e-Learning Powered by Intelligent Tutoring

Vladimir Goodkovsky, PhD

University of Virginia

USA

vladimir@virginia.edu

Abstract: The paper offers a way of empowering existing e-learning content, systems, and platforms with intelligent tutoring to provide the most effective solutions in education and training to accelerate successful learning. A hierarchy of main learning and tutoring models has been proposed. A basic tutoring functionality introduced includes passive and active tutoring manners, where said active tutoring manner represents interplay of presenting, testing and diagnosing modes of dynamic adaptive sequencing of learning activities/resources. The developed solutions are focused on logical aspects of tutoring, leaving all other (media, content, delivery and management) aspects to be realized by traditional components of e-learning systems.

Introduction

e-Learning is rapidly developing and growing in popularity. But despite its obvious strengths, e-Learning has still serious weaknesses.

Indisputable strengths of e-Learning are:

• Possibility to create, present and play with an impressive and engaging interactive multimedia content;
• Possibility of sophisticated organization and easy navigation through well organized content;
• Availability on the market a wide spectrum of learning (content) management, authoring tools and delivery platforms;
• Anytime and anywhere delivery of content, navigation and testing;
• Reusability and standardization in the progress.

Meanwhile, e-Learning as a branch of instructional technology represents a complex multidisciplinary field, where different people with different education, backgrounds, mindsets and goals are supposed to cooperate. It is where traditional "instructionalists" meet radical "technologists". As a matter of fact, the instructionalists have not prepared a strong theoretical basis for the technologists. That is why most technologists prefer to stay aside of instructional issues developing "pedagogy-free" technologies, tools and standards. As a result, known e-Learning technologies are instruction-free and oriented mainly on authoring, managing and delivering multimedia content for learners' self-study. Yep, it is simple, cheap, but not really effective solution. Users are increasingly not happy with it.

It is also known and proven that the most effective instruction is a one-on-one tutoring. The only problem with tutoring used to be its affordability: a few learners can afford it. But by now the situation is different: modern technologies provide a high level of scalability, which seems to be a ready-made technological basis for providing billions of learners with an effective tutoring services.

The rest of the paper considers how to empower existing e-learning with said tutoring in more detail.

Tutoring: Why?

A short answer: just because good tutoring provides a 2-sigma shift (which means 98% success) in average mastery (see Bloom 1984).

In more detail, an ideal tutoring in e-learning environment is able potentially:

• On authoring phase,
  • to organize e-learning content in mutually connected, consistent, sufficient, flexible and efficient way for the best learning possible;
  • to address a spectrum of personal abilities, preferences and learning styles of the learners;
  • to diversify content presentation to meet said spectrum;
• On learning run-time phase,
  • to provide a dynamic detailed picture of current learner model (a learning progress);
  • to make tutoring decisions critical for learning success;
  • to provide the learner with performance and achievement feedbacks;
  • to provide real-time control over e-learning tailored to the dynamic learner model by
    • dynamic restriction of learner's free navigation through available learning resources and/or
    • dynamic sequencing of available e-learning activities/resources for the learner,
  • to collect statistics for revealing not very effective parts of courseware;
• On re-authoring phase,
  • to re-author not very effective parts of courseware based on collected statistics.

e-Learning powered with such tutoring services would be a really enjoyable activity, providing a systematic success, boosting satisfaction, inspiration and confidence of the learner.

Said tutoring services are domain/audience-specific and need to be re-designed for each instructional unit. Their cost is added to total e-learning cost. That stops most practitioners. That is why the challenge is to develop a unified tutoring, which can be re-used for any domain/audience without big deal of redesign, resulting in really cost-effective instructional units.

Developing such unified cost-effective intelligent tutoring services was a goal of our work for many years.

Intelligent Tutoring: What it is

Intelligent Tutoring is a multidisciplinary field. Traditionally it is considered as an application of Artificial Intelligence in education/training targeted to developing Intelligent Tutoring Systems (ITS). Widely known classical definition of ITS specifies their main components: Expert system in domain under study, Learner model, Instructional module and Interface ( Wenger 1987), ( Burns & Capps 1988). This early definition is not very explanatory and constructive for our goal. Moreover it focuses on procedural knowledge / skills learning only.

Intelligent Tutoring: Definition

A more explanatory and constructive definition of intelligent tutoring systems has been developed in our work. (See figure 1 below). It represents a nested hierarchy of the following models:

A learning activity model, which is represented in a very generic hierarchical form as follows:

1. Domain Situation is a snapshot of the domain under learner's study;
2. Domain Space is an aggregation (a set or continuum) of all said domain situations;
3. Domain Task(s) to be solved/performed in the domain by the learner represents a pair(s) of a domain situation given to the learner and a domain situation required to be achieved by the learner;
4. Domain Expert is able to solve the task(s) in the domain by driving the given domain situation to the required one by acting in the domain space. To solve task(s) as the expert is a goal of the learner and a tutor.

A tutoring activity model, which is based upon the above learning activity model (1-4), is represented with the same hierarchical pattern (Situation-Space-Task-Expert):

5. Tutoring Situation represents a snapshot of the learner, where the learner is defined as an extension of the expert with possible deviations from an expert-like normative activity;
6. Tutoring Space aggregates all possible tutoring situations (learner's snapshots, including expert snapshots and possible deviations from them);
7. Tutoring Task to be solved in the tutoring space is a pair of given tutoring situation (the learner) and required tutoring situation (the domain expert);
8. Tutoring Expert (or namely Intelligent Tutoring System ) is able to solve the tutoring task by driving given tutoring situation (the learner) to required one (the domain expert) by sequencing said domain situations and tasks of said learning activity of the learner.

A management model

•  An administrator solves a number of organizational tasks including control over access of learners to specific courses. (These tasks are not considered in this paper).

Figure1: Hierarchy of learning and tutoring models

Intelligent Tutoring: Coverage

The introduced above hierarchy is very generic and works for Computer-Based and Web-Based Training systems, Simulator/Game-based ITS and Expert System-based ITS.

Computer/Web-Based Training (CWBT) systems, which are most widely used by now, are able to:

1. recreate specific (multimedia) learning domain Situations and by this way
2. represent the entire domain Space under study;
3. present domain specific Tasks/questions as well as
4. pre-store correct answers (as a trivial representation of an Expert's solution of the task(s)/question(s));
5. pre-store incorrect answers (as a trivial extension of the expert model up to a learner model) representing the tutoring Situations;
6. represent the entire tutoring Space by collecting all pre-stored learner's responses/answers on the domain situations and tasks/questions;
7. represent the tutoring Task (what is the next domain situation/task to present to the learner?) and
8. represent and run a script/flowchart pre-sequencing the next domain situations and tasks (as a trivial tutoring Expert model).

Known Simulators/Games are often able to model a plurality or continuum of learning situations in a domain under study. Corresponding Simulator/Game-based ITSs are able to:

1. present a plurality/continuum of specific domain Situations to the learner and by this way to
2. represent the entire domain Space under study;
3. present specific domain Task(s) to the learner (e.g., by specifying a goal of the game);
4. pre-store a correct solution/performance (as a simple Expert-gamer model) as well as
5. may pre-store incorrect solution/performance (as an extension of the expert-gamer model into the learner model) representing the tutoring Situation;
6. represent an entire tutoring Space (including all pre-stored learner's responses on domain predefined situations and solutions/performances of tasks);
7. represent tutoring Task (at least a task to compare learner's performance with pre-stored expert-gamer's performance), and
8. simulate the tutor-Expert (at least by realizing comparison of learner's and expert's performances, providing a shallow feedback and a total score).

Known Expert Systems are able to generate an expert solution of the domain tasks. Corresponding ITS based on Expert Systems (Anderson et al.1995) are able to:

1. present the learning domain Situations
2. the entire domain Space and
3. Task(s) in the domain under study;
4. perform an Expert solution of the domain task(s). In contrast to previous CWBT and Simulators / Games-based ITS, the expert solution of each task is generated automatically, not pre-stored;
5. represent a tutoring Situation as a snapshot of   a current solution of the domain task by the learner;
6. represent a tutoring Space as a variety of all tutoring situations;
7. represent tutoring Tasks (at least the tasks to compare a learner's solution with an automatically generated expert's solution in real time and to provide a learner with shallow tutoring feedback and a total score);
8. realize the tutor-Expert's solution of these tutoring tasks.

Known Learning Management Systems (LMS) have to control over all above-mentioned subordinated systems (plus over some others), particularly to provide access of each particular learner to each specific unit of instruction (to the tutoring Experts).

Intelligent Tutoring: What is Missing

Known models of the domain situation (1), space (2), task (3), and expert (4) are pretty well developed by now. Learning management systems (9) are in good shape too. In contrast, the tutoring situation (5), tutoring space (6), tutoring task (7) and tutor-expert (8) models are rather under-developed in theory and trivial in practice. Yep, they are trivial or even omitted in practice because they are too complex in theory. Practitioners just do not mess with such challenges. To make intelligent tutoring attractive and affordable for practitioners, the simple formal models of tutoring situation, tutoring space, tutoring task and expert-tutor of generative type have to be developed.

The good news is that the well-developed domain situation/space/task/expert models (1-4) can serve as a solid foundation for building said under-developed tutoring models (5-8).

Intelligent Tutoring: State of the Art

The existing variety of approaches (too many to be referenced) to ITS research, design and development can be categorized as a Bottom-Up approach or Generalization. Indeed:

• known Case-Based techniques are based on machine learning/generalization of a tutoring behavior demonstrated by human tutor (works only for trivial tutoring spaces);
• known Rule-Based techniques are based on human heuristic generalization of a tutoring behavior and
• known Model-Based techniques are based on theoretical models of a generalized tutoring behavior.

Due to complexity of real tutoring spaces, researchers working within the Bottom-Up approach focus on separate aspects/parts, not on the entire tutoring process/picture. They rely mostly on personal experience, best practices, specific pedagogical paradigms, as well as on Artificial Intelligence techniques and available technologies. No doubt this is a very strong foundation. But by definition, the Bottom-Up generalizations are always limited and error-prone. Moreover, ITSs developed on this way are usually very complex, very specific and not reusable.

Intelligent Tutoring: Advocated Approach

In order to minimize inevitable generalization heuristics in development of practical tutoring models (5-8) and to meet, justify and align known results of Bottom-Up generalizations, the Top-Down approach has been proposed. In essence, it represents a systematic specification of approved and recognized theoretical meta-models of modern exact sciences to the tutoring situations, spaces, tasks and methods.

Such a Top-Down approach is very challenging because it is targeted to creation of the most generic and powerful models of tutoring. It is also a very conservative approach because it is based on already approved and recognized theoretical frameworks. Simply put, it is just a right way of modeling, which is native to technology. For more information on this subject see (Goodkovsky 1997).

Unfortunately by now, exact sciences are not distilled and organized into simple generic models to use. Known tendency of scientific knowledge integration is not so strong as the opposite tendency of producing more and more specific knowledge in depth. That is why one needs to study a lot of specific sciences to reveal such unified models. This is a serious challenge for all researchers, especially for those who work in interdisciplinary fields such as instructional design.

Anyway, the truth is that only synergy of Bottom-Up generalization and Top-Down specification is able to provide the best simple and reusable basis for cost-effective intelligent tutoring in e-learning environment.

Intelligent Tutoring: What is Proposing

What is proposing is a new technology of powering existing e-learning with cost-effective Intelligent Tutoring built upon the following unified models in context of the above proposed hierarchy:

Tutoring Space Model (6):

• State Space:
  • A set of Learning Objectives representing predefined learnable aspects/parts of said domain Expert model (4)   (for example: competences, abilities, knowledge, skills, aptitudes, ...) :
    • Each Learning Objective's achievement states: Ground, Supplied and Achieved;
    • Strategy specified by prerequisite Relations/Beliefs among the learning objectives;
  • Learner Personal Data Framework;
• Behavioral Space:
  • Tutoring Assignments of domain Situations (1) and Tasks (3);
  • Expected Responses of the learner (4, 5) on each Tutoring Assignment;
• State-Behavior Relations/Beliefs.

Tutoring Situation Model (5):

• State Space:
  • Learner's Progress through the Learning Objectives:
    • Each Learning Objective's state (Ground, Supplied, Achieved) achievement Beliefs;
  • Learner Personal Data (Requirements, Preferences, adapting Parameters);
• Behavioral Space:
  • Tutoring Assignments (1, 3) presented;
  • Expected Responses (4,5) obtained;

Tutoring Task Model (7):

• What is given to the ITS:
  • The Tutoring Space Model (6);
  • The Tutoring Situation Model (5);
• What is required from the ITS:
  • To provide the learner with a tutoring feedback on learning performance and achievement;
  • To sequence specific domain Situations (1) and Tasks (3) and by this way to drive the learner to Achieved states of Learning Objectives (4) by the most effective way (saving on resources and/or time).

Tutoring Expert Model (8)

• Logic Layer (represented by a Unified Generator of Intelligent Tutoring (Goodkovsky 2004));
• Media Layer (Content Delivery):
  • Learning Domain Interactive Multimedia:
    • Domain Situations (Presentations, Simulations, Games, ...);
    • Domain Tasks (problems, questions, test items, exercises, ...)
    • Expert Responses (correct answers / actions / solutions ...);
    • Learner Responses (correct + incorrect answers/actions/solutions...);
  • Tutoring Persona to communicate with the learner, which can be represented with a feedback window, a talking head, or something like that.
• Logic-Media Converter to facilitate communication between said layers.

What has been Done

Introduced models have been under our extensive research, development, testing, and honing for decades. Their prototypes have been used in several generations of ITS toolkits and applications. More than 50 courses have been developed, tested and implemented in education, industry, business, and defense.

Since 2004, a new Technological Suite for powering e-learning systems by intelligent tutoring is under development. It is supposed to be SCORM compliant and include:

• (Optional) Analyst Tool to support easy Job/Task analysis and represent said domain Situations (1), Space (2), Tasks (3), and Expert (4) activity.
• Author Tool/Wizard to support easy creation of entire courseware for intelligent tutoring by filling in the blanks of proposed modeling frameworks (1-7),
• Unified Generator of Intelligent Tutoring (Goodkovsky 2004) to playback created courseware in real time of interaction with the learner,
• Intelligent Learning/Content Management System (9).

How to Develop e-Learning to be Powered by Intelligent Tutoring

For Passive Tutoring, e-Learning of task solving activity can be developed by following steps of authoring:

• Defining a separate unit of instruction around said task solving activity;
• Decomposing said domain task into a set of domain sub-tasks (3), which the learner is supposed to address during the task solving activity;
• Specifying at least one expert's response (4) on each sub-task from said set of sub-tasks;
• (Optional yet recommended) specifying possible non-expert's responses (5) on each sub-task;
• (Optional) specifying a set of specific learning objectives (4) of the unit (e.g., competences, abilities, knowledge, skills), where each objective should be detailed enough to be covered with one sub-task;
• (Optional) specifying a tutoring strategy (6) by defining prerequisite relations among the objectives;
• Assigning a specific performance feedback (7) on each specified learner's response of each sub-task;
• (Optional) pre-defining an achievement feedback (7) on learner's progress through the set of sub-tasks of the entire unit.

For Active Tutoring, traditional e-learning courses can be developed by following steps of authoring:

• Defining a separate unit of instruction, which is supposed to have its own learning objectives, learning resources and testing items;
• Specifying a set of specific learning objectives (4) of the unit, where each objective should be detailed enough to be tested with one test item;
• Collecting/developing a set of testing items (3). At minimum, each testing item should be able to test achievement of one objective and all the test items should be able to test all the objectives. Optionally, to make testing the most flexible, adaptive and effective, it is recommended to add extra test items in the test with diversified coverage of learning objectives [by 2, 3, 4,...objectives].
• Specifying a tutoring strategy (6) by defining prerequisite relations among the objectives;
• Collecting/developing a set of learning resources (1). At minimum, each learning resource should be able to supply achievement of one objective and all set of learning resources should be able to supply all the objectives. Optionally, to support the most adaptive choice and presentation of learning resources, it is recommended to upgrade a minimal set of learning resources with extra learning resources with diversified supply of learning objectives, difficulty levels, learning styles, and so on.

As one can see, making e-learning courseware ready for intelligent tutoring means making it mutually connected, consistent, sufficient and efficient for learning. That is exactly what is necessary to ensure quality of any e-learning, even for self-study in pre-tutoring era.

Intelligent Tutoring: How it Works

The ITSs based upon proposed modeling frameworks filled in with specific content are able to realize the following unified functionality:

Passive tutoring (without interventions) the learner performing a specific task by :

• Recognizing a current sub-task the learner is addressing;
• Capturing a response of the learner on this sub-task;
• Recognizing captured response by its comparison with at least one (expert's) response predefined in the expert model (4);
• Providing the learner with predefined performance feedback based upon results of the above recognition;
• (Optional) updating said learner model (5) characterizing learner's current progress through learning objectives and corresponding situation in said tutoring space;
• (Optional) making tutoring decisions based upon updated learner model and predefined positions in the tutoring space;
• (Optional) providing the learner with a predefined achievement feedback based upon position achieved by the learner in the tutoring space.

Active Tutoring (with interventions) the learner by sequencing said domain situations (1) and tasks (3) represented with available learning resources and test items includes 3 loops:

• A loop of adaptive (domain situation by situation) presenting with switching to a testing loop, when accumulated uncertainty about mastery of presented resources exceeds a customized threshold, wherein selecting each next resource takes into account its difficulty and a current difficulty threshold of the learner, as well as learner's personal requirements and preferences;
• A loop of adaptive (task by task) testing with incrementing the current difficulty threshold each time when learner's response is correct and switching to a diagnosing loop, when accumulated blaming beliefs generated by faults made by the learner exceeds a customized threshold;
• A loop of adaptive (task by task) diagnosing of fault-causing objectives with switching to the presenting loop and dropping the current difficulty threshold of the learner, when blaming beliefs accumulated by one objective exceed the same beliefs of all other objectives by a customized threshold;

These 3 loops together form a complete set of basic tutoring modes (presenting, testing and diagnosing) able to drive the learner to 100% achievement of all predefined learning objectives.

Mixed tutoring with managed interventions (for example for job support and/or knowledge management):

• Passive tutoring until a serious fault in task solving or by learner's demand;
• Active tutoring until 100% remediation of all fault-causing objectives (e.g., knowledge, skills,...);
• Resuming passive tutoring after successful remediation.

All developed solutions demonstrated easy authoring and the most effective dynamic adaptive sequencing of tutoring modes (presenting, testing and diagnosing loops) and learning resources within each mode, provided 100% mastery due to embedded diagnosing and remediation, as well as were highly appreciated by end-users. Unfortunately our marketing skills/efforts were not sufficient to get funding for the next logical step: commercial breakthrough.

Summary

Despite the fact that tutoring is a really complex activity, there is a possibility to generalize its known representations, to find relevant theoretical frameworks, and to specify them down to theoretically sound simple models of intelligent tutoring activity for e-learning technologies. This possibility has been explored and hierarchy of learning/tutoring models has been developed.

Introduced models completely separate a logic (control) from media (presentation), facilitate evaluating and ensuring quality of e-learning. They allow re-use already available e-learning media content, multimedia-oriented authoring tools and content delivery platforms and upgrade them with innovative control solutions. Generic (domain/learner-independent) parts of the models proposed can be used as authoring frameworks for easy design of traditional e-learning of high quality, as well as innovative cost-effective e-learning powered with intelligent tutoring.

Proposed solutions are generic enough to cover a wide variety of existing CBT, WBT and Intelligent Tutoring Systems. In this wide scope, the proposed solutions indicate the path for smooth transition of modern e-learning technologies from simple (like in CBT) to adaptive (like in ITS) sequencing of learning activities/resources within each instructional unit. By this way, the proposed solutions are able to bridge the existing gap between known Learning Management and Content Delivery Systems.

Moreover, the proposed solutions represent solid candidates for adaptive intelligent sequencing and navigation solutions in learning technology standards, particularly SCORM.

Literature References

Anderson, J. R., Corbett, A. T., Koedinger, K. R., & Pelletier, R. (1995). Cognitive Tutors: Lessons learned. The Journal of the learning Sciences, 4, 167-207.

Bloom, B. S. (1984) The 2 sigma problem: The search for methods of group instruction as effective as one-to-one tutoring. Educational Researcher, 13 (6):4-16, 1984.

Wenger E. (1987). Artificial Intelligence and Tutoring Systems. Morgan Kaufmann. Los Altos, CA.

Burns H. L. & Capps C. G. (1988). Foundations of intelligent tutoring systems: an introduction. Foundations of intelligent tutoring systems (Eds. Polson M. C. & Richardson J. J.). Lawrence Erlbaum, London, pp1-19. 

Goodkovsky, V.A. & Kazennov, A.Y. (1992) Intelligent Tutoring Systems: Theory, Technology, and Practice. East-West conference on Emerging Computer Technologies in Education . Moscow, RU: ICSTI. 23-127.

Goodkovsky, V.A. (1996). Intelligent Tutor: Shell, Toolkit & Technology. ITS'96 Workshop on Architecture and Methods for designing cost-effective and reusable ITSs , Montreal, Canada, June 1996.

Goodkovsky, V.A. (1997). Intelligent Tutor: Top-down Approach to Intelligent Tutoring System Design. Developing Technical Standards for Learning Technology. Learning Technology Standards Committee (WG P1484). http://ltsc.ieee.org/archive/harvested-2003-10/miscellaneous/goodkov/goodkov.htm

Goodkovsky, V.A. (2004). Unified Generator of Intelligent Tutoring . No: 10/909,101. US Patent Application.

Rob Coper. (2001). Modeling of units of study from pedagogical perspective the pedagogical meta-model behind EML. Educational Technology Expertise Centre. Open University of the Netherlands. First draft, version 2. June 2001. http://eml.ou.nl/introduction/articles.htm