Komodo 3 release

[Ajedrecista]

Hi Don:

Thanks for answering so fast. Your effort must be huge, but the reward is this fantastic engine. Gaviota TB support sounds good (6-man Gaviota would be superb!). With the next release, Komodo will be among the best engines and maybe a true contender to the World Title when multicore support will be added. Good luck!

You posted that maybe +30 Elo is gained from Komodo 2.03 but I have read in other fora that if you say +30 it is +30 as minimum, and more likely +35 or +40. It means that Komodo 3 will be near 2990 in IPON rating list (now Komodo 2.03 is at 2952). It would be impressive! Wait and see while Ingo starts the matches.

Regards from Spain.

Ajedrecista.

[Don]

Hi Don:

Thanks for answering so fast. Your effort must be huge, but the reward is this fantastic engine. Gaviota TB support sounds good (6-man Gaviota would be superb!). With the next release, Komodo will be among the best engines and maybe a true contender to the World Title when multicore support will be added. Good luck!

You posted that maybe +30 Elo is gained from Komodo 2.03 but I have read in other fora that if you say +30 it is +30 as minimum, and more likely +35 or +40. It means that Komodo 3 will be near 2990 in IPON rating list (now Komodo 2.03 is at 2952). It would be impressive! Wait and see while Ingo starts the matches.

Regards from Spain.

Ajedrecista.

It's very difficult to predict the rating gain. In some releases we measure more than we actually have gotten, it other releases we get more than we think we should on the rating lists. Usually it has been more than we think but not always.

In our private testing at 30 seconds + 0.5 increment it is coming out at 40 ELO as measured against 3 foreign programs and that is probably the basis for what I said. To be conservative we expect to actually get half of that, or 20 ELO points. We would be very disappointed if it was less than that. It's difficult to predict how our changes scale to longer time controls but for the most part an improvement at game in 15 seconds and beyond is a real improvement at any longer time controls but it may not be by the same amount. Much below game in 15 seconds and we can be surprised.

However we also improved our time control algorithm for non-fischer time controls. We noticed that our results on the IPON were much higher that our results on the other lists and this is one major difference in how the testing is done. We have never bothered to optimize the time control (other than Fischer) and it's very simple and basic and we discovered that we are losing at least 10 ELO due to that. So it should go up 30 ELO on most lists and 20 on IPON and there is a chance it will be a few more than that on both list. It could be less too, but this would surprise and disappoint us.

For reference, the vast majority of improvements have come from evaluation changes. Larry already mentioned that we looked at compensation for a pawn and to make a long story short we lowered the value of a pawn slightly combined with other things. But we did many other evaluation changes too.

Komodo does have what we consider original ideas that we came up with on our own, although some of them have made their way to other programs or were discovered independently by others. So Komodo 3 added another nice idea that I don't think anyone else does and that was something like 5-10 ELO. We also got a minor improvement with help from Richard Vida and we found a bug (oversight) in his program which helps his program by a small amount too.

One thing that I just checked yesterday was that our tactical performance has improved. People have been complaining that Komodo is tactically weaker than the top programs and even programs significantly weaker than Komodo. So I ran the Arasan13 test and compared against other programs and Komodo 2.03 and much to my surprise we have improved - and non-trivially. We did not set out to do that, it just happened as a result of other improvements. It's still below par compared to other programs but the difference is not quite as embarrassing. Critter of course blows them all away in tactics (I did not check Houdini.)

[Ajedrecista]

Hi Don:

Thanks for answering so fast. Your effort must be huge, but the reward is this fantastic engine. Gaviota TB support sounds good (6-man Gaviota would be superb!). With the next release, Komodo will be among the best engines and maybe a true contender to the World Title when multicore support will be added. Good luck!

You posted that maybe +30 Elo is gained from Komodo 2.03 but I have read in other fora that if you say +30 it is +30 as minimum, and more likely +35 or +40. It means that Komodo 3 will be near 2990 in IPON rating list (now Komodo 2.03 is at 2952). It would be impressive! Wait and see while Ingo starts the matches.

Regards from Spain.

Ajedrecista.

It's very difficult to predict the rating gain. In some releases we measure more than we actually have gotten, it other releases we get more than we think we should on the rating lists. Usually it has been more than we think but not always.

In our private testing at 30 seconds + 0.5 increment it is coming out at 40 ELO as measured against 3 foreign programs and that is probably the basis for what I said. To be conservative we expect to actually get half of that, or 20 ELO points. We would be very disappointed if it was less than that. It's difficult to predict how our changes scale to longer time controls but for the most part an improvement at game in 15 seconds and beyond is a real improvement at any longer time controls but it may not be by the same amount. Much below game in 15 seconds and we can be surprised.

However we also improved our time control algorithm for non-fischer time controls. We noticed that our results on the IPON were much higher that our results on the other lists and this is one major difference in how the testing is done. We have never bothered to optimize the time control (other than Fischer) and it's very simple and basic and we discovered that we are losing at least 10 ELO due to that. So it should go up 30 ELO on most lists and 20 on IPON and there is a chance it will be a few more than that on both list. It could be less too, but this would surprise and disappoint us.

For reference, the vast majority of improvements have come from evaluation changes. Larry already mentioned that we looked at compensation for a pawn and to make a long story short we lowered the value of a pawn slightly combined with other things. But we did many other evaluation changes too.

Komodo does have what we consider original ideas that we came up with on our own, although some of them have made their way to other programs or were discovered independently by others. So Komodo 3 added another nice idea that I don't think anyone else does and that was something like 5-10 ELO. We also got a minor improvement with help from Richard Vida and we found a bug (oversight) in his program which helps his program by a small amount too.

One thing that I just checked yesterday was that our tactical performance has improved. People have been complaining that Komodo is tactically weaker than the top programs and even programs significantly weaker than Komodo. So I ran the Arasan13 test and compared against other programs and Komodo 2.03 and much to my surprise we have improved - and non-trivially. We did not set out to do that, it just happened as a result of other improvements. It's still below par compared to other programs but the difference is not quite as embarrassing. Critter of course blows them all away in tactics (I did not check Houdini.)

Hi again:

IMO it is better to be conservative with the Elo gain issue (it is strange as minimum that someone announces +100 gain when really is +50). So I completely agree with you here (it also avoids disappointments). Congrats for the improve in Arasan test suite.

Your post is very complete! I did not expecte it. You explain the things quite clear: I knew some of them, but others were out of my knowledge until now. I am curious in your unique new idea that boosts 5 or 10 Elo your engine (no need to explain it). And it is very smart that both Komodo team and Mr. Vida have helped each other for improving your engines. I celebrate it because IMO it is better to help and receive help for improving not just in chess but in everything else. Please keep the good work!

Regards from Spain.

Ajedrecista.

[vijayendran]

Thanks for the 32 bit Komodo 3 !
Keep up the wonderful work.

Best Wishes,
Vijayendran

[vijayendran]

Which is the best recommended polyglot opening book for komodo ?

Vijayendran

[FWCC]

Congratulations,Komodo is No.2 behind Houdini on the IPON single core engines.It seems the MP version should be a killer!Good Luck and good work!

[Don]

Which is the best recommended polyglot opening book for komodo ?

Vijayendran

You can use the komodo book that Salvo did for us. There are a lot of polyglot books around and we did not carefully evaluate them for komodo, but this was made specifically for an ancient version of Komodo and probably would still be a good choice.

I'm going to put it up on the site shortly.

Don

[vijayendran]

You can use the komodo book that Salvo did for us. There are a lot of polyglot books around and we did not carefully evaluate them for komodo, but this was made specifically for an ancient version of Komodo and probably would still be a good choice.

I'm going to put it up on the site shortly.

Don

Okay, Thank you.

[ernest]

Komodo strange small bug:
infinite analysis, with MPV=3

[d]3r1b2/6pk/8/4BQPP/p7/7K/8/3r4 b - - 0 1
Analysis by Komodo32 3 32bit:
1. -+ (-#1998): 46...g6 47.Qxg6#
2. +/- (1.39): 46...Kg8 47.h6 R1d7 48.Qe6+ Kh8 49.g6 a3 50.Kg2 Rd2+ 51.Kf3 R2d3+ 52.Kf4 R3d4+ 53.Bxd4 Rxd4+ 54.Kf5 Rd8 55.Qe5 Rc8 56.Kg4 Rd8 57.hxg7+ Bxg7 58.Qa5 Re8 59.Qxa3 Rb8 60.Qa2 Rb4+ 61.Kf3 Be5 62.Qe6
3. +/- (1.39): 46...Kh8 47.h6 a3 48.Kg4 R1d7 49.g6 Kg8 50.Qe6+ Kh8 51.Kf3 Ra7 52.hxg7+ Rxg7 53.Qf5 Re8 54.Kg4 Rd8 55.Bxg7+ Bxg7 56.Qa5 Re8 57.Qxa3 Rd8 58.Qa5 Rb8 59.Qh5+ Kg8 60.Qd5+ Kh8 61.Qa5

[Don]

This is a known bug that has been reported many times. I'm working on fix for it for the next release.

Komodo strange small bug:
infinite analysis, with MPV=3

[d]3r1b2/6pk/8/4BQPP/p7/7K/8/3r4 b - - 0 1
Analysis by Komodo32 3 32bit:
1. -+ (-#1998): 46...g6 47.Qxg6#
2. +/- (1.39): 46...Kg8 47.h6 R1d7 48.Qe6+ Kh8 49.g6 a3 50.Kg2 Rd2+ 51.Kf3 R2d3+ 52.Kf4 R3d4+ 53.Bxd4 Rxd4+ 54.Kf5 Rd8 55.Qe5 Rc8 56.Kg4 Rd8 57.hxg7+ Bxg7 58.Qa5 Re8 59.Qxa3 Rb8 60.Qa2 Rb4+ 61.Kf3 Be5 62.Qe6
3. +/- (1.39): 46...Kh8 47.h6 a3 48.Kg4 R1d7 49.g6 Kg8 50.Qe6+ Kh8 51.Kf3 Ra7 52.hxg7+ Rxg7 53.Qf5 Re8 54.Kg4 Rd8 55.Bxg7+ Bxg7 56.Qa5 Re8 57.Qxa3 Rd8 58.Qa5 Rb8 59.Qh5+ Kg8 60.Qd5+ Kh8 61.Qa5