AlphaStar is first trained in a supervised way using the publicly available replays from SCII matches. Even if you just consider self-play, I don't see how you

AlphaStar is first trained in a supervised way using the publicly available replays from SCII matches. Even if you just consider self-play, I don't see how you would create a simulation that is realistic enough so that you could generalize afterwards to reality. RL today has an extremely hard time dealing with different distributions. You can see the amount of work OpenAI [1] had to do to go from a virtual hand+rubik's cube to real ones, even though you can make a very accurate simulation much more easily. [1] https://openai.com/blog/solving-rubiks-cube/ https://openai.com/blog/solving-rubiks-cube/