A Comparative Evaluation of Generative Adversarial Networks for Channel Modeling in Coded MIMO-OFDM Systems

In 5G and beyond communication, the combination of deep learning (DL) and digital twin (DT) channel model makes the system improve efficiency. By using the generative model of DL, the specified statistical distribution, the channel feature as channel impulse response is generated from the measurement data. This paper compares the schemes of Generative adversarial networks (GAN), Variational Auto-encoder GAN (VAE-GAN) and Wasserstein GAN with gradient penalty (WGAN-GP) in acquiring the BERs of SISO/MIMO system with different delays from the real data. From the simulation results, the channel characters are re-edified exactly referred to as the original channel, especially in terms of WGAN-GP. Moreover, of the LDPC decoding algorithms as Linear Approximation Min-Sum (LAMS) and Belief Propagation (BP), the performance of LAMSLDPC is the best in case of employing WGAN-GP in MIMOOFDM system. If the SNR is increased (SNR > 20dB), the better BERs of the WGAN-GP are. Therefore, we propose to exploit LDPC-LAMS combined to WGAN-GP in our coded MIMOOFDM system.