Combining HYPER and Alchemy systems for learning from noisy data

Andrej Oblak (2009) Combining HYPER and Alchemy systems for learning from noisy data. EngD thesis.

PDF Download (1376Kb)

Abstract

This thesis refers to the field of machine learning. It concentrates on the use of inductive logic programming (ILP) for learning hypotheses from noisy data. We used ILP program HYPER, which is not actually able to learn from noisy data, so for this purpose we had to modify it accordingly. We named this modified version HYPER/N. But since HYPER/N also learns hypotheses, which are a result of some random dependencies in learning data, we had to evaluate learned hypotheses with Alchemy - a system for statistical relational learning. We kept only the hypotheses with highest estimates and discarded all the remaining ones. We tested learning with HYPER/N and Alchemy on synthetic data (standard ILP learning problems and weather domain) with added noise and on real weather data. We were satisfied with the results despite some problems, which we have to solve in the future. This work is organized as follows. First we describe the basics of inductive logic programming and ILP program HYPER. Then we describe the basics of Markov logic networks, which are important for the understanding of Alchemy, the description of which follows next. We also describe results of learning with Alchemy on standard ILP learning problems. In the last chapter, we describe HYPER/N, its connection with Alchemy's weight learning and results of learning on synthetic data with added noise and also on real weather data.

Actions (login required)