Komodo MCTS Progress.

[Gary Internet]

How strong is KMCTS 2273 which is currently playing in "Battle of the Leelas" on Chess.com?

I remember Mark saying in the TCEC chat a few weeks ago that 2246 that played in the TCEC Cup was roughly 25 Elo stronger than 2227 that played in the TCEC Premier Division in Season 14.

How does 2273 compare to either of those versions?

[Paloma]

only mark and larry will know this.

[lkaufman]

How strong is KMCTS 2273 which is currently playing in "Battle of the Leelas" on Chess.com?

I remember Mark saying in the TCEC chat a few weeks ago that 2246 that played in the TCEC Cup was roughly 25 Elo stronger than 2227 that played in the TCEC Premier Division in Season 14.

How does 2273 compare to either of those versions?

25 elo was probably an overestimate for 2246, but for 2273 it is a reasonable figure on one or four threads. However we are now able to use the available 46 cores for 46 threads (previously we only used 32 threads) so that might be helping a bit in the chess.com event.

[Werewolf]

How strong is KMCTS 2273 which is currently playing in "Battle of the Leelas" on Chess.com?

I remember Mark saying in the TCEC chat a few weeks ago that 2246 that played in the TCEC Cup was roughly 25 Elo stronger than 2227 that played in the TCEC Premier Division in Season 14.

How does 2273 compare to either of those versions?

25 elo was probably an overestimate for 2246, but for 2273 it is a reasonable figure on one or four threads. However we are now able to use the available 46 cores for 46 threads (previously we only used 32 threads) so that might be helping a bit in the chess.com event.

It seems to me Komodo MCTS has a unique place in the world. There are now four possible "types" of engine:

Alpha Beta + hand-tuned Evaluation Function (99% of all engines)
Alpha Beta + NN (Giraffe only?)
MCTS + NN (Alpha Zero + Lc0)
MCTS + hand-tuned Evaluation Function (Komodo MCTS only?)

I hope you don't decide to go NN - this experiment with Komodo is very interesting.

I forget how your MCTS differs from Lc0, but I think you said there is a difference.

[CMCanavessi]

You forgot about Scorpio

[lkaufman]

How strong is KMCTS 2273 which is currently playing in "Battle of the Leelas" on Chess.com?

I remember Mark saying in the TCEC chat a few weeks ago that 2246 that played in the TCEC Cup was roughly 25 Elo stronger than 2227 that played in the TCEC Premier Division in Season 14.

How does 2273 compare to either of those versions?

25 elo was probably an overestimate for 2246, but for 2273 it is a reasonable figure on one or four threads. However we are now able to use the available 46 cores for 46 threads (previously we only used 32 threads) so that might be helping a bit in the chess.com event.

It seems to me Komodo MCTS has a unique place in the world. There are now four possible "types" of engine:

Alpha Beta + hand-tuned Evaluation Function (99% of all engines)
Alpha Beta + NN (Giraffe only?)
MCTS + NN (Alpha Zero + Lc0)
MCTS + hand-tuned Evaluation Function (Komodo MCTS only?)
I hope you don't decide to go NN - this experiment with Komodo is very interesting.

I forget how your MCTS differs from Lc0, but I think you said there is a difference.

Well, the details of Komodo MCTS are very different from Lc0, but conceptually it is similar except we substitute short alpha-beta searches for the NN calls. We are definitely trying out NN, but we won't switch to it unless there is a clear benefit. Theoretically, we might end up with Komodo MCTS being best for machines without a good GPU, Komodo NN (or whatever we call it) being best for those with a good GPU, and regular Komodo as well. But maintaining three separate engines is probably not very practical. We might have to make some tough decisions. Or maybe some hybrid will win out. Infinite possibilities!

[dkappe]

Larry,

Have you tried a matchup between lc0 with 112x9 distilled net vs kmcts on 1 cpu?