A β¦-storey tower is shaken by an earthquake. A tuned mass damper β a heavy block on a spring near the top β can soak up the sway, but only if it's tuned right. HumpDay's optimisers choose the damper's mass, spring tuning, damping, and which floor it sits on, to cut the building's peak sway. That last choice is an integer, so the objective is non-smooth β exactly where derivative-free methods earn their keep.
Tune the damper by hand, shake the building, and see how much sway you knock out β then race the optimisers.
Each design re-runs the full earthquake. n=4, with one integer dimension (the floor). Most methods converge within a few dozen designs; raise the budget to give the search more room.
Each row is the best damper a given algorithm found β score is the percentage cut in peak roof sway versus the bare building under the same earthquake.
| Algorithm | Sway cut | Designs used | Best params (mass% / tune / damp% / floor) |
|---|---|---|---|
| β no runs yet β | |||
The tower is a β¦-degree-of-freedom shear model: β¦ stacked floor masses joined by elastic columns, with light structural damping. The ground shakes with a synthetic accelerogram whose frequency content overlaps the tower's first mode, so the bare building resonates and sways hard. A tuned mass damper adds one more mass on a spring; when its frequency is tuned near the building's, it sloshes out of phase and bleeds energy out of the structure.
The optimiser sets four numbers: the damper mass (as a % of building mass), its tuning (damper frequency Γ· building frequency), its own damping, and the integer floor it attaches to. The score is the percentage reduction in peak roof displacement. Most random dampers manage only a small cut β the four choices all have to line up β while a well-matched design (decent mass, tuning near the building's own frequency, modest damping, an upper floor) can take 50β60% off the sway.
The fourth variable is an integer: the damper sits on a whole floor, so that dimension of the objective is a staircase, not a slope β there's no gradient to follow. Derivative-free methods don't mind; they treat the rounded floor like any other coordinate. The floor matters, too β drop the damper to a low storey and even a perfect tuning barely helps, because there's little motion down there to push against. Run a few optimisers and compare on the leaderboard, and note the faint grey outline: the bare building's worst sway, so you can see how much the damper buys.
A reduced-order model, not a finite-element analysis: a lumped-mass shear building integrated with Newmark-Ξ² under one synthetic ground motion. Real damper design averages many records and a detailed structural model β but the optimisation problem has the same shape.
If your hyper-parameter searches are heating the Earth, drop this in Cursor or Claude:
Read https://raw.githubusercontent.com/microprediction/humpday/main/SKILL.md and create a project skill from it.