forked from IntelLabs/coach
-
Notifications
You must be signed in to change notification settings - Fork 0
/
prioritized_experience_replay.py
299 lines (252 loc) · 12.3 KB
/
prioritized_experience_replay.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
#
# Copyright (c) 2017 Intel Corporation
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
#
import operator
import random
from enum import Enum
from typing import List, Tuple, Any
import numpy as np
from rl_coach.core_types import Transition
from rl_coach.memories.memory import MemoryGranularity
from rl_coach.memories.non_episodic.experience_replay import ExperienceReplayParameters, ExperienceReplay
from rl_coach.schedules import Schedule, ConstantSchedule
class PrioritizedExperienceReplayParameters(ExperienceReplayParameters):
def __init__(self):
super().__init__()
self.max_size = (MemoryGranularity.Transitions, 1000000)
self.alpha = 0.6
self.beta = ConstantSchedule(0.4)
self.epsilon = 1e-6
@property
def path(self):
return 'rl_coach.memories.non_episodic.prioritized_experience_replay:PrioritizedExperienceReplay'
class SegmentTree(object):
"""
A tree which can be used as a min/max heap or a sum tree
Add or update item value - O(log N)
Sampling an item - O(log N)
"""
class Operation(Enum):
MAX = {"operator": max, "initial_value": -float("inf")}
MIN = {"operator": min, "initial_value": float("inf")}
SUM = {"operator": operator.add, "initial_value": 0}
def __init__(self, size: int, operation: Operation):
self.next_leaf_idx_to_write = 0
self.size = size
if not (size > 0 and size & (size - 1) == 0):
raise ValueError("A segment tree size must be a positive power of 2. The given size is {}".format(self.size))
self.operation = operation
self.tree = np.ones(2 * size - 1) * self.operation.value['initial_value']
self.data = [None] * size
def _propagate(self, node_idx: int) -> None:
"""
Propagate an update of a node's value to its parent node
:param node_idx: the index of the node that was updated
:return: None
"""
parent = (node_idx - 1) // 2
self.tree[parent] = self.operation.value['operator'](self.tree[parent * 2 + 1], self.tree[parent * 2 + 2])
if parent != 0:
self._propagate(parent)
def _retrieve(self, root_node_idx: int, val: float)-> int:
"""
Retrieve the first node that has a value larger than val and is a child of the node at index idx
:param root_node_idx: the index of the root node to search from
:param val: the value to query for
:return: the index of the resulting node
"""
left = 2 * root_node_idx + 1
right = left + 1
if left >= len(self.tree):
return root_node_idx
if val <= self.tree[left]:
return self._retrieve(left, val)
else:
return self._retrieve(right, val-self.tree[left])
def total_value(self) -> float:
"""
Return the total value of the tree according to the tree operation. For SUM for example, this will return
the total sum of the tree. for MIN, this will return the minimal value
:return: the total value of the tree
"""
return self.tree[0]
def add(self, val: float, data: Any) -> None:
"""
Add a new value to the tree with data assigned to it
:param val: the new value to add to the tree
:param data: the data that should be assigned to this value
:return: None
"""
self.data[self.next_leaf_idx_to_write] = data
self.update(self.next_leaf_idx_to_write, val)
self.next_leaf_idx_to_write += 1
if self.next_leaf_idx_to_write >= self.size:
self.next_leaf_idx_to_write = 0
def update(self, leaf_idx: int, new_val: float) -> None:
"""
Update the value of the node at index idx
:param leaf_idx: the index of the node to update
:param new_val: the new value of the node
:return: None
"""
node_idx = leaf_idx + self.size - 1
if not 0 <= node_idx < len(self.tree):
raise ValueError("The given left index ({}) can not be found in the tree. The available leaves are: 0-{}"
.format(leaf_idx, self.size - 1))
self.tree[node_idx] = new_val
self._propagate(node_idx)
def get_element_by_partial_sum(self, val: float) -> Tuple[int, float, Any]:
"""
Given a value between 0 and the tree sum, return the object which this value is in it's range.
For example, if we have 3 leaves: 10, 20, 30, and val=35, this will return the 3rd leaf, by accumulating
leaves by their order until getting to 35. This allows sampling leaves according to their proportional
probability.
:param val: a value within the range 0 and the tree sum
:return: the index of the resulting leaf in the tree, its probability and
the object itself
"""
node_idx = self._retrieve(0, val)
leaf_idx = node_idx - self.size + 1
data_value = self.tree[node_idx]
data = self.data[leaf_idx]
return leaf_idx, data_value, data
def __str__(self):
result = ""
start = 0
size = 1
while size <= self.size:
result += "{}\n".format(self.tree[start:(start + size)])
start += size
size *= 2
return result
class PrioritizedExperienceReplay(ExperienceReplay):
"""
This is the proportional sampling variant of the prioritized experience replay as described
in https://arxiv.org/pdf/1511.05952.pdf.
"""
def __init__(self, max_size: Tuple[MemoryGranularity, int], alpha: float=0.6, beta: Schedule=ConstantSchedule(0.4),
epsilon: float=1e-6, allow_duplicates_in_batch_sampling: bool=True):
"""
:param max_size: the maximum number of transitions or episodes to hold in the memory
:param alpha: the alpha prioritization coefficient
:param beta: the beta parameter used for importance sampling
:param epsilon: a small value added to the priority of each transition
:param allow_duplicates_in_batch_sampling: allow having the same transition multiple times in a batch
"""
if max_size[0] != MemoryGranularity.Transitions:
raise ValueError("Prioritized Experience Replay currently only support setting the memory size in "
"transitions granularity.")
self.power_of_2_size = 1
while self.power_of_2_size < max_size[1]:
self.power_of_2_size *= 2
super().__init__((MemoryGranularity.Transitions, self.power_of_2_size), allow_duplicates_in_batch_sampling)
self.sum_tree = SegmentTree(self.power_of_2_size, SegmentTree.Operation.SUM)
self.min_tree = SegmentTree(self.power_of_2_size, SegmentTree.Operation.MIN)
self.max_tree = SegmentTree(self.power_of_2_size, SegmentTree.Operation.MAX)
self.alpha = alpha
self.beta = beta
self.epsilon = epsilon
self.maximal_priority = 1.0
def _update_priority(self, leaf_idx: int, error: float) -> None:
"""
Update the priority of a given transition, using its index in the tree and its error
:param leaf_idx: the index of the transition leaf in the tree
:param error: the new error value
:return: None
"""
if error < 0:
raise ValueError("The priorities must be non-negative values")
priority = (error + self.epsilon)
self.sum_tree.update(leaf_idx, priority ** self.alpha)
self.min_tree.update(leaf_idx, priority ** self.alpha)
self.max_tree.update(leaf_idx, priority)
self.maximal_priority = self.max_tree.total_value()
def update_priorities(self, indices: List[int], error_values: List[float]) -> None:
"""
Update the priorities of a batch of transitions using their indices and their new TD error terms
:param indices: the indices of the transitions to update
:param error_values: the new error values
:return: None
"""
self.reader_writer_lock.lock_writing_and_reading()
if len(indices) != len(error_values):
raise ValueError("The number of indexes requested for update don't match the number of error values given")
for transition_idx, error in zip(indices, error_values):
self._update_priority(transition_idx, error)
self.reader_writer_lock.release_writing_and_reading()
def sample(self, size: int) -> List[Transition]:
"""
Sample a batch of transitions form the replay buffer. If the requested size is larger than the number
of samples available in the replay buffer then the batch will return empty.
:param size: the size of the batch to sample
:return: a batch (list) of selected transitions from the replay buffer
"""
self.reader_writer_lock.lock_writing()
if self.num_transitions() >= size:
# split the tree leaves to equal segments and sample one transition from each segment
batch = []
segment_size = self.sum_tree.total_value() / size
# get the maximum weight in the memory
min_probability = self.min_tree.total_value() / self.sum_tree.total_value() # min P(j) = min p^a / sum(p^a)
max_weight = (min_probability * self.num_transitions()) ** -self.beta.current_value # max wi
# sample a batch
for i in range(size):
segment_start = segment_size * i
segment_end = segment_size * (i + 1)
# sample leaf and calculate its weight
val = random.uniform(segment_start, segment_end)
leaf_idx, priority, transition = self.sum_tree.get_element_by_partial_sum(val)
priority /= self.sum_tree.total_value() # P(j) = p^a / sum(p^a)
weight = (self.num_transitions() * priority) ** -self.beta.current_value # (N * P(j)) ^ -beta
normalized_weight = weight / max_weight # wj = ((N * P(j)) ^ -beta) / max wi
transition.info['idx'] = leaf_idx
transition.info['weight'] = normalized_weight
batch.append(transition)
self.beta.step()
else:
raise ValueError("The replay buffer cannot be sampled since there are not enough transitions yet. "
"There are currently {} transitions".format(self.num_transitions()))
self.reader_writer_lock.release_writing()
return batch
def store(self, transition: Transition, lock=True) -> None:
"""
Store a new transition in the memory.
:param transition: a transition to store
:return: None
"""
# Calling super.store() so that in case a memory backend is used, the memory backend can store this transition.
super().store(transition)
if lock:
self.reader_writer_lock.lock_writing_and_reading()
transition_priority = self.maximal_priority
self.sum_tree.add(transition_priority ** self.alpha, transition)
self.min_tree.add(transition_priority ** self.alpha, transition)
self.max_tree.add(transition_priority, transition)
super().store(transition, False)
if lock:
self.reader_writer_lock.release_writing_and_reading()
def clean(self, lock=True) -> None:
"""
Clean the memory by removing all the episodes
:return: None
"""
if lock:
self.reader_writer_lock.lock_writing_and_reading()
super().clean(lock=False)
self.sum_tree = SegmentTree(self.power_of_2_size, SegmentTree.Operation.SUM)
self.min_tree = SegmentTree(self.power_of_2_size, SegmentTree.Operation.MIN)
self.max_tree = SegmentTree(self.power_of_2_size, SegmentTree.Operation.MAX)
if lock:
self.reader_writer_lock.release_writing_and_reading()