CPS: Medium: Causal Reinforcement Learning for Smart Building CPS

About This Grant

The building sector is responsible for 28% of the U.S. primary energy use and 35% of the carbon emissions. Yet most buildings continue to rely on outdated control systems that are difficult to customize and fail to adapt to the changing occupant needs or environmental conditions. This project supports the development of intelligent building technologies that reduce energy use while maintaining comfort, helping address the national goals helping reduce operational costs and improve the economic efficiency of building operations at scale. The research introduces a new control framework that uses building knowledge and available building information to enable energy-efficient, autonomous operation with minimal human intervention. Compared to the conventional systems that require significant engineering effort to scale across different buildings, the proposed approach offers a novel flexible alternative. The project also includes education and outreach efforts that engage K-12 teachers, and undergraduate students through hands-on learning, curriculum development, and research experiences in smart building technologies and cyber-physical systems. The research focuses on developing a causal reinforcement learning (RL) framework that integrates structured knowledge about building systems and occupant behavior to accelerate learning and improve control performance. The core innovation is the use of structural causal models (SCMs) that describe cause-effect relationships within the building environment. These models, built from building topology and expert knowledge, guide RL agents to learn more efficiently and safely from limited sensor data. The project includes three technical components: (1) scalable methods for constructing SCMs through topology reduction; (2) causal RL algorithms that use these models to optimize energy and comfort; and (3) adaptive strategies that adjust control policies based on the building characteristics and occupant preferences, implemented on low-cost sensing stations. The framework will be validated in both simulation and a real-world office testbed. Expected outcomes include reductions in energy consumption, improved occupant experience, and a generalizable framework for intelligent control in data-constrained, heterogeneous systems. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.