g oal -o riented c onceptualization of p rocedural k nowledge martin možina, matej guid, aleksander...
TRANSCRIPT
GOAL-ORIENTED CONCEPTUALIZATION
OF PROCEDURAL KNOWLEDGEMartin Možina, Matej Guid, Aleksander Sadikov, Vida Groznik, Ivan Bratko
Artificial Intelligence LaboratoryFaculty of Computer and Information Science
University of Ljubljana, Slovenia
ITS 2012
CONCEPTUALIZATION OF PROCEDURAL KNOWLEDGE
ORIGINAL THEORY PROBLEM SOLUTION.......................................................................
axiomslawsformulasrules of the game…
path: requires excessive computation, difficult to memorize
CONCEPTUALIZEDDOMAIN THEORY
DECLARATIVE KNOWLEDGE PROCEDURAL KNOWLEDGE
WHAT? HOW?basic domain knowledge goal-oriented rules
CONCEPTUALIZATION OF PROCEDURAL KNOWLEDGE
ORIGINAL THEORY PROBLEM SOLUTION.......................................................................
axiomslawsformulasrules of the game…
path: requires excessive computation, difficult to memorize
CONCEPTUALIZEDDOMAIN THEORY
DECLARATIVE KNOWLEDGE PROCEDURAL KNOWLEDGE
WHAT? HOW?basic domain knowledge goal-oriented rules
CONCEPTUALIZATION OF PROCEDURAL KNOWLEDGE
ORIGINAL THEORY PROBLEM SOLUTION.......................................................................
axiomslawsformulasrules of the game…
path: requires excessive computation, difficult to memorize
CONCEPTUALIZEDDOMAIN THEORY
basic rules of chesspiece movementsthe 50-move rule…the “right” corner conceptbasic strategy…
procedures: IF-THEN rulessimple and compact ruleseasy to memorize…intuitive knowledgeintermediate goals…
DECLARATIVE KNOWLEDGE PROCEDURAL KNOWLEDGE
PROBLEM STATE SPACE
::
::
::
. . .
. . .
start node
goal nodes
(too) longsolution path
. . .
LEARNING INTERMEDIATE GOALS
::
::
::
. . .
. . .
goal nodes ofintermediate goals
start nodes ofintermediate goals
. . .
KNOWLEDGE ELICITATION WITH ABML
IF ... THEN ...IF ... THEN ......
ABMLargument-based machine learning
experts’ arguments
constrain learning
obtained models are consistent
with expert knowledge
Možina M. et al. Fighting Knowledge Acquisition Bottleneck with Argument Based Machine Learning. ECAI 2008.
arguments
critical examplescounter examples
experts introduce
new concepts (attributes)
human-understandable models
(suitable for teaching)
BENEFITS OF ABML FOR KNOWLEDGE ELICITATION
IF ... THEN ...IF ... THEN ......
ABMLargument-based machine learning
explain single example easier for experts to articulate knowledge
“critical” examples expert provides only relevant knowledge
“counter” examples detect deficiencies in explanations
Možina M. et al. Fighting Knowledge Acquisition Bottleneck with Argument Based Machine Learning. ECAI 2008.
arguments
critical examplescounter examples
GOAL-ORIENTED RULE LEARNING
GOAL EVALUATION:• is the goal achievable?• does the goal always lead to progress?
GOAL-ORIENTED
RULE LEARNING
Computer (to the expert):
“What goal would you suggest for white in this position?
What are the reasons for this goal to apply in this position?”
GOAL-ORIENTED RULE LEARNING: A “CRITICAL” EXAMPLE
Computer (to the expert):
“What goal would you suggest for white in this position?
What are the reasons for this goal to apply in this position?”
GOAL-ORIENTED RULE LEARNING: A “CRITICAL” EXAMPLE
The expert (a FIDE master):
“White can squeeze black king’s area. It is possible to build a barrier and squeeze the area available to the black king.”
Computer (to the expert):
“What goal would you suggest for white in this position?
What are the reasons for this goal to apply in this position?”
GOAL-ORIENTED RULE LEARNING: A “CRITICAL” EXAMPLE
The expert (a FIDE master):
“White can squeeze black king’s area. It is possible to build a barrier and squeeze the area available to the black king.”
GOAL-ORIENTED RULE LEARNING: A “COUNTER” EXAMPLE
Computer found an example where current goal “squeeze black king's area”does not lead to progress.
1.Kf5-g5: mate in 8 moves (optimal execution)1.Bg4-e2: mate in 10 moves (worst execution)
In this case, the expert found this execution of the goal to be perfectly acceptable.
HUMAN-UNDERSTANDABLE PLAY X OPTIMAL PLAY
no progress….
Computer:“Would you admonish a student if he or she played 1.Bg4-e2 in this position?”
CONCEPTUALIZATION OF DOMAIN KNOWLEDGE: CHESS ENDGAME
goal-oriented instructions
example games with goal-oriented instructions
KBNK – the most difficult of elementary chess endgames:several recorded cases when even grandmasters failed to win
the result of conceptualization: Hierarchy of (only) 11 GOALS
TEACHING MATERIALS (1): TEXTBOOK INSTRUCTIONS
TEACHING MATERIALS (2): EXAMPLE GAMES WITH GOALS
A GRANDMASTER FAILED TO WIN ...
A grandmaster of chess failed to win the following endgame…
A GRANDMASTER FAILED TO WIN ...
GM Kempinski (white) – GM Epishin (black), Bundesliga 2001
A GRANDMASTER FAILED TO WIN ...
GM Kempinski (white) – GM Epishin (black), Bundesliga 2001
A GRANDMASTER FAILED TO WIN ...
GM Kempinski (white) – GM Epishin (black), Bundesliga 2001
A GRANDMASTER FAILED TO WIN ...
GM Kempinski (white) – GM Epishin (black), Bundesliga 2001
A GRANDMASTER FAILED TO WIN ...
GM Kempinski (white) – GM Epishin (black), Bundesliga 2001
A GRANDMASTER FAILED TO WIN ...
GM Kempinski (white) – GM Epishin (black), Bundesliga 2001
A GRANDMASTER FAILED TO WIN ...
GM Kempinski (white) – GM Epishin (black), Bundesliga 2001
A GRANDMASTER FAILED TO WIN ...
GM Kempinski (white) – GM Epishin (black), Bundesliga 2001
A GRANDMASTER FAILED TO WIN ...
GM Kempinski (white) – GM Epishin (black), Bundesliga 2001
A GRANDMASTER FAILED TO WIN ...
GM Kempinski (white) – GM Epishin (black), Bundesliga 2001
A GRANDMASTER FAILED TO WIN ...
GM Kempinski (white) – GM Epishin (black), Bundesliga 2001
A GRANDMASTER FAILED TO WIN ...
GM Kempinski (white) – GM Epishin (black), Bundesliga 2001
A GRANDMASTER FAILED TO WIN ...
GM Kempinski (white) – GM Epishin (black), Bundesliga 2001
A GRANDMASTER FAILED TO WIN ...
GM Kempinski (white) – GM Epishin (black), Bundesliga 2001
A GRANDMASTER FAILED TO WIN ...
GM Kempinski (white) – GM Epishin (black), Bundesliga 2001
A GRANDMASTER FAILED TO WIN ...
GM Kempinski (white) – GM Epishin (black), Bundesliga 2001
A GRANDMASTER FAILED TO WIN ...
GM Kempinski (white) – GM Epishin (black), Bundesliga 2001
A GRANDMASTER FAILED TO WIN ...
GM Kempinski (white) – GM Epishin (black), Bundesliga 2001
A GRANDMASTER FAILED TO WIN ...
GM Kempinski (white) – GM Epishin (black), Bundesliga 2001
A GRANDMASTER FAILED TO WIN ...
GM Kempinski (white) – GM Epishin (black), Bundesliga 2001
A GRANDMASTER FAILED TO WIN ...
GM Kempinski (white) – GM Epishin (black), Bundesliga 2001
A GRANDMASTER FAILED TO WIN ...
GM Kempinski (white) – GM Epishin (black), Bundesliga 2001
A GRANDMASTER FAILED TO WIN ...
GM Kempinski (white) – GM Epishin (black), Bundesliga 2001
A GRANDMASTER FAILED TO WIN ...
GM Kempinski (white) – GM Epishin (black), Bundesliga 2001
A GRANDMASTER FAILED TO WIN ...
GM Kempinski (white) – GM Epishin (black), Bundesliga 2001
A GRANDMASTER FAILED TO WIN ...
… but why our students didn’t?
INTERMEDIATE GOAL: BUILD A BARRIER
INTERMEDIATE GOAL: BUILD A BARRIER
INTERMEDIATE GOAL: BUILD A BARRIER
INTERMEDIATE GOAL: BUILD A BARRIER
INTERMEDIATE GOAL: BUILD A BARRIER
INTERMEDIATE GOAL: BUILD A BARRIER
INTERMEDIATE GOAL: BUILD A BARRIER
INTERMEDIATE GOAL: BUILD A BARRIER
RESULTS OF A PILOT EXPERIMENT
PARTICIPANTS• three chess beginners of different
strengths
TEACHING MATERIALS• goal-oriented textbook instructions• example games with instructions
Phase II: Examination of teaching materialsparticipants were given access to teaching materials
Phase III: Playouts against optimally defending computerafter each game (but not during the games!) the materials were accessible to participants
The students learned the skill operationally in up to an hour’s time of studyingthe instructions and testing their skill in actual problem solving (playing the endgame).
Phase I: Three trial KBNK gamesparticipants were unable to deliver checkmate
GOAL-ORIENTED CONCEPTUALIZATION OF PROCEDURAL KNOWLEDGE
GOAL-ORIENTED CONCEPTUALIZATION
OF PROCEDURAL KNOWLEDGE
human assimilable simple and compact easily executable
ABML: powerful knowledge elicitation method
next step: ITS for teaching chess endgames
learning intemediate goals
+
? http://www.ailab.si/matej/
dr. Matej Guid. Artificial Intelligence Laboratory, Faculty of Computer and Information Science, University of Ljubljana. Research Page: www.ailab.si/matej