Learning Time Series by Complex-valued Neural Network with Short-Term Memory

J. Shima, M. Nishida, S. Kikuchi, H. Saito, and M. Nakanishi (Japan)


Recurrent neural network, Complex-valued neuron, Time series learning.


This paper proposes a complex-valued recurrent neu ral network with short-term memory (CRNM), which is a simple recurrent neural network with short-term memory (SRNSM) with complex values, to deal with time series learning. It is known that a complex-valued neuron can prevent learning stagnation by the inter action of the real part and the imaginary part of the complex value. We experiment three kinds of time series learning in computer simulations. Experimen tal results show that CRNM attains higher success rate and superior generalization ability to all time se ries, against Elman’s network, multilayer network us ing complex neurons with local feedback (MNCF) and SRNSM.

Important Links:

Go Back