Xi Chen

Abstract:  
Tracking activities in multi-resident smart homes using non-intrusive ambient sensors poses a major challenge: associating anonymous data to specific residents. To overcome the representational limitations of traditional filtering methods, this thesis reformulates this task as a sequential decision-making problem based on Markov Decision Processes. Within this framework, we first propose supervised approaches. These include NEP+Search, which uses next-event prediction to evaluate the consistency between an event and a resident's history, thereby constructing a reward signal to guide the search, as well as Behavior Cloning, which directly predicts assignment actions end-to-end. These methods significantly outperform classical models during complex concurrent activities. Next, to improve generalization, we explore generative approaches with LADA for text-conditioned zero-shot reasoning, and LLM+BC for the fine-tuning of lightweight language models via LoRA, achieving state-of-the-art performance even on highly noisy data. Empirically evaluated on the CASAS, MARBLE, and MuRAL datasets, this research offers a comprehensive progression from mathematical modeling to advanced generative AI solutions, ensuring accurate, scalable, and privacy-preserving tracking.

Keywords:
Multi-resident data association, Smart home, Markov Decision Process, Large language models, Ambient intelligence, Deep learning, Human Activity Recognition.