def get_action(self, obs_dict, step_cnt):
"""
get actions
:param detector_obs_list:
:param fighter_obs_list:
:param joint_obs_dict:
:param step_cnt:
:return:
"""
detector_action = []
fighter_action = []
for y in range(self.fighter_num):
tmp_course = obs_dict['fighter'][y]['course'] # (1, )
tmp_pos = obs_dict['fighter'][y]['pos'] # (2, )
tmp_r_visible_pos = obs_dict['fighter'][y][
'r_visible_pos'] # (10, 2)
tmp_l_missile = obs_dict['fighter'][y]['l_missile'] # rule use
tmp_s_missile = obs_dict['fighter'][y]['s_missile'] # rule use
tmp_j_visible_fp = obs_dict['fighter'][y][
'j_visible_fp'] # rule use
tmp_j_visible_dir = obs_dict['fighter'][y][
'j_visible_dir'] # (10, 1)
tmp_g_visible_pos = obs_dict['fighter'][y][
'g_visible_pos'] # (10, 2) # rule use
tmp_striking_id = obs_dict['fighter'][y]['striking_id'] # (10, 1)
# model obs change, 归一化
course = tmp_course / 359.
pos = tmp_pos / self.size_x
r_visible_pos = tmp_r_visible_pos.reshape(
1, -1)[0] / self.size_x # (20,)
j_visible_dir = tmp_j_visible_dir.reshape(1, -1)[0] / 359 # (10,)
striking_id = tmp_striking_id.reshape(1, -1)[0] / 1
obs = np.concatenate(
(course, pos, r_visible_pos, j_visible_dir, striking_id),
axis=0)
logger.debug('obs: {}'.format(obs))
true_action = np.array([0, 1, 0, 0], dtype=np.int32)
if obs_dict['fighter'][y]['alive']:
# rule policy
true_action = fighter_rule(tmp_course, tmp_pos, tmp_l_missile,
tmp_s_missile, tmp_r_visible_pos,
tmp_j_visible_dir, tmp_j_visible_fp,
tmp_g_visible_pos, step_cnt)
logger.debug('true action rule out: {}'.format(true_action))
# model policy
if not any([any(r_visible_pos >= 0), any(j_visible_dir >= 0)]):
tmp_action = self.maddpg.select_action(y, obs)
logger.debug('tmp action: {}'.format(tmp_action))
# 添加动作, 将动作转换为偏角
tmp_action_i = np.argmax(tmp_action)
logger.info('tmp action i: {}'.format(tmp_action_i))
true_action[0] = action2direction(true_action[0],
tmp_action_i, ACTION_NUM)
logger.info('true action: {}'.format(true_action))
fighter_action.append(true_action)
fighter_action = np.array(fighter_action)
return detector_action, fighter_action
def get_action(self, obs_dict, step_cnt):
"""
get actions
:param detector_obs_list:
:param fighter_obs_list:
:param joint_obs_dict:
:param step_cnt:
:return:
"""
detector_action = []
fighter_action = []
for y in range(self.fighter_num):
tmp_course = obs_dict['fighter'][y]['course'] # (1, )
tmp_pos = obs_dict['fighter'][y]['pos'] # (2, )
tmp_r_visible_pos = obs_dict['fighter'][y][
'r_visible_pos'] # (10, 2)
tmp_j_visible_fp = obs_dict['fighter'][y][
'j_visible_fp'] # (10, 1)
tmp_l_missile = obs_dict['fighter'][y]['l_missile'] # rule use
tmp_s_missile = obs_dict['fighter'][y]['s_missile'] # rule use
tmp_j_visible_fp = obs_dict['fighter'][y][
'j_visible_fp'] # rule use
tmp_j_visible_dir = obs_dict['fighter'][y][
'j_visible_dir'] # (10, 1) # rule use
tmp_g_visible_pos = obs_dict['fighter'][y][
'g_visible_pos'] # (10, 2)
# model obs change, 归一化
course = tmp_course / 359.
pos = tmp_pos / self.size_x
r_visible_pos = tmp_r_visible_pos.reshape(
1, -1)[0] / self.size_x # (20,)
j_visible_fp = tmp_j_visible_fp.reshape(1, -1)[0] / 359. # (10,)
g_visible_pos = tmp_g_visible_pos.reshape(
1, -1)[0] / self.size_x # (20,)
obs = np.concatenate(
(course, pos, r_visible_pos, j_visible_fp, g_visible_pos),
axis=0)
logger.debug('obs: {}'.format(obs))
true_action = np.array([0, 1, 0, 0], dtype=np.int32)
if obs_dict['fighter'][y]['alive']:
# model policy
tmp_action_i = self.maddpg.select_action(y, obs)
logger.debug('tmp action i: {}'.format(tmp_action_i))
# rule policy
true_action = fighter_rule(tmp_course, tmp_pos, tmp_l_missile,
tmp_s_missile, tmp_r_visible_pos,
tmp_j_visible_dir, tmp_j_visible_fp,
tmp_g_visible_pos, step_cnt)
logger.debug('true aciton rule out: {}'.format(true_action))
# 添加动作 todo
true_action[2] = np.argmax(
tmp_action_i) if IS_DISPERSED else range_transfer(
tmp_action_i, 11)
if true_action[2] == 11:
logger.info('agent {}: right'.format(y + 1))
logger.wait()
logger.info('true action: {}'.format(true_action))
fighter_action.append(true_action)
fighter_action = np.array(fighter_action)
return detector_action, fighter_action
# temp
true_action = np.array([0, 1, 0, 0], dtype=np.int32)
if not red_obs_dict['fighter'][y]['alive']:
# 如果有智能体已经死亡，则默认死亡动作输出
action_list.append(
np.array([-1 for _ in range(ACTION_NUM)],
dtype=np.float32))
else:
# model policy
tmp_action = maddpg.select_action(y, obs)
logger.debug('tmp action: {}'.format(tmp_action))
# rule policy
true_action = fighter_rule(tmp_course, tmp_pos,
tmp_l_missile, tmp_s_missile,
tmp_r_visible_pos,
tmp_j_visible_dir,
tmp_j_visible_fp,
tmp_g_visible_pos, step_cnt)
logger.debug(
'true aciton rule out: {}'.format(true_action))
# 添加动作
true_action[2] = np.argmax(
tmp_action) if IS_DISPERSED else range_transfer(
tmp_action, ACTION_NUM)
logger.debug('true action: {}'.format(true_action))
action_list.append(tmp_action)
red_fighter_action.append(true_action)
# env step
logger.info('agent0 true action: {}'.format(