Efficient learning of personalized visual preferences in daylit offices: An online elicitation framework

摘要

Human preference-based control in buildings may achieve maximum occupant satisfaction as well as energy savings. Adaptive and online learning methods are needed for learning human preferences, using the minimum amount of data possible with quantified uncertainty. This paper presents an online visual preference elicitation learning framework, developed for efficiently learning occupants' visual preference profiles and hidden satisfaction utility functions in daylit offices. A combination of Thompson sampling and pure exploration (uncertainty learning) methods was used in the sequential elicitation framework, to determine the set of successive visual preference queries (for visual conditions) with most information gain. In this way, a balance between exploration and exploitation is realized and the area around the satisfaction utility maximum is learned with minimum number of steps.Experiments with human subjects were conducted in private sidelit offices to demonstrate the feasibility and performance of the learning framework. A single-variable model (vertical illuminance) was used to demonstrate the method and visualize results. The entropy of the distribution of the most preferred visual condition is computed for each learned preference profile to quantify the certainty and evaluate the learning efficiency. Our method learns most individual visual preferences to an acceptable certainty level within one day, and indicates the need for personalization. Finally, we discuss the integration of visual preferences into control applications. A switching algorithm is proposed, shifting iteratively between the learning and control modes depending on the certainty of the learned preference model. This work contributes to developing comprehensive online learning methods towards preference-based tuned indoor environments.