Langley has done research in a wide range of areas involving machine learning, which allows him to provide perspective on his work. In this paper, he describes a learning program, SAGE, that learned how to solve simple procedural tasks. He describes the basic architecture of SAGE, by which it starts with a general set of production rules for a task and then forms more specific versions by observing when the original rules lead to dead ends. Langley then shows in detail how SAGE works on the “slide-jump puzzle” (a simple coin moving puzzle) and also describes how it was applied to algebra problems and learning seriation strategies.

This paper is interesting in that it discusses a kind of learning not often addressed in the literature--learning gradually over time from a number of different cases. Furthermore, SAGE does not expect examples specially selected to teach it a concept. Instead, it simply looks at the examples that arise while solving a problem. Most machine learning work has involved systems that are given carefully selected examples by a teacher. On the other hand, the domains used by Langley are all quite simple, and it is not clear how the methods described will scale up to more complex domains. Nonetheless, this paper will be useful in an examination of realistic learning techniques.