Assignment Problems

 elay
Eranda C

Abstract
Assignment problems arise in dierent situations where we have to nd an
optimal way to assign n objects to m other objects in an injective fashion.
Depending on the objective we want to optimize, we obtain dierent problems
ranging from linear assignment problems to quadratic and higher dimensional
assignment problems.
The assignment problems are a well studied topic in combinatorial opti-
mization. These problems nd numerous application in production planning,
telecommunication, VLSI design, economics, etc.
We introduce the basic problems classied into three groups: linear as-
signment problems, three and higher dimensional assignment problems, and
quadratic assignment problems and problems related to it. For each group of
problems we mention some applications, show some basic properties and de-
scribe brie y some of the most successful algorithms used to solve these prob-
lems.

1 Introduction
Assignment problems deal with the question how to assign n objects to m other
objects in an injective fashion in the best possible way. An assignment problem
is completely specied by its two components: the assignments - which represent
the underlying combinatorial structure, and the objective function to be optimized
which models \the best possible way".
In the classical assignment problem one has m = n and most of the problems with
m > n can be transformed or are strongly related to analogous problems with m = n.
Therefore, we will consider m = n through the rest of this chapter, unless otherwise
specied.
 This research has been supported by the Spezialforschungsbereich F 003 \Optimierung und
Kontrolle", Projektbereich Diskrete Optimierung.
y Technical University Graz, Institute of Mathematics, Steyrergasse 30, A-8010 Graz, Austria,
E-mail: [email protected]

1
From the mathematical point of view an assignment is a bijective mapping of a nite
set N = f1; 2; : : :; ng into itself, i.e., a permutation  assigning some j = (i) to
each i 2 N . The set of all permutations (assignments) of n items will be denoted by
Sn and has n! elements. Every permutation  of the set N = f1; : : :; ng corresponds
uniquely to a permutation matrix X = (xij ) with xij = 1 for j = (i) and xij = 0
for j 6= (i). Thus a permutation matrix X = (xij ) can be dened as a matrix which
fullls the following conditions, so-called assignment constraints
X
n
xij = 1 for all j = 1; : : :; n
i=1
X
n
xij = 1 for all i = 1; : : :; n (1)
j =1
xij 2 f0; 1g for all i; j = 1; : : :; n
By replacing the conditions xij 2 f0; 1g by xij  0 in (1), we get a doubly stochastic
matrix . The set of all doubly stochastic matrices forms the assignment polytope
PA . Due to a famous result of (Birkho, 1946), the assignment polytope PA is the
convex hull of all assignments, or equivalently, every doubly stochastic matrix can
be written as convex combination of permutation matrices.
The concept of an assignment is strongly related to another well known concept
in graph theory and in combinatorial optimization, matching in bipartite graphs.
A bipartite graph G is a triple (V; W ; E ), where the vertex sets V and W have no
vertices in common and the edge set E is a set of pairs (i; j ) where i 2 V and j 2 W .
A subset M of E is called a matching , if every vertex of G is incident with at most
one edge from M . The cardinality of M is called cardinality of the matching. The
maximum matching problem asks for a matching with as many edges as possible. A
matching M is called a perfect matching , if every vertex of G is incident with exactly
one edge from M . Evidently, every perfect matching is a maximum matching. A
perfect matching in a bipartite graph G = (V; W ; E ) with V = fv1 ; v2; : : :; vng,
W = fw1; w2; : : :; wng can be represented by a permutation M of f1; 2; : : :; ng such
that M (i) = j if and only if (vi ; wj ) 2 M . Hence a perfect matching in a bipartite
graph is an assignment. p
(Hopcroft and Karp, 1973) gave an O(jE j jV j)-algorithm which p constructs a perfect
matching if it exists. (Even and Tarjan, 1975) gave an O( jV jjE j) algorithm for
the maximum ow problem on unit capacity simple networks, algorithm which can
also be applied to nd a matchingpof maximum cardinality in a bipartite graph.
(Alt et al., 1991) gave an O(jV j1:5 jE j= log jV j) implementation for the Hopcroft-
Karp algorithm. Based on ideas similar to those in (Hopcroft and Karp, 1973), a
fast randomized Monte-Carlo algorithm is given by (Mulmuley et al., 1087). This
2
algorithm nds a perfect matching at costs of a single matrix inversion. The reader
is referred to the bibliography in (Burkard and Cela, 1999) for further reference
pointers related to algorithms for cardinality matching problems.

2 Linear Assignment Problems

The linear assignment problem (LAP) is one of the oldest and most studies prob-
lems in combinatorial optimization. Many dierent algorithms have been developed
to solve this problem. Also other aspects of the problem as the asymptotic behav-
ior or special cases have been thoroughly investigated. The reader is referred to
(Dell'Amico and Martello, 1997) for a comprehensive annotated bibliography and
to (Burkard and Cela, 1999) for a recent review on assignment problems.
2.1 Problem denition and applications
Recall the original model where n items are to be assigned to n other objects in
the best possible way. Let cij be the cost incurred by the assignment of object i
to object j . We are looking for an assignment  which minimizes the overall cast
P n c . Thus, the linear assignment problem (LAP) is given as follows
i=1 i(i)
X
n
min c ;
2Sn i=1 i(i)
(2)

where Sn is the set of permutations of f1; 2; : : :; ng. Based on the description (1) of
the set of all assignments (see Section 1), the LAP can also be formulated as follows:
P
min ij cij xij over all matrices X = (xij ) which fulll (1).
Due to Birkho's result we can relax the conditions xij 2 f0; 1g to xij  0 and
obtain the linear programming formulation of the LAP. Any basic solution of this
linear program corresponds to a permutation matrix.
X
n
min cij xij
i=1
X
n
xij = 1 j = 1; : : :; n (LP)
i=1
Xn
xij = 1 i = 1; : : :; n
j =1
xij  0 i; j = 1; : : :; n.

3
 As we will mention in the next section many algorithms for the LAP are based
on linear programming techniques and consider often the dual linear program:
X
n X
n
max ui + vj
i=1 j =1
ui + vj  cij i; j = 1; : : :; n (3)
ui; vj 2 IR i; j = 1; : : :; n,
where ui and vj , 1  i; j  n, are dual variables.
Among the numerous applications of the LAP the so-called personnel assignments
are the most typical. In the personnel assignment we want to assign people to
objects, e.g. jobs, machines, rooms, to other people etc. Each assignment has a
\cost" and we want to make the assignment so as to minimize the overall sum of
the costs. For example one company might want to assign graduates to vacant jobs
In this case the cost cij is given by cij = ?pij where pij is the prociency index
for placing candidate i to job j , and thePgoal is to assign each candidate i to some
vacancy (i) such P that the overall cost i ci(i) is minimized, or equivalently, the
overall prociency i pi(i) is maximized.
There are many other applications of the linear assignment problem e.g. in locating
and tracing objects in space, scheduling on parallel machines, inventory planning,
vehicle and crew scheduling, wiring of typewriters etc. The reader is referred to
(Ahuja et al., 1995) and (Burkard and Cela, 1999) for a detailed description of some
applications of the LAP and literature pointers to other applications.

2.2 Algorithms for the LAP

The LAP can be solved eciently, and the design of ecient solution methods for
this problem has been an object of research for many years. There exists an amazing
amount of algorithms, sequential and parallel, for the LAP, ranging from primal-
dual combinatorial algorithms, to simplex-like methods. The worst-case complexity
of the best sequential algorithms for the LAP is O(n3 ), where n is the size of the
problem. From the computational point of view very large scale dense assignment
problems with about 106 nodes can be solved within a couple of minutes by sequential
algorithms, see (Lee and Orlin, 1994).
There is a number of survey papers and books on algorithms, among others (Derigs,
1985), (Dell'Amico and Toth, 2000) and the book on the rst DIMACS challenge
edited by (Johnson and McGeoch, 1993). Among papers reporting on computational
experience we mention (Carpaneto et al., 1988; Lee and Orlin, 1994; Dell'Amico and
Toth, 2000) and some of the papers in (Johnson and McGeoch, 1993).
4
Most sequential algorithms for the LAP can be classied into primal-dual algorithms
and simplex-based algorithms. Primal-dual algorithms work with a pair consisting of
an infeasible solution xij , 1  i; j  n, of LP (called primal solution), and a feasible
solution ui ; vj , 1  i; j  n of the dual (3) (called dual solution). These solutions
fulll the complementarity slackness conditions:
xij (cij ? ui ? vj ) = 0 ; for 1  i; j  n (4)
These solutions are updated iteratively until the primal solution becomes feasible,
while keeping the complementary slackness conditions fullled and the dual solution
feasible. At this point the primal solution would be optimal, according to duality
theory.
Dierent primal-dual algorithms dier on 1) the way they obtain a starting pair of
a primal and a dual solution fullling the conditions described above, and 2) the
way the solutions are updated. A starting dual solution can be obtained as in the
Hungarian method by setting ui := minfcij : 1  j  ng, for 1  i  n, and then
vj := minfcij ? ui: 1  i  ng, for 1  j  n. An infeasible primal starting solution
could be given by a matching of maximal cardinality in the bipartite graph G =
(V; W ; E ), where V = W = f1; 2; : : :; ng, and E = f(i; j ):cij := cij ? ui ? vj = 0g.
Then, set xij := 1 if (i; j ) is an edge of the matching and xij = 0, otherwise. Clearly,
the pair of solutions obtained in this way fulll the complementarity slackness con-
ditions. One way of updating the pair of solutions is the shortest augmenting path
method. This method gives raise to a whole class of algorithms which meet the
best known time complexity bound for the LAP, namely O(n3 ). For a given pair
of solutions as above construct a weighted directed bipartite graph G~ = (V; W ; E~ )
with arc set E~ = D [ R with set of forward arcs D = f(i; j ): (i; j ) 2 E; xij = 0g and
set of backward arcs R = f(j; i): (i; j ) 2 E;  xij = 1g. The weights of the backward
arcs are set equal to 0, whereas the weights of the forward arcs are set equal to
the corresponding reduced costs cij . Then select a node r in V which has not been
assigned yet, and solve the single-source shortest path problem, i.e., compute the
shortest paths from r to all nodes of G~ . The shortest among all paths from r to some
free node in W is used to augment the current primal solution by swapping the free
and matched edges. The dual solution and the reduced costs are then accordingly
updated. It can be shown that after n augmentations an optimal primal solution
results, see (Derigs, 1985).
There are various shortest augmenting path algorithms for the LAP. Basically they
dier in the way they determine a starting pair of primal and dual solutions, and
by the subroutine they use for computing the shortest paths. Most of the existing
algorithms use the Dijkstra algorithm for the shortest path computations.
Simplex-based algorithms are special implementations of the primal or the dual
simplex algorithm for linear programming applied to LP. Simplex-based algorithms
5
for the LAP are specic implementations of the network simplex algorithm. The
latter is a specialization of the simplex method for linear programming to network
problems. The specialization relies on exploiting the combinatorial structure of
network problems to perform ecient pivots acting on trees rather than on the
coecient matrix.
It is well known that there is a one-to-one correspondence between primal (integer)
basic solutions of the LAP and spanning trees of the bipartite graph G related to
assignment problems as described in Section 1. Moreover, given a spanning tree,
one can uniquely determine the values of the corresponding dual variables so as
to fulll the complementarity slackness conditions, as soon as the value of one of
those variables is xed (arbitrarily). Every integer primal feasible basic solution is
highly degenerate because it contains 2n ? 1 variables and n ? 1 of them are equal
to 0. Hence degeneracy poses a problem, and the rst simplex-based algorithms for
the LAP were exponential. The rst steps towards the design of polynomial-time
simplex-based algorithms were made by introducing the concept of so-called strongly
feasible trees , introduced by (Cunningham, 1976). There are implementations of
simplex-based algorithm for the LAP which match the best known time complexity
bound of O(n3 ). The reader is referred to (Burkard and Cela, 1999) for references
and further details.
More recently (Ramakrishnan et al., 1993) applied an interior point algorithm to the
LAP and got promising results, in particular for large size instances, see (Johnson
and McGeoch, 1993).
Since the late 1980s a number of parallel algorithms for the LAP has been proposed.
The speed-up achieved by such algorithms is limited by the sparsity of the cost
matrices and/or the decreasing load across the iterations. For a good review on
parallel algorithms for the LAP and network ow problems in general the reader is
referred to (Bertsekas et al., 1995).
2.3 Asymptotic behavior and probabilistic analysis
When dealing with the asymptotic behavior of the LAP, it is always assumed that
the cost coecients cij are independent random variables (i.r.v.) with a common
prespecied distribution. The main question concerns the behavior of the expected
optimal value of the problem as its size tends to innity.
For cost coecients cij being i.r.v. with a uniform distribution on [0; 1] it has been
shown that the optimal value of the LAP remains within constant bounds as the
size n of the problem tends to innity. The best upper bound equals 2 and is due to
(Karp, 1987). The best lower bound equals 1.51 and is due to (Olin, 1992). Although
the gap between the current lower and upper bounds on the expected optimal value
of the LAP is large, it is believed that the expected value is close to 1:6 or more
6
exactly 62 in the case of independent cost coecients cij uniformly distributed on
[0; 1]. For a discussion in some details and for more references see (Burkard and
Cela, 1999).
A more general scenario where the coecients cij are i.r.v. with a common arbitrary
general distribution has been investigated by (Frenk et al., 1987) and (Olin, 1992).
Under mild assumption on the probability distribution of the coecients the authors
derive constant upper and lower bounds on the expected optimal value of the LAP.
In the case of coecients cij being i.r.v. uniformly distributed on [0; 1], the LAP
can be solved in expected O(n2 log n) time by a randomized algorithm proposed
by (Karp, 1980). Faster randomized algorithms which produce a solution whose
objective function value is within a constant factor of the optimal objective function
value have been proposed by several authors. There is for instance an algorithm of
(Karp et al., 1994) which runs in expected linear time (O(n)) and provides a solution
within a factor of 3 + O(n?a)) of the optimum with probability 1 ? O(n?a ), where a
is some xed positive number. The best known linear assignment problem with an
objective function dierent from that of the LAP is the bottleneck linear assignment
problem (BLAP)
min max c :
 1in i(i)
(5)
considered originally by (Fulkerson et al., 1953). This problem occurs e.g. in the
assignment of jobs to parallel machines so as to minimize the latest completion
time. One of the rst algorithms proposed for the BLAP is the so-called threshold
algorithm . The threshold algorithm chooses a cost element cij =: K - the threshold
value - and constructs a matrix C dened as follows
( 1 if c > K
ij
cij :=
0 if cij  K
Then the algorithm checks whether the bipartite graph with adjacency matrix C
contains a perfect matching or not. The algorithms repeats than this procedure
for a new (and eventually smaller) value of the threshold K . The smallest value
K for which the corresponding bipartite graph contains a perfect matching, is the
optimum value of the BLAP. The best time complexity known today amounts to
O(npnm) and is due to (Punnen and Nair, 1994). (Here m is the number of nite
elements in the coecient matrix (cij ) which would correspond to the number of
edges of the bipartite graph in the graph theoretical setting.)
A randomized algorithm with quadratic expected running time (O(n2)) has been
given by (Pferschy, 1996). A computational study on the comparison of dierent
deterministic algorithms for the BLAP has been given by (Pferschy, 1997).
7
Another linear assignment problem introduced by (Burkard and Rendl, 1991) is the
so-called lexicographic bottleneck assignment problem (LexBAP). In the LexLAP we
want to nd a permutation (assignment)  which lexicographically minimizes c
over all permutation , where c is the vector of costs ci(i), 1  i  n, sorted
non-increasingly.
(Martello et al., 1984) have considered the so-called balanced assignment problem
(BalAP). Given a real n  n matrix C = (cij ), the balanced assignment problem can
be formulated as  
min

max c
i i(i)
? min c
i i(i)
:
The problem can be solved eciently in O(n4 ) time.
A more general linear assignment problem which includes as special cases the linear
assignment problem (2) and the bottleneck assignment problem (5) is the algebraic
assignment problem (AAP) introduced by (Burkard et al., 1977). In the AAP the
coecients cij are elements of a totally ordered semigroup (H; ; ) with composition
 and order relation . The AAP can then be formulated as follows:
min c1(1)  c2(2) 


 cn(n) :
2Sn
(6)

The AAP can be solved eciently if the order relation and the composition fulll
some natural algebraic properties. For further results and reference pointers consult
the survey on algebraic optimization by (Burkard and Zimmermann, 1982).

2.4 Available computer codes and test instances

FORTRAN listings of codes for the LAP and the LBAP can be found in the book
by (Burkard and Derigs, 1980). The code for the LAP is a primal-dual algorithm
based on shortest path computations done by a version of Dijkstra's algorithm.
Source codes of another primal-dual algorithm for the LAP and the LBAP can
be downloaded from http://207.158.230.188/assignment.html. One can choose
among a C++, a PASCAL, and a FORTRAN implementations of an algorithm of
(Jonker and Volgenant, 1986).
A (compressed) FORTRAN source le - called 548.Z - of an implementation of the
Hungarian algorithm, due to (Carpaneto and Toth, 1980), can be downloaded from
ftp://netlib.att.com in /netlib/toms. Other listings of FORTRAN codes for
the LAP can be found in (Carpaneto et al., 1988). The codes are available from the
oppy disk included in the book.
The C code of an ecient implementation of the scaling push-relabel algorithm of
(Goldberg and Kennedy, 1995) for the LAP can be downloaded from Goldberg's
8
network optimization library at
http://www.neci.nj.nec.com/homepages/avg/soft.html.
Finally, listings of 5 FORTRAN codes of auction algorithms for the LAP can be
found in Bertsekas' homepage at
http://web.mit.edu/dimitrib/www/auction.txt.

Test instances of the LAP can be downloaded as ascii les from the homepage of
the OR-Library maintained by J. Beasley at
http://mscmga.ms.ic.ac.uk/pub.
Other test instances can be obtained from the ELIB library at
ftp://ftp.zib.de/pub/Packages/mp-testdata/assign/index.html.

Clearly, since the LAP can be formulated as as a minimum cost ow problem,

algorithms developed for the later can also be applied to the LAP. However, such
algorithms are not supposed to exploit the specic features of the LAP, and hence
may not be competitive with algorithms developed especially for the LAP. Besides
Goldberg's network optimization library and Bertsekas' homepage, other codes for
network optimization can be found in Netlib at
http://www.OpsResearch.com/OR-Links/index.html.
Furthermore, C codes of implementations of the primal and the dual network simplex
algorithm, due to Lobel, can be obtained through
http://www.zib.de/Optimization/index.de.html.

3 Multidimensional Assignment Problems

3.1 General Remarks and Applications
Multi-dimensional (sometimes referred as multi-index ) assignment problems (MAP)
are natural extensions of the linear assignment problem. They have been considered
for the rst time by (Pierskalla, 1967). The most prominent representatives of this
class are axial and planar 3-dimensional assignment problems to be considered in the
next section. The MAP asks for d ? 1 permutations 1 ; 2; : : :; d?1 which minimize
the following objective function:
X
n
min
 ; ;:::;
ci1 (i)2 (i):::d?1 (i) :
1 2 d?1 i=1
In terms of graphs a multidimensional assignment problem can be described as
follows: Let a complete d-partite graph G = (V1; V2; : : :; Vd; E ) with vertex sets Vi ,
jVij = n, i = 1; 2; : : :; d, and edge set E be given. A subset X of V = Sdi=1 Vi is a
9
clique , if it meets every set Vi in exactly one vertex. A d-dimensional assignment is
a partition of V into n pairwise disjoint cliques. If c is a real valued cost function
dened on the set of cliques of G = (V1; V2; : : :; Vd; E ), the d-dimensional assignment
problem asks for a d-dimensional assignment of minimum cost. Special cases where
the costs c of a clique are not arbitrary, but given as a function of elementary costs
attached to the edges of the complete d-partite graph (eg. sum costs, star costs, tour
costs or tree costs) have been investigated and the performance of simple heuristics
has been analyzed in these cases. For more information see e.g. (Burkard and Cela,
1999) and the references therein.
Multidimensional assignment problems in their general form have found some ap-
plications as a means to solve data association problems in in multi-target tracking
and multi-sensor surveillance. The data association problem consists in partitioning
the observations into tracks and false alarms in real time. General classes of these
problems can be formulated as multidimensional assignment problems. Other appli-
cations of MAPs are related to track initiation, track maintenance, and multi-sensor
tracking. Another interesting MAP arises in the context of tracking elementary par-
ticles. For more information on these two applications see (Burkard and Cela, 1999)
and the references therein.

3.2 Axial 3-Dimensional Assignment Problems

Consider n3 cost coecients cijk . The axial 3-dimensional assignment problem (3-
DAP) can then be stated as
n X
n X
X n
min cijk xijk
i=1 j =1 k=1
X n X n
s.t. xijk = 1 ; i = 1; 2; : : :; n ;
j =1 k=1
X n X n
xijk = 1; i = 1; 2; : : :; n ; (7)
i=1 k=1
X n X n
xijk = 1; i = 1; 2; : : :; n ;
i=1 j =1
xijk 2 f0; 1g 8 1  i; j; k  n :
We can think of cijk as the cost of assigning job j to be performed by worker i in
machine k. It follows that xijk = 1, if job j is assigned to worker i in machine k,
and xijk = 0, otherwise.
10
Equivalently, a 3-DAP can be described with the help of two permutations  and
X
n
min
; 2S
ci(i) (i): (8)
n
i=1
Thus this problem has (n!)2 feasible solutions. It has been shown that the dimension
of the axial 3-index assignment polytope (i.e., the convex hull of feasible solutions to
problem (7)) is n3 ? 3n +2. Furthermore several classes of facet dening inequalities
have been identied and ecient algorithms for the separation of these facets have
been developed. For more information on these topic see the description in (Burkard
and Cela, 1999) and consult the references cited therein.
In contrast to the linear assignment problem LAP the 3-DAP cannot be solved
eciently and branch and bound algorithms are the mostly used algorithms to solve
this problem. The lower bounds are usually computed by solving some Lagrangean
relaxation of the 3-DAP by subgradient optimization approaches, see e.g. (Balas and
Saltzman, 1991). The authors introduce in (Balas and Saltzman, 1991) also a non-
trivial branching strategy which exploits the structure of the problem and allows to
x several variables at each branching node.
A heuristic for solving the 3-DAP has been proposed by (Pierskalla, 1967).
Finally, there exists a number of eciently solvable special cases of the 3-DAP, e.g.
if the cost coecients are taken from a 3-dimensional Monge array or if the cost
coecients are decomposable, i.e., cijk = ui vj wk and ui ; vj , and wk are nonnegative.
The reader is referred to (Burkard and Cela, 1999) and the references therein for
more information on this topic.
3.3 Planar 3-Dimensional Assignment Problems
Let cijk , 1  i; j; k  n, be n3 cost coecients. The planar 3-dimensional assignment
problems (3-PAP) is stated as follows:
n X
n X
X n
min cijk xijk
i=1 j =1 k=1
s.t.
X
n
xijk = 1 ; j; k = 1; 2; : : :; n ; (9)
i=1
Xn
xijk = 1 ; i; k = 1; 2; : : :; n ;
j =1
X n
xijk = 1 ; i; j = 1; 2; : : :; n ;
k=1

11
 xijk 2 f0; 1g i; j; k = 1; 2; : : :; n :
The 3-PAP has interesting applications in time tabling problems. See (Euler and Le
Verge, 1996) for a recent study on time tables and related polyhedra.
It is easily seen that the feasible solutions of the 3-PAP correspond to Latin squares.
Thus, the number of feasible solutions of a 3-PAP of size n equals the number of
Latin squares of order n, and hence increases very fast. Similarly to the 3-DAP also
the 3-PAP is a hard problem and cannot be solved eciently. There are not many
algorithms known for the 3-PAP. Besides two branch and bound algorithms due to
(Vlach, 1967) and (Magos and Miliotis, 1994) there is also a tabu search algorithm
for the 3-PAP due to (Magos, 1996). It uses a neighborhood structure based on the
relationship between the 3-PAP and the latin squares. See e.g. (Burkard and Cela,
1999) for more information on algorithms for the 3-PAP.

4 The Quadratic Assignment Problem

The quadratic assignment problem (QAP) was introduced in 1957 by Koopmans
and Beckmann as a model for a plant location problem. Since then the QAP has
been object of intensive investigations concerning dierent aspects of the problem
ranging from algorithms to asymptotic behavior and special cases. In contrast to its
linear counterpart the QAP can not be solved eciently. However there are some
restricted cases of the QAP which can be solved in polynomial time. Such special
cases are QAPs whose coecient matrices show special combinatorial properties,
e.g. Monge and Monge-like properties. The interested reader is referred to (Cela,
1998) for a detailed discussion on eciently solvable special cases of the QAP.
Also from the practical point of view the QAP is widely considered as one of the
hardest problems in combinatorial optimization and there are instances of size 24
which can not yet be solved to optimality in reasonable computational time, e.g. the
Nugent instance of size [24], see QAPLIB (Burkard et al., 1997)
The reader is referred to (Burkard and Cela, 1997) for a comprehensive annotated
bibliography, to (Cela, 1998) for a recent monograph, and to (Burkard et al., 1998)
for a recent review on quadratic assignment problems and for reference pointers.
4.1 Problem denition and applications
Among the most common and best known applications of the QAP are those which
arise in a facility location context. This is one reason for choosing the facility location
terminology to introduce the problem.
Consider the problem of allocating n facilities to n locations, with costs depending
on the distance between the locations and the ow between the facilities, plus costs
12
associated with a facility being placed at a certain location. The objective is to assign
each facility to a location such that the total cost is minimized. More specically,
we are given three n  n real matrices A = (aij ), B = (bkl) and C = (cik ), where
aij is the ow between the facility i and facility j , bkl is the distance between the
location k and location l, and cik is the cost of placing facility i at location k. The
Koopmans-Beckmann version of the QAP can be then formulated as follows:
0n n 1
X X X
n
min @
2Sn
aij b i  j + ai i A
( ) ( ) ( ) (10)
i=1 j =1 i=1
where Sn is the set of all permutations of N . A product aij b(i)(j ) is the cost of
assigning facility i to location (i) and facility j to location (j ).
A more general version of the QAP was introduced by Lawler in 1963. In this version
we are given a four-dimensional array D = (dijkl ) of coecients instead of the two
matrices A and B , and the problem can be stated as
0n n 1
X X X
n
min @
2Sn
dij i  j + ci i A :
( ) ( ) ( ) (11)
i=1 j =1 i=1

Besides applications in facility location the QAP nds applications in VLSI design,
computer manufacturing, scheduling, process communications, backboard wiring
and typewriter keybord design. We describe here a less known application on turbine
balancing. Hydraulic turbine runners as used in electricity generation consist of a
cylinder around which a number of blades are welded at regular spacings. Due to
inaccuracies in the manufacturing process, the weights of these blades dier slightly,
and it is desirable to locate the blades around the cylinder in such a way that
the distance between the center of mass of the blades and the axis of the cylinder
is minimized. This problem was introduced by Mosevich in 1986 and has been
formulated as a QAP by (Laporte and Mercure, 1988). The places at regular spacings
on the cylinder are modeled by the vertices v1; : : :; vn of a regular n-gon on the unit
circle in the Euclidean plane. Thus
    
v = sin 2i ; cos 2i ; 1  i  n.
i n n
The masses of the n blades are given by the positive reals 0 < m1  m2 


 mn .
The goal is to assign each mass mi to some vertex v(i) in such a way that the center
of gravity !
Xn sin( 2ni )
m(i) ;
i=1 cos( 2ni )
13
of the resulting mass system is as close to the origin as possible. Minimizing the
Euclidean norm of the above vector is equivalent to minimizing the expression
n X
X n  2(i ? j ) 
m(i)m(j) cos n : (12)
i=1 j =1
Thus we have a quadratic assignment problem,
 where  the coecient matrices A and
B have a very special structure: aij = cos 2(i?nj) and bij = mi
mj .
Finally, notice that there are a number of other well known combinatorial optimiza-
tion problems which can be formulated as QAPs, e.g. graph partitioning, maximum
clique, the traveling salesman problem, the linear arrangement problem, and the
minimum weight feedback arc set problem.
To conclude this section notice that the QAP is a \very hard" problem from the
theoretical point of view. Not only that the QAP cannot be solved eciently but it
even cannot be approximated eciently within some constant approximation ratio.
Furthermore, also nding local optima is in general not a trivial task. For informa-
tion and references concerning the computational complexity of the QAP see (Cela,
1998).

4.2 Alternative formulations of the QAP

There exist several equivalent formulation of the QAP. Dierent formulations stress
dierent characteristics of the problem and lead to dierent solution approaches.
Recall that there is a one-to-one correspondence between the permutations of N =
f1; 2; : : :; ng and the n  n permutation matrices dened in Section 1. Let Xn be the
set of n  n permutation matrices. In terms of permutation matrices QAP(A,B,C)
can be formulated as the following quadratic integer program:
X n X
n X n X
n X
n
min aij bkl xik xjl + cij xij (13)
i=1 j =1 k=1 l=1 i;j =1
s.t. (xij ) 2 Xn (14)
Let us dene an inner product between n  n matrices as follows
n X
X n
hA; Bi := aij bij ;
i=1 j =1

Clearly, we have XAX T = (a(i)(j )), for some n  n matrix A, a permutation  2 Sn

and the associated permutation matrix X 2 Xn . Thus the QAP (13)-(14) can be
14
formulated alternatively as
min hA; XBX T i + hC; X i (15)
s.t. X 2 Xn :

Finally consider the trace formulation of the QAP. The trace tr(A)Pof an n  n
matrix A = (aij ) is dened as sum of its diagonal elements: tr(A) = ni=1 aii . Let
us denote B  := XB tX t, where X is the permutation matrix corresponding to .
We get
 Xn
 X
n
tr(AB ) = aij bji = aij b(i)(j) ;
i;j =1 i;j =1
bij P
since = b(i)(j ), for i; j 2 f1; 2; : : :; ng. Since tr(CX t) = ni=1 ci(i) , the QAP in
(15) can be formulated as
min tr(AXB T + C )X T (16)
s.t. X 2 Xn :
4.3 Linearizations
A rst attempt to solve the QAP would be to eliminate the quadratic term in the
objective function (13), in order to transform the problem into a (mixed) 0-1 linear
program (MILP). The linearization of the objective function is usually achieved by
introducing new variables and new linear (and binary) constraints. The very large
number of new variables and constraints, however, poses an obstacle for eciently
solving the resulting linear integer programs. The optimal value of an LP relaxation
of some MILP formulation is a lower bound for the QAP. In this context the \tight-
ness" of the continuous relaxation of the linear integer program mentioned above is
a desirable property. Several linearizations of the QAP have been proposed in the
literature, e.g. by (Lawler, 1963), by (Kaufmann and Broeckx, 1978) (this lineariza-
tion has the smallest number of variables and constraints), by (Frieze and Yadegar,
1983) and by (Adams and Johnson, 1994). The last linearization unies most of the
previous linearizations and is important for getting lower bounds. The QAP with
array of coecients C = (dijkl ) is proved to be equivalent to the following mixed 0-1
linear program
n X
X n
min dijkl yijkl
i;j =1 k;l=1
s.t. (xij ) 2 Xn ;
15
 X
n
yijkl = xjl ; j; k; l = 1; : : :; n;
i=1
Xn
yijkl = xjl ; i; j; l = 1; 2; : : :; n;
k=1
yijkl= yjilk ; i; j; k; l = 1; : : :; n;
yijkl 0; i; j; k; l = 1; 2; : : :; n;
where each yijkl represents the product xik xjl . Although as noted by (Adams and
Johnson, 1994) a signicant smaller formulation in terms of both the variables and
constraints could be obtained, the structure of the continuous relaxation of the above
formulation is favorable for solving it approximately by Lagrangean dual methods.
The theoretical strength of the linearization (17) relies on the fact that the con-
straints of the continuous relaxations of previous linearizations can be expressed as
linear combinations of the constraints of the continuous relaxation of (17). More-
over, many of the previously published lower-bounding techniques can be explained
based on the Lagrangean dual of this relaxation. For more details on this topic we
refer to Section 4.4.2.
4.4 Lower bounds
Lower bounding techniques are used within implicit enumeration algorithms, such
as branch and bound, to perform a limited search of the set of feasible solution, until
an optimal solution is found. Numerous bounding techniques have been developed
for the QAP so far. The most successful bounding techniques for the QAP can
be classied in 3 groups: Gilmore-Lawler type lower bounds, bounds based on LP
relaxation, eigenvalue related bound
4.4.1 Gilmore-Lawler type lower bounds
The basic idea of these type of lower bounds goes back to the early 60s when
(Gilmore, 1962) and (Lawler, 1963) developed the so-called Gilmore-Lawler bound
(GLB) for the QAP. Nowadays the Gilmore-Lawler type lower bounds and especially
the GLB are the most frequently used bounds within branch and bound algorithms
for the QAP. The most advantageous property of these bounds is that they can be
computed eciently. The main drawback is the fast deterioration of their quality
with increasing problem size.
To compute the GLB for a given QAP of size n one has to solve n2 + 1 LAPs, n2 of
them of size n ? 1 and the last one of size n, as described below.
Consider an instance of the Lawler QAP (11) with coecients D = (dijkl ). For each
ordered pair of indices (i; k), 1  i; k  n, solve an LAP with coecient matrix
16
D(i;k) = (dijkl ) under the additional constraint that (i) = k. Let us denoted by
lik the optimal solutions of the above mentioned LAP. Finally, solve an LAP with
coecient matrix (lij + cij ); its optimal value is the GLB. Since the LAP can be
solved eciently, also the GLB can be computed eciently.
Several Gilmore-Lawler type lower bounds have been developed aiming at improving
the quality of the GLB. One of the ideas on which such bounds are based are the so-
called reduction methods. These methods decompose each quadratic cost coecient
into several terms so as to guarantee that some of them end up in being linear cost
coecients and can be moved to the linear term of the objective function. This
would yield a tighter lower bound because the LAP can be solved exactly.
More recently another bounding procedure which shares the basic idea of the GLB
has been proposed by (Hahn and Grant, 1998). This procedure combines GLB ideas
with reduction steps in a general framework. The resulting bound (HGB) shows a
good trade o between computation time and bound quality when tested in instances
from QAPLIB, see (Burkard et al., 1997).
4.4.2 Bounds based on linear programming relaxations
Consider a mixed integer linear programming (MILP) formulations of the QAP.
Clearly, the optimal solution of the continuous relaxation of an MILP formulation
is a lower bound for the optimal value of the corresponding QAP. Moreover, each
feasible solution of the dual of this relaxation is also a lower bound. Even for QAPs
of moderate size it is practically impossible to solve to optimality the LP relaxations
of any MILP formulation. Generally, such formulations are highly degenerated and
already for QAPs of moderate size the memory requirements become prohibitive
as shown by (Resende et al., 1995). The LP relaxations of the MILPs are there-
fore approximately solved by applying subgradient optimization based techniques
or Lagrangean relaxations. Adams et al. have shown that a particular Lagrangean
relaxation of the MILP (17) can be solved eciently for each xed set of Lagrangean
multipliers, see (Adams and Johnson, 1994). The Lagrangean multipliers are then
updated iteratively in the fashion of a dual ascent procedure. The strength of the
method of Adams and Johnson relies on the fact that it can produce all Gilmore-
Lawler-like bounds described in Section 4.4.1 and others, but the HGB, for dierent
settings of the Lagrangean multipliers.
Recently (Karisch et al., 1999) have shown that although HGB cannot be obtained
by applying the algorithm of Adams and Johnson, both AJB and HGB can be
obtained as feasible solutions of the dual of the continuous relaxation of the MILP
formulation (17). Karisch et al. propose an iterative algorithm to approximately
solve this dual, and show that AJB, HGB, and all other Gilmore-Lawler-like bounds
can be obtained by applying this algorithm with specic settings for the control
17
parameters.

4.4.3 Eigenvalue based lower bounds

These bounds were introduced by (Finke et al., 1987), and can be applied to the
Koopmans-Beckmann QAP in (10). They are based on the relationship between the
objective function value of the QAP in the trace formulation (16) and the eigen-
values of its coecient matrices. When designed and implemented carefully, these
techniques produce bounds of good quality in comparison with Gilmore-Lawler-like
bounds or, more generally, with bounds based on linear relaxations. However, these
bounds are quite expensive in terms of computation time requirements and are,
therefore, not appropriate for use within branch and bound algorithms.

4.5 Exact solution methods

Since QAP is a hard problem from the theoretical (and also from the practical) point
of view and no ecient algorithms are known for this problem. All existing exact
algorithms are in principle enumeration methods like branch and bound, cutting
plane algorithms and branch and cut algorithms.

4.5.1 Branch and bound algorithms

Nowadays branch and bound algorithms appear to be the most ecient exact algo-
rithms for solving the QAP.
The most ecient branch and bound algorithms for the QAP employ the Gilmore-
Lawler bound (GLB). The reason is that other bounds which outperform GLB in
terms of bound quality are simply too expensive in terms of computation time. More
recently some eorts have been made to employ other Gilmore-Lawler-like bounds.
The bound of HGB has been used in a branch and bound algorithm by (Hahn et al.,
1998) and the results are promising.
Three types of branching strategies are mostly used for the QAP: single assignment
branching , pair assignment branching and branching based on relative positioning .
The most ecient strategy is the single assignment branching. It assigns a facility to
a location in each branching step, i.e., each problem is divided into subproblems by
xing the location of one of the facilities which are not assigned yet. Several rules for
the choice of the facility-location pair to determine the subproblems of a new level
of the search tree have been proposed by dierent authors. The appropriate rule
usually depends on the bounding technique. If the GLB is employed the branching
rule is frequently formulated in terms of the reduced costs of the last assignment
problem solved to bound the subproblem which is currently being branched.
18
As for the selection rule there seems to be no clear winner among dierent strate-
gies tested for the QAP, ranging from problem-independent depth or breadth rst
search to instance dependent criteria related to the maximization of lower bounds
or reduced costs.
More recently a number of parallel branch and bound algorithms have been devel-
oped for the QAP, e.g. by (Pardalos and Crouse, 1989), (Bruengger et al., 1997),
and (Clausen and Perregaard, 1997).

4.5.2 Cutting plane methods

The cutting plane methods in general can be classied into traditional cutting plane
methods and polyhedral cutting plane methods . Traditional cutting plane algorithms
for the QAP have been developed by dierent authors, e.g. (Bazaraa and Sherali,
1982), (Balas and Mazzola, 1984a; Balas and Mazzola, 1984b)), and (Kaufmann and
Broeckx, 1978). These algorithms make use of mixed integer linear programming
(MILP) formulations for the QAP which are suitable for Benders' decomposition.
Generally, the time needed for these methods to converge is too large, and hence
these methods may solve to optimality only very small QAPs. However, heuristics
derived from cutting plane approaches produce good suboptimal solutions in early
stages of the search.
Also polyhedral cutting planes or branch and cut algorithms make use of MILP
formulations of the QAP. Additionally, polyhedral cutting plane methods make use
of a class of (nontrivial) valid or facet dening inequalities known to be fullled by
all feasible solutions of the original problem. Some properties and few facet dening
inequalities of the QAP polytope are already known, but still polyhedral cutting
plane methods for the QAP are not yet backed by a strong theory. Some eorts to
design branch and cut algorithms for the QAP have been made by (Padberg and
Rijal, 1996) and (Kaibel, 1997). The numerical results are encouraging, although
the developed software is of preliminary nature, as claimed by the authors.

4.6 Heuristics
The QAP is a very dicult problem from the practical point of view, instances of
dimension n > 20 being still not practical to solve because of very high computa-
tion time requirements. This is probably the reason why the literature abounds in
heuristics which are the only available algorithms to provide good quality solutions
for the QAP in a reasonable computational time. The reader is referred to (Burkard
et al., 1998) for numerous literature pointers to numerous heuristics applied to the
QAP.
19
The numerous heuristic approaches developed for the QAP can be classied in con-
struction methods, limited enumeration methods, local search algorithms including
improvement methods, tabu search and simulated annealing, genetic algorithms,
greedy randomized adaptive search procedures (GRASP), and ant systems. In the
case of the QAP there is no widely accepted winner among these strategies. The con-
struction methods, the limited enumeration methods, and the improvement methods
seem, however, to be outperformed by the other heuristic approaches.
We refer to other chapters of this handbook for a detailed description of the general
metaheuristic approaches mentioned above. In the following we brie y mention the
most frequently used neighborhood structures included in local search algorithms
for the QAP.
Frequently used neighborhoods for the QAP are the pair-exchange neighborhood
and the cyclic triple-exchange neighborhood. In the case of pair-exchanges the
neighborhood of a given solution (permutation) consists of all permutations which
can be obtained from? the

given one by applying a transposition to it. The size of this
n
neighborhood is O( 2 ). In the case of cyclic triple-exchanges, the neighborhood of
a solution (permutation)  consists of all permutations obtained from  by? a
cyclic
exchange of some triple of indices. The size of this neighborhood is O( n3 ). In
general cyclic triple-exchanges do not lead to better results when compared with
pair-exchanges.
4.7 Available computer codes for the QAP
(Burkard et al., 1997) have compiled a library of QAP instances (QAPLIB) which is
widely used to test bounds, exact algorithms, and heuristics for the QAP. Many of
these instances have not been solved to optimality yet, the most celebrated among
them being the instances of (Nugent et al., 1969) of size larger than 25. QAPLIB
can be found at
http://www.opt.math.tu-graz.ac.at/ekarisch/qaplib.
Two codes for computing lower bounds are also available from the QAPLIB web
page: a FORTRAN code due to (Burkard and Derigs, 1980) to compute the GLB for
instances of size up to 256, and another FORTRAN code to compute the elimination
bound (ELI) for symmetric QAP instances of size up to 256. Finally a FORTRAN
code of the branch and bound algorithm developed by (Burkard and Derigs, 1980)
can also be downloaded from the QAPLIB web page.
Recently, (Espersen et al., ) have developed QAPpack which is a JAVA package
containing a branch and bound algorithm to solve the QAP. In QAPpack a number
of bounds based on linearization are implemented: the Gilmore-Lawler bound, the
bound of (Carraresi and Malucelli, 1994), the bound of (Adams and Johnson, 1994),
the bound of (Hahn and Grant, 1998), and the bound of (Karisch et al., 1999). The
20
implementation is based on the dual framework provided by (Karisch et al., 1999).
QAPpack can be found at http://www.imm.dtu.dk/ete/QAPpack.
There are also some codes of heuristics available. The (compressed) FORTRAN
source le - 608.Z - of a heuristic due to (West, 1983), can be downloaded at
ftp://netlib.att.com in /netlib/toms.
The source les (compressed tar-les) of two FORTRAN implementations of GRASP
for dense QAPs by (Resende et al., 1996) and sparse QAPs by (Pardalos et al., 1997)
can be downloaded from Resende's web page at http://www.research.att.com/emgcr/src/index.ht
The source le of a FORTRAN implementation of the simulated annealing algorithm
of (Burkard and Rendl, 1984) can be downloaded from the QAPLIB web page.
The source le of a C++ implementation of the simulated annealing algorithm of
(Connolly, 1990), due to Taillard, can be downloaded from Taillard's web page at
http://www.idsia.ch/eeric/codes.dir/sa qap.c. Also a source le of a PAS-
CAL implementation of the robust tabu search algorithm by (Taillard, 1991) can be
found at Taillard's web page.

4.8 Asymptotic behavior

While being a very dicult problem both from the theoretical and from the practical
point of view, the QAP shows an interesting asymptotic behavior, which suggests
that under certain probabilistic conditions on the problem data, QAPs which are
large enough are trivial to solve. Namely, it can be shown that the ratio between the
\best" and \worst" values of the objective function approaches 1, as the size of the
QAP approaches innity. Thus the relative error of every heuristic method vanishes
as the size of the problem tends to innity, i.e., every heuristic nds almost always
an almost optimal solution when applied to QAP instances which are large enough.
A number of authors have investigated the asymptotic behavior of the QAP. (Burkard
and Fincke, 1983; Burkard and Fincke, 1985) and (Frenk et al., 1985) have shown the
convergence of the above mentioned ratio to 1 in probability. Later (Szpankowski,
1995) improved the convergence to almost surely. Summarizing we get the following
theorem.
Theorem 4.1 Consider
  a sequence of QAPs Pn , for n 2 IN, with n  n coecient
matrices A n = aijn and B = bijn . Assume that aijn and bijn , n 2 IN, 1 
( ) ( ) ( ) ( ) ( )

i; j  n, are independently distributed random variables on [0; M ], where M is a

positive constant. Moreover, assume that entries aijn , n 2 IN, 1  i; j  n, have a
( )

common distribution, and entries bijn , n 2 IN, 1  i; j  n, have also a common

( )

distribution (which does not necessarily coincide with that of a(ijn) ). Furthermore,
assume that these random variables have nite expected values, variances and third

21
moments.
(n) (n)
Let opt and wor denote  (n) and a worst solution of Pn with their objective
 an optimal
(n)
function values z opt and z wor
 (n)  (n) , respectively. Then the following equality holds
!1 z opt =z wor = 1
almost surely nlim
The asymptotic behavior of the QAP has been exploited by (Dyer et al., 1986) to
analyze the performance of branch and bound algorithms for QAPs with coecients
generated randomly as described above. Dyer et al. have shown that any branch
and bound algorithm that uses single assignment branching and employs a bound
obtained by solving the continuous relaxation of the linearization of Frieze and Yade-
gar would branch on at least n(1?o(1))n=4 nodes with probability tending to 1 as the
size n of the problem tends to innity.
4.9 The biquadratic assignment problem
A generalization of the QAP arises if we consider objective functions of higher degree
and obtain in this way cubic, biquadratic and generally N -adic assignment problems
as introduced by (Lawler, 1963). In this way we get among other also the biquadratic
assignment problem , denoted by BiQAP and stated as follows:
n X
X n X
n X n
min aijkl bmpstxim xjp xks xlt
i;j =1 k;l=1 m;p=1 s;t=1
s.t. X = (xij ) 2 Xn;
where A = (aijkl ) and B = (bmpst) are two n4  n4 arrays.
An application of the BiQAP arises in Very Large Scale Integrated (VLSI) circuit
design. A detailed description of the mathematical modeling of the VLSI problem as
a BiQAP is given by (Burkard et al., 1994). Similarly to the QAP also the BiQAP is
a hard problem and cannot be solved eciently. Gilmore-Lawler-like lower bounds,
branch and bound algorithms and some local search heuristics as well as a GRASP
implementation for the BiQAP can be found in the literature. See (Burkard and
Cela, 1997) or (Cela, 1998) for more information and reference pointers.
4.10 The Bottleneck QAP
Another problem related to the QAP is the bottleneck quadratic assignment problem
(BQAP), obtained by substituting the sum by a max operation in the objective
function of the QAP:
min maxfaij b(i)(j ): 1  i; j  ng :
2Sn

22
The rst occurrence of the BQAP in the literature is due to (Steinberg, 1961) and
arises as an application in backboard wiring while trying to minimize the maximum
length of the involved wires. A well studied problem in graph theory which can be
modeled as a BQAP is the bandwidth problem. In the bandwidth problem we are
given an undirected graph G = (V; E ) with vertex set V and edge set E , and seek a
labeling of the vertices of G by the numbers 1; 2; : : :; n, where jV j = n, such that the
maximum distance of 1-entries of the resulting adjacency matrix from the diagonal
is minimized, i.e., the bandwidth of the adjacency matrix is minimized.
Analogously to the QAP also the BQAP is a hard problem which cannot be solved
eciently. Some enumeration algorithms to solve BQAP to optimality have been
proposed by (Burkard, 1974). Those algorithms employ a Gilmore-Lawler-like bound
for the BQAP.
(Burkard and Fincke, 1982) investigated the asymptotic behavior of the BQAP and
proved results analogous to those obtained for the QAP: Under certain probabilistic
constraints on the problem data, the relative dierence between the worst and the
best value of the objective function approaches 0 with probability tending to 0 as
the size of the problem approaches innity.

23
References
Adams, W. P. and Johnson, T. A. (1994). Improved linear programming-based
lower bounds for the quadratic assignment problem. In Pardalos, P. M. and
Wolkowicz, H., editors, Quadratic Assignment and Related Problems, volume 16
of DIMACS Series on Discrete Mathematics and Theoretical Computer Science,
pages 43{75. AMS, (Providence, RI).
Ahuja, R. K., Magnanti, T. L., Orlin, J. B., and Reddy, M. R. (1995). Applications
of network optimization. In Ball, M. O., Magnanti, T. L., Monma, C. L., and
Nemhauser, G. L., editors, Network Models - Handbooks of Operations Research
and Management Science 7, pages 1{83. Elsevier, Amsterdam.
Alt, H., Blum, N., Mehlhorn, K., andp Paul, M. (1991). Computing maximum cardi-
1:5
nality matching in time o(n m= log n). Information Process. Letters, 37:237{
240.
Balas, E. and Mazzola, J. B. (1984a). Nonlinear programming: I. linearization
techniques. Mathematical Programming, pages 1{21.
Balas, E. and Mazzola, J. B. (1984b). Nonlinear programming: Ii. dominance rela-
tions and algorithms. Mathematical Programming, 30:22{45.
Balas, E. and Saltzman, M. J. (1991). An algorithm for the three-index assignment
problem. Operations Research, 39:150{161.
Bazaraa, M. S. and Sherali, H. D. (1982). On the use of exact and heuristic cutting
plane methods for the quadratic assignment problem. Journal of Operations
Research Society, 33:991{1003.
Bertsekas, D. P., non, D. A. C., Eckstein, J., and Zenios, S. (1995). Parallel comput-
ing in network optimization. In Ball, M. O., Magnanti, T. L., Monma, C. L.,
and Nemhauser, G. L., editors, Network Models - Handbooks in Operations Re-
search and Management Science 7, pages 330{399. Elsevier, Amsterdam, The
Netherlands.
Birkho, G. (1946). Tres observaciones sobre el algebra lineal. Rev. univ. nac.
Tucuman (A), 5:147{151.
Bruengger, A., Clausen, J., Marzetta, A., and Perregaard, M. (1997). Joining forces
in solving large-scale quadratic assignment problems in parallel. In Proceedings
of the 11-th IEEE International Parallel Processing Symposium (IPPS), pages
418{427.
24
Burkard, R. E. (1974). Quadratische bottleneckprobleme. Operations Research
Verfahren, 18:26{41.
Burkard, R. E. and Cela, E. (1997). Quadratic and three-dimensional assignments.
In Dell'Amico, M., Maoli, F., and Martello, S., editors, Annotated Bibliogra-
phies in combinatorial Optimization, pages 373{392. Wiley, Chichester.
Burkard, R. E. and Cela, E. (1999). Linear assignment problems and extensions. In
Pardalos, P. and Du, D.-Z., editors, Handbook of Combinatorial Optimization
4. Kluwer Academic Publishers. to appear.
Burkard, R. E., Cela, E., and Klinz, B. (1994). On the biquadratic assignment
problem. In Pardalos, P. M. and Wolkowicz, H., editors, Quadratic Assignment
and Related Problems, volume 16 of DIMACS Series on Discrete Mathematics
and Theoretical Computer Science, pages 117{147. AMS, (Providence, RI).
Burkard, R. E., Cela, E., Pardalos, P., and Pitsoulis, L. (1998). The quadratic
assignment problem. In Pardalos, P. and Du, D.-Z., editors, Handbook of Com-
binatorial Optimization 3, pages 241{338. Kluwer Academic Publishers.
Burkard, R. E. and Derigs, U. (1980). Assignment and Matching Problems: Solution
Methods with FORTRAN Programs. Springer, Berlin.
Burkard, R. E. and Fincke, U. (1982). On random quadratic bottleneck assignment
problems. Mathematical Programming, 23:227{232.
Burkard, R. E. and Fincke, U. (1983). The asymptotic probabilistic behavior of
the quadratic sum assignment problem. Zeitschrift fur Operations Research,
27:73{81.
Burkard, R. E. and Fincke, U. (1985). Probabilistic asymptotic properties of some
combinatorial optimization problems. Discrete Applied Mathematics, 12:21{29.
Burkard, R. E., Hahn, W., and Zimmermann, U. (1977). An algebraic approach to
assignment problems. Mathematical Programming, 12:318{327.
Burkard, R. E., Karisch, S. E., and Rendl, F. (1997). Qaplib - a quadratic
assignment problem library. Journal of Global Optimization, 10:391{403.
An on-line version is available via World Wide Web at the following URL:
http://www.opt.math.tu-graz.ac.at/ekarisch/qaplib/.

Burkard, R. E. and Rendl, F. (1984). A thermodynamically motivated simulation

procedure for combinatorial optimization problems. European Journal of Op-
erational Research, 17:169{174.

25
Burkard, R. E. and Rendl, F. (1991). Lexicographic bottleneck problems. Operations
Research Letters, 10:303{308.
Burkard, R. E. and Zimmermann, U. (1982). Combinatorial optimization in lin-
early ordered semimodules: a survey. In Korte, B., editor, Modern Applied
Mathematics, pages 392{436. North Holland, Amsterdam.
Carpaneto, G., Martello, S., and Toth, P. (1988). Algorithms and codes for the
assignment problem. Annals of Operations Research, 13:193{223.
Carpaneto, G. and Toth, P. (1980). Solution of the assignment problem. ACM
Transactions on Mathematical Software, 6:104{111.
Carraresi, P. and Malucelli, F. (1994). A reformulation scheme and new lower bounds
for the qap. In Pardalos, P. and Wolkowicz, H., editors, Quadratic Assignment
and Related Problems, volume 16 of DIMACS Series in Discrete Mathematics
and Theoretical Computer Science, pages 147{160. AMS, Providence, RI.
Cela, E. (1998). The Quadratic Assignment Problem: Theory and Algorithms.
Kluwer Academic Publishers, Dordrecht.
Clausen, J. and Perregaard, M. (1997). Solving large quadratic assignment problems
in parallel. Computational Optimization and Applications, 8:111{127.
Connolly, D. T. (1990). An improved annealing scheme for the qap. European
Journal of Operational Research, 46:93{100.
Cunningham, W. H. (1976). A network simplex method. Mathematical Program-
ming, 11:105{116.
Dell'Amico, M. and Martello, S. (1997). Linear assignment. In Dell'Amico, M., Maf-
oli, F., and Martello, S., editors, Annotated Bibliographies in Combinatorial
Optimization, pages 355{372. Wiley, Chichester.
Dell'Amico, M. and Toth, P. (2000). Algorithms and codes for dense assignment
problems. To appear.
Derigs, U. (1985). The shortest augmenting path for solving assignment problems -
motivation and computational experience.
Dyer, M. E., Frieze, A. M., and McDiarmid, C. J. H. (1986). On linear programs
with random costs. Mathematical Programming, 35:3{16.
26
Espersen, T., Karisch, S. E., Cela, E., and Clausen, J. QAPPACK- a java package for
solving quadratic assignment problems. working paper, Department of Math-
ematical Modelling, Technical University of Denmark, Denmark, and Institute
of Mathematics, Technical University Graz, Austria.
Euler, R. and Le Verge, H. (1996). Time-tables, polyhedra and the greedy algorithm.
Discrete Applied Mathematics, 65:207{221.
Even, S. and Tarjan, R. E. (1975). Network ow and testing graph connectivity.
SIAM Journal on Computing, 4:507{518.
Finke, G., Burkard, R. E., and Rendl, F. (1987). Quadratic assignment problems.
Annals of Discrete Mathematics, 31:61{82.
Frenk, J. B. G., van Houweninge, M., and Kan, A. H. G. R. (1985). Asymptotic
properties of the quadratic assignment problem. Mathematics of Operations
Research, 10:100{116.
Frenk, J. B. G., van Houweninge, M., and Rinnooy Kan, A. H. G. (1987). Order
statistics and the linear assignment problem. Computing, bf 39:165{174.
Frieze, A. M. and Yadegar, J. (1983). On the quadratic assignment problem. Discrete
Applied Mathematics, 5:89{98.
Fulkerson, R., Glicksberg, I., and Gross, O. (1953). A production line assignment
problem. Technical Report RM{1102, The Rand Corporation, Sta. Monica, CA.
Gilmore, P. C. (1962). Optimal and suboptimal algorithms for the quadratic assign-
ment problem. SIAM Journal on Applied Mathematics, 10:305{313.
Goldberg, A. V. and Kennedy, R. (1995). An ecient cost scaling algorithm for the
assignment problem. Mathematical Programming, 75:153{177.
Hahn, P. and Grant, T. (1998). Lower bounds for the quadratic assignment problem
based upon a dual formulation. Operations Research, 46:912{922.
Hahn, P., Grant, T., and Hall, N. (1998). Solution of the quadratic assignment prob-
lem using the hungarian method. European Journal of Operational Research,
108:629{640.
Hopcroft, J. E. and Karp, R. M. (1973). An n 2 algorithm for maximum matchings
5

in bipartite graphs. SIAM Journal on Computing, 2:225{231.

27
Johnson, D. S. and McGeoch, C. C., editors (1993). Network Flows and Matching
- First DIMACS Implementation Challenge, volume 12 of DIMACS Series in
Discrete Mathematics and Theoretical Computer Science. AMS, Providence,
RI.
Jonker, R. and Volgenant, A. (1986). Improving the hungarian assignment algo-
rithm. Operations Research Letters, 5:171{175.
Kaibel, V. (1997). Polyhedral Combinatorics of the Quadratic Assignment Problem.
PhD thesis, Universitat zu Koln, Germany.
Karisch, S. E., Cela, E., Clausen, J., and Espersen, T. (1999). A dual framework
for lower bounds of the quadratic assignment problem based on linearization.
Computing, 63:351{403.
Karp, R. M. (1980). An algorithm to solve the m  n assignment problem in expected
time o(mn log n). Networks, 10:143{152.
Karp, R. M. (1987). An upper bound on the expected cost of an optimal assignment.
In Discrete Algorithms and Complexity, pages 1{4. Academic Press, Boston.
Karp, R. M., Kan, A. H. G. R., and Vohra, R. V. (1994). Average case analysis of
a heuristic for the assignment problem. Mathematics of Operations Research,
19:513{522.
Kaufmann, L. and Broeckx, F. (1978). An algorithm for the quadratic assignment
problem using benders' decomposition. European Journal of Operational Re-
search, 2:204{211.
Laporte, G. and Mercure, H. (1988). Balancing hydraulic turbine runners: A
quadratic assignment problem. European Journal of Operational Research,
35:378{382.
Lawler, E. L. (1963). The quadratic assignment problem. Management Science,
9:586{599.
Lee, Y. and Orlin, J. B. (1994). On very large scale assignment problems. In Hager,
W. W., Hearn, D. W., and Pardalos, P. M., editors, Large Scale Optimization:
State of the Art , pages 206{244. Kluwer Academic Publishers, Dordrecht, The
Netherlands.
Magos, D. (1996). Tabu search for the planar three-index assignment problem.
Journal of Global Optimization, 8:35{48.

28
Magos, D. and Miliotis, P. (1994). An algorithm for the planar three-index assign-
ment problem. European Journal of Operational Research, 77:141{153.
Martello, S., Pulleyblank, W. R., Toth, P., and de Werra, D. (1984). Balanced
optimization problems. Operations Research Letters, 3:275{278.
Mulmuley, K., Vazirani, U. V., and Vazirani, V. V. (1087). Matching is as easy as
matrix inversion. Combinatorica, 7:105{113.
Nugent, C. E., Vollmann, T. E., and Ruml, J. (1969). An experimental comparison
of techniques for the assignment of facilities to locations. Journal of Operations
Research, 16:150{173.
Olin, B. (1992). Asymptotic Properties of Random Assignment Problems. PhD the-
sis, Division of Optimization and Systems Theory, Department of Mathematics,
Royal Institute of Technology, Stockholm.
Padberg, M. W. and Rijal, M. P. (1996). Location, Scheduling, Design and Integer
Programming. Kluwer Academic Publishers, Boston.
Pardalos, P. and Crouse, J. (1989). A parallel algorithm for the quadratic assignment
problem. In Proceedings of the Supercomputing Conference 1989, pages 351{360.
ACM Press.
Pardalos, P. M., Pitsoulis, L. S., and Resende, M. G. C. (1997). Fortran subroutines
for approximate solution of sparse quadratic assignment problems using grasp.
ACM Transcations on Mathematical Software, 23:196{208.
Pferschy, U. (1996). The random linear bottleneck assignment problem. RAIRO
Operations Research, 30:127{142.
Pferschy, U. (1997). Solution methods and computational investigations for the
linear bottleneck assignment problem. Computing, 59:237{258.
Pierskalla, W. P. (1967). The tri-substitution method for the three-multidimensional
assignment problem. Canadian ORS Journal, 5:71{81.
Punnen, A. P. and Nair, K. P. K. (1994). Improved complexity bound for the
maximum cardinality bottleneck bipartite matching problem. Discrete Applied
Mathematics, 55:91{93.
Ramakrishnan, K. G., Karmarkar, N. K., and Kamath, A. P. (1993). An approximate
dual projective algorithm for solving assignment problems. In Johnson, D. S.
29
 and McGeoch, C. C., editors, Network ows and matching{ First DIMACS Im-
plementation Challenge, volume 12 of DIMACS Series in Discrete Mathematics
and Theoretical Computer Science, pages 431{449. AMS, Providence, RI.
Resende, M. G. C., Pardalos, P. M., and Li, Y. (1996). Fortran subroutines for
approximate solution of dense quadratic assignment problems using grasp. ACM
Transcations on Mathematical Software, 22:104{118.
Resende, M. G. C., Ramakrishnan, K. G., and Drezner, Z. (1995). Computing lower
bounds for the quadratic assignment problem with an interior point algorithm
for linear programming. Operations Research, 43:781{791.
Steinberg, L. (1961). The backboard wiring problem: A placement algorithm. SIAM
Review, 3:37{50.
Szpankowski, W. (1995). Combinatorial optimization problems for which almost
every algorithm is asymptotically optimal! Optimization, 33:359{367.
Taillard, E. (1991). Robust tabu search for the quadratic assignment problem.
Parallel Computing, 17:443{455.
Vlach, M. (1967). Branch and bound method for the three-index assignment prob-
lem. Ekonomicko-Matematicky Obzor, 12:181{191.
West, D. H. (1983). Algorithm 608: Approximate solution of the quadratic assign-
ment problem. ACM Transactions on Mathematical Software, 9:461{466.

30