Although mechanized induction is relevant for science and technology, standard methods, such as similarity-based learning, are seriously limited for learning hard concepts from poorly understood domains. Several important domains, such as protein folding, would benefit from more advanced treatment of data and partial knowledge to construct new features that help similarity- based learning. The goal of this project is to improve methods for feature construction, using increased interaction between computer and user to take advantage of the strengths of each. Two thrusts of this project are more dynamic construction and more interactive refinement. For dynamic construction, the problem of constructing representations is decomposed into several subproblems for forming and selecting components for new features. Components are dynamically determined by information contained in automated analysis of data and knowledge. For interactive refinement, the approach allows partial specification of "hunches", which can represent domain-dependent or independent heuristics. These fragments of knowledge can be updated by the learning systems and re-examined by the user.
