EVM Loss: A Loss Function for Training Neural Networks in Communication Systems

Neural networks and their application in communication systems are receiving growing attention from both academia and industry. The authors note that there is a disconnect between the typical objective functions of these neural networks with regards to the context in which the neural network will eventually be deployed and evaluated. To this end, a new loss function is proposed and shown to increase the performance of neural networks when implemented in a communication system compared to previous methods. It is further shown that a 'split complex' approach used by many implementations can be improved via formalisation of the 'concatenated complex' approach described herein. Experimental results using the orthogonal frequency division multiplexing (OFDM) and spectrally efficient frequency division multiplexing (SEFDM) modulation formats with varying bandwidth compression factors over a wireless visible light communication (VLC) link validate the efficacy of the proposed method in a real system, achieving the lowest error vector magnitude (EVM), and thus bit error rate (BER), across all experiments, with a 5 dB to 10 dB improvement in the received symbols EVM overall compared to the baseline implementation, with bandwidth compressions down to 40% compared to OFDM, resulting in a spectral efficiency gain of 67%.