-
Notifications
You must be signed in to change notification settings - Fork 0
/
search_tree.py
528 lines (422 loc) · 18 KB
/
search_tree.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
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
import time
import numpy as np
from numba import njit, int32, int8, uint64, void, int16, float64, boolean, prange
from numba.types import UniTuple
from numba.experimental import jitclass, structref
from numba.extending import overload, overload_method
from numba import types
from bitboard_utils import get_moves_index, place_disks, possible_moves, count_bits, make_move
from constants import MCTS_BATCH_SIZE
MAX_NODES = 5000000
class SearchTree(structref.StructRefProxy):
"""
Proxy class representing a search tree for the Monte Carlo Tree Search (MCTS) algorithm.
This class only stores all the data used to perform the MCTS.
Attributes:
- nodes_count (int32): Total number of nodes in the tree.
- root_id (int32): ID of the root node.
- parent (np.ndarray): Array storing parent node IDs for each node.
- first_child (np.ndarray): Array storing the ID of the first child node for each node.
- num_children (np.ndarray): Array storing the number of children for each node.
- moves (np.ndarray): Array storing moves associated with each node.
- player_boards (np.ndarray): Array storing bitboards of player positions for each node.
- opponent_boards (np.ndarray): Array storing bitboards of opponent positions for each node.
- num_visits (np.ndarray): Array storing the number of visits for each node.
- reward (np.ndarray): Array storing the accumulated reward for each node.
Note:
player_boards and opponent_board are always stored from the current player perspective.
This means that boards are inverted at every depth of the tree.
player_board always store the board state of the current player and vice-versa
"""
def __new__(cls):
self = search_tree_ctor()
return self
@property
def nodes_count(self):
return get_nodes_count(self)
@property
def root_id(self):
return get_root_id(self)
@property
def parent(self):
return get_parent(self)
@property
def first_child(self):
return get_first_child(self)
@property
def num_children(self):
return get_num_children(self)
@property
def moves(self):
return get_moves(self)
@property
def player_boards(self):
return get_player_boards(self)
@property
def opponent_boards(self):
return get_opponent_boards(self)
@property
def num_visits(self):
return get_num_visits(self)
@property
def reward(self):
return get_reward(self)
@njit(cache=True)
def get_nodes_count(tree: 'SearchTree') -> int:
return tree.nodes_count
@njit(cache=True)
def get_root_id(tree: 'SearchTree') -> int:
return tree.root_id
@njit(cache=True)
def get_parent(tree: 'SearchTree') -> np.ndarray:
return tree.parent
@njit(cache=True)
def get_first_child(tree: 'SearchTree') -> np.ndarray:
return tree.first_child
@njit(cache=True)
def get_num_children(tree: 'SearchTree') -> np.ndarray:
return tree.num_children
@njit(cache=True)
def get_moves(tree: 'SearchTree') -> np.ndarray:
return tree.moves
@njit(cache=True)
def get_player_boards(tree: 'SearchTree') -> np.ndarray:
return tree.player_boards
@njit(cache=True)
def get_opponent_boards(tree: 'SearchTree') -> np.ndarray:
return tree.opponent_boards
@njit(cache=True)
def get_num_visits(tree: 'SearchTree') -> np.ndarray:
return tree.num_visits
@njit(cache=True)
def get_reward(tree: 'SearchTree') -> np.ndarray:
return tree.reward
search_tree_fields = [
('nodes_count', int32),
('root_id', int32),
('parent', int32[::1]),
('first_child', int32[::1]),
('num_children', int32[::1]),
('moves', int8[::1]),
('player_boards', uint64[::1]),
('opponent_boards', uint64[::1]),
('num_visits', int32[::1]),
('reward', int32[::1])
]
@structref.register
class SearchTreeTypeTemplate(types.StructRef):
def preprocess_fields(self, fields):
return tuple((name, types.unliteral(typ)) for name, typ in fields)
structref.define_boxing(SearchTreeTypeTemplate, SearchTree)
SearchTreeType = SearchTreeTypeTemplate(search_tree_fields)
@njit(void(SearchTreeType), cache=True)
def reset(tree):
"""
Resets the search tree to initial state.
Parameters:
tree (SearchTree): The search tree instance to reset.
"""
tree.nodes_count = 1
tree.root_id = 0
tree.parent[0] = -1
tree.first_child[0] = -1
tree.num_children[0] = -1
tree.num_visits[0] = 0
tree.reward[0] = 0
@njit(int32(SearchTreeType, uint64, uint64), cache=True)
def define_root(tree, player_board, opponent_board):
"""
Defines the root node of the search tree with initial player and opponent bitboards.
Parameters:
tree (SearchTree): The search tree instance.
player_board (uint64): Bitboard representing the current player's disks.
opponent_board (uint64): Bitboard representing the opponent's disks.
Returns:
int32: ID of the root node.
"""
reset(tree)
tree.player_boards[tree.root_id] = player_board
tree.opponent_boards[tree.root_id] = opponent_board
return tree.root_id
@njit(boolean(SearchTreeType, int32), cache = True)
def parent_skiped(tree, node_id):
"""
Checks if the parent node's skipped his turn.
Parameters:
tree (SearchTree): The search tree instance.
node_id (int32): ID of the current node.
Returns:
boolean: True if the parent node's skipped his turn, False otherwise.
"""
parent_id = tree.parent[node_id]
if parent_id == -1:
return False
all_disks = tree.player_boards[node_id] | tree.opponent_boards[node_id]
all_parent_disks = tree.player_boards[parent_id] | tree.opponent_boards[parent_id]
return all_disks == all_parent_disks
@njit(boolean(SearchTreeType, int32), cache = True)
def is_terminal(tree, node_id):
"""
Checks if the node is terminal (i.e., has no children).
Parameters:
tree (SearchTree): The search tree instance.
node_id (int32): ID of the node to check.
Returns:
boolean: True if the node is terminal (has no children), False otherwise.
"""
return tree.num_children[node_id] == 0
@njit(boolean(SearchTreeType, int32), cache = True)
def is_fully_expanded(tree, node_id):
"""
Checks if all children of the node have been visited at least once.
Parameters:
tree (SearchTree): The search tree instance.
node_id (int32): ID of the node to check.
Returns:
boolean: True if all children are fully expanded, False otherwise.
"""
first_child = tree.first_child[node_id]
if first_child == -1:
return False
num_children = tree.num_children[node_id]
for i in range(num_children):
if tree.num_visits[first_child + i] <= 0:
return False
return True
@njit(UniTuple(uint64, 2)(SearchTreeType, int32, int8), cache = True)
def compute_boards(tree, node_id, move):
"""
Computes new player and opponent bitboards after making a move.
Parameters:
tree (SearchTree): The search tree instance.
node_id (int32): ID of the node.
move (int8): Move to make.
Returns:
UniTuple(uint64, 2): Tuple containing new player bitboard and new opponent bitboard.
"""
if move == -1:
return tree.player_boards[node_id], tree.opponent_boards[node_id]
else:
player_bb = tree.player_boards[node_id]
opponent_bb = tree.opponent_boards[node_id]
selected_square = uint64(1) << move
flipped_disks = place_disks(selected_square, player_bb, opponent_bb)
new_player_board = player_bb | (selected_square | flipped_disks)
new_opponent_board = opponent_bb ^ flipped_disks
return new_player_board, new_opponent_board
@njit(int32(SearchTreeType, int32), cache = True)
def expand(tree, node_id):
"""
Expands a node by adding child nodes corresponding to possible moves.
Parameters:
tree (SearchTree): The search tree instance.
node_id (int32): ID of the node to expand.
Returns:
int32: ID of the newly expanded child node.
Note:
Skipped turns are handled in this function. If the current player has no possible moves and his parent
didn't already skipped his turn, then the current player is given 1 child which is a dummy move (ie: -1).
Otherwise, if the parent has already skipped his turn, the current node is set as terminal.
"""
if tree.first_child[node_id] != -1:
# If children are already initialized, return a randomly selected unvisited child
first_child_id = tree.first_child[node_id]
num_children = tree.num_children[node_id]
children = slice(first_child_id, first_child_id + num_children)
unvisited_children = np.where(tree.num_visits[children] == 0)[0]
move_index = np.random.choice(unvisited_children)
new_node_id = first_child_id + move_index
move = tree.moves[new_node_id]
tree.opponent_boards[new_node_id], tree.player_boards[new_node_id] = compute_boards(tree, node_id, move)
return new_node_id
empty_squares = (tree.player_boards[node_id] | tree.opponent_boards[node_id]) ^ 0xFFFFFFFFFFFFFFFF
possible_moves_bb = possible_moves(tree.player_boards[node_id], tree.opponent_boards[node_id], empty_squares)
moves = get_moves_index(possible_moves_bb)
nb_moves = moves.shape[0]
if nb_moves == 0 and not parent_skiped(tree, node_id):
nb_moves = 1
moves = np.array([-1], dtype=np.int8)
elif nb_moves == 0:
tree.num_children[node_id] = 0
return node_id
if tree.nodes_count + nb_moves > MAX_NODES:
print("Max Nodes Reached")
raise Exception(f'The tree reached the maximum number of nodes authorized -> {MAX_NODES}')
first_child_id = tree.nodes_count
tree.nodes_count += nb_moves
tree.first_child[node_id] = first_child_id
tree.num_children[node_id] = nb_moves
for i in range(nb_moves):
idx = first_child_id + i
tree.moves[idx] = moves[i]
tree.parent[idx] = node_id
tree.first_child[idx] = -1
tree.num_children[idx] = -1
tree.num_visits[idx] = 0
tree.reward[idx] = 0
move_index = np.random.choice(nb_moves)
move = moves[move_index]
new_node_id = first_child_id + move_index
tree.opponent_boards[new_node_id], tree.player_boards[new_node_id] = compute_boards(tree, node_id, move)
return new_node_id
@njit(int32(SearchTreeType, int32, float64), cache=True, fastmath=True)
def best_child(tree, node_id, c_param=1.4):
"""
Returns the ID of the best child node based on UCB1 score.
Parameters:
tree (SearchTree): The search tree instance.
node_id (int32): ID of the node to select the best child from.
c_param (float64): Exploration parameter for UCB1 calculation (default is 1.4).
Returns:
int32: ID of the best child node.
"""
first_child = tree.first_child[node_id]
num_children = tree.num_children[node_id]
log_total_visits = np.log(tree.num_visits[node_id])
sqrt_log_total_visits = np.sqrt(2 * log_total_visits)
best_score = -99999
best_child_index = -1
for i in range(num_children):
child_idx = first_child + i
visits = tree.num_visits[child_idx] + 1e-10
score = tree.reward[child_idx] / visits + c_param * sqrt_log_total_visits / np.sqrt(visits)
if score > best_score:
best_score = score
best_child_index = i
return int32(first_child + best_child_index)
@njit(int32(SearchTreeType, int32, float64), cache = True)
def tree_policy(tree, node_id, c_param):
"""
Implements the tree policy for selecting child nodes based on UCB1 scores.
Parameters:
tree (SearchTree): The search tree instance.
node_id (int32): ID of the current node.
c_param (float64): Exploration parameter for UCB1 calculation.
Returns:
int32: ID of the selected child node.
"""
while not is_terminal(tree, node_id):
if not is_fully_expanded(tree, node_id):
return expand(tree, node_id)
else:
node_id = best_child(tree, node_id, c_param=c_param)
return node_id
@njit(int8(SearchTreeType, int32), cache = True)
def default_policy(tree, node_id):
"""
Simulates a random play-out from a given node until the end of the game.
Parameters:
tree (SearchTree): The search tree instance.
node_id (int32): ID of the node to start the simulation from.
Returns:
int8: Outcome of the simulation (1 for win, -1 for loss).
Note:
The win or loss is from the POV of the current node. This is important for the backup part.
"""
current_player_bb = tree.player_boards[node_id]
opponent_bb = tree.opponent_boards[node_id]
initial_player_color = 1
pass_count = 0
while pass_count < 2:
empty_squares = (current_player_bb | opponent_bb) ^ 0xFFFFFFFFFFFFFFFF
possible_moves_bb = possible_moves(current_player_bb, opponent_bb, empty_squares)
valid_moves = get_moves_index(possible_moves_bb)
if valid_moves.shape[0] == 0:
pass_count += 1
current_player_bb, opponent_bb = opponent_bb, current_player_bb
initial_player_color = -initial_player_color
continue
pass_count = 0
move = np.random.choice(valid_moves)
selected_square = uint64(1) << move
flipped_disks = place_disks(selected_square, current_player_bb, opponent_bb)
new_player_board = current_player_bb | (selected_square | flipped_disks)
new_opponent_board = opponent_bb ^ flipped_disks
current_player_bb, opponent_bb = new_opponent_board, new_player_board
initial_player_color = -initial_player_color
if count_bits(current_player_bb) > count_bits(opponent_bb):
return initial_player_color
return -initial_player_color
@njit(void(SearchTreeType, int32, int32), cache = True)
def backup(tree, node_id, reward):
"""
Backpropagates the result of a simulation up the tree.
Parameters:
tree (SearchTree): The search tree instance.
node_id (int32): ID of the node to start the backpropagation from.
reward (int32): Reward from the simulation (-1 for loss, 1 for win).
"""
while node_id != -1:
tree.num_visits[node_id] += 1
if reward == 1:
tree.reward[node_id] += reward
reward = -reward
node_id = tree.parent[node_id]
@njit(void(SearchTreeType, int32, int32, float64), parallel = True, cache = True)
def search_batch(tree, root, nb_rollouts, c_param):
"""
Performs batched Monte Carlo Tree Search (MCTS) on a given search tree starting from a specified root node.
Parameters:
tree (SearchTreeType): The search tree instance to use for MCTS operations.
root (int32): The node ID of the root in the search tree.
nb_rollouts (int32): Number of rollouts (simulations) per batch operation.
c_param (float64): Exploration parameter for the UCB1 formula in MCTS.
"""
for _ in range(MCTS_BATCH_SIZE):
selected_node = tree_policy(tree, root, c_param)
for _ in prange(nb_rollouts):
reward = default_policy(tree, selected_node)
backup(tree, selected_node, -reward)
@njit(int8(SearchTreeType, uint64, uint64, int32, int32, float64), parallel = True, cache = True)
def search(tree, current_player_board, opponent_board, nb_iterations, nb_rollouts, c_param):
"""
Performs Monte Carlo Tree Search (MCTS) on a given game state represented by the current player's board
and the opponent's board.
Parameters:
tree (SearchTreeType): The search tree instance to use for MCTS operations.
current_player_board (uint64): Bitboard representing the current player's positions.
opponent_board (uint64): Bitboard representing the opponent's positions.
nb_iterations (int32): Number of iterations to run the MCTS algorithm.
nb_rollouts (int32): Number of rollouts (simulations) per iteration.
c_param (float64): Exploration parameter for the UCB1 formula in MCTS.
Returns:
int8: The best move determined by MCTS.
Note:
The reward from the Default Policy is inverted because we want it from the POV of the parent.
This is because the result should reflect how good was the move that led to that Node (ie: The node on which the simulation was done)
"""
root = define_root(tree, current_player_board, opponent_board)
for _ in range(nb_iterations):
# Step 1: Selection
selected_node = tree_policy(tree, root, c_param)
if nb_rollouts > 1:
for _ in prange(nb_rollouts):
reward = default_policy(tree, selected_node)
backup(tree, selected_node, -reward)
else:
# Step 2: Simulation
reward = default_policy(tree, selected_node)
# Step 3: Back propagation
backup(tree, selected_node, -reward)
return tree.moves[best_child(tree, root, c_param=0)]
# Allows to use SearchTree as a constructor in a jitted function
@njit(SearchTreeType(), cache=True)
def search_tree_ctor():
st = structref.new(SearchTreeType)
# Initialize the struct fields
st.nodes_count = 0
st.root_id = 0
st.parent = -np.ones(MAX_NODES, dtype=np.int32)
st.first_child = -np.ones(MAX_NODES, dtype=np.int32)
st.num_children = -np.ones(MAX_NODES, dtype=np.int32)
st.moves = -np.ones(MAX_NODES, dtype=np.int8)
st.player_boards = np.zeros(MAX_NODES, dtype=np.uint64)
st.opponent_boards = np.zeros(MAX_NODES, dtype=np.uint64)
st.num_visits = -np.ones(MAX_NODES, dtype=np.int32)
st.reward = -np.ones(MAX_NODES, dtype=np.int32)
return st
@overload(SearchTree)
def overload_SearchTree():
def impl():
return search_tree_ctor()
return impl