def get_action(self, root_id, board, turn, enemy_turn):
if turn != enemy_turn:
pi = self.player.get_pi(root_id, board, turn, tau=0.01)
self.player_pi = pi
self.player_visit = self.player.get_visit()
action, action_index = utils.argmax_onehot(pi)
else:
pi = self.enemy.get_pi(root_id, board, turn, tau=0.01)
self.enemy_pi = pi
self.enemy_visit = self.enemy.get_visit()
action, action_index = utils.argmax_onehot(pi)
return action, action_index
def get_pi(self, root_id, tau):
self._init_mcts(root_id)
self._mcts(self.root_id)
visit = np.zeros(self.board_size**2, 'float')
for action_index in self.tree[self.root_id]['child']:
child_id = self.root_id + (action_index, )
visit[action_index] = self.tree[child_id]['n']
pi = visit / visit.sum()
if tau == 0:
pi, _ = utils.argmax_onehot(pi)
return pi
def get_action(self, root_id, board, turn, enemy_turn):
if turn != enemy_turn:
if isinstance(self.player, agents.ZeroAgent):
pi = self.player.get_pi(root_id, tau=0)
else:
pi = self.player.get_pi(root_id, board, turn, tau=0)
else:
if isinstance(self.enemy, agents.ZeroAgent):
pi = self.enemy.get_pi(root_id, tau=0)
else:
pi = self.enemy.get_pi(root_id, board, turn, tau=0)
action, action_index = utils.argmax_onehot(pi)
return action, action_index
def get_action(self, root_id, board, turn, enemy_turn):
if turn != enemy_turn: # 플레이어 턴일때
if isinstance(
self.player, agents.ZeroAgent
): # isinstance() : 내장함수, 첫번째 파라미터의 객체가 두번째 파라미터의 클래스에 해당하는지 확인한다
pi = self.player.get_pi(root_id, tau=0)
else:
pi = self.player.get_pi(root_id, board, turn,
tau=0) # 플레이어는 human이므로 이 부분이 실행됨
# for monitor
self.monitor.get_pi(root_id, tau=0)
else: # 적 턴일때
if isinstance(self.enemy, agents.ZeroAgent):
pi = self.enemy.get_pi(root_id, tau=0) # 적 플레이어는 이 부분이 실행됨
else:
pi = self.enemy.get_pi(root_id, board, turn, tau=0)
# action : boardsize**2크기의 array에서 착수위치만 onehot 인코딩됨
# action_index : onehot인코딩된 위치
action, action_index = utils.argmax_onehot(pi)
return action, action_index
def get_pi(self, root_id, tau):
self._init_mcts(root_id)
self._mcts(self.root_id)
visit = np.zeros(self.board_size**2, 'float')
policy = np.zeros(self.board_size**2, 'float')
for action_index in self.tree[self.root_id]['child']:
child_id = self.root_id + (action_index, )
visit[action_index] = self.tree[child_id]['n']
policy[action_index] = self.tree[child_id]['p']
self.visit = visit
self.policy = policy
pi = visit / (
visit.sum() + 1e-8
) # normally visit.sum() is not zero because of expansion
if tau == 0:
pi, _ = utils.argmax_onehot(pi)
return pi
def get_pi(self, root_id, tau):
# MCTS
self._init_mcts(root_id)
self._mcts(self.root_id)
# 초기화
visit = np.zeros(self.board_size**2, 'float')
policy = np.zeros(self.board_size**2, 'float')
for action_index in self.tree[self.root_id]['child']:
child_id = self.root_id + (action_index, )
visit[action_index] = self.tree[child_id][
'n'] # MCTS에서 각 자식 노드의 방문 횟수
policy[action_index] = self.tree[child_id][
'p'] # 정책, 승리 가능성이 높은 위치가 높은 값을 가짐
self.visit = visit
self.policy = policy
pi = visit / visit.sum()
if tau == 0:
pi, _ = utils.argmax_onehot(pi) # 최대값을 가진 위치 중 하나만 onehot인코딩된다
return pi