Learning to cooperate in decentralized wireless networks

Several key wireless communication setups call for coordination capabilities between otherwise interfering transmitters. Coordination or cooperation can be achieved at the

expense of channel state information exchange. When such information is noisy, the derivation of robust decision-making algorithms is unfortunately known to be very challenging via conventional optimization method. In this paper we introduce a learning-based framework which allows the agents, aka. the transmitters, to produce as-relevant-as-possible messages to each other on the basis of arbitrarily partial and noisy local channel

state information. The messages are produced via distributed deep neural networks (DNNs) which are trained for a specific coordination purpose. The message-passing DNNs are completed with decision-making DNNs which are trained for a network

metric maximization. Promising preliminary results are obtained in the context of sum-rate maximizing decentralized power control.