Abstract : DARPA's Bootstrapped Learning (BL) program was a research effort to build and evaluate an electronic learner (e-student) that can be instructed by a human in the style of human-mentored learning. The BAE Systems' team was responsible for evaluation, curriculum construction, and instructional mechanisms of the BL program. An independent Learning Team was responsible for the machine learning (ML) algorithms. This report presents BAE Systems' results in developing instructional materials to test the e-student, evaluating the e-student's learning results against control human subjects, and developing instructional techniques and mechanisms to teach the e-student. Overall program results demonstrate that natural instruction is a powerful and concise alternative to typical instructional input provided to current ML systems; they also point to the feasibility of constructing individual learning methods that can take advantage of multiple types of instructional input. In the future, we recommend focusing on learning and performance in a specific domain (e.g., ISR analysis) as it would (1) provide the opportunity to overcome the limitations discussed in this report, (2) stretch the bootstrapping approach to an extended training regimen (extended over both time and conceptual coverage), and (3) provide the opportunity to assess success in a concrete domain with specific success criteria.
[1]
Clayton T. Morrison.
Augmenting Instructable Computing with Planning Technology
,
2009
.
[2]
J. Curtis,et al.
Developing A Hidden Domain for Human and Electronic Students
,
2011
.
[3]
Arne Jönsson,et al.
Wizard of Oz studies: why and how
,
1993,
IUI '93.
[4]
Ciaran O'Reilly,et al.
Learning from Human Teachers: Issues and Challenges for ILP in Bootstrap Learning
,
2010
.
[5]
Daniel Bryce,et al.
MABLE: a framework for learning from natural instruction
,
2009,
AAMAS.
[6]
Arne Jönsson,et al.
Wizard of Oz studies -- why and how
,
1993,
Knowl. Based Syst..
[7]
John Mohammed,et al.
Developing an International Space Station curriculum for the Bootstrapped Learning program
,
2010,
2010 IEEE Aerospace Conference.
[8]
Sub.
Novice Human Teachers of a Virtual Toddler : A Case Study
,
2009
.
[9]
M. Gluck,et al.
Learning and Memory: From Brain to Behavior
,
2007
.
[10]
Dewayne E. Perry,et al.
Designing human benchmark experiments for testing software agents
,
2011
.