Date of Graduation

Summer 2026

Degree

Master of Science in Computer Science

Department

Computer Science

Committee Chair

Razib Iqbal

Abstract

Smart environments equipped with distributed sensors and actuators continuously generate event streams that reflect both user behavior, such as movement patterns, appliance usage, and room transitions, and environmental context, such as temperature, lighting, and humidity changes. The complex, heterogeneous time-series data from such event streams can be effectively used for predicting actuation. Existing automation systems often follow IF-THEN style rule-based automation where a specific sensor input triggers a predefined action, and the recent machine learning based approaches typically consider current sensors states instead of temporal sequence of past events. While these approaches are sufficient for simple scenarios, evolving user preferences and environmental contexts make reliable actuation challenging. This leads to the necessity of an adaptive automation system where the system is not only aware of user preferences and environmental contexts but also considers the temporal context for decision making. This research aims to develop a data-driven framework that automatically learns sensor actuator relationships for adaptive smart environment automation. To achieve this, the proposed framework directly learns actuator behaviors from raw sensor streams without predefined rules or user annotations. This thesis formulates predictive smart-home actuation as a state conditioned temporal prediction problem. The framework predicts both actuator identity and actuator state from recent sensor-actuator event streams and a global actuator-state vector. Experiments across five real-world smart-environment datasets, including two office deployments and three residential apartment deployments, show that temporal models outperform non-sequential baselines. Transformer with Actuator State Embedding (Transformer ASE) achieves the strongest joint actuator-state prediction accuracy in most datasets, while Gated Recurrent Unit (GRU) performs best in the densest dataset, demonstrating that model selection depends on event density and deployment characteristics. Edge measurements on Raspberry Pi class hardware show that the evaluated models can support low-latency inference without compromising accuracy.

Keywords

smart environment, actuation prediction, GRU, transformer, temporal modeling, sensor-actuator network

Subject Categories

Artificial Intelligence and Robotics | Computer Sciences | Data Science | Human-Computer Interaction

Copyright

© Md Abdur Rahman Fahad

Download

Available for download on Sunday, July 30, 2028

Open Access

Share

COinS