You are on page 1of 9

IEEE TRANSACTIONS ON EDUCATION, VOL. 51, NO.

4, NOVEMBER 2008 439

Efficient Fuzzy Evolutionary Algorithm-Based


Approach for Solving the Student
Project Allocation Problem
Dipti Srinivasan, Senior Member, IEEE, and Lily Rachmawati, Student Member, IEEE

Abstract—This paper presents a solution framework for the Typically, a list of projects is proposed by the university’s aca-
student project allocation (SPA) problem which is based on evo- demic staff and by researchers from institutions within the uni-
lutionary algorithms (EAs), a biologically inspired stochastic, versity. Students are required to select a specific number of these
population-based search paradigm. Project-based assessment is a
common component of engineering courses that are conducted in
projects, list them in order of preference, and submit the list to
universities around the world. In their final year of study, a list of their department. The challenge of finding a satisfactory match
projects is made available by the academic staff and students are between students and the available projects, given overlapping
required to select a specific number of options from this list. The student choices and other university requirements, brings about
department then assigns a suitable project to each student such that what is known as the student project allocation (SPA) problem.
preferred projects can be allocated to as many students as possible. This problem has repeatedly proven to be computationally de-
While student interest is the primary criteria, several additional
factors need to be considered such as project prerequisites, load manding and time-consuming. The sheer size of the solution
balancing of staff commitments, and other specific university space means that the probability of finding an optimal solution
requirements. The allocation problem can therefore be seen as a using conventional resource allocation approaches is very small
complex multiobjective problem with multiple constraints. The [1]–[6]. A “satisfactory” allocation is one that can be deemed
EA-based project allocation system was recently developed and both effective and efficient [1], [2], such that it optimizes stu-
implemented in a large university department to automate this
dent satisfaction while dealing suitably with the combinatorial
process, and to improve the matching of students to their desired
projects. The solution which provides the highest level of satisfac- complexity of the problem.
tion in meeting the varied objectives is then used to allocate projects In large university departments, the allocation of projects to
to students. This new automated system is not only able to achieve students is typically performed manually by a human decision
a very high level of user satisfaction, but is also able to do so in a maker (DM) in a process that requires several iterations and
very short time, resulting in significant time savings. takes several weeks to complete. Over the past few years, sev-
Index Terms—Evolutionary algorithm (EA), hybrid fuzzy evolu- eral techniques have been proposed to solve such complex re-
tionary algorithm, student project allocation (SPA) problem. source allocation problems [1]–[4]. However, most of these ap-
proaches cannot be directly applied to the SPA problem in large
universities because there are simply too many conflicting goals
I. INTRODUCTION
and constraints. The purpose of this paper is to present a novel
ROJECT-BASED assessment is a common component of
P engineering courses conducted in universities around the
world, with final-year students often being required to embark
method of addressing the SPA problem, taking into account all
of the considerations that have been raised so far: computational
efficiency, student satisfaction, and departmental satisfaction.
on a project that typically spans two semesters. The aim of this Keeping in mind the limited number of published works in
component is to encourage students to learn how to apply skills the field of SPA, the following is a review of promising de-
acquired in the classroom and to contemplate innovative ways velopments that have occurred in the last few years. An expert
of solving problems. The final-year projects (FYPs) are usu- system in this particular application domain was designed and
ally open-ended in nature, giving the students flexibility in ju- presented by Teo and Ho in [5]. This knowledge-based model
diciously selecting viable alternatives, and challenging them to was applied to a single-objective SPA problem on a dataset in-
innovate in terms of initiating new concepts and designs [7]. volving 413 projects and 330 pairs of students who were allowed
The benefits of this exercise include the intrinsic rewards that to submit a list of up to ten projects in order of preference. The
come with solving real-life problems and the inculcation of the algorithm operated in a sequential manner through the list of
skills necessary for independent learning. It is believed that such preferences. A project preferred by a student-pair is allocated di-
projects lay the foundation for lifelong learning. rectly when there is no competition for that project. Otherwise,
the project is allocated randomly to one of the competing stu-
Manuscript received November 23, 2006; revised July 25, 2007. Current ver- dent-pairs. Thus, the algorithm proceeds with the next-ranked
sion published November 5, 2008. preferences for student pairs yet to be allocated any projects.
The authors are with the Department of Electrical and Computer Engi-
neering, National University of Singapore, Singapore 119260, Singapore The algorithm managed to allocate a project within the preferred
(e-mail: dipti@nus.edu.sg). list of 10 projects to 79.9% of the student pairs.
Color versions of one or more of the figures in this paper are available online In [6], Anwar and Bahaj proposed an integer programming
at http://ieeexplore.ieee.org.
Digital Object Identifier 10.1109/TE.2007.912537
approach which consumes less computational time. Their in-

0018-9359/$25.00 © 2008 IEEE


440 IEEE TRANSACTIONS ON EDUCATION, VOL. 51, NO. 4, NOVEMBER 2008

teger programming algorithm solved the same SPA problem in- II. THE STUDENT PROJECT ALLOCATION PROBLEM
stance as in [5] in just 6.66 minutes. However, on the down side,
Mathematically, the SPA problem is very similar to that de-
fewer student pairs were allocated projects that were ranked
scribed in [5] and [6]. The allocation of the FYPs is carried out
highly in their preferences using this method than by using that
as follows. A list of projects are proposed by academics within
of Teo and Ho in [5].
the department and by professionals working in several research
Both methods make significant strides towards feasibly
centres within the university. Faculty members propose projects
solving the SPA problem using nonhuman means, yet neither
by accessing a form where they provide the project title, a brief
provides the desirable compromise between computation com-
description, keywords, and information on any special skills or
plexity and student satisfaction. Further, SPA problems often
interests required. Each project has a unique project ID, and is
involve additional objectives, which are in conflict with student
recorded in a database. The list of projects is released to stu-
satisfaction. An important example of such an objective is the
dents, who are asked to select eight projects, rank their choices,
institutional requirement that there be a fair distribution of
and submit them online. Each student also has an associated
projects across staff supervisors. The lexicographic approach
cumulative average point (CAP), which is a measure of aca-
amenable to the methods in [5] and [6] is less appealing be-
demic performance (based on a scale of 0–5). The project allo-
cause this approach imposes a rigid preference structure. Little
cation exercise is carried out after all the students in the cohort
flexibility with respect to preference of tradeoff among the
have submitted the lists of their preferred projects. Typically, the
objectives is possible. This paper presents an evolutionary al-
number of students ranges between 500–650, and each project
gorithm (EA)-based project allocation system which is capable
allocation exercise involves 85–100 academic supervisors and
of efficiently finding an allocation scheme characterized by a
600–800 projects. The allocation results are announced to the
high student satisfaction level. Further, the EA may be easily
students one month before the start of the academic year.
modified to accommodate multiple conflicting objectives in a
The SPA problem is described as follows.
manner that allows the DM to specify his/her preference.
Let
The variant of the EA which caters to multiobjective SPA
1)
problems employs fuzzy aggregation to evaluate the desirability
be the set of students, , and the corresponding CAP
of tradeoff between objectives. The state of the art in evolu-
of the student;
tionary multiobjective optimization pursues a set of nondom-
2) be the set of subdepart-
inated solutions displaying different tradeoffs, and requires a
ments;
DM to choose from this set of optimal solutions [8]–[12]. The
3) be the set of staff supervi-
multiobjective formulation of the SPA problem presented here
sors;
deviates from the norm in employing a fuzzy, nonlinear aggrega-
4) be the set of projects
tion of objectives to evaluate the quality of candidate solutions.
offered.
The departure is motivated by the fact that in most multiobjec-
Each student is allowed to ballot for projects in order of
tive SPA problems, the DM has some notion of the acceptable
preference. Given these parameters, the goal is to find the pair
tradeoffs between objectives. Embedding this knowledge into
, for all , such that each student is assigned
the search process increases both its efficiency and effective-
a different project. Also, in such a single-objective formulation,
ness. The notion of acceptable tradeoffs is imprecise. Hence, it
the result of the project allocation process must optimize the
is best captured in terms of fuzzy rules and fuzzy membership
overall level of student satisfaction This problem can be defined
functions. The objectives are aggregated into a measure of the
as follows:
quality of candidate solutions based on the rules, membership
functions and operators. The resulting hybrid fuzzy-EA system
utilizes the search capability of EA and the mathematical fa-
cilities of fuzzy logic to express the imprecise knowledge of (1)
the DM. The hybrid fuzzy multiobjective EA (F-MOEA) thus
guides the solution to regions in the objective space where trade- where is the project allocated to student and
offs between objectives are desirable to the DM. is a function which returns the order of in ’s ballot list.
For completeness, both the single-objective optimization of If is not found in the list, a constant , where is
student satisfaction and the multiobjective optimization are the maximum number of projects a student may ballot for, is
presented in this paper. According to the particular needs of returned. A decrease in indicates that, on average, students
the institution, either approach could be readily adopted. The are assigned projects higher in their ranking of preferences.
EA-based project allocation system that has been used for the In actual practice, a more reasonable allocation would involve
past three years has not only resulted in a multifold reduction satisfying several objectives and criteria, some of which could
in time and improved matching of students to their desired be conflicting with each other. In this paper, the multiobjective
projects, but has done so while balancing various objectives formulation of the SPA problem involves four additional depart-
and constraints. The Sections II–V describe the context of the mental objectives. These objectives are described as follows.
current study, and demonstrate the workings and utility of the 1) Variation in loading across the subdepartments.
proposed approach. An application of the EA to three real Subdepartments may offer differing numbers of projects
datasets demonstrates the ability of the proposed approach to depending on their available resources; for instance, a sub-
find a project allocation that satisfies the criteria set. department with fewer staff supervisors will offer fewer
SRINIVASAN AND RACHMAWATI: EFFICIENT FUZZY EVOLUTIONARY ALGORITHM-BASED APPROACH 441

projects. In order to optimize the allocation of resources 4) Spread of academic achievement of assigned students
dedicated to supervision, it is desirable to have proportion- across academic supervisors.
ately similar loading across subdepartments. The expected Just as it is desirable to have an even spread of good stu-
number of students assigned to the subdepartment , is dents across subdepartments, department administrators
defined as follows: also find it desirable to balance the spread of academi-
cally-good students across academic supervisors. Let
be the mean CAP of students assigned to projects under
(2)
advisor . The spread of student academic performance
across staff members is defined as follows:
where is the total number of students and the total
number of projects. Let be the number of students assigned
to the subdepartment , a measure of loading spread
across the subdepartments is defined as follows:

(3) III. AN EA FOR SINGLE-OBJECTIVE SPA PROBLEMS


Evolutionary algorithms have earned wide popularity due
2) Variation in loading across staff supervisors.
to their ability to solve complex optimization problems in the
Depending on their resources, workload and schedule, su-
presence of multiple objectives and constraints [8]–[12]. The
pervisors may each propose a different number of projects.
EA-based solution approach presented in this paper was there-
Frequently, it is the case that a large number of students se-
fore developed and implemented for automating the project
lect projects offered by a small number of popular staff su-
allocation process. The allocation was based on students’ choice
pervisors as their most preferred option. This leaves other
of eight projects in their interest area. Before this algorithm was
staff who offers less popular projects with few or no stu-
implemented three years ago, the process of allocating projects
dent. Thus, efforts to balance the workload of staff mem-
to students was done manually, using a heuristic approach.
bers (such that the staff-student ratio is minimized) often
come into conflict with efforts to optimize the allocation of Finding a good solution was a very time consuming process,
projects according to the preferences of the students. The taking up to a month each time. An expert system based on
expected number of students each staff member oversees heuristics similar to that employed by Teo and Ho [5] was also
can be defined as follows: developed and tested before the EA-based allocation algorithm
was implemented. The proposed algorithm not only resulted in
higher success rates in allocating projects, but also allowed the
(4)
DM to obtain and analyze numerous allocation combinations
before deciding on the best combination to adopt.
where is the total number of students and the total
The EA is so-called because of its mimicry of the biolog-
number of projects.
ical processes of selection, recombination and mutation in com-
Let be the number of students assigned to staff advisor
putational operations, so as to arrive at optimal solutions to
. A measure of the spread of loading across staff members
a problem. The computational operators of selection, recom-
is given by
bination and mutation are applied iteratively to a set of pos-
sible solutions in order to move incrementally towards an op-
timum, as defined by the objective function. Selection discards
(5) low-performing solutions; recombination hybridizes a number
of high-performing solutions to obtain even better solutions; and
3) Spread of academic achievement of assigned students mutation introduces small solution variations that allow for a
across subdepartments. local search of better solutions in the neighborhood. The three
It is in the interest of the department to have a uniform main processes are repeated until a stopping criterion (usually
spread of academic achievement among the students as- the maximum number of iterations) is satisfied (Fig. 1).
signed to each subdepartment. This is to avoid a situation To solve the SPA problem using an EA, a representation
where a majority of the high-performing students are clus- scheme which facilitates the computational evolutionary oper-
tered in one particular subdepartment and low-performing ators is proposed as follows. Each candidate solution denotes a
students are clustered in another. Let be the mean CAP complete project assignment for the cohort and is represented
across the cohort and be the mean CAP of students as- as a string of positive integers denoting project IDs (Fig. 2).
signed to projects in subdepartment . The spread of stu- Each position in the string corresponds to a student in the cohort
dents with respect to academic achievement is defined to whom the project represented by the integer is allocated.
as follows: Initially, a number of such candidate solutions are generated
using a two-stage process. Let the number of solutions within
the population be denoted as . To generate a candidate
solution, students chosen in random succession are assigned a
(6) project picked randomly from within their own eight choices.
442 IEEE TRANSACTIONS ON EDUCATION, VOL. 51, NO. 4, NOVEMBER 2008

Fig. 1. Block diagram of EA.

Fig. 2. Candidate solution representation (P denotes the project ID assigned


to student n).

Fig. 4. Modification to new solutions (mutation).

in which case the position is left unassigned. For the same posi-
tion in , the complementary string is chosen and copied with
the same restriction that project IDs already assigned to other
positions in cannot be reused.
This operation is repeated for all positions in the string while
maintaining corresponding lists of unallocated projects, one
each for and . To avoid positional bias, the order in which
positions in and are chosen is again randomized. Students
Fig. 3. Hybridization of two candidate solutions (crossover). corresponding to empty positions in and are assigned
projects randomly from the list of unallocated projects. At the
end of the hybridization process, two new complete project
Some students are not assigned any projects by the end of this assignments for the cohort, and , undergo further modifi-
stage because of the competition for popular projects. These cation with a probability of as illustrated in Fig. 4. For each
students are allocated projects randomly from the list of free new solution, students are randomly selected and assigned
projects. The randomness helps alleviate possible bias arising a project from the corresponding students’ ballot list. If the
from the order in which students are encoded in the represen- string contains another student who has been assigned to the
tation scheme. Such initialization ensures that each student is same project ID, this student is assigned a new project that
allocated a unique project in the candidate solutions. is randomly selected from the list of available projects. This
The stopping criterion in the proposed EA is defined as the modification tweaks new solutions in order to find assignment
maximum number of iterations and is denoted by MAXITER, schema that are absent in the evolved population. The search
which is set to zero at initialization. Following the initialization is beneficial when good assignment schema are found and
of candidate solutions, a binary tournament with replacement retained in the successive iterations.
is applied to the group of candidate solutions. At each round, After the modification process, new solutions that have exact
two candidate solutions are chosen randomly and evaluated in duplicates in the evolved population of solutions are discarded.
terms of , which is the objective function in the optimization. The binary tournaments, hybridization and modification, are
The solution with the lower value of is selected to be the first repeated until new solutions are obtained. Before pro-
parent . The random selection of two candidate solutions is ceeding to the next iteration, both sets of old solutions
performed again to obtain the second parent . and new solutions are evaluated and sorted in terms of
The pair, , is then subjected to a process of hybridiza- their performance based on objective . The top 20% of old
tion (Fig. 3) with a probability of to generate two new can- solutions are then combined with the top 80% of new solutions
didate solutions, and . To construct and , one solu- to make up candidate solutions in the next iteration. Retaining
tion from and is randomly chosen for each position in a number of best candidate solutions from the old population
the string. The project ID assigned to the position in the chosen helps improve the convergence properties of the EA, while
string is copied to the corresponding position in , unless that making room for new candidate solutions in the evolutionary
project ID has already been assigned to another position in , process [8], [9].
SRINIVASAN AND RACHMAWATI: EFFICIENT FUZZY EVOLUTIONARY ALGORITHM-BASED APPROACH 443

TABLE I
OBJECTIVES IMPROVEMENT AND DETERIORATION WITH SINGLE-OBJECTIVE OPTIMIZATIONS

IV. A HYBRID FUZZY EA FOR MULTIOBJECTIVE improvement in the other three criteria. Similar effects are ob-
SPA PROBLEMS served in the single-objective optimization of , , and .
The investigation demonstrated that no conflict is present be-
In this section, the single-objective EA for solving the SPA tween the last four criteria.
problem (presented in Section III) is modified to accommo- The SPA problem is therefore posed as a biobjective opti-
date four other departmental objectives. The departmental cri- mization problem with as the first objective function. As the
teria may be described as linguistic, perception-based and ex- four criteria are not identical, concurrent, effective optimization
perience-based and can be incorporated into the EA using the of the four requires that all four criteria ( , , , and ) are
representation and approximate reasoning facilities provided by reflected in the fitness function of the EA. This is achieved by
fuzzy logic [13], [14]. In particular, the objectives are repre- combining the four criteria in a weighted sum, the second ob-
sented as fuzzy variables, which act as inputs to a fuzzy infer- jective function,
ence system (FIS) evaluating the fitness of the associated candi-
date solution. The reason for incorporating fuzzy logic in this
manner is twofold. First, fuzzy logic offers an effective rep- (8)
resentation and reasoning framework in evaluating the quality
of candidate solutions. Second, the fitness function guides the The weights were selected to reflect the relative
search to desirable regions of the fitness space [14] based on the emphasis the DM places on each of the criteria , , ,
prespecified preference of the user. and . In the implementation described by this paper, the fol-
In the heuristic-based method applied to the department’s lowing were selected: , . The
SPA problem in the past, the multiple objectives were optimized weights corresponding to supervisor satisfaction are set higher
in consecutive stages of allocation. Such a lexicographic ap- than those corresponding to balance across the subdepartments
proach is less than ideal, as separately optimizing the different as by optimizing the former, the latter is also improved. Because
objectives does not give the DM a perspective of the character- subdepartmental balance may be improved without sacrificing
istics of possible tradeoffs. balance across supervisors, nonzero weights are attached to
The hybrid fuzzy-GA approach allows the judgment of the and to attain further possible improvement in the two criteria.
human DM to motivate the definition of the fitness function. The For each candidate solution, and are computed and aggre-
human DM considers various objectives and their combined in- gated by the FIS (Fig. 5) to arrive at a fitness value that reflects
teraction and aggregates the level of satisfaction for all objec- the DM’s evaluation of the corresponding solution.
tives to evaluate a candidate solution’s overall quality. Fuzzy The values of objective functions and are fuzzified ac-
logic was introduced to capture the DM’s perceptions and rea- cording to the membership functions defined for each respective
soning in this process and provide an aggregator for the values objective. With the fuzzy quantities as arguments, the fuzzy in-
of the objective functions. The effect of the incorporation of ference engine evaluates the fuzzy fitness score based on fuzzy
fuzzy logic in this manner is the convergence of solutions to- rules and operator definitions. The fitness score is obtained by
wards areas where the DM deems a tradeoff between objectives defuzzification based on the membership function defined for
to be favorable. fitness values [13]. The database stores the various membership
To solve the multicriteria problem defined in Section II, a functions, fuzzy operators and fuzzy rules. A detailed model of
FIS computes the fitness function in the following manner. In the FIS used in this framework is presented in subsection A,
Section II, five criteria ( , , , , and ) were described. which follows.
Results from ten runs of such single-objective optimizations are
presented in Table I. Single-objective optimizations of each the A. FIS Model
five criteria reveal that while a minimization of consistently The FIS in the proposed approach defines a mapping func-
results in direct degeneration of the other objectives, such con- tion from the objective space to the fitness space,
flict is not present between , , , and . which is also known as “the decision space.” Let be a candi-
Likewise, single-objective optimization of the criteria con- date solution, be the values of objective
sistently results in the degeneration of , but brings about small functions and fit be the fitness of
444 IEEE TRANSACTIONS ON EDUCATION, VOL. 51, NO. 4, NOVEMBER 2008

Fig. 5. Block diagram of the fitness function.

Fig. 6. Membership functions for f and f .

where and fit are real numbers. The mapping of to the func- act as the prototype. The rules specify in the consequent
tion is nonlinear and its definition consists of the following. the outputs for a particular prototypical combination of
1) Membership functions for each of the V inputs. Within input data. Averaging constitutes interpolation to obtain
the universe of discourse of input , fuzzy sets are output for nonprototypical inputs. The weights reflect the
defined by the corresponding membership functions . degree of similarity between the input data and the defined
2) A set of fuzzy rules of the form prototypes. For a given input , the output
of the FIS is given by the following fitness score:

(10)

where is the -th fuzzy set of the -th input, and


is a real number. The conjunctive term “and” may also be where is the activation strength of rule
replaced by the disjunctive term “or.” (11)
3) Mathematical operators for disjunction, conjunction
and implication for evaluating the rules. Disjunction For a disjunction of terms, the operator is replaced by .
and conjunction represent an aggregation of the perfor- The design of the FIS components (membership functions
mance of a particular candidate solution on the different and rules) is based on a DM’s expert knowledge. The mem-
objective functions. Disjunction is employed to model bership functions for each input reflects the DM’s perception
independence between the satisfaction-level of objectives of the value of each individual objective. The antecedent part of
concerned whereas conjunction represents interdepen- the fuzzy rule expresses combinations of objective values and
dence. In the case of independence, a particular amount the consequent part reflects the DM’s preference for the asso-
of improvement in one of the objectives concerned in- ciated combination. This preference reflects the importance of
creases the satisfaction of the DM by the same amount, objectives relative to all other combinations of objectives. The
irrespective of the performance in the other objectives. In rule-base could be constructed either manually, when not many
the case of interdependence, the gain in the DM’s satis- rules are involved, or by conducting a survey of the DM’s pref-
faction depends both on the magnitude of improvement erences and then employing available techniques to construct an
and the associated performance in the other objectives. FIS from the data.
In this paper, addition and multiplication are employed The membership functions for each input and output are
for disjunction and conjunction [14]. Implication is designed according to the DM’s perception of the quantities.
achieved with a weighted-average of the consequents Three membership functions corresponding to three granular-
in the fuzzy rules. Fuzzy inference may be viewed as a ities [small (S), medium (M), and large (L)] were designed
prototype-matching exercise, in which input data which for objectives and (Fig. 6). Fuzzy rules were manually
corresponds to unity membership in the defined fuzzy sets designed, resulting in the fitness landscape illustrated in Fig. 7.
SRINIVASAN AND RACHMAWATI: EFFICIENT FUZZY EVOLUTIONARY ALGORITHM-BASED APPROACH 445

On all three data sets, the expert system based on heuris-


tics achieved a lower success rate compared to the EA-based
methods (Table II). Although the use of the expert system
(which sequentially allocates the first choices, then the second
and so on) resulted in similar or larger percentage of students
being assigned their first choice, the EAs produced a more
balanced allocation that maximized overall levels of student
satisfaction. A larger percentage of students were allocated one
of their first three choices as a result of the single-objective EA
(75.7%, 85.16%, and 81.12% for DS-1, DS-2, and DS-3, re-
spectively) than through the heuristic method (70.2%, 73.30%,
and 70.92% for DS-1, DS-2, and DS-3, respectively). Cor-
responding figures for the biobjective F-MOEA are 74.26%,
Fig. 7. Fitness landscape. 80.26%, and 78.7% for DS-1, DS-2, and DS-3, respectively.
While the single-objective EA produces several solutions at
each iteration (the best of which can be used for project alloca-
A high level of fitness represents a desirable combination of tion), the solutions obtained from the fuzzy multiobjective algo-
and values. rithm show a simultaneous optimization of various objectives.
The landscape also captures the preference of over There were different tradeoffs between the objectives, one of
(a 10% improvement in is rewarded with higher scores than which was chosen based on the DM’s preference.
the same improvement in , other variables being constant). A number of results with slightly differing tradeoff character-
The fitness function favours the propagation of individuals istics were obtained at the end of the experiments. The results
with favorable tradeoffs of and via the mechanism for DS-3 are plotted in Fig. 8 in terms of the numeric values of
of selection at both the pre-recombination and population and as defined in Sections II and IV, respectively. Note
replacement stages. Further fine-tuning may be required if that the results were concentrated on a small area in the objec-
the FIS does not focus the search on the desired region of the tive function space and that all the final solutions were of high
objective space. As a rule of thumb, shifting the membership quality, with a success rate ranging from 92.0% to 97.2%. A so-
of fuzzy sets describing the desired optimum (i.e., “small” for lution with a reasonable tradeoff between and was finally
minimization problems and “large” for maximization problem) selected for implementation.
towards the ideal optima increases the relative importance of The multiobjective variant (F-MOEA) can be said to find a
the objective. balance between the various objectives mentioned in Section II.
Across the three data sets, the biobjective F-MOEA is observed
V. RESULTS AND DISCUSSION to attain lower allocation success in comparison to the single-ob-
This section presents the results obtained on three datasets jective EA. The detriment in student satisfaction is traded-off
(DS-1, DS-2, and DS-3) using the single-objective algorithm for a more equitable spread across the supervisors and subde-
presented in Section III, as well as the multiobjective fuzzy algo- partments. The allocation results from the single-objective EA
rithm described in Section IV. Further, the heuristic presented are characterized by an uneven spread of students across subde-
by Teo and Ho [5] was implemented and applied to the same partments and across academic supervisors. One subdepartment
datasets. The datasets DS-1, DS-2, and DS-3 were obtained was assigned too many students, while another had too few stu-
from three cohorts of 575, 519, and 544 students, respectively. dents. Additionally, there was a poor balance of load across staff
The number of projects offered by academic staff members were members. The maximum student to staff ratio observed is 10 : 1
629, 764, and 762, respectively. The allocation effort involved (two cases) and the minimum is 4 : 1 (five cases). In contrast, the
four subdepartments with about 90 academic supervisors. F-MOEA results in an almost perfect equalization of the load
The experiments were conducted on a Pentium IV 3.0 GHz across subdepartments and staff supervisors. Fig. 9 shows the
CPU. Each run of the algorithms on these data sets produced a deviation of the obtained student-staff assignments from the ex-
population of 150 candidate solutions over MAXITER pected number, averaged over the results obtained from the three
iterations, which consumed an average of 50 minutes of CPU data sets.
time. Results of the application of these algorithms are presented In allocation schemes produced by the single-objective EA,
in Table II. The success rate is measured in terms of student sat- considerable variation in the average CAP of assigned students
isfaction, and defined as the percentage of students who were al- across the four subdepartments is observed. The range is 0.25
located a project of their choice. The average computation time and the standard deviation 0.10. The corresponding figures for
taken by the expert system-based algorithm was 24 hours, while staff members are 2 and 0.3039, respectively. The F-MOEA,
the time taken by the single-objective EA was 50 minutes. The on the other hand, yielded a more uniform distribution of CAP
F-MOEA took longer (three hours) due to the additional calcu- among the academic supervisors. Across subdepartments, the
lations necessary to evaluate the various objectives. In terms of average CAP of assigned students had a range of 0.05 and a stan-
computing time, all these approaches compare very well to the dard deviation of 0.02. Across the staff supervisors, the average
manual approach which took approximately four weeks to find CAP had a range of 1.2 and a standard deviation of 0.17, indi-
a reasonably good solution. cating a fairer distribution compared to the other two algorithms.
446 IEEE TRANSACTIONS ON EDUCATION, VOL. 51, NO. 4, NOVEMBER 2008

TABLE II
PROJECT ALLOCATION RESULT: PERCENTAGE OF STUDENTS ALLOCATED PROJECTS OF THEIR CHOICE

The F-MOEA also shows distinct advantages over the


two-stage optimization strategy proposed by Anwar and Bahaj
[6] in which the two objectives are optimized separately.
This approach is suitable when a number of hard boundaries
exist. However, SPA problems often involve criteria with soft
negotiable boundaries. For instance, a slight deterioration in
the staff loading balance may be acceptable if this allows for
a significant improvement in student satisfaction. Also, when
N objectives are involved, the approach put forth by Anwar
and Bahaj in [6] requires at least N stages of single-objective
optimization. In contrast, the multiobjective EA proposed in
Fig. 8. The f and f scores of final candidate solutions for DS-3.
this paper requires only one stage of optimization.

VI. CONCLUSION
This paper has presented a FYP allocation system that is
based on EAs. It has been developed and implemented in a large
university department in Singapore to automate the process
of project allocation, and to improve the process of matching
students to their desired projects. The project allocation system
assigns a suitable project to each student, maximizing a combi-
nation of several criteria including student preferences, project
prerequisites, and staff loading. The proposed single-objective
and multiobjective algorithms were extensively tested on real
datasets and their performance was compared with a heuris-
tics-based expert system approach. The algorithms obtained
excellent results within a reasonable time frame. Despite its
longer computation time and slightly lower success rate, the
Fig. 9. Deviation in the number of assigned students from the expected number F-MOEA achieved a balanced distribution of students across
across staff supervisors. subdepartments and supervisors.
Although the two variants of the EA were developed for a spe-
cific SPA problem, the algorithm can be adapted for other SPA
By incorporating the DM’s perceptions and reasoning in eval- problems with little modification. In contrast, the manual design
uating the fitness of candidate solutions, the multiobjective EA of the FIS in the multiobjective variant may prove to be a bot-
guided the population stochastically to the desired regions in the tleneck for an SPA problem with more objectives. The authors
objective space. Further, the simultaneous optimization allowed suggest that future efforts be channelled towards supporting the
the exploration of tradeoffs between the objectives, which may automated design of the fuzzy rule base. Such support could be
differ from one problem instance to another. The results indicate in the form of surveys of the preferences of DMs such that mem-
the clear superiority of this algorithm over the heuristics-based bership functions and fuzzy rules can be developed using data
expert system and the single-objective EA. mining techniques.
SRINIVASAN AND RACHMAWATI: EFFICIENT FUZZY EVOLUTIONARY ALGORITHM-BASED APPROACH 447

REFERENCES [11] N. Srinivas and K. Deb, “Multiple-Objective optimization with non-


dominated sorting genetic algorithms,” Evol. Comput., vol. 2, no. 3,
[1] T. Ibaraki and N. Katoh, Resource Allocation Problems: Algorithmic pp. 1556–1561, 1995.
Approaches. Cambridge, MA: MIT Press, 1988. [12] E. Zitzler and L. Thiele, “Multiobjective evolutionary algorithms: A
[2] T. Ichimori, “On min-max integer allocation problems,” Oper. Res., comparative case study and the strength Pareto approach,” IEEE Trans.
vol. 32, no. 2, pp. 449–450, 1984. Evol. Comput., vol. 3, no. 4, pp. 257–271, Nov. 1999.
[3] G. A. P. Kindervater and H. W. J. M. Trienekens, “Experiments with [13] L. A. Zadeh, “Knowledge representation in fuzzy logic,” IEEE Trans.
parallel algorithms for combinatorial problems,” Eur. J. Oper. Res., Knowl. Data Eng., vol. 1, no. 1, pp. 89–100, Mar. 1989.
vol. 33, no. 1, pp. 65–81, Jan. 1988. [14] S. Dick, “Toward complex fuzzy logic,” IEEE Trans. Fuzzy Syst., vol.
[4] T. C. Koopmans, “Efficient allocation of resources,” Econometrica, 13, no. 3, pp. 405–414, Jun. 2005.
vol. 19, no. 4, pp. 455–465, Oct. 1951.
[5] C. Y. Teo and D. J. Ho, “A systematic approach to the implementation
of final year project in an electrical engineering undergraduate course,”
IEEE Trans. Educ., vol. 41, no. 1, pp. 25–30, Feb. 1998. Dipti Srinivasan (S’91–M’97–SM’02) received the Ph.D. degree in engi-
[6] A. A. Anwar and A. S. Bahaj, “Student project allocation using integer neering from the National University of Singapore (NUS), Singapore, in 1994.
programming,” IEEE Trans. Educ., vol. 46, no. 3, pp. 359–367, Nov. She worked as a Postdoctoral Researcher at the University of California,
2003. Berkeley, from 1994 to 1995 before joining NUS, where she is an Associate
[7] A. Khambadkone, A. A. Kassim, and G. Samudra, “Design of assess- Professor in the Department of Electrical and Computer Engineering. She was
ment for project modules with dissertation type of session,” in Proc. a Visiting Professor with the Department of Electrical Engineering, Indian In-
Int. Conf. Engineering Education, Manchester, U.K., Aug. 2002, pp. stitute of Science, Bangalore, India, from 1998 to 1999. Her research interest
1–4. is in application of soft computing techniques to engineering optimization and
[8] D. A. Van Veldhuizen and G. B. Lamont, “Multiobjective evolutionary control problems.
algorithms: Analyzing the state of the art,” Evol. Comput., vol. 8, no.
3, pp. 125–148, 2000.
[9] J. Horn and N. Nafpliotis, “Multiobjective optimization using the
niched Pareto genetic algorithm,” Illinois Genetic Algorithms Lab, Lily Rachmawati (S’06) received the B.Eng. degree from the National Univer-
Univ. Illinois, Urbana-Champaign, 1995, IlliGAL Rep. 93005. sity of Singapore (NUS), Singapore, in 2004.
[10] K. Deb, A. Pratap, S. Agarwal, and T. Meyarivan, “A fast and elitist She is working towards the Ph.D. degree at NUS. Her research interests in-
multiobjective genetic algorithm: NSGA-II,” IEEE Trans. Evol. clude multiobjective optimization, evolutionary computing techniques, and hy-
Comput., vol. 6, no. 2, pp. 182–197, Apr. 2002. brid computational intelligence techniques.

You might also like