"""Implementation of a space that represents closed boxes in euclidean space."""
from __future__ import annotations
from typing import Any, Iterable, Mapping, Sequence, SupportsFloat
import numpy as np
from numpy.typing import NDArray
import gymnasium as gym
from gymnasium.spaces.space import Space
def _short_repr(arr: NDArray[Any]) -> str:
"""Create a shortened string representation of a numpy array.
If arr is a multiple of the all-ones vector, return a string representation of the multiplier.
Otherwise, return a string representation of the entire array.
Args:
arr: The array to represent
Returns:
A short representation of the array
"""
if arr.size != 0 and np.min(arr) == np.max(arr):
return str(np.min(arr))
return str(arr)
def is_float_integer(var: Any) -> bool:
"""Checks if a variable is an integer or float."""
return np.issubdtype(type(var), np.integer) or np.issubdtype(type(var), np.floating)
[docs]class Box(Space[NDArray[Any]]):
r"""A (possibly unbounded) box in :math:`\mathbb{R}^n`.
Specifically, a Box represents the Cartesian product of n closed intervals.
Each interval has the form of one of :math:`[a, b]`, :math:`(-\infty, b]`,
:math:`[a, \infty)`, or :math:`(-\infty, \infty)`.
There are two common use cases:
* Identical bound for each dimension::
>>> Box(low=-1.0, high=2.0, shape=(3, 4), dtype=np.float32)
Box(-1.0, 2.0, (3, 4), float32)
* Independent bound for each dimension::
>>> Box(low=np.array([-1.0, -2.0]), high=np.array([2.0, 4.0]), dtype=np.float32)
Box([-1. -2.], [2. 4.], (2,), float32)
"""
def __init__(
self,
low: SupportsFloat | NDArray[Any],
high: SupportsFloat | NDArray[Any],
shape: Sequence[int] | None = None,
dtype: type[np.floating[Any]] | type[np.integer[Any]] = np.float32,
seed: int | np.random.Generator | None = None,
):
r"""Constructor of :class:`Box`.
The argument ``low`` specifies the lower bound of each dimension and ``high`` specifies the upper bounds.
I.e., the space that is constructed will be the product of the intervals :math:`[\text{low}[i], \text{high}[i]]`.
If ``low`` (or ``high``) is a scalar, the lower bound (or upper bound, respectively) will be assumed to be
this value across all dimensions.
Args:
low (SupportsFloat | np.ndarray): Lower bounds of the intervals. If integer, must be at least ``-2**63``.
high (SupportsFloat | np.ndarray]): Upper bounds of the intervals. If integer, must be at most ``2**63 - 2``.
shape (Optional[Sequence[int]]): The shape is inferred from the shape of `low` or `high` `np.ndarray`s with
`low` and `high` scalars defaulting to a shape of (1,)
dtype: The dtype of the elements of the space. If this is an integer type, the :class:`Box` is essentially a discrete space.
seed: Optionally, you can use this argument to seed the RNG that is used to sample from the space.
Raises:
ValueError: If no shape information is provided (shape is None, low is None and high is None) then a
value error is raised.
"""
assert (
dtype is not None
), "Box dtype must be explicitly provided, cannot be None."
self.dtype = np.dtype(dtype)
# determine shape if it isn't provided directly
if shape is not None:
assert all(
np.issubdtype(type(dim), np.integer) for dim in shape
), f"Expected all shape elements to be an integer, actual type: {tuple(type(dim) for dim in shape)}"
shape = tuple(int(dim) for dim in shape) # This changes any np types to int
elif isinstance(low, np.ndarray):
shape = low.shape
elif isinstance(high, np.ndarray):
shape = high.shape
elif is_float_integer(low) and is_float_integer(high):
shape = (1,)
else:
raise ValueError(
f"Box shape is inferred from low and high, expected their types to be np.ndarray, an integer or a float, actual type low: {type(low)}, high: {type(high)}"
)
# Capture the boundedness information before replacing np.inf with get_inf
_low = np.full(shape, low, dtype=float) if is_float_integer(low) else low
self.bounded_below: NDArray[np.bool_] = -np.inf < _low
_high = np.full(shape, high, dtype=float) if is_float_integer(high) else high
self.bounded_above: NDArray[np.bool_] = np.inf > _high
low = _broadcast(low, self.dtype, shape)
high = _broadcast(high, self.dtype, shape)
assert isinstance(low, np.ndarray)
assert (
low.shape == shape
), f"low.shape doesn't match provided shape, low.shape: {low.shape}, shape: {shape}"
assert isinstance(high, np.ndarray)
assert (
high.shape == shape
), f"high.shape doesn't match provided shape, high.shape: {high.shape}, shape: {shape}"
# check that we don't have invalid low or high
if np.any(low > high):
raise ValueError(
f"Some low values are greater than high, low={low}, high={high}"
)
if np.any(np.isposinf(low)):
raise ValueError(f"No low value can be equal to `np.inf`, low={low}")
if np.any(np.isneginf(high)):
raise ValueError(f"No high value can be equal to `-np.inf`, high={high}")
self._shape: tuple[int, ...] = shape
low_precision = get_precision(low.dtype)
high_precision = get_precision(high.dtype)
dtype_precision = get_precision(self.dtype)
if min(low_precision, high_precision) > dtype_precision:
gym.logger.warn(f"Box bound precision lowered by casting to {self.dtype}")
self.low = low.astype(self.dtype)
self.high = high.astype(self.dtype)
self.low_repr = _short_repr(self.low)
self.high_repr = _short_repr(self.high)
super().__init__(self.shape, self.dtype, seed)
@property
def shape(self) -> tuple[int, ...]:
"""Has stricter type than gym.Space - never None."""
return self._shape
@property
def is_np_flattenable(self):
"""Checks whether this space can be flattened to a :class:`spaces.Box`."""
return True
[docs] def is_bounded(self, manner: str = "both") -> bool:
"""Checks whether the box is bounded in some sense.
Args:
manner (str): One of ``"both"``, ``"below"``, ``"above"``.
Returns:
If the space is bounded
Raises:
ValueError: If `manner` is neither ``"both"`` nor ``"below"`` or ``"above"``
"""
below = bool(np.all(self.bounded_below))
above = bool(np.all(self.bounded_above))
if manner == "both":
return below and above
elif manner == "below":
return below
elif manner == "above":
return above
else:
raise ValueError(
f"manner is not in {{'below', 'above', 'both'}}, actual value: {manner}"
)
[docs] def sample(self, mask: None = None) -> NDArray[Any]:
r"""Generates a single random sample inside the Box.
In creating a sample of the box, each coordinate is sampled (independently) from a distribution
that is chosen according to the form of the interval:
* :math:`[a, b]` : uniform distribution
* :math:`[a, \infty)` : shifted exponential distribution
* :math:`(-\infty, b]` : shifted negative exponential distribution
* :math:`(-\infty, \infty)` : normal distribution
Args:
mask: A mask for sampling values from the Box space, currently unsupported.
Returns:
A sampled value from the Box
"""
if mask is not None:
raise gym.error.Error(
f"Box.sample cannot be provided a mask, actual value: {mask}"
)
high = self.high if self.dtype.kind == "f" else self.high.astype("int64") + 1
sample = np.empty(self.shape)
# Masking arrays which classify the coordinates according to interval type
unbounded = ~self.bounded_below & ~self.bounded_above
upp_bounded = ~self.bounded_below & self.bounded_above
low_bounded = self.bounded_below & ~self.bounded_above
bounded = self.bounded_below & self.bounded_above
# Vectorized sampling by interval type
sample[unbounded] = self.np_random.normal(size=unbounded[unbounded].shape)
sample[low_bounded] = (
self.np_random.exponential(size=low_bounded[low_bounded].shape)
+ self.low[low_bounded]
)
sample[upp_bounded] = (
-self.np_random.exponential(size=upp_bounded[upp_bounded].shape)
+ high[upp_bounded]
)
sample[bounded] = self.np_random.uniform(
low=self.low[bounded], high=high[bounded], size=bounded[bounded].shape
)
if self.dtype.kind in ["i", "u", "b"]:
sample = np.floor(sample)
return sample.astype(self.dtype)
def contains(self, x: Any) -> bool:
"""Return boolean specifying if x is a valid member of this space."""
if not isinstance(x, np.ndarray):
gym.logger.warn("Casting input x to numpy array.")
try:
x = np.asarray(x, dtype=self.dtype)
except (ValueError, TypeError):
return False
return bool(
np.can_cast(x.dtype, self.dtype)
and x.shape == self.shape
and np.all(x >= self.low)
and np.all(x <= self.high)
)
def to_jsonable(self, sample_n: Sequence[NDArray[Any]]) -> list[list]:
"""Convert a batch of samples from this space to a JSONable data type."""
return [sample.tolist() for sample in sample_n]
def from_jsonable(self, sample_n: Sequence[float | int]) -> list[NDArray[Any]]:
"""Convert a JSONable data type to a batch of samples from this space."""
return [np.asarray(sample, dtype=self.dtype) for sample in sample_n]
def __repr__(self) -> str:
"""A string representation of this space.
The representation will include bounds, shape and dtype.
If a bound is uniform, only the corresponding scalar will be given to avoid redundant and ugly strings.
Returns:
A representation of the space
"""
return f"Box({self.low_repr}, {self.high_repr}, {self.shape}, {self.dtype})"
def __eq__(self, other: Any) -> bool:
"""Check whether `other` is equivalent to this instance. Doesn't check dtype equivalence."""
return (
isinstance(other, Box)
and (self.shape == other.shape)
# and (self.dtype == other.dtype)
and np.allclose(self.low, other.low)
and np.allclose(self.high, other.high)
)
def __setstate__(self, state: Iterable[tuple[str, Any]] | Mapping[str, Any]):
"""Sets the state of the box for unpickling a box with legacy support."""
super().__setstate__(state)
# legacy support through re-adding "low_repr" and "high_repr" if missing from pickled state
if not hasattr(self, "low_repr"):
self.low_repr = _short_repr(self.low)
if not hasattr(self, "high_repr"):
self.high_repr = _short_repr(self.high)
def get_precision(dtype: np.dtype) -> SupportsFloat:
"""Get precision of a data type."""
if np.issubdtype(dtype, np.floating):
return np.finfo(dtype).precision
else:
return np.inf
def _broadcast(
value: SupportsFloat | NDArray[Any],
dtype: np.dtype,
shape: tuple[int, ...],
) -> NDArray[Any]:
"""Handle infinite bounds and broadcast at the same time if needed.
This is needed primarily because:
>>> import numpy as np
>>> np.full((2,), np.inf, dtype=np.int32)
array([-2147483648, -2147483648], dtype=int32)
"""
if is_float_integer(value):
if np.isneginf(value) and np.dtype(dtype).kind == "i":
value = np.iinfo(dtype).min + 2
elif np.isposinf(value) and np.dtype(dtype).kind == "i":
value = np.iinfo(dtype).max - 2
return np.full(shape, value, dtype=dtype)
elif isinstance(value, np.ndarray):
# this is needed because we can't stuff np.iinfo(int).min into an array of dtype float
casted_value = value.astype(dtype)
# change bounds only if values are negative or positive infinite
if np.dtype(dtype).kind == "i":
casted_value[np.isneginf(value)] = np.iinfo(dtype).min + 2
casted_value[np.isposinf(value)] = np.iinfo(dtype).max - 2
return casted_value
else:
# only np.ndarray allowed beyond this point
raise TypeError(
f"Unknown dtype for `value`, expected `np.ndarray` or float/integer, got {type(value)}"
)
