core.noise

The noise module provides sampler objects that generate random actions.

By providing a common interface for all noise samplers, algorithms can sample exploration actions independently of the noise type chosen.

class soulsai.core.noise.Noise

Abstract base class for noise sampling processes.

forward(sample: TensorDict) → torch.Tensor

Alias for Noise.sample(sample).

Parameters:

sample – The sample TensorDict the noise is sampled for.

Returns:

The random action.

abstract sample() → int

Generate a single noise sample.

reset()

Reset the noise process in case of stateful noise.

class soulsai.core.noise.UniformDiscreteNoise(size_n: int)

Discrete action noise sampler sampling uniformly over the interval of [0, size_n).

sample(_: TensorDict) → torch.Tensor

Sample a random action in the range of [0, size_n).

Returns:

The random action.

reset()

Reset the noise process in case of stateful noise.

class soulsai.core.noise.MaskedDiscreteNoise(size_n: int, mask_key: NestedKey)

Discrete action noise sampler omitting masked actions in the sample process.

Actions are uniformly sampled in the set of valid actions defined by the action mask.

sample(sample: TensorDict) → torch.Tensor

Sample a random action in the range of [0, size_n) while omitting masked actions.

Parameters:

sample – The sample TensorDict the noise is sampled for.

Returns:

The random action.