def decode_move_index(self, index):
"""Turn an integer index into a board point."""
if index == self._pass_idx:
return Move.pass_turn()
row = index // self._board_size
col = index % self._board_size
return Move.play(Point(row=row + 1, col=col + 1))
def remove_dead_stones_and_score(game_state):
b_resign = False
w_resign = False
if game_state.last_move.is_resign:
if game_state.next_player == Player.black:
w_resign = True
else:
b_resign = True
if b_resign or w_resign:
return GameResult(
b_resign, w_resign,
0.0, 0.0,
game_state.komi(), game_state.board)
end_game = game_state
while end_game.last_move == Move.pass_turn():
end_game = end_game.previous_state
final_board = remove_dead_stones(end_game)
final_state = GameState.from_board(
final_board,
game_state.next_player,
game_state.komi())
final_state = final_state.apply_move(Move.pass_turn())
final_state = final_state.apply_move(Move.pass_turn())
return compute_game_result(final_state)
def decode_sgf_move(sgf_move, num_rows):
if sgf_move == '':
return Move.pass_turn()
assert len(sgf_move) == 2
col = ALPHABET.index(sgf_move[0]) + 1
row = num_rows - ALPHABET.index(sgf_move[1])
return Move.play(Point(row, col))
def test_move_is_hashable(self):
moves = {
Move.play(Point(1, 1)): 1,
Move.resign(): 2,
}
self.assertEqual(1, moves[Move.play(Point(1, 1))])
self.assertEqual(2, moves[Move.resign()])
def test_create_game_from_board(self):
board = Board(5, 5)
board.place_stone(Player.black, Point(2, 2))
board.place_stone(Player.black, Point(4, 4))
game = GameState.from_board(board, Player.white)
self.assertEqual(Player.white, game.next_player)
self.assertFalse(game.is_valid_move(Move.play(Point(2, 2))))
self.assertTrue(game.is_valid_move(Move.play(Point(3, 3))))
def read_move(self):
is_play = self.read_bool()
is_pass = self.read_bool()
is_resign = self.read_bool()
if is_play:
row = self.read_int()
col = self.read_int()
return Move.play(Point(row=row, col=col))
if is_pass:
return Move.pass_turn()
assert is_resign
return Move.resign()
def test_new_game(self):
start = GameState.new_game(19)
next_state = start.apply_move(Move.play(Point(16, 16)))
self.assertEqual(start, next_state.previous_state)
self.assertEqual(Player.white, next_state.next_player)
self.assertEqual(Player.black, next_state.board.get(Point(16, 16)))
def test_ko_as_array(self):
start = GameState.new_game(19)
# .wb.
# wb.b
# .wb.
# ....
game = start.apply_move(Move.play(Point(1, 3)))
game = game.apply_move(Move.play(Point(1, 2)))
game = game.apply_move(Move.play(Point(2, 2)))
game = game.apply_move(Move.play(Point(2, 1)))
game = game.apply_move(Move.play(Point(3, 3)))
game = game.apply_move(Move.play(Point(3, 2)))
game = game.apply_move(Move.play(Point(2, 4)))
# W takes the ko
game = game.apply_move(Move.play(Point(2, 3)))
ko_array = game.ko_points_as_array()
self.assertEqual((19, 19), ko_array.shape)
self.assertEqual(1, ko_array[1, 1])
self.assertEqual(0, ko_array[2, 1])
self.assertEqual(0, ko_array[5, 5])
def test_legal_move_mask(self):
start = GameState.new_game(19)
# .wb.
# wb.b
# .wb.
# ....
game = start.apply_move(Move.play(Point(1, 3)))
game = game.apply_move(Move.play(Point(1, 2)))
game = game.apply_move(Move.play(Point(2, 2)))
game = game.apply_move(Move.play(Point(2, 1)))
game = game.apply_move(Move.play(Point(3, 3)))
game = game.apply_move(Move.play(Point(3, 2)))
game = game.apply_move(Move.play(Point(2, 4)))
# W takes the ko
game = game.apply_move(Move.play(Point(2, 3)))
legal_moves = game.legal_moves_as_array()
self.assertEqual((19 * 19 + 1, ), legal_moves.shape)
illegal_indices = [
# Suicide
19 * 0 + 0,
# Stones here
19 * 0 + 2,
19 * 0 + 1,
19 * 1 + 0,
19 * 2 + 2,
19 * 2 + 1,
19 * 1 + 3,
19 * 1 + 2,
# ko
19 * 1 + 1,
]
for i, val in enumerate(legal_moves):
if i in illegal_indices:
self.assertEqual(0, val, "{} should be illegal".format(i))
else:
self.assertEqual(1, val, "{} should be legal".format(i))
def encode(self, game_state):
board_tensor = np.zeros(self.shape())
if game_state.next_player == Player.black:
board_tensor[8] = 1
else:
board_tensor[9] = 1
for r in range(self._board_size):
for c in range(self._board_size):
p = Point(row=r + 1, col=c + 1)
go_string = game_state.board.get_string(p)
if go_string is None:
if game_state.does_move_violate_ko(Move.play(p)):
board_tensor[10][r][c] = 1
else:
liberty_plane = min(4, go_string.num_liberties) - 1
if go_string.color == Player.white:
liberty_plane += 4
board_tensor[liberty_plane][r][c] = 1
return board_tensor
def test_komi(self):
start = GameState.new_game(19, 0.5)
next_state = start.apply_move(Move.play(Point(16, 16)))
self.assertAlmostEqual(0.5, next_state.komi())
def test_ko(self):
start = GameState.new_game(19)
# .wb.
# wb.b
# .wb.
# ....
game = start.apply_move(Move.play(Point(1, 3)))
game = game.apply_move(Move.play(Point(1, 2)))
game = game.apply_move(Move.play(Point(2, 2)))
game = game.apply_move(Move.play(Point(2, 1)))
game = game.apply_move(Move.play(Point(3, 3)))
game = game.apply_move(Move.play(Point(3, 2)))
game = game.apply_move(Move.play(Point(2, 4)))
# W takes the ko
game = game.apply_move(Move.play(Point(2, 3)))
# B can't take back
self.assertTrue(game.does_move_violate_ko(Move.play(Point(2, 2))))
self.assertFalse(game.is_valid_move(Move.play(Point(2, 2))))
# "ko threat"
game = game.apply_move(Move.play(Point(19, 19)))
game = game.apply_move(Move.play(Point(18, 18)))
# B can take now
self.assertFalse(game.does_move_violate_ko(Move.play(Point(2, 2))))
self.assertTrue(game.is_valid_move(Move.play(Point(2, 2))))
def test_last_move(self):
start = GameState.new_game(19)
next_move = Move.play(Point(16, 16))
state = start.apply_move(next_move)
self.assertEqual(Move.play(Point(16, 16)), state.last_move)
def test_move_number(self):
start = GameState.new_game(19)
self.assertEqual(0, start.num_moves)
game = start.apply_move(Move.play(Point(1, 3)))
game = game.apply_move(Move.play(Point(1, 2)))
self.assertEqual(2, game.num_moves)
def test_is_valid_move(self):
start = GameState.new_game(19)
state = start.apply_move(Move.play(Point(16, 16)))
self.assertTrue(state.is_valid_move(Move.play(Point(16, 17))))
self.assertFalse(state.is_valid_move(Move.play(Point(16, 16))))
def select_move(self, game_state):
start = time.time()
self.root = None
if self._ladder_rollouts > 0:
self.root = self.read_ladders(game_state, self._ladder_rollouts)
if self.root is None:
self.root = self.create_node(game_state, add_noise=True)
num_rollouts = 0
while num_rollouts < self._num_rollouts:
to_expand = set()
batch_count = 0
while batch_count < self._batch_size:
# Find a leaf.
node = self.root
move = self.select_branch(node)
while node.has_child(move):
node.add_virtual_loss(move)
node = node.get_child(move)
move = self.select_branch(node)
node.add_virtual_loss(move)
batch_count += 1
to_expand.add((node, move))
batch_num_visits = len(to_expand)
new_children = self.create_children(to_expand)
for new_child in new_children:
new_child.parent.record_visit(new_child.move, new_child.value)
num_rollouts += batch_num_visits
# Now select a move in proportion to how often we visited it.
visit_counts = self.root.visit_counts
expected_values = calc_expected_values(self.root.total_values,
visit_counts)
tiebreak = 0.499 * (expected_values + 1)
decide_vals = visit_counts + tiebreak
for move_idx in np.argsort(decide_vals):
visit_count = visit_counts[move_idx]
if visit_count > 0:
sys.stderr.write('{}: {:.3f} {}\n'.format(
format_move(self._encoder.decode_move_index(move_idx)),
expected_values[move_idx], visit_count))
temperature = self._temp_schedule.get(game_state.num_moves)
if temperature > 0:
move_indices, = np.where(visit_counts > 0)
raw_counts = decide_vals[move_indices]
p = np.power(raw_counts, 1.0 / temperature)
p /= np.sum(p)
move_index = np.random.choice(move_indices, p=p)
else:
move_index = np.argmax(decide_vals)
self._log_pv(self.root)
chosen_move = self._encoder.decode_move_index(move_index)
sys.stderr.write('Select {} Q {:.3f}\n'.format(
format_move(chosen_move), expected_values[move_index]))
end = time.time()
sys.stderr.write('Decided in {:.3f}s\n'.format(end - start))
sys.stderr.flush()
if expected_values[move_index] < self._resign_below:
sys.stderr.write('Resigning because Q {:.3f} < {:.3f}\n'.format(
expected_values[move_index], self._resign_below))
return Move.resign()
if self._gracious_winner is not None:
if game_state.last_move is not None and game_state.last_move == Move.pass_turn(
):
pass_idx = self._encoder.encode_move(Move.pass_turn())
if visit_counts[pass_idx] >= 2 and \
expected_values[pass_idx] > self._gracious_winner:
sys.stderr.write('Pass has Q {:.3f}\n'.format(
expected_values[pass_idx]))
return Move.pass_turn()
return chosen_move
