The problem
Write one control policy for a building’s heating and cooling that keeps occupants comfortable and cuts energy use and carbon, and make it work in buildings and climates it has never seen.
What I built
- Held-out test sets. I ran the hosts’ building simulator to generate 6 new buildings in 5 climates for 30 days each: 30 scenarios the policy never trained on.
- A physics-informed controller. Zone PI loops, night and weekend setback, occupancy-aware pre-start and a holiday detector, improved over eight versions with Optuna.
- Selection on validation. I chose final submissions on the unseen-building scores, not the public leaderboard.
The hardest part
Not overfitting to the handful of buildings in the public data. I read the simulator’s source to understand the physics, then tested every change on newly generated buildings and weather. Along the way I found a billing bug in the simulator. I reported it to the hosts instead of exploiting it, and they fixed it.
Result
A best public score of 0.01244 against 0.01853 for the best public notebook, and well ahead of the organisers’ baseline. Round 1 ends on 13 November 2026 and the final on 15 January 2027.
Updates
No logged updates yet. New work on this project will appear here and in the activity log.