The ability to manage the distributed functionality of large multi-vendor networks will be an important step towards ultra-dense 5G networks. Managing distributed scheduling functionality is particularly important, due to its influence over inter-cell interference and the lack of standardization for schedulers. In this paper, we formulate a method of managing distributed scheduling methods across a small cluster of cells by dynamically selecting schedulers to be implemented at each cell. We use deep reinforcement learning methods to identify suitable joint scheduling policies, based on the current state of the network observed from data already available in the RAN. Additionally, we also explore three methods of training the deep reinforcement learning based dynamic scheduler selection system. We compare the performance of these training methods in a simulated environment against each other, as well as homogeneous scheduler deployment scenarios, where each cell in the network uses the same type of scheduler. We show that, by using deep reinforcement learning, the dynamic scheduler selection system is able to identify scheduler distributions that increase the number of users that achieve their quality of service requirements in up to 77% of the simulated scenarios when compared to homogeneous scheduler deployment scenarios.

中文翻译：

---

使用深度学习的动态调度程序管理

管理大型多供应商网络的分布式功能的能力将是迈向超密集 5G 网络的重要一步。由于分布式调度功能对小区间干扰的影响以及调度器缺乏标准化，因此管理分布式调度功能尤为重要。在本文中，我们通过动态选择要在每个小区实施的调度程序，制定了一种管理跨小区小集群的分布式调度方法的方法。我们使用深度强化学习方法，根据从 RAN 中已有数据观察到的网络当前状态，确定合适的联合调度策略。此外，我们还探索了三种训练基于深度强化学习的动态调度器选择系统的方法。我们比较了这些训练方法在模拟环境中的性能，以及同构调度器部署场景，其中网络中的每个单元都使用相同类型的调度器。我们表明，通过使用深度强化学习，动态调度程序选择系统能够识别调度程序分布，与同类调度程序部署场景相比，在多达 77% 的模拟场景中，这些分布增加了实现其服务质量要求的用户数量.