Learning Relations: Basing Top-Down Methods on Inverse Resolution

Top-down algorithms for relational learning specialize general rules until they are consistent, and are guided by heuristics of different kinds. In general, a correct solution is not guaranteed. By contrast, bottom-up methods are well formalized, usually within the framework of inverse resolution. Inverse resolution has also been used as an efficient tool for deductive reasoning, and here we prove that input refutations can be translated into inverse unit refutations. This result allows us to show that top-down learning methods can be also described by means of inverse resolution, yielding a unified theory of relational learning.