New-generation wireless networks are designed to support a wide range of services with diverse key performance indicators (KPIs) requirements. A fundamental component of such networks, and a pivotal factor to the fulfillment of the target KPIs, is the virtual radio access network (vRAN), which allows high flexibility on the control of the radio link. However, to fully exploit the potentiality of vRANs, an efficient mapping of the rapidly varying context to radio control decisions is not only essential, but also challenging owing to the interdependence of user traffic demand, channel conditions, and resource allocation. Here, we propose CAREM, a reinforcement learning framework for dynamic radio resource allocation in heterogeneous vRANs, which selects the best available link and transmission parameters for packet transfer, so as to meet the KPI requirements. To show its effectiveness, we develop a testbed for proof-of-concept. Experimental results demonstrate that CAREM enables an efficient radio resource allocation under different settings and traffic demand. Also, compared to the closest existing scheme based on neural network and the standard LTE, CAREM exhibits an improvement of one order of magnitude in packet loss and latency, while it provides a 65% latency improvement relatively to the contextual bandit approach.
翻译:设计新一代无线网络是为了支持范围广泛的服务,提供各种关键业绩指标(KPI)要求,这种网络的一个基本组成部分,以及实现目标KPI的一个关键因素,是虚拟无线电接入网络(vRAN),这种网络在控制无线电连接方面具有高度的灵活性;然而,为了充分利用VRANs的潜力,对无线电控制决定迅速变化的背景进行有效测绘不仅至关重要,而且由于用户交通需求、频道条件和资源分配的相互依存性,因此也具有挑战性。在这里,我们提议CAREM,一个加强的多式VRANs动态无线电资源分配的学习框架,选择可提供的最佳链接和传输参数,以便满足KPI的要求。为了显示其有效性,我们开发了一个验证概念的测试台。实验结果显示,CAREM在不同的环境和交通需求下能够有效地分配无线电资源。此外,与以神经网络和标准LTE为基础的最接近的现行计划相比,CAREM展示了对包损失和耐用量规模的一级的改进,同时提供了相对范围的改进范围为65%。