def reset(self):
self.root_node = MCTSNode()
self.root_node.is_root = True
self.root_node.is_leaf = True
self.training_data = []
self.training_score = []
self.training_move = []
self.simulate_board = LambdaGoBoard(self.board_size)
self.go_board = LambdaGoBoard(self.board_size)
def rollout(self, root_node, stop_at_leaf=False):
node = root_node
v = -1
finish = False
while not finish:
if node.is_end():
break
v = node.best_child()
if node.children[v] is None:
env = MISEnv() if use_dense else MISEnv_Sparse()
env.set_graph(node.graph)
next_graph, r, done, info = env.step(v)
node.children[v] = MCTSNode(next_graph,
self,
idx=v,
parent=node)
if stop_at_leaf:
finish = True
node = node.children[v]
# backpropagate V
V = node.state_value()
while node is not root_node:
V += 1
self.update_parent(node, V)
node = node.parent
self.root_max = max(self.root_max, V)
return V
def simulate_best_move(self, color, pos_filter=None):
# create a new node as root node, we may need to reuse node from original tree in the future:
self.root_node = MCTSNode()
self.root_node.set_simulate_board(self.go_board)
current_color = LambdaGoBoard.get_color_value(color)
# root node is the node before current player play the stone,
# so the color of root node should be color of current color
self.root_node.player_color = current_color
self.root_node.is_root = True
self.expand_mcts_node(self.root_node, None)
right_move = self.mcts_search(self.root_node, color)
return right_move
def search(self, graph, iter_num=10):
root_node = MCTSNode(graph, self)
ans = []
for i in range(iter_num):
r = self.rollout(root_node)
if self.performance: print(r)
ans.append(r)
return ans
def search_for_exp(self, graph, time_limit=600, min_iter_num=100):
now = time.time()
root_node = MCTSNode(graph, self)
ans = []
cnt = 0
while cnt < min_iter_num or time.time() - now < time_limit:
r = self.rollout(root_node)
ans.append(r)
cnt += 1
return ans
def train(self, graph, TAU, batch_size=10, iter_p=2, stop_at_leaf=False):
self.gnnhash.clear()
mse = torch.nn.MSELoss()
env = MISEnv() if use_dense else MISEnv_Sparse()
env.set_graph(graph)
graphs = []
actions = []
pis = []
means = []
stds = []
done = False
while not done:
n, _ = graph.shape
node = MCTSNode(graph, self)
means.append(node.reward_mean)
stds.append(node.reward_std)
pi = self.get_improved_pi(node,
TAU,
iter_p=iter_p,
stop_at_leaf=stop_at_leaf)
action = np.random.choice(n, p=pi)
graphs.append(graph)
actions.append(action)
pis.append(pi)
graph, reward, done, info = env.step(action)
T = len(graphs)
idxs = [i for i in range(T)]
np.random.shuffle(idxs)
i = 0
while i < T:
size = min(batch_size, T - i)
self.optimizer.zero_grad()
loss = torch.tensor([0], dtype=torch.float32)
for j in range(i, i + size):
idx = idxs[j]
Timer.start('gnn')
p, v = self.gnn(graphs[idx], True)
Timer.end('gnn')
n, _ = graphs[idx].shape
# normalize z with mean, std
z = torch.tensor(((T - idx) - means[idx]) / stds[idx],
dtype=torch.float32)
loss += torch.tensor(mse(z, v[actions[idx]]) - \
(torch.tensor(pis[idx], dtype=torch.float32) * torch.log(p + EPS)).sum(), dtype=torch.float32, requires_grad = True)
loss /= size
loss.backward()
self.optimizer.step()
i += size
def best_search1(self, graph, TAU=0.1, iter_p=1):
self.gnnhash.clear()
env = MISEnv() if use_dense else MISEnv_Sparse()
env.set_graph(graph)
ma = 0
reward = 0
done = False
while not done:
n, _ = graph.shape
node = MCTSNode(graph, self)
pi = self.get_improved_pi(node, TAU, iter_p=iter_p)
ma = max(ma, self.root_max + reward)
action = np.random.choice(n, p=pi)
graph, reward, done, info = env.step(action)
return ma, reward
def build_tree(self, num_simulations=5000):
# first, for every possible next code, we perform one simulation
for possible_code in self.possible_codes.get_all():
# print(len(self.possible_codes))
node = MCTSNode(code=possible_code)
node.perform_simulation(self.possible_codes,
next_code=possible_code)
self.nodes.append(node)
# then, we perform a simulation for remaining number of tries
# nodes to perform simulations are chosen by roulette wheel selection
num_simulations -= len(self.possible_codes)
for i in range(num_simulations):
# if i % 1000 == 0:
# print(i)
node = self._get_best_node()
node.perform_simulation(self.possible_codes,
next_code=node.code)
class MCTS(object):
def __init__(self, side, search_size, logger):
self.tree_manager = TreeManager()
self.tree_manager.c_uct = 2
self.actual_side = side
self.side = 1 - side
self.search_size = search_size
self.logger = logger
def init(self, actions, env):
self.root = MCTSNode(self.tree_manager, self.side, descendant=actions)
self.env = env
def uct(self):
self.logger.debug(f'=== UCT start === Step {self.env.game.step_count}')
self.logger.debug(f'game map {self.env.game.map}')
for i in range(self.search_size):
self.logger.debug('=== Search start ===')
node = self.root
env = self.env.clone()
player_move = 2
self.logger.debug(f'game map {env.game.map}')
self.logger.debug(f'real side {self.actual_side}')
self.logger.debug(f'step {env.game.step_count}')
self.logger.debug('== Select ==')
# Select
while node.untried_actions == [] and node.children != []:
# node is fully expanded and non-terminal
node = node.select()
env.take_side_action(node.side, node.action)
self.logger.debug(f'Select side {node.side}, '
f'action {node.action}')
player_move -= 1
if player_move == 0:
env.step()
self.logger.debug('Select env step')
player_move = 2
self.logger.debug('== Expand ==')
# Expand
if node.untried_actions:
# if we can expand (i.e. state/node is non-terminal)
action = random.choice(node.untried_actions)
env.take_side_action(1 - node.side, action)
self.logger.debug(f'Expand side {1 - node.side}, '
f'action {action}')
player_move -= 1
if player_move == 0:
env.step()
self.logger.debug('Expand env step')
player_move = 2
node = node.add_node(1 - node.side, action,
env.get_side_action(node.side))
self.logger.debug(f'Expand add node, side {node.side}, '
f'action {node.action}')
side = 1 - node.side
roll_out = 0
self.logger.debug('== Rollout ==')
# Rollout
while env.get_side_action(side):
# while state is non-terminal
actions = env.get_side_action(side)
action = random.choice(actions)
env.take_side_action(side, action)
self.logger.debug(f'Rollout side {side}, action {action}')
side = 1 - side
player_move -= 1
if player_move == 0:
env.step()
self.logger.debug('Rollout env step')
roll_out += 1
player_move = 2
winner = env.get_winner()
self.logger.debug(f'winner {winner}')
if winner >= 0:
if winner == node.side:
z = 1
else:
z = -1
else:
z = 0
self.logger.debug(f'z {z}')
# Backpropagate
while node is not None:
node.update(z, 0)
self.logger.debug(f'side {node.side}, '
f'action {node.action}, z {z}')
node = node.parent
z *= -1
self.logger.debug('=== Search end ===')
sorted_nodes = self.root.sorted_child()
self.logger.info(f'=== Side {self.actual_side} ===')
_count = 0
for node in sorted_nodes:
if _count < 3:
self.logger.info(node.str())
_count += 1
assert node.check_vls()
selected = sorted_nodes[0]
self.root = selected
return selected.action
def init(self, actions, env):
self.root = MCTSNode(self.tree_manager, self.side, descendant=actions)
self.env = env
class MCTSRobot(object):
def __init__(self,
name='DefaultMCTSRobot',
layer_number=19,
old_model=None,
search_time=100,
board_size=19,
komi=7.5,
train_iter=2):
self.name = name
self.layer_number = layer_number
self.search_time = search_time
self.train_iter = train_iter
self.board_size = board_size
self.komi = komi
self.stone_number = self.board_size * self.board_size
self.simulate_board_list = []
self.max_play_move = 1024
self.ColorBlackChar = 'b'
self.ColorWhiteChar = 'w'
self.BlackIdentify = np.ones((self.board_size, self.board_size),
dtype=int)
self.WhiteIdentify = np.ones(
(self.board_size, self.board_size), dtype=int) * -1
self.PosArray = []
for row in range(self.board_size):
for col in range(self.board_size):
self.PosArray.append((row, col))
self.PosArray.append(None)
self.reset()
if old_model is None:
self.model = DualHeadModel(self.name,
self.board_size,
layer_number=self.layer_number)
else:
print('Trying to load old model for continue training:' +
'./model/' + old_model)
self.model = DualHeadModel(self.name,
self.board_size,
model_path='./model/' + old_model,
layer_number=self.layer_number)
def reset(self):
self.root_node = MCTSNode()
self.root_node.is_root = True
self.root_node.is_leaf = True
self.training_data = []
self.training_score = []
self.training_move = []
self.simulate_board = LambdaGoBoard(self.board_size)
self.go_board = LambdaGoBoard(self.board_size)
def reset_board(self):
self.go_board.reset(self.board_size)
def apply_move(self, color, pos):
# print ('aplying move:' + color + ' in the position ' + str(pos))
current_data = []
current_board = self.go_board.get_board()
current_data.append(current_board)
if LambdaGoBoard.get_color_value(color) == LambdaGoBoard.ColorBlack:
current_data.append(self.BlackIdentify)
elif LambdaGoBoard.get_color_value(color) == LambdaGoBoard.ColorWhite:
current_data.append(self.WhiteIdentify)
else:
raise Exception('Incorrect color character')
self.training_data.append(current_data)
if pos is None:
current_move = np.zeros((self.board_size * self.board_size + 1),
dtype=int)
current_move[self.board_size * self.board_size] = 1
self.training_move.append(current_move)
else:
(row, col) = pos
current_move = np.zeros((self.board_size * self.board_size + 1),
dtype=int)
current_move[row * self.board_size + col] = 1
self.training_move.append(current_move)
is_valid = self.go_board.apply_move(color, pos)
# if not is_valid:
# print ('# incorrect move:' + color + ' pos:' + str(pos) + ' Reason:' + str(reason))
# else:
# print ('# correct move:' + color + ' pos:' + str(pos) + ' ')
# start_time = time.time()
# self.board.update_score_board()
# end_time = time.time()
# print ('total update time:' + str(end_time - start_time))
# print(str(self.board))
def showboard(self):
# self.go_board.update_score_board()
return str(self.go_board)
def simulate_best_move(self, color, pos_filter=None):
# create a new node as root node, we may need to reuse node from original tree in the future:
self.root_node = MCTSNode()
self.root_node.set_simulate_board(self.go_board)
current_color = LambdaGoBoard.get_color_value(color)
# root node is the node before current player play the stone,
# so the color of root node should be color of current color
self.root_node.player_color = current_color
self.root_node.is_root = True
self.expand_mcts_node(self.root_node, None)
right_move = self.mcts_search(self.root_node, color)
return right_move
def mcts_search(self, root_node, color):
right_move = (None, 0)
for i in range(self.search_time):
# node_visited = []
# start_time = time.time()
self.search_to_expand(root_node)
# end_time = time.time()
# print ('Search to Expand time:' + str(end_time - start_time))
# print ('searching......, iter:' + str(i))
# value = self.expand_mcts_node(node_visited)
# self.node_back_up(node_visited, value)
right_node = self.lookup_right_node(self.root_node)
right_move = (right_node.move, right_node.get_value())
self.display_result(color, right_move, right_node)
# print ('Found right move:' + str(right_move[0]) + 'with value:' + str(right_move[1]))
# print ('Visited count: ' + str(right_node.visit_count))
# time.sleep(5)
return right_move
def search_to_expand(self, current_node):
# if not current_node.is_leaf:
# # it is not a leaf node, search the best one and call search_to_expand again
best_value = -10000
best_node = None
for child in current_node.children:
# print ('searching...........' + str(child.move))
# print ('current value of best node:' + str(child.current_value))
# print ('average value of best node:' + str(child.average_value))
# print ('total value of best node:' + str(child.total_value))
# print ('visit count of best node:' + str(child.visit_count))
# print ('policy value of best node:' + str(child.policy_value))
# print ('------------------------------------')
node_value = child.get_value()
if node_value > best_value:
best_value = node_value
best_node = child
# print ('found best node:' + str(best_node.move) + ' with value:' + str(best_value))
# print ('current value of best node:' + str(best_node.current_value))
# print ('average value of best node:' + str(best_node.average_value))
# print ('total value of best node:' + str(best_node.total_value))
# print ('visit count of best node:' + str(best_node.visit_count))
# print ('policy value of best node:' + str(best_node.policy_value))
if best_node is None:
raise Exception('Node without valid child')
else:
if best_node.is_leaf:
# print('trying to expand node.' + str(best_node.move))
# start_time = time.time()
value = self.expand_mcts_node(best_node, current_node)
# end_time = time.time()
# print (' node Expand time:' + str(end_time - start_time))
else:
# print('searching into best child:' + str(best_node.move) + '.........................................')
value = self.search_to_expand(best_node)
# print('value from child is:' + str(value))
# print('number of children of current node:' +str(len(current_node.children)))
current_node.visit_count = current_node.visit_count + 1
current_node.total_value = current_node.total_value + value
current_node.average_value = (
current_node.total_value /
current_node.visit_count) / self.stone_number
current_node.current_value = current_node.average_value
return value
# else:
# value = self.expand_mcts_node(current_node)
# return value
def expand_mcts_node(self, current_node, parent_node):
# if parent_node is None:
# # no parent node, it is root node, we can use the value in current node, which is root node that was initialized.
# current_color = current_node.player_color
# else:
# # copy data from parent node
# parent_color = parent_node.player_color
# current_color = LambdaGoBoard.reverse_color_value(parent_color)
# # enemy_color = parent_node.player_color
# current_node.player_color = current_color
# current_node.simulate_board = LambdaGoBoard(parent_node.simulate_board.board_size)
# current_node.simulate_board.copy_from(parent_node.simulate_board)
# current_node.simulate_board.apply_move(LambdaGoBoard.get_color_char(current_color), current_node.move)
current_color = current_node.player_color
child_color = LambdaGoBoard.reverse_color_value(current_color)
# current_data_list = []
current_data = []
current_board = current_node.simulate_board.get_board()
current_data.append(current_board)
if current_color == LambdaGoBoard.ColorBlack:
current_data.append(self.BlackIdentify)
elif current_color == LambdaGoBoard.ColorWhite:
current_data.append(self.WhiteIdentify)
else:
raise Exception('Incorrect color character')
# for i in range(361):
# current_data_list.append(current_data)
# predict_start_time = time.time()
result = self.model.predict(current_data)
# predict_end_time = time.time()
# print (' Predict time:' + str(predict_end_time - predict_start_time))
policy_array = result[0][0]
current_value = result[1][0]
move_and_policy_value = zip(self.PosArray, policy_array)
# # need to consider whether sorting this array helps to improve the MCTS searching speed.
# move_and_policy_value.sort(key=lambda x:x[1], reverse=True)
# simulate_start_time = time.time()
for single_move_and_policy in move_and_policy_value:
move = single_move_and_policy[0]
policy_value = single_move_and_policy[1]
new_child = MCTSNode()
new_child.simulate_board = LambdaGoBoard(self.board_size)
new_child.simulate_board.copy_from(current_node.simulate_board)
is_valid, reason = new_child.simulate_board.apply_move(
LambdaGoBoard.get_color_char(current_color), move)
if is_valid:
new_child.move = move
new_child.policy_value = policy_value
new_child.player_color = child_color
current_node.children.append(new_child)
# simulate_end_time = time.time()
# print (' Simulating time:' + str(simulate_end_time - simulate_start_time))
current_node.is_leaf = False
current_node.visit_count = current_node.visit_count + 1
current_node.total_value = current_node.total_value + current_value
current_node.average_value = (
current_node.total_value /
current_node.visit_count) / self.stone_number
current_node.current_value = current_node.average_value
return current_value
def lookup_right_node(self, current_node):
most_visited_count = -1
best_policy = -1
most_visited_node = None
for child in current_node.children:
if child.visit_count > most_visited_count:
most_visited_count = child.visit_count
best_policy = child.policy_value
most_visited_node = child
elif child.visit_count == most_visited_count:
if child.policy_value > best_policy:
best_policy = child.policy_value
most_visited_node = child
return most_visited_node
def select_move(self, color):
right_move = self.simulate_best_move(color)
self.go_board.apply_move(color, right_move[0])
# print ('# selected move:' + str(right_move))
return right_move[0]
def display_result(self, color, right_move, right_node):
display_string = "# Player: "
if LambdaGoBoard.get_color_value(color) == LambdaGoBoard.ColorBlack:
display_string = display_string + "Black "
else:
display_string = display_string + "White "
move_string = ' Move:' + str(right_move[0]) + ' '
value_string = ' Value:' + str(right_move[1]) + ' '
visit_count_string = ' Count:' + str(
right_node.visit_count) + ' '
node_value_string = ' NodeValue:' + str(
right_node.average_value) + ' '
policy_value_string = ' Policy:' + str(
right_node.policy_value) + ' '
display_string = display_string + move_string[
0:20] + visit_count_string[0:20]
display_string = display_string + value_string[
0:25] + node_value_string[0:25] + policy_value_string[0:25]
if LambdaGoBoard.get_color_value(color) == LambdaGoBoard.ColorBlack:
print(display_string)
print('# ')
else:
print('# ')
print(display_string)
if abs(right_node.average_value) > 1:
print('# incorrect node:')
print(str(right_node))
time.sleep(20)
def train(self, board_states, move_sequence, score_board):
train_data_len = len(self.training_data)
train_move_len = len(self.training_move)
if not train_data_len == train_move_len:
raise Exception(
'Inconsist state, training data and training move not in same length'
)
self.go_board.update_score_board()
result_score = self.get_score()
if result_score > self.komi:
for i in range(len(self.training_move)):
if i % 2 == 1:
# black win while current move is for white, revert it
self.training_move[i] = self.training_move[1] * -1 + 1
else:
for i in range(len(self.training_move)):
if i % 2 == 0:
# white win while current move is for black, revert it
self.training_move[i] = self.training_move[1] * -1 + 1
result_score_board = self.go_board.score_board
for i in range(train_data_len):
self.training_score.append(result_score_board)
print('# robot ' + self.name + ' is in training.........')
self.model.train(self.training_data,
self.training_score,
self.training_move,
steps=self.train_iter)
def new_game(self):
self.reset_board()
self.simulate_board_list = []
for i in range(self.board_size * self.board_size):
self.simulate_board_list.append(LambdaGoBoard(self.board_size))
def get_current_state(self):
return self.go_board.board
def get_score_board(self):
if not self.go_board.score_board_updated:
self.go_board.update_score_board()
return self.go_board.score_board
def get_score(self):
if not self.go_board.score_board_updated:
self.go_board.update_score_board()
return self.go_board.score_board_sum
def expand_mcts_node(self, current_node, parent_node):
# if parent_node is None:
# # no parent node, it is root node, we can use the value in current node, which is root node that was initialized.
# current_color = current_node.player_color
# else:
# # copy data from parent node
# parent_color = parent_node.player_color
# current_color = LambdaGoBoard.reverse_color_value(parent_color)
# # enemy_color = parent_node.player_color
# current_node.player_color = current_color
# current_node.simulate_board = LambdaGoBoard(parent_node.simulate_board.board_size)
# current_node.simulate_board.copy_from(parent_node.simulate_board)
# current_node.simulate_board.apply_move(LambdaGoBoard.get_color_char(current_color), current_node.move)
current_color = current_node.player_color
child_color = LambdaGoBoard.reverse_color_value(current_color)
# current_data_list = []
current_data = []
current_board = current_node.simulate_board.get_board()
current_data.append(current_board)
if current_color == LambdaGoBoard.ColorBlack:
current_data.append(self.BlackIdentify)
elif current_color == LambdaGoBoard.ColorWhite:
current_data.append(self.WhiteIdentify)
else:
raise Exception('Incorrect color character')
# for i in range(361):
# current_data_list.append(current_data)
# predict_start_time = time.time()
result = self.model.predict(current_data)
# predict_end_time = time.time()
# print (' Predict time:' + str(predict_end_time - predict_start_time))
policy_array = result[0][0]
current_value = result[1][0]
move_and_policy_value = zip(self.PosArray, policy_array)
# # need to consider whether sorting this array helps to improve the MCTS searching speed.
# move_and_policy_value.sort(key=lambda x:x[1], reverse=True)
# simulate_start_time = time.time()
for single_move_and_policy in move_and_policy_value:
move = single_move_and_policy[0]
policy_value = single_move_and_policy[1]
new_child = MCTSNode()
new_child.simulate_board = LambdaGoBoard(self.board_size)
new_child.simulate_board.copy_from(current_node.simulate_board)
is_valid, reason = new_child.simulate_board.apply_move(
LambdaGoBoard.get_color_char(current_color), move)
if is_valid:
new_child.move = move
new_child.policy_value = policy_value
new_child.player_color = child_color
current_node.children.append(new_child)
# simulate_end_time = time.time()
# print (' Simulating time:' + str(simulate_end_time - simulate_start_time))
current_node.is_leaf = False
current_node.visit_count = current_node.visit_count + 1
current_node.total_value = current_node.total_value + current_value
current_node.average_value = (
current_node.total_value /
current_node.visit_count) / self.stone_number
current_node.current_value = current_node.average_value
return current_value
class MCTSP(object):
def __init__(self):
import threading
self.pool = Pool('tankai', 0)
self.pool.setup()
self.pool.reg_task([random_rollout])
self.backup_lock = threading.Lock()
self.select_lock = threading.Lock()
self.log_info = {}
def setup(self, side, search_size, logger):
self.tree_manager = TreeManager()
self.tree_manager.c_uct = 2
self.actual_side = side
self.side = 1 - side
self.search_size = search_size
self.logger = logger
def reset_root(self, actions):
self.root = MCTSNode(self.tree_manager, self.side, descendant=actions)
def select(self, node):
self.select_lock.acquire()
action_list = []
# Select
while node.untried_actions == [] and node.children:
# node is fully expanded and non-terminal
node = node.select()
action_list.append((node.side, node.action))
self.select_lock.release()
return action_list, node
def _backup(self, node, flag, result):
z = result['z']
# backpropagate from the expanded node and work back to the root node
while node is not None:
if (node.is_leaf() and flag) or node.is_root():
# node is add in the front of the backup
# or node is the root, which won't apply vl
node.update(z, 0, {})
else:
node.update(z, 3, {})
node = node.parent
z *= -1
self.backup_lock.acquire()
self.jobs.pop(str(result['action_list']))
self.pbar.update()
self.backup_lock.release()
def backup(self, result):
self.select_lock.acquire()
uid = result['uid']
node, flag = self.backup_dict[uid]
if flag:
# add child and descend tree
side, action = result['action_list'][-1]
node = node.add_node(1 - node.side, action, result['descendants'])
self._backup(node, flag, result)
self.backup_dict.pop(uid, None)
self.select_lock.release()
def uct(self, env):
import sys
import pickle
import zlib
import random
import time
import uuid
self.info = {}
self.info['results'] = []
self.jobs = {}
self.backup_dict = {}
self.pbar = tqdm(total=self.search_size, file=sys.stdout)
_env = zlib.compress(pickle.dumps(env, -1))
for i in range(self.search_size):
wait = True
while wait:
node = self.root
untried_flag = False
action_list, node = self.select(node)
if node.untried_actions:
untried_flag = True
# Expand
if untried_flag:
# if we can expand (i.e. state/node is non-terminal)
action = random.choice(node.untried_actions)
action_list.append((1 - node.side, action))
if str(action_list) in self.jobs:
time.sleep(0.01)
else:
self.backup_lock.acquire()
self.jobs[str(action_list)] = 1
tmp_node = self.root
if untried_flag:
for a in action_list[:-1]:
tmp_node = tmp_node.select_child_by_action(a[1])
tmp_node.apply_vl(3)
else:
for a in action_list:
tmp_node = tmp_node.select_child_by_action(a[1])
tmp_node.apply_vl(3)
self.backup_lock.release()
wait = False
uid = str(uuid.uuid4())
self.backup_dict[uid] = [node, untried_flag]
self.pool.apply_async(random_rollout, (uid, _env, action_list, 0),
self.backup)
self.pool.join()
self.pbar.close()
sorted_nodes = self.root.sorted_child()
_count = 0
for node in sorted_nodes:
if _count < 3:
self.logger.info(node.str())
_count += 1
assert node.check_vls()
selected = sorted_nodes[0]
return selected
def reset_root(self, actions):
self.root = MCTSNode(self.tree_manager, self.side, descendant=actions)