Blind equalization with recurrent neural networks using natural gradient

Neural networks have recently been used in blind equalization extensively due to their capability to discover complex decision regions. This paper introduces a novel approach to adaptive channel equalization with recurrent neural network (RNN) for a QSPK signal constellation. The proposed method utilises an FIR based natural gradient in conjunction with a scale factor to update the weights. The use of the natural gradient in RNN for weight update is two-fold: stabilizing the weights without normalization and establishing the network's capability to perform blind equalization. The work targets wireless communications in non linear channels for M-PSK and M-QAM modulation schemes. Computer simulations show that the natural gradient offers a stable training to RNN, where the weights are small in size and vary slowly with time.