Differential-Evolution-based Weights Fine Tuning Mechanism for GRU to Predict 5G Traffic Flow

Abstract

Gated recurrent unit (GRU) neural network has been widely used in mobile communication traffic prediction. This is because that GRU can obtain acceptable prediction accuracy with lower execution costs. However, the accuracy of the GRU still depends on the interaction between the backpropagation (BP) and the activation function and the optimizer that it will lead to a chance of overfitting because the weights of neurons are too concentrated on certain features. This study uses differential evolution (DE) to fine tune the weights in GRU. The simulation results show that using of our proposed method can obtain higher accuracy rate than the original GRU through the global search of DE.