Auction-Driven Multiuser Beamforming with Deep Learning

Consider a multi-user wireless information transfer system where a multiple-antenna access point (AP) aims to sell its downlink radio access to the end users in an auction premise so that the social welfare is maximized. To this end, the AP holds auctions in which each user bids for its desired minimum signal-to-interference-plus-noise power ratio. Based on the user's channel state information, bids, and service demands, the AP seeks the optimal set of users for which information streams are to be allocated through beamforming. We formulate the optimization problem of finding the optimal allocation rule, apply brute-force search approach to find all the feasible allocation sets using uplink-downlink duality-based algorithm (UDD), and obtain the optimal allocation rule. To circumvent the time-greediness of conventional optimization methods such as semi-definite relaxation or UDD, we propose a deep neural network architecture, and use it to solve the optimization problem with a good accuracy. The training data is collected by solving offline plenty of network realizations via the application of the UDD algorithm.