Incremental learning with partial instance memory

Marcus A. Maloof and Ryszard S. Michalski

Agents that learn on-line with partial instance memory reserve some of the previously encountered examples for use in future training episodes. We extend our previous work by combining our method for selecting extreme examples with two incremental learning algorithms, AQ11 and GEM. Using these new systems, AQ11-PM and GEM-PM, and the task computer intrusion detection, we conducted a lesion study to analyze trade-offs in performance. Results showed that, although our partial-memory model decreased predictive accuracy by 2%, it also decreased memory requirements by 75%, learning time by 75%, and in some cases, concept complexity by 10%, an outcome consistent with earlier results using our partial-memory method and batch learning.

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

Maloof, M.A., & Michalski, R.S. (2002). Incremental learning with partial instance memory. Foundations of intelligent systems, Lecture Notes in Artificial Intelligence, Vol. 2366, 16–27. Berlin: Springer-Verlag. (Proceedings of the Thirteenth International Symposium on Methodologies for Intelligent Systems, Lyon, France, June 27–29)

@incollection{maloof.lnai.02,
  author = "Maloof, M.A. and Michalski, R.S.",
  title = "Incremental learning with partial instance memory",
  booktitle = "Foundations of intelligent systems",
  series = "Lecture Notes in Artificial Intelligence",
  volume = 2366,
  year = 2002,
  pages = "16--27",
  publisher = "Springer-Verlag",
  address = "Berlin",
  note = "Proceedings of the Thirteenth International Symposium on
    Methodologies for Intelligent Systems, Lyon, France, June 27--29"
}