TalkChess.com

Can't wait for Komodo 14 release so Komodo 12 becomes free and I can see what is all the fuss about MCTS

Larry, are you and Mark also working on a NN version?

dannyb wrote: ↑Sat Oct 19, 2019 10:18 am Larry, are you and Mark also working on a NN version?

Well, we have a version of KomodoMCTS that can use and benefit from a NN, but not yet by enough to justify releasing it. If it needs a good GPU and a third-party NN, the benefit needs to be substantial. We don't yet have a decent NN of our own.

lkaufman wrote: ↑Fri Oct 18, 2019 4:46 pm

leavenfish wrote: ↑Fri Oct 18, 2019 5:35 am

lkaufman wrote: ↑Wed Oct 16, 2019 6:27 pm a significant improvement in MCTS playing strength on six or more threads (we got a plus 20 elo result in 1300 bullet games on 8 threads)

Why?
I mean, how does a +6 on a single thread translate to +20 on 8? Why not +6 on 1 and 3 and 8 threads?
I can understand that more core = more strength, but this I do not understand.

There are some parameter values that depend on the number of threads, in both regular and MCTS mode. Sometimes what works best on 1 thread is not very good on 8, for example. So although the overall engine improvement was about six elo, by modifying the setting of one or more parameters in MCTS mode with many threads we got a significant further improvement.

I am wondering if this sort of 'exponential benefit' or further improvement you speak of in going from 1 thread to 8 or more, theoretically might not be seen when using K 13.2 for simple multi-pv analysis of positions, but seen only in play mode?

leavenfish wrote: ↑Sat Oct 19, 2019 7:01 pm

lkaufman wrote: ↑Fri Oct 18, 2019 4:46 pm

leavenfish wrote: ↑Fri Oct 18, 2019 5:35 am

lkaufman wrote: ↑Wed Oct 16, 2019 6:27 pm a significant improvement in MCTS playing strength on six or more threads (we got a plus 20 elo result in 1300 bullet games on 8 threads)

Why?
I mean, how does a +6 on a single thread translate to +20 on 8? Why not +6 on 1 and 3 and 8 threads?
I can understand that more core = more strength, but this I do not understand.

There are some parameter values that depend on the number of threads, in both regular and MCTS mode. Sometimes what works best on 1 thread is not very good on 8, for example. So although the overall engine improvement was about six elo, by modifying the setting of one or more parameters in MCTS mode with many threads we got a significant further improvement.

I am wondering if this sort of 'exponential benefit' or further improvement you speak of in going from 1 thread to 8 or more, theoretically might not be seen when using K 13.2 for simple multi-pv analysis of positions, but seen only in play mode?

In MCTS mode multi-pv simply displays available information, it has no effect on the analysis and takes (almost) no extra time. In regular mode it is very likely that nearly every parameter is set wrongly for multi-pv, since it has huge effects and we don't test that way. But since we recommend using the MCTS mode when using multi-pv (especially if set to 3 or more, for 2 it's a close call), this is moot if you are taking our advice.