side1_step_reward = 0
side2_step_reward = 0
for y in range(side1_detector_num):
side1_step_reward += o_detector_reward[y]
for y in range(side1_fighter_num):
side1_step_reward += o_fighter_reward[y]
for y in range(side2_detector_num):
side2_step_reward += e_detector_reward[y]
for y in range(side2_fighter_num):
side2_step_reward += e_fighter_reward[y]
side1_total_reward += side1_step_reward
side2_total_reward += side2_step_reward
# print('Round %d, Step %d:' % (round_cnt, step_cnt))
# print('Side 1 reward: %d, Side 2 reward: %d' % (side1_step_reward, side2_step_reward))
if env.get_done():
print('Round %d done at step %d!' % (round_cnt, step_cnt))
if o_game_reward > e_game_reward:
print('Side 1 WIN!!!')
side1_win_times += 1
elif o_game_reward < e_game_reward:
print('Side 2 WIN!!!')
side2_win_times += 1
else:
print('DRAW!!!')
draw_times += 1
print(
'Side 1 total step reward: %d, Side 2 total step reward: %d'
% (side1_total_reward, side2_total_reward))
print('Side 1 round reward: %d, Side 2 round reward: %d' %
(o_game_reward, e_game_reward))
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
# step and get next obs
env.step(red_detector_action, red_fighter_action, blue_detector_action, blue_fighter_action)
next_red_obs_dict, next_blue_obs_dict = env.get_obs()
next_fighter_tmp_obs = next_red_obs_dict['fighter']
next_enemy_tmp_obs = next_red_obs_dict['enemy']
next_obs_list.append(next_fighter_tmp_obs)
next_en_obs_list.append(next_enemy_tmp_obs)
# calculate reward
red_detector_reward, red_fighter_reward, red_game_reward, blue_detector_reward, blue_fighter_reward, blue_game_reward = env.get_reward()
detector_reward = red_detector_reward
fighter_reward = red_fighter_reward
reward_list.append(sum(fighter_reward))
done = env.get_done()
step_cnt += 1
print('step: ', step_cnt)
if done:
t = {} # 暂时没想好什么处理
break
elif len(obs_list) >= BATCH_SIZE:
# print('start train')
fighter_model.learn(obs_list, en_obs_list, next_obs_list, next_en_obs_list, memory_mask1_list, memory_mask2_list, action_mask_list,
reward_list, action_head0_list, action_head1_list, action_head2_list, probs_head0_list, logits_head1_list, probs_head2_list, behavior_value_list)
train_epoch += 1
time_to_save = True
# summ = fighter_model.output_summ()
# fighter_model.train_writer.add_summary(summ, step_cnt)