Open Side Menu Go to the Top
Register
Are rollouts still useful Are rollouts still useful

01-11-2023 , 08:06 AM
If you are analyzing a position with a modern strong engine are rollouts still necessary?
Is a rollout every going to change the evaluation anymore?
When people had no bots, and when the early bots were not super strong; sure rollouts were required.
But have we reached the point where rollouts are no longer needed and the engine is just always correct?
There would seem to be a point where that would be true. Are we there yet? I am just wondering whether I am completely wasting my time when I rollout a position anymore.
Are rollouts still useful Quote
01-11-2023 , 12:13 PM
No, we're not there yet. XG at the ++ level is very accurate, but rollouts will still change its evaluation for a number of plays, particularly back games. There are also certain types of positions where even an XG rollout will not give the right answer (although humans can solve them easily.)
Are rollouts still useful Quote
01-11-2023 , 05:57 PM
There are certain types of positions that the bots are consistently off one way or the other. Even simple looking stuff. In holding games, XG tends to overrate the holding game side.

4-ply:


++ Eval:


Rollout:



Containment positions where you hit a late shot and are trying to contain that hit checker is another. It almost always underestimates the side doing the containing. Often by .1 or more. So you have to roll that stuff out. The rollout is probably a little suspect too, but it should certainly be more accurate than the ++ eval.

On the other hand, if the bot tells me I took a 1.2 in a blitz position, I'm not rolling that one out...

Last edited by _Z_; 01-11-2023 at 06:05 PM.
Are rollouts still useful Quote
01-11-2023 , 06:07 PM
One interesting thing I heard someone say and I think it's right or mostly right, is that if increase the level of analysis, say you go from 4-ply to a ++ eval, and the equity increases, then that's a pretty reliable indicator that the equity is going to increase even more from a rollout, as in the above example.

So I think there is a lot of room for improvement in the bots, and that's before getting into the fact that a lot of the bots' analysis is based on cubeless money play and then that is adjusted for the match situation. So that can introduce some errors at least at the margins.
Are rollouts still useful Quote

      
m