Gaussian Process models for ubiquitous user comfort preference sampling; global priors, active sampling and outlier rejection

This paper presents a ubiquitous thermal comfort preference learning study in a noisy environment. We introduce Gaussian Process models into this field and show they are ideal, allowing rejection of outliers, deadband samples, and produce excellent estimates of a users preference function. In addition, informative combinations of users preferences becomes possible, some of which demonstrate well defined maxima ideal for control signals. Interestingly, while those users studied have differing preferences, their hyperparameters are concentrated allowing priors for new users. In addition, we present an active learning algorithm which estimates when to poll users to maximise the information returned.