Partial-Memory Learning for Static and Changing Concepts

Marcus A. Maloof

Systems that learn with partial instance memory store some of the examples encountered in the input stream. Such learners then use these examples in future training episodes. Our approach is to select examples from the boundaries of concept descriptions, under the assumption that they map and enforce these boundaries. We have explored this idea with batch and incremental versions of the AQ rule-learning algorithm, and we have evaluated our method using static and changing concepts. We have also applied our systems to a variety of real-world domains, such as computer intrusion detection.

In this talk, I will begin with a brief introduction to foundational topics in machine learning. This will support the main part of the talk on our work on learning with partial instance memory. I will present experimental results for two problems: computer intrusion detection and the STAGGER concepts, a synthetic problem in which concepts change over time. I will conclude with directions for future research and with highlights of other research interests.

This talk is based on work with Ryszard Michalski

Slides from the talk available in PostScript (gzipped) and PDF.