Refinement Action-Based Framework For Utilization Of Softcomputing In Inductive Learning

In this thesis we describe a novel approach to application of Evolutionary Algorithms (EAs) into the field of Inductive Logic Programming (ILP). One of the main issues of ILP is the time complexity that comes out of the large size of ILP search space. Improving the search efficiency is therefore the main objective of this thesis. To reach this, EAs were used in this work as they have proven to be efficient solvers for many optimization problems in many applications before. The target of the thesis is to design a system that would use EAs to speed up the search process in ILP while also enabling to use full potential of ILP including the possibilities of first order logic as well as the refinement operators. In the same time we aim at developing a system that would not be too problem-specific and would be both user and implementation friendly. Unlike the traditional approaches that focus on evolving populations of logical clauses, our refinement-based approach uses the evolutionary optimization process to search for sequences of refinement actions that iteratively adapt the initial working clause so that it changes into a well classifying clause. Utilization of context-sensitive refinements (adaptations) helps the search operations to produce mostly syntactically correct concepts and enables using available background knowledge both for efficiently restricting the search space and for directing the search. To efficiently use the context information about the dataset we propose a specific structure to represent it and also an algorithm that automatically induces this structure from data prior to the start of the learning phase. The fact that the user is not required to manually define the language bias by writing down the semantic and syntactic restrictions speeds up the whole process of solving ILP tasks and gives our system the user friendliness that is needed for applying this one system to various ILP problems. In general, the EA-based ILP system presented in this work is more flexible, less problem-specific and the framework can be easily used with any stochastic search algorithm within the ILP domain. Experimental results on several data sets verify the usefulness of this approach. The benchmarking shows that our system is capable to achieve similar or even better results than other state-of-the-art ILP systems.