Augmenting Instructable Computing with Planning Technology

Advances in human-instructable computing are contributing to a new breed of computer systems that can be taught by natural instruction rather than requiring direct programming. The current approach in the MABLE “electronic student” emphasizes the interface that maps different modes of instruction to machine learning algorithms that can learn the concepts and task knowledge being taught. While the interface provides more natural interaction with the system, there are still many constraints put on how the teacher teaches, in particular in what the teacher can assume about MABLE’s ability to compose previously learned concepts. We present a method for automatically translating MABLE’s learned task knowledge into a STRIPS planning domain, and planner-generated plans back into MABLE’s knowledge representation. In this way, existing planning technology is used to augment MABLE’s problem solving ability. This allows us to relax the requirement that the teacher explicitly teach every composite procedure and also provides a role for planning to contribute directly to learning in a more capable student.