A Deep-Learning Based Approach for Wireless Information and Power Transfer in Massive MIMO Systems.

Wireless power transfer (WPT) is a promising energy harvesting technique conceived for extending the battery-recharge period of energy-constrained nodes in wireless networks. WPT is more practically realizable in a massive multiple-input-multiple-output (m-MIMO) communication system, thanks to the possibility of multiple antenna transmission and application of energy beamforming. In this work, we consider the downlink m-MIMO communication and aim to design an efficient wireless information and energy transfer strategy. Specifically, we propose to jointly optimize the energy and information transmission durations at the transmitter along with the amount of power to be transferred for maximizing the users minimum rate subject to minimum harvested energy constraint. The problem is not convex, so a suitable approximate or heuristic method for learning the parameters such as Bayesian optimization, random search, etc. can be used. Some of these methods have been investigated in the literature with acceptable overall performance. However, in this project, we propose to solve the problem using machine learning methods. Specifically, we aim to investigate various deep learning architectures suitable for this problem. We would like to highlight that building a machine learning model needs significant computing and time resources as it has to be trained and tested on huge amounts of data. The process of data generation is of critical importance. The data should be representative of the possible scenarios and must be generated in a way that it should help the model to generalize. Similarly, the solutions corresponding to the generated data must be computed with an already existing approach. These approaches must be tuned to ensure their results are as close to the reality as possible. Further, it is imperative to verify the integrity of the generated data/solution to avoid problems while training the models. Note that real-world data generation and experiments is a tedious task and requires specialized hardware to be used by trained personnel and is out of scope of this work