Three expensive runs, then only what the bound can't rule out: picking a hanger width

Keeping the cheap model's own penalty makes the regret bound 2.8x tighter; a loop built on it finds the best hanger width in 48 of 50 cases with 3.9 of 7 runs.

The hanger-codesign programme asks how good a cloth simulator must be before the hanger it designs for a laundry robot is the right one. The last step proved a bound on how much a cheap simulator's choice can cost, and found it too loose to certify anything the fast 2D model picked. This step finds out why it was loose and turns the fixed version into a procedure.

What the old bound threw away. The cheap choice's regret on the expensive model is exactly the change in the paired error between the two designs, minus how much worse the cheap model thinks the true best design is. The first bound dropped that second term. For the 2D chain it is large, because the chain is sharp: it thinks the widths next to its pick are clearly worse. Keep it, bound the error's change by a slope K times the distance from the cheap pick, and maximise over the unknown best design. The result needs only the cheap scores and K.

How much tighter. On the 16 cases where the 2D chain's pick is not the 3D sheet's best, the realised regret is a median 34% of the new bound against 12% of the old one: 2.8 times tighter. With K estimated from just three expensive runs (the cheap pick and its two neighbours), the bound still covers the realised regret in 47 of 50 tier pairs, and all 50 if the estimate is doubled.

The loop. Run the expensive model at those three widths, estimate K, run every width the bound cannot rule out, update K, repeat. Over the 50 pairs from milestones 1b to 1d (no new simulation):

strategyexpensive runs (of 7)finds the bestmean regret
trust the cheap model123 of 500.028 cm
best of the three nearby runs3.047 of 500.0022 cm
screen and verify3.948 of 500.0010 cm
run everything750 of 500

Both misses are the softest cloth (stretch stiffness 40 N/m), where the best width sits 1 cm from the 2D pick and the error curves away, so the slope near the pick understates it. Doubling the screen catches them but, on the 2D chain, runs all seven widths.

Limits. Seven candidates is a small set, so the saving is a few runs; it should matter more in the continuous design loop that comes next. The 50 cases overlap, so there are no intervals, and the "expensive" model is still our own converged sheet, not a validated solver. Full tables: the project README, "Milestone 5b".

Simulation only; produced with AI assistance.

All writing · More in Research findings · All tools and research