TY - GEN
T1 - Reinforcement learning for hybrid beamforming in millimeter wave systems
AU - Peken, Ture
AU - Tandon, Ravi
AU - Bose, Tamal
N1 - Funding Information:
This work was partially supported by US NSF through grant CNS-1715947. It was also partly supported by the Broadband Wireless Access and Applications Center (BWAC); NSF Award No. 1265960.
Publisher Copyright:
© 2019 International Foundation for Telemetering. All rights reserved.
PY - 2019
Y1 - 2019
N2 - The use of millimeter waves (mmWave) for next-generation cellular systems is promising due to the large bandwidth available in this band. Beamforming will likely be divided into RF and baseband domains, which is called hybrid beamforming. Precoders can be designed by using a predefined codebook or by choosing beamforming vectors arbitrarily in hybrid beamforming. The computational complexity of finding optimal precoders grows exponentially with the number of RF chains. In this paper, we develop a Q-learning (a form of reinforcement learning) based algorithm to find the precoders jointly. We analyze the complexity of the algorithm as a function of the number of iterations used in the training phase. We compare the spectral efficiency achieved with unconstrained precoding, exhaustive search, and another state-of-art algorithm. Results show that our algorithm provides better spectral efficiency than the state-of-art algorithm and has performance close to that of exhaustive search.
AB - The use of millimeter waves (mmWave) for next-generation cellular systems is promising due to the large bandwidth available in this band. Beamforming will likely be divided into RF and baseband domains, which is called hybrid beamforming. Precoders can be designed by using a predefined codebook or by choosing beamforming vectors arbitrarily in hybrid beamforming. The computational complexity of finding optimal precoders grows exponentially with the number of RF chains. In this paper, we develop a Q-learning (a form of reinforcement learning) based algorithm to find the precoders jointly. We analyze the complexity of the algorithm as a function of the number of iterations used in the training phase. We compare the spectral efficiency achieved with unconstrained precoding, exhaustive search, and another state-of-art algorithm. Results show that our algorithm provides better spectral efficiency than the state-of-art algorithm and has performance close to that of exhaustive search.
UR - http://www.scopus.com/inward/record.url?scp=85090974845&partnerID=8YFLogxK
UR - http://www.scopus.com/inward/citedby.url?scp=85090974845&partnerID=8YFLogxK
M3 - Conference contribution
AN - SCOPUS:85090974845
T3 - Proceedings of the International Telemetering Conference
SP - 138
EP - 147
BT - 55th Annual International Telemetering Conference, ITC 2019
PB - International Foundation for Telemetering
T2 - 55th Annual International Telemetering Conference: Cultivating the Next Generation of Range Engineers, ITC 2019
Y2 - 21 October 2019 through 24 October 2019
ER -