Komodo 12.2 Announcement

[Leo]

Its great that it will perform better on 44 cores.

[mjlef]

Komodo 12.2 has been released at komodochess.com. As usual it is free for subscribers and discounted 20% for purchasers of Komodo 11 or higher. It includes both the normal and MCTS versions. This is a rather major release, for several reasons.

The strength of the MCTS version is vastly improved, up around 200 elo points at blitz levels, perhaps 150 at much longer time controls. Our predicted ratings for it on the CCRL 40/4 scale are 3320 on 1 thread, 3350 on 2, and 3380 on 4 threads. We now believe that the MCTS version is stronger than the normal one when both are using MultiPV = 3 or more, and would rank in the top five independent engines if substituted for normal Komodo.

The low thread limit for MCTS is gone, we believe it can benefit from up to 64 threads, although this has not yet been proven beyond 30 threads.
The many-years old slow startup of Komodo on big machines with big hash tables is solved! Also Komodo is now able to make better use of big hardware with more than one “NUMA Node”. These are important for major tournaments like CCC and TCEC, and most people with 12 or more cores should see some benefits.

The playing strength of normal Komodo has been improved about 10 elo on four threads at blitz time controls, somewhat less on single thread. In general MP play has been improved noticeably since last (12.1.1) release.
Several minor bugs have been fixed.

We are considering splitting normal and MCTS Komodo into two separate engines in the future, but if so subscribers will be eligible to receive both. As to whether the MCTS version will surpass the normal one and become our main (or even only) version, only time will tell. Now that the gap is down to about 100 elo, it is at least a realistic possibility.

Thank you for this new version.

However, there seems to be a problem/bug with the displayed score in the MCTS version when outputting the final info line after the search stops.
Didn't check the A/B version.

Example 1

Code: Select all

info depth 22 seldepth 29 time 58059 score cp -112 nodes 34122 nps 587 pv f1f4 c6a4 f4g4 e7f6 f5h5 f8f7 g4g6 e4g5 h5h7 g5h7 g6f7 b7f7 b3f7 c3b2 b1b2 a4c6 d4f6 h7f6 g2g3 g7g5 b2c1
info nodes 35333
info depth 22 seldepth 29 time 60001 nodes 35333 nps 588 score cp -28
bestmove f1f4 ponder c6a4

Example 2

Code: Select all

info depth 22 seldepth 16 time 58044 score cp 992 nodes 34904 nps 601 pv g2g7 f7e8 d6e7
info nodes 36421
info depth 22 seldepth 16 time 60001 nodes 36421 nps 607 score cp 45
bestmove g2g7 ponder f7e8

Please note the difference in the final score compared to the former info line.

Komodo MCTS was running in a terminal under Linux, 1024 MB MCTS Hash, MCTS Explore set to 12, 1 min per position.
It also happens after a 'go infinite' and 'stop'.

That does look odd. I will take a look. I have not seen that before. Was this from the start position?

[peter]

Komodo 12.2 has been released at komodochess.com.

Thanks for the fast development, Larry.
Komodo MCTS is becoming a more and more interesting Analysis- tool especially in MV- mode.

If you don't have to care about time to depth as for number of MV- lines, it's good to see all the interesting candidate moves with variants and evals, even if there isn't as big an advantage in search between primary and non primary lines like in A-B- search, you yet can see the relations of evals and moves along and between the variants and watch the changes of those after giving in one after another forward and backward.

There simply isn't any need for adding MV- lines or reducing their number in Forward- Backward- Analysis anymore.

[Joerg Oster]

Komodo 12.2 has been released at komodochess.com. As usual it is free for subscribers and discounted 20% for purchasers of Komodo 11 or higher. It includes both the normal and MCTS versions. This is a rather major release, for several reasons.

The strength of the MCTS version is vastly improved, up around 200 elo points at blitz levels, perhaps 150 at much longer time controls. Our predicted ratings for it on the CCRL 40/4 scale are 3320 on 1 thread, 3350 on 2, and 3380 on 4 threads. We now believe that the MCTS version is stronger than the normal one when both are using MultiPV = 3 or more, and would rank in the top five independent engines if substituted for normal Komodo.

The low thread limit for MCTS is gone, we believe it can benefit from up to 64 threads, although this has not yet been proven beyond 30 threads.
The many-years old slow startup of Komodo on big machines with big hash tables is solved! Also Komodo is now able to make better use of big hardware with more than one “NUMA Node”. These are important for major tournaments like CCC and TCEC, and most people with 12 or more cores should see some benefits.

The playing strength of normal Komodo has been improved about 10 elo on four threads at blitz time controls, somewhat less on single thread. In general MP play has been improved noticeably since last (12.1.1) release.
Several minor bugs have been fixed.

We are considering splitting normal and MCTS Komodo into two separate engines in the future, but if so subscribers will be eligible to receive both. As to whether the MCTS version will surpass the normal one and become our main (or even only) version, only time will tell. Now that the gap is down to about 100 elo, it is at least a realistic possibility.

Thank you for this new version.

However, there seems to be a problem/bug with the displayed score in the MCTS version when outputting the final info line after the search stops.
Didn't check the A/B version.

Example 1

Code: Select all

info depth 22 seldepth 29 time 58059 score cp -112 nodes 34122 nps 587 pv f1f4 c6a4 f4g4 e7f6 f5h5 f8f7 g4g6 e4g5 h5h7 g5h7 g6f7 b7f7 b3f7 c3b2 b1b2 a4c6 d4f6 h7f6 g2g3 g7g5 b2c1
info nodes 35333
info depth 22 seldepth 29 time 60001 nodes 35333 nps 588 score cp -28
bestmove f1f4 ponder c6a4

Example 2

Code: Select all

info depth 22 seldepth 16 time 58044 score cp 992 nodes 34904 nps 601 pv g2g7 f7e8 d6e7
info nodes 36421
info depth 22 seldepth 16 time 60001 nodes 36421 nps 607 score cp 45
bestmove g2g7 ponder f7e8

Please note the difference in the final score compared to the former info line.

Komodo MCTS was running in a terminal under Linux, 1024 MB MCTS Hash, MCTS Explore set to 12, 1 min per position.
It also happens after a 'go infinite' and 'stop'.

That does look odd. I will take a look. I have not seen that before. Was this from the start position?

No. But it also happens with the startpos.

Code: Select all

uciok
setoption name MCTS Hash value 512
info string MCTS Hash table size is now 512 meg
setoption name Use MCTS value true
ucinewgame
go movetime 16000
info depth 9 seldepth 3 time 2001 score cp 52 nodes 99 nps 49 pv e2e4 e7e5 g1f3 d7d5
info depth 10 seldepth 5 time 4002 score cp 54 nodes 157 nps 39 pv e2e4 e7e5 g1f3 b8c6 f1c4 g8f6
info depth 11 seldepth 7 time 6003 score cp 51 nodes 270 nps 44 pv e2e4 e7e5 g1f3 b8c6 f1c4 g8f6 b1c3 f8c5
info depth 12 seldepth 8 time 8005 score cp 46 nodes 448 nps 55 pv e2e4 e7e5 g1f3 b8c6 f1c4 g8f6 b1c3 f8c5 e1g1
info depth 13 seldepth 8 time 10007 score cp 44 nodes 528 nps 52 pv e2e4 e7e5 g1f3 b8c6 f1c4 g8f6 e1g1 f8c5 f1e1
info depth 13 seldepth 8 time 12008 score cp 44 nodes 640 nps 53 pv e2e4 e7e5 g1f3 b8c6 f1c4 g8f6 e1g1 f8c5 f1e1
info depth 14 seldepth 8 time 14009 score cp 44 nodes 782 nps 55 pv e2e4 e7e5 g1f3 b8c6 f1c4 g8f6 e1g1 f8c5 f1e1
info nodes 947
info depth 14 seldepth 11 time 16002 nodes 947 nps 59 score cp 12
bestmove e2e4 ponder e7e5

[menniepals]

Jeremy Bernstein makes the Android versions for us, it's just a question of when he can find the time to do it.

I hope he finds time. I know it is not only I that is waiting.

[peter]

There simply isn't any need for adding MV- lines or reducing their number in Forward- Backward- Analysis anymore.

Only one weakness of this advantage is the lack of keeping "found" moves and evals in hash in backward- analysis qualitatively, compared to A-B- komodo and other good A-B-engines.

[Scally]

Can someone provide a binary for the Raspberry Pi please? Obviously we are talking about an older free version...

Thanks

Al.

[mjlef]

There simply isn't any need for adding MV- lines or reducing their number in Forward- Backward- Analysis anymore.

Only one weakness of this advantage is the lack of keeping "found" moves and evals in hash in backward- analysis qualitatively, compared to A-B- komodo and other good A-B-engines.

I think with a more complex MCTS node allocation scheme I could improve that. I will add it to my list. How many moves back should we try to keep? The MCTS node tree takes up a lot of memory, of course.

[peter]

How many moves back should we try to keep?

Well, I'd say as much as possible, if it's about the best moves all along the lines only, of course.

Seriously, I know, that komodo's hash- aging is a special one, e.g. changing the number of MV-lines during running analysis makes the GUIs delete the hash already more or less completely, that's another reason I enjoy not to have to change that at Forward- Backward with MCTS anymore.

It's a matter of wideness of tree as for similarly good evaluated candidates along the lines of course, but default komodo is still much better in backward hashing, probably a matter of other way of search in MCTS- mode by itself.

E.g. LC0 isn't as good neither in that matter as for my personal pov as good A-B- engines are.

Edit: Had MCTS Explore set too high for my first trials probably, not having seen at once, that it's now not about Exploit anymore, having reduced that to Default seems to make backward hashing better already again.

[Joerg Oster]

That does look odd. I will take a look. I have not seen that before. Was this from the start position?

Mark, I also just tried the Windows version.
Same issue.

Code: Select all

uciok
setoption name Use MCTS value true
ucinewgame
isready
readyok
go infinite
info depth 10 seldepth 5 time 2002 score cp 33 nodes 157 nps 78 pv e2e4 e7e5 g1f3 b8c6 f1c4 f8e7
info depth 12 seldepth 9 time 4003 score cp 32 nodes 340 nps 84 pv e2e4 e7e5 g1f3 b8c6 f1b5 a7a6 b5c6 d7c6 e1g1 c8e6
info depth 13 seldepth 10 time 6004 score cp 29 nodes 592 nps 98 pv e2e4 e7e5 g1f3 b8c6 f1b5 a7a6 b5c6 d7c6 e1g1 c8e6 d2d4
info depth 14 seldepth 10 time 8004 score cp 28 nodes 740 nps 92 pv e2e4 e7e5 g1f3 b8c6 f1b5 a7a6 b5c6 d7c6 e1g1 c8e6 d2d4
stop
info nodes 830
info depth 14 seldepth 10 time 9721 nodes 830 nps 85 score cp 8
bestmove e2e4 ponder e7e5