Direct and indirect transfer: explorations in concept formation
Joel David Martin · 1992
Rapid learning in a diverse, changing environment requires transfer of prior experience. This dissertation argues that there are two fundamentally different types of transfer that can and should be tightly integrated in machine learning systems. On the one hand, direct transfer finds and applies prior experience that is most relevant to a new instance, i.e., it does not conflict with the new instance. Any differences between this prior experience and the instance result from random variation not from consistent regularities in the world, and hence may be safely ignored. Direct transfer allows the learner to incrementally refine its knowledge about regularities in its world. Indirect transfer, on the other hand, finds and applies prior experience that is analogous to a new instance. The differences between indirectly relevant prior experience and the new instance result from important, consistent regularities and may not be ignored. However, the prior experience and the instance are analogous and share abstract properties. In our work, indirectly relevant experience and instances share statistical properties concerning the variability of attribute values. Indirect transfer using variability information helps the learner determine what aspects of a new instance reflect novel regularities and what aspects reflect random variation. These two approaches to transfer, when integrated, allow more extensive use of prior experience than does either alone. Furthermore, the integration of the methods improves both. In this dissertation, we present (a) an architecture for integrating direct and indirect transfer, (b) six general principles to guide the development of transfer systems, (c) TWILIX, a concept formation system that implements the six transfer principles, and (d) an initial set of empirical evaluations of TWILIX.