← All demos the budget solve · bb_core.js

The budget frontier

Expected maximum as a function of the path-time budget, adaptive against both baselines.

Every point is a full solve: dual bisection on the shadow price, wavefront policies from the obstacle problem, Frank–Wolfe mixing. The one-shot baseline keeps everyone until its best single screening date and then retains an upper tail; static thinning spends the same budget on a random subset chosen at time zero. Change the intensity and the sweep reruns.

100
computing…
What to look for: no baseline may exceed the adaptive curve by more than that point's certified suboptimality, and one-shot must never fall below static thinning; a larger crossing would falsify the solver, since both baselines are feasible policies for the same problem. The true ordering is adaptive ≥ one-shot ≥ static, with ties at both ends: at tight budgets the best screening date is time zero, when every path still sits at the same starting point, and at slack budgets the whole cloud is funded, so all three policies keep everyone. The adaptive curve should also be concave — each extra unit of budget buys less than the one before.

Discretization: 40 time steps on a 241-point state grid, chosen so the ten-point sweep completes in about a second. Finer grids shift the curves by a few hundredths without changing their order.