def policy(self, game):
with self._graph.as_default():
predictions = self._model.predict(
-game.board().reshape(1, 15, 15, 1))
available = numpy.zeros((225, 1))
positions = util.list_positions(game.board(), renju.Player.NONE)
for pos in positions:
available[pos[0] * 15 + pos[1]] = 1
arr = predictions.T + available * 0.1
code_move = numpy.argmax(arr)
print(self._name + ':', util.to_move([code_move // 15,
code_move % 15]))
return arr
def process_move(self, move):
# check and record move, if not finished
if not self.game.is_posible_move(move):
self.board_canvas.print_message("Move is invalid!")
return False
self.board_canvas.place_move(move, self.game._player.__repr__())
self.board_canvas.print_message(self.game._player.__repr__() +
" move is: " + util.to_move(move))
if not self.game.move(move):
self.finish_game("Player " + self.game._player.__repr__() +
" wins")
return False
return True
def choose_move(board):
positions = util.list_positions(board, renju.Player.NONE)
model = load_model('modelusual.h5')
model_graph = tf.get_default_graph()
model_graph_2 = tf.get_default_graph()
model_2 = load_model('modelsuper.h5')
SUMEKENOV = CnnAgent(color='black',
name='SUMEKENOV',
model=(model, model_graph))
SUMEKENOV_2 = CnnAgent(color='black',
name='SUMEKENOV_2',
model=(model_2, model_graph_2))
arr = SUMEKENOV.policy(game)
arr_2 = SUMEKENOV_2.policy(game)
arr_final = arr + arr_2
mv = numpy.argmax(arr_final, axis=1)
return util.to_move(mv)
def policy(self, game):
if (self.flag):
newBoard = numpy.zeros((1, 15, 15, 1))
for i in range(15):
for j in range(15):
if (game.board()[i][j] == -1):
newBoard[0][i][j][0] = 0.5
if (game.board()[i][j] == 1):
newBoard[0][i][j][0] = 1
with self._graph.as_default():
predictions = self._model.predict(newBoard)
available = numpy.zeros((225, 1))
positions = util.list_positions(game.board(), renju.Player.NONE)
for pos in positions:
available[pos[0] * 15 + pos[1]] = 1
arr = predictions.T * available
code_move = numpy.argmax(arr)
# maxmove = 0
# move = 0
# xy = 0
# for m in predictions[0]:
# if ((m > maxmove) and (available[xy][0])):
# move = xy
# maxmove = m
# xy += 1
# code_move = maxmove
print(self._name + ':',
util.to_move([code_move // 15, code_move % 15]))
return arr
else:
move = input()
pos = util.to_pos(move)
probs = numpy.zeros(game.shape)
probs[pos] = 1.0
return probs
def policy(self, game):
if (self._color == 'black' and len(
util.list_positions(game.board(), renju.Player.NONE)) == 225):
res = numpy.zeros((225, 1))
res[142] = 1
return res.reshape((1, 225))
predictions = numpy.zeros((1, 225))
for model in self._models:
with model[1].as_default():
predictions += model[0].predict(
-game.board().reshape(1, 15, 15, 1))
available = numpy.zeros((225, 1))
positions = util.list_positions(game.board(), renju.Player.NONE)
for pos in positions:
available[pos[0] * 15 + pos[1]] = 1
arr = predictions.T + available
code_move = numpy.argmax(arr)
if (self._verbose):
print(self._name + ':',
util.to_move([code_move // 15, code_move % 15]))
return arr
def policy(self, game):
from time import time
if (self._color == 'black' and len(
util.list_positions(game.board(), renju.Player.NONE)) == 225):
res = numpy.zeros((225, 1))
res[112] = 1
return res.reshape((1, 225))
board = numpy.copy(-game.board())
available = numpy.zeros(225)
checker = numpy.zeros(225)
for parsed_pos in util.list_positions(game.board(), renju.Player.NONE):
pos = parsed_pos[0] * 15 + parsed_pos[1]
parsed_pos = tuple(parsed_pos)
board[parsed_pos] = 1
if (util.check(board, parsed_pos)):
checker[pos] += 1
board[parsed_pos] = -1
if (util.check(board, parsed_pos)):
checker[pos] += 1
board[parsed_pos] = 0
available[pos] = 1
start = time()
with self._graph.as_default():
predictions = self._model.predict(board.reshape(1, 15, 15, 1))[0]
arr = (predictions * available) * (1 + checker)
code_move = numpy.argmax(arr)
if (self._verbose):
print(self._name + ':',
util.to_move([code_move // 15, code_move % 15]),
time() - start)
return arr.reshape(1, 225)
def policy(self, game):
start = time()
if (self._color == 'black' and len(
util.list_positions(game.board(), renju.Player.NONE)) == 225):
res = numpy.zeros((225, 1))
res[112] = 1
return res.reshape((1, 225))
predictions = 0
temp_high = 0
if self._color == 'black':
with self._black_graph.as_default():
predictions = self._black_model.predict(
-game.board().reshape(1, 15, 15, 1),
batch_size=1,
verbose=0)[0]
else:
with self._white_graph.as_default():
predictions = self._white_model.predict(
-game.board().reshape(1, 15, 15, 1),
batch_size=1,
verbose=0)[0]
res = numpy.zeros(225)
n = numpy.zeros(225)
def rollout(board,
temp_high,
color,
checking_pos,
max_high=self._high,
gamma=self._gamma):
temp_predictions = 0
if color == 'black':
with self._black_graph.as_default():
temp_predictions = self._black_model.predict(board.reshape(
1, 15, 15, 1),
batch_size=1,
verbose=0)[0]
else:
with self._white_graph.as_default():
temp_predictions = self._white_model.predict(board.reshape(
1, 15, 15, 1),
batch_size=1,
verbose=0)[0]
temp_pos = numpy.random.choice(225, p=temp_predictions)
temp_parsed_pos = numpy.unravel_index(temp_pos, (15, 15))
if (board[temp_parsed_pos] != 0):
if (temp_high % 2):
res[checking_pos] += gamma**(temp_high - 1)
else:
res[checking_pos] -= gamma**(temp_high - 1)
return
if (util.check(board, temp_parsed_pos)):
if (temp_high % 2):
res[checking_pos] -= gamma**(temp_high - 1)
else:
res[checking_pos] += gamma**(temp_high - 1)
return
board[temp_parsed_pos] = 1
next_color = 'white'
if color == 'white':
next_color = 'black'
board[temp_parsed_pos] = -1
if (temp_high < max_high):
rollout(board, temp_high + 1, next_color, checking_pos)
return
i = 0
while (time() - start < self._timeout and i < self._samples):
i += 1
pos = numpy.random.choice(225, p=predictions)
parsed_pos = numpy.unravel_index(pos, (15, 15))
board = numpy.copy(-game.board())
n[pos] += 1
if (board[parsed_pos] != 0):
res[pos] -= 1
continue
if (util.check(board, parsed_pos)):
res[pos] = res.max() + 1
break
board[parsed_pos] = 1
color = 'white'
if self._color == 'white':
color = 'black'
board[parsed_pos] = -1
rollout(board, 1, color, pos)
values = (res + 10 * predictions) / (1 + n) + (res > 0.5 * n)
if (self._verbose):
code_move = numpy.argmax(values.reshape(1, 225))
print(self._name + ':', util.to_move([code_move // 15, code_move % 15]), \
'working time:', time() - start, 'iterations:', i)
#values -= values.mean()
#values /= values.std()
return values.reshape(1, 225)
def policy(self, game):
if (self._color == 'black' and len(
util.list_positions(game.board(), renju.Player.NONE)) == 225):
res = numpy.zeros((225, 1))
res[142] = 1
return res.reshape((1, 225))
predictions = 0
temp_high = 0
if self._color == 'black':
#positions = util.list_positions(game.board(), renju.Player.NONE)
#if (len(positions) == 225):
with self._black_graph.as_default():
predictions = self._black_model.predict(
-game.board().reshape(1, 15, 15, 1),
batch_size=1,
verbose=0)[0]
else:
with self._white_graph.as_default():
predictions = self._white_model.predict(
-game.board().reshape(1, 15, 15, 1),
batch_size=1,
verbose=0)[0]
max_sum_log = -100000000000000
best_move = -1
results = dict()
danger = set()
def check_state_white(board, pos, checking_pos, temp_prob, temp_high,
results):
# print(board, pos)
temp_high += 1
parsed_pos = numpy.unravel_index(pos, (15, 15))
#if (util.check(-board, parsed_pos)):
#if self._color == 'white':
#checking_pos = -1
#return
#results[0] = checking_pos
#results[1] = 0
#return
if (temp_high > self._high):
if (checking_pos not in danger):
if (results[checking_pos] < temp_prob):
results[checking_pos] = temp_prob
return
board[parsed_pos] = 1
white_predict = 0
with self._white_graph.as_default():
white_predict = self._white_model.predict(
board.reshape(1, 15, 15, 1))[0]
top_args_white = numpy.argsort(white_predict)[::-1][:self._width]
for pos in top_args_white:
if (util.check(-board, parsed_pos)):
if (self._color == 'black'):
danger.add(checking_pos)
continue
if checking_pos not in danger:
check_state_black(
numpy.copy(board), pos, checking_pos,
temp_prob + numpy.log(white_predict[pos]), temp_high,
results)
def check_state_black(board, pos, checking_pos, temp_prob, temp_high,
results):
# print(board, pos)
# global max_sum_log, best_move
temp_high += 1
parsed_pos = numpy.unravel_index(pos, (15, 15))
#if (util.check(-board, parsed_pos)):
# if self._color == 'black':
#checking_pos = -1
#return
#results[0] = checking_pos
#results[1] = 0
#return
if (temp_high > self._high):
if (checking_pos not in danger):
if (results[checking_pos] < temp_prob):
results[checking_pos] = temp_prob
return
board[parsed_pos] = -1
black_predict = 0
with self._black_graph.as_default():
black_predict = self._black_model.predict(
board.reshape(1, 15, 15, 1))[0]
top_args_black = numpy.argsort(black_predict)[::-1][:self._width]
for pos in top_args_black:
if (util.check(-board, parsed_pos)):
if (self._color == 'white'):
danger.add(checking_pos)
continue
if checking_pos not in danger:
check_state_white(
numpy.copy(board), pos, checking_pos,
temp_prob + numpy.log(black_predict[pos]), temp_high,
results)
top_args = numpy.argsort(predictions)[::-1][:self._width]
# print(top_args)
for pos in top_args:
results[pos] = -100000000000
if (self._verbose):
print(numpy.unravel_index(pos, (15, 15)), end=' - ')
if (self._color == 'black'):
check_state_white(numpy.copy(-game.board()), pos, pos,
numpy.log(predictions[pos]), temp_high,
results)
else:
check_state_black(numpy.copy(-game.board()), pos, pos,
numpy.log(predictions[pos]), temp_high,
results)
max_sum_log, best_move = -100000000, -1
for move in results:
if move not in danger:
if max_sum_log < results[move]:
max_sum_log = results[move]
best_move = move
res = numpy.zeros((225, 1))
res[best_move] = 1
code_move = numpy.argmax(res.reshape(1, 225))
if (self._verbose):
print(self._name + ':',
util.to_move([code_move // 15, code_move % 15]))
return res.reshape(1, 225)
def policy(self, game):
start = time()
if (self._color == 'black' and len(
util.list_positions(game.board(), renju.Player.NONE)) == 225):
res = numpy.zeros((225, 1))
res[112] = 1
return res.reshape((1, 225))
predictions = 0
temp_high = 0
checker = numpy.zeros(225)
available = numpy.zeros(225)
board = numpy.copy(-game.board())
for parsed_pos in util.list_positions(game.board(), renju.Player.NONE):
pos = parsed_pos[0] * 15 + parsed_pos[1]
parsed_pos = tuple(parsed_pos)
board[parsed_pos] = 1
if (util.check(board, parsed_pos)):
checker[pos] += 1
board[parsed_pos] = -1
if (util.check(board, parsed_pos)):
checker[pos] += 1
board[parsed_pos] = 0
available[pos] = 1
if self._color == 'black':
with self._black_graph.as_default():
predictions = self._black_model.predict(board.reshape(
1, 15, 15, 1),
batch_size=1,
verbose=0)[0]
else:
with self._white_graph.as_default():
predictions = self._white_model.predict(board.reshape(
1, 15, 15, 1),
batch_size=1,
verbose=0)[0]
logs = numpy.zeros(225, dtype=numpy.float32)
res = numpy.zeros(225, dtype=numpy.float32)
n = numpy.zeros(225, dtype=numpy.float32)
def rollout(board,
temp_high,
color,
checking_pos,
sum_logs,
max_high=self._high,
gamma=self._gamma,
fine=self._fine,
bonus=self._bonus):
temp_predictions = 0
if color == 'black':
with self._black_graph.as_default():
temp_predictions = self._black_model.predict(board.reshape(
1, 15, 15, 1),
batch_size=1,
verbose=0)[0]
else:
with self._white_graph.as_default():
temp_predictions = self._white_model.predict(board.reshape(
1, 15, 15, 1),
batch_size=1,
verbose=0)[0]
temp_pos = numpy.random.choice(225, p=temp_predictions)
temp_parsed_pos = numpy.unravel_index(temp_pos, (15, 15))
if (board[temp_parsed_pos] != 0):
if (temp_high % 2 == 0):
res[checking_pos] += fine * gamma**(temp_high - 1)
else:
res[checking_pos] -= bonus * gamma**(temp_high - 1)
return
board[temp_parsed_pos] = 1
next_color = 'white'
if color == 'white':
next_color = 'black'
board[temp_parsed_pos] = -1
if (util.check(board, temp_parsed_pos)):
if (temp_high % 2 == 0):
res[checking_pos] -= fine * gamma**(temp_high - 1)
else:
res[checking_pos] += bonus * gamma**(temp_high - 1)
return
if (temp_high < max_high):
rollout(board, temp_high + 1, next_color, checking_pos,
sum_logs + numpy.log(temp_predictions[temp_pos]))
else:
if logs[checking_pos]:
logs[checking_pos] = max(
logs[checking_pos],
sum_logs + numpy.log(temp_predictions[temp_pos]))
else:
logs[checking_pos] = sum_logs + numpy.log(
temp_predictions[temp_pos])
return
i = 0
while (time() - start < self._timeout and i < self._samples):
i += 1
pos = numpy.random.choice(225, p=predictions)
parsed_pos = numpy.unravel_index(pos, (15, 15))
n[pos] += 1
if (board[parsed_pos] != 0):
res[pos] -= 1
continue
board[parsed_pos] = 1
color = 'white'
if self._color == 'white':
color = 'black'
board[parsed_pos] = -1
if (util.check(board, parsed_pos)):
res[pos] += res.max() + 1000
break
rollout(numpy.copy(board), 1, color, pos,
numpy.log(predictions[pos]))
board[parsed_pos] = 0
for indx in numpy.where(logs == 0.0):
logs[indx] = logs.min() - 100.0
value = (logs) * available * (1 + checker)
#print(value)
value /= value.sum()
if numpy.max(value) > 0.8:
code_move = numpy.argmax(value)
else:
if (self._verbose):
print('----!!!----')
code_move = numpy.argmax((n + res) * available * (1 + checker))
ans = numpy.zeros(225)
ans[code_move] = 1
if (self._verbose):
for elem in numpy.where(res != 0)[0]:
print(util.to_move(numpy.unravel_index(int(elem), (15, 15))),
'R:', res[int(elem)], 'V:', value[int(elem)], 'N:',
n[int(elem)])
# code_move = numpy.argmax(values.reshape(1, 225))
print(self._name + ':', util.to_move([code_move // 15, code_move % 15]), \
'working time:', time() - start, 'iterations:', i)
return ans.reshape(1, 225)
def policy(self, game):
self._start_time = time()
if not self._color:
if ((225 - len(util.list_positions(game.board(), renju.Player.NONE))) % 2 == 1):
self._color = 'white'
else:
self._color = 'black'
if (len(util.list_positions(game.board(), renju.Player.NONE)) == 225):
res = numpy.zeros((225, 1))
res[112] = 1
return res.reshape((1, 225))
done = True
if (self._root and len(util.list_positions(game.board(), renju.Player.NONE)) < 224):
to_do_moves = game._positions[-2:]
for move in to_do_moves:
if not self.make_move(move[0] * 15 + move[1]):
done = False
else:
done = False
if self._verbose:
print(done)
checker = np.zeros(225)
available = numpy.zeros(225)
self._board = np.copy(-game.board())
for parsed_pos in util.list_positions(self._board, renju.Player.NONE):
pos = parsed_pos[0] * 15 + parsed_pos[1]
parsed_pos = tuple(parsed_pos)
self._board[parsed_pos] = 1
if (util.check(self._board, parsed_pos)):
checker[pos] += 1
self._board[parsed_pos] = -1
if (util.check(self._board, parsed_pos)):
checker[pos] += 1
self._board[parsed_pos] = 0
available[pos] = 1
if not done:
self._root = Node(self._board, self._color, self._node_model_black, self._node_model_white)
self.tree_search()
values = (self._root._N > self._iters / 5) * self._root._R / (1 + self._root._N) * available * (1 + 10 * checker)
if np.max(values) > self._min_prob:
code_move = np.argmax(values)
else:
code_move = np.argmax((self._root._N) * available * (1 + 10 * checker))
res = np.zeros(225)
res[code_move] = 1
if (self._verbose):
for elem in np.where(self._root._R != 0)[0]:
try:
print(util.to_move(numpy.unravel_index(int(elem), (15, 15))), 'R:', self._root._R[int(elem)], 'V:', values[int(elem)])
except:
print(np.where(self._root._R != 0))
continue
print(self._name + ':', util.to_move([code_move // 15, code_move % 15]), \
'working time:', time() - self._start_time, 'iterations:', self._iters)
return res.reshape(1, 225)