Case studies¶
Each one is a script in
scripts/, and each page
gives the one command that reproduces it and shows what that command printed when this site was
built. The BOPTEST heat pump is the exception: it runs in causaldyn-bench against a live
emulator, and its page quotes the committed results of that run.
| case study | the decision | command |
|---|---|---|
| Pricing under confounding | steer a KPI to target from logs whose action was aimed at the customers most likely to churn | uv run python scripts/flagship_demo.py |
| Flatten the curve | the least intervention that holds an epidemic at hospital capacity | uv run python scripts/epidemic_demo.py |
| A pendulum from a confounded log | raise a pendulum and hold it, from a log whose torque an operator used to cancel the wind | uv run python scripts/pendulum_demo.py |
| Media budgets | which channel to spend on, and when, from logs planned against demand | uv run python scripts/mmm_demo.py |
| Marketplace dispatch | which drivers to send where, and the surge prices that fall out of the same optimisation | uv run python scripts/run_marketplace_demo.py |
| Driver supply, end to end | an incentive between two zones, from confounded logs to a certified plan run on the true plant | uv run python scripts/spine_demo.py |
| Confounded incentives, closed loop | an incentive under hidden confounding, when churn costs more than wasted budget | uv run python scripts/run_dynamic_confounding_demo.py |
| A heat pump from a weather-compensated log | hold a house at a target above its comfort bound, on a live building emulator, from the log a weather-compensated controller left behind | just boptest-prescribe-live in causaldyn-bench |
Run them from a checkout after uv sync. None of the scripts enables x64, so JAX runs them at its
default precision; see the dtype policy.