Paul McGlynn gt8984a@prism.gatech.edu
Summary of "Artificial Intelligence and Tutoring Systems"
by Etienne Wenger, Chapter 2, "Basic Issues"
1. Introduction
A. point: decompose "knowledge communication system"
B. jargon
knowledge communication systems
tutoring sys
student
knowledge/expertise in some domain
CAI - computer aided instruction
presents a curriculum
static pre-stored presentation
ITS - about the same as KCS
models an active expert
contains knowledge to be communicated
can apply knowlege within domain
C. characteristics of intelligence
knowledgable
able to learn
able to adapt
able to apply knowledge
able to reason about "symbols"
D. components
domain knowledge
student model
pedagogical knowledge
interface
E. evaluation of article
purposely not specified
the learning and/or underlying pedagogical theory
the domain/topic/curriculum content
the target student population
population characteristic (novice, expert, ...)
the target setting (classroom, lab, industry, etc.)
implied to be classroom?
several possibilties described
the pedagogical goals (if any) of the system
describe typical interaction scenario
2. Domain knowledge: the object of communication
A. The functions of expert module
generate solutions to problems
generate (multiple) solution paths
explicit standard of evaluation - direct knowledge comparison
B. Aspects of communicability
domain knowledge related to pedagogy
relations between knowledge items
relative acquisition difficulty of knowledge items
explanation rationales = goals + causes
explanation = analogy or taxonomy
transparency - can student observe expert's internal functions?
psychological plausibility - does expert reason like human?
particular viewpoint
biased by knowledge representation language
choice of atomic primitives
expert & student should share or understand viewpoints
good teachers adapt to compensate for differences
3. Student model: the recipient of communication
A. difficult to model student
hard for people to understand each other
limited communication channel between human & computer
perfect model not required for reasonable pedagogical decisions
B. information
observe student behavior
interpret ("understand") actions
reconstruct student knowledge from interpretations
directly compare student knowledge to expert knowledge
high resolution comparison allows finely directed tutoring
sources of incorrect or suboptimal behavior
incomplete knowledge
incorrect versions of knowledge
model can specify incorrect knowledge, accompanied by
remedial actions
explanatory information
C. representation
expert's knowledge "language" often insufficient for student
must accomodate incorrect knowledge
construct both expert & student languages from primitives
neutral primitives
language does not specify correctness
can't account for all observed errors in advance
error primitives
observe & catalog errors first
use cataloged errors as language primitives
limits language
can use experience of experts to build catalogs
executability/runnability
student model contains all pertinent information
includes information outside of domain of expert
execute model to predict particular student's behavior
D. the diagnostic model: accounting for data
diagnosis
reconstruct student's goal structure
model student's knowledge
automated theory formation - find theory to fit data
automatic programming - program to implement behavior
"direction" of analysis
top-down - observe methods used to achieve goals
bottom-up - observe single steps
driving force behind analysis
model - modify model to fit data change
data - model constructed from data primitives
problems
model size requires huge searches
uncertainty
observable behavior is "tip of the iceberg"
>1 misconception can cause several outputs
incorrect
correct
"noise"
limitations of modelling language
students are inconsistent
students' behavior changes as they learn
type of information available for diagnosis
passive - only observe behavior and infer
active - direct interaction to resolve uncertainties
affects choice of next exercise, etc.
inferential - student does not directly aid diagnosis
interactive - student is asked to resolve uncertainties
people can't always explain themselves
adequately
correctly
limited by natural language understanding
4. Pedagogical knowledge: the skill of communication
A. didactic process
embedded & distributed in system or distinct module
interaction of specialized rules
principles subsequently interpreted into decisions
increases chance of use in several domains
descisions made by reference to student model & domain knowledge
global decisions: sequence of lessons
local decisions: intervention
when - let student search vs. interrupt
what
guidance
explanations
remediation
Pegagogy more difficult than subjects applied to
B. degrees of control - fixed or adaptive
strict monitoring - system reacts to student but keeps control
mixed initatiative - student & system share control
coaching - student controls
5. Interface: the form of communication
A. importance
interface affects understandabilty of topic
interface affects student's acceptance of system
interface capabilities may drive whole system
B. student should have accurate perception of system's capabilities
C. natural language
text understanding
text generation
voice recognition?
voice generation?
D. graphics
6. Summary & conclusion
boundaries between components are indistinct
hard-coded design vs. system flexibility