EVALUATION OF LOSS FUNCTION OF RADIO MODULATION TRAINING DATASET IN AUTOMATIC MODULATION CLASSIFICATION FOR IN-DEPTH ARCHITECTURE NEURAL NETWORKS MODEL

Abstract

In automatic modulation classification (AMC), loss function plays an important role to determine the errors between the outputs of training neural networks for every single training and given target value. Loss function is crucial not only for attaining the objective of minimising loss but also for model parameter estimation and initial assessment phases of the model since its value contributes considerably to the sum of the cost function. In this study, the loss function will be evaluated based on different tuned models of neural networks and how it may affect the performance of the accuracy model by using a combination of deep neural networks (DNNs) of convolutional neural networks (CNNs) and gated recurrent units (GRUs). The results show that by changing the parameters of filter size and number of filters per layer network, the loss function is moderately reduced to an average loss of 1.3 for all four tests carried out during training the DNNs. Overall accuracy performance of 70 - 76% was achieved by utilising a low-specification and low-performance graphics processing unit (GPU) designed specifically for AMD Radeon R5 M330 graphics engine for network model training hardware resources. From the results, the performance accuracy of 76% was achieved by using batch size of 2,000, filter sizes of 1 x 7 and 1 x 8, and 50, 100 and 220 filters per layer in the hierarchy layer of the neural network.
Keywords: Automatic modulation classification (AMC); end-to-end learning; deep learning; deep neural networks (DNNs); loss function.

Description