Student modeling using intelligent agents in a web-based ... - CiteSeerX

Student modeling using intelligent agents in a web-based teaching and learning environment Clara-Inés Peña*

Jose-L Marzo Josep-Lluis de la Rosa University of Girona - Spain {clarenes,marzo,peplluis}@eia.udg.es

Abstract In this paper, we present an approach for student modeling by means of intelligent agents at USD (“Unitats de Suport a la Docència”["Teaching Support Units"]), a tutoring system that has been designed and used to support distance learning on the web. Considering that the aim of pedagogical design of intelligent tutors is to engage students in sustained reasoning activity and to interact with them based on a deep understanding of their behavior, we want to create an intelligent student agent that by means of CaseBased Reasoning techniques be capable of dynamically building a representation of the student’s interests and characteristics (subjective likes) taking into consideration learning styles. At this point, we will examine opportunities to improve the teaching and to motivate the students to learn what they want, in a userfriendly environment that suits their learning style. The system perceives the first student profile approximation by evaluating the diagnostic questionnaire proposed by the learning style model applied. Subsequently, this profile should be fine-tuned through the student interaction at the USD learning environment and applying CBR techniques. Therefore, the USD system will adapt the learning scenarios (teaching contents, navigation tools and navigation strategies) according to the type of the student that uses it. Keywords. Student modeling, CBR, intelligent agents, learning styles, intelligent tutors, adaptive instructional hypermedia.

1. Introduction The Internet provides an infrastructure that supports unprecedented communication capabilities and opportunities for collaboration. In the field of education, the Internet allows collaboration between various domain experts and teachers in designing novel approaches to teaching and co-operation among teachers who can share instructional material. It offers a vast *

store of information that can be accessed in a structured manner or explored in an unstructured manner, providing opportunities for designing tutoring systems with diverse pedagogic strategies. The freedom and flexibility offered by the Internet can, however, turn into an huge waste of time, effort and resources, if the nature of educational processes and the capabilities of educational technologies are not adequately considered while designing a tutoring system. The aim of this paper is to show the use of intelligent agents to model the student in the USD teaching and learning environment [7] applying Case-Based Reasoning techniques and taking into consideration student learning styles. Continuous assistance for students during learning is also addressed. The current USD system is an adaptable platform that allows teachers to create and manage teaching units for sequential or free navigation. A teaching unit consists of a set of HTML pages that contains the didactic material. The navigation structure is given by means of a directed graph set up by the teacher following his/her curricular criteria. Students may customize their learning environment by selecting the working language and the size, shape and position of the icons. The navigation environment provides various tools: to navigate the learning contents in a guided or free way (using forward or backward buttons, a contents tree, or a glossary of terms); to make self assessed, interactive exercises; to communicate with student and teacher communities (e-mail, chat, BSCW-forum); to make bookmarks; to print out learning contents; to check student progress; etc. In this system, the student modeling is used to offer teaching units adapted to the student’s learning style while considering the media format of the didactic content, the available interactivity elements, the presentation style for the information, the best subject content for a particular situation, the most suitable navigation technique and the best navigation tool for comfortable surfing among subjects. Knowing that the cognitive science offers important value to the design of intelligent tutors, we realized a

She is a Computer Science researcher at the Universidad Industrial de Santander Colombia

deep analysis of some learning styles models to choose the best to apply to web-delivered courseware for Engineering and Computer Sciences education as our USD system it is, and found the Felder and Silverman Learning Style Model (FSLSM) [6] that has been tested widely in similar scenarios [3] to categorize students according to their skills in processing, perceiving, entering, organizing and understanding the learning information. For the initial student classification, we applied the Index for Learning Styles (ILS), a diagnostic instrument of the FSLSM model that assesses preferences on four dimensions (active/reflective, sensing/intuitive, visual/verbal, and sequential/global). The system offers each student a personal agent (user agent) that will know his/her likes and preferences (subjectivity) and will be able to recommend automatically, suitable teaching units, navigation techniques and navigation tools in USD teaching/learning environment. This approach will allow the system to be more competitive because working with subjectivity will permit it to acquire knowledge that can be transported and used profitably by any teaching unit or didactic content offered. Thus, the system will offer educational solutions in terms of each requirement. By means of the student’s interaction with the learning environment, the system will create and store a student profile with his/her preferences (subjective likes), asking its agents and orientating the teaching contents according to the preferences of the current student. Each student’s agent will proactively increase the knowledge that it has about his/her student by applying his own strategies. It will be able to orientate the motivation information to the student’s needs and to improve the system performance by using different motivation techniques, collaborative filtering and selfcriticism. At current, we had applied the ILS questionnaire to a group of seventy students and we are realizing simulations with the HabitatPro environment [4], a tool for contents personalization and market prospecting using Case-Based Reasoning and Fuzzy Logic AI techniques, to generate the recommended information for building and maintaining the USD student model.

2. Student modeling using Case-Based Reasoning Techniques 2.1 Introduction The main idea of CBR is to solve a new problem by retrieving a previous similar situation and by reusing information and knowledge of that situation [1]. Finding a similar past case and reusing its solution in the new

problem situation solves a new problem. In CBR terminology, a case usually denotes a problem situation. A previously experienced situation, which has been captured and learned in a way that can be reused in the solving of future problems, is referred to as a past case, previous case, store case, or retained case. Correspondingly, a new case or unsolved case is the description of a new problem to be solved. Case-Based Reasoning is – in effect – a cyclic and integrated process of solving a problem, learning from this experience, solving a new problem, and so on. If two problems resemble each other, their solutions are similar and therefore it is possible to apply an adaptation of the former solution to the current problem.

2.2 Subjectivity comprehension: images and opposite images In order to get an agent to act correctly in the best interests of the individual that it represents, it is indispensable that it incorporates knowledge somehow on his/her likes, preferences and personality. This knowledge is, in general, extremely diffuse and contradictory (subjective) and therefore difficult to represent and manage. Our user agent obtains this subjective knowledge by means of images and opposite images of the students that it represents. Student modeling in the USD system is based on three concepts: •

•

•

The electronic environment where the teaching and learning activity is carried out and where the aim of personalization in the interaction with the student is identified clearly. The intelligent agents satisfy this necessity by living, learning and interacting among them and with the exterior, within this electronic environment. The products or the teaching contents, navigation techniques and navigation tools offered by the environment. Different types of students will have distinct preferences for the educational units on offer and in consequence, the requirement for a personalized relationship between the environment and the student grows. The student or the person that is interested in the teaching units. This is the entity that needs personalized treatment, which can be acquired using intelligent agents that model the student’s behavior and act as an intermediary between them and the learning environment.

2.2.1 Characterization of the subjective particularities of the educational units by means of pairs attribute-value. The agent’s personalization

techniques involved in the student modeling center on the concept of pairs attribute-value widely used in Artificial Intelligence for knowledge representation. The attribute applicable to a teaching unit or to a student is equivalent to a property or a characteristic. We considered for didactic contents, some of the following possible attributes: the media format for presentation (i.e. graphic, text, hypertext, audio, etc), the type of interactivity that it offers (i.e. sensitive maps, simulations, exercises), the type of didactic contents used to explain the situations (i.e. objectives, resumes, examples, synthesis, lectures), etc. Each attribute can have a value among a group of possible values. If we define the group of products, we are identifying the products to which certain sets of attributes are applicable. In the USD system we just considered a global group product named Teaching Unit with the following attribute-value structure for each dimension of the FSLSM learning style model: GROUP: TEACHING UNIT Attributes: instructional strategies, instructional complementary materials, interactive and assessment elements, media format and navigation tools. Attribute-values: Instructional strategies: lesson objectives, case studies, lectures, nucleus of knowledge, conceptual maps and synthesis. Instructional complementary materials: examples, animations. Interactive elements: simulations, interactive graphics, and glossaries. Assessment elements: self-assessment exercises, and open-answer exercises. Media formats: hypertext, text slide-shows or multimedia slide-shows, graphic media clips, digital movies, audio media clips or lineal texts. Navigation tools: punctuals, structurals, and for collaborative work. The values that the attributes can take are in general of a subjective nature, because the meaning of each one depends on the person that uses or defines it (an student could learn better using a teaching unit in a graphic format than one in a text format) and therefore sophisticated techniques are needed for their manipulation. In this case, adding values of this type does not make sense. The algorithms used for this purpose in the system are based on AI techniques like Case-Based Reasoning and fuzzy logic. We define the image of a product as a set of pairs attribute-values that characterize it. For instance, the image of the Introduction to Computers Teaching Unit for an active student (according to directions from the

FSLSM model) could contain the following attributevalue pairs: Instructional strategy: nucleus of knowledge. Instructional complementary materials: examples. Interactive elements: simulations. Assessment elements: self-assessment exercises or open-answer exercises. Media formats: lineal texts. Navigation tools: punctuals, structurals and for collaborative work. While for a sensitive student it might be: Instructional strategy: case studies or conceptual maps Instructional complementary materials: none Interactive elements: simulations or interactive graphics. Assessment elements: open-answer exercises. Media formats: hypertext, multimedia slide-shows, graphic media clips, digital movies, audio media clips or lineal texts. Navigation tools: punctuals, structurals and for collaborative work. Starting with the concept of product images, it is possible to define the new concept of distance between products using fuzzy logic techniques. This distance allows you to obtain, from the images of two products, a numerical value that represents the degree of similarity existing between the products. This distance is a function: dp : P x P Æ R Where P is the set of images of the products and R the set of real numbers. 2.2.2 Characterization of the subjective particularities of the students by means of the triple elements attribute-value-weight. A set of pairs attribute-value related to a student could reflect his/her preferences with respect to the teaching units and the learning environment. For instance, an active student (categorized by Felder) could be characterized for a specific teaching unit with the following set of pairs: {Instructional strategy/ nucleus of knowledge, Instructional complementary materials/examples, Interactive elements/simulations, Assessment elements/open-answer exercises, Media formats/hypertext, lineal texts, Navigation tools/ backward and forward arrows, forum}

The representation of the student’s preferences by means of pair’s attribute-value is not efficient because it does not collect the intensity of the preferences or the importance that the student gives to each of the attributes. To solve this problem, we introduce the concept of an associated weight to an attribute and to a student. Each student will give to each attribute his/her own weighting, which will indicate the importance the attribute has for him/her when assigning a degree of preference to a product. The set of weights that can be associated to the attributes is configurable. In this system we used the following weights: W= {Indifferent, Less Important, Medium Important, Important, Very Important and Necessary} Corresponding to the values 0, 1, 2, 4, 8 and 1000 respectively. At this point we can observe that W will always contain two special constants: Indifferent with value zero and Necessary with a very big value. From this point, we define the image of the student for a group, as a set of the triple-elements: attribute/value/weight that encodes the student’s preferences and the importance given to the values of the attributes for the products of the group. By introducing the weight parameter, we can use the following examples to show the characteristics of the Teaching Units preferred by a student with an active learning style and by a student with a reflexive learning style. Teaching Unit for an active student: {Instructional strategy/ nucleus of knowledge/very important Instructional complementary materials/examples/important Interactive elements/simulations/important Assessment elements/open-answer exercises/important Media formats/lineal texts/very important Navigation tools/ backward and forward arrows, forum/necessary} An active learner tries to acquire the knowledge by doing, he/she likes to work in groups (i.e. the forum tool could be necessary) and is comfortable navigating the contents by means of the direct guidance navigation technique, whereby the backward and forward navigation arrows are also necessaries. Teaching Unit for a reflexive student

{Instructional strategy/nucleus of knowledge/indifferent Instructional complementary materials/examples/very important Interactive elements/interactive graphics/very important Assessment elements/self-assessment exercises/important Media formats/multimedia slide-shows/very important Navigation tools/backward and forward arrows, email/very important} A reflexive learner processes the information introspectively, acquires the knowledge better by means of graphical contents, thinks a lot before acting and prefers to work alone or in pairs. By extending the concept of distance between products introduced above, and generalizing it to include the weights of the attributes we can define two new concepts: 1.

The distance between a student and a Teaching Unit We define d: P x C Æ R

Where P is the set of teaching Unit images and C the set of student images. Given an image c of a student and an image p of a Teaching Unit, the function d (p, c) considers simultaneously all the attributes used in the two images, along with their values and weights to return a numerical value with the distance between the student and the product. Therefore function d takes the subjective information represented in the student and teaching unit images to obtain a concrete numerical measure of the affinity between them. 2.

The distance between two students

We define dc : C x C Æ R Where C is the student images set. Given two student images, c and d, the function d (c , d) considers simultaneously all the attributes used in the two images along with their values and their weights to return a numerical value that represents the distance between the two students. In a similar way to the function d (p , c), the dc function takes the subjective information represented in the student images to obtain a concrete numerical measure of the affinity between them. The applications of these two functions (1 and 2 above) are immediate. With them, it will be possible to tell a student about the didactic materials that will work best for him/her, to put students in groups according to related preferences (clustering) or to use the information studied by a student in a teaching unit to promote the

same didactic materials for other students of similar learning styles (collaborative filtering). Another type of application of great importance for these functions is the analysis and the prospecting of new teaching units. 2.2.3. Upgrading student images. As we said previously, an essential characteristic of intelligent agents consists of the capacity to learn from its interactions with other agents and with the environment. In our case, in which the agents incorporate knowledge about the preferences and the personality of the students, the learning process will consist of continuous tuning of the images of the students so that gradually they reflect their likeness more faithfully. In this way, each time the student follows the Teaching Unit his/her image will be updated and adjusted to a new situation. Not only will the system maintain something like an average value for each one of the attributes, but it will also upgrade automatically the weights, so that the attributes for which the student always chooses the same values (or nearly the same) will have higher weights, while the attributes for which the student shows no marked preference for a particular value or range of values will have a lower weight. The magnitude of the image upgrade will be controlled depending on different factors: a)

The demonstrated student interest: If a student largely carries out the learning activities proposed in a teaching unit that has some particular attributes, the system should suppose that he/she likes this unit and therefore the magnitude of his/her image upgrade for that specific unit, should be bigger than for another teaching unit that presents the same learning contents but with different attributes and which the student has never entered or if he/she has entered, has only carried out a minimum of the learning activities proposed there. b) The quality of the student interaction: the magnitude of the student image upgrade will be related to the quality or quantity of interactions that he/she has had within the learning environment. c) The type of teaching unit: the system can more faithfully upgrade the image of a student who chooses a teaching unit rich in contents and diverse learning activities. d) The time: it has to be remembered that people's preferences change with time. Therefore the upgrade of the student image should be more representative if a moderate time has passed since the last upgrade. e) The student’s preferences: the system should be sufficiently open to allow the student to change his/her image according to his/her own preferences.

2.2.4. Images and opposite-images. Previously it was seen how it is possible to upgrade the images of the students to reflect the fact that they show interest in a specific type of teaching unit, carrying out a high percentage of the proposed learning activities there, either visiting a certain number of pages, carrying out exercises of X or Y degrees of difficulty, participating in the programmed chats or contributing and analyzing information through the discussion forums, or using certain navigation tools. All these types of actions have something in common: they provide positive information, that is to say, information on what the student likes. It has already been observed how the system can take advantage of this type of information to learn from the student's likes to upgrade his/her image. Just as important as knowing what they like in order to provide them with suitable teaching materials, is to know what they dislike, because it is, for example, boring him/her with information that is of no interest or not the kind of didactic that suits them. There are different ways to get this kind of information: • •

Identifying the types of teaching units offered to the student repeatedly without arousing his/her interest. Asking the students directly

We come then, to the problem of how to use the negative information provided by the student. This negative information is always of the same type: the student rejects a teaching unit presented in a particular style with more or less intensity. One way to manage this type of information consists of some kind of negative adjustment of the student image, defining for example his/her opposite-image. The opposite-images are also groups of triple-elements attribute/value/weight associated to each student. Therefore, the oppositeimage of a student c is the image of an imaginary student c' with preferences that are totally opposed to those of c. The way to manage the negative information is now clear: every time that a student rejects a teaching unit with a type of specific learning content, the student's opposite-image will be upgraded, just as if he/she had shown interest in it. The magnitude of the upgrade of the opposite-image will be proportional to the magnitude of the rejection. In this way, positive information will upgrade the image and negative information will upgrade the opposite-image. Consequently, using the function of distances between students and teaching units, not only can the teaching units be ranked from the one that is most adjusted to the student’s learning style and which the student feels comfortable working with, to the one that is least adjusted to the student’s learning style (using the

image), but also from the one that the student least dislikes to the one that he/she most dislikes (using the opposite-image). Although the introduction of the opposite-images contributes to the best representation of the system, it also complicates a little some of the concept we mentioned previously. For example, at present we have 2 possible distances between a student and a type of teaching unit and 4 distances among students. This makes it necessary to define the global distances between student and teaching units and among students, i.e. a careful combination of the particular distances. The mathematical formulation adapted for information filtering and student profiling can be observed in [5].

the information agents that create and maintain the student model and filter the learning information offering to the represented entity only the most relevant. Actually we are realizing standalone simulations for student modeling by directly interacting with the HabitatPro environment using previous student learning characteristics perceived by the application of the ILS questionnaire of the FSLSM model. The following figure shows the MAS-PLANG input/output schema.

3. Summary and future trends Considering that being self-reflexive and explicit about the role of learning styles can make more rewarding and enhance student learning at the same time, we had applied a methodology based on CBR and Fuzzy Logic techniques to model the student that uses USD teaching and learning environment taking into account his/her learning style following the directions of the FSLSM model adopted. In our system, the student modeling is used to offer teaching units adapted to the student’s learning style while considering the media format of the didactic contents, the available interactivity elements, the presentation style of the information, the best subject content for a particular situation, the most suitable navigation technique and the best navigation tool for comfortable surfing among subjects. By means of intelligent agents, the system obtains the students subjective knowledge creating and maintaining images and opposite images of them. The agent’s personalization techniques involved in the student modeling center on the concept of pairs attribute-value widely used in Artificial Intelligence for knowledge representation but including the concept of an associated weight to and attribute and to a student. From this point, the image of the student for a teaching unit is defined as a set of the triple-elements attribute/value/weight that encodes the student’s preferences and the importance given to the values of the attributes for the elements of the teaching unit to be adapted. To completely implement the approach proposed in this paper, we are developing a multi-agent system named MAS-PLANG [2] using two levels of agents: the assistants or intelligent tutors that supervise the student actions on the learning environment to offer, surprise, recommend or motivate the student with suitable didactic contents according to his/her learning style and

Figure 1. The MAS-PLANG input-output schema

References [1] A. Aamodt, E. Plaza, “Case-Based Reasoning: Foundational Issues, Methodological Variations, and System Approaches”, AI Communications, IOS Press, Vol. 7:1, pp. 39-59, 1994. [2] C. I. Peña, J. L. Marzo, “Adaptive Intelligent Agent Approach to Guide the Web Navigation on the PLAN-G Distance Learning Platform”. IEE Colloquium "Lost in the Web - Navigation on the Internet", London, November 1999. [3] C. A. Carver, R. A. Howard, and W. D. Lane, "Addressing Different Learning Styles Through Course Hypermedia", IEEE Transactions on Education, 42(1), February 1999, pp. 33-38. [4] Habitat-ProTM Environment, Agents Inspired Technologies S.A, University of Girona, Girona, Spain, 2001. http://www.agentsinspired.com. [5] M. Montaner, B. López, J.L. de la Rosa, “Developing CBR in Recommender Agents”, Kluwer Academic Publishers, Netherlands, 2001. [6] M. R. Felder and L. Silverman, “Learning and Teaching Styles in Engineering Education”. In Engineering Education 78(7), 1988, pp. 674-681. [7] R. Fabregat, J.L. Marzo, C.I. Peña, "Teaching Support Units", Computers and Education in the 21st Century: Kluwer Academic Publishers, 2000.