Difference between -visits and -playouts

[tomecki]

I'm trying to get a better understanding of how LC0 works but haven't been able to find the answer to this anywhere.  From what I understand -playouts is the number of nodes that get expanded in the Monte Carlo search, and this number is what determines LC0s strength of "Play against LC0" site.  I see that -visits is used on the self-play client and seems to do something similar.  Can anyone please explain the difference?  Is -visits only used for generating self-play games? Thanks.

[jkiliani]

The difference between playouts and visits is tree-reuse. Any node that is part of the reused tree after a move is counted as a visit, but not as a playout. So the visit count is always at least equal to the playout count, and in case of a large proportion of tree-reuse, much larger.

[Gian-Carlo Pascutto]

The tree gets reused from one move to the next.

Imagine you run 1000 playouts every move, and start on the root
position. You put 500 into 1. e4, and then effectively 1. e4 gets played.
On the next move, you do 1000 playouts again.

The position after 1. e4 now has had 1000 playouts, but will have 1500
visits, because 500 have been carried over from the previous move.

playouts = equal effort each move
visits = equal data each move

Restricting these (instead of just letting the engine use the available
time to do as much as possible) always means restricting the engine in
some way. There is some evidence equal visits each move is a more
efficient use of resources than equal playouts. It avoids spending a lot
of effort when the moves are forced anyway.

Restricting via visits or playouts is useful for self-play because the
speed of people's machines differs. In a real game, you don't want
restrictions but just do as much as possible as time allows.

--
GCP