| 103D6 |
|---|
| A Trial of Time Series Prediction with Bi-directional Computation Style: Enrichment of Inner Information Representations and Its Analysis |
| AUTHORS |
| Hiroshi Wakuya, Katsunori Shida (Saga University) |
| ABSTRACT |
| A bi-directional computing architecture for time series prediction proposed recently computes not only the future prediction transformation but also the past prediction one, and its idea is applied to several prediction tasks. According to the previous studies, bi-directionalization of the computing architecture makes possible to improve prediction performances of time-variant phenomena with different kinds of data sets. In spite of such experimentally-confirmed advantages, their detailed mechanism has not been clear yet. Then, in order to solve this problem, responses of the dynamic neurons, which play an important role in temporal signal processing in the bi-directional model, are investigated based on the principal component analysis approach. This attempt comes from the motive of i) an estimation of the dimensions of the inner information representations, ii) an extraction of the dominant signal components from the original dynamic neurons' behavior, and so on. As a result, an enrichment of the inner information representations and sufficiently-preserved features of the applied time series are confirmed, and this fact enables the bi-directional model to get a better score than the conventional uni-directional one. Furthermore, the development process of the bi-directional model reveals the importance of these factors through the computer simulations. |
Back to Program
Time Table
Top of This Site
Annual Conference 2002