CArLA - Intelligent agent as support for learning and teaching process Siniša Parović, Slavomir Stankov, Branko Žitko Faculty of Natural Sciences, Mathematics and Education Nikole Tesle 12, 21000 Split, Croatia Phone: (385) 21-38 51 33-105, Fax: (385) 21-38 54 31 E-mail: slavomir.stankov{sinisa.parovic, branko.zitko}@pmfst.hr Abstract - Intelligent tutoring systems (ITS) are computer based intelligent systems for support and improvement of learning and teaching process. Tutor – Expert System (TEx-Sys) is a intelligent hypermedia authoring shell, and it is used for editing ITS. In process of learning, students continually try to assimilate more knowledge in less time. As a result student does not adopt whole teaching unit which had unsatisfactory prerequisite for adopting new teaching unit, so this mode of learning does not aim to complete knowledge adoption. For that reason ITS produces necessity for assessment in student process of learning. Learning agent in Learning module of TExSys tries to help student in his process of learning. Guiding by using questions for checking student level of adopted knowledge is one of the learning agent feature. 1.

Introduction

Definition of an agent is not simple and it depends of its creator. Reason of that is a big difference between certain agents (different domain, way of acting, goal) so it is very difficult to give one definition for all developed agents. Because of that, after global description of agent techniques, some groups of agent will be described more detailed. According to Russel agent is created from architecture and program [4]. Architecture is hardware components which program uses for collecting information about environment (sensors) and acting upon that environment using effectors. From software point of view, agent is a (sub) program which has specific

plan of action defined on limited domain and pattern of acting which allows changing its own interaction. We can say that all agents are programs but all programs are not agents. Wooldridge and Jennings [4] mention weak and strong approach of agents. Weak approach describes agents as group of hardware (or more often software) units which properties are autonomy, social ability, reactivity and proactivity. Strong approach gives agents human characteristics like knowledge, faith, intention, obligation, rationality and so on (mental states). Agent should have some characteristics more often connected with human intelligence: learning, accommodation, independence, creativity... There are conditions which agent must satisfy to be called an agent. These major properties (mentioned by lot of authors) are: • autonomy • social ability • reactivity • proactivity Autonomy means that agent must have an ability of its own acting in solving specific problem. Issue of autonomy can be eased by fact that many of agents consult with other agents (human or computer). Social ability is very important agent property. Communication with other agents in both ways (giving and receiving information) is very important. Reactivity is a characteristic of many computer systems, and means ability of agent to act in environment and by perception to react on changes.

Proactivity is ability of an agent to act, not only as passive actor which reacts on changes in environment, but as an active element which take over initiative when necessary. Agent becomes a source of happening and not just an object of manipulation. Other properties which describe behavior of an agent are mentioned in literature like temporarily, act on behalf of someone else (subservient), rationality, learn from environment by perception, mobility, accommodation, personality based on mental states. 1.1. Intelligent agent Intelligent agent is an entity capable to solve given problems with high degree of autonomy, led by some kind of intelligence. [5] Although imprecise, this intuitive definition of an intelligent agent is the most important for understanding the term of intelligent agent. Agent is focused on solving problems but there are systems which serve to solve problems by not using the agent techniques. These systems usually have strictly defined principals of acting (they don't have ability of accommodation to environment) and they don't belong to systems made by using agent techniques. However, agent is an entity capable to interact with environment. For managing in changeable environment agent needs environment accommodation ability – a property often called intelligence. Intuitive definition presents agent as an entity which: has some goal, in achieving goal acts autonomicly and is capable to make decisions. From definition we can see that intelligent agents for implementation often use Artificial Intelligence principals. 1.2. Pedagogical agents Agent which consist a set of normative teaching goals and plans for achieving these goals (e.g., teaching strategies), and associated resources in the learning environment is pedagogical agent. Tutoring agents are entities whose ultimate purpose is to communicate with the student in order to

efficiently fulfill their respective tutoring function, as part of the pedagogical mission of the system. So, educational applications, in general, are based on Tutor and Mentor agents. Especially WEB applications are used to implement distance learning. We can have different types of Pedagogical Agents: Tutor, Mentor, Assistance, Web (agents working on Internet applications) or mixed agents which can teach and learn. Pedagogical agents may be divided into: goal driven (tutor, mentor, and assistance) and utility driven (Web agents). The Table 1 shows the characteristics of pedagogical goals driven agents. Tutor Mentor Assistant Knowledge about environment Domain Expertise Student Model Pedagogical aspects

Strong

Strong

Medium

Strong

Strong

Strong

Weak to nil Weak to Strong Medium nil

Strong Medium

Table 1. Characteristics of goals driven pedagogical agents [4] Multi-agent architecture is accepted for design and implementation of Intelligent Tutoring Systems (ITS) because of easier following pedagogical standards. Basic idea was supporting ITS with agents which has pedagogical knowledge. The fundamental reason for introducing agents as elements of tutoring knowledge is their capability of communication and interaction (agents can adapt and learn during an instructional session). These characteristics are fundamental for agent’s survival in an educational environment. Agent must act in a world populated by other agents, because many of an agent’s goals require help of another agent (for example Tutor agent can require help of Interface agent). So, pedagogical agents can act as virtual tutors, virtual students, or virtual learning

companions that can help students in the learning process. 1.3. Pedagogical Agents with Animated Presentation Pedagogical agents can have specific user interface which allows presenting of animation. There are several motivations for using an animated presentation agent for teaching/learning propose: • add expressive power to a system’s presentation skills; • help the students to perform procedural tasks by demonstrating them; • serve as a guide through the elements of the scenario (simulations); • engage students without distracting or distancing them from the learning experience (this kind of an agent is the best choice for presenting of physical or chemical experiments). 1.4. Intelligent agent CArLA (Courseware Agent or Learning Assistant) CArLA is a new feature added in Learning and Teaching module of TEx-Sys system, and acts as assistant in learning process. Purpose of CArLA is to help users (students) and bring hypermedia authoring shell TEx – Sys to higher level of usefulness. Learning assistant acts in three different steps. First step is preparation for learning process. Like in regular learning process, in this step the only user is a teacher. Second step is standard learning process where computer teacher tries to present domain knowledge (from knowledge base) to student. Last step is student testing. Student and teacher in this step get information how successful learning process was. This is some kind of feedback from system to student. That shows student capability to go forward to the next learning unit or he should relearn the same unit. 2.

Preparation for learning process

Intuitively is understandable that in this part of learning process only user is a teacher. In

this step the teacher (pedagogical expert) prepares learning assistant base (agent can not work without this base). Learning assistant base is basic source of information about domain knowledge. Teacher using his experience or pedagogical knowledge chooses strategy of showing nodes from semantic net of domain knowledge.

Figure 1. LACreate – tool for preparing Learning Assistant base Teacher who creates learning assistant base uses tool LACreate (Learning Assistant Create) (Figure 1). LACreate tool is very simple. Teacher chooses knowledge base, makes lesson by choosing nodes, and saves prepared base (learning assistant base). Administrator takes care that this tool uses only teacher. Learning assistant base contains following information’s: 1. name and path of knowledge base (semantic network of domain knowledge), 2. list of nodes (not necessarily all from knowledge base) and order of showing, 3. question and answer about chosen node, 4. wrong answer action (return to unlearned node, correct answer). It is important to say that fields with questions and answers can contain "Auto" value ("Auto" is default value). In that case CArLA,

during the learning process, generates question (randomly) about shown node and find answer on that question. Path to knowledge base: C:\Knowledge_Bases\Racunalo_kao_sustav\Racunalo_kao_sustav.kb Node No ID Question Answer Action name 1

8

Input units

Auto

Auto

Return to Unlearned Node

2

20

Mouse

Auto

Auto

Return to Unlearned Node

3

42

CDROM

What is the full name of ROM acronym?

Read only memory

Return to Unlearned Node

4

21

Optical scanner

Auto

Auto

Correct Answer

Table 2. Example of learning assistant base In this case CArLA presents one of the most important characteristics – ability to learn domain knowledge. Question and answer are saved and learning agent uses them for testing students. Before creating intelligent agent CArLA, system TEx – Sys was not able to give student unusual question during the test (in Testing or Quiz module). All questions are generated based on a question types. Now, with teacher's help, system can test student using these unusual questions. All modules for testing (Testing and Quiz module) should use these questions. 3.

Presentation of domain knowledge

Agent CArLA in this step of learning process will be described in two phases: starting of CArLA and helping during learning process. Starting is typical for agents. CArLA, using its sensors, analyzes student's actions and in case of mistake reacts. During the learning process, agent presents nodes from learning assistant base using following teachers orders

and helps student (sensors are still active and agent acts by messages in case of mistake). 3.1. Starting There are two basic types of starting CArLA agent: starting by student (if he thinks that he isn't able to learn presented unit of domain knowledge successfully) and auto starting. Auto starting is one of the most important characteristics of intelligent agents. As previously mentioned agent has sensors and uses it for checking system in background. If agent feels that system needs him, agent activates himself (Agent CArLA follows this procedure). CArLA is created to help students in learning process, so naturally its sensors are placed to follow student's work. If sensors register difficulties during this work, they will send signal to agent for auto starting. Agent CArLA has three types of sensors: 1. sensor for percepting time which student spends on learning of selected node from knowledge base (semantic network with frames), 2. sensor for percepting student's moves through knowledge base, 3. sensor for percepting student's watching of structural attributes. subroutine ShowNode: NewTime = Time If (NewTime – OldTime < 5) Then If Warning = True Then StartingLearningAssistant Else Warning = True Kraj ako End If OldTime = New Time If ((NumberAttributes > 1) And (OverviewAttributes = False)) Then PokretanjeLearningAssistenta End If OverviewAttributes = False

Table 3. Subroutine for starting Learning Assistant Following the time spent for learning selected node, agent tries to forbid student a fast pass through learning unit, because that way of

learning doesn't give good results. If student spends less time than necessary for learning (default value is 5 sec), sensor sends warning to student. If student repeats time sensor error, agent CArLA begins with auto starting (Table 3), (agent thinks that student is not able to learn unit alone). Following the moving through semantic network agent tries to minimize possibility that student gets lost in semantic network. If student chooses the same node more then two times, sensors will result with auto starting of CArLA. One of criteria to a good preparation of knowledge base is also a lots of structural attributes (pictures, Web sites, presentations...) connected to nodes in semantic network. Structural attributes are very helpful to students for better understanding of presented issue (node). Sensor follows watching of structural attributes and reacts if node has more than one structural attribute and if student doesn't look at it at all. 3.2. Acting during the presentation of domain knowledge Agent CArLA, using Learning Assistant base (described before), leads student through the knowledge base following the order that teacher finds the most successful. It is supposed that this way of learning is more successful. Agent acts only if it is started (by student or by auto starting). Big problem in Learning and Teaching module of TEx- Sys was that students were not interested in this kind of learning. This problem resulted with a very small degree of absorbed knowledge. Learning Assistant will try, with unusual questions, to increase student's interests in domain knowledge, and also will make sure that student does not skip unlearned unit. During the learning process, agent acts in background. It leaves the initiative to student and acts with advices if necessary. The student, of course, doesn't have complete initiative (he is forbidden to use list of nodes for moving through knowledge base and he is

forced to follow order of nodes that teacher saved in Learning Assistant base).

Figure 2. CArLA in Learning and Teaching module In this step of learning process, agent CArLA acts in two ways. First is acting by warnings. Sensors are still active and agent helps student to reduce number of errors with messages. Second way of acting is learning with student. In background agent needs to generate question (if in learning assistant base question's value is "Auto") and, after analyzing domain knowledge, to find an answer to generate question. During its work, agent CArLA saves all information’s in history file. Analyzing this file teacher can see with what part of domain knowledge students have problems, and what kind of problems. Information from this file can be useful in student modeling and during evaluating process. 4.

Test of knowledge after learning

After learning, as it was pointed out, we approach to student knowledge testing. On results of knowledge test depends student passing over to next learning unit or he will be directed on re-studying the same learning unit. After first and second phase of learning, Learning Assistant has saved question, for every node and answer on it. In final phase agent presents mentioned questions. If student does not know answer. Agent offers him a

help. Choosing Help option (Figure 3) instead of entering the answer, offers him a help in a form of four possible answers where only one is correct.

Figure 3. Test of knowledge with Help option Now student, using elements of recognition, can pick correct answer. It is obvious that in the first method shows better student knowledge and it should be evaluated by using different coefficient (appropriate coefficient will be gained after testing more users). At the end of testing, student is informed about whole test process: number of test questions, number of correct and wrong answers, and important answer to question, is student ready for learning next learning unit or relearning is necessary. This last information does not only depend on proportion of correct and wrong answers; it also depends on teacher’s consideration of particular node importance. In other words, nodes in Learning Assistant base with assigned Correct Answer action obviously, on opinion of the creator of named base, are less important then nodes with assigned Return to Unlearned Node action. So, student has to answer correctly at least half of questions based on node with assigned Return to Unlearned Node action to step over to the next learning unit. That is on questions placed in Learning Assistant base. Result of agent work is not only the mark of student success in learning, but there are also many quality information saved in History file. Information in that file can be used for student evaluation and questions used by

agent can also be used for testing whole learning unit. We need to work out on algorithm for using saved questions. 5.

Conclusion

Created agent CArLA needs few upgrades in future. Agent must be capable to split knowledge base to small learning units and use courseware strategy (order of learning units, required knowledge for each learning unit). Learning units will be presented one by one, and agent will take care about which student is successively learned which unit. Agent will forbid student to learn unit if he did not successive passed previous learning units. Agent can be designed to analyze semantic network and make his own criteria of presenting nodes from that knowledge base. Agent can, by using of induction or deduction technique, analyze semantic network and make learning assistant base. After all, it is important to notice that agent CArLA needs testing to better defining of parameters (time needed for learning selected node, percent of knowledge witch student mast show for successively passing learning unit). Testing will also show is this new technique in Learning and Teaching module of Tutor – Expert System better than this module without usage of CArLA. 6.

References [1] Godamski, Adam Maria: Agents and Intelligence;1998. http://www.erg.casaccia.enea.it/ing /tispi/gadomski/gad-ag.html [2] Horvat, Katarina: Projecting Dynamic Courseware in Hypermedia Authoring Shell TEx-Sys; Graduate thesis, mentor dr. sc. Slavomir Stankov; Faculty of Natural Sciences, Mathematics and Education University of Split, Split, Croatia, 2002 (in Croatian). [3] Lieberman, Henry: Autonomous Interface Agents; Electronic

Publications, CHI, 1997. http://www.acm.org/sigchi/chi97/pr oceedings/paper/hl [4] Martins Giraffa, Lucia Maria; Viccari, Rosa Maria: The Use of Agents Techniques on Intelligent Tutoring Systems; RIBIE 98, IV Congresso da Rede Iberoamericana de Informática Educativa, 1998. http://www.c5.cl/ieinvestiga/actas/ri bie98/156.html [5] Motik, Boris: Using Intelligent Agents as Improvement of Interaction with the Information Infrastructure; M. Sc. Thesis; Faculty of Electrical Engineering and Computing, Zagreb, Zagreb, Croatia, 1999 (in Croatian). [6] Stankov, Slavomir: Isomorphic Model of the System as the Basis of Teaching Control Principles in an Intelligent Tutoring System; PhD Diss; Faculty of Electrical Engineering, Mechanical Engineering and Naval Architecture, University of Split, Split, Croatia, 1997 (in Croatian). [7] Stankov, Slavomir; Glavinić, Vlado; Granić, Andrina; Rosić, Marko: Intelligent Tutor Systems – Research, Development and Application; Magazine Edupoint, Zagreb, 2001 (in Croatian). http://edupoint.carnet.hr/casopis/br oj-01 [8] Stankov, Slavomir; Rosić, Marko; Rakić, Krešimir: Testing and Evaluating using Quiz in Intelligent Tutoring Systems; Zbornik radova MIPRO, Opatija, 2001 (in Croatian). [9] Veljan, Darko: Combinatory and Discrete Mathematics; Algoritam, Zagreb, Croatia, 2001 (in Croatian). [10] Žitko, Branko: Approachto Project of Computer Based Intelligent Tutoring Systems; Graduate thesis, mentor dr. sc. Slavomir Stankov; Faculty of Natural Sciences, Mathematics and Education University of Split, Split, Croatia, 2002 (in Croatian).

[11] Parović, Siniša: Intelligent Agent as Support in Learning and Teaching Process; Graduate thesis, mentor dr. sc. Slavomir Stankov; Faculty of Natural Sciences, Mathematics and Education University of Split, Split, Croatia, 2002 (in Croatian).