Grant: Discovery Launch Supplement

The integration of recommendation systems is crucial in living area systems characterized by a high level of cooperation between humans and machines. The objective of recommendation systems in smart buildings is to optimize energy consumption, improve occupant comfort, and enhance overall building efficiency. These systems rely on diverse data sources such as sensors, weather information, energy usage patterns, and user preferences, to generate personalized recommendations. Recent studies have emphasized the significance of incorporating occupants' perspectives and feedback to enhance these systems. The proposed research program goes beyond merely integrating occupants' feedback, by involving them from the initial stages of platform development via a unified and innovative interactive and cooperative machine learning (ML) framework. Indeed, in the proposed research, humans play a central role in "training" the models, contributing to both data annotation and improving the models' generalization capabilities. Furthermore, our research program stands out by proposing a recommendation system that considers multiple objectives, such as user's goals, comfort, age, and gender, to adapt the model's design. Moreover, while most existing approaches lack explainability, this research focuses on developing explainable recommendations while addressing privacy concerns. The recommendation platform will be validated in different contexts and constructed based on information obtained from three main components that we will develop: 1) activities recognition and prediction, 2) Non-intrusive Load monitoring (NILM), and 3) load forecasting. The development of each component presents its own set of challenges. For instance, activity recognition and prediction will be addressed at different levels of complexity, ranging from individual activities to intersections of multiple activities. One major challenge in smart building applications in general and activity recognition, in particular, is the collection of the required labels. Indeed, the development of powerful supervised ML techniques often demands a substantial amount of labeled data, which is crucial for learning high-quality models. This challenge will be tackled through cooperative ML, and the development of explainable unsupervised learning and unsupervised model adaptation techniques to cope with the scarcity of labeled data. Additionally, current ML models for load forecasting and NILM struggle to capture patterns for all prediction periods. To enhance the performance of ML models, we will explore various valuable sources of information, such as weather, occupancy, and activity predictions, while also incorporating interactive learning. The ultimate goal is to address the fundamental question of how artificial intelligence and humans can collaborate to create a new generation of energy management systems. The proposed research will open new avenues in ML and applications related to energy and buildings.