Version of Record: https://www.sciencedirect.

com/science/article/pii/S0305054819300528
Manuscript_8983af47bec1dedc6f7dfd96c9c9cea9

Lower bounds and compact mathematical

formulations for two soft constraints for
university examination timetabling problems.
∗1 † 2 ‡2
Taha Arbaoui , Jean-Paul Boufflet , and Aziz Moukrim
1
Laboratoire d’Optimisation des Systèmes Industriels, Institut Charles Delaunay
(ICD-LOSI), Université de Technologie de Troyes, 12 rue Marie Curie, CS 42060, 10004,
Troyes, France
2 Sorbonne universités, Université de Technologie de Compiègne, Laboratoire Heudiasyc,

UMR CNRS 7253, 57 avenue de Landshut, CS 60319, 60203 Compiègne, France

Abstract
The examination timetabling problem (ETT) can be described as a
set of exams to be scheduled over an examination session while respecting
numerous hard and soft constraints. In this paper we consider the spacing
soft constraints that seek to prevent students sitting more than one exam
per day. Out of consideration for candidates, these soft constraints often
feature in real ETT problems that academic institutions seek to solve.
Work on ETT has tended to focus on heuristic approaches, and little effort
has gone into developing lower bounds, although both are of practical
and theoretical interest. For this study we consider formulations of these
soft constraints as defined in the ITC2007 examination timetabling track.
In existing mathematical formulations of these spacing soft constraints
the number of equations is of the order of the square of the number of
exams, and current solvers may face problems at runtime because of their
large memory requirement. In this study we present a generic model
for computing lower bounds, together with more compact formulations
where the number of equations is of the order of the number of exams.
Computational results on spacing soft constraints that seek to prevent
students sitting more than one exam per day are an improvement on
results obtained so far on lower bounds, and our new formulations yield a
more compact model that gives better results than those given by existing
formulations.

Keywords: Examination timetabling problems, Spacing soft constraints,

Lower bounds, Compact mathematical formulation
∗ taha.arbaoui@utt.fr
† jean-paul.boufflet@hds.utc.fr
‡ aziz.moukrim@hds.utc.fr

© 2019 published by Elsevier. This manuscript is made available under the Elsevier user license
https://www.elsevier.com/open-access/userlicense/1.0/
1 Introduction
Academic institutions periodically face the problem of organizing examination
sessions. Drawing up an examination timetable is a complex, time-consuming,
tedious task for practitioners, who are often faced with conflicting demands from
the institution, from teachers and from students. The examination timetabling
problem (ETT) can be described as a set of exams to be scheduled into a set
of periods while respecting numerous hard and soft constraints. The quality
of a solution that satisfies all the hard constraints is measured by quantifying
violations of soft constraints. A large number of variants of ETTs can be found
in the literature because every institution has its own set of hard constraints
and uses a different set of soft constraints to assess solutions.
In 1997, Burke et al. [10] presented an introduction to automated exami-
nation timetabling using information collected from UK Universities, where the
authors catalogued and structured common types of constraints, and described
popular approaches used at the time. In a later comprehensive survey, Schaerf
[31] presented school, course and examination timetabling problems. In these
two studies variants of ETTs were discussed and solutions were reviewed and
compared.
A number of methodologies were developed to tackle specific problems, but
scientific comparisons between different methodologies are difficult, given that
the problems presented do not have the same sets of hard and soft constraints.
However, comparisons became easier when benchmarks with standard variants
started to be used, such as those from the universities of Toronto [11], Notting-
ham [16] and Melbourne [24]. In 2009, Qu et al. [29] summarized earlier surveys
on examination timetabling [10, 12, 15, 31] and presented a state of the art of
solution methods. Benchmarks available at that time were presented, and ap-
proaches and results on the benchmark datasets were discussed. The second In-
ternational Timetabling Competition (ITC2007) examination timetabling track
[22] introduced a problem description which covers many real-world situations.
These benchmarks have been made available to the research community.
Examination timetabling has been intensively studied over the last decades
using a large variety of optimisation techniques for preprocessing and for solving.
Gogos et al. [20] and Arbaoui et al. [6] proposed a preprocessing step to
reveal hidden hard constraints that can be deduced prior to solving. The initial
data can then be enriched and infeasibilities can be detected sooner during the
solving process.
A large number of solution methods based on heuristics, meta-heuristics
and their hybridizations, and hyper-heuristics can be found in the literature. A
selection of investigated methods includes: Graph Ordering Heuristic [2], Tabu
Search [21], Simulated Annealing [17], Great Deluge [26], Hill Climbing with
or without Late Acceptance [8], Bin Packing Heuristic [32], Evolutionary and
Nature Inspired Algorithms [3] and Hyper-Heuristic [9]. Recently, Alzaqebah
et al. [4] obtained the best result found so far for one instance of the Toronto
benchmark using a hybrid bee colony approach, simulated annealing, and late
acceptance hill climbing.

2
 Exact methods and their hybridizations with heuristics have been studied
to solve examination timetabling problems. Hybridizing integer programming
and a decomposition approach was investigated by Qu et al. [30]. An integer
programming phase for assigning exams to rooms was used by Gogos et al. [20].
For a specific problem, MirHassani [25] proposed a Mixed Integer Program-
ming (MIP) model. Recently, a column generation approach was proposed by
Woumans et al. [33]. To increase the spacing of exams for students, multiple
versions of an exam to be planned are allowed, and a trade-off between spacing
for students and numbers of exam versions is optimized.
The ITC2007 competition provides a framework which describes seven soft
constraints widely encountered and a dataset based on current real-world in-
stances [22]. The results on ITC2007 instances have been steadily improved over
the last decade through the application of heuristic approaches that consider
the objective function with all seven terms. A simple yet effective hill-climbing
heuristic was proposed in [13], which achieved competitive results. Recently,
in [7], the authors proposed a simulated annealing, where better results were
obtained once parameters had been finely tuned.
However, real-world problems may contain a subset of these soft constraints.
Following the competition, the Yeditepe dataset proposed new real instances
where two terms are considered [28].
In this paper we consider the two soft constraints that aim at minimizing
number of students sitting more than one exam per day. We chose this frame-
work because it provides a firm basis for establishing lower bounds and new
mathematical models for these two spacing soft constraints that are often used
in academic institutions.
Assessing lower bounds on soft constraints that aim at minimizing the num-
ber of students sitting more than one exam per day has both a theoretical and
a practical interest, in providing practitioners with insights on unavoidable vio-
lations and determining the distance from the bounds of any proposed heuristic
solution. To the best of our knowledge, Arbaoui et al. [5] presented the first
work on spacing soft constraint lower bounds for ITC2007 problems.
In contrast to the numerous heuristic based approaches, few exact meth-
ods have been reported. McCollum et al. [23] proposed a new mathematical
model for the ITC2007 examination timetabling track that provides a mean-
ingful basis for hard and soft constraints. As the authors make plain, their
model was not designed to solve sizeable real instances, but it was created only
for modeling purposes. However, it does optimally solve two instances of the
Yeditepe dataset [28]. Running their model on sizeable real instances using cur-
rent solvers is unfortunately not easy, because of the large numbers of variables
and constraints. Fonseca et al. [18] proposed an improved version of this model
that made a better use of memory and they encountered fewer out-of-memory
events on sizeable instances. Arbaoui et al. [6] ran an improved formulation
on each term considered individually, and for certain terms optimal values were
attained.
Soghier et al. [32] investigated bin packing heuristics and proposed an adap-
tive hybrid hyper-heuristic approach. This way of approaching ETT treats ex-

3
ams as items and rooms/periods as bins. Data-dependent Dual-Feasible Func-
tions (DDFF) have previously been proposed for building lower bounds for bin-
packing problems (see Carlier et al. [14]). In [6], the authors proposed a valid
inequality based on an adaption of a DDFF that proved to be effective on small
instances.
Applying valid inequalities while reducing the number of variables and con-
straints has been seen to be helpful in improving exact methods based on integer
programming formulations [6, 18].
In this study we present improvements for computing lower bounds on the
two spacing soft constraints used to prevent students sitting more than one exam
per day, and we present more compact formulations for these soft constraints
that can result in a MIP model that consumes less memory.
The remainder of the paper is organized as follows. In Section 2 we out-
line the hard constraints of the ITC2007 examination timetabling competition
and the two spacing soft constraints used to prevent students sitting more than
one exam per day. In Section 3 we propose three ways of computing the min-
imum number of spacing soft constraint violations for every clique of a given
size, and the generic MIP model that computes the associated lower bounds.
In Section 4 we recall earlier contributions that allowed the numbers of hard
constraints to be reduced, we present our new formulations of the spacing soft
constraints, and we compare the sizes of different MIP formulations. We focus
on two terms, and for the purposes of comparison we use the improved approach
[22] made available to the community by Tomàs Müller [1], the winner of the
ITC2007 examination timetabling track. In Section 5 we present and comment
on computational results for the two soft constraints applied on the ITC2007
and Yeditepe datasets. Our concluding remarks are made in Section 6.

2 Framework of the study

This section describes the hard constraints and the two spacing soft constraints
used to minimize the number of students sitting more than one exam per day,
as defined in the examination track of the Second International Timetabling
Competition (ITC2007) [22] used in our study.
Input data consists of exams, rooms, and days/periods. Every exam has a
duration, and must be allocated to a period and a room. The set of students
enrolled for each exam is given. Each room has a seating capacity. Exams can
share a room, unless there is an exclusivity hard constraint (see below) and
provided that the capacity of the room is respected. The examination session
is composed of days defined over a specified length of time. Each day has a
number of periods and each period has a duration.
The standard hard constraints are:
Conflicts: A student can sit only one exam at a time.
Period length: The exam duration must be less than or equal to the period
duration.

4
Room capacity: The capacity of any room cannot be exceeded at any period.
and some exams may be subject to the following hard constraints:
Room exclusivity: An exam must take place in a room on its own.
Time-ordering: There is an order between a pair of exams that must be re-
spected.
The time-ordering hard constraints are the After constraint (exam i has to
be scheduled after exam j), the Exclusion constraint (exams i and j have to
be scheduled at different periods), and the Coincidence constraint (exams i
and j must be scheduled at the same period).
We will now consider the Two-In-a-Row and Two-In-a-Day spacing soft
constraints used to minimize the number of students sitting more than one
exam per day. These are constraints widely encountered in practice:

Two-In-a-Row: For a student, exams that are scheduled back-to-back on the

same day should be avoided.
Two-In-a-Day: For a student, exams that are scheduled on the same day but
not back-to-back should be avoided.

Data and weights relating to assessing these soft constraints are provided in
the input file of each instance.
Mc Collum et al. presented in [23] a dataset based on current real-world in-
stances, together with a mathematical model. This provides a stable framework
made available to the community for studying lower bounds and new formu-
lations of soft constraints that have relevance for a large number of academic
institutions. The notation that we use below for constants, parameters, and
variables closely follows the original notation in the model proposed by [23].

3 Two-In-a-Row & Two-In-a-Day lower bounds

We propose improvements for calculating lower bounds for Two-In-a-Row (2R)
and Two-In-a-Day (2D) spacing soft constraints, where the corresponding terms
to be minimized are denoted C 2R and C 2D . We consider them both individually
and together.
To the best of our knowledge, the first attempt to design lower bounds for
these soft constraints was proposed by Arbaoui et al. [5]. Their idea was to use
cliques calculated on the conflict graph G(E, A), where E is the set of exams,
and for exams i and j, there is an edge [i, j] ∈ A if at least one student takes the
two exams. Each edge [i, j] is weighted by wij , the number of students taking
exams i and j.
The authors proposed three limits on the size of a clique, beyond which at
least one violation of these spacing constraints occurs. Cliques whose size ex-
ceeds those limits are selected. For every clique, contributions for penalties of

5
these spacing soft constraints are computed using a MIP formulation that gives
optimal values. The contributions of two cliques with no common edge can be
summed. To compute a lower bound, a family of edge-disjoint cliques is built ap-
plying a greedy heuristic without any guarantee of optimality. Moreover, since
selected cliques have no common edges, some edges cannot be counted as viola-
tions. Furthermore, the greedy heuristic does not take account of Coincidence
constraints between exams.
Computing minimum numbers of violations is discussed below in Section 3.1.
The limits make it possible to know whether a clique induces at least one soft
constraint violation, but they do not make it possible to compute minimum
numbers of violations.
In Section 3.2 we propose a generic MIP model for computing lower bounds
for Two-In-a-Row and Two-In-a-Day soft constraints (individually or together).
The formulation we propose allows lower bounds to be computed optimally using
minimum numbers of violations and individual contributions of selected cliques.
This formulation ensures that a violation cannot be counted twice, while taking
into account Coincidence constraints between exams.

3.1 Minimum number of violations

For a clique c of size k within the conflict graph, we propose computing λR (k),
λD (k) and λRD (k), that is to say the minimum numbers of Two-In-a-Row,
Two-In-a-Day, and Two-In-a-Row AND Two-In-a-Day spacing soft constraint
violations. Every spacing soft constraint violation for exams i and j is assessed
using the weight wij of the corresponding edge [i, j]. The number of soft con-
straint violations equals the number of weighted edges to be summed to give
either the Two-In-a-Row term or the Two-In-a-Day term of the objective func-
tion.
A day is said to be of type Di if it has i periods, and nDi is the number of
days of type Di. Not all the possible Di’s exist for a particular instance. We
denote as δ the set of numbers of periods corresponding to the types of days
in an instance: so, for example, δ = {2, 3, 4} meansPthat the instance contains
days of types D2, D3 and D4. Additionally, ϕ = d∈(δ\{1}) nDd denotes the
sum of numbers of D2, D3 and D4 days. Note that there are no D5 days in
either the ITC2007 or the Yeditepe instances.
A clique corresponds to pairwise adjacent exams in the conflict graph, so
these exams must be allocated to different periods. The three limits on the size
of a clique beyond which at least one violation of Two-In-a-Row, or Two-In-a-
Day, or Two-In-a-Row AND Two-In-a-Day occurs are the following:
Xi
LR = nDi
2
i∈δ
D D1
L = n + 2ϕ
RD D1
L = n +ϕ
LR : on Di days 2i exams can be allocated without any Two-In-a-Row
 

6
violations, and so one exam can be allocated to D1 and D2 days, and two
exams can be allocated to D3 and D4 days with an empty period between the
two exams.
LD : one exam can be allocated to a D1 day and two exams can be allocated
back-to-back to D2, D3 and D4 days without any Two-In-a-Day violations.
LRD : at most one exam per day can be allocated without any Two-In-a-Row
or Two-In-a-Day violations.
We have LRD ≤ LR , and if there are days of types D3 or D4, then
RD
L ≤ LD .
To compute λR (k), λD (k) and λRD (k), that is to say the minimum numbers
of edges to be used to count penalties, the idea is to determine the type of day
and the type of period that an exam needs to be allocated to in order to create
as few violations as possible.
For Two-In-a-Row soft constraints, if k ≤ LR , then no violation occurs.
If LR < k ≤ LR + nD2 + nD4 , we have LR allocated exams without any
violations, one in D1 and D2 days, and two in D3 and D4 days (empty period
between exams), after which each exam allocated to D2 or D4 days creates one
Two-In-a-Row violation, so we have k − LR Two-In-a-Row violations. When
k = LR + nD2 + nD4 exams are allocated, note that there is an empty period
in D3 and D4 days.
If LR +nD2 +nD4 < k, we have LR +nD2 +nD4 allocated exams, after which
each exam allocated to D3 or D4 days creates

two Two-In-a-Row violations.
We have nD2 + nD4 + 2 k − LR + nD2 + nD4 violations.
Regarding λR (k), the following proposition therefore holds:

Proposition 1 Given a clique c of size k, there will be at least λR (k) Two-In-

a-Row violations where λR (k) is equal to:
if k ≤ LR

0

k − LR if LR < k ≤ LR + nD2 + nD4

nD2 + nD4 + 2 k − LR + nD2 + nD4



otherwise

For Two-In-a-Day soft constraints, if k ≤ LD then no violation occurs.

If LD < k ≤ LD + nD3 + nD4 , we have LD allocated exams without any
violations, we have two back-to-back exams in D2, D3 and D4 days, and each
exam allocated subsequently to D3 or D4 days will create one Two-In-a-Day
violation (three exams are consecutive). We thus have k − LD violations.
If LD + nD3 + nD4 < k, we have LD + nD3 + nD4 allocated exams, with
only D4 days having an empty period, and each exam allocated will create two
Two-In-a-Day violations. We thus have nD3 + nD4 + 2 k − LD + nD3 + nD4

violations.
So, when computing λD (k) the following proposition holds:

Proposition 2 Given a clique c of size k, there will be at least λD (k) Two-In-

7
a-Day violations where λD (k) is equal to:
if k ≤ LD

0

k − LD if LD < k ≤ LD + nD3 + nD4

 D3
n + nD4 + 2 k − LD + nD3 + nD4



otherwise

For Two-In-a-Row AND Two-In-a-Day soft constraints, if k ≤ LRD

then no violation occurs.
If LRD < k ≤ LRD +ϕ, we have LRD exams allocated without any violations,
we have one exam per day, and each exam allocated subsequently to D2, D3, or
D4 days will create a violation (at most ϕ). We thus have k − LRD violations.
If LRD +ϕ < k ≤ LRD +ϕ+nD3 +nD4 , we have LRD +ϕ allocated exams (two
exams in D2, D3 and D4 days), after which each exam allocated to D3 or D4
days will create two additional violations. We thus have ϕ + 2 k − LRD + ϕ

violations.
If LRD + ϕ + nD3 + nD4 < k, we have LRD + ϕ + nD3 + nD4 allocated
exams, with only D4 days having an empty period, and each exam allo-
cated subsequently
to D4 days will create three

violations. We thus have
ϕ + 2 nD3 + nD4 + 3 k − LRD + ϕ + nD3 + nD4 violations.
When computing lambda λRD (k), the following proposition therefore holds:

Proposition 3 Given a clique c of size k, there will be at least λRD (k) Two-
In-a-Row or Two-In-a-Day violations where λRD (k) is equal to:
if k ≤ LRD


 0

 k − LRD if LRD < k ≤ LRD + ϕ




ϕ + 2(k − (LRD + ϕ)) if LRD + ϕ < k ≤ LRD +ϕ + nD3 + nD4

 ϕ + 2 nD3 + nD4




 otherwise
RD D3 D4



+3 k− L +ϕ+n +n


Figure 1 provides an illustrative example. On the left, a clique c of k = 5

exams, {A, B, C, D, E}, is displayed with weighted edges (wij are the numbers
of students taking i and j exams).
In Figure 1 (top right), the examination timetabling period has two days, one
D2 and one D3, so ϕ = 2. Values for the limits are (LR = 3) < k, (LD = 4) < k,
and (LRD = 2) < k. It can be seen that clique c contributes to Two-In-a-Row
or Two-In-a-Day penalties, because at least one of these spacing soft constraint
violations occurs, and so at least one edge has to be used in the computation.
However, the numbers of violations of these spacing constraints are not known.
Applying the propositions:

• λR (5) = 3 = nD2 + nD4 + 2 k − LR + nD2 + nD4

• λD (5) = 1 = k − LD

• λRD (5) = 4 = ϕ + 2(k − (LRD + ϕ))

8
 D2 D3
E Clique c
ϕ=2
100 A
100
LR = 3 LD = 4 LRD = 2
1
100 λR (5) = 3 λD (5) = 1 λRD (5) = 4
2R 2D
D 2 1 Penalty weights: w = 10 w =1

100 Optimal allocations of exams within days:

1 feasible swaps
100

100 B
A B C D E
C

Figure 1: Example of a clique c for which Two-In-a-Row or Two-In-a-Day soft

constraint violations occur.

If exams in every clique of size k = 5 are scheduled in the examination

timetabling period, there will be three Two-In-a-Row violations, one Two-In-
a-Day violation, and four Two-In-a-Row or Two-In-a-Day violations. These
values correspond to numbers of weighted edges that contribute to the compu-
tation of corresponding penalties.

3.2 Generic formulation for C 2R , or C 2D , or C 2R + C 2D LBs

We propose a generic formulation to optimally compute lower bounds for Two-
In-a-Row (C 2R term), or Two-In-a-Day (C 2D term), or Two-In-a-Row AND
Two-In-a-Day (C 2R + C 2D ) spacing soft constraints. The formulation ensures
that an edge (a soft constraint violation) cannot be counted twice while taking
into account Coincidence constraints between exams in the various cliques.
We denote as S a set of cliques c selected such that (|c| = k) is larger than
a chosen limit (see Section 3.1) and E(S) is the set of edges of all cliques in S.
The idea is to optimally select edges e = [i, j] of cliques c ∈ S which can
contribute once to lower bounds using λR (k), λD (k) and λRD (k), and, for every
clique c, using the contributions to the penalties Two-In-a-Row, or Two-In-
a-Day, or Two-In-a-Row AND Two-In-a-Day, denoted as T R (c), T D (c) and
T RD (c).
To obtain the T R (c), T D (c) and T RD (c) optimal values we use a MIP model
presented in [6]. Problems to be solved closely resemble set cover problems. All
the exams in a clique have to be allocated to different periods. Subsets of one,
two, three, or four exams are used to generate sets of permutations. According
to its size a permutation can be allocated to days of types D1, D2, D3 and D4.
Every permutation of a subset of exams can be assessed, since Two-In-a-Row
or Two-In-a-Day violations may occur. The model finds an optimal selection
of some assessed permutations such that at most all the days are covered and

9
that all the exams in the clique are allocated.
It is not necessary to generate every possible permutation of a subset of
exams, because many of them will have the same penalty induced by soft con-
straint violations. Breaking symmetries saves many variables, but the number
of permutations to be used can be reduced still further. Among the permuta-
tions of a subset of exams to be allocated to a type of day a single assessed
permutation is retained, since if this subset of exams is selected to cover one of
the days of this type, the permutation used has a minimal penalty.
These insights can be illustrated using the clique c shown in Figure 1. Let
us consider subsets of one, two, and three exams to be allocated to D2 and D3
days.
For each day of type D2 and each pair of different exams i and j, we have the
following two permutations: (i, j) for exam i in the first period and exam j in the
second period, and (j, i) when the order is reversed. These two permutations
have the same penalty induced by the Two-In-a-Row spacing soft constraint
violation. We denote this “same induced penalty” as (i, j) ∼ (j, i). Permutations
with one exam i also need to be considered. Given that in an optimal solution
one exam may be allocated to a D2 day, we have (i, ) ∼ ( , i).
For a D3 day and one exam i we have (i, , ) ∼ ( , i, ) ∼ ( , , i). For two
exams i and j we have (i, j, ) ∼ (j, i, ) ∼ ( , i, j) ∼ ( , j, i), and (i, , j) ∼ (j, , i).
For three exams i, j, k we have (i, j, k) ∼ (k, j, i), and (i, k, j) ∼ (j, k, i), and
(j, i, k) ∼ (k, i, j).
We can further reduce the number of permutations. As an example, given
a subset of three exams, let us assume that T RD (c) is to be assessed (Two-In-
a-Row AND Two-In-a-Day), and that the retained permutations are (i, j, k),
(i, k, j), and (j, i, k). Next, let us assume that in an optimal solution this subset
of exams is allocated to a D3 day. Since the solution is optimal, the permutation
used to cover a D3 day has a minimum penalty. Therefore, penalties can be
computed for permutations (i, j, k), (i, k, j) and (j, i, k) and the permutation
with the smallest penalty will be retained.
Let us take the example shown in Figure 1 with a subset of exams {C, D, E},
and with weights w2R = 10 and w2D = 1. Consequently, (C, D, E) has a penalty
of 22, (C, E, D) has a penalty of 31, and (D, C, E) has a penalty of 31. The single
permutation to be retained in order to create a decision variable is (C, D, E).
In Figure 1 the optimal allocations of exams are displayed on the lower right,
with feasible swaps indicated by curved arrows (symmetries).
We obtain T R (c) = 30, T D (c) = 2 and T RD (c) = 32, the minimum contri-
butions for penalties.
The generic formulation that we propose for optimally selecting edges of
assessed cliques is the following:
Minimize:
X
wij (w2R Xe + w2D Ye ) (1)
e=[i,j]∈E(S)

10
 subject to:

∀e ∈ E(S) Xe + Ye ≤ 1 (2)
X
∀c ∈ S Xe ≥ λR (c) (3)
e=[i,j]∈ε(c)
X
∀c ∈ S Ye ≥ λD (c) (4)
e=[i,j]∈ε(c)
X
∀c ∈ S (Xe + Ye ) ≥ λRD (c) (5)
e=[i,j]∈ε(c)
X
∀c ∈ S w2R wij Xe ≥ T R (c) (6)
e=[i,j]∈ε(c)
X
∀c ∈ S w2D wij Ye ≥ T D (c) (7)
e=[i,j]∈ε(c)
X
∀c ∈ S wij (w2R Xe + w2D Ye ) ≥ T RD (c) (8)
e=[i,j]∈ε(c)
(
∀c ∈ S ∀e = [i, j] ∈ ε(c) such that j ∈ E coin Xe = Xe′
′ ′ coin ′ ′
(9)
∀j such that [j, j ] ∈ H ∀e = [i, j ] ∈ E(S) Ye = Ye′

Xe , Ye ∈ {0, 1} (10)
Decision variable Xe = 1 if edge e contributes to a Two-In-a-Row violation,
zero otherwise, and decision variable Ye = 1 if edge e contributes to a Two-In-
a-Day violation, zero otherwise.
For all cliques in S, an edge cannot be counted twice for Two-In-a-Row and
Two-In-a-Day violations. These disjunctions are enforced using Equations (2).
For every clique, minimum numbers of edges involved are enforced using
Equations (3)-(5) (see Section 3.1). For every clique, minimum contributions to
penalties are enforced using Equations (6)-(8).
Since the endpoints of edges are exams, we can consider Coincidence con-
straints, if there are any, between exams in selected cliques. Coincidence con-
straints are enforced using Equations (9). First, consider two exams j and j ′
subject to a Coincidence constraint (i.e. [j, j ′ ] ∈ H coin ). Next, consider an
exam i, an edge e = [i, j], and assume that edge e is selected for one of the
spacing violations: edge e′ = [i, j ′ ] also has to be selected.
A preprocessing step proposed in [6] ensures that the two edges e = [i, j] and
e′ = [i, j ′ ] exist by propagating Coincidence constraints on the conflict graph
G(E, A). Note that the edges added by applying this preprocessing step have
no students in common (wij = 0) and have no impact in relation to penalties.
Equation (1) assesses the minimum weighted sum of edges e ∈ E(S) while
respecting for every clique c the minimum numbers of Two-In-a-Row, Two-In-

11
a-Day, and Two-In-a-Row AND Two-In-a-Day soft constraints violations, the
minimum contributions to penalties, and the Coincidence constraints between
exams.
An illustrative example for a conflict graph G(E, A) with exams
{A, B, C, D, E, F, G, H} is shown top left in Figure 2. We assume weights
w2R = 10 and w2D = 1. There is a common edge [A, B], weighted by
wAB = 1, between cliques c and c′ , that cannot be counted twice, and there is
a coincidence constraint between exams E and G. For the sake of explanation
cliques c and c′ are similar, the exams {F, G, H} playing the same role in clique
c′ as the exams {C, D, E} in clique c (see also Figure 1).

E coincidence G Propagation of
E H E coincidence H
100 A 100 0
100 0
1 100 1
100 100
D 2 1 2 G D 0 G

100 100 0
1 1
100 100
B
100 100
C F C F
Clique c Clique c’
An optimal allocation of exams Three optimal allocations of exams of cliques
of clique c within days: with respect to coincidence constraint:
A B C D E value: 32 G F H
An optimal allocation of exams A B E D C
of clique c’ within days:
A B F G H value: 32 H G F
A B D E C LB: 63
An optimal allocations of exams of cliques
without respecting coincidence constraint:
F H G
F G H A B C D E
A B C D E LB: 54

Figure 2: Example for Two-In-a-Row and Two-In-a-Day lower bound.

In Figure 2, consider the feasible optimal allocations of exams in cliques c

and c′ (on the left, below the conflict graph). The optimal penalty is 32 for
both cliques (see also Figure 1).
The greedy heuristic builds a family of edge-disjoint cliques composed of one
of the two cliques with a LB value of 32 for the Two-In-a-Row and Two-In-a-
Day (C 2R + C 2D ) lower bound.
Without considering the coincidence constraint, our formulation computes a
lower bound with a value of 54 = 10 + 22 + 22, which breaks down as follows: 10
for exams {A, B} allocated to D2, 22 for exams {C, D, E} allocated to D3, and
22 for exams {F, G, H} allocated to D3. These values correspond to an optimal
allocation of the exams in the two cliques, as shown bottom left in Figure 2.
Our formulation therefore gives a better lower bound.

12
 With preprocessing, we now consider the coincidence constraint between
exams E and G.
Figure 2 shows (top right) the new edges we add to the conflict graph. For
example, edge [E, H] with wEH = 0 is added to the conflict graph. There are
no students taking both exams E and H, but exams E and G are subject to
a coincidence constraint, and some students are taking both exams G and H,
meaning that exams E and H cannot be placed in the same period.
Let us consider Equations (9), and focus first on the Xe variables asso-
ciated with a Two-In-a-Row soft constraint violation. All of the equations
X[D,E] = X[D,G] , X[C,E] = X[C,G] , X[G,H] = X[E,H] , X[F,G] = X[E,F ] must
be enforced (edges e and e′ ). For example, if edge e = [D, E] is used to as-
sess a soft constraint violation (X[D,E] = 1), edge e = [D, G] must also be
used (X[D,G] = 1), which enforces the relative positions of exams {D, E, G} in
relation to Two-In-a-Row soft constraint violations. Using Equations (9), we
also enforce the relative positions of exams in relation to Two-In-a-Day soft
constraint violations. By Equations (2) we cannot have Xe = Ye . Hence, the
coincidence constraint is enforced without explicitly considering days or periods.
Considering the coincidence constraint, our formulation computes a lower
bound with a value of 63 = 10 + 22 + 31, which is a better lower bound.
In Figure 2 (bottom right) there are three examples of optimal allocations
of exams {A, B, C, D, E, F, G, H} with respect to the coincidence constraint
between exams E and G.
We denote as B the generic formulation that consists of Equations (1) to (10).
Numbers of variables are of the order of the number of edges O(|A|), and
numbers of constraints are of the order of the numbers of edges and cliques
O(|A| + |S|).
Our proposed generic formulation B can be used to compute lower bounds for
either C 2R , C 2D , or C 2R + C 2D . Cliques with edges in common can contribute,
so no family of edge-disjoint cliques built using a greedy heuristic would be
able to provide a better evaluation. The formulation also allows Coincidence
constraints to be enforced.

4 New formulation of spacing soft constraints

In this section we recall previous contributions that sought to reduce numbers
of hard constraints. We propose compact formulations for Two-In-a-Row and
Two-In-a-Day soft constraints, and we compare sizes of MIP formulations in
Table 1.
Although it is able to solve smaller instances, the formulation proposed in
[23] requires a large amount of memory. Formulations proposed in [6, 18] aim
at reducing the numbers of hard and soft constraints. We denote M the formu-
lation presented in [6], F ′ the formulation proposed by [18] and O the original
formulation described in [23].
In practice it was observed by Fonseca et al. [18] for formulation F ′ , and

13
by Arbaoui et al. [6] for formulation M, that these formulations could not
easily be run on the current generation of solvers. Instantiating sizeable real
problems with Two-In-a-Row or Two-In-a-Day soft constraints requires too
much memory. Formulations O, F ′ and M have as many Two-In-a-Row or
Two-In-a-Day soft constraints and variables as there are edges in the conflict
graph G(E, A).
The equations used by all previous formulations for assessing the C 2R and
2D
C terms are of two kinds. The first kind, based on edges [i, j] ∈ A such that
wij 6= 0, detects whether a violation occurs within days, while the second counts
the numbers of students involved.
Based on the work of Glover [19], we propose a more compact formulation
for Two-In-a-Row and Two-In-a-Day soft constraints.
Considering the exams that conflict with a given exam i, i.e. the neighbors
N (i) of i in the conflict graph G(E, A), the idea is to first compute an upper
bound (UB) of the number of students that can be counted if Two-In-a-Row or
Two-In-a-Day soft constraint violations occur. This UB can then be used to
aggregate the assessments of these soft constraints.
For an exam i, the value Ti is an upper bound of the number of students to
be counted:
X
∀i ∈ E Ti = wij (11)
j∈N (i)
If exam i is not allocated to a period p, then at most all the j ∈ N (i) exams
will be allocated to p. Assuming exam i is allocated to a period q 6= p such that
Two-In-a-Row or Two-In-a-Day soft constraint violations occur, then Ti is an
upper bound of the number of students to be counted.
We denote the number of exams as nE , the number of periods as nP , and
the number of days as nDay .

4.1 Revisiting the Two-In-a-Row soft constraints

For the Two-In-a-Row spacing soft constraint, the idea is to assess all the soft
constraint violations that occur when an exam i is allocated to a period p on a
certain day and some of its neighbors j ∈ N (i) are allocated to period p + 1 on
the same day. We propose the following formulation:
XX
C 2R = w2R Rip (12)
i∈E p∈P

∀i ∈ E ∀p ∈ P yp(p+1) = 1 

(13)
X
P P
wij Xj(p+1) − Ti (1 − Xip ) ≤ Rip 

[i,j]∈A
P
where Boolean decision variables Xip = 1 iff exam i is in period p, zero other-
wise. The variable Rip ∈ N assesses the number of students to be counted for
the Two-In-a-Row penalty if exam i is allocated to period p, and some exams

14
j ∈ N (i) are allocated to period p + 1 on the same day (parameter yp(p+1) = 1).
Considering Equations (13), if some exams P j ∈ N (i) are allocated to period
P P
p + 1, for these exams we have Xj(p+1) = 1 and [i,j]∈A wij Xj(p+1) ≤ Ti . When
P
Xip = 0, exam i is not allocated to period p, and we have Rip = 0 since we
are minimizing the C 2R term and Rip ∈ N. When Xip P
= 1, exam i is allocated
to period p, and all Two-In-a-Row penalties are counted for all adjacent exams
(i.e. [i, j] ∈ A) allocated to period p + 1. Hence, Equation (12) counts the C 2R
term.
We now have O(nE nP ) Equations (12) and (13) compared to O((nE )2 (nP )2 )
for formulations O, F ′ and M (see [6, 18, 23]) using O(nE nP ) integer variables.
The number of integer variables can be further reduced to O(nE ). An exam
i is allocated to a single period p, and therefore all but one of Rip are null.
Equations (12) and (13) can be advantageously rewritten as:
X
C 2R = w2R Ri (14)
i∈E

∀i ∈ E ∀p ∈ P yp(p+1) = 1 

(15)
X
P P
wij Xj(p+1) − Ti (1 − Xip ) ≤ Ri 

[i,j]∈A
P P
P
We have one Xip= 1 and [i,j]∈A wij Xj(p+1) ≤ Ri , meaning that one single
P
Equation (15) may set Ri ≥ 0, and for all other Equations (15) Xiq = 0 (p 6= q).
Consequently, all Equations (15) are valid.
For Two-In-a-Row soft constraints, formulations O, F ′ and M require
O((nE )2 ) Boolean variables, while we have a single integer variable Ri for each
exam i. Earlier formulations have O((nE )2 (nP )2 ) equations, while we now have
O(nE nP ) equations.

4.2 Revisiting the Two-In-a-Day soft constraints

For the Two-In-a-Day spacing soft constraint, the idea is to assess all the soft
constraint violations that occur when an exam i is allocated to a period p on a
certain day and some of its neighbors j ∈ N (i) are allocated to periods q > p+1
on the same day.
We denote as Di ∈ N the variable that assesses the Two-In-a-Day penalties
for an exam i allocated to a period p and some exams j ∈ N (i) allocated to
periods q > p + 1 such ypq = 1 (same day). We propose:
X
C 2D = w2D Di (16)
i∈E

∀i ∈ E ∀p ∈ P 

(17)
X X
P P
wij Xjq − Ti (1 − Xip ) ≤ Di 

q∈P, q>p+1, ypq =1 [i,j]∈A

15
 For Equations (17), if some exams j ∈ N (i) are allocated to a period q >
P
p + 1 such that ypq = 1 (same day), for these exams we have Xjq = 1 and
P P
P P
q∈P, q>p+1, ypq =1 [i,j]∈A wij Xjq ≤ Ti . When Xip = 0, exam i is not allocated
to period p, we have Di = 0, since the C 2D term is minimized and Di ∈ N.
P
When Xip = 1, exam i is allocated to period p, and all Two-In-a-Day penalties
are counted for all adjacent exams allocated to periods q on the same day (i.e.
ypq = 1), but not back-to-back (such that q > p + 1 ) with period p. A single
period p is used to allocate exam i, one Equation (17) may set Di ≥ 0, and all
other Equations (17) are valid. Therefore, Equation (16) counts the C 2D term.
For Two-In-a-Day soft constraints, formulations O and M require O((nE )2 )
Boolean variables, formulation F ′ requires O(nE nDay ) additional Boolean vari-
ables, while we have a single integer variable Di for each exam i. We have
O(nE nP ) Equations (16) and (17), compared to O((nE )2 (nP )2 ) Equations for
formulations O and M. Formulation F ′ requires O(nE nDay ) additional equa-
P
tions to link the additional Boolean variables to Boolean decision variables Xip
E 2 Day
and O((n ) n ) Equations for assessing, while our proposed formulation uses
O(nE nP ) Equations (16).
We denote as L the formulation that consists of equations of formulation M
for hard constraints, and we use Equations (14) to (17) to assess the Two-In-a-
Row or Two-In-a-Day soft constraints.

4.3 Comparing formulations

Earlier formulations could not easily be run because of memory overload. Ta-
ble 1 compares numbers of variables, numbers of Hard constraints, and numbers
of Soft constraints to assess C 2R and C 2D terms.
Let nS denote the number of students, nR the number of rooms, nDur the
number of different durations for exams, nHaf t the number of pairs [i, j] of ex-
ams that correspond to after constraints, nEaf t the number of exams subject to
after constraints, nHcoin the number of pairs [i, j] of exams that correspond to
coincidence constraints, nEcoin the number of exams subject to coincidence con-
straints, nHexcl the number of pairs [i, j] of exams that correspond to exclusion
constraints, nEsole the number of exams subject to room exclusive constraints.
P
All the formulations have the same numbers of Boolean variables Xip (an
R PR
exam allocated to a period), Xir (an exam allocated to a room), and Xipr (an
exam allocated to a period and a room) used to enforce hard constraints.
For Conflict constraints (see H-Constraints), formulation M has O(nE nP )
equations, whereas formulations O and F ′ have O(nS nP ) equations. In general
nE ≪ nS , formulation M has fewer Conflict constraints.
After and Coincidence hard constraints: formulation F ′ introduced O(nE )
new integer variables (see Integer Hard), and O(nE ) additional constraints are
P
required to link these variables to Xip (see H-Constraints and Link). Formula-
tion O uses O(n (n ) ) and O(nHcoin nP ) constraints, formulation F ′ uses
Haf t P 2

O(nHaf t ) and O(nHcoin ) constraints, and formulation M uses O(nEaf t nP ) and

O(nEcoin nP ) constraints without introducing additional variables.

16
 O F′ M L
Variables
nE nP nE nP nE nP nE nP
Boolean Hard +nE nR +nE nR +nE nR +nE nR
+nE nP nR +nE nP nR +nE nP nR +nE nP nR
Integer Hard nE

(nE )2 (nE )2 (nE )2

Boolean Soft +nDur nP nR +nDur nP nR +nDur nP nR nDur nP nR
+nE nDay
Integer Soft nP nR nP nR nP nR nP nR
+nE
H-Constraints
Conflict nS nP nS nP nE nP nE nP
Link nE
After nHaf t (nP )2 nHaf t nEaf t nP nEaf t nP
Coincidence nHcoin nP nHcoin nEcoin nP nEcoin nP
Exclusion nHexcl nP n Hexcl nP
Room Exclusive nE nEsole nP nR nE nEsole nP nR nEsole nP nR nEsole nP nR
S-Constraints
C 2R (nE )2 (nP )2 (nE )2 (nP )2 (nE )2 (nP )2 nE nP
C 2D (nE )2 (nP )2 (nE )2 nDay (nE )2 (nP )2 nE nP
+nE nDay

Table 1: Comparing O, F ′ , M and L formulations.

Formulation M is chosen because it has smaller numbers of hard constraints.

Existing formulations have as many spacing soft constraints as the number
of edges in the conflict graph G(E, A) for assessing the Two-In-a-Row or Two-
In-a-Day terms. Formulation L has as many spacing soft constraints as the
number of exams.
Table 1 shows that the proposed formulation L has fewer variables and
fewer constraints than existing formulations, and so less memory is required to
instantiate sizeable real-world problems.

5 Results
In this section we present results obtained on the ITC2007 and Yeditepe in-
stances. The lower bound results for Two-In-a-Row and Two-In-a-Day soft
constraints are presented. These results were obtained using sets of maximal
cliques. We also use the sum of two terms to compare the proposed formulation
L to existing formulations O, F ′ and M, since both terms aim at preventing
students sitting more than one exam per day.
Tests were done using a CPLEX 12.5 MIP solver with a single thread and
MipEmphasis parameter set to feasibility, using C++ compiled with gcc ver-

17
sion 4.4.7, on a machine with an Intel Xeon E5-2670 and 8 GB of RAM. The
computing times are reported in seconds.

nE nHcoin |A| ω LR LD LRD

1 607 2 10308 20 29 - 29
2 870 4 4466 15 24 26 13
3 934 82 13887 21 24 24 12
4 273 4 5792 17 14 14 7
5 1018 19 4890 13 28 28 14
6 242 18 2293 13 8 - 8
7 1096 13 12102 16 40 - 40
8 598 5 9213 17 41 79 40
9 169 2 1193 10 13 - 13
10 214 53 2340 18 22 22 12
11 934 82 13887 21 17 18 9
12 78 2 635 12 7 - 7
yue20011 126 0 1397 14 12 - 12
yue20012 141 0 1792 17 12 - 12
yue20013 26 0 84 6 4 - 4
yue20021 162 0 2320 16 14 - 14
yue20022 182 0 2781 20 14 - 14
yue20023 38 0 145 6 4 - 4
yue20031 174 0 2323 14 12 - 12
yue20032 210 0 3046 16 12 - 12
An LB computable: 15/20

Table 2: Characteristics of ITC2007 and Yeditepe instances. LR , LD and LRD

limits.

Characteristics of ITC2007 and Yeditepe datasets

Table 2 shows characteristics of the ITC2007 and Yeditepe datasets where
preprocessing is done. Preprocessing is performed in less than two minutes for
each instance (see [6]).
We report instance labels, numbers of exams nE and numbers of Coincidence
constraints nHcoin . Column |A| reports numbers of edges of G(E, A) conflict
graphs, and column ω reports maximum clique sizes. We use the code presented
in [27] to compute maximum clique sizes and sets of maximal cliques used to
compute lower bounds. Columns LR , LD and LRD report the limits of sizes of
cliques above which there is at least one violation of the Two-In-a-Row, Two-
In-a-Day, or Two-In-a-Row AND Two-In-a-Day soft constraints (see Section
3.1).
Instances 1, 6, 7, 9 and 12 have no D3 and no D4 days, and so these
instances have no Two-In-a-Day soft constraint. None of the Yeditepe instances
has a Two-In-a-Day soft constraint. For these instances, column LD contains -.
Instances 1, 5, 7, 8 and 9 are such that ω is less than each limit, so we cannot
compute a lower bound.
Instance labels for which a lower bound can be computed are in bold print.

Two-In-a-Row & Two-In-a-Day lower bounds

Computable lower bounds are given in Table 3. Columns |A| and |S| report

18
the number of edges, and the size of the set of maximal cliques. For every
sizeable ITC2007 instance, the set of maximal cliques S is computed in less
than two minutes, then, for all selected cliques, we compute T R (c), T D (c) and
T RD (c) contributions using the MIP summarized in Section 3.2 in less than one
hour. For model B the number of variables is of the order of O(|A|) and the
number of constraints is of the order of O(|A| + |S|) (see Section 3.2).
The first work on lower bounds was applied on the first eight ITC2007 in-
stances, and column LB [5] reports the values for instances 2, 3, 4 and 6, for
which C 2R + C 2D lower bounds can be computed.
B
|A| |S| LB [5] U Bh LB t
2 4466 21 10 10 10 10
3 13887 853 330 3300 670 5
4 5792 9935 291 16369 1620 9694
6 2293 12638 1740 5500 2600 70
10 2340 8 - 0 0 3
11 13887 2262 - 13890 3970 5821
12 635 4501 - 2613 2030 3
yue20011 1397 249 - 52 19 10
yue20012 1792 431 - 119 30 18
yue20013 84 40 - 29 13 4
yue20021 2320 139 - 116 14 15
yue20022 2781 151 - 195 34 23
yue20023 145 41 - 56 20 6
yue20031 2323 187 - 156 24 15
yue20032 3046 637 - 430 74 66

Table 3: Results for lower bounds on ITC2007 and Yeditepe instances.

To obtain an upper bound on C 2R + C 2D terms we used UniTime 4.0 [1],

while setting the weights of the other terms in the objective function to zero.
The method we used is an improved version of the work presented by the winner
of the ITC2007 examination timetabling track [26]. Column U Bh reports the
values. Columns LB and t report results using formulation B and computing
times.
Optimality is attained by formulation B for all instances. Other than for in-
stances 4 and 11, the computing times for formulation B are short. For instances
2 and 10, optimal values for the C 2R + C 2D terms are obtained (LB = U Bh ).
The lower bound given by formulation B is a strict improvement on the
results presented in [5] (by up to 456% for instance 4). For all Yeditepe instances,
a LB can be quickly computed. Knowing a minimum cost for these spacing soft
constraints is of practical interest. Instances yue20013 and yue20023 can be
optimally solved with costs of 29 and 56 respectively (see [28]), we notice that
the gaps are 55% and 64%.
Examination timetabling problems are hard problems that have been widely
investigated in the literature using heuristic methods. Results on benchmarks
are regularly improved applying these methods.
Approaches for computing lower bounds for Two-In-a-Row and Two-In-a-
Day soft constraints have been investigated less extensively. Bridging the gap

19
between theory and practice remains challenging. As a comparison of columns
LB and U B shows, there is still room for improvement.
We proposed an exact method that provides better results than selecting
edge-disjoint cliques applying a greedy heuristic. The proposed generic for-
mulation B uses minimum numbers of spacing soft constraint violations, and
contributions of cliques. Formulation B ensures that each edge is used once,
while enforcing coincidence constraints on exams. For spacing soft constraints
the formulation allows lower bounds to be computed optimally. Results are
better than previously proposed lower bounds.

Comparing formulations O, F ′ , M and L on C 2R + C 2D

Results for formulations O, F ′ , M and L are given in Table 4. Instance
labels are shown in the first column, and column U Bh reports values obtained
after letting the [1] code run for one hour. Columns U B and t show values
obtained (UB) and computing times (t) for formulations. A one-hour time limit
is used, and where this limit is reached the computing time is shown as -. Where
no solution can be found, and where the 8 GB size limit is reached, the values
NF (Not Found) and OOM (Out-Of-Memory) are shown respectively. For
formulations O, F ′ , M and L, the best results are shown in bold print.
O F′ M L
U Bh UB t UB t UB t UB t
1 0 NF - NF - NF - NF -
2 10 OOM 590 - 140 - 235 -
3 3300 OOM NF - NF - NF -
4 16369 NF - NF - NF - NF -
5 0 0 78 135 - 0 77 0 1834
6 5500 5460 - 4910 - 5160 - 4580 -
7 0 OOM 0 685 0 687 0 1414
8 0 0 231 0 1246 0 300 0 410
9 0 0 21 0 735 0 17 0 246
10 0 0 22 0 204 0 54 0 242
11 13890 OOM NF - NF - NF -
12 2913 4725 - 3185 - 3675 - 3115 -
yue20011 52 158 - 112 - 84 - 73 -
yue20012 119 298 - 184 - 190 - 162 -
yue20013 29 29 5 29 4 29 2 29 3
yue20021 116 182 - 243 - 149 - 125 -
yue20022 195 7046 - NF - 396 - 365 -
yue20023 56 56 9 56 - 56 5 56 46
yue20031 156 274 - 316 - 322 - 255 -
yue20032 430 8386 - 657 - 648 - 550 -

Table 4: Comparison of formulations O, F ′ , M and L on ITC2007 and Yeditepe

instances.

In [6] we investigated cliques and Data-dependent Dual-Feasible Functions

(DDFF) valid inequalities and we achieved better results. Formulation L also
uses these valid inequalities.
Formulations F ′ and M proposed in [6, 18] aim at reducing numbers of
hard and soft constraints. The table shows that using 8 GB, the improved
formulations did not cause any out-of-memory events. These formulations can

20
be run on standard computers.
For instances 1, 3, 4 and 11, no solution is attained within the time limit,
regardless of the considered formulation. For instances 5, 7, 8, 9 and 10, all
compact formulations gave solutions, with no students taking two exams on the
same day (zero cost), and the solutions were obtained in no more than half an
hour of computing time.
For instances 6 and 12 of the ITC2007 dataset, better results are attained
using formulation L. For all Yeditepe instances, formulation L gives either opti-
mal solutions or results that are better than those given by other formulations.
Formulation M gives the best result only for instance 2. Formulation F ′ does
not give a solution for the yue20022 instance.
Given the difficulty of timetabling problems, running MIP formulations on
sizeable real problems can still be challenging. However, comparing the results
of heuristic and formulation L, we have U Bh = 5500 and U B = 4580 for
instance 6. A better result is achieved within the one hour time limit.
Table 4 clearly shows that from formulation O to formulation L improve-
ments were obtained. The proposed formulation L reduces the number of vari-
ables and constraints for Two-In-a-Row and Two-In-a-Day spacing soft con-
straints. It is much more compact and can be run on current generations of
solvers to provide results within a reasonable computing time.

6 Conclusion
In this paper we proposed improvements to lower bounds and new mathematical
formulations for the spacing soft constraints that seek to prevent students sitting
more than one exam per day.
We proposed a generic MIP model that computes lower bounds by optimally
selecting edges of a set of cliques using their individual contributions while taking
into account hard constraints relating to exam coincidence. Better lower bounds
were achieved than in previous results in the literature.
In existing mathematical formulations of spacing soft constraints the number
of equations is of the order of the square of the number of exams. We proposed
new formulations of the spacing soft constraints that use as many soft constraints
as the number of exams. Our new formulations yield a model that is more
compact and better results were achieved compared to existing mathematical
formulations.

Acknowledgment
This work was carried out in the framework of the Labex MS2T, which was
funded by the French Government, through the program “Investments for the
future” managed by the National Agency for Research (Reference ANR-11-
IDEX-0004-02). We would also like to thank the referees for their insightful
comments that helped us improving the quality of this paper.

21
References
[1] University Timetabling, comprehensive academic scheduling solutions.
http://www.unitime.org/, April, 2018.
[2] Syariza Abdul-Rahman, Edmund K. Burke, Andrzej Bargiela, Barry Mc-
Collum, and Ender Özcan. A constructive approach to examination
timetabling based on adaptive decomposition and ordering. Annals of
Operations Research, 218(1):3–21, 2014.

[3] Salwani Abdullah, Hamza Turabieh, Barry McCollum, and Paul McMul-
lan. A tabu-based memetic approach for examination timetabling prob-
lems. In Proceedings of the 5th International Conference of Rough Set and
Knowledge Technology RSKT 2010, Beijing, China, October 15-17, volume
6401 of Lecture Notes in Computer Science, pages 574–581, 2010.

[4] Malek Alzaqebah and Salwani Abdullah. Hybrid bee colony optimization
for examination timetabling problems. Computers & Operations Research,
54:142–154, 2015.

[5] Taha Arbaoui, Jean-Paul Boufflet, and Aziz Moukrim. An analysis frame-
work for examination timetabling. In Proceedings of the Sixth International
Symposium on Combinatorial Search (SoCS 2013), pages 11–19, July 2013.
Leavenworth, WA, USA.

[6] Taha Arbaoui, Jean-Paul Boufflet, and Aziz Moukrim. Preprocessing and
an improved MIP model for examination timetabling. Annals of Operations
Research, 229:1–22, 2015.

[7] Michele Battistutta, Andrea Schaerf, and Tommaso Urli. Feature-based

tuning of single-stage simulated annealing for examination timetabling.
Annals of Operations Research, 252(2):239–254, 2017.

[8] Edmund K. Burke and Yuri Bykov. The late acceptance hill-climbing
heuristic. European Journal of Operational Research, 258(1):70–78, 2017.

[9] Edmund K. Burke, Michel Gendreau, Matthew Hyde, Graham Kendall,

Gabriela Ochoa, Ender Özcan, and Rong Qu. Hyper-heuristics: A sur-
vey of the state of the art. Journal of the Operational Research Society,
64(12):1695–1724, 2013.

[10] Edmund K. Burke, Kirk Jackson, Jeff Kingston, and Rupert Weare. Auto-
mated university timetabling: The state of the art. The Computer Journal,
40(9):565–571, 1997.
[11] Edmund K. Burke, James P. Newall, and Rupert F. Weare. A memetic
algorithm for university exam timetabling. In international conference on
the practice and theory of automated timetabling, pages 241–250. Springer,
1995.

22
[12] Edmund K. Burke and Sanja Petrovic. Recent research directions in auto-
mated timetabling. European Journal of Operational Research, 140(2):266–
280, 2002.
[13] Yuri Bykov and Sanja Petrovic. A step counting hill climbing algorithm
applied to university examination timetabling. Journal of Scheduling,
19(4):479–492, 2016.

[14] Jacques Carlier, François Clautiaux, and Aziz Moukrim. New reduction
procedures and lower bounds for the two-dimensional bin packing problem
with fixed orientation. Computers & Operations Research, 34(8):2223–
2250, 2007.

[15] Michael W. Carter and Gilbert Laporte. Recent developments in practical

course timetabling. In Proceedings of the Second International Conference
of Practice and Theory of Automated Timetabling, volume 1408 of Lecture
Notes in Computer Science, pages 3–19. 1998.

[16] Michael W. Carter, Gilbert Laporte, and Sau Yan Lee. Examination
timetabling: Algorithmic strategies and applications. Journal of the
Operational Research Society, 47(3):373–383, 1996.

[17] Meryem Cheraitia and Salim Haddadi. Simulated annealing for the
uncapacitated exam scheduling problem. International Journal of
Metaheuristics, 5(2):156–170, 2016.

[18] George H. G. Fonseca and Haroldo G. Santos. A New Integer Linear Pro-
gramming Formulation to the Examination Timetabling Problem. In The
6th Multidisciplinary International Conference on Scheduling: Theory and
Applications (Mista 2013), pages 345–355, August 2013. Gent, Belgium.

[19] Fred Glover. Improved linear integer programming formulations of nonlin-

ear integer problems. Management Science, 22(4):455–460, 1975.

[20] Christos Gogos, Panayiotis Alefragis, and Efthymios Housos. An Im-

proved Multi-staged Algorithmic Process for the Solution of the Examina-
tion Timetabling Problem. Annals of Operations Research, 194:203–221,
2012.

[21] Graham Kendall and Naimah Mohd Hussin. A tabu search hyper-heuristic
approach to the examination timetabling problem at the mara university
of technology. In International Conference on the Practice and Theory of
Automated Timetabling, pages 270–293. Springer, 2004.
[22] Barry McCollum, Paul McMullan, Edmund K. Burke, Andrew J.
Parkes, and Rong Qu. The Second International Timetabling
Competition: Examination Timetabling Track. Technical Report
QUB/IEEE/TECH/ITC2007/Exam/v4.0, Queen’s University, Belfast,
2007.

23
[23] Barry McCollum, Paul McMullan, Andrew J. Parkes, Edmund K. Burke,
and Rong Qu. A New Model for Automated Examination Timetabling.
Annals of Operations Research, 194:291–315, 2012.
[24] Liam T.G. Merlot, Natashia Boland, Barry D. Hughes, and Peter J.
Stuckey. A hybrid algorithm for the examination timetabling problem.
In International Conference on the Practice and Theory of Automated
Timetabling, pages 207–231. Springer, 2002.
[25] S.A. MirHassani. Improving paper spread in examination timetables using
integer programming. Applied Mathematics and Computation, 179(2):702
– 706, 2006.

[26] Tomás Müller. ITC2007 Solver Description: a Hybrid Approach. Annals

of Operations Research, 172:429–446, 2009.

[27] Patric R.J. Östergård. A fast algorithm for the maximum clique problem.
Discrete Applied Mathematics, 120(1–3):197 – 207, 2002.

[28] Andrew J. Parkes and Ender Özcan. Properties of yeditepe examina-

tion timetabling benchmark instances. In Proc. of the 8th international
conference on the practice and theory of automated timetabling (PATAT
2010), pages 531–534, August 2010. Queen’s University Belfast, Northern
Ireland.

[29] Rong Qu, Edmund K. Burke, Barry McCollum, Liam T.G. Merlot, and
Sau Y Lee. A survey of search methodologies and automated system devel-
opment for examination timetabling. Journal of scheduling, 12(1):55–89,
2009.

[30] Rong Qu, Fang He, and Edmund K. Burke. Hybridizing integer pro-
gramming models with an adaptive decomposition approach for exam
timetabling problems. The 4th Multidisciplinary International Scheduling:
Theory and Applications, pages 435–446, August 2009. Dublin, Ireland.

[31] Andrea Schaerf. A survey of automated timetabling. Artificial intelligence

review, 13(2):87–127, 1999.

[32] Amr Soghier and Rong Qu. Adaptive selection of heuristics for assigning
time slots and rooms in exam timetables. Applied Intelligence, 39:1–13,
2013.

[33] Gert Woumans, Liesje De Boeck, Jeroen Beliën, and Stefan Creemers. A
column generation approach for solving the examination-timetabling prob-
lem. European Journal of Operational Research, 253:178–194, 2016.

24