Power Allocation in mmWave Cell-F ree Massive MIMO with User Mobility Using Deep Learning

We investigate a deep learning method to allocate the downlink transmission power in mmWave cell-free massive multiple-input multiple-output (MIMO), an NP-hard problem. A deep learning method has the advantage that it has significantly lower computational complexity than the non-DL heuristics that are typically used for this task. We consider an indoor office scenario with user equipment (UEs) moving at pedestrian speeds. The max-min power allocation policy is adopted since it guarantees a minimum service quality for all UEs. We choose a long short-term memory (LSTM) network, because it takes into account the correlations between successive power allocation instances. The LSTM network is trained and tested using datasets generated by means of the bisection algorithm. The numerical results, obtained for a particular 3GPP scenario, show that the LSTM network approximates the power allocation of the bisection algorithm very closely. In addition, we found the differences between the two methods regarding the spectral efficiency per UE to be negligible.