Open Beckschen opened 6 years ago
IPG is embedded in npo.py
(TRPO/Q-prop)
You need to specify the parameters correctly to enable it in launcher_utils.py
, or make your own script to run it.
@SystemSigma Hello, thanks a lot. Indeed IPG is embedded in npo.py just as TRPO and PPO. How to specifiy parameters to enable it in launcher_utils.py? There are many kinds of key.
I think qprop_nu
is the parameter you are looking for, since IPG seems to be implemented over qprop.
@SystemSigma __ Thanks for your quick help. Actually I still don't know what is the process to set IPG:( My purpose is set a simple environment running with IPG or Qprob. But still no idea how to do. And another question. Have you ever run the experiment of qprob? how. And what is the .sh files under ~/launcher/expriments used for? Should I execute the hyperparameter manually? Thanks. Merry christmas:)
algo_gym_stub
is the launch file to the stubbed call to the algos train methods, i use this to run my environments.
launcher_utils
provides many parameters, including the algo name, which for IPG should be actrpo
, but im not sure.. If you study a bit the code it will be more clear.
Qprop runs fine for me in this way.
I can find the algorithms like trpo and ddpg in the file of algos. And the Readme tells that rllab++ has finished the implementation of IPG(intepolated policy gradient). But I didn't find where it is.