Lies, damn lies and preferences: A Gaussian process model for ubiquitous thermal preference trials

This paper presents a study of user comfort levels using an ubiquitous interface. The aim is to analyse the comfort function of an individual as opposed to previous approaches that look at the average human being. The data is analysed using Gaussian Process regression which allows several mechanisms to be exploited. These include regression on the data to give an estimate of a users comfort function. The prediction variance is also estimated and outlier influence can be reduced easily. In addition, a natural means of combining the preferences of users falls out of the approach. The combination algorithm takes into account fairness tempered by the quality of the user' preference estimates. Empirical results show that the combined preferences have a well defined maxima which can be used as a control signal for a HVAC system. The Gaussian Process approach is hierarchical and interestingly, while those users studied have differing preferences, their hyperparameters (at the second level of the hierarchy) are concentrated; i.e. there is a strong commonality across individuals in this domain.