Social learning SoftThermostat for commercial buildings

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

“…Daum et al [18], discusses the conflict between different users preferences noting the diversity of preferences and problems faced when trying to combine them into one function with a well-defined peak. Song et al [20] present a system with a similar aim to that here which learns user preferences and combines them for use in the BMS system. While our approach allows combination of preferences and active learning as in [18,20], it differs in that we use a purely Bayesian approach based on the what the samples themselves reveal rather than assuming a model for the data a priori 3 ; a functional logistic model relationship in [18] or using a collaborative rule engine in the case of [20].…”

“…Song et al [20] present a system with a similar aim to that here which learns user preferences and combines them for use in the BMS system. While our approach allows combination of preferences and active learning as in [18,20], it differs in that we use a purely Bayesian approach based on the what the samples themselves reveal rather than assuming a model for the data a priori 3 ; a functional logistic model relationship in [18] or using a collaborative rule engine in the case of [20]. Zhao et al [21] proposes two zones, a comfort and discomfort zone and recursively estimates the boundary set between these.…”

“…The cost functions in this figure are well behaved in terms of being used as control signals; they have a maximum and they fall away from that maximum smoothly without discontinuities. 20 19 We insert a minus here for convenience so that all combination functions are to be maximised. 20 The exception are the tails at the extremity (30°C) which occur because of the lack of data at that point; causingŶ 3 0 to return to zero as mentioned in Section 5.2.1; note this can easily be removed using a sanity check.…”