Convergence criteria for arm3d? about rllab-curriculum

[florensacc]

Yes, it could take over a couple of days in a machine like yours to reach the reported performance. What do you mean exactly by "reruns"?

[therealchuckliu]

Ok, that's good to know. Just eyeballing the log output it seems the same couple tests are being run repeatedly and labeled as either itr #0 or #1

[florensacc]

You might be referring to the inner training: as described in the paper, every iteration executes "n" TRPO steps. Therefore, if your n=2 you should be seeing many itr #0, itr #1 and then repeat. In between them, you should see a log of the outer iteration progress.

[therealchuckliu]

Oh I see, looked at the code more carefully and makes sense now. Will totally take over a day to finish running, I think I'll set up an EC2 instance instead XD. Thanks!