A reinforcement learning approach to queue-aware scheduling in full-duplex wireless networks

摘要

Full-duplex communications promise to double the throughput of a wireless network, so long as the resulting interferences can be combated. Nonetheless, already dealing with the intricacy of determining base station-to-user radio channels, full duplex wireless networks need additional information on the channels in between all the user equipment. This information is necessary to determine user radio conditions, and thereafter efficiently allocate resources. A signaling overhead is likely to burden the user equipment, which already lack any methods to estimate and convey such user-to-user channel states. In this paper, we aim to circumvent the complexities and requirements of traditional scheduling techniques by introducing a reinforcement learning based scheduling algorithm for full-duplex wireless networks. This scheduling approach does not need to estimate user-to-user channels, and rather learns how to best allocate the network?s radio resources. We show that our proposal can match scheduling with complete channel state information in terms of user equipment throughput, and that it performs well under multiple testing scheduling scenarios: increased user equipment numbers, randomized user equipment demand, and user equipment clustering among others.