Quote:
Originally Posted by Regret$
IIRC from the interviews with the devs for openai, they actually do simulations where they present/test two midgame options and then simulate millions of times. Presumably this means they have stopped many different games where the choice was presented, ranged or melee, and ranged won over millions of simulations. Unsure if that is exactly true but that was the inference from the interview.
close, they use random forest - so its not a sim saying to try both. it just tries random things, records the results over thousands/millions iterations, and comes to the GTO solution