Page 17 of 18
Re: Komodo 12 and MCTS
Posted: Sat Jun 02, 2018 8:58 pm
by shrapnel
glennsamuel32 wrote: ↑Fri Jun 01, 2018 5:54 pm
Perhaps Daniel's project below can aid in implementing NN ?
quote]
What NN ? Marsell and Leo will say they have never seen NN or AlphaZero, so it does not exist !
Re: Komodo 12 and MCTS
Posted: Sat Jun 02, 2018 9:00 pm
by shrapnel
glennsamuel32 wrote: ↑Fri Jun 01, 2018 5:54 pm
Perhaps Daniel's project below can aid in implementing NN ?
What NN ? Marsell and Leo will say they have never seen NN or AlphaZero, so it does not exist !
Re: Komodo 12 and MCTS
Posted: Sun Jun 03, 2018 4:18 am
by mjlef
schack wrote: ↑Thu May 31, 2018 4:48 pm
Q for Larry and Mark: are hash table requirements less if using MCTS?
This is not easy to answer, since in an MP system, every processor will be making lots of short searches. For now I suggest leaving the Hash size whatever size you would normally use at the time control (the setHash.tx file included with the download explains how to pick a good hash size with a few searches on your specific machine and setup). The MCTS search uses additional memory (currently 320 megs but this will be user settable when we release an update). The default is big enough for a single Thread to search for hours, but power users want to be able to analyze with tons of cores for a very long time).
Mark
Re: Exploiting versus exploring
Posted: Sun Jun 03, 2018 4:32 am
by mjlef
main line wrote: ↑Thu May 31, 2018 11:06 am
lkaufman wrote: ↑Thu May 17, 2018 5:39 am
Very long infinite analysis won't even work with MCTS on right now, it will just stop analyzing when reaching memory limits set by the program (not by user). We don't have any basis for recommending a value other than the default for longer time contnrols or analysis; the MCTS mode is just too new. Don't bother setting Skill for anything other than default 20 with MCTS on right now, it will only have small and somewhat unpredictable effects. Most of the UCI options just don't work yet with MCTS on, it's just too new. Syzygy also doesn't work yet with MCTS on. Hash size, table memory, Ponder, and Threads (up to 3) should work ok with MCTS. We'll gradually make more UCI options work properly with MCTS over time. Threads beyond 3 is top priority.
Look at this
Monte Carlo can't find Qc4!! played by AlphaZero. And Monte Carlo analysis doesn't work
I just checked the latest Komodo development version and it find Qd4 almost instantly. We have improved it a lot. It will be a free update for Komodo 12 owners.
Mark
Re: Exploiting versus exploring
Posted: Sun Jun 03, 2018 3:41 pm
by Master Om
mjlef wrote: ↑Sun Jun 03, 2018 4:32 am
main line wrote: ↑Thu May 31, 2018 11:06 am
lkaufman wrote: ↑Thu May 17, 2018 5:39 am
Very long infinite analysis won't even work with MCTS on right now, it will just stop analyzing when reaching memory limits set by the program (not by user). We don't have any basis for recommending a value other than the default for longer time contnrols or analysis; the MCTS mode is just too new. Don't bother setting Skill for anything other than default 20 with MCTS on right now, it will only have small and somewhat unpredictable effects. Most of the UCI options just don't work yet with MCTS on, it's just too new. Syzygy also doesn't work yet with MCTS on. Hash size, table memory, Ponder, and Threads (up to 3) should work ok with MCTS. We'll gradually make more UCI options work properly with MCTS over time. Threads beyond 3 is top priority.
Look at this
Monte Carlo can't find Qc4!! played by AlphaZero. And Monte Carlo analysis doesn't work
I just checked the latest Komodo development version and it find Qd4 almost instantly. We have improved it a lot. It will be a free update for Komodo 12 owners.
Mark
Can u please explain me how to use this MCTS in Fritz 16 ?
The dedicated monte carlo doesnot work which works with Rybka/Fritz 16 which plays randommly and builds a tree.
can u guide me please how to use this MCTS ? No of cores ? and all ?
Re: Exploiting versus exploring
Posted: Sun Jun 03, 2018 4:15 pm
by zenpawn
mjlef wrote: ↑Sun Jun 03, 2018 4:32 am
I just checked the latest Komodo development version and it find Qd4 almost instantly. We have improved it a lot. It will be a free update for Komodo 12 owners.
Mark
Do you mean 23.Qc4 ? That's the move AlphaZero played vs Stockfish 8.
Re: Exploiting versus exploring
Posted: Sun Jun 03, 2018 4:17 pm
by mjlef
shrapnel wrote: ↑Fri Jun 01, 2018 6:56 am
main line wrote: ↑Thu May 31, 2018 11:06 am
Monte Carlo can't find Qc4!! played by AlphaZero. And Monte Carlo analysis doesn't work
Naturally it won't. It's not going to be found by an Engine like Komodo which is standing on one leg, literally !
No NN, no brilliant moves.
It spots on my system.
Really ? after how many hours of Analysis and how many tweaks to Settings ?
If one is prepared to spend hours tweaking Komodo/Houdini one can find most of the moves made by the Powerful AlphaZero ; BUT not consistently and it takes time . Remember that AlphaZero made all those brilliant moves consistently UNDER ONE MINUTE.
Comparing Komodo to AlphaZero is like comparing the brightness of the Moon to the Sun !
We plan on releasing Komodo 12.1 shortly as a free upgrade for Komodo 12 owners. It finds the Alpha Zero move from the position posted within 3 seconds (it was probably faster, but I just used the PV display that updates every 3 seconds and I was too lazy to try a 1 second search). I ran it for two minutes on a 2 core laptop just to make sure it did not change its mind. I have not tried a longer run using Komodo 12 to see if it also finds the moves, but another poster here said it does.
We have never tuned Komodo to try and match any moves by Alpha Zero. I actually never even tried this position until someone posted it here. Please stop making things up. I encourage people be be skeptical when new things come along, but your statements are easily proven false. Instead of speculating, how about trying it? Or just ask me or Larry?
Mark
Re: Exploiting versus exploring
Posted: Sun Jun 03, 2018 4:24 pm
by mjlef
Master Om wrote: ↑Sun Jun 03, 2018 3:41 pm
mjlef wrote: ↑Sun Jun 03, 2018 4:32 am
main line wrote: ↑Thu May 31, 2018 11:06 am
Monte Carlo can't find Qc4!! played by AlphaZero. And Monte Carlo analysis doesn't work
I just checked the latest Komodo development version and it find Qd4 almost instantly. We have improved it a lot. It will be a free update for Komodo 12 owners.
Mark
Can u please explain me how to use this MCTS in Fritz 16 ?
The dedicated monte carlo doesnot work which works with Rybka/Fritz 16 which plays randommly and builds a tree.
can u guide me please how to use this MCTS ? No of cores ? and all ?
'
Sure. It is a bit confusing. There was an older MCTS option in the past explained below.
On the "Analysis" menu of Fritz 16 is a "Monte Carlo" icon. Do not use that. This was for a version of Fritz/Rybka that did a form of MCTS. Instead, click on the engine name, "Advanced" then "Engine Parameters" and check the "Use MCTS" box. The OK your way out. Komodo will then use MCTS during its search or analysis. You might find it convenience to save the change as another engine. So one normal Komodo 12 and an MCTS Komodo 12. Chessbase told me they did that in the Fritz GUI they use with the Komodo 12 package they sell.
For number of cores/thread, if you are using Komodo 12, you can set it to 3, since that version will only use up to 3 threads. For the Komodo 12.1 free update, you can use up to 12 (if your machine has that many), since we improved the MP scaling in MCTS mode. Actually, just set cpus/core/threads to whatever you machine has, since Komodo will automatically limit the cores based on the version installed.
Mark
Re: Komodo 12 and MCTS
Posted: Sun Jun 03, 2018 6:45 pm
by Hurnavich
Hi, ok so box checked and running MCTS i'm thick here does the display output show me something different or just shows the best move found like normal K
Thanks
Re: Exploiting versus exploring
Posted: Sun Jun 03, 2018 7:27 pm
by Master Om
mjlef wrote: ↑Sun Jun 03, 2018 4:24 pm
Master Om wrote: ↑Sun Jun 03, 2018 3:41 pm
mjlef wrote: ↑Sun Jun 03, 2018 4:32 am
I just checked the latest Komodo development version and it find Qd4 almost instantly. We have improved it a lot. It will be a free update for Komodo 12 owners.
Mark
Can u please explain me how to use this MCTS in Fritz 16 ?
The dedicated monte carlo doesnot work which works with Rybka/Fritz 16 which plays randommly and builds a tree.
can u guide me please how to use this MCTS ? No of cores ? and all ?
'
Sure. It is a bit confusing. There was an older MCTS option in the past explained below.
On the "Analysis" menu of Fritz 16 is a "Monte Carlo" icon. Do not use that. This was for a version of Fritz/Rybka that did a form of MCTS. Instead, click on the engine name, "Advanced" then "Engine Parameters" and check the "Use MCTS" box. The OK your way out. Komodo will then use MCTS during its search or analysis. You might find it convenience to save the change as another engine. So one normal Komodo 12 and an MCTS Komodo 12. Chessbase told me they did that in the Fritz GUI they use with the Komodo 12 package they sell.
For number of cores/thread, if you are using Komodo 12, you can set it to 3, since that version will only use up to 3 threads. For the Komodo 12.1 free update, you can use up to 12 (if your machine has that many), since we improved the MP scaling in MCTS mode. Actually, just set cpus/core/threads to whatever you machine has, since Komodo will automatically limit the cores based on the version installed.
Mark
I was also thinking same on that monte carlo button. Thanks for the detailed reply. AM finding it interesting.
Last time I used komodo was when late Mr Don Dailey offered me komodo (was CCT i think) as i was a tester for him. He gave me free.
Mr Houdart was also too modest to offer me Houdini 6 free when he came to know its bit difficult to buy for me.
Hope If I could get Komodo 12.1 coz am little bit interested in MCTS only.
Dont know if chessbase will update.
Regards
Om