Rate Adaptation for Downlink Massive MIMO Networks and Underlaid D2D Links: A Learning Approach

In this paper, a novel learning-based rate adaptation mechanism is proposed for a downlink massive multiple-input-multiple-output (MIMO) network with underlaid device-to-device (D2D) links, where the link signal-to-interference-plus-noise ratio (SINR) cannot be accurately predicted before transmission, even its distribution statistics are unknown at the very beginning. Specifically, two coexistence schemes are considered: (1) the D2D links only reuse the downlink subframes; and (2) the D2D receivers also join the uplink channel estimation of the associated cells. For the second scheme, the downlink interference to the D2D receivers is suppressed at the cost of channel training overhead. The geographic distributions of the selected downlink and D2D users in each frame are modeled as two independent stochastic processes with unknown statistics. As a result, the distribution of interference power is unknown to the transmitters. In order to facilitate robust rate allocation, we first derive the asymptotic expressions of downlink and D2D signal-to-interference-plus-noise ratios (SINRs) for sufficiently large antenna number, and show that their distributions can be approximated by Gaussian or exponential random variables. Subsequently, distributive learning algorithms are proposed to evaluate the means and variances of these random variables. This enables the BSs and D2D transmitters to determine the transmission rates under a constraint on packet outage probability.