TITLE: Representing and learning search bias for ILP
PRESENTER: Mark Reid
AFFILIATION: PhD Research Student, CSE, UNSW
DATE: Friday 22nd March 2002
TIME: 12:00noon - 1.00pm
PLACE: CSE K17 1st Floor Seminar Room
The main problem my research addresses is: What can be done to tame the complexity of the search spaces encountered in inductive logic programming? In particular, I am interested in this question in the context of learning "whole theories" as opposed to learning these theories "clause-by-clause". The most common approach for constraining search in ILP has been through the use of language biases. These allow a domain expert to restrict an ILP system to evaluate only those hypotheses that are deemed to be sensible by the expert. While effective, care must be taken not to under- or over-restrict the search space with a badly chosen language bias.
Orthogonal to language biases are search biases. These describe the order in which to search a hypothesis space once it has been defined by a language bias. This talk will outline my present and planned work on an ILP system that can improve its search bias across a suite of related learning tasks. I will focus on two of the more interesting aspects of the present system: its use of a randomised best-first search strategy and a graph-based representation of logic programs and their features.
Disclaimer: This is most definitely a "work-in-progress" talk. Any results, implied or otherwise, will be purely coincidental. That said, useful suggestions and criticisms will be most appreciated.
School of Computer Science & Engineering, UNSW.