Air Conditioner Control Learning Users' Sensations Based on Reinforcement Learning and Its Scalability Improvement

摘要

This study proposes the air conditioner (AC) control system based on users' sensations. The purpose of the system is to improve low-performance ACs in terms of energy efficiency and comfortableness performance. The system consists of wireless sensor nodes and user nodes such as PCs and smartphones, and it is applicable to already installed ACs. The users enter their sensation such as cold, good, a little hot and very hot through the user node, and the system determines the appropriate control according to the users' sensations. The control signal is transmitted via an equipped IR remote controller. The appropriate control policy is determined based on Q-Learning, which is a reinforcement learning method. In this study, we propose several types of methods and investigate effective methods in terms of energy efficiency and comfortableness performance. For multiple users, several methods for integrating users' sensations are presented. These make our proposed system applicable to a large number of users. Further, in order to reduce the energy consumption and the number of users' inputs, several types of reward functions are presented. In addition to the reward functions proposed in, in this paper, we propose a new reward function, which improves the performance for the large number of users. In the simulation, five types of methods including a new proposal in this paper, are evaluated in terms of the time needed for providing a comfortable environment and the energy consumption. We clarify the effective methods among them and the tendency of the scalability against the number of users.