jcwleo / mario_rl Goto Github PK

exploration의 문제인지는 모르겠지만 특정 부분에서 더이상 넘어가지 못하는 문제 발견.
일단 PPO에서는 해당 구간을 넘어가지 못함.
다시 A2C를 이용하여 적용해 볼 예정.

Errors in mario_ppo.py. Fix please:

mario_ppo.py:174: RuntimeWarning: invalid value encountered in greater
return (p.cumsum(axis=axis) > r).argmax(axis=axis)

Traceback (most recent call last):
File "mario_ppo.py", line 413, in
rewards = np.hstack(rewards) * reward_scale
NameError: name 'reward_scale' is not defined

I tried to run mario_a2c.py, mario_ppo.py and mario_curio.py but for non of them I cannot improve the reward.
Did you use the same hyper-parameters as in the files to conduct the evaluation? (i.e. number of workers, learning rate)
Which version of the libraries did you use ?

For instance, A2C without ICM: (after 3M time-steps)

Traceback (most recent call last):
File "mario_ppo.py", line 412, in
rewards = np.hstack(rewards) * reward_scale
NameError: name 'reward_scale' is not defined

A2C로 학습이 끝나면 PPO도 적용해서 성능 비교를 할 예정