Komodo Sub

[Henk]

Yes best to stop with alpha beta like algorithms when A0 approach is clearly better.

Dive into deep learning and neural networks.

But maybe we don't have enough computer power yet for training them.

I don't have a GPU on my computer.

[leavenfish]

Anyway, when we do release, expect a faster BMI version included in the package (for Haswell and newer hardware supporting PEXT/BMI2), and more strength. We are pretty close but until the gain are found found and confirmed, and testing is complete, I do not know when we will release.

Mark

My guess since you are chasing elo...a release will not happen until Team Stockfish has their official release. That way you can test and see if you are close or past them in terms of elo.

Really, I think what you are seeing (myself included) since progress is slow as you say, is that subscribers would very much like to have access to more of the 'test' versions.

People who pay for an engine (not say with a GUI, just engines) want to tinker and test for their money. Just my two cents...as a subscriber.

[mjlef]

Anyway, when we do release, expect a faster BMI version included in the package (for Haswell and newer hardware supporting PEXT/BMI2), and more strength. We are pretty close but until the gain are found found and confirmed, and testing is complete, I do not know when we will release.

Mark

My guess since you are chasing elo...a release will not happen until Team Stockfish has their official release. That way you can test and see if you are close or past them in terms of elo.

Really, I think what you are seeing (myself included) since progress is slow as you say, is that subscribers would very much like to have access to more of the 'test' versions.

People who pay for an engine (not say with a GUI, just engines) want to tinker and test for their money. Just my two cents...as a subscriber.

We have considered releasing more development versions with things like access to more search and evaluation components. On the plus side it may be that users can find better settings than us. On the minus, I would have to try and explain what all these things are doing, and they would not be well tested with the common GUIs since that takes up a lot of time. We are more likely to do that when more people ask for them.

[carldaman]

Anyway, when we do release, expect a faster BMI version included in the package (for Haswell and newer hardware supporting PEXT/BMI2), and more strength. We are pretty close but until the gain are found found and confirmed, and testing is complete, I do not know when we will release.

Mark

My guess since you are chasing elo...a release will not happen until Team Stockfish has their official release. That way you can test and see if you are close or past them in terms of elo.

Really, I think what you are seeing (myself included) since progress is slow as you say, is that subscribers would very much like to have access to more of the 'test' versions.

People who pay for an engine (not say with a GUI, just engines) want to tinker and test for their money. Just my two cents...as a subscriber.

We have considered releasing more development versions with things like access to more search and evaluation components. On the plus side it may be that users can find better settings than us. On the minus, I would have to try and explain what all these things are doing, and they would not be well tested with the common GUIs since that takes up a lot of time. We are more likely to do that when more people ask for them.

Sounds like a good idea, Mark. Users like to experiment.

[MikeB]

No worries here Mark. Will be a loyal subscriber until you guys decide to retire. Hopefully not anytime soon. 😊

[Ovyron]

Sounds like a good idea, Mark. Users like to experiment.

Yes! I have read that there are rumors about people that like running those experiments, not necessarily to create settings stronger than default, but to create engine personalities that play attractive games

It would also be excellent if we could get some free Komodo 9 with some access to search and evaluation components, so that users interested in this (but not in the elo improvement) wouldn't have to wait till version 11/12 goes free.

[Nordlandia]

Stockfish Dev still struggles in Queen vs Minor imbalances.

[pgn][Event "DESKTOP-CBOUVEB, Rapid 15m+30s"]
[Site "https://lichess.org/LZmpisC5"]
[Date "2018.01.16"]
[Round "1"]
[White "Komodo 11.2.2 64-bit"]
[Black "Stockfish 130118 64 POPCNT"]
[Result "1-0"]
[WhiteElo "?"]
[BlackElo "?"]
[Variant "From Position"]
[TimeControl "-"]
[ECO "?"]
[Opening "?"]
[Termination "Normal"]
[FEN "r1bqk1nr/ppppp2p/8/8/8/8/PPPPPPPP/RNB1KBNR w KQ - 0 1"]
[SetUp "1"]
[Annotator "lichess.org"]

1. d3 Nf6 2. Nf3 Rg8 3. g3 d6 4. h3 b6 5. Bg2 Bb7 6. c4 Kd7 7. Nc3 Qf8 8. e4 Nh5 9. Rg1 e5 10. Nh4 Ng7 11. Be3 Re8 12. O-O-O Kc8 13. Kb1 Qe7 14. Nd5 Qd8 15. Nf5 Kb8 16. g4 Ne6 17. Bf3 c6 18. Nc3 Ref8 19. Be2 c5 20. Rdf1 Nd4 21. f4 Qd7 22. Nh6 Rh8 23. f5 b5 24. Rf2 b4 25. Nd1 Bc6 26. g5 Qb7 27. Bg4 Bxe4?! { (0.00 → 0.54) Inaccuracy. Best move was b3. } (27... b3 28. a3 Nc2 29. Kc1 Bxe4 30. dxe4 Qxe4 31. Bf3 Qxc4 32. Be2 Qe4 33. Bf3 Qc4) 28. dxe4 Qxe4+ 29. Kc1 b3 30. a3 Nc2 31. Bf3 Qxc4 32. Be2 Qe4 33. Rg4 Qb7 34. f6 d5?! { (0.07 → 0.85) Inaccuracy. Best move was Rd8. } (34... Rd8) 35. Bd2?! { (0.85 → 0.07) Inaccuracy. Best move was Bxc5. } (35. Bxc5 Qc7 36. Bxf8 Ne3+ 37. Kd2 Nxg4 38. Rf3 Qc2+ 39. Ke1 Rxf8 40. Nxg4 d4 41. Rd3 Kc7) 35... c4 36. f7 c3?! { (-0.20 → 0.72) Inaccuracy. Best move was d4. } (36... d4 37. Rf5) 37. bxc3 e4 38. Kb2 Ka8 39. Rf6 Qe7 40. Ra6 e3?! { (0.65 → 1.42) Inaccuracy. Best move was Rb8. } (40... Rb8 41. Bf4) 41. Nf5 Qe5 42. Nfxe3 Nxe3 43. Bxe3 Rxf7 44. Rd4 Qf5 45. Rd2 Re8 46. Bb5 Re6 47. Ra4 Qf3?! { (1.42 → 1.97) Inaccuracy. Best move was Qxh3. } (47... Qxh3 48. Bc4) 48. Bf2 Rfe7?! { (1.78 → 2.70) Inaccuracy. Best move was Kb8. } (48... Kb8 49. Rb4 Rb7 50. h4 a6 51. Bf1 Rxb4 52. axb4 h6 53. gxh6 Rxh6 54. Kxb3 Rc6 55. Be2) 49. Rad4 a6 50. Bf1 Re5?! { (2.34 → 3.21) Inaccuracy. Best move was Qh1. } (50... Qh1 51. Bd3) 51. h4 a5?! { (2.76 → 3.44) Inaccuracy. Best move was R7e6. } (51... R7e6) 52. Kxb3 Kb7 53. Bd3 R5e6?! { (2.99 → 3.75) Inaccuracy. Best move was Re8. } (53... Re8) 54. Ra4 h6?! { (3.59 → 4.45) Inaccuracy. Best move was Kb8. } (54... Kb8 55. Rxa5) 55. g6 Re2 56. Kc2 Rxd2+ 57. Kxd2 { Black resigns. } 1-0[/pgn]

https://lichess.org/LZmpisC5

[kk]

My testing is revealing a distinct difference between Houdini 6.0.3 and Komodo 11.2.2 whereby Houdini is always quicker to find a longer mating line, and reporting it. I run Houdini on a dual core i7 Surface whereas Komodo runs on a HP Omen with 4 core i7 and still Komodo may take considerably more time to report a mating line. This has happened several times only in analysis, I haven't determined if this is the case for game play mode too.

I've also determined that when there is an overwhelming advantage for the side Komodo is playing it appears to neglect updating the lines and doesn't appear to report consistently as it continues processing through the plies. Where an overwhelming advantage is reported by Komodo it doesn't appear to continue progressing through the plies as it normally would, and i can only guess that is the case as the depth of analysis is not reported as quickly as it normally does in game play mode. Whereas Houdini in the same situation appears to report as it usually would in game mode.



Has anyone else experienced this unusual behaviour.

[Paloma]

Here komodo has a similar behavior.
depth 31 in 1:00
depth 32 in 2:40
...
depth 35 in 8:06
...
depth 38 in 28:13
depth 39 in 40:12 and still standing on this depth for 10 hours.

eval +8.66
no more progress so far

Task manager shows 50% utilization
Komodo 11.2.2, i7 6700 @4.0GHz, 2T, 8GB Hash

[d] 1r3rk1/5p1p/pp1p1p1Q/2q2P2/2P1p1b1/P1PB1R2/6PP/1R5K w - - 0 6

Analysis by Komodo 11.2.2 64-bit:

6.Bxe4 Rfe8 7.Bd5 Bxf3 8.gxf3 Re5 9.Rg1+ Qxg1+ 10.Kxg1 Rxf5 11.f4 Re8 12.Qh3 Re1+ 13.Kg2 Re2+ 14.Kf3 Ree5
Weiß steht klar auf Gewinn: +- (6.56 ++) Tiefe: 32 00:02:40 848MN, tb=591671
...
6.Bxe4 Qe5 7.Re3 Bxf5 8.Bxf5 Qxf5 9.Rg3+ Qg6 10.Rxg6+ fxg6 11.Qf4 Rb7 12.Qd4 Re8 13.Kg1 Rbe7 14.Qxf6 Re6
Weiß steht klar auf Gewinn: +- (8.49 ++) Tiefe: 38 00:28:13 9587MN, tb=3581086
6.Bxe4 Qe5 7.Re3 Bxf5 8.Bxf5 Qxf5 9.Rg3+ Qg6 10.Rxg6+ fxg6 11.Qf4 Rb7 12.Qd4 Re8 13.Kg1 Rbe7 14.Qxf6 Re6
Weiß steht klar auf Gewinn: +- (8.72) Tiefe: 38 00:34:20 11653MN, tb=3921838
6.Bxe4 Qe5 7.Re3 Bxf5 8.Bxf5 Qxf5 9.Rg3+ Qg6 10.Rxg6+ fxg6 11.Qf4 Rb7 12.Qd4 Re8 13.Kg1 Rbe7 14.Qxf6 Re6
Weiß steht klar auf Gewinn: +- (8.66 --) Tiefe: 39 00:40:12 13699MN, tb=4522611

[kk]

Hi Herbert,
Your example very much illustrates one of the points I had experienced.
I am aware that for every ply that is completed something like an exponential increase in time is required to complete the next ply....however, i can leave the processing overnight and see no update in reporting of analysis where the score for the side to move is high.

However in a great number of cases I tend to leave both Komodo and Houdini to analyse positions to around 40 ply or a couple of hours at most and the proposed move correlates well between the two machines and the score at that ply is often very close. I have witnessed Houdini reaching the best book move (according to Hiarcs G++) first and early but where Komodo make take a little longer in most cases it does get there. There are examples of Komodo making it to the book move first. There are also examples of neither program agreeing with Hiarcs G++ but both programs finding the same deviating move, this may be after an overnight analysis, but wqould suggest maybe Hiarch G++ doesn't have all the answers?