state = next_state
step += 1
if step > config.start_train_step and train_mode:
agent.epsilon = 0
# 학습 수행
loss, maxQ, r_i, loss_rl, loss_fm = agent.train_model_RND()
loss_list.append(loss)
max_Q_list.append(maxQ)
r_i_list.append(r_i)
loss_rl_list.append(loss_rl)
loss_fm_list.append(loss_fm)
# 타겟 네트워크 업데이트
if step % (config.target_update_step) == 0:
agent.update_target()
# 네트워크 모델 저장
if step % config.save_step == 0 and step != 0 and train_mode:
agent.save_model(config.load_model, train_mode)
reward_list.append(episode_rewards)
episode += 1
# 게임 진행 상황 출력 및 텐서 보드에 보상과 손실함수 값 기록
if episode % config.print_episode == 0 and episode != 0:
print(
"step: {} / episode: {} / reward: {:.2f} / loss_tot: {:.4f} / loss_rl: {:.4f} / loss_fm: {:.6f} / maxQ: {:.2f} / reward_i: {:.6f}"
.format(step, episode,
np.mean(reward_list), np.mean(loss_list),
config.lamb * np.mean(loss_rl_list),
action="store_false",
default=True,
help="Do not perform the testing phase")
parser.add_option("-V",
"--video-record",
dest="record",
action="store",
default="video.mp4",
type="string",
help="Records")
options, args = parser.parse_args()
with tf.Session() as sess:
from agent import (init_phase, bootstrap_phase, learning_phase,
testing_phase, update_target, make_video)
init_phase(sess)
if options.bootstrap:
bootstrap_phase(sess)
update_target(sess)
if options.learning:
learning_phase(sess)
if options.testing:
testing_phase(sess)
if options.record:
make_video(sess, "videos/final.mp4", 15)