an ordinary grad student can't reasonably train an SC2 agent to do anything on their 1 GPU workstation.
that's the motive -- an RL environment that's a little more interesting than mountaincar or Atari, and that doesn't require a $500 software license like the MuJoCo based physics sims, but that is simple enough to do something on a small machine, and can scale up to a really interesting and big strategic game-playing challenge.
Comments
Okay... but all of those differences make this game EASIER than SC2, so why work on a reduced difficulty problem now.
an ordinary grad student can't reasonably train an SC2 agent to do anything on their 1 GPU workstation.
that's the motive -- an RL environment that's a little more interesting than mountaincar or Atari, and that doesn't require a $500 software license like the MuJoCo based physics sims, but that is simple enough to do something on a small machine, and can scale up to a really interesting and big strategic game-playing challenge.