Robert N. Banks
This paper concerns the "disjunctive concept problem" - the problem of designing a system that will correctly learn disjunctive rules or concepts. The reason that this presents a problem is briefly explained, within the framework of learning from examples. A new, general-purpose solution to the problem is then described. It incorporates two novel features. Firstly it uses a lattice data structure - loosely modelled on an ATMS, which efficiently maintains all plausible solutions in parallel. Secondly, it splits the data into subsets, forming hypotheses based on the initial subset. These hypotheses are refined, or eliminated by subsequent subsets of data. This solution was implemented, and the implementation tested in one domain, that of learning rules for the diagnosis of acute abdominal pain. It was found that heuristics were needed to prune the number of hypotheses to a manageable size. These heuristics, and other issues which this analysis brings to light, are discussed.
This paper is not available online