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

Area of Science:

• Electrical Engineering
• Computer Science
• Optical Communications

Background:

• Neural networks are increasingly used in communication systems.
• A gap exists between neural network training objectives and real-world deployment performance.
• Current methods for handling complex numbers in neural networks can be suboptimal.

Purpose of the Study:

• To propose a new loss function for neural networks in communication systems.
• To introduce an improved 'concatenated complex' approach for neural network implementation.
• To evaluate the performance of the proposed methods in a wireless visible light communication (VLC) system.

Main Methods:

• Developed a novel loss function tailored for communication system objectives.
• Formalized and implemented a 'concatenated complex' approach for neural networks.
• Conducted experiments using orthogonal frequency division multiplexing (OFDM) and spectrally efficient frequency division multiplexing (SEFDM) modulation.
• Tested the system over a wireless visible light communication (VLC) link with varying bandwidth compression.

Main Results:

• The proposed loss function significantly improved neural network performance compared to baseline methods.
• The 'concatenated complex' approach demonstrated superior results over the 'split complex' method.
• Achieved the lowest error vector magnitude (EVM) and bit error rate (BER) across all tested conditions.
• Demonstrated a 5 dB to 10 dB improvement in received symbol EVM.
• Attained a spectral efficiency gain of 67% with 40% bandwidth compression compared to OFDM.

Conclusions:

• The novel loss function and 'concatenated complex' method offer substantial performance gains for neural networks in communication systems.
• The proposed techniques are validated in a practical wireless VLC environment.
• This work provides a more effective approach for deploying neural networks in spectrally efficient communication systems.