-
Notifications
You must be signed in to change notification settings - Fork 1
/
mcts.py
232 lines (183 loc) · 7.24 KB
/
mcts.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
import pyximport
pyximport.install()
# from nn_evaluate import evaluate
from evals import arbiter_draw, fifty_move_draw, three_fold_repetition
from time import time
from position import Position
from side import Side
import numpy as np
import math
def has_moves(pos):
moves = pos.generate_moves_all(legal=True)
return len(list(moves)) > 0
def game_over(pos):
return not has_moves(pos) or arbiter_draw(pos)
class Node(object):
def __init__(self):
self.N = 0
class Edge(object):
def __init__(self):
self.N = 0
self.Q = 0
class MCTS(object):
def __init__(self, s0, max_simulations=800, c=math.sqrt(2), w_r=0.5, w_v=0.75, w_a=None):
self.max_simulations = max_simulations
self.simulations = 0
# Root state to search from
self.s0 = s0
# Exploration parameter. Is 0.05 in ExIT.
# Not used in AlphaGoZero, presumably same in AlphaZero.
# Theoretical value is sqrt(2).
# ExIT use c = 0.05 but they use RAVE which affects exploration.
self.c = c
# c_RAVE is 3000 in ExIT.
# AlphaZero paper thinks RAVE is unnecessary when using Policy Networks.
# ExIT paper thinks RAVE helps efficiency early on.
# Mixing parameter for rollout vs NN value estimate
self.w_r = w_r
# UCT bonus towards action prior. AlphaZero essentially uses this as
# their main UCT term with a slightly different formula, so this is
# essentially the exploration parameter there.
# AlphaGoZero seemed to use w_a = 5 * additional term sqrt(node.N).
# ExIT paper recommends w_a = avg simulations per action in the root
# e.g. 100 in game of Go when doing 10k simulations.
self.w_a = w_a if w_a is not None else self.max_simulations / 35
# Value bonus for UCT. Not used in Alpha*. ExIT used 0.75.
self.w_v = w_v
self.tree = dict()
# Heuristic function to evaluate a non-terminal node
# Approximates V_MCTS*(s)
self.Qnn = None
# Heuristic function approximating argmax_a MCTS(a|s)
self.Pnn = None
def set_Qnn_function(self, fn):
self.Qnn = fn
def set_Pnn_function(self, fn):
self.Pnn = fn
def search(self):
pos0 = Position.from_fen(self.s0)
while self.time_available():
self.simulate(Position(pos0))
return self.select_move(pos0, 0)
def time_available(self):
return self.simulations < self.max_simulations
def simulate(self, position):
self.simulations += 1
# print(self.simulations)
states_actions = self.sim_tree(position)
z = self.sim_default(Position(position)) if self.w_r > 0 else 0
self.backup(states_actions, z)
def sim_tree(self, position):
states_actions = []
while not game_over(position):
s = self.get_state(position)
if s not in self.tree:
self.new_node(s)
states_actions.append((s, None))
break
a = self.select_move(position, self.c)
states_actions.append((s, a.to_uci))
position.make_move(a)
return states_actions
def default_policy(self, position):
legal = self.get_actions(position=position)
return np.random.choice(legal)
def sim_default(self, position):
while not game_over(position):
a = self.default_policy(position)
position.make_move(a)
stm = position.side_to_move()
if stm == Side.B and position.in_check_simple(): return 1.0
elif stm == Side.W and position.in_check_simple(): return -1.0
else: return 0.0
def new_node(self, s):
node = Node()
node.N = 0
self.tree[s] = node
for a in self.get_actions(s=s):
edge = Edge()
edge.Pnn = 0 if self.w_a else 0
edge.Qnn = self.Qnn(s, a) if self.w_v else 0
edge.N = 0
edge.Q = 0
self.tree[(s, a.to_uci)] = edge
def uct_value(self, s, a, c, w_a, w_v):
node = self.tree[s]
edge = self.tree[(s, a.to_uci)]
uct = c * math.sqrt(math.log(node.N) / edge.N) if edge.N else math.inf
# May want to tune temperature of the softmax generating P towards
# optimizing this bonus using Pnn. ExIT use temp 1.0. AlphaGoZero used
# 0.67 for temp.
policy_prior_bonus = w_a * edge.Pnn / (edge.N + 1) if w_a else 0
value_prior_bonus = w_v * edge.Qnn if w_v else 0
return uct + policy_prior_bonus + value_prior_bonus
def Q(self, s, a):
edge = self.tree[(s, a.to_uci)]
return edge.Q
def pv(self, s0):
pos = Position.from_fen(s0)
actions = []
s = self.get_state(pos)
while s in self.tree:
a = self.select_move(pos, 0)
pos.make_move(a)
actions.append(a)
s = self.get_state(pos)
return actions
def select_move(self, position, c):
# When c == 0 should we select just based on largest N(s,a)?
# Perhaps extend the search if argmax Q(s,a) and argmax N(s,a) disagree
s = self.get_state(position)
legal = self.get_actions(position=position)
if position.white_to_move():
uct_vals = [self.Q(s, a) + self.uct_value(s, a, c, self.w_a, self.w_v) for a in legal]
best_ind = np.argmax(uct_vals)
else:
uct_vals = [self.Q(s, a) - self.uct_value(s, a, c, self.w_a, self.w_v) for a in legal]
best_ind = np.argmin(uct_vals)
return legal[best_ind]
def backup(self, states_actions, z):
for s, a in states_actions:
node = self.tree[s]
node.N += 1
if a is not None:
edge = self.tree[(s, a)]
edge.N += 1
v = (1 - self.w_r) * edge.Qnn + self.w_r * z
edge.Q += (v - edge.Q) / edge.N
def get_actions(self, s=None, position=None):
pos = position if position else Position.from_fen(s)
return list(pos.generate_moves_all(legal=True))
def get_state(self, position):
return position.fen(timeless=True) + ' 0 0'
def eval_approx(s, a):
pos = Position.from_fen(s)
pos.make_move(a)
return -evaluate(pos) / 1000
def mcts_selfplay(s0):
mcts = MCTS(s0, max_simulations=800, c=math.sqrt(2), w_r=1, w_v=0.0, w_a=0)
# mcts.set_Qnn_function(eval_approx)
move = mcts.search()
print(move)
print(mcts.pv(s0))
return move
if __name__ == '__main__':
import cProfile
import pstats
pos = Position.from_fen("r4rk1/1pp1qppp/p1np1n2/2b1p1B1/2B1P1b1/P1NP1N2/1PP1QPPP/R4RK1 w - - 0 10")
def goprofile():
s0 = Position().fen()
now = time()
mcts_selfplay(s0)
end = time() - now
print(end, 's')
goprofile()
# print(pos)
# def goprofile():
# for i in range(10000000):
# pos.generate_moves_all(legal=True)
# goprofile()
# cProfile.run("goprofile()", filename="outprofile")
# pstats.Stats("outprofile").strip_dirs().sort_stats("time").print_stats(15)
# print()
# pstats.Stats("outprofile").strip_dirs().sort_stats("cumulative").print_stats(15)