colosseum.agent.actors.base

 1import abc
 2import random
 3from typing import TYPE_CHECKING
 4
 5import dm_env
 6import numpy as np
 7
 8from colosseum.utils.acme.specs import MDPSpec
 9
10if TYPE_CHECKING:
11    from colosseum.mdp import ACTION_TYPE
12
13
14class BaseActor(abc.ABC):
15    """
16    The `BaseActor` class is the abstract class for the actor component of a reinforcement learning agent, which
17    handles the interactions with the MDPs.
18    """
19
20    @abc.abstractmethod
21    def __init__(self, seed: int, mdp_specs: MDPSpec):
22        """
23        Parameters
24        ----------
25        seed : int
26            The random seed.
27        mdp_specs : MDPSpec
28            The full specification of the MDP.
29        """
30        self._mdp_spec = mdp_specs
31        self._seed = seed
32
33        self._rng = np.random.RandomState(seed)
34        self._rng_fast = random.Random(seed)
35
36    @abc.abstractmethod
37    def select_action(self, ts: dm_env.TimeStep, time: int) -> "ACTION_TYPE":
38        """
39        Parameters
40        ----------
41        ts : dm_env.TimeStep
42            The TimeStep for which the agent is required to calculate the next action.
43        time : int
44            The current time of the environment. In the episodic case, this refers to the in-episode time, whereas in
45            the continuous case this refers to the total number of previous interactions.
46
47        Returns
48        -------
49        action : ACTION_TYPE
50            The action that the agent suggests to take given the observation and the time step.
51        """