Thesis Topic Details

Topic ID:
535
Title:
Inductive Logic Programming and Minimum Message Length Inference
Supervisor:
Mike Bain
Research Area:
Associated Staff
Assessor:
Claude Sammut
Topic Details
Status:
Active
Type:
Research
Programs:
CS CE BIOM BINF SE
Group Suitable:
No
Industrial:
Pre-requisites:
DN average or better
Description:
The Minimum Message Length (MML) principle is a universal principle
applicable in machine learning, statistics, econometrics and data
mining. It is Bayesian (meaning that it takes advantage of available
prior knowledge) and firmly based in information theory. In its most
general form, MML can in principle be used to infer any arbitrarily
complex computable function that might underly a body of data. It can be
thought of as a quantitative form of Ockham's razor.

In this work, we consider modelling data using logic programs, which
gives us a rather rich language of which decision trees and graphs are a
special case. We re-visit earlier work of Muggleton, Srinivasan and
Bain (1992) which used Minimum Description Length (MDL), a related
subsequent idea to MML. Several pairs of papers (Quinlan and Rivest,
1989; Wallace and Patrick, 1993) (Kearns, Mansour and Ng, 1997;
Viswanathan, Wallace, Dowe and Korb, 1999) show the advantages of
refining MML coding schemes. The benefits of faster machines should
also be a substantial bonus in re-visiting this work.

The project could be divided into at least two parts. One part of the
project is concerned with optimal MML encoding - this will entail both
optimal encoding of logic programs and optimal encoding of the data
given the relevant logic programs. Another part of the project will
concern, given a body of data, searching through the space of logic
programs to find the one which minimises the length of the two-part MML
message. Possible search techniques include random coding (Wallace),
simulated annealing, genetic algorithms, etc.

Time permitting, possible extensions to the project include inference of
constraint logic programs.
Comments:
--
Past Student Reports
 
No Reports Available. Contact the supervisor for more information.

Check out all available reports in the CSE Thesis Report Library.

NOTE: only current CSE students can login to view and select reports to download.