Article ID | Journal | Published Year | Pages | File Type |
---|---|---|---|---|
406654 | Neural Networks | 2012 | 6 Pages |
Abstract
Introducing a forgetting factor allows a support vector machine to solve time-varying problems adaptively. However, the exponential forgetting factor proposed in an earlier work does not ensure convergence of average generalization error even for a simple linearly separable problem. To guarantee convergence, we propose a factorial forgetting factor which decays factorially over time. We approximately derive the average generalization error of the factorial forgetting factor as well as that of the exponential forgetting factor using a simple one-dimensional problem, and confirm our theory by computer simulations. Finally, we show that our theory can be extended to arbitrary types of forgetting factors for simple linearly separable cases.
Keywords
Related Topics
Physical Sciences and Engineering
Computer Science
Artificial Intelligence
Authors
Hiroyuki Funaya, Kazushi Ikeda,