def default_policy(self, node: MCTSTreeNode, env: Env):
action = random.sample(node.legal_actions, 1)[0]
while not env.is_over():
# step forward
next_state, next_player_id = env.step(action, False)
if not env.is_over():
action = random.sample(next_state['legal_actions'], 1)[0]
# game over
reward = env.get_payoffs()[0]
return reward
def default_policy(self, node: MPMCTSTreeNode, env: Env):
if env.is_over():
return env.get_payoffs()
player_id = env.get_player_id()
#print(player_id)
state = env.get_state(player_id)
action, _ = self.drqn_agents[player_id].eval_step(state)
while not env.is_over():
#print(action)
# step forward
next_state, next_player_id = env.step(action, False)
if not env.is_over():
#action = random.sample(next_state['legal_actions'],1)[0]
action, _ = self.drqn_agents[next_player_id].eval_step(
next_state)
# game over
return env.get_payoffs()
def expand_action_on_node(self, node: MCTSTreeNode, action, env: Env):
# 进行step，获得进行这次action的legal_actions
next_state, next_player_id = env.step(action, False)
new_node = MCTSTreeNode(key=action,
action=action,
legal_actions=next_state['legal_actions'],
game_over=env.is_over(),
parent=node)
node.children[action] = new_node
return new_node
def tree_policy(self, root_node: MCTSTreeNode, env: Env):
untried_action = self.get_untried_action(root_node)
if untried_action is not None:
# 在env上执行action
node = self.expand_action_on_node(root_node, untried_action, env)
else:
# select max-UCB value node
node = self.get_max_UCB_child_node(root_node, CP_VAL)
# step
next_state, next_player_id = env.step(node.action, False)
node.legal_actions = next_state['legal_actions']
node.game_over = env.is_over()
return node