IDLab, Department of Information Technology, IMEC, Ghent University, Technologiepark Zwijnaarde 15, B-9052 Ghent, Belgium.
Sensors (Basel). 2021 Oct 21;21(21):6977. doi: 10.3390/s21216977.
Nowadays, broadband applications that use the licensed spectrum of the cellular network are growing fast. For this reason, Long-Term Evolution-Unlicensed (LTE-U) technology is expected to offload its traffic to the unlicensed spectrum. However, LTE-U transmissions have to coexist with the existing WiFi networks. Most existing coexistence schemes consider coordinated LTE-U and WiFi networks where there is a central coordinator that communicates traffic demand of the co-located networks. However, such a method of WiFi traffic estimation raises the complexity, traffic overhead, and reaction time of the coexistence schemes. In this article, we propose Experience Replay (ER) and Reward selective Experience Replay (RER) based Q-learning techniques as a solution for the coexistence of uncoordinated LTE-U and WiFi networks. In the proposed schemes, the LTE-U deploys a WiFi saturation sensing model to estimate the traffic demand of co-located WiFi networks. We also made a performance comparison between the proposed schemes and other rule-based and Q-learning based coexistence schemes implemented in non-coordinated LTE-U and WiFi networks. The simulation results show that the RER Q-learning scheme converges faster than the ER Q-learning scheme. The RER Q-learning scheme also gives 19.1% and 5.2% enhancement in aggregated throughput and 16.4% and 10.9% enhancement in fairness performance as compared to the rule-based and Q-learning coexistence schemes, respectively.
如今，使用蜂窝网络许可频谱的宽带应用发展迅速。出于这个原因，长期演进非许可（LTE-U）技术有望将其流量卸载到非许可频谱。然而，LTE-U 传输必须与现有的 WiFi 网络共存。大多数现有的共存方案考虑协调的 LTE-U 和 WiFi 网络，其中有一个中央协调器，用于通信共存网络的业务需求。然而，这种 WiFi 流量估计方法增加了共存方案的复杂性、业务开销和反应时间。在本文中，我们提出了基于经验回放（ER）和奖励选择性经验回放（RER）的 Q-learning 技术，作为非协调的 LTE-U 和 WiFi 网络共存的解决方案。在提出的方案中，LTE-U 部署了一个 WiFi 饱和感测模型来估计共存 WiFi 网络的业务需求。我们还在非协调的 LTE-U 和 WiFi 网络中实现的基于规则和基于 Q-learning 的共存方案之间进行了性能比较。仿真结果表明，RER Q-learning 方案比 ER Q-learning 方案收敛更快。与基于规则和 Q-learning 的共存方案相比，RER Q-learning 方案分别在总吞吐量上提高了 19.1%和 5.2%，在公平性方面提高了 16.4%和 10.9%。
