Você está na página 1de 28

Comput Manag Sci (2010) 7:19–46

DOI 10.1007/s10287-007-0066-8

ORIGINAL PAPER

A mixed-integer mathematical modeling approach


to exam timetabling

Salem M. Al-Yakoob · Hanif D. Sherali ·


Mona Al-Jazzaf

Published online: 1 December 2007


© Springer-Verlag 2007

Abstract This paper explores mathematical programming models for an exam


timetabling problem related to Kuwait University (KU). In particular, we consider
two subproblems: (a) the ExamTimetabling Problem (ETP), which is concerned with
assigning exams to designated exam-periods and classrooms, and (b) the Proctor
Assignment Problem (PAP), which deals with the assignment of proctors to exams.
While this exam timetabling problem is ubiquitous in many academic institutions
worldwide, idiosyncrasies of the problem related to gender-based policies and having
multiple exam centers at KU require novel models. A mixed-integer exam timetabling
programming model (ETM) is developed for ETP, which takes into account restric-
tions related to exam-period conflicts, facility and human resources, and commuting
and traffic considerations. Assuming that exam-periods are specified for all exams
as determined by ETM, another mixed-integer programming model is formulated
for PAP, denoted by PAM, which incorporates the proctors’ preferences for specific
days and exam-periods. Computational results are reported and analyzed for solv-
ing ETM and PAM directly using the CPLEX Optimization Software (version 9.0),
and for implementing a specialized sequential LP-based heuristic for solving PAM.
The results obtained significantly improve upon those derived via the existing manual
approach implemented at KU, in terms of eliminating conflicts as well as from the
overall efficiency and equity points of view.

S. M. Al-Yakoob (B) · M. Al-Jazzaf


Department of Mathematics and Computer Science, College of Science,
Kuwait University, P.O. Box: 5969, Safat 13060, State of Kuwait, Kuwait
e-mail: salem@al-yakoob.com

H. D. Sherali
Grado Department of Industrial and Systems Engineering (0118),
Virginia Polytechnic Institute and State University, Blacksburg, VA 24061, USA
e-mail: hanifs@vt.edu

123
20 S. M. Al-Yakoob et al.

Keywords Exam timetabling · Mixed-integer programming ·


Mathematical programming · Optimization

JEL Classification 90C90 · 90C11

1 Introduction

This paper is concerned with the development of mathematical programming models


for the Exam Timetabling Problem (ETP) and the Proctor Assignment Problem (PAP)
at Kuwait University, which is composed of thirteen colleges, each having a number of
departments. This problem, as well as many other academic scheduling and timetabling
problems such as class scheduling and faculty and teaching assistant scheduling, are
challenging and complex tasks in relatively large universities such as KU. The highly
combinatorial nature of such problems makes ad hoc manual approaches inefficient,
expensive, and often accompanied by serious deficiencies such as exam and class
conflicts, dissatisfaction of faculty members in regard to having consecutive exams
scheduled at two different campuses, clustering exams in certain exam-periods of some
days, as well as parking and traffic congestion. Deficiencies associated with the man-
ual scheduling and timetabling approaches are likely to negatively impact the overall
academic process at KU, which is currently further aggravated due to the increasing
number of accepted students, and certain newly imposed gender-based policies.
The exam timetabling problem at KU is concerned with assigning exams for dif-
ferent sections of offered classes to a limited number of time periods (exam-periods),
while specifying the classrooms (laboratories) for each assigned exam in a conflict-
free fashion. Exams are held in different campuses at KU and, hence, the travel time
between campuses must be taken into consideration. Due to the recent implementa-
tion of a newly imposed gender-based policy, some exams are for female sections,
some are for male sections, and the rest are for joint sections. Also, we need to assign
proctors to exams, while taking into consideration restrictions and preferences of
proctors related to specific exam-periods. For example, some proctors are graduate
students who have final exams during certain exam-periods. Furthermore, due to the
gender-based policy, it is sometimes more convenient to assign female proctors to
female sections. Also, exams need to be reasonably spread over all the available exam-
periods in order to avoid the unavailability of classrooms and proctors, and to provide
students with sufficient time between exams. Hence, the overall exam timetabling and
proctor assignment problem at KU is addressed via two subproblems, denoted by ETP
and PAP, where the first is concerned with assigning exams to available exam-periods
and classrooms, and the second deals with assigning proctors to exams given the exam
time-periods as determined by the solution to Problem ETP.
The exams need to be scheduled while satisfying certain constraining rules and
regulations pertaining to facility and human resources, as well as maintaining certain
features concerning the offering patterns of exams. Some pertinent examples are given
next.
(a) Availability of classrooms
The total number of exams scheduled during a given exam-period cannot exceed
the available facilities such as classrooms (laboratories).

123
A mixed-integer mathematical modeling approach to exam timetabling 21

(b) Location of campuses


Exams of classes that are taken concurrently by students should not be scheduled
consecutively over different campuses. Consecutive scheduling of exams of such clas-
ses often lead to traffic congestion, parking problems, and considerable inconvenience
for students.
(c) Avoiding exam conflicts
Exam conflicts occur if exams of two classes that are taken concurrently by a group
of students are assigned to the same exam-period. Naturally, this should be avoided.
(d) Exam spreading
Exams should be spread over the available exam-periods in an efficient manner
allowing students reasonable time durations between exams and avoiding the cluster-
ing of exams over certain exam-periods.
(e) Consideration of proctors’ preference
Proctors’ preferences for certain exam-periods should be taken into consideration.
Dissatisfaction of proctors and discrepancies in satisfaction levels among proctors is
very common. Hence, proctors’ preferences need to be quantified and incorporated in
the developed models.
(f) Gender-related policies
It is sometimes desirable or required to assign female proctors to exams corre-
sponding to female sections.
The combinatorial nature of the overall problem due to the large number of clas-
ses offered over different campuses (exam centers), and involving male, female, and
joint classes, make the current manual timetabling of exams at KU inefficient and
problematic as discussed in the next section.
Currently, the assignment of exams to exam-periods at KU is performed by indi-
vidual colleges. This approach lacks proper coordination efforts between the different
colleges, which is imperative in order to avoid exam conflicts and to enhance exam
offering patterns. In general, generating exam schedules for each college is achieved
in five stages as follows.

(a) The Office of the Assistant Dean for Students’ Affairs prepares basic statistics
concerning the exams of the college. These statistics include: number of exams,
pertinent departments, number of students in classes, facility resources (rooms
and laboratories), etc.
(b) Based on the statistics obtained in Part (a), exams for the various classes are
assigned to exam-periods in two steps. First, exams of multiple-section classes
are scheduled, taking into consideration facility resources. Second, exam-periods
for the remaining classes are specified so that no two classes having different
time-slots are assigned to the same exam-period. This is done to avoid exam
conflicts, i.e., students having more than one exam during a given exam-period.
(c) Based on the statistics obtained in Part (a), subsets of classes are established, each
of which is composed of classes taken simultaneously by a group of students.
These subsets are used to avoid exam conflicts.
(d) Exams are assigned to exam-periods and suitable classrooms by the Office of
the Assistant Dean for Students’ Affairs. This assignment is performed in an ad
hoc fashion. Pertinent information is then sent to the individual departments.

123
22 S. M. Al-Yakoob et al.

(e) Finally, individual departments assign proctors to exams, without considering


proctors’ preferences for specific days and exam-periods. Although satisfying
such preferences of the proctors is a secondary goal to generating “good” con-
flict-free exam schedules, considering this issue in a modeling approach can
greatly enhance the quality and satisfaction of the generated schedules.

The current manual approach for generating exam schedules at KU has many draw-
backs and shortcomings, some of which are identified below.
(a) Exam conflicts
Exam conflicts often occur due to the lack of coordination among colleges that offer
classes taken by common groups of students. Also, the exam scheduling approach
outlined previously may lead to exam conflicts within the same college. Generat-
ing conflict-free exam schedules is an arduous combinatorial task that can be better
achieved via proper modeling and analytical tools.
(b) Exam-driven class registrations
A major disadvantage of the current approach is that students often sign-up for classes
based on the exam-periods of these classes. Hence, certain feasible combinations of
classes are precluded to prevent conflicts. This has led many students to take more
than the average number of years to graduate.
(c) Consecutive exams over two different campuses
Scheduling two exams that are taken by a common group of students consecutively
over two different campuses is undesirable, and results in traffic congestion, parking,
and delay problems.
(d) Inefficient utilization of facility resources
Manual scheduling approaches often lead to the inefficient use of classrooms and
parking facilities. For example, exams are often assigned to unsuitable classrooms in
terms of size.
(e) Dissatisfaction of proctors
Preferences of proctors for specific days and exam-periods are often overlooked by
the individual departments when assigning proctors to exams. This typically leads to
dissatisfaction among proctors.
The problem considered in this paper is of great importance to students, faculty
members, and decision-makers at KU, and as evident from the current exam schedul-
ing approach, generating effective exam schedules is a step forward in further enhanc-
ing the educational process at KU. Furthermore, this research effort makes a useful
contribution to the development of quantitative methods for tackling an academic
scheduling and timetabling problem that incorporates many novel features such as
proctors’ preferences, traffic considerations, and certain imposed gender policies.
The proposed mathematical models and solution algorithms can be used to generate
conflict-free exam schedules before or after the student registration period. The models
may be incorporated in a decision-support system to generate exam schedules having
many desirable features for KU that avoid the shortcomings identified above, as well
as for other academic institutions in the Gulf region having similar exam timetabling
environments.
The remainder of this paper is organized as follows. In Sect. 2, we review the liter-
ature related to academic timetabling and scheduling. In Sect. 3, we present modeling

123
A mixed-integer mathematical modeling approach to exam timetabling 23

preliminaries that will be used to formulate mixed-integer programming models in the


subsequent section to solve the overall exam timetabling problem at Kuwait Univer-
sity. In Sect. 4, we first formulate models for Problems ETP and PAP, to respectively
assign exams to exam-periods and classrooms, and to assign proctors to scheduled
exams. Computational results and a comparison of the proposed modeling approach
with the current manual scheduling approach are presented in Sect. 5. Finally, Sect. 6
closes the paper with a summary and some concluding remarks.

2 Related literature

Many mathematical programming and heuristic-based approaches have been proposed


in the literature for solving a variety of timetabling problems, although no specific
method can be applied universally due to the specifics of individual problems. For
comprehensive review papers, see for example, Burke et al. (1996, 2001); Burke and
Petrovic (2002); Carter and Laporte (1996); Sandhu (2001), and Schaerf (1999).
Exam timetabling problems faced by academic institutions differ from one institu-
tion to another. Although most institutions face a common task, which is to allocate
sessions and rooms to exams in a conflict-free fashion, each institution might need to
address some unique restriction depending on established policies. Burke et al. (1996)
survey several soft-constraints that have been used for modeling purposes at univer-
sities in the United Kingdom. In the remainder of this section, we present literature
pertinent to the exam timetabling problem.
Birbas et al. (1997) presented a 0-1 integer programming model for the timetabling
problem faced by Greek high schools. Other related models were formulated by
Daskalaki et al. (2004) and Daskalaki and Birbas (2005) for a timetabling problem
related to Greek universities. In the latter paper, a two-stage relaxation procedure is
utilized by the authors. In the first stage, the computationally burdensome constraints
are initially relaxed. In the second stage, the relaxed constraints are introduced afresh
and the timetable is generated on a day-by-day basis to find local optima to the original
problem.
Mirrazavi et al. (2003) developed a two-phase approach that first allocates rooms to
classes and second, assigns time-periods in which to teach the different classes. Both
phases have been modeled as integer goal programs, and use a preprocessing module
to aid the two-phase approach in removing redundancies prior to optimization.
Cheng et al. (2003) formulated a student timetabling problem as a multi-commodity
flow problem. The authors claimed that the formulated network model can be easily
solved by existing approximation methods for such problems.
The proctor assignment problem can be viewed as an employee timetabling prob-
lem. In general, employee timetabling problems are concerned with the assignment of
employees to specific shifts. In our problem, the employees are proctors and the shifts
are exam-periods. Several mathematical programming models have been proposed
for generating employee schedules. The assignment of instructors in executive devel-
opment programs was discussed by Mukherjee and Gilbert (1997) who proposed a
0-1 mixed-integer programming model. Because of the intractability of this model, four
heuristics were developed and investigated based on different Lagrangian relaxations

123
24 S. M. Al-Yakoob et al.

of the problem. Beaumont (1997a) formulated a mixed-integer programming model


and described the solution of a problem of scheduling a workforce so as to meet
demand that varies markedly with the time of day and moderately with the day of
week. In addition, Beaumont (1997b) has proposed another mixed-integer program-
ming approach for modeling cyclical rosters for scheduling staff employees. Mulli-
nax and Lawley (2002) developed a linear integer programming model that assigns
patients to nurses while balancing nurse workloads. Solutions generated by networks
of scheduling agents have been investigated by Meisels and Kaplansky (2003). The
agents implement an exhaustive search over a portion of the solution space and send
messages to a central agent, which then resolves global constraint violations. Meisels
et al. (1996) and Meisels and Kaplansky (2004) have studied approaches based on
representing employee timetabling problems as networks with side-constraints.
The satisfaction of proctors’ preferences to the extent possible is an essential
objective that one of our proposed models aims to achieve. Many models related
to employee preferences have been addressed in the literature. Al-Yakoob and Sherali
(2006a) developed mixed-integer programming models for the assignment of faculty
members to classes. These models aim to minimize the dissatisfaction levels of fac-
ulty members in a fair fashion. Also, Al-Yakoob and Sherali (2006b) have studied
an employee scheduling problem where subsets of employees need to be matched
to subsets of work centers based on employees’ expressed preferences for work
centers.
Although exam timetabling problems have been widely addressed in the literature,
none of these fully consider all the aspects that are pertinent to the problem under con-
sideration. As seen later in this paper, the issues related to gender policies, different
locations of campuses, preferences of proctors, and parking and traffic congestion lead
to mixed-integer formulations that require novel heuristic procedures to solve them.

3 Modeling preliminaries

3.1 Notation and assumptions

Let |A| be the cardinality of any set A. Let N be the set of days during which exams
will be held, indexed by n = 1, . . . , |N |. The specific value of |N |can be determined
based on the number of exams that must be taken and previous experience, or it can
be better derived via the proposed modeling approach by examining different values
of |N |.
Currently, there are four exam-periods in a given day specified as follows: 8:00
a.m. to 10:00 a.m., 11:00 a.m. to 1:00 p.m., 2:00 p.m. to 4:00 p.m., and 5:00 p.m.
to 7:00 p.m. Hence there are 4|N | total exam-periods. Let T denote the set of exam-
periods, indexed by t = 1, . . . , |T | = 4|N |, where t = 4n − 3, 4n − 2, 4n − 1 and
4n respectively index the exam-periods of day n ∈ N . Also, for n ∈ N , we define
 n = {4n − 3, 4n − 2, 4n − 1, 4n}.
Let G denote the set of colleges whose exams are to be scheduled, indexed by
g = 1, . . . , |G|, and let G1 = {(g 1 , g2 ) : g1 , g 2 ∈ G, g1 < g 2 , where g1 and g2 are
located in different campuses}. Let D g be the set of departments in college g ∈ G,

123
A mixed-integer mathematical modeling approach to exam timetabling 25

indexed by d = 1, . . . , |D g |. Let C g, d be the set of classes offered by department


d ∈ D g of college g ∈ G, indexed by c = 1, . . . , |C g, d | . The number of offered
sections of class c ∈ C g, d is denoted by O g, d, c . Let A g, d, c be the average number
of students that are registered in a section of a class c ∈ C g, d . The number A g, d, c
may be obtained by dividing the total number of students taking class c ∈ C g, d by
the number O g, d, c .
Note that a given section of a class may be composed of all female or male students,
or a combination of both. Let F g, d, c , M g, d, c and J g, d, c respectively denoted the
number of female, male, and joint sections of class c ∈ C g, d , for department d ∈ D g
of college g ∈ G.
For g ∈ G, let R g be the set of classroom types that are available at College g,
indexed by r = 1, . . . , |R g | . Note that a classroom type is characterized by its capac-
ity and perhaps other pertinent features. The capacity of a classroom of type r ∈ R g
is denoted by ω g, r , and the number of available classrooms of type r is given by
N g, r . For a given college g ∈ G, department d ∈ D g and class c ∈ C g, d , we define
K g, d, c ⊆ {1, . . . , |R g |} to be a subset of classroom types in which an exam for a
section of class c can be held. Note that, for example, an exam of a class having 50
students can be held in a classroom having a capacity of 100 seats. However, to the
extent possible, our proposed modeling approach will avoid such inefficient assign-
ments. Also, for a classroom of type r ∈ R g , let V g, d, r ⊆ {1, . . . , |C g, d |} be a subset
of classes offered by department d ∈ D g of college g ∈ G whose exams can be held
in a classroom of type r ∈ R g .
In the current practice, proctors for exams of a given department belong to the same
department. However, allowing proctors of one department to proctor exams for other
departments requires a simple modification to the proposed model PAM that will be
developed in Sect. 4 for assigning proctors to exams. Therefore, we discard college
and department indices when defining proctor-related notation. Hence, for example,
C g, d and O g, d, c become C and O c , respectively. Let P denote the set of proctors,
indexed by p = 1, . . . , |P|.
Partition the set P into P1 and P2 , where the former represents the set of proctors
who are graduate students and the latter gives the complement of P1 in P. For later
use, we let p = 1, . . . , |P1 | index the set of proctors who are graduate students and
we let p = |P1 | + 1, . . . , |P1 | + |P2 | = |P| index the set of proctors who are not
graduate students.
A proctor may prefer to avoid proctoring two exams consecutively. Let P3 denote
the set of proctors who do not wish to proctor two exams consecutively, and let P4
be the complement of P3 in P. Note that the maximum number of exams that can
be assigned to a proctor for any given day n ∈ N is two. Let L 1 and L 2 denote the
maximum number of exams that can be possibly assigned to a proctor belonging to
P1 and P2 , respectively, within the entire duration spanning all exam-periods. The
specific values of L 1 and L 2 are decided by the respective departments, depending
on considerations similar to whether a proctor is a graduate student or a full-time
employee, etc.
For exam-period t ∈ T , let P1t ⊆ P1 denote the set of proctors who are graduate
students and have exams during exam-period t. Accordingly, we let P2t be the com-
plement of P 1t in P1 , which represents the set of proctors who are graduate students

123
26 S. M. Al-Yakoob et al.

and available to proctor exams offered during exam-period t. Hence, the total number
of proctors that are available during exam-period t ∈ T is given by |P2t | + |P2 |.
Let T p ⊆ T denote the set of exam-periods for which proctor p is available. Note
that for p ∈ P 2 , T p = T , unless some exam-periods are blocked for other consider-
ations. For n ∈ N , let  n, p =  n T p denote the set of exam-periods of a given day
n ∈ N for which proctor p is available. Also, for c ∈ C, let P c ⊆ P denote the set of
proctors who are available during the exam-period of the exam associated with class c.
For exam-period t ∈ T , let C t ⊆ C denote the set of classes whose exams are offered
during exam-period t, and let C p,t ⊆ C denote the set of classes whose exams can be
monitored by proctor p during exam-period t. Note that for p ∈ P 2 , C p,t = C t . Let
P F ⊆ P and P M ⊆ P respectively, denote the sets of female and male proctors and
let M J ⊆ P denote the set of proctors who can proctor male and joint sections. Also,
define Q i = {( p1 , p 2 ) : where p1 , p 2 ∈ Pi with p1 < p 2 }, for i = 1, 2.

3.2 Special sets of classes

Exams of certain classes should not be held during the same exam-period to avoid
conflicts, or during the same day to enhance exam offering patterns. Accordingly, spe-
cial sets of classes are constructed and a penalty is associated with each set to reflect
the significance of not having exams of classes from this set scheduled during the
same exam-period or the same day, with a higher penalty value indicating a relatively
greater corresponding undesirability.
These special sets are defined as follows:
(A) We define S1 , . . . , Sm to be sets of classes from all colleges such that each set
contains classes that can be taken concurrently by a group of students, and are
formulated based on established student curricula. Note that the sets S1 , . . . , Sm
are not necessarily pairwise disjoint. For college g ∈ G and department d ∈ D g ,
g,d 
let Si = Si C g,d , ∀ i = 1, . . . , m. Let M Si denote the set of students who
are taking the classes in Si .
(B) For g ∈ G, d ∈ D g let U g,d = {(c1 , c2 ) : c1 , c2 ∈ C g,d , c1 < c2 , where both
classes c1 , and c2 are taught by the same faculty member and each requires the
presence of this faculty member during the exam}.

3.3 Proctors’ preferences and a dissatisfaction cost function

The preferences of the proctors for specific exam days and exam-periods are essen-
tial information that needs to be appropriately and fairly considered when assigning
proctors to exams. Hence, every proctor fills out a questionnaire pertaining to the pref-
erences for specific days and exam-periods and also the desirability to work during
two consecutive exam-periods. The questionnaire consists of three parts (A), (B) and
(C) as detailed below.
(A) General information
This part records preferences of proctors for working during two consecutive exam
periods.

123
A mixed-integer mathematical modeling approach to exam timetabling 27

(B) Preferences for specific days


Each proctor is instructed to provide a permutation of the set of days N . The first
day of this permutation represents the highest preference and the last day represents
the lowest preference. For later use, we associate the number “1” with the first choice,
“2” with the second choice, and so on. Let k1 ( p, n) be a positive integer number
representing the preference associated with proctor p ∈ P for day n ∈ N .
(C) Preferences for specific exam-periods
For a given exam day n ∈ N each proctor is required to give a permutation of the
set of exam-periods of day n. The first exam-period of this permutation represents the
highest preference and the last exam-period represents the lowest preference. For later
use, we associate the permutation (1, 2, 3, 4) for exam-periods corresponding to the
first choice of days, (5, 6, 7, 8) with the second choice, and so on. Let k2 ( p, n, t) be
a positive integer number representing the preference associated with proctor p ∈ P
for exam-period t ∈  n of day n ∈ N .
Let I = {( p, n, t) : p ∈ P, n ∈ N and t ∈  n, p } . Hence, the set I is restricted to
incorporate only valid combinations of ( p, n, t). Note that the set I may also exclude
some other specific valid combinations of ( p, n, t), if so desired.
Let N + denote the set of positive integers. Define a dissatisfaction function DC :
I → N + by DC( p, n, t) = k1 ( p, n)+k2 ( p, n, t). Note that the function DC( p, n, t)
represents the overall dissatisfaction level of proctor p ∈ P for proctoring an exam
during exam-period t of day n.

4 Formulation of ETM and PAM

4.1 Model ETM

The model variables and constraints are respectively defined in Sects. 4.1.1 and 4.1.2,
and the overall formulation of ETM is given in Sect. 4.1.3. Incorporating enhancing
issues as soft-constraints are then presented in Sect. 4.1.4.

4.1.1 Problem variables

In order to formulate the problem described in the foregoing sections, we define the
following integer variables.
For g ∈ G and d ∈ D g , let


⎨ 1 if the exam associated with class c ∈ C g, d is held during the
X g, d, c, t = exam -period t ∈ T,

0 otherwise.

Also, for g ∈ G and d ∈ D g , let Yg,d,c,r,t be the number of sections associated


with class c ∈ C g,d whose exams are assigned to classrooms of type r ∈ R g during
exam-period t ∈ T . Hence, for g ∈ G, d ∈ D g , and c ∈ C g,d , we have Yg,d,c,r,t = 0,
for all r ∈
/ K g,d,c and for all t. Note that Yg,d,c,r,t is bounded above by N g,r .

123
28 S. M. Al-Yakoob et al.

4.1.2 Problem constraints

The model incorporates the following constraints:


(A) Exam offering requirements
The following constraint requires that the exam for each class shall be held in one
exam-period. Scheduling the exam simultaneously for all sections of any given class
prevents possible leakage of the exam questions.

X g, d, c, t = 1, ∀ g ∈ G, d ∈ D g , c ∈ C g,d . (1.1.1)
t∈ T

Two exams that are taken by a common group of students should not be held dur-
ing the same exam-period. This conflict-free requirement is guaranteed by Constraint
(1.1.2) for the sets Si , for i = 1, . . . , m.
  
X g, d, c, t ≤ 1, ∀ t ∈ T and i = 1, . . . , m. (1.1.2)
g∈ G d∈ D g c∈ S g,d
i

Also, for g ∈ G, d ∈ D g , classes c1 and c2 such that (c1 , c2 ) ∈ U g,d cannot be


offered during the same exam-period, as enforced by the following constraint.

X g, d, c1 ,t + X g, d, c2 ,t ≤ 1, ∀ g ∈ G, d ∈ D g , (c1 , c2 ) ∈ U g, d
and t ∈ T.
(1.1.3)
(B) Availability of classrooms
Exams must be held in suitable classrooms as guaranteed by Constraint (1.2.1),
subject to the availability of the facility resources as enforced by Constraint (1.2.2).

Yg, d, c, r, t = O g, d, c X g, d, c, t , ∀ g ∈ G, d ∈ D g , c ∈ C g, d and t ∈ T.
r ∈ K g, d, c
(1.2.1)
 
Yg, d, c, r, t ≤ N g, r
, ∀ g ∈ G, r ∈ R g
and t ∈ T. (1.2.2)
d∈ Dg c∈ V g, d, r

(C) Availability of proctors


The following constraint ensures that the total number of exams offered by depart-
ment d ∈ D g of college g ∈ G during a given exam-period t ∈ T does not exceed
the available number of proctors of department d during exam-period t, given by
g, d, t g, d
| P2 | + | P2 | .
 g, d, t g, d
O g, d, c X g, d, c, t ≤ | P2 | + | P2 |, ∀ g ∈ G, d ∈ D g and t ∈ T.
c∈ C g, d
(1.3.1)
(D) Enhancing exam timetables
Since the time between two consecutive exam-periods is one hour and due to poten-
tial traffic congestion, it is required not to hold different exams that belong to different

123
A mixed-integer mathematical modeling approach to exam timetabling 29

colleges consecutively if the exams can be taken concurrently. This is guaranteed by


Constraint (1.4.1).
 
X g1 , d1 , c1 , t + X g2 , d2 , c2 , t+1 ≤ 1, ∀ (g1 , g2 ) ∈ G1, d1 ∈ D g1 ,
g1 , d1 g2 , d2
c1 ∈ Si c2 ∈ Si

d2 ∈ D g2 , n ∈ N , t ∈ {4n − 3, 4n − 2, 4n − 1} and i = 1, . . . , m. (1.4.1)

The maximum number of exams that may be taken on a given day by any student
is two, as assured by the following constraint.
   
X g, d, c, t ≤ 2, ∀ n ∈ N and i = 1, . . . , m. (1.4.2)
g∈ G d∈ D g c∈ S g, d t ∈  n
i

Constraint (1.4.3) assures that the maximum number of exams that may be taken
on two consecutive days is l, where l may take the value of two or three as determined
by the concerned decision-maker.
   
X g, d, c, t ≤ l, ∀ n ∈ {1, . . . , |N | − 1} and
g∈ G d∈ D g c∈ S g, d t ∈ 4n−3,...,4n+4
i
i = 1, . . . , m. (1.4.3)

The next constraint obviates assigning any two exams from the set Si , for i =
1, . . . , m, with one on the last exam-period of one day, and the other on the first
exam-period of the next day.
     
X g1 , d1 , c1 , 4n + X g2 , d2 , c2 , 4n+1 ≤ 1,
g1 ∈ G d1 ∈ D g1 c ∈ S g1 , d1 g2 ∈ G d2 ∈ D g2 c ∈ S g2 , d2
1 i 2 i
∀ n ∈ {1, . . . , |N | − 1} and i = 1, . . . , m. (1.4.4)

(E) Reducing parking and traffic congestion


Parking and traffic congestion can be considerably reduced if the number of stu-
dents taking exams are reasonably spread over all the available exam-periods. This
might be achieved by imposing a maximum bound on the number of students involved
in exams during each exam-period.
For each college g ∈ G the following constraints accomplish the intended spread
by imposing an upper bound (denoted by C Q g ) on the number of students taking
exams offered by college g during each exam-period t.
 
(A g, d, c O g, d, c ) X g, d, c, t ≤ C Q g , ∀ g ∈ G, and t ∈ T. (1.5.1)
d∈ Dg c∈ C g, d

Remark 4.1.1 The value of C Q g can be computed based on the total capacities of all
classrooms that are available at college g ∈ G, and the number of available parking
and traffic facilities during each exam-period.

123
30 S. M. Al-Yakoob et al.

4.1.3 ETM as a feasibility problem

Consider the feasible region (FR) defined as follows.


FR = {(X, Y ) : where
X g, d, c, t ∈ {0, 1}, ∀ g ∈ G, d ∈ D g , c ∈ C g,d and t ∈ T,
Yg, d, c, r, t ∈ {0, 1, . . . , N g, r }, ∀ g ∈ G, d ∈ D g , c ∈ C g,d , r ∈ R g and t ∈ T,
and (X, Y ) satisfies Constraints (1.1.1)–(1.1.3), (1.2.1), (1.2.2), (1.3.1), (1.4.1)–(1.4.4)
and (1.5.1)}.
In this case, ETM is concerned with finding a feasible solution in FR. Note that,
any feasible solution in FR satisfies all the requirements discussed in Sect. 4.1.2. In
the case of infeasibility of FR, we can successfully increment the value of N by one
until feasibility is obtained. Another approach to deal with the infeasibility of FR is
to treat the enhancing issues delineated in Part (D) of Sect. 4.1.2 as soft-constraints as
discussed in the next section.

4.1.4 Incorporating enhancing issues as soft-constraints

The enhancing issues presented in Part (D) of Sect. 4.1.2 can be alternatively incorpo-
rated in ETM as soft-constraints as discussed next. For i ∈ {1, . . . , m}, consider the
set Si . Let Bt,1 i be a binary variable that equals 1 if two exams of different colleges
in different campuses are taken by students during exam-periods t and t + 1, where
t ∈ {4n − 3, 4n − 2, 4n − 1} of day n ∈ N . For n ∈ N , let Bn, 2 be a nonnegative
i
integer variable representing the number of exams exceeding two that are taken by
students on a given day n. Note that since there are four exam-periods on any given day,
2 ∈ {0, 1, 2}. Let B 3 be a nonnegative integer variable representing the number
Bn, i n, i
of exams exceeding l [as defined for (1.4.3)] that are taken by students during exam-
periods t ∈ {4n − 3, . . . , 4n + 4} over days n and n +1 for each n ∈ {1, . . . , |N |−1}.
Since, there are eight exam-periods on two consecutive days, Bn, 3 ∈ {0, . . . , 8 − l}.
i
4
Let Bn, i be a binary variable that equals 1 if two exams, one on the last exam-period
(t = 4n) of any given day n and the other on the first exam-period (t = 4n + 1) of
the next day n + 1, are taken by students, ∀ n ∈ {1, . . . , |N | − 1}.
Then, Constraints (1.4.1),…,(1.4.4) can be alternatively formulated as follows.

⎛⎛ ⎞ ⎞
⎜⎜   ⎟ ⎟
Bt,1 i ≥ ⎝⎝ X g1 , d1 , c1 , t + X g2 , d2 , c2 , t+1 ⎠ −1⎠ , ∀ (g1 , g2 ) ∈ G1,
g , d1 g , d2
c1 ∈Si 1 c2 ∈Si 2

d1 ∈ D g1 , d2 ∈ D g2 , t ∈ {4n − 3, 4n − 2, 4n − 1}, n ∈ N and i = 1, . . . , m.


(1.6.1)
⎛⎛ ⎞ ⎞
⎜⎜     ⎟ ⎟
i ≥ ⎝⎝ X g, d, c, t ⎠ −2⎠ , ∀ n ∈ N and i = 1, . . . , m.
2
Bn,
g∈ G d∈ D g c∈ S g, d t ∈  n
i
(1.6.2)

123
A mixed-integer mathematical modeling approach to exam timetabling 31

⎛⎛ ⎞ ⎞
⎜⎜     ⎟ ⎟
3
Bn, i ≥ ⎝⎝ X g, d, c, t ⎠ − l ⎠ ,
g∈ G d∈ D g c∈ S g, d t ∈ {4n−3,...,4n+4}
i
∀ n ∈ {1, . . . , |N | − 1} and i = 1, . . . , m. (1.6.3)
⎛⎛
⎜⎜   
i ≥ ⎝⎝
4
Bn, X g1 , d1 , c1 , 4n
g1 ∈ G d1 ∈ D g1 c ∈ S g1 , d1
1 i
⎞ ⎞
   ⎟ ⎟
+ X g2 , d2 , c2 , 4n+1 ⎠ − 1⎠ ,
g2 ∈ G d2 ∈ D g2 c2 ∈ Si
g2 , d2

∀ n ∈ {1, . . . , |N | − 1} and i = 1, . . . , m. (1.6.4)

Then, the modified version of ETM, denoted by ETM1, can be stated as fol-
lows, where δi , i = 1, . . . , 4 are respectively penalties associated with the variables
Bt,1 i , Bn,
2 , B 3 , and B 4 . Specific values of δ , . . . , δ should be determined by the
i n, i n, i 1 4
concerned decision-makers depending on the importance of satisfying the enhancing
issues discussed in Part (D) of Sect. 4.2.1.
ETM1:
Minimize

  
m  
m
δ1 Bt,1 i + δ2 2
Bn, i
n∈N t ∈{4n−3, 4n−2, 4n−1} i=1 n∈N i=1
 m  m
+ δ3 3
Bn, i + δ4 4
Bn, i,
n∈ {1,...,|N |−1} i=1 n∈ {1,...,|N |−1} i=1

subject to
(1.1.1)–(1.1.3), (1.2.1), (1.2.2), (1.3.1),(1.5.1), (1.6.1)–(1.6.4),

X g, d, c, t ∈ {0, 1}, ∀ g ∈ G, d ∈ D g , c ∈ C g, d and t ∈ T,


Yg, d, c, r, t ∈ {0, 1, . . . , N g, r }, ∀ g ∈ G, d ∈ D g , c ∈ C g, d , r ∈ R g and t ∈ T,
Bt,1 i ∈ {0, 1}, ∀ t ∈ {4n − 3, 4n − 2, 4n − 1}, n ∈ N , and i = 1, . . . , m,
i ∈ {0, 1, 2}, ∀ n ∈ N and i = 1, . . . , m,
2
Bn,
i ∈ {0, . . . , 8 − l}, ∀ n ∈ {1, . . . , |N | − 1} and i = 1, . . . , m,
3
Bn,
i ∈ {0, 1}, ∀ n ∈ {1, . . . , |N | − 1} and i = 1, . . . , m.
4
Bn,

Remark 4.1.2 Note that due to the minimization objective function in Problem ETM1,
the constraints (1.6.1)–(1.6.4) will be active (and relevant) only when the respective
right-hand sides take positive values; else, the corresponding variable on the left-hand
side will be zero. Moreover, for the same reason, these variables will automatically

123
32 S. M. Al-Yakoob et al.

be integral for binary X-values, even if declared as continuous within any software
package used to solve ETM1.

4.2 Model PAM

In this section, we formulate a mixed-integer programming model (PAM) for assign-


ing proctors to exams taking into consideration their preferences for specific days and
exam-periods. In this case, exam-periods for all the exams are fixed as determined
from the solution of ETM. A detailed formulation of PAM is given in Sect. 4.2.1 and
related discussion is provided in Sect. 4.2.2.

4.2.1 Model formulation

In order to formulate the problem, we define the following integer variables. For c ∈ C
and p ∈ P c , we define the following integer variables.

1 if proctor p is assigned to an exam associated with a section of class c,
ac, p =
0 otherwise.

Recall that ETM specifies the exam-period for each exam, based on which we define
the following parameter. For c ∈ C and t ∈ T , let

1 if the exam associated with class c is assigned to exam-period t,
λc, p =
0 otherwise.

Given a pair of proctors ( p1 , p2 ) ∈ Q i , let the variable w p1 , p2 represent the abso-


lute difference between the dissatisfaction levels of proctors p1 and p2 , for i = 1, 2.
Then the proctor assignment model (PAM) is formulated as follows.
PAM:

 
2 
Minimize p + π w p1 , p2 ,
p∈P i=1 ( p1 , p 2 )∈ Q i
subject to

a c, p = O c , ∀ c ∈ C, (2.1)
p∈Pc

λ c, t a c, p ≤ 1, ∀ p ∈ P and t ∈ T p , (2.2)
c∈C p, t
 
λ c, t a c, p ≤ 2, ∀ p ∈ P and n ∈ N , (2.3)
t ∈  n, p c ∈ C p, t
 
λ c, t a c, p ≤ L 1 , ∀ p ∈ P1 , (2.4.1)
t ∈T p c∈C p, t
 
λ c, t a c, p ≤ L 2 , ∀ p ∈ P2 , (2.4.2)
t ∈T c∈C t

123
A mixed-integer mathematical modeling approach to exam timetabling 33

 
λ c, t a c, p + λ c, t+1 a c, p ≤ 1, ∀ t ∈ {4n − 3, 4n − 2, 4n − 1},
c∈C p, t c∈C p, t+1

n ∈ N , p ∈ P3 , (2.5)
  
p = DC( p, n, t) λ c, t a c, p , ∀ p ∈ P, (2.6)
n∈N t ∈  n, p c ∈ C p, t

w p1 , p 2 ≥  p1 −  p 2 , ∀ ( p1 , p 2 ) ∈ Q i , i = 1, 2, (2.7)
w p1 , p 2 ≥  p2 −  p 1 , ∀ ( p1 , p 2 ) ∈ Q , i = 1, 2, i
(2.8)
a c, p ∈ {0, 1} , ∀ c ∈ C and p ∈ P c ,
 p ≥ 0, ∀ p ∈ P,
w p1 , p 2 ≥ 0, ∀ ( p1 , p 2 ) ∈ Q i , for i = 1, 2.

4.2.2 Constraints

For a given class c ∈ C, Constraint (2.1) assures that there are O c proctors to mon-
itor all the exams of the sections associated with class c. Constraint (2.2) guarantees
that a proctor may be assigned to at most one exam of a section of a class in any
time-period. A proctor may be assigned to at most two exams per day as enforced
by Constraint (2.3). The maximum loads of proctors p ∈ P1 and p ∈ P2 over all
exam-periods are respectively enforced by Constraints (2.4.1) and (2.4.2). Restric-
tions related to not proctoring during two consecutive exam-periods for proctors in
P3 are enforced by Constraint (2.5). The dissatisfaction level of a proctor is computed
in Constraint (2.6), and then Constraints (2.7) and (2.8) give the absolute difference
between dissatisfaction levels for pairs of proctors.

4.2.3 Objective function

The objective function of the model seeks to minimize the total dissatisfaction and
inequity costs associated with assigning proctors to exams. The total dissatisfaction
cost is given by the first part of the objective function, and the sum of differences in
dissatisfaction levels between proctors is given by the second part of the objective
function. A weight factor π is associated with the second objective function term to
reflect its relative priority with respect to the first term.

4.2.4 Other equity modeling considerations

Alternative modeling approaches for the proctors’ equity considerations are presented
next. Consider the following constraint.

 p1 −  p2 = θ p+1 , p2 − θ p−1 , p2 , ∀ ( p1 , p 2 ) ∈ Q i , i = 1, 2, (2.7.1)

where θ p+1 , p2 , θ p−1 , p2 ≥ 0.


Let PAM1 denote the modified equivalent version of PAM where Constraints (2.7)
and (2.8) are replaced with (2.7.1) and the second term of the objective function of

123
34 S. M. Al-Yakoob et al.


2 
PAM is replaced with π = [θ p+1 , p2 − θ p−1 , p2 ]. Let i ∈ {1, 2} and
i=1 ( p1 , p 2 )∈ Q i
( p1 , p 2 ) ∈Qi . For a given solution of PAM1 where  p1 ≥  p2 , the most attrac-
tive values for θ p+1 , p2 and θ p−1 , p2 are θ p+1 , p2 =  p1 −  p2 and θ p−1 , p2 = 0 . On
the other hand, if  p1 <  p2 , then the most attractive values for θ p+1 , p2 and θ p−1 , p2
are θ p+1 , p2 = 0 and θ p−1 , p2 =  p2 −  p1 . Therefore, combining both cases, this
new modeling approach also attempts to minimize the absolute difference between
dissatisfaction levels for pairs of proctors.
Now, let PAM2 denote the model derived from PAM by replacing Constraints (2.7)
and (2.8) with the following constraints:

θi max ≥  p , ∀ p ∈ Pi , for i = 1, 2, (2.8.1)


θi min ≥  p , ∀ p ∈ Pi , for i = 1, 2, (2.8.2)

where θi max , θi min ≥ 0 , and by replacing the second term of the objective func-
2
tion of PAM with the term π [ θi max − θi min ]. Note that for i = 1, 2, θi max
i=1
and θi min respectively represent the maximum and minimum dissatisfaction levels
of proctors in Pi . Hence, PAM2 attempts to reduce the spread in the dissatisfaction
levels among the proctors in each set Pi , i = 1, 2 as opposed to minimizing the sum
of the absolute differences between dissatisfaction levels for pairs of proctors. Hence,
it involves O(|Pi |), rather than O(|Pi |2 ), constraints. Computational results related to
solving PAM, PAM1, and PAM2 will be presented in Sect. 5.

4.3 Gender-related issues

Due to the gender-based policy, it is sometimes more convenient to assign female


proctors to female sections. Hence, ETM and ETM1 can be modified to assign, to the
extent possible, female sections to female proctors.
Let the set of female proctors available to proctor exams of classes of department
d ∈ D g at college g ∈ G be given by P F, g, d . Then the following constraint assures
that the number of female proctors available to proctor exams for classes of department
d ∈ D g at college g ∈ G exceeds the number of exams for female class sections held
during exam-period t.


F g, d, c X g, d, c, t ≤ |P F, g, d |, ∀ g ∈ G, d ∈ D g , t ∈ T. (3.1)
c ∈C g,d

Some female proctors strongly prefer proctoring female sections only; hence, PAM
(and its foregoing variants) can be modified to avoid assigning male and joint sections

123
A mixed-integer mathematical modeling approach to exam timetabling 35

to these proctors, as assured by Constraint (3.2) below


λc, t ac, p ≥ M c + J c , ∀ t ∈ T, c ∈ C t . (3.2)

p∈ MJ P c

5 Computational results and a comparison with a manual timetabling approach

Next, we define a glossary of some additional notation that will be used in this section.
Denote:

• M : Linear relaxation of any model M.


• v(M) : Optimal objective function value for any model M.
• Opt-gap: Optimality gap.
• RT: Run time in seconds where all runs have been made on a Pentium IV com-
puter having 1.99 GB of RAM and using CPLEX-MIP-9.0 (See for example ILOG
(2005)), where all procedures have been coded in JAVA.
• RT⊗ : Run time elapsed in seconds using CPLEX-MIP-9.0 until out-of-memory
difficulties resulted.
• MET: The current manual scheduling approach for Problem ETP that is presently
used at KU to generate exam schedules, as described in Al-Jazzaf (2006).

The following statistics are compiled when using the manual approach (MET).

• C1MET : Total number of exam conflicts. Recall that an exam conflict occurs if two
exams of classes that are taken concurrently by a group of students are assigned to
the same exam-period.
• C2MET : Total number of occurrences of coinciding exam periods for classes taught
by the same faculty member associated with the set U g,d .
• C3MET : Number of occurrences where groups of students take three exams on one
day.
• C4MET : Number of occurrences where groups of students take four exams on two
consecutive days.
• C5MET : Number of occurrences where groups of students take one exam on the last
exam-period of a given day and another exam on the first exam-period of the next
day.
• RMET : Number of occurrences where an exam is assigned to an unsuitable classroom.

The following are additional notation that are specific to PAM.

• vub (PAM) : Best solution (upper bound) found for PAM.


• vlb (PAM)
 : Best lower bound found for PAM.
• α PAM = p.
p∈P

2 
• β PAM = w p1 , p2 .
i=1 ( p1 , p2 )∈Q i

123
36 S. M. Al-Yakoob et al.

Table 1 Test problems


Test cases for Test cases for Term
problem ETP problem PAP

p1 q1 Fall 2002–2003
p2 q2 Spring 2002–2003
p3 q3 Fall 2003–2004
p4 q4 Spring 2003–2004
p5 q5 Fall 2004–2005
p6 q6 Spring 2004–2005

• µiPAM : The mean for the individual proctors’ dissatisfaction levels obtained via PAM
for proctors in the set Pi , for i = 1, 2.
• σiPAM : The standard deviation for the proctors’ individual dissatisfaction levels
obtained via PAM for proctors in the set Pi , for i = 1, 2.
• MPA : The manual scheduling approach that is presently used at KU to generate
proctors’ schedules as described in Al-Jazzaf (2006).

The following statistics are compiled when using the manual approach (MPA).

• α MPA : The summation of the individual dissatisfaction levels for all proctors.
• β MPA : The summation of the absolute differences between the dissatisfaction levels
of proctors.
• µMPA (i): The mean for the individual proctors’ dissatisfaction levels for proctors in
the set Pi , for i = 1, 2.
• σ MPA (i): The standard deviation for the individual proctors’ dissatisfaction levels
for proctors in the set Pi , for i = 1, 2.

Note that we considered six test cases related to Problem ETP and six test cases
related to Problem PAP. Details related to these test cases are available in Al-Jazzaf
(2006). For the sake of simplicity in reference, Table 1 associates each test case with
an indexed symbol.

5.1 Computational experience in solving ETM

We began by solving ETM directly via CPLEX-MIP-9.0 for the College of Science.
We selected this particular college because it is one of the largest colleges at Kuwait
University in terms of number of classes offered. Tables 2 and 3 present computa-
tional results related to solving ETM and ETM for the College of Science based on
test cases p1 , . . . , p6 . Since all the constraints (1.6.1)–(1.6.4) of Sect. 4 were dealt
with as hard-constraints in ETM, the resulting model becomes a feasibility problem,
and hence, v(ETM) = v(ETM) = 0 .
Note that we were able to easily solve ETM directly using CPLEX for the six test
cases. It turned out that the root node relaxation of the problem yielded a feasible
integer solution.

123
A mixed-integer mathematical modeling approach to exam timetabling 37

Table 2 Computational results related to solving ETM for the College of Science

Test problem Rows Columns Nonzero entries RT

p1 26,151 42,280 349,840 0.48


p2 26,453 44,800 334,646 0.25
p3 26,100 44,120 346,080 0.59
p4 25,494 46,840 321,320 0.39
p5 26,584 45,800 348,880 0.50
p6 26,553 48,600 338,740 0.48

Table 3 Computational results related to solving ETM for the College of Science

Test problem Rows Columns No. of integer variables Nonzero entries RT

p1 20,768 42,960 42,960 326,544 0.95


p2 21,851 44,800 44,800 334,646 0.77
p3 21,096 44,760 44,760 324,202 0.77
p4 20,661 46,840 46,840 300,782 0.83
p5 21,196 45,800 45,800 324,664 0.98
p6 21,519 48,680 48,680 317,228 0.97

Table 4 Computational results related to solving ETM1 for the College of Science

Test problem Rows Columns No. of integer variables Nonzero entries RT

p1 24,791 49,708 42,960 357,628 0.83


p2 25,982 51,772 44,800 366,332 0.72
p3 25,020 51,452 44,760 353,772 0.72
p4 24,414 52,916 46,840 328,476 0.83
p5 25,264 52,408 45,800 356,328 0.86
p6 25,353 54,924 48,680 345,864 0.84

Likewise, we were also able to readily solve ETM1 for the College of Science,
where it turns out that v(ETM) = v(ETM) = 0 for all the test problems. Table 4
presents statistics related to solving ETM1.
Next, we attempted to solve ETM and ETM1 directly using CPLEX-MIP-9.0 for
two colleges (the College of Science and the College of Education) simultaneously.
However, no meaningful solution was obtained due to out-of-memory difficulties.
Table 5 presents results related to solving ETM for the two colleges based on test
problems p1 and p6 . We comment here that ETM can be solved jointly for multiple
colleges at Kuwait University using a sequential variable fixing heuristic similar to
the one that will be developed in Sect. 5.3 to solve PAM. An actual implementation
of this heuristic is deferred to future research due to the intricate information that has
to be compiled and analyzed for all the colleges.

123
38 S. M. Al-Yakoob et al.

Table 5 Computational statistics related to ETM for the College of Science and the College of Education
for two academic terms

Test problem Rows Columns No. of integer variables Nonzero entries RT⊗

p1 38,077 49,440 49,440 748,126 300.38


p6 37,453 54,780 54,780 705,691 300.27

Table 6 Number of sections


Department Number of sections
offered by the various
departments in the College of
Mathematics and Computer Science 139
Science for the Fall 2002–2003
term Chemistry 51
Physics 50
Earth and Environmental Sciences 34
Statistic and Operations Research 38
Biological Sciences 93

Table 7 Computational results related to solving PAM

Test problem Rows Columns Nonzero entries v(PAM) RT

q1 1,737 1,828 10,662 2,042.6 0.20


q2 1,741 1,789 10,550 1,918.9 0.20
q3 1,450 1,684 9,790 1,914.1 0.22
q4 1,803 2,039 11,785 1,433.8 0.20
q5 1,308 1,565 8,702 2,590 0.20
q6 1,995 2,110 12,192 1,622.7 0.20

5.2 Computational experience in solving PAM

In this section, we present computational results related to solving PAM, PAM1, and
PAM2 for the Department of Mathematics and Computer Science at the College of
Science. We have selected this particular department because it is the largest depart-
ment in the College of Science in terms of the number of offered classes and sections.
Table 6 below presents the total number of sections offered by all departments of the
College of Science for the Fall 2002-2003 term. Also, in solving these models, we
initially used just the maximum load restriction L 2 for p ∈ P2 as defined in Sect. 3.
However, we encountered noticeable load discrepancies among some of the proctors.
To minimize these discrepancies, we imposed a lower bound, denoted by L min , on the
total load of proctors to ensure fairness of exam assignments in all the runs reported
below.
A direct solution of PAM and PAM1 via CPLEX was out-of-reach due to out-of-
memory difficulties for all test cases. (The minor modeling difference between these
two formulations did not reveal any computational effects.) Solutions of the linear
relaxations of PAM and PAM1 were readily obtained, where, as expected, v(PAM) =
v(PAM1) for all test cases. Tables 7 and 8 present computational results for PAM.

123
A mixed-integer mathematical modeling approach to exam timetabling 39

Table 8 Computational results related to solving PAM

Test problem Rows Columns No. of integer No. of continuous Nonzero RT⊗
variables variables entries

q1 1,513 1,828 1,360 468 10,007 31,173.08


q2 1,520 1,789 1,321 468 9,901 16,226.63
q3 1,295 1,684 1,246 438 9,136 28,707.44
q4 1,552 2,039 1,522 517 10,891 42,570.53
q5 1,216 1,565 1,204 361 8,394 7,294.43
q6 1,747 2,110 1,604 506 11,352 16,544.28

Table 9 Computational results related to solving PAM2

Test problem Rows Columns Nonzero entries v(PAM2) RT

q1 929 1,396 8,174 1908.10 0.20


q2 933 1,357 8,062 1659.34 0.19
q3 698 1,281 7,472 1846.59 0.19
q4 917 1,563 9,053 1389.79 0.20
q5 706 1,238 6,836 1569.58 0.20
q6 1,123 1,643 9,506 1535.116 0.20

Table 10 Computational results related to solving PAM2

Test problem Rows Columns No. of integer No. of continuous Nonzero v(PAM2) RT
variables variables entries

q1 705 1,396 1,360 36 7,519 1925.9 472.25


q2 712 1,357 1,321 36 7,413 1670.9 0.20
q3 543 1,281 1,246 35 6,818 1857.0 42.25
q4 666 1,563 1,522 41 8,159 1398.9 0.23
q5 614 1,238 1,204 34 6,528 1573.0 0.23
q6 875 1,643 1,604 39 8,666 1, 603 7, 733.03

Next, Tables 9 and 10 provide statistics related to solving PAM2 and PAM2 using
CPLEX.
A direct solution of PAM2 was attainable for all test cases, noting that the linear
relaxation of the model is fairly tight as indicated from the objective function values
of PAM2 and PAM2. Also, note that the run time for solving PAM2 varies from .20
seconds for Test Problem q2 to 472.25 seconds for Test Problem q1 .
Now, let PAM3 denote a modification of PAM that eliminates equity considerations,
other than minimizing the total dissatisfaction, by:
(a) discarding the w -variables; (b) discarding the second term of the objective function,
2 
namely w p1 , p2 , and (c) discarding Constraints (2.7) and (2.8).
i=1 ( p1 , p2 )∈ Q i

123
40 S. M. Al-Yakoob et al.

Table 11 Computational results related to solving PAM3

Test problem Rows Columns Nonzero entries v(PAM3) RT

q1 865 1,392 8,046 1,843 0.19


q2 869 1,353 7,934 1,665 0.20
q3 636 1,277 7,348 1,810 0.20
q4 843 1,559 8,905 1,339 0.20
q5 646 1,234 6,716 1,482 0.20
q6 1, 053 1,639 9,366 1,473 0.20

Table 12 Computational results related to solving PAM3

Test Rows Columns No. of integer No. of continuous Nonzero v(PAM3) RT


problem variables variables entries

q1 641 1,392 1,360 32 7,391 1,843 0.20


q2 648 1,353 1,321 32 7,285 1,665 0.20
q3 481 1,277 1,246 31 6,694 1,810 0.20
q4 592 1,559 1,522 37 8,011 1,339 0.20
q5 554 1,234 1,204 30 6,408 1,482 0.20
q6 805 1,639 1,604 35 8,526 1,473 0.20

We were now able to solve PAM3 to optimality, where it turned out that the
objective function values for PAM3 and PAM3 are identical and the root node relaxa-
tion of the problem yielded a feasible mixed-integer solution for all the test problems
(see Tables 11, 12). Hence, this indicates that the objective term and constraints per-
taining to balancing proctors’ discrepancies tremendously encumbers the performance
of the branch-and-bound procedure used by CPLEX to solve PAM. In this respect,
PAM2 offers a good compromise in accommodating equity issues.

5.3 LP-based heuristic procedure for solving PAM

The out-of-memory difficulties associated with solving PAM for all test problems
primarily stems from the large number of binary variables and complex constraint
structures. This motivates the development of a heuristic procedure similar to that
adopted in Al-Yakoob and Sherali (2006a) to solve a class-faculty assignment prob-
lem, which generates a solution in an iterative fashion where each iteration involves
solving a modified version of PAM on which the integrality of only a subset of the
integer variables is enforced. This heuristic procedure is described next.
In PAM, let the vector a be partitioned as a = (a1 , a2 . . . , aC ), where for c ∈
{1, . . . , C} , ac denotes the vector of a-variables associated with class c . Partition the

set {1, . . . , C} into ≥ 1 components, denoted by 1 , . . . , , where j =
 j=1
{1, . . . , C} and j k = ∅ , ∀ j, k ∈ {1, . . . , } , and j = k (Recommendations

123
A mixed-integer mathematical modeling approach to exam timetabling 41

for suitable partitioning schemes are addressed in the sequel below). Furthermore,
we will let PAM(1, opt-gap1 ) denote the relaxation of PAM for which integrality is
enforced only on the a-variables that correspond to c ∈ 1 and an optimal solution
is required to be found within a tolerance opt-gap1 of optimality. For l = 2, . . . , ,
we define PAM(l, opt-gapl ) to be the relaxation of PAM that has the following char-
acteristics:

l−1
(a) The values of all the a-variables that correspond to c ∈ j are fixed as obtained
j=1
from PAM(l − 1, opt-gapl−1 ) .
(b) The integrality on the a-variables that correspond to c ∈ l is enforced, and is
relaxed for the remaining variables corresponds to j , j > l .
(c) An optimal solution is required to be found within a tolerance opt-gapl of opti-
mality.
The sequential fixing of variables heuristic SFVH then proceeds as follows:
Initialization: Let ≤ C be some integer number. Let l = 1 and let opt-gap1 be
some selected optimality tolerance. SolvePAM(1, opt-gap1 ) .
Main Step: If the solution to PAM(l, opt-gapl ) is integral; then terminate the algo-
rithm; the proposed solution is that obtained from solving PAM(l, opt-gapl ). If l =
and the foregoing condition is not satisfied, then PAM(l, opt-gapl ) is infeasible; ter-
minate the heuristic—no feasible solution has been found. Otherwise, increment l by
one and solve the PAM(l, opt-gapl ). Repeat the Main Step.
A partitioning scheme for a = (a1 , a2 , . . . , a C )
The actual value of depends on the problem data, the required overall optimality
of solution, and the maximum time-frame during which a solution should be obtained.
In general, smaller values of should produce better quality solutions, but would
be more computationally demanding. Note that based on our extensive computational
testing and analysis, we observed that in some cases, enforcing the integrality of one
extra variable significantly changes the solvability of the problem. Also, it seems rea-
sonable to set 1 to be the set of classes that are taken by the first and second year
students to the extent possible. This stems from the fact that these classes mostly
involve a large number of sections.
Lower and upper bounds
Recall that at termination of Heuristic SFVH, at iteration l = l ∗ , say, the final

PAM(l ∗ , opt-gapl ) yields an overall feasible solution whenever it produces an opti-
mum, and hence, the objective function value of this problem constitutes the best avail-
able upper bound. Moreover, the objective function value of PAM(1, opt-gap1 ) gives a

lower bound on the overall problem. Hence, vub (PAM) = v(PAM(l ∗ , opt-gapl )) and
vlb (PAM) = v(PAM(1, opt-gap )). 1

Let v(PAMl ) = v(PAM(l, opt-gapl)), for l = 1, . . . , l ∗ and let perct-opt


(vub , vlb ) = 100 1 − vub (PAM )−vlb (PAM)
vub (PAM)
. Tables 13 and 14 present statistics related
to solving PAM using Procedure SFVH based on the six test problems for the depart-
ment of Mathematics and Computer Science. Note that the linear relaxation of PAM
was readily solved for all test problems as illustrated in Table 7. Also, Table 13 indi-
cates that Procedure SFVH produced solutions having optimality percentages ranging

123
42 S. M. Al-Yakoob et al.

Table 13 Statistics related to solving PAM using Procedure SFVH

Test problem v(PAM) No. of integer RT v(PAM) No. of integer RT


for l = 1 variables for l = 2 variables

q1 2,918 474 5,588.83 3,213 1,011 3,058.36


q2 2,921.84 314 90.55 3,023.3 849 101.81
q3 2,965.5 393 10,451.25 3,576 848 35.69
q4 2,487.13 326 534.61 2,639 905 27,830.42
q5 3,090 1,006 1,103.19 4,548 1,204 182.78
q6 2,151.48 270 464.05 2,846.5 506 29,871.84
Test problem v(PAM) No. of integer RT v(PAM) No. of integer RT
for l = 3 variables for l = 4 variables

q1 4,924 1,360 1,317.63 – – –


q2 4,600 1,321 1,803.06 – – –
q3 5,017 1,246 9,296.86 – – –
q4 3,026.65 1,414 8,055.20 4,648 1,522 0.20
q5 – – – – – –
q6 3,152.7 1,194 709.77 3,476 1,604 8.23
Test Problem vub (PAM) Total RT perct_opt(vub , vlb )

q1 4,924 9,964.82 59.2


q2 4,600 1,995.42 63.5
q3 5,017 19,784.11 59.1
q4 4,648 36,420.43 53.5
q5 4,548 1,285.97 67.9
q6 3,476 31,053.89 61.8

from 53.5 to 67.9%. Note that these optimality percentages are based on the initial LP
relaxation values, which might be weak and therefore not truly indicative of the quality
of solutions attained (see further comparisons in Sect. 5.4.2). For test problem q5 , a
solution was obtained using two iterations, while for test problems q1 , q2 , and q3 ,
solutions were obtained using three iterations, and for test problems q4 , and q6 , four
iterations were needed to reach the stated solutions.
Now, let PDPAM2
M
and PDPAM3
M
respectively denote the total difference of discrepancies
between proctors based on the solution to PAM2 and PAM3. Table 14 compares results
obtained via Procedure SFVH and PAM2 and PAM3.
While procedure SFVH produced higher total dissatisfaction levels than those
obtained via PAM2 and PAM3, with its full consideration of equity, it outperformed
PAM2 and PAM3 in the sense that it generated total pairwise differences of discrep-
ancies that are much smaller than those obtained based on the outcome of these other
modified models.
Remark 5.1 The difficulty in solving PAM (and PAM1) might be due to symmetry
effects related to groups of identical proctors within each set P1 and P2 . A more sophis-
ticated data manipulation could attempt to identify such groups and then augment the

123
A mixed-integer mathematical modeling approach to exam timetabling 43

Table 14 Total proctors’ discrepancies obtained via Procedure SFVH and PAM2 and PAM3

Test problem q1 q2 q3 q4 q5 q6

Procedure SFVH p 2,252 1,918 2,090 1,563 2,429 1,785
p∈P
2 
w p1 , p2 2,672 2,682 2,927 3,085 2,119 1,691
i=1 ( p1 , p2 )∈ Q i
Total 4,924 4,600 5,017 4,648 4,548 3,476

PAM2 p 1,864 1,601 1,821.9 1,340 1,487 1,507
p∈P
PDPAM2
M 9,776.0 11,211 5,178 6,996 7,681 6,128
Total 11,640 12,812 6,999.9 8,336 9,168 7,635

PAM3 p 1,843 1,665 1,810 1,339 1,482 1,473
p∈P
PDPAM3
M 11,635 11,170 9,057 7,718 8,068 10,333
Total 113,478 12,835 10,867 9,057 9,550 11,806

Table 15 Values of C1MET , C2MET , C3MET , C4MET , C5MET , and RMET obtained via the manual approach
MET

Test problem C1MET C2MET C3MET C4MET C5MET RMET

p1 15 0 11 3 1 12
p2 17 2 9 11 2 18
p3 16 6 8 1 0 17
p4 16 1 8 3 0 13
p5 17 4 15 3 2 12
p6 15 5 11 3 0 12

model formulation with some suitable symmetry-defeating constraints to enhance


model solvability (Sherali and Smith 2001). This is deferred to future research.

5.4 Comparison of the manual scheduling approaches MET and MPA and ETM and
PAM, respectively

In this section, we compare the overall performance of the manual approach in present
use at KU and the proposed modeling approach. In Sect. 5.4.1 a comparison of ETM
with the manual approach MET is presented, while a comparison of PAM with the
manual approach MPA is presented in Sect. 5.4.2.

5.4.1 Comparison of the manual approach MET and ETM

Table 15 displays the values of C1MET , C2MET , C3MET , C4MET , C5MET , and RMET obtained
by the manual approach MET.

123
44 S. M. Al-Yakoob et al.

Table 16 Improvement of proctors’ preferences achieved via the proposed modeling approach

 MPA PAM   MPA PAM     


α MPA +β MPA − α PAM +β PAM
Test problem 100 α α MPA
−α 100 β β MPA
−β
100 α MPA +β MPA

q1 34.3 68.1 73.9


q2 30.6 80.8 72.8
q3 37.6 80.5 72.6
q4 42.3 80.4 74.8
q5 14.4 79.7 65.5
q6 45.4 88.9 81.2

Table 17 Means and standard deviations of the proctor dissatisfaction levels obtained manually and via
PAM

Test problem µMPA (i) µPAM (i) σ MPA (i) σ PAM (i)

i i i i

1 2 1 2 1 2 1 2

q1 40.5 111.6 6 74.7 24.5 30.4 3 5.7


q2 52 88.6 22 62.5 17 27.5 11 5.5
q3 39 112.7 15.5 71 10 31.3 3.5 6.3
q4 32.3 81.1 21.8 46.2 9.9 28.5 4.6 6.2
q5 29.3 104.6 11.5 91.7 3 28.7 3.6 5.6
q6 46.5 99.6 26.8 54.1 23 27.5 6.7 3.1

Note that the values of all the statistics given in Table 15 vanish to zero when gen-
erating exam timetables via ETM, indicating that exam schedules generated via ETM
overcome the shortcomings of the schedules generated via the manual approach MET.

5.4.2 Comparison of the proctors’ preferences achieved via the manual approach
and PAM

This section presents a comparison of the proctors’ dissatisfaction levels achieved via
the manual approach MPA and PAM.
Tables 16, 17 reveal that even though the provable optimality percentages associ-
ated with solving PAM using Procedure SFVH were relatively low (probably due to
the weak initial LP relaxation-based lower bound for assessing this gap), the modeling
approach by far outperformed the current manual approach from both an individual
and fairness (or equity) points of view. Furthermore, not only are the mean dissatisfac-
tion levels higher using the manual approach, but the manual approach also produces
 MPA 
β −β PAM
greater disparity among proctors, as evidenced by the values of 100 β MPA % in
Table 16, and of σ and σ in Table 17.
MPA PAM

123
A mixed-integer mathematical modeling approach to exam timetabling 45

6 Summary and conclusions

This paper presents mixed-integer programming models for an exam timetabling prob-
lem related to Kuwait University (KU). This problem, as well as other scheduling and
timetabling problems such as class scheduling, and faculty and teaching assistant
scheduling, are challenging and complex tasks faced by relatively large universities
such as KU. The highly combinatorial nature of such problems makes ad hoc man-
ual approaches inefficient, expensive, and often accompanied by serious deficiencies
resulting in exam and class conflicts, dissatisfaction of proctors, consecutive exams
over two different campuses, clustering exams in certain exam-periods of some days,
and parking and traffic congestion. A two-stage modeling approach has been pro-
posed in this paper for the overall exam timetabling and proctor assignment problem
at KU. For the first stage, a mixed-integer programming model, ETM, was developed
to assign the exams of sections of classes to exam-periods, taking into consideration
conflict-free requirements as well as commuting issues. The assignment of proctors
to exams for each department was then handled via another mixed-integer program-
ming model, PAM, given the exam-periods as determined via the solution to ETM.
Model PAM attempts to perform an assignment taking into consideration the proctors’
preferences for specific days and exam-periods.
Model ETM was solved via a direct application of the commercial package CPLEX-
MIP (version 9.0) for the College of Science at Kuwait University, and exam timetables
generated manually and via ETM for the various departments at the College of Science
for six academic terms were compared and analyzed. Solving PAM via CPLEX-MIP
(version 9.0) was out-of-reach for six test cases related to the Department of Math-
ematics and Computer Science at the College of Science, but an alternative model
(PAM2) that formulates the equity issue by attempting to minimize the spread of
dissatisfaction levels among the proctors was readily solved to optimality. This moti-
vated the development of an iterative heuristic procedure for the original Model PAM
in which the integrality of only a subset of the integer variables is enforced at any
given iteration. The results obtained far outperformed the current manual approach
from both an individual and fairness (or equity) points of view.

Acknowledgments This research is supported by Kuwait University under Grant No. [YS 05/04]. We also
express our gratitude to the Kuwait Foundation for the Advancement of Sciences (KFAS) for their generous
support. Special thanks are also due to Miss. Renju Lekshmi for her help in the CPLEX computational
implementation of the developed models and heuristics. We also thank the referees for their constructive
comments that helped improve the presentation in this paper.

References

Al-Jazzaf M (2006) A mixed-integer mathematical modeling approach to exam timetabling at Kuwait


University. Master’s Thesis, Department of Mathematics and Computer Science, Kuwait University
Al-Yakoob SM, Sherali HD (2006a) Mathematical programming models and algorithms for a class-faculty
assignment problem. Eur J Oper Res 173: 488–507
Al-Yakoob SM, Sherali HD (2006b) A column generation approach for an employee scheduling problem
with multiple shifts and work locations. J Oper Res Soc 1–10
Beaumont N (1997a) Scheduling staff using mixed integer programming. Eur J Oper Res 98: 473–484

123
46 S. M. Al-Yakoob et al.

Beaumont N (1997b) Using mixed-integer programming to design employee rosters. J Oper Res Soc 48:
585–590
Birbas T, Daskalaki S, Housos E (1997) Timetabling for greek high schools. J Oper Res Soc 48: 1191–1200
Burke EK, Bykov Y, Petrovic S (2001) A multicriteria approach to examination timetabling. In: Burke
E, Erban W (eds) Practice and theory of automated timetabling III, (LNCS 2079). Springer, Berlin,
pp 104–117
Burke EK, Elliman DG, Ford PH, Weare RF (1996) Examination timetabling in British universities: a sur-
vey. In: Burke E, Ross P (eds) Practice and theory of automated timetabling I, (LNCS 1153). Springer,
Berlin, pp 76–90
Burke EK, Petrovic S (2002) Recent research directions in automated timetabling. Eur J Oper Res 140:
266–280
Carter MW, Laporte G (1996) Recent developments in practical examination timetabling. In: Burke E,
Ross P (eds) Practice and theory of automated timetabling I, (LNCS 1153). Springer, Berlin, pp 3–21
Cheng E, Kruk S, Lipman M (2003) Flow formulations for the student scheduling problem. In: Burke E,
De Causmaecker P (eds) Practice and theory of automated timetabling IV, (LNCS 2740). Springer,
Berlin, pp 299–309
Daskalaki S, Birbas T (2005) Efficient solutions for a university timetabling problem through integer pro-
gramming. Eur J Oper Res 160: 106–120
Daskalaki S, Birbas T, Housos E (2004) An integer programming formulation for a case study in university
timetabling. Eur J Oper Res 153: 117–135
ILOG Inc. (2004) CPLEX optimization solver reference manual
Meisels A, Gudes E, Solotorevsky G (1996) Employee timetabling, constraint networks and knowledge-
based rules: a mixed approach. In: Burke E, Ross P (eds) Practice and theory of automated timetabling
I, (LNCS 1153). Springer, Berlin, pp 93–105
Meisels A, Kaplansky E (2003) Scheduling agents-distributed timetabling problems. In: Burke E,
De Causmaecker P (eds) Practice and theory of automated timetabling IV, (LNCS 2740). Springer,
Berlin, pp 166–177
Meisels A, Kaplansky E (2004) Iterative restart technique for solving timetabling problems. Eur J Oper Res
153: 41–50
Mirrazavi SK, Mardle SJ, Tamiz M (2003) A two-phase multiple objective approach to university timet-
abling utilizing optimization and evolutionary solution methodologies. J Oper Res Soc 54: 1155–1166
Mukherjee AK, Gilbert KC (1997) Lagrangian heuristics for instructor scheduling in executive development
programmes. J Oper Res Soc 48: 373–382
Mullinax C, Lawley M (2002) Assigning patients to nurses in neonatal intensive care. J Oper Res Soc 53:
25–35
Sandhu KS (2001) Automating class schedule generation in the context of a university timetabling infor-
mation system, Phd dissertation, School of Management, Nathan Campus, Griffith University
Schaerf A (1999) A survey of automated timetabling. Artif Intell Rev 13: 87–127
Sherali HD, Smith JC (2001) Improving discrete model representations via symmetry considerations.
Manage Sci 47(10): 1396–1407

123

Você também pode gostar