Machine Learning In Programming by Demonstration: Lessons Learned from CIMA

Acquisition, Learning, and Demonstration: Automating Tasks for Users

Volume

Papers from the 1996 AAAI Spring Symposium

Contents

An agent that learns tasks from the user faces several problems: learning executable procedures; identifying relevant features of data in complex environments; filtering out irrelevant user actions; keeping the user in control without miring the user in the details of programming; and utilizing all forms of instruction the user might give including examples, ambiguous hints and partial specifications. This paper discusses the design and preliminary implementation of a system that aims toward these goals. Several lessons for system designers arise from this experience: the quality of user interaction is more important than the power of inference, though plausible inferences surely improve user interaction; involving end-users in all phases of design is critical to justifying design decisions; micro-theories of users, actions and data can take the place of domain theory in domains that have no formal theory; and, finally, a conventional concept (classification) learner can be adapted to utilize micro-theories to learn deterministic, operational action descriptions from examples, hints and partial specifications.