See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/221204639

Solving the Sports League Scheduling Problem with Tabu Search

Conference Paper · August 2000

DOI: 10.1007/3-540-45612-0_2 · Source: DBLP

CITATIONS READS

41 2,255

2 authors, including:

Jin-Kao Hao
University of Angers
493 PUBLICATIONS 11,870 CITATIONS

SEE PROFILE

All content following this page was uploaded by Jin-Kao Hao on 20 May 2014.

The user has requested enhancement of the downloaded file.

 Solving the Sports League Scheduling Problem with
Tabu Search

1 2
Jean-Philippe Hamiez and Jin-Kao Hao
1
LGI2P / Ecole des Mines d'Alès – EERIE
Parc Scientifique Georges Besse – 30035 Nîmes Cedex 01 (France)
[email protected]
2
LERIA / Université d’Angers – 2, Bd. Lavoisier – 49045 Angers Cedex 01 (France)
[email protected]

Abstract. In this paper we present a tabu approach for a version of the Sports
League Scheduling Problem. The approach adopted is based on a formulation
of the problem as a Constraint Satisfaction Problem (CSP). Tests were carried
out on problem instances of up to 40 teams representing 780 integer variables
with 780 values per variable. Experimental results show that this approach
outperforms some existing methods and is one of the most promising methods
for solving problems of this type.

1 Introduction

Many sports leagues (e.g. soccer, hockey, basketball) must deal with scheduling
problems for tournaments. These scheduling problems contain in general many
conflicting constraints to satisfy and different objectives to optimize, like
minimization of traveling distance [2], only one match per team and per day,
stadium unavailability at particular dates, minimum number of days between a home
match and its corresponding away match, etc. Generating satisfactory schedules
with respect to these conditions and objectives is therefore a very difficult problem
to solve.
Many studies have been carried out to try to solve these problems with a variety
of approaches and varying degrees of success: integer linear programming [7][12],
constraint programming [10][17], local search (simulated annealing [19], tabu
search [23], hybrid approach [4]).
This paper deals with a specific Sports League Scheduling Problem (SLSP)
described by K. McAloon, C. Tretkoff and G. Wetzel in [11]. After having obtained
poor results in integer linear programming tests, they experimented with constraint
programming, an approach that produced better results. Finally, with a basic local
search algorithm, they produced the same results as ILOG Solver does, but with less
computing time.
C.P. Gomes, B. Selman and H.A. Kautz [9] obtained better results than those of
McAloon et al. using constraint programming. With a randomized version of a
deterministic complete search they solved problems involving a greater number of
teams.
 J.C. Régin proposed two approaches with constraint programming for the SLSP
[15][16]. The first one, using powerful filtering algorithms [3][13][14], produced
better results than those of McAloon et al. and those of Gomes et al. in terms of
execution time and robustness, since it solved problem instances of a larger size.
Using a second approach, he transformed the SLSP into an equivalent problem by
adding an implicit constraint. With a new heuristic and specific filtering algorithms,
he improved on his own results.
Finally, let us mention the work done by G. Wetzel and F. Zabatta [22]: using
multiple threads on a 14 processor Sun system they obtained better results than the
first approach of Régin. They were, however, not able to solve problems as large as
those which Régin solved with his second approach.
The goal of this study is to propose an advanced local search approach based on
tabu search (TS) [8] for the SLSP. References of the study are results presented in
[11], [15] and [16].
The paper begins by formally describing the SLSP (§2). After modeling the
problem as a constraint satisfaction problem (CSP) [20] (§3), we present our tabu
algorithm (§4) and compare its results with those of [11], [15] and [16] (§5). Before
concluding, we discuss some observations made during this work (§6).

2 Problem Description

In the rest of the paper, we will deal with the following constraints and definitions,
the same as in [11]:
• There are |T| teams (|T| even), where T is the set of all teams. All teams play each
other exactly once (half competition);
• The season lasts |T| - 1 weeks;
• Every team plays one game in every week of the season;
• There are |T| / 2 periods and, each week, one game is scheduled in every period;
• No team plays more than twice in the same period over the course of the season.
The problem then is to schedule the tournament with respect to all these
constraints.
Table 1 below shows an example of a valid schedule for |T| = 8 teams labeled
from 1 to 8; there are 7 weeks and 4 periods.

Table 1. Example of a valid schedule for 8 teams

Weeks
1 2 3 4 5 6 7
1 1 vs 2 1 vs 3 5 vs 8 4 vs 7 4 vs 8 2 vs 6 3 vs 5
Periods

2 3 vs 4 2 vs 8 1 vs 4 6 vs 8 2 vs 5 1 vs 7 6 vs 7
3 5 vs 6 4 vs 6 2 vs 7 1 vs 5 3 vs 7 3 vs 8 1 vs 8
4 7 vs 8 5 vs 7 3 vs 6 2 vs 3 1 vs 6 4 vs 5 2 vs 4
As shown in Table 1, a configuration may be represented as a two-dimensional
array with weeks in columns and periods in rows. Each column satisfies the
cardinality constraint: each team appears exactly once, i.e. all the teams are
different. In each row, no team appears more than twice. There is also a global
constraint on the array: each match only appears once, i.e. all matches are different.

3 Problem Formulation

To represent the SLSP we follow the constraint programming approach: we consider

it as a constraint satisfaction problem.

3.1 Constraint Satisfaction Problem - CSP

A constraint satisfaction problem [20] is defined by a triplet (X, D, C) with:
• A finite set X of n variables: X = {X1, ... , Xn};
• A set D of associated domains: D = {D1, ... , Dn}. Each domain Di specifies the
finite set of possible values of the variable Xi;
• A finite set C of p constraints: C = {C1, ... , Cp}. Each constraint is defined for a
set of variables and specifies which combinations of values are compatible for
these variables.
Given such a triplet, the problem is to generate a complete assignment of the
values to the variables, which satisfies all the constraints: such an assignment is said
to be consistent. Since the set of all assignments (not necessarily consistent) is
defined by the Cartesian product D1 × ... × Dn of the domains, solving a CSP means
to determine a particular assignment among a potentially huge number of possible
assignments.
The CSP, as it has been formalized, is a powerful and general model. In fact, it
can be used to conveniently model some well-known problems such as k-coloring
and satisfiability, as well as many practical applications relating to resource
assignment, planning or timetabling.

3.2 Formulation of the SLSP as a CSP

We will use the following notations to represent the SLSP as a constraint
satisfaction problem:
• P: set of periods, |P| = |T| / 2;
• W: set of weeks, |W| = |T| - 1;
• tn: team number n, tn ∈ T, 1 ≤ n ≤ |T|;
• x(tm, tn): schedule of the match tm vs. tn. Values of this variable type are of
(pm, n, wm, n) pattern, meaning that the match is scheduled in period pm, n and week
wm, n.
The set of variables is naturally X = {x(tm, tn), 1 ≤ m < n ≤ |T|} and all domains
are equal to D = {(pm, n, wm, n), pm, n ∈ P, 1 ≤ pm, n ≤ |P|, wm, n ∈ W, 1 ≤ wm, n ≤ |W|};
∀ x ∈ X, Dx = D. The set C of constraints contains the following three types of
constraints:
• Uniqueness of all teams in each week. For each team tn, tn ∈ T, 1 ≤ n ≤ |T|, we
impose the constraint: WEEK(tn) ⇔ wm, n ≠ wq, n, ∀ (m, q) ∈ [1; |T|] , m ≠ n, q ≠ n
2

and m ≠ q;
• No more than two matches for each team in the same period. For each team tn,
tn ∈ T, 1 ≤ n ≤ |T| and each period, we impose the constraint:
PERIOD(tn) ⇔ |{pm, n = pq, n, (m, q) ∈ [1; |T|] , m ≠ n, q ≠ n et m ≠ q}| ≤ 1;
2

• All matches are different. For each match < tm, tn >, (tm, tn) ∈ T², tm ≠ tn, we
impose the constraint: ALLDIFF(< tm, tn >) ⇔ < tm, tn > ≠ < tq, tr >, ∀ (tq, tr) ∈ T²,
t q ≠ t r.
The WEEK and ALLDIFF constraints are always satisfied in our algorithm.

4 Solving the SLSP with Tabu Search

Tabu search is an advanced local search method using general mechanisms and rules
as guidelines for smart search. Readers may find formal description of the method
in [8]. We now define the components of our tabu algorithm for the SLSP, called
TS-SLSP.

4.1 Search Space – Configuration

As represented in Table 1, a configuration is a complete assignment of D = {(pm, n,
wm, n), pm, n ∈ P, 1 ≤ pm, n ≤ |P|, wm, n ∈ W, 1 ≤ wm, n ≤ |W|} items to variables of
X = {x(tm, tn), 1 ≤ m < n ≤ |T|}. Thus, a configuration is a |W| * |P| sized table, whose
items are integer couples (m, n), 1 ≤ m < n ≤ |T|. For |T| = 40 teams, this represents a
problem with 780 variables and 780 values per variable.
There are |T| / 2 * (|T| - 1) matches to be scheduled. A schedule can be thought of
as a permutation of these matches. So, for |T| teams, the search space size is
[|T| / 2 * (|T| - 1)]! In other words, the search space size grows as the factorial of the
square of |T| / 2.

Fig. 1. Constructing initial configuration (|T| = 6)

4.2 Initial Configuration
In local search algorithms, the initial configuration specifies where the search begins
in the search space. One can start with a random configuration by first creating all
the |T| / 2 * (|T| - 1) matches with respect to the ALLDIFF constraint, then randomly
assigning a match to each (w, p) couple, w ∈ W, p ∈ P.
We chose another way to build an initial configuration, inspired from [18]:
construct a complete graph with the first |T| - 1 teams as vertices while placing the
vertices in order to form a regular (|T| - 1)-gon (edges represent matches). Color the
edges around the boundary using a different color for each edge. The remaining
edges can be colored by assigning to each one the same color as that used for the
boundary edge parallel to it, see left drawing in Fig. 1. At each vertex there will be
exactly one color missing and these missing colors are different. The edge of the
complete graph incident to the last vertex (i.e. the last team) can be colored using
these missing colors, see right drawing in Fig. 1. Finally, fill in week “i”, in the
initial configuration, with edges colored “i”.
This initial configuration has the property of satisfying the WEEK and ALLDIFF
constraints. The algorithm will try to satisfy the PERIOD constraint.

4.3 Neighborhood
Let s be a configuration from the search space S. The neighborhood N: S → 2 is an
s

application such as for each s ∈ S, s' ∈ N(s) if and only if s and s' only differ by
values of a single couple of variables, with at least one in conflict (a variable is said
to be conflicting if it is involved in an unsatisfied constraint).
A neighboring configuration of s can be obtained by making a single swap of
current values of two arbitrary variables, with at least one conflicting in s. A move
from configuration s to a neighboring configuration s' can then be described by a
couple < x(tm, tn), x(tq, tr) > i.e. swapping two matches (see Fig. 2). In addition, swaps
are made in the same week to keep the WEEK constraint satisfied.
During the search the neighborhood size always evolves with the number of
conflicts.

Weeks Weeks
1 2 3 4 5 1 2 3 4 5
1,2 2,6 3,4 5,6 2,4 1 1 1,2 2,6 3,4 5,6 1,5
Periods

4,6 1,3 2,5 1,4 3,6 2 2 4,6 1,3 2,5 1,4 3,6
3,5 4,5 1,6 2,3 1,5 3 3 3,5 4,5 1,6 2,3 2,4
Configuration s Configuration s'
Fig. 2. Neighborhood illustration (6 teams)
4.4 Evaluation Function
To compare, in terms of quality, two configurations s and s’ from S, we define an
evaluation function which is an order relation for S.
Let OccPS(p, tn) be the occurrence number of team tn, in period p, in configuration
s. The evaluation function f(s) is the total number of excess appearances of all teams
in all periods (let us call it fP(s)) i.e. the minimum number of variables to be changed
to satisfy the PERIOD constraint:

P T
f(S) = fp(S) = ∑ ∑ P
(s, n, p),
p = 1n = 1 (1)
0 if OccPS (p, t n ) ≤ 2,
(s, n, p) = 
P î OccPS (p, t n ) − 2 otherwise .

Solving the SLSP means finding a configuration s* ∈ S such as f(s*) = 0.

4.5 Neighborhood Evaluation

The tabu algorithm considers in general at each step the whole neighborhood. So it
is imperative to be able to quickly evaluate the cost of neighboring configurations.
To do this, we used a technique inspired by [6].
Let δ be a |X| * |X| matrix. Each entry δ[x(tm, tn), x(tq, tr)] represents the effect of
the chosen move (swapping matches < tm, tn > and < tq, tr >) on the evaluation
function. Thus, the cost of s’ ∈ N(s) is immediately obtained by adding the proper
entry of δ to f(s) (in O(1)). To get the best neighbor, one search only a subset of the
δ matrix in O(|N(s)|) time. After a move, the δ matrix is updated in O(|X| |X|) time
in the worst case.

4.6 Tabu List Management

A fundamental component of TS is tabu list: a special, short-term memory that
maintains a selective history, composed of previously encountered configurations or
more generally pertinent attributes of such configurations. The aim of a tabu list is
to avoid cycling and go beyond local optima.
Entries in our tabu list are paired matches. Indeed, we make match swaps. After
swapping, the matched couple is classified tabu for the next k iterations (k, called
tabu tenure, is problem dependent), which means that reverse swap is forbidden
during that period.
Tabu tenure plays a very important role for the performance of a tabu algorithm.
If overestimated, it will perhaps incorrectly prohibit the visiting of unexplored
configurations and the method’s capacity to explore the neighborhood will be
reduced. If underestimated, the method may get trapped in local optima.
 After having tried various tabu tenures (randomized and bounded, dynamic, static
etc.), we chose a randomized one. For s ∈ S, we define the tabu tenure ks by:
ks = rand(g) where rand(g) is a random integer from [1; g], g ∈ [4; 100].
To implement the tabu list, we use a |X| * |X| matrix with entries corresponding to
moves. Each entry stores the current iteration number plus the tabu tenure. With this
data structure, a single comparison with the current iteration number is sufficient to
know if a move is tabu.
Nevertheless, note that a tabu move leading to a configuration better than the best
configuration found so far is always accepted (aspiration criterion).

Input: |T|; {number of teams}

Output: a valid schedule or "No valid schedule found";
var f, f*; {evaluation function and its best value
encountered so far}
s, s*; {current configuration and the best
configuration encountered so far}
begin
initialize the tabu list to empty;
generate initial configuration s; {§4.2}
s* := s;
f* := f(s);
while not Stop Condition do
make a best move m in the same week such that m is
not tabu or satisfies the aspiration criterion;
introduce m in the tabu list;
if (f(s) < f*) then
s* := s;
f* := f(s);
else if f* has not been improved for a sufficient
time then
s = s*; {intensification}
set the tabu list to empty; {diversification}
if (f* = 0) then {valid schedule found}
return s*;
return "No valid schedule found";
end
Fig. 3. The TS-SLSP algorithm

4.7 Diversification – Intensification

Intensification and diversification are useful techniques for improving the search
power of a tabu algorithm.
Intensification stores interesting properties of the best configurations found so far
and to be used later. It may initiate a return to attractive regions to search them more
thoroughly. Diversification tries to direct the search to unexplored configurations.
Our TS algorithm includes a single intensification: after a certain number of
non-improving iterations, we return to the best configuration found so far. The
difficulty then is to formally determine when to go back to this best configuration,
because an underestimated interval might inadvertently stop a possibly promising
search. Values of this parameter were determined empirically in [30; 1 000].
We make a diversification immediately after an intensification process by
removing all tabu statuses to enable previously classified tabu moves.

4.8 General Algorithm

The TS-SLSP algorithm (see Fig. 3) begins with an initial configuration in a search
space S built with respect to WEEK and ALLDIFF constraints.
Then it proceeds iteratively to visit a series of locally best configurations
following the neighborhood. At each iteration, a best move m (in the same week to
keep the WEEK constraint satisfied) is applied to the current configuration, even if
it does not improve the current configuration in terms of the value of the evaluation
function. Intensification / diversification steps are performed after a sufficient
number of non-improving moves.
Stop Condition. The algorithm stops if f(s) = 0 (a valid schedule is found) or if a
given limit is reached concerning the number of moves.

5 Computational Results

In this section, we compare results of our TS algorithm with those of [11], [15]
and [16]. Our tests were carried out on a Sun Sparc Ultra 1 (256 RAM, 143 MHz).
TS-SLSP1 is implemented in C (CC compiler with -O5 option). Tests were carried
out on problem instances including 6 to 40 teams (only those greater than 14 are
shown). The TS-SLSP algorithm was allowed to run until a maximum number of 10
million iterations.

5.1 Comparison Criteria

We used three main criteria to do the comparative study. These criteria are given
here in decreasing order of importance:
• Number |T| of teams: one algorithm is said to be more efficient than another one,
if it solves the SLSP with a higher |T|;
• Number of moves: the algorithm’s effort to find a solution, machine independent;
• Running time: the CPU time spent by an algorithm to carry out a given number of
moves, machine dependent.

1
For the source code of TS-SLSP, contact the authors.
5.2 Comparative Results
McAloon et al. [11] recall that Integer Linear Programming (ILP) with CPLEX [5]
was not able to solve the SLSP for |T| = 14 teams, even when avoiding symmetries.
Nevertheless, the method is able to provide solutions for |T| ≤ 12. They also propose
a constraint programming algorithm, implemented under ILOG Solver, which
solved the |T| = 14 problem in 45 minutes (Ultra Sparc), and was the first result
obtained with constraint programming. However, their Constraint Logic
Programming approach failed for |T| = 14. Finally, with a basic local search
algorithm, they solved the |T| = 14 problem in 10 minutes.
Gomes et al. [9] used constraint programming to solve the SLSP up to 18 teams
(in more than 48 hours) with a deterministic complete search. By including
randomness in their algorithm, they reached the same result more quickly:
approximately 22 hours for 18 teams.
In 1998, Régin presented improved results with a more elaborate approach based
on constraint propagation [15] (let us call the approach CP1). This approach
integrates advanced techniques including symmetries elimination and new powerful
filtering algorithms based on arc consistency [3][13][14]. CP1 solves much larger
problem instances since it is able to produce valid schedules for |T| = 24 in 12 hours.
In yet another study, by using multiple threads on a 14 processor Sun system,
Wetzel and Zabatta [22] obtain better results, since schedules were generated for 26
and 28 teams.
Recently, Régin proposed another improved approach [16] (let us call it CP2)
integrating among others, a heuristic based on Euler's theorem for Latin squares.
The SLSP is also transformed into an equivalent problem by adding a dummy week
in order to quickly deduce some inconsistencies. This later approach gives the best-
known results for the SLSP, since it produces solutions for 40 teams in 6 hours.
Table 2. Comparative results of TS-SLSP and well-known methods. A “-“ sign means no
result is available
CP1 CP2 TS-SLSP
Success
|T| Time Backtracks Time (s) Backtracks Time Moves
Ratio (%)
16 4.2 s 1 112 0.6 182 100 0.5 s 4 313
18 36 s 8 756 0.9 263 100 0.3 s 2 711
20 < 6 min 72 095 1.2 226 100 23.9 s 149 356
22 10 h 6 172 672 1.2 157 100 34.3 s 163 717
24 12 h 6 391 470 10.5 2 702 90 5 min 1 205 171
26 - - 26.4 5 683 50 10.7 min 2 219 711
28 - - 316 32 324 50 12.5 min 2 043 353
30 - - 138 11 895 10 22 min 3 034 073
32 - - - - 10 49 min 6 387 470
34 - - - - 30 25 min 2 917 279
36 - - - - 10 1.5 h 9 307 524
40 - - 6h 2 834 754 10 54.3 s 68 746
 Table 2 shows the results of TS-SLSP together with those of CP1 and CP2.
Columns 2-5 give respectively, for CP1 and CP2, execution times and numbers of
backtracks. The last three columns give respectively the success ratio (number of
successful runs / total number of runs), mean times and average numbers of moves
for successful runs of TS-SLSP.
From the Table 2, we may make several remarks. First, we observe that TS-SLSP
is much more powerful than CP1 both in terms of the size of problem instances
solved and computing efforts. Compared with CP2, TS-SLSP also manages to find
valid schedules for problem instances going up to 40 teams. We notice however that
the resolution becomes much more difficult for |T| > 28. We notice also that the
running time of the TS-SLSP algorithm for |T| = 40 is not consistent with other
results. This is simply explained by the stochastic nature of the TS-SLSP algorithm.
Globally, the results of TS-SLSP are consistent and its running times remain
reasonable.

6 Discussions

How to handle the different constraints of a constrained problem is an important

issue that must be answered by any neighborhood algorithm. For the SLSP, recall
that there are three types of constraints (see §3.2): WEEK, PERIOD and ALLDIFF.
As we saw previously in §4, TS-SLSP works with a limited search space whose
configurations satisfy already the WEEK and ALLDIFF constraints. The goal of
TS-SLSP is then to satisfy the PERIOD constraint.
We also experimented with another constraint handling technique, in which the
search space is defined by all the configurations satisfying the ALLDIFF constraint,
and the search starts with a randomized initial configuration (see §4.2). Then the
algorithm tries to satisfy the WEEK and PERIOD constraints. In this case, the
evaluation function f '(s) used to guide the search is defined by a weighted2
summation function of the violations of the WEEK constraint fW(s), formally
described by formula 2 below3, and PERIOD constraint fP(s) (see formula 1 in §4.4):
∀ s ∈ S, f '(s) = 2 * fW(s) + fP(s).

W T
f W (S) = ∑ ∑ W
(s, n, w),
w = 1n = 1 (2)
0 if OccWS (w, t n ) ≤ 1,
(s, n, w) = 
W î OccWS (w, t n ) − 1 otherwise .

2
It seems that the best weight values are 1 for PERIOD and 2 for WEEK.
3
OccWS(w, tn) is the occurrence number of team tn on week w in configuration s.
 The neighborhood is extended to all possible exchanges of matches (with at least
one in conflict) in the same week or in the same period or in different periods and
weeks, see Fig. 4.
The tabu tenure is weighted and dynamic: k ’S = α * [fW(s) + fP(s)] + rand(g),
α ∈ [0.1; 1]. Intensification and diversification processes are used in the same way
as in TS-SLSP.
Weeks
1 2 3 4 5 6 7
1 1 vs 3 1 vs 2 5 vs 8 1 vs 5 4 vs 8 2 vs 6 3 vs 5
Periods

2 3 vs 4 2 vs 8 1 vs 4 6 vs 8 2 vs 5 1 vs 7 6 vs 7
3 5 vs 6 4 vs 6 2 vs 7 4 vs 7 3 vs 7 3 vs 8 1 vs 8
4 7 vs 8 5 vs 7 3 vs 6 2 vs 3 1 vs 6 4 vs 5 2 vs 4
Fig. 4. Extended neighborhood illustration

Experimentation using this technique produces results similar to those of CP1.

Although it fails to solve the |T| > 22 problem, this technique is very fast for 22
teams: it provides solutions in less than 28 minutes (34 146 moves). When only
dealing with the PERIOD constraint, TS-SLSP outperforms results of this technique,
since it solves the SLSP up to 40 teams. The main difference between the two
approaches is the structure of the initial configuration. The technique starting with
an extra constraint to verify (the WEEK constraint) must explore a larger search
space than the other technique.

7 Conclusion

In this paper we presented a tabu algorithm (TS-SLSP) for the sports league
scheduling problem. The algorithm is based on a CSP formulation of the SLSP and
includes the following main features:
• a simple swap neighborhood;
• efficient data structures for fast neighborhood search;
• a dynamic tabu tenure;
• a simple intensification-diversification process.
TS-SLSP was tested on problem instances going up to 40 teams. The
experimental results show that TS-SLSP largely outperforms some previously
developed approaches (ILP [11], basic LS [11] and constraint programming [9][15]).
Indeed, while these approaches are limited to instances of 24 teams, TS-SLSP is
able to find a schedule for instances going up to 40 teams. This result compares well
with the best-known approach, which combines well-elaborated constraint
propagation algorithms and a non-trivial formulation of the initial problem [16].
At the same time, the computing times required by TS-SLSP are much greater
than those obtained using the most efficient algorithm. Nevertheless, there is
certainly plenty of room for improvement in TS-SLSP. One possible enhancement
would be to integrate more elaborate intensification and diversification mechanisms,
another would be to have a close study of configuration structures in order to devise
clever neighborhoods and constraint handling techniques. It would also be
interesting to envisage a combination of some advanced constraint propagation
techniques with tabu search.
This work underlines once again the importance of efficient data structures for
the high performance of a TS algorithm. It also confirms that parameter setting is
crucial for obtaining quality solutions.
The TS-SLSP may also be adaptable to other sports scheduling problems e.g. the
minimization of breaks (two consecutive home or away matches) [21] and those
studied in [17]. Let us also mention that TS has already proven efficient for solving
some sports scheduling problems, with constraints of various types. A hybrid tabu
search has been proposed [4] to respond to the National Hockey League’s
scheduling problem, which underlines the effectiveness of a TS / genetic algorithm
combination; the skilful mix of elements of the two methods produces better results
than those of separately considered methods. All these studies suggest that a
metaheuristic approach (e.g. TS, simulated annealing, genetic hybrid) has good
potential for solving various planning and scheduling problems related to sporting
events.4

Acknowledgments

We would like to thank J.C. Régin for having sent us his latest results and the
referees for their relevant remarks.

References

1. Ackley, D.H.: A Connectionist Machine for Genetic Hillclimbing. Boston, MA: Kluwer
Academic Publishers (1987)
2. Bean, J.C., Birge, J.R.: Reducing Travelling Costs and Player Fatigue in the National
Basketball Association. Interfaces 10 / 3 (1980) 98–102
3. Bessière, C., Régin, J.C.: Arc Consistency for General Constraint Networks: Preliminary
Results. In: Proceedings of the Fifteenth International Joint Conference on Artificial
Intelligence. Nagoya, Aichi, Japan: International Joint Conference on Artificial
Intelligence, Inc. (1997) 398–404
4. Costa, D.: An Evolutionary Tabu Search Algorithm and the NHL Scheduling Problem.
INFOR 33 / 3 (1995) 161–178
5. CPLEX Optimization: Using the CPLEX©Callable Library. Tahoe, NV, Inc. (1995)
6. Ferland J.A., Fleurent, C.: Genetic and Hybrid Algorithms for Graph Coloring. Annals of
Operations Research: Metaheuristics in Combinatorial Optimization 63 / 1. Hammer,
P.L. et al. (Eds) (1996) 437–461
7. Ferland J.A., Fleurent, C.: Allocating Games for the NHL Using Integer Programming.
Operations Research 41 / 4 (1993) 649–654
8. Glover, F., Laguna, M.: Tabu Search. Boston, MA: Kluwer Academic Publishers (1997)

4
Additional studies are reported in [4][19][23].
9. Gomes C.P., Selman B., Kautz H.A.: Boosting Combinatorial Search Through
Randomization. In: Proceedings of the Fifteenth National Conference on Artificial
Intelligence. AAAI Press / The MIT Press, Madison, WI (1998) 431–437
10. Henz, M.: Scheduling a Major College Basketball Conference – Revisited. To appear in:
Operations Research 49 / 1 (2001)
11. McAloon, K., Tretkoff, C., Wetzel, G.: Sports League Scheduling. In: Proceedings of the
Third ILOG Optimization Suite International Users’ Conference. Paris, France (1997)
12. Nemhauser, G.L., Trick, M.A.: Scheduling a Major College Basketball Conference.
Operations Research 46 / 1 (1998) 1–8
13. Régin, J.C.: A Filtering Algorithm for Constraints of Difference in CSPs. In:
Proceedings of the Twelfth National Conference on Artificial Intelligence, Vol. 1.
Seattle, Washington: AAAI Press (1994) 362–367
14. Régin, J.C.: Generalized Arc Consistency for Global Cardinality Constraint. In:
Proceedings of the Thirteenth National Conference on Artificial Intelligence / Eighth
Conference on Innovative Applications of Artificial Intelligence, Vol. 1. Portland,
Oregon: AAAI Press (1996) 209–215
15. Régin, J.C.: Modeling and Solving Sports League Scheduling with Constraint
Programming. First Congress of the French Operational Research Society (ROADEF).
Paris, France (1998)
16. Régin, J.C.: Sports Scheduling and Constraint Programming. INFORMS. Cincinnati,
Ohio (1999)
17. Schaerf, A.: Scheduling Sport Tournaments Using Constraint Logic Programming.
Constraints 4 / 1 (1999) 43–65
18. Schreuder, J.A.M.: Constructing Timetables for Sport Competitions. Mathematical
Programming Study 13 (1980) 58–67
19. Terril, B.J., Willis, R.J.: Scheduling the Australian State Cricket Season Using Simulated
Annealing. Journal of the Operational Research Society 45 / 3 (1994) 276–280
20. Tsang, E.P.K.: Foundations of Constraint Satisfaction. London and San Diego: Academic
Press (1993)
21. Werra (de), D.: Geography, Games and Graphs. Discrete Applied Mathematics 2 (1980)
327–337
22. Wetzel G., Zabatta F.: CUNY Graduate Center CS Technical Report (1998)
23. Wright, M.: Timetabling County Cricket Fixtures Using a Form of Tabu Search. Journal
of the Operational Research Society 45 / 7 (1994) 758–770

Appendix:

The table in the next page gives a solution to the Sports League Scheduling Problem
found by TS-SLSP for 40 teams.
 Weeks
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39
1 21 32 3
4
4
5
5
6
6
7
7
8
8
9
9
10
10
11
11
12
32
40
27
39
14
15
15
16
16
17
17
18
18
19
19 20 21
20 21 22
22
23
23
24
24
25
25
26
1
13
27
28
28
29
29
30
30
31
12
40
32
33
33
34
34
35
35
36
36
37
37
38
38
39
14
26
3 1
2 39 2 3 19 5 6 22 8 9 10 11 12 13 14 15 7 38 18 10 20 21 22 23 24 25 26 27 4 29 30 7 32 33 34 17 36 1 2
4 5 6 31 8 9 34 11 12 13 14 15 16 17 18 28 40 21 31 23 24 25 26 27 28 29 30 16 32 33 19 35 36 37 40 39 37 38
8 5
3 34 10 11 3 4 23 6 7 4 9 2 7 12 13 14 15 16 13 1 19 16 21 22 15 24 25 22 27 28 25 30 10 2 33 34 5 6 3
39 36 37 8 9 31 11 12 17 14 23 20 17 18 19 20 21 26 40 24 29 26 27 36 29 30 35 32 33 38 35 18 28 38 39 31 32 37
5 6
4 37 7 1 25 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 9 26 27 28 29 30 31 32 1 2 3 4
38 39 8 40 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 40 33 34 35 36 37 38 39 33 34 35 36
9 7
5 33 8 9 1 2 3 4 5 6 7 8 30 10 11 12 13 14 15 16 17 18 19 20 6 22 23 24 25 26 27 28 29 30 1 2 18 4 5
37 38 39 10 11 12 13 14 15 16 17 36 19 20 21 22 23 24 25 26 27 28 29 40 31 32 33 34 35 36 37 38 39 31 32 40 34 35
7 8
6 35 9 10 11 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 2 17 18 19 20 21 22 23 24 25 26 27 28 1 16 3 4 5 6
36 37 38 39 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 40 28 29 30 31 32 33 34 35 36 37 38 39 29 40 31 32 33 34
4 9
7 38 1 2 12 13 27 2 3 8 5 6 11 8 9 10 11 12 17 14 15 20 17 18 19 20 21 26 23 24 29 26 14 32 3 4 1 10 7
35 6 7 38 39 40 15 16 13 18 19 16 21 22 23 24 25 22 27 28 25 30 31 32 33 34 31 36 37 34 39 40 37 29 30 35 28 33
18 10
8 24 11 12 13 14 24 1 2 30 4 5 33 7 8 33 10 11 12 13 14 6 16 17 3 19 20 21 22 23 9 1 2 3 4 5 15 7 8
34 35 36 37 38 30 16 17 40 19 20 40 22 23 39 25 26 27 28 29 39 31 32 9 34 35 36 37 38 15 25 26 27 28 29 21 31 32
10 11
9 32 12 17 14 26 16 21 1 23 3 25 5 2 7 36 9 6 11 12 13 10 15 16 17 14 1 20 21 4 5 2 3 8 9 6 7 8 13
33 34 31 36 40 38 35 18 37 20 39 22 27 24 40 26 31 28 29 30 35 32 33 34 39 15 37 38 18 19 24 25 22 23 28 29 30 27
11 12 13 14 15 16 17 18 19 1 2 3 4 34 6 7 8 9 10 11 12 13 14 5 16 17 18 19 20 1 2 3 4 5 6 7 8 9 10
Periods

10 31 32 33 34 35 36 37 38 39 20 21 22 23 40 25 26 27 28 29 30 31 32 33 40 35 36 37 38 39 21 22 23 24 25 26 27 28 29 30
12 13
11 30 14 15 16 17 18 28 20 21 1 10 3 4 5 6 4 8 9 19 11 12 10 14 23 16 17 18 1 2 3 13 5 6 7 8 9 2 11
31 32 33 34 35 36 40 38 39 22 15 24 25 26 27 31 29 30 22 32 33 37 35 28 37 38 39 19 20 21 40 23 24 25 26 27 36 29
13 14
12 29 15 16 17 18 19 20 21 22 31 1 2 3 4 5 6 7 39 9 10 11 12 13 14 15 16 1 2 3 4 5 6 7 8 9 10 11 12
30 31 32 33 34 35 36 37 38 40 24 25 26 27 28 29 30 40 32 33 34 35 36 37 38 39 17 18 19 20 21 22 23 24 25 26 27 28
14 15
13 28 16 13 18 19 20 17 22 2 24 4 1 6 3 4 5 10 7 8 9 14 11 12 13 18 8 2 3 22 1 6 7 15 5 10 11 12 9
29 30 35 32 33 34 39 36 19 38 21 26 23 28 29 30 27 32 33 34 31 36 37 38 35 40 16 17 39 23 20 21 40 27 24 25 26 31
15 22
14 27 17 18 4 20 21 19 23 24 25 26 13 1 2 3 16 5 6 7 8 9 4 11 12 26 2 3 28 5 6 31 8 9 10 11 12 13 1
40 29 30 7 32 33 37 35 36 37 38 14 28 29 30 19 32 33 34 35 36 40 38 39 27 14 15 31 17 18 34 20 21 22 23 24 25 39
16 17
15 26 18 24 20 21 22 23 24 25 26 27 28 29 1 2 3 4 5 6 7 8 9 10 1 2 3 4 5 6 7 8 9 10 11 12 13 19 15
27 28 40 30 31 32 33 24 35 36 37 38 39 30 31 32 33 34 35 36 37 38 39 11 12 13 14 15 16 17 18 19 20 21 22 23 40 25
17 18
16 25 19 20 21 22 5 24 25 26 27 28 29 30 31 1 2 3 4 18 6 3 8 1 2 7 4 5 6 7 8 9 31 11 12 13 14 15 16
26 27 28 29 30 10 32 33 34 35 36 37 38 39 32 33 34 35 23 37 40 39 9 10 40 12 13 14 15 16 17 36 19 20 21 22 23 24
6 19
17 36 23 21 22 23 15 25 26 27 28 29 9 31 32 9 1 2 3 4 5 15 1 2 21 4 5 6 7 8 24 10 11 12 13 14 3 16 20
25 40 27 28 29 39 31 32 33 34 35 18 37 38 24 34 35 36 37 38 30 7 8 30 10 11 12 13 14 39 16 17 18 19 20 33 22 40
19 20
18 23 21 22 23 24 25 26 27 28 29 30 31 32 35 34 37 1 2 3 4 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18
24 25 26 27 28 29 30 31 32 33 34 35 36 40 38 40 36 37 38 39 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22
20 21
19 22 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 1 2 1 2 3 4 5 6 7 8 10 11 11 12 13 14 15 16 17 18 19
23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 38 39 3 4 5 6 7 8 9 10 40 40 13 14 15 16 17 18 19 20 21
21 16
20 41 20 19 2 15 1 7 29 3 23 12 6 5 33 8 35 17 8 5 16 7 13 15 18 3 19 25 9 10 31 4 1 4 2 35 6 14 17
28 26 29 9 37 14 10 40 18 39 13 21 24 37 25 39 20 31 36 27 38 34 34 33 11 36 32 11 12 32 22 27 26 30 38 30 24 23

View publication stats