When do we get to pick the submission for the final evaluation on all environments for 2 tracks?
Hey @jyotish, does the submission we pick have to be something we submitted previously, or it can be a new submission after round 2 ends? It doesn’t really make sense to pick something we previously submitted because those are for sample efficiency only.
By default the number of ml.p3.2xlarge instances one can use for training is 0. And one wants to use more to train on 16 env parallelly, one needs to contact supports for an increase in limit. I contacted support and they said that it will take a while. Since this problem will apply to everyone, is there any way the organizer can make the process of increasing limit faster?
Thank you for the information
Agree, came across many sample efficient algorithms that require more than 2 hours to run 8M steps
Currently, I see that round 2 has only 6 procgen games instead of 16. Is that temporary or will that be until the end of round 2?
Thank you very much!
Is it possible for me to collaborate with another competitor and form a team in round 2?
I’m a little bit confused. My mean reward is 8.134, yet my mean normalized reward is 0. I have 3 questions:
- What are R_min and R_max used to calculate the normalized score for coinrun?
- How many episodes are used to calculate the evaluation scores?
- How are these evaluation episodes (levels) different from the episodes given in training?
What’s the number of levels allowed for training in round 1? Are the evaluation levels same as training level for round 1? If they are not the same, how are they different?
What are the evaluation environments for round 1?