European Journal of Operational Research 153 (2004) 92–101

www.elsevier.com/locate/dsw

Global constraints for round robin tournament scheduling

a,*
Martin Henz uller b, Sven Thiel
, Tobias M€ c

a
School of Computing, National University of Singapore, Singapore 117543, Singapore
b
Programming Systems Lab, Saarland University, 66041 Saarbr€ucken, Germany
c
Max-Planck-Institut f€ur Informatik, Im Stadtwald, 66123 Saarbr€ucken, Germany

Abstract

In the presence of side-constraints and optimization criteria, round robin tournament problems are hard combi-
natorial problems, commonly tackled with tree search and branch-and-bound optimization. Recent results indicate that
constraint-based tree search has crucial advantages over integer programming-based tree search for this problem do-
main by exploiting global constraint propagation algorithms during search. In this paper, we analyze arc-consistent
propagation algorithms for the global constraints ‘‘all-different’’ and ‘‘one-factor’’ in the domain of round robin
tournaments. The best propagation algorithms allow us to compute all feasible perfectly mirrored pattern sets with
minimal breaks for intermural tournaments of realistic size, and to improve known lower bounds for intramural
tournaments balanced with respect to carry-over effects.

2003 Published by Elsevier B.V.

Keywords: Timetabling; Constraints satisfaction; Graph theory

1. Introduction field are constructive in a sense that interesting

properties of tournaments are identified and
In round robin sport competitions, each team then––by employing graph-theoretical and com-
plays each other team a fixed number of times binatorial arguments––methods to construct cor-
and the matches are organized in rounds. Round responding tournaments are described.
robin schedules can be characterized as one- This works well for highly regular tournaments.
factorizations of complete graphs and are studied However, in the presence of irregular constraints
in graph theory and combinatorial design. Nu- which occur in tournament planning practice and
merous results have been obtained on variants of which are difficult to capture as properties of
the round robin scheduling problem, including graphs, constructive methods fail and the problem
intermural tournaments, facility-sharing tourna- degenerates to a combinatorial search problem.
ments and bipartite tournaments; extensive refer- Techniques that have been used to solve such
ences are given in [2,23]. The techniques in this problems include integer programming [15,20], lo-
cal search [24] and constraint programming
[7,11,19]. Constraint programming has been shown
*
Corresponding author. recently to outperform integer programming on
E-mail address: [email protected] (M. Henz). practical problems [7,8]. Constraint programming

0377-2217/$ - see front matter
2003 Published by Elsevier B.V.

doi:10.1016/S0377-2217(03)00101-2
 M. Henz et al. / European Journal of Operational Research 153 (2004) 92–101 93

allows to systematically exploit the round robin bound for intramural tournaments balanced with
and other constraints, often leading to relatively respect to carry-over effects.
small search trees for medium-sized tournaments.
To solve large problems, stronger propagation
algorithms for pruning the search trees become 2. Round robin tournaments and graphs
important. In finite domain programming systems
such as Ilog Solver [10] and Mozart [13], where In round robin sport competitions, each team
constraints are encoded as propagators that op- plays each other team a fixed number of times g
erate on a constraint store which stores domains of during the competition. Let us first assume g ¼ 1,
variables, arc-consistency is the strongest kind of thus we are dealing with single round robin tour-
propagation that can be achieved for a given naments (SRR). A temporally dense single round
constraint. The most important constraints for robin (DSRR) for n teams distributes the nðn
1Þ
round robin tournaments are matches over a minimal number of rounds such
that every team plays at most one match per
• the all-different constraint, which expresses that round. If n is even, the number of rounds is n
1.
the rows in a tournament contain every team A DSRR with an odd number of teams consists of
only once, and n rounds in each of which n
1 teams play and
• the one-factor constraint, called symmetric all- one team does not. This team is said to have a bye.
different by R
egin [17], which expresses that In the following, we are limiting ourselves to an
every column groups the teams into matches. even number of teams, since the problem for an
odd number of teams n
1 can be reduced to the n
R
egin gives arc-consistent propagation algo- case by introducing an additional team that always
rithms for both problems, which we will review in ‘‘plays’’ against the team with a bye. When g ¼ 2,
Section 5, after introducing graph-theoretical no- we speak of a dense double round robin (DDRR).
tation in Section 2, presenting the basic ideas of
constraint programming in Section 3, and giving a Teams Rounds
formal description of the two constraints in Sec-
tion 4. 1 2 3 4 5
A practical consideration in modeling of com- 1 2 4 6 3 5
binatorial search problems using constraint pro- 2 1 3 5 6 4
gramming is the trade-off between the strength of 3 5 2 4 1 6
propagation at each node of the search tree and its 4 6 1 3 5 2
computation time. For example, a naive non-arc- 5 3 6 2 4 1
consistent propagation algorithm for all-different 6 4 5 1 2 3
constraint sometimes outperforms the arc-consis-
tent one; in such cases, the decrease in the size of
the search trees does not outweigh the increase of The planning of a DSRR consists of assigning
time that is spent at each of their nodes. This sit- for each round an opponent team to each team.
uation is common in constraint programming and Often other decisions have to be taken as well,
necessitates an experimental evaluation for a given such as the place in which the matches are carried
application domain. out. For intermural tournaments, this amounts to
An extensive experimental evaluation of prop- the question whether a team plays home or away.
agation algorithms for the round robin tourna- For intramural tournaments, a court may need to
ments is given in Section 6. Using the empirically be selected. We first concentrate on opponent team
best propagation algorithms, we are able to com- assignment and discuss intermural tournaments in
pute all feasible perfectly mirrored pattern sets Section 6.
with minimal breaks for intermural tournaments The single round robin schedule on the right
of realistic size, and to improve a known lower shows a valid assignment of opponent teams for
94 M. Henz et al. / European Journal of Operational Research 153 (2004) 92–101

n ¼ 6 and g ¼ 1 in each round. The value in row t ðsx1 ; . . . ; sxk Þ 2 c holds. The set of all solutions to a
and column r tells the team against which team t constraint problem P is denoted by solðPÞ.
plays in round r. The constraint programming approach to solv-
In order to characterize the mathematical ing combinatorial search problems such as round
properties of round robin schedules, we need to robin scheduling problems works as follows. En-
introduce some terminology on graphs. code the problem as a constraint satisfaction
Let G ¼ ðV ; EÞ be an undirected graph with problem P, find a new problem P0 that has the
vertex set V and edge set E where ðx; xÞ 62 E (no same set of solutions by applying so-called con-
self-loops). The degree of G is the maximal number sistency techniques. Now augment P0 in two ways,
of edges incident to some vertex in G. G is called by adding a new constraint c0 and its negation,
complete, if there is an edge from any vertex to any respectively, to P0 . To the resulting problems
other. A factor of G is a subgraph of G with vertex P0 þ c0 and P0 þ :c0 , apply again consistency
set V . A factorization of G is a set of factors of G techniques, find new constraints c00 for each of the
which are pairwise edge-disjoint and whose union problems, and so on. This process leads to a binary
of edges is E. A set M  E is called a matching in search tree at whose leaves are either problems
G, if no two distinct edges in M share a common that have no solution, or problems where D con-
endpoint. We call a vertex v matched by M if it is tains only singletons, which directly correspond to
incident to some edge in M, and free otherwise. A solutions.
matching M is called perfect if it covers all vertices There are many degrees of freedom in this
of G, i.e. there are no free vertices. A perfect process, including the original encoding of the
matching is also called a one-factor, because it is a problem, the consistency techniques to be applied,
factor with degree 1. A one-factorization of G is a the choice of new constraints at each step and the
factorization of G consisting of one-factors. order in which the resulting search tree is explored.
A one-factorization of the complete graph with The success of constraint programming relies on
n nodes (n being even) corresponds to a DSRR for good choices for all these components. However,
n teams as follows. Every node i represents a team, we are here mainly concerned with consistency
every one-factor represents a round, and an edge techniques. For the other aspects of constraint-
ði; jÞ in a one-factor r fixes a match between teams based round robin scheduling, see [7,8].
i and j in round r. The properties of factorizations A propagation technique is a function that
guarantee that every team plays every other team maps constraint problems P ¼ ðX ; D; CÞ to new
exactly once. This fact is employed in constructive constraint problems P0 ¼ ðX ; D0 ; CÞ, where for
methods for tournament planning (see references every x 2 X , D0x  Dx , and where solðPÞ ¼ solðP0 Þ.
in [23]). A CSP is arc-consistent with respect to the
constraint c on variables x1 ; . . . ; xk , if for each
index i 2 f1; . . . ; kg, and each value v 2 Dxi , there
3. Constraint programming exists an element ðvx1 ; . . . ; vxi
1 ; v; vxiþ1 ; . . . ; vxk Þ 2 c.
A CSP is arc-consistent, if it is arc-consistent with
We represent round robin tournament prob- respect to all of its constraints. Arc-consistent
lems as a constraint satisfaction problem (CSP). A propagation is a propagation technique that turns
CSP is a triple P ¼ ðX ; D; CÞ, where X is a finite a given CSP into an arc-consistent CSP.
set of variables, and D assigns to each variable
x 2 X a finite domain Dx of possible values. Each
element c of C expresses a constraint on some 4. Constraints for round robin scheduling
variables x1 ; . . . ; xk , and thus c  Dx1      Dxk .
A solution s to the constraint problem P The canonical constraint satisfaction problem
assigns to each variable x a value sx 2 Dx such for opponent team assignment in DSRR repre-
that each constraint is satisfied. This means that sents the target timetable by an n  ðn
1Þ matrix
for every constraint c on variables x1 ; . . . ; xk , o of variables, whose variables ot;r tell the opponent
 M. Henz et al. / European Journal of Operational Research 153 (2004) 92–101 95

team against which team t plays in round r. More eqðxi ; j; rj;i Þ for 1 6 j < i 6 m ð7Þ
formally, we define a DSRR problem as
PDSRR ¼ ðX ; D; CÞ, where X contains all variables (the constraint eq reflects the equality of the
in o and Dot;r ¼ f1; . . . ; ng for every team t and first two arguments into the third argument),
round r. The set C contains the following con- (2) arc-consistent propagation with respect to
straints: these constraints, plus arc-consistent propaga-
tion with respect to the redundant constraint
all-differentðot;1 ; . . . ; ot;n
1 Þ;
all-differentðx1 ; . . . ; xm Þ, and
for every t 2 f1; . . . ; ng; ð1Þ (3) arc-consistent propagation with respect to the
and one-factor constraint itself.

one-factorðx1;r ; . . . ; xn;r Þ; The propagation behavior of (1) is strictly

for every r 2 f1; . . . ; n
1g; ð2Þ weaker than arc-consistent propagation for the
one-factor constraint.
where the constraints all-different and one-factor
are defined as follows:
Example 1. For m ¼ 6, let Dx1 ¼ f2; 3; 4; 5; 6g,
all-differentðx1 ; . . . ; xm Þ Dx2 ¼ f1; 2; 3; 6g, Dx3 ¼ f1; 2; 4; 5; 6g, Dx4 ¼ f1;
¼ fðv1 ; . . . ; vm Þ 2 Dx1      Dxm j 3; 6g, Dx5 ¼ f1; 2; 3; 5g, Dx6 ¼ f1; 2; 3; 5g. Arc-
consistency with respect to the neq constraints
8i;j;i6¼j  vi 6¼ vj g; ð3Þ
removes 2 from Dx2 and 5 from Dx5 , and arc-
consistency with respect to the eq constraints
one-factorðx1 ; . . . ; xm Þ
removes 6 from Dx4 , 2 from Dx5 and 5 from Dx6 .
¼ fðv1 ; . . . ; vm Þ 2 Dx1      Dxm j Arc-consistency with respect to the neq and eq
8i;j  vi 6¼ i ^ vi ¼ j $ vj ¼ ig: ð4Þ constraints fails to reach arc-consistency with re-
spect to the one-factor constraint, which further
Propagation algorithms for all-different and one-
removes 2, 3 and 6 from Dx1 , 1 and 3 from Dx2 , 1, 2
factor vary in strength. For the all-different con-
and 6 from Dx3 , and 1 and 3 from Dx6 .
straint, we consider the following two propagation
algorithms:
Adding the redundant constraint all-differ-
entðx1 ; . . . ; xm Þ with arc-consistent propagation, as
(1) splitting the constraint up into mðm
1Þ in-
done in propagation algorithm (2), improves the
equality constraints of the form xi 6¼ xj ac-
propagation behavior in some cases. In Example 1,
cording to its definition in formula (3)
arc-consistent propagation with respect to this
(arc-consistency with respect to such inequali-
constraint and the neq and eq constraints achieves
ties is trivial), and
arc-consistency with respect to the one-factor
(2) arc-consistent propagation with respect to the
constraint. The next example shows that this is not
all-different constraint itself.
always the case.
For the one-factor constraint, we consider three
Example 2. For m ¼ 6, let Dx1 ¼ f2; 3; 6g, Dx2 ¼
propagation algorithms:
f1; 6g, Dx3 ¼ f1; 4; 5g, Dx4 ¼ f3; 5g, Dx5 ¼ f3; 4g,
Dx6 ¼ f1; 2g. Here, arc-consistency with respect to
(1) after introducing a half-matrix of auxiliary
the neq and eq constraints and the all-different
variables ri;j , where 1 6 i < j 6 m, whose do-
constraint is not able to remove any values from
mains are Dri;j ¼ f0; 1g, arc-consistent propa-
any domain, whereas arc-consistency with respect
gation corresponding to the constraints
to the one-factor constraint removes 2 and 6 from
xi 6¼ i for 1 6 i 6 m; ð5Þ Dx1 , 1 from Dx2 , 4 and 5 from Dx3 , 3 from Dx4 , 3
from Dx5 , and 1 from Dx6 , thus fixing the one-fac-
eqðxi ; j; ri;j Þ for 1 6 i < j 6 m; ð6Þ tor 1–3, 2–6 and 4–5.
96 M. Henz et al. / European Journal of Operational Research 153 (2004) 92–101

We conclude from these examples that arc- • Let ðv1 ; . . . ; vm Þ 2 one-factorðx1 ; . . . ; xm Þ denote
consistent propagation for the one-factor con- a solution. The set M ¼ ffxi ; xvi gj1 6 i 6 mg is
straint deserves consideration and––assuming that well-defined, a subset of E and a perfect match-
there is an efficient algorithm for it––has the po- ing in G.
tential for improving round robin scheduling be- • The matching M 0 corresponds to the solution
yond the addition of an arc-consistent redundant ðv01 ; . . . ; v0m Þ of the constraint where v0i is the
all-different constraint. index of the mate of the node xi in the matching
Most previous work on constraint-based tour- M 0.
nament planning [7,8,11,19] used only algorithm 1
for the all-different constraint and algorithm 1 for This observation motivates the following defi-
the one-factor constraint. Trick [22] suggests to use nitions: We call an edge e of G matchable if there is
algorithm 2 for the all-different constraint and al- a perfect matching M in G containing e and un-
gorithm 2 for the one-factor constraint. Our goal is matchable otherwise.
to evaluate the propagation algorithms to achieve
guidelines for using propagation algorithms in Lemma 2 (R
egin [17]). The following algorithm
round robin scheduling. is a propagation technique for arc-consistent prop-
agation with respect to the constraint one-
factorðx1 ; . . . ; xm Þ. Iterate over the domains Dx1 ; . . . ;
5. Propagation algorithms Dxm to enforce constraints
i 62 Dxi for 1 6 i 6 m and
Consider the constraint all-differentðx1 ; . . . ; xm Þ.
The value graph of this constraint is the bipartite i 2 Dxj $ j 2 Dxi for 1 6 i < j 6 m ð8Þ
S
graph G ¼ ðV ; EÞ with V ¼ fx1 ; . . . ; xm g [ i Dxi and construct the variable graph G for the con-
and E ¼ ffxi ; vgjv 2 Dxi g. straint. For every pair of values i and j, such that
fxi ; xj g is not a matchable edge in G, remove i from
Lemma 1 (R
egin [16]). The following algorithm is Dxj and remove j from Dxi .
a propagation technique for arc-consistent propa-
gation with respect to the constraint all-different The problem of finding the unmatchable edges
ðx1 ; . . . ; xm Þ. Construct the value graph of the con- was presented in [1, Exercise 12.42]. It was stated
straint. For every variable x and value v, such that that the problem can be solved with a variation of
fx; vg is not a matchable edge in G, remove v from EdmondsÕ blossom-shrink algorithm [5,6]. R
eginÕs
Dx . algorithm follows these ideas. The running time of
the resulting algorithms is OðmsÞ, where m denotes
R
egin gives an algorithm for identifying non- the number of nodes and s the number of edges
matchable edges in a value graph with m variables of the variable graph. In round robin tournaments,
and k values with complexity Oðkm3=2 Þ. For round m is bounded by the number of teams n and s is
robin tournaments, both k and m are bounded by bounded by n2 , resulting in a complexity of the
the number
pffiffiffi of teams n, resulting in a complexity of propagation algorithm of Oðn3 Þ.
Oðn2 nÞ.
Now, consider the constraint one-factor
ðx1 ; . . . ; xm Þ. The variable graph of this constraint is 6. Experimental evaluation
the graph G ¼ ðV ; EÞ with the vertex set V ¼
fx1 ; . . . ; xm g and the edge set E ¼ ffxi ; xj gjj 2 Dxi g. For the experimental evaluation, we use the
Looking at the definition of the one-factor programming system Mozart 1.1.0 [13] for the
constraint, it is easy to derive a one-to-one corre- concurrent constraint language Oz [21], which
spondence between the solutions of the con- provides extensive support for finite domain con-
straint and the perfect matchings in its variable straint programming. We implemented the prop-
graph G: agation algorithms all-different and one-factor
 M. Henz et al. / European Journal of Operational Research 153 (2004) 92–101 97

described in the previous sections using the LEDA only be solved using arc-consistent propagation
library [12] and made them available to the Mozart for all-different, and thus we fix this propagation
system through MozartÕs Constraint Propagator algorithm in the rest of this section.
Interface [14]. The run times given in this section
are always the average user time of five runs on a 6.1. Unconstrained single round robin tournaments
256 MB 400 MHz Pentium II PC running Linux.
The coefficient of deviation (standard deviation/ Here, we compare the performance of arc-con-
arithmetic mean) was always below 3%. sistent one-factor versus the encoding using neq
The C++ and Oz source code for generating and eq and the redundant all-different constraint.
and running these benchmarks is available at [9]. In this and all following benchmarks, we use
The goals of the experimental evaluation are as constraint-based tree search by using constraints
follows: of the form ot1 ;r ¼ t2 as branching constraints c (see
Section 3), which means that we enumerate the
• evaluate the usefulness of the arc-consistent opponent variables. We order the variables round-
one-factor constraint for round robin applica- wise. At each node, we enumerate the first variable
tions by comparing the sizes of search trees that has a non-singleton domain. The value t2 is
and run times resulting from searches that use always the smallest element of this domain. More
the new constraint with searches that use arc- sophisticated enumeration techniques such as first-
consistent all-different or simply the encoding fail do not improve the search. Table 1 com-
using neq and eq, pares the encodings using neq/eq from Section 4,
• investigate the range of round robin problems Formula (4), the addition of a redundant arc-
to which arc-consistent one-factor provides ad- consistent all-different constraint as discussed in
vantages over other techniques, and Example 1, and the algorithm one-factor, for
• evaluate the efficiency of the arc-consistent one- finding the first DSRR.
factor constraint for practical applications. We observe that for these kinds of benchmarks,
arc-consistent one-factor achieves optimal propa-
We observe from all experiments that the use of gation in a sense that there are no failures in the
the arc-consistent all-different constraint is crucial. search trees, whereas the encoding using neq/eq
It typically leads to a reduction of the size of the requires search. From 26 teams onward, this
search tree and the runtime by one or two orders method fails to produce solutions within reason-
of magnitude. Large tournament problems can able time. Arc-consistent all-different occasionally

Table 1
Benchmarks on unconstrained DSRR
n neq/eq all-different one-factor
F UT F UT F UT
6 0 0.030 0 0.048 0 0.020
10 3 0.428 1 0.558 0 0.220
14 62 3.18 0 4.35 0 1.13
18 20 5.82 0 8.86 0 4.37
22 675 81.3 2 25.0 0 12.7
26 ? ? 2 57.1 0 32.0
30 ? ? 0 122.0 0 71.2
34 ? ? 17 241.0 0 145.0
38 ? ? 5 429.0 0 272.0
42 ? ? 4 742.0 0 484.0
n: number of teams; F : number of failures in the search tree; UT : user time. The ? symbols indicate that 30 minutes of runtime were
exceeded.
98 M. Henz et al. / European Journal of Operational Research 153 (2004) 92–101

requires a bit of search, but the performance dif- the two techniques increases with the problem size.
ference to arc-consistent one-factor is not dramatic Note that for the double round robin problems,
here. the reduction of the search tree afforded by the
redundant all-different constraint does not justify
6.2. Tightly constrained round robin tournaments its computational effort.

The next set of benchmarks looks at tightly 6.3. Minimizing carry-over

constrained problems. We constrain the problems
by randomly forbidding opponent teams, until We consider sports, in which a match between
there are very few or no solutions to the problem. two teams a and b has an impact on the perfor-
Table 2 compares the three remaining propagation mance of these teams in the next round, an effect
algorithms for finding all solutions, or proving called carry-over. In such a sport, each sequence of
unsatisfiability. The problems s  are single round two teams should appear at most once in such a
robin problems as described throughout the paper, schedule, leading to a schedule balanced with re-
whereas the problems d  are double round robin spect to carry-over effects. For example, this is not
problems. For the latter, we have instead of an all- the case for the schedule on page 3; the sequence
different constraint per team (requiring that each (3, 5) appears three times. Since this ideal is not
team plays each other team once) constraints that always achievable, the goal is to minimize the
force the number of occurrences of all other teams carry-over effect using a cost function. Russell [18]
in each row of the o matrix to be 2. There is no gives a constructive method for generating tour-
restriction on the distance between first leg and naments with no carry-over effect where n is a
return match. power of two, and conjectures the non-existence of
Although the results vary considerably, we note such tournaments in all other cases. He gives a
that in these tightly constrained problems, arc- constructive method to generate tournaments with
consistent one-factor results in a reduction of the small carry-over effects leading to a carry-over
size of search trees by a factor of up to 10, and of effect of 60 for n ¼ 6, 138 for n ¼ 10 and 196 for
the runtime by even more. The difference between n ¼ 12 according to a canonical cost measure.

Table 2
Benchmarks on tightly constrained DSRR and DDRR
File n S neq/eq all-different one-factor 2
F UT F UT F UT
s_6_yes 6 4 7 0.088 5 0.124 4 0.060
s_8_yes 8 5 47 0.540 24 0.606 10 0.200
s_10_yes 10 1 37 0.704 17 0.686 1 0.106
s_12_yes 12 1 3216 77.9 1452 64.1 179 6.26
s_14_yes 14 1 8407 242.0 1328 75.3 527 20.4
s_6_no 6 0 4 0.048 4 0.066 4 0.022
s_8_no 8 0 13 0.172 12 0.246 6 0.086
s_10_no 10 0 20 0.530 13 0.646 6 0.168
s_12_no 12 0 241 6.64 111 5.37 25 0.794
s_14_no 14 0 537 16.7 182 10.9 69 2.54
s_16_no 16 0 1467 64.5 213 18.0 86 5.37
s_18_no 18 0 593 38.6 95 9.57 30 2.29
s_20_no 20 0 ? ? 2755 314.0 254 23.0
d_6_yes 6 4 21 0.066 4 0.084 2 0.020
d_8_yes 8 32 5776 19.2 2736 28.1 226 0.93
d_10_yes 10 2 76646 409.0 38251 677.0 5956 35.6
n: number of teams; S: number of solutions; F : number of failures in the search tree; UT : user time.
 M. Henz et al. / European Journal of Operational Research 153 (2004) 92–101 99

With constraint-based branch-and-bound using straints [7]. However, not all such pattern sets lead
arc-consistent propagation for one-factor con- to schedules even if there are no other side con-
straints, we are able to prove the optimality of his straints. De Werra [4] gives a construction of fea-
schedule for n ¼ 6. However, for n ¼ 10, we obtain sible pattern sets with 3n
6 breaks, proves their
a schedule with a better cost (136) after 26.2 sec- optimality and states that no constructive method
onds. The best values that we obtain is 128 for is known to enumerate all feasible pattern sets. In
n ¼ 10 shown on the previous page, which is ob- this situation, it is useful to enumerate feasible
tained using a randomized search strategy (the pattern sets, which––surprisingly––has to our
runtime of about 30 minutes is therefore not very knowledge not been tackled so far. For this task
informative). Trick [22] reports an even better cost we use the constraint model given in [7] and add
(122) after 1 day of runtime, also using constraint the model for opponents given in Section 4. The
programming. For n ¼ 12, we improve the best table on the right gives the number of pattern sets
known schedule (cost 196) given in [18] and for n 6 18; the pattern sets are listed in [9]. The
achieve a schedule with a cost of 188, which is column P gives the number of pattern sets that
given in [9]. fulfill the pattern set constraints, the column F
gives the number of pattern sets that fulfill the
Teams Rounds model for opponents. For all n 6 16, there exists a
1 2 3 4 5 6 7 8 9 schedule for every computed pattern set. For
n ¼ 18, there are four cases, for which we could
1 6 5 10 9 4 8 7 2 3 neither prove infeasibility nor generate a schedule.
2 7 4 9 10 6 3 8 1 5
3 8 9 7 6 5 2 4 10 1
4 9 2 8 7 1 10 3 5 6 n P F
5 10 1 6 8 3 7 9 4 2
6 1 7 5 3 2 9 10 8 4 4 0 0
7 2 6 3 4 10 5 1 9 8 6 1 1
8 3 10 4 5 9 1 2 6 7 8 4 2
9 4 3 2 1 8 6 5 7 10 10 15 4
10 5 8 1 2 7 4 6 3 9 12 56 10
14 210 17
Small carry-over schedule for n ¼ 10. 16 792 46
In this study, the use of arc-consistent propa- 18 3003 84
gation for the one-factor constraint is again cru-
cial.
For n 6 16 the use of the arc-consistent one-
factor constraint or the redundant arc-consistent
6.4. Feasible pattern sets
all-different constraint for one-factor was not
crucial and the pattern sets were obtained faster
Here, we consider intermural dense double
with trivial propagation for the one-factor con-
round robins (DDRR) ðg ¼ 2Þ, where the second
straint. For n ¼ 18, the arc-consistent propagation
part of the double round robin repeats the first
for the one-factor constraint allows to prove the
part with opposite venues. A team is said to have a
infeasibility of 17 pattern sets; without it, the
break, if it either plays two consecutive matches
number of pattern sets generated is 101.
home or away. We consider the problem of finding
all schedules that minimize the overall number of
breaks. A widely accepted method of searching for 6.5. Intermural tournaments
intermural tournaments is to generate pattern sets
first [7,8,15,20]. These are sets of home/away pat- The intermural benchmarks in this section show
terns that satisfy simple row and column con- that the pruning obtained from arc-consistent
100 M. Henz et al. / European Journal of Operational Research 153 (2004) 92–101

Table 3
Benchmarks on intermural tournaments
File n S neq/eq all-different one-factor 2
F UT F UT F UT
i_8_yes 8 7 0 0.138 0 0.182 0 0.142
i_12_yes 12 3 0 0.440 0 0.510 0 0.436
i_16_yes 16 4 6 1.30 4 1.51 4 1.22
i_20_yes 20 10 35 4.23 22 4.63 22 3.61
acc97/98 9 179 273 19.5 273 28.4 268 22.1
n: number of teams; S: number of solutions; F : number of failures in the search tree; UT : user time.

one-factor not always outweighs its computational different constraint is crucial for efficient solution
effort. The best strategy for intermural tournament of all tournament scheduling problems considered.
problems is to first find so-called pattern sets [3]. Arc-consistent propagation for the one-factor
Table 3 shows the run times for finding all solu- constraint is essential for intramural tournaments.
tions of intermural tournament problems, all of For large unconstrained and tightly constrained
which except the last one are randomly con- single and multiple round robin tournaments, we
strained as in the previous section. The last prob- observe a typical reduction of the search tree and
lem is the ACC 1997/98 problem [8,15], which is runtime by one order of magnitude. Intermural
tightly constrained by a variety of conditions. tournaments do not benefit much from the arc-
Since we are not concerned with the computation consistent one-factor constraint.
of the pattern sets, we fix a particular pattern set Using these algorithms, we could establish new
for the benchmarks, except for the ACC 1997/98 lower bounds for the minimization of carry-over
problem. The numbers for this last problem in- effects for intramural single round robin tourna-
clude the effort for generating pattern sets, since ments and enumerate the all feasible pattern sets
the one-factor constraint already achieves some for intramural tournaments with a minimal num-
additional pruning during pattern set generation. ber of breaks for up to 16 teams. For 18 teams,
We observe that the effort for neither arc-con- there are four open cases.
sistent one-factor nor arc-consistent all-different is
justified for these intermural tournaments. The
Acknowledgements
reason is that pattern sets already enforce one-
factor to such an extent that arc-consistent one-
The authors would like to thank Ernst Althaus,
factor achieves almost no additional pruning of
Kurt Mehlhorn and Tony Tan for many helpful
the search trees. Note that the arc-consistent one-
discussions, Jan A.M. Schreuder for providing the
factor constraint could be sped up in this case by
motivation for the work on pattern sets, Michael
exploiting that the variable graph is pffiffiffibipartite, Trick for pointing out omissions and mistakes,
which would lead a complexity of Oðn2 nÞ, similar
Marleen van Brandenburg for carrying out initial
to the arc-consistent all-different.
experiments on round robin tournament bench-
marks, and the reviewers for pointing out the
potential of the optimization for intermural tour-
7. Conclusion naments.

We analyzed the use of the global constraints

References
all-different and one-factor for constraint-based
search for round robin tournament schedules. We [1] R.K. Ahuja, T.L. Magnanti, J.B. Orlin, Network Flows:
conclude from an extensive experimental evalua- Theory, Algorithms and Applications, Prentice Hall,
tion that arc-consistent propagation for the all- Englewood Cliffs, NJ, 1993.
 M. Henz et al. / European Journal of Operational Research 153 (2004) 92–101 101

[2] I. Anderson, Combinatorial Designs and Tournaments, arbr€ucken, Swedish Institute of Computer Science, Stock-
Oxford University Press, 1997. holm, and Universit
e Catholique de Louvain, 1999.
[3] W.O. Cain Jr., The computer-assisted heuristic approach [14] T. M€ uller, Constraint extensions tutorial, http://
used to schedule the major league baseball clubs, in: S.P. www.mozart-oz.org, Programming Systems Lab, Sa-
Ladany, R.E. Machol (Eds.), Optimal Strategies in Sports, arbr€ucken, Swedish Institute of Computer Science, Stock-
number 5 in Studies in Management Science and Systems, holm, and Universit
e Catholique de Louvain, 1999.
North-Holland, Amsterdam, 1977, pp. 32–41. [15] G.L. Nemhauser, M.A. Trick, Scheduling a major college
[4] D. de Werra, Scheduling in sports, in: P. Hansen (Ed.), basketball conference, Operations Research 46 (1) (1998)
Studies on Graphs and Discrete Programming, North- 1–8.
Holland, Amsterdam, 1980, pp. 381–395. [16] J.-C. R
egin, A filtering algorithm for constraints of
[5] J. Edmonds, Maximum matching and a polyhedron with 0, difference in CSPs, in: Proceedings of the AAAI 12th
1-vertices, Journal of Research of the National Bureau of National Conference on Artificial Intelligence, AAAI
Standards 69B (1965) 125–130. Press, 1994, pp. 362–367.
[6] J. Edmonds, Paths, trees, and flowers, Canadian Journal [17] J.-C. R
egin, The symmetric alldiff constraint, in: T. Dean
on Mathematics 23 (1965) 449–467. (Ed.), Proceedings of the International Joint Conference on
[7] M. Henz, Constraint-based round robin tournament plan- Artificial Intelligence, Stockholm, Sweden, vol. 1, Morgan
ning, in: D. De Schreye (Ed.), Proceedings of the Interna- Kaufmann Publishers, San Mateo, CA, 1999, pp. 420–
tional Conference on Logic Programming, Las Cruces, 425.
New Mexico, The MIT Press, Cambridge, MA, 1999, pp. [18] K.G. Russell, Balancing carry-over effects in round robin
545–557. tournaments, Biometrika 67 (1) (1980) 127–131.
[8] M. Henz, Scheduling a major college basketball confer- [19] A. Schaerf, Scheduling sport tournaments using constraint
ence––revisited, Operations Research 49 (1) (2001). logic programming, Constraints 4 (1) (1999) 43–65.
[9] M. Henz, T. M€ uller, S. Thiel, M. van Brandenburg, Bench- [20] J.A.M. Schreuder, Combinatorial aspects of construction
marks and results for round robin tournaments, http:// of competition Dutch professional football leagues, Dis-
www.comp.nus.edu.sg/~henz/roundrobin_ crete Applied Mathematics 35 (1992) 301–312.
benchmarks, 2000. [21] G. Smolka, The Oz programming model, in: J. van
[10] ILOG Inc., Mountain View, CA 94043, USA, http:// Leeuwen (Ed.), Computer Science Today, Lecture Notes
www.ilog.com, ILOG Solver 4.0, Reference Manual, in Computer Science 1000, Springer-Verlag, Berlin, 1995,
1997. pp. 324–343.
[11] K. McAloon, C. Tretkoff, G. Wetzel, Sports league [22] M.A. Trick, A schedule-then-break approach to sports
scheduling, in: Proceedings of the 1997 ILOG Optimiza- timetabling, in: E. Burke, W. Erben (Eds.), Practice and
tion Suite International UsersÕ Conference, Paris, July Theory of Automated Timetabling III, Selected Papers of
1997. PATAT 2000, Konstanz, Germany, Lecture Notes in
[12] K. Mehlhorn, S. N€aher, LEDA: A Platform for Combi- Computer Science 2079, Springer-Verlag, Berlin, 2000.
natorial and Geometric Computing, Cambridge University [23] W.D. Wallis, One-Factorizations, Kluwer Academic Pub-
Press, Cambridge, 1999. lishers, Dordrecht, 1997.
[13] Mozart Consortium, The Mozart Programming System. [24] J.P. Walser, Domain-Independent Local Search for Linear
Documentation and system available from http:// Integer Optimization, PhD thesis, Universit€at des Saarlan-
www.mozart-oz.org, Programming Systems Lab, Sa- des, Saarbr€ ucken, August 1998.