Module moog.action_spaces.grid
Discrete grid action space for controlling agent avatars.
Expand source code
"""Discrete grid action space for controlling agent avatars."""
from . import abstract_action_space
from dm_env import specs
import numpy as np
class Grid(abstract_action_space.AbstractActionSpace):
"""Discrete grid action space.
This action space has 5 actions {left, right, up, down, do-nothing}. These
actions control either the force or the velocity of the agent(s).
"""
_ACTIONS = (
np.array([-1, 0]), # Move left
np.array([1, 0]), # Move right
np.array([0, -1]), # Move down
np.array([0, 1]), # Move up
np.array([0, 0]), # Do not move
)
def __init__(self, scaling_factor=1., action_layers='agent',
control_velocity=False, momentum=0.):
"""Constructor.
Args:
scaling_factor: Scalar. Scaling factor multiplied to the action.
agent_layer: String or iterable of strings. Elements (or itself if
string) must be keys in the environment state. All sprites in
these layers will be acted upon by this action space.
control_velocity: Bool. Whether to control velocity (True) or force
(False).
momentum: Float in [0, 1]. Discount factor for previous action. This
should be zero if control_velocity is False, because imparting
forces automatically gives momentum to the agent(s) being
controlled. If control_velocity is True, setting this greater
than zero gives the controlled agent(s) momentum. However, the
velocity is clipped at scaling_factor, so the agent only retains
momentum when stopping or changing direction and does not
accelerate.
"""
self._scaling_factor = scaling_factor
if not isinstance(action_layers, (list, tuple)):
action_layers = (action_layers,)
self._action_layers = action_layers
self._control_velocity = control_velocity
self._momentum = momentum
self._action_spec = specs.DiscreteArray(len(self._ACTIONS))
def step(self, state, action):
"""Apply action to environment state.
Args:
state: Ordereddict of layers of sprites. Environment state.
action: Numpy float array of size (2). Force to apply.
"""
self._action *= self._momentum
self._action += Grid._ACTIONS[action]
self._action = np.clip(
self._action, -self._scaling_factor, self._scaling_factor)
for action_layer in self._action_layers:
for sprite in state[action_layer]:
if self._control_velocity:
sprite.velocity = self._action / sprite.mass
else:
sprite.velocity += self._action / sprite.mass
def reset(self, state):
"""Reset action space at start of new episode."""
del state
self._action = np.zeros(2)
def random_action(self):
"""Return randomly sampled action."""
return np.random.randint(len(Grid._ACTIONS))
def action_spec(self):
return self._action_specClasses
class Grid (scaling_factor=1.0, action_layers='agent', control_velocity=False, momentum=0.0)-
Discrete grid action space.
This action space has 5 actions {left, right, up, down, do-nothing}. These actions control either the force or the velocity of the agent(s).
Constructor.
Args
scaling_factor- Scalar. Scaling factor multiplied to the action.
agent_layer- String or iterable of strings. Elements (or itself if string) must be keys in the environment state. All sprites in these layers will be acted upon by this action space.
control_velocity- Bool. Whether to control velocity (True) or force (False).
momentum- Float in [0, 1]. Discount factor for previous action. This should be zero if control_velocity is False, because imparting forces automatically gives momentum to the agent(s) being controlled. If control_velocity is True, setting this greater than zero gives the controlled agent(s) momentum. However, the velocity is clipped at scaling_factor, so the agent only retains momentum when stopping or changing direction and does not accelerate.
Expand source code
class Grid(abstract_action_space.AbstractActionSpace): """Discrete grid action space. This action space has 5 actions {left, right, up, down, do-nothing}. These actions control either the force or the velocity of the agent(s). """ _ACTIONS = ( np.array([-1, 0]), # Move left np.array([1, 0]), # Move right np.array([0, -1]), # Move down np.array([0, 1]), # Move up np.array([0, 0]), # Do not move ) def __init__(self, scaling_factor=1., action_layers='agent', control_velocity=False, momentum=0.): """Constructor. Args: scaling_factor: Scalar. Scaling factor multiplied to the action. agent_layer: String or iterable of strings. Elements (or itself if string) must be keys in the environment state. All sprites in these layers will be acted upon by this action space. control_velocity: Bool. Whether to control velocity (True) or force (False). momentum: Float in [0, 1]. Discount factor for previous action. This should be zero if control_velocity is False, because imparting forces automatically gives momentum to the agent(s) being controlled. If control_velocity is True, setting this greater than zero gives the controlled agent(s) momentum. However, the velocity is clipped at scaling_factor, so the agent only retains momentum when stopping or changing direction and does not accelerate. """ self._scaling_factor = scaling_factor if not isinstance(action_layers, (list, tuple)): action_layers = (action_layers,) self._action_layers = action_layers self._control_velocity = control_velocity self._momentum = momentum self._action_spec = specs.DiscreteArray(len(self._ACTIONS)) def step(self, state, action): """Apply action to environment state. Args: state: Ordereddict of layers of sprites. Environment state. action: Numpy float array of size (2). Force to apply. """ self._action *= self._momentum self._action += Grid._ACTIONS[action] self._action = np.clip( self._action, -self._scaling_factor, self._scaling_factor) for action_layer in self._action_layers: for sprite in state[action_layer]: if self._control_velocity: sprite.velocity = self._action / sprite.mass else: sprite.velocity += self._action / sprite.mass def reset(self, state): """Reset action space at start of new episode.""" del state self._action = np.zeros(2) def random_action(self): """Return randomly sampled action.""" return np.random.randint(len(Grid._ACTIONS)) def action_spec(self): return self._action_specAncestors
- AbstractActionSpace
- abc.ABC
Methods
def random_action(self)-
Return randomly sampled action.
Expand source code
def random_action(self): """Return randomly sampled action.""" return np.random.randint(len(Grid._ACTIONS)) def reset(self, state)-
Reset action space at start of new episode.
Expand source code
def reset(self, state): """Reset action space at start of new episode.""" del state self._action = np.zeros(2) def step(self, state, action)-
Apply action to environment state.
Args
state- Ordereddict of layers of sprites. Environment state.
action- Numpy float array of size (2). Force to apply.
Expand source code
def step(self, state, action): """Apply action to environment state. Args: state: Ordereddict of layers of sprites. Environment state. action: Numpy float array of size (2). Force to apply. """ self._action *= self._momentum self._action += Grid._ACTIONS[action] self._action = np.clip( self._action, -self._scaling_factor, self._scaling_factor) for action_layer in self._action_layers: for sprite in state[action_layer]: if self._control_velocity: sprite.velocity = self._action / sprite.mass else: sprite.velocity += self._action / sprite.mass
Inherited members