maddpg.memory.push(obs_list[:10], action_list[:10],
next_obs_list[:10], red_fighter_reward)
maddpg.memory.push(obs_list[10:], action_list[10:],
next_obs_list[10:], blue_fighter_reward)
# if done, perform a learn
if env.get_done():
if maddpg.episode_done > maddpg.episodes_before_train:
logger.info('done and training now begins...')
c_loss, a_loss = maddpg.update_policy()
agent0_c_loss = float(c_loss[0].data.cpu().numpy())
agent0_a_loss = float(a_loss[0].data.cpu().numpy())
# save loss
if not IS_TEST:
loss_w(
agent0_c_loss,
'train/{}/pics/agent0_c_loss.txt'.format(
PICS_PATH))
loss_w(
agent0_a_loss,
'train/{}/pics/agent0_a_loss.txt'.format(
PICS_PATH))
break
# if not done learn when learn interval
if maddpg.episode_done > maddpg.episodes_before_train and (
step_cnt % LEARN_INTERVAL == 0):
logger.info('training now begins...')
c_loss, a_loss = maddpg.update_policy()
agent0_c_loss = float(c_loss[0].data.cpu().numpy())
agent0_a_loss = float(a_loss[0].data.cpu().numpy())
# save loss
if not IS_TEST:
# model obs change, 归一化
course = tmp_course / 359.
pos = tmp_pos / size_x
r_visible_pos = tmp_r_visible_pos.reshape(1, -1)[0] / 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] / size_x # (20,)
# todo 归一化
obs = np.concatenate((course, pos, r_visible_pos, j_visible_fp, g_visible_pos), axis=0)
# store
dqn.memory.push(obs_list[y], action_list[y], copy.deepcopy(obs), fighter_reward[y])
# if done, perform a learn
if env.get_done():
# detector_model.learn()
dqn.learn()
logger.info('episode: %d, reward = %f' % (i_episode, total_reward))
logger.info('e_greedy: %f' % dqn.epsilon)
# store total_reward
reward_w(total_reward, 'train/dqn/pics/reward.txt')
break
# if not done learn when learn interval
if (step_cnt > 0) and (step_cnt % LEARN_INTERVAL == 0):
dqn.learn()
step_cnt += 1
# store loss
loss_w(dqn.cost_his, 'train/dqn/pics/loss.txt')
logger.info('**********train finish!**************')
axis=0)
next_obs_list.append(obs)
# store
maddpg.memory.push(obs_list, action_list, next_obs_list,
fighter_reward)
# if done, perform a learn
if env.get_done():
if maddpg.episode_done > maddpg.episodes_before_train:
logger.info('done and training now begins...')
c_loss, a_loss = maddpg.update_policy()
agent0_c_loss = float(c_loss[0].data.cpu().numpy())
agent0_a_loss = float(a_loss[0].data.cpu().numpy())
# save loss
loss_w(agent0_c_loss,
'train/maddpg3/pics/agent0_c_loss.txt')
loss_w(agent0_a_loss,
'train/maddpg3/pics/agent0_a_loss.txt')
break
# if not done learn when learn interval
if maddpg.episode_done > maddpg.episodes_before_train and (
step_cnt % LEARN_INTERVAL == 0):
logger.info('training now begins...')
c_loss, a_loss = maddpg.update_policy()
agent0_c_loss = float(c_loss[0].data.cpu().numpy())
agent0_a_loss = float(a_loss[0].data.cpu().numpy())
# save loss
loss_w(agent0_c_loss, 'train/maddpg3/pics/agent0_c_loss.txt')
loss_w(agent0_a_loss, 'train/maddpg3/pics/agent0_a_loss.txt')
step_cnt += 1
tmp_img_obs = tmp_img_obs.transpose(2, 0, 1)
tmp_info_obs = red_obs_dict['fighter'][y]['info']
fighter_model.memory.push(obs_list[y]['screen'],
obs_list[y]['info'],
action_list[y],
copy.deepcopy(tmp_img_obs),
copy.deepcopy(tmp_info_obs),
fighter_reward[y])
# if done, perform a learn
if env.get_done():
# detector_model.learn()
fighter_model.learn()
logger.info('episode: %d, reward = %f' %
(i_episode, total_reward))
logger.info('e_greedy: %f' % fighter_model.epsilon)
# store total_reward
reward_w(total_reward, 'train/simple/pics/reward.txt')
break
# if not done learn when learn interval
if (step_cnt > 0) and (step_cnt % LEARN_INTERVAL == 0):
# detector_model.learn()
fighter_model.learn()
step_cnt += 1
# store loss
loss_w(fighter_model.cost_his, 'train/simple/pics/loss.txt')
logger.info('**********train finish!**************')