An Ebola outbreak spreads through a population (an SEIR model). You can deploy control effort — case isolation, contact tracing, safe burials, ring vaccination — which cuts transmission but is costly to sustain. The optimiser sets how hard to push in each of 8 time windows to minimise total harm = deaths + control cost. The lesson it learns is the epidemiologist's: don't blanket-lock everything, and don't act late — hit it hard during the early growth phase, then ease off.
Set the control intensity for each window and see the outbreak respond.
8-D problem (control intensity per window). Score is the % of total harm avoided versus letting the outbreak run unchecked.
Each row is the best response a given optimiser found — % of harm (deaths + control cost) avoided versus no intervention. Compare the control profiles: the good ones spike early, then back off.
| Algorithm | Harm avoided | Policies | Control per window |
|---|---|---|---|
| — no runs yet — | |||
The outbreak follows an SEIR model: Susceptible → Exposed (incubating) → Infectious → Removed (recovered or, for Ebola, often died). Left unchecked it burns through the population, infecting a large fraction and — at a 60% case-fatality — killing many. Control effort in a window cuts transmission there, but every unit of effort costs (economic disruption, resources), so blanketing all eight windows at full strength is hugely expensive for little extra benefit.
The optimiser tunes the eight intensities to minimise total harm. The response it finds is sharp and non-obvious: almost nothing in the first window (there's little to suppress yet), then maximum effort through the growth phase to break the exponential, then stand down once it's contained. Act too late and the curve has already exploded; hold control forever and you pay for suppression you no longer need. The red area is the infectious curve; the faint red line is the unchecked outbreak; the blue backdrop is how hard control is applied over time.
A compartmental toy model (R₀ ≈ 2.2, ~9-day incubation, ~10-day infectious period, 60% CFR), not a calibrated forecast of any real outbreak — but the control-timing trade-off it captures is real.
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.