Policy

class Policy(feature_dim: int, action_space: gym.spaces.space.Space, **params)

Bases: pandemonium.utilities.registrable.Registrable

Base abstract class for decision making rules

A policy is a function \(\pi: \mathcal{X} \mapsto \mathcal{A}\) that maps from the space of features onto the space of actions.

Methods Summary

__call__(self, \*args, \*\*kwargs)	Call self as a function.
act(self, \*args, \*\*kwargs)	Samples an action from a distribution over actions
action_filter(self, state)	Filters the actions available at a given state
dist(self, \*args, \*\*kwargs)	Produces a distribution over actions

Methods Documentation

__call__(self, \*args, \*\*kwargs)

Call self as a function.

act(self, \*args, \*\*kwargs) → Tuple[ForwardRef(‘Action’), Dict[str, torch.distributions.distribution.Distribution]]

Samples an action from a distribution over actions

action_filter(self, state)

Filters the actions available at a given state

Closely relates to option initiation sets:

>>> [o for o in options if o.initiation(state) == 1]

See also

Interest Functions by K. Khetarpal et al. 2020 https://arxiv.org/pdf/2001.00271.pdf

See also

Ray RLlib has a working implementation https://ray.readthedocs.io/en/latest/rllib-models.html#variable-length-parametric-action-spaces

dist(self, \*args, \*\*kwargs) → torch.distributions.distribution.Distribution

Produces a distribution over actions