Professional Documents
Culture Documents
CntrlEngg (Optimization) RecentAdvancesAndHistoricalDevelopmentOfVectorOptimization WStadler
CntrlEngg (Optimization) RecentAdvancesAndHistoricalDevelopmentOfVectorOptimization WStadler
CntrlEngg (Optimization) RecentAdvancesAndHistoricalDevelopmentOfVectorOptimization WStadler
For information about Vols. 1-156, please contact your bookseller or Springer-Verlag
Vol. 157: Optimization and Operations Research. Proceedings 1977 Vol. 184: R. E. Burkard and U. Derigs, Assignment and Matching
Edited by R. Henn. B. Korte. and W. Oettll. VI, 270 pages. 1978. Problems: Solution Methods with FORTRAN-Programs. VIII, 148
pages. 1980.
Vol. 158: l. J. Cherene, Set Valued Dynamical Systems and Econom,c
Flow. VIII. 83 pages. 1978. Vol. 185: C. C. von Weizsacker, Barriers to Entry. VI, 220 pages. 1980.
Vol. 159: Some Aspects of the Foundations of General Equilibrium
Vol. 186: Ch.-l. Hwang and K. Yoon, Multiple Attribute Decision
Theory: The Posthumous Papers of Peter J. Kalman. Edited by J. Green.
Making - Methods and Applications. A State-of-the-Art-Survey. XI,
VI, 167 pages. 1978.
259 pages. 1981.
Vol. 160: Integer Programming and Related Areas. A Classified Vol. 187: W. Hock, K. Schittkowski, Test Examples for Nonlinear
Bibliography. Edited by D. Hausmann. XIV, 314 pages. 1978. Programming Codes. V. 178 pages. 1981.
Vol. 161: M. J. Beckmann, Rank in Organizations. VIII, 164 pages.1978.
Vol. 188: D. Bos, Economic Theory of Public Enterprise. VII, 142 pages.
Vol. 162: Recent Developments in Variable Structure Systems, Eco· 1981.
nomics and Biology. Proceedings 1977. Edited by R. R. Mohler and
Vol. 189: A. P. LUthi, Messung wirtschaftlicher Ungleich~eit. IX, 287
A. Ruberti. VI, 326 pages. 1978.
pages. 1981.
Vol. 163: G. Fandel, Optimale Entscheidungen In Organisationen. Vol. 190: J. N. Morse, Organizations: M~ltiple Agents with Multiple
VI, 143 Seiten. 1979. Criteria. Proceedings, 1980. VI, 509 pages. 1981.
Vol. 164: C. l. Hwang and A. S. M. Masud, Multiple Objective De- Vol. 191: H. R. Sneessens, Theory and Estimation of Macroeconomic
cision Making - Methods and Applications. A State-of-the-Art Survey. Rationing Models. VII, 138 pages. 1981.
XII, 351 pages. 1979.
Vol. 192: H. J. Bierens: Robust Methods and Asymptotic Theory in
Vol. 165: A. Maravall, Identification in Dynamic Shock-Error Models. Nonlinear Econometrics. IX, 198 pages. 1981.
VIII, 158 pages. 1979.
Vol. 193: J. K. Sengupta, Optimal Decisions under Uncertainty. VII,
Vol. 166: R. Cuninghame-Green, Minimax Algebra. XI, 258 pages.
156 pages. 1981.
1979.
Vol. 194: R. W. Shephard, Cost and Production Functions. XI, 104
Vol. 167: M. Faber, Introduction to Modern Austrian Capital Theory.
pages. 1981.
X, 196 pages. 1979.
Vol. 195: H. W. Ursprung, Die elementare Katastrophentheorie. Eine
Vol. 168: Convex Analysis and Mathematical Economics. Proceedings
Darstellung aus der Sicht der Okonomie. VII, 332 pages. 1982.
1978. Edited by J. Kriens. V, 136 pages. 1979.
Vol. 196: M. Nermuth, Information Structures in Economics. VIII, 236
Vol. 169: A. Rapoport et aI., Coalition Formation by Sophisticated pages. 1982.
Players. VII, 170 pages. 1979.
Vol. 197: I nteger Programming and Related Areas. A Classified
Vol. 170: A. E. Roth, Axiomatic Models of Bargaining. V, 121 pages. Bibliography. 1978 - 1981. Edited by R. von Randow. XIV, 338
1979. pages. 1982.
Vol. 171: G. F. Newell, Approximate Behavior of Tandem Queues. XI, Vol. 198: P. Zweifel, Ein okonomisches Modell des Arztverhaltens.
410 pages. 1979. XIX, 392 Seiten. 1982. .
Vol. 172: K. Neumann and U. Steinhardt, GERT Networks and the Vol. 199: Evaluating Mathematical Programming Techniques. Pro-
Time-Oriented Evaluation of Projects. 268 pages. 1979. ceedings, 1981. Edited by J.M. Mulvey. XI, 379 pages. 1982.
Vol. 173: S. Erlander, Optimal Spatial Interaction and the Gravity Vol. 200: The Resource Sector In an Open Economy. Edited by
Model. VII, 107 pages. 1980. H. Siebert. IX, 161 pages. 1984.
Vol. 201: P. M. C. de Boer, Price Effects in Input-output-Relations:
Vol. 174: Extremal Methods and Systems Analysis. Edited by A. V.
Fiacco and K. O. Kortanek. XI, 545 pages. 1980. A Theoretical and Empirical Study for the Netherlands 1949-1967.
X, 140 pages. 1982.
Vol. 175: S. K. Srinivasan and R. Subramanian, Probabilistic Analysis
Vol. 202: U. Witt, J. Perske, SMS - A Program Package for
of Redundant Systems. VII, 356 pages. 1980.
Simulation and Gaming of Stochastic Market Processes and Learning
Vol. 176: R. Fare, Laws of Diminishing Returns. VIII, 97 pages. 1980. Behavior. VII. 266 pages. 1982.
Vol. 203: Compilation of I nput-Output Tables. Proceedi ngs, 1981.
Vol. 177: Multiple Criteria Decision Making-Theory and Application.
Edited by J. V Skolka. VII, 307 pages. 1982.
Proceedings, 1979. Edited by G. Fandel and T. Gal. XVI, 570 pages.
1980. Vol. 204: K. C. Mosler, Entscheidungsregeln bei Rislko: Multivariate
stochastische Dominanz. VII, 172 Seiten. 1982.
Vol. 178: M. N. Bhattacharyya, Comparison of Box-Jenkins and Bonn
Monetary Model Prediction Performance. VII, 146 pages. 1980. Vol. 205: R. Ramanathan, Introduction to the Theory of Economic
Growth. IX, 347 pages. 1982.
Vol. 179: Recent Results in Stochastic Programming. Proceedings,
1979. Edited by P. Kall and A. Prekopa. IX, 237 pages. 1980. Vol. 206: M. H. Karwan, V. Lotfi, J. Teigen, and S. Zionts, Redundancy
in Mathematical Programming. VII, 286 pages. 1983.
Vol. 180: J. F. Brotchie, J. 'Iv. Dickey and R. Sharpe, TOPAZ - General
Planning Technique and its Applications at the Regional, Urban, Vol. 207: Y. Fujimori, Modern Analysis of Value Theory. X, 165 pages.
and Facility Planning Levels. VII, 356 pages. 1980. 1982.
Vol. 181: H. D. Sherali and C. M. Shetty, Optimization with Dislunctive Vol. 208: Econometric Decision Models. Proceedings, 1981. Edited
Constraints. VIII, 156 pages. 1980. by J. Gruber. VI, 364 pages. 1983.
Vol. 182: J. Wolters, Stochastic Dynamic Properties of Linear Eco- Vol. 209: Essays and Surveys on Multiple Criteria Decision Making.
nometric Models. VIII, 154 pages. 1980. Proceedings, 1982. Edited by P. Hansen. VII, 441 pages. 1983.
Vol. 183: K. Schittkowski, Nonlinear Programming Codes. VIII, 242 Vol. 210: Technology, Organization and Economic Structure. Edited
pages. 1980. by R. Sato and M.J. Beckmann. VIII, 195 pages. 1983.
294
Recent Advances
and Historical Development
of Vector Optimization
Proceedings of an International Conference
on Vector Optimization
Held at the Technical University of Darmstadt, FRG
August 4-7, 1986
Spri nger-Verlag
Berlin Heidelberg New York London Paris Tokyo
Editorial Board
Managing Editors
Editors
This work is subject to copyright. All rights are reserved, whether the whole or part of the material
is concerned, specifically the rights of translation, reprinting, re-use of illustrations, recitation,
broadcasting, reproduction on microfilms or in other ways, and storage in data banks. Duplication
of this publication or parts thereof is only permitted under the provisions of the German Copyright
Law of September 9, 1965, in its version of June 24, 1985, and a copyright fee must always be
paid. Violations fall under the prosecution act of the German Copyright Law.
© Springer-Verlag Berlin Heidelberg 1987
2142/3140-543210
PREFACE
I. HISTORICAL RETROSPECT
W. STADLER:
"Initiators of Multicriteria Optimization" 3
E. BEDNARCZUK:
"Well Posedness of Vector Optimization Problems" 51
J.M. BORWEIN:
"Convex Cones, Minimality Notions, and Consequences" 62
A.R. DA SILVA:
"Evaluation Functionals are the Extreme Points of a Basis
for the Dual of c~[a,bl" 86
R. DURIER:
"Sets of Efficiency in a Normed Space and Inner Product" 114
J. GWINNER:
"A Result of Farkas Type and Duality in Vector
Optimization" 137
S. HELBIG:
"Parametric Optimization with a Bottleneck Objective and
Vector Optimization" 146
VI
J. JAHN:
"Duality in Partially Ordered Sets" 160
I. KALISZEWSKI:
"Generating Nested Subsets of Efficient Solutions" 173
U. KRAUSE:
"Hierarchical Structures in Multicriteria Decision Making" 183
R. LUCCHETTI:
"Well Posedness, Towards Vector Optimization" 194
L.I. POLISHCHUCK:
"On Some Applications of Stochastic Dominance in
Multiobjective Decision-Making" 208
P. SERAFINI:
"Some Considerations About CompUtational Complexity for
Multi Objective Combinatorial Problems" 222
A. STERNA-KARWAT:
"On the Existence of Cone-Efficient Points" 233
A. WIECZOREK:
"Pseudo-Utilities" 241
J. ESTER:
"A Fuzzy Concept of Efficiency" 257
G. ISLEI:
"An Approach to Measuring Consistency of Preference
Vector Derivations Using Least Square Distance" 265
H.R. WEISTROFFER:
"A Flexible Model for Multi-Objective Optimization" 311
H.A. ESCHENAUER:
"Multicriteria Optimization Procedures in Application on
Structural Mechanics Systems" 345
V. RELATED TOPICS
W. Stadler
Division of Engineering
San Francisco State University
1600 Holloway Avenue,
San Francisco, CA 94132, U.S.A.
Suwmary
1. Introduction
The basic problem statement is: Obtain "optimal" decision (s) x EX for
g (x) subject to x EX.
In the preceding the following notation has been used for the
natural order on JRn: For x, y ERn,
x Y iff xi =Y i , ViE I;
(see Figure 1). That is, the linear order $ on the real numbers
'/'(2) OR,:s)
•
192
g(·Hg,(·),9.(·»
~---1f--t-+----...lL~
9,
Figure 2. Decisions with many Desiderata
d ~ d ==* d'" d.
That is, one simply uses the maximal values over the range of g(.) to
induce maximal decisions.
W( .) : (lRN ; ~) -+ (lR s)
with
2. A Historical Sketch
a.nd in the science.s did not begin in earnes.t until the 1970s. This
latter evolutjon of the subject is traced in Refs. 3 and 4.
(1) Ref. 6, p. 21. "It is required to find a point (x,y) such that
in whatever direction we take an infinitely small step, P and n do not
increase together but that, while one increases, the other decreases."
He terms the collection of these points the contract curve,
10
~ ap _ k an Cln
ax - ax and k Cly'
(jp ClIT
Cly ax'
(~ - c ~) dx + (ClP - c ClIT) dy 0
Clx Clx Cly Cly ,
again yielding points of the contract curve. The result is the same if
agent Y is used.
(4) Ref. 6, pp. 22 - 23. He indicated that ~~ > 0 would imply that
both agents could increase their utilities; that if ~~ < 0, motion is
impossible in any direction, thus implying that this is a sufficient
condition for (Pareto) optimality (the condition is only necessary,
in general). Finally, on p. 25, he indicates that the maxima and
minima of P(x,y) and n(x,y) represent demarkation points for the 'pure'
and 'impure' parts of the contract curve.
Let the two goods be given as X,Y and let a specific allocation
for a consumer be denoted by wi = (xi'Yi)' i 1,2. Assume that the
initial allocations of the consumers are woi (xoi'Y oi ) and that their
trade is constrained by obstacle surfaces fi(w i ) = O. The ophelimity
(preference) of each consumer is characterized by a family of
indifference surfaces whose integration along some specific path
(when possible) yields the index of ophelimity I = F(~) (a nonunique
utility function ~(.), since any monotonically increasing transformation
F(·) thereof is another utility function with the same indifference
curves). Pareto coined the term ophelimity to express the economic
aspects of consumer satisfaction, since he believed that the term
"utility" had long served to describe the broader sociological aspects
of satisfaction.
and hence,
~'li th the good X taken as money, we have Px 1 for the price of X and
for the price of the good Y at equilibrium. If the prices are constant,
then it furthermore follows that
~i (w. )
Y 1.
~i (w.)
X 1.
14
Pareto always emphasized his debt to those who worked before him,
even when he considered their theories wrong or inadequate. To his
detractors he wrote: "Perfecting a theory is completely different from
seeking to destroy it by foolish and pedantic sUbtleties."
Marshall and his pupil Pigou represented the total welfare of the
society by the sum
where
4. Game Theory
E
n
{n ERN n 2: 0, n
1
+ ... + n N ,; 1},
(i) n EEn' 1;' ,1;" EEl;' f(I;' ,n) 2:A, f(I;" ,n) 2:A ~ f(l;,n) 2:A,
(ii) I; EEl;' n' ,n" EEn' f(l;,n') ';A, f(l;,n") ,;A -" f(l;,n) ';A,
With this assumption, von Neumann then proves the following theorem.
A similar situation was dealt with nicely when Kuhn and Tucker
discovered that the optimality conditions attributed to them had been
stated and proven earlier by W. Karush (1939) in an unpublished
~aster's thesis, and by F. John (1948) in a paper appropriately
titled "Extremum Problems with Inequalities as Subsidiary Conditions".
Kuhn (Ref. 14) wrote a historical paper in which these related
efforts are put into proper perspective and he rectifies his earlier
omission by publishing Karush's results as an appendix to his paper.
The paper by Koopmans (Ref. 15) is long and the present cow~ents
can only hint at the extent of his results. Suppose that the total
output of a commodity n is given by
k
1:: a .x.,
j=1 nJ J
v = A x
"'" --'
18
y - y ~ Q.
to avoid having to deal with cones whose vertex is not at the origin.
~~~
~·~UM~------------------------------------Y,
(0)
-Y2
Figure 4. The Attainable Cone (A) and the Local Possible Cone (D~ at y
0, p»Q.
Fifty years later, Kuhn and Tucker (Ref. 18) extended the methods
of linear programming to nonlinear programming including inequality
constraints. Prompted, in part, by Koopmans' paper on efficient
solutions, they also posed a vector maximum problem. YH th only
inequality constraints considered, they define the vector maximum
problem as: "Find x satisfying f (x) ~ 0 such that g (x) > g (x) for no x
20
o
<i>x
(a)
o
<i>x " 0,
OT
<i>x
x 0, x~ O.
(b) o
<i>u ~ 0, OT
<i>x
a 0, a ~ O.
x = Q n f- 1 (7) ) {x EX
z
21
g(o) : X ~ Y
and denote
h-(x*(o))
x
= x*(x).
0x<I>«x,z;(.»;~) :$;0 x+ ~,
0x<I>«X,z;(.»;x) =0:
(ii)
for all z * ( .) :> 0*
z'
ff*
z
0z*<I>«X,z;(.»: z;(.» = Z;[f(X)] = o.
7. Conclusions
References
11. Borel, E., The Theory of Play and of Integral Equations with Skew
Symmetric Kernels (in French), Comptes Rendu5 Academie des Sciences,
Vol. 173, pp. 1304 -1308. English translation in Econometrica, Vol.
21, No.1, pp. 97 - 100, January 1953.
25
12. von Neumann, J., On the Theory of Parlor Garnes (in German),
Hathematische Annalen, Vol. 100, pp. 295 - 320, 1923.
Jam Smith
1723 = 1790
28
Francis Edgeworth
18~5 = 1926
30
Vilfredo Pareto was born in Paris, France, on July 15, 1848. His
father, the Marchese Raffaele Pareto had fled there from Italy for
political reasons. In 1858 they returned to Italy and Vilfredo
completed primary school and eventually attended the Polytechnic
Institute of Turin where he attained a degree in civil engineering.
JL848 = JL923
32
Emille BOIrcn
1871 = 1956
34
He was one of the many scientists who fled Nazi Germany along with
such greats as Kac, John and Bochner. At the outbreak of World War II,
he was asked to participate in war related projects, worked as a
consultant for Los Alamos in the construction of the atomic bomb from
1943 on, continued to work on many advisory committees after the war,
and became a member of the Atomic Energy Commission in 1954. He died
three years later, in Washington, D.C., on February 8, 1957.
35
John Von
1903 = 1957
36
From 1944 to 1954 he was with the Cowles Commission for Research in
Economics at the University of Chicago; in 1948, he became its director
and he joined the faculty as a Professor of Economics. In 1955 the
Cowles Commission moved to Yale University where it became the Cowles
Foundation for Research in Economics; Koopmans joined the Yale faculty
as Professor of Economics at the same time. He served as Director of
the Cowles Foundation from 1961 to 1967.
He was author and editor of numerous books and papers ranging from
Tanker Freight Rates and Tankship Building (1939) to Activity Analysis
of Production and Allocation (1951). He continued to serve on the Yale
faculty until his retirement in 1981 and he remained actively involved
in the Cowles Foundation until his death in New Haven, Connecticut, on
February 26, 1985.
37
1910 = 1985
38
Georg Cantor was born 3. March, 1845 in St. Petersburg. The family
moved to Germany in 1856. He wanted to study mathematics, his father
insisted on engineering. In 1862 his father relented. Georg started
studying in 1862 in Zurich and in 1863 continued in Berlin where Kummer,
WeierstraB and Kronecker were teaching.
1845 = 1918
40
868 = 9-42
42
Harold Kuhn was born in Santa Monica, California, on July 29, 1925.
After a brief interruption of his studies by service in the U.S. Army
(1944 -1946), he graduated from the California Institute of Technology
in 1947 with a B.S. in Mathematics. He then continued his studies at
Princeton, where he received his M.A. and Ph.D. degrees in 1948 and
1950, respectively. He was a Fulbright Research Scholar in Paris from
1950 to 1951. In 1952 he joined the mathematics department of Bryn Mawr
College and he left there in 1959 for a year's study at the London
School of Economics as a National Science Foundation Senior Post
Doctoral Fellow. He returned to Princeton in 1960 as an Associate
Professor of Mathematical Economics; he was appointed to the rank of
Professor in 1963.
1905 =
46
Hurwicz was born on August 21, 1917, in Moscow, Russia, where his
parents had fled from Poland during World War I. The family returned
to Poland after the war. After his basic education, he eventually
studied at the University of Warsaw. He left Poland in 1938 to s·tudy
economics at the London School of Economics (1938-1939), at the
Institut des Hautes Etudes Internationales in Geneva (1939 -1940),
Harvard University (1941) and he eventually completed his studies at
the University of Chicago in 1942.
LeoniJ H1ULlrwicz
1917 =
II. MATHEMATICAL THEORY
WELL POSEDNESS OF VECTOR OPTIMIZATION PROBLEMS
Ewa Bednarczuk
Systems Research Institute
Polish Academy of Sciences
01-447 Warsaw, Newelska 6
Summary
1. Introduction
f (x) E wE (f ( r »
where weak efficiency is understood in the sense of a given cone C.
52
M = {x E: r: fIx) E:wE(f(r»}
f (x , u) E: wE (f (r (u) , u) )
f(x) E:wE(f(r(u))
Yo if, for each open set Q containing n (Yo)' there exists a neigh-
bourhood W of Yo such that n (y) E: Q for y E: W.
The following proposition gives an equivalent formulation of well
posedness of a (VOP)u.
Proof (1)-+(2). Suppose on the contra~y that (2) does not hold. If
N(U)=O, then (VOP)u is not well posed and the implication is proved.
Hence, assume that N(.u)/O. Now, there are a o E:N(u), and a sequence
{xn }, xn E: r (u) \ Mao (u), n=1, 2, ••• , ~!! f (x n ) =a o such that {xn } does
not contain any convergent subsequence or contains a subsequence with
the limit point not belonging to r (u). In both cases, there exists
a subsequence {xnk} of {xn } such that
Example. Let X=R, Y=R2, r=r (u) =R, f:R + R2, f (x) = (sinx,cosx). The set
f(r) is simply the unit circle of the euclidean plane and wE(f(r» =
=E(f(r» is the part of the circle contained between points (-1,0) and
(0,-1). The resulting solution set is of the form
FIG.1
• • • •
---- ----
' . . . . . t---- ------
/
M-M(u)
FIG.2
If the problem (VOP)u is well posed the set M(u) need not be compact as
shows the following example.
are compact. This follows from the fact that for different a the sets
~(a) are disjoint.
semicontinuous at u. Yo i N(U).
Now, suppose on the contrary that
By the definition, there exists yos f(r(u» such that Yo s Yo + intC.
By the lower semicontinuity of fr, there exists a sequence
wm E r(uk ), m=1,2, ..• , such that f(wm) E fr(u k ) and Yo = lim f(wm).
m m m+oo
Since Yo=Yo+c, c E int C, hence, for all m sufficiently large
f(zm)=f(wm)+c m, for some c m belonging to int C. Thus,
f(zm) E f(wm)+intC contradictory to the assumption that zmEM(Uk).
m
Now, the only possibility is that x k E r(uk ) " r (u) for all k suffi-
ciently large. Since x k i M(u), it must be f(xk)E N(Uk ) \ N(U).
Since N is assumed to be u.s.c. at u
the sequence f(x k ) contains
a convergent subsequence with the limit point as N(u). Without loss
of generality we can assume lim f(xk)=a. Hence, according to the
definition {xk } is a minimi~tng sequence, and by well posedness of
the problem {xk } contains a convergent subsequence with the limit
point Xo belonging to r(u). By the continuity of f, f(xo)=a,
and consequently xos M(u)" F, contradictory to the assumption
M(U)" F=f/l.
Q.E.D.
Remark 1. In Theorem 1 we consider the problems (VOP)u for which
N(u)=wE(f(r(u))) and this fact is essential for the proof. Now we can
ask a question whether Theorem 1 remains true if instead of (VOP)u
we consider vector optimization problems (VOP)~ of finding all
elements x of r(u) such that
fIx) S E(f(r(u))) ,
f
where g:RxR+R is defined as
g(x,u)
u, g(' ,u)
are closed. Now consider the solutions
r -x+1
'1
for
for
x~o
x~O
Proof. The first part of the present proof is the same as the
first part of the proof of Theorem 1. Only the case x k E r (uk) /"I r (u)
for all k sufficiently large should be analysed here again. Since
f(x k ) E fr(u), the sequence {f(x k )} contains a convergent subsequence
with the limit point YOE fr(u). But Yo actually belongs to N(u).
The proof of this fact is exactly the same as the proof of the same
fact in the first part of the proof of Theorem 1. Hence, {x k } is
a minimizing sequence and contains a convergent subsequence with the
limit point xo' Finally XoEM(U)/"IF, which leads to a contradiction.
Q.E.D.
For fully parametric problems (VOP)u of finding all elements
x E r (u) such that
59
f(x,u) e: wE(f(r(u»,u),
inf{f (x) Ix e: r} ,
inf{f(x) Ix e: ru},
upper semicontinuous at u.
The above definition was used in Bednarczuk 119841 to formulate
sufficient conditions for upper semicontinuity of solutions in scalar
case. The next theorem is a scalar counterpart of Theorem 1.
6. Conclusions
References
J.n. Borweln
Dalhousie University
HaUfax. Nova Scotia
1. Introduction
2. Cones with structure
3. nlnimaUty notions
4. Lattice complementarity
5. Conclusion
THREE THEnES
2.2. PROPOSITIOI
1) Intervals are bounded Iff S Is normal In In this holds
ll1 S Is pointed (sn-S - Ol-
11) lorm and order bounded sets coincide Iff S Is normal
with nonempty Interior.
111) S Is normal 111 S Is weakly normal 1U. S+ Is
generating (s+- s+ • y.) U1 s+ Is a (strict) Hone: for
some 8)0
f(x) Ss g(x)
2c. Bases
2d. Lattices
2.1l. PROPOSITION
o A Danien laUice is order complete.
it) An order complete space is a lattice iff the cone is
IlneraUng.
Ui) A closed cone with base Is latticlal iff B is a
(ChoqueU simplex. In Un eVlry closed pOinted cone Is
Danlen but only the usual simplicial cones are lattices.
(Sel Plressini(fi7).)
tv) A (countably)-arder completa Banach lattice is Daniell
H1 it contains no lattice copy of 1m (10. (See Schaefer
(74), Undanstrauss and Tzafrtrt(7B).)
v) An absolute lattice Is Rlesz 111 Intervals are bounded.
(See Borwein and Yost(B4a).)
vt) A Banach lattice is Daniell (order-conUnuous) iff It
has weak compact intervals iff intervals are strongly
exposed. (See Borwein(B6b). also, Undenstrauss and
Tzafrlrl(78)J
68
rp - 1(8)
(C-x) n -5 c S
-s
11) An 5-least element (strIng minimum. dlmlnaUng) poInt
If iI cllsed set C is a pltnt x In C vlth
(C - X) c 5
(C - X) n inH-S) C ~~~~,~,S"";~,,,;",,;,,~~S~
"""
~"""
"" "' \ C '""""
" '"/;
" ',,
" " ' " """" "
",,,,, " " "
',',',"',,',',', ,',','... ',','".',','
"""""""'"
X
These paints are exactly K-efficient pOints of C. where
K :- {OJ U int S. At the expense of using non-closed cones theIr
theory can be subsumed. (See Borweln(80b).) Also. theIr only
real virtue is the ease of analysis. In the picture. zero is a
strong minimum while x is a domin4/ed weak minimum'
Compare ProposlUon l.4.
cone(C-x) n -9 c S
proper
Proof. strict
73
-I.
bounded functtons In Y admit no strlcUy monotone functtonals.
In particular S (n+ admits no strlcUy monotone
funcUonals. Thus no scalarlzaUon result applies. Despite this.
S Is an order complete Banach laUlce cone vlth tnterlor. Thus
It Is a significant problem to characterize some or all of the
5-properly enlcient paints of subsets of I. (n.
vh.re
76
F(q): = Ix 1 0: T(x)+q 1 OJ
In 1ft type (5) or (H+) matrices are those vith It mast one
positive entry per rov and Ire called pre-Leonttef.
78
is equivalent to solving
Also
Tx ~ lx for all x! 0
T(X)AX ~ 0 ImpUes x S o.
Xo:= LE(x: Tx ! q. x ! f} .
Then
TKo S qVTf .
space (e.g. Loo' [[0.11. c.). The open Leonlief model leilds to iI
6.REFEREIICES
A.R. da Silva
Instituto de Matematica
Universidade Federal do Rio de Janeiro
Brazil
Abstract
nals are the extreme points of a basis for the dual cone of the cone
1. Introduction
Let C [a,b] be the real vector space formed by all real continuous
1/2
(f € C[a,b])
K : = {f € C [a,bJ: f::: O}
has no inner point (for a proof see e.g. BOHL [lJ page 38) . In this
(g € CI [a,bJ)
ordering cone
w*-compact basis for the dual cone Kl* with the evaluation functio-
Let w; [a,bJ denote the real vector space of all generalised derivable
measure on [a,b] .
As it is well-known in the case f is a continuously differentiable
Lemma 2.1
Let ( Cl [a, bJ, II. Ill) (resp. Kl ) be the real n<lrmed linear space (resp.
Proof:
whence
that implies
that is Bo (f) c Kl •
space has a non-empty interior its dual cone has a w*-compact basis.
is a w*-compact basis for K* (For a proof see e.g. JAMESON [5J page
123) •
Theorem 2.2
Let (C l [a,bJ, II. Ill) (resp. Kl ) be the real normed linear space (resp.
Then the dual cone Ki has a w*-compact basis B* with the evaluation
et(g):=g(t).
Proof:
Now let
we have
Therefore
let (f) I = If (t ) I
o 0
~Mllflll,
whence
91
Now the net (t d ) c::: [a,b] has a convergent subnet still denoted by (t d ).
f (s) = lim f (t d )
d
= 1/1 (f) •
w*
Now let A*: = COiiV (F*).
w*
(*) ~ £ COiiV (A* U (-A*)) .
In fact, otherwise there exist nets (t A) c::: [O,lJ and (n A) c::: A* (resp.
= lim [2t A - 1J .
A
Therefore
92
So we have
<I> = w* - lim n I
A' A
Therefore <I> € A*, in contradiction to the choice of <I>. Consequently
In particular, we have
so since <I>€B*c.K*
1
Now let
In fact,
what implies
whence
B*.
o Let ~ be a multiplicative functional and be v and ~ linear
functionals in B* such that
o
~ = (1/2) v + (1/2) ~.
On one hand,
Since s is multiplicative
(II) F*c:::ext(B*).
F* = ext (B*) ,
References
[3J Holmes R.B. (1975) Geometric Functional Analysis and its A~p1ic~
4, 31-32.
ro, Brazil.
Polarities and Stability in Vector Optimization
Abstract :This paper studies the closedness, upper and lower semicontinuity of the solution and
value multifunctions of a parametrized multiobjective problem . The analysis is based upon the use
of polarities in a characterization of the set of maximal points .
Introduction:
In this paper we study the behavior of the set of solutions and the set of
values for a parametrized vector maximization problem. This topic has been
investigated before by several authors as Sawaragi / Nakayama / Tanino [12]
Lucchetti [9] ,PenotlSterna-Karwat [11] . A special attention has been given
recently to the upper semicontinuity of solutions by Bednarczuk [1] .To our
knowledge, Penot and Sterna-Karwat provide the most general results on the
lower semicontinuity and graph-closed ness . Here we slightly weaken their
hypothesis on the behavior of the constraints and objective functions .For
instance in Theorem 2.2 , Conditions 2.9 and 2.10 are substituted for the
closed ness and the lower semicontinuity of G , and in Theorem 2.3 our second
assumption is weaker than the continuity of f .The relation that we consider
on the value space is only supposed to be transitive and we use a notion of
maximality which is sometimes referred to, as efficiency up to
indifference .This choice seems to be natural in the case of a non symmetric
97
Let be Y a set equipped with a transitive relation denoted by !;; ( read "smaller
98
complementary relations. The symbol ::::: stands for the "equivalence" relation
(j;; and ;d ) and c for the "strict" relation (j;; and ~) relative to j;;. We will use
relations It: and :tI. For any relation 0 on Y and for each x,y e Y , we will
yeA, Yo j;; y entails y j;; Yo . If besides yoeA then it is called a maximal point
[O]A = n {Ox : Xe A }.
The subset [O]A of Y is called the J:l.Q..l.ar. of A . In what follows we will take
for 0 , the relation It: ("non strictly smaller than") which may be defined by
The main interest of this relation is that the corresponding polar [1t:]A of A
is exactly the set of points which are non dominated by A .Thus we get the
following formula for the set of maximal points of A
j;; -max A = [1t:]A n A (1.1 )
(1.2)
Indeed the inclusion ::::> is obvious and conversely if z ~ [It: ]cl¢A there exists
yecl¢A such that zcy , implying by (1.3) the existence of x eA such that zcx.
In the sequel we shall need the following characterization of transitivity .
!;-max{!;A) c A and using (1.1) we get !;-max (~A) = [1t:]~A n A. From (ii)
Xlt: z . Then we have x e !;-max{x,z} and x ~ !;-max ~{ x,z} since xcy with
ye !;{ x,z} . @
of relations on Y, 0 will denote the multifunction from WxY to Y such that for
each (w,Y) € WxY ,
o (w ,Y) = { z€ Y:Z € Ow Y } .
We consider for each w€ W the problem:
lJ>(w)
max {f(w,x) : x € X }
where f is a function from WxX to IR .The solution set of such a problem may
limit of F at x :
Y EO ( LimX-I y- F)x
Y EO ( LimX+ I y- F)x
tending to x admits a subnet (Xj)j€J such that there exists a net (Yj)j€J tending
to Y with (Xj,Y)j€J in the graph of F .The link with graph closed ness and lower
semicontinuity is the following : The inclusions
Then
and
T n a(UxV) = +. (2.6)
By virtue of (2.5) there exists a neighborhood U' of w, such that U' c U and for
Substituting for a the relation It: we get sufficient conditions for the
our next result that if It: is a closed set in WxYxY , G is graph-closed and
Then
.E.!:.QQ.! : Assumptions (2.8) and (2.9) allow to apply Proposition 2.1 . We have
(2.11 )
Note that the first condition of Theorem 2.2 means that the graph of the
multifunction r;t: is closed in the product space WxYxY . When we consider a
fixed quasi ordering given by a convex cone C in a topological vector space
this amounts to say that C \ (C n -C) is open .
As seen at the beginning of this section ,the solution set of a scalar
maximization problem is equal to the set of maximal points of the feasible
set Dw with respect to the relation defined on X by (2.1) . Thus in the scalar
case the solution multifunction is graph closed at w as soon as (2.8) is
satisfied for the relation (2.1) and (2.9), (2.10) are satisfied with D instead
of G .These conditions are fulfilled for instance if we suppose f continuous on
{w}xDw and D continuous at w as in Theorem 8 of [8] and also under weaker
assumptions, as in Corollary 5.4 of [2] .
In the non scalar case the graph-closedness of the solution multifunction
may be derived directly from Theorem 2.2 . Let us denote the solution set of
the problem (l1> w) by :
Then
( LimW- X- S)w c Sw
J
104
f.rQ.Q1 : According to (2.13), we may use Theorem 2.2 (with 0 substituted for
G), provided that we prove the graph-closed ness of the multifunction
i;Cfw' WxX ==~ X defined with the aid of (2.14). Suppose that p cfw q , that is,
f(w,p) C w f(w ,q). By (i), there exist neighborhoods T, U, V of w, f(w,p) and
(2.15)
v€V such that f(t,r)!;t u and V!;t f(t,s) . In view of (2.15) and the
transitivity of!; t ' (1.8), (1.9), this amounts to f(t,r) C t f(t,s) for each
such that y=f(w,x} , and by (2.17) there exist neighborhoods T,U of wand x
V n (!;;f(t,u» *- + (2.18)
each net (w j,Xj}jEI in the graph of 0 with (w j}jEI tending to w , there exists a
, O}w c Ow
(LimW-X- (2.20)
Then
GT n v :# t (2.22)
is non empty .We shall prove that there exists XE X such that for every
neighborhoods T of wand U of x
(2.24)
ATx n Ux = t (2.25)
that x EDw . Thus for every neighborhoods T,U,V of respectively W,X and y we
have from (2.23) and (2.24)
f(TxU) n V:# t .
Hence , from (2.21) we have y E !; wf(w ,x) with xED w or equivalently
yE !;wGw. @
Then
~-max G is u.s.c. at w .
Theorem 3.1 recovers in the scalar case Theorem 7 from Hogan [8] saying that
max f(w,Dw) is continuous as soon as D is continuous at w,uniformly compact
multifunction .
Then
.E.!:.Q.Qi : As in Theorem 2.3 ,assumptions (i) and (ii) imply that the relation
defined on X using (2.14) satisfies (3.1) . Moreover (iii),(iv) are equivalent to
conditions (3.2) and (3.3) with D instead of G .Thus using the formulation
(2.13) of Sw, as the set of maximal points of Dw for the relation (2.14) we
The following simple example shows that apparently "good assumptions" are
not sufficient to ensure the lower semicontinuity of the value or the solution
multifunction . Suppose that W = IR and X = Y = 1R2 , f(w,x) = x for each WEIR
and that 1R2 is quasi ordered by the half space {(r1,r2) : r1 ~ O} .we take for Dw
multifunctions reduce to the singleton (w,O) while at ° they are equal to the
segment [(0,0),(0,1)]
In the sequel we shall give two kinds of results ensuring the lower
semicontinuity of the value multifunction. The first one is a variant for the
case of maximality up to indifference of Proposition 3.1 of [11] .
Given a transitive relation !:;; on Y ,we say that Y is !:;; -normal if for every
y€ Y there exists a neighborhood basis ~ (y) such that each V€ ~ (y) satisfies
(LimWxY-Y-
, ;d)(w,y) C ;dwY for each y€Y; (4.2)
Then
(4.1) . By the lower semicontinuity of G (4.3), and (4.4) ,for each neighborhood
(4.7)
compactoid at w I (4.5) the filter ff = (ff TV h,v admits a cluster point zl such
3 teFT (4.8)
Therefore, using (4.2) and (4.3), zl ;dwZo and zleGw which entails, by the
neighborhood T of w,
3 teFT ~t-max Gt n Vo ¢~
contradicting (4.6). @
family {Athe:w of convex cones (with vertex at the origin) is defined. We shall
Aat = At \ {-At}·
We say that a multifunction G :W ==4 Y is attracted by the family {~the:w at
~rmax Gt n V ¢ ~ .
111
Ihen
Remark: In the case where At=A for each t EW (4.9) is automatically satisfied
f.[QQf : Suppose that ZOE (~w-max Gw) \ ( LimW+ y- ~-max G)w . Ihen there
J
w such that
(4.14)
Ao = int( n Aot : t E12 ) is non empty. Let roE Ao and to>O be such that to roE U .
In view of (4.14), for each t EI2 nH, there exists YtEGt such that
Zo +~roE Yt - Aot
Since Act is a convex cone we have, for each t E12 n H , Aot + Ac c Aot and
112
(4.15)
References :
[4] Dolecki S., Continuity of bilinear and non bilinear polarities, Int. School of
Math. G.Stampacchia ,Erice 1984 (R.Conti, E. De Giorgi, F. Giannessi, eds.)
Springer Verlag ,1986.
polarities, to appear.
Roland DURIER
Laboratoire d'Analyse Numerique
Universite de Bourgogne
B.P. 138 - 21004 - DIJON CEDEX
Abstract : In a normed space X the distances to the points of a given set A being
considered as the objective functions of a multicriteria optimization problem, we de-
fine four sets of efficiency (effiCient, strictly efficient, weakly efficient and pro~
perly efficient points). Instead of studying properties of the sets of efficiency ac-
cording to properties of the norm, we investigate an inverse problem : deduce proper-
ties of the norm of X from properties of the sets of efficiency, valid for every finite
subset A of X.
We thus obtain some new characterizations of inner product spaces. The tools are
classical James' characterizations of inner product spaces and a description of each
set of efficiency which looks like the description of the convex hull of A.
The same machinery allows to study polyhedral spaces where only partial results
are obtained. Some characterizations of strictly convex spaces are also given.
INTRODUCTION
- 1 - PRELIMINARIES.
The members of c(A) (resp. E(A), C(A)), are called strictly efficient
(resp. efficient, weakly efficient) points.
Let A be a finite subset of X
M(A) is the set of x in X for which there exists a family of positive
numbers (wa)a E A such that, for every y in X,
We list now some known facts about the sets of efficiency in a normed
space. The first proposition deals with the direct problem, while the
others deal with the inverse problem.
We denote by coCA), coCA) and ri coCA) the convex hull, the closed
convex hull and the relative interior of the convex hull of A.
Proof: Some of thes~ assertions are classic (see [7], [14], [12], [2]).
Proofs of relations about c(A), E(A) can be found in [6] (Proposition
1.3. and Corollary 4.1.). Proofs of relations about M(A) can be found
in [3] (Remark 3.6. and Corollary 4.3.). Another proof of (5) will be
given in Section 5. 0
Remark 1.3. Proposition 1.2. suggests inverse results which will be ob-
tained further. The necessary condition of (2) will be shown to imply
that X is an inner product space, provided the dimension of X exceeds
two, an unavoidable restriction from (4). The necessary condition of (3)
focusing on two-points sets will be shown to imply that X is strictly
convex. On the other hand we shall show that the necessary condition of
(5) does not imply that X is polyhedral.
The next Proposition 1.4. summarizes known characterizations of inner
product spaces, using assumptions on every subset A of X, or at least in-
finite subsets.
Proof. For (1), see [14], Theorem 5. ; the assumption may be replaced by
C(H)c H for every closed hyperplane H of X.
For (2), see [15], Theorem 1.19. and for (3), see [1] (17.10). 0
Proof. For (i) see [2], Proposition 3. and for (ii) see [13], Proposition
4.1. 0
Theorem 2.4.
(1) Let A be a compact subset of X. Then x E X is in C(A) if and only
if, for every oES,
A n (x + P -0) ~ 0
(2) Let A be a finite subset of X. Then x E X is
(i) in c(A) if and only if, for every oES,
An(x + P- o ) ~ 0,
(ii) in E(A) if and only if, for every oES,
An(x+Qo)~0 or Acx + fl_ o .
(3) Let A be a finite subset of X and let X be finite dimensional.
Then x E X is in M(A) if and only if, for every o E S,
A n (x + Uo ) ~ 0 or A c x + P -0 .
a1Uo¢'~'o'p E oy(a),
a 1 P _ 0 ¢'~ '0' p E 0 y ( a) ,
(p,o)
(p, 0)
>
°
> 0,
we get : o 1N(A) if and only if there exists, in the dual space of X, a
closed half-space containing every oy(a) ( a E A) , one at least of these
sets being contained in the open half-space.
From Theorem 3.5. of [3], this last assertion is equivalent to
01M(A). 0
When X is an inner product space, the former description of C(A), c(A)
and E(A) (resp. M(A)) can be viewed as a description of coCA) (resp.
Au ri co(A)). The geometrical character of conditions of Theorem 2.4. is
attractive and turns out to be useful.
Some partial results about E(A), essentially in ffi2 with the 11- norm
or with a polyhedral norm are given in [9] and in [18]. These papers deal
with the so-called location theory; the problem consists, for example,
in locating a facility shared by many users in such a manner that the fa-
cility is as close as possible to each of the users.
(C) means for every closed hyperplane H, there exists C in 5 such that
HcPonP_ o •
Some easy characterizations of inner product spaces can now be
produced
C(A) = coCA),
or E(A) = coCA),
or c(A) = coCA),
or, i f X is finite dimensional,
M(A) = A uri coCA).
Proof: All these conditions imply coCA) c C(A), except the last one which
implies ri coCA) c C(A). o
Theorem 4.1. If
(i) \j o E S, P-0 c Q 0 or P -0 c int Q o '
or (i i) for every two-points set A, coCA) cc(A),
or(iii) for every two-points set A, E(A) c co(A),
or (i v) for every two-points set A, E(A) c c(A),
or (v) for every three-points set A, C(A) c E(A),
then X is strictly convex.
Proof :
1. If a segment [u, u-ao] (with a > 0 and 0 E S) is included in S,
then [u, u-ao[ is included in P- o ' but does not meet Q o .
If a segment [u, u+80] (with B > 0 and 0 E S) is included in S,
then [u, u+80[ is included in P- o but does not meet int Qo .
Sufficiency of (i) is proved by absurd.
I y - a I = I x - a I = I y - b I = I x - b I = 1 and I y I < I x I •
Thus x is not in E(A) and consequently we have a three-points subset A
such that C(A).¢E(A). 0
Suppose now that a non empty finite set A exists with M(A) i E(A) ;
then there exists x satisfy-ing x E E(A) and xjtM(A). We can suppose x = 0,
with a translation of A. According to Theorem 2.4. and the remarks above,
we have :
(*) An U8 0, A;iP_ 8 and An Q8 i 0,
or the same relations with -8.
It is sufficient to investigate consequences of (*). They are
- 6 - CONCLUDING REMARKS.
REFERENCES
1. Introduction
2. Separation Assertions
1)
In more general spaces the continuity can be proved.
131
SA. : = cl C + A.k,
o A. E R, (2.1)
and each fixed z E Z there exist A., A.' E R such that z ~ SA. and z E SA. ' •
For the functional s we are interested in one can take the following
modified Minkowski-functional
Both classes of that cut are nonempty which was shown above. The
"upper" class is connected according to
because of
(2.3)
s(z) <: r ~ z ~ C
°
+ rk . (2.7)
In the case r =0 we conclude from (2.4) - (2.7) all the assertions (iv)
concerning separation. Since r E R is arbitrary we obtain from (2.8)
the assertion (iii), i.e. s(Z) = R. This completes the proof.
Remarks.
Z U (c 1 C - Ak 0 ) ,
AER
such that
(ii) s(Z) =R
Z U (int K - AkO)
7I.ER
3. Duality Assertions
Always weak duality holds since the dual problem (D) is of the
"Lagrange type". Moreover there exists a strong duality theorem and,
under strong assumptions, an inverse duality theorem, too (/5/, /6/,
/8/) .
135
We set
° if u E V,
L(u) 1 + = otherwise
and obtain
hence f (xO) ED. Since by the weak duality for any h ED the relation
h-f(xO) ER~"{O} must fail, f(xO) is dual efficient.
4. References
/3/ ELSTER, K.-H./IWANOW, E.: tiber Fortschritte auf dem Gebiet der
mehrkriteriellen Entscheidungsprobleme. Wis-s. Zei tschr. TH
Ilmenau 11. (1985) 2, 15 - 36.
/4/ ESTER, J./SCHWARTZ, B.: Ein erweitertes Effizienztheorem. Math.
Operationsforschung u. Statist., sera optimization 1! (1983),
331 - 342.
/5/ GERSTEWITZ, C./IWANOW, E.: Dualitat fur nichtkonvexe Vektor-
Optimierungsprobleme. Wiss. Zeitschr. TH Ilmenau 31 (1985) 2,
61 - 81. -
/6/ GERTH, C./WEIDNER, P.: Nonconvex separation theorems and some
applications to the vector optimizatiori (To appear in JOTA).
/7/ G~PFERT, A.: Multicriteria duality, examples and advances:
Lecture Notes in Econ. and Math. Systems Nr. 273, Springer
Verlag 1986, 52 - 58.
/8/ JAHN, J.: Mathematical Vector Optimization in Partially Ordered
Linear Spaces. Meth. u. Verf. der math. Physik Bd. 31. Vlg.
Peter Lang, Frankfurt/M. (1986).
/9/ KL~TZLER, R.: Dualitat bei diskreten Steuerproblemen. Math.
Operationsforsch. u. Statist., sera optimization 12 (1981) 3,
411-420. -
/10/ WEIDNER, P.: Charakterisierung von Mengen effizienter Elemente
in linearen Raumen auf der Grundlage allgemeiner Bedingungen.
Diss. A, MLU Halle-Wittenberg (1995).
A RESULT OF FARKAS TYPE AND DUALITY
IN VECTOR OPTIMIZATION
Joachim Gwinner
Fachbereich Mathematik, Technische Hochschule Darmstadt
SchloBqartenstr. 7. D-6100 Darmstadt
1. INTRODUCTION
The p~oo6 of this minmax type result can be given using the separation
theorem [see Gw-1, §2]; Theorem 2.1 can also be directly derived from a
more general minmax theorem due to Fuchssteiner and Konig [Fu-KO], based
itself on a general Hahn-Banach principle.
P~oo6. Defining
tp(x) max{F(x,Z>1 Z e Zo}
y(x) max{G(x,y)1 y e Yo}
(2.4) is equivalent to
Z e Z )
o
140
result.
(3.7)
G(x, 'Y2)+ F(X, ,z) ~ <~2'z>
Multiplying (3.5) and (3.6) by -1 and adding up all three inequalities,
we come up with the claimed inequality <A _A z>:>. o.
2 l' q.e.d.
This section presents our strong duality theorem for the dual pair of
veotor programs introduced in §3. Here we reqnire Karlin's CQ (2.3)
or equivalently Slater's CQ (2.7) as hypothesis.
The pltoo6 essentially follows the lines of Oettli [Oe-1] and consists
of three parts.
1 .) We show: ~ is V-minimal ~~ e W.
By (3.3) we have
system
y e Y • (Vx e X) G(x.y) ~ 0 .
Now vary A e E*
z only; this implies z1 and we obtain
hence I e W
We remark that this strong duality result extends earlier work of Aubin
[Au-1]. Corley [Co-1]. and using the convex core topology also the neces-
sary optimality conditions of Kirsch-Warth-Werner [K-W-W, II §3]. where,
however. affine equality constraints are also considered.
where
(x e X I (V Y e y) G ( x • y) ~ O}. (5.1 )
Now let S be a generating set of X, i.e. S is a subset of Y such
that x e X belongs to il: if and only if
(vy e S) G(x,y) ~ 0
We note that Slater's CQ (2.7) implies that S= contains the zero ele-
ment only and thus excludes equality constraints.
With the splitting
S< SUR,
x {x e X I (Vy e S= U R) G( X t y) ~ O} ,
Y R +5
Lemma 5.1. Let S< be defined as above and let (H) be satsfied. Then
there exists some x e X such that G(x,y) < 0 holds for all yeS.
For the p~oo6 we refer to [Bo-Wo] or to the proof of the more closely
related Lemma 6.1 in [Gw-1].
Now we can present the promised strong duality result.
Theorem 5.2. Let S< be defined as above and let (H) be satisfied.
Let Xe V. Then X is V-minimal, if and only if X is W-maximal.
6. CONCLUDING REMARKS
REFERENCES
Abstract
I. Introduction
Y € M, Y $ v - Y = v.
Y € M, Y ~ v - 3 Yi Vi.
i€{1, .•• ,I}
Let S be the set of all weakly efficient and Seff the set of all effi-
cient points of M. Then consider the vector optimization problem
147
OP(q): min
subject to y € M.
:= inf (q,y)
y€M
and
;= {y € M I (q,y) = E }.
q
Let LM
be the set of all parameters q € RI, for which the problem OP(q)
has a solution, i.e.
The next theorem shows the relationship between VOP and OP(q) and bet-
ween the sets Seff (resp. S) and Pq :
148
Remark: BROSOWSKI and CONCI [4] proved part c) for a vI in the spe-
cial case of a set M described in section V.
which implies u i < vi flir i=1, ... ,1. This contradicts v € Seff (resp.
v € S). In the case of efficiency of v, we have
({v} - C) n M = {v},
This result shows that even for non-convex and non-compact sets M the
efficient and weakly efficient points can always be characterized by an
optimization problem, which is nearly a discrete Tschebyscheff-approxi-
mation problem. Other authors corne to similiar conclusions:
For this, let M be a compact set and put Y1* := min Y1 and K1 := M. Then
define recursively for i=2, ••• ,1 YEM
Ki - 1
Yj* for j
Ki := {y < i}
E
Yj
Yi* := min Yi·
y€K i
It follows:
min
YEM
Remark: The main idea of the proof below is due to COLGEN and SCHNATZ.
In [5], they showed a similiar result for linear optimization with vari-
able objective function.
k
I := {i E {1, ..• ,1} I (Vi) is unbounded},
y.
151
v + AY E M for all A 2: o.
Hence, the sequence (vk ) is bounded from above. Applying lemma 111.1.,
we get a contradiction to the compactness of P q for q = -v, V € Seff. 0
This theorem leads to a necessary and sufficient criterion for the non-
emptiness of Seff.
(q,u) ~ (q,v)
o = b(2)
and
bet) for t E [0, 1 ] •
LM = u r (v) •
VES eff
u r (v) •
YES
Example III. 6. :
a)
y."--~
IV. Continuity properties of the optimal value and the optimal set
Theorem IV.1.: The mapping E is continuous in q E int LM, where int de-
notes the topological interior.
Proof: Let E > 0 be given. Furthermore, let a > 0, such that a $ E/2
and Ua(q) eLM' where Uo(q) := {p E R
1
max Ip.-q. I < a}.
i=1, .•. ,1 ~ ~
156
Define a set K by
Pi < qi + I)
for i=1, •.• ,1,
qi + Yi ,;:; Eq + I)
which implies
Hence,
For this,
It should be mentioned that theorems IV.1. and IV.2. hold without as-
sumption to M. Another continuity property of P is shown in
Proof: If P
is not upper-semicontinuous in q € M, then there exist an L
open neighbourhood W of Pq and a sequence (qk) in M, which converges L
to q, such that P k\W ~ ~. Let v k € P k\W. If (vk ) is a bounded sequen-
q q
ce, tben we get a contradiction with theorem IV.2 .. Therefore, let (vk )
be an unbounded sequence with ~vkn2 + 00. Since the optimal-value-mapping
E is continuous (note that LM
= BI is an open set if Seff ~ ~), for
sufficiently large k
Hence, the sequence (vk ) is bounded from above. Applying lemma 111.1.,
the optimal sets P p ' p € L
M, are not compact. With corollary 111.3.,
this is a contradiction to Seff ~ ~. D
v. A scalarization of VOP
A: TxRn ->R
a: T ->R
I
Pi: R ->R for i=1, •.. ,1
1
For each A E A := {p E R I Pl O} consider the minimization problem
MP(A): min z
subject to
A(t,x) s a(t) for all t E T
Pi(x) - Ai s Z for i=1, ••• ,l.
Denote by EA and P A the optimal value and the optimal set of MP(A).
With M p(Z), it is obvious that
For that reason, we can apply the results of the former sections:
A) Characterization theorem
BROSOWSKI characterizes in [3J weakly p-efficient points of Z by a
Kolmogoroff-type-theorem. In the linear case, i.e. that Z is described
by linear inequalities and Pi are linear for i=1, ••. ,l, a Kuhn-Tucker-
type-theorem holds for weakly p-efficiency (see [3]). Our characteri-
zation theorem reads as follows (compare corollary 11.2.):
References
[1] V.J. BOWMAN: On the relationship of the Tchebycheff norm and the ef-
ficient frontier of multi-criteria objectives
in: THIRIEZ and ZIONTS, Multiple Criteria Decision Ma-
king, Springer-Verlag, Berlin, 1976
[2] B. BROSOWSKI:Parametric semi-infinite optimization
Peter Lang-Verlag, Frankfurt/Main, 1982
[3] B. BROSOWSKI:A criterion for efficiency
Preprint, University of Frankfurt/Main, 1986
[4] B. BROSOWSKI and A. CONCI: On vector optimization and parametric pro-
gramming, Segundas Jornados Latino Americanas de Mate-
matica Aplicada 2, 1983, 483-495
[5} R. COLGEN and K. SCHNATZ: Continuity properties in semi-infinite
parametric linear optimization, Numerical Functional
Analysis and Optimization 3, 1981, 451-460
[6} J. JAHN: Parametric approximation problems arising in vector
optimization, to appear in IOTA
[7] C. K6LLNER: Ein interaktives Verfahren zur Losung finiter, li-
nearer Vektoroptimierungsprobleme, Diplomthesis, Uni-
versity of Frankfurt/Main, 1987
DUALITY IN PARTIALLY ORDERED SETS
Johannes Jahn
Institute of Applied Mathematics
University of Erlangen-Nurnberg
Martensstrasse 3, 8520 Erlangen, F. R. Germany
pEP, p S P ==> p S p.
The set P is also called the primal set and the set D is said to be
the dual set of P. In the next section we characterize minimal ele-
ments of the primal set P by maximal elements of the dual set D.
Recall that an element d E D is called a maximal element of D, if
-
d ED, d::; d => d S d.
D : = {d E X I pEP => P :f d or d p} .
If the dual set D is nonempty, then we have for each pEP and d E D
the relation p :f d or d ::; p. In optimization theory this result is
often called a weak duality theorem. In the context of this paper it is
an immediate consequence of the definition of the dual set D.
2. Duality Relations
Lemma 1:
Let (X,S) be a partially ordered set, let P be a nonempty subset of X,
and let D be defined as in (1). If x E P n D, then x is a minimal ele-
ment of the set P and also a maximal element of the set D.
Proof:
Since xED, it follows that for each pEP with P ~ x we obtain x ~ p.
This implies that x E P is a minimal element of P. Next, take an arbi-
trary element d E D with x S d. x belongs to P and therefore we get
d S x from the def ini tion of D. Consequently, xED is a maximal ele-
ment of D. c
Theorem 2:
Let (X,~) be a partially ordered set, let P be a nonempty subset of X,
and let D be defined as in (1). Then each minimal element of the set
P is also a maximal element of the set D.
Proof:
Let pEP be any minimal element of the set P, i.e. for each pEP
with P ~ P we have p ~ p. So, p belongs to the set D, and by Lemma
p is also a maximal element of the set D. c
Theorem 3:
Let (X,~) be a partially ordered set, let P be a nonempty subset of X,
-
and let D be defined as in (1 ) • For each maximal element d E D of the
set D let the following implication hold:
Proof:
Let d E D be an arbitrary maximal element of the set D. Then we have
for each d E D with d S d also d S d. Consequently, by the implication
(2) d belongs to the set P, and Lemma 1 leads to the assertion. D
Lemma 4:
Let (X,S) be a partially ordered real linear space, let P be a non-
empty subset of X. Moreover, let there exist some hEX with Ox S h
and h ~ OX' and let an arbitrary d E D be given.
(c) If the space is equipped with a topology and the set P+ given by
(3) is closed, then the implication (2) is satisfied.
Proof:
Let d E D be arbitrarily given and assume that d $ P.
(b) Since the set P+ is convex and algebraically closed, the set X\P+
is algebraically open (e.g., compare [9, p. 13] for a similar
result). Notice that
(c) Because the set P+ is assumed to be closed, the set X\P+ = D\P is
open, and the assertion is a consequence of part (a) of this lemma.
o
Remark 5:
If, in addition to the assumptions of Lemma 4, the partial order is
also antisymmetric, then the ordering cone ex := {x EX I Ox $ x} in-
duced by the partial order $ is pointed, and therefore in this case
the set P+ given by (3) equals the algebraic sum of the sets P and ex'
i.e. P+ = P + ex' In order to see this equality notice that
P + ex'
The assumption in Lemma 4, (c) that P+ = P + ex is closed is a stand-
ard assumption used for the proof of a converse strong duality theo-
rem in vector optimization (e.g., see [8, Thm. 2.5 and Lemma 2.4]).
V P E P\{d}}. (5)
Lemma 6:
Let the assumption (4) be satisfied, and let P be any nonempty subset
of x. Then 15 c D where 15 is given by (5) and D denotes the dual set of
P.
166
Proof:
In the case of 0 = 0 the assertion is obvious. If 0 is nonempty, take
any d E D. Then we consider two cases:
(a) Assume that there exists some l E C*'X* with
led) s l(p) for all pEP.
Then we get for all p E P that p ~ d or d p. This means d E D.
(b) If there exists some l E Cx*\ {Ox*} with
l (d) < l(p) for all p E P\{d} ,
it follows for all pEP that p ~ d or d p. So we obtain d E D
as well. c
Lemma 7:
Let the assumption (4) be satisfied, and, in addition, let X be
locally convex. Let P be a nonempty subset of X for which the set
P+C x is closed and convex. Assume that there exists a continuous
linear functional l *' E Cx*' * with - co < inf l *' (p).
pEP
Moreover, let the following condition be satisfied:
For each minimal element p of the set P there exists
either a continuous linear functional l E C~* with
l(p) ~ l(p) for all pEP (6)
or a continuous linear functional l E Cx*\{Ox*} with
l(p) < l(p) for all p E P\{p}.
Then D c 0 where 0 is given by (5) and D denotes the dual set of P.
Proof:
Since the assertion is evident for D = 0, in the following we assume
+
D 0 and we pick an arbitrary d E D. Then we have for all pEP
P ~ d or d = p which implies that d *
P + (Cx\{Ox}). Next we consider
two cases:
(a) If we assume that d E P + Cx' we get d E P. Since d E P n D,
by Lemma 1 d is a minimal element of the set P. Then by the
condition (6) d belongs to D.
167
Remark 8:
The assumption in Lemma 7 that the infimum of a continuous linear
functional ill E C~* on P is finite is fulfilled for instance, if the
primal set has a lower bound, say E x. In this case P c x
+ Cx {x}
and
inf l
II
(p) ~ l
II
(x) > -
A
=.
pEP
Remark 9:
The condition (6) holds for instance, if the set P is strictly convex.
A set P which is assumed to have a nonempty interior is called
striatly aonvex, if for every P1' P2 E P with P1 P2 +
AP1 + (1-A)P2 E int{P) for all A E(O,1).
In fact, if p is any minimal element of the set P, i.e.
{{p} - Cx) n P = {p}, by Eidelheit's separation theorem (e.g., see
[9, Thm. 3.16]) there exists some l E Cx*\{Ox*} with
l{p) < l{p) for all p E int{P).
168
Remark 10:
If the set P is a convex polytope in ~m equipped with the compo-
nentWise partial order, then by a known result of Gale [3, p. 308]
for each minimal element p of P there exists a vector l E R n with
positive components such that
lTp ~ lTp for all pEP.
Corollary 11:
Let the assumptions of Lemma 7 be satisfied. Then an element of the
primal set P is minimal if and only if it is a maximal element of the
dual set
Proof:
With Lemma 6 and Lemma 7 we obtain that the dual set 0 can be written
as above. Then we get the assertion from Theorem 2 and Theorem 3 to-
gether with Lemma 4, (c) and Remark 5. 0
min ex
subject to
Ax ~ b
x ~ 0 ~n
P := {exi Ax ~ b, x ~ 0 ~n}. (8 )
Lemma 12:
Let the assumption (7) be satisfied, and assume that the set P given
by (8) is bounded. Then the dual set of P reads
Proof:
Since the primal set P is a convex polytope, by the Lemmas 6 and 7
(in connection with Remark 10) the dual set of P can be written as
Consequently, for b +
o~ the dual multi-objective linear programming
problem is equivalent to the following problem formalized by
max Ub
subject to
ATUT.t ~ cTt
uT.t ~ °mP
t 1 ,···,.tm > °
U EIRmxp.
4. Conclusion
References
[1] H.W. Corley, "Duality Theory for the Matrix Linear Programming
Problem", J. Math. Ana~. App'L 104 (1984) 47-52.
[2] K.-H. Elster and A. Gopfert, "Recent Results on Duality in Vector
Optimization", in this volume.
[3] D. Gale, The Theory of Linear Economic Mode~s (McGraw-Hill,
New York, 1960).
[4] Ch. Gerstewitz and E. Iwanow, "Dualitat fur nichtkonvexe Vektor-
optimierungsprobleme", Wiss. Z. TH Itmenau 31 (1985) 61-81.
[5] J. Gwinner, "A Result of Farkas Type and Duality in Vector Opti-
mization", in this volume.
[6] E.H. Ivanov and R. Nehse, "Some Results on Dual Vector Optimiza-
tion Problems", Optimization 16 (1985) 505-517.
[7] E. Iwanow, "Verallgemeinerte Konvexitat und die Dualitat in der
Vektoroptimierung", Wiss. Z. TH I~menau 31 (1985) 39-44.
[8] J. Jahn, "Duality in Vector Optimization", Math. Programming 25
(1983) 343-353.
[9] J. Jahn, Mathematicat Vector Optimization in Partiatty Ordered
Linear Spaces (Peter Lang, Frankfurt, 1986).
[10] M.G. Krein and M.A. Rutman, "Linear Operators Leaving Invariant
a Cone in a Banach Space" (Uspehi Mathematiceskih Nauk (N.S.) 3,
no. 1 (23), 3-95 (1948)), AMS T:t'anstation Series 1. Volume 10,
Providence (1962), p. 199-325.
[11] H.C. Lai and C.P. Ho, "Duality Theorem of Nondifferentiable
Convex Multiobjective Programming", J. Optim. Theory App~. 50
(1986) 407-420.
[12] C. Malivert, "Dualite en Programmation lineaire Multicritere",
Math. Operationsforsch. Statist. Sere Optim. 15 (1984) 555-572.
[13] H. Nakayama, "Geometric Consideration of Duality in Vector Opti-
mization", J. Optim. Theory App~. 44 (1984) 625-655.
[14] H. Nakayama, "Duality Theory in Vector Optimization: An Over-
view", in: Y.Y. Haimes and V. Chankong (eds~, Decision Making
with Muttipte Objectives (Springer, Berlin, 1985).
[15] H. Nakayama, "Lagrange Duality and its Geometric Interpretation",
in: P. Serafini (ed.), Mathematics of Mu~ti Objective Optimi-
zation (Springer, Wien, 1985), p. 105-127.
[16] V. Postolica, "Vectorial Optimization Programs with Multifuncti-
ons and Duality", Ann. Sci. Math. Quebec 10 (1986) 85-102.
172
Abstract
1. Introduction
Yi-Yi
" M
Yj-Yj
~
the assumption about the existence of y is needed,makes both
programs very attractive. Moreover, these programs permit one to get some
extra information about properly efficient solutions determined by them,
which might be useful to a decision maker. Such information is provided
by the "if" parts of Theorem 2.1 and Theorem 2.2: for each y E. Z which
solves either pA or PA not only do we know that there exists a finite
number M (because is properly efficient) but also a certain upper
bound on this number is established.
Below, whenever y~ Z is properly efficient, we shall assume that
M is the smallest possible number for which the inequality of Defini-
tion 2.1 is fulfilled. In other words we shall refer to a slightly
modified definition of M.
Yi-Yi
M max max min
y i;Yi-Yi>O j;Yj-Yj>O Yj-Yj
where Y€. Z.
Obviously, with respect to proper efficiency both definitions are
equivalent. The difference is Definition 2.2 addresses one particular
number whereas Definition 2.1 addresses rather an interval of numbers.
One confines oneself to properly efficient solutions not only
because this set of solutions is technically more easy to deal with
but also because improperly efficient solutions in the context of
multiple objective decision making are regarded as anomalous. Indeed,
interpreting coordinates Yi of elements Y£ Z as values of objective
functions fi(x}, i=1, ••• ,k over certain set S, for an improperly
efficient solution Y €. Z
there exists an objective function fi for which the marginal
gain can be made arbitrarily large to ~ach of the marginal losses
in other criteria (compare [4]).
It is quite natural then to confine ones attention to properly
efficient solutions for which
for each objective function fi the gain relative to at least one
of the losses in other criteria is' not larger than a certain
prescribed value.
In other words one may be interested in properly efficient solutions
for which M is not greater than a given number. The motivation of
this is that one may want to avoid solutions for which the gain for
each objective function fi relative to at least one of the losses in
177
efficient set, but the notion of the central part is difficult to make
precise (compare Fig. 1 a,b,c). If Z is non-convex the situation is
even less clear as is shown by the example.
a) b) c)
Fig. 1. Values of M: a - high, b - moderate, c - small.
3 for O~Yl<3,
1 for 3~Yl~4,
2 for 4<Yl~S,
2 for
1 for
4 8
Fig. 2.
179
On the ot~er hand, however, the solution resulting form the maxi-
mization of 2::y. over Z, which is often assumed to be a good
i=1 ~
starting point for interactive decision making processes, will be
always included in the first subset generated by one of the algorithms
proposed above. This fact is implicit in Lemma 2.2. To prove this lemma
we will need the following fact.
Lemma 2.1. Given a nonempty set Z, there is at most one properly
efficient element with M<1.
Yi-Yi
, 1
whenever
Yj-Yj
and therefore My~1.
If My< 1 then by Lemma 2.1 there is no other Y (. Z, Y efficient,
such that My <1. Thus, Y has the smallest value of M.
If M-=1 then observe that for each Y £ Z, yfy
Y
180
Yi-Yi
and this entails that for each Y € Z, yiy
Yi-Yi
min ~ 1
j;y·-Y·>O Yj-Yj
J J
which completes the proof.
For the case k=2 one can also make use of values of M to get
the approximate shape of the efficient frontier of the set Z. Suppose
that for a certain value of M an efficient element yEo Z has been
determined by the algorithm pA or PA' Then, by Definition 2.1,
Y2-Y2
or - _ - - ~ ~1 for all Y E'. z, ytY. This can be
Y1 -Y 1
represented graphically as in Fig. 3 - all elements Y E. Z must lie on
or below the line (YI-Yj);:; H(Y2- Y t
or (Y2- Y2) ~ H(Y 1 -Yl) which
gives a rough approximation of Z (the shaded region in Fig. 3).
Repeating this procedure for another efficient element, sa:' y, and
taking the common part of approximations resulting form y and y we
get a nore accurate approximation.
Fig. 3.
Example 2.2. Consider the set Z given in Example 2.1. Suppose that
the following efficient elements of Z have been determined
Fig. 4.
3. Conclusions
References
Ulrich Krause
Fachbereich Mathematik und Informatik
Universitat Bremen, 2800 Bremen, West Germany
1. Introduction
Confronted with decisions involving many criteria in real lif. people
sometimes apply a rankin~ of importance to the criteria, instead of
trying to balance the criteria on some common scale. The principle
"first things first" is at the heart of any hierarchical procedure
and one may ask for the reasons people are applying it. Beside em-
pirical investigation, an answer may also be given by looking analy-
tically for possible appealing features which hierarchical structures
could possess. In this respect, not only the individual decision
taker is of interest, but also his cultural environment, including
e.g. the dictionary or the decimal system both being based on lexi-
cographical ordering. It is the main result of the present paper
that hierarchical decision procedures may be characterized by some
specific stability property. Roughly speaking, the latter means that
a decision taken in favor of one alternative against the other, is
not upset by a small perturbation in the assessment of any of the
criteria. For the result, we have to assume the decision rule under
consideration yields decisions in sufficiently many cases of conflict
between the different criteria. Without this assumption the stability
property is not very demanding and it is shared by various different
rules.
one hand simplifies the mathematics and gives on the other hand more
room to applications. The reason for the latter being that not in
every application real numbers are available. Obviously, for taking
decisions within A not the function F:R n + R is really needed, but
only the quasi-ordering defined by x~y iff F(x)~F(y). A quasi-
ordering~ on IRn is a binary relation~ on IRn which is reflexive
(x.{x for all x), transitive (x4y and y.{z imply x~z) and total
(x~y or Y~ x). By -< we denote strict ordering, i.e. x<y means x<4y
and not y~ x. By'" we denote indifference, i.e. x""'y means x~y and
y~x. It should be noted that~ is not assumed to be anti-symmetric
(x4 y and y4. x imply x=y). Thus more generally what we are looking
for is a stable quasi-ordering ~ on IR n , meaning there exists some
o<d~l such that x4y is equivalent to ,x~ ,y, for all x,y ElR n and all
small perturbations l. Before turning to the final version of the
stability property, it is useful to look at the following alternative
descriptions of its present version.
Lemma. For: any quas i-order; ng 4 on IRn the foll owi ng properti es are
equivalent.
(i) 4 is stable
(i i ) x~y <= > l X ~ l Y
for all X,YER n , all scale transformations l defined by lz=(a 1 z 1+b 1 ,
.•. , anzn+b n ) with ai,b i E IR arbitrary, but ai>o for i=l, ... ,no
( iii) If for x, y , u , vEIR n s i gn (y i-x i ) =s i gn ( v i - u i) hold s for all i,
the n x -< y i seq u i val en t to u -< V. ( The reb y for r EIR, s i gn r =-fr-r if r", 0
and sign 0=0.)
Proof.
(i) => (ii): By accumulation of small perturbations. Let ai)o, biEIR
i =1 , ... ,n. De fin e for kE {I ,2 ,3 , ...• }
bi t bi
= ~(ai - 1) if a i ",l and 0i=r- if a i =l.
1
By assumption 4 is stable for some o<d~l. Choose k such that for
Ei' 0i defi ned above one has IEi I, I 0i I<d. Then
TZ = (E1z1+01, ... ,Enzn+on) defines a small perturbation and by iter-
ating it k-times the scale transformation l is obtained,
Cb ...• 't)z=(a1z1+b1' ... ,anzn+bn)=lz. Therefore (i) implies
x 4 y <= > Tx 4 Ty <=> ('(. '( ) x ~ ( T• '( ) y <= > ... <=> LX -4 l Y .
(ii) => (iii): Let x,y,u,vER n and suppose sign(y.-x.)=sign(v.-u.)
1 1 1 1
for all i. Define
186
vi-u i
a.
1
=- --
Yi-Xi
if and a i =1 otherwise. Then a i >0 for all i.
3. Hierarchical structures
n
Let (Xi'~i) for i=l, ... ,n be quasi-ordered sets and let X = IT X.
i =1 1
be the Cartesian product. A typical element of X we denote by
x=(x 1 , ... ,x n ). A quasi-ordering ~on X we call hierarchical, or of
hierarchical structure, if there exist
some number m € {l, ... ,n},
an injection a : {l, ... ,m} + {l, ... ,n}
and a mapping t, :{l, ... ,m} + {-<., }-}
such that for any X,YEX the relation x<..y holds i f and only i f
there exists some iE{1, .•. ,m} with xo(i)t,(i)o(i) Yo(i) and
x o(j )("\Jo( j) Yo( j ) fo r all jE {l, ... , i -1 }.
n
on X = IT Xi is called strictly increasing, if xi~iYi for all
i=l
i implies x~y and X-4.y holds provided there is in addition some
j with xj"'(jYj' Converse1Y~ consider a quasi-ordering40n X which
is stable and strictly increasing. By the theorem~ must be hier-
archical and because it is strictly increasing we must have m=n
and 6(i)=.( for all i. Hence{ muH be lexicographical. Thus the
theorem yields immediately the following corollary.
Corollary 2.
(a) Any quasi-ordering on Rn, n~2, which is stable (with respect to
S on R) must be hierarchical. It is stable and strictly increasing
if and only if it is lexicographic.
(b) Any stable aggregation rule F:R n - R, n~2, must be trivial in
the sense that F is constant or is a function of exactly one coordi-
nate only.
Before turning to the proof of the theorem we like to add some re-
marks concerning the relationship of the corollaries to two dif-
ferent branche's in the literature.
Step 1. First we show that there exists L'.E{<, '>-}, such that for
all x,yEX we have:
xi -<i Yi for all iEU, ... ,n} => xL'.y.
Suppose the implication does not hold. Then there exist X,YEX
satisfying xi <iYi for all i but xrJy. Let x,yO be arbitrary.
Since by assumption (X i ,4 i ) has at least 3 indifference classes,
we can find u,v,zEX satisfying the following relations:
Zi ~i u i .( i v. , for xi .(i Yi
~~ Ch 0 0 s e L'.E { .t..., :>} a c cor din g to s t e pl. Fix t his L'. and calla
subset f/I"'I~ U, •.. ,n} decisive (cf.[l],[2]) if there exist
x,yEX such that the following is true:
si(x,y)=l for all iEI and Sj(x,y)=-l for all HI =>XL'.Y.
Due to stability of ~ , for I decisive the above implication holds
for any x,yEX. By step 1, {l, ... ,n} is decisive. Hence there exiSts
a minimal decisive subset I of {l, ... ,n}. We shall show that I is
a singleton. Suppose the contrary, that is there exists an hEI
with I\{h} '" f/I. Since (Xi'~i) has at least 3 indifference classes,
we can find x,y,zEX satisfying the following relations:
191
Zh -<h xh <.
h Yh for the h choosen
Xj ..( j Yj -< j Zj for all jEI\{h}
Yk -< k zk -< k xk fa)\' all kE{l, ... ,n}\I (which may be 0).
Concerning x,y we have Sj(x,y)=l for all JEI and sk(x,y)=-l for all
keI. Hence XAY because I is decisive. Concerning X,Z we have
Sj(x,z)=l for all jEI\{h} and Sj(x,Z)=-l for all kU\{h}. Since I
is a minimal decisive subset, I\{h} can not be decisive and we thus
can not have XAZ. With respect to y,z finally we have sh(z,y)=l and
Sj(z,y)=-l for all J*i. Since I is minimal, {i} can not be decisive
and we can not have ZAy. Thus we arrive at XAY, not XAZ, not ZAY.
For A=<" this amounts to x..t..y, z{x, y~z which contradicts transi-
tivity. For A=>- we have y.l.,x, x{ z, z4y which too contradicts trans-
itivity. Therefore we must have I = {h} and I is a singleton.
In that case the theorem is proven with m=l. Otherwise there exist
x,yEX such that x a(l),va(l)Ya(l)butX{Y. Lets for simplicity assume
that a(l)=l. Choose aEX l and define for U,VEX'=X 2X... xX n
u4'v if and only if (a,u)~(a,v) .
Obviously 4' is a quasi-ordering on X'. Defining si on X' for
i=2, ... ,n analogously to si on X it follows that si(u,v)=si«a,u),(a,v»
for U,VEX'. Since ~ is stable, 4' must be stable too. Furthermore,
~' is not trivial for the following reason. For x,yEX choosen above
we may write x=(x 1 ,u),y=(Y1'v) with u,vEX'. Because of Xl ~lY1 the
stability of=, implies that (xl'u)..«Y1'v) holds iff (a,u)«a,v)
holds. Now x<y, and hence u <'v. Thus ~' is not trivial. There-
fore the beginning of step 4 we may apply also to '" on X'. Thus
we may conclude as follows. Let x,yEX and write
x=(x 1 ,u), y=(Yl'v) with u,vEX'. If Xl -l Y1 then stability of ~
implies that x-<y is equivalent to u -<'v. Therefore from what we
had at the beginning it follows that (with a(l)=l)
x<y <=> (x 1 l:1(1)l Y1) or (xl "'lY1 and u <'v). If applied to ~'
on X' we obtain a(2)E{1, ... ,n}\{a(l)} and ~(2)E{.(,,>}such that
u <'v <=>(Xa(2)~~H/a(2» or (x a(2) ""a(2)Ya(2) and u -('v).
Putting together (and inserting a(l) for 1) we obtain that x<y holds
if and only if
(Xa(l)~(~~l)Ya(l» or (xa(l)~a(l)Ya(l) and xa(2)~(~~2)Ya(2»
and X
a(2)
~ y
a(2) a(2) and u .( 'v).
This process we may continue until we arrive after m iterations
with m~n at a trivial quasi-ordering. This completes the proof of
the theorem.
References
[1] Arrow, K.J., "Social Choice and Individual Values", 2nd ed.,
Yale University Press, New Haven and London, 1963 (first
edition 1951).
[2] Fishburn, P.C., Axioms for lexicographic preferences, Rev.
Econ.Studies 42 (1975), 415-419.
[3] Fuchs, L., "Partially Ordered Algebraic Structures", Pergamon
Press, Oxford etc., 1963.
[4] Krause, U., Recht und Gerechtigkeit, Okonomie und Gesellschaft
Jahrbuch 2: Wohlfahrt und Gerechtigkeit (1984), 158-172.
[5] Roberts, F.S., "Measurement Theory", Addison-Wesley, Reading,
1979.
193
Roberto Lucchetti
Department of ~1athematics, University of Milano
Via Saldini 50, 20133 Milano, Italy
Introduction. The paper has two goals: to revisit the concept of well
posedness in some fields where it has already been defined and to give
a tentative definition of well posedhess for vector optimization pro-
blems. As far as the first part is concerned, I shall refer to other,
more specific papers, for a deeper analysis: here I want only to put a
few of properties of well posedness in some evidence: in particular I
shall spend some words about the ordinal character of the property, star-
ting from Patrone 1987, because of the originality of this approach,
that seems to me of some interest. This analysis leads to a definition
of well posed problems that does not involve the given objective fun-
ction, but only the preference system with respect to which the opti-
mization process must be done. To be more detailed about the organiza-
tion of the paper, section two will deal with minimum problems, section
three with variational inequalities and section four with saddle point
problems. The first part is then terminated with section five, where
some words are spent to find a common origin between previous defini-
tions and some comments are presented about the Nash equilibrium pro-
blem: this is the starting point for the second part, which is develo-
ped in section six. Here a tentative definition is given for well posed-
ness in vector optimization problems, furthermore I shall present some
simple examples and initial remarks. A deeper study of the definition,
a further analysis showing how these (and other) properties here pre-
sented for minimum problems can be translated in the vector optimiza-
tion case will be made somewhere else.
mes (P) will also denoted by (f,K) to emphasize the data of the problem.
Definition 2.1 {Tyhonov 1966) The problem (f,K) is (Tyhonov) well posed
if:
there exists Xo E K such that f{xo)~f(x) ~ x E K and
v {x } E K such that{f{x »+ f{x o ), then{x }+ xo.
n n n
For motivations,comments and further references see,for instance, Nas-
hed 1982,Lucchetti 1983,Dontchev-Zolezzi:here I make only the following
remarks: the definition actually requires uniqueness of the minimum po-
intra definition of generalized wellposedness (g.w.p.) can be given as-
king that every minimizing sequence has a subsequence converging to a mi-
nimum point;this implies compactness of the solution set:see for instan-
ce Lucchetti 1982.Finally an alternative definition is given in Bednar-
czuk(1984) requiring only upper semicontinuity of the level set multi-
function of f , restricted to the set K.This definition is very close
to g.w.p. as it easy to see. Observe also that in definition 2.1 one
can pay attention only to sequences{x} such that{f{x )}is decreasing
n n
to f (x o ) •
Example 2.2. (Rockafellar 1970). Let f: mn + (_oo, +~ be a convex,lo-
wer semicontinuous function,with one and only one minimum point.Then
(f,mn ) is T.w.p .. Observe that this result can be false for an analo-
gous function defined on an infinite dimensional Hilbert space.
Example 2.3. Let X be a Banach space,x E X,K a closed convex subset of
X. Let flu) = iu- xl. This is the so called best approximation problem.
It is easy to show the following statements:
if X is reflexive,then f has a minimum point on K, for every x and K,
if X is strictly convex,then f has at most one minimum point on K;
if moreover the norm on X is Kadec,that is{x}+ weakly x,{lx I}+ Ix
n n
imply {xn}+x, then (f,K) is T.w.p.
Less easy,but very relevant,is showing the o~posite result: namely if the
best approximation problem, for a fixed x E X,is T.w.p. on every closed
convex set K, then X necessarily fulfills the three mentioned proper-
ties{Holmes '1972, where T.w.p. is called strong solvability). There are
many characterizations of T.w.p.,mainly in the convex case: many of
them are collected, for instance, in Lucchetti 1983. Here I recall only
one of them,both for its elegance and simplicity and for later uses.
196
Theorem 2.4 (Furi-Vignoli 1970) Let X be a complete metric space and f:X
7m lower semicontinuous and bounded from below. Then diam {x : fIx) ~
nuous function with one and only one minimum point on every closed con-
vex set K. Suppose that for every {K } converging to Ko in the Hausdorff
n
sense we have that {m } converges to mo. Then (f,K) is T.w.p. for all K.
n
Theorem 2. 10 (Lucchetti 1984) Let f :XxP +m be uniformly continuous
on bounded sets,f(.,p)a convex function for every p,let K :P +X_be a lo-
wer and upper semicontinuous multifunction at the point po.Consider the
problem (Pp) : minimize f(.,p) on Kp.
Let Mp the solution set of (Pp). If (f(.,po) ,Kpo) is T.w.p.,then M is
upper semicontinuous at Po.
Some comments about previous theorems are needed.The first one shows
how T.w.p. can improve convergence of solutions: it is a standard argu-
ment that in the hypotheses and notations of theorem 2.8,without T.w.p.
it is only possible to get weak convergence of m to mo 1hence T.w.p.
n
improves it to strong convergence.Theorem 2.9 is a kind of converse of
theorem 2.8 since it shows that if we want strong convergence of the
solutions,then we must deal with T.w.p. problems.Both theorems 2.8 and
2.9 have,in my opinion, an important theoretical meaning;namely they
show a connection between T.w.p. and continuous dependence on the da-
ta under perturbations(sometimes this property is called Hadamard w.p.).
Theorem 2.10 can be considered as a similar result,but with a different
emphasisjhere T.w.p. is assumed only for the limit problem PPo:hence
the result can be considered as easier to handle. Moreover it clarifies,
in the (restricted) setting previously presented,famous Berge's theorem
(Berge 1963) and its successors,as it shows how T.w.p.,rather than com-
pactness of the constraint set, plays a central role in the upper semi-
continuity property of the solution set. By the way,obs~ethat T.w.p.
is a property of both the function and the constraint set,considered
together. Previous theorems are only some examples of results in this
direction:some more are in the papers already cited and, for instance,
in Lucchetti1982,Bednarczuk1982,1984.Some consequences of theorem 2.8
applied to the best approximation problem are presented in Lucchetti1985.
The next aspect related to T.w.p. I shall consider here is the follo-
wing:how many T.w.p. problems are there in some specified classes of
198
minimum problems? Let us call GURn ) (or simply G) the set of functions
f: mn + [_00 , +00) which are convex and lower sernicontinuous. On G we
consider Kuratowski convergence.It is well known that G then becomes a
metrizable compact space. Inside G we consider the set N = {f £ G such
that f is Frechet differentiable and (f(.) _(y,.)~n) is T.w.p. for e-
very y£ mn }.
Theorem 2.11 (Patrone 1986) N is dense in G and G-N is a first catego-
ry set.
Last theorem then shows that a very particular subset of T.w.p. func~
tions is a "large" set inside G. The paper proving the theorem was ori-
ginated by a previous one (Lucchetti-Patrone 1978) where it is possi-
ble to find other results in this direction, for different classes of
problems,and also the infinite dimensional case is investigated:it
should be noticed that here Kuratowski-Mosco convergence does not pro-
vide the same strong result. Last property I present in this section
is the so called ordinal property. I mention that in game theory or in
mathematical economics we often have some preference systems, among
the data of the model, (defined on certain sets),rather than a repre-
sentative function of these preferences. Then it is not a nonsense to
ask whether T.w.p. of an optimization problem depends or not on the par-
ticular function we choose to represent the preference system. We have
the following:
Theorem 2.12 (Patrone 1987) Let X a connected topological space , ~ a
total preorder on X, f,g : X +m continuous functions both represen-
ting :$ • Then (f,X) is T.w.p. if and only if (g,X) is T.w.p.
The connectedness hypothesis can be accepted without any trouble, whi-
le continuity seems to be rather unnatural: in minimum problems it is
lower semicontinuity that people accept without comments. But it is an
obvious remark that if f has on X one and only one minimum point xo'
then we can substitute the problem (f,X) by the equivalent problem
(g,X) where g(x) = fIx) for every x r Xo and g(x o ) = f(xo) - 1. We do
not change the underlying preference system ,we do not destroy lower
semicontinuity and we get a (trivial!) T.w.p. problem. However this is
a rather stupid way to go on. We can observe the g has the (strange)
property that every minimizing sequence eventually coincides with the
minimum point. Then, from now on, we consider problems without this
199
This section and the next one are really a quick overview of w.p. in
variational inequalities and in saddle point problems; this is due to
two main reasons: these topics are not so strictly related to vector
optimization as the minimum problem ;moreover they are not so widely
studied and of cornmon knowledge. It should be at first observed that
the minimum case is in both extensions at the basis of the given defi-
nitions. To state the problem let X be a reflexive Banach space and K
a closed convex set of X; let A: X ... X* be a monotone operator and let
f:X .... JR be a convex function. Here the problem (V.I) is to find
x E: K such that (Ax, x-y) + f (x) - f (y) :;; 0 for every y E: K.
Let H (a) ={ x E: K : (Ax,x-y) + f (x) - f (y) :;; a Ix - y I for every y }
Then we have the following:
Definition 3.1 (Lucchetti-Patrone 1981) We say that the problem (V.I)
200
is w.p. if H(a) # ~ for every a > 0 and diarnH (a) ... 0 as a .... o.
It is apparent that theorem 2.4 inspired the definition, if the sets
H(a) are considered as a kind of level sets for the variational inequa-
lity. Furthermore it is well known that a convex minimum problem can
be stated in terms of a variational inequality by means of the deriva-
tive of the functional to be minimized: hence next theorem can repre-
sent a further motivation for the given definition:
Theorem3.2 (Lucchetti-Patrone 1981) If A = g' for some function g which
is convex, lower semicontinuous and Gateaux differentiable and if f + g
is bounded from below, then (V.lj is w.p. if and only if (f+g,K)is T.w.p.
Next, I present one example and two theorems on w.p. variational ine-
qualities. Both theorems deal with a linear operator : the first one
exibits some conditions which are equivalent to w.p.,while the second
one focuses attention on an equation, rather than an inequality. Namely
(V.I) reduces to find x such that Ax = f when the set K is the whole X
and f is linear.
Example 3.3 (Lucchetti-Patrone 1981) If A is hemicontinuous and stron-
gly monotone, if f is convex lower semicontinuous and not identically
+ 00 , then (V.I) is w.p ••
Theorem 3.4( Lucchetti-Patrone 1982-83) Let X be a Hilbert space and A
a linear bounded operator. Let f E: X. Then the three following conditions
are equivalent:
(V. I) is w. p. on every closed convex set K
- (V. I) has one and only one solution on every closed convex set K
- A is coercive( namely there is a > 0 such that (Ax,x)~ alxl 2 for all x)
Theorem 3.5 (Lucchetti-Patrone 1982-83) Under the same hypotheses of the-
orem 3.4, suppose moreover that K = X • Consider the problem: find x E: X
such that Ax = f (as a V.I) . Then the following are equivalent:
- A is a homeomorphism
(V. I) is w.p.
for every g E: X,Ax = g is w.p. (as a V.I ).
Observe that this last theorem deals, as theorems 2.8,2.9,2.10, with
the problem of relating well posedness with continuous dependence on
the data of the solution sets:here this fact is stated by means of the
continuity of the inverse operator A- 1 .
201
In this section I shall have in mind the saddle point problem as a spe-
cialization to the zero sum case of t~e Nash equilibrium prob~em of a
two players non cooperative game. This can give the idea of extending
concepts to this last case: in fact something is done, but in my opi-
nion the general case must be studied more deeply; so here I confine
myself to the zero sum case. Next section will contain some remarks on
the Nash problem. As usual, some terminology is needed. Let X,Y be me-
tric spaces, f: XxY +~ a given function. The saddle point problem (S)
is to find (x,y) EXXY such that f(x,y);'; f(x,y);'; f(x,y) 'i XEX,'i YEY.
Let h(x) = inf f(x,y), g(y) = s~pf(x,y) and w = g - h. Observe that
y
Wf;O; a sequence {(x ,y )} is said maximinimizing if {w(x ,y )}+ O.
n n n n
Finally let E (a) = { (x,y) E XxY: f (x,y) f; sup f (x,y) - a,f (x,y);'; inff (x,y)
x
+a } •
adjoint ,-A and B are positive, a and b are elements of X and Y respec-
tively. Let f(x,y) = (Ax,x) + (By,y) - 2(a,x) - 2(b,y) + (x,My).
Then the following statements are equivalent:
(S) is w.p. on every HxK,Hc::: X, K c: Y closed convex sets
- (S) has one and only one solution on every HxK
- -A and B are coercive operators.
This is the analogous of theorem 3.4 in previous section. Next theorem
deals with the ordinal property already investigated in the minimum case.
Theorem 4. 5 (Patrone 1987) Let X, Y connected,~ a total preorder on XxY,
f1 and f2 continuous functions both representing ~ • Then (S) with f1
is w.p. if and only if (S) with f2 is w.p .•
I want to conclude this part by saying that perhaps it should be of so-
me interest to investigate density and genericity properties for w.p.
(V.I) and (S) problems and connections between w.p. and continuous de-
pendence on the data: I do not know results in this direction.
I draw here some conclusions arising from previous sections and I in-
troduce w.p. in the context of vector optimization problems (V.O ): de-
finition of w.p. for (V.O) and some consequences are object of the next
section. To find out a possible schema for w.p. we can see that there
is the introduction of a notion of (value of the problem and of) appro-
ximate solution sets, with increasing precision in the approximation.
Hence w. p. will mean:
there is a solution of the problem and
every sequence of approximate solutions converges to the solution, when
the error in the approximation goes to zero.
An alternative to the second condition is:
I do not want to introduce here too many notations and too heavy termi-
nology: as we only give an attempt to start with the argument, we can
confine ourselves to a simple setting. Thus (V.O problem can be sta-
ted in such a way: suppose we have a function f: X -+- Y, where X, Yare
204
Banach spaces and there exists a closed convex cone C in Y which gene-
rates a preorder in the natural way: x ~ y iff y-x E C. Given a set K c
C X the vector optimization problem is: to find
x ~ K such that (f (x) - C ) " f ( K ) {f (xl).
The first step for giving a definition of w.p. is to define inf f(K).
Given a set AcY the most natural idea for defining the set inf A is
to keep all the elements which are minimal for the closure of A; namely:
Definition 6.1 x E inf A if x E A and there exist~ no y f x such that
x- YE C.
References
Leonid I. Polishchuck
Institute of Economics and Industrial Engineering
Siberian Branch of the USSR Academy of Sciences
Prospect Lavrentjeva 17, 630090 Novosibirsk 90, USSR
1. Introduction
{x E X I (2)
209
Vi 1, •.. ,n,
We indicate here two most typical examples. The first one relates
to the social welfare problem (hereafter SW-interpretation), when
variables xi are comparable quantitative welfare characteristics (say,
incomes) of n individuals. As a second example can serve the problem
of decision making under uncertainty (UDM-interpretation): in this
case the components of x represent the performance characteristics of
stochastic system in n potentially possible and equiprobable states
of nature 1 ).
n
~ 1 {~(x1, ••• ,xn) = L: f(x i ) I f : R .... lR is a
i=1 (3)
monotonically increasing function}.
where 11 = (1, •.• ,1) and x,y are the permutations of x,y components
respectively such that 1 ~ x
relation also is a special case of stochastic dominance: L = R~ with
2
n x.
{(j) (x) =1: f (x.~1I) I f R + R is a convex
i=1 (6)
continuous function}.
efficiency m(x) = *
guided in line with the Lorenz domination by the average (expected)
(x·1I).
(7)
D(X*) (8)
individual by /:, > 0, the income of the j-th individual may increase
not more than by (Pi/Pj)/:'. With regard to what has been said the
variable Pi can be considered as a comparative characteristic of
mutual efforts necessary to ensure the i-th individual's welfare.
p.x*>O. ( 11 )
p'x* ~O
5) • o (11 ' )
*
~ ~+ j 1, ... ,n. (12 )
x • 11 x*· ~
Let
xO = {p EJR~! p. x ~ 1, v x EX} ( 1 4)
nondominated:
215
t - max,
rda) x = at, (15 )
X E X9).
n
f(x) E (17)
m(x) i=1
f(x) = ( 1 8)
m(x)
n
f(x) E ( 19)
m(x) i,j=1
m(x,z) = z (20)
Notes
5. This proposition and the next ones are given here without proofs
which can be found in [8].
models [17].
References
Paolo Serafini
University of Udine
Department of Mathematics and Computer Science
Via Zanon 6, 1-33100 Udine
1 INTRODUCTION
In the field of vector (or multi objectivel optimization there has been a relatively little
interest in solving combinatorial or discrete problems. During the 70's just a few papers
have been published on multi objective (m.o.) integer linear programming. However no
special emphasis was put on the important aspect of computational complexity. This can
be certainly ascribed to the fact that the theory of NP-completeness was developing at a
fast pace in those same years.
A pioneering paper by P. Hansen 161 appeared in 1980; he attacked the shortest path
problem with two objectives (actually several versions of it) with a truly computational
attitude. Following that paper only a few papers appeared adopting the same perspective,
almost always on shortest path problems (17,17,13,14,81). Other papers on m.o. integer
programming can also be quoted (11,21,4,9,16,31).
From all these results one receives the impression that the computational complexity
of m.o. problems has not been formulated yet in precise terms in spite of the fact that real
problems naturally expressed as combinatorial m.o. problems are rather common. Just to
quote an example consider problems defined on graphs with costs and failure probabilities
associated to the graph components, and requirements of finding a certain sub graph with
low cost and high reliability.
It is also important to define precisely the basic concepts in order to increase the
'awareness' of anyone facing a m.o. combinatorial problem, as it has already happened in
scalar optimization after the development of the theory of NP-completeness.
Hence the aim of this paper is at making some preliminary considerations which can
be helpful in facing m.o. combinatorial problems with a computational perspective.
223
:4 GENERAL CONSIDERATIONS
given I:X _ Zm
"find" minimal elements in the subset F f X
always exists and sometimes can be easily computed. In case the number K enters a
complexity measure, it is useful to remember that ILl = 0(log2 K) and hence K = 0(2ILI).
One consideration seems to be particularly important: the number m of objectives
enters the encoding L and it is likely to be present in complexity measures. It is crucial
to note that, in contrast to the asymptotic meaning of a complexity measure, m is never
too large, least of all tends to infinity. Of course we are speaking of those m.o. problems
with truly incomparable objectives and not, for instance, of those problems facing the
minimization of a vector in R m as a m.o. minimization on all components, in which case
m can be rather large. Therefore we will consider m as a fixed parameter.
Now the original problem will be particularized into a list of several versions, each one
ofthem giving a precise meaning to the term "find". Our aim is at examining these versions
and at possibly selecting one of them as a standard version to measure complexity. This
list comprises nine versions and by no means pretends to be exhaustive. Other significant
versions could be envisaged and attached to the list.
Given any instance (I, F)
Vi : find all minimal elements;
224
V2 : find all minimal values and at least one minimal element for each minimal value;
Va : find all minimal values;
V. : find at least p > 1 minimal values or, in case p minimal values do not exist, report
this fact and find all minimal values. Furthermore find at least one minimal element
for each found minimal value;
Vs : find at least p > 1 minimal values or, in case p minimal values do not exist, report
this fact and find all minimal values;
V6 : find one minimal element;
V7 : find one minimal value;
Va : given z e zm, does there exist z e F: /(z) ~ z?;
Vg : given z, we zm, W > 0, does there exist z e F: w/(z) ::; wz?
Vl is sometimes regarded as the genuine m.o. problem and consequently some state-
ments about computational complexity have been derived with respect to this version for
particUlar problems (for the Shortest Path problem see [6,8]). As a matter of fact this is
always an intractable problem: just consider a constant objective function and Vl calls
for a complete enumeration of the feasible set F. However, and this is a point apparently
missed by the above quoted authors, Vl is not the m.o. counterpart of a scalar opti-
mization problem, which requires the computation of only one minimal element among
possibly many minimal elements, all of them with the same objective function value. In
this sense it is V2 the natural counterpart of a scalar optimization problem.
Quite often it is realized that the computational effort demanded by V2 is too high
and so one is just trying to solve V4 or even V6 • In particular, the parameter p in V4
should be regarded as part of the instance description. Theoretically no a priori upper
bound exists for p and this makes V4 not so different from V2 (as we shall see soon).
Va, Vs and V7 are usual evaluation versions.
The remaining Va and V9 are recognition versions. In particular Va is the exact
counterpart of the recognition version of a scalar problem (see [151), whereas V9 is the
recognition version of a scalarized m.o. problem.
It should be noted that V2 and Va can be intractable as well for the simple reason they
could easily require an exponential amount of time. In fact the number of nondominated
values is 0(Km-l) in the worst case, where K is the bound previously introduced, and
computing all these values requires obviously a proportional amount of time. Therefore,
unless K is a polynomial function of the input size, no algorithm designed to solve V2 or
Va can be efficient. This dependence on the range of the objective function is an intrinsic
feature of multi objective problems. In view of these observations we shall assume that
the bound K is a polynomial function of the input string length ILl.
Then it is possible to prove some polynomial reductions between some versions:
Va oc V2 , Vs oc V., V7 oc V6, V4 oc V2 , Vs oc Va, V6 oc V2 , V7 oc Va, V6 oc V41 V7 oc Vs ,
Vg oc Va: easy to show;
V6 oc V7 : this is indeed possible in most combinatorial problems;
V2 oc Va, V4 oc Vs : as above;
V7 oc V9 : by binary search;
V2 oc Va : consider a box with side lenghts K containing all feasible values. Then by binary
search along one coordinate, keeping fixed the other coordinates (as in the constraint
method) it is possible to determine a nondominated value. Hence O(K m - 1 log2 K) calls
of Va are required in the worst case. Recall that this reduction is in any case of complexity
O(Km-l);
225
if the approach to a m.o. problem does call for some interaction with the decision maker,
it is sensible to expect V8 playing a fundamental role in soliciting the decision maker's
preferences.
Hence Irom now on, when speaking 01 a m.o. optimization problem, we wall alwa1ls
reler to Va.
Of course a problem known to be NP-complete when m = 1 will remain NP-complete
for m > 1, if, as it usually happens, it will still be in the class NP. The unfortunate fact is
that most polynomial single objective problems become NP-complete for m > 1. This will
be matter of investigation in the next sections. In particular we shall investigate the role
of total unimodularity and matroidal structure which are the most common structures
leading to polynomial algorithms for single objective problems.
3 TOTAL UNIMODULARITY
V(P) ~ F ~ P (1)
where V (P) is the set ofvertices of a polyhedron P defined by a set of linear inequalities.
Then there is no difference in solving either the combinatorial Problem A:
min ex
s.t. x E F
min ex
s.t. x E P.
For instance, if F is a subset of zn, (1) holds iff the linear inequalities defining P
form a totally unimodular matrix with arbitrary integers as r.h.s. coefficients [20].
At first glance it would seem that the inclusions (1) benefit m.o. problems as well,
since they refer to constraints only. However this is not the case.
Let I(x) = Ox and let Q be the conical hull of the rows of OJ let Q* be the negative
dual of Q (i.e. Q* := {x : xy ~ 0 Tty E Q}). Then x is minimal for Problem A if and only
if (x + Q*) n F = x. Similarly x is minimal for Prolem B if and only if (x + Q*) n P = x.
The latter condition is stronger in general if Q* is not a halfspace, that is if there are
more objectives. Consider the following simple example: F = {0,1?, Cl = (-l,S) and
C2 = (1, -4). Then Q* is the conical hull of (S,l) and (4,1). It is easily seen that the
origin is nondominated for Problem A and dominated for Problem B.
Another property of scalar linear programming inherited by some combinatorial prob-
lems, i.e. the fact that neighbourhoods of adjacent vertices are exact (local minima are
necessarily global minima) no longer holds. Consider a simple example with F = {O, I?,
Cl = (-2,1) and C2 = (1, -2). Then Q* is the conical hull of (1,2) and (2,1) and so the
origin is locally but not globally optimal.
These considerations have a serious impact on the type of properties and results
needed when dealing with more objectives. In fact a current line of research consists in
227
characterizing the polyhedron of feasible solutions (as done by Edmonds for the match-
ing problem or by other authors for the Travelling Salesman Problem [12]) so that the
inclusions (1) hold, at least near the optimum. But, for m.o. objective problems, this
is a useless effort, as explained by the previous considerations and also by the following
general statement.
Consider the m.o. problem
"min" Cz
s.t. z e F
where V(P) f;; F f;; P and P is a polyhedron assigned through linear inequalities.
Then this problem is NP-complete, since it has binary knapsack as a particular case:
c1z ~ Kl ( objective)
-r:'z ~ -K2 ( constraint )
ze{O,I}'" ( F = V(P) with P the unit cube)
Note that
"min" Cz
s.t. z e V(P)
is NP-complete as well and this result casts a shadow on the idea of looking for minimal
elements of m.o. linear programming among the vertices of the feasible polyhedron, an
idea which has received a considerable attention in the past.
Since the above statement is a general one, the question arises whether particular
problems exploiting total unimodularity for their solution, like Shortest Path and Assign-
ment, are still difficult. The negative answer is that both are NP-complete for m > 1,
even with nonnegative costs.
We shall limit ourselves to consider first the Biobjective Shortest Path and then the
Biobjective Assignment, both with nonnegative costs. As mentioned in the introduction
the Biobjective Shortest Path has been investigated rather extensively ([6,8,14]).
Biobjective Shortest Path is equivalent to Constrained Shortest Path; as the latter
is claimed to be NP-complete ([51 p. 214) Biobjective Shortest Path is NP-complete as
well. However the reference provided in [51 is a private communication and so it may be
interesting to give here a simple proof of this fact.
The proof goes through a transformation from the following Binary Knapsack:
"given c1 ~ 0, c2 ~ 0, K 11 K 2, does there exists z e {O, l}n such that c1 z ~ K 1 and
c2z ~ K2 7"
Consider now the network with nodes Vo, ... , V", and a pair of arcs from Vk - 1 to Vk
for k = 1, ... , n with costs (c~, 0) and (O,~) for each pair. Then the question "is there a
path from Vo to Vn such that c1z ~ Kl and c2z :5 Ek c% - K2 7" is equivalent to Binary
Knapsack.
It is immediate to see that the problem is in NP and so it is NP-complete. However
things are not so bad because a pseudopolynomial algorithm is immediately suggested for
it from the previous transformation from a pseudopolynomial problem. The algorithm we
shall describe can be applied to those instances with nonnegative costs for at least one of
the objectives.
228
4 MATROIDS
Let (E, M) be a matroid and M' be the family of maximal subsets (see [11]). Let
c : E -+ Z be given costs. It is well known that the problem
min" c(e)
AEM'L.t
eEA
Let us now consider the biobjective case: «given cost c : E -+ Z2, bound K E Z2 is
there a maximal A E ,M' s.t. EeeA c(e) ~ K 1"
This is clearly an NP problem. Its NP-completeness it easily proved by recalling that
linearly constrained matroidal optimization problems are NP-complete (see !2j) and by
observing that these problems are equivalent to biobjective matroidal optimization.
Besides this result there is another negative fact. Note that E is partially ordered by
the cost and one is tempted to apply the greedy algorithm to any topological sorting of
E in order to find a nondominated solution (recall that a topological sorting of E is any
total order < on E such that c(a) ~ c(b) ~ a < b). Unfortunately the following holds:
Fact 1 : the greedy algorithm applied to a topological sorting of E does not neces-
sarily output a nondominated solution.
Proof: Just consider as a counterexample the graphic matroid for the complete graph
K. with costs C(V1,V2) = (3,0); C(V1,VS) = (0,3); C(V1,V.) = (1,1); C(V2,VS) = (4,4);
C(V2,V.) = (2,2); c(vs,v.) = (2,2). The greedy algorithm applied to the topologi-
cal sorting (V1IV4), (vs,v.), (V2,V.), (V1,VS), (V1IV2), (V2,VS), gives the spanning tree
{(V1, v.)( Vs, V.)(V2, v.)} with total cost (5,5). However this solution is dominated by the
tree {(V1, V.)(V1, V2)(V1, vs)} with cost (4,4). Moreover, for arbitrary costs, the solution
can be arbitrarily bad. •
A converse of the above statement does hold however, and this result can provide
some insight into the problem.
Fact 2 : for any nondominated solution A there exists a topological sorting for which
the greedy algorithm outputs A.
Proof: we may associate to the set E a graph G in the following way: the nodes
correspond to the elements E and an arc exists from ej to ej iff c( ed ~ c( ej ). Obviously G
is a directed acyclic graph. Given a maximal subset A ~ E and its complement B := E\A,
we may form a new graph GA by adding to G the following set of arcs
FA := ((a,b) : a E C(b) n A Vb E B}
where C(b) is the circuit generated by b. Then proving the thesis is equivalent to proving
that G A is acyclic: in fact GA represents a partial order in this case and any topological
sorting of it is the claimed sorting.
Hence let us suppose that GA contains a directed cycle. Since G is acyclic, the cycle
must contain some arc of FA' More precisely, due to the fact that G is transitively closed,
the cycle may be assumed to consist alternatively of one arc of G and one of FA and so it
is of the form
with aj E A and bj E B. We may also suppose without loss of generality that this is the
cycle through a 1 with the fewest arcs.
This last property enables us to obtain a new maximal subset A' by swapping the ai's
with the bi's (that this is indeed possible can be proven by using the same arguments as for
the two matroid intersection algorithm !11], to which this construction is quite similar).
Moreover, since c(bd ~ c(aiH) for i = 1, ... ,k -1 and c(bk ) ~ c(at), A' dominates A
contrarily to our assumption. I
Thus we have identified two families of maximal subsets, namely M*, the family of
maximal nondominated subsets of E, and M+, the family of maximal subsets A for which
230
G A is acyclic (or equivalently there exists a topological sorting of E such that the greedy
algorithm outputs A). Moreover Fact 1 and Fact 2 have proved that the strict inclusion
oM* c oM+ holds.
Let us now consider also the family oM' of maximal subsets A which are locally non-
dominated, in the sense that they are not dominated by any other maximal subset obtained
by swapping an element b f. A with any a e O(b). Obviously if a subset A is locally dom-
inated GA does have a cycle. So the inclusion oM+ ~ oM' is proved. Furthermore the
inclusion is strict as proven by the following example: Let oM be the graphic matroid for
=
the graph (V, E) with V {Vl,V2,VS,V4,V5} and
Let the costs be Cn = (2,2), ClS = (1,5), C2S = (1,1), Cu = (3,3), C15 = (0,4), C45 =
(1,1). Let A = {(V2,VS), (Vl,VS), (VltV4), (V4,V5)}' Then A is locally nondominated but·
GA is not acyclic.
The interesting conclusion is that, even if the concept of maximal subset obtained via
the greedy algorithm applied to some topological sorting of E is weaker than the concept
of nondominated solution, it is anyway stronger than the one of locally nondominated
solution. The reason is that, as apparent from the proof of Fact 2, a larger neighbourhood
is considered by requiring acyclicity of GA. The important fact is that exploration of this
larger neighbourhood is not computationally expensive; in fact it can be carried out with
complexity O(lEI2).
If we consider other ways of ordering the elements of E, no better conclusions can be
drawn. Obviously the greedy algorithm applied to any lexicographic ordering of E does
produce a nondominated solution. However, roughly speaking, these solutions are not the
ones we are looking for, since they are by definition too much biased in favour of one
particular objective.
Let us also consider the ordering of E induced by IIclloo (break any tie via topological
sorting) or by IIc-clioo for some constant c. We recall that minimization of II/(:I:)-cil oo is a
well known scalarization procedure yielding all nondominated solutions by suitably varying
c. Nonetheless the greedy algorithm applied to this ordering of E does not necessarily
output a nondominated solution, as seen by reconsidering the counterexample provided
in Fact 1.
5 OONOLUSIONS
The previous sections showed that in m.o. objective combinatorial problems one is
typically faced with NP-hard problems. Hence it is almost compulsory to attack these
problems through heuristics combined with clever implicit enumeration techniques.
But this is not the only reason. In fact the modelling itself of a real problem through
a m.o. mathematical model is due to a lack of information about the actual preferences of
the decision maker. Therefore the interest in having guaranteed optimal solutions should
shift in favour of having a rather large number of acceptably 'good' solutions.
Moreover the common approach of getting solutions through some kind of interaction
with the decision maker is also to some extent a heuristic procedure.
It is perhaps convenient to bring together these two kinds of solution approximation
into a unique procedure. In other words one should direct the ideas for assessing the
231
decision maker's preferences into the procedure generating solutions. For example, since
generating solutions of a combinatorial problem usually corresponds to visiting a very
large search tree, possible cutoffs of the search tree could be also controlled by heuristic
demands for speeding up the computation and by the interaction with the decision maker.
For linear discrete problems which are polynomial in case of a single objective, it may
be convenient to solve the scalarized version V9 (a weighted linear combination) and then,
once discovered some nondominated solutions on the boundary of the convex hull of I(F),
to build some implicit enumeration scheme to 'peel off' the external strata of I(F) in
order to discover other hidden nondominated solutions. Again this can be done through
some interaction with the decision maker.
This is a research field almost unexplored at the time. A preliminary work in this sense
exploiting duality concepts developed by the author can be found in [18,19]. Certainly this
kind of research would turn more into a patient handcraft work on particular problems
rather than produce general theoretical breakthroughs. However the practical utility of
such an approach should be quite promising and worth the effort.
6 REFERENCES
(1) Bitran,G.R., Theory and algorithm. for linear multiple objective program, with zero-one vari·
able" Math. Progr. 17 (1979) 362-390.
(2) Camerini,P.M., and Vercellis,C., The mafroidal knap.ack: a claBB of (often) well·.olvable
problemr, Operations Research Letters 3 (1984) 157·162.
(3) Chalmet,L.G., Lemonidis,L., and Elzinga,D.J., An algorithm for the bicriterion integer pro·
gramming problem, European J. of Operations Resear (1986) 292·300.
(4) Dyer,M.E., and WaIker,J., A note on bicriterion programming, Math. Progr. 27 (1983)
355-361.
(5) Garey,M.R., and Johnson,D.S., Oomputer. and intractability, Freeman, San Francisco, USA,
1979.
(6) Hansen,P., Bicriterion path problemr, in Multiple criteria deciBion making, Fandel,G., and
Gal,T., editors, 109-127, Springer, Berlin, 1980.
(7) Hartley,R., Survey of algorithm. for vector optimi.ation problemr, in Multi.objective deci.ion
making, French,S., et aI., editors, Academic Pw York, 1983.
(8] Henig,M.I., The ,horte,t path problem with two objective function" European J. of Operations
Research 25 (1985) 281·291.
(9) Ignizio,J.P., Multiobjective mathematical programming via the Multiples model and algorithm,
European J. of Operations Research 22 (1985) 338·346.
(10) Kung,H.T., Luccio,F., and Preparata,F., On finding the ma:tima of a .et of vector" J. of
ACM 22 (1975) 469-476.
(11) Lawler,E., Combinatorial optimization: network, and matroid., Holt, Rinehart and Winston,
New York, 1976.
(12) Lawler,E., Lenstra,J.K., Rinnooy Kan,A.H.G. and Shmoys,D.B., The traveling .aleBman
problem, Wiley, Chichester, UK, 1985.
(13) Martins,E.Q.V., On a multicriteria .horte.t path problem, European J. of Operations Research
16 (1984) 236-245.
232
Alicia Sterna-Karwat
Department of Mathematics, Monash University
Clayton 3168, Australia
1. Introduction
if Xl - ~ € c\2(C).
Thus, according to the definition of efficiency, if A £ X, e € A is
efficient in A if X € A and X - e € C together imply e - X € C [2,3]. The
set of the efficient points in A will be denoted by E(A;C). Such points are
described in [1,10-12] as maximal up to indifference.
such that the remaining part C2 ~ C is still a convex· cone and satisfies (i).
Moreover. the removed part C1 is not a vector subspace but its closure is a
1.2 Definition. We say that a convex cone C belongs to the class ~ if the
following condition holds: for every vector subspace L contained in X. C nL
is a vector subspace whenever C n L is a vector subspace.
L ~ 2(C) [11].
Proof. Suppose that 2(C) = C n 2(C) and C (~. Then there exist a vector
The converse of Proposition 2.2 is not true (see the last paragraph in this
section).
The following theorem proved in [11] shows that an arbitrary convex cone can
be uniquely decomposed as a union (hence as a sum) of a convex cone from the class
~ and a convex cone whose closure is a vector subspace. which meet at the origin.
236
a ~ a(C). Put K(C) =~(C)(C). Lemma 2.1 in [11] ensures that K(C) is a
K(C) = C1 . Consequently, using (i) and (iv) for both decompositions, one obtains
iCC) £ K(C) £ n i(C l.) £ iCC). Hence K(C) = i(C), which again by Theorem 2.4,
i € I
means that C € '€. []
2.6 Corollary. If C is a convex cone such that C\{O} is evenly convex (i.e.
C\{O} is an intersection of open half spaces) then C € '€.
The proof follows from Proposition 2.2 since the condition (+) implies
iCC} = C n iCC}. D
The proof follows from the fact that if K = UtB then K n i(K} {O} and
t~O
Moreover, let us observe that for (i) closed cones, (ii) cones satisfying
condition (+) of Proposition 2.7 (in particular for cones for which continuous
C-strictly positive functionals exist), (iii) cones with bases, (iv) cones whose
closures are pointed, the transfinite sequence {~(C}} stops at the first place,
i.e. K(C} = T(C} = C n iCC} = iCC). However, in Rn , the convex cone
C = Cl U ... ~,where k ~ n and
3.1. Theorem [12]. Let X be a locally convex t.v.s. and let C € ~. Then
239
wi th the norm IIxll = }; IXn I. then for every convex corie C and for every convex
3.2 Theorem [11]. Let X be a t.v.s., let C £ X be a convex cone and let
A £ X. Then
(i) E{A;C) = E{A;K{C» n E{A;p{C».
(ii) If C € ~, then E{A;C) = E{A; p{C».
References
[1] J.M. Borwein, On the existence of Pareto efficient points, Maths. Oper.
Res. 8 (1983), 64-73.
[2] H.W. Corley, An existence result for maximization with respect to cones,
J. Optim. Theory Appl. 31 (1980), 277-281.
[3] R. Hartley, On cone-efficiency, cone-convexity and cone-compactness,
SIAM J. Appl. Math. 34 (1978), 211-222.
[4] G.B. Hazen, T.L. Morin, Optimality conditions in nonconical multiple
objective programming, J. Optim. Theory Appl. 40 (1983), 25-60.
[5] M.I. Henig, Existence and characterization of efficient decisions with
respect to cones, Mathematical Programming 23 (1982), 111-116.
[6] R.B. Holmes, Geometric Functional Analysis and Its Applications,
Springer-Verlag, New York, 1975.
[7] J. Jahn, Mathematical Vector Optimization in Partially Ordered Linear
spaces, Peter Lang, Frankfurt am Main, 1986.
[8] J.L. Kelley, I. Namioka et al., Linear Topological Spaces, Princeton,
D. Van Nostrand Co., Inc., 1963.
[9] V.L. Klee, Convex sets in linear spaces, Duke Math. J. 18 (1951),
443-466.
[10] A. Sterna-Karwat, On existence of cone-maximal points in real topological
linear spaces, Israel J. Math. 54 (1986) 33-41.
[11] A note on convex cones in topological vector spaces,
Bull. Austral. Math. Soc. 35 (1987) 97-109.
240
Acknowledgement.
The author would like to thank Jean-Paul Penot for his comments which improved
the presentation of the paper.
Current address.
Department of Mathematics, The University of Newcastle, New South Wales 2308,
Australia.
PSEUDO-UTILITIES
Andrzej Wieczorek
Institute of Computer Science
Polish Academy of Sciences
P.O. Box 22, 00-901 Warsaw
The relation between the case of a single relation and the "parame-
tric case" resembles, to some extent, the relation between the usual
problem of the existence of a utility representation of a preference
relation and its "measurable" version, the latest considered by Aumann
[1969], Wieczorek [1980] and many others.
I. BASIC DEFINITIONS
if the formula above holds for a function U:X.xX ~ R then we say that
U is ~enerated by a utility u.
One can immediately see that the usual problem of finding a utility
representation of > can be split, by means of the previously defined
243
3. p(x,y) := if x > y
:= 0 otherwise;
Obviously, having passed through (1) and (2) we get a utility re-
presentation of > but notice that, generally, finding a utility re-
presentation of > does not have to go through (1) and (2). The fol-
lowing observation shows that this procedure may be sometimes even
misleading: even if > admits a utility representation u and p is
a pseudo-utility representation of >, p need not be generated by
u nor by any other utility representation of >. A trivial example
of this kind is the usual order > in the real line and its pseudo-
representation p defined by (3).
k
sgn p, max(p,O), P , for any positive integer k,
r~(x,y) := r(x,y).~(x>y)
For every tET, let X(t) = X and let >t be a binary relation in
X(t) such that the joint graph
~ := {(t,x,y) I x >t y }
is relatively open in
247
For every tET, let X(t) c: X be convex and let >t be a binary
relation in X(t) such that the jOint graph
0/ := {(t,x,y) Ix >t y}
is relatively open in
-
and (C.) of elements of
~
e such that
n U{K.xC.li
~ ~
1,2, ... }.
If for every tET and yEX(t) the set Ut(y) :={xEX(t)1 x >t y} is
convex then there exists a continuous function lP : ~ .... [0,1] such that
for every tET, lP(x,y) is a pseudo-utility representation of >t'
quasi-concave in x.
Since (*) has been already discussed, we only notice that (**) holds,
for instance, whenever X is also compact (see A.6,p. 31 in Wieczorek
[1986]).
is continuous;
d. u. is continuous, and
]
e. for every s E nisi the set
Mj (s) := {~I (s,~) E fj and uj(s with Sj replaced by~)
is contractible.
Conditions (c) and -(e) do not look very nicely but they were adopted
by Debreu just for greater generality. Notice that for every j, (b), (d)
and
c-. rj is lower semi-continuous;
imply (c). Since obviously
a-. S. is a compact convex set in a Euclidean space;
]
is a special case of (a), (e) would be, for every j, an immediate con-
sequence of (a-) and
e-. u j is quasi-concave in Sj'
b. fj is closed;
c. rj is lower semi-continuous and has nonempty convex sections
The crucial step in the proof of Theorem (11) (this would be also
the case in the proof of Theorem (12) if the authors were not restric-
ting themselves just to Euclidean spaces) relies on the use, for every
j, of a function f.: n.s.xs. ~ R with the following properties:
J 1. 1. J
13.a. fj(S'~) > 0 if and only if the player j can improve upon at
s by changing the jth coordinate of s to ~;
13.c. for every s E nisi' {~E Sjl (s, ~) E Zj} is nonempty and con-
~ (version: ~).
Let me set the matters more precisely (I shall begin with a defini-
tion and continue with a theorem) ;
Theorem (14), with conditions (c-), (e") replacing (c) (but not the
version) is in Wieczorek [1986], Theorem 1.4, p. 11 or Ichiishi,
Theorem 4.7.2, p. 70.
1, ••• ,n}
One can also think of other versions of Theorem (14)~ the appearing
conditions would then guarantee the existence of an (s*,s*) E Z by
means of a Lefschetz type theorem.
15. PROBLEM. Given sets S1' ••• 'Sn in a topological vector space
(version: ~~) and a relation =
Wj nisixs j when does there
exist a continuous function f J. : n.s.xs. ~ R such that
~ ~ J
i'j = {( s, ~ ) If j (s, ~) > O} which is quasi-concave in ~ (version:
such that for all s E n.s.
~ ~~
~ a,{~lpj(s,~) ~a} is
acyclic or empty).
~
252
REFERENCES
R.J. Aumann, 1969, Measurable utility and the measurable choice theorem,
in: La Decision (Editions du Centre National de la Recherche
Scientifique), 15-26
Fan Ky, 1952, Fixed point and minimax theorems in locally convex topo-
logical linear spaces, Proc. Nat. Acad. Sci. USA 38, 271-275
P. Fishburn, 1970, Utility theory for decision making, Wiley, New York
A.Ya. Kiruta, A.M. Rubinov and E.B. Yanovskaya, 1980, Optimal~nyr vybor
raspredeleni~ v slozhnykh social~no-ekonomicheskikh zadachakh, Nauka,
Leningrad
J.F. McClendon, 1984, Minimax theorems for ANR~s, Proc. Amer. Math.
Soc. 90, 149-154
;Jochen Ester
Technische Universitat Karl-Marx-Stadt
PSF 964, Karl-Marx-Stadt, 9010, DDR
O. MOTIVATION
We have been working in the field of MCDM since the early seventies.
With the aim of practical applications and technical and economic
effects we have achieved some results in theory, methodology and
applications. Recently, we could help engineers, economists and
others in solving the follOWing practical problems:
o optimal design of freezing systems,
0 optimal design of mechanical elements,
0 optimal design of textile articles,
0 optimal selection in training of high-performance swimmers,
0 optimal distortion of the command variable in position
feedback control,
o optimal control of robots.
Our last result in development of computer-tools for Computer Aided
Decision Making is an interactive program system POLYP.BC for micro-
computers. At present, we have reached the following stage:
MADM: Incomplete weighting method, Prevalence method, Method with
fuzzy dominating sets, Generalized Hodges-Lehmann criterion,
MODM: Global stochastic search, Antiparallel gradients method,
Generalized antiparallel gradients method (interactive),
Interactive method for linear t~ODM-problems, some single
criterion optimization methods.
These methodS reqUire a lot of important subroutines, e.g. normali-
zation, ordering (partial ordering, fuzzy partial ordering) and
methods for scalariz~ng functions and model-building, which are
implemented as single modules of the system. Due to the simplicity
of the structure, of the connections and the interfaces between all
modules in POLYP.BC, new modules can be written and connected with
the system.
Our experience is that there is no method, which can be used for
each problem or for each deciSion maker.
In many cases we could find out that the experts in modelling of the
258
( q1 ...
q )T = q shall be maximized.
m
X~Rn _ decision space,
fo r instance,
~(q) = m~x minyux(x)~(X,q» •
Ai = f q Iti ( q )}
and we choose
0 for qi ~ 0
[
jUi(q) = q1 i for 0 ~ qi ' 1
for 1 ~ qi
From Fuzzy Theory we know a lot of aggregation rules, for instance,
(i) • • • intersection
( iii) .. . . . . . • intermediate
(iv) m~xfi) • • • • • • •••• • • • • • union (7)
section
(vi) p(minY'i»+(1-P)(~L (Ui»,O ~ p ,£ 1 • 0 • intermediate-
• 0 •
i i '!: intersection
(vii) p(maX(Mi»+(1-p)(~[
i I . i
(M
1 i
» ,0 ' p 61 • • • • • • intermediate-
union
(i)-(iv) can be covered by
m
I'G( p ,g, r y«> = ( L
i=1
gi(rfi)P
m
)1/p (8 )
...
i=1
-oo<p<+oo,
~ = ~ ".~m)T 0
mixyt~) =fj = 1
r can be found by
r i = 1fU~ , i=1(1)m 0 (11)
What is the character of the reference point?
o
_ _ ifnon~Rnve.x dominating set
of p- in the case of
/ . strict "or"
convgx dominating set
'11 ~. strict "and"
'\.
of ft: in the case of
fJ' ~
~~------~~~--~-.~-------
Figure 1: Dominating sets
()
x-w
Figure 2: Niveau surfaces of~G(p), r,g are fixed
eP = U eP
g g
(13)
It holds
I. ePf PM' for -00< p < +00 (14)
263
40 HIERARCHICAL STRUCTURE
It is very easy to build up a hierarchical model of efficiency or
quality of a decision o If, for instance, G is given in the following
form
G = un
j iEl j
Ai =
j
Gj U (18 )
G j = {q lfj<t(q»} (19)
L
kfI
gS(r s / s = 1
k k
( 22)
s
Using (22) we do not change the direction in which lies the refe-
rence point, but its fictive locus in this direction.
5. SUMMARY
a) Admissibility and/or certainty can be handled as a single
criterion.
b) The logical property of the aggregation rule can be controlled
by a parameter p.
c) Given reference points and given weighting coefficients can be
taken into consideration.
d) With P and the reference point we can control the global pro-
perties (locus) of the solution.
e) With the weighting coefficients we can control the substitution-
rates in the locus of the solution.
f) The solution is an optimal fuzzy solution and it is paretooptimal.
g) By a hierarchical structure we can bUild up a more complicated,
multilevel decision making model.
REFERENCES
/1/ Ester,J.:Troltzsch,F. (1986). On generalized notions of efficiency
in MCDM, Syst.Anal.Model.Simul., 3(2)
/2/ Werners,B. (1984). Interaktive EntscheidungsunterstDtzung durch
ein flexibles mathematisches Programmierungssystem,Minerva,MOnchen
/3/ Peschel,M. (1980).Ingenieurtechn.Entscheidungen, Technik, Berlin
/4/ Ester,J.;HolzmOller,R. (1986).About some methods and applications
in MCDM, Syst.Ana~.Model.Simul., 3(5)
/5/ Dubois,D.:Prade,H. (1980). Fuzzy sets and systems - theory and
applications, Academic Press, New York
AN APPROACH TO r.1EASURDIG CONSISTENCY OF
PREFERENCE VECTOR DERIVATIONS
USING LEAST SQUARE DISTANCE
Gerd Islei
Manchester Business School
Booth Street West, Manchester M15 6PB
There is a wide interest in the use of judgement models for real life
decision making. However recent research in the psychology of decision
making has pointed to serious shortcomings which are often not recog-
nized by the practitioner and are also neglected by the analyst. We
shall examine some. conceptual inferences by looking at different ver-
sions of calculating preferences and discuss their usefulness for real
life applications.
(with other words, all entries of the matrix can be deduced from n-1
pairwise comparisons) •
The eigenvector method (Saaty (1980)) does not require consistency but
assumes that judgement matrices are reciprocal (i.e. a ij = 1/a ji for all
i,j=1, • • • , n) and requires that all entries of the upper triangular
half of the matrix be given (i.e. n(n-1)/2 comparisons have to be made).
life case study evaluates 8-10 choices with respect to 8-10 attributes,
i.e. requires a total number of 200-300 comparisons for the E.M.).
There are two more observations from our applications which should be
mentioned
(i) The D.M. often considered the spread of the E.M. results to be too
large (Lockett, Stratford, Cox, Hetherington and Yallup (1986»
A1 =a 12 A2 A1 = a 13 A3 A2 =a 23 A3
A1=2 A2 A = 5 A3 A = 7 A3
1 2
This can be recorded in the following judgement matrices
A1 a 11 a 12 a 13 2 5
A2 a 21 a 22 a 23 A2 1/2 7
A3 a 31 a 32 a 33 A31/5 1/7
268
+)
These jUdgements can also be represented as geometric graphs the
following way :
A1
A3 ~---. .- - -... A2
7
The reason for this is that L.S.M. only uses direct comparisons but ig-
nores (what we call) indirect comparisons (i.e. it ignores that for ex-
ample from A1 = 2 A2 and A2 = 7 A3 we can deduce the indirect comparison
A1 = 14 A3 )·
let us now take a closer look at the principle on which the calculation
of the E.M. is based.
(2) (2)
a 32 (2)
a 3l a 33
(2) _
with all - a ll 'a ll + a 12 'a 2l + a 13 'a 3l
(2)
a 12 = a l ,. a 12 a 12 'a 22 + a 13 'a 32
+
etc
But if we take a look for example at the element a~~)of the general
3 x 3 matrix
all,all'all
+ all'a12'a2l
+ all'a13,a3l
increasing weight
(ii) Inconsistent loops are handled unsatisfactorily
e.g. in our concrete example the entry a~~)contains the loops
an' a 32 o a 21 5-1/7'1/2 0.35
that is, it contains the same loop twice only traversed in opposite di-
rection (as indicated by the arrows) .
(1) Results of the E.M. for judgement matrices with inconsistent in-
puts exhibit a comparatively wide spread
(2) The E.M. can produce odd results
Example 1
A1 A2 A3 A4 AS Solution E.M.
--+ A2 7 3 8 A = 37.43%
2
A3 2 1/3 1/4 5 A = 13.42%
3
----to A4 8 4 5 A4 = 38.69%
It has been claimed (ibid.) that the E.M. produces the "intuitively"
correct ranking of the attributes A2 and A4 (namely A4 >A 2 ). We shall
use the following (partial) geometric graph of this problem to examine
this claim.
-1
+) Note that for all xE;R we have: x;x l!! 1 with equality iff x=1
271
A = 1 -1 A4 1.00 A4
2
A4= 8 -1 /7 A2 1. 14 A2
A = 4 -1 /3 A2 1. 33 A2
4
A = 5·1/8 A2 0.63 A2
4
A = 1 •1 A2 1.00 A2
4
A = 1•1 A2 1.00 A2
4
with average A4 =1.02 11. 2 (i. e. A4 >A 2 ) ++)
+) Note that in the case of this second order approximation the dir-
e~t compariso~ a?4=1 appears twice in the calculation, i.e. it is
g~ven more we~ghE.
++) Note that this peculiar discrepancy to the previous result occurs
because the arithmetic mean is taken over (basically multiplica-
tive) comparisons which should be combined multiplicatively.
272
ExamEle 2
A1 A2 A3 A4 AS A6 Solution E.M.
-
A1 1 4 3 1 3 4 A1 =32.08%
AS 1/3 5 5 3 AS =23.74%
We conclude that the E.M. can produce 'odd' results simply because it
uses an arithmetic average for ratio comparisons, hence including in-
consistent loops in a very unsatisfactory way.
they correspond to the same frame of reference+). They also take ac-
count of tradeoffs (i.e. comparisons of order 2) between these frames
of reference (i.e. rows/columns). However in our experience they do
not consider tradeoffs of tradeoffs (i.e. higher order comparisons).
Nevertheless our applications also suggest that slight differences in
the results are rather less important in discriminating between differ-
ent approaches than a potential to reduce data-requirements and to per-
mit a more comprehensive feedback for the D.M.
A1 2 5 A1 2 14
In the case of matrix(II) the entries of the upper triangular half re-
present the following equations :
A1 - 2 A2 0 (1)
A1 14 A3 0 (2)
A2 - 7 A3 0 (3)
A1 + A2 + A3 = 1 (c)
Now taking up our empirical argument it is not too difficult to real-
ise that these equations represent planes in a three-dimensional space
with coordinates A 1 , A 2 , A3 . For example, the equation A1 - 2 A2 =O
A1 - 2 A2 0 (4)
A1 - 5 A3 0 (5)
A2 - 7 A3 0 (6)
A1 + A2 + A3 (c)
is not consistent.
Therefore the solution in this case should lie inside the shaded tri-
angle of Figure 2. The optimal solution will be the point of the tri-
angle which has least square distance from the planes (4) - (6). (For
details and further discussion see Islei and Lockett (1986).
A 1 = 53.64%
Obviously our method does not require all direct comparisons for a cal-
culation but can be used as soon as the available direct comparisons
permit us to calculate at least one indirect comparison between any two
attributes. We are thus able to generate a succession of results by
including more and more comparisons in our calculation.
and obtain :
compo 1 - 5 compo 1 - 6 compo 1 - 7
Al = 5.51% A = 5.58% A = 5.90%
1 1
A2 = 37.16% A2 = 38.81% A2 = 38.06%
A = 11. 53% A = 12.10% A = 12.04%
3 3 3
A = 43.05% A4 = 40.70% A4= 40.16%
4
A = 2.75% A = 2.82% A = 3.85%
5 5 5
The results of this analysis for compo 1-9 and compo 1-10 are shown in
TABLE 1 and 2 respectively. (See Appendix for some comments on this
error analysis) .
and for the average distance of the solution to the direct comparisons.
5. A Practical Application
Now we shall look at some figures from a typical real life case study
and show how the error analysis was used to analyse inconsistencies.
This application required 9 attributes and the following judgement mat-
rix was supplied for them
A1 A2 A3 A4 A5 A6
A1 1 7 7 1 1 3 7
A2 1/7 1 5 1/7 1/5 1/3
A3 1/7 1/5 1/9 1/7 1/5
A4 7 9 1 3 5 7
A5 1 ·5 7 1/3 1 3 5
A6 1/3 3 5 1/5 1/3 1 3
A7 1/7 1/7 1/5 1/3
A8 1/7 7 3 1/5 1/5 1/5 5
A9 9 9 3 1/5 7
However when we presented to the D.M. the results of G.L.S. and the
error analysis (TABLE 3 shows the printout for the complete matrix) ,
indicating that comparison a 69 =S clearly stood out (i.e. that the in-
clusion of this comparison generated a considerable noise), he real-
ised that he should have input a 69 =1/S.
The new results for G.L.S. with error analysis (for the complete mat-
rix) can be found in TABLE 4.
A comparison of these results with the previous ones shows that rank
reversals occurred for the three leading attributes with either method.
+)
However only E.M. produced a rank reversal for the attributes AS' A6 .
The D.M. was quite satisfied with the new results and accepted the
noise as indicated in our error analysis.
In fact the solution for G.L.S. after input of 2(n-1) comparisons was
We have done many more applications using our analysis with very simi-
6. Conclusion
Using our method there are basically two procedures to assess inconsis-
tencies.
The next figures give the latest calculated results for the attributes
A 1 , . . . , A5 (in percent).
This is followed by the standard error E(i) for each attribute Ai (the
figures in the brackets are the standard errors expressed as percent
of the actual results).
Below this are listed the distances of each individual input to the
final solution,
We are also able to use the average distance of all direct comparisons
as an indicator of the inconsistencies.
+) Note that this distance is not precisely the Euclidean distance but
is enlarged by a factor of 100 and retains a sign (indicating if the
input comparison has to be increased (+) or decreased (-) in order to
be more compatible with the calculated solution).
280
REFERENCES
Cogger, K.O. and Yu, P.L. (1983), "Eigen Weight Vectors and Least
Distance Approximation for Revealed Preference in Pairwise Weight
Ratios", ORSA/TIMS Conference.
Coombs, C.H. (1964), "A Theory of Data", .Tohn ltliley, New York.
Islei, G. and Lockett, A.G. (1986), "An Approach to Preference Vector
Derivation Using Geometric Least Square", Proceedings of VIIth
International Conference of M.C.D.M.
Kahneman, D. and Tversky, A. (1982), "Judgement under Uncertainty
Heuristics and Biases", Cambridge University Press, Cambridge.
Keeney, R.L. and Raiffa, H. (1976), "Decisions with Multiple Object-
ives : Preferences and Value Tradeoffs", John Wiley, New York.
Lockett, A.G. and Hetherington, B. (1983), "Subjective Data and MCDM",
in "Essays and Surveys on Multiple Criteria Decision Making", Ed.P.
Hansen, Springer Verlag, pp.247-259.
Lockett, A.G., Stratford, M., Cox, B., Hetherington, B., and Yallup,P.
(1986) "Modelling a Research Portfolio Using AHP: A Group Decision
Process", R&D Management, May.
Saaty, T.L. (1980), "The Analytic Hierarchy Process", McGraw-Hill,
New York.
Saaty, T.L. and Vargas, L.G. (1984), "Inconsistency and Rank Preserva-
tion", Journal of Math. Psych., 28, pp 205-214.
281
(0,0,1)
/
/
(0,1,0)
-- - /
/
/
/ ~ Incon~1.tent. Ca ••
/
/
/
282
TABLE 1
1 ) A1 5.88%
2 ) A2 36.50%
3 ) A3 12.40%
4 ) A4 41.79%
5 ) A5 3.43%
STANDARD ERROR
E( 1 )- 0.31 5.2%)
E( 2 ) .. 1.46 4.0%)
E( 3 ) .. 0.95 7.7%)
E( 4 )= 1.23
E( 5 ) 0.58 .. 2.9%)
17.0%)
TABLE 2
1 A1 6.89%
2 A2 37.94%
3 A3 13.97%
4 A4 36.23%
5 A5 4.97%
STANDARD ERROR
E( 1 ) .. 1. 23 17.8%)
E( 2 )= 0.80 2.1%)
E( 3 )- 2.31 16.6%)
E( 4 )- 5.64 15.6%)
E( 5 )= 1.81 36.4%)
TABLE 3
·· ·• 32 )) =.. 77
D[a( 1 2 )]- -1-.79 YOUR INPUT a( 1
D[a( 1 3 )]=-0.66 YOUR INPUT a( 1
D[a( 1 4 )]--1.56
D[a( 1
· 5 )].. 3.59
YOUR
YOUR
INPUT
INPUT
a(
a(
1
1
•
•
4
5
)- 1
)- 1
D[a( 1 6 )] .. -3.15 YOUR INPUT a( 1 ) .. 3
··
D[a( 1 7 )]'" -1.25 YOUR • 6 ) .. 7
INPUT a( 1 • 7
D[a( 1 8 )]- -3.75 YOUR INPUT a( 1 ) .. 7
D[a( 1 9 ) ] - 1.53 • 8
YOUR INPUT a( 1 • 9 )- 1
D[a( 2
D[a( 2
· ....
3 )]- -2.45
4 )] 1.48
YOUR
YOUR
INPUT
INPUT
a(
a(
2
2
•
•
3
4
) .. 5
) ... 1428571
Dla( 2 5 ) ] 1.69 YOUR INPUT a( 2 )- .2
··
• 5
Dla( 2
D[a( 2 ..
6 ) ]= 1.24
7 ) ]
.. 0.39
YOUR
YOUR
INPUT
INPUT
a(
a(
2
2
•
•
6
7
)= .3333333
)- 1
Dla( 2 8 ) ] 3.58 YOUR INPUT a( 2 • 8 ) ... 1428571
D[a( 2
Dla( 3 ..
9 ) ] = 2.64
4 ) ] 1.02
YOUR
YOUR
INPUT
INPUT
a(
a(
2
3
• 9
4
)= .1111111
) ... 1111111
Dla( 3 5 ) ] = 1.37 YOUR INPUT a( 3 ) - .1428571
··
Dla( 3 6 )] = 1.43 • 5
YOUR INPUT a( 3 • 6 ) - .2
D[a( 3 7 )] = -0.42 YOUR INPUT a( 3 • 7 )- 1
Dla( 3
D[a( 3 ·· 9 )]..
8 ) ]- 1.16
1.50
YOUR
YOUR
INPUT
INPUT
a(
a(
3
3
•
•
8
9
)- .3333333
)- .1111111
Dla( 4 5 )]- -6.63 YOUR INPUT a( 4 • 5 ) .. 3
D[a( 4
Dla( 4 , · 6 )]= -5.34
7 )] = -0.93
YOUR
YOUR
INPUT
INPUT
a(
a(
4
4
•
,
6
7
)- 5
) .. 7
Dla( 4
Dla( 4 ,
D[a( 5
· 9 )]..
8 )]- -2.21
3.09
YOUR
YOUR
INPUT
INPUT
a(
a(
4
4
• 8
9
)- 5
)- 1
6 )]- -4.76 YOUR INPUT a( 5 6 ) .. 3
D[a( 5 7 )] .. -1.15 YOUR INPUT a( 5 7 ) .. 5
D[a( 5
D[a( 5
Dla( 6 · ..
8 )]= -3.64
9 ) ] 8.01
7 )]= -0.72
YOUR
YOUR
INPUT
INPUT
a(
a(
5
5
,
•
8
9
)- 5
)= .3333333
YOUR INPUT a( 6 7 )- 3
Dla( 6 8 )]=-4.50 YOUR INPUT a( 6 , 8 ) .. 5
Dla( 6 , 9 )]=-14.81 YOUR INPUT a( 6 ) .. 5
D[a( 7 8 ) ]= 2.64 • 9 ) ... 2
YOUR INPUT a( 7 8
Dla( 7 9 )] = 1.55 YOUR INPUT a( 7 , 9 ) ... 1428571
Dla( 8 , 9 ) ]- 4.06 YOUR INPUT a( 8 , 9 ) ... 1428571
AVERAGE DISTANCE = 2.82
284
'I'ABLE 4
DISTANCE OF THIS COMPARISON TO LAST SOLUTION D[a( 8 , 9 )] = 2.54
1 A1 18.75%
2 A2 3.73%
3 A3 2.66%
4 A4 21.41%
5 A5 13.57%
6 A6 7.87%
7 A7 3.21%
8 A8 5.64%
9 A9 23.15%
STANDARD ERROR
E( 1 )= 0.69 3.7%)
E( 2 ) = 0.25 6.7%)
E( 3 ) = 0.34 12.8%)
E( 4)= 1.18 5.5%)
E( 5 )= 3.10 22.8%)
E( 6 )= 0.52 6.7%)
E( 7 )= 0.17 5.3%)
E( 8 )= 1.63 28.9%)
E( 9 )= 1.59 6.9%)
DISTANCE OF COMPARISONS TO SOLUTION
D[a( 1 2 )]= -1.04 YOUR INPUT a( 1 , 2 )= 7
D[a( 1 , 3 )] =' 0.02 YOUR INPUT a( 1 3 )= 7
D[a( 1 4 )]= -1.89 YOUR INPUT a( 1 4 )= 1
D[a( 1 5 )] = 3.66 YOUR INPUT a( 1 5 )= 1
D[a( 1 6 )]= -1.54 YOUR INPUT a( 1 6 )= 3
D[a( 1 7 )]= -0.53 YOUR INPUT ae 1 , 7 )= 7
D[a( 1 8 »)= -2.94 YOUR INPUT a( 1 8 )= 7
D[a( 1 9 »)= -3.12 YOUR INPUT a( 1 9 )= 1
D[a( 2 , 3 )] = -1.87 YOUR INPUT a( 2 , 3 )= 5
D[a( 2 4 )] = 0.66 YOUR INPUT a( 2 , 4 )= .1428571
D[a( 2 5 )] = 1.00 YOUR INPUT a( 2 5 )= .2
D[a( 2 6 )] = 1.05 YOUR INPUT a( 2 6 )= .3333333
D[a( 2 7 )] = 0.37 YOUR INPUT a( 2 , 7 )- 1
D[a( 2 8 )]= 2.90 YOUR INPUT a( 2 8 )= .1428571
D[a( 2 9 )] = 1.15 YOUR INPUT a( 2 9 ) = .1111111
D[a( 3 4 )] = 0.28 YOUR INPUT a( 3 4 ) = .1111111
D[a( 3 5 )] = 0.71 YOUR INPUT a( 3 5 )= .1428571
D[a( 3 6 )] = 1.06 YOUR INPUT a( 3 6 )= .2
D[a( 3 7 )]= -0.39 YOUR INPUT a( 3 7 )= 1
D[a( 3 8 )]= 0.74 YOUR INPUT a( 3 8 )= .3333333
D[a( 3 9 ) ] = 0.08 YOUR INPUT a( 3 , 9 )= .1111111
D[a( 4 5 )]= -6.10 YOUR INPUT a( 4 , 5 ) = 3
D[a( 4 6 )]= -3.52 YOUR INPUT a( 4 6 )= 5
D[a( 4 7 )]= -0.15 YOUR INPUT a( 4 7 )= 7
D[a( 4 8 ») = -1. 33 YOUR INPUT a( 4 8 )= 5
D[a( 4 9 )]= -1.23 YOUR INPUT a( 4 9 )= 1
D[a( 5 6 ) ] = -3.18 YOUR INPUT a( 5 6 )= 3
D[a( 5 7 )]= -0.49 YOUR INPUT a( 5 7 )= 5
D[a( 5 8 »)= -2.87 YOUR INPUT a( 5 8 )= 5
D[a( 5 9 )] = 5.55 YOUR INPUT a( 5 9 ) = .3333333
D[a( 6 7 )]= -0.56 YOUR INPUT a( 6 7 )= 3
D[a( 6 8 )]= -3.99 YOUR INPUT a( 6 8 )= 5
D[a( 6 9 ») = 3 .18 YOUR INPUT a( 6 9 )= .2
D[a( 7 8 )]= 2.04 YOUR INPUT a( 7 8 ) = .2
D[a( 7 9 ) ] = -0,09 YOUR INPUT a( 7 9 )= .1428571
D[a( 8 9 )] = 2.31 YOUR INPUT ae 8 9 )= .1428571
AVERAGE DISTANCE = 1. 77
AGGREGATION PROCEDURES FOR HIERARCHICALLY GROUPED
DECISION ATTRIBUTES WITH APPLICATION TO
CONTROL SYSTEM PERFORMANCE EVALUATION
Integrated
Microcomputer-based
Process Control System
(5)
0, if Yi
:::;. i'i '.
-,
(y;
k
- i~)/(~'.'
~ ~
- y~),
~
if Yi < Yi < ~"i '. (6)
-<
-II
1, if Yi Yi;
(7)
b«Yi-Yi,min)/(Yi-Yi,min)-I) , if Yi,min ~
Yi ~ y!1
Ui(Yi) (Yi-Yi)/(Yi-Yi) , if y!1 <
Yi < -n
Yi (8)
a(Yi-Yi)/(Yi,max-Yi)+I, if Y~1 ~ Yi ~
Yi,max
(9)
definitions of this class; its level sets coincide with the shifted po-
sitive cone RP+ that defines the partial preordering of the attribute
space. This function has several interpretations. One of them is a L-
-shaped utility function, consistently summarizing the information con-
tained in the points Yi,min' yi, yi,
Yi,max' and thus serving as an
approximation to the preferences of the decision maker. Observe that
this particular function is a cardinal utility function, since it is
defined by ratios of intervals and thus independent of any affine trans-
formations of outcomes; therefore, it can be even used in interpersonal
comparisons of utility in a committee.
Another interpretation of this function is a transformed and we-
ighted (with changing, but piece-wise constant weighting coefficients)
Chebyshev or 100 norm of the difference between the point Ymax and the
actual attribute assessment y = f(x). To ilustrate this interpretation
more clearly, consider a case when the points Ymin' y' are not specifi-
ed and denote y" = y; the function (8,9), after an affine transformati-
on, simplifies then to the form:
(11)
-
Since we assume that Yi < Yi,max for all i = l, ••. p, the achieve-
ment function corresponds in this case to the weighted Chebyshev norm
with changed sign: however, the weights Ai are not specified explicite-
ly by the decision maker, but defined implicitely through his statement
of aspiration levels y as compared to the upper bound point Ymax. In-
terpreted as an achievement function, this function was used, for exam-
ple, by Wierzbicki (1984) and Lewandowski, Johnson, Wierzbicki (1986);
as a Chebyshev norm, by Nakayama (1984) and Steuer and Choo (1983). The
case when a decision maker specifies a reservation and an aspiration
point was also investigated, though not in the form of the achievement
function (8,9), by Gorecki et al. (1984), and by Weistroffer (1984).
The simplification (10,11) of the achievement function (8,9) is
not the only one possible. If the deCision maker knows both upper bo-
und point Ymax and the lower bound point Ymin but sp~cifies only one
reference point y, this point should be interpreted as an aspiration
level point rather than reservation level point; but in mathematical
modifications of (8,9) we must treat it as reservation levels, y y',
and let the aspiration levels coincide with the upper bounds, y" = Ymax
293
p
s(y,y',~") =u(y) = (minu.(y.) + (c/p) E u.(y.»/(l +c) (12)
l~iS.p ~ ~ i=l ~ ~
Step 2: Betting and saaling attributes. The following criteria were se-
lected for the evaluation of the hierarchical structural form of a con-
trol system (on the lower level of attributes):
T.1.
TABLE OF ASPIRATIONS
E CRITERIA C1 C2 C3
j +
4-k
DECISION
MAKERS
p Tr 0
D1
y' 5 7 0
Y" 7 9 @)
y' 3 6,5 6
D2
-"
y 6 Q 9
-y'
D3
0 (0 2
-"
y 0 7 4
D4
y' G) 5 ~~
-Y" G) CD CD
D5
y' 7 8 8
-"
y 8 9 9
TABLE OF ATTRIBUTE-ALTERNATIVE ASSIGNEMENTS,INITIAL SETTING OF ASPIRATION LEVELS AND VALUES OF ACHIEVEMENT FUNCTIONS
8(Y,y',y")
j+
k = 1 k = 2 k = 3 k = 4 k = 5
-I- •
~TERNATIVES
~ 7,1j / 7V/ / c=O c=1 c=O c=1 c=O c=1 c=O c=1 c=O c=1
A2 (Fig. 3b) Y 5,55 9,65 6,7 -4,4 2,3 -3,1 13,6 -2,1
Step 3: InitiaZ setting of aspi~ation ZepeZs. During this step, the co~
(13)
A
where kj' k j denote indices of extreme value assignements (circles in
Table T.2.); thus
(14)
y"
(15)
DI. (16)
J
where:
'V
Yjk = Yjk + y-J"'k) /2
(-, (17)
All the calculated values are indicated in Table T.l. A high value of
the indicator (DI j >l) implies in principle a discussion between commi-
ttee members and possible correction of th7 assigned aspiration levels.
Occasionally (as is the case here with 0 criterion) the nature of the
criterion implies an apriori the possibility of divergent approaches
to the allocation of aspiration levels.
300
--"
--"
--"
0)
"
~
0< K1 !: 1
K2~O
-" K1 K2"
--" [[J
n =12(20) r=6(10)
'l\
~
JIIJJ0/S=2
r=6(10)
b)
0< K, ~,
0< K ~ 1
n=12(20)
C)
case "0" from T.3. In order to reduce the uncertainty of attribute va-
lues assignement, we consider here two values. The first value
y'(i,j,k) = y' (i,j) is obtained directly from analytical calculations
for the attribute j and alternative i (independently from the committee
member'k) for parameters resulting from the test case "0". The second
value y"(i,j,k) = y"(i,j) is obtained under the assumption that loading
factors were increased, as compared to the test case "0", by 5%. These
values are shown in Table T.4.
A final value for attribute assignement is obtained by taking the
mean:
y(i,j,k) y(i,j) 0.5(y'(i,j) + y"(i,j» (18)
T.3.
~19~~~'~~">T
NUMBER OF
A B C D E
CLOSED-LOOP 16 50 60 80 90
NUMBER OF
120 260 440 700 750
TRANSMITERS
NUMBER OF
SIGNALLING AND 70 180 310 400 450
ALARMS
NUMBER OF
FINAL CONTROL 44 230 500 800 950
ELEMENTS
T.4.
ALTERNATIVES P Tr <')
Step 8. Group ranking. Table T.S. contains the individual ranking lists
of alternative solutions formed on the basis of achievement function
values in Step 6.
T.S.
D4 3 13,6 11,6 A2 , A3 , A1
Documentation is complete.
y' 4 9 5 5 9 9 8 1 I
Al - CS 1 -6,6 -3,0 -7,5 -3,4 -9,0 -4,5 -9,0 -4,3 -6,0 -2,9
II
y 5 10 6 6 10 10 9 2
y' 9 1 5 9 9 1 7 4
A2 - CS 2 II
-8,5 -3,9 -8,0 -3,7 -9,0 -4,4 -8,8 -4,4 -9,0 -4,4
y 10 2 6 10 lO 2 8 5
y' 8 6 10 1 1 6 1 9
A3 - CS 3 .-8,5 -4,1 -8,3 -3,9 -9,0 -4,6 -9,0 -4,5 -8,7 -4,2
II
y 9 7 10 2 2 7 2 10 c.>
~
+k y' 7
8 7 8 G) 3 CD 3
D1
-II -(8)
y. 9 8 7 5 G) 4
0
y' 7 5 5 6 4
D2
-II
CD 8 ®
y 9 8 7 9 8 7 7
CD
y' 6 9 6 @ @ @ I@
D3 -II
0
Y 8 9 6 10 @ @ (!p)
CD
y' 9 9 @ 6 6 8 10
D4 _II
CD
y 9 9 9 @ 7 9 10
CD
y' @ § 9 .8 8 9 (0
D5 -II
CD
y 160) (Fa) I(ro) 9 fa) (3) 10 !(l)
307
control actions.
Documentation is incomplete.
T.7.
Documentation is unacceptable.
Step 8: Group ranking. Table T.7. shows individual ranking list of al-
ternative solutions according to achievement function values and an
aggregated ranking based on averaging achievement across the committee.
7. CONCLUSIONS
8. REFERENCES
H. Roland Weistroffer
School of Business
Virginia Commonwealth University
Richmond, Virginia 23284-0001, USA
Introduction
.?. 0, (1)
where x = [xl' x2' "" x n ) ERn is the vector of decision variables, each
xERn representing a possible solution to the problem, the f i : Rn~ R, i = 1, " ' , k,
are the objective functions, and the gj: Rn 7 R, j = 1, " ' , m, are the constraint
functions.
312
The best compromise solution sought, would have to be feasible (i.e., satisfy
the given constraints) and efficient (or non-dominated or Pareto optimum), that is,
there should be no other feasible solution at which all the objective functions
achieve equal or better values and at which at least one objective function achieves
a better value. An iterative and interactive technique can be used to determine the
best compromise solution within the efficient and feasible solution set.
The model and technique proposed in this paper is based on the reference point
approach briefly described in the following section.
The reference point approach (Wierzbicki 1982, Lewandowski and Grauer 1982,
Grauer 1983) basically works as follows: the decision maker provides reference
values, i.e., objective function values that he considers acceptable for the
decision problem, without knowing whether these values are realizable or not. A
computer program is then used to minimize the under-achievements with respect to
those reference values that were picked too high, and to maximize the over-
achievements with respect to those reference values that were picked too low, i.e.
minimize si ~ fi(x), i = 1, 2, ••• , k, (2)
where the reference values si may be larger or smaller than the corresponding
objective function values fi (x). The minimization is carried out via some order
preserving scalarizing function (or surrogate objective function) and subject to the
given constraints gj (x) 2. 0, j = 1, 2, ••• , m. The resultant solution and the
corresponding objective function values are shown to the decision maker, who may now
decide to raise or lower some of the reference values, based on new insight as to
what is realizable. A new solution is found, and the process continues until the
decision maker feels that the solution obtained is the best compromise solution (or
very close to it).
A special case of this approach arises when the reference values are con-
sistently larger than the corresponding actually achieved objective function
values. The reference values may then be viewed as targets or goals or desired
values. After finding a solution by minimizing the under-achievements, the decision
maker is asked to lower some of the desired values to a more realistic level, and
the process is repeated as before. The approach can now be regarded as a general-
ized form of goal programming. This approach lets the decision maker set his
initial expectations very high (i.e. specify rather optimistic desired values), and
then iteratively and interactively induces him to lower these expectations to real-
istic levels. The interactive goal programming method of Weistroffer (1983) is
based on this approach. STEM (Benayoun, de Montgolfier, Tergny and Laritchev 1971)
also follows the philosophy of decreasing expectations, but does not ask the
decision maker to reduce any desired values, but rather to change some of the
objective functions into additional constraints.
313
In the introduction to this paper it was pointed out that the decision maker
may not always be able to distinguish between objectives and constraints. For that
reason. the following alternative problem formulation to (1) is proposed:
maximize [f 1(x). f 2 (x) • •••• fk(x)]T
functions are identical in this formulation, both are depicted by fi (x), and are
referred to as criterion functions. Under the usual assumptions that the objective
functions are continuous and that the feasible set is compact, (3) is equivalent to
(1): The compactness and continuity conditions imply the existence of upper and
lower bounds for the objective functions and the d i and r i may be set equal to those
bounds. The d i and ri for the constraint functions may be set to zero.
The d i can be viewed as desired values and the ri can be viewed as required
values.
(4)
where[xl' x2' "., x n 1 = xeRn is the vector of decision variables, each element of
Rn representing a specific solution, fj: Rn ~ R, j = 1, 2, "" k are the criterion
functions, d j , j = 1, 2, " ' , k are the corresponding desired values, rj' j = 1, 2,
.", k are the corresponding required values, WjE(O, 11 expresses the importance of
getting close to the desired value d j rather than merely meet the required value rj
for each criterion j E{1, 2, ...
, k} (Wj determines to what degree a criterion
function is to be considered an objective or a constraint). The scalarizing
315
Some Remarks
desirable. From a numerical point of view, some experimentation with the wi values
for strict constraints may be required. The author is currently working on
designing software to implement the ideas described in this paper.
References
P. L. YU
and
1. S. Chien
School of Business
University of Kansas
Lawrence, Kansas 66045
Abstract
1. Introduction
In Multiple Criteria Decisien Analysis, 'we are usually given a set of goal
functions or criteria and a set of alternatives., with which we try to find the
most "valuable" alternative according to some preference defined on the outcome
space. The most valuable alternative is usually the one which maximizes some
value function defined on the outcome space. It can be an element of some nondom-
ina ted solutions or some satisfying solutions. (See [18, 21] and those quoted
therein for instances.) One of the common properties in this kind of approach is
that the outcome space and the alternative set are fixed. However, many decision
1. This research has been supported by NSF of U.S.A. Grant No. IST-8418863.
318
problems which are involved with multiple criteria do not necessarily have a fixed
set of outcome space and alternative sets. It is well known, and has been stud-
ied, that if one would set a little more difficult goal, his performance results
usually could be better (See Section 5)-. That is, by stretching their ability a
little bit human beings can achieve more. Thus, the outcome space can be
stretched a little bit to contain some preset goal levels. It i.s also well known
that our performance tends to respond differently -to different sets of goals.
Therefore, the selection of the goal functions can be very important in achieving
the overall performance over some period of time.
The purpose of this article is to investigate the foundation for effective
goal settings as to maximize the overall performance over a period of time. The
goals are treated as variable, rather than predetermined. The problem, however,
is still closely related to multiple criteria decision making problems (MCDM). It
simply involves more complexity in the setup than ordinary MCDM.
such as Tom's girlfriends, parents, or financial problems, etc., may disturb his
attention and drive for improvement.)
(vi) Should the Coach help Tom develop a habit far training himself to a-
chieve his goals as scheduled with 100% concentration in order to become an Olym-
pic star?
In this paper, we shall discuss and answer the above questions and offer a
general principle for effective goal setting. The principle is not only appli-
cable to the above athlete problem, but also to many other decision problems or
situations in which proper goal setting can be used to improve overall perfor-
mance, including the problems of assembly line workers., receptionists, managers,
researchers and corporate presidents. The principle can be used for self goal
setting to improve self-performance and be used for goal setting for other persons
to improve their performance.
In order to describe and illustrate the foundation of effective goal setting,
we shall briefly review some fundamentals of human behavior mechanisms and sketch
their relation to the problem of effective goal set'ting in Section 2. In Section
3, we shall briefly review the concept of habitual domains and use the concept to
discuss the processes af implanting, accepting and habituating of ideas in goal
setting.
In Section 4, we lay down a general principle of effective goal setting as a
complex mUltiple criteria optimal control system. Using habitual domain concepts,
in Section 4.1, we first classify decision problems into four classes: simple
routine, mixed routine, fuzzy and challenging. The general principle is then
described in Section 4.2, utilizing the concepts of behavior mechanism discussed
in Section 2. Further discussion on the control parameters of goal setting and
supportive systems are given respectively in Section 4.3 and 4.4. Finally, the
principle is mathematically formulated in Section 4.5.
There is rich literature in goal setting contributed by organization behav-
iorists and industrial psychologists. For instances, see [I, 3-16] and those
quoted therein. Because of their emphasis on statistical empirical studies (a
habitual domain), they primarily focus on simple routine or mixed routine pro-
blems, not including fuzzy or challenging problems. Strictly speaking, their
results cannot be used for a broad general class of goal setting problems. How-
ever, as their results are, in general, supporting the principle described, we
shall briefly sketch their empirical study results in Section 5. Thus, the reader
can use our principle to explain their results as verification exercises.
2. A Behavior Mechanism
-..
co
c:
r
l-
e
c 51
-,(14)
-c;
~
Il e e I
I- : :
c I
u
I c:CI !:!
.....
.2 - -I
I> ..
-
e
en
>.
.r:.
I C ~I
Ci)
I ~ ell
...,
I tl ~ I
Q..
(13)
1< -gl
1_ <J
.. en (12)
Ci)
~
.....I
~
......----,---.... External
Experience/ Being
Reinforcement Observed
(II) Each human being has a set of goals or equilibrium points to reach and
maintain (goal setting). A list of gods is given in Table 1 (p. 8). The
status or stat·e of each goal is constant.1y (conS'ciously or unconsciously)
monitored (state valuation). When a significant "unfavorable" deviation
between the perceived value and the ideal value exists, each goal can pro-
duce a "charge" (a tension or urge). The totality of the charges produced
by all goals which form a hierarchical structure is called charge struc-
ture. This charge structure can change rapidly with time, information
inputs, and one's psychological state and conditien. (See Boxes 2-5; also
Hypotheses 5-6 of [17,18].) Note that when a goal (e.g., health) is main-
tained at the tolerable interval of the ideal, without other effort, the
goal may produce very little or no charge and will be neglected.
vat ion; but avoidance justification simply diffuses or dissipates the char-
ges. In the latter case, charges are not equal to force, drive or motiva-
tion.
(VII) All functions/components of the flow chart are interconnected. Through
time, they interact with one another. For instance, once an action is
taken (Box 9), say publishing an article, the event and its consequence
will be observed and registered in the decision maker's memory (Box 1) and
also will likely be observed and interpreted by other parties (Boxes 12 and
13) which ·may, in turn, react upon the original decision maker (Boxes 7 and
11). (See Hypothesis 8 in [17,18] for information inputs.)
til>n; (ii:) the subject should have sufficient ability and confidence that the
goals can be obtained in the due time, which induces active problem solving rather
than avoidance justification; and (iii) the supportive systems (including reward-
ing systems and monitoring systems) and the environment should be such that the
easiest way to release the high level of charges is to perform the desired
job-related works to reach the preset goals.
In our at'hlete's problem, if John (loachcannot inspire Tom Student to a high
level of charges or burning desire ,to become an Olympic star, or if Tom Student
cannot mass enough confidence that he can become a star in due time, or if Tom
Student is in an environment in which other more serious events keep on bothering
him (such as, diseases and/or health problems, 'disputes with girlfriends, serious
financial disasters, etc.), then we cannot expect that Tom S,tudent can contin-
uously put his attention to athletic training to become an Olympic star. Even if
he could occasionally think about ,the event. because of other priorities or avoid-
ance justification he could not mass enough effort and time to go through the
training to achieve the goal of becoming an Olympic star.
3. Habitual Domsins
The dynamic mechanism, Although changing with time and with one's psycho-
logical states, can stabilize and, unless extraordinary events occur, each indi-
vidual can have stable habitual patterns for processing information. This obser-
vation is captured by the concept of habitual domains.
It has been recognized that each human being has habitual ways to respond to
stimuli. These are sometimes called conditioned or programmed behaviors. For
example, each individual has habitual ways of eating, dressing, and speaking.
Some habitually emphasize economical gains, while others habitually emphasize
social reputation. Some are habitually persistent in their goal pursuit, while
others habitually change their objectives. Some are habitually optimistic and see
the po'sitive side, while others are habitually pessimistic and see the negative
side. Some habitually pay attention to details; while others only to generali-
ties.
job seeking, house purchasing, dealing with friends, etc. also follow. In the
following discussion, BD are used without specifying 'the corresponding functions
or events, as only the general properties of HD will be discussed.
Qt. Hore precisely, R(It,Ot) is the set- of ideas/actions that can be cumulatively
generated by any sequence of operators from the set of 0t which act on It and the
resulting ideas/actions frem the operations. As an example, let It = {O,l} and 0t
- {+} ("+" is the ordinary mathematical addition symbol). Then R(It,Ot) is the
set of all nonnegative integers. If 0t .. {+,-}, then R(It,Ot) is the set of all
integers. In the athlete problem, the idea of becoming an Olympic star can gener-
ate "glory, pride-, wealth, hard' work, discipline ••• " as a subset of the reachable
domain.
Potential and reachable domains are a1ese1y related. We say that {It,Ot}
is a generator of PDt iff PDt" R(It,Ot)' and {It,Ot} is a basis for PDt iff
{It ,Ot} is a generator of PDt and no strict subset of {It ,Ot} is a generator of
PDt·
Finally we observe that HOt' as defined, is dynamic and changes with
time. All four elements of HOt can evolve with time: this makes analysis of HD t
fairly difficult. Fortunately, it can be shown that the HOt can reach a stable
state. Thus, human behaviors are still, to a certain degree, predictable.
In [2,181, it is shown that PDt and APt depend on the activation of pat-
terns of neural circuits. Under suitable conditions, these patterns will satisfy
a system of differential equations with stable steady solutions. Thus, PDt and
APt can reach their stable states unless extraordinary events arrive continuously
(which rarely is the case).
Once PDt and APt reach their stable states, one can expect ADt to occur
with some regularity. In such stable states of HD t , one ean expect habitual ways
of thinking, responding, and Teacting to stimuli and events to occur. Thus, per-
sonality, attitude, and conditioned or programmed behavior will be formed for each
individual. Such formation has a great impact on decision making styles and opti-
mization theories. The existence of habitual domains has significant implications
for high stake decision problems, optimal solutions, gaming and conflict dissolu-
tion, career management and leadership. For the details see [2,18-201. We shall
discuss only those related to effective goal setting.
Understanding the HOs of ourself and others is very essential for effective
goal setting. For instance, in the athlete problem (I), if John Coach wants to
successfully implant a burning desire (high level of charge) in Tom Student's mind
to become an Olympic star, he must first explore Tom's HO. What are the active
ideas and interests in Tom's mind? What are his CUTrent goals for life? How much
confidence does he have in himself and in the coach's suggestions? What are his
channels of external information? Without a careful study, the implanting may
fail. For example, suppose that Tom's father is a successful owner of a restaur-
326
ant whom Tom respects very highly. Tom wants to be a millionaire by getting into
the restaurant business after his graduation. He does not have high hopes and
interest in becoming a sport star because he does not believe that he could become
a star and thus a milli@naire~ If the 'coaCh doea not know the above, he may just
try to persuade Tom by saying how glortous and fulfilling it wiLl be to become an
Olympic star without suggesting -to Tom that an Oly.mpi~ star can also become a
millionaire and without building up the confidence for Tom himself to believe that
he could be a star. If this is the case, we see that the coach's probability of
success cannot be too great. If the coach knows Tom's HD~ he may visit his father
and layout the glorious picture in which Tom is an Olympic star. He may show the
father why Tom can be a star and hint that many Ol"}'1llpic stars did become million-
aires and have happy lives. Once the fathe.r is convinced., he could talk to Tom.
Through his father's channel, Tom may build up more confidence in himself and more
interest in becoming a star.
Note that newly arrived messages or concepts ean be rejected quickly if
they are not consistent with the existing HD and the subject does not have an
"open minded" operator in his MD. Thi's is due to the least resistance principle
(it is easier to reject than to aecept, because it requires encoding, association
and integration to accept the new ~oncepts and to maintain logical consistency).
In fact, it always takes time to implant new ideas and have the ideas accepted.
In studying effective goal .setting, the concept of the cores of HDs become
very important. Roughly speaking, the core of the HD is the set of ideas or con-
'cepts that are very strongly implanted in the HD and can be easily retrieved to
interpret the arriving events or to perform the needed work. As a core is a con-
cept on how strongly the ideas are encoded and how easily retrieved, while a gen-
erator or basis is a concept on those ideas that can potentially activated, we
notice that the core can usually just be a proper subset of a generator or a ba-
sis. For example, when we talk about relatives and friends, the set of {the par-
ents, ·the wife or husband, the sons and daughters, the business colleagues, the
neighbors} can 'be a core; ~ut the set cannot be a generator or a basis for all
re'latives and friends. In the athlete problem, when talking about future ambi-
tion, {a millionaire, restaurant} can be the core of Tom's current HD; but the set
cannot be a generator or a basis for all potential ideas about future ambition.
New ideas can be implanted into HD and become elements of the core.
In order to make the concept of core precise. let I be a given idea and
let be a particular event or problem under consideration.
Define the activation density of idea I in event E at time t by:
atU, E ) = P,robt-[I activates IEgets attention] (1)
which is the conditional probability for I to be activated given that our atten-
tion is paid to E at time t.
Note that the activation density is a function of time and
327
Depending on the regularity and availability that ideas, concepts and ope-
rations are needed to successfully perform jobs or solve problems, we can roughly
classify the decision problems into four classes.
Class I. Simple Routine Problems. In this class, the set of ideas, con-
cepts and operations which are needed to successfully solve problems is almost
fixed and well ordered. Assembly line work, typiug, sleeping, eating, driving,
etc. can belong to this class. With proper training, most 'people can acquire this
set of ideas, concepts and operations and incorporate them into the core of HD for
solving their problems. Thus, habituation of the order of operations becomes an
important part in performing and solving the problems.
Class II. Mixed Routine Problems. To successfully execute or solve this
class of problems, one needs to solve variable combinations of simple routine
problems. The work of repa'lr shops, receptionists, waitresses, clerks, and regu-
lar dinner prepara~ion belong to this class. In this class, the set of ideas,
concepts and operations needed to successfully solve the individual component
problems are known, and can be acquired and habituated as the core of HD by proper
training. However" the best order to perform the job may not be fixed. There are
also time allocation problems over the individual component problems. One notices
that the problems of this class are more complex than those of Class I. The core
of HB in this class, in general, is an a-core with a smaller than that of Class
1.
Class III. Fuzzy Problems. There are no sure set of ideas, concepts and
operations that can guarantee the success of a solution, even though the set can
be fuzzily known. Sales problems, advertisement problems, ordinary research and
development problems, simple conflict problems with friends, problems of preparing
for an important dinner party, etc. are some examples of this class of problems.
There may exist a set of ideas, concepts and operations which are commonly but
fuzzily known to be good for solving the problems. This set of ideas may be ac-
quired and habituated through training and learning as to become a large propor-
'tion of the core of HD (i.e. a-core with high value of a) for solving the prob-
lems. However, this set alone is usually not adequate to surely and successfully
solve the problems. There is always a need of new ideas or concepts, which may be
acquired from the potential domain by active thinking and association, to success-
fully solve the problems. In 'general, a-core, with a large a, of the decision
maker's HD is not adequate for solving the problems. In fact, rigid and inflex-
ible HD (i.e. the a-core is almost fixed even if we lower the value of a) may
pTove to be detrimental to solving the fuzzy problems.
Class IV. Challenging Problems. These problems cannot be successfully
solved by a-core of HD, no matter how low we reduce the value of a. These kinds
329
formers' charge., and, pe1:haps, to matimize the favorability of the working en-
vironment. A mathematical formulation of this 'complex optimal control system will
be described in Section 4.5, after the preparat'i'on of Sections 4.3 - 4.4.
Let us first integrate the above principle with the behavior mechanism as
depicted in Flow Chart 2. In the flow chart we assume that physiological condi-
tions and external information inputs, except supportive systems, have been neu-
tralized and do not create high levels of charges or dis.turbances on the perfor-
mers. Thus they are put in the insignificant sub-box "others" of Box (4).
,---
I ,..-~-I-..., ..-~-I-...,(2)
Supportive
Systems
I
I
I
I
I
,
I
r - - - - - - , (7)
I
I Working
I r - - - -.......- - - - , ( 4 )
Environment
I
r----t---..------i
I I Others
,, L----'------r---'
Confidence
Training,
I Assistance,
I r---- Abi 11 ty
& Supervising
: .. ---1---,----1...----,
t- I Avoldonce
, Justification
1..------'----.,r----....
(9)
Drive and Monitoring
Performance
Flow Chart 2.
Note that in Flow Chart 2, the boxes in the right column (i.e. Boxes (7) -
(10» are components of the supportive system; while those in the left column
(Le. Boxes (1) - (6» are the modified "behavior mechanism". Let us explain the
principle using Flow Chart 2 as follows:
Boxes (1) - (2). Goal Setting and State Valuation. Although the func-
tioning of these two boxes are as in Flow Chart 1, they are subject to the influ-
ence and control of our purposeful goal setting and supportive system in addition
to physiological monitoring and other external information inputs. In Box (1),
"goal setting" has been used as a purposeful control on job performance as well as
331
its general meaning of setting the ideal value of living goals beyond those rela-
ted to job performance (1. e. including t-hose listed in Table 1.)
Boxes (3) - (4). Charge Structures and Attention. In order to obtain
attention allocation, the purposeful goal setting must be able to create the rela-
tively highest level of charges on the performer's chaTge structures. That is,
the charge of "not reaching the performance goals" is so much higher than that of
"reaching the performance goals", that no life goal functions other than the per-
formance goals can create a more significant charge. When the performance goals
can create a sufficiently high level of Charge, while all other life goals cannot
do so, the performers can have a burning desire to achieve the set performance
goals. To achieve this objective, we need to (A) create the rosy and promising
picture of the state when the performance goals ,are accomplished (positive re-
wards); (B) let the performers realize the sad and depressing state when the per-
formance goals are ~ accomplished (negative rewards or punishment); (C) set up
barriers so that job-unrelated messages or disturbances would not reach the per-
formers and create a high level of charge; and (D) develop good working environ-
ment in which the performers enjoy their work and in whi~h accidents are unlikely
to occur.
Note that in Box (4), there is a sub-box, Attention to Others, which in-
cludes physiological monitoring and external information inputs unrelated to job
performance. The more attention' paid to "others", the less attention paid to
job-related work. It may represent a leakage of time and effort with respect to
performance. Note that 100% of attention over a iong period of time to job--
related work is impossible, because men as living systems at least need to eat,
sleep, etc. to release the high level of charges cr.eated by physiological goal
functions (see [17 _,18] for further discussion).
Boxes (5) - (6). Positive Problem Solving or Avoidance Justification, and
Drive and Performance. As mentioned in Section 2, positive problem solving is to
work hard to accomplish the preset goal values, while avoidance justification is
to lower the ideal goal value or the significance of the problems. Thus, after
attention, the attitude of positive problem solving will produce "drive" and per-
formance; while that of avoidance justification will usually not. Besides person-
ality, confidence, hope and the ability to accomplish the ideal goal value in an
acceptable time horizon plays an important role in deciding which attitude will be
taken. Using supportive systems, especially training, as'sistance and supervising
(Box 8), to increase the performer's confidence and ability is an important means
to enhance the attitude of positive problem solving and to improve the perfor-
mance.
Boxes (7) - (10). Supportive Systems These four boxes are some important
devices for supportive systems. They help create or modulate charge structures as
to make performers pay more atteution to job-related problems, cultivate a posi-
332
tive problem solving attitude, improve coafidence and ability to accomplish the
ideal goal value, and to actually achieve the set goal value. We shall further
explore the impact of these supportive devices on performance in Section 4.4.
Note that by utilizing Flow Chart 2, one eould desc~ibe a dynamic system
for the problem of effective goal setting (see Section 4.5.) ~ith proper assump-
tions such as additivity with respect to time, the dynamic system can even be
expressed in terms of di:fferential or integral equations. To avoid detraction, we
shall not stop here to do so.
As we want to create a high level of charge and maximize the time allocated
to job-related work, when selecting goals we shall pay attention to performers'
habitual domains. To be effective, the selected goal should be able to strike
resonance on the performers' HDs. If the performer is striving for survival and
needs money, then .the goals directly related to monetary reward will catch his
attention. On the other hand, if the performer is loaded with money and looking
for social recognition, then the goals related to social recognition may catch his
attention. If the performer has already reached a mature state of self-actualiza-
tion and can fai.rly easily detach himself from the attraction of money and social
recognition, then the goals related to independence and freedom of the self may
catch his attention.
In Table 1, we list a fairly inclusive collection of goal functions that
may affect our behavior and charge structures. One way to visualize the work of
these goal functions is to open a telephone directory or to think of various in-
dustries and businesses which try to satisfy some peoples' needs generated by
these goal functions. However, to transform the goal functions listed in Table 1
into a set of vivid goals that can strike resonance in peoples' HDs, we need to
actively study the interested peoples' HDs as well as use our ability of associa-
tion and creativity. For instance, "salary" may be more striking than "accumula-
tion of wealth"; "firing from job" than "self-importance feeling"; "sex" than
"giving birth to the next generation", etc. The billboard which says "Driving Is
Like Making Love, Don't Rush to Finish It" may be more effective to people than
one which says "Don't Speed". We shall leave these to the reader's creative imag-
ination.
In addition, to be effective the selected goals must be clear, easily mea-
surable and closely related to the overall performance that we are looking for.
In the athlete problem, the time needed for 100 meters is clear, easily measurable
and directly related the overall performance; while the streamlining of movement
333
is directly related but less clear and more difficult to measure even with the aid
of video tapes. The clear and measurable goals make monitoring and feedback eas-
ier, which can help performers' "state valuat:ioo" to cr.eate charges (Box (3» and
to enhance the confidence and positive problem solving (Box (5».
Finally, we note that for Class I-II problema. as the set of ideas, con-
cepts and operations for successfully executing the jobs is known and almost
fixed, the selection of clear and easily measl1'l'able goals is lDUch easier than that
of Class III and IV. In the case that no clear and measurable goals can be found,
one may try "I will spend two hours in the morning en some section of the problem,
two hours in the afternoon on other section of the problem" (Le. we use time
allocation as goals; "two hours" is arbitrary, just for an example).
Through implanting, accepting and habituation, the selected goals will have
impact on our behavior and overall performance. As time passes, it may be a good
idea to review whether the habituated goa'ls are still effective for overall per-
formance.
Once the set of goals are selected, we need to specify at what level and
within what time frame (minutes, hours, days, weeks, etc.) in order to achieve
optimal performance.
The more specific and difficult the level, within the acceptable range, the
more likely is the setting to create a high level of charge and obtain the atten-
tion allocation. As a consequence, more drive and performance will be produced.
In the extreme, in Class I of simple routine problems, industrial engineers have
used motion and time studies to set the "optimal" level of production outcome for
each worker, which has proved to be very successful in so called "scientific man-
agement".
Note that the high level of goals can create a high level of charges, which
may be short-lived if the workers do not have the confidence and ability to a-
chieve it (through avoidance justification (Box (5». Proper training, assistance
and progress report are some important means for maintaining the confidence to
achieve the set goal levels. Again through implanting, training, accepting and
habituating, the performers of Class I-II (simple routine - mixed routine prob-
lems) can come to accept and commit themselves to achieving the set goal levels
and improve their proficiency in job performance.
For the problems in Class III-IV (fuzzy and challenging problems) the time
to achieve the set goals may be highly uncertain which may thwart the confidence
to continue the work. In addition to a good support system, the following may be
helpful in maintaining the confidence and drive to achieve the goals: (i) decom-
posing the overall project into a number of subprojects and trying to complete the
334
subprojects one at 3 time; (i~) using self-suggestion to see the rosy and promis
ing picture of completing the project to reinforce the burning desire (charge) to
achieve the goals; and (iii) caLling for belp from experts and other useful sour-
ces. In the athlete problem, the coach can set ~e goal level progressively,
starting with 10.8 seconds, then 10.6, 10.• 4., 10.2 secous, ••• etc.; he and the
student can renew and reinforce the rosy picture of becoming an Olympic star to
encourage themselves.
Finally, in ~etting the goal level for other people to perform, partici-
pation of the workers or performers may be important. ~his is especially so when
the problems are of Class III and IV (fuzzy and challenging). Participation al-
lows us to better see the HDs of performers and bet-ter under·stand the problems.
In addition, because of participati-on., the perfoT1llerc may obtain a feeling of
importance and be more willing to .commit themselves to achieve the set goal le-
vels. However, for simple or lIIi:xed routine prob-Iems (Class I or II), as the set
of ideas, concepts and operations needed for successful execution of the job is
known, more objec·tive assessment for .achievement level's. can be obtained (for in-
stance, motion and time studies) which can be used to set the goal level effect-
ively. In these cases, the impact .Q!I overall performance by worker participation
in setting goal level msy not be significant.
Training and assistance can increase. the perforaers. ability and confidence
to achieve the preset goal levels; while supervising ean, in addition, create
their charges and attention to job-related work- Note that training involves
implanting, accepting and habituating processea of ideas, concepts and operations,
which is very important for the beginners to achieve the preset goal levels, es-
pecially those doing simple and mixed routine works. Assistance is usually needed
when unusual problems occur in the routine work. Timely assistance can reduce the
performers' frustration and leakage of attention to work, in addition to increas-
ing their ability to cope with unusual problems.
Human beings have an innate need for external information to fulfill their
life goals. Information on their performance and rewards are important to ·them
for assuring themselves how'well they have been doing (related to self-important
feeling and social approval category of Table 1). This information is an import-
ant input to State Valuation (Box ~2», which can affect charge structures, atten-
tion, confidence, drive and overall performsnce.
In the athlete problem, keeping record o~ the p~gress of Tom Student and
showing it to him may provide a main thrust to increase his drive for further
improvement. Note that more detailed monitoring and feedback could also improve
the training and ability to perform.
In fuzzy and challenging problems (Classes III-IV), the monitor and feed-
back may be more difficult. People may be periodically self-assessing as to as-
certain the progress and situation of the problems.
We work to fulfill our life goals (Table 1). The reward of the work is
thus vital in providing the charge, attention and drive to achieve the preset goal
levels. Usually, the larger the reward, the larger the charge and drive to accom-
plish the goals.
Note that the reward systems should be broadly considered. These include
monetary payment, authority, power, position and promotion. Not only positive
rewards but negative rewards need to be considered. To be effective, the reward
systems need to create for the performers a large difference between the charge of
~ achieving the goals and that of achieving the goals. The large difference,
which signifies the importance of achieving the goals, can command the attention
336
not sustain. This problem can be severe especially for fuzzy and challenging
problems. In the athlete ·problem. if Tom Student has no high regard for the re-
ward of becoming an Olympic sta~. or has no confidence in obtaining it. then most
likely he could not generate the continued high level of drive necessary to become
an Olympic star.
with
if at time t attention is paid to the work
otherwise
Note that the total attention time allocation to the work during the time
337
At = J
T
at(X t , Yt , Zt; Ut , Vt;£t) dt (7)
°
Now let the efficiency and effectiveness in perf~rming the work be repre-
sented by
Et ....lIt = Et(Xt,Yt,Zt; Ut,V t , £t) (8)
Then the principle of effective goal setting for the time horizon [O,T] can be
speci~ied as (i) selecting p clear and measurable goals and determining their
achievement levels, Ut £ U; and (ii) selecting t~e .supportive systems Vt £ V, so
that the total attent-ion time a-llocation to the work "'r and the efficiency and
effectiveness to perform the work ET are maximized, where ~ and ET satisfy the
dynamics of (3) - (8).
Many solution concepts can be utilized for solving the above two criteria
problems. For instances see [18,21].
Note that the favorability of the workingenvirO'l:lment, which may be impor-
tant to fuzzy and challenging problems (Classes III and IV), can be similarly
formulated as an objec~ive function. We shall leave it to the reader.
Also, instead of two criteria,
and ET , maximization, we may consider
~
only one criterion by aggregating the two criteria into one. For instance, one
may simply consider the performance index of
T
PT ~ at • Et(Xt'Yt,Zt;Ut,Vt;£t)dt
when proper assumptions are made to make the above integral meaningful. Again, we
shall leave this to the reader for further exploration.
Finally supportive systems can be costly. The reader may want also to
consider cost in the formulation.
Assuming sufficient ability, Locke and Bryan's research in 1966 [10], using
70 university students to perform the Complex Computation Task, found that there
was a positive relationship between goal difficulty and task performance. Impli-
citly, it was assumed that the goal is internally accepted by the individual. In
the 1981 review, Locke, Shaw, Saari and Latham [12], concluded that, since 1969,
48 studies partly or wholly supported the hypothesis that hard goals lead to bet-
ter performance than medium or easy goals. 9 studies failed to support it. Ia
general, it is valid to say that difficult, challenging, but realistic goals will
enhance task performance. Although we know the positive relationship between goal
difficulty and task performance, we are not sure whether the performance will
increase at an increasing, decreasing, or constant rate with the increase in the
degree of goal difficulty.
By asking 10 male and 10 female paid college student volunteers from Uni-
versity of Maryland to perform a simple addition task, Bryan and Locke [1], con-
cluded that specific goals can serve to raise the motivation of subjects to a
higher level. The assigning of specific and hard but realistic goals to the sub-
jects raised the task performance and favored the development of more positive
attitudes toward the task. Dossett, Latham and Mitchell [3], with female clerical
personnel as subjects solving simple arithmetic problems, found that specific
goals led to higher performance than did the "do best" goals. Similar results
were' found in Locke [9] and Locke and Bryan [10] studies, by using college stu-
dents as subjects to do simple arithmetic problems. In their 1981 review [12],
Locke, Shaw, Saari and Latham stated that individuals with specific and hard or
challenging goals outperform individuals with specific easy goals, "do best"
goals, or no goals. When goals are set in quantitative terms, they tend to be
clearer, and more specific to the individual.
339
Knowledge «>f re'sults has long been an important variable in goal setting
theory. Many studies have been done to determine its effect on task perform-
ance. Cummings, Schwab, and Rosen'.s [3] field experiment with college students as
subjects performing simple addition task showed that correct knowledge of results
(KR) increased goal level significantly above that generated by no KR. Dossett,
Latham and Mitchell [4], through asking 60 female clerical personnel to perform
simple addition, concluded that, with goal difficulty held constant, there was no
significant effect found for KR on task performance. Similar results were found
in Locke [9] and Locke and Bryan's studies [10]. Both studies used college stu-
dents to perform simple arithmetic work. However, given specific, hard goals, a
significant goal effect was found between KR and no KR subjects. In conclusion,
[12] stated that KR appears to be necessary if goals are to improve performance.
KR achieves its motivational effects through the incorporating of goal setting and
that in the absence of goal setting KR alone is not sufficient to improve perfor-
mance.
5.5 Rewards
5.7 Supportiveness
The experiment conducted by Locke, Fredrick, and Bobko [11] by using col-
lege undergraduates as subjects to perform the job of giving uses for common ob-
jects indicated that ability, self-efficacy, goals, and task strategies were all
related to task performance. No reliable individual difference factors, other
than ability, have emerged in the goal-setting literature, probably because most
of the studies have used assigned goals [12]. Other things being equal, greater
ability should lead to better task performance.
fects. There is probably an optimal time span for the setting of goals that may
depend on both the individual and the task situation [121.
6. Conclusion
References
2. Chan. S. J., and Yu, P. L., "Stable Habitual Domains: Existence and
Implications," Journal of Mathematical Analysis and Applications, Vol.
110, No.2, 1985, PP. 469-482.
3. Cummings, L. L., Schwab, D. P., and Rosen, M., "·Performance and Knowledge
Of Results As Determinants Of Goal Setting," Journal of Applied Psy-
chology, 1971, 55, PP. 526-530.
10. Locke, E. A., and kyan, J. F., "The Effects of Goal-Setting, Rule-Learn-
ing, and Knowledge of Score on Performance," American Journal of Psy-
chology, 1966, 79, PP. 451-457.
342
ll. Locke, E. A., Frederick~ E., and Bobko, P., "Effect of Self-Efficacy,
Goals, and Task Strategies. on.Task Performance,"Journal of Aplied
Psychology, 1984, 69, PP. 241-251.
12, Locke, E. A., Shaw, K. N., Saari, L. M., and Latham, G. R., "Goal Setting
and Task Performa'llce: 1969-19Se.~" Psychological Bulletin, 1981, 90,
PP. 125-152.
15. Ronan, W. W., and Latham, G. P., "Effects of Goal Setting and Supervision
on Worker Behavior in an Industrial Situation," Journal of Applied
Psychology, 1973, 58, 302-307.
17. Yu, P. L., "Behavior Bases and Habitual Domains of Hu1JIaU Decision/-
Behavior - An Integration of Psychology, Optimi2stion Theory and Com-
mon Wisdom," International J-ournal .of Systems, Measurement and Deci-
sions, 1981, 1, PP. 39-62.
19. Yu, P. L., Behavior Mechanism and strategical becisions, the Foundation
and Applications of Knowing Yourself and Others, Academic Sinica
Press, Institute of Economics, Taipei, Taiwan, 1985.
20. Yu, P. L., and Huang, S. D., Behavior Bases and Habitual Domain Analysis
- Theory and Applications, (to be published).
21. Zeleny, M., Multiple Criteria Decision Making~_~cGraw Hill, New York, New
York, 1982.
IV. ENGINEERING APPLICATIONS
MULTICRITERIA OPTIMIZATION
PROCEDURES IN APPLICATION
ON STRUCTURAL MECHANICS SYSTEMS
by
Hans A, Eschenauer
Research Laboratory for Applied Structural OptImization
at the
Institute of Mechanics and Control Engineering
University of Siegen FR GERMANY
J
1. INTRODUCTION
IIII Sh ap e conditions
:\
Constraints 9. (~) sO
Feasibility
Study
,,,..,. ;'''";';"~
Global behavior
Project -Compliance
I--
/
Definition Phase with the most
important
demands
\
Phase Local behavior
-Compliance with
I
certain demands
1/
for single
components
(subsystems)
Design Phase
(Subsystems)
Appl ication: Shape
Optimization
I
Manufacturing
Erection
Tesling
349
2. MATHEMATICAL FUNDAMENTALS
f objective function,
This is called:
.;.;
~-
,'"
,
I I
I
/ ax
1/
'/
Ug. '2.. 1 ;
,/ E66~e~en~ vee~on
.- '"
"" and eolian.
=--...Bcr--'~~ .:: -: ~ __
351
Criterion SRoce
The main program SAPOP prepares the data exchange between the
optimization algorithms and the structural analysis. The actual
optimization process is coordinated by the selected optimiza-
tion algorithms. The respective structural analysis program is
constantly integrated into the computations being made in order
354
Optimization
Algorilhms
Model of
actual design .r.
SEOLI 2
SEJlML I--=-J!_""I OptimlZotlon
SLIP40 Modelling
~ I--'~-:-i.-g~-:-i-I FUNC
t:JLe!2l".
GREGA
LPNLP
SONP
EXIBEM
E.S!l
L (x,a., e,w)
a----
with
q 2
P1 = 1: [h.(x)]
i=1 ~-
P2 = 1: [g . (x) ]
J -
2
C
a
= {j ej > o} (3.1)
J'€C a
P3 1: [g.(x)]
J -
2
Cb = {j I ej o und g. >
J - o} .
j€C b
a) Minimization phase
minimization problem.
L
a
(A~)
~
= Min {L
A. a
(x. + A.S.)} •
-~ ~-~
~
x'+ 1 = -~ * s .•
x. + A.~
-~ -~
x
k+ 1 x* with La(~*,~k,~k,~k) = Min {La(~,~k,~k,~k)}
xEIRn
T
L (x) RJ L (x.) + 'VL (x. )
a - a -~ a-~
-1
s. - H. 'VL (x.) (3.3)
-~ =~ a -~
b) Correction Phase
(3.4)
VL(~*,~*,~*) VL (x*,a*,~*,w) = 0 .
a - - - -
with j E C "
a"
~ k k+1
{
0 for ~. + 2 w2 g. (x ) ~ 0
~. ~+1 = J J -
J ~.k + 2 w2 ~ gj(~ k+1 ) others;
J
with j E Cb :
{
0 for g. (x ~+ 1 ) ~ 0
~. ~+1 J -
J ~ ~+1
2 w3 g. (x ) others.
J -
358
Ini tializin9_
~,~,~,~
Minimization p-hase
!';Ik+ 1~ Min La ( !,;, ~ Ikl, ~(kl , ~(kl)
~
Correction p-hase
Determining of the values for
~(k+1~ Q(k+1l ,~(k+1l
{
w. for wi wimax
k+1 lmax
w. i=1,2,3.
1 k others
YW i
dy dy
~ Y + ~ - + c (3.4)
dcp dcp
with
T
Y (rN cpcp , rQ cp , rMcpcp , u, w, X) state vector,
c load vector,
~, ~ matrices.
resp. LLQ
cp
+ 4]..1.4 Q
cp o (3. 5b)
~i = ~i ~i-1 • (3.7)
Here, the ~i are the modified state vectors and ~i the transfer
matrix. With r~gard to the condition of compatibility, whereby
the same forces and moments must be transferred and the same
deformations must exist when two elongated twills have a
common intersection point, the transfer procedure between
boundary i=O and i=n of a shell of revolution is
~n = ~l. ~i ~O =~ ~O (3.8)
B
11
u'" + B
22
V· + (B
12
+2B
33
) (u' +v") + K11 W'"
x + K22~Y' +
- p
(3.9)
.,
B 22 v + B33 v " +(B33+B12)~' + K22Wy + K33 W; +(K 33 +K 12 ) WX -
- S22(W y +w) = 0 .
~(~) d - f 0 (3.10)
m
L [w J' f. (x) ] (3.11)
j=1 ] -
where m
o < Wj < 1 , L w.
J
j=1
-
(f.(x) - y.) 2
r = 2 (3.13b)
] - ]
EUCLIDian metric,
z. (x) - 13 ~- 0 x E IR n + 1 , (3.16)
J -
j=1, •.• ,m
w, 1
Wi=:r
Y, I, I, I, I,
(4. 1 )
a
R* E [" := {~(E ) : IR 2 -+
3
IR } ,
_---tl~_~2
middle
surface
~1 < -R < -u
R (4.2c)
369
(4.3)
I Approach funcl,on
I B{1;·.~)
~l S P :s. ~u
I IOphm,,"',on Strategy
L pI .!.{~))
I
~
Scalar Optimization Problem
Min pl.!.{~)J ~ E IR"
optimal
hk{~) = 0 . k =1 .. ... mh => compromise solution &.
9l (~) sO. I =1 .. ..• mg
!5. l ~ ~ :s ~u
(~
t1 - t2
,
t(~,~) = t1 - (~-~1 )
- ~2
~1 ~ ~ < ~2
~1
t3 - t4
t(~,~) t3 - (~-~ 1 )
- ~4
~2 ~ ~ ~ <!l3 (4.4)
~3
n;
t(~,~) t4 < <
~3 ~
"2
p(x.~)
T= 650 kN
r
... p(x.~) = Po + P, cos ~ +
60+---
50 +----\l----j--
40 +----+,~-t-
,----,----,
I I
20
10
G
o ,----l---4---I---+--9-j.0--1~00::-;(kNl
The example of the conveyer belt drum is also used for testing
various optimization algorithms combined with strategies. It
is generally not possible to give an optimization procedure
and an optimization strategy in such a way that they definitely
lead to a good result in all cases. For this reason some
combinations for particular cases of application, among others
the conveyer belt drum, were evaluated via quality criteria
in numerous investigations [6]. The criteria consider four
different aspects:
optimality,
velocity,
quality,
security.
373
Q = L. 1w.P.
.1.
i=1, ••• ,4 L w.1. = 1 • (4.5)
1.
COT MM OW OF COT MM OW OF
5. CONCLUSION
The software package SAPOP was developed as an optimization
procedure on the basis of the three columns "Structural
Analysis", "Optimization Algorithm", and "Optimization Model-
ling". SAPOP connects problem-independent Optimization and
Structural Analysis Methods by a so-called problem program.
An essential part of the problem program is the transformation
procedure in order to transform a Vector-Optimization-Problem
into a Substitute Scalar Problem (SSP). The application and
tests of the efficiency of the developed optimization methods
on a special task of Structural Mechanics, e.g. shape opti-
mization of a shell structure (conveyer belt drum) is shown.
Finally, several combinations of Optimization Strategies and
Optimization Algorithms have been tested by establishing
valuation criteria exemplified by a conveyor belt drum. This
is the first step towards establishing a knowledge based
system for an expert system.
APPRECIATION
The author wishes to express his appreciation and thanks to
Miss Baerbel Huette who typed this chapter with great patience.
375
REFERENCES
O. Introduction
with
a(v,w) ...
r
J (L
m QV ()w
k ( x ) - - + q(x)v(x)w(x) }dx ,
i=l ()XjaXi
B
(f,v) .. ) f(x)v(x)dx ,
B
1
K" (VEW 2 (B) : v(x) 2: f(x) XE B, v i r " g}.
m
The set Be R, m 2: 1, denotes a bounded domain with piecewise smooth
boundary r .. aB and the functions involved are assumed to be suffi-
ciently smooth to have a corresponding regular solution u of problem
(1.1) (cf.C13]). In this case the solution of the variational inequa-
lity (1.1) is equivalent to the solution of the extremal problem
ueK J (u) ... min ( J (v) I v E K )- (1. 2)
with
1
J(v) .. -a(v,v) - (f,v).
2
1
If u E W (B) is a solution of (1.1> (resp. of (1.2», then the domain B
2
decomposes into the following two s\.\bsets:
.. = f(x)
..
B ( X E B u(x) } (Coincidence set) ,
B
+
... { x E B U ()I ) ). t(x) } B \ B
2
(cf.[17J) is valid. i.e. let u e. C (B) f'I C(B) be a function satisfying
the following two inequalities.
The general way for solving infinite extremal problems (1.2), however,
is their approximation by finite programming problems. In this con-
crete case the approximation is done by the method of finite elements,
1
where the appro>:imation of the space W (B) is convenient by using
2
linear elements (cf.[9J, [10J).
Let B a Finite-Element-Model of B which arises from a regular trian-
h
gulation of the domain B, then the following notations are used:
i
x - vertices of the triangulation B , N - number of vertices, w (x) -
h I
basic functions of the Finite-Element-Space V - span(w , ••• ,w ) and h
h 1 N
denotes the discretization parameter. Since the finite dimensional
N
function space V . is isomorph with the Euclidean space R the follo-
h
wing identification is possiblel
T 1 N T N
V t V v := (v , •••• v ) .. (v (x ), ••• ,v (x » E R •
h h 1 N h h
respectively
Ir } .
N N
K '" { vER L v w (x) .::: f(>: ) )( E B v '" g.1 ~ (1. 5)
N i-1 i h'
f(>:
i
)
'" f 1 .. 9 (>{
i
) . g. ,
1
(f ,w.
1
)
'" f 1. , f . (f , •••• f )
1 N
T
,
a(w . ,w.
1 J
) . a
i j
A '" {a }
ij i ,j"'l
N
If Au ~ Av and u
i
S. v
i
i £. Ir , then u ~ v .
The solution u of (1.3) satisfies the necessary and sufficient opti-
h
mality conditions which have the following forml
( Au) 2: fiE I \ I r
h i i
(1. 7)
«Au ) f ) (u 'f ) =0 i £ 1\ Ir
h i i ih i
both methods have the properties which are in the following described.
To ensure "good" convergence order of this method a direction search
problem is considered which determines the global best descent direc-
tion. Therefore, the solution of the direction search program at the
k
iteration point v is determined by the direction
k "'k k
r := v v
"'Ie
where v is the solution of the following linear equation system:
"'k k
(Av ) f i ; It.(v )
i i
"'Ie Ie k
v == v i 4i I t. (v )
i
"'k
v g1 i e Ir
i
I: I:
with I£(v ) = { i £ I \ Ir : v
i
'f
i
;5;, E. }.
Ie "'I: Ie I: "'k
The direction r := v - v at the point v £ K, where v is
N
generated by (2.1), is feasible and is the global best descent
at the linear manifold
'"
K
N
(v
I:
) == { v£R
N
v
i
= v
k
i
i E It. (v
k
) , v
i
.. gi i t Ir } .
k
The direction r is the local best descent direction of the regula-
rized functional
1 1 -1
J A-1 (v) = -(v,Av> -1 - (f,v>A-1 = -(v,v> - <A f,v>
2 A 2
k
on the linear manifold K (v). Because of the positive definiteness of
N
the matrix A the norms <.,.) and <.,.> _1 are spectral equivalent.
A
Now, in order to compute the iteration point
1:+1 I: I: k
v := v + t r
384
k
the step-size t is determined as maximal as possible, i.e.
k+1 k k k k
J(v ) .. min (J(v + tr ) t 2: 0, v + tr E K } (2.2)
N
k k k k
t .. min {1, min {(f - v )/r r < 0 }} (2.3)
k i i i i
i;I£(v)
k k k k
t .. min ( i , min k{(f(x) - v (x»/r (x) r (x) < 0 }} (2.4)
i;BE.(v )
where
r
k
(x) .. I: r w (x) ...
N
i=l
k
i i
1::
i=l
N "'k
(v
i
- v
k
i
)w (x)
i
BE(v
k
) .. h
k
(x f; B : v (x) - f(x) < f. } .
Because of the construction of the algorithm the following properties
k
are satisfied for the iteration points v •
k
For the sequence (v ) the following relations are ensured:
k+1 k
(1) J(v ) < J(v) "fk,
k+1 k
IE(v ) ::> I E (v ) 'f k,
k k+1
<i v) If I (v ) ... I (u ), then v .. u
o 0 h h
385
k
Because of the monotonous increase of the £- active index set I£(v )
the convergence of this method of feasible directions is not guaran-
teed for any feasible starting point. By using the discrete maximum
principle it is possible to prove some properties of monotonicity of
k "'k
the sequences (v) and (v ) which guarantee the convergence of this
algorithm.
o
Let the starti ng point v ~ K be chosen as solution of the linear
N
equation system
0
(Av ) 0 i E I \ Ir
i
-
(2.5)
0
v g. i " Ir
i 1
with
'"
g. 2: g. £ Ir
1 1
Because of the fact that the matrix A is the discrete analogon of the
elliptical differ.ential operator L problem (2.5) describes a discre-
tized homogenuous differential equation with boundary conditions of
o
first kind. Therefore, v always exists.
Theorem 2.3.1
o
If the starting point v is chosen by (2.5), the following monoto-
nic properties are satisfied.
k k-l
(1) (Av ) 2: f i E I \ ( Ir v I £ (v », V k ~ 1,
i i
(i 1> v
k
2: 1.1
h
¥ k ~ 0 .
Proof. The proof is given with the help of complete induction with
respect to the iteration steps. Because of the linearity of the opera-
o
tor A and because of the assumptions on the starting point v the
following estimations are true:
(1.1
h i
) - V
0
.. g1 v
0
i g. - g.
1 1
.. 0
i " Ir ,
(1.1
h i
) - V
0
i
.. f
i
- v
0
i
i f
i
f
i
.. 0 i E I
0
(1.1
h
)
(A(u
h
- v
0
»
i
(Au
h i
) - (Av
0
)
i
(Au )
h i
.. f
i
i 0
i , I \ (I r v I (1.1 ».
o h
386
o
(u) - (v) i 0 i f. I \ (I r v I (u »
h i i o h
and
o o
u - v i 0 i.e. u i v
h h
1 o
(Av) '" (A (t (v
0 "'0
- v ) + v »
0 0 0 "'0 0
.. t (Av ) + (1-t ) (Av ) ...
i i i i
(2.6)
o "'0 0 0
... t (Av) ... t f i .If i e I \ (I r v I e. (v ».
o
Consider (1.6) and t f. [O,lJ then it follows
1 o
(Av) ! f i ( I \ (I r v I E. (v ».
i i
1
Setting v .. u - v the following inequality is true
h
o
v .. (u ) - v ... gi - (g + t (g - g » ..
i h i i i i i
o
(l-t ) (g - g ) i 0 for i f. I r
i i
o 0
v (u) - v ... (u) - v i 0 for i e I < (v ).
i hi i hi i ..
o
For i E 1\ (I r v If(v » follows from (1.7) and (2.6)
1 o 0
(Av) ... (Au) - (Av) .. f - t f = (1-t )f So 0 •
i h i i i i
k-l k-2
Lemma 2.1. provides I£(v ) ~ I£(v ) and therefore
k k-1
(Av ) l f for i E I \ (I r lJ I t (v »
i
k
Now, consider v LI
h
- v the relations
v S. 0 i E Ir and v ~~ 0
i i
are satisfied again.
With (1.7) and (2.7) it follows
k k
(Av) (Au) - (Av ) f - (Av ) s.f ·-f =0
h i i i i i
k-1
for (I \ (I r v I € (v ».
k k-1 k-1
(Av) S. (Av ) ¥ i (I \ (I r lJ I f. (v ) ), 'rf k
k k-1
This property leads to the inequality v ~ v -'rf k.
Furthermore, the following relation with respect to the total
step length holds:
k k k
(Av) = t (v ) f "'I i £ I \ (I r U Ie (v », .!f k ,
I( k+l
this leads to t(v ) S. t(v ) "'I k.
k
Therefore, the total step length t(v ) can be used as a measure
for the convergence of the method.
(:I
Remark 2.4.: If v is chosen as in (2.5) it can be shown in the same
"'k
manner that the sequence (v ) has the following property:
UsiAg the arguments of Theorem 2.3. and Remark 2.4. the following
statement about the convergence of the algorithm can be proved:
Theorem 2.4.1
o
Let the starting point v be chosen as in (2.5). Then there
exists an index k with
o
and a i u u i Se ,
h
k
t L 1 - £ (for a given £),
s s
k
since the condition t - 1 is equivalent to
ko +1 "'ko
v '" v and
o o
91: Choose v ~ K (Av) '" a
N i
"'k k
(Av ) f i ; I E (v ) ,
i i
"'k k Ie
v '" v i E I£(v ) ,
i i
"'k
v g. i ~ I r
i 1
389
931 Determine
t
I:
.- min { 1 , min
i , I t (v
{(op
i
- v
I:
i
)/r
I:
i
r
I:
i
<a )} .
941 Put v
1:+1
.= v
I:
+ t r
I: I:
, I: 1= k+1.
I: I:
951 If t ~ 1 - f stop 1 v i s f - opti mal, otherwi se go to 92.
51
Because of the fact that in the step-size problem (2.4) the restric-
tion set of problem (2.3) is included, the step length by using Algo-
rithm B can not be larger then by using Algorithm A assuming the
k
direction r is the same in both cases. Therefore, the following
connection between the total step length of both algorithms can
be provedl
I: I:
(v) and (v) denote the sequences which are generated by Algo-
A B
rithm A and B respectively.
I: k k+l 1:+1
If t(v ) ~ t(v ) then t(v ) ~ t(v )
A B A B
o 0
Because of the fact that both algorithms start with t(v ) .. t(v ) .. 0
A B
k 10
and terminates with t(v 0) .. t(v ) = 1, the faster convergence of
A B
Algorithm A in face of Algorithm B can be concluded from Lemma 2.2 ••
Theorem 3.1.:
k I
II v
h
- v
h'
I 100 (3.1)
and
"'k "'1
II v
h
- v
h'
I
I
I
'00 (3.2)
k I
(i) for the step length: t - t
h h'
k+1 1+1 c hl'1
(ii) for the iterates: Ilv v h ' 1100 So 4
h
where ~ .. min {p,q},
i.e., if the previous iterates are closely related then the follo-
wing iterates do it too.
Proof:
i
(i) For the simplicity it is assumed that B ~ B then every x £ B
h' h h
also belongs to B • Denote
h'
k
t .. «v k ) f )/«v ) - (v
"'k
) )
i h i i h i h i
respectively
391
1 1 "'1
t == «v ) 'f ) / ( (v ) - (v ) ),
i' h' i ' i' h' i" h' i '
i i '
the step length at the point x respectively at x
Then the step length of the method is determined by
k 1
t '" mi n { t liE I } respectively t = min { t i'E 1')0.
h i h' i'
k 1 1
Without loss of generality let t ~ t be and t is determined at
h h' h'
the step i'. Obviously the following inequality holds:
o
k 1
t - t S. t - t (3.3)
h h' io i~
k "'1 "'k 1
- (v ) (v ) + (v ) (v ) )0 ..
h I h' i' h i h' i '
1 k 1 "'1 "'k
'" - {[(v ) - (v ) ].pi + [(v ) - (v ) ] f +
n h i h'i' h'l' hi i
"'k 1 "'1 1 "'1 1
+ (v ) (v ) - (v ) (v ) + (v ) (v )
h i h' i ' h' i ' h' I ' h' i' h' i '
k "'1
- (v ) (v ) )0 ==
h i h' I '
1 k 1 "'1 "'k
n
{[(v
h i
) - (v
h' i '
) ] f.1 + [(v
h' I '
) - (v
h
)
i ] 'fi +
1 "'1
Because of the continuity of the functions~, v and v In B there
exists a constant c Independently of h such that
392
c 1 c "'1 c
l'f I ~ -
i 2
I (v
h' i '
) ~
2
and I (v
h'
)
i
~ -2
Moreover, because of the construction of the search direction we have
k "'k k
(v ) - (v ) (r ) > 0
h i h i h i
and
1 "'1 1
(v ) - (v ) .. - (r ) >0
h' i ' h' i ' h' i '
such that
A
there exists a constant c independently of hand h' with
lin ~ c. Therefore,
I( 1 "'1 "'k
It - t ~ cc I (v ) - (v ) I+ cc I (v ) - (v ) I
i i ' h i h' i' h' i ' h i
and together with the assumptions (3.1> and (3.2) the estimation
A P .., q
It t ~ ccc h + ccc h 5. c h""
i i ' 1 2 3
1 1 k k 1 k
+ t
h'
(v
h'
- v
h
) + v
h
(t
h'
- t
h
)
The step lengths are bounded and the concerned functions are conti-
nuous over B, therefore a constant c exists independently of h such
that
k 1 k 1
It ~ c , It :c;. c: 'I Ilv
h
: I s. c: and I Iv II 00 ~ c.
h h' 00 h'
393
is true and together with (3.1),(3.2) and part (i) of the theorem the
proof is complete;
P q
I I v k+1 - v 1+1 I I 00 ~ c (c h I" + c h + c h ) ~
c hl>'
h h' 3 1 2 4
Remark 3.1.1 The assumptions (3.1) and (3.2) are never of restricted
kind. The meaning of condition (3.1) is that the iteration solu-
tions constructed by the algorithm are closely related and Theorem
3.1. shows that the following iteratws have the same property.
Condition (3.2) is a demand for the approximation of an elliptic
boundary value problem. Hence the existence of q is guaranteed.
k
Remark 3.2.1 The proof was given for the case that v is constructed
k
by Algorithm A~ If v is generated by Algorithm B in the proof it
is taken into consideration that the step lengths t and t a r e
ii'
to calculate not only at the discretization points but also at an
arbitrary point Xi B. But for the step length at the point x I: B
the same estimation is true as at the discretization point.
The()rem 3.2.1
k-1 1-1 k
Ilv - v and t (v ) < 1 •
h h' h
k A
k k
I Eo (v ) ;6 0 , I E(v ) c I E (v )
h h h
such that I E (v
k
h
) c I E (v
1
h'
) ¥ h' < h .
394
k k
Proof. Since t(v ) < 1 the coincidence set I£(v ) is not empty and
h h
k k
an index i ~ 1£ (v ) exists for which (v )
o h h io
k k ...
,
A
I £ (v ) { i E I I (v )
'Pi S. £.
). with 0 < E. < E. (3.4)
h h i
k
A k
Since i E I £ (v ) the set It(v ) is not empty. Using Theorem 3.1. , an
index r 0 h
= min (p,q)
h
(q analogous to (3.2» exists such that
k I k I JA
Ilv
h
- v
h'
II
GO
S. ChI' i. e. I (v
h i
) - (v )
h' i '
S. ch
A
k
Thus, for i E I £. (v ) the inequality
h
(v
I
h' i '
) - 'f
i '
S. (v
k
)
h i
+ ch
,All
- 'P S. E + ch
i
JA
S. f.
Remark 3.3: The statement of Theorem 3.2. is also true for the case
k
that v is generated by Algorithm B. Therefore, the index set
A
k
IE,(V ) is to replace by a set
h
... k k
BE (v ) = ( )C ~ B I V (x) - 'P(x) S. E. ). , (3.5)
h h
The following lemma describes the variation of the total step length
if the discretization parameter is modified.
k
Let v and v be generated by Algorithm Band let
h h'
k 1
t(v ) 5. t(v h' ) be.
h
k+1 1+1
t(v 5. t(v holds.
h h'
k k k ..... k k
I.(v );Ii 0, Ir(v ) c I.(v ) such that I.(v ) c IE,(v ) ¥h'( h.
Co h c. h co h c. h h'
Ie Ie
Now, the algorithm continues with Ie(v ) .'" Ie (v ), i.e. , at the
h h
k Ie
iteration step k the set Ie(v h) is replaced by the subset Ie(v h)'
The iteration sequence which is generated by this modification is
"Ie
denoted by (v). This modified algorithm stops after finite
h
Ie
number of steps (infact of (3.4) the modified method with IE. (v )
h
is interpretable as an algorithm with t - enlargement).
396
Lemma 3.3.:
"'k
The sequence (v ) which is generated by the modified Algorithm B
h
converges to an £ - optimal solution of problem (1.3) after a
finite number of steps, i.e., there exists an index k such that
o
and 0 s. u u
h
S. £e •
Theorem 3.3.:
Ie
Proof. For arbitrary h' the sequence ( v ) , generated by Algorithm A,
h'
is compared with (v
h
..... k
h
) . Since
) with f. < £ can be chosen as in
I£(v
k A
(3.4) , Lemma 3.1- and 2.2 provide the following estimation for the
step length at the O-th step:
o 0 0 ..... 0
t (v ) 2: t (v ).
h' h
1 A1
This leads to the total step length t(v ) 2: t(v ).
h' h
k ..... k
Now, at the step Ie let t(v ) 2: t(v ) be. As an immediate consequen-
h' h
ce of Lemma 3.2. and 2.2.
Ie ..... Ie
t(v 0 ) 2: t(v 0) =1
h' h
4. Numerical Results
2 2 2 0 .. 5
B { X E R r :c (>: + x) < 0.5 } •
1 2
o
The starting point is v (x) '" O. the domain B is discretized by a
regular triangulation with the grid constant h.
Table below contains the iteration number k for the computation of
problem (1.3). (1.4) with Algorithm A for several discretization
parameters h (€ '" 0.001).
The iteration numbers for other numerical methods solving this problem
with the same precision are described in Table 2.
Table 1:
Table 21
0.1 109 4 3 74
0.05 397 5 3 76
0.025 1519 8 6 80
0.0125 6109 9 10 97
Literatur
P.D. Robinson
College of Science, Sultan Qaboos university,
PO Box 32486 Al Khoud, Sultanate of Oman.
and
P.K. Yuen
School of Mathematical Sciences,
University of Bradford, Bradford, U.K.
I. INTRODUCTION
where
J J(V,U) -<V,AU> + <V,f> + <g,U>, (6)
S S(V,U) <AU - f, A*V - g>, (7)
Au - f = 0, A*v - g = 0, (9)
(v also assumed unknown), and we note that
J(v,u) = <g,u>. (10)
If
U = u + ou, V = v + OV (11)
then
J(V,U) - <g,u> = -<ov,Aou> (12)
and as the right-hand side of (12) has indefinite sign then J(V,U)
does not provide a direct bound for <g,u>. However if the trial
vectors U and V are constrained to satisfy
t(AU-f) + t- 1 (A*V-9) = tAou + t- 1A*OV = 0, t>O, (13)
and the resulting functional J is designated J+, then
J+ -<g,u> = t- 2 <ov,A*OV> ~ 0. (14)
Similarly if the constraint
t(AU-f) - t- 1 (A*V-9) = tABu - t- 1A*Bv = 0, t > 0, (15)
specifies J_ , then,
J_ -<g,u> = -t- 2 <ov,A*OV> S 0. (16)
Thus we have the constrained bounds
J+(V,U) ~ <g,u> ~ J_(V,U), (17)
which may be difficult to implement directly as the inverses of A
and A* are assumed unknown.
The canonical structure of these bounds is exhibited by working
in terms of the new vector-variables:
X+Y = tU, x+y = tu, X-Y = t- 1v, x-y = t- 1v, p+q = tf, p-q = t- 1 g (18)
and the operators
L = i(A+A*), M = l(A-A*), (19)
these latter denoting the self-adjoint and skew-self-adjoint parts of
A. The equation-pair (9) becomes
Lx + My - P = 0, Ly + Mx - q = 0, (20)
and we have
J(V,U) = I(X,Y) = -<X,LX> + <Y,LY> - 2<X,MY> + 2<X,p> - 2<Y,q>. (21)
Since from (4)
<Z,LZ> ~ b<Z,Z>, b>O, (22)
the functional I(X,Y) is convex in Y and concave in X. It thus
possesses the saddle property characteristic of dual variational
bounding principles [~]. In fact
J+ = 1+ = I(X,Y): pI/~X = °<=> LX + MY p = 0, (23)
and
J_= I = I(X,Y): aI/~Y °<=> LY + MX - q = 0. (24)
402
Specifically
I+ = <Y,LY> - 2<Y,q> + «MY-p), L- 1 (MY-P»
I(X,Y) + «LX+MY-p), L- 1 (LX+MY-P», X arbitrary, (25)
and similarly
rot .... -1 tJ
I_ - I(X,Y) - «LY+MX-q), L (LY+MX-q», Y arbitrary. (26)
1
Should L- be known, then (25) and (26) furnish direct upper and lower
bounds on <g,u>. Otherwise they can be weakened by use of the
inequality
(27)
Transformation back to the original vector-variables, and optimization
with respect to the disposable parameter t, finally yields the cons-
traint-free result stated in (5) - (8) above [~].
It is interesting to note that in the especially simple case when
A is self-adjoint and when g = f the bounds in (5) reduce to the result
R(U) + b-111AU - fll2 ~ <f,u> ~ R(U) (28)
involving the Rayleigh functional
R(U) = -<U,AU> + 2<f,U> (29)
of which (6) is'a generalization.
If a weaker condition than (4), namely
II AU U ~ 1311 UII, ~ > 0, (30)
is assumed, then it is possible to derive the weaker bounds
J + j ~-1 C ~ <g,u> ~ J - l~-l C (31)
by direct Schwarz inequality methods [~].
Bounds on other functionals of u can also be derived in some
situations. For example if
<U,(A ± B)U> ~ E±<U,U>, (32)
then we can show that
J + 1:. IIAU - f) + (A*V-2BU)\l2 ~ <u,Bu> ~ J -.2:. II (AU-f) (A*V-2BU)U 2
4E 4E+ (33 )
where
,.,
J - <V,AU> + <V,f> + <U,BU>. (34)
3. NONLINEAR EQUATIONS
4. SOME APPLICATIONS
J
~
u(x) + k(x,y)u(y)dy fIx)
(41)
a
J
~
<V,U> = V(x)U(x)dx , (42)
we can take
g(x) = k(x' ,x) (43)
giving
<g,u> = f(x') - u(x'), (44)
so that bounds on <g,u> yield pointwise points on u(x'). This
generalizes previous work for integral equations with symmetric kernels
[~]. Integro-differential equations can be tackled in a similar way by
first converting them into equations like (41) using appropriate Green's
functions.
404
Jx 2u(y)dy =
]
1
u'(x) + t X2, 0 < X < 1,
x (45)
u(O) =0
for which no analytical solution seems available. The equivalent inte-
gral equation is
x 1
u(x) + i J z2dzf u(y)dy = -x
1 3
3
, o ~ x ~ 1 , (46)
o z
The accuracy here is much better than can be obtained using a con-
ventional 200-step finite-difference Simpson-quadrature approach.
For problems where u is an n-vector with components (u1 ,u 2 ' •••
un)' we can take g = (0,0 ••• 1 .•• 0), so that <g,u> is any desired com-
ponent of u. Thus in principle bounds can be obtained on solutions of
difference equations, or on solutions of discretized differential
problems.
Adaptation of bivariational methods to programming problems
involving inequality constraints is possible [~].
~5
REFERENCES
1. Sewell, M.J. and Noble, B., Proc. Roy. Soc. (London) A361,
293-324 (1978).
2. Robinson, P.D. and Yuen, P.K., S.I.A.M. J. Numer. Anal.
23, 1230-1240 (1976).
3. Barnsley, M.F. and Robinson, P.D., Proc. Roy. Soc. (Edinburgh)
75A, 109-118 (1976).
4. Barnsley, M.F. and Robinson, P.D., J. lnst. Maths. Applics.,
9, 485-504 (1977).
5. pack,-D.C., Cole, R.J. and Mika, J., I.M.A.J. Appl. Maths.,
32, 253-266 (1984).
6. Robinson, P.D. and Yuen, P.K., Bivariational Methods for
Hammerstein Integral Equations, submitted for publication.
7. Robinson, P.D. and Barnsley, M.F., S.I.A.M. J. Numer. Anal.
16, 135-144 (1979).
Vol. 211: P. van den Heuvel. The Stability of a Macroeconomic Vol. 237: Misspecification Analysis. Proceedings, 1983. Edited by
System with Quantity Constraints. VII. 169 pages. 1983. T. K. Dijkstra. V, 129 pages. 1984.
Vol. 212: R. Sato and T. NOno. Invariance Principles and the Vol. 238: W; Domschke, A. Drexl, Location and Layout Planning.
Structure of Technology. V. 94 pages. 1983. IV, 134 pages_ 1985.
Vol. 213: Aspiration Levels in Bargaining and Economic Decision Vol. 239: Microeconomic Models of Housing Markets. Edited by
Making. Proceedings. 1982. Edited by R. Tietz. VIII. 406 pages. 1983. K. Stahl. VII, 197 pages. 1985.
Vol. 214: M. Faber. H. Niemes und G. Stephan. Entropie. Umwelt- Vol. 240: Contributions to Operations Research. Proceedings,1984.
schutz und Rohstoffverbrauch. IX. 181 Seiten.1983. Edited by K. Neumann and D. Pallaschke. V, 190 pages. 1985.
Vol. 224: M. Kiy. Ein disaggregiertes Prognosesystem fUr die Bundes- Vol. 249: Spatial Price Equilibrium: Advances in Theory, Ccmputation
and Application. Proceedings, 1984. Edited by P. T. Harker. VII, 277
republik Deutschland. XVIII. 276 Seiten. 1984.
pages. 1985.
Vol. 225: T. R. von Ungern-Sternberg. Zur Analyse von Markten mit
Vol. 250: M. Roubens, Ph. Vincke, Preference Modelling. VIII.
unvollstandiger Nachfragerinformation. IX. 125 Seiten. 1984
94 pages. 1985.
Vol. 226: Selected Topics in Operations Research and Mathematical Vol. 251: Input-Oulput Modeling. Proceedings, 1984. Edited by
Economics. Proceedings. 1983. Edited by G. Hammer and D. A. Smyshlyaev. VI, 261 pages. 1985.
Pallaschke. IX. 478 pages. 1984.
Vol. 252: A. Birolini, On the Use of Stochastic Processes in Modeling
Vol. 227: Risk and Capital. Proceedings. 1983. Edited by G. Bam- Reliability Problems. VI, 105 pages. 1985.
berg and K. Spremann. VII. 306 pages. 1984.
Vol. 253: C. Wothagen, Economic Th~ry and Intemational Trade in
Vol. 228: NonlinearModels of Fluctuating Growth. Proceedings.1983. Natural Exhaustible Resources. VI. 172 pages. 1985.
Edited by R. M. Goodwin. M. Kruger and A. Vercelli. XVII. 277 pages.
1984. Vol. 2,54: S. Mulier, Arbitrage Pricing of Ccntingent Claims. VIII, 151
pages. 1985.
Vol. 229: Interactive Decision Analysis. Proceedings. 1983. Edited
by M. Grauer and A.P. Wierzbicki. VIII. 269 pages. 1984. Vol. 255: Nondifferentiable Optimization: Motivations and Applica-
tions. Proceedings, 1984. Edited by V. F. Demyanov and D. Paiissch-
Vol. 230: Macro-Economic Planning with Conflicting Goals. Proceed- ke. VI, 350 pages. 1985.
ings. 1982. Edited by M. Despontin. P. Nijkamp and J. Spronk. VI. Vol. 256: Convexity and Duality in Optimization. Proceedings, 1984.
297 pages. 1984. Edited by J. Ponstein. V, 142 pages. 1985.
Vol. 231: G. F. Newell. The M/M/= Service System with Ranked Ser- Vol. 257: Dynamics of Macrosystems. Proceedings, 1984. Edited by
vers in Heavy Traffic. XI. 126 pages. 1984. J.-P. Aubin, D. Saari and K. Sigmund. VI, 280 pages. 1985.
Vol. 232: l. Bauwens. Bayesian Full Information Analysis of Vol. 258: H. Funke, Eine allgemeine Theorie der Polypol- und
Simultaneous Equation Models Using Integration by Monte Carlo. Oligopolpreisbildung. III, 237 pages. 1985.
VI. 114 pages. 1984.
Vol. 259: Infinite Programming. Proceedings, 1984. Edited by
Vol. 233: G. Wagenhals. The World Copper Market XI. 190 pages. E.J. Anderson and A.B. Philpott. XIV, 244 pages. 1985.
1984.
Vol. 260: H.-J. Kruse, Degeneracy Graphs and the Neighbourhood
Vol. 234: B. C. Eaves. A Course in Triangulations for Solving Problem. VIII, 128 pages. 1986.
Equations with Deformations. III. 302 pages. 1984.
Vol. 261: Th.R.Gulledge, Jr., N.K. Womer, The Economics of Made-
Vol. 235: Stochastic Models in ReliabilityTheory. Proceedings. 1984. to-Order Production. VI, 134 pages. 1986.
Edited by S. Osaki and Y. Hatoyama. VII. 212 pages. 1984. Vol. 262: H. U. Buhl, A Neo-Classical Theory of Distribution and
Wealth. V. 148 paQltS. 1986.
Vol. 236: G. Gandolfo. P.C. Padoan. A Disequilibrium Model of
Real and Financial Accumulation in an Open Economy. VI,172 pages. Vol. 263: M. Schafer, Resource Extraction and Market Structure. XI,
1984. 154 pages. 1986.
Vol. 264: Models of Economic Dynamics. Proceedings, 1983. Ed~ed Vol. 293: D. Batten, J. Casti, B. Johansson (Eds.), Economic Evolu-
by H.F. Sonnenschein. VII, 212 pages. 1986. tion and Structural Adjustment. Proceedings, 1985. VI, 382 pages.
1987.
Vol. 265: Dynamic Games and Applications in Economics. Ed~ed by
T. 8a§ar. IX, 288 pages. 1986. Vol. 294: J. Jahn, W. Krsbs (Eds.), Recent Advances and Historical
Developmeni of Vector Optimization. VII, 405 pages. 1987.
Vol. 266: Multi-Stage Production Planning and Inventory Control.
Edited by S. Axsater, Ch. Schneeweiss and E. Silver. V, 264 pages.
1986.
Vol. 267: R. Bemelmans, The Capacity Aspect of Inventories. IX, 165
pages. 1986.
Vol. 268: V. Firchau, Information Evaluation in Cap~al Markets. VII, 103
pages. 1986.
G.Gandolfo
International Economics
1986. 133 figures. XXVI, 804 pages. ISBN 3-540-16707-2
Contents: The Pure Theory ofInternational Trade: Introduction. The clas-
sical (Ricardo-Torrens) theory of comparative costs. The neoclassical theory
of international trade. The Heckscher-Ohlin model. Tariffs, protection,
economic integration. International trade and economic growth. Some refi-
nements ofthe orthodox theory. The "new" theories of international trade.
Neo-Ricardian theories of international trade. - Bibliography.
International Monetary Theory: The foreign exchange market. Balance of
payments and national accounts. The role of the exchange rate in the
adjustment process in a partial equilibrium framework. The role of income
changes in the adjustment process. The absorption approach and interac-
tions between exchange-rate and income in the adjustment process. Money
and other assets in the adjustment process under fixed exchange rates.
Money and other assets in the adjustment process under flexible exchange
rates. International capital movements and other problems. FIxed versus
flexible exchange rates. International liquidity and international fmancial
markets. The problem of the integration between the pure theory of inter-
national trade and international monetary economics. - Bibliography. -
Name Index. - Subject Index.
W. Whitmore, Jr.