MULLEN@SUMEX-AIM.ARPA (01/25/84)
From: Juanita Mullen <MULLEN@SUMEX-AIM.ARPA> [Reprinted from the Stanford SIGLUNCH distribution.] Friday, January 27, 1984 Chemistry Gazebo, between Physical & Organic Chemistry 12:05 SPEAKER: Tom Dietterich, HPP Stanford University TOPIC: Learning with Constraints In attempting to construct a program that can learn the semantics of UNIX commands, several shortcomings of existing AI learning techniques have been uncovered. Virtually all existing learning systems are unable to (a) perform data interpretation in a principled way, (b) form theories about systems that contain substantial amounts of state information, (c) learn from partial data, and (d) learn in a highly incremental fashion. This talk will describe these shortcomings and present techniques for overcoming them. The basic approach is to employ a vocabulary of constraints to represent partial knowledge and to apply constraint-propagation techniques to draw inferences from this partial knowledge. These techniques are being implemented in a system called, EG, whose task is to learn the semantics of 13 UNIX commands (ls, cp, mv, ln, rm, cd, pwd, chmod, umask, type, create, mkdir, rmdir) by watching "over-the-shoulder" of a teacher.