Investigation of Hypothesis-Driven Constructive Induction in AQ17-HCI

Abstract : Investigation of the AQ17-HCI method within the general purpose constructive-induction system AQ17 began with experimentation involving its supporting -induction system AQ15 and a target concept from the ROBOTS domain in the EMERALD system. With the aid of the diagrammatic visualization (OlAV) technique, AQ15 results from incremental training example sets were clearly displayed on the screen contrasting the learned concept against the target concept. AQ15 had searched successfully through the space of possible descriptions with its ultimate goal of inducing a positive output hypothesis that matched exactly the target concept. Investigation of the constructive-induction capabilities of AQ17-HCI proceeded with an analysis of output hypotheses from the attribute-removal module AQ17(HCI/ar). The analysis disclosed that AQ17 established the attributes occurring in the positive output hypothesis as relevant based on an attribute-utility measure. It discarded all other attributes as irrelevant. As a consequence, only relevant attributes appear in the negative output hypothesis which may lead to an overlap of the hypotheses: certain combinations of attribute values occurring in the positive output hypothesis may also occur in the negative output hypothesis. As part of the analysis, I applied a method described by Wnek, which I dubbed the altered positive output hypothesis (APOH) method and which uses what I call hypothesis-driven deduction (HOD). The APOH method yielded a final positive hypothesis which matched exactly the target concept without the incremental addition of training examples. My investigation led me to a comparison of the APOH method with Wnek's Vs* (version space with the STAR) method. When the methods initially encounter a coverage overlap by positive and negative rules, each method proceeds with a successful effort to output a final hypothesis fo