Modified binary PSO training of recurrent neural network for 1/n rate convolutional decoders

In this paper, a new approach based on combination of modified binary particle swarm optimization (MBPSO) and recurrent neural network (RNN) for decoding of 1/n rate convolutional codes has been developed. For reducing complexity of Viterbi Algorithm (VA) some novel pseudo optimum iterative algorithms based on minimization of noise energy function (NEF) have been proposed however all of them may be trapped in local minima and require high decision delay which is impossible for practical implementation. It is shown in this paper that hybrid gradient descent method and MBPSO can perform better bit error rate (BER) and low latency in decoding procedure compared to VA performance. This results demonstrate that modified version of binary PSO guarantees fast convergence and reaches to global optimum.