On 26th June 2026, Toki Miyake (M1), Department of Electrical Engineering and Information Systems, received DICOMO 2026 Outstanding Presentation Award.
Outstanding Presentation Award
The DICOMO (Multimedia, Distributed, Cooperative, and Mobile) Symposium is one of Japan's largest domestic symposia in the fields of distributed, mobile, and ubiquitous computing, held annually by the Information Processing Society of Japan since 1997. The Outstanding Presentation Award is presented each year to a selection of particularly excellent presentations chosen from among more than 200 submissions.
About awarded research
In the control of building heating, ventilation, and air-conditioning (HVAC) systems, it is essential to suppress excessive energy consumption while maintaining occupant comfort. Model-free reinforcement learning (MFRL), which can acquire a control policy through trial and error without requiring a physical model, is a promising approach; however, its exploration in the early stages of learning is inefficient, and the vast number of trials needed to reach practical performance has been a barrier to deployment in real-world environments. In this research, we proposed a method that harnesses the common-sense knowledge held by large language models (LLMs) as a "teacher": in the early stages of learning, the agent imitates the control recommended by the LLM to rapidly ramp up performance, and once the agent has become proficient in its environment and begins to outperform the LLM, it transitions smoothly to autonomous reinforcement learning. In simulation, the method accelerated learning by up to approximately 27 times relative to conventional approaches. In a demonstration experiment in an actual occupied room, it reduced deviation from the comfort range to roughly one-fifth that of conventional control after only 15 days of learning, achieving control performance that surpassed even the LLM teacher on its own.
Your impression & future plan
In carrying out this research, I would like to express my heartfelt gratitude to Professor Yoshihiro Kawahara, Project Assistant Professor Hiroaki Murakami, and all the other faculty members who provided dedicated guidance on a daily basis, as well as to everyone in the laboratory who kindly assisted with my presentation practice. I am also deeply grateful to everyone at Daikin Industries, Ltd. for their collaboration on this joint research. I feel that being able to convey the appeal of the research in an accessible way is what led to this award, and I am truly delighted. Encouraged by this recognition, I will continue to work all the harder to develop this method into an air-conditioning control technology that is genuinely useful in society.
