Login| Sign Up| Help| Contact|

Patent Searching and Data


Title:
RECURRENT NEURAL NETWORK
Document Type and Number:
Japanese Patent JPH0784974
Kind Code:
A
Abstract:

PURPOSE: To provide a recurrent neural network which can generate and process nonlinear vibration of free frequency of a free waveform, specifies the details of a determining method for internal parameters as a formula, and easily generates nonlinear vibration having target physical characteristics at a low frequency of trial.

CONSTITUTION: The recurrent neural network is in two-layered hierarchical structure and each layer consists of an equal number of elements; and the neural network is equipped with a 1st layer 100 consisting of linear form neurons xi, a 2nd layer 200 consisting of nonlinear form neurons xi, a feedforward combination part 300 which multiplies the output of the 1st layer 100 by wji and inputs the result to the 2nd layer 200, a recurrent combination part 310 which multiplies the output of the 2nd layer 200 by λi and gives primary delay of a time constant τi, and then inputs the result to the 1st layer 200, and a self-combination part 320 which multiplies the output of the form neurons xi of the 1st layer 100 by μi and gives primary delay of the time constant τi, and then inputs the result to the form neurons xi again, and operates according to the specific formula.


Inventors:
YOSHIDA KAZUMI
KISHI NORIMASA
Application Number:
JP23019593A
Publication Date:
March 31, 1995
Filing Date:
September 16, 1993
Export Citation:
Click for automatic bibliography generation   Help
Assignee:
NISSAN MOTOR
International Classes:
G06F15/18; G06N3/04; G06N3/063; (IPC1-7): G06F15/18
Attorney, Agent or Firm:
Junnosuke Nakamura (1 outside)