TY - GEN
T1 - Enhancing HVAC Control Efficiency
T2 - European Conference on Machine Learning and Principles and Practice of Knowledge Discovery in Databases, ECML PKDD 2024
AU - Kadamala, Kevlyn
AU - Chambers, Des
AU - Barrett, Enda
N1 - Publisher Copyright:
© The Author(s), under exclusive license to Springer Nature Switzerland AG 2024.
PY - 2024
Y1 - 2024
N2 - This paper explores the application of imitation learning (IL) and reinforcement learning (RL) in HVAC control. IL learns to perform tasks by imitating a demonstrator, utilising a dataset of demonstrations. However, the performance of IL is highly dependent on the quality of the expert demonstration data. On the other hand, RL can adapt control policies based on different objectives, but for larger problems, it can be sample inefficient, requiring significant time and resources for training. To overcome the limitations of both RL and IL, we propose a combined methodology where IL is used for pre-training and RL for fine-tuning. We introduce a fine-tuning methodology to HVAC control inspired by a robot navigation task. Using the 5-Zone residential building environment provided by Sinergym, we collect state-action pairs from interactions with the environment using a rule-based policy to create a dataset of expert demonstrations. Our experiments show that this combined methodology improves the efficiency and performance of the RL agent by 1% to 11.35% compared to existing literature. This study contributes to the ongoing discourse on how imitation learning can enhance the performance of reinforcement learning in building control systems.
AB - This paper explores the application of imitation learning (IL) and reinforcement learning (RL) in HVAC control. IL learns to perform tasks by imitating a demonstrator, utilising a dataset of demonstrations. However, the performance of IL is highly dependent on the quality of the expert demonstration data. On the other hand, RL can adapt control policies based on different objectives, but for larger problems, it can be sample inefficient, requiring significant time and resources for training. To overcome the limitations of both RL and IL, we propose a combined methodology where IL is used for pre-training and RL for fine-tuning. We introduce a fine-tuning methodology to HVAC control inspired by a robot navigation task. Using the 5-Zone residential building environment provided by Sinergym, we collect state-action pairs from interactions with the environment using a rule-based policy to create a dataset of expert demonstrations. Our experiments show that this combined methodology improves the efficiency and performance of the RL agent by 1% to 11.35% compared to existing literature. This study contributes to the ongoing discourse on how imitation learning can enhance the performance of reinforcement learning in building control systems.
KW - Continuous HVAC control
KW - Imitation learning
KW - Reinforcement learning
UR - http://www.scopus.com/inward/record.url?scp=85203876664&partnerID=8YFLogxK
U2 - 10.1007/978-3-031-70378-2_16
DO - 10.1007/978-3-031-70378-2_16
M3 - Conference Publication
AN - SCOPUS:85203876664
SN - 9783031703775
T3 - Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
SP - 256
EP - 270
BT - Machine Learning and Knowledge Discovery in Databases. Applied Data Science Track - European Conference, ECML PKDD 2024, Proceedings
A2 - Bifet, Albert
A2 - Krilavičius, Tomas
A2 - Miliou, Ioanna
A2 - Nowaczyk, Slawomir
PB - Springer Science and Business Media Deutschland GmbH
Y2 - 9 September 2024 through 13 September 2024
ER -