def mcts(payload):
root = Node(None, None)
my_id = (payload['role_id'] + 2) % 3
next_id = (payload['role_id'] + 1) % 3
next_next_id = (payload['role_id'] + 2) % 3
my_card_ = payload['hand_card'][payload['role_id']]
my_card_.sort()
my_card = card_list_to_dict(card_to_list(change_card_form_reversal(my_card_)))
next_card_ = payload['hand_card'][next_id]
next_card_.sort()
next_card = card_list_to_dict(card_to_list(change_card_form_reversal(next_card_)))
next_next_card_ = payload['hand_card'][next_next_id]
next_next_card_.sort()
next_next_card = card_list_to_dict(card_to_list(change_card_form_reversal(next_next_card_)))
last_move_, last_p_ = get_last_move(payload['role_id'], next_id, next_next_id, payload['last_taken'])
last_move = change_card_form_reversal(last_move_)
last_p = (last_p_ + 2) % 3
moves_num = len(get_moves(my_card, last_move))
state = State(my_id, my_card, next_card, next_next_card, last_move, -1, moves_num, None, last_p)
root.set_state(state)
computation_budget = 1000
for i in range(computation_budget):
expand_node = tree_policy(root, my_id)
reward = default_policy(expand_node, my_id)
backup(expand_node, reward)
best_next_node = get_bestchild_(root)
move = best_next_node.get_state().action
return move
def get_next_state_with_random_choice(self, untried_move):
# 下家变自家，下下家变下家，自家变下下家
valid_moves = get_moves(self.my_card, self.last_move)
moves_num = len(valid_moves)
i = np.random.choice(moves_num)
tmp = valid_moves[i].copy()
if untried_move is not None:
tmp = untried_move
while self.is_buchu(tmp) and self.last_pid == self.my_id:
i = np.random.choice(moves_num)
tmp = valid_moves[i].copy()
move = []
next_next_card = self.my_card.copy()
for k in [str(i) for i in range(3, 14)] + ['1', '2', '14', '15']:
move.extend([int(k)] * tmp.get(k, 0))
next_next_card[k] -= tmp.get(k, 0)
my_id = (self.my_id + 1) % 3
my_card = self.next_card.copy()
next_card = self.next_next_card.copy()
# 判断出完牌游戏是否结束
winner = self.my_id
for lis in next_next_card.values():
if lis != 0:
winner = -1
break
last_move = move.copy()
last_p = self.my_id
# 如果选择不出， 下家的last_move等于自家的last_move
if len(move) == 0:
last_p = self.last_pid
last_move = self.last_move.copy()
if len(move) == 0 and self.last_pid == my_id:
last_move = []
valid_moves_ = get_moves(my_card, last_move)
moves_num_ = len(valid_moves_)
next_state = State(my_id, my_card, next_card, next_next_card,
last_move, winner, moves_num_, move, last_p)
return next_state
def expand(self):
if self.state.try_flag == 0:
valid_moves = get_moves(self.state.my_card, self.state.last_move)
for move in valid_moves:
self.state.init_untried_actions(move)
self.state.try_flag = 1
moves_num = len(self.state.untried_actions)
i = np.random.choice(moves_num)
untried_move = self.state.untried_actions[i].copy()
while self.state.is_buchu(untried_move) and self.state.last_pid == self.state.my_id:
i = np.random.choice(moves_num)
untried_move = self.state.untried_actions[i].copy()
new_state = self.get_state().get_next_state_with_random_choice(untried_move)
del self.state.untried_actions[i]
sub_node = Node(self, new_state)
self.add_child(sub_node)
return sub_node
def choose(self, state):
min_crads = min([sum(p.get_hand_card()) for p in self.game.players])
if min_crads > 7:
# 获得手牌
hand_card = self.get_hand_card()
# 拆牌器和引擎用了不同的编码 1 -> A, B -> *, R -> $
trans_hand_card = [
card_list[i] for i in range(15) for _ in range(hand_card[i])
]
# 获得上家出牌
last_move = [
card_list[i] for i in range(15)
for _ in range(state.last_move[i])
]
# 拆牌
D = Decomposer()
combs, fine_mask = D.get_combinations(trans_hand_card, last_move)
# 根据对手剩余最少牌数决定每多一手牌的惩罚
left_crads = [sum(p.get_hand_card()) for p in self.game.players]
min_oppo_crads = min(
left_crads[1],
left_crads[2]) if self.player_id == 0 else left_crads[0]
round_penalty = 15 - 12 * min_oppo_crads / 20
# 寻找最优出牌
best_move = None
best_comb = None
max_value = -np.inf
for i in range(len(combs)):
# 手牌总分
total_value = sum([cards_value[x] for x in combs[i]])
small_num = 0
for j in range(0, len(combs[i])):
if j > 0 and action_space[j][0] not in ["2", "R", "B"]:
small_num += 1
total_value -= small_num * round_penalty
for j in range(0, len(combs[i])):
# Pass 得分
if combs[i][j] == 0 and min_oppo_crads > 4:
if total_value > max_value:
max_value = total_value
best_comb = combs[i]
best_move = 0
# 出牌得分
elif combs[i][j] > 0 and (fine_mask is None
or fine_mask[i, j] == True):
# 特判只有一手
if len(combs[i]) == 1 or len(
combs[i]) == 2 and combs[i][0] == 0:
max_value = np.inf
best_comb = combs[i]
best_move = combs[i][-1]
move_value = total_value - cards_value[
combs[i][j]] + round_penalty
if move_value > max_value:
max_value = move_value
best_comb = combs[i]
best_move = combs[i][j]
if best_move is None:
best_comb = [0]
best_move = 0
# 最优出牌
best_cards = action_space[best_move]
move = [best_cards.count(x) for x in card_list]
# 输出选择的牌组
# print("\nbest comb: ")
# for m in best_comb:
# print(action_space[m], cards_value[m])
# 输出 player i [手牌] // [出牌]
print("Player {}".format(self.player_id),
' ',
Card.visual_card(hand_card),
end=' // ')
print(Card.visual_card(move), "From RuleBasedModel")
return move, None
# start = time.time()
# 定位current_node
cards_out = self.game.cards_out
length = len(cards_out)
# 判断是否定位到current_node的flag
flag = 0
if self.new_game is False:
# 前两步对手选择的move
out1 = self.list_to_card(cards_out[length - 2][1])
out2 = self.list_to_card(cards_out[length - 1][1])
for child in self.current_node.get_children():
if self.compare(child.state.action, out1):
self.current_node = child
flag = 1
break
if flag == 1:
for child in self.current_node.get_children():
if self.compare(child.state.action, out2):
self.current_node = child
flag = 2
break
my_id = self.player_id
if flag != 2:
self.new_game = False
root = Node(None, None)
self.current_node = root
# 下家id
next_id = (my_id + 1) % 3
# 下下家id
next_next_id = (my_id + 2) % 3
my_card = self.card_list_to_dict(self.get_hand_card())
# 下家牌
next_card = self.card_list_to_dict(
self.game.players[next_id].get_hand_card())
# 下下家牌
next_next_card = self.card_list_to_dict(
self.game.players[next_next_id].get_hand_card())
last_move = self.trans_card(Card.visual_card(self.game.last_move))
last_p = self.game.last_pid
moves_num = len(get_moves(my_card, last_move))
state = State(my_id, my_card, next_card, next_next_card, last_move,
-1, moves_num, None, last_p)
self.current_node.set_state(state)
# 搜索
computation_budget = 2000
for i in range(computation_budget):
expand_node = tree_policy(self.current_node, my_id)
reward = default_policy(expand_node, my_id)
backup(expand_node, reward)
best_next_node = get_bestchild(self.current_node, my_id)
move = best_next_node.get_state().action
self.current_node = best_next_node
new_move = self.card_to_list(move)
hand_card = []
for i, n in enumerate(Card.all_card_name):
hand_card.extend([n] * self.get_hand_card()[i])
print("Player {}".format(self.player_id), ' ', hand_card, end=' // ')
print(Card.visual_card(new_move), "From MctsModel")
# end = time.time()
# dur = end - start
# print('cost: {}'.format(dur))
return new_move, None
def choose(self, state):
start = time.time()
# 定位current_node
cards_out = self.game.cards_out
length = len(cards_out)
# 判断是否定位到current_node的flag
flag = 0
if length > 2:
# 前两步对手选择的move
out1 = self.list_to_card(cards_out[length - 2][1])
out2 = self.list_to_card(cards_out[length - 1][1])
for child in self.current_node.get_children():
if self.compare(child.state.action, out1):
self.current_node = child
flag = 1
break
if flag == 1:
for child in self.current_node.get_children():
if self.compare(child.state.action, out2):
self.current_node = child
flag = 2
break
my_id = self.player_id
if flag != 2:
root = Node(None, None)
self.current_node = root
# 下家id
next_id = (my_id + 1) % 3
# 下下家id
next_next_id = (my_id + 2) % 3
my_card = self.card_list_to_dict(self.get_hand_card())
# 下家牌
next_card = self.card_list_to_dict(
self.game.players[next_id].get_hand_card())
# 下下家牌
next_next_card = self.card_list_to_dict(
self.game.players[next_next_id].get_hand_card())
last_move = self.trans_card(Card.visual_card(self.game.last_move))
last_p = self.game.last_pid
moves_num = len(get_moves(my_card, last_move))
state_ = State(my_id, my_card, next_card, next_next_card,
last_move, -1, moves_num, None, last_p)
self.current_node.set_state(state_)
# 搜索
computation_budget = 2000
for i in range(computation_budget):
expand_node = tree_policy(self.current_node, my_id)
reward = default_policy(expand_node, my_id)
backup(expand_node, reward)
best_next_node = get_bestchild_(self.current_node)
move = best_next_node.get_state().action
self.current_node = best_next_node
new_move = self.card_to_list(move)
hand_card = []
for i, n in enumerate(Card.all_card_name):
hand_card.extend([n] * self.get_hand_card()[i])
print("Player {}".format(self.player_id), ' ', hand_card, end=' // ')
print(Card.visual_card(new_move))
end = time.time()
dur = end - start
# print('cost: {}'.format(dur))
return new_move, None
