GAD Editorial · Published · · English study · Chinese summary · Study dated 2026-10-05
An occupant-centred building system can become surprisingly demanding of its occupants. Repeated questions about temperature may produce useful labels, but every request also interrupts someone. A model that learns efficiently should therefore be evaluated through the burden of supplying information as well as the quality of its predictions.
A study first submitted on 16 September 2023 investigates active learning for personalised thermal-comfort models. It selects informative conditions for feedback and evaluates the approach in an EnergyPlus simulation of a real testbed, supplemented by preference data from 58 study occupants. The reported benefits belong to that evaluation; they do not establish equivalent savings or satisfaction in a newly occupied building. Study and dated record
Define what the answer means
For a hypothetical workplace pilot, distinguish a person's immediate preference from a diagnosis of the room. Wanting a cooler environment does not, by itself, identify which architectural or operational condition should change. Keep the wording and timing of each question consistent enough to interpret the response.
GAD proposes documenting what the system asks, why that moment was selected and what use is permitted for the answer. Give participants a clear way to decline or stop responding. Collect only the information necessary for the study, and establish consent, access and retention arrangements before collecting personal feedback.
A non-response should remain a non-response. Treating silence as comfort would reward a system for missing people who are busy, absent or unwilling to participate.
Measure the questions that were not asked
An active-learning method selects some conditions and leaves others unlabelled. A useful evaluation therefore needs a separate check on what the selection process misses. Compare its sampling with an appropriate independent reference, including periods or locations that receive fewer prompts.
Record total requests, completed responses and reported burden separately. A lower count of collected labels could indicate greater efficiency, but it could also reflect disengagement. The interpretation depends on the mechanism behind the reduction.
Inspect disagreement and unusual conditions rather than collapsing everything into a building-wide score. A method that serves the majority well may still leave a small group poorly represented. For design teams, those exceptions can point towards local spatial issues worth investigating, without proving their cause.
Keep learning separate from intervention
Begin evaluation in an observational or simulated setting with the necessary permissions. Changes to occupied-building controls require a separately approved plan, suitable limits and a responsible facilities team. Research evidence about asking fewer questions does not itself authorise changes to equipment or conditions.
GAD has not reproduced the study. Its contribution to an architectural brief is a sharper definition of human-centred performance: consider how feedback is requested, whose experience remains visible and what happens when the model is uncertain. The best result is not simply fewer prompts. It is a more dependable understanding of comfort without making participation unnecessarily onerous.