Proceedings of 2013 RISP International Workshop on Nonlinear Circuits, Communication and Signal Processing
Volume, Number, Page
Vol. 1
No. 1
pp. 409-411
Published date
Mar. 4, 2013
Publisher
Japanese:
English:
Conference name
Japanese:
English:
2013 RISP International Workshop on Nonlinear Circuits, Communications and Signal Processing
Conference site
Japanese:
English:
Big Island, Hawaii
Abstract
Recurrent neural network (RNN) had been applied for equal- ization of nonlinear communication channel. However the error surface of RNN contains local minima, so a gradient de- scent algorithm can easily get stuck and produce sub-optimal solution. A global optimization algorithm called Differen- tial Evolution (DE) was used to train RNN for this task and shown to achieve better result than the gradient-based Real Time Recurrent Learning (RTRL).