The use of Agents techniques on Intelligent Tutoring Systems


Lucia Maria Martins Giraffa
Informatics Institute /PUCRS CPGCC/UFRGS
CBL - Leeds Universitygiraffa@inf.pucrs.br

Rosa Maria Viccari
Informatics Institute /UFRGS
CPGCC/UFRGS 
CBL - Leeds University
rosa@inf.ufrgs.br

 
The aim of the paper is give an overview about Pedagogical Agents, presenting taxonomy, properties and its application to design and implement Intelligent Tutoring Systems.

1. Introduction
 

In recent years, the term agent has been incorporated to the vocabulary of mainstream Computer Science (CS), used to name from simple system process to highly skilled software/hardware resembles, denoting an entity created to perform either a specific task or set of tasks.

Such a wide notion is straighten in Artificial Intelligence (AI), where researchers are interested in a notion that confers to the agent the property of being intelligent. This tendency occurs also in case of Intelligent Tutoring Systems (ITS) and in Intelligent Learning Environments (ILE) when we can consider agents as Pedagogical Agents.

We can have different types of Pedagogical Agents: Tutor, Mentor, Assistance, MOO, Web (agents working on INTERNET applications), Learnerís agents and mixed agents (which can teach and learn).

Pedagogical agents have a set of normative teaching goals and plans for achieving these goals (e.g., teaching strategies), and associated resources in the learning environment [34].

Tutoring agents are entities whose ultimate purpose is to communicate with the student in order to efficiently fulfil their respective tutoring function, as part of the pedagogical mission of the system [19]. So, educational applications, in general, are based on Tutor and Mentor agents. Even on WEB applications, where the focus is on distance learning [17; 18; 23; 24; 26].

There is no consensual definition about agents on AI field. Researchers from connected areas (like education and psychology) interested in have knowledge about it will find the concepts sprayed on very technical reports and papers just into AI meetings or specific magazines. So, the aim of the paper is to give an overview about basic agentís concepts, focussing on Pedagogical Agents and their application to designing and implementing ITS, directed to researches interested in agentís techniques.

Section 2 summarises the different approaches to agents in AI and connected areas. Section 3 focus on pedagogical agents. Section 4 discusses the use of animated presentation agents for ITS. Section 5 discusses the use of agents to built ITS in a social approach and what king of educational contributions that can arise using a multi-agent architecture to design and implement educational systems. Section 6 presents some considerations about our experience and perspectives of using this approach to improve educational software.
 
 

2. Agents
 

According to Russel [30], architecture and program compose an agent. By architecture, he means, the hardware components, in general. So, an agent is anything that can be viewed as perceiving its environment through sensor and acting upon that environment through effectors. But, from the software viewpoint, an agent is substantially a program which has a specific plan of action defined in a limited domain and a behaviour pattern which allows it to change at the right moment its own interaction with the world depending on stimuli from the environment [2]. We can say that all software agents are programs, but not all programs are agents.

Wooldridge and Jennings [37] divided the notions of agents in two ways: a weak and a strong approach.


In AI literature, some authors point out different desirable properties for agents:
 
 

Some times, also, the term High-Level agents is used to distinguish agents from other software and hardware components. This "high-level'' is considered present in symbolic representations and/or cognitive-like functions such as planning capabilities, natural language processing abilities and so on [31; 32].

To sum up, agents must have these properties: reactive, autonomous, goal-driven or utility driven, temporally, continuos, communicative, learning, mobile, flexible, character. It will depend on the environment and its connected design.

Nevertheless, there is not, in the research community, a consensual definition of.

The figure 1 shows our graphical scheme to classify agents. This classification was inspired on [7].

More recently appears the notion of Pedagogical Agents because many systems for educational purpose have adopted the agentsí paradigm in order to better explore interaction and dynamics changing on teaching-learning environment. These Pedagogical Agents incorporate multiple characteristics of Entertainment Agents and some of them are on the board of Artificial Life [9].
 
 
 

The following section will make considerations about these special agents on educational applications.
 

3. Pedagogical Agents
 

In order to facilitate the assignment of pedagogical tasks in a tutoring session context, the multi-agent-system concept has been adapted to ITS, giving the idea of an agent-based approach to represent the pedagogical knowledge and its use in a tutoring context.

The fundamental reason for introducing agents as tutoring knowledge elements is their capabilities of communication and interaction (agents can adapt and learn during an instructional session). These characteristics are fundamental for agentís survival in an educational environment.

According to Vassileva [35], an agent must act in a world populated by other agents, because many of an agentís goals require the help of another agent. In this way, relationships among agents can be viewed as another kind of resources for achieving goals.

Pedagogical agents may be divided into: goal driven (tutor, mentor, assistance) and utility driven (Moo and Web agents). The table 1 shows the characteristics of pedagogical goalsí driven agents.

The utility driven agents are used by pedagogical purpose like labour agents, i.e., help students to find things (specific software, files, directories, scheduling group meeting, reminding the deadline of homework, etc.): they execute tasks. The Moo agents have good flexibility and have a textual (same context) approach. The web-based agents have mobility and operate on different contexts (Textual, hypermedia and virtual reality, for example). In this paradigm the interaction occurs based on co-operative or competitive process in which human and agents communicate and perform activities.
 


Table 1: Characteristics of pedagogical goalsí driven agents

 
Tutor
Mentor
Assistant
Knowledge about

environment

S
S
M
Domain Expertise
S
S
S
Student Model
S
M
W to Nil
Pedagogical aspects
S
M
W to Nil

Obs.: S: strong, M : medium, W : weak


The goal based agents decide their actions (to achieve the goal) based on information described by desirable situations.

So, pedagogical agents can act as virtual tutors, virtual students, or virtual learning companions that can help students in the learning process. The agent may be a character to which states can be attributed, namely mental states. These mental states are beliefs, desires, intentions, commitments and expectations. Such concepts are similar to the ones' humans possess.

Pedagogical agents have some fundamental properties: autonomously (realise a set of goals), social ability, proactiveness, and persistence. Some of them can be reactive, continuously performing, capable to learn and a character represents most of them.

This classification arises from a detailed analysis made by the authors based on ITS designed with agents approach. This analysis is summarised on Table 2, where agentsí properties found in all these systems were joined and compared. The educational purpose of all these environments was analysed in order to identify the three roles played by the goal driven pedagogical agents.

Table 2 : Agentsí properties x educational purpose


Educational

purpose

Agentsí

Properties

Tutor Mentor Assistance
Autonomously 1,2,5,6,8,10,11,13,14,15 3, 7 4, 9
Social ability 1,2,5,6,8,10,11,12,13,14,15 3, 7 4, 9
Proactiveness 1,2,3,4,5,6,8,10,11,12,13,14,15 3, 7 9
Reactive 1,2,5,6,7,8,10,14 3 4, 9
Persistent or

continuous

1,5,6,8,10,11,12,

13,14,15

3, 7 4, 9
Temporally 2    
Character 1,7,5,6,10,12,14,15   9
Learning 5    
Flexible 5, 10    

Obs.: 4 is a MOO system , 5 and 15 are a Web system

1 Persona - ITS to instruct the user how to use a modem [1]

2 ITS society to teach specific subjects [3]

3 ITS - to teach music concepts [4]

4 Assistant - ITS to help students [5]

5 ITS - to teach concepts of radiology in medicine [7]

6 ADELE - Personal tutor for each student on the WEB [17]

7 ITS - to teach Biology for children [ 16]

8 ITS - to teach some Math concepts [20]

9 ITS - to select studentís preferences [22]

10 STEVE - ITS to teach some LAB resources [28]

11 ITS - to teach Math concepts [29]

12 EduAgents - agents that help (teach) students how to solve elementary equation problems [13]

13 Co-operation environment with agents to help student learn about technology [15]

14 Multi-Ecological - ILE to student learn concepts about pollution [9]

15 Eletro-Tutor - ITS to teach some concepts about electricity [26].

Some experiments using VR-Virtual Reality are adding some new possibilities to investigate the interaction between student and learning environments [28; 34].

The fact of using characters and/or VR does not implies much difference to classify a pedagogical agent. The difference as arises from the purpose (implicit when we design the agent): it is goal driven or based on utility. Agents like ADELE and STEVE work like tutors in spite of their appearance.

The educational purpose of utility-based agents (MOO and Web) to work like a tutor. It happens because their strategies focus on teaching students specific skills by using a direct approach, like occur in 4, 5 and 15 presented above.
 
 

4. Pedagogical Agents with Animated Presentation

There are several motivations for using an animated presentation agent for teaching/learning propose:

Add expressive power to a systemís 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. [10; 11; 12; 27].

According to Elliot [6], using robust artificial personalities in tutoring agents allows us to study the implementation constraints, effectiveness, and appealingness of social interaction between tutoring agents and students.

The agents are capable of expressing these emotions through a variety of theoretical paths resulting in various multimedia manifestations. Emotions arise naturally in many human social situations as a by-product of goal-driven and principled (or unprincipled) behaviour, simple preferences, and relationships with other agents. Through the use of social and emotionally intelligent reasoning components it may be possible for automated tutoring systems to make use of a subset of these techniques formally associated only with human teachers.

Giving different personality types to a tutoring agent provide some additional tools with which work toward a number of goals. It allows us, for example, to present the same material in different ways. An example is the Multi-Ecological system [9], which has animated agents with specific personalities (implicit into the modelling) to improve the interaction between student and system.

5. ITS built in a social approach using agents

Wang and Chang, [36] say that the social learning systems are emerging learning environments that allow multiple students and agents to work at the same computer or across connected machines under different protocols of learning activities.

Recent developments on ITS consider, also, a co-operative approach between the learner and the system. Many research groups created environment-using simulation where the teaching-learning process is simulated by a set of agents in order to have interaction among them and allow us to observe the dynamics changes that occur during the interaction process. One example is presented by Moussalle [21], where all the agents must be opened to change strategies and believes/knowledge about how to solve problems. Another example is the simulation created by Frasson [7] where two agents are simulated in the computer: a tutor and a troublemaker. This second agent is a particular companion who has a specific behaviour: it deliberately misleads the solution to systematically test the studentís self-confidence and his knowledge. The learner debates the solution with the troublemaker in a process controlled by a tutor. Should there be an impasse the tutor can intervene by giving a hint, the correct solution, etc.

According with Oliveira, Viccari and Coelho [25], if teaching is considered a problem-solving activity, then the principal contribution of Distributed Artificial Intelligence - DAI - is the possibility of improving the problem-solving performance of the system. Such an improvement is possible if the task can be divided in sub-tasks with some degree of locality and parallelism. In fact, there are many works presented in the literature showing that such a division is possible. That is called the distributed perspective to implement DAI-based learning environments. Usually, the architectures developed with this approach are variations on the traditional, functional, architecture for ITS (domain model, student model, and teaching model), where one or more specialised agents implement each function. Since each agent has a well-defined function, its role within the system does not change - in other words, the system's organisation is fixed. These characteristics are often found in Distributed Problem Solving systems. Combining the contributions from the different agents generates the overall teaching plan. Control is distributed with respect to the generation of the plan, not to its execution; i.e., there is one central communication channel. Therefore, from the student's viewpoint, the system is an "individual," although internally it is a society.

A teaching/learning environment may be regarded in a social way, i.e., a society made of various autonomous agents (human and/or artificial). Some of them playing the role of tutors and other the role of learners (with the possibility of role exchange), and all involved in building a common corpus of knowledge about some particular domain. Each agent may or may not be subdivided in internal agents (subagents, local agents): if it is the case, we can talk about an external society and an internal society.

Such a view raises a number of theoretical and practical issues, and shifts the discussion towards topics like co-operative learning, the application of Vigotsky and Piaget's theories, and others. One important point is that the behaviour of each agent affects the behaviour of the society as a whole, and vice-versa. A quick learner and/or a person with learning disabilities may have some influence in the rhythm of the other learners. Such a differentiated behaviour may lead to situations demanding, for example, the interference of another agent (a tutor or another learner). The need of mechanisms to detect and deal with this kind of situations will impact the architecture and behaviour of the agents and of the external society.

Again, according to [25], it is important to notice that the two perspectives depicted here are not mutually exclusive. As we can see in the examples, the social perspective regards the computational systems as embedded in an external society, composed of multiple human/artificial agents. The distributed perspective is concerned with the internal organisation of the system. While the later focuses on engineering aspects, the former brings up to mind the questions of which educational principles we should apply, and in what kind of real environment our systems are to be applied.

For many authors like Mitsuru and Cook, social environments mean collaborative work. Mitssuru said: one of the major educational significance of the collaborative learning is to enhance the participantsí motivation to awaken mature reflections on their own understanding and externalise the result of it; Cook said that collaborative learning is not always effective. So, the analysis of dialogue data has the purpose to then uncover a correlation between a teacherís goals, intentions, and interventions for creative reflection and the studentís attempts at creative reflection.

However, we must make a distinction between collaborative and co-operative learning. For us, a co-operative learning means that two or more students have the same task and they divided the task in sub-tasks and each one do something, but the others cannot make changes on others work. Then, the final solution is a patchwork of the individual solutions. Each student has the same task and the final solution is something made by everybody. Each student can suggest his/her own solution and the group discuss about it. The final solution arises from the group interactive work. Both situations adopt the social perspective.
 

6. Conclusions

The use of Artificial Intelligence techniques is not new and has been taking place in different research groups. As Khuwaja [14] remarked, even though Intelligent Tutoring Systems have been implemented with relative success they are not practical enough to be used in the real world. This situation can change if we introduce new methodologies with multiple domains, which will allow us to apply them at schools as well as in training schemes in companies.

The restrictions of these systems can be overcome when we attempt to integrate the notion of co-operation to the teaching-learning process using multi-agents focus where the techniques and methods allow us to work in a co-operative way taking into account external human agents and internal agents modelled in the machine (computer).

So the potential use of this resource in an educational environment allows us to investigate aspects of the environment dynamics and the userís behaviour that were not possible through conventional techniques. An example is the use of different strategies to obtain the solution to a specific problem.

Intelligent software agent technology has been suggested as a promising approach to extend intelligent tutoring systems in such way that the need for social context for learning can be fulfilled.
 

7. References
 

1. Andre, E.; Muller, J.; Rist, T. Life-Like Presentations Agents: A New perspective for Computer Based Technical Documentation. AI-ED97: Eighth World Conference on Artificial Intelligence in Education - Workshop V: Pedagogical Agents, 8., 1997. Proceedings... Kobe: Japan, 1997.

2. Colazzo, L.; Silvestri, L. The pragmatics of the Tutor:A proposal of modelling. AI-ED97 : Eighth World Conference on Artificial Intelligence in Education - Workshop V : Pedagogical Agents, 8.,1997. Proceedings... Kobe: Japan, 1997.

3. Costa,E.B.;Perkusich,A. A Multi-Agent Interactive Learning Environment Model.AI-ED97 : Eighth World Conference on Artificial Intelligence in Education - Workshop V : Pedagogical Agents, 8.,1997. Proceedings... Kobe: Japan, 1997.

4. Cook,J. An Empirically Based mentoring Agent for Supporting Music Composition Learning. AI-ED97 : Eighth World Conference on Artificial Intelligence in Education - Workshop V : Pedagogical Agents, 8.,1997. Proceedings... Kobe: Japan, 1997.

5. Dillenbourg, P; Jermann, P.; Schneider, D.; Traum, D.; Buiu, C. The Design of MOO Agents: Implications from an Empirical CSCW Study.AI-ED97: Eighth World Conference on Artificial Intelligence in Education - Workshop V: Pedagogical Agents, 8., 1997. Proceedings... Kobe: Japan, 1997.

6. Elliott, C. Affective Reasoner Personality Models for Automated Tutoring Systems. AI-ED97 : Eighth World Conference on Artificial Intelligence in Education - Workshop V : Pedagogical Agents, 8.,1997. Proceedings... Kobe: Japan, 1997.

7. Franklin,S.; Graesser,A. Is it an Agent, or just a Program?: a taxonomy for Autonomous Agents. (http://www.msci.menphis.edu/~franklin/AgentProg.html)

8. Frasson,C.; Mengelle,T.; Aimeur,E. Using Pedagogical Agents in a Multi-Strategic Intelligent Tutoring System. AI-ED97 : Eighth World Conference on Artificial Intelligence in Education - Workshop V : Pedagogical Agents, 8.,1997. Proceedings... Kobe: Japan, 1997.

9. Giraffa, L.M.M.; Nunes,M.A.; Viccari,R.M. Multi-Ecological: an Intelligent Learning Environment using Multi-Agent architecture. MASTAí97: Multi-Agent System: Theory and Applications. Proceedings... Coimbra: DE-Universidade de Coimbra, 1997.

10. Hayes-Roth,B.et alli. Directed Improvisation. Stanford: CA,1994. (Technical Report KSL-94-61)

11. Hayes-Roth,B.et alli. Directed Improvisation with animated puppets. CHIí95: Conference on Human-Computer Interaction, 1.,1995.Proceedings... Denver: 1995.

12. Hayes-Roth,B.; Van Gent,R.; Huber,D. Acting in Character. In: Creating Personalities for Synthetic Actors: towards autonomous personality agents. Trappl,R.; Petta,P. (Eds.). Berlin: Springer Verlag, 1997.

13. Hietala,P.; Niemirepo,T. Collaboration with Software Agents: What if the Learning Companion Agent Make Errors? AI-ED97 : Eighth World Conference on Artificial Intelligence in Education - Workshop V : Pedagogical Agents, 8.,1997. Proceedings... Kobe: Japan, 1997.

14. Khuwaja.R.; Desmarais,M.;Cheng,R. Intelligent Guide: Combining User Knowledge Assessment with pedagogical Guidance. International Conference on Intelligent Tutoring Systems - ITSí96, 3.,196. Proceedings ... Berlin: Springer-Verlag Verlag: 1996.

15. Leroux ,P. Co-operation between a Pedagogical Assistant, a Group of Learners and a Teacher. In: Euro AI-Ed - European Conference on Artificial Intelligence in Education,1996. Brna,P et alii (Eds.). Proceedings...Lisboa: Colobri,1996.

16. Lester, J. et alli. Mixed Initiative Problem Solving with Animated Pedagogical Agents. AI-ED97 : Eighth World Conference on Artificial Intelligence in Education - Workshop V : Pedagogical Agents, 8.,1997. Proceedings... Kobe: Japan, 1997.

17. Lewis, W J.; Shaw,E. Using Agents to Overcome Deficiencies in Web-Based Courseware. AI-ED97: Eighth World Conference on Artificial Intelligence in Education - Workshop V : Pedagogical Agents, 8.,1997. Proceedings... Kobe: Japan, 1997.

18. Miyahara and Okamoto. Collaborative Filters Towards Distributed Work/Learning Environment. AI-ED97 : Eighth World Conference on Artificial Intelligence in Education - Workshop V : Pedagogical Agents, 8.,1997. Proceedings... Kobe: Japan, 1997.

19. Mitsuru,I. Et alii. Opportunistic Group Formation. AI-ED97 : Eighth World Conference on Artificial Intelligence in Education - Workshop V : Pedagogical Agents, 8.,1997. Proceedings... Kobe: Japan, 1997.

20. Morin, J.F.; Lelouche,R. Tutoring Knowledge Modelling as Pedagogical Agents in an ITS. AI-ED97 : Eighth World Conference on Artificial Intelligence in Education - Workshop V : Pedagogical Agents, 8.,1997. Proceedings... Kobe: Japan, 1997.

21. Moussale, N.M.; Viccari,R.M.;Correa,M. Intelligent Tutoring Systems Modelled Through the Mental States. In: Lectures Notes on Artificial Intelligence - SBIAí96. Borges,D. and Kaestner,C.(Eds.). Berlin: Springer Verlag, 1996.

22. Murray,W.R. Knowledge-based Guidance in the CAETI Center Associate. AI-ED97 : Eighth World Conference on Artificial Intelligence in Education - Workshop V : Pedagogical Agents, 8.,1997. Proceedings... Kobe: Japan, 1997.

23. Nakabayashi ET alii. Architecture of an Intelligent Tutoring System on the WWW. AI-ED97: Eighth World Conference on Artificial Intelligence in, 8., 1997. Proceedings... Kobe: Japan, 1997.

24. Ogata and Yano. Knowledge Awareness Filtering toward Efficient Collaborative Learning. AI-ED97: Eighth World Conference on Artificial Intelligence in Education -, 8., 1997. Proceedings... Kobe: Japan, 1997.

25. Oliveira,F.M.;Viccari,R.M. Are ..or social systems? In: Euro AI-Ed - European Conference on Artificial Intelligence in Education,1996. Brna,P et alii (Eds.). Proceedings...Lisboa: Colobri,1996.

26. Silveira,R.A.; Viccari,R.M.. Projeto Eletrotutor: Desenvolvimento e Avaliação de Ambientes Inteligentes de Ensino-Aprendizagem. CLEI-PANEL'97:XXIII Conferencia Latinoamericana de Informática, 23, 1997. ProceedingsÖ Valparaiso: CLEI, 1997

27. Petta, P.; Trappl, R. Why to Create Personalities for Synthetic Actors. In: Creating Personalities for Synthetic Actors: towards autonomous personality agents. Trappl,R.; Petta,P. (Eds.). Berlin: Springer Verlag, 1997.

28. Rickel,J.; Johnson, W.L. Intelligent Tutoring in Virtual Reality: A Preliminary report. AI-ED97 : Eighth World Conference on Artificial Intelligence in Education - Workshop V : Pedagogical Agents, 8.,1997. Proceedings... Kobe: Japan, 1997.

29. Ritter,S. Communication, Co-operation and Competition Among Multiple Tutor Agents. AI-ED97 : Eighth World Conference on Artificial Intelligence in Education - Workshop V : Pedagogical Agents, 8.,1997. Proceedings... Kobe: Japan, 1997.

30. Russel, S.; Norvig,P. Artificial Intelligence: A modern Approach. New Jersey:Prentice Hall,1996.

31. Shoham,Y. Agent-oriented programming. Stanford: CSD-Stanford University, 1990. (Technical Report STAN-CS-1335-90)

32. Shoham,Y. Agent-oriented programming. Artificial Intelligence, 60,1,1993.

33. Smith,D.C.; Cypher,A.; Spohrer,J. KindSim: Programming Agents Without a Programming Language. Communication of the ACM, 37,7,1994.

34. Thalmann,D.; Noser,H.; Huang,Z. Autonomous Virtual Actors Based on Virtual Sensors. In: Creating Personalities for Synthetic Actors: towards autonomous personality agents. Trappl,R.; Petta,P. (Eds.). Berlin: Springer Verlag, 1997.

35. Vassileva,J. Goal-Based Pedagogical Agents. AI-ED97 : Eighth World Conference on Artificial Intelligence in Education - Workshop V : Pedagogical Agents, 8.,1997. Proceedings... Kobe: Japan, 1997.

36. Wang,W.C.; Chan,T.M. Experience of Designing an Agent-Oriented Programming Language for Developing Social Learning Systems. AI-EDí97:World Conference on Artificial Intelligence in Education, 8,,1997. Proceedings...Tokyo: IOS Press,1997.

37. Wooldridge,M.;Jennings,N.R. Intelligent Agents: Theory and Practice.London: QM&Wc, 1995

(http://www.doc.mmu.ac.uk/STAFF/mike/ker95.ps)