It's amazing how many comments suggest training a neutral net for prediction. Additional fitting through elo ratings I believe are suggested, as is dynamic tuni

It's amazing how many comments suggest training a neutral net for prediction. Additional fitting through elo ratings I believe are suggested, as is dynamic tuning. To me, having played games with rollback, this takes away one of the key benefits of simplistic rollback: predictability and consistency. Sure it's not the same as offline, but your brain gets pretty good at understanding when it happens and even at predicting it. If an opponent is stationary a few more frames than feels right, or keeps moving, you know to correct for an actions they've already sent - and it's a good estimation most of the time. Over fitting with neural nets IMO removes this consistency without providing much benefit, plus if you've got a strong neural net you might as well train locally against that first.