search

nexus-rl / project-x

0 stars 0 forks source link

State setters #5

Open some-rando-rl opened 2 years ago

some-rando-rl commented 2 years ago

What state setters do we want to use?

If we want to pull from replay data, how should we go about choosing what states make sense?

What proportion of states should start from kickoff?

Do we want to anneal states over time (e.g. start heavy on kick-offs so that it learns how to do them properly, then add in other setters over time)?

pizzalord22 commented 2 years ago

we probably want a lot of random states combined with states that encourage behavior that we want to see.

lydonator commented 2 years ago

I think there's a case for 95% of states that are entirely above goal height (from SSL replays), the remainder being special cases. Reason: We just don't need states on the ground because that's where everything ends up.