Komodo 13 released

[cma6]

"We think the KRPPvsKRP and KPPPvs KPP are the most useful 7 piece Syzygy files to use first."
Andrew, may I ask from where are you getting these 7-man Syzygy TB files?
Thanks, CMA

Andrew
[/quote]

We think the KRPPvsKRP and KPPPvs KPP are the most useful 7 piece Syzygy files to use first. Note that it is quite possible to promote a pawn to a queen or other piece, and this can make the engine think it is bad if that specific table is missing, since to the engine, it looks like it has done from a win to a position where it is just vary ahead. I wanted to mention that since it can look odd when you analyze.
[/quote]

[mjlef]

Both Hash tables are used. Hash is used by the short alpha-beta searches used to estimate winning probability and the policy (how likely a move is to be played) and MCTS Hash is used to store the Monte Carlo Tree. For regualr home machines the values should be good enough unless you plan to do very long analysis, in which case increase MCTS Hash.

Hi Mark:
For long correspondence analysis with 64 GB of RAM on system (36-core dual-Xeon) where 4 GB is used for normal Komodo hash, what MCTS hash would you recommend and any change in 4 GB for Komodo overall?
Also with respect to Syzygy settings, should one stick with previously recommended Probe depth of 3 and Probe limit of 6?
[/quote]

MTCS Hash setting is going to depend on the machine. You can do a run for say a minute and note the nps shown (with Threads = 35 or so...I like to keep a thread free so the GUI and other programs do not get bogged down). Assuming say 40 nps per Thread, that gives 1400 nps for 35 of them. That is what the nps should show on the search (yours will vary) Each MCTS tree entry is about 40 bytes, and an hour has 3600 seconds, so 1400 * 40 * 3600 is about 200 megs an hour. so doubling that for situations where the node rate shoots up more (like when many positions lead to mate or 3 rep or 50 move rule), that should give you about 10 hours of analysis time before it might get full. You will have to see your actual nps to do these estimates, of course.

Syzygy is going to depend on the speed of your drive. A fast SSD can probably handle a Probe Depth of 2 and a Probe Limit of 6 or 7. If it is a regular hard drive, set Smart Syzygy = true, and Probe Limit to 6 or 7 as appropriate, then Play with Probe Depth, increasing it until it no longer is bogged down. You will want to run an endgame position on that to make the probes more likely. Note SSD and NVME drives vary in speed a lot, so you will just have to play with it a bit. I set the defaults assuming people have a regular hard drive.

Mark

[mjlef]

All Komodo are much below Stockfish right now. Around 60 elo at very least in TCEC condition.

That's what i thought.
So is komodo trying to hold down the elo of its standard engine just to give time for mcts to catch up with it?

I assume you were joking. No, we are not "holding down the elo" of anything. The MCTS work has been more interesting so we spend more time with it, but there was also a considerable increase in the alpha-beta search elo in K 13.01. We work on regualr alpha-beta when ideas pop up.

Mark

[cma6]

Hi Mark:
For long correspondence analysis with 64 GB of RAM on system (36-core dual-Xeon) where 4 GB is used for normal Komodo hash, what MCTS hash would you recommend and any change in 4 GB for Komodo overall?
Also with respect to Syzygy settings, should one stick with previously recommended Probe depth of 3 and Probe limit of 6?
[/quote]

MTCS Hash setting is going to depend on the machine. You can do a run for say a minute and note the nps shown (with Threads = 35 or so...I like to keep a thread free so the GUI and other programs do not get bogged down). Assuming say 40 nps per Thread, that gives 1400 nps for 35 of them. That is what the nps should show on the search (yours will vary) Each MCTS tree entry is about 40 bytes, and an hour has 3600 seconds, so 1400 * 40 * 3600 is about 200 megs an hour. so doubling that for situations where the node rate shoots up more (like when many positions lead to mate or 3 rep or 50 move rule), that should give you about 10 hours of analysis time before it might get full. You will have to see your actual nps to do these estimates, of course.

Using Aquarium GUI, after 1 minute in a typical endgame position with 6-man Syzygy installed, SF shows 45,680 kN/s and 260K TB access; Komodo shows 40,000 kN/s and 827K TB access; Komodo MCTS does not show any speed nor any TB access.
Do you think that this is an Aquarium issue or perhaps Komodo MCTS does not give TB access?
Since I may not be able to get any MCTS speed reading, what would your best guess for a good MCTS hash to use other than the default 128K?

[mjlef]

Using Aquarium GUI, after 1 minute in a typical endgame position with 6-man Syzygy installed, SF shows 45,680 kN/s and 260K TB access; Komodo shows 40,000 kN/s and 827K TB access; Komodo MCTS does not show any speed nor any TB access.
Do you think that this is an Aquarium issue or perhaps Komodo MCTS does not give TB access?
Since I may not be able to get any MCTS speed reading, what would your best guess for a good MCTS hash to use other than the default 128K?

K MCTS does use Syzygy (as long as you have it on and the SyzygyPath set). I would estimate 40 nps per thread, so maybe 1600 nps on your machine. I would then estimate about 1600*40*60 or about about 4 megs per minute of search, so maybe 230 megs per hour of search. I do not know why Aquarium is not showing you the nps, but maybe below some value it does not display. You could double click the Komodo executable, then type these, substituting whatever number of Thread you use for the n below:

setoption name Use MCTS value true
setoption name Threads value n
go movetime 30000

note the nps on the last line it shows after the 30 second search. This does not use Syzygy but the nps difference would be small with that anyway. Take that nps and multiply it by 2400 then divide by 1000000 to get an estimate of million Monte Carlo nodes per minute. then adjust for the total search time you want to run.

Mark

[cc2150dx]

Using Aquarium GUI, after 1 minute in a typical endgame position with 6-man Syzygy installed, SF shows 45,680 kN/s and 260K TB access; Komodo shows 40,000 kN/s and 827K TB access; Komodo MCTS does not show any speed nor any TB access.
Do you think that this is an Aquarium issue or perhaps Komodo MCTS does not give TB access?

AFAIK Aquarium will only display KN/s if the speed value is equal or greater than 1 KN/s, if not it displays 0 KN/s. Also, K MCTS only shows TB hits when finished playing its move (not during).

I hope this makes sense.

[cma6]

Using Aquarium GUI, after 1 minute in a typical endgame position with 6-man Syzygy installed, SF shows 45,680 kN/s and 260K TB access; Komodo shows 40,000 kN/s and 827K TB access; Komodo MCTS does not show any speed nor any TB access.
Do you think that this is an Aquarium issue or perhaps Komodo MCTS does not give TB access?

AFAIK Aquarium will only display KN/s if the speed value is equal or greater than 1 KN/s, if not it displays 0 KN/s. Also, K MCTS only shows TB hits when finished playing its move (not during).

Mark,
The implication of your comment is that MCTS in my Aquarium setup is < 1 KN/s. Does that not mean that there is no point in using MCTS since the speed is so slow?

[schack]

Above Larry estimated the increase for the AB version as 10 elo points. Are you saying this is "considerable?"

I'm not being snarky - I'm honestly curious.

[quote=mjlef post_id=798208 time=1557258219 user_id=771]
[quote=S.Taylor post_id=798135 time=1557213345 user_id=245]
[quote=JJJ post_id=798132 time=1557210167 user_id=7438]
All Komodo are much below Stockfish right now. Around 60 elo at very least in TCEC condition.
[/quote]
That's what i thought.
So is komodo trying to hold down the elo of its standard engine just to give time for mcts to catch up with it?
[/quote]

I assume you were joking. No, we are not "holding down the elo" of anything. The MCTS work has been more interesting so we spend more time with it, but there was also a considerable increase in the alpha-beta search elo in K 13.01. We work on regualr alpha-beta when ideas pop up.

Mark
[/quote]

[lkaufman]

Using Aquarium GUI, after 1 minute in a typical endgame position with 6-man Syzygy installed, SF shows 45,680 kN/s and 260K TB access; Komodo shows 40,000 kN/s and 827K TB access; Komodo MCTS does not show any speed nor any TB access.
Do you think that this is an Aquarium issue or perhaps Komodo MCTS does not give TB access?

AFAIK Aquarium will only display KN/s if the speed value is equal or greater than 1 KN/s, if not it displays 0 KN/s. Also, K MCTS only shows TB hits when finished playing its move (not during).

Mark,
The implication of your comment is that MCTS in my Aquarium setup is < 1 KN/s. Does that not mean that there is no point in using MCTS since the speed is so slow?

Komodo MCTS can play/analyze extremely well doing less than 1 kn/s, because each node evaluated is evaluated at a level that would probably give even Magnus a tough time at some rapid level. The NPS figure can't be compared meaningfully with any other engine.

[cma6]

Using Aquarium GUI, after 1 minute in a typical endgame position with 6-man Syzygy installed, SF shows 45,680 kN/s and 260K TB access; Komodo shows 40,000 kN/s and 827K TB access; Komodo MCTS does not show any speed nor any TB access.
Do you think that this is an Aquarium issue or perhaps Komodo MCTS does not give TB access?
Since I may not be able to get any MCTS speed reading, what would your best guess for a good MCTS hash to use other than the default 128K?

K MCTS does use Syzygy (as long as you have it on and the SyzygyPath set). I would estimate 40 nps per thread, so maybe 1600 nps on your machine. I would then estimate about 1600*40*60 or about about 4 megs per minute of search, so maybe 230 megs per hour of search. I do not know why Aquarium is not showing you the nps, but maybe below some value it does not display. You could double click the Komodo executable, then type these, substituting whatever number of Thread you use for the n below:

setoption name Use MCTS value true
setoption name Threads value n
go movetime 30000

note the nps on the last line it shows after the 30 second search. This does not use Syzygy but the nps difference would be small with that anyway. Take that nps and multiply it by 2400 then divide by 1000000 to get an estimate of million Monte Carlo nodes per minute. then adjust for the total search time you want to run.
Mark
I get 2.92 mm Monte Carlo nodes/minute. But what does that tell me about MCTS hash size, e.g., if my average search is 10 minutes, how large should my MCTS hash be?