the expert module - techomepage.cem.itesm.mx/juresti/its/diapositivas/tema 3 - the exper… ·...
TRANSCRIPT
The Expert Module
Cs5034
Material preparado por: Dr. Jorge Adolfo Ramírez Uresti
Introduction
2
Two places of intelligence in an ITS Knowledge of the subject domain
Pedagogical knowledge
“Humans cannot tutor effectively in a domain they are not expert, and there are also inarticulate experts who make terrible instructors.” – Anderson (1988)
Rev. 200811
Introduction ...
3
ITSs typically are incomplete Provide only part of the instruction
Supplemented by human teachers
Ideally should have abundance of knowledge Human expert 10 years of experience
Developing the domain knowledge is: Labour-intensive
Knowledge needs to be discovered and codified
Over 50% of effort in building an ITS
Two types of domain knowledge: Knowledge about the domain itself
Knowledge about how to be proficient in the domain
Rev. 200811
Introduction ...
4
Three options to encode knowledge Black box model
Not actually codifying human knowledge
Formulas that produce same results
Expert systems Extract knowledge from a human expert
Devising a way of codifying and applying knowledge
Applying the knowledge does not have to correspond to the way a human applies it
Cognitive models Make the expert system a simulation
Applies knowledge in the same way a human does it
Rev. 200811
Introduction ...
5
Pedagogical effectiveness
Implementation effort
more
more
Black box
models
Expert
systems
Rev. 200811
Relation of Expert Modules to Expert
Systems
6
Expert systems
(methodology defined)
Expert module
of ITS
Black box
models
Cognitive
models
Qualitative
Process
models
Experts systems
(Criterion defined)
Expert system criterion: System that achieves high-quality performance
Rev. 200811
Black Box Models
7
Generates the correct input-output behaviour over a
range of tasks in the domain
Can be used as a judge of correctness
Internal computations are
Not available to the student
Of no use when teaching
Rev. 200811
Black Box Models ...
8
Example: SOPHIE SPICE simulator
Helped in troubleshooting circuits
Not possible to explain its decisions
Example: game of chess Brute force methods
Provides good advice on a move
Cannot explain why that is a good move
Rev. 200811
Black Box Models ...
9
Black Box used in reactive tutors
Better than nothing
Expert systems can be easily converted into tutors
Rev. 200811
Black Box Models ...
10
Issue-based tutoring
Make patterns defined on the students’ and experts’
behaviour
Attach an instruction to those patterns
Example: WEST
Student no bump and Tutor bump
Tutor interrupts with an explanation of “bumping”
Rev. 200811
Black Box Models ...
11
INPUT
(e.q., Game
Board)
STUDENT
BLACK
BOXOUTPUT
(e.q., Bump)
OUTPUT
(e.q., Count)
TUTORIAL INTERVENTIONRev. 200811
Black Box Models ...
12
Issue-based tutoring...
Useful even for non Black Box models
More economical and efficient to code interventions
Not important to know details
Disadvantages
Students may think a concept should not be applied at all
Cannot explain misconceptions
Rev. 200811
Black Box Models ...
13
Surface level versus Deep Tutoring
Surface Level a) “The side-angle-side rule requires two congruent
sides and a congruent angle; you have only given
one congruent side and a congruent angle”
b) “Try to prove AB AB”
c) “To apply the side-angle-side postulate you
have to establish AB is congruent to itself. You
cannot simply assume it.”
d) “ Whenever you are trying to prove triangles
congruent it is a good idea to prove that shared
sides are congruent to themselves. This will give
you a pair of corresponding parts
Deep Level
A
C D
Rev. 200811
Glass Box Expert Systems
14
Great quantity and humanlike nature of knowledge
Knowledge acquisition is time-consuming
Useful for ITSs when the domain expert is also an expert teacher
Rev. 200811
Glass Box Expert Systems ...
15
Example: GUIDON Based on MYCIN – diagnosis of bacterial infections
450 if-then rules encode probabilistic reasoning for medical diagnosis
Uses t-rules for instruction Based on a differential between the expert’s behaviour and the
student’s behaviour
Rules are defined on the expert’s reasoning processes
Problems: Exhaustive backward search – not like humans reason
Many MYCIN rules are too complex to be taught
Rev. 200811
Glass Box Expert Systems ...
16
IF
The infection which requires therapy is meningitis
Organisms were not seen in the stain of the culture
The type of infection is bacterial
The patient does not have a head injury defect
The age of the patient is between 15 and 55 years
THEN
The organisms that might have been causing the
Infection are diplococus-pneumoniae(.75) and
neisseria-meningitidis(.74)
Typical MYCIN ruleRev. 200811
Glass Box Expert Systems ...
17
IF
The number of factors appearing in the domain
wich need to be asked by the student is zero
The number of subgoals remaining to be determined
before the domain rule can be applied is equal to 1
THEN
Say: subgoal suggestion
Discuss the (sub)goal with the student in a
goal-directed mode
Wrap up the discussion of the domain being considered
GUIDON’s tutorial rulesRev. 200811
Glass Box Expert Systems ...
18
Example: NEOMYCIN Different control structure on domain knowledge
Domain independent set of rules about how to use the domain rules
Current active set of hypotheses contained in a new data structure called differential
Designed to reflect characteristics of human short-term memory
Lesson: pay attention to the knowledge in the expert module and in the way it is deployed (same restrictions as humans)
Rev. 200811
Cognitive Models
19
Develop a simulation of human problem solving in a domain Knowledge is decomposed into meaningful, humanlike
components
Knowledge is deployed in a humanlike manner
Problems: Develop is time-consuming and constrained
Running a cognitive model may be slow
Decide which psychological components are essential for tutoring
Rev. 200811
Cognitive Models ...
20
Types of knowledge to tutor: Procedural
Knowledge about how to perform a task
Example: calculus, algebra
Declarative Set of facts appropriately organized to reason with them
More general and not specialized for particular use
Example: geography
Causal Allows to reason about the behaviour of a device
Example: troubleshooting
Rev. 200811
Procedural Knowledge
21
Usually a rule-based system
Production system
If-then rules matched to working memory of facts
Working memory -> short-term human memory limitations
Recognize-act cycle -> basic data-driven character of human cognition
Rev. 200811
Procedural Knowledge ...
22
Sub{} SatisfactionCondition:TRUE
L1: {}--> (ColSequence RightmostTOPcell
RightmostBottomCell RightmostAnswerCell)
COLSEQUENCE (TC BC AC) Satisfaction Condition: (Blank? (Next TC))
L2: {}--> (SubCol TC BC AC)
L3: {}--> (ColSequence (Next TC)(Next BC) Next AC))
SubCol (TC BC AC) Satisfaction Condition: (NOT (Blank? AC))
L4: {(Blank? BC)}--> (WriteAns TC AC)
L5: {(Less? TC BC)}--> (Borrow TC )
L6: {}--> (Diff TC BC AC)
Multiple column substraction skillRev. 200811
Procedural Knowledge ...
23
Use of rule-based representations
Students make errors when trying to repair their procedures at impasses created by missing rules
Eliminating rules -> predicts human errors
Each rule is an independent piece of knowledge
Loss of rules corresponds to human errors
Each rule can be communicated to a student independent of total problem structure
Rules can be used to represent the student’s knowledge state – set of production rules.
Rev. 200811
Procedural Knowledge ...
24
Model tracing Observe student’s surface behaviour
Try to match student’s actions to rules firing on a rule-based system
Continue interaction based on the SM generated from the tracing
Uses: Provide immediate feedback on errors
Interrogate students about their intentions
Rev. 200811
Declarative Knowledge
25
Use when student must: Understand basic principles and facts of a domain
Reason with these generally – knows how to justify his actions
Not concerned student becomes facile at any one application of the knowledge
Knowledge base is separate from inference procedures –knowledge and control are separated
Can be combined with procedural knowledge Knowledge and control are together
Rev. 200811
Declarative Knowledge ...
26
Example: SCHOLAR
Goal: to communicate information about South American geography
Used a semantic net representation
“Close to internal knowledge structure of humans”
Nodes
Representing various concepts
Linked by various relationships – define fundamental inference processes
Rev. 200811
27
STATE
-- -- -- -- --
LATITUDE
-- -- -- -- --
-- -- -- -- --
-- -- -- -- --
CONTINENT
-- -- -- -- --
-- -- -- -- --
-- -- -- -- --
COUNTRY
(SUPERC (State Independent))
(SUPERP Continent)
----------------------------------------
Examples Argentina
Bolivia Brazil ..................
Uruguay U.S. Venezuela)
URUGUAY
(SUPERC Country)
-- -- -- -- --
-- -- -- -- --
SOUTHAMERICA(SUPERC Continent)
(Countries Argentina
-------------------------
Uruguay Venezuela) Rev. 200811
Declarative Knowledge ...
28
Schema representation
Contains slots
Knowledge is entered into slots
Erroneous knowledge creates “bugs”
Rev. 200811
Declarative Knowledge ...
29
Evaporation
Actors
Source: Large-body-of-water
Destination: Air mass
Factors
Temperature(Source)
Temperature(Destination)
Proximity(Source, Destination)
Functional-relationship
Positive(Temperature(Source))
Positive(Temperature(Destination))
Result
Increase(Humidity(Destination))
Rev. 200811
Declarative Knowledge ...
30
Schema representation ...
Student is supposed to have the general inference
procedures to reason about knowledge
Task: represent knowledge in a form that these procedures can be
invoked by student
Problems
Rule “If the student can use knowledge, he knows it” does
not apply
Schema must “know” how student makes inferences
Rev. 200811
Declarative Knowledge ...
31
Socratic dialogue
To understand inferences ITS needs to understand dialogues
Give the control to the tutor
Asks questions
Instruction and assessment is done through these questions
Tutor probes the student’s ability to apply knowledge
Rev. 200811
32
WHY: Do you think AMAZON jungle has heavy rainfall or light rainfall?
(Ask for a prediction about a particular case)
STUDENT: Heavy rainfall
WHY: Why does the Amazon have heavy rainfall?
(Ask for prior cause)
STUDENT: Because it´s near the Atlantic
WHY: Yes, the Amazon jungle is near the equatorial Atlantic. How does
that affect rainfall in the Amazon jungle?
(Ask for intermediate cause)
STUDENT: The water in the current is warm.
•
•
•
WHY: Do you think that any place with mountains has heavy rainfall?
(Form a general rule for insufficient cause)
STUDENT: Yes
WHY: Southern California has mountains. Why doesn t Southern California have
heavy rainfall?
(Pick a counter-example for an insuffient cause)
Rev. 200811
Declarative Knowledge ...
33
IF
The student gives an explanation of one or more
Factors that are not sufficient
THEN
Formulate a general rule for asserting that the factors
given are sufficient
Ask the student if the rule is true
Rule for implementing Socratic method
Rev. 200811
Declarative Knowledge ...
34
Socratic dialogue...
Conditions of rules refer to underlying knowledge rather than
surface behaviour
Rules involve a mix of knowledge, assessment and instruction
Disadvantage: natural language
Rev. 200811
Qualitative Process Models
35
Knowledge that underlies our ability to mentally simulate and
reason about a dynamic process
Troubleshooting behaviour – reasoning through the causal structure of a
device
Reasoning about the causal structure of the world
Usually uses equations to model a device
Reasoning about the device involves tracing the constraints among the
equations
Rev. 200811
Qualitative Process Models ...
36
Used when the goal is to use knowledge for
troubleshooting
Like black boxes but can explain how it reasoned by using
formulas
Rev. 200811
Example: LECOBA
37
Domain: Binary Boolean Algebra (Basic)
Cognitive model – procedural knowledge Students learn how to simplify Boolean expresions
In order to learn, students must: Read material
See examples of problems being solved by an expert
Solve exercises
Get feedback Well done!
Tell them when they have done something different from an expert solution
Rev. 200811
Example: LECOBA…
38
A program for simplifying Boolean Expressions was developed
Allowed to simplify in several modes
Novice to expert
Returned a path to see how knowledge was applied (rules)
Allowed to apply knowledge in exactly the same way a student had done it (compare or simulate a student)
The “simplifier” allowed to assess student’s work and to demostrate expert behaviour
Rev. 200811
Example: LECOBA …
39
Simplifier was unable to explain its knowledge
Used Model Tracing
Knowledge about the domain was coded in several texts linked to the current topic Introduction
Operations with 0 & 1
Basic Theorems
Introduction to heuristics
Basic Laws
Complex heuristics
Rev. 200811
Example: LECOBA …
40
Topics explained by the tutor were very dull
Student had to read all the material
Rev. 200811
Example: LECOBA …
41
Exercises and problems were pre-defined and clasiffied in
several groups
Tutor only chose from a set of exercises and decided if
they were to be used as an example or as a problem
Rev. 200811
Example: LECOBA …
42
Tutor’s and Companion’s dialogs were hardcoded
Decision on how to talk to the student were related to
his global knowledge level
Rev. 200811
Example: LECOBA …
43
Tutor and Companion knowledge to give a justification
was pre-defined in text files
Rev. 200811