d3rlpy.preprocessing.MinMaxRewardScaler¶

class d3rlpy.preprocessing.MinMaxRewardScaler(dataset=None, minimum=None, maximum=None)[source]¶

Min-Max reward normalization preprocessing.

\[r' = (r - \min(r)) / (\max(r) - \min(r))\]

from d3rlpy.algos import CQL

cql = CQL(reward_scaler="min_max")

You can also initialize with d3rlpy.dataset.MDPDataset object or manually.

from d3rlpy.preprocessing import MinMaxRewardScaler

# initialize with dataset
scaler = MinMaxRewardScaler(dataset)

# initialize manually
scaler = MinMaxRewardScaler(minimum=0.0, maximum=10.0)

cql = CQL(scaler=scaler)

Methods

Parameters

dataset (Optional[d3rlpy.dataset.MDPDataset]) –
minimum (Optional[float]) –
maximum (Optional[float]) –

fit(episodes)[source]¶

Estimates scaling parameters from dataset.

Parameters: episodes (List[d3rlpy.dataset.Episode]) – list of episodes.
Return type: None

fit_with_env(env)¶

Gets scaling parameters from environment.

Note

RewardScaler does not support fitting with environment.

Parameters: env (gym.core.Env) – gym environment.
Return type: None

get_params(deep=False)[source]¶

Returns scaling parameters.

Parameters: deep (bool) – flag to deeply copy objects.
Returns: scaler parameters.
Return type: Dict[str, Any]

get_type()¶

Returns a scaler type.

Returns: scaler type.
Return type: str

reverse_transform(reward)[source]¶

Returns reversely processed rewards.

Parameters: reward (torch.Tensor) – reward.
Returns: reversely processed reward.
Return type: torch.Tensor

transform(reward)[source]¶

Returns processed rewards.

Parameters: reward (torch.Tensor) – reward.
Returns: processed reward.
Return type: torch.Tensor

transform_numpy(reward)[source]¶

Returns transformed rewards in numpy array.

Parameters: reward (numpy.ndarray) – reward.
Returns: transformed reward.
Return type: numpy.ndarray

Attributes

TYPE: ClassVar[str] = 'min_max'¶