Beruflich Dokumente
Kultur Dokumente
Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at http://www.jstor.org/page/
info/about/policies/terms.jsp
JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide range of content
in a trusted digital archive. We use information technology and tools to increase productivity and facilitate new forms of scholarship.
For more information about JSTOR, please contact support@jstor.org.
INFORMS is collaborating with JSTOR to digitize, preserve and extend access to Interfaces.
http://www.jstor.org
This content downloaded from 125.20.9.248 on Fri, 21 Aug 2015 06:02:37 UTC
All use subject to JSTOR Terms and Conditions
INTERFACES
Vol.
11, No.
October
5,
Sciences
Institute of Management
0092-2102/81/1105/0O84S01.25
1981, The
Copyright ?
1981
HEURISTIC "OPTIMIZATION":
WHY, WHEN, AND HOW TO USE IT
Stelios H.
Florida
International
University,
Zanakis
of Business
College
Miami,
Management,
Florida
Administration,
of
Department
33199
and
James R. Evans
University
of Cincinnati,
College
of Business
Administration,
Ohio 45221
Department
of Quantitative
Cincinnati,
Analysis,
Abstract.
solutions
mentable
and rigorously
for heuristics.
tematically
and proper
for heuristic
paper presents a framework
in attitudes
this change
towards heuristics,
This
respect
examining
"optimization"
their desirable
by sys
features,
usage.
"heuristic"
derives from the Greek "heuriskein,"
meaning "to dis
In that sense, a heuristic aims at studying the methods and rules of discovery
[Polya, 1947, pp. 112-113] or assisting in problem solving, which is a process of
but not immediately attainable aim
trying to attain a preconceived
systematically
seen heuristics as procedures to
have
Researchers
[Polya, 1962, 1963]. Operations
to obtain
activities
reduce search in problem-solving
[Tonge, 1961] or a means
a
To
time
within
limited
solutions
[Lin, 1975].
practitioners,
computing
acceptable
The word
cover."
heuristics
are
simple
often
procedures,
guided
by
solutions
common
sense,
that
are meant
to
easily
and
to difficult
problems,
optimal
provide
quickly.
... often outlined and
This branch of study is "not very clearly circumscribed,
seldom presented in detail [Polya, 1947]," quick and dirty [Woolsey,
1975a], not
the
mathematical
theorems,
convergence,
optimality,
etiquette (proofs,
abiding by
in analytic purity in the high
that "are conceived
etc.), in contrast to algorithms
in the dark
citadels of academic research, heuristics are midwifed
by expediency
corners of the practitioner's
lair, . . . and are accorded lower status [Glover, 1977]."
For many years heuristics have encountered the hostility, scorn, and condemnation of
tools of the trade, as
them as the poor man's
high academic priests, "treating
with
should
be
that
1977]."
[Eilon,
deep suspicion
regarded
expediencies
as
to
evidenced
this
started
however,
by the growing
change,
picture
Recently,
number of related articles in scholarly reputable journals. The Operations Research
Society of America awarded the 1977 Lancaster Prize to two papers on heuristic
1977; and Karp,
1977]. The
[Cornuejols, Fisher, and Nemhauser,
optimization
Sciences
1980 Innovative Education Award went to
Institute for Decision
American
a paper on heuristic problem
Why then this change?
solving
[Brightman,
84
INTERFACES
OPTIMIZATION;HEURISTICS
October
This content downloaded from 125.20.9.248 on Fri, 21 Aug 2015 06:02:37 UTC
All use subject to JSTOR Terms and Conditions
1981
for probabilistic
that elevated them from
analysis of heuristic methods
to a hot topic of sophisticated analysis now respected in the OR/MS
poor relatives
academic
community.
As more
instances where
is desirable
and advanta
geous:
(1) Inexact or limited data used to estimate model parameters may inherently
contain errors much larger than the "suboptimality"
of a good heuristic. We have
seen monumental
efforts in academia and industry to develop and/or use ILP codes,
for example, to solve large-scale capital budgeting problems with intangible benefit
estimates or risk analysis. In one case [Zanakis, 1977], an existing 0-1 heuristic was
run to produce in 1/10 of CPU time
easily modified within a day and successfully
a
99.7% of the "optimal"
solution to
308x372
capital budgeting problem with
elsewhere with a special
intangible benefits, which had been solved "optimally"
a
six-month developmental
effort.
purpose algorithm after
(2) A simplified model is used, which is already an inaccurate representation of
solution only academic. In both cases 1
the real problem, thus making the "optimal"
exact
and 2, a fast near-optimal
solution makes more sense than a time-consuming
solution to an inexact problem.
INTERFACES
October 1981
85
This content downloaded from 125.20.9.248 on Fri, 21 Aug 2015 06:02:37 UTC
All use subject to JSTOR Terms and Conditions
but it is computationally
Unattractive due to
(4) An exact method is available,
time and/or storage requirements. Large real-world complex problems may
excessive
prevent an optimizer from finding an optimal or even a feasible solution within a
on the other hand, can produce at least feasible so
reasonable effort. Heuristics,
lutions with minimal
manpower
be appropriate. However,
for many real-world
above often suggest a heuristic method.
applications,
the
should
(3) Speed, i.e., reasonable computing times that do not grow polynomially
even exponentially
as problem size increases.
*Ed. Note:
I observed
this
10 years
ago
inManagement
86
Science,
1971,
Vol.
17, No.
or
7, p. 500.
INTERFACESOctober 1981
This content downloaded from 125.20.9.248 on Fri, 21 Aug 2015 06:02:37 UTC
All use subject to JSTOR Terms and Conditions
(4) Accuracy,
i.e., small average and mean square errors; these will increase
chances
the
of a good (poor) solution. The degree of acceptable accuracy
(decrease)
should be dictated by the problem and the user.
should obtain good solutions,
in reasonable
i.e., the method
(5) Robustness,
a
for
wide
times,
variety of problems and not be too sensitive to changes in paramet
ers. This is not true of many branch-and-bound methods where a small change in a
parameter may double or triple running times.
?
a formidable
feasible
(6) Accept multiple starting points, not necessarily
task in large and severely constrained problems.
(7) Produce multiple solutions
(ideally in a single run) by properly selecting
input parameters or perturbing the final solution. This allows the user to select the
result that is most accurate or "satisficing"
(in case of unquantifiable
intangibles).
A new perturbation heuristic, added to several existing 0-1 LP heuristics,
captured
about 65% of possible improvement to optimality,
thus increasing average accuracy
to 99.8%
ability with
1981].
take advantage of search "learning"
and avoid
some ratio measure of objective function im
effort.
true optimum from a series of iterative solutions
An ideal heuristic is one that has all of the above properties. It should be noted
that problem-dependent
that take advantage of the special structure of a
heuristics,
Evans
and
Cullen,
1977], are more efficient than general mathemati
problem [e.g.,
cal programming heuristics, but their use is limited only to a specific class of prob
lems. A classical example
is scheduling heuristics vs integer linear programming
heuristics applied to scheduling problems.
INTERFACES
October 1981
87
This content downloaded from 125.20.9.248 on Fri, 21 Aug 2015 06:02:37 UTC
All use subject to JSTOR Terms and Conditions
1977; Fisher,
1980].
analysis, where one assumes a probability distribution of
statistical properties of a heuristic
[Karp, 1976, 1977].
functions can provide tractable results, which is a limitation of
(c) Probabilistic
data to establish
problem
Only simple density
this approach.
(d) Statistical
true (unknown) optimal
widely,
large. Because of its simplicity and the fact that it has not been publicized
this estimation procedure is summarized in an appendix.
CONCLUSIONS
The need for and usage of good heuristics in both academia and businesses will
continue increasing fast. When confronted with real world problems, a researcher in
of seeing his product,
academia experiences at least once the painful disappointment
a theoretically sound and mathematically
procedure, not used by its
"respectable"
ultimate user. This has encouraged researchers to develop new improved heuristics
thus spreading further their usage in
and rigorously
evaluate their performance,
practice, where heuristics have been advocated for a long time. More work is needed
case [Zanakis,
in the area of heuristic optimization, particularly in the multiobjective
to
world
meet
real
the
fail
the
codes
available
where
problem needs.
1981],
seriously
88
INTERFACESOctober 1981
This content downloaded from 125.20.9.248 on Fri, 21 Aug 2015 06:02:37 UTC
All use subject to JSTOR Terms and Conditions
References
Problem
H.,
1980, "Teaching
Quantitative
by General
Techniques
Solving Heuristics,"
AIDS 12th National
pp. 116-118.
Proceedings
Conference,
of Test Problems
Com
for Linear Programming
Chames,
A., et al.,
1974, "On Generation
Codes,"
munications
Vol.
17, No.
10, pp. 583-586.
oftheACM,
of a New Heuristic
for the Traveling-Salesman
Christofides,
N.,
1976, "Worst-Case
Problem,"
Analysis
Brightman,
MS
Cooke,
University,
February.
for Bounds
of Random
Biometrika
Variables,"
Vol.
66,
pp.
to
G. L.,
of Bank Accounts
G., Fisher, M. L., and Nemhauser,
1977, 1979, "Location
Float: An Analytic
Science
Optimize
Study of Exact and Approximate
Algorithms,"
Management
Vol. 23, pp. 789-810
Vol. 25, pp. 808-809).
Science
(Errata inManagement
D. G.,
Solution Values
for Estimating
for Large Combinato
1977, "Procedures
Dannenbring,
Optimal
rial Problems,"
Science Vol. 23, pp. 1273-1283.
Management
Cornuejols,
Eilon,
Evans,
Optimization,"
a Framework
Proceedings
Problem:
Fisher, M. L.,
1-17.
Management
1980, "Worst-Case
Analysis
of Heuristic
Algorithms,"
Some
Annual
AIIE
and an
Properties
Science
Vol.
26, pp.
J. M.
and Meyer,
R. R.,
Sufficient
Conditions
for Integer Program
1976, "New
Optimality
and Their Application,"
Science Technical
of Wiscon
Computer
Report No. 278, University
October.
sin, Madison,
for Combinational
Problems: An
Garey, M. R. and Johnson, D. S., 1976, "Approximation
Algorithms
Fleisher,
ming
Annotated
in Algorithms
and Complexity,
Bibliography,"
F. and Sommer, D. C,
of Rounding
1975, "Pitfalls
Vol. 6, pp. 211-220.
Sciences
lems," Decision
for Integer Programming
F., 1977, "Heuristics
Glover,
Using
Glover,
Vol.
8, pp.
Constraints,"
Decision
Prob
Sciences
156-166.
Vol.
Golden,
Golden,
7, pp. 209-225.
for Large Combinatorial
Optimum
A Venture
Problems:
into a New Dimension,"
Journal
1980, "Solving
Large-Scale
Research
Society Vol. 31, pp. 217-225.
Vol.
of Combinational
Networks
1975a, "On the Computational
Problems,"
Karp, R.,
Complexity
pp. 45-68.
Ignizio,
J. P.,
of
Operational
5,
Karp,
R. M.,
of Partitioning
for the Traveling-Salesman
1977, "Probabilistic
Analysis
Algorithms
Problem
in the Plane," Mathematics
Research
Vol. 2, pp. 209-224.
of Operations
View of Management
December
Levitt, T., 1978, "A Heretical
Fortune,
'Science',"
18, pp. 50-52.
Karp,
Lin,
Lin,
Programming
R. L.,
1979,
as an Aid
Integer Programming
Algorithms,"
V. A. and Whybark,
D. C,
Mabert,
Vol. 8, pp. 167-177.
to Solve It, Princeton
1947, How
Polya, G.,
Polya,
G.,
1962, Mathematical
to Network
"Controlled
Discovery,
Press, Princeton.
University
Vol.
New York.
I,Wiley,
New York.
Vol.
II, Wiley,
1965, Mathematical
Polya, G.,
Discovery,
Test Problem Generator,"
Rardin, R. L. and Lin, B. W.,
1980, "The RIP Random
Integer Programming
School of Industrial & Systems Engineering,
Institute of Technology,
June.
Georgia
J. B. and Suzuki,
of Nonlinear
Test Problems,"
Com
Rosen,
S., 1965, "Construction
Programming
munications
oftheACM
Vol.
8, No.
2, p.
113.
INTERFACESOctober 1981 89
This content downloaded from 125.20.9.248 on Fri, 21 Aug 2015 06:02:37 UTC
All use subject to JSTOR Terms and Conditions
Silver,
to Heuristic
Methods,"
Problem,"
unpublished
in Management
F. M.,
Science,"
1961, "The Use of Heuristic
Management
Tonge,
Programming
Vol. 7, pp. 231-237.
Science
A Quick
R. E. D. and Swanson, H. S., 1975a, Operations
Research for Immediate Application:
Woolsey,
and Dirty Manual,
Harper & Row, New York.
to Do Integer Programming
R. E. D.,
1975b, "How
H. M. Salkin, ed., Addison-Wesley,
13.
Chap.
S. H.,
0-1 Linear Programming:
Zanakis,
1977, "Heuristic
Vol. 24, pp. 91-104.
Science
Methods,"
Management
Woolsey,
in Integer
Program
ming,
An
Experimental
Comparison
Zanakis,
Zanakis,
of Three
Weibull
Weibull
to Implementing
1980, "From Teaching
D., and Silver, E.,
of Translation,"
Vol.
10, No. 6, pp. 103-110.
Interfaces
to a
Method
with an Application
Large-Scale
Integer Goal Programming
in Organizations:
Multicriteria
Decision
Location/Allocation
Problem,"
Multiple
Making,
S. H.,
Austin,
L., Nowading,
Problems
Inventory Management:
S. H.,
Zanakis,
1981, "A
Facility
Agents
Zanakis,
with Multiple
Criteria,
and Gillenwater,
S. H.,
Working
Paper, West
Virginia
J. Morse,
J.,
1981,
College
ed. Springer-Verlag.
Improved Heuristics
of Graduate
for
Integer
and 0-1
Programming,
Studies.
90 INTERFACESOctober 1981
This content downloaded from 125.20.9.248 on Fri, 21 Aug 2015 06:02:37 UTC
All use subject to JSTOR Terms and Conditions
APPENDIX
Let Z, ^ Z2 ^ .... ^ Zn be an ordered random sample of iterative solutions
(For a maximization
(objective function values) to a large minimization
problem.
..
i=l, 2,
.,n.) A point estimate of the
problem,
simply replace Z? by -Zn+1_f,
unknown true optimum can be obtained by one of the following simple estimators:
Z = 2Z,-Z2 (1)
used by Dannenbring
[1977], or
[1979a]
and Zanakis
Z* =
and Gillenwater
2Zj -(e-1)
[1981],
(2)
or
by Cooke [1979]. Note that these estimators do not require any restrictive
assumption and apply to any distribution with a location or threshold parameter.
From a storage requirement viewpoint
their order of preference
is Z, Z, Z* but
proposed
This
is expected
estimator.
100(l-e~w)%
because
confidence
by Golden
proposed
< a < Zj
6, (4)
Zj
where
of the different
interval on the
and Alt
[1979]:
=Z[o.63nl
+ i ""*
(5)
=
Here, [
] denotes rounding down and af Z (Z or Z* will be more accurate).
This requires the (empirically
that the random sample came
justified) assumption
from a three-parameter Weibull
cumulative distribution
F(Z)
Z>a,
(6)
parameters a (location), b (scale), and c (shape). Golden and Alt [1979] rec
ommend refining this confidence
interval by using maximum
likelihood estimates
(MLE). The added accuracy, however, will not justify the incremental effort be
cause: (1) MLE's do not have a closed form, thus requiring an iterative procedure on
a computer; (2) it will require the additional estimation of the shape parameter c,
which is the most difficult of the three parameters to estimate; (3) estimator b is a
with
INTERFACESOctober 1981
91
This content downloaded from 125.20.9.248 on Fri, 21 Aug 2015 06:02:37 UTC
All use subject to JSTOR Terms and Conditions