Page 1 of 4

Komodo MCTS

Posted: Fri Jun 07, 2019 1:07 pm
by cma6
I ran Komodo MCTS overnight on late middlegame position with multi-PV =2 in Aquarium GUI. But, as usual, 30 ply is the deepest the search will go. Engine is given 4096 MB RAM. Is 30 ply the limit for Komodo MCTS?

Re: Komodo MCTS

Posted: Fri Jun 07, 2019 3:22 pm
by Master Om
Till now I have found its good for nothing. I am yet to find positions where Komodo is better to SF in any aspect. MCTS appears to be good on paper but isn't working in komodo atleast. It didn't help me in my CC games analysis till now. Still am using it if i could find any.

Re: Komodo MCTS

Posted: Fri Jun 07, 2019 3:43 pm
by Laskos
Master Om wrote:
Fri Jun 07, 2019 3:22 pm
Till now I have found its good for nothing. I am yet to find positions where Komodo is better to SF in any aspect. MCTS appears to be good on paper but isn't working in komodo atleast. It didn't help me in my CC games analysis till now. Still am using it if i could find any.
Did you try Leela in CC games analysis? How do you find it?

Re: Komodo MCTS

Posted: Fri Jun 07, 2019 4:47 pm
by lkaufman
cma6 wrote:
Fri Jun 07, 2019 1:07 pm
I ran Komodo MCTS overnight on late middlegame position with multi-PV =2 in Aquarium GUI. But, as usual, 30 ply is the deepest the search will go. Engine is given 4096 MB RAM. Is 30 ply the limit for Komodo MCTS?
"ply" has no real meaning in MCTS mode, we just display a number that is calculated from the number of Monte Carlo nodes, it's quite arbitrary. If ply stops increasing, it must be because you filled up the MCTS hash so nodes don't increase. Try a much larger setting for MCTS hash.

Re: Komodo MCTS

Posted: Sat Jun 08, 2019 3:40 am
by ChiefPushesWood
lkaufman wrote:
Fri Jun 07, 2019 4:47 pm
"ply" has no real meaning in MCTS mode, we just display a number that is calculated from the number of Monte Carlo nodes, it's quite arbitrary. If ply stops increasing, it must be because you filled up the MCTS hash so nodes don't increase. Try a much larger setting for MCTS hash.
Lqrry, is the MCTS hash limited to 2048?

Chief

Re: Komodo MCTS

Posted: Sat Jun 08, 2019 4:48 pm
by lkaufman
ChiefPushesWood wrote:
Sat Jun 08, 2019 3:40 am
lkaufman wrote:
Fri Jun 07, 2019 4:47 pm
"ply" has no real meaning in MCTS mode, we just display a number that is calculated from the number of Monte Carlo nodes, it's quite arbitrary. If ply stops increasing, it must be because you filled up the MCTS hash so nodes don't increase. Try a much larger setting for MCTS hash.
Lqrry, is the MCTS hash limited to 2048?

Chief
That was a bug in Komodo 13.01 that was fixed in Komodo 13.02.

Re: Komodo MCTS

Posted: Sun Jun 09, 2019 12:26 pm
by cma6
Tanks, LK, I increased MCTs hash from 16 to 2048, keeping regular hash at 4096. That allowed "MCTS ply" to increase from 30 to 35 on overnight run. How large an MCTS hash would be too large?

Re: Komodo MCTS

Posted: Sun Jun 09, 2019 12:36 pm
by mjlef
cma6 wrote:
Sun Jun 09, 2019 12:26 pm
Tanks, LK, I increased MCTs hash from 16 to 2048, keeping regular hash at 4096. That allowed "MCTS ply" to increase from 30 to 35 on overnight run. How large an MCTS hash would be too large?
It will depend on the specific machine, number of Threads, how long you wish to analyze and to some degree upon the position(s). You could just try doubling it to 4096 and see what happens, or follow the instructions in the Readme file we include with Komodo to see how to estimate memory usage. You should also update to Komodo 13.02 (just redownload it from ur site to get the latest release), which will let you set a higher value.

MCTS programs need to hold the game tree in memory, so for vary long analysis on machines with lots of cores, they need lots of memory.

Mark

Re: Komodo MCTS

Posted: Sun Jun 09, 2019 1:21 pm
by cma6
Thanks, Mark, I am experimenting now with ver. 13.02 with MCTS hash at 4096 MB. I will often analyze for hours with 35 cores.
BTW, I see that default Syzygy probe depth is now 8; whereas 3 was what we had been using. What is best for very LTC?

Re: Komodo MCTS

Posted: Sun Jun 09, 2019 2:15 pm
by mjlef
cma6 wrote:
Sun Jun 09, 2019 1:21 pm
Thanks, Mark, I am experimenting now with ver. 13.02 with MCTS hash at 4096 MB. I will often analyze for hours with 35 cores.
BTW, I see that default Syzygy probe depth is now 8; whereas 3 was what we had been using. What is best for very LTC?
I set conservative values for Syzygy Probe Depth as the defaults, assuming many people have a regular hard drive. If you have a SSD, I suggest turning off Smart Syzygy, and setting Syzygy Probe Depth to 2. That will use Syzygy the most. set Syzygy Probe Limit to whatever size Syzygy files you have.

Smart Syzygy when on does this: It always probes 5 piece positions at all depths, but limits 6 and 7 piece lookups to whatever Syzygy Probe Depth is. Since the 5 piece tables used during the search are abut 500 mgs, they will be quickly buffered into system memory. The idea is to give people with slower hard drives the ability to still benefit as much as possible. If you have a SSD, most likely Smart Syzygy should be off.

As more and more new machines come with SSD, I am likely to make the default not use Smart Syzygy.