Heuristics such as the Occam Razor's principle have played a significant role in reducing the search for solutions of a learning task, by giving preference to most compressed hypotheses. For some application domains, however, these heuristics may become too weak and lead to solutions that are irrelevant or inapplicable. This is particularly the case when hypotheses ought to conform, within the scope of a given language bias, to precise domain-dependent structures. In this paper we introduce a notion of inductive learning through constraint-driven bias that addresses this problem. Specifically, we propose a notion of learning task in which the hypothesis space, induced by its mode declaration, is further constrained by domain-specific denials, and acceptable hypotheses are (brave inductive) solutions that conform with the given domain-specific constraints. We provide an implementation of this new learning task by extending the ASPAL learning approach and leveraging on its meta-level representation of hypothesis space to compute acceptable hypotheses. We demonstrate the usefulness of this new notion of learning by applying it to two class of problems - automated revision of software system goals models and learning of stratified normal programs.
展开▼