-
Notifications
You must be signed in to change notification settings - Fork 1
/
mcts_c4.py
329 lines (286 loc) · 11.8 KB
/
mcts_c4.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
#!/usr/bin/env python
# -*- coding: utf-8 -*-
__author__ = 'arenduchintala'
import numpy as np
import pdb
import sys
class COLOR:
PURPLE = '\033[95m'
CYAN = '\033[96m'
DARKCYAN = '\033[36m'
BLUE = '\033[94m'
GREEN = '\033[92m'
YELLOW = '\033[93m'
RED = '\033[91m'
BOLD = '\033[1m'
UNDERLINE = '\033[4m'
END = '\033[0m'
EPS = 1e-4
def color_board(i, x, y, highlist_pos, color):
if (x, y) in highlist_pos:
if color is None:
s = COLOR.UNDERLINE + str(int(i)) + COLOR.END
else:
s = color + str(int(i)) + COLOR.END
else:
s = str(int(i))
if i == 0:
s = '.'
elif i == 1:
s = COLOR.RED + s + COLOR.END
else:
s = COLOR.YELLOW + s + COLOR.END
return s
def display_board(state, pos, color):
assert type(pos) is set
s = ''
for r in range(state.board.shape[0]):
s += ' '.join([color_board(i, r, idx, pos, color) for idx, i in enumerate(state.board[r, :])]) + '\n'
return s
class SearchTree(object):
def __init__(self, game):
self.game = game
self.iters = 1000
self.nodes_seen = {}
self.gamma = 1.0
def search(self, root_node):
self.nodes_seen[str(root_node.state)] = root_node
for _ in range(self.iters):
if _ % 100 == 0:
sys.stdout.write('.')
sys.stdout.flush()
# print('************************ search iter' + str(_) + '*************************')
search_sequence = []
node, search_sequence = self.selection(root_node)
if not node.state.is_terminal:
node = self.expansion(node)
search_sequence.append(node)
reward, winner, winner_pos = self.rollout(node)
self.backup(node, reward, winner, search_sequence)
# print('********************** end search iter' + str(_) + '*******************')
print('')
# self.display_child_values(root_node, 0.0)
# print('diplay completed!!!!!!!!!!')
best_action, best_node = self.select_child(root_node, 0.0)
return best_action
def selection(self, node):
search_sequence = [node]
while node.completed_expansion:
assert not node.state.is_terminal
action, node = self.select_child(node, 0.5)
search_sequence.append(node)
return node, search_sequence
def expansion(self, node):
assert not node.completed_expansion
assert not node.state.is_terminal
child_node = self.expand_child(node)
return child_node
def rollout(self, node):
state = node.state
steps = 0
while not state.is_terminal:
pa = self.game.possible_actions(state)
selected_action = pa[np.random.choice(len(pa))]
state = self.game.next_state(state, selected_action)
steps += 1
if steps > game.max_steps:
print('exceeded max steps')
pdb.set_trace()
return self.gamma ** steps, state.winner, state.winner_pos
def backup(self, node, reward, winner, search_sequence):
# print('------------------backup-----------------')
for node in search_sequence:
node.rewards[winner] += reward
node.visits += 1
# print('-----------------------------------------')
return True
def display_child_values(self, node, exp_param):
combined, values, ucb, actions, rewards, reward_sum, visits = self.child_scores(node, exp_param)
f = [(a[1], a[0], c, v, u, w, sw, vs) for a, c, v, u, w, sw, vs in zip(actions, combined, values, ucb, rewards, reward_sum, visits)]
s = '\n'.join(['combined %0.2f' % c +
' value: %0.2f' % v +
' ucb: %0.2f' % u +
' position:' + str(a0) + ',' + str(a1) +
' rewards:' + w + ' reward_sum:' + sw +
' visits:' + str(vs) for
a1, a0, c, v, u, w, sw, vs in sorted(f)])
print(s)
return s
def child_scores(self, node, exp_param):
n_visits = float(node.visits)
cn_player = 3 - node.state.player # make selection from current node to child node
values = np.zeros(len(node.expanded_actions))
ucb = np.zeros(len(node.expanded_actions))
actions = [None] * len(node.expanded_actions)
rewards = []
reward_sum = []
visits = []
for idx, (a, cn) in enumerate(node.expanded_actions.items()):
cn_visits = sum(cn.rewards)
cn_value = (float(cn.rewards[cn_player]) / float(cn_visits))
cn_ucb = np.sqrt(np.log2(n_visits) / cn_visits)
assert not np.isnan(cn_value)
assert not np.isnan(cn_ucb)
values[idx] = cn_value
ucb[idx] = cn_ucb
actions[idx] = a
rewards.append(' '.join(['%.2f' % w for w in cn.rewards]))
reward_sum.append('%.2f' % sum(cn.rewards))
visits.append(cn.visits)
combined = (1.0 - exp_param) * values + exp_param * ucb
return combined, values, ucb, actions, rewards, reward_sum, visits
def select_child(self, node, exp_param):
combined, values, ucb, actions, rewards, reward_sum, visits = self.child_scores(node, exp_param)
max_idx = np.random.choice(np.flatnonzero(combined == combined.max()))
best_action = actions[max_idx]
return best_action, node.expanded_actions[best_action]
def expand_child(self, node):
assert len(node.unexpanded_actions) > 0
action, _ = node.unexpanded_actions.popitem()
new_state = self.game.next_state(node.state, action)
new_rewards = [EPS, EPS, EPS]
new_state_unexpanded_actions = {a: None for a in self.game.possible_actions(new_state)}
new_state_expanded_actions = {}
if str(new_state) in self.nodes_seen:
expanded_node = self.nodes_seen[str(new_state)]
else:
expanded_node = SearchNode(new_state, new_state_unexpanded_actions, new_state_expanded_actions, new_rewards)
self.nodes_seen[str(new_state)] = expanded_node
node.update_expansion(action, expanded_node)
return expanded_node
class SearchNode(object):
def __init__(self, state, unexpanded_actions, expanded_actions, rewards):
self.state = state
self.unexpanded_actions = unexpanded_actions
self.expanded_actions = expanded_actions
self.rewards = rewards
self.visits = 0
self.completed_expansion = False
def update_expansion(self, action, child_node):
assert action not in self.expanded_actions
self.expanded_actions[action] = child_node
self.completed_expansion = len(self.unexpanded_actions) == 0
return True
def __str__(self,):
s = str(self.state) + '\n'
s += 'visits:' + str(self.visits) + '\n'
s += 'rewards:' + ' '.join(['%.2f' % i for i in self.rewards]) + '\n'
s += 'u:' + str(len(self.unexpanded_actions)) + ' e:' + str(len(self.expanded_actions)) + '\n'
s += 'exp complete:' + str(self.completed_expansion) + '\n'
return s
class Game(object):
left = [(0, 0), (0, -1), (0, -2), (0, -3)]
right = [(0, 0), (0, 1), (0, 2), (0, 3)]
down = [(0, 0), (1, 0), (2, 0), (3, 0)]
leftdown = [(0, 0), (1, -1), (2, -2), (3, -3)]
rightdown = [(0, 0), (1, 1), (2, 2), (3, 3)]
def __init__(self, size):
self.size = size
self.max_steps = size[0] * size[1]
def start(self,):
b = np.zeros(self.size, dtype=int)
player = 2
return State(player, None, b)
def possible_actions(self, state):
a = []
for c in range(state.board.shape[1]):
for r in reversed(range(state.board.shape[0])):
if state.board[r, c] == 0:
a.append((r, c))
break
return a
def next_state(self, state, action):
new_board = state.board.copy()
new_player = 3 - state.player
new_board[action[0], action[1]] = new_player
new_state = State(new_player, action, new_board)
winner, winner_pos = self.check(new_state)
pa = self.possible_actions(new_state)
is_terminal = winner != 0 or len(pa) == 0
new_state.set_is_terminal(is_terminal, winner, winner_pos)
return new_state
def __left_shift(self, state):
r, c = state.position
i = 0
while 0 <= c + i and state.board[r, c + i] == state.board[r, c]:
i -= 1
i += 1
return r, c + i
def __lefttop_shift(self, state):
r, c = state.position
i = 0
while 0 <= c + i and 0 <= r + i and state.board[r + i, c + i] == state.board[r, c]:
i -= 1
i += 1
return r + i, c + i
def __righttop_shift(self, state):
r, c = state.position
i = 0
j = 0
while c + i <= 6 and 0 <= r + j and state.board[r + j, c + i] == state.board[r, c]:
i += 1
j -= 1
j += 1
i -= 1
return r + j, c + i
def check(self, state):
for name, shift, vec in zip(['left', 'down', 'lefttop', 'righttop'],
[self.__left_shift, None, self.__lefttop_shift, self.__righttop_shift],
[Game.right, Game.down, Game.rightdown, Game.leftdown]):
if shift is not None:
r, c = shift(state)
else:
r, c = state.position
cl = [(i[0] + r, i[1] + c) for i in vec if (0 <= i[0] + r < self.size[0] and 0 <= i[1] + c < self.size[1])]
if len(cl) == 4:
_cl = list(zip(*cl))
cc = state.board[_cl].tolist()
if cc == [state.player] * 4:
return state.player, set(cl)
return 0, set([])
class State(object):
def __init__(self, player, position, board):
self.board = board
self.position = position
self.player = player # player whos move lead to this state, i.e. the last player
self.is_terminal = False
self.winner = None
self.winner_pos = None
def set_is_terminal(self, is_terminal, winner, winner_pos):
self.is_terminal = is_terminal
self.winner = winner
self.winner_pos = winner_pos
def __str__(self,):
s = 'player:' + str(self.player) + '\n'
s += 'is_terminal:' + str(self.is_terminal) + '\n'
s += 'board:' + '\n'
for r in range(self.board.shape[0]):
s += ''.join([str(int(i)) for idx, i in enumerate(self.board[r, :])]) + '\n'
return s.strip()
def mcts_search(state, game):
search_tree = SearchTree(game)
unexpanded_actions = {a: None for a in game.possible_actions(state)}
expanded_actions = {}
root_node = SearchNode(state, unexpanded_actions, expanded_actions, [EPS, EPS, EPS])
best_action = search_tree.search(root_node)
return best_action
if __name__ == '__main__':
game = Game((6, 7))
state = game.start()
winner = 0
while len(game.possible_actions(state)) > 0 and winner == 0:
print(display_board(state, set([state.position]), None))
print('player' + str(3 - state.player) + ':')
if 3 - state.player == 1:
action_map = {i[1]: i for i in game.possible_actions(state)}
i = -1
while i not in action_map:
i = input('select action: ' + ','.join([str(k) for k, v in action_map.items()]) + ':')
i = int(i)
selected_action = action_map[int(i)]
else:
selected_action = mcts_search(state, game)
state = game.next_state(state, selected_action)
winner, winner_pos = game.check(state)
print(display_board(state, set(winner_pos), COLOR.GREEN))
print('game over!')