01---An Attribute-based Ant Colony System for Adaptive Learning Object Recommendation

Available online at www.sciencedirect.com

www.elsevier.com/locate/eswa

Expert Systems with Applications 36 (2009) 3034–3047

Expert Systemswith Applications

An attribute-based ant colony system for adaptive learningobject recommendation

Yao Jung Yang a,b,*, Chuni Wu a

a Department of Information Management, Hsing-Kuo University of Management, No. 89, Yuying Street, Tainan 709, Taiwanb Department of Information Technology, Soochow University, No. 1, Shizi Street, Jiangsu, China

Abstract

Teachers usually have a personal understanding of what ‘‘good teaching” means, and as a result of their experience and educationallyrelated domain knowledge, many of them create learning objects (LO) and put them on the web for study use. In fact, most studentscannot find the most suitable LO (e.g. learning materials, learning assets, or learning packages) from webs. Consequently, many research-ers have focused on developing e-learning systems with personalized learning mechanisms to assist on-line web-based learning and toadaptively provide learning paths. However, although most personalized learning mechanism systems neglect to consider the relationshipbetween learner attributes (e.g. learning style, domain knowledge) and LO’s attributes. Thus, it is not easy for a learner to find an adap-tive learning object that reflects his own attributes in relationship to learning object attributes. Therefore, in this paper, based on an antcolony optimization (ACO) algorithm, we proposed an attributes-based ant colony system (AACS) to help learners find an adaptivelearning object more effectively. Our paper makes three critical contributions: (1) It presents an attribute-based search mechanism to findadaptive learning objects effectively; (2) An attributes-ant algorithm was proposed; (3) An adaptive learning rule was developed to iden-tify how learners with different attributes may locate learning objects which have a higher probability of being useful and suitable; (4) Aweb-based learning portal was created for learners to find the learning objects more effectively.� 2008 Elsevier Ltd. All rights reserved.

Keywords: Adaptive learning; Ant colony optimization; Learning style

1. Introduction

In the classic teacher-centered situation, the course isbuilt from the content defined by the teacher or author,and most teachers or educators agree that, in the designand development of educational material, attention mustbe focused on learner characteristics and requirementsand defined in terms of content and learning style. Teachersusually have a personal understanding of what ‘‘goodteaching” means as a result of their experience and educa-tional related domain knowledge level in education, and

0957-4174/$ - see front matter � 2008 Elsevier Ltd. All rights reserved.

doi:10.1016/j.eswa.2008.01.066

* Corresponding author. Address: Department of Information Man-agement, Hsing-Kuo University of Management, No. 89, Yuying Street,Tainan 709, Taiwan. Tel.: +886 6 2871511; fax: +886 6 2870917.

E-mail addresses: [email protected] (Y.J. Yang), [email protected] (C. Wu).

they create learning resources to put on the web for thepurpose of study. In fact, most students cannot find themost suitable learning objects from the web because eachLO has different attributes (e.g. learning object level, learn-ing type), and each individual learner also has differentcharacteristics or attributes (e.g. learning style, domainknowledge level). Adaptive learning provides an alternativeto the traditional ‘‘one size fits all” approach and has dri-ven the development of teaching and learning towards adynamic learning environment. Thus, getting an adaptiveLO to suit learners personalized needs is an importantissue. Two major problems arise here: the ‘‘one size fit all”approach gives the same learning materials to each learner(Brusilovsky, 2001; Stewart, Cristea, Brailsford, &Ashman, 2005), and the immense amount of informationavailable leads to information overload (Berghel, 1997).Thus, adaptive learning has gained more attention in recent

Y.J. Yang, C. Wu / Expert Systems with Applications 36 (2009) 3034–3047 3035

years (Manouselis & Sampson, 2003; Melis et al., 2001;Weber, 1999). Several intelligent proposals have beendeveloped: dynamic learning recommendation (Chen,Huang, & Chu, 2005; Huang, Chen, Huang, Jeng, &Kuo, 2007), intelligent learning contents suggestions(Huang, Chen, Kuo, & Jeng, 2007), adaptive pedagogicalpath (Semet, Jamont, Biojout, Lutton, & Collet, 2003;Semet, Lutton, & Collet, 2003), and adaptive learning(Canales, Pena, Peredo, Sossa, & Gutierrez, 2007; Wang,Wang, & Huang, 2007). Establishing the learning path oflearner is certainly not a new approach as indicated above;but learner characteristics and the learning behaviors oflearners have resulted in the development of an adaptivesystem.

In Fig. 1, a scenario shows that a learner named Daniel

wants to get an SQL relative learning object. Which LO isthe best choice? To determine this, we must consider twocore factors: the learners’ attributes and the attributes ofthe learning object. Each learner has different attributes,such as learning style and domain knowledge level (e.g.apprentice, beginner, intermediate, expert); each LO hasits own attributes also, such as different types (e.g. text,video, etc.) and different levels (e.g. introductory, profes-sional). Kolb (1974), Felder and Silverman (1988) and Lar-kin-Hein and Bundy (2001) indicated that students learn inmany different ways. Some learn by seeing and hearing,others by feeling and doing; some focus on acting. Thus,learning styles should be considered as in order to developa dynamic adaptive learning environment (Kettel, Thom-son, & Greer, 2000; Robert, Lorne, & Maung, 1990). Inthe scenario, matching the learner’s requirements accord-ing his attributes, in order to assist him in finding the adap-tive learning objects efficiently is a significant problem. Inthe last decade, swarm intelligence, which can provide animpressive level of adaptability for learners in this sort ofdynamic learning environment has been reported to have

Fig. 1. Which one has the most a

applications in many fields. Ant colony system (ACS; seeDorigo & Gambardella, 1997) aim at exploring an optimallearning object and predicating several possible pedagogi-cal items. Employing an extended ant colony systemapproach based on Kolb’s learning style model, the attri-bute-based ant colony system (AACS) is proposed to con-struct a search mechanism for finding a suitable learningobject. AACS is different from many existing user modelsbecause it relies on the attributes of the learning objectsof each learner in order to find the optimal learning object.The remainder of this paper is organized as follows: Sec-tion 2: introduces Kolb’s learning style model and ant col-ony system and optimization algorithm. Section 3:proposes an attributes-based ant colony system (AACS)for an adaptive learning object search mechanism. Section4: makes experimental evaluations for parameters tune andheuristic information. Section 5: creates an application forAACS and a web portal for the adaptive learning objectstorage and retrieval. Section 6: conclusion.

2. Relative work

In the last decade, the idea of the learning object (LO),which can simplify and make reusable a large package oflearning resources has emerged. IEEE, the IMS GlobalLearning Consortium, the advanced distributed learning(ADL) co-laboratory, and others (MASIE Center, 2002)have produced learning objects standardized works. Cur-rently, the international specification, Sharable ContentsObjects References Model (SCORM) (ADL, 2004), thatis based on the results of work done by the above men-tioned groups, is widely used in the e-learning ecosystem.Thus, these related learning objects can be reorganized intoa course sequence (ADL, 2004; IMS, 2003). Accordingly,many web-based tutoring systems have developed sophisti-cated solutions for customizing learners’ learning needs to

daptive LOs for the learner?

Fig. 2. Kolb’s learning styles.

3036 Y.J. Yang, C. Wu / Expert Systems with Applications 36 (2009) 3034–3047

yield meaningful learning relationships. Course mainte-nance systems, adaptive courseware generations, anddynamic courseware generations (Brusilovsky & Vassileva,2003) are the core approaches developed in this field.

The afore-mentioned approaches are based on usinglearning activities and education items to predict optimalpaths. Thus, if this model can be associated with a sophis-ticated adaptation technique, such as ant colony optimiza-tion (ACO), it will provide an intelligent analysis of thesolution-well-trodden learning objects for learners.

2.1. Kolb’s learning style model

David Kolb published his learning styles model in 1984.The model gave rise to related terms such as Kolb’s expe-riential learning theory (ELT), and Kolb’s learning stylesinventory (LSI). Kolb’s learning theory sets out four dis-tinct learning styles (or preferences), which are based ona four-stage learning cycle.

As Fig. 2 shows, Kolb explains that different people nat-urally prefer a certain individual learning style. Here arebrief descriptions of the four Kolb learning styles.

1 SCORM (2004). Advance Distributed Learning Network Shareable

Content Object Reference Model, retrieved October 10, 2004 from http://www.adlnet.org/

2 LOM (2004).IEEE WG 12: Learning Object Metadata, retrievedOctober 10, 2004 from http://ltsc.ieee.org/wg12/index.html

2.1.1. Diverging (feeling and watching – CE/RO)

These people are able to look at things from differentperspectives. They are sensitive. They prefer to watchrather than do, tending to gather information and useimagination to solve problems. The learners who are cate-gorized as ‘‘Diverging” style learners prefer learningobjects in the form of animation, charts, graphs, flowcharts or symbols.

2.1.2. Assimilating (watching and thinking – AC/RO)

The Assimilating learning preference is for a concise, log-ical approach. Ideas and concepts are more important thanpeople. These people require good clear explanations ratherthan practical opportunity. The learners who are catego-rized as ‘‘Assimilating” style learner prefer learning objectsin the form of the audio, video, lectures, verbal tutorials, etc.

2.1.3. Converging (doing and thinking – AC/AE)

People with a Converging learning style can solve prob-lems and will use their learning to find solutions to practi-cal concerns. The learners who are categorized as‘‘Converging” style learners prefer learning objects in theform of text-based materials, such as Microsoft office(PowerPoint, Word, Excel), web pages, etc.

2.1.4. Accommodating (doing and feeling – CE/AE)

People with an Accommodating learning style will tendto rely on others for information rather than carry out theirown analysis. The learners who are categorized as ‘‘Accom-modating” style learner prefer learning objects in experi-ence shared and practice activities (for example, tutorialson the web, web pages, SCORM1, LOM2, etc.) rather thanin the form of animation, charts, graphs, flow charts orsymbols.

Table 1ACO algorithm according to chronological order of appearance

ACOalgorithm

TSP Main references

Ant system(AS)

Yes Dorigo (1992) and Dorigo, Maniezzo, & Colorni(1991a,b, 1996)

Elitist AS Yes Dorigo (1992) and Dorigo, Maniezzo, & Colorni(1991a,b, 1996)

ANT-Q Yes Gambardella & Dorigo (1995) and Dorigo &Gambardella (1996)

Ant colonysystem

Yes Dorigo & Gambardella (1997a,b)

Max–Min AS Yes Stutzle & Hoos (1996, 2000) and Stutzle (1999)Rank-based

ASYes Bullnheimer, Hartl, and Strauss (1997, 1999)

ANTS No Maniezzo (1999)Hyper-cube

ASNo Blum, Roil, & Dorigo (2001) and Blum & Dorigo

(2004)

In the column TSP we indicate whether this ACO algorithm has alreadybeen applied to the traveling salesman problem.

Y.J. Yang, C. Wu / Expert Systems with Applications 36 (2009) 3034–3047 3037

2.2. Ant colony optimization (ACO) algorithm for finding

adaptive Learning Objects

Ant colony optimization is a metaheuristic in which acolony of artificial ants cooperates in finding goodsolutions to difficult optimization problems. A metaheuisticis a set of algorithm concepts that can be used to defineheuristic methods applicable to a wide set of different prob-lems. The use of metaheuristics has significantly increasedthe possibility of finding high quality solutions to hard,practically relevant combinatorial optimization problemswithin a reasonable time. The traveling salesman problem(TSP) plays an import role in ACO research: the firstACO algorithm, call ant system, proposed by Dorigo,was successfully applied in tacking the well-known TSP(Dorigo & Gambardella, 1997), as well as many of theACO algorithms proposed subsequently, in shown in Table1 (Dorigo & Stutzle, 2004).

The ant colony system (ACS) is a particular algorithmof ACO which is based on agents that simulate the natu-ral behavior of ants, develop mechanisms for cooperation,and assist them in using experience (Dorigo & Gambard-ella, 1997) to find the shortest path between a food sourceand the nest. ACS is a population-based heuristics thatenables the exploration of the positive feedback whereasthe ants are able to communicate (ants lay pheromonefor indirect communication, so called stimergy3) informa-tion concerning food source via an aromatic essence. Theants lay pheromone and heuristic information to marktrails. As the paths are visited by other ants, some ofthe trails may be reinforced and others paths may beallowed to evaporate. Pheromone trails can be observedvia the number of ants passing through the trail. Whenthere are more pheromones on a path, there is larger

3 The original definition of stimergy (see Grasse, 1959; p. 79).

probability that other ants will use that path, and there-fore the pheromone trail on such a path will grow fasterand attract more ants to follow (so called positive feed-back). An iterative local search algorithm tries to searchthe current paths to neighboring paths until a better solu-tion is found. Informally, an ACO algorithm can be imag-ined as the following algorithm:

procedure ACO algorithm

Construct SolutionApply Local SearchUpdate Pheromone Trails

End

end

Here are brief descriptions of the ACO algorithm.

2.2.1. Construction Solution

The problem can be represented in a finite set of compo-nents C, set C ¼ fc1; c2; . . . ; cNcg, where Nc is the number ofthe components, and the states of the problem are definedin term of sequence x = hci,cj, . . ., ch, . . .i of finite lengthover the elements of C. The set of all possible states isdenoted by X. The set of solution S is subset of X (i.e.,S # X). Given this formulation, the ants build solutionsby performing randomized walks on the completely con-nected graph GC = (C,L) whose nodes are the componentsC, and the set L fully connects the components C. the GC iscalled the construction graph and the elements of L arecalled the connections. At each construction step, ant k

applies a probabilistic action choice rule, called random

proportional rule, to decide which node to visit next. In par-ticular, the transition probability with which ant k, cur-rently at node i, chooses to go to node j is

P kijðtÞ ¼

½sijðtÞ�a½gijðtÞ�bP

l2Nki

½silðtÞ�a½gilðtÞ�b ; 8j 2 Nk

i ; k 2 1 � m ð2:1Þ

where gij = 1/dij is a heuristic value that is available a pri-ori, a and b are two parameters which determine the rela-tive influence of the pheromone trails and the heuristicinformation, and Nk

i is the feasible neighborhood of ant kwhen located at node i and the set of nodes that ant k

has not visited yet (the probability of choosing a node out-side Nk

i is 0). By this probabilistic rule, the probability ofchoosing a particular arc (i, j) increases with the value ofthe associated pheromone trails sij and according to theheuristic information value gij. The role of parameters aand b determine the consideration of pheromone trailsand heuristic bias.

4 FORPA is an adaptive learning portal which is a real instance forAACS mechanism.

3038 Y.J. Yang, C. Wu / Expert Systems with Applications 36 (2009) 3034–3047

2.2.2. Update of Pheromone trails

After all the ants have constructed their tours, the pher-omone trails are updated. This is done by first lowering thepheromone value on all arcs by a constant factor, and thenadding pheromone on the arcs the ants have crossed intheir tours. Pheromone evaporation is implemented by

sij ð1� qÞsij; 8ði; jÞ 2 L; ð2:2Þwhere 0 < q 6 1 is the pheromone evaporation rate. Afterevaporation, all ants deposit pheromone on the arcs theyhave crossed in their tour:

sij sij þXm

k¼1

Dskij; 8ði; jÞ 2 L; ð2:3Þ

where Dskij is the amount of the pheromone ant k deposits

on the arcs it has visited. It is defined as follows:

Dskij ¼

1=Ck; if arcði; jÞ belongs to T k;

0; otherwise;

(ð2:4Þ

where Ck, the length of the tour Tk build by the kth ant, iscomputed as the sum of the lengths of the arcs belonging toTk.

2.2.3. Heuristic information

It is represented as a consciousness constant, which isthe heuristic preference of moving from nodei to the nextnode j. It is set to 1/dij, where dij is the distance betweennode i and node j.

2.2.4. Local searchPheromone evaporation and pheromone deposits take

place only on the arcs belonging to the best-so-far tour,when ant located at node i, ant k, moves to node j chosenaccording to the so-called pseudorandom proportional rule,given by q

j ¼ argmaxh2Nki

sih½gih�b

n o; if q 6 q0;

J otherwise;

(ð2:5Þ

where q is a random variable uniformly distributed in [0, 1],q0 (0 6 q0 6 1) is a parameter, and j is a random variableselected according to the probability distribution given byEq. (2.1) (with a = 1). In other words, with probability q0

the ant makes the best possible move as indicated by thelearned pheromone trails and heuristic information (theant is exploiting the learned knowledge), while with proba-bility (1 � q0) it performs a biased exploration of the arcs.Thus, the update in ACS is implemented by the followingequation:

sij ð1� qÞsij þ qDsbsij ; 8ði; jÞ 2 T bs ð2:6Þ

where Dsbsij ¼ 1=Cbs, which Cbs is the length of the tour Tbs.

It is important to note that in the pheromone trails update,both evaporation and new pheromone deposit, only appliesto the arcs of Tbs and not to all the arcs.

Many reasearchers have given ACO their attention todevelop sophisticated models and to extend its algorithm.

The Elitist Ant System (EAS) was introduced in Dorigo(1992) and Dorigo et al. (1991a, 1996). The idea is to pro-vide strong additional reinforcement to the arcs belongingto the best tour found since the start of the algorithm; notethat this additional feedback can be viewed as additionalpheromone deposited by an additional ant called best-so-far ant. A Rank-Based Ant System was proposed by Bulln-heimer et al. (1999). In Rank-Based Ant System, each antdeposits with its rank. Cordon, Herrera, Fernandez deViana, and Moreno (2000) used the transition rule andpheromone evaporation mechanism to improve the ants’solutions. Most recently, Semet, Lutton, et al. (2003) andSemet, Jamont, et al. (2003) applied the ACO heuristicsto an e-learning pedagogic material navigation problem,and Jamont, Collect, Lutton, Biojout, and BourgeoisRepublique (2005) experimented with an ‘‘ant-hill” methodwhich laid the pheromone depending on how students val-idated an item, so as to optimize learning paths with differ-ent students who have different views. Thus, the ACOmethod seems well suited for tackling adaptive learningobject searches in a dynamic learning environment.

3. Attribute-based ant colony system (AACS)

In this section, we present attribute-based ant colonysystem (AACS) as a method of finding a learning objectand then provide recommendations for the adaptive learn-ing object for learners. AACS is derived from an extensionof the ant colony system that updates the trails’ phero-mones from different knowledge levels and different stylesof a group’s learners to create a powerful and dynamiclearning object search mechanism. In order to achieve this,there are three prerequisites to applying AACS, that (a) theadaptive learning portal knows the learner’s attributes, andthe attributes include the learner’s knowledge level andlearning style (the learners have been referred to previouslyusing Kolb’s learning style). (b) The learner’s attributes(include Kolb’s learning style and learner’s knowledgelevel) and LO’s attributes (learning object type and learn-ing object level) have been annotated by teachers or con-tent providers. For example, learners who are categorizedas ‘‘Diverging” style learners prefer learning objects inthe form of audio or video (i.e., audio, video, animation),rather than Text-style learning objects. The AACS is fur-ther used as a prerequisite to match the relationshipsbetween learners and the learning object.

3.1. Learner’s activities and attributes recorded into theLearning Object Repository

In FORPA4 (Yang & Song, 2006), the learning portalcan gather the learning activities of each learner in an on-line course and then record their paths and attributes into

Table 2The learner’s attributes and LO’s attributes

Learner attributes Learning object attributes

Kolb’s learningstyle

Knowledgelevel

Learning objecttype

Learningobject level

1.Diverging 1.Apprentice 1.Graphic(image,charts, symbol),

1.Initial

2.Assimilating 2.Beginner 2.Video(audio,animation)

2.Introductory

3.Converging 3.Intermediate 3.Text(word,PowerPoint,Excel)

3.Advance

4.Accommodating 4.Expert 4.XML(Web,SCORM, LOM)

4.Professional

Y.J. Yang, C. Wu / Expert Systems with Applications 36 (2009) 3034–3047 3039

a learner transaction database (Learning Object Reposi-tory). Given the definition of a learning object relationship,let R = {r1, r2, . . ., rn}, R is indicated as a finite set of learn-ing objects relationships, which is expressed as the follow-ing traversal sequence of a learner. For example:R = {ra, rb, rc, rd}, (learning objects {a,b,c,d} to simplify),and we have paired these objects as transaction data (adatabase record) that is verified as a forward sequence of{(a,b), (a,c), (c,d), (d,a), (d,b), (b,c)}, where pair (a,b)denotes the relationship between the learning objects a

and b, and the learning object b is the next learning objectfrom a. Furthermore, the transaction data of the learnersand learning object relationship can be formally stored asa transition matrix in a learning object repository.

3.2. The Attribute ants and pheromone update rule

Since learning objects may be continuously added intoobject repositories by learning content providers at anytime, each learner has his own attributes, and the learningobject also has its own attributes, as shown in Table 2, wepropose the idea of an ‘‘attribute” ant to make it easy for alearner to get the most suitable learning object based on hisown attributes and LO’s attributes.

The ‘‘attributes” ant mainly uses an extended learnermodel and object level to involve the adaptive rule in adynamic learning setting. Following this rule, the systemcan improve the quality of the pheromone as well as helplearners easily find their own adaptive learning object. InAACS, we define two rules to reinforce pheromone updateas following:

Rule 1: IF learner’s attributes MATCH the LO’s attributes

THEN Daemon action with MRstyle and MRlevel.Rule 2: IF learner’s attributes PARTIALLY MATCH the

LO’s attributes THEN Daemon action with MRstyle

or MRlevel.

There is a formal definition of a ‘‘MATCH” and ‘‘PAR-TIALLY MATCH”:

1. IF learner learning style = learning object type ANDDomain Knowledge Level = learning object levelTHENwe

call it a ‘‘MATCH”.2. IF (learner learning style = learning object type AND

Domain Knowledge Level h i learning object level) OR(learner learning style h i learning object type ANDDomain Knowledge Level = learning object level) THENwe call it a ‘‘PARTIALLY MATCH”.

In AACS, whether or not the heuristic information canbe reinforced relies on the ‘‘attribute” ant. We give a formaldefinition of ‘‘attribute” ant: let k ant, which has one of theKolb’s learning styles and its own domain knowledge level,generate a fixed amount of pheromone when it travelsalong a node. If the k ant’s attributes ‘‘MATCH” or area ‘‘PARTIALLY MATCH” with the node attributes

(learning object attributes), then the nodes which are onthe traveled path created by an ant k may obtain an extrapheromone after each time unit. In order to further explainthe idea of the ‘‘attribute” ant, we illustrate using an exam-ple for explaining the adaptive rule. As show in Fig. 3,when a learner searches from the node a to its neighbornodes, the next node is possibly b, c and d. However,involving the adaptive rule, the learner with an ‘‘Assimilat-

ing” learning style and his domain knowledge level marked‘‘Beginner”, the node c with a ‘‘Video” type and object levelmark ‘‘Introductory”. Thus, the intensity of node c has ahigher probability of being recommended than node b

and d for the learner with ‘‘Assimilating” style and ‘‘Begin-

ner” level.

3.3. The proposed algorithm

We propose the attribute-based ant colony system(AACS) is an extension of the ant colony system, and theparameters and functions used in this paper are the sameas those defined in ACO. However, we applied two rulesfor pheromone update and heuristic information.

3.3.1. The heuristic information

The heuristic value gij is a normalized value function ofthe queue length qij (the learner waiting to be processed) onthe node connecting the node i with its neighbor j The heu-ristic information is defined as the following:

gij ¼ 1� ðqij �MRijÞXNi

l¼1

qil

, !ð3:1Þ

gij gives a quantitative measure associated with the nodewaiting time and match ratio (MR, given by Eq. (3.7)).An ant’s decisions are therefore taken on the basis of thecombination of long-term learning process and an instanta-neous heuristic prediction, and it is implied in the definitionas the required capability of a learner to achieve the nextlearning objective. In other words, a higher value of gij

means there is a higher probability of being able to choosethe high match ratio nodes, and to get the approximatelyoptimal adaptive LOs.

Fig. 3. A MATCH, PARTIALLY MATCH scenario.

procedure Attribute_ACS/* main procedure*/

Initialize_Parameterwhile (condition not terminate) do

Construct_SolutionHeuristic_Decision_RuleUpdate_Pheromone_Trailsif (Match the adaptive rule) then

Daemon_Actionsend while

Recommended_Learning_Objectend procedure

3040 Y.J. Yang, C. Wu / Expert Systems with Applications 36 (2009) 3034–3047

3.3.2. Pheromone trail update

The relational strength between the ith node and the jthnode is pheromone trail intensity sij. The incremental inten-sity Dsij(t) is that which locates pheromone trail value atthe time t, and is updated as the following formula:

sijðtÞ ¼ qsijðt � 1Þ þ DsijðtÞ ð3:3Þ

where q is the evaporate ratio of the trail at an interval timeunit. If an ‘‘attribute” ant has chosen the jth node afterlocating node i and laid its pheromone trail, the pheromonelevels on the jth node should be updated and the contribu-tions of all ‘‘attribute” ants, and the amounts of the pher-omone laid by the ants is defined as below,

sijðtÞ ¼ qsijðt � 1Þ þ w � Dsstyleij ðtÞ þ Dslevel

ij ðtÞ ð3:4Þ

where Dsstyleij ðtÞ and Dslevel

ij ðtÞ are the variable amounts ofpheromone deposited in the arc(i, j), and represent thoselearner who ‘‘MATCH” their attributes (attribute ants)with learning object attributes. Additionally, we give aweighted variable w to adjust the match ratio of the learn-ing style factor. Thus, the number of attributes ants has tobe considered; the adaptive solution is the following:

Dsk;styleij ðtÞ ¼

Xm

n¼0

ððm� nÞ � Q �MRstyleij ðtÞÞ ð3:5Þ

where m is the number of MRk;styleij ðtÞ on the arc(i, j)

Dsk;levelij ðtÞ ¼

Xm

n¼0

ððm� nÞ � Q �MRlevelij ðtÞÞ ð3:6Þ

where c is a constant number to adjust actual learning ob-ject search situations, m is the number of MRk;style

ij ðtÞ on thearc(i, j), MRstyle is the match ratio for learner learning style,and MRlevel is the match ratio for learning object level, thedefinition of match ratio given as follows:

LetX8j2Ni

MRij ¼X8j2Ni

MRstyleij þ

X8j2Ni

MRlevelij ð3:7Þ

where MR Styleij ¼ ðSi � T jÞ=ðP8r2NiðSi � T rÞ þ 1Þ

MR Levelij ¼ ðKi � LjÞ=ðX8r2Ni

ðKi � LrÞ þ 1Þ;

and

MRstyleij ¼ e�MR Styleij ; ð3:8Þ

MRlevelij ¼ e�MR Levelij ð3:9Þ

where S is the learner style, T is the learning object type, K

is the learner knowledge level, and L is the learning objectlevel.

3.3.3. Transition probability of attributes ants

The probability P kijðtÞ given in Eq. (2.1) is defined as a

compromise value of sij and gij where the k ant move for-ward from a learning object i to the next object j, j is oneof the neighbors of i, j 2Nk(i). The role of parameter aand b is the following: If a = 0, this corresponds to a classicstochastic greedy algorithm. If b = 0, only pheromoneamplification is at work, that is, only pheromone is used,without any heuristic bias. This generally leads to poorresults, in particular, for values of a > 1, it leads to therapid emergence of a stagnation situation, that is, a situa-tion in which all the ants follow the same path and con-struct the same tour, which, in general, is stronglysuboptimal (Dorigo, 1992: Dorigo et al., 1996).

3.3.4. The adaptive procedure for AACS algorithm

We proposed that the AACS algorithm has five mainprocedures for resolving the problem of adaptive paths,and pseudo-code of the attribute-based ant colony systemis shown as the following:

procedure Initial_Parameter

set the Q for trail constant intensity, heuristic fac-tor, a, b and evaporate rate q, and setting the pher-omone trails to a initial value s0> 0;Initializes learner attribute;set diverging = 1, Assimilating = 2, converging = 3,accommodating = 4;set apprentice = 1, beginner = 2, intermediate = 3,expert = 4;initializes learning object attribute;set Graphic = 1, Video = 2, Text = 3, XML = 4;set initial = 1, introductory = 2, advance = 3,professional = 4;end Initial_Paramete

procedure Construct_Solution

select a start node x1 and put into construct nodelist xL

while (xL R Sc and Nki –/Þ do

j Select_Next_Node (xL, R)

put j into xL;end while

if xL 2 S then return xL

else abort

end-if

end Construction_Solution

procedure Heruistic_Decision_Rule

let S = current_learner_learning_style forS = 1, . . .4,let T = current_learning_object_type forT = 1, . . .4,let K = current_learner_domain_knowl-edge_level for K = 1, . . .4,let L = current_learning_object_level forL = 1, . . .4,Computing MR Styleij ¼ ðSi � T jÞ=ð

P8r2NiðSi�

T rÞ þ 1ÞMR Levelij ¼ ðKi � LjÞ=ð

P8r2

N iðKi � LrÞ þ 1Þ,MRstyle

ij ¼ e�MR Styleij ,

MRlevelij ¼ e�MR Level

ij

LetP8j2Ni

MRij ¼ w �P8j2Ni

MRstyleij þP

8j2NiMRlevel

ij

/* where w is a parameter for enhance learningstyle weighted, w > 1 */foreach xij do

get candidate j from Nki

computing the heuristic information gij ¼ 1�ððqij �MRijÞ=

PNil¼1qilÞ

P kijðtÞ ¼

½sðij tÞ�a½gijðtÞ�bPl2Nk

i

½silðtÞ�a½gilðtÞ�b;

8j 2 Nki ; k 2 1 � m

end foreach

end Heuristic_Decision_Rule

Y.J. Yang, C. Wu / Expert Systems with Applications 36 (2009) 3034–3047 3041

procedure Update_Pheromone_Trails

set the adding pheromone with evaporate factorDsij (1 � q)*sij "(i, j) 2L where 0 < q 6 1,computing all ants deposit pheromonesij sij þ

Pmk¼1Dsij; 8ði; jÞ 2 L

end Update_Pheromone_Trails

procedure Daemon_Actions

foreach (attributes ant MATCH the adaptiverule)computing global pheromone update with attri-bute match ratio

Dsk;styleij ðtÞ ¼

Pmn¼0ðm� nÞ � Q �MRk;style

ij ðtÞDsk;level

ij ðtÞ ¼Pm

n¼0ðm� nÞ � Q �MRk;levelij ðtÞ

sijðtÞ ¼ qsijðt � 1Þ þ Dsstyleij ðtÞ þ Dslevel

ij ðtÞend foreach

end Daemon_Actions

procedure Recommended_Learning_Object

List the learning object from node list xL

Return adaptive learning object listend Recommended_Learning_Object

Below are brief descriptions of the main procedure func-tion in AACS algorithm pseudo-code

Initialize_Parameter: the procedure which initializesparameters of AACS, Q, a, b, q, learning style, domainknowledge level, object type and object level.

Construct_Solution: the procedure for construct solu-tion, where Sc is the set of candidate solutions, and Nk

i isset of the candidate nodes next to i for ant k, the randomwalk of ants is biased by pheromone trails which are asso-ciated with a connection between nodes i andj. The ant k

selects the next nodes j from neighbor list based on the

Fig. 4. The result of (a) higher attribute-ant match ratio efficiently causes daemon pheromone update. (b) Varying the number of ants used: the plot givethe match ratio values deviation from the optimal tour as a function of the number of iterations for using a number of ants varying from 10 ants to 200ants.

LearningResource

(LR)

Login/Register

Author

Learner

Get

Isa Isa Isa

LRObjective

Isa Isa

Easy

Expert

AdvanceIsa

Isa

ResourceFormat/Ty

pe

Isa Isa

Web Page

Document

Media

Package

(HTML/XML)

(WORD,PPT)

(VIDEO/AUDIO)

(SCORM,LMS)

Isa

Isa

Isa

Profile

Login/Questionnaire

Isa

Login/Answer

Record

FacilitateLearning

Style

DomainKnowledgeBackground

Isa

Assimilating

Diverging

Converging

Accommodating

Expert

Intermidate

Beginner

Publish Teacher Domain Expert

IsaIsa Isa

Isa

Kolb Learning Style

Learner Attribute

Author Attribute

ResourceLevel

Easy Intermeate Advance

ExpertIsa

Isa

IsaIsa

Isa

Resource AttributeIsa

Isa

Isa

LRsOntology

Repository

Store

LRsFuzzy-Ontology

DatabaseJustified

Store

Ontology Search

FuzzyRule

Module

OntologySearchAgent

Ontology Fuzzy abase/Rule&Search

E-Education Portal

Isa

Store/Reference

References References

Call Find

References

References

LearningObjects

Repository A portal using AACStechnique

Fig. 5. An adaptive portal (FORPA) architecture using AACS technique.

3042 Y.J. Yang, C. Wu / Expert Systems with Applications 36 (2009) 3034–3047

Fig. 6. A example for Oracle Database Courseware applied AACS system.

5 For interpretation of color in Fig. 4, the reader is referred to the webversion of this article.

Y.J. Yang, C. Wu / Expert Systems with Applications 36 (2009) 3034–3047 3043

Select_Next_Node function, which mainly focused on thetransition probability value P k

ij, as shown in Eq. (2.1), afterthat, choose the next nodej from candidate nodes and putinto optimal solution node list xL.

Heruistic_Decision_Rule: the procedure which computesthe heuristic information gij and match ratio, where MRstyle

is the match ratio for learner learning style, and MRlevel isthe match ratio for learning object level.

Update_Pheromone_Trails: this procedure computes thegeneral pheromone update sij after evaporation, where Dsij

is the amount of pheromone ant k deposits on the arcs ithas visited.

Daemon_Actions: the procedure which computes theglobal pheromone update value Dsij, where Dsstyle

ij ðtÞ andDslevel

ij ðtÞ are the variable amounts of pheromone depositedin the arc(i, j), and represent those learners who ‘‘MATCH”

their attributes (attribute ants) with learning objectattributes.

Recommended_Learning_Object: the procedure whichgets the adaptive learning object from solution nodes listxL, and then returns to the system.

4. Experimental evaluations of trails’ intensities

This section describes the simulation of tuning parame-ters in AACS. Pilot experiments were conducted to opti-mize the parameters in order to match the actualsituation. We observe the variation of the intensity oftwo trails. As shown in Fig. 4, the blue5 curve which utilizes

Fig. 7. FOAPA portal gets the adaptive learning object according to the learner’s attributes.

3044 Y.J. Yang, C. Wu / Expert Systems with Applications 36 (2009) 3034–3047

traditional pheromone rule, does not consider affects fromlearning style, and adds the population to take variablesof trail intensity s into account in each time t. Consideringthe solid green curve as shown in Fig. 4 that uses our pro-posed rule (in Section 3.3.2, Eqs. (3.3)–(3.9)) and consid-ering the learning style of the learner, we dynamicallyadd to the population in which the learning object hasthe same style each time. It was found that this curvehas exponential growth. That is, if a node (as a learningobject) has more learners with the same style travelingalong a path with AACS operating, the higher the matchratio (MB) will become, and the node is possibly a candi-date for the best choice. In Fig. 4, the experiment showsthat higher MB value caused the attribute-ant to get theadaptive learning object efficiently, and the proposed ruleto update the trail intensity has a significant enoughamount of pheromone to affect the probability distribu-tion along the paths as opposed to the traditional phero-mone update rule.

5. Application: a web portal for adaptive learning path

recommendation

The architecture of a web-base adaptive learning portalwas shown in Fig. 5, The modules of the FOAPA aredescribed below.

5.1. Login/Questionnaire

Upon entering the Portal for the first time, each learner/author is prompted with a short questionnaire, in order to

determine his or her characteristics, learning style, anddomain knowledge level. This profile is automaticallyupdated, taking into account the learner interactions withthe Portal.

5.2. Learners

The learners need to be able to authenticate themselvesvia login module, as well as to define, edit and save theirprofile for adaptive personalized services. Then, learnerscan, via the AACS agent and match rule module, retrieveLOs according to their specific requirements, learningstyles and domain knowledge levels.

5.3. Authors

The authors need to be able to publish or upload theirLOs in a commonly accessible format, so that they canbe effectively, efficiently and adaptively searched andretrieved in the different contexts of use described in LO’s

attribute. Different LOs are described in metadata whichare related to each resource and presented according to dif-ferent learner profiles, thus meeting the requirements forlearners-adaptive personalized learning.

5.4. AACS agent

The AACS Agent is responsible for compiling the queryresult from the learning objects repository, ranking themaccording to user preferences and match ratio, as suppliedby the match rule module. In particular, all the learning

Fig. 8. A screenshot for learner get the recommendation learning object.

Y.J. Yang, C. Wu / Expert Systems with Applications 36 (2009) 3034–3047 3045

objects were using Dublin6 (2004) metadata tags the learn-ing object, and it a standard ontology object presentationform.

5.5. Match rule module

The match rule module is responsible for defining criteriafor searching and ranking learning objects. Through therules which are defined in Section 3.3.2, the attributes-antcan access the adaptive learning object via the AACS algo-rithm quickly and efficiently.

5.6. Ontology object repository

Before storing into the ontology object repository, allthe learning objects must be tagged with Dublin (2004)metadata tags, details about which have been publishedas ANSI/NISO Standard Z39.85-2007.

5.7. Learning objects database

We create table schema for storage of the learner’s attri-butes and LO’s attribute which can be accessed via theAACS agent and match rule module. The AACS agent usesattribute-ant behavior to get the adaptive learning object

6 DCMI (2007). Dublin Core Metadata Initiative, retrieved July 2, 2007from http://dublincore.org.

via the match rule, and then returns the optimal resultfor the learner.

We provide the ORACLE OCP (oracle certificated pro-fessional) courseware to be a real instance for AACS sys-tem, the example was shown in Fig. 6. In this example,shows that a learner how to get the adaptive learning objectusing the AACS mechanism, the higher MB value will leadto the high probability to get the most suitable learningobject.

The FORPA system screenshot was shown in Figs. 7and 8.

6. Conclusion

This work has proposed an attribute-based ant colonysystem (AACS) and introduced an instance of an adaptivelearning portal, named FORPA. It is now available athttp://192.192.111.113. AACS offer efficient heuristic infor-mation and match rules for enhancing the discovery of anadaptive learning object. It also provides an idea, namedattributes-ant, which combined the Kolb’s learning stylemodel and the learner’s domain knowledge level with learn-ing objects attributes to provide an adaptive solution forlearners. The FORPA, a web-based learning portal withthe AACS application, has been developed by the authorsto support the adaptive learning device. The key conceptfor developing an adaptive learning system comes primarilyfrom recognizing the different attributes of different learn-ers. From the computer-supported learning perspective,there are two critical points: how a learning system launches

3046 Y.J. Yang, C. Wu / Expert Systems with Applications 36 (2009) 3034–3047

an adaptive learning object and how different attributes oflearning objects can conform to learner’s attributes and asa consequence, lead learners down a well-trodden path. Inaddition, the attributes-ant colony system proposed in thisstudy has been demonstrated as a successful solution foradaptive learning object search mechanisms in an e-learningdomain. The findings suggest that different object attributesand learner’s attributes have enhanced the opportunity foradaptive learning that leads to improved learning outcomes.

In the future, we would like to apply the AACS to var-ious domain applications, such as supply-chain systems,intelligent education systems, visitor tour paths, job sche-dule systems, and so on.

References

ADL (2004). Advance distributed learning initiative. Shareable contentobject reference model (SCORM) 204 (3rd ed.). Overview. ADL WebSite <http://www.adlnet.gov/scorm> Accessed 18.06.07.

Berghel, H. (1997). Cyberspace 2000: Dealing with information overload.Communications of the ACM, 40(2).

Blum, C., & Dorigo, M. (2004). The hyper-cube framework for ant colonyoptimization. IEEE Transactions on Systems, Man, and Cybernetics-

Part B, to appear.Blum, C., Roil, A., & Dorigo, M. (2001). HC-ACO. The hyper-cube

framework for ant colony optimization. In Proceedings of MIC’2001—

Metaheuristics International Conference (vol. 2 pp. 399–403).Brusilovsky, P. (2001). Adaptive hypermedia. In Alfread Kobas (Ed.),

User modeling and user adapted interaction. Ten year anniversary issue

(Vol. 11, pp. 87–110).Brusilovsky, P., & Vassileva, J. (2003). Course sequencing techniques for

large-scale web-based education. International Journal of Continental

Engineering Education and Lifelong Learning, 13(1–2), 75–94.Bullnheimer, B., Hartl, R. F., & Strauss, C. (1997). A new rank based

version of the ant system – A computational study. Technical Report,University of Viena, Institute of Management Science.

Bullnheimer, B., Hart1, R. F., & Strauss, C. (1999). A new rank-nasedversion of the ant system: A computational study. Central European

Journal for Operations Research and Economics, 7(1), 25–38.Canales, A., Pena, A., Peredo, R., Sossa, H., & Gutierrez, A. (2007).

Adaptive and intelligent web based education system: Towards anintegral architecture and framework. Expert Systems with Applications,

33, 1076–1089.Chen, J. N., Huang, Y. M., & Chu, W. C. (2005). Applying dynamic

fuzzy Petri net to web learning system. Interactive Learning, 13(3),159–178.

Cordon, O., Herrera, F., Fernandez de Viana, I., & Moreno, L. (2000). Anew ACO model integrating evolutionary computation concepts: Thebest-worst ant system. In Proceedings of the 2nd international workshop

on ant algorithms, Brussels, Belgium (pp. 22–29).Dorigo, M. (1992). Optimization, learning and natural algorithms. PhD

thesis, Dipartimento di Elettronica, Politecnico di Milano, Milan.Dorigo, M., & Gambardella, L. M. (1996). A study of some properties of

Ant-Q. In H. Voigt, W. Ebeling, I. Rechenberg, & H. Schwefel (Eds.),Proceeding of PPSN-IV, Forth International Conference on Parallel

Problem Solving from Nature. Lecture Notes in Computer Science (Vol.1141, pp. 656–665). Berlin: Springer-Verlag.

Dorigo, M., & Gambardella, L. M. (1997). Ant colony system: Acooperative learning approach to the traveling salesman problem.IEEE Transaction on Evolutionary Computation, 1(1), 53–66.

Dorigo, M., & Gambardella, L. M. (1997a). Ant colonies for the travelingsalesman problem. BioSystems, 43(2), 73–81.

Dorigo, M., & Gambardella, L. M. (1997b). Ant colony system: Acooperative learning approach to the traveling salesman problem.IEEE Transactions on Evolutionary Computation, 1(1), 53–66.

Dorigo, M., Maniezzo, V., & Colorni, A. (1991a). Positive feedback as asearch strategy. Technical report 91-016 revised, Dipartmento diElettonica, Politecnico di Milano, Milan.

Dorigo, M., Maniezzo, V., & Colorni, A. (1991b). The Ant System: Anautocatalytic optimizing process. Technical report 91-016 revised,Dipartmento di Elettonica, Politecnico di Milano, Milan.

Dorigo, M., Maniezzo, V., & Colorni, A. (1996). Ant System: Optimiza-tion by a colony of cooperating agents. IEEE Transactions on System,

Man, and Cybernetics - Part B, 26(1), 29–41.Dorigo, M., & Stutzle, T. (2004). Ant colony optimization. Bradford Books

(Hardcover-Jul 1, 2004).Felder, R. M., & Silverman, L. K. (1988). Learning and teaching styles in

engineering education. Engineering Education, 78(7), 674–681.Gambardella, L. M., & Dorigo, M. (1995). Ant-Q: A reinforcement

learning approach to the traveling salesman problem. In A. Prieditis &S. Russell (Eds.), Proceedings of the Twelfth international Conference

on Machine Learning (ML-95) (pp. 252–260). Palo Alto, CA: MorganKaufmann.

Grasse, P. P. (1959). La reconstruction du nid er les coordinationsinterindividulles chez Bellicositermesnatalensis er Cubitermes sp. Latheorie de la stigmergie: Essai d’interpretation du comportement destermites constructeurs. Insectes Sociaux, 6.

Huang, Y. M., Chen, J. N., Huang, T. C., Jeng, Y. L., & Kuo, Y. H.(2007). Standardized course generation process using dynamic fuzzyPetri nets. Expert Systems with Applications, 34(1).

Huang, Y. M., Chen, J. N., Kuo, Y. H., & Jeng, Y. L. (2007). Anintelligent human-expert forum system based on fuzzy informationretrieval technique. Expert Systems with Applications, 34(2).

IMS Global Learning Consortium, Inc. (2003). IMS learning designbest practice and implementation guide. Version 1.0 final specifica-tion. <http://www.imsglobal.org/learningdesign/ldvlp0/imsld_bestvlp0.html> Accessed 10.05.06.

Jamont, Y., Valigiani G., Collect, P., Lutton, E., Biojout, R., & BourgeoisRepublique, C. (2005). Experimenting with a real-size man-hill tooptimize pedagogical paths. In SAC’05 proceedings (pp. 13–17). Mars,Santa Fe, NM, USA.

Kettel, L., Thomson, J., & Greer, J., (2000). Generating indivisualizedhypermedia applications. In P. Brusilovski (Ed.), Proceddings of ITS-

2000 workshop on adaptive and intelligent web-based education systems

(pp. 28–36), Montreal.Kolb, D. A. (1974). Learning style inventory technical. Boston McBer &

Co.Larkin-Hein, T., & Bundy, D. D. (2001). Research on learning style:

Applications in the physics and engineering classroom. IEEE Trans-

action on Education, 44(3), 276–281.Maniezzo, V. (1999). Exact and approximate nondeterministic tree-search

procedures fir the quadratic assignment problem. INFORMS Journal

on Computing, 11(4), 358–369.Manouselis, N., & Sampson, D. (2003). Agent-based e-learning course

recommendation: Matching learner characteristics with content attri-butes. International Journal of Computers and Applications (IJCA),

25(1).MASIE Center (2002). Making sense of learning specifications &

standards: A decision marker’s guide to their adoption. IndustryReport. <http://www.masie.com/standards/S3_Guide.pdf> Accessed18.06.07.

Melis, E., Andres, E., Budenberder, J., Frishauf, A., Goguadse, G.,Libbrecht, P., et al. (2001). ActiveMath: A generic and adaptive web-based learning environment. International Journal of Artificial Intelli-

gence in Education, 12(4), 385–407.Robert, P. B., Lorne, O., & Maung, K. S. (1990). The importance of

learning style in end-user training. MIS Quarterly, 14(1), 101–119.Semet, Y., Jamont, Y., Biojout, R., Lutton, E., & Collet, P. (2003).

Artificial ant colony and e-learning: An optimization of pedagogicalpaths. HCI.

Semet, Y., Lutton, E., & Collet, P. (2003). Ant colony optimization for e-learning: Observing the emergence of the pedagogic suggestions. IEEE

Swarm Intelligence Symposium, April (pp. 46–52).

Y.J. Yang, C. Wu / Expert Systems with Applications 36 (2009) 3034–3047 3047

Stewart, C., Cristea, A., Brailsford, T., & Ashman, H. (2005). Authoringonce, delivering many: Creating reusable adaptive courseware. In Web-

based education conference WBE’05, Grindelwald, Switzerland.Stutzle, T. (1999). Local search algorithms for the combinatorial prob-

lems: Analysis, improvements, and new applications. DISKI (vol. 220).Sankt Augustin, Germany, Infix.

Stutzle, T., & Hoos, H. H. (1996). Improving the ant system: A detailedreport on the MAX -MIN Ant System. Technical report AIDA-96-12,FG Intellektik, FB Information, TU Darmstadt, Germany.

Stutzle, T., & Hoos, H. H. (2000). MAX–MIN Ant System. Future

Generation Computer System, 16(8), 889–914.

Wang, T.-I., Wang, K.-T., & Huang, Y.-M. (2007). Using a style-basedant colony system for adaptive learning. Expert Systems with Appli-

cations. Available online.Weber, G. (1999). Adaptive learning systems in the world wide web. In: J.

Kay (Ed.), Proceedings of the international conference on user modeling,

UM99, Banff, Canada, June 20–24, 1999 (pp. 371–378). Wien,NewYork: Springer.

Yang, Y. J., & Song, K. (2006). FOAPA: A fuzzy ontology e-educationadaptive portal architecture for e-learner. In ICS 2006 proceedings of

the international computer symposium 2006 (pp. 949–957).