A Unified Theory of Learning, Doing and Teaching Hierarchical Tasks:
The Learn-Do-Teach Challenge for Collaborative Agents

What does it mean to "know how to do" something, such as preparing a meal, rotating the tires on a car, or shutting down a power system?

For a human, this usually means being able not only to perform the task, but also to teach it to someone else, and often that the task was learned from someone else in the first place. For machines, however, this is generally not true. Machines have long been able to perform complex tasks based on their programming without being able either to learn such tasks from humans or teach them to humans.

The objective of this research is to develop a unified theory of learning, doing and teaching complex hierarchical tasks. The figure above shows what an agent based on a unified theory should be able to do: (a) learn a new task from a human, (b) do the task, and then teach it to (c) another copy of itself (i.e., based on the same theory) and (d) another human. For details on (e) and (f), see technical report.

Download Technical Report (PDF)