Adaptive Tutoring Systems M Sasikumar sasikumar@iitb.ac.in - - PowerPoint PPT Presentation

Sasikumar M

1

Adaptive Tutoring Systems

M Sasikumar sasikumar@iitb.ac.in thelittlesasi@gmail.com

Sasikumar M

2

Overview

• What is ATS
• What is ITS
• Knowing vs Teaching
• Course plan, assessments, etc
• Architecture, components and issues
• ...Case study of LCM tutor
• Case study of Stepmat & Language learning

Sasikumar M

3

ATS

Direct customised instruction and/or feedback to learner. Adaptive tutoring systems

• ne to one, adaptive => personalised

Focus can be: instruction, problem solving instruction – personalised/adaptive instruction – tailor instruction as per the user problem solving – ITS – our main focus in the course.

Sasikumar M

4

ATS aspects

Both needs a student model with different

• concerns. In 1, it is learner style and other

relatively static aspects. In 2, it is more fine grained and very dynamic. Focus on design and development of systems in these categories, and review relevant educational aspects.

Sasikumar M

5

Personalise what?

• Anything!
• Curriculum
• Content – language, style, examples
• Pedagogy – definition-example-explanation?
• Learning style – visual, hands-on, auditory
• UI – what to display where and how
• Pace, Location and time
• Problems
• Explanations
• Feedback
• Intervention strategies

Sasikumar M

6

Why personalise?

• Traditional instruction – one size fits all – not

effective enough

• Use of computer and Internet enables a solution!
• Learning is personal!
• E-learning has not much followers, despite

claimed advantages.

• Negative factors
• Personalisation could alleviate them.

Sasikumar M

7

Not a new interest

• Despite the lack of

technology, the concern is many decades old.

• Programmed instruction,

Bloom's mastery learning, Kellers personalised system

• f instruction, etc
• Listed among engineering

grand challenges!

For years, researchers have debated whether phonics or whole-word recognition is the best way to teach children how to

• read. Various experts can be found who will advocate
• ne approach or the other.

Sasikumar M

8

Sasikumar M

9

Personalise?

• Matching content to learner style: good or bad?
• Learning theories recommend presenting

information in multiple perspectives.

• Existence of mental representations of

environment [Cognitivism].

• Dewey's insight that learning process is of

unique and individual nature.

Sasikumar M

10

Our scope

• Two streams in this course:
• Adaptive instruction
• Intelligent tutoring
• Important and difficult to realise in general.

Sasikumar M

11

ITS

• Names
• ICAI – intelligent computer aided instruction
• Knowledge based tutoring systems
• ITS name coined by JS Brown
• Early 70s start from AI angle. Peak period in 70-
• 80s. Human tutor as model.
• One-to-one tutoring demonstrated significant

performance gain (Bloom, 84) – can we get even a part of that with ITS?

• Now re-activated due to growth of e-learning,

web and high performance computing systems.

Sasikumar M

12

What is ITS?

• Intelligent
• In what way?
• Interaction
• Diagnosis
• Pedagogy
• etc
• Tutoring
• Not teaching!
• One to one

assumed

• Narrowly focussed,
• ften a single type
• f problem.

System

Sasikumar M

13

ITS and e-learning

• Infinite repetitions of problem solving sessions

at individual pace

• High degree of personalisation
• Studies show good performance improvement

with such approaches.

• Particularly important in the increasing

acceptance of constructivist view of learning.

• But:
• Hard to build: ITS is AI-complete
• Expensive in system capability

Sasikumar M

14

ITS as intelligent teaching

• Organisation of content
• Granularity
• Meta-data/annotations
• Selection of content for individual learner
• Goal ordering
• Remedial teaching
• Refer to specialised alternate resources.

Sasikumar M

15

ITS as tutoring

• Specific problem solving skills
• Tutoring depends on the nature of problem

solving

• Equipment diagnosis/classification
• Iterative processes
• Processes
• System design and implementation critically

dependent on this.

Sasikumar M

16

MYCIN to GUIDON

• MYCIN – among the earliest expert systems
• Very popular, and spurred most of the

innovations in the field of expert systems.

• Domain: diagnosis of bacterial infections of

blood, and suggest treatment.

• Evaluation showed performance comparable to

that of experienced doctors.

• Why not use it to teach new doctors?

Sasikumar M

17

..

• MYCIN to be the 'expertise' provider
• A teacher module will:
• Pick a patient to be diagnosed
• Learner pretends to be the doctor, ask for

symptoms.

• Asks for investigations (e.g. blood count) and
• bservations (nature of work, etc)
• Compares these with what MYCIN would do for the

same case

• Uses difference to “instruct” the learner.

Sasikumar M

18

...

• MYCIN had the 'usual' explanation facilities
• Why do you want to know?

– I am trying to see if you have this disease

• How did you infer x

– I used rule 213 which says the following.

• These could be 'adapted' to instruct the learner
• f what is wrong
• So it seemed!

Sasikumar M

19

...

• MYCIN had rules to diagnose, not explicit

justifications and strategies.

• When learner deviates from the MYCIN route, it

was hard to know if it is right or wrong.

• Rules encoded different types of knowledge,

not distinguishable.

Sasikumar M

20

...

• If person is less than 8 years, do not prescribe

tetracycline.

• Significance of 8 years?
• What is the effect, if this is violated?
• Are there situations where this can be broken?

Sasikumar M

21

...

• If person is female
• And age is between 15 and 50
• And pregnant
• Then swelling of knees could be due to

pregnancy

• An attempt to make MYCIN look smart!
• But confusing to a learner as to the purpose of

the rule.

Sasikumar M

22

...

• Similarly rules embodying definitions, common

norms (age vs alcohol use), priorities, etc.

• A thorough revamp of the knowledge base was

needed to make it 'respectable' as a tutor.

• Clancey documents the journey in his seminal

papers – worth a read.

Sasikumar M

23

Course Plan

• Contact sessions: Tue and Fri, 2.00 to 3.30pm
• Assessment:
• 4 assignments, mid-term test, final exam (standard

weightage)

• Along the lines of LCM, build one simple ITS.

Choose a domain and stick to it till end.

• Build a student model and problem selection for

assignment 1.

• Implement an ITS using CTAT framework.
• Implement a pedagogy model for the assignment 2.

Sasikumar M

24

Broad Scope...

ITS Adapt Instn Learning Theories Learner Styles Knowledge Representation Machine Learning Validation Of Systems Case Studies Student Model Tutor/ Pedagogy Model Domain Model Development Tools CTAT Python

Sasikumar M

25

Resources

• Building Intelligent Interactive Tutors, Beverly

Park Wolf, Morgan Kaufman, 2009

• Most of the aspects covered in the course can be

seen here.

• A lot of papers and articles
• Almost all will be available through the Moodle

portal

• You are welcome to suggest relevant articles,

resources, etc

• Post on Moodle; will review and add to the pool.

Sasikumar M

26

ITS components

Student Model Domain Model Tutoring model UI

Sasikumar M

27

Student model

• A number of aspects can be included here
• Learner background [maths, language, science, etc]
• Learning style
• Psychological profile [motivation, initiative,

collaboration, etc]

• Knowledge wrt what is to be learned
• Interaction history with the system so far
• etc.
• Selection of what to include, depends on other

models

Sasikumar M

28

Domain model

• The space of learning
• Contents depend on what is to be taught/

learned.

• Facts
• Concepts
• Processes
• Relationships
• Strategies
• ....

Sasikumar M

29

...Domain model

• With each, we need
• Relevant content to explain it

– For different student profiles?

• Dependent units
• How to test if learned
• General misconceptions
• Common errors
• How to detect them
• How to correct them
• Usually results in a lot of hard-coded system

Sasikumar M

30

... Domain model

• As far as possible to be externalised to make it

visible to tutor model, and link with student model

• If-then rules with a generic interpreter
• Databases with a semantic layer
• Ontologic representation/concept map
• Some tradeoff in flexibility...

Sasikumar M

31

... Domain model [bugs]

• Tutoring requires understanding of correct and
• therwise behaviour
• What makes a learner take a wrong/different

path

• Misconceptions in a relation
• Wrong relationship/formula
• Buggy steps in a process
• Externalised domain knowledge needed to

capture these effectively.

Sasikumar M

32

... Domain model [process]

• Set of sub-processes
• Variability in ordering
• What is permitted when?
• All else could be “errors”.
• e.g. Subtraction of 2-digit numbers: column
• rdering
• e.g. Different investigation sequence in a

diagnosis task

Sasikumar M

33

... Student model

• Representing knowledge acquired.
• Initial model: subset of what is to be learned.
• Knowledge as “additive”
• Student knowledge could be different in

appearance from the 'desired' knowledge.

• Misconceptions, incorrect relationships,

different organisations, etc

• What is “wrong” to be corrected?!

Sasikumar M

34

... Student model

• Some sort of overlay representation
• As a diff of target knowledge
• Target knowledge to capture acceptable flexibility

(step ordering, alternatives, etc)

• Then deviations can be considered as errors.
• In concept graph
• Learned concepts can be coloured.
• Different colours for different 'level's of learning.
• How to indicate levels

– Bloom taxonomy – What can he do, vs what he should be able to do.

Sasikumar M

35

Tutor model

• Active model among the three. Need to drive a

lot of decisions.

• What do we do next?
• Go to next problem? Which problem?
• Introduce remedial resources?
• Having trouble? Should I ping?
• Going off-track? Should I tell him? How? Now?
• Relatively less understood, and most rigid so

far...

• Double-loop structure of Kurt VanLehn

Sasikumar M

36

Does it work?

• Not much in the evaluation of effectiveness...
• Some empirical studies.
• Sherlock – 20-25 hrs with the system equivalent

to experience gained in 4 years of apprenticeship.

• PUMP algebra tutor also demonstrated

significant performance improvement, even in lateral areas.