Abstract:
A common approach to machine learning aims at the synthesis of object-level
macro-operators representing sequences of actions. We argue that a greater
generality could be achieved by learning instead meta-level operators
representing sequences of partially defined (or generic) inference steps
that lead to the
discovery of object-level operators or concepts. The implementation
of these ideas relies on a hierarchy of formalized contexts. Inference
steps are
represented as deduction traces that can be both derived from and forced
back into a new kind of reflective contexts. While learning follows from
hopping up
a level in this hierarchy, reuse implies hopping down again. We illustrate
this approach with an example showing how to learn and apply an instance
of a
rule involving the compilation of heterogeneous knowledge.
In: Proceedings of the International and Interdisciplinary Conference
on Modeling and Using Context (CONTEXT-97), Rio de Janeiro, Brazil, February
4-6, Federal University of Rio de Janeiro Ed., pp. 177-188.
Back to CONTEXT-97 Program