Few debates devolve more quickly into bickering than those about the thermostat. Some like it hot. Some like it cold. And others prefer it somewhere in between. Wouldn’t it be great, then, if AI could find an agreeable middle ground? Researchers at Purdue University believe it can, and they have laid out a novel framework for HVAC systems in new research.
A paper on the preprint server Arxiv.org (“Learning Personalized Thermal Preferences via Bayesian Active Learning with Unimodality Constraints“) describes a recommender system — a type of machine learning model that attempts to predict a person’s likes or dislikes — for thermal preferences. The researchers explain: “The objective of this paper is to sequentially pose intelligent queries to occupants in order to optimally learn the indoor room temperature values which maximize their satisfaction. This framework is an important step toward the development of intelligent HVAC systems which would be able to respond to individual occupants’ personalized thermal comfort needs.”
A lack of relevant corpora poses a challenge in creating an “intelligent” temperature recommender system, the team notes. Data must be collected from a given HVAC-equipped building’s occupants, and hyperlocal surveys aren’t particularly scalable or cost-effective. An alternative, the researchers posit, is a system that “sequentially” poses questions to office workers and “optimally” learns from their responses a range of ideal temperatures.
So how does it work? Folks in the building answer the question “How satisfied are you with current thermal conditions?” every half hour in one of three ways: “I’m satisfied with current condition,” “I prefer warmer,” or “I prefer cooler.” Each successive Q&A round narrows down the range of temperatures; in the second round, for example, the system looks for temperatures 3 degrees (Centigrade) above or below the current room temperature in increments of 0.5 degrees.
To validate their design, the paper’s authors first set their framework loose on three synthetic occupants’ preference data — occupants they imagined worked in private rooms and had preferred temperatures between 22.1 degrees and 25 degrees, for the sake of the experiment. They fed the “responses” from the virtual test subjects to the AI system, which after just six questions was able to infer the maximum preferred room temperatures for all three.
Next, the team deployed the system in a real-world private office (in West Lafayette, Indiana) with six test subjects, each of whom visited one room (set to 21 degrees) every day over the course of several days. Every 30 minutes, they were posed temperature preference questions via a website, and their responses were used to predict new sets of temperatures. In the end, after between five and 10 queries per person, the recommender managed to settle on a likely two-degree range for all six subjects with 95 percent certainty.
The team believes that their approach could increase building occupants’ “satisfaction” and reduce energy waste.
“In this paper, we have focused on the development of [a] simple, robust, low-cost, easy to compute and easy to implement personalized … framework which focuses on room air temperature as its most important feature,” they wrote. “This framework is an important step toward the development of intelligent HVAC systems which would be able to respond to individual occupants’ personalized thermal comfort needs.”