1
0
mirror of https://github.com/gryf/coach.git synced 2025-12-17 11:10:20 +01:00

Create a dataset using an agent (#306)

Generate a dataset using an agent (allowing to select between this and a random dataset)
This commit is contained in:
Gal Leibovich
2019-05-28 09:34:49 +03:00
committed by GitHub
parent 342b7184bc
commit 9e9c4fd332
26 changed files with 351 additions and 111 deletions

View File

@@ -19,12 +19,13 @@ from typing import List
import numpy as np
from rl_coach.core_types import RunPhase, ActionType
from rl_coach.exploration_policies.exploration_policy import ExplorationPolicy, ExplorationParameters
from rl_coach.exploration_policies.exploration_policy import ContinuousActionExplorationPolicy, ExplorationParameters
from rl_coach.spaces import ActionSpace, BoxActionSpace, GoalsSpace
# Based on on the description in:
# https://math.stackexchange.com/questions/1287634/implementing-ornstein-uhlenbeck-in-matlab
class OUProcessParameters(ExplorationParameters):
def __init__(self):
super().__init__()
@@ -39,7 +40,7 @@ class OUProcessParameters(ExplorationParameters):
# Ornstein-Uhlenbeck process
class OUProcess(ExplorationPolicy):
class OUProcess(ContinuousActionExplorationPolicy):
"""
OUProcess exploration policy is intended for continuous action spaces, and selects the action according to
an Ornstein-Uhlenbeck process. The Ornstein-Uhlenbeck process implements the action as a Gaussian process, where
@@ -56,10 +57,6 @@ class OUProcess(ExplorationPolicy):
self.state = np.zeros(self.action_space.shape)
self.dt = dt
if not (isinstance(action_space, BoxActionSpace) or isinstance(action_space, GoalsSpace)):
raise ValueError("OU process exploration works only for continuous controls."
"The given action space is of type: {}".format(action_space.__class__.__name__))
def reset(self):
self.state = np.zeros(self.action_space.shape)