hypocaust
A building climate controller for the NeurIPS 2026 Smart Buildings challenge, keeping every zone comfortable while using less energy.
Control Systems
Smart Buildings
Solo Competitor
Python, NumPy, CMA-ES
Each zone runs a PI controller toward an economic setpoint, while the air handlers and hot water follow actual demand and set back when spaces are empty. Adaptive recursive least squares adjusts the gains as the building's response changes.
Tuning against the official scorer is slow, so I built a local harness that matches it to four significant figures and tuned the controller with CMA-ES. The testing showed that trading comfort for energy savings never paid off in the score.
A public leaderboard score of 0.01281, against 0.01420 and 0.01582 for the public reference controllers. Lower is better.

