Novel Shuffle Algorithm for Efficient Training of Neuron Network Digital Predistortion

This paper presents a shuffle algorithm to optimize the training sequence for the stochastic and the mini-batch training of neuron network (NN) digital predistortion (DPD) technique. The proposed training sequence can ensure an efficient convergence by reducing the correlation of continuous inputs. Experiments are validated through an augmented real-valued time-delay neural network (ARVTDNN). Experiment results show that the proposed training sequence can significantly improve the convergence speeds and the linearization performance.