| Hypothesis-Driven Constructive Induction in AQ17-HCI: A Method and Experiments (1994) | |||||||||||||||||
Abstract | |||||||||||||||||
| A method for consreactive induction is described that generates new problem-relevant atu'ibutes by analyzing and abstracting iteratively created inductive concept hypotheses. The method, called HCI (hypothesis-driven consreactive induction), first creates a set of initial rules from the training examples using a standard AQ rule-learning algorithm. The rules generated are then analyzed and evaluated according to a rule quality criterion. The analy. sis determines which of the original attributes are irrelevant, and reduces the original representanon space. It also determines the best-performing rules for each decision class, and assembles them into sets that are assigned names, and treated as new attributes. These new attributes are then used to reformulate the training examples from the previous step, and the whole inductive process is repeated. This iterative process stops when the performance accuracy of the last generated rules exceeds a predefined performance threshold. In several experiments on learning various well-defined transformations, the AQ17-HCI system implementing the method consistently and significantly outperformed, in terms of the predictive accuracy, the AQ15 role learning system, GREEDY3 and GROVE decision | |||||||||||||||||
Publication details | |||||||||||||||||
| |||||||||||||||||