How strong MCTS bots work these days

[hyperpape]

There's an interesting discussion going on on the computer go mailing list recently. Hopefully it's common knowledge that MCTS isn't really composed of "random" playouts. Both at the initial tree building phase and the playout phase, engines use go knowledge to decide what moves to use. Nonetheless, I found this thread surprising: it seems that several good engines are much smarter than I realized.

What do I mean by that? Aya, for instance, can beat GnuGo with only 350 playouts, while Oakfoam needs only 700. Or measured another way, with 1200 playouts, Aya can be 1k on KGS, with 2500, it can be 1 dan.

Many program authors haven't chimed in, but there does seem to be a consensus that the best approach relies on intelligent tree selection.

Check out the original post (http://dvandva.org/pipermail/computer-g ... 06320.html) or view the whole thread (http://dvandva.org/pipermail/computer-g ... .html#6320).

[leichtloeslich]

Aya, for instance, can beat GnuGo with only 350 playouts, while Oakfoam needs only 700. Or measured another way, with 1200 playouts, Aya can be 1k on KGS, with 2500, it can be 1 dan.

As you said yourself, MCTS bots don't use random playouts ("light playouts"), but use more or less extensive go knowledge when choosing candidate moves ("heavy playouts").

So the number of playouts isn't really that informative if we have no idea how performance heavy a playout is. For example, I read somewhere[1], that Zen uses relatively few very knowledge-heavy playouts.

Also, what does it mean that a bot is KGS 1k with 1200 playouts?
Seems to me that statement should somehow include information on time settings/hardware used.

[1] In Investigating the Limits of Monte Carlo Tree Search Methods in Computer Go from http://webdocs.cs.ualberta.ca/~mmueller/publications.html it says on page 8:

While the techniques used in ZEN have not been published, it’s [sic!] authors have publicly described ZEN as using knowledge-heavy, slow but very well-informed playouts.