The Heuristic Problem-Solving Approach

Author(s): L. R. Foulds
Source: The Journal of the Operational Research Society , Oct., 1983, Vol. 34, No. 10
(Oct., 1983), pp. 927-934
Published by: Palgrave Macmillan Journals on behalf of the Operational Research
Society

Stable URL: http://www.jstor.com/stable/2580891

REFERENCES
Linked references are available on JSTOR for this article:
http://www.jstor.com/stable/2580891?seq=1&cid=pdf-
reference#references_tab_contents
You may need to log in to JSTOR to access the linked references.

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 [email protected].

Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at
https://about.jstor.org/terms

Operational Research Society and Palgrave Macmillan Journals are collaborating with JSTOR to
digitize, preserve and extend access to The Journal of the Operational Research Society

This content downloaded from

80.112.186.47 on Tue, 07 Jan 2025 15:21:43 UTC
All use subject to https://about.jstor.org/terms
J. Opl Res. Soc. Vol. 34, No. 10, pp. 927-934, 1983 0160-5682/83 $3.00 + 0.00
Printed in Great Britain. All rights reserved Copyright ? 1983 Operational Research Society Ltd

The Heuristic Problem-Solving Approach

L. R. FOULDS
University of Florida

For a variety of reasons, the finding of an optimal solution is impractical for m

A common way of overcoming this unhappy state of affairs is the develop
(approximate) methods. The purpose of this paper is to discuss some of the issues
an approach?that is, the use of a method which, on the basis of experience of jud
to yield good solutions but which cannot guarantee optimality. The use of such m
by the emergence of the theory of AfP-completeness, i.e. the study of the compl
which is briefly introduced. A number of heuristic methods are presented in ord
of the ideas discussed. Heuristic procedures are classified according to design. Som
of both how to design effective heuristics and how to use heuristics in the real w

INTRODUCTION

An algorithm for a problem is a procedure which is guaranteed to converge to a

satisfactory solution to the problem. Some industrial problems are so complex that the
standard models of operational research are inappropriate. This sometimes occurs because:
(1) the number of pieces of information needed to describe the problem is enormous; (2)
the problem has features which are difficult to quantify or involve a conflict of objectives;
(3) it may be difficult to collect accurate data.
An O.R. consultant will provide no comfort whatsoever when telling a businessman that
there does not exist an efficient method for finding an optimal solution to his problem.
Despite this, the scheduling, routing, allocation, or whatever still has to be done.
If this is the state of affairs, he can consider a number of strategies: (1) find an efficient
algorithm for the problem; (2) show that only special cases of the problem are of interest
and find an efficient algorithm for them; (3) relax some of the constraints of the problem
and develop an algorithm for the relaxed (easier) problem; (4) construct an algorithm that
runs quickly on most of the problems likely to be encountered; (5) give up the quest for
optimality and provide a method which runs quickly and produces useful but not
necessarily optimal solutions.
Many practical O.R. problems are what is termed NP?complete?this concept is
discussed later. It is extremely unlikely that aim (1) can then be achieved. Aims (2) and
(3) are usually more appropriate. However, Churchman1 has quite rightly warned of the
dangers of discarding the problem at hand for a surrogate which is not equivalent but
which can be solved by known methods. This practice has raised the criticism that
sometimes O.R. practitioners "bend a clienfs problem" so that it can be solved by a
standard method, thereby producing solutions of little interest to the client. Aim (4) is often
attempted by O.R. consultants and often achieves customer satisfaction. The simplex
algorithm of linear programming is a good illustration. Unfortunately, sometimes such an
algorithm cannot be found. The purpose of this paper is to discuss what can be done when
we are reduced to aim (5).
This paper discusses the strategy of developing and using approximate methods. In order
to motivate this use, we begin in the next section with an introduction to the theory of
complexity.

THE COMPLEXITY OF PROBLEMS

Combinatorial mathematics is often described as the selection and arran

objects. In the past, most of the work in this area has been on the e
of possible selection or arrangements. Recently there has been inc
finding the best possible selection or arrangement. Each possible co
927

This content downloaded from

80.112.186.47 on Tue, 07 Jan 2025 15:21:43 UTC
All use subject to https://about.jstor.org/terms
 Journal of the Operational Research Society Vol. 34, No. 10

some kind of score, and an optimization process is carried out. This relatively new area
of mathematics is termed combinatorial optimization. It is evident that many problems
faced by O.R. practitioners fail within this compass, including: sequencing, scheduling,
routing, layout and design problems. There is usually a finite number of feasible solutions
to each instance of these problems. Therefore it seems intuitive that the fundamental
theorem of combinatorial optimization could be employed: examine every feasible solution
and choose the best. Unfortunately, there are usually far too many solutions for this
approach to be practical. As an example, consider the travelling salesman problem. Here
a salesman must find a closed tour, visiting each of a given number n, of cities exactly
once, travelling minimum distance. If the distance function is asymmetric, there are n\
possible tours. If each tour could be evaluated in a billionth ofa second, it would still take
over 1600 years to examine them for /7 = 21. (Other solution procedures for this problem
will be discussed later.)
We have to be able to do better than complete enumeration. In order to be able to
compare alogrithms, we need some way of judging the effectiveness of an algorithm.
Edmonds2 provided the following convention, which has been generally accepted.

An algorithm is considered to be effective if it can guarantee to solve any

instance of the problem for which it was designed by performing a number of
elementary computational steps where this number can be expressed as a
polynomial function of the size of the problem.

Computer scientists often use the term 'polynomially-bound time', as it is assumed that
computational time is linearly proportional to the number of elementary computational
steps. The size of a specific instance of a problem is defined to be the number of symbols
required to describe it.
Given a combinatorial optimization problem, it is natural to ask whether a
polynomially-bounded time algorithm has been found for it. The set of problems for which
the answer is "yes" is denoted by P (for polynomial). There are a few problems for which
it has been shown that membership in P is impossible. There exists a third set of problems
whose status is unknown?it is not known whether they are in P or not. There is at least
one algorithm known for each of these problems, but no known effective algorithms have
as yet been reported. This set is denoted by NP (for non-deterministic polynomial).
The following caveat should be made here. There are problems in P which, for practical
size (say n), are intractable and problems not in P which are tractable. For instance, a
problem in P of size n = 1000 is intractable if its solution time is ofthe order of n9 (= IO27).
But a problem not in P of the same size with solution time of the order e?001n is tractable.
Some tensor inversion problems have the former character, and some mathematical
programming problems the latter. The relative efficiency of the ellipsoidal and simplex
algorithms for linear programming further brings out this point.
NP contains a subset of problems termed NP-complete. Each problem in the set has the
following properties: (1) it belongs to NP; (2) if an effective algorithm exists for it, then
an effective algorithm exists for every problem in NP, i.e. P = NP. This means that
NP-complete problems are the hardest in NP. Many of the problems in NP have defied
the attempts to find effective algorithms of some of the best mathematicians over the past
30 years. There is also more objective circumstantial evidence that P ^ NP. Thus it seems
unlikely that an effective algorithm exists for any of the NP -complete problems. Hence
the fact that a problem is NP -complete is considered by some to be justification for
approximate procedures to be applied to it. Many commonly encountered O.R. problems
are TVP-complete. So we now turn our attention to the aim of developing approximate
procedures with good performance guarantees which are also effective in the sense defined
earlier.

THE HEURISTIC APPROACH

As early as 300 A.D. Pappas, writing on Euclid, suggested the appr

methods?which are easy to use but which do not guarantee optimalit

928

This content downloaded from

80.112.186.47 on Tue, 07 Jan 2025 15:21:43 UTC
All use subject to https://about.jstor.org/terms
 L. R. Foulds?The Heuristic Problem-Solving Approach

of Descartes and Leibniz, the subject became known as heuristics. Allied with logic,
philosophy and psychology, it has the aim of investigating the methods of invention and
discovery. The name itself was derived from the Greek heuriskein?to discover. Today the
term heuristic means a method which, on the basis of experience or judgement, seems likely
to yield a good solution to a problem but which cannot be guaranteed to produce an optimum.
Most people use heuristics every day:

*In order to guard against a chill, take a warm coat to work when the weather forecast
is bad.
*In order to ensure a supply of small change on hand, go to the bank and change money
regularly.
*In order to be certain of being able to buy theatre tickets, book early.
Heuristics are common in business too:

*Schedule the shortest jobs first.

*Never lend money so that the level of monthly repayment is more than 25% of inc
*Perform tasks in order of importance.

There has been much work published on heuristic problem solving, and a few of
more important contributions will now be briefly discussed. Further references to o
more specialized work appear later. Pioneering work was done by Newell and Simon
the computer simulation of human thinking. They devised a computer program called Th
General Problem Solver (GPS) to stimulate human problem-solving behaviour. It app
operators to objects to attain goals. Its processes are geared towards three types of g
(transformation, operator application and difference reduction) and methods for attainin
them. The authors furnish an example in which the goal is to transform one symbolic lo
expression into another. The operations are certain allowed manipulations, and the ob
are the expressions themselves. Tests for differences that must be recognized between pa
of expressions and these entities are fed into the computer. The steps used by the compu
and by a human subject in solving the problem are summarized in a table. It show
number of similarities and differences. They also presented this work at a conferenc
Germany.4 Newell and Simon5 went on to develop the idea of problem-solving mach
They discuss computer programs that not only play games but which also loc
warehouses and prove theorems in geometry. Newell and Simon6 also proposed
detailed and powerful variable iteration technique for solving test problems by computer

A CLASSIFICATION OF HEURISTICS

There are four basic strategies for heuristic procedures. Many m

combination of more than one of these strategies.

1. The construction strategy

The input for methods based on this strategy is nothing more than t
a specific instance of the problem. A solution is built up one comp
A construction strategy begins by examining this data and attem
element which is likely to be a valuable part of a very good final soluti
additional elements of a solution are added. The better constructi
some kind of 'look-ahead' mechanism. That is, additions are made,
appear a good idea at the time, but because they are likely to be of
complete solution. Once the final solution has been built, up, it m
improvements can be easily effected. Thus the strategy is often ap
the construction method. This strategy is worthwhile when it is r
generate feasible solutions to the problem.

2. The improvement strategy

The input for methods based on this strategy is a solution to the pro

929

This content downloaded from

80.112.186.47 on Tue, 07 Jan 2025 15:21:43 UTC
All use subject to https://about.jstor.org/terms
 Journal of the Operational Research Society Vol. 34, No. 10

is then progressively improved by the application of a series of modifications. In some

instances it may be impossible to make much progress in this way, and yet the final product
may still be far from optimality. Because of this, some improvement strategies are
far-sighted in the following sense. Some iterations of the 'improvement' process may
actually be allowed to bring about a worsening in solution value if it can be seen that this
will create a situation where worthwhile gains can be made. This strategy is worthwhile
when it is relatively easy to generate starting solutions. A variety of solutions can be used
as input and the best final result chosen. Sometimes the strategy is used to convert an
infeasible solution into a feasible one.

3. The component analysis strategy

Some problems are so large or so complicated that the only practical approach is
break them up into manageable portions. Sometimes these portions are then dealt wi
independently by heuristics or even algorithms. The solutions for the portions are th
married to form some master plan. Of course, it may be extremely difficult to piece togethe
the solutions to the different components into an acceptable plan. If the components
be ordered in some sort of logical sequence, it usually makes sense to examine them
the same order. This ordering is often based on some time-scale which is an integral p
ofthe problem. For instance, in investment problems a natural sequencing may be to defin
a component as one year's activities. The output of the analysis of one component ma
be a valuable input for the analysis of later components.

4. The learning strategy

Methods based on this strategy often use a tree-search diagram to chart their progress.
That is, the different options which appear at various stages are represented by different
branches ofa tree. The sequences of choices actually made can be traced by a path through
the tree. The choice of which branch to take is guided by learning from the outcome
earlier decisions. The early termination of a branch and bound search is an example o
this strategy. The reader is referred to the works of Muller-Merbach.7,8 The form
classifies heuristic techniques morphologically, and the latter contains a useful section
heuristic methods.

EXAMPLES OF HEURISTICS

To make the ideas of the last section more lucid we present a number o
illustrating the different strategies. To make this meaningful we must
problem for which a heuristic discussed is designed. In order to avoid lengthy
of a variety of problems, we introduce but one problem. We use as our v
section a problem referred to in the second section; the travelling sales
(T.S.P.). This problem can be formulated as;
Minimize

? ?d^ (i)
subject to
n

Zx//=1> 7 = U2, . ..,?. (2)

?*f/=l, i = l,2,...,/t. (3)

ai~ aj + nxij<n ? \, i ? \,2,. . . ,n,
j = 1,2,.. .,n,

i *h (4)
930

This content downloaded from

80.112.186.47 on Tue, 07 Jan 2025 15:21:43 UTC
All use subject to https://about.jstor.org/terms
 L. R. Foulds?The Heuristic Problem-Solving Approach

for some real numbers q{.

Xij = 0 or 1, i = 1,2,. .. ,n,

y'=l,2,...,/i, (5)
where

Xij = 1 if the salesman travels directl

= 0 otherwise;

dy = the distance from city / to city

n = the number of cities.

The problem (1), (2),. .., (5) is TVP-complete, and thus heuristic solution procedures are
in order. There is no suggestion that the heuristics about to be presented are superior t
all others for the T.S.P. They are introduced, not as the most efficient T.S.P. procedures
but in order to illustrate the ideas of the last section.

1. The construction strategy

Christofides9 has developed a construction strategy heuristic for the symmetric T.S.P
i.e. where d^= dyi for all ij pairs. The method begins by first building up what is calle
a minimal spanning tree. In terms ofthe T.S.P. a spanning tree is a set of legs (direct paths
between cities) with the property that it is possible to find a path of legs between ever
pair of cities. A minimal spanning tree is a spanning tree of minimum total distanc
(defined as the sum of the distances of the legs). A minimal spanning tree will not contain
any circuits of legs as it is assumed that all distances are non-negative and thus any l
of a circuit could be dropped without penalty. The minimal spanning tree is constructe
one leg at a time, as follows. The legs are assembled in non-increasing order of distanc
Each leg is examined in turn and accepted as part of the minimal tree unless it creates
circuit with those legs already accepted. The minimal spanning tree is complete whe
(n ? 1) legs have been accepted. Then the cities which are connected to an odd numb
of cities are identified. It is straightforward to show that there must be an even numb
of these cities. These cities are then matched in pairs according to the following criterion:
the sum of the distances of the legs corresponding to the matched pairs is a minimum
These legs are then added to the minimum spanning tree. This will create circuits and may
even result in duplicated legs. The travelling salesman's tour traverses each unique leg i
this final set exactly once and each duplicated leg exactly twice. The reader is referred
Christofides' paper9 for details of how to find the actual tour through these legs and f
improvement strategies to be applied to it. The paper also proves that the tour constructed
will have total distance no more than half as long again as the optimum.

2. The improvement strategy

Improvement strategies for the T.S.P. belong to what is called the tour perturbation
class. They are applied to a given feasible solution. Such a solution is produced by a
construction strategy heuristic. One simple strategy of this type is to identify every differe
pair of legs in the tour. For every such pair P, a search is made for the pair of legs, n
currently in the tour, which, if exchanged with P, bring about the greatest decrease in tou
distance. Of all the possible exchanges, the most beneficial exchange is actually mad
generating an improved solution. The process is then repeated. When no further beneficial
exchange can be found, the process is terminated. This strategy can be generalized to t
exchange of r legs at a time. Indeed a tour is termed r-optimal if no improvement can
found by removing any r legs and replacing them with any other r legs.
Lin and Kerningham10 have proposed an improvement strategy heuristic for the
symmetric T.S.P. It uses the interchange philosophy just discussed and employs a rationale
very similar to the r-optimal concept. The procedure starts with a solution that is generate
pseudo-randomly and continues improving the solution until no further improvement
931

This content downloaded from

80.112.186.47 on Tue, 07 Jan 2025 15:21:43 UTC
All use subject to https://about.jstor.org/terms
 Journal of the Operational Research Society Vol. 34, No. 10

available. It does this iteratively using the r-optimal approach, but r is not fixed. The key
to the success of the procedure is the following theorem.

If a sequence of numbers has a positive sum, there is a cyclic permutation of

these numbers such that every partial sum is positive.

This allows the procedure to consider only sequences of interchanges whose partial sum
of gains is always positive (improving). The authors present a detailed statement of their
heuristic and a numerical example. The computational experience they report is very
impressive for small problems (with no more than 42 cities). Here the optimal solution can
be obtained by branch and bound enumeration in reasonable computational time as a
check on the quality of heuristic solutions. It appears that the probability of obtaining an
optimal solution by their heuristic is close to one. Computational time is 0(n2).

3. The component analysis strategy

As has been mentioned, Christofides' method9 gives the assurance that solution distance
will never be over 50% more than the optimal distance. A different, and often more
realistic, guarantee can be considered: that the solution will not often deviate far from
optimality. This guarantee is useful for scenarios where travelling salesman tours or vehicle
schedules are being planned on actual planar maps. It is usually the case in such instances
that the set of city locations satisfy tests of random location, i.e. the cities give the
appearance of being randomly located. In 1958 Beardwood et al.16 proved a theorem about
the statistical properties of optimal tours through a randomly distributed set of cities.
Karp11 used this result to produce a component analysis strategy for the T.S.P. It begins
by breaking up the map containing all the cities into a number of components. Each
component contains relatively few cities?so few that an exhaustive search can be made
(in reasonable computing time) for the optimal tour of the cities in it. Each of these
subtours then has one, two, three, or four legs removed from it in an effort to combine
the subtours into one master tour. The method for achieving this is a variant ofthe method
which produced the minimal spanning tree in Christofides' method.9 As the number of
cities increases, the probability of Karp's method producing a solution no more than a
given percentage above optimality increases.

4. The learning strategy

Krolak et al.12 have proposed a learning strategy for the T.S.P. It is a person-computer
approach that is iterative in nature. Basically, the computer generates a solution, the
person interacts by making improvements, the computer generates another solution, and
so on.

We turn now to the problem of how to design an effectiv

THE DESIGN OF HEURISTICS

When confronted with the problem of designing a heuristic for

often very useful to set aspiration levels before plunging into the a
is, what type of performance guarantees must the heuristic possess?
apply to the actual problems under study. This usually means la
just small test problems. Desirable characteristics include:

1. execution in reasonable computational time;

2. solutions which are close to optimality;
3. only a small probability of any one solution being far below
4. simplicity of both design and computational requirements.

Characteristics 2 and 3 suggest a worst-case performance boun

and Johnson13 have provided an annotated bibliography up to 1
existing heuristics.

932

This content downloaded from

80.112.186.47 on Tue, 07 Jan 2025 15:21:43 UTC
All use subject to https://about.jstor.org/terms
 L. R. Foulds?The Heuristic Problem-Solving Approach

Returning to the point of guaranteeing good performance on realistic and not just small
test problems, we turn to the question of validation. Sometimes this can be achieved by
constructing large-scale problems with known optimal solutions. That is, this solution is
defined and then the rest of the data for the problem is "built around it" ensuring
optimality. Although such problems are inherently artificial, careful construction can
provide useful test problems. Despite the previous remark, it is often still quite useful to
compare heuristic performance with that of an algorithm on a wide variety of small and
medium-sized test problems if possible.
When validation is difficult, but it is relatively easy to generate feasible solutions, the
following design approach of Baum and Carlson14 may be valuable. Called the Better Than
Most approach, B.T.M., it randomly generates solutions, selecting those that are feasible.
The best solutions among these are selected according to the optimality criteria of the
analyst, where some of these criteria may be difficult to quantify. Baum and Carlson
discuss the obvious question of how large a sample must be taken in order to guarantee
that the best solution identified will be, with a specified probability, within a given
percentage of optimality.
This approach leads to the question of when to terminate a heuristic that generates one
or more feasible solutions early in its execution and then spends most of its time searching
for improvements. McRoberts15 outlines and provides examples of the application of
extreme value distributions as mechanisms for estimating optimum limits. Given such
estimates, he further illustrates the development of rules applicable to the decision of
whether to continue or terminate.
When optimality criteria are difficult to quantify, it may be reasonable to contemplate
the design of a heuristic which involves interaction with an analyst during its executio
Using his intuition and experience, he may be able to guide the heuristic's search proce
far more efficiently than a predetermined set of rules. There is the further strategy o
relaxing certain difficult constraints and having the analyst choose from a variety
solutions proposed by the interactive problem. The analyst then, on the basis of his choice,
may be able to set slightly more constrained problems for the interactive heuristic. These
approaches have been successfully used to solve large-scale vehicle scheduling problems
heuristically.

THE USE OF HEURISTICS

Good heuristics have obvious advantages over the more standar

ational research:

1. They can often be implemented by a practitioner who lacks advanced training, by virtue
of their simplicity.
2. They require relatively modest computational time.
3. They are flexible, often allowing use as planning tools where many scenarios and set
of non-quantifiable constraints must be considered.

One disadvantage is the fear that solutions far below optimality will be produced. It
often not clear how significant the deficiency is. There can be further problems with
implementation. Most ofthe literature on the topic deals with successfully conceptualizing
heuristic solutions for idealized models. (It is a shame that more negative results are no
reported as these often contribute to the learning process of others as well.) Many of t
heuristics posed do not pass the test of being subjected to real-world problems becaus
of model deficiency. There are often complications not present in the model for which the
heuristic is designed. Thus successful heuristic implementation often lags behind heuristic
development by a considerable degree.
In order to safeguard against the error of introducing a heuristic at the expense o
existing or new standard techniques, the management should ask the following question

1. Will we spend more than we save by developing and implementing a new heuristic?
2. Will the decision process be unnecessarily slowed down by its introduction?
933

This content downloaded from

80.112.186.47 on Tue, 07 Jan 2025 15:21:43 UTC
All use subject to https://about.jstor.org/terms
 Journal of the Operational Research Society Vol. 34, No. 10

3. Is there a chance that work quality will be reduced to a level that negates the savings
brought about by introduction?
4. Are the results, when actually implemented, not really significantly better than those
produced before introduction?

If the answer to any of these questions is in the affirmative, the manager should seriously
question the value of the introduction of the heuristic proposed.

CONCLUSIONS

We have attempted to justify the use of the heuristics in situations wher

of optimality is either impossible or meaningless. Heuristic design strateg
and illustrated via the T.S.P. problem. Good heuristics are fast, sim
of producing high quality solutions most ofthe time. As there is a danger
heuristics without all of these characteristics, the analyst should never b
first heuristic that comes to mind. Sometimes the more standard tec
represent a better way. Instances of the failure to accept this have at
heuristic approach into disrepute. Because the approach is so problem-
at present one of the standard O.R. techniques. Yet this should not preve
accepted by practitioners as a valid decision-making tool. Situati
practical application of O.R., where lack of reliable data reigns supre
the heuristic approach is the only realistic alternative, and therein lies it

REFERENCES

1C. W. Churchman (1970) Operations research as a profession. Mgmt Sci. 17, B37-B
2J. Edmonds (1965) Paths, trees and flowers. Can. J. Math. 17, 449-467.
3A. Newell and H. A. Simon (1961) Computer simulation of human thinking. Scienc
4 A. Newell and H. A. Simon (1961) A program that simulates human problem solving. Pro
Automata. Oldenburg, Munich.
5A. Newell and H. A. Simon (1964) Problem-solving machines. Int. Sci. Technol. 36,
6A. Newell and H. A. Simon (1965) Simulation of human processing of informatio
Computing (J. Heller, P. D. Lax, D. H. Lehmer and R. D. Richtmyer, Eds). Slaught Mem
10, Am. math. Mon. 72, No. 2, Part 11, 111-118.
7H. Muller-Merbach (1976) A morphological classification of heuristic techniques. Z. O
8H. Muller-Merbach (1969) Operations Research. Methods and Models for Optimal Plan
Vahlen, Berlin.
9N. Christofides (1976) Worst case analysis of a new heuristic for the travelling salesman
Carnegie-Mellon University, Pittsburgh, Pennsylvania.
10 S. Lin and B. Kerningham (1973) An effective heuristic algorithm for the travelling sa
Res. 21, 498.
11 R. M. Karp (1976) The probabilistic analysis of some combinatorial search algorithms. In
Directions and Recent Results in Algorithms and Complexity (J. B. Traub, Ed.). Acade
12 P. Krolak, W. Felts and G. Marble (1971) A man-machine approach toward solving th
problem. CACM 14, 327-334.
13 M. R. Garey and D. S. Johnson (1976) Approximation algorithms for combinatorial p
bibliography. In Proc. Symp. on New Directions and Recent Results in Algorithms and Com
Ed), pp. 41-52. Academic Press, New York.
14 S. Baum and R. Carlson (1979) On solutions that are better than most. Omega 7, 24
15 K. L. McRoberts (1971) A search model for evaluation of combinatorially explosive p
1331-1349.
16 J. Beardwood, J. H. Halton and J. M. Hammersley (1959) The shortest path through many points. In Pro
Camb. Phil. Soc. 55, 299-327.

934

This content downloaded from

80.112.186.47 on Tue, 07 Jan 2025 15:21:43 UTC
All use subject to https://about.jstor.org/terms