def test_simulation_can_recover_from_sucide_move_white(self):
model = self.model
board, player = game_init()
x = randrange(SIZE)
y = randrange(SIZE)
for i in range(SIZE):
for j in range(SIZE):
make_play(0, SIZE, board) # Black does not play playing
if i == x and j == y:
make_play(0, SIZE, board) # pass on one intersection
else:
make_play(i, j, board)
make_play(0, SIZE, board) # Black does not play playing
policies, values = model.predict_on_batch(board)
policy = policies[0]
policy[y * SIZE + x], policy[SIZE * SIZE] = policy[
SIZE * SIZE], policy[y * SIZE + x] # Make best move sucide
mask = legal_moves(board)
policy = ma.masked_array(policy, mask=mask)
self.assertEqual(np.argmax(policy),
y * SIZE + x) # Best option in policy is sucide
tree = new_tree(policy, board)
chosen_play = select_play(policy,
board,
mcts_simulations=128,
mcts_tree=tree,
temperature=0,
model=model)
# First simulation chooses pass, second simulation chooses sucide (p is still higher),
# then going deeper it chooses pass again (value is higher)
self.assertEqual(chosen_play, SIZE * SIZE) # Pass move is best option
def genmove(self, color):
policies, values = self.model.predict_on_batch(self.board)
policy = policies[0]
value = values[0]
if self.resign and value <= self.resign:
x = 0
y = SIZE + 1
return x, y, policy, value, self.board, self.player
if not self.mcts_tree or not self.mcts_tree['subtree']:
self.mcts_tree = new_tree(policy, self.board)
index = select_play(policy, self.board, self.mcts_simulations,
self.mcts_tree, self.temperature, self.model)
logger.info("Generated index %s", index)
x, y = index2coord(index)
# show_tree(x, y, self.mcts_tree)
policy_target = np.zeros(SIZE * SIZE + 1)
for _index, d in self.mcts_tree['subtree'].items():
policy_target[_index] = d['p']
self.board, self.player = self.play(color, x, y)
return x, y, policy_target, value, self.board, self.player