Machine Learning Structure for Box-plus Operation with Soft Information Processing

This paper studies machine learning (ML) structure for box-plus operation as an element of future error correction based on artificial intelligence (AI) using soft information processing. Box-plus operations are important components for check-node (CND) of error correction codes based on bipartite graphs. We propose an ML structure for box-plus operation and observe the behavior of the structure by investigating all possible structures constructed from 6 neurons in the hidden layer. We utilize activation function: (i) hyperbolic tangent (tanh) function for the hidden layer since it maps any real-valued input to the 1 to 1 range for soft information, and (ii) linear function only at the output layer. We evaluate the performances of the proposed structure using a series of computer simulations. We found that a single hidden layer with many neurons is better than multiple hidden layers with less number of neurons, because the box-plus operation is not too complex. We expect that these results open new ground for the development of future error correction codes based on ML.