Inductive logic programming using bounded hypothesis space

Abstract

Inductive Logic Programming (ILP) systems apply inductive learning to an inductive learning task by deriving a hypothesis which explains the given examples. Applying ILP systems to real applications poses many challenges as they require large search space, noise is present in the learning task, and in domains such as software engineering hypotheses are required to satisfy domain specific syntactic constraints. ILP systems use language biases to define the hypothesis space, and learning can be seen as a search within the defined hypothesis space. Past systems apply search heuristics to traverse across a large hypothesis space. This is unsuitable for systems implemented using Answer Set Programming (ASP), for which scalability is a constraint as the hypothesis space will need to be grounded by the ASP solver prior to solving the learning task, making them unable to solve large learning tasks. This work explores how to learn using bounded hypothesis spaces and iterative refinement. Hypotheses that explain all examples are learnt by refining smaller partial hypotheses. This improves the scalability of ASP based systems as the learning task is split into multiple smaller manageable refinement tasks. The thesis presents how syntactic integrity constraints on the hypothesis space can be used to strengthen hypothesis selection criteria, removing hypotheses with undesirable structure. The notion of constraint-driven bias is introduced, where hypotheses are required to be acceptable with respect to the given meta-level integrity constraints. Building upon the ILP system ASPAL, the system RASPAL which learns through iterative hypothesis refinement is implemented. RASPAL's algorithm is proven, under certain assumptions, to be complete and consistent. Both systems have been applied to a case study in learning user's behaviours from data collected from their mobile usage. This demonstrates their capability for learning with noise, and the difference in their efficiency. Constraint-driven bias has been implemented for both systems, and applied to a task in specification revision, and in learning stratified programs.