Federated Reinforcement Learning for the Building Facilities
Federated Reinforcement Learning for the Building Facilities
复制标题
DOI：
10.1109/coins54846.2022.9854959
复制
发表时间：
2022-08
期刊：
影响因子：
--
通讯作者：
Koki Fujita;Shugo Fujimura;Yuwei Sun;H. Esaki;H. Ochiai
中科院分区：
文献类型：
--
作者：
Koki Fujita;Shugo Fujimura;Yuwei Sun;H. Esaki;H. Ochiai
近年来，已经引入了利用AI和IoT的系统。物联网的发展增强了人与物之间的联系，便利性正在提高。人工智能还用于自动化人类执行的任务，并控制各种任务。它也开始应用于建筑设施，并且存在建筑物相互协作的情况。在本文中，我们解决的问题，控制建筑设施。建筑物配备了空调、蓄电池和太阳能电池板。其目标是控制暖通空调系统考虑到交通的人和蓄电池的状态。由于每个建筑物都有不同的目标情况，因此为每个建筑物找到最佳策略非常重要。我们的目标是通过使用强化学习来解决这个问题，并开发一个框架，可以通过简单的奖励函数来学习各种策略。在这项研究中，我们已经实验表明，控制是最佳的节能方案。对于建筑设施，我们提出了各种基本的奖励函数，并确认了通过组合这些函数可以学习灵活的策略。此外，我们还证明了联邦学习可以加速学习收敛，同时保护建筑物之间的隐私。
In recent years, systems utilizing AI and IoT have been introduced. The development of IoT enhances the relationship between people and things, and convenience is improving. AI is also utilized to automate tasks that have been performed by humans, and to control various tasks. It is also beginning to be applied to building facilities, and there are situations where buildings interact cooperatively with each other. In this paper, we address the issue of controlling building facilities. Buildings are equipped with air conditioners, storage batteries, and solar panels. The goal is to control HVAC system considering the traffic of people and the state of storage batteries. Since each building has different target situations, it is important to find the optimal policies for each building. We aim to solve this problem by using reinforcement learning and to develop a framework that can learn various policies by the simple reward functions. In this study, we have experimentally shown that the control is optimal for power saving scenarios. For building facilities, we proposed various basic reward functions and also confirmed that flexible policies can be learned by combining these functions. Furthermore, we show that the learning convergence can be accelerated by federated learning while preserving privacy among the buildings.