Repository navigation
Expand file tree
/
Copy pathtest_mcts.py
More file actions
194 lines (173 loc) · 7.18 KB
/
Copy pathtest_mcts.py
File metadata and controls
194 lines (173 loc) · 7.18 KB
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
# Copyright 2018 Google LLC
#
# 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 copy
import unittest
import numpy as np
import coords
import go
from go import Position
import mcts
from tests import test_utils
ALMOST_DONE_BOARD = test_utils.load_board('''
.XO.XO.OO
X.XXOOOO.
XXXXXOOOO
XXXXXOOOO
.XXXXOOO.
XXXXXOOOO
.XXXXOOO.
XXXXXOOOO
XXXXOOOOO
''')
TEST_POSITION = go.Position(
board=ALMOST_DONE_BOARD,
n=105,
komi=2.5,
caps=(1, 4),
ko=None,
recent=(go.PlayerMove(go.BLACK, (0, 1)),
go.PlayerMove(go.WHITE, (0, 8))),
to_play=go.BLACK
)
SEND_TWO_RETURN_ONE = go.Position(
board=ALMOST_DONE_BOARD,
n=75,
komi=0.5,
caps=(0, 0),
ko=None,
recent=(go.PlayerMove(go.BLACK, (0, 1)),
go.PlayerMove(go.WHITE, (0, 8)),
go.PlayerMove(go.BLACK, (1, 0))),
to_play=go.WHITE
)
class TestMctsNodes(test_utils.MiniGoUnitTest):
def test_action_flipping(self):
np.random.seed(1)
probs = np.array([.02] * (go.N * go.N + 1))
probs = probs + np.random.random([go.N * go.N + 1]) * 0.001
black_root = mcts.MCTSNode(go.Position())
white_root = mcts.MCTSNode(go.Position(to_play=go.WHITE))
black_root.select_leaf().incorporate_results(probs, 0, black_root)
white_root.select_leaf().incorporate_results(probs, 0, white_root)
# No matter who is to play, when we know nothing else, the priors
# should be respected, and the same move should be picked
black_leaf = black_root.select_leaf()
white_leaf = white_root.select_leaf()
self.assertEqual(black_leaf.fmove, white_leaf.fmove)
self.assertEqualNPArray(
black_root.child_action_score, white_root.child_action_score)
def test_select_leaf(self):
flattened = coords.to_flat(coords.from_kgs('D9'))
probs = np.array([.02] * (go.N * go.N + 1))
probs[flattened] = 0.4
root = mcts.MCTSNode(SEND_TWO_RETURN_ONE)
root.select_leaf().incorporate_results(probs, 0, root)
self.assertEqual(root.position.to_play, go.WHITE)
self.assertEqual(root.select_leaf(), root.children[flattened])
def test_backup_incorporate_results(self):
probs = np.array([.02] * (go.N * go.N + 1))
root = mcts.MCTSNode(SEND_TWO_RETURN_ONE)
root.select_leaf().incorporate_results(probs, 0, root)
leaf = root.select_leaf()
leaf.incorporate_results(probs, -1, root) # white wins!
# Root was visited twice: first at the root, then at this child.
self.assertEqual(root.N, 2)
# Root has 0 as a prior and two visits with value 0, -1
self.assertAlmostEqual(root.Q, -1/3) # average of 0, 0, -1
# Leaf should have one visit
self.assertEqual(root.child_N[leaf.fmove], 1)
self.assertEqual(leaf.N, 1)
# And that leaf's value had its parent's Q (0) as a prior, so the Q
# should now be the average of 0, -1
self.assertAlmostEqual(root.child_Q[leaf.fmove], -0.5)
self.assertAlmostEqual(leaf.Q, -0.5)
# We're assuming that select_leaf() returns a leaf like:
# root
# \
# leaf
# \
# leaf2
# which happens in this test because root is W to play and leaf was a W win.
self.assertEqual(root.position.to_play, go.WHITE)
leaf2 = root.select_leaf()
leaf2.incorporate_results(probs, -0.2, root) # another white semi-win
self.assertEqual(root.N, 3)
# average of 0, 0, -1, -0.2
self.assertAlmostEqual(root.Q, -0.3)
self.assertEqual(leaf.N, 2)
self.assertEqual(leaf2.N, 1)
# average of 0, -1, -0.2
self.assertAlmostEqual(leaf.Q, root.child_Q[leaf.fmove])
self.assertAlmostEqual(leaf.Q, -0.4)
# average of -1, -0.2
self.assertAlmostEqual(leaf.child_Q[leaf2.fmove], -0.6)
self.assertAlmostEqual(leaf2.Q, -0.6)
def test_do_not_explore_past_finish(self):
probs = np.array([0.02] * (go.N * go.N + 1), dtype=np.float32)
root = mcts.MCTSNode(go.Position())
root.select_leaf().incorporate_results(probs, 0, root)
first_pass = root.maybe_add_child(coords.to_flat(None))
first_pass.incorporate_results(probs, 0, root)
second_pass = first_pass.maybe_add_child(coords.to_flat(None))
with self.assertRaises(AssertionError):
second_pass.incorporate_results(probs, 0, root)
node_to_explore = second_pass.select_leaf()
# should just stop exploring at the end position.
self.assertEqual(node_to_explore, second_pass)
def test_add_child(self):
root = mcts.MCTSNode(go.Position())
child = root.maybe_add_child(17)
self.assertIn(17, root.children)
self.assertEqual(child.parent, root)
self.assertEqual(child.fmove, 17)
def test_add_child_idempotency(self):
root = mcts.MCTSNode(go.Position())
child = root.maybe_add_child(17)
current_children = copy.copy(root.children)
child2 = root.maybe_add_child(17)
self.assertEqual(child, child2)
self.assertEqual(current_children, root.children)
def test_never_select_illegal_moves(self):
probs = np.array([0.02] * (go.N * go.N + 1))
# let's say the NN were to accidentally put a high weight on an illegal move
probs[1] = 0.99
root = mcts.MCTSNode(SEND_TWO_RETURN_ONE)
root.incorporate_results(probs, 0, root)
# and let's say the root were visited a lot of times, which pumps up the
# action score for unvisited moves...
root.N = 100000
root.child_N[root.position.all_legal_moves()] = 10000
# this should not throw an error...
leaf = root.select_leaf()
# the returned leaf should not be the illegal move
self.assertNotEqual(leaf.fmove, 1)
# and even after injecting noise, we should still not select an illegal move
for i in range(10):
root.inject_noise()
leaf = root.select_leaf()
self.assertNotEqual(leaf.fmove, 1)
def test_dont_pick_unexpanded_child(self):
probs = np.array([0.001] * (go.N * go.N + 1))
# make one move really likely so that tree search goes down that path twice
# even with a virtual loss
probs[17] = 0.999
root = mcts.MCTSNode(go.Position())
root.incorporate_results(probs, 0, root)
leaf1 = root.select_leaf()
self.assertEqual(leaf1.fmove, 17)
leaf1.add_virtual_loss(up_to=root)
# the second select_leaf pick should return the same thing, since the child
# hasn't yet been sent to neural net for eval + result incorporation
leaf2 = root.select_leaf()
self.assertIs(leaf1, leaf2)