Generalization and noise

Abstract

This paper describes a research project which aims at applying Machine Learning (ML) techniques to ease Knowledge Acquisition (KA) for Knowledge Based systems. Since noise in real life data has a drastic effect on ML, we examine in detail problems connected with noise. The learning system integrates two apparently distinct approaches: the numeric approach and the symbolic approach. It uses a filtering mechanism that is driven by statistical information and by comparison between several sources of knowledge (multi-expertise and experts-users “cross-examination” of input). The system also attempts to generate concepts which are resilient to noise and to improve the language of description. While it is usually thought that noise prevents using ML techniques in real applications, we attempt to show that on the contrary existing techniques can be stretched to cope with noise and to obtain better results than traditional KA techniques.