A3C: state_processor.py about reinforcement-learning HOT 6 OPEN

These are all very good points and it does seem that you can improve the performance of the agent. However, the original intent of the A3C algorithm (as well as DQN and all its predecessors) is to create one algorithm that can play a variety of games. In this case, the StateProcessor is designed for all Atari games.

Your custom feature engineering works fine for Breakout, but does it work for all Atari games without manual intervention?

from reinforcement-learning.

Thank you

Exactly, I want to use specific game knowledge to enhance the training process. If it works, then I will think how to "guess" the game knowledge implicitly.

The problem now is: Performance of DQN (and A3C) is not compatible with the papers! Kindly see
#30

from reinforcement-learning.

from reinforcement-learning.

Hi, feed_dict in Tensorflow seems very slow. Using sess.run to process each state seems not a good idea.

from reinforcement-learning.

How else do you... use Tensorflow?

from reinforcement-learning.

Maybe it does not need to use Tensorflow to process each state. Using numpy or other libraries can avoid use feed_dict frequently.

from reinforcement-learning.