Logic ecosystems meet meta-interpretivative learning: design and experiments on 2p-Kt

This thesis is rooted in the field of Inductive Logic Programming (ILP), and, in particular, Meta-Interpretative Learning (MIL). ILP is a branch of Machine Learning where the Artificial Intelligence tries to induce Horn clauses from a given background knowledge and some positive/negative examples. The goal of this thesis is the development of a system for assisting interpretative learning algorithms. In order to achieve that, we extend the 2p-Kt logic ecosystem for symbolic artificial intelligence, with meta-rules support. We then design and implement a system of pluggable components aiming to assist the various steps of ILP algorithms, such as generalization of induced rules and refinement of theories. The results are: a 2p-Kt based library of various generalization, validation and refinement strategies, a brand new algorithm inspired by Metagol (named MetaPatrol) and a test suite. The system consists of a 2p-Kt module supporting the definition of meta-rules; as well as the generalization, validation, and refinement of induced theories as first class mechanisms. Overall, the system enables the engineering of ILP solutions by combining multiple strategies for the many aspects of MIL.