gryf/coach
1
0
Fork 0
mirror of https://github.com/gryf/coach.git synced 2025-12-17 11:10:20 +01:00
Code Issues Projects Releases Wiki Activity

TD3 (#338)

Browse Source
This commit is contained in:
Gal Leibovich
2019-06-16 11:11:21 +03:00
committed by GitHub
parent 8df3c46756
commit 7eb884c5b2
107 changed files with 2200 additions and 495 deletions
Download Patch File Download Diff File Expand all files Collapse all files

46
rl_coach/exploration_policies/additive_noise.py
View File

@@ -17,7 +17,6 @@
from typing import List
import numpy as np
import scipy.stats
from rl_coach.core_types import RunPhase, ActionType
from rl_coach.exploration_policies.exploration_policy import ContinuousActionExplorationPolicy, ExplorationParameters
@@ -31,8 +30,9 @@ from rl_coach.spaces import ActionSpace, BoxActionSpace
class AdditiveNoiseParameters(ExplorationParameters):
def __init__(self):
super().__init__()
self.noise_percentage_schedule = LinearSchedule(0.1, 0.1, 50000)
self.evaluation_noise_percentage = 0.05
self.noise_schedule = LinearSchedule(0.1, 0.1, 50000)
self.evaluation_noise = 0.05
self.noise_as_percentage_from_action_space = True
@property
def path(self):
@@ -48,17 +48,19 @@ class AdditiveNoise(ContinuousActionExplorationPolicy):
2. Specified by the agents action. In case the agents action is a list with 2 values, the 1st one is assumed to
be the mean of the action, and 2nd is assumed to be its standard deviation.
"""
def __init__(self, action_space: ActionSpace, noise_percentage_schedule: Schedule,
evaluation_noise_percentage: float):
def __init__(self, action_space: ActionSpace, noise_schedule: Schedule,
evaluation_noise: float, noise_as_percentage_from_action_space: bool = True):
"""
:param action_space: the action space used by the environment
:param noise_percentage_schedule: the schedule for the noise variance percentage relative to the absolute range
of the action space
:param evaluation_noise_percentage: the noise variance percentage that will be used during evaluation phases
:param noise_schedule: the schedule for the noise
:param evaluation_noise: the noise variance that will be used during evaluation phases
:param noise_as_percentage_from_action_space: a bool deciding whether the noise is absolute or as a percentage
from the action space
"""
super().__init__(action_space)
self.noise_percentage_schedule = noise_percentage_schedule
self.evaluation_noise_percentage = evaluation_noise_percentage
self.noise_schedule = noise_schedule
self.evaluation_noise = evaluation_noise
self.noise_as_percentage_from_action_space = noise_as_percentage_from_action_space
if not isinstance(action_space, BoxActionSpace):
raise ValueError("Additive noise exploration works only for continuous controls."
@@ -68,19 +70,20 @@ class AdditiveNoise(ContinuousActionExplorationPolicy):
or not np.all(-np.inf < action_space.low) or not np.all(action_space.low < np.inf):
raise ValueError("Additive noise exploration requires bounded actions")
# TODO: allow working with unbounded actions by defining the noise in terms of range and not percentage
def get_action(self, action_values: List[ActionType]) -> ActionType:
# TODO-potential-bug consider separating internally defined stdev and externally defined stdev into 2 policies
# set the current noise percentage
# set the current noise
if self.phase == RunPhase.TEST:
current_noise_precentage = self.evaluation_noise_percentage
current_noise = self.evaluation_noise
else:
current_noise_precentage = self.noise_percentage_schedule.current_value
current_noise = self.noise_schedule.current_value
# scale the noise to the action space range
action_values_std = current_noise_precentage * (self.action_space.high - self.action_space.low)
if self.noise_as_percentage_from_action_space:
action_values_std = current_noise * (self.action_space.high - self.action_space.low)
else:
action_values_std = current_noise
# extract the mean values
if isinstance(action_values, list):
@@ -92,15 +95,18 @@ class AdditiveNoise(ContinuousActionExplorationPolicy):
# step the noise schedule
if self.phase is not RunPhase.TEST:
self.noise_percentage_schedule.step()
self.noise_schedule.step()
# the second element of the list is assumed to be the standard deviation
if isinstance(action_values, list) and len(action_values) > 1:
action_values_std = action_values[1].squeeze()
# add noise to the action means
action = np.random.normal(action_values_mean, action_values_std)
if self.phase is not RunPhase.TEST:
action = np.random.normal(action_values_mean, action_values_std)
else:
action = action_values_mean
return action
return np.atleast_1d(action)
def get_control_param(self):
return np.ones(self.action_space.shape)*self.noise_percentage_schedule.current_value
return np.ones(self.action_space.shape)*self.noise_schedule.current_value