Optimization of a neural decoder for short BCH codes under critical communication regime

Abstract

In the present paper, a decoding strategy is introduced in the context of error-correcting codes where a neural network is trained to predict error patterns using simultaneously the information from the modules and the syndromes of the received vectors. In the proposed decoder, the most reliable positions are iteratively selected to be the erroneous bits of the estimated error pattern, so these are later subtracted from the received vector before a new decoding is performed. For the prediction of the error pattern, a deep neural network with reduced complexity is designed. The experiments carried out for short BCH codes transmitting via the AWGN channel show that the performances obtained with this decoding strategy surpass those obtained exclusively with the neural network.