Export

Requires pip install "ehr2rl[d3rlpy]". See Export to d3rlpy for discrete and continuous actions.

ehr2rl.to_d3rlpy(dataset, provenance_path=None)

Convert an EHRDataset to d3rlpy’s MDPDataset.

Integer actions whose trajectories declare metadata["action_sizes"] are exported as one discrete action with the full declared action space, so actions that never occur in the data are still part of it. Several action columns are combined into one joint index in row-major order; recover the per-column bins with numpy.unravel_index(action, action_sizes). Floating-point actions are exported unchanged as continuous actions.

Parameters:

  • dataset (EHRDataset) – Non-empty dataset with rewards already shaped.

  • provenance_path (str | Path | None) – Where to write a JSON provenance sidecar. Every trajectory must carry the same metadata["provenance"].

Returns:

One episode per trajectory.

Return type:

d3rlpy.dataset.MDPDataset

Raises:

  • ImportError – If the d3rlpy extra is not installed.

  • ValueError – If the dataset is empty, multi-column integer actions lack action_sizes, sizes differ between trajectories, or an action is outside its declared size.

  • TypeError – If provenance_path is given and a trajectory has no provenance.