The critical question about massive MIMO systems highlighted in the article: Massive MIMO: "Ten Myths and One Critical Question" by authors: Emil Bjornson, Erik G. Larsson, and Thomas L. Marzetta, is the use of the mode of FDD which generates problems like pilots overhead , channel estimation, feedback overhead,....
Many works have used artificial intelligence, particularly Deep Learning, to solve the problems of FDD mode in massive MIMO systems.
So my reflection which is at the same time a question for the whole community is the following: With the use of Deep Learning, if the system is sufficiently trained:
1- Will these problems always exist?
2- Won't the system be automatic so that it will self-configure to channel conditions?
The propagation channels are changing whenever a device moves a distance ≈ wavelength/4. The measurements campaigns regarding channel prediction that I have seen are not able to predict channels beyond wavelength/2. I doubt that deep learning can solve this since wireless propagation is extremely complex - this is why ray-tracing is so computationally complex. It is relatively easy to generate coverage maps that predict the average channel quality over an area but very hard to predict the small-scale fading variations around the average.
However, one can indeed use machine learning to identify high-level channel properties. In addition to the average channel quality, one could learn second-order statistics of the channel fading. That can be utilized to reduce the pilot overhead, similar to what JSDM and other theory-driven frameworks are doing. I believe this is what the vendors are already trying to do in their massive MIMO products. But as far as I know, the performance is still much worse than with TDD.
Thank you Emil Björnson for this very clear answer. I always had doubts about that but now it's clear in my head.