developing a hidden domain for human and electronic students
TRANSCRIPT
Developing A Hidden Domain for Human and Electronic Students Presented at IJCAI ALIHT Workshop July 2011
Jeremy Ludwig, Alex Davis Stottler Henke Associates
San Mateo, CA
{ludwig, davis} @stottlerhenke.com
Mark Abrams, Jon Curtis Cycorp
Austin, TX
{mabrams, jonc} @cyc.com
DISTRIBUTION STATEMENT A: Approved for Public Release, Distribution Unlimited
Bootstrap Learning Objective: Develop an electronic student (e-student) to learn from focused human-style instruction. Example: Drawing the letter “p”
Bootstrap Learning Framework
!Figure 1: The BLADE Framework
2.2 Human Learning and Teaching
"#$%!#&'!()*+,!-.),&%/0+/&1!+,!%2'&)3%,&!+'')*%'!#&'!#,'&/)-&./!0.''%''%'!+44!/%4%5+,&!-+0+6#4#&#%'1!+,3!#&'!7.+4!#'!&.!4%+/,!)'#,7!&(%!'+*%!#,'&/)-&#.,!*%&(.3'!)'%3!6%&8%%,!()*+,'9!:'!0+/&!.;!&(%!<5+4)+&#.,!=%+*1!8%!+/%!,.&!+44.8%3!+--%''!&.!%2'&)3%,&!#*04%*%,&+&#.,!3%&+#4'!&.!%,')/%!.)/!6%,-(*+/$'!+/%!),6#+'%39!>%!$,.8!&(+&!&(%! 4%+/,%/'! +/%! 3%'#7,%3! &.! 6%! 3.*+#,2#,3%0%,3%,&1! +,3! &(+&! &(%?! +/%! '0%-#+4#@%3! &.! 0+/&#-)4+/! A+&)/+4!B,'&/)-&#.,!C%&(.3'!DABC'E!/+&(%/!&(+,!0+/&#-)4+/!0/.64%*!3.*+#,'9!=(%! +/%+! .;! -.*0)&%/! &)&./#,7! -+,! 6%! '%%,! +'! +,! #,5%/'%! 0/.64%*! &.! 8(+&! 8%! +/%! #,5%'+&#,79! B,!
0+/&#-)4+/1! &(%! +/%+! .;! &%+-(+64%! +7%,&'! 6%+/'! '.*%! ')/;+-%! '#*#4+/#&?! &.!F"9! B,! &(#'! ;#%431! ()*+,! '&)3%,&'!&%+-(! 4%+/,#,7! +7%,&'! #,!./3%/! &.! #*0/.5%! &(%#/! .8,!),3%/'&+,3#,7!.;! -.,-%0&'! DG"%+/,#,7!6?!=%+-(#,7HE9!I,%!%J+*04%!#'!&(%!GF%&&?K'!F/+#,H!'?'&%*!LM+5#'!%&!+491!NOOPQ9!R.8%5%/1!#,!&(%'%!'?'&%*'!&(%!#*0./&+,-%!#'!04+-%3!.,!(.8!8%44!&(%!()*+,!#,'&/)-&./!4%+/,'1!,.&!.,!&(%!-+0+6#4#&#%'!.;!&(%!4%+/,#,7!+7%,&9!
2.3 Human Case Studies
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
3. PHASE I STUDY: BLOCKSWORLD
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
&(/%%!64.-$'!+,3!&(%,!&.!6)#43!+!'#*04%!G3../8+?H!.)&!.;!64.-$'!#,!&(#'!%,5#/.,*%,&9!=(%!&+/7%&!8+'!+!'#*04%!'&/)-&)/%!.;!&8.!'&+-$'!.;!64.-$'!&.00%3!8#&(!+!4#,&%49!>%!)'%3!+!>#@+/3!.;!I@!D>I@E!LM+(46+-$!%&!+491!UVVPQ!'&?4%!*%&(.3.4.7?1!8(%/%!&(%!&%+-(%/'K!,+&)/+424+,7)+7%!#,'&/)-&#.,'!8%/%!&/+,'4+&%3!#,&.!0/%-#'%!&%/*'!DB"E!;./!&(%!%2'&)3%,&!6?!+!()*+,!#,&%/0/%&%/9!!
From: Robert D. Grant, David DeAngelis, Dan Luu, Dewayne E. Perry and Kathy Ryall. TOWARDS EVALUATING HUMAN-INSTRUCTABLE SOFTWARE AGENTS, IHCI 2011, Rome, July 2011
Role of Curriculum 1. Instruct the e-student 2. Test the e-student Advantages of a defined set of curricula: • Algorithm development • E-student testing • Developing a repository
HD Requirements 1. Open to learning from experience 2. Need to learn from experience to succeed 3. Explicit reasoning 4. Elements easily visible to and encoded by an
automated system 5. Problems come up reasonably often in real
life
Role of Human Comparison Curriculum is necessary and sufficient Provide baseline for expected e-student
performance
Human Comparison Requirements 6. Include problems that will challenge the
human students. 7. But not too challenging 8. Same problems will be presented to human
students and the e-student.
Ground Station Domain
Student fills the role of ground station operator, controlling satellite communication
Laddered Curriculum Each rung in the ladder: • Is single a concept (e.g. procedure, pattern) • Contains multiple lessons by telling,
example, or feedback
Rungs build on top of one another to form a ladder Units are used to group related rungs
Phase II Laddered Curriculum
Abnormal Component
Fault 1 Fault 2 Fault 3 Fault 4 Fault 5 Fault 6
Unit 1 Abnormal
Unit 2 Identify
Fault 1 Fault 2 Fault 3 Fault 4 Fault 5 Fault 6 Unit 3 Repair
Fault 1 Fault 2 Fault 3 Fault 4 Fault 5 Fault 6 Unit 4 Fault -> Repair
Diagnose and Repair
Unit 5 Procedure
Example e-student Trace Lesson: HD-Abnormal-CBE-LearningLesson [ Utter( name=Utter-3574, timestamp=Timestamp(system=lesson,offset=379), utterance=LessonTeaches(name=LessonTeaches-3573,arg1=WhenTrue(name=WhenTrue-3572,arg1=Abnormal)), addressee=Student(name=theStudent)) ] [ Utter( name=Utter-3591, timestamp=Timestamp(system=lesson,offset=574),utterance=Method(name=Method-3590,arg1=byExample), addressee=Student(name=theStudent)) ] [ Imperative( name=Imperative-3635, timestamp=Timestamp(system=lesson,offset=808), request=startSimulation(name=startSimulation-3634,modifiers=[0],faults=["Antenna_Motion_Error"]), addressee=World(name=theWorld)) ] [ WorldDone( name=WorldDone-3644, timestamp=Timestamp(system=lesson,offset=1952),imperative=Imperative-3635,source=World(name=theWorld), addressee=Teacher(name=theTeacher)) ] [ Perception( name=Perception-3646, timestamp=Timestamp(system=lesson,offset=2025), gainedPercepts=[ ComponentConnection(
… for about 70 more pages
Phase III Laddered Curriculum
Abnormal Component
Fault 1 Fault 2 Fault 3 Fault 4 Fault 5 Fault 6
Unit 1 Abnormal
Unit 2 Identify
Intrack Azimuth Fault 3 Fault 4 Fault 5 Fault 6 Unit 3 Repair
Fault 1 Fault 2 Fault 3 Fault 4 Fault 5 Fault 6 Unit 4 Fault -> Repair
Diagnose and Repair
Unit 5 Procedure
Task Change Unit 6 Triggers Indicator Change
Perform Diagnose and Repair
Unit 7 Monitor
Update Perform Diagnose and Repair
Monitor Bias
Performance Results
Results are not intended for direct comparison: • Does not include all units • Does not include injection penalties
Future Work Significant future work exists in the challenges
contained in this curriculum that were tested on humans but not addressed by the e-student:
• Complicated repair procedures • Analogous repair procedures • Context-based monitoring • Concept revision
Acknowledgements We want to thank our colleagues at BAE Systems, Bester Tracking Systems, Cycorp, Stottler Henke Associates, and University of Texas at Austin for their significant input throughout the curriculum development process: Manfred Bester, David DeAngelis, Robert Grant, Dan Luu, John Mohammed, Dewayne Perry, Benjamin Rode, Howard Reubenstein, Kathy Ryall, and Candy Sidner, This material is based upon work supported by the United States Air Force Research Laboratory (AFRL) under Contract No. FA8650-07-C-7722. The views expressed are those of the author and do not reflect the official policy or position of the Department of Defense or the U.S. Government.