Proceedings of 2009 4th International Conference on Computer Science & Education

Design of Personalized Test Paper Generating System

of Educational Telenet Based on Genetic Algorithm

Yong Ou-Yang Hong-Fang Luo

School of Computer Science School of Computer Science
Hubei University of Technology Hubei University of Technology
Wuhan, China Wuhan, China
oyywuhan@l63.com hmily1946@163.com

Abstract—When generating test paper, traditional distant II. MATHEMATICAL MODEL FOR PAPER GENERATING
education online examination system either merely extracts the PROBLEM
same test questions from item bank or extracts test questions at
random, or select test questions according to some classifications Composition of one set of test paper involves many factors,
of student. It can not individually varies and give custom made and each test question includes many attributes. Attributes
scheme according to student’s learning situation. This paper associated with paper generating includes time, question type,
improved the traditional genetic algorithm and proposed a new chapter, knowledge point, difficulty coefficient, level and score
method of test paper generating based on student’s learning these seven items. Time is what students needed to finish test
situation. It can generate personalized test paper dynamically question in normal. Question type is divided into multiple
that satisfying every student according to student’s online self- choice, judgment question and essay question. Chapter and
condition in distance education network and factors reflected knowledge point identify concrete content that examined by
during learning such as psychological status, comprehension and test question. Whether test question is easy or difficulty is
application level. decided by difficulty coefficient. Level is defined according to
requirements such as concept, theoretical analysis and
Index Terms—Distance education, test paper generating,
application. Score value of test question is marked by score.
genetic algorithm
The seven attributes above can be got when one test
I. INTRODUCTION question is extracted from item bank. A piece of test paper is
Educational telenet has developed greatly and prevailed in composed of n test questions which extracted from item bank,
recent years. But it lacks of personalization and differentiation, and an attributive matrix(as in (1)) of n*7 test questions can be
therefore, putting up personalized and intelligent educational got, of which each column represents one attribute, and each
environment scientifically according to different users has row one test question.
became a hot issue of research [1][2][3]. A lot of research in
this field has been done and favourable results has been
⎡ a11 a12 a13 a14 a15 a16 a17 ⎤
achieved, such as conceptual model[4] providing guiding ⎢a
opinions to students and group decision evaluating education ⎢ 21 a22 a23 a24 a25 a26 a27 ⎥⎥
website by Hwang[5], and a personalized online learning k = ⎢a31 a32 a33 a34 a35 a36 a37 ⎥ . (1)
system based on fuzzy item response theory[6] by Chih-Min in ⎢ ⎥
2008. ⎢ ... ... ... ... ... ... ... ⎥
⎢⎣an1 an2 an3 an4 an5 an6 an7 ⎥⎦
As one of the most important parts in intelligent distance
education, online examination system is a key point to
Obviously, this is the target state matrix in solving paper
research. Its function mainly includes online examination
generating problem, which should meet some constraints as in
management, intelligent test paper composition and
(2),(3),and (4).
achievement evaluation, of which intelligent test paper
composition is the main embodiment of personalized service. n
Now most existing online examination systems, which merely total _ time = ∑ a i1 . (2)
extract test papers from item bank according to requirements i =1
input by teachers in advance and certain strategies of
generating paper, can not extract paper suiting each student n
according to students’ specific learning situation. This paper total _ score = ∑ a i7 . (3)
proposed an intelligent test paper composition method based on i =1

genetic algorithm, which extracts suitable test questions

dynamically from item bank according to students’ online a i3 ∈ [chapter _ range] . (4)
learning situation and meet students’ personalized test
requirements well.

170 978-1-4244-3521-0/09/$25.00 ©2009 IEEE

 Proceedings of 2009 4th International Conference on Computer Science & Education
The test content and difficulty should be different according
to different student, constraints that the attributes from the 4th
column to the 6th column should meet, are established
dynamically according to student knowledge learning model
before running the algorithm of test paper generating .
Modeling learning situation of each student, student
knowledge learning model reflects students’ performance in
intelligent educational environment during learning.
It regards 6 evaluation values and 15 quantities as reference
factors to evaluate student’s learning situation, the former
includes memorization, comprehension, application, analysis,
integration and assessment; the latter contains performance of
the test, age, educational background, physiological conditions,
learning environment, mood, learning efficiency, etc. It
receives estimation of degree that student masters in terms of
concept, theory and application. Detailed content can be
referenced in [7]. We can refer from the above that test paper
generating problem is an optimal solving problem of multi-
target, and the optimal solving is not unique.
III. DESIGN OF GENETIC ALGORITHM FOR TEST PAPER
GENERATING PROBLEM
A. Traditional test paper genetic algorithm
Traditional test paper generating such as random extraction
and backtracking, with low success rate of test paper
generating, high space cost and time consumption, is suitable
for item bank system with small scale. Genetic algorithm was
proposed by J.H.Holland, an American scholar. Simulating
biological evolution process in nature, genetic algorithm is a
computing model with simple practice, strong robustness,
global optimization and fast convergence. It is also suitable for
parallel processing, and all these features are suitable for
solving test paper generating problem.The concrete process that
adopts genetic algorithm is as following:
1) To choose certain quantity of chromosome to compose
an initial population.
2) To adopt some kind of fitness function to compute
inferiors and superiors of individuals in the population. The
higher fitness is, the better variety is.
3) To choose chromosome to do genetic operation such as
replication, crossover and mutation according to fitness value.
4) To modify illegal solution, compute fitness for new
individuals, and check the new population whether satisfies
test paper generating constrains or not, if it is so, then output
the solution, or else replace parent individuals by new
individuals to compose population of the next generation, then
go to step 2).
B. Improvement of Genetic Algorithm
In traditional genetic algorithm, the total number of “1”in
each string of initial population equals n the total number of
test questions. But it may change after genetic operation such
as crossover and mutation and become an illegal solution. At
this time, it is needed to correct the solution, that is, doing
computation and making the number of “1” in the string into
“n”. Generally speaking, this process is complex and of an
great increase in computation. New individuals replace old
978-1-4244-3521-0/09/$25.00 ©2009 IEEE
individuals totally, so early maturity may come out and result
into local optimum. In addition, requirements for the 7
attributes are different, error is not expected for test score,
certain requirements are desired as for other attributes.
Improvements of the traditional genetic algorithm are made as
following:
1) Coding scheme: Extraction for each test question is
regarded as gene, “1”represents some test question is selected
while “0” not. Different combination of gene is considered as
chromosome which supports blocking according to function.
For example, the chromosome 1000010101-101010-101101-
11011 shows that test questions of the 1st, 6th, 8th and 10th are
extracted 10 item banks to compose the test paper. In this way,
the whole binary string is divided into different function
blocks. Each function block, which adopts independent binary
code, responds to one kind of question type, in which
“1”represents the test question is selected while “0” not and the
t number of “1” is the number of this kind of question type.
Obviously, the initial population generated by the rule above
has already satisfied requirements for test question type, score
and time to complete test paper.
2) Fitness function: Fitness function is used to judge
virtues or defect degree of individuals in the population.
Genetic algorithm takes advantage of fitness value to guide
search, rather than continuity and differentiability of fitness
function and other auxiliary information. The definition of
fitness function is as in (5).
m
F = 1/ (1 + ∑ k i ei ),i ∈ [1, n] . (5)
i =1
in which ei responds to error that the ith test paper generating
factor corresponding to test paper constraint, generating ki is
weight coefficient and ki>0. The fitness function with weighted
error reflects characteristics solving for test paper generating.
The smaller the error that test paper individuals responds to
each constraint of test paper generating , the bigger the fitness
value is, and the closer the individuals close to test paper
generating.
3) Crossover operation: Crossover is the most important
operation of genetic algorithm. There are many crossover
operators, single-point crossover operator which is the most
commonly used is adopted in this paper. And internal crossover
operator also adopted to make sure that number of question
choosing is not changed. In other words, to do crossover
operation in a single chromosome rather than two
chromosomes, as fig.1 shows.
cross-point
1 0111 0 1 0 0 0 1 11 0 1 1
Figure 1. crossover operation
4) Mutation operation: Intra-group mutation of two bit is
adopted for mutation operator, namely to mutate two bits in a
171
 Proceedings of 2009 4th International Conference on Computer Science & Education
group of question type at one time, change one bit from “1” to
“0” and other from “0” to “1”, and coding block of each
chromosome mutate independently. Execution of mutation
operator depends on mutation probability which is very small
in general. Execution process of mutation operator is as below:
a) To choose some bit randomly in chromosome and to
reverse it.
b) To reverse in the coding block where the bit is
included in, search for a bit with opposite value and whose
place is closest to the bit, and reverse the value of the bit
found, as fig.2 shows.
001 1 1 0 11 001 0 1 0 11
001 0 1 1 11
Figure 2. Mutation operation
In this way, requirements of constraints, question type and
question number are ensured.
5) Elimination operation: In easy genetic algorithm,
individuals of child generation after crossover take place of
individuals of parent generation that take part in crossover,
abandoning information of parent generation fast. But the new
individuals generated by crossover are not always better than
parent generation, which is not good for fast convergence of
algorithm. For the sake of efficient crossover and make sure of
both the new individuals and parent individuals participate in
evaluation and selection, N individuals, which are selected
from 2N individuals with roulette selection method, compose
new population. N represents scale of population.
6) Terminal conditions: Basis for finishing iterative
process:
a) The population meets constraints of test paper
generating;
b) The fitness values of two successive are less than the
specified value;
c) The iterative value is maximum value.
It can be seen that the number of “1”in each chromosome
won’t be changed after improved genetic operation on
chromosome in population of each generation, avoiding
modifying illegal solution and reducing operations greatly. At
the same time the parent generation and the child generation
participate in struggle for existence, giving more chance to
better individuals that reproduce to the next generation and
higher probability to get a global optimal solution. Details of
the algorithm can be seen in fig.3.
IV. CONCLUSIONS
This paper takes advantage of genetic algorithm to design a
personalized online examination and test paper generating
system, which evaluates students’ mastery degree of
knowledge comprehensively and accurately. At the same time,
the improved genetic algorithm with high executive speed and
172 978-1-4244-3521-0/09/$25.00 ©2009 IEEE
efficiency, which is used to compose test paper that is more
suitable for testing students’ knowledge level in test paper
generating system, so as to achieve the goal to assistant
teaching with intelligence and personality.
Algorithm:Improved Genetic Algorithm
Begin:
Step 1: Initial parameters of genetic algorithm, let t=0, M is
scale of population, K is the number of chromosome in
recent generation which will be replicated to the next
generation;
Step 2: To generate initial population (t=1);
Step 3: To compute fitness function value P(gi) for each
chromosome;
Step 4: To sort chromosomes in descending order according to
P(gi);
Step 5: To check whether satisfying the condition to finish
iterative process or not, if it so end execution and exit
the algorithm;
Step 6: To replicate the preceding K chromosomes in
population (t) to temporary storage area;
Step 7: Do partheno-genetic operation on each chromosome gi
in population (t);
Step 8: To compute fitness of each daughter chromosome，
and to replicate it to temporary storage area;
Step 9: To sort K parent chromosomes and M child
chromosomes generated in descending order according
to fitness, the former is memorized in step 6 and the
latter is generated in step 7;
Step 10: To choose the preceding M chromosomes from
temporary storage area and regards them as group(t ← t
+ 1) of the next generation.
Step 11: Go back to step 5;
End;
Figure 3. Improved genetic algorithm
ACKNOWLEDGMENT
The authors want to thank Zhong-bao Ma for discussions,
valuable suggestions and revisions that contributed to the
development of this work. We are also grateful to Professor
Chun-zhi Wang,and to the anonymous referees for their useful
comments.
REFERENCES
[1] G. Eason, B. Noble, and I. N. Sneddon, “On certain integrals of
Lipschitz-Hankel type involving products of Bessel functions,” Phil.
Trans. Roy. Soc. London, vol. A247, pp. 529–551, April 1955.
(references)
[2] J. Clerk Maxwell, A Treatise on Electricity and Magnetism, 3rd ed., vol.
2. Oxford: Clarendon, 1892, pp.68–73.
[3] I. S. Jacobs and C. P. Bean, “Fine particles, thin films and exchange
anisotropy,” in Magnetism, vol. III, G. T. Rado and H. Suhl, Eds. New
York: Academic, 1963, pp. 271–350.
[4] K. Elissa, “Title of paper if known,” unpublished.
[5] R. Nicole, “Title of paper with only first word capitalized,” J. Name
Stand. Abbrev., in press.
[6] Y. Yorozu, M. Hirano, K. Oka, and Y. Tagawa, “Electron spectroscopy
studies on magneto-optical media and plastic substrate interface,” IEEE
Transl. J. Magn. Japan, vol. 2, pp. 740–741, August 1987 [Digests 9th
Annual Conf. Magnetics Japan, p. 301, 1982].
 Proceedings of 2009 4th International Conference on Computer Science & Education

[7] M. Young, The Technical Writer's Handbook. Mill Valley, CA: University Science, 1989.

978-1-4244-3521-0/09/$25.00 ©2009 IEEE 173