reference/generated/torchrl.collectors.bootstrap_estimate

Run in Google Colab

Colab

Download Notebook

Notebook

View on GitHub

GitHub

bootstrap_estimate

class torchrl.collectors.bootstrap_estimate(values: Tensor, dim: int = -1, *, statistic: Literal['mean', 'median', 'iqm'] = 'iqm', confidence: float = 0.95, num_resamples: int = 1000, generator: Generator | None = None)

Reduce values along dim to an estimate with a percentile bootstrap interval.

Parameters:

• values (torch.Tensor) – the samples, e.g. per-episode returns.

• dim (int, optional) – the dimension holding the samples. Defaults to -1.

Keyword Arguments:

• statistic (str, optional) – "mean", "median" or "iqm" (the mean of the samples left after dropping the lowest and highest 25%). Defaults to "iqm".

• confidence (float, optional) – the coverage of the interval, in (0, 1). Defaults to 0.95.

• num_resamples (int, optional) – the number of bootstrap resamples, each drawn with replacement. Defaults to 1000.

• generator (torch.Generator, optional) – the generator used to draw resamples, for reproducible intervals. Defaults to None.

Returns:

a TensorDict with entries "estimate", "ci_low" and "ci_high" whose batch size is the shape of values without dim.

Examples

>>> import torch
>>> from torchrl.collectors import bootstrap_estimate
>>> returns = torch.tensor([[0.0, 0.0, 0.0, 100.0], [5.0, 5.0, 5.0, 5.0]])
>>> summary = bootstrap_estimate(returns, statistic="iqm")
>>> summary["estimate"]
tensor([0., 5.])
>>> summary["ci_low"][1], summary["ci_high"][1]
(tensor(5.), tensor(5.))