Intelligent Tutoring Systems: Architecture and Characteristics Abstract Indira Padayachee University of Natal, Durban, Information Systems & Technology, School of Accounting & Finance padayacheei@nu.ac.za This paper provides a close examination of specific intelligent tutoring system (ITS) architectures spanning the period 1988-1999. ITSs are classified into three categories, namely traditional three-model, classical four-model and new-generation architectures for the purposes of this study. Similarities and differences between architectures of the same category, and between different category architectures are discussed. The study depicts the influence of application domains, learning and instruction trends, as well as software development advances on ITS architecture and behaviour. This research aims to examine various ITS architectures and its actual behaviour to establish a set of generic characteristics and behaviour. These characteristics are useful for comparing and evaluating existing ITSs, and can guide the design of new ITSs. Key Words: Intelligent tutoring system, Artificial intelligence, Intelligent Tutoring System architectures, Intelligent tutoring system characteristics. Computing Review Categories: K.3.1, K.3.2 1. Introduction Intelligent Tutoring Systems (ITSs) are instructional systems that use artificial intelligence (AI) techniques in computer programs to facilitate learning. These systems are based on cognitive psychology as an underlying theory of learning, which deals mainly with issues such as knowledge representation and organisation within the human memory as well as the nature of human errors [Shute & Psotka, 1996]. The intelligent tutoring systems adopt a mixed-initiative teaching dialogue, which allows the system to initiate interactions with the learner, as well as interpret and respond meaningfully to learner-initiated interactions [Garito, 1991; Beverly Park Woolf, University of Massachusetts, 1998]. There exists a number of research papers that provide detailed descriptions of intelligent tutoring system architectures developed for specific application domains. These papers are useful in that they assist in understanding the functionality and operability mechanics of ITSs, and stimulate further development. However, there has been limited effort in examining both system architecture and behaviour, to ascertain common characteristics of ITSs. This research aims to examine various ITS architectures and its actual behaviour to establish a set of generic characteristics and behaviour. These characteristics are useful for comparing and evaluating existing ITSs, and can guide the development of new systems. The study also considers factors that have influenced ITS architectures, and discusses the similarities and differences between ITS architectures. 2. Intelligent Tutoring System architectures This section examines selected intelligent tutoring systems spanning the period 1988-1999, and classifies them into three categories, namely three-model, four-model and new-generation architectures. The ITS architectures have in the main been named after their respective designers for easy referencing. 2.1 Three-model architectures for ITSs A three-model architecture typically comprises three major building blocks or components, namely the systems domain expertise, student knowledge and skill, and tutoring expertise. Two examples representing the three-model architecture are discussed below. 1
2.1.1 Derry et al architecture Derry, Hawkes & Ziegler [1988] proposed an ITS architecture comprising three major components, namely a tutoring model, expert domain model, and student knowledge model. Each of these components will be discussed below. Tutoring model This model makes use of heuristically guided routines to perform three levels of instructional activity, namely planning an individual s route through a curriculum (the agenda); planning lessons (using action schemata); and online tutorial intervention. At each of these instructional levels, student performance data can be compiled and made available to other levels. This model uses three sub-components, namely a curriculum planner, lesson planner and intervening monitor for performing the three levels of instructional activity. Expert domain model This model provides information to guide routines in the tutoring model. Student knowledge model This model also informs the routines of the tutoring model. Derry et al [1988] combines the planning and opportunistic architectures for intelligent tutoring, by using global models of domain knowledge to structure the student s learning experience, and bases didactic decisions on the learner s operational model of performance. This architecture has a well-defined tutorial component embodying curriculum planning, lesson planning, and tutorial monitoring and intervention with the ultimate goal of moving the student toward the expert model of knowledge. The functionality of the expert domain and student knowledge models appears to be limited to providing information to routines in the tutoring model. A shortcoming of this architecture is the omission of the user interface component and allied usability issues. A major strength of this architecture is the use of heuristically guided routines in the tutoring model, which provides dynamic student-centred tutoring. This architecture is recommended for problem-solving or procedural task application domains. 2.1.2 Siemer & Angelides architecture The next architecture described by Siemer & Angelides [1998] also supports a basic three-model structure, comprising domain expertise, student knowledge and skill, and tutoring expertise. The point of departure between this architecture and the previous one is that it explicitly identifies the processes manipulating the three knowledge bases in each of the models, and incorporates an additional process to co-ordinate the three models. This general intelligent tutoring system architecture is illustrated in figure 1.1. The three major architectural components and the overall system control are described below. The domain model The domain model contains knowledge relating to the subject matter. The intelligent tutoring system utilises its domain knowledge to reason with and solve problems, or to answer questions posed by students. Different knowledge representations of the same domain knowledge may be incorporated to support alternative teaching strategies. Domain knowledge processes, referred to as the expertise provide for the content of tutorial interactions. The tutoring model The tutoring model provides the knowledge needed to attain teaching goals. It should have: control over the sequence and selection of subject material to be presented to the student; response mechanisms to answer learners questions with appropriate answers; and knowledge of when learners need help, in the course of solving a problem or practising a skill, and what type of help to offer. To achieve this, the tutoring model needs to embrace different teaching strategies (styles of delivery). The tutoring knowledge processes referred, to as the didactic aspect is responsible for selecting teaching goals, and for determining appropriate teaching strategies for learners on the basis of their student models, learner s needs and/or preferences, learning experiences, the domain of discourse and the instructional objectives of the intelligent tutoring system. 2
The student model The student model represents the learner s emerging knowledge and skills. Information such as learning preferences, past learning experiences and advancement may als o be relevant in adapting the teaching process. It may also record the learner s errors and misconceptions. The student knowledge processes, known as diagnostics, analyse the behaviour of the student. Overall system control The overall system control process is needed to co-ordinates the three-knowledge models to provide student-centred tutoring. For example, an intelligent tutoring system has to select teaching strategies and presentations for each subject area, in accordance with an individual learner s needs and preferences stored in the student model. Additional flexibility may also be provided with a student or system initiated help system. This architecture provides a comprehensive description of the knowledge and functional requirements related to the various architectural components comprising an ITS. It does not delve into implementation issues or operability mechanics. This architecture extends the concept of lesson planning and dynamic adaptation thereof in the previous architecture, to embrace multiple knowledge representations and teaching strategies. Each of the models are welldefined in terms of the knowledge bases and processes supported. Furthermore, an additional process is identified for managing and co-ordinating the activities of the three models, which is not evident in the previous architecture. Since this is a general Intelligent Tutoring System architecture, it can be adopted and/or adapted for use in any application domain In summary, the three-model architecture represents the traditional architecture of ITSs comprising three main components that are commonly referred to as domain model, student model, and tutoring model. The three-tier architecture of ITSs made way for the four-model ITS architecture, which is discussed next. 2.2 Four- model architecture for ITSs The four-model architecture retained the three major components of the traditional three-model architecture, and added the user interface as a fourth component. This architecture became the classical standard architecture for ITSs. Dede s architecture A typical example of a four-model architecture comprising the knowledge base, student model, pedagogical module, and user interface is described below [Dede, 1986]. A brief discussion on each of these components follows: Knowledge base A tutor or coach incorporates declarative (what), procedural (how), and metacognitive (thinking about what and how) knowledge. This component is synonymous with the domain model of other architectures. Student model An internal model representing cognitive processes (such as information retrieval, calculation and problem solving), meta-cognitive strategies (for example, learning from errors) and psychological attributes (developmental level, learning style, and interests) are maintained for each learner. Pedagogical module This module is similar to its counterpart the tutoring model in other architectures. It uses a model of the learner's present comprehension to select an efficient path through its knowledge representation to generate expert behaviour by the learner. It employs different teaching strategies on the basis of an evolving student model, and an underlying instructional theory that determines which pedagogical means is most efficient to accomplish a given end, alternative approaches to dialogue management (adjusting to different learning styles), and domain-dependent teaching heuristics. A teaching module is recommended to facilitate integration and co-ordination of the functions of the other components. User interface This component integrates three types of information that are needed in carrying out a dialogue: knowledge about patterns of interpretation (to understand a speaker) and action (to generate utterances) within dialogues; domain knowledge needed for communicating content; and knowledge needed for communicating intent. 3
The four-model architecture makes a number of important contributions in that it: incorporates a separate user interface component; embodies cognitive processes, meta-cognitive strategies and learning styles in the student model; and includes domain dependent teaching heuristics in the pedagogical model, all of which are hitherto unmentioned. The user interface is regarded as an internal and integral component of the ITS architecture whereas other architectures viewed this component as external to an ITS. This inclusion has positive consequences in that user-interface design and usability issues have become part of ITS development concerns. Another important contribution is the use of a pedagogical component, which combines instructional theory with pedagogical strategies and dialogue management for providing instruction to learners. This architecture was coined as classical in that it became the industry standard for ITS construction. 2.3 New-generation architectures for ITSs New-generation architectures represent a departure from the traditional three-model and classical four-model architectures in that they integrated software development advances, as well as new learning and instructional theories into the design of ITSs. Two of these architectures are described below. 2.3.1 Multi-agent architecture A multi-agent architecture called MATHEMA [Costa & Perkusich, 1996] illustrated in figure 1.2 is used as a basis for the design of a computer-based intelligent learning environment, which comprises the following six components: 1. An external motivator (representing human external entities that motivate the learner to work in MATHEMA, for example, a teacher); 2. A human learner; 3. A micro-society of artificial tutoring agents (MAR), that may co-operate among themselves to achieve problem solving activities in a formal and well-structured knowledge domain, divided into different microdomains, each one covering micro-specialities. 4. A human experts society (HES), working as sources of knowledge to MAR; 5. An interface agent between a human learner and MAR, responsible for communication and which includes a mechanism of selection for tutor agent (supervisor); and 6. A communication agent providing interaction between MAR and HES, and responsible for the communication and maintenance of MAR. The main idea underlying this architecture is to integrate human learners in a micro-society of artificial agents with the objective of promoting their learning. Grandbastien [1999] is of the opinion that agent-based architectures are more flexible in that new artificial tutoring agents may be added, existing artificial tutoring agents may be modified and/or deleted without negatively impacting on the operation of the other components within the architecture. Hence modular development would be enhanced, and module reusability would be promoted. This architecture promotes the notion of a computer-based intelligent learning environment, which includes external human motivators and a human expert society. Due recognition is thus given to the environmental context in which learning takes place. This architecture encompasses components of the traditional & classical architectures, albeit in a unique structure and representation. The domain model is embedded in MAR and HES, the role and functions of the tutoring model is distributed among a micro-society of artificial tutoring agents, the user interface component is represented as interface agent and the human learner is represented as a component of the learning environment, and not as a student model as with other architectures. This architecture is recommended for specific, formal and well-structured knowledge domains such as the classical logic domain. 2.3.2 Self s architecture Self [1999] revisits the conventional tripartite division of ITSs into the domain, student and tutoring model from a constructivist learning perspective. The new tripartite model of the architecture of computer-based learning environments includes the standard ITS architecture as a subset, and is displayed in figure 1.3. The proposed new generation three-model architecture for intelligent tutoring systems comprises the situation, interaction and affordance models. Each of these components is dis cussed below. Situation model The contexts and the dynamics of the learning process are embodied within a situation model, which contains descriptions of resources (although it may also contain representations of aspects of the domain of knowledge), which are available in a learning situation as opposed to a pure domain model, which contain descriptions of target knowledge. A model of domain knowledge may thus be perceived as a subset of the broader notion of a situation model. 4
Interaction model An interaction process model focuses on interaction sequences by considering the learner s actions, the contexts in which they occurred, and the learner s cognitive structures at the time. Here again, the notion of an interaction process model is perceived as a superset of an ITS-style student model. Affordance model An affordance model is developed in terms of items of knowledge, which may be learned through particular events (for example, an event such as the presentation of remediational feedback affords the learning of the item of knowledge being remediated). The affordance model is thus broader than the model of teaching as curriculumbased planning. Self extends the scope of the traditional three-model architecture as a result of new insights gained into the complex nature of the learning process, and strongly recommends that learning resources, interaction sequences, and items of knowledge be modelled into the respective architectural components. Self pays due diligence to, and explicitly incorporates the context of learning into the architecture of ITSs which has not been alluded to in other architectures. The issue of context in learning has become an important research area since an understanding of contexts is needed in order to design better and more usable ITSs [Patel & Russell, 1998]. Self s architecture challenges traditional learning theories and may be regarded as a watershed for the development of future intelligent tutoring systems based on constructivistic processes involved in learning. New-generation architectures for ITSs have emerged from the need to build specific functionality for specialised application domains, embrace important trends in software development, namely modular and incremental development, global sharing of knowledge, as well as incorporate current trends in learning and instruction. 3. Genetic characteristics of Intelligent Tutoring Systems From the architectures examined, it is evident that intelligent tutoring systems possess a number of generic characteristics and behaviour, which are related to specific architectural components. Table 1.1 outlines five architectural components and their associated characteristics. It is envisaged that these characteristics would be useful for the comparison and evaluation of exis ting ITSs, as well as guide the design of new ITSs, It should, however, be noted that the architecture and behaviour of ITSs are to some degree influenced/constrained by specific application domains (e.g. geography, programming, mathematics, etc.) and design paradigms (e.g. problemsolving monitors, coaches, diagnostic tutors, microworld etc.). Hence, some of the characteristics may with justification, not be incorporated in certain architectures. 5
Table 1.1 Generic Characteristics/Behaviour of an ITS ITS Architecture Domain Model Tutoring Model Student Model System Control User interface Characteristics/Behaviour Intelligent tutoring systems should: Possess system domain knowledge to ma ke inferences or solve problems; Provide explanations of problem solutions; Give alternative explanations of the same concept; Answer arbitrary questions from the student; Incorporate knowledge about common misconceptions and missing concepts. Possess system teaching goals & plans; Provide alternative teaching strategies; Be guided by an underlying instructional theory; Tailor system s teaching strategies with student s needs; Allow student to initiate instructional activities; Provide contextualised, doma in-relevant and engaging learning activities; Diagnose misconceptions and learning needs; Intervene if the student appears to be having difficulty; Relate a diagnosed error to a misconception or a missing concept; Incorporate remedial strategies in order to provide alternative remedial teaching styles. Maintain information about the student s knowledge, and skills (current and advancing) in the student model; Store information on the student s cognitive processes; Store information on student s learning preferences and/or past learning experiences in the student model, if the need arises; Monitor and assess student performance and update student model. Provide helpful feedback on student input; Treat all detected errors; Respond if it cannot diagnose an error; Intervene to remediate a misconception or a missing concept; Adapt to student s level of advancement; Adapt to the needs and preferences of the student. Promote ease of use; Incorporate natural interaction dialogues; Ensure that the dialogue is task-oriented and adaptive; Possess an effective screen design; Embrace a variety of interaction styles. Sources: [Beverly Park Woolf, University of Massachusetts [1998], Costa & Perkusich, 1996]; [Dede, 1986]; Derry et al [1988]; Dix et al [1993]; Gold [1998]; Self [1999]; and Siemer & Angelides [1998]. 6
4. Conclusion A close examination of specific three-model, four-model and new-generation architectures for intelligent tutoring systems (ITSs) has been undertaken spanning the period 1988-1999. ITS architectures are, to some degree, related to and influenced by one or more of the following factors: application domain, design paradigms, architectural styles such as the plan-based architecture and/or opportunistic architecture, software development advances such as agentbased architectures, and modern learning and instructional theories. Given all these factors/influences, a number of different architectures have emerged, each bearing some resemblance to others, yet possessing some unique characteristic(s)/functionality. An important outcome of this investigation is the unveiling of a number of generic characteristics and behaviour that should be provided by the architecture of an ITS. These characteristics play an important role in both the design and evaluation of ITS systems. Potential further areas for research are the use of this model incorporating both architecture and behaviour to design new intelligent tutoring systems, as well as to evaluate existing ITS architectures for conformity. References Beverly Park Woolf, University of Massachusetts. Nov 1998. Training & Development. 52(11). Costa, E. de B. and A. Perkusich. 1996. Modelling the co-operative interactions in a teaching/learning situation. In C. Frasson, G. Gauthier and A. Lesgold (Eds) Lecture notes in computer science: intelligent tutoring systems, Proceedings of ITS'96, third international conference on intelligent tutoring systems, Montreal: Springer- Verlag. Dede, C. 1986. A review and synthesis of recent research in intelligent computer-assisted instruction. International man-machine studies, 24, 329-353. Derry, S.J., Hawkes, L.W. and U. Ziegler. 1988. A plan-based opportunistic architecture for intelligent tutoring. In Proceedings of ITS-88, first international conference in intelligent tutoring systems, Montreal: University of Montreal. Dix, A., Finlay, J., Abowd, G. and R. Beale. 1993. Human-computer interaction. Hemel, Hampstead: Prentice- Hall. Garito, M.A. 1991. Artificial intelligence in education: evolution of the teaching-learning relationship. British journal of educational technology, 22(1), 41-47. Gold, S.C. 1998. The design of an ITS-based business simulation: A new epistemology for learning. Simulation & Gaming. 29(4). Grandbastien, M. 1999. Teaching expertise is at the core of ITS research. International journal of artificial intelligence in education, 10, 335-349. Patel, A., Russell, D. 1998. An initial framework of contexts for designing usable intelligent tutoring systems. Information Services & Use, 18(1/2). Self, J.A. 1999. The distinctive characteristics of intelligent tutoring systems research: ITSs care, precisely. International journal of artificial intelligence in education, 10, 350-364. Shute, V. J. and J. Psotka. 1996. Intelligent tutoring systems: past, present and future. In D. Jonassen (Ed.) Handbook of research on educational communications and technology. NY: Macmillan. pp.570-600. Siemer, J. and M.C. Angelides. 1998. A comprehensive method for the evaluation of complete intelligent tutoring systems. Decision support systems, 22, 85-102. 7
Appendix User interface Domain Model Domain Knowledge Expertise Overall system control Student model Student knowledge diagnosis Teaching Model Teaching strategies Teaching goals Knowledge didactics Figure 1.1: Siemer s & Angelides s general intelligent tutoring system architecture Source: Siemer & Angelides[1998:87] External Motivator Human Learner Interface Agent Human experts society Communication Agent Microsociety of artificial tutoring agents Figure 1.2: Costa s & Perskuchisk s Architecture of MATHEMA for IT Learning Environment Source: Costa & Perkusich[1996:170 ] Situation Model Domain model Interaction Model Student Model Affordance Model Tutoring Model Figure 1.3 Self s ITS components Source: Self [1999:15] 8