Confidence-Aware Behavior Learning for Proactive Smart Home Automation

Description

Research Problem. Most smart home systems today still wait for the user to do something. The user presses a button, sets a schedule, or writes a simple rule such as "turn the light on when motion is detected". These systems do not learn how the user actually lives, so lights and devices are often left on when nobody needs them, and the user keeps controlling things manually. Previous smart home projects at our own university, such as IoT Smart Home Controllers (Sweiti and Fannoun, 2021) and Smart Home Lighting System (Jaafreh and Almtour, 2022), as well as the Smart Home Automation System published by Zaro et al. (2021), succeeded in connecting sensors and devices to the cloud. However, they still depend on the user to give commands and do not learn long-term behaviour. Some newer systems use machine learning to predict what the user will do next, but they usually act on every prediction, even when the model is unsure. When the house makes a wrong move, for example turning off a light while someone is still reading, users quickly stop trusting it. In this project we study whether a smart home can learn a user's daily routine and also know when it is confident enough to act alone. Our research question is: can a lightweight prediction model with a confidence check reduce wasted energy and manual commands without increasing wrong automatic actions, while running on low-cost hardware? Research Aim. We aim to build and test a smart home system that learns user habits, predicts device usage, and gradually moves from asking the user for confirmation to acting on its own as it proves to be reliable. Research Objectives. • Build a sensing system that collects motion, light, energy, and door access data, and filters noisy sensor readings. • Turn the collected data into time-based features such as hour of the day, weekday or weekend, occupancy, and recent device states. • Train a Gradient Boosting model to predict which device actions the user is likely to take. • Design a "trust ladder" decision policy: at first the system asks the user before acting, and once its confidence and past accuracy are high enough, it acts automatically. • Test the full system on a small-scale smart home prototype. Research Contribution. Rule-based and schedule-based systems cannot adapt, and many learning-based systems act without checking how sure they are. Our main contribution is a confidence-aware approach in which the system earns more freedom only after showing that its predictions are correct. We want to find out whether this gives a better balance between automation, energy saving, and user trust, without needing powerful hardware. In our first project report, the system was designed to act directly on predictions. While planning the implementation, we realised that one wrong automatic action could make the user lose trust in the whole system. This led us to the trust ladder idea that this research is built on.