Deep Learning in Syndrome Decoding for Distributed Computer Systems

Abstract

A solution of a self-diagnosis problem for distributed computer systems consists in determining fault-free and faulty nodes of the system by the given syndrome. This problem can be reduced to the classification problem, which can be efficiently solved by deep learning algorithms. The article contains the statement and constraints of a syndrome decoding problem, the description of the developed algorithm of syndrome decoding using a convolutional neural network and the algorithm of teaching sets generation. A software implementation of the developed algorithms was fulfilled using DeepLearnToolBox of the Matlab interactive environment. We conducted experiments on the teaching sets with various number of nodes in distributed computer system and various number of faulty nodes. The hyperparameters of the convolutional neural network: length of a teaching set, number of teaching epochs, step size of a convolution kernel, number and sizes of kernels in a convolutional layer, number of convolutional neural network layers were experimentally selected. The algorithm effectiveness was evaluated by the dependence of the number of correctly detected nodes on the number of faulty nodes in a distributed computer system. Conducted experiments show that this algorithm should be used in distributed computer systems where number of faulty nodes did not exceed 30% of the system size. Despite of small length of the teaching set, the convolutional neural network keeps a good generalization ability.