Nonmonotonic inductive logic programming by instance patterns
Chongbing Liu, Enrico Pontelli · 2007
In this paper, we present a new approach, called NM-ILP-IP, for inductive learning in the context of nonmonotonic logic frameworks. This approach is based on the notations of concept instances and instance patterns introduced in [13]. When a strictly correct Horn theory cannot be induced, this approach induces a normal logic program, by specializing a previously learned overly-general theory. The advantages of this approach over others include: (a) it does not rely on existing ILP systems, and it avoids many of the effectiveness and efficiency drawbacks of ordinary ILP systems; (b) no theorem prover is needed during the learning process; (c) it introduces negation as failure (NAF) of existing predicates and introduces new abnormality predicates only when necessary, making the final theory more compact.