Open Bootstrapped Learning Project from BBN Technologies

Research sponsored by DARPA.

www.bbn.com

Contribute to the Project

We're betting that some of you have great ideas about natural instruction and human-like learning for computers. Contribute your results, test curricula, or learning domains, and they will be added to the collection, so others can find out how their system measures up. Help us build the community and contribute today!

To contribute, email us at oblp@bbn.com

Mentoring a Computer

If computers could learn anything that a human mentor could teach them:

We wouldn't need to write new software every time we wanted a new capability.
Mentors wouldn't need to be computer programmers.
Computer systems could keep up with advances in technology, techniques, or tactics by learning new tricks.

The Project

The Bootstrapped Learning Project is a DARPA research program for developing ways for machines to learn from naive human teachers, through a mixture of normal human interactions such as the sharing of tips, correcting, and showing examples. Learn more.

If only I could learn how
to change a tire...

Community

Participate in the Bootstrapped Learning Community

Are you doing research in artificial intelligence, machine learning, or computer knowledge?
Are you interested in pushing the state-of-the-art about how computers can learn?
Do you want to help realize a future where regular people can teach computers in the same way that they mentor other people?

Here's your chance!

The Open Bootstrapped Learning Project is collecting curricula and domains for the empirical analysis of machines that learn from human instruction. The Bootstrapped Learning Framework provides an agent-based infrastructure for scripted teaching interactions, curricula for evaluating teachable systems, and a baseline virtual student to make it easier to interface your system with the framework.

If you are designing a teachable system, download curricula and see how your learner measures up. If you are creating tests for a teachable learner, share them with others here and help shape the future of natural instruction research.

Human Baseline Curricula

Take these tests to help verify the spectrum concept and gather baseline data on how humans perform, to measure machines against! Each takes about 10 minutes, and the data collected is anonymous.

Materials

Available Materials

cartoon of a robot playing with a soccer ball

Ready to play ball?

Framework:

Beta release (from BL Cup competition)
1.0 release (2010 release)

Curricula

Robocup Keepaway: curriculum description
- Out of Bounds (coming soon)
- Taker: Triangle bounds (coming soon)
- Keeper: Where to pass (coming soon)
- Taker: Who to guard (coming soon)
- Keeper: Where to move (coming soon)
- Taker: Guard or take? (coming soon)
- Keeper: When to pass (coming soon)

Related Publications

Spectrum Curricula for Measuring Teachability, Jacob Beal, Alice Leung, and Robert Laddaga, AAMAS 2010 workshop "Agents Learning Interactively from Human Teachers," May, 2010.
Curricula and Metrics to Investigate Human-Like Learning, Jacob Beal, Paul Robertson, and Robert Laddaga, AAAI 2009 Spring Symposium "Agents that Learn from Human Teachers", March, 2009.

Must...pull...harder!