On-Line Learning Algorithms.
|
Accession number;99A0938108
|
| Title;On-Line Learning Algorithms. |
| Author;
TAJIKA ICHIRO
(Tohoku Univ.)
|
Journal Title;Record of Electrical and Communication Engineering Conversazione, Tohoku University
|
Journal Code:F0511A
|
ISSN:0385-7719
|
|
VOL.68;NO.1;PAGE.165-168(1999)
|
| Figure&Table&Reference;REF.9 |
| Pub. Country;Japan |
| Language;Japanese |
| Abstract;In this paper, we investigate some subjects concerning the mechanism of learning process based on the two kinds of learning models, the PAC learning model and the on-line prediction model. First, we provide a framework in which we use the mutual information between a target concept and a hypothesis to measure the accuracy of a hypothesis. Using this measure, we introduce a notion of mutual information gaining(MI-gaining) algorithms, and explore their relation to PAC-learning algorithms. Furthermore, we give a boosting scheme that transforms a weak MI-gaining algorithm into a strong MI-gaining algorithm. Secondly, we extend an on-line prediction model, in which an algorithm predicts a binary value by combining the predictions of several prediction strategies, by introducing the notion of the confidence parameter for prediction. That is, both the prediction algorithm and experts are required to make bets on their predictions. We give an efficient on-line prediction algorithm that achieves almost the same performance as the best expert. Thirdly, we generalize the on-line portfolio selection model, in which an algorithm invests several stocks each trading day and competes with the best investment strategy that maintains a fixed investment proportions on the whole trading days, to the one where investment strategies partition the whole trading days into several segments and switch investment proportions for each segment. We give an efficient on-line investment algorithm that acieves almost the same performance as the best investment strategy. (author abst.) |
|
|
|
Related Articles;
|
|
