/; x C q .i /T y .i / .!/ (6.103a)
i D1
s.t. H .i / a.!/; x Fi .!/ C M .i / y .i / .!/ D h.i / a.!/; x a.s.,
i D 1; : : : ; B (6.103b)
CF.!/ D R .a.!/; x/ a.s. (6.103c)
x 2 D; y .!/ 0 a.s.;
.i /
i D 1; : : : ; B: (6.103d)
Numerical Implementation
X
R
min E.c T x C q T y/ D c T x C pr q T y r (6.104a)
rD1
T x C Wy 1 D h1
:: :: :: (6.104b)
: : :
Tx C Wy D hR
R
x 2 D; y r 0; r D 1; : : : ; R: (6.104c)
6.2 Stochastic Linear Programming Techniques 287
Here, the vectors and matrices c; q; T; W; h1 ; : : : ; hR contain the data of the (dis-
cretized) stochastic structural optimization problem, and the vectors x; y 1 ; : : : ; y R
involve the variables and slack variables of the given problem. Finally, p 1 ; : : : ; p R
denote the probabilities of the R realizations of the (discretized) stochastic optimal
design problem.
Chapter 7
Stochastic Structural Optimization
with Quadratic Loss Functions
7.1 Introduction
Problems from plastic analysis and optimal plastic design are based on the convex,
linear or linearized yield/strength condition and the linear equilibrium equation for
the stress (state) vector. In practice one has to take into account stochastic variations
of several model parameters. Hence, in order to get robust optimal decisions, the
structural optimization problem with random parameters must be replaced by an
appropriate deterministic substitute problem. A direct approach is proposed based
on the primary costs (weight, volume, costs of construction, costs for missing
carrying capacity, etc.) and the recourse costs (e.g. costs for repair, compensation
for weakness within the structure, damage, failure, etc.). Based on the mechanical
survival conditions of plasticity theory, a quadratic error/loss criterion is developed.
The minimum recourse costs can be determined then by solving an optimization
problem having a quadratic objective function and linear constraints. For each
vector a./ of model parameters and each design vector x, one obtains then an
explicit representation of the “best” internal load distribution F . Moreover, also the
expected recourse costs can be determined explicitly. It turns out that this function
plays the role of a generalized expected compliance function involving a generalized
stiffness matrix. For the solution of the resulting deterministic substitute problems,
i.e., the minimization of the expected primary cost (e.g. volume, weight) subject to
expected recourse cost constraints or the minimization of the expected total primary
and recourse costs, explicit finite dimensional parameter optimization problems
are obtained. Furthermore, based on the quadratic cost approach, lower and upper
bounds for the probability of survival can be derived.
In optimal plastic design of mechanical structure [38] one has to minimize a
weight, volume or more general cost function c, while in limit load analysis [74] of
plastic mechanical structures the problem is to maximize the load factor —in both
cases—subject to the survival or safety conditions, consisting of the equilibrium
equation and the so-called yield (feasibility) condition of the structure.
X
B
G0 .x/ D i 0 .!/Li Ai .x/ (7.1a)
i D1
G0 D G0 .a; x/ WD (7.1b)
reads
FQiL a.!/; x Fi FQiU a.!; x ; i D 1; : : : ; B; (7.4a)
where the bounds FQiL ; FQiU containing the plastic capacities with respect to axial
forces and moments are given by
0 1 0 1
NiplL
a.!/; x yiL a.!/ Ai .x/
B C B C
B C B C
FQiL a.!/; x WD B Mipl a.!/; x C D B yiU a.!/ Wipl .x/ C (7.4b)
@ A @ A
Mipl a.!/; x yiU a.!/ Wipl .x/
0 1 0 1
U
Nipl a.!/; x yiU a.!/ Ai .x/
B C B C
B C B C
FQiU a.!/; x D B Mipl a.!/; x C D B yiU a.!/ Wipl .x/ C ; (7.4c)
@ A @ A
Mipl a.!/; x yiU a.!/ Wipl .x/
Here,
8 9
< Nipl
L
a.!/; x Mipl a.!/; x =
i a.!/; x D min
L ; (7.4g)
: Ni 0 Mi 0 ;
8 9
< Nipl
U
a.!/; x Mipl a.!/; x =
Ui a.!/; x D min ; (7.4h)
: Ni o Mi 0 ;
According to (7.4a, f), the feasibility condition for the vector F of interior loads
(member forces and moments) can be represented uniformly by the conditions
.i /
FilL a.!/; x Hl Fi FilU a.!/; x ; i D 1; : : : ; B; l D 1; 2; : : : ; l0 C 3;
(7.5a)
.i /
where the row 3-vectors Hl and the bounds FilL ; FilU ; i D 1; : : : ; B; l D
1; : : : ; l0 C 3, are defined by (7.4a–c) and (7.4f–h). Let e1 DW H1T ; e2 DW H2T ; e3 DW
H3T denote the unit vectors of R3 .
Defining the .l0 C 3/ 3 matrix H .i / by
0 1
e1T
B e2T C
B C
B C
B e3T C
H .i / WD B
B H4 .Ni 0 ; Mi 0 /
C
C (7.5b)
B C
B :: C
@ : A
Hl0 C3 .Ni 0 ; Mi 0 /
and the .l0 C 3/-vectors FiL D FiL a.!/; x ; FiU D FiU a.!/; x by
0 1 0 QU 1
FQiL Fi
B h1 L C B h1 U C
B i C B i C
FiL WD B : C ; FiU WD B : C; (7.5c)
@ : A: @ : : A
hl0 L
i hl0 Ui
jzil j 1; i D 1; : : : ; B; l D 1; : : : ; l0 C 3; (7.8a)
.i /
H Fi Filc
zil D zil Fi I a.!/; x D l ; i D 1; : : : ; B; l D 1; : : : ; l0 C 3: (7.8b)
%il
The quotient zil ; i D 1; : : : ; B; l D 1; : : : ; l0 C 3, denotes the relative deviation
.i /
of the load component Hl Fi from its “ideal” value Filc with respect to the radius
%il of the feasible interval ŒFilL ; FilU . According to (7.8a, b), the absolute values
jzil j of the quotients zil should not exceed the value 1. The absolute value jzil j of
the quotient zil denotes the percentage of use of the available plastic capacity by the
corresponding load component. Obviously, jzil j D 1; jzil j > 1, resp., means maximal
use, overcharge of the available resources.
Consider now the .l0 C 3/-vectors
.i / .i /
H1 Fi Fi1c H2 Fi Fic2
zi WD .zi1 ; zi 2 ; : : : ; zil0 C3 /T D ; ; ;
%i1 %i 2
.i / !T
Hl0 C3 Fi Filc0 C3
: (7.8c)
%il0 C3
With
0 1 0 1 0 1
%i1 %i1 0 ::: 0 Fi1c
B %i 2 C B 0 %i 2 ::: 0 C BFc C
B C B C c B i2 C
%i WD B :: C ; %id WD B : :: :: C i
; F WD B: C (7.8d)
@ : A @ :: : : A @ :: A
%il0 C3 0 0 : : : %il0 C3 Filc0 C3
we get
zi D %1
id .H Fi Fi /:
.i / c
(7.8e)
.i / .i /
!T
ti mli mri H4 Fi ; Hl C3 Fi
zi D ; ; ; ;:::; 0 U : (7.9b)
Ai yi Ai yi yNic Ai yi yNic h4 i
U U U U
hl0 C3 i
According to the methods introduced in [97, 99, 100], the fulfillment of the
survival condition for elastoplastic frame structures, hence, the equilibrium condi-
tion (7.2) and the feasibility condition
(7.6) or (7.8a, b), can be described by means
of the state function s D s a.!/; x defined, in the present case, by
n ˇ ˇˇ
ˇ
s D s a.!/; x WD min s W ˇzil Fi I a.!/; x ˇ 1 s; i D 1; : : : ; B;
o
l D 1; 2; : : : ; l0 C 3; CF D P a.!/; x : (7.10)
Hence, the state function s is the minimum value function of the linear program
(LP)
min s (7.11a)
s.t.
ˇ ˇ
ˇ ˇ
ˇ ˇ
ˇzil Fi I a.!/; x ˇ 1 s; i D 1; : : : ; B; l D 1; : : : ; l0 C 3 (7.11b)
ˇ ˇ
CF D P a.!/; x : (7.11c)
Since the objective function s is bounded from below and a feasible solution .s; F /
always exists, LP (7.11a–c) has an optimal solution .s ; F / D s a.!/; x ;
!
F a.!/; x for each configuration a.!/; x of the structure.
7.2 State and Cost Functions 295
Consequently, for the survival of the structure we have the following criterion,
cf. [99]:
Theorem 7.1 The elastoplastic frame structure having configuration .a; x/ carries
the exterior load P D P .a; x/ safely if and only if
s .a; x/ 0: (7.12)
!T
T T
z F I a.!/; x WD z1 F I a.!/; x ; : : : ; zB F I a.!/; x ; (7.13b)
If we put
sO D 1 C s or s D sO 1; (7.14)
Remark 7.1 According to (7.15a, b) and (7.12), the safety or survival condition of
the plane frame with plastic material can be represented also by
sO .a; x/ 1:
The following inequalities for norms or power/Hölder means kzk in RB.l0 C3/ are
well known [25, 57]:
1 1
kzk1 kzk1
B.l0 C 3/ B.l0 C 3/
1
p kzk2 kzk1 kzk2 ; (7.17a)
B.l0 C 3/
where
v
0 C3
u B l0 C3
X
B lX uX X
kzk1 WD jzil j; kzk2 WD t z2il : (7.17b)
i D1 lD1 i D1 lD1
1 1
sO .a; x/ p sO2 .a; x/ sO .a; x/ sO2 .a; x/; (7.18a)
B.l0 C 3/ B.l0 C 3/
where sO2 D sO2 .a; x/ is the modified state function function defined by
n o
sO2 .a; x/ WD min z.F I a; x/2 W CF D P .a; x/ : (7.18b)
Obviously, we have
q
sO2 .a; x/ D G1 .a; x/; (7.18c)
where G1 .a; x/ is the minimum value function of the quadratic program
X 0 C3
B lX
min zil .Fi I a; x/2 : (7.19)
CFDP a.!/;x i D1 lD1
The inequalities in (7.18a) show that for structural analysis and optimal design
purposes we may work also with the state function sO2 D sO2 .a; x/ which can be
defined easily by means of the quadratic program (7.19).
7.2 State and Cost Functions 297
Obviously,
are possible convex functions measuring the costs resulting from the position zil
of a load component HQ l Fi relative to the corresponding safety interval (plastic
.i /
If different weights are used in the objective function (7.19), then for the bars we
obtain, cf. (7.8c), the cost functions
X
B X
B
G1 WD kWi 0 zi k2 D zTi WiT0 Wi 0 zi : (7.21a)
i D1 i D1
Defining
0 1 0 1
W10 0 ::: 0 z1
B 0 W20 ::: 0 C B z2 C
B C B C
W0 WD B : :: :: :: C; z WD B : C; (7.21b)
@ :: : : : A @ :: A
0 0 : : : WB0 zB
we also have
G1 D G1 .a; xI F / D zT W0T W0 z
D kW0 zk22 D kzk22;W0 ; (7.21c)
Using the weighted quadratic cost function (7.20c), the state function sO2 D sO2 .a; x/
is replaced by
n o
sO2;W .a; x/ WD min z.F I a; x/ W CF D P .a; x/
0 2;W0
np o
D min G1 .a; xI F / W CF D P .a; x/ : (7.21e)
Since
1
sO2;W .a; x/ W 0 sO2 .a; x/ or sO2 .a; x/ sO .a; x/: (7.21g)
0
W 0 2;W0
7.3 Minimum Expected Quadratic Costs 299
Putting
0 1 0 1 0 1
H .1/ F1c %1
B H .2/ C BFc C B %2 C
B C c B 2 C B C
H WD B :: C ; F WD B : C ; % WD B : C; (7.21h)
@ : A @ :: A @ :: A
H .B/ FBc %B
If
FiL C FiU
F c D HFc with F c WD (7.22b)
2 i D1;:::;B
as in the case of no interaction between normal forces and moments, see (7.4a–c),
and in the case of symmetric yield stresses
we also have
G1 a.!/; xI F D .F F c /T H T %1 T 1
d W0 W0 %d H.F F /:
c
(7.22d)
For simplification, we assume first in this section that the total cost representa-
tion (7.22d) or (7.22e) holds.
According to the equilibrium condition (7.2), the total vector F of generalized
forces of the members fulfills
CF D P .a.!/; x/:
F D F .a.!/; x/
300 7 Stochastic Structural Optimization with Quadratic Loss Functions
of the structure can be obtained by solving the following optimization problem with
quadratic objective function and linear constraints
for the random configuration a.!/; x we get the minimum expected (total)
quadratic costs
G 1 D G 1 .x/; x 2 D; (7.25a)
where G 1 .x/ may be obtained by interchanging expectation and minimization
G 1 .x/ D G1 .x/ WD E minfG1 .a.!/; xI F / W CF D P .a.!/; x/g: (7.25b)
hence,
min .F F c /T H T %1 T 1
d W0 W0 %d H.F F /;
c
(7.26)
CFDP .a.!/;x/
with quadratic objective function and linear constraints with respect to F can be
solved by means of Lagrange techniques. We put
W D W .a; x/ WD H T %1 T 1
d W0 W0 %d H (7.27)
0 D rF L D 2W .F F c / C C T ; (7.28b)
0 D r L D CF P: (7.28c)
7.3 Minimum Expected Quadratic Costs 301
1
F D F c W 1 C T (7.28d)
2
and
1
P D CF D CFc CW 1 C T ; (7.28e)
2
hence,
1
F D F c W 1 C T D F c W 1 C T .CW 1 C T /1 .CF c P /: (7.28f)
2
where “tr” denotes the trace of a matrix. The minimal expected value G1 is then
given by
G1 .x/ D EG1 a.!/; x
T 1 1
D E CF c .a.!/; x P a.!/; x/ CW a.!/; x C T
CF c a.!/; x P a.!/; x
1 1 c
D E tr CW a.!/; x C T CF a.!/; x P a.!/; x
T
CF c a.!/; x P a.!/; x : (7.29a)
Since the vector P D P .a.!/; x/ of external generalized forces and the vector
of yield stresses U D U .a.!/; x/ are stochastically independent, then in case
yiL D yiU ; i D 1; : : : ; B, we have
where
and
We compare now, especially in case F c D 0, formula (7.28g) for the costs G1 D
G1 .a; x/ with formula
WD uT P
u WD Kel1 P
is the vector of displacements, and Kel denotes the stiffness matrix in case of an
elastic structure. Obviously, the cost function G1 D G1 .a; x/ may be interpreted
as a generalized compliance function, and the m m matrix K D K.a; x/ can be
interpreted as the “generalized stiffness matrix” of the underlying plastic mechanical
structure. If we suppose that
where
W% .a; x/ WD %1 T 1
d .a; x/W0 W0 %d .a; x/: (7.31b)
Though also in this case problem (7.31a) can be solved explicitly, the resulting
total cost function has a difficult form. In order to apply the previous technique,
the vector F c is approximated—in the least squares sense—by vectors HF with
F 2 R3B . Hence, we write
F c HFc ; (7.32a)
We obtain
F c D F c .F c / WD .H T H /1 H T F c : (7.32c)
e.F c / WD kHF c F c k
D H.H T H /1 H T I F c : (7.32d)
With the vector F c D F c .F c / the total costs G1a D G1a a.!/; xI F can be
approximated now, see (7.22a), by
G1a a.!/; xI F WD .HF HFc /T W% a.!/; x .HF HFc /
D .F F c /T H T W% a.!/; x H.F F c /
D .F F c /T W a.!/; x .F F c /; (7.33a)
304 7 Stochastic Structural Optimization with Quadratic Loss Functions
Obviously, the approximate cost function G1a D G1a .a; xI F / has the same form
as the cost function G1 D G1 .a; xI F / under the assumption (7.22b), see (7.22d).
Hence, the minimum cost function G1a D G1a .a; x/ can be determined by solving,
cf. (7.23),
and the expected cost function G1 D G1 .x/ representing the expected total
weighted quadratic costs resulting from a violation of the feasibility condition
(7.4a, f), we get [46, 50] the optimization problem
X
B
G0 .x/ WD i 0 Li Ai .x/ (7.35d)
i D1
with i 0 WD Ei 0 .!/, and G1 D G1 .x/ is defined by (7.29a) or (7.29b).
Due to (7.20c) and (7.21a–c), the upper cost bound 1 can be defined by
1 WD g1 G1max ; (7.35e)
where g1 > 0 is a certain reliability factor, and G1max denotes the maximum of
the total cost function G1 D G1 .z/ on the total admissible z-domain Œ1; 1.l0 C3/B .
Hence,
X
B
G1max WD max kWi 0 zi k2
z2Œ1;1.l0 C3/B
i D1
X
B
D max kWi 0 zi k2
zi 2Œ1;1.l0 C3/
i D1
X
B
D max kWi 0 e .j / k2 ; (7.35f)
1j 2l0 C3
i D1
1
p sO2 .a; x/ sO .a; x/ sO2 .a; x/;
B.l0 C 3/
Due to the first definition of G1 D G1 .a; x/ by (7.18c) and (7.19), related to the
case W0 D I , we also have
P G1 a.!/; x 1 ps .x/ P G1 a.!/; x B.l0 C 3/ : (7.37b)
G .x/
P G1 a.!/; x g1 1 1 ; (7.38b)
g1
where the expectation G1 .x/ D EG1 a.!/; x is given by (7.29a) or (7.29b). The
probabilistic constraint
P G1 a.!/; x g1 ˛min (7.39a)
p
for the quadratic mean rate sO2 D G1 .a; x/ of minimum possible use of plastic
capacity within the plane frame with configuration .a; x/ implies ps .x/ ˛min
for g1 D 1, cf. (7.37b). Hence, due to (7.38b), constraint (7.39a) and therefore
ps .x/ ˛min can be guaranteed then by the condition
see (7.35b).
Here, the weight or scale matrices Wi 0 and the weight or cost factors i 0 ; i D
1; : : : ; B, must be selected such that thepdimensions of G0 and G1 coincide. For
example, if Wi 0 D I; i D 1; : : : ; B, and G1 .a; x/ is then the quadratic mean rate
7.4 Deterministic Substitute Problems 307
ref
with a certain weight or cost reference value G0 .
Minimizing now the expected total costs
subject to the equilibrium conditions (7.2) and the remaining condition for the
decision variables
x 2 D; (7.40c)
min E.G0 .a.!/; x/ C G1 .a.!/; xI F .!//: (7.41)
CF.!/DP a.!/;x a:s:
x2D
According to the definitions (7.35d) of G0 and (7.25b) of G1 , problem (7.42) can
be represented also by
min G0 .x/ C G1 .x/ : (7.43)
x2D
308 7 Stochastic Structural Optimization with Quadratic Loss Functions
We first suppose that the structure consists of uniform material with a symmetric
random yield stress in compression and tension. Hence, we assume next to
with
K0 a.!/; x D .H T %1 T 1
d W0 W0 %d H /
1
where
1
˚ 1
K̊ 0 .x/ WD H T%.x/ T
˚ 1
d W0 W0 %.x/ d H : (7.46c)
Moreover, we get
1 U 1
U a.!/; x WD K a.!/; x D y .!/2 K̊.x/
1
D K̊.x/1 : (7.47c)
yU .!/2
In case of a random weight matrix W0 D W0 a.!/ , for U .x/ we also obtain a
representation of the type (7.47d), provided that (i) the random variables W0 a.!/
and yU .!/ are stochastically independent and (ii) K̊.x/ is defined by
0 !1 11
K̊.x/ WD @E C K̊ 0 W a.!/ ; x C T A : (7.47e)
we find
!
1
Z.x/ D .z1 ; z2 ; : : : ; zm / WD E K̊.x/1 B.x/
yU .!/2
!
1
DE K̊.x/1 b1 .x/; K̊.x/1 b2 .x/; : : : ; K̊.x/1 bm .x/ : (7.49d)
yU .!/2
With Eq. (7.50) for zj ; j D 1; : : : ; m, the expected cost function G1 .x/ can be
represented now by
s.t.
tr.z1 ; z2 ; : : : ; zm / 1 (7.52b)
!
1
K̊.x/zj D E bj .x/; j D 1; : : : ; m (7.52c)
yU .!/2
x 2 D; (7.52d)
s.t.
!
1
K̊.x/zj D E bj .x/; j D 1; : : : ; m (7.53b)
y .!/2
U
x 2 D: (7.53c)
If a representation of
1
U.x/ D EU a.!/; x D EK a.!/; x D ˇ.!/K̊.x/1 ;
see (7.29d), (7.30a, b), of the type (7.47d) does not hold, then we may apply the
approximative procedure described in the following.
First, the probability distribution Pa./ of the random vector a D a.!/ of model
parameters is approximated, as far it concerns the subvector aI D aI .!/ of a D
a.!/ of model parameters arising in the matrix
K D K a.!/; x D K aI .!/; x ;
by a discrete distribution
X
N
POaI ./ WD ˛s "a.s/ (7.54)
I
sD1
.s/
having realizations aI taken with probabilities ˛s ; s D 1; : : : ; N .
Then, the matrix function U D U.x/ can be approximated by
X
N
UO .x/ WD ˛s K .s/ .x/1 ; (7.55a)
sD1
312 7 Stochastic Structural Optimization with Quadratic Loss Functions
where
.s/ .s/
K .s/ .x/ WD K.aI ; x/ D CK 0 .aI ; x/C T ; (7.55b)
see (7.30b). Consequently, the expected cost function G1 D G1 .x/ is approxi-
mated by
T
GO1 .x/ WD trUO .x/EP a.!/; x P a.!/; x
X
N
T
D ˛s trK .s/ .x/1 EP a.!/; x P a.!/; x : (7.56)
sD1
.s/ .s/ 1
z.s/ D .z1 ; z2 ; : : : ; z.s/
m / WD K .x/ B.x/
.s/
.s/ 1
D K .x/ b1 .x/; K .s/ .x/1 b2 .x/ : : : ; K .s/ .x/1 bm .x/ ; (7.57)
.s/
where B D B.x/ is defined again by (7.49a). Thus, for the columns zj ; j D
1; : : : ; m, we obtain the conditions
.s/
K .s/ .x/zj D bj .x/; j D 1; : : : ; m; (7.58)
X N
GO1 .x/ D
.s/ .s/
˛s tr.z1 ; z2 ; : : : ; z.s/
m /; (7.59)
sD1
.s/
where zj ; j D 1; : : : ; m; s D 1; : : : ; N , are given by (7.58).
Because of the close relationship between the representations (7.59) and (7.51)
O
for G1 ; G1 , approximate mathematical optimization problems result from (7.59)
which are similar to (7.52a–d), (7.53a–c), respectively.
Ui ˙ L Q U .x/ ˙ Q L
i .x/
i
D y i ; (7.61b)
2 2
where, cf (7.4b, c, g, h),
ji jAi .x/ iU Wipl .x/
Q
i .x/ WD min ; ; D L; U: (7.61c)
Ni 0 Mi 0
follow from (7.8f), (7.8g), resp., by inserting formula (7.60) and extracting then
the random variable .!/. Because of (7.62a–c), the generalized stiffness matrix
K D K.a; x/ can be represented again in the form (7.47a), hence,
K a.!/; x D y2 .!/K̊.x/; (7.63a)
where
T
F̊ c .x/ WD F̊ c1 .x/T ; : : : ; F̊ cB .x/T : (7.63c)
314 7 Stochastic Structural Optimization with Quadratic Loss Functions
Thus, due to (7.32b, c), for the vector F c D F c a.!/; x defined by (7.32a–c)
we find
F c a.!/; x D .!/F c .x/ (7.63d)
with
Inserting now (7.63a, d) into formula (7.34b), for the (approximate) minimum total
costs we finally have, cf. (7.47c), (7.48),
1
1
G1a a.!/; x D tr K̊.x/ .!/CF c
.x/ P a.!/; x
.!/2
T
.!/CF c .x/ P a.!/; x
1 T
1
D tr K̊.x/ P a.!/; x P a.!/; x
.!/2
T
1 1
tr K̊.x/ CFc .x/P a.!/; x
.!/
CP a.!/; x F c .x/T C T C tr K̊.x/1 CFc .x/F c .x/T C T :
(7.64)
The minimum expected cost function G1a .x/ is then given, cf. (7.48), by
1
G1a .x/ DE trK̊.x/1 B.x/
.!/2
1
E trK̊.x/1 CF c .x/P .x/T C P .x/F c .x/T C T
.!/
Ctr K̊.x/1 CF c .x/F c .x/T C T ; (7.65)
where P D P .x/; B D B.x/ are again given by (7.49a, b). Obviously, also in the
present more general case G1a can be represented by
(7.66b)
see (7.51). Hence, due to the close relationship between (7.49d), (7.51) and
(7.66a, b), the deterministic substitute problems stated in Sect. 7.4 can be treated
as described in Sect. 7.5.
g1 .P / g1 (7.68a)
at a so–called design point P , see [38, 46, 50]. Since g1 D g1 .P / is convex,
we have
Moreover,
P WD EP.!/ (7.70d)
In practice, the following two cases are taken into account [38, 46, 50]:
Case 1: Linearization at P WD P
Under the above assumptions in this case we have
0 1
B g1 g1 .P / C
pQs .x0 I g1 / D ˚ B
@ r
C ps .x0 I g1 /:
A
rP g1 .P /T cov P ./ rP g1 .P /
(7.71)
Case 2: Linearization at a Boundary Point P of Œg1 .P / g1
Here it is g1 .P / D g1 and therefore
0 1
B rP g1 .P /T .P P / C
pQs .x0 I g1 / D ˚ B
@ r
C:
A (7.72)
rP g1 .P / cov P ./ rP g1 .P /
T
min c T P; (7.74)
.C FQ c P /T K 1 .C FQ c P /g1
see (7.72), (7.73), the best upper bound pQs .x0 ; g1 / can be obtained by solving the
minimization problem
The new approach for optimal design of elasto-plastic mechanical structures under
stochastic uncertainty is illustrated now by means of the 12-bar truss according to
Fig. 7.2.
Suppose that L D 1;000 mm, E D 7,200 N/mm2 is the elastictiy modulus,
and the yield stresses with respect to tension and compression are given by
7.7 Numerical Example: 12-Bar Truss 319
P1y Š N.; 2 /
having a normal distribution with mean and variance 2 . The standard deviation
is always 10% of the mean .
The numerical results presented in this section have been obtained by
Dipl.Math.oec. Simone Zier, Institute for Mathematics and Computer Applications,
Federal Armed Forces University, Munich.
The equilibrium matrix C of the 12-bar truss is given by
0 1
0 0 0 1:0 0 0 0 0 0:894427 0 0 0:707107
B 0 0 0 0 1:0 0 0 0 0:447214 0 0 0:707107 C
B C
B C
B 0 1:0 0 0 0 0 0 0 0 0:894427 0:707107 0 C
B C
B 0 0 0 0 1:0 0 0 0 0 0:447214 0:707107 0 C
C DB
B 0
C:
C
B 0 1:0 1:0 0 0 0:707107 0 0 0 0:707107 0 C
B C
B 0 0 0 0 0 1:0 0:707107 0 0 0 0:707107 0 C
B C
@ 1:0 1:0 0 0 0 0 0 0:707107 0 0 0 0:707107 A
0 0 0 0 0 1:0 0 0:707107 0 0 0 0:707107
(7.80)
cf. (7.30d). Here, w0i is the element of the 1 1 weight matrix Wi 0 , and %i D %i .x/
is defined, cf. (2.9b), by
FiU FiL
%i .x/ D D y Ai .x/; i D 1; : : : ; B: (7.81b)
2
Defining
2
%i .x/
wQ i D wQ i .x/ WD ; i D 1; : : : ; B; (7.81c)
w0i
K.y ; x/ D CK 0 .y ; x/C T
0
wQ 4 C 0:8wQ 9 C 0:5w Q 12 Q 9 C 0:5w
0:4w Q 12 0 0
B 0:4w Q 9 C 0:5w Q 12 Q 5 C 0:2w Q 9 C 0:5w Q 12 w Q5
B w 0
B Q 2 C 0:8wQ 10 C 0:5w Q 11 Q 10 0:5w Q 11
B 0 0 w 0:4w
B
B 0 w Q5 0:4w
Q 10 0:5w Q 11 w Q 5 C 0:2w Q 10 C 0:5w Q 11
DB
B w Q4 0 0:5w Q 11 0:5w Q 11
B
B 0 0 0:5w Q 11 0:5w Q 11
B
@ 0:5w Q 12 0:5w Q 12 w Q2 0
0:5w Q 12 0:5w Q 12 0 0
1
wQ4 0 0:5w Q 12 0:5w Q 12
0 0 0:5w Q 12 0:5w Q5 C
C
0:5w Q 11 Q 11 w Q2 C
0:5w 0 C
C
0:5wQ 11 0:5w Q 11 0 0 C
C:
Q3 C w
w Q 4 C 0:5w Q 7 C 0:5w
Q 11 Q 7 0:5w
0:5w Q 11 0 0 C
C
0:5wQ 7 0:5w Q 11 Q 6 C 0:5w
w Q 7 C 0:5w Q 11 0 w Q6 C
C
0 0 Q1 C w
w Q 2 C 0:5w Q 8 C 0:5wQ 12 0:5w Q 8 C 0:5w Q 12 A
0 w Q6 0:5w Q 8 C 0:5w Q 12 Q 6 C 0:5w
w Q 8 C 0:5w Q 12
(7.82)
In the present case the cost factors i 0 in the primary cost function G0 .x/ D G0 .x/,
cf. (4.1a), are defined by
1 1
i 0 WD D 6; 4 104
V0 mm3
7.7 Numerical Example: 12-Bar Truss 321
a b
4.0 e+07
1 o
8000
1
o
1
3.5 e+07
1
o
cross−sectional areas
6000
1 *
* 3 o
1
total volume
*
3.0 e+07
3
1 * 3 4
* 3 4 o
4000
1 * 3 4
3 4
2.5 e+07
* o
3 4
* 4
3
2000
7 7 7 7 7 7 7 7
4 o
2.0 e+07
4
+
o
5
9
6
2 +
o
5
9
6
2 +
o
5
9
6
2 +
o
5
9
6
2 +
o
5
9
6
2 +
o
5
9
6
2 +
o
5
9
6
2 +
o
5
9
6
2 o
0
Fig. 7.3 Optimal design using (MEC). (a) Optimal cross-sectional areas. (b) Total volume
w0i D 100:
In Fig. 7.3a, b the optimal cross-sectional areas Ai ; i D 1; : : : ; 12, and the total
volume are shown as functions fo the expectation P 1y D EP1y .!/ of the random
force component P1y D P1y .!/. With increasing expected force P 1y , the cross-
sectional areas of bar 1,3,4,8,12 are increasing too, while the other remain constant
or are near zero. The resulting optimal design of the truss can be seen in Fig. 7.4.
Here, bars with cross-sectional areas below Amin D 100 mm2 have been deleted.
322 7 Stochastic Structural Optimization with Quadratic Loss Functions
1.0
1 o
a b c
5 e+07 7 e+07
o
cross−sectional areas
0.5
1 * o
total volume
1 * 3
* 3 o
4
0.0
1 * 3 x x x x x x x x
3 4 o
1 3* 4
1 3* 4 o
−0.5
3* 4
3* 4 7 7 7 7
7
3 e+07
7 7
4 7 o
4
−1.0
+
o
5
9
6
2 +
o
5
9
6
2 +
o
5
9
6
2 +
o
5
9
6
2 +
o
5
9
6
2 +
o
5
9
6
2 +
o
5
9
6
2 +
o
5
9
6
2 o
0
150000 250000 350000 450000 150000 250000 350000 450000 150000 250000 350000 450000
expectation of the load expectation of the load expectation of the load
Fig. 7.5 Optimal design using (ECBO). (a) Optimal cross-sectional areas. (b) Expected minimum
volume G 0 . (c) Probability of failure
In Figs. 7.3a and 7.5a by the symbol “” the almost equal optimal cross-sectional
areas of bar 8 and 12 are marked.
The probability of failure of an (MEC)-optimal truss is always zero showing also
the robustness of the optimal 6-bar truss according to Fig. 7.4.
The related numerical results obtained for the expected cost based optimization
problem (ECBO) are presented in Fig. 7.5a–c. Here, the optimal cross-sectional
areas, the expected minimum volume and the related probability of failure is
represented again as functions of the expected form P 1y . The resulting optimal
design is the same as in (MEC), where in this case the probability of failure is also
zero, which confirms again the robustness of this optimal design.
Chapter 8
Maximum Entropy Techniques
we consider, cf. Sect. 1.2, the expectation or mean value minimization problem
where
Z
v.; x/ D v.P!Q ; x/ WD Ev.!;
Q x/ D v.!; x/.d!/: (8.2b)
Here, “E” denotes the expectation operator, and the probability distribution on
the measurable space .˝; A/ denotes the true probability distribution P!Q WD
of the random element !. We may assume that lies in a certain given set
of probability measures on A (a priori information about ). In the following we
suppose that all integrals, expectations, probabilities, resp., under consideration
exist and are finite.
and take an "-optimal decision x 2 D" .ˇ/ with an appropriate bound " > 0.
Here, the set D" .ˇ/ of "-optimal decisions is defined by
Of course,
Remark 8.1 Hypothesis-finding in case that there is some a priori information, but
no sample information ! N D .!1 ; : : : ; !N / of !:
Q
In the above case a so-called e-estimate ˇ of P!Q can be applied which is defined
as follows:
Definition 8.1 Let e W ! R denote a function on the product of the
Q
given set of probability measures—containing the true distribution of !—such
that e.; / can be considered as a measure for the error selecting the distribution
, while P!Q D is the true distribution. An eestimate of P!Q is then a distribution
ˇ 2 such that
denotes therefore the maximum expected loss if P!Q D is the true distribution of
Q and the decision maker uses a certain ˇ-optimal decision. Because of v.; x/
!,
v ./; x 2 D, cf. (8.5), we have
and
is the maximum error relative to the decision making problem .˝; D; v/, if any ˇ-
optimal decision is applied, while P!Q D is the true distribution of !.
Q Obviously,
Based on the loss set V , the functions H D H" .; ˇ/ and v ./ can be
represented also as follows:
Z
H" .; ˇ/ D supf v.!/.d!/ W v 2 V" .ˇ/g; (8.13c)
Z
v ./ WD inf f v.!/.d!/ W v 2 V g: (8.13d)
Remark 8.2 According to the above assumptions, the loss set V lies in the
space L1 .˝; A; / of all -integrable functions f D f .! for each probability
distribution D ; D ˇ under consideration.
Remark 8.3 Identifying the decision vector x 2 D with the related loss function
v D v.!; x/, the set D of decisions can be identified with the related loss set V .
Hence, we can consider the loss set V as the generalized admissible set of decision
or design vectors. On the other hand, the optimal decision problem under stochastic
uncertainty can also be described by the set ˝ of elementary events and a certain set
V of measurable real functions f D f .!/ on ˝ playing the role of loss functions
related to the decision vector x f ./.
We have then the following properties:
Lemma 8.1 Suppose that D" .ˇ/ 6D ; for all " > 0:
I)" ! H" .; ˇ/ is monotonous increasing on .0; C1/I
II)H" .; ˇ/ v ./; " > 0I
III)If the loss set V is convex, then " ! H" .; ˇ/ is concave;
IV) If v.ˇ; x/ v .ˇ/ C "N for all x 2 D and a fixed "N > 0, then H" .; ˇ/ D
supfv.; x/ W x 2 Dg; " "NI
V) The assertions (1)–(4) hold also for " 0, provided that D0 .ˇ/ 6D ;.
Proof Because of D" .ˇ/ 6D ;; " > 0, the maximum expected loss H" .; ˇ/ is
defined for all " > 0 . I) The monotonicity of " ! H" .; ˇ/ follows from D" .ˇ/
Dı .ˇ/, if " < ı: II) The inequality v.; x/ v ./; x 2 D; yields H" .; ˇ/ D
supf.; x/ W x 2 D" .ˇ/g v ./. III) Let be 1 > 0; 2 > 0; 0 ˛ 1
and x1 2 D"1 .ˇ/; x2 2 D"2 .ˇ/: Because of the convexity of the loss set V , there
exists x3 2 D; such that v.; x3 / D ˛v.; x1 / C .1 ˛/v.; x2 /. This yields then
v.ˇ; x3 / D ˛v.ˇ; x1 / C .1 ˛/v.ˇ; x2 / ˛.v .ˇ/ C "1 / C .1 ˛/.v .ˇ/ C
"2 / D v .ˇ/ C "N with "N D ˛"1 C .1 ˛/"2 . Hence, x3 2 D"N .ˇ/ and therefore
H"N .; ˇ/ v.; x3 / D ˛v.; x1 / C .1 ˛/v.; x2 /. Since x1 ; x2 were chosen
arbitrarily, we get now H"N .; ˇ/ ˛H"1 .; ˇ/ C .1 ˛/H"2 .; ˇ/. The rest of the
assertion is clear.
Remark 8.4 According to Lemma 8.1(V) we have H" .; ˇ/ H0 .; ˇ/; " 0,
provided that D0 .ˇ/ 6D ;:
328 8 Maximum Entropy Techniques
and
As shown by the following examples, there are families .xˇ" /">0 of "-optimal
decisions xˇ" with respect to ˇ; hence, xˇ" 2 D" .ˇ/; " > 0; such that
v.; xˇ" / H0 .; ˇ/ C ı for all 0 < " < "0 ; (8.15)
1 1
V D convf.1; 0/T ; .2; 0/T ; .0; 2/T ; .0; 1/T gnconvf. ; /T ; .0; 1/T g;
2 2
8.1 Uncertainty Functions Based on Decision Problems 329
where “conv” denotes the convex hull of a set. Selecting D .0; 1/T and ˇ D
. 12 ; 12 /T , we get
1
v .ˇ/ D ; H" .; ˇ/ D v.; xˇ" / D 2.v .ˇ/ C "/ D 1 C 2"
2
with xˇ" D .0; 2.v .ˇ/ C "//T D .0; 1 C 2"/T 2 D" .ˇ/ for 0 < " < 12 . On the other
hand, D0 .ˇ/ D convf. 12 ; 12 /T ; .1; 0/T gnf. 21 ; 12 /T g, and therefore H0 .; ˇ/ D 12 :
Hence, (8.15) holds, i.e., H" .; ˇ/ D .; xˇ" / D 1 C 2" > 1 D H0 .; ˇ/ C ı; " > 0,
with ı D 12 .
Remark 8.5 As it turns out later, the instability (8.15) follows from the fact that V
is not closed.
Example 8.2 Let ˝ D f!1 ; !2 g, and suppose that D V is given by V D
f.0; 0/T g [ fz 2 R2C W z1 z2 1g. Moreover, ˇ D .1; 0/T and 2 R2C;1 with 2 > 0.
Then, v .ˇ/ D 0 and D0 .ˇ/ D f.0; 0/T g. Furthermore, H" .; ˇ/ D C1; and
xˇ" D ."; 1" /T 2 D" .ˇ/, where v.; xˇ" / D 1 " C 2 1" ; " > 0. Thus, H0 .; ˇ/ D 0,
and also (8.15) holds:
1
v.; xˇ" / D 1 " C 2 > H0 .; ˇ/ C ı D ı
"
2
for all 0 < " < ı
and each (fixed) ı > 0:
Remark 8.6 Here, V is closed, but it is not convex, which is the reason for the
instability (8.15) in the present case.
A necessary and sufficient condition excluding the instability (8.15) of the two-
step procedure .I; II / procedure:
F" D supfF .x/ W x 2 D" .ˇ/g D maxfF .x/ W x 2 D" .ˇ/g; 0 " "N; (8.16)
where, because of the continuity of F, the maximum is taken. In the second part we
show that F" # F0 for " # 0: Due to the monotonicity of " ! F" and F" F0 ; " >
0, we have lim F" F0 . Moreover, with (8.16), for 0 " "N and each c 2 R it
"#0
holds
4c D f" W 0 " "N; there is x 2 D"N .ˇ/ with v.ˇ; x/ v .ˇ/ C " and F .x/ cg:
(8.17b)
8.2 The Generalized Inaccuracy Function H.; ˇ/ 331
Since D"N .ˇ/ is compact, sequence .x k / has an accumulation point x 0 2 D"N .ˇ/,
and the continuity of x ! F .x/ and x ! v.ˇ; x/ on D"N .ˇ/ yield the existence
of a subsequence .x kj / of .x k /; such that F .X kj / ! F .x 0 / and v.ˇ; x kj / !
v.ˇ; x 0 /; j ! 1: From 8.18 we get then v.ˇ; x 0 / v .ˇ/ C "0 and F .x 0 / c:
Hence, F"0 c and therefore "0 2 4c ;, because we have 0 " ", N since 0
"kj "N: The above considerations yield that 4c is closed for all c 2 R. Assuming
that there is cQ 2 R; such that lim F" cQ > F0 , we have F" cQ > F0 for all " > 0.
"#0
However, this yields then 4cQ D f0 " "N W F" cQ Wg D .0; "N, which contradicts
to the closedness of 4c for each c 2 R shown above. Thus, lim F" D F0 .
"#0
Remark 8.7 According to (8.1–8.3), the second part of the above Theorem 8.1 is
used mainly for F .x/ D v.; x/. In this case we have:
inf fv.; x/ W x 2 D" .ˇ/g " inf fv.; x/ W x 2 D0 .ˇ/g for " # 0: (8.19)
Obviously, the above result can be formulated also with the loss set V in the
following way:
Corollary 8.1 I) Let V" .ˇ/ 6D ;; " > 0 and suppose that there is "N > 0 such
that V"N .ˇ/ is compact and f ! f; f ! ˇf; f 2 V"N .ˇ/ are real valued and
continuous functions on V"N .ˇ/. Then v .ˇ/ 2 R; V0 .ˇ/ 6D ;; and H.; ˇ/ D
H0 .; ˇ/I
II) Replacing f ! f; f 2 V"N .ˇ/ by an arbitrary continuous function F W
V"N .ˇ/ ! R, and keeping the remaining assumptions in (I), then supfF .f / W
f 2 V" .ˇ/g # supfF .f / W f 2 V0 .ˇ/g for " # 0:
Proof The assertion follows immediately from V" .ˇ/ D fv.:; x/ W x 2 D" .ˇ/g:
Let denote P!Q D the true distribution of !,Q and suppose that the hypothesis
“P!Q D ˇ” has been accepted. Moreover, assume that D" .ˇ/ 6D ; for all " > 0; This
holds if and only if v .ˇ/ > 1 or v .ˇ/ D 1 and then v.ˇ; x/ D 1 for
an x 2 D. Using a decision x 2 D" .ˇ/, then we have a loss from fv.; x/ W x 2
D" .ˇ/g, and
denotes the maximum, minimum, resp., expected loss, if the computation of an "-
optimal decision is based on the hypothesis “P!Q D ˇ”, while P!Q D is the true
Q Corresponding to Lemma 8.1 on H" .; ˇ/, we can
probability distribution of !.
show this result:
Lemma 8.3 Suppose that D" .ˇ/ 6D ;; " > 0. Then,
I) " ! h" .; ˇ/, with h" .; ˇ/ D inffv.; x/ W x 2 D" .ˇ/g, is monotonous
decreasing on .0; C1/;
II) h" .; ˇ/ v ./ for all " > 0;
III) " ! h" .; ˇ/; " > 0 is convex, provided that the loss set V is convex;
IV) If v.ˇ; x/ v .ˇ/ C "N for all x 2 D and a fixed "N > 0, then h" .; ˇ/ D
v ./; " > "N;
v) The assertions (a)–(c) hold also for " 0, in case that Do .ˇ/ 6D ;:
Lemmas 8.1 and 8.3 yield then this corollary:
Corollary 8.2 For two numbers "1 ; "2 > 0 ( 0, if Do .ˇ/ 6D ;; resp:) we have
Proof If "1 "2 , then H"2 .; ˇ/ H"1 .; ˇ/ h"1 .; ˇ/, and in case "1 > "2 it
is h"1 .; ˇ/ h"2 .; ˇ/ H"2 .; ˇ/ according to (a) of Lemmas 8.1 and 8.3.
Hence, the limit lim h" .; ˇ/ D sup h" .; ˇ/, exists, and corresponding
"#0 ">0
to (8.14a) we define
For the functions h.; ˇ/; H.; ˇ/ defined by (8.14a) and (8.21) the following
result holds:
Theorem 8.2 I) Let D" .ˇ/ 6D ; for " > 0. Then
Proof I) Lemma 8.3 yields h.; ˇ/ D sup h" .; ˇ/ v ./. Because of (8.20)
">0
we have h" .; ˇ/ H"N .; ˇ/; " > 0 for each fixed "N > 0. From this we obtain
h.; ˇ/ D sup h" .; ˇ/ H"N .; ˇ/; "N > 0, hence, h.; ˇ/ inf H" .; ˇ/ D
">0 ">0
H.; ˇ/:
II) According to Theorem 8.2.1 and the Definition of H.; /, we get v ./
H.; / H" .; / D supfv.; x/ W x 2 D" ./g D supfv.; x/ W v.; x/
8.2 The Generalized Inaccuracy Function H.; ˇ/ 333
v ./ C "g v ./ C " for each " > 0. Thus, H.; / D v ./, and due to the
first part we also have v ./ D h.; /.
For the geometrical interpretation of the values h.; ˇ/; H.; ˇ/ consider now
the transformed loss set
If the loss set V is convex, then the linear transformation V;ˇ D T;ˇ .V /
of V with respect to T;ˇ W f ! .f; ˇf /T is again a convex set. Hence,
v 0 .ˇ/ D inffz2 W z 2 V;ˇ g; which means that v .ˇ/ can be interpreted as the
second coordinate of one of the deepest points of V;ˇ . In the same way, v ./ is
the first coordinate one of the points lying on the left boundary of V;ˇ .
According to Fig. 8.2, the values h.; ˇ/, H.; ˇ/ and also the divergences
I.; ˇ/ WD H.; ˇ/ H.; /, J.; ˇ/ WD h.; ˇ/ h.; / can be interpreted
corresponding to V;ˇ in this way:
V WD Cf ;
be defined by
Here, “clconv” denotes the closed, convex hull of a set. We still put
P
n
f .1=n/ 2 R, we find vf .ˇ/ ˇk f .1=n/ < C1. In addition, because
kD1
of zk 0; k D 1; 2; : : : ; n for z 2 Cf and with ˇ 0 we get zk ˇk ˇ T z
v .ˇ/ C " for each z 2 V" .ˇ/; hence, 0 zk .1=ˇk /.vf .ˇ/ C "/; provided
that ˇk > 0:
IV) Since ff .˛/ W ˛ 2 riRnC;1 g ff .˛/ W ˛ 2 RnC;1 g and ˛ ! f .˛/ D
.f .˛1 /; : : : ; f .˛n //T ; ˛ 0 is a continuous mapping for real f .0/ , we find
that ff .˛/ W ˛ 2 riRnC;1 g is bounded as a subset of the compact set ff .˛/ W
˛ 2 riRnC;1 g. Due to [137], Theorem 17.2 we obtain then Cf D clconvff .˛/ W
˛ 2 riRnC;1 g D conv.clff .˛/ W ˛ 2 riRnC;1 g/ D convff .˛/I ˛ 2 RnC;1 gI
indeed, if f .˛ / ! z; ! 1 with ˛ 2 riRnC;1 ; then, due to the compactness
of RnC;1 we have a subsequence .˛ j / of .˛ /, such that ˛ j ! ˛ 2 RnC;1 ; j !
1: Because of f .˛ j / ! z; j ! 1 and the continuity of f , we get then
f .˛/ D z; hence, z 2 ff .˛/ W ˛ 2 RnC;1 g and therefore, as asserted, clff .˛/ W
˛ 2 riRnC;1 g D ff .˛/ W ˛ 2 RnC;1 g: Being the convex hull of a compact
P
set, Cf is also a compact set. For z 2 Cf we have z D i f .˛ i /; i
i D1
P
P
0; ˛ 2 i
RnC;1 ; i D 1; 2; : : : ; and D 1: Thus, zk D
i
i f .˛ki /
i D1 i D1
P
P
f. i ˛ki / D f .˛k / with ˛ D i ˛ i 2 RnC;1 : Hence, we have therefore
i D1 i D1
found an ˛ 2 RnC;1 such that z f .˛/:
V) Due to the representation of an element z 2 Cf stated in part (II), we have zk D
P
n P
n
lim zk ; where zk D i f .˛ki / with ˛ i 2 riRnC;1 and i 0; i D
!1 i D1 i D1
1: As above, for each D 1; 2; : : : we have the relation z f .˛ /; with ˛ D
Pn
i ˛ i : However, the sequence .˛ / has an accumulation point ˛ in RnC;1 I
i D1
We show now that ˛ 2 riRnC;1 : Assuming that ˛k D 0 for an index 1 k n,
with a sequence ˛ j ! ˛; j ! 1 we get the relation ˛kj ! ˛k D 0 and
therefore f .˛kj / ! f(0)D C1; j ! 1: However, this is not possible, since
j j
zk f .˛k / and .z / is a convergent sequence. Thus, ˛k > 0. Furthermore,
from zkj f .˛kj /; j D 1; 2; : : : ; zkj ! zk ; ˛kj ! ˛k ; j ! 1 we finally
obtain zk f .˛k /, hence, z f .˛/ with an ˛ 2 riRnC;1 :
The above lemma yields now several consequences on H" .; ˇ/; h" .; ˇ/ W
Corollary 8.3 For each 2 RnC;1 the value vf ./ has the representation
For H.; ˇ/ D H .f / .; ˇ/ and h.; ˇ/ D h.f / .; ˇ/, with V D Cf , from
Lemma 8.4 we get this result:
.f /
Corollary 8.4 I) If f .0/ 2 R, then H .f / .; ˇ/ D H0 .; ˇ/ and h.f / .; ˇ/ D
h0 .; ˇ/ D inffT f .˛/ W ˇ T f .˛/ D vf .ˇ/g for all ; ˇ 2 RnC;1 :
.f /
X
n X
n
vf .ˇ/ D inff ˇk f .˛k / W ˛k > 0; k 6D ; ˛k D 1g; (8.25)
kD k6D
and in the case f .0/ 2 R the inequality “˛k > 0” may be replaced also by “˛k
0”. If f .0/ D C1 and in addition f is strictly monotonous decreasing on Œ0; 1,
then V0 .ˇ/ D ;.
Indicating the dependence of the set Cf , cf. (8.23), as well as the values vf .ˇ/
.n/ .n/
on the index n (= number of elements of ˝) by means of Cf , vf .ˇ/, resp.,
.n/ .n/
then vf .ˇ/ D inffˇ z W z 2
T
Cf g.
Moreover, for a ˇ 2 RnC;1
with ˇk D 0,
O
1 k n, and using the notation ˇ D .ˇ1 ; : : : ; ˇk1 ; ˇkC1 ; : : : ; ˇn /T , Eqs. (8.24)
and (8.25) yield
vf
.n/
.ˇ/ D vf
.n1/ O
.ˇ/: (8.26)
A relationship between vf ./ and vf .e/; e D .1=n; : : : ; 1=n/T is shown in the
following:
Corollary 8.8 For all 2 RC;1
n
we have vf ./ f .1=n/ vf .1=n; : : : ; 1=n/:
Proof Equation (8.24) in Corollary 8.3 yields vf ./ T f .e/ D f .1=n/: Select
then an arbitrary " > 0. According to (8.24) there exists an ˛ " 2 riRnC;1 , such that
P
n P
n
v .e/ C " e T f .˛ " /: Hence, v .e/ " C f .˛k" / " C f . .1=n/˛k" / D
kD1 kD1
" C f .1=n/: Since " > 0 was chosen arbitrarily, we have vf .e/ f .1=n/ and
therefore vf .e/ f .1=n/ vf ./:
Note that the assertion in this corollary can also be found in [3]. Having some
properties of H .n/ and h.n/ , we determine now these functions for some important
special cases of f . Next to we consider, see Fig. 8.3, the family .fb /b0 , defined by
(cf. [3])
1
1b .1 t 1b / ; t 0 for b > 0; b 6D 1
fb .t/ D
log t ; t 0 for b D 1:
0
0 1 2 3 4
Cf 1 Cf0 Cf1 Cf2
2
1 X
n
k 1=b 1b
v.b/ ./ D k .1 . n / / for b > 0; b 6D 1
1b P 1=b
kD1 k
kD1
(8.27b)
X
n
v.1/ ./ D k log.1=k / (8.27c)
kD1
./ D O T fb .˛O .b/ / with O D .1 ; : : : ; m /T ; ˛O .b/ D .˛1 ; : : : :; ˛m
.b/ .b/ T
v.b/ /
(8.27d)
8.2 The Generalized Inaccuracy Function H.; ˇ/ 339
and
8
1=b P 1=b
n
ˆ
ˆ k for b > 0; b 6D 1
< k =
.b/
˛k D kD1
ˆ and k D 1; : : : ; m;
:̂
k for b D 1
X
n
V.o/o./ D ffo .˛/ W ˛ 2 RC;1
n
; k ˛k D max k g; 2 RnC;1 ;
1kn
kD1
(8.27e)
.b/
V.b/o ./ D ffb .˛ /g with ˛
.b/ .b/
D .˛1 /T ; 2 riRnC;1 ; b>0 (8.27f)
.b/ 1
V.b/o ./ D fz W zk D fb .˛k /; k > 0 and zk D ; k D 0g (8.27g)
1b
D ffb .˛Q .b/ /g; 2 RnC;1 nriRnC;1 ; 0 < b < 1; and certain ˛Q .b/
V.b/o ./ D ;; 2 RC;1
n
nriRnC;1 and b 1: (8.27h)
Proof Consider first the case b D 0. From (8.23) we easily find that C.o/ D
clconvff0 .˛/ W ˛ 2 riRnC;1 g D f.1 ˛k /kD1;:::;n W ˛ 2 RnC;1 g: Because
of (8.24) we further have v.o/ ./ D inff1 T ˛ W ˛ 2 RnC;1 g D 1 supfT ˛ W
˛ 2 RC;1 g D 1 max k : Hence, V.o/o./ D fz 2 C.o/ W T ˛ D
n
1kn
max k ; ˛ 2 RnC;1 g; which shows the assertion for b D 0. Thus, let now b > 0
lkn
and k > 0; k D 1; : : : ; m; mC1 D : : : D n D 0: From (8.24) and (8.26) we get
.m/ O
D inffO T fb .˛/ O
.n/
then v.b/ ./ D v.b/ ./ O W ˛O 2 riRm
C;1 g with D .1 ; : : : ; m /
T
min O T fb .˛/
O (8.28)
s.t. ˛O 2 riRm
C;1
P m
O u/ D O T fb .˛/
is then given by L.˛; O C u. ˛k 1/. Moreover, the optimality
kD1
conditions—without considering the constraints ˛k > 0; k D 1; : : : ; m—read
@L X
m
0 D D ˛k 1 (8.29)
@u
kD1
@L
0 D D k Df b .˛k / C u
@˛k
340 8 Maximum Entropy Techniques
Inserting Df b .t/ D t b ; t > 0 for b > 0; b 6D 1 and Df 1 .t/ D 1=t; t > 0 for
b D 1 into (8.29), yields
8
1=b P 1=b
m
ˆ
ˆ k for b > 0; b 6D 1
< k =
.b/ kD1
˛k D ˛k D
ˆ and k D 1; 2; : : : ; m:
:̂
k for b D 1
.b/
Since ˛k > 0; k D 1; : : : ; m, also the conditions ˛k > 0; k D 1; : : : ; m, hold.
.b/ .b/
Moreover, ˛O .b/ D .˛1 ; : : : ; ˛m /T is the unique solution of (8.28), since ˛ !
.n/ .m/ OT
T f .˛/ is strictly convex on riRm C;1 . Hence, v.b/ ./ D v.b/ ./ D fb .˛O .b/ /:
Using the convention 0 .C1/ D 0, yields the rest of (8.27b). In addition, this
proves also part (II). For showing (III) and therefore also (I), let 2 riRnC;1 and
z 2 V.b/o ./; b > 0: We remember that fb .0/ D C1 for b 1 and fb .0/ 2 R for
.b/ .b/ .n/
0 b < 1: Part (II) yields then ˛ .b/ D .˛1 ; : : : ; ˛n /T 2 riRnC;1 and v.b/ ./ D
T fb .˛ .b/ /, hence, fb .˛ .b/ / 2 V.b/o ./ .6D ;/:
According to Lemma 8.4, for 0 < b < 1, b 1, resp., there is ˛ 2 RnC;1 ,
˛ 2 riRnC;1 ;, resp., such that z fb .˛/ (to each z 2 V.b/o .//: This immediately
.n/
yields v.b/ ./ D T z D T fb .˛/; since fb .˛/ 2 C.b/ : Thus, z D fb .˛/; because
of k > 0; k D 1; : : : ; n: Assuming ˛ 6D ˛ .b/ ; for ˛Q D 12 ˛ C 12 ˛ .b/ we get on the
one hand ˛Q 2 riRnC;1 , hence, fb .˛/
Q 2 C.b/ ; b > 0, and on the other hand we have
.n/ .n/
v.b/ ./ T fb .˛/
Q < 12 T fb .˛/ C 12 T fb .˛ .b/ / D v.b/ ./; since fb is strictly
convex for b > 0. Consequently, ˛ D ˛ .b/ and therefore V.b/o ./ D ffb .˛ .b/ /g; b >
0; 2 riRnC;1 : Now consider 1 > 0; : : : ; m > 0; mC1 D : : : D n D 0:
.n/ .m/ O
Again from part (II) we get v ./ D v .b/ ./ D T fb .˛O .b/ /: We put ˛Q .b/ D
.b/
.b/ .b/
.˛1 ; : : : ; ˛m ; 0; : : : ; 0/T : Let 0 < b < 1. Because of f .0/ D 1=.1b/ 2 R, due to
Lemma 8.4 we obtain C.b/ D convffb .˛/ W ˛ 2 RnC;1 g, hence, fb .˛Q .b/ / 2 V.b/o ./;
since ˛Q .b/ 2 RnC;1 and T fb .˛Q .b/ / D O T fb .˛O .b/ / D v.b/ ./:
.n/
Consider now z 2 V.b/o ./: According to Lemma 8.4, part V , there is ˛ 2 RnC;1
with z fb .˛/: Because of O T zO D T z D v.b/ ./ D T fb .˛/ D O T fb .˛/
.n/
O
we have fb .˛/ 2 V.b/o ./ and zO D fb .˛/; O since k > 0; k D 1; 2; : : : ; m:
Assuming ˛O 6D ˛O .b/ and considering then under this assumption D 12 ˛ C 12 ˛Q .b/ ;
we get fb . / 2 C.b/ ; since 2 RnC;1 and v.b/ ./ T fb . / D O T fb .O / <
.n/
1 OT
O C 1 O T fb .˛O .b/ / D v ./: However, this yields a contradiction, hence,
.n/
2
fb .˛/ 2 .b/
it holds ˛O .b/ D ˛.
O Obviously, each ˛ 2 RnC;1 fulfilling this equation is contained in
V.b/o ./.
Thus, V.b/o ./ D ffb .˛/ W ˛ 2 RnC;1 ; ˛O D ˛O .b/ g D ffb .˛ .b/ /g D
P
m
.b/
f.fb .˛ .b/ /T ; 1=.1 b/; : : : ; 1=.1 b//T g; since ˛k D 1 and therefore
kD1
˛k D 0; k D m C 1; : : : ; n: Finally, let b 1: Due to fb .˛O .b/ / D
8.2 The Generalized Inaccuracy Function H.; ˇ/ 341
.fb .˛O .b/ /T ; C1; : : : ; C1/T and C.b/ Rn , we find fb .˛O .b/ / 62 C.b/ : Suppose
that z lies in V.b/o ./: According to Lemma 8.4 .f .0/ D C1/ there is then
an ˛ 2 riRnC;1 with z f .˛/ and therefore v.b/ ./ D T z D O T zO D
.n/
Pm
T fb .˛/ D O T fb .˛/:O Because of ˛ 2 riRnC;1 , it is ˛k < 1 and therefore
kD1
˛O 6D ˛O .b/ : On the other hand we have D 12 ˛ C 12 ˛Q .b/ 2 riRnC;1 : This yields
v.b/ ./ T fb . / D O T fb .O / < 12 O 0 fb .˛/
O C 12 O T fb .˛/
.n/ .n/
O .b/ / D v.b/ ./; which is
again a contradiction. Consequently, V.b/o ./ D ; for b 1 and 62 riRnC;1 :
P
n
Remark 8.8 I) Obviously, v.1/ ./ D k log 1k is the (Shannon-) entropy of
kD1
the discrete distribution .
II) Assume that 1 > 0; : : : ; m > 0; mC1 D : : : D n D 0: For b > 0; b 6D 1,
from (8.27b) we get
1 X X
m m
v.b/ ./ D k .1 .k 1=b = k 1=b /1b /
1b
kD1 kD1
1 X
m X
m
D .1 k 1=b =. k 1=b /1b /
1b
kD1 kD1
1 X
m
1 X n
D .1 k 1=b /b / D .1 . k 1=b /b /: (8.30)
1b 1b
kD1 kD1
Hence,
1
v.b/ ./ D .1 M.1=b/ .//;
1b
P
n
II) If 0 < b < 1, then with ˛ .b/ D .˛k 1=b = ˛k 1=b /kD1;:::;n , we get
kD1
P
n
H .b/ .; ˇ/ D h.b/ .; ˇ/ D k fb .˛k .b/ /
kD1
P
n P
n
D 1
1b .1 k .ˇk 1=b = ˇk 1=b /1b /for all; ˇ 2 RnC;1 :
kD1 kD1
III) If b D 1, then
X
n
H .1/ .; ˇ/ D h.1/ .; ˇ/ D k log.1=ˇk /; ; ˇ 2 RnC;1
kD1
is the Kerridge-Inaccuracy for the hypothesis “P! D ˇ”, while P! D is the true
distribution. However, this justifies the notation generalized inaccuracy function for
H.; ˇ/ and h.; ˇ/.
In the following we derive a representation of H" .; ˇ/ and H.; ˇ/ which can be
used also to find sufficient conditions for H.; ˇ/ D H0 .; ˇ/. For this we make
the following assumptions on the loss set V , cf. (8.13a, b), of the decision problem
.˝; D; v/ and the probability measure ; ˇ on A:
V is a convex subset of L1 .˝; A; /\L1 .˝; A; ˇ/, where 1 < v .ˇ/ < C1:
8.2 The Generalized Inaccuracy Function H.; ˇ/ 343
and
Z
g" .f / D f .!/ˇ.d!/ .v .ˇ/ C "/; f 2 L1 .˝; A; ˇ/; " 0;
F is an affine, real valued functional on V and g" an affine, real valued functional
on the linear space X D L1 .˝; A; ˇ/: Moreover, it holds
H" .; ˇ/ D min.sup .v.; x/ av.ˇ; x// C a.v .ˇ/ C "/ (8.32)
a0 x2D
Due to v .ˇ/ > 1; we also have D"=2 .ˇ/ 6D ;: Hence, there is f1 2 V such
that ˇf1 v .ˇ/C 2" < v .ˇ/C" and therefore g" .f1 / < 0 for all " > 0: According
to our assumptions, (8.31) has a finite infimum, hence, all assumptions in the
344 8 Maximum Entropy Techniques
Proof Let h denote the right hand side of (8.33). Then, h H.; ˇ/: Put now
Suppose now that D0 .ˇ/ 6D ; and H"N .; ˇ/ < C1 for an "N > 0, hence,
H" .; ˇ/ 2 R for 0 < " < ":N Because of H0 .; ˇ/ D supfF .f / W g0 .f / D 0; f 2
V g, for the consideration of H0 .; ˇ/, in stead of (8.31), we have to consider the
optimization problem
min F .f / (8.34)
s.t. g0 .f / D 0; f 2 V: (8.35)
A Kuhn–Tucker-coefficient related to (8.34) is, cf. e.g. [137], Section 28, a value
a0 ; such that
hence,
max f
bez. ˇf D v .ˇ/; f 2 V
then
As in the preceding section, assume that P!Q D is the true probability distribution
of !Q and let denote P!Q D ˇ a certain hypothesis on the true distribution I For all
" > 0 suppose that D" 6D ;: Selecting a decision x 2 D" , then with respect to the
true distribution , with respect to the hypothesis ˇ, resp., we have the error
resp., where e1 .; x/, e2 .; x/, are defined only if v ./ 2 R, v .ˇ/ 2 R;
respectively.
Evaluating this error still by means of a function , then
e
I;" .; ˇ/ D supf.e.; x// W x 2 D" .ˇ/g; e D e1 ; e2 ;
e
J;" .; ˇ/ D inf f.e.; x// W x 2 D" .ˇ/g; e D e1 ; e2
denotes the maximum, minimum error relative to ; for the computation of an "-
optimal decision based on the hypothesis ˇ, while P!Q D is the true distribution.
Hence, as far as the limits under consideration exist, we define the class of
generalized divergences by
and
We consider now I e ; J e ; e2 for some special cases of the cost function : For this
purpose we set
I) If .t/ D t; t 2 R; then
e1
I;" .; ˇ/ D sup e1 .; x/ D sup .v.; x/ v .// D H" .; ˇ/ v ./
x2D" .ˇ/ x2D" .ˇ/
e2
I;" .; ˇ/ D sup e2 .; x/ D sup .v.; x/ v .ˇ// D H" .; ˇ/ v ./
x2D" .ˇ/ x2D" .ˇ/
e1
J;" .; ˇ/ D inf e1 .; x/ D inf .v.; x/ v .// D h" .; ˇ/ v ./
x2D" .ˇ/ x2D" .ˇ/
e2
J;" .; ˇ/ D inf e2 .; x/ D inf .v.; x/ v .ˇ// D h" .; ˇ/ v ./:
x2D" .ˇ/ x2D" .ˇ/
Furthermore, we have
e2
I;" .; ˇ/ D sup j e2 .; x/ j
2D" .ˇ/
V D Cf ; f D fb ; b 0
II)
1 X X
n n
k ..k /b .ˇk =
1=b 1=b
I .b/ .; ˇ/ D J .b/ .; ˇ/ D ˇk /1b /
1b
kD1 kD1
for b > 0; b 6D 1I
III)
X
n
I .1/ .; ˇ/ D J .1/ .; ˇ/ D k log.k =ˇk /:
kD1
Remark 8.12 Obviously, we see now that I .1/ .; ˇ/ D J .1/ .; ˇ/ is the die
Kullback-divergence between and ˇ, which justifies now the notation generalized
divergence for Ie ; Je .
According to Corollary 8.10 we have I.; ˇ/ D 0 H) J.; ˇ/ D 0; and
I.; ˇ/ D J.; ˇ/ D 0 for ˇ D : However, I.; ˇ/ D 0 or J.; ˇ/ D 0 does not
imply ˇ D in general.
Example 8.3 Putting D .1=n; : : : ; 1=n/T in the above Corollary 8.11a, then for
all ˇ 2 RnC;1 we get
X
n
1
I .0/ .; ˇ/ D 1=n inf f ˛k W ˛ 2 RnC;1 ; ˇ T ˛ D max ˇk g
n 1kn
kD1
D 1=n 1=n D 0;
X
n
1
J .0/ .; ˇ/ D 1=n supf ˛k W ˛ 2 RnC;1 ; ˇ T ˛ D max ˇk g
n 1kn
kD1
D 1=n 1=n D 0:
The representations of H.; ˇ/ and h.; ˇ/ given in the Theorems 8.4 and 8.5
yield the following representation of I,J:
Corollary 8.12 I) If H"N .; ˇ/ < C1 for an "N > 0 and v ./ 2 R; then
UI;l .ˇ/.UJ;l .ˇ/; resp:/ D f 2 caC;1 .˝; A/ W v ./ 2 R; I.; ˇ/ <
resp. J.; ˇ/ < g:
Corollary 8.13 Let I.; ˇ/ D H .b/ .; ˇ/ H .b/ .; /; ; ˇ 2 RnC;1 ; b > 0: Then
UI;l .ˇ/ is convex for all > 0; ˇ 2 RnC;1 and each b > 0I UI;r ./ is convex for all
> 0; 2 RnC;1 and each 1=2 b 1:
Remark 8.15 Generalizations of the Kullback-Divergence
R Pure mathematical generalizations of the Kullback-divergence I .; ˇ/ D
.l/
As was shown in the literature, see e.g. [53, 67, 81], in the information-theoretical
foundation of statistics, the following minimization problem plays a major role:
Here, I .1/ .; ˇ/ denotes the die Kullback-divergence between a given probability
measure ˇ
m. measure on A/ and 2 C; where C is a certain subset of f 2
caC;1 .˝; A/ W
mg:
In order to find a further relation between the divergences I; J and the squared
Euclidean distance in a Hilbert space H , we replace the divergence I .l/ .; ˇ/ in
(8.39) by k ˇ k2 , where ; ˇ, C , resp., are considered as elements, a subset of
a Hilbert space H , then we obtain the optimization problem
min k ˇ k2 (8.40)
s.t. 2 C:
where < ; ˇ > denotes the scalar product in the Hilbert space H . Putting
d 2 .; ˇ/ Dk ˇ k2 ; then (8.41) is equivalent with
where in (8.41) and in (8.42) the equality sign (and therefore, of course, the theorem
of Pythagoras) holds, if ˇ0 lies in the relative algebraic interior of C . As was shown
in [35], the optimization problem (8.39) can also be interpreted as a (generalized)
projection problem, sind a solution ˇ0 (8.39) can be characterized by the condition
analogous to (8.42).
We show now that a corresponding result can be obtained also for the minimiza-
tion problems
and
where I.; ˇ/; J.; ˇ/ denote the divergences according to (8.38a), (8.38b).
Let denote a convex subset of caC;1 ; such that H.; ˇ/; h.; ˇ/ are defined
and H.; ˇ/ 2 R; h.; ˇ/ 2 R; v ./ 2 R for all ; ˇ 2 : Moreover, let be ˇ a
fixed element of and C a subset of : Now, a “projection” of ˇ onto C is defined
as follows:
Definition 8.6 A solution ˇ0 of (8.44), (8.45), resp., is called an I -, an J -
projection, resp., of ˇ onto C .
Some properties of I -, J -projections are given in the following:
Theorem 8.8 Suppose that C is convex, and ! H.; 0 /, ! h.; 0 /, resp.,
is affine linear on for 0 D ˇ and all 0 2 C: Moreover, assume that for all
; 0 2 C the continuity condition H.; 0 C t. 0 // ! H.; 0 /, h.; 0 C
t. 0 // ! h.; 0 / holds for t # 0:
I) A necessary condition for an I -, J -projection ˇ0 , resp., of ˇ onto C is then the
condition (analogous to (8.42), (8.43))
resp., where the sign “D” holds, provided that ˇ0 lies in the relative algebraic
interior of C .
II) If limt #0 1t I.ˇ0 ; ˇ0 C t. ˇ0 // D limt #0 1t .H.ˇ0 ; ˇ0 C t. ˇ0 //
H.ˇ0 ; ˇ0 // D 0,
1 1
lim J.ˇ0 ; ˇ0 C t. ˇ0 // D lim .h.ˇ0 ; ˇ0 C t. ˇ0 // h.ˇ0 ; ˇ0 // D 0;
t #0 t t #0 t
resp., for all x 2 C and a ˇo 2 C; then (8.46), (8.47) is also sufficient for an
I -, J -projection ˇ0 of ˇ onto C .
describes the information available on P!Q D . For a given hypothesis P!Q D ˇ the
I-projection of ˇ onto C./ describes then the nearest element of C D C./ to ˇ,
and
˚
denotes the distance between ˇ and C./ (often identified with ). Hence, an
increasing distance I.?; ˇ/ between ˇ and C./ means a decreasing quality of the
hypothesis P!Q D ˇ:
Corresponding to [81] we introduce therefore the following notion:
Definition 8.7 The value I.?; ˇ/ D I.?; ˇI / is called the minimum useful
discrimination information—relative to the decision problem .˝; D; v/—against
the (zero-)hypothesis P!Q D ˇ.
Remark 8.16 Useful Discrimination-Information
The notion useful discrimination information emphasizes the fact that the
generalized divergence I.; ˇ/ measures the difference between the distributions
and ˇ relative to a (subsequent) decision problem .˝; D; v/; see also the definition
of economic information measures used in [53, 88].
We show now some properties of the function ! I.?; ˇI /:
Lemma 8.5 Let be convex,
a linear parameter space and ! R I.; ˇ/ convex
on : Furthermore, let
0 D f 2
W there is 2 ; such that T .!/.d!/ D
g: Then
0 is convex, and ! I.?; ˇI / is convex on
0 :
8.3 Generalized Divergence and Generalized Minimum Discrimination. . . 355
Proof Let R1 ; 2 2
0 and 0 < ˛ < 1: Then there R are elements 1 ; 2 2 ; such
that
R i D T .!/ i .d!/; i RD 1; 2: This yields T .!/.˛1 C .1 ˛/2 /.d!/ D
˛ T .!/1 .d!/ C .1 ˛/ T .!/2 .d!/ and ˛1 C .1 ˛/2 2 , hence, ˛1 C
.1 ˛/2 2
0 : Furthermore,
which yields the rest of the assertion, since, up to the above conditions, 1 ; 2 were
arbitrary.
Remark 8.17 Convexity of I.:; ˇ/: Because of the concavity of ! v ./ D
H.; /, the function ! I.; ˇ/ is convex, provided that H.; ˇ/ D H0 .; ˇ/ D
supff W f 2 V0 .ˇ/g:
If
is a finite dimensional space, then the convexity of ! I.?; ˇI / on
0
yields the continuity of this function—at least—on the relative interior ri
0 of
0 : If
the function ! I.?; ˇI / is continuous on a sufficiently large range of definition,
then
provided that ON ! a.s., where O1 ; O2 ; : : : is a sequence of estimation functions
for . In this case we interpret then
as an estimate of I.?; ˇ/: For the testing of hypotheses we have then, cf. [81] the
following procedure:
Definition 8.8 Reject the (null-) hypothesis P!Q D ˇ; if IO.?; ˇ/ is significantly
large.
For illustration of this test procedure we give still the following example:
Example 8.5 Let ˝ D
D D D R and v.!; x/ D .a.!/x b.!//2 ; where
x 2 R and a./; b./ are square integrable random variables. Then,
ˇ
ˇ 2
I.; ˇ/ D a2 ab =a2 ab =a2 ;
R
where a2 D a.!/2 .d!/ etc. If now T .!/ D a.!/b.!/; then I.?; ˇ/ D
n ˇ ˇ 2 R
o ˇ ˇ
inf a2 ab =a2 ab =a2 W a.!/b.!/.d!/ D D inf u.=uab =a2 /2 ;
u>0
ˇ ˇ
hence, I.?; ˇ/ D 0; provided that sign D sign.ab / and I.?; ˇ/ D 4 j ab j; if
ˇ
sign D sign.ab /:
References
1. Allgöwer F et al (eds) (2000) Nonlinear model predictive control. Birkhäuser Verlag, Basel
2. Aoki M (1967) Optimization of stochastic systems - topics in discrete-time systems.
Academic, New York
3. Arimoto S (1971) Information-theoretical considerations on estimation problems. Inf Control
19:181–194
4. Arimoto S (1996) Control theory of non-linear mechanical systems. Clarendon Press, Oxford
5. Aström KJ (1970) Introduction to stochastic control theory. Academic, New York
6. Aström KJ, Wittenmark B (1995) Adaptive control. Addison-Wesley, Reading
7. Augusti G, Baratta A, Casciati F (1984) Probabilistic methods in structural engineering.
Chapman and Hall, London
8. Aurnhammer A, Marti K (2000) OSTP/fortran-program for optimal stochastic trajectory
planning of robots. UniBw München, Munich
9. Aurnhammer A (2004) Optimale Stochastische Trajektorienplanung und Regelung von
Industrierobotern. Fortschrittberichte VDI, Reihe 8, Nr. 1032. VDI-Verlag GmbH, Düsseldorf
10. Barucha-Reid AT (1972) Random integral equations. Academic, New York
11. Bastian G et al (1996) Theory of robot control. Springer, Berlin
12. Bauer H (1968) Wahrscheinlichkeitstheorie und Grundzüge der Masstheorie. Walter de
Gruyter, Berlin
13. Bauer H (1996) Probability theory. De Gruyter, Berlin
14. Berger J (1985) Statistical decision theory and Bayesian analysis. Springer, New York
15. Bernhardt R, Albright SL (1993) Robot calibration. Chapman and Hall, London
16. Bertsekas DP (1973) Stochastic optimization problems with nondifferentiable cost
functionals. J Optim Theory Appl 12:218–231
17. Biles WE, Swain JJ (1979) Mathematical programming and the optimization of computer
simulations. Math Program Stud 11:189–207
18. Block C (2008) Aktive Minderung personeninduzierter Schwingungen an weit gespan-
nten Strukturen im Bauwesen. In: Fortschrittberichte VDI, Reihe 11, Schwingungstechnik,
No. 336. VDI-Verlag GmbH, Düsseldorf
19. Bobrow JE, Dubowsky S, Gibson JS (1985) Time-optimal control of robotic manipulators
along specified paths. Int J Robot Res 4(3):3–17
20. Bobrow JE (1988) Optimal robot path planning using the minimum-time criterion. IEEE J
Robot Autom 4(4):443–450
21. Bohlin R (1999) Motion planning for industrial robots. Chalmers University, Goteborg
22. Box GE, Draper NR (1987) Empirical model-building and response surfaces. Wiley,
New York
23. Box GEP, Wilson KG (1951) On the experimental attainment of optimum conditions. J R Stat
Soc 13:1–45
24. Bucher CG, Bourgund U (1990) A fast and efficient response surface approach for structural
reliability problems. Struct Saf 7:57–66
25. Bullen PS (2003) Handbook of means and their inequalities. Kluwer Academic, Dordrecht
26. Bunke H (1972) Gewöhnliche Differentialgleichungen mit zufälligen Parametern. Akademie-
Verlag, Berlin
27. Chankong V, Haimes YY (1983) Multiobjective decision making. Oxford, North Holland
28. Chen Y-Ch (1991) Solving robot trajectory planning problems with uniform cubic B-splines.
Optim Control Appl Methods 12:247–262
29. Carathéodory C (1918) Vorlesungen über reelle Funktionen. Teubner, Leipzig
30. Cocchetti G, Maier G (1988) Static shakedown theorems in piecewise linearized plasticity.
Arch Appl Mech 68:651–661
31. Corradi L, Zavelani A (1974) A linear programming approach to shakedown analysis of
structures. Comput Methods Appl Mech Eng 3,37–53
32. Craig JJ (1988) Adaptive Control of Mechanical Manipulators. Addison-Wesley, Reading
33. Csiszar I (1967) Information-type measures of difference of probability distributions and
indirect observations. Stud Sci Math Hung 2:299–318
34. Csiszar I (1967) On topological properties of f-divergences. Stud Sci Math Hung 2:329–339
35. Csiszar I (1975) I-divergence geometry of probability distributions and minimization prob-
lems. Ann Probab 3:146–158
36. Damkilde L, Krenk S (1997) Limits - a system for limit state analysis and optimal material
layout. Comput Struct 64(1–4):709–718
37. Dieudonné J (1969) Foundations of modern analysis. Academic, New York
38. Ditlevsen O, Madsen HO (1996) Structural reliability methods., Wiley, New York
39. Doltsinis I (1999) Elements of plasticity. WIT Press, Southampton
40. Dreyfus SE (1964) Some types of optimal control of stochastic systems. J SIAM Ser A
Control 2(1):120–134
41. Dreyfus SE (1965) Dynamic programming and the calculus of variations. Academic,
New York
42. Dreyfus SE, Law AM (1977) The art of dynamic programming. Academic, New York
43. Dullerud GE, Paganini F (2000) A course in robust control theory. Springer, New York
44. Ermoliev Yu, Wets R (eds) (1988) Numerical techniques for stochastic optimization. Springer
series in computational mathematics, vol 10. Springer, Berlin
45. Findeisen R, Raff T, Allgöwer F (2007) Sampled-data nonlinear model predictive control
for constrained continuous time systems. In: Tarbouriech S et al (eds) Adaptive strategies in
control systems with input and output constraints. Springer, Berlin, pp 207–235
46. Frangopol DM (1995) Reliability-based optimum structural design. In: Sundarajan C (ed)
Probabilistic structural mechanics handbook, Chapman and Hall, New York, pp 352–387
47. Gänssler P, Stute W (1977) Wahrscheinlichkeitstheorie. Springer, Berlin
48. Galambos J, Simonelli I (1996) Bonferroni-type inequalities with applications. Springer,
New York
49. Garcia CE (1989) Model predictive control: theory and Practice – a survey. Automatica
25:335–349
50. Gasser M, Schuëller GI (1998) Some basic principles in reliability-based optimisation (RBO)
of structures and mechanical components. In: Marti K, Kall P (eds) Stochastic programming
methods and technical applications. Lecture notes in economics and mathematical systems,
vol 458. Springer, Berlin, pp 80–103
51. Gessing R, Jacobs LR (1984) On the equivalence between optimal stochastic control and
open-loop feedback control. Int J Control 40(1):193–200
52. Gessner P, Spreman K (1972) Optimierung in Funktionenräumen. LNMES, vol 64. Springer,
Berlin
53. Good IJ (1969) What is the use of a distribution? In: Krishnaiah PR (ed) Multivariate
analysis. Proceedings of the 2nd international symposium on multivariate analysis. Academic,
New York, pp 183–203
References 359
54. Haftka RT, Gürdal Z, Kamat MP (1990) Elements of structural optimization. Kluwer,
Dordrecht
55. Han R, Reddy BD (1999) Plasticity-mathematical theory and numerical analysis. Springer,
New York
56. Hardy GH, Littlewood JE, Polya G (1964) Inequalities. University Press, Cambridge
57. Hardy GH, Littlewood JE, Pólya G (1973) Inequalities. Cambridge University, London
58. Haubach-Lippmann C (1998) Stochastische Strukturoptimierung flexibler Roboter.
Fortschritt-berichte VDI, Reihe 8, Nr. 706. VDI Verlag, Düsseldorf
59. Hodge PG (1959) Plastic analysis of structurals. McGraw-Hill, New York
60. Holmes RB (1972) A course on optimization and best approximation. Lecture notes in
mathematics, vol 257. Springer, Berlin
61. Holtgrewe D (1996) Adaptive regelung flexibler roboter. Igel Verlag, Paderborn
62. Hoppen P (1992) Autonome mobile roboter. B.I. Wissenschaftsverlag, Mannheim
63. Howard RA (1971) Dynamic probabilistic systems. Wiley, New York
64. Hwang YK, Ahuja N (1992) Gross motion planning - a survey. ACM Comput Surv
24(3):219–291
65. Isermann R (1988) Identifikation dynamischer Systeme. Springer, Berlin
66. Jacobsen ShH, Schruben Lee W (1989) Techniques for simulation response optimization,
Oper Res Lett 8:1–9
67. Jaynes ET (1968) Prior probabilities. IEEE Trans Syst Sci Cybern 4:227–241
68. Johanni R (1988) Optimale Bahnplanung bei Industrierobotern. Fortschrittberichte VDI,
Reihe 18, Nr. 51. VDI-Verlag, Düsseldorf
69. Kall P (1974) Approximations to stochastic programs with complete fixed recourse. Numer
Math 22:333–339
70. Kall P (1976) Stochastic linear programming. Econometrics and operations research, vol XXI.
Springer, Berlin
71. Kall P, Wallace SW (1994) Stochastic programming. Wiley, Chichester
72. Kaliszky S (1989) Plasticity: theory and engineering applications. Elsevier, Amsterdam
73. Kalman RE, Falb PL, Arbib MA (1969) Topics in mathematical system theory. McGraw-Hill,
New York
74. Kemenjarzh JA (1996) Limit analysis of solids and structures. CRC, Boca Raton
75. Kerridge DF (1961) Inaccuracy and inference. J R Stat Soc Ser B 23:184–194
76. Khalil HK (1992) Nonlinear systems. MacMillan, New York
77. Kirsch U (1993) Structural optimization. Springer, Berlin
78. Kleijnen JPC (1987) Statistical tools for simulation practitioners. Marcel Decker, New York
79. König JA (1987) Shakedown of elastic-plastic structures. Elsevier, Amsterdam
80. Ku R, Athans M (1973) On the adaptive control of linear systems using the open-loop
feedback optimal approach. IEEE Trans Autom Control 18:489–493
81. Kullback S (1959) Information theory and statistics. Wiley, New York
82. Lawo M (1987) Optimierung im konstruktiven Ingenieurbau. Vieweg, Braunschweig/
Wiesbaden
83. Lawo M, Thierauf G (1980) Stabtragwerke, Matrizenmethoden der Statik und Dynamik.
Teil I: Statik. F. Vieweg, Braunschweig
84. Lin K (1997) Zur adaptiven und langzeitprädikativen Regelung mit einem statistischen
Optimierungskriterium. Dissertation, TU Hamburg-Harburg
85. Luenberger DG (1969) Optimization by vector space methods. Wiley, New York
86. Luenberger DG (1973) Introduction to linear and nonlinear programming. Addison-Wesley,
Reading
87. Madsen HO, Krenk S, Lind NC (1986, 2006) Methods of structural safety. Dover, New York
88. Marschak J (1971) Economics of information systems. J Am Stat Assoc 66:192–219
89. Marti K (1975) Convex approximations of stochastic optimization problems. Meth Oper Res
20:66–76
360 References
112. Marti K, Stein I (2012) Computing stochastic optimal feedback controls using an iterative
solution of the hamiltonian system. In: Topping BHV, Ivanyi P (eds) Proceedings of the 14th
international conference on civil, structural and environmental engineering computing. Civil-
Comp Press, Stirlingshire, Paper 109. http://dx.doi.org/10.4203/ccp.102.109
113. Martino JP (1993) Technological forecasting for decision making, 3rd edn. McGraw-Hill,
New York
114. Miesbach S (1995) Bahnführung von Robotern mit Neuronalen Netzen. Dissertation, TU
München, Fakultät für Mathematik
115. Montgomery CA (1984) Design and analysis of experiments. Wiley, New York
116. Myers RH (1971) Response surface methodology. Allyn and Bacon, Boston
117. Nagarajaiah S, Narasimhan S (2007) Optimal control of structures. In: Arora JS (ed)
Optimization of structural and mechanical systems. World Scientific, Hackensack, pp 221–
244
118. Natanson LP (1977) Theorie der Funktionen einer reellen Veränderlichen. Verlag Harri
Deutsch, Zr̈ich, Frankfurt
119. Neal BG (1965) The plastic methods of structural analysis. Chapman and Hall, London
120. Nielsen MP (1984) Limit analysis and concrete plasticity. Prentice Hall, Englewood Cliffs
121. Øksendal B (2000) Stochastic differential equations. Springer, Berlin
122. Ostrowski A (1959) Über Eigenwerte von Produkten Hermitescher Matrizen. In: Abhandlun-
gen aus dem mathematischen Seminar der Universität Hamburg, vol 23(1). Springer, Berlin,
pp 60–68
123. Pai DK, Leu MC (1991) Uncertainty and compliance of robot manipulators with application
to task feasibility. Int J Robot Res 10(3):200–212
124. Perez A (1967) Information-theoretic risk estimates in statistical decisions, Kybernetika 3:1–
21
125. Pfeiffer F, Johanni R (1987) A concept for manipulator trajectory planning. IEEE J Robot
Autom 3(2):115–123
126. Pfeiffer F, Reithmeier E (1987) Roboterdynamik. Teubner, Stuttgart
127. Pfeiffer R, Richter K (1990) Optimal path planning including forces at the gripper. J Intell
Robot Syst 3:251–258
128. Prager W, Shield RT (1967) A general theory of optimal plastic design. J Appl Mech
34(1):184–186
129. Prekopa A, Szantai T (1978) Flood control reservoir system design using stochastic program-
ming. Math Program Stud 9:138–151
130. Qu S (1995) Optimal Bahnplanung unter Berücksichtigung stochastischer Parameter-
schwankungen. Fortschrittberichte VDI, Reihe 8, Nr. 472. VDI-Verlag, Düsseldorf
131. Qu Z, Dawson DM (1996) Robust tracking control of robot manipulators. IEEE Press,
New York
132. Rackwitz R, Cuntze R (1987) Formulations of reliability-oriented optimization. Eng Optim
11:69–76
133. Redfern DA, Goh CJ (1996) Feedback control of state constrained optimal control problems.
In: Dolezal J, Fidler J (eds) System modelling and optimization. Chapman and Hall, London,
pp 442–449
134. Reif K (1997) Steuerung von nichtlinearen Systemen mit Homotopie-Verfahren. Fortschritt-
berichte VDI, Reihe 8: Mess-, Steuerungs- und Regelungstechnik, Nr. 631. VDI Verlag,
Düsseldorf
135. Richter H (1966) Wahrscheinlichkeitstheorie. Springer, Berlin
136. Richalet J, Rault A, Testud JL, Papon J (1978) Model predictive heuristic control: applications
to industrial processes. Automatica 14:413–428
137. Rockafellar RT (1970) Convex analysis. University Press, Princeton
138. Sawaragi Y, Nakayama H, Tanino T (1985) Theory of multiobjective optimization. Academic,
New York
139. Schacher M (2011) Stochastisch optimale Regelung von Robotern. Fortschritt-Berichte VDI,
Reihe 8, Mess-, Steuerungs- und Regelungstechnik, Nr. 1200. VDI Verlag GmbH, Düsseldorf
362 References
140. Schilling RJ (1990) Fundamentals of robotics. Analysis and control. Prentice Hall, London
141. Schilling K et al (1997) The european development of small planetary mobile vehicle. Space
Technol 17(3/4):151–162
142. Schilling K, Flury W (1990) Autonomy and on-board mission management aspects for the
Cassini Titan probe. Acta Astronaut 21(1):55–68
143. Schlöder JP (1998) Numerische Methoden zur Behandlung hochdimensionaler Aufgaben
der Parameteridentifizierung. Dissertation, Universität Bonn, Math.-Naturwissenschaftliche
Fakultät
144. Schoofs AJG (1987) Experimental design and structural optimization. Doctoral Thesis,
Technical University of Eindhoven
145. Schröer K (1993) Identifikation von Kalibrierungsparametern kinematischer Ketten. Carl
Hanser, München Wien
146. Schuëller GI (1987) A critical appraisal of methods to determine failure probabilities. Struct
Saf 4:293–309
147. Schuëller GI, Gasser M (1998) Some basic principles of reliability-based optimization (RBO)
of structure and mechanical components. In: Marti K, Kall P (eds) Stochastic programming-
numerical techniques and engineering. LNEMS, vol 458. Springer, Berlin, pp 80–103
148. Sciavicco L, Siciliano B (2000) Modeling and control of robot manipulators. Springer,
London
149. Shin KG, McKay ND (1985) Minimum-time control of robotic manipulators with geometric
path constraints. IEEE Trans Autom Control 30:531–541
150. Slotine J-J, Li W (1991) Applied nonlinear control. Prentice-Hall, Englewood Cliffs
151. Smirnov WI (1966) Lehrgang der Höheren Mathematik, Teil IV. Deutscher Verlag der
Wissenschaft, Berlin
152. Smith DL (ed) (1990) Mathematical programming methods in structural plasticity. CISM
courses and lectures, vol 299. Springer, New York
153. Smith DL, Munro J (1976) Plastic analysis and synthesis of frames subject to multiple
loadings. Eng Optim 2:145–157
154. Soong TT (1988) Active structural control in civil engineering. Eng Struct 10:74–84
155. Soong TT (1990) Active structural control: theory and practice. Wiley, New York
156. Soong TT, Costantinou MC (1994) Passive and active structural vibration control in civil
engineering. CISM courses and lectures vol 345. Springer, New York
157. Spencer B, Nagarajaiah S (2003) State of the art of structural control. J Struct Eng ASCE
129(7):845–856
158. Spillers WR (1972) Automated structural analysis: an introduction. Pergamon, New York
159. Stengel RF (1986) Stochastic optimal control: theory and application. Wiley, New York
160. Stöckl G (2003) Optimaler Entwurf elastoplastischer mechanischer Strukturen unter
stochastischer Unsicherheit. Fortschritt-Berichte VDI, Reihe 18, Nr. 278. VDI-Verlag GmbH,
Düsseldorf
161. Stone HW (1987) Kinematic modeling, identification, and control of robotic manipulators.
Kluwer Academic, Boston
162. Theil H (1968) Optimal decision rules for government and industry. North-Holland,
Amsterdam
163. Thoft-Christensen P, Baker MJ (1982) Structural reliability theory and its applications.
Springer, Berlin
164. Tin-Loi F (1990) On the optimal plastic synthesis of frames. Eng Optim 16:91–108
165. Tin-Loi F(1995) Plastic limit analysis of plane frames and grids using GAMS. Comput Struct
54:15–25
166. Tong YL (1980) Probability inequalities in multivariate distributions. Academic, New York
167. Tse E, Athans M (1972) Adaptive stochastic control for a class of linear systems. IEEE Trans
Autom Control 17(1),38–51
168. Walter W (2000) Gewöhnliche Differentialgleichungen. Springer, Berlin
169. Weinmann A (1991) Uncertain models and robust control. Springer, New York
References 363
170. Weinzierl K (1995) Konzepte zur Steuerung und Regelung nichtlinearer Systeme auf der
Basis der Jacobi-Linearisierung. Verlag Shaker, Aachen
171. Yang JN, Soong TT (1988) Recent advance in active control of civil engineering structures.
Probab Eng Mech 3(4):179–188
172. Zier S (2011) Auslegung von Stab- und Balkentragwerken aus ideal elastisch-plastischem
Material unter stochastisch variabler Lasteinwirkung. Fortschritt-Berichte VDI, Reihe 18,
Mechanik/Bruchmechanik, Nr. 331. VDI Verlag GmbH, Düsseldorf
Index
Jacobian, 245
Factor Jensen’s inequality, 18
cost, 1
demand, 1
noise, 1 Lagrangian, 13
of production, 1, 8 Least squares estimation, 24
weight, 6 Limited sensitivity, 21
Failure, 8 Limit state function, 8, 9
approximate expected or recourse cost Linear programming, 8
constraints, 20 Lipschitz-continuous, 27
costs, 10, 12 Loss function, 18
domain, 9
expected costs of, 12
mode, 9 Manufacturing, 2
probability of, 12, 14 Manufacturing errors, 1
of the structure, 21 Markov-type inequality, 34
Failure/survival domains, 28 Material parameters, 1
Function(s) Maximum costs, 6
approximation of expected loss, 24 Mean value function, 24, 27
approximation of performance, 21 Mean value theorem, 19, 27
Index 367
Thickness, 8 Variable, 1, 8, 30, 34, 63, 67, 73, 76, 87, 103,
Total weight, 8 104, 113, 154, 166, 173, 177, 185,
Tschebyscheff-type inequality, 31 253, 256, 273, 281, 283, 286, 287,
Two-step procedure, 2 290, 299, 300, 306, 307, 309, 310,
312, 313, 355
design, 1
Uncertainty, 2 Vector
economic, 10 input, 1
model, 10 nominal, 2
physical, 10 optimization problem, 4
probabilistic, 3 Volume, 8
statistical, 10
stochastic, 3
Unit matrix, 248, 249 Weak functional-efficient, 5
Useful discrimination information, Weight factors, 6
354 Worst case, 6