Experience-Driven Wireless D2D Network Link Scheduling: A Deep Learning Approach
Experience-Driven Wireless D2D Network Link Scheduling: A Deep Learning Approach
复制标题
DOI：
10.1109/icc.2019.8761818
复制
发表时间：
2019-05
期刊：
影响因子：
--
通讯作者：
Shuai Zhang;Wenlong Shen;Max Zhangt;Xianghui Cao;Yu Cheng
中科院分区：
文献类型：
--
作者：
Shuai Zhang;Wenlong Shen;Max Zhangt;Xianghui Cao;Yu Cheng
近年来，由于网络设备数量的增加和部署设置的多样化，设备到设备（device-to-device, D2D）网络的协议设计重新引起了人们的关注。在无线环境中，大多数网络优化任务基本上都是np困难问题，因为管理干扰会引入组合复杂性。现有的方法使用一般启发式算法来解决底层图问题。它们虽然高效、简单，但不能适应服务提供者不断变化的需求和优先级，不能利用过去的数据来识别和开发其中的信息。在本文中，我们研究了一个代表性的网络优化任务，即通过单无线电、单通道D2D网络中的链路调度来最大化基于吞吐量的系统效用，并提出了一种基于学习的方法来利用过去的经验来生成一个好的调度策略。我们结合了循环神经网络（RNN）提供的模式匹配能力和强化学习（RL）在不断变化的环境中的灵活性。该算法在现有的软件框架下实现，并通过数值实验进行了验证。我们发现它的整体解质量与现有的各种网络规模的启发式算法相当，并且报告了在显著降低计算时间的情况下提高的系统吞吐量。
The protocol design of device-to-device (D2D) networks have regained research interest in recent years, due to the increasing number of networking devices and the diverse deployment settings. Most of the network optimization tasks are fundamentally difficult NP-hard problems in wireless settings, because managing interference introduces combinatorial complexity. Existing approaches use general heuristic algorithms for the underlying graph problems. While efficient and simple, they are not adaptive to the changing requirement and priorities of the service providers, and make no use of the past data to recognize and exploit the information within. In this paper, we study a representative network optimization task of maximizing the throughput-based system utility through link scheduling in a single-radio, single-channel D2D networks, and propose a learning-based method to leverage past experience to generate a good scheduling policy. We combine the pattern matching capabilities provided from recurrent neural networks (RNN) and the flexibility in changing environment from reinforcement learning (RL). The algorithm is implemented with existing software frameworks and tested with numerical experiments. We find that its overall solution quality is comparable to existing heuristics with various network scales, and report an improved system throughput with significant lower computation time.