End-to-End Autoencoder Design for MU-MIMO Systems With Low Pilot Overhead

Multiuser multiple-input multiple-output (MU-MIMO) systems are critical for 5G and beyond networks due to their potential to improve spectral efficiency. However, traditional model-based designs often struggle to cope with the complexity and variability of real-world wireless channels, especially in mobile environments. In MU-MIMO systems, this challenge is further aggravated by the need for frequent pilot transmissions, which increase control signaling and reduce data throughput. This paper proposes an end-to-end autoencoder-based architecture that jointly optimizes the transmitter and receiver through supervised learning directly from data, enabling robust communication. A forward-backward channel estimation technique is introduced to interpolate sparse pilot-based estimates, significantly reducing the pilot overhead without explicit channel feedback. Simulations under the 3GPP TR 38.901 fading model show that the proposed system achieves accurate channel estimation with less than 3% pilot overhead, at the cost of a fixed 10 ms delay, targeting enhanced mobile broadband (eMBB) scenarios where throughput is prioritized over ultra-low latency. The estimator outperforms traditional least-squares estimation by 6.6 dB in mean square error, while the end-to-end system reduces the symbol error rate (SER) by up to 3 dB compared to the baseline 8-PSK scheme with zero-forcing equalization.