Optimal Control 1st Edition Richard Vinter (Auth.) 2024 Scribd Download
Optimal Control 1st Edition Richard Vinter (Auth.) 2024 Scribd Download
com
https://ebookgate.com/product/optimal-control-1st-
edition-richard-vinter-auth/
https://ebookgate.com/product/optimal-control-an-
introduction-1st-edition-arturo-locatelli/
https://ebookgate.com/product/optimal-control-3rd-edition-frank-
l-lewis/
https://ebookgate.com/product/counterexamples-in-optimal-control-
theory-semen-ya-serovaiskii/
https://ebookgate.com/product/batch-distillation-simulation-
optimal-design-and-control-second-edition-diwekar/
Dynamic programming and optimal control 3rd Edition
Dimitri P. Bertsekas
https://ebookgate.com/product/dynamic-programming-and-optimal-
control-3rd-edition-dimitri-p-bertsekas/
https://ebookgate.com/product/optimal-control-of-singularly-
perturbed-linear-systems-and-applications-automation-and-control-
engineering-1st-edition-zoran-gajic/
https://ebookgate.com/product/robust-industrial-control-systems-
optimal-design-approach-for-polynomial-systems-1st-edition-
michael-j-grimble/
https://ebookgate.com/product/solution-to-modern-control-
systems-11th-edition-richard-c-dorf/
https://ebookgate.com/product/bleach-vol-29-kubo/
Modern Birkhauser Classics
Optimal Control
Modern Birkhäuser Classics
Richard Vinter
Birkhäuser
Boston • Basel • Berlin
Richard Vinter
Department of Electrical Engineering
Imperial College of Science, Technology, and Medicine
London SW7 2BT
United Kingdom
Originally published in the series Systems & Control: Foundations & Applications
Optimal Control
With 12 Figures
Birkhauser
Boston Basel Berlin
Richard Vinter
Department of Electrical Engineering
Imperial College of Science, Technology,
and Medicine
London SW7 2BT
UK
All rights reserved. This work may not be translated or copied in whole or in part without
the written permission of the publisher (Birkhauser Boston, c/o Springer-Verlag New York,
Inc., 175 Fifth Avenue, Ndw York, NY 10010, USA), except for brief excerpts in connection
with reviews or scholarly analysis. Use in connection with any form of information storage
and retrieval, electronic adaptation, computer software, or by similar or dissimilar methodol-
ogy now known or hereafter developed is forbidden.
The use of general descriptive names, trade names, trademarks, etc., in this publication, even
if the former are not especially identified, is not to be taken as a sign that such names, as
understood by the Trade Marks and Merchandise Marks Act, may accordingly be used freely
by anyone.
ISBN 0-8176-4075-4
ISBN 3-7643-4075-4 SPIN 10748579
What control strategy will transfer a space vehicle from one circular or-
bit to another in minimum time or in such a manner as to minimize fuel
consumption? How should a batch distillation column be operated to max-
imize the yield, subject to specified constraints on purity of the product?
Practical questions such as these underlie the field of Optimal Control.
In the language of mathematics, Optimal Control concerns the properties
of control functions that, when inserted into a differential equation, give
solutions which minimize a "cost" or measure of performance.' In engineer-
ing applications the control function is a control strategy. The differential
equation describes the dynamic response of the mechanism to be controlled,
which, of course, depends on the control strategy employed.
Systematic study of optimal control problems dates from the late 1950s,
when two important advances were made. One was the the Maximum Prin-
ciple, a set of necessary conditions for a control function to be optimal. The
other was Dynamic Programming, a procedure that reduces the search for
an optimal control function to finding the solution to a partial differential
equation (the Hamilton-Jacobi Equation).
1This is a simplification: the field also concerns optimization problems with dynamic
constraints which might be functional differential equations, difference equations, partial
differential equations, or take other forms.
xiv Preface
Nonsmooth functions surfaced once again when efforts were made to ex-
tend the applicability of necessary conditions such as the Maximum Prin-
ciple. A notable feature of the Maximum Principle (and one that distin-
guishes it from necessary conditions derivable using classical techniques),
is that it can take account of pathwise constraints on values of the control
functions. For some practical problems, the constraints on values of the
control depend on the vector state variable. In flight mechanics, for exam-
ple, the maximum and minimum thrust of a jet engine (a control variable)
will depend on the altitude ( a component of the state vector). The Max-
imum Principle in its original form is not in general valid for problems
involving state dependent control constraints. One way to derive necessary
conditions for these problems, and others not covered by the Maximum
Principle, is to reformulate them as generalized problems in the Calculus
of Variations, the cost integrands for which include penalty terms to take
account of the constraints. The reformulation comes at a price, however.
To ensure equivalence with the original problems it is necessary to employ
penalty terms with discontinuous derivatives. So the route to necessary
conditions via generalized problems in the Calculus of Variations can be
followed only if we know how to adapt traditional necessary conditions to
allow for nonsmooth cost integrands.
Two important breakthroughs occurred in the 1970s. One was the end
product of a long quest for effective, local descriptions of "non-smooth"
functions, based on generalizations of the concept of the "subdifferential"
of a convex function, to larger function classes. F. H. Clarke's theory of gen-
eralized gradients, by achieving this goal, launched the field of Nonsmooth
Analysis and provided a bridge to necessary conditions of optimality for
nonsmooth variational problems (and in particular optimal control prob-
lems reformulated as generalized problems in the Calculus of Variations).
The other breakthrough, a somewhat later development, was the concept
of Viscosity Solutions, due to M. G. Crandall and P.-L. Lions, that provides
a framework for proving existence and uniqueness of generalized solutions
to Hamilton-Jacobi equations arising in Optimal Control.
Over the last two decades, viscosity techniques have had a growing fol-
lowing. Applications of Viscosity Methods are now routine in Stochastic
Control, Mathematical Finance, Differential Games and other fields be-
xvi Preface
A subsidiary purpose is to meet the needs of readers with little prior ex-
posure to modern Optimal Control who seek quick answers to the questions:
what are the main results, what were the deficiencies of the "classical" the-
ory and to what extent have they been overcome? Chapter 1 provides, for
their benefit, a lengthy overview, in which analytical details are suppressed
and the emphasis is placed instead on communicating the underlying ideas.
Rockafellar and J. B. Wets [125] and Clarke et al. [53], and the somewhat
older book by J.-P. Aubin and H. Frankowska [12]. It is expected, how-
ever, that readers, whose main interest is in Optimization rather than in
broader application areas of Nonsmooth Analysis, which require additional
techniques, will find these chapters helpful, because of the strong focus on
topics relevant to Optimization.
Optimal Control is now a large field, and our choice of material for in-
clusion in this book is necessarily selective. The techniques used here to
derive necessary conditions of optimality are within a tradition of research
pioneered and developed by Clarke, loffe, Loewen, Mordukhovich, Rockafel-
lar, Vinter, and others, based on perturbation, elimination of constraints
and passage to the limit. The necessary conditions are "state of the art,"
as far as this tradition is concerned. Alternative approaches are not ad-
dressed, such as that of H. Sussmann [132], a synthesis of traditional ideas
for approximating reachable sets and of extensions to the Warga's theory of
derivate containers, which permit a relaxation of hypotheses under which
the Maximum Principle is valid in some respects and leads to different kinds
of necessary conditions for problems in which the dynamic constraint takes
the form of a differential inclusion. The topic of higher order necessary con-
ditions, addressed for example in [155] and [102], is not entered into, nor
are computational aspects, examined for example in [113], discussed.
Above all, I express my profound gratitude to my wife, Donna, for her en-
couragement and support throughout this seemingly endless writing project
and for helping me, finally, bring it to a close.
The motion of the vehicle during the maneuver is governed by the rocket
thrust and by the rocket thrust orientation, both of which can vary with
Minimize - r(tf)
over radial and tangential components of the thrust history,
(T,.(t),Tt(t)), 0 < t < t1, satisfying
r(t) - u,
it (t) = v2(t)/r(t) - p/r2(t) +Tr(t)/m(t),
v(t) _ -u(t)v(t)/r(t) + Tt(t)/m(t),
m(t) _ -('Ymax/Tmax)(T, (t) +T2(t))1I2,
(T, 2(t)
m t2(t))1/2 < Tmax,
m(0) = mo, r(0) = ro, u(0) = 0, v(0) = µ/ro,
u(tf) = 0, v(t f) = p/r(tf).
Figure 1.2 shows a control strategy for the following values of relevant
dimensionless parameters:
Tmax/1?0 = 0.1405,
7ma / a µ/r0 0.07487, 3.32 .
µ/ro ro/
For this strategy, the radius of the terminal circular orbit is
r(tf) = 1.5ro.
The arrows indicate the magnitude and orientation of the thrust at times
t = 0, 0.1tf, 0.2t1,... , t f. As indicated, full thrust is maintained. The
thrust is outward for (approximately) the first half of the maneuver and
inward for the second.
Suppose, for example,,that the attracting center is the Sun, the space
vehicle weighs 10, 000 lb, the initial radius is 1.50 million miles (the radius
of a circle approximating the Earth's orbit), the maximum thrust is 0.85 lb
(i.e., a force equivalent to the gravitational force on a 0.85 lb mass on the
surface of the earth), the maximum rate of fuel consumption is 1.81 lb/day,
and the transit time is 193 days. Corresponding values of the constants are
Then the terminal radius of the orbit is 2.44 million miles. (This is the
radius of a circle approximating the orbit of the planet Mars.)
4 1. Overview
mgx(s) = 1mv2(s)
("loss of potential energy equals gain of kinetic energy"). But, denoting the
time variable as t, we have
1 + jdx(s)/dsI2
v(s) =
dt(s)/ds
J0
Bf
dt = f 1 + Idx(s)/dsl2
v(s)
ds.
SN = { ai0i(x) : a E RN} .
i=1
The widely used finite element methods are modern implementations of
Galerkin's method.
The Action Principle. Let x(t) be the vector of generalized coordinates
of a conservative mechanical system. The Action Principle asserts that x(t)
8 1. Overview
Here T (x, a) is the kinetic energy and V (x) is the potential energy. Suppose,
for example, x = (r, B), the polar coordinates of an object of mass m moving
in a plane under the influence of a radial field (the origin is the center of
gravity of a body, massive in relation to the object). See Figure 1.6. Then
2 m(r2 + r2e2)
and
V (r) = -K/r,
for some constant K. The action in this case is
(m2t) + r2(t)82(t)] - K/r(t) I dt.
J
r motion
attractive center
The Action Principle has proved a fruitful starting point for deriving the
dynamical equations of complex interacting systems, and for studying their
qualitative properties (existence of periodic orbits with prescribed energy,
etc.).
Necessary Conditions
Consider the optimization problem
(In this equation, Lx and L are the gradients of L(t, x, v) with respect to
the second and third arguments, respectively.)
The Euler Equation (1.1) is, under appropriate hypotheses, a necessary
condition for an arc x to be a minimizer. Notice that, if the minimizer
± is a C2 function, then the Euler Equation is a second-order, n-vector
differential equation:
Lvt (t, x(t), x(t)) + Lvy (t, x(t), x(t)) x(t)
+Lvv(t,x(t),x(t)) X (t) = Lx(t,x(t),x(t)).
A standard technique for deriving the Euler Equation is to reduce the
problem to a scalar optimization problem, by considering a one-parameter
family of variations. The Calculus of Variations, incidentally, owes its name
to these ideas. (Variations of the minimizing arc cannot reduce the cost;
conditions on minimizers are then derived by processing this information,
with the help of a suitable calculus to derive necessary conditions of op-
timality.) Because of its historical importance and its continuing influence
on the derivation of necessary conditions, we now describe the technique
in detail.
Fix attention on a minimizer x. Further hypotheses are required to derive
the Euler Equation. We assume that there exists some number K such that
jL(t, x, v) - L(t, y, w) < K(lx - yI + Iv - wl) (1.2)
Otherwise expressed
T
f e-1 [L(t, x(t) + ey(t), (t) + ey(t)) - L(t, x(t), x(t))]dt >0.
10 1. Overview
a relationship that holds for all continuous arcs w satisfying (1.3). To ad-
vance the analysis, we require
with respect to the strong L2 topology, we readily deduce that (1.6) implies
(1.5) also when w is a L2 function. This fact is used shortly.
Define the constant elements ej E L2 ([S, T]; Rn), j = 1, . . . , n to have
components:
ei (t) _
1 if i = j i = 1, ... , n.
0 ifi#j
Since a function is a constant function if and only if it can be expressed
as a linear combination of the e2(t)s, the properties asserted in the lemma
can be rephrased:
a E V,
< V, W >L2 = 0
<e3 ,w>L2=0,
since ej E V. This last condition can be expressed as
fT
w(t)dt = 0.
s
In view of our earlier comments,
<a,w>L2= 0.
This contradicts (1.7). The lemma is proved.
In view of (1.4), the lemma informs us that there exists a vector d such
that
- J t L (s, x(s), i(s))ds + x(t), x(t)) = d a.e. (1.8)
s
Since Lx (t, x(t), x(t)) is integrable, it follows that t -+ L (t, x(t), x(t)) is
almost everywhere equal to an absolutely continuous function and
dt
(L x(t) - L) = d L i(t) + L x (t) - Lx X (t) - L x (t)
1.2 The Calculus of Variations 13
(In the above relationships, L, Lv, etc., are evaluated at (.t (t), xk (t)).) We
deduce (1.9).
A more sophisticated analysis leads to an "almost everywhere"' version of
this condition for autonomous problems, when the minimizer x in question
is assumed to be merely absolutely continuous.
Variations of the type x(t) +ey(t) lead to the Euler--Lagrange Condition.
Necessary conditions supplying additional information about minimizers
have been derived by considering other kinds of variations. We note in
particular the Weierstrass Condition or the Constancy of the Hamiltonian
Condition:
Suppressing the (t, x(t)) argument in the notation and expressing the Weier-
strass Condition as
L(v) - L(i(t)) > L,, (i) (v - i(t)) for all v E Rn,
we see that it conveys no useful information when L(t, x, v) is convex with
respect to the v variable: in this case it simply interprets Lv as a subgradient
of L in the sense of convex analysis. In general however, it tells us that L
coincides with its "convexification" (with respect to the velocity variable)
along the optimal trajectory; i.e.,
L(t, x(t), (t)) = L(t, x(t), i(t)).
Here L(t, x,.) is the function with epigraph set co {epi L(t, x,.)}. See Figure
1.7.
L(t,X(t),v)
convex hull of
epi
L(t,X(t),X(t))
p(t) x(t) - L(t, x(t), x(t)) = maXVER^ {p(t) . v - L(t, x(t), v)} .
1.2 The Calculus of Variations 15
vxH(t, x, p) I x=.t(t),p=p(t)
= vx (p . X(t, x, p) - L(t, x, x(t, x, p))) I x=2(t),p=p(t)
= p - X. (t, x, p) - Lx (t, x, x(t, x, p))
- LT, (t, x, X (t, x, P)) - Xx (t, x, p) I x=z(t),p=p(t)
= (p - Lv (t, x, X (t, x, p))) . xx (t, x, p) - Lx (t, x, X (t, x, P)) I x=t (t),p=p(t)
= 0 -,ro(t).
(The last step in the derivation of these relationships makes use of (1.11)
and (1.12).) We have evaluated the x-derivative of H:
vxH(t,x(t),P(t)) = -P(t).
As for the p-derivative, we have
direction (p(S),-p(T))
Suppose that Hypotheses (i) and (ii) are satisfied. Then there exists an ab-
solutely continuous arc p : [S, T] -* Rn satisfying the following conditions.
d 2t2t
y2(t) = y2(t) a.e.
It follows that, for some constant c,
/'
dt y2 (t) = c + J y2 (s)ds for all t.
0
(i) L is continuous,
(ii) L(t, x,.) is convex for all t E [S, T] and x E Rn, and
(iii) there exist constants c > 0, d > 0, and a > 0 such that
L(t, x, v) > clvll+a - d for all t E [S, T], x E Rn, and v E Rn.
thus
Ot(t, x) - H(t, x, -4' (t, x)) = 0 for all t E (S, T), X E Rn.
(It is helpful in the present context to consider a variational problem in
which the right endpoint is unconstrained and a function of the terminal
value of the arc is added to the cost.)
22 1. Overview
Proof. Take any Lipschitz continuous arc x : [S, T] -* R" for which x(S)
x0. Then t -* 0(t, x(t)) is a Lipschitz continuous function and
O(S, xo) f T d
dt
0(t, x(t))dt + g(x(T))
(we have used the facts that x(S) = xo and 4(T, x) = g(x))
s
But the first term on the right satisfies
It follows that
We see from (1.16) and (1.17) that :x has cost ¢(S, xo) and, on the other
hand, any other Lipschitz continuous arc satisfying the constraints of prob-
lem (Q) has cost not less than t(S, xo). It follows that O(S, xo) is the min-
imum cost and x is a minimizer.
Proposition 1.4.2 Let V be the value function for (Q). Suppose that
(i) V is a continuously differentiable function;
(ii) for each (t, x) E [S, T] x R'n, the optimization problem (Qt,.,) has a
minimizer that is continuously differentiable.
Then V is a solution to (HJE).
Proof. Take (t, x) E [S, T] x R'n, T E [t, T], a continuously differentiable
function y : [t, T] -4 Rn, and a continuously differentiable minimizer y for
(Qt,x). Simple contradiction arguments lead to the following relationships.
and
t+E t+
E-1 [V (t + E, X + y(s)ds) - V (t, x)] + e-1 J L(s, V(s), y(s))ds = 0.
Jt t
The state trajectory depends on our choice of control function u(t), S < t <
T, and the initial state x(S). The object is to choose a control function u(.)
and initial state to minimize the value of the cost g(x(S),x(T)) resulting
from our choice of u(.) and x(S).
Frequently, the initial state is fixed; i.e., C takes the form
C = {xo} x C1 for some xo and some Cl C Rn.
In this case, (P) is a minimization problem over control functions. However,
allowing freedom in the choice of initial state introduces a flexibility into
the formulation that is useful in some applications. A case in point is ap-
plication of Optimal Control to some problems of exploitation of renewable
resources, where an endpoint constraint set of interest is
C = { (xo, xl) : xo = x1 j.
This "periodic" endpoint constraint corresponds to the requirement that
the population which is being harvested is the same at the beginning and
the end of the harvesting cycle ("sustainable harvesting"). The size of the
initial population is determined by optimization.
There follows a statement of the Maximum Principle, whose discovery
by L. S. Pontryagin et al. in the 1950s was an important milestone in the
emergence of Optimal Control as a distinct field of research.
Theorem 1.5.1 (The Maximum Principle) Let (u, x) be a minimizer
for (P). Assume that
(i) g is continuously differentiable,
(ii) C is a closed set,
(iii) f is continuous, f (t, . , u) is continuously differentiable for each (t, u),
and there exist e > 0 and k(.) E Ll such that
If(t,x,u)- f(t,x',u)J < k(t)lx-x'l
for all x, x' E x(`t) + eB and u E U, a.e., and
(iv) Gr U is a Borel set.
Then there exist an arc p E W',' ([S, T]; Rn) and A > 0, not both zero, such
that the following conditions are satisfied:
and
R(t, x(t), p(t), u(t)) = maUx7-l(t, x(t), p(t), u) a.e.,
Minimize fs L(t,x(t),i(t))dt
over arcs x satisfying (1.19)
i(t) E U,
(x(S),x(T)) E C .
(It is left to the reader to check that, if the condition A > 0 is violated,
then, in either case, (1.22) and (1.23) imply (q, a, A) = 0, in contradiction
of (1.21).)
It is known that a - 0. It follows that a(.) - A. In terms of p(t)
A-lq(t) and these conditions can be expressed
p(t) = L.(t, x(t), x(t)) a.e. (1.24)
(p(S), -p(T)) = , (1.25)
(-p(t),x(t)) = VX,vfl(t,x,p(t),x(t))
and
W(t, x(t), p(t), Xk(t)) = maUx{?-1(t, x(t), p(t), v)}.
*****
*****
Mitä tämä todistaa? Mitäpä muuta kuin sitä, että Axel Ahlbergin
taide alunpitäen tavallaan on ollut lähempänä niitä periaatteita ja
ihanteita, jotka ovat nykyaikaisen teatteritaiteen uudistuspyrintöjen
pohjana, lähempänä kuin monien sellaisten, jotka
todellisuustuntuisen realismin merkeissä ovat taidettaan kehitelleet.
Mutta ennen kaikkea se todistaa, että Axel Ahlberg on oikea
näyttelijä. "Schauspielen heisst: Menschen darstellen, dichterisch
wahre Menschen" — näytteleminen on ihmisten esittämistä,
runoilijan silmillä nähtyjen tosi-ihmisten. Näitä tosi-ihmisiä on
Ahlbergin näyttämökuvain galleriassa pitkä ja unohtumaton sarja.
*****
ROXANE
Mitä?
CYRANO
Hatun töyhtösulkaa."
3.
Adolf Lindfors.
*****
*****
*****
*****
On luonnollista, että pienimpiä yksityiskohtia ja hiuksenhienoja
vivahteita hyväkseen käyttävä ja niihin perustuva "siselöimistyö"
vaatii laajaa esivalmistelua. Niinpä saattaakin jokaisesta Lindforsin
luomasta lukea — ikäänkuin rivien välistä —, että luontevan
liukkaitten liikkeitten, välähdyksinä vaihtuvien ilmeitten ja mainioitten
äänenpainojen takana on mitä tarkin harkinta ja huolellisin
valmistelu. Lindfors perustaa hahmottelunsa tunnollisiin tutkimuksiin.
Hän käyttää hyväkseen tarjolla olevat kirjalliset lähteet
syventyäkseen henkilönsä aikakauteen, hän miettii ja punnitsee sen
ajatusmaailman rakennetta, jossa kuvattava liikkuu, ja vasta tällaisen
tarkan analyysin jälkeen hän ryhtyy muovailutyöhönsä, jossa ei
mikään saa jäädä sattuman varaan. Niinpä saakin hän aikaan sen,
että esim. hänen verraton Napoleoninsa Sardoun "Rouva
Suorasuussa" virittää katsojaan oikean historiallisen tunnelman.
Hänen liikkeensä, asemansa, eleensä, ryhtinsä — koko kuva on aivan
niiden vaikutelmien mukainen, jotka muotokuvat, elämäkerrat ja
anekdootit ovat mieliimme syövyttäneet.
4.
Otto Närhi.
*****
Mimiikki oli ehkä kaikkein tehoisin myöskin Närhin taiteen
ilmaisukeinoista. Hän sai kasvojen eleisiin mahtumaan kokonaisia
tunteitten maailmoja. Hänen miimillinen asteikkonsa oli erinomaisen
laaja ja hän hallitsi sen salamannopeasti vaihtuvia vivahteita aivan
suvereenisesti. Symboolisilla eleillä hän osasi tulkita mielentilat,
maalailevilla havainnollistaa lauselmien sisällyksen ja alleviivaavilla
huomauttaa sanojen ja lauseitten merkityksellisyyttä ja tärkeyttä. —
Mutta myöskin äänikielensä hän oli kehittänyt taipuisan tottelevaksi
aseeksi. Hänen lausunnassaan oli iskevää nasevuutta, mutta
"liukuvan puheen" taidon hän myöskin oli omistanut. Hän puhui
pitämättä puheita. Tahi — kuten Coquelin vanhemmalla oli tapana
sanoa —: Hän ei näyttämöllä puhunut, vaan hän "sanoi"; puhuminen
on nimittäin vain sanojen kuuluviksi tekemistä äänen avulla,
"sanominen" tähtää määrättyyn sisällykseen. Saattaa puhua hyvinkin
paljon ja kuitenkin sanoa mitättömän vähän… Närhi ymmärsi
korkeus- ja mataluussuhteitten, valojen ja varjojen jakoperusteiden
tärkeyden ja osasi niiden varaan lausuntansa rakentaa.
*****
*****
*****
Sellaista oli tämän miehen taide, miehen, joka seisoo aivan yksin
suomalaisen teatterin nuoressa historiassa. Moni suomalainen
näyttelijä — ajattelen esim. parhaassa miehuudessaan manalle
mennyttä Kaarlo Braxénia — on astellut samansuuntaisia teitä kuin
Otto Närhi, mutta ei yksikään niitä polkuja yhtä pitkälle kuin hän.
5.
Mimmi Lähteenoja.
*****
*****
6.
Kirsti Suonio.
*****
1.
"Venetsian kauppias."
*****
*****