End-to-End MIMO Systems with Conditional Generative Adversarial Networks

Conventional autoencoder-based systems, comprising a neural-network encoder at the transmitter and a neural-network decoder at the receiver, often face limitations in re-alistic signal transmission due to its reliance on differentiable channel models. In response to this limitation, an end-to-end framework has emerged, employing a conditional generative adversarial network (CGAN) for channel learning. The CGAN not only learns to represent channel effects through feature extraction but also facilitates gradient back-propagation between the receiver and the transmitter, thereby enhancing the system's adaptability. This paper extends the CGAN-based end-to-end system from single input single output channels to multiple input multiple output (MIMO) scenarios. This CGAN-based MIMO system demonstrates promising performance comparable to the autoencoder communication systems, highlighting the potential of CGANs as effective alternatives for original channels.