A Knowledge Level Analysis of Learning Programs

This chapter develops a taxonomy of learning methods using techniques based on Newell’s knowledge level. Two properties of each system are defined: knowledge level predictability and knowledge level learning. A system is predictable at the knowledge level if the principle of rationality can be applied to predict its behavior. A system learns at the knowledge level if its knowledge level description changes over time. These two definitions can be used to generate the three-class taxonomy. The taxonomy formalizes the intuition that there are two kinds of learning systems: systems that simply improve their efficiency (symbollevel learning; SLL) and systems that acquire new knowledge (knowledge-level learning; KLL). The implications of the taxonomy for learning research are explored. Automatic programming research can provide ideas for SLL. Development of methods for KLL must rely either on the development of a principle of plausible rationality or on the construction of learning methods that work well only for certain kinds of environments. Explanation-based generalization and chunking methods address only SLL and do not provide solutions to the problems of KLL.