An Analysis of Task Learning by Instruction

Jim Blythe
American Assocation for Artificial Intelligence (AAAI), 2005.

Abstract

Many useful planning tasks are handled by plan execution tools, such as PRS, that expand procedure definitions and keep track of several interacting goals and tasks. Learning by instruction is a promising approach to help users modify the definitions of the procedures. However, the impact of the set of possible instructions on the performance of such systems is not well understood. We develop a framework in which instruction templates may be characterized in terms of syntactic transforms on task definitions, and use it to explore the properties of coverage, ambiguity and efficiency in the set of instructions that are understood by an implemented task learning system. We determine what kind of ambiguity is affected by the instruction set, and show how context-dependent interpretation can increase efficiency and coverage without increasing ambiguity.

PDF version.


Jim Blythe