def run(agent1_name,
agent2_name,
map_name,
round_num,
max_step,
random_pos=False):
"""
:param agent1_name: 红方名称
:param agent2_name: 蓝方名称
:param map_name: 地图名称
:param round_num: 对战局数
:param max_step: 单局最大step
:param random_pos: 随机起始位置
:return: agent1_win_times, agent2_win_times, draw_times, agent1_crash_times, agent2_crash_times, agent1_timeout_times, agent2_timeout_times, agent1_launch_failure_times, agent2_launch_failure_times
"""
side1_win_times = 0
side2_win_times = 0
draw_times = 0
log_flag = agent1_name + '_vs_' + agent2_name
agent1_launch_failed = False
agent2_launch_failed = False
round_cnt = 0
agent1_crash_list = []
agent2_crash_list = []
agent1_timeout_list = []
agent2_timeout_list = []
# file path constructing
map_path = 'maps/' + map_name + '.map'
agent1_path = 'agent/' + agent1_name + '/agent.py'
agent2_path = 'agent/' + agent2_name + '/agent.py'
if not os.path.exists(map_path):
print('Error: map file not exist!')
exit(-1)
if not os.path.exists(agent1_path):
print('Error: agent1 file not exist!')
exit(-1)
if not os.path.exists(agent2_path):
print('Error: agent2 file not exist!')
exit(-1)
# make env
env = Environment(map_path,
'raw',
'raw',
max_step=max_step,
render=True,
random_pos=random_pos,
log=log_flag)
# get map info
size_x, size_y = env.get_map_size()
side1_detector_num, side1_fighter_num, side2_detector_num, side2_fighter_num = env.get_unit_num(
)
# create agent
agent1 = AgentCtrl(agent1_name, size_x, size_y, side1_detector_num,
side1_fighter_num)
agent2 = AgentCtrl(agent2_name, size_x, size_y, side2_detector_num,
side2_fighter_num)
if not agent1.agent_init():
print('ERROR: Agent1 ' + agent1_name + ' init failed!')
agent1.terminate()
agent2.terminate()
agent1_launch_failed = True
if not agent2.agent_init():
print('ERROR: Agent2 ' + agent2_name + ' init failed!')
agent1.terminate()
agent2.terminate()
agent2_launch_failed = True
# 若此处一方启动失败，则认为该方全败，启动失败计round_num次，若双方启动失败，则认为双方平局round_num次，其他与前述相同。
if agent1_launch_failed and agent2_launch_failed:
return 0, 0, round_num, 0, 0, 0, 0, round_num, round_num
elif agent1_launch_failed:
return 0, round_num, 0, 0, 0, 0, 0, round_num, 0
elif agent2_launch_failed:
return round_num, 0, 0, 0, 0, 0, 0, 0, round_num
# execution
# input("Press the <ENTER> key to continue...")
for x in range(round_num):
if x != 0:
env.reset()
step_cnt = 0
round_cnt += 1
while True:
step_cnt += 1
# get obs
side1_obs_dict, side2_obs_dict = env.get_obs()
# get action
agent1_action, agent1_result = agent1.get_action(
side1_obs_dict, step_cnt)
if agent1_result == 0:
side1_detector_action = agent1_action['detector_action']
side1_fighter_action = agent1_action['fighter_action']
elif agent1_result == 1:
agent1_crash_list.append(round_cnt)
elif agent1_result == 2:
agent1_timeout_list.append(round_cnt)
agent2_action, agent2_result = agent2.get_action(
side2_obs_dict, step_cnt)
if agent2_result == 0:
side2_detector_action = agent2_action['detector_action']
side2_fighter_action = agent2_action['fighter_action']
elif agent2_result == 1:
agent2_crash_list.append(round_cnt)
elif agent2_result == 2:
agent2_timeout_list.append(round_cnt)
# execution
if agent1_result == 0 and agent2_result == 0:
env.step(side1_detector_action, side1_fighter_action,
side2_detector_action, side2_fighter_action)
elif agent1_result != 0 and agent2_result != 0:
env.set_surrender(2)
elif agent1_result != 0:
env.set_surrender(0)
else:
env.set_surrender(1)
# get done
if env.get_done():
# reward
o_detector_reward, o_fighter_reward, o_game_reward, e_detector_reward, e_fighter_reward, e_game_reward = env.get_reward(
)
if o_game_reward > e_game_reward:
side1_win_times += 1
elif o_game_reward < e_game_reward:
side2_win_times += 1
else:
draw_times += 1
break
agent1.terminate()
agent2.terminate()
return side1_win_times, side2_win_times, draw_times, len(
agent1_crash_list), len(agent2_crash_list), len(
agent1_timeout_list), len(agent2_timeout_list), 0, 0
round_cnt = 0
agent1_crash_list = []
agent2_crash_list = []
# input("Press the <ENTER> key to continue...")
for x in range(args.round):
side1_total_reward = 0
side2_total_reward = 0
if x != 0:
env.reset()
step_cnt = 0
round_cnt += 1
while True:
time.sleep(step_delay)
step_cnt += 1
# get obs
side1_obs_dict, side2_obs_dict = env.get_obs()
# get action
try:
side1_detector_action, side1_fighter_action = agent1.get_action(
side1_obs_dict, step_cnt)
# todo 更改我方动作
# for i in range(len(side1_fighter_action)):
# if args.agent1 == 'fix_rule':
# # 规则
# side1_fighter_action[i]['r_fre_point'] = i + 1
# else:
# # model
# # side1_fighter_action[i][1] = i+1
# side1_fighter_action[i][1] = random.randint(1, 10)
# # side1_fighter_action[i][2] = 11
red_detector_action = []
fighter_model = dqn.RLFighter(ACTION_NUM)
# execution
for x in range(MAX_EPOCH):
step_cnt = 0
env.reset()
episodic_reward = 0
while True:
obs_list = []
action_list = []
red_fighter_action = []
# get obs
if step_cnt == 0:
red_obs_dict, blue_obs_dict = env.get_obs()
# get action
# get blue action
blue_detector_action, blue_fighter_action = blue_agent.get_action(
blue_obs_dict, step_cnt)
# get red action
obs_got_ind = [False] * red_fighter_num
for y in range(red_fighter_num):
true_action = np.array([0, 1, 0, 0], dtype=np.int32)
if red_obs_dict['fighter'][y]['alive']:
obs_got_ind[y] = True
# tmp_img_obs = red_obs_dict['fighter'][y]['screen']
# tmp_img_obs = tmp_img_obs.transpose(2, 0, 1)
tmp_info_obs = red_obs_dict['fighter'][y]['info']
tmp_action = fighter_model.choose_action(tmp_info_obs)
max_step=MAX_STEP,
random_pos=True)
# get map info
size_x, size_y = env.get_map_size() # size_x == size_y == 1000
red_detector_num, red_fighter_num, blue_detector_num, blue_fighter_num = env.get_unit_num(
)
red_detector_action, blue_detector_action = [], [] # temp
for i_episode in range(MAX_EPOCH):
step_cnt = 0
env.reset()
total_reward = 0.0 # 每回合所有智能体的总体奖励
rr = np.zeros((FIGHTER_NUM, )) # 每回合每个智能体的奖励
# get obs
red_obs_dict, blue_obs_dict = env.get_obs() # output: raw obs结构体
obs_dict_list = [red_obs_dict, blue_obs_dict]
while True:
# obs_list = []
obs_list = [] # len == n agents
action_list = [] # # len == n agents
fighter_action = [] # # len == n agents
# get red action
for obs_dict in obs_dict_list:
for y in range(red_fighter_num):
tmp_course = obs_dict['fighter'][y]['course'] # (1, )
tmp_pos = obs_dict['fighter'][y]['pos'] # (2, )
tmp_l_missile = obs_dict['fighter'][y][
'l_missile'] # (1, )
step_cnt = 0
num_agents = 10
num_units = 20
env.reset()
obs_list, en_obs_list = [], []
behavior_value_list = []
action_head0_list, action_head1_list, action_head2_list = [], [], []
probs_head0_list, logits_head1_list, probs_head2_list = [], [], []
memory_mask1_list, memory_mask2_list, action_mask_list = [], [], []
next_obs_list, next_en_obs_list = [], []
reward_list = []
while True:
red_detector_action, red_fighter_action = [], []
red_obs_dict, blue_obs_dict = env.get_obs()
fighter_tmp_obs = np.asarray(red_obs_dict['fighter'], dtype=np.float32) # 在同质化的设定中，只有 fighter
enemy_tmp_obs = np.asarray(red_obs_dict['enemy'], dtype=np.float32)
fighter_visible_enemys_dict = red_obs_dict['fighter_visible_enemys_dict']
fighter_data_obs_list = red_obs_dict['fighter_raw']
# get obs
if step_cnt == 0:
red_obs_dict, blue_obs_dict = env.get_obs() # 此处得到 obs
# get red and blue action
blue_detector_action, blue_fighter_action = blue_agent.get_action(blue_obs_dict, step_cnt)
obs_got_ind = [False] * red_fighter_num
red_obs, red_en_obs = np.asarray(red_obs_dict['fighter']), np.asarray(red_obs_dict['enemy'])
agent_obs = np.zeros([num_agents] + list(red_obs.shape))
blue_fighter_num)
red_detector_action = []
fighter_model = dqn.RLFighter(ACTION_NUM)
# execution
for x in range(MAX_EPOCH):
step_cnt = 0
env.reset()
while True:
obs_list = []
action_list = []
red_fighter_action = []
# get obs
if step_cnt == 0:
red_obs_dict, blue_obs_dict = env.get_obs() # 此处得到 obs
# get action
# get blue action
blue_detector_action, blue_fighter_action = blue_agent.get_action(
blue_obs_dict, step_cnt)
# get red action
obs_got_ind = [False] * red_fighter_num
for y in range(red_fighter_num):
true_action = np.array([0, 1, 0, 0], dtype=np.int32)
if red_obs_dict['fighter'][y]['alive']:
obs_got_ind[y] = True
tmp_img_obs = red_obs_dict['fighter'][y]['screen']
tmp_img_obs = tmp_img_obs.transpose(2, 0, 1)
tmp_info_obs = red_obs_dict['fighter'][y]['info']
tmp_action = fighter_model.choose_action(
tmp_img_obs, tmp_info_obs)