In MCTS, you expect that given adequate time, the move with the highest win percentage will also have the most visits. Is this also true for ELF? If so, how long would you expect it to take for the two to converge? I have failed to see it after a few thousand playouts, but I suppose that’s not that many.

_________________
Occupy Babel!

A few thousand is tiny. If I want to do some serious analysis of a position I'll give Elf 50-100k playouts.

I don't think this is necessarily true, only if search did not find new moves in upper levels (or important followups below them) for a while, and there are no close candidates.

But generally, both visits and winrates will lag behind latest knowledge, only in different ways (visits can only accumulate slowly, winrates can only change by averaging into them).

Good point, moha. In this case, however, it's a relatively quiet opening position, with no major changes in the win percentages for each move as the simulations proceed. I'll try to post examples soon.

Uberdude: I'm aware it's a smallish simulation, but this is a "low level" property. By that, I mean I'm not asking "why isn't the engine seeing a particular move or evaluating things in the right way?" What I'm asking is "why isn't it adequately exploring the move it gives the highest percentage chance?" I would think it would not have required that many simulations. Maybe I need to work some examples with the formula from Remi's paper to get a better intuition for how fast a move will be explored.

_________________
Occupy Babel!