s2, r, terminal = env.step(a)
# store experience
if SAVE_EXPERIENCE:
if np.abs(r) > 0.0:
er.add_experience(s, a, r, terminal, s2)
else:
if np.random.random() < 0.0018:
er.add_experience(s, a, r, terminal, s2)
s = s2
score += r
font = pygame.font.SysFont(None, 18)
text = font.render("Score: %.2f" % score, True, BLACK)
gameDisplay.blit(text, (DISPLAY_SHAPE[0] / 2 - 30, 60))
# Update Display
pygame.display.update()
gameover(score)
while True:
gameLoop()
if SAVE_EXPERIENCE:
er.save(EXP_REPLAY_FILE)
pygame.quit()
quit()
epsilon -= epsilon_decay
if config.total_step % config.args.online_update_freq == 0:
train_batch = replay.sample(config.args.batch_size)
loss = qnet.learn_on_minibatch(train_batch, config.args.gamma)
sys.stdout.write(
"\rTrain step at {}th step | loss {} | epsilon {}".format(
config.total_step, loss, epsilon))
sys.stdout.flush()
if config.total_step % config.args.target_update_freq == 0:
# print("Update target net")
qnet.update_target_model(config.args.tau)
config.total_step += 1
total_reward += reward
state = newstate
if done:
break
replay.add(replay_ep.buffer)
print("\nDone epoch in {} steps, {} random steps, Total reward: {}".format(
config.total_step - start_step, num_random_step, total_reward))
if (config.episode % config.args.save_model_freq == 0
and config.total_step > config.args.num_pretrain_step):
qnet.save(config.args.output_dir)
config.save()
replay.save()
print("Save model at {}".format(config.args.output_dir))
