search

rll / rllab

rllab is a framework for developing and evaluating reinforcement learning algorithms, fully compatible with OpenAI Gym.
Other
2.9k stars 801 forks source link

Best way to record average return vs iterations? #187

Open yx222 opened 6 years ago

yx222 commented 6 years ago

Hi,

I'm struggling to find the best way to record the progress of policy optimisation using existing functionality in RLLAB. I could use the function to save iteration snapshot, but that's a bit too 'heavy' since I would only like to record a couple of statistics such as average discounted return, etc at every iteration and see how they evolve (i.e. is the agent really making good progress in learning?)

Is there an existing functionality to support that lightweight logging of these statistics or is my best bet to make some changes in the 'algoirhtm' class to do some bespoke logging?

Any suggestion would be much appreciated!

cheers,

Yunlong

nosyndicate commented 6 years ago

I guess you need to change the algo class, basically find out the logger and corresponding quantities, and remove the ones you do not want to have, then it will not show in the iteration snapshot.