A Class of Recurrent Neural Network (RNN) Architectures with SOM for Estimating MIMO Channels

Abstract

Artificial Neural Network (ANN)s can handle Multi Input Multi Output (MIMO) channel prediction and estimation. With suitable modification, ANNs also tackle time - varying properties of the wireless links. But these turn out to be cumbersome to configure and train for which alternative ANN architectures are required for such applications. The immediate option that emerges is the Recurrent Neural Network (RNN) which has the capacity to deal with time - dependent inputs. But a problem is observed with respect to the approach in which RNNs are trained to deal with signals with real and imaginary components. Signals with bifurcated real and complex components help the RNN to learn better. But for tightly coupled transmissions which exists most of the times in wireless channels, the performance of such RNNs suffer. The present work attempts to reduce this tradeoff and adopts a split - activation RNN training approach with exclusive blocks for in-phase and quadrature components. The responses of such blocks, are combined and optimized with Self Organizing Map (SOM). The results show better performance and ease of implementation than MLPs with temporal characteristics.