公开(公告)号：JPH07152714A

公开(公告)日：1995-06-16

申请号：JP18371794

申请日：1994-08-04

Applicant: KOREA ELECTRONICS TELECOMM

Inventor： SAI CHINEI ,  KICHI TOSHIKAGE

IPC: G06G7/60 ,  G06F15/18 ,  G06N3/04

Abstract: PURPOSE: To provide a regression model capable of preventing rapid performance decline as the order of supplied data becomes higher, obtaining a fast learning speed and obtaining desired performance without falling into a local minimum point. CONSTITUTION: The regression model is constituted of a self dividing network 10 and a feed forward mapping network 20, the self dividing network 10 divides an input space into non-overlapping local areas, and in the meantime, in response to the output of the self dividing network 10, the feed forward mapping network 20 obtains partial linear maps for the respective divided input spaces. Thus, compared to a local averaging method, excellent performance is obtained in a less medium variable scale and a local minimum point problem is solved since the network is automatically constituted so as to obtain the desired performance.

公开(公告)号：JPH07152717A

公开(公告)日：1995-06-16

申请号：JP18860694

申请日：1994-08-10

Applicant: KOREA ELECTRONICS TELECOMM

Inventor： SAI CHINEI ,  KICHI TOSHIKAGE

IPC: G06G7/60 ,  G06F15/18 ,  G06N3/08

Abstract: PURPOSE: To be practical from the side of a learning data number and an element function by applying a linearity learning method between an output layer and a hidden layer and a nonlinearity learning method between the hidden layer and input and estimating a network variable. CONSTITUTION: In the stage S1 of learning composed of the input layer, the hidden layer and the output layer, a new input learning pattern is introduced to a nonlinearity estimation network, the output of the nonlinearity estimation network is obtained in the stage S2 and the error is obtained. Then, whether or not the error is larger than a prescribed error reference value is discriminated in the stage 3. In the case that the error is not larger than the reference value, as the stage S5, whether or not to input the entire learning data to the nonlinearity estimation network is discriminated and the mean square error of the nonlinear estimation network is obtained in the case of inputting them. Then, as the stage S6, whether or not the mean square error is larger than a prescribed stipulated error is discriminated, and in the case that the mean square error is larger than the stipulated error, the error reference value is reduced for a prescribed error reduction value and a first stage is returned.