Machine Learning Based Digital Pre-Distortion in Massive MIMO Systems: Complexity-Performance Trade-offs

In this paper, we study the trade-off between complexity and performance in massive MIMO systems with neural-network based digital pre-distortion (NN-DPD) blocks at the base station. In particular, we consider a multi-user massive MIMO system with per-antenna NN-DPDs, each with an adjustable NN architecture in terms of the size and the number of NN hidden layers. We first analyze the system performance in terms of compensation of the non-linear hardware distortion for different levels of NN-DPD complexity and the number of antennas. We illustrate the required level of complexity in the trained NN-DPD blocks to approach the performance of an ideal conventional DPD. The statistics of the signal to interference and noise plus distortion ratio for a randomly located UE are selected as the performance metrics. We then assume a limited total digital computation power to be allocated among the NN-DPD blocks and propose to select the NN-DPD architecture of each TX branch based on the channel conditions of its corresponding antenna. To illustrate the importance of such a smart DPD resource allocation, we have analyzed the performance of a massive MIMO system with different NN-DPD architecture selection strategies. Numerical results indicate that by adopting the smart NN-DPD resource allocation, a significant boost in the system performance can be achieved, making room for reducing the overall system cost when scaling a massive MIMO system.