2002 Fiscal Year Final Research Report Summary
A method for constructing learning networks containing linguistic knowledge
| Project/Area Number |
13650448
|
|---|
| Research Category |
Grant-in-Aid for Scientific Research (C)
|
| Allocation Type | Single-year Grants |
|---|
| Section | 一般 |
|---|
| Research Field |
System engineering
|
|---|
| Research Institution | Kyushu University |
|---|
Principal Investigator |
JIN Chunzhi Kyushu University, Faculty of Information Science and Electrical Engineering, Professor, システム情報科学研究院, 助手 (90274555)
|
|---|
| Co-Investigator(Kenkyū-buntansha) |
WADA Kiyoshi Kyushu University, Faculty of Information Science and Electrical Engineering, Professor, システム情報科学研究院, 教授 (60125127)
|
|---|
| Project Period (FY) |
2001 – 2002
|
| Keywords | Learning Networks / Noise / Linguistic Knowledge / Dynamic Systems |
|---|
| Research Abstract |
We can not avoid the noise problem in system identification and control. So, knowing of the noise characteristics and investigation on propagation features of the noisy signals through networks are needed when identifying or controlling systems by network manners. Also, a network construeting method that make good use of human knowledge and experiments is necessary, so that construct a suitable network well matching the identification and control specifications from various networks containing various activation functions and various connections.
In this research, the following studies are carried out for system identification and control using learning networks :
1) Developing RBP network and its constructing method : a new type of network named RBP (Radial Basis Function-Perceptron) network that combines RBF network and Perceptron network, and its constructing method are presented. RBP network has both advantages of RBF network and Perceptron network, the learning speed is fast, and the generalization ability is good.
2) Analysis of propagation characteristics of probabilistic signal through networks : probabilistic characteristics of the noise signal through the network are expressed by each order momentum, the propagation characteristics are investigated, and a calculation method of each order momentum on any node is presented.
3) Estimation method for noise covariance : covariance is one of most important indicator to evaluate the noise. We propose a new estimation method by introducing forward and backward auxiliary stochastic quantities to improve numerical feature.
|
