Komodo CCT release

[Tom Likens]

I'm not using every dark corner of the language, I just wanted to learn C++ and was planning a major code cleanup anyway. In particular this is more like a C++ program than C++11 and the position state is now a class, I like the new for loops which is c++11 specific and templates are nice. Probably more than anything I wanted to use the c++ standardized threading model and the "chrono" stuff for consistent time handling which is quite awesome. There is not that much difference in windows threads and pthreads but it's great having completely portable source code.

I'm mulling over the switch to C++11 for exactly the reasons you mention. My issue is whether or not to start with a new codebase or retroactively fit the changes into what I have now. In many ways starting with a new codebase would be cleaner (and it might even be faster to finish).

[Jesse Gersenson]

Wow, +38 elo on IPON. Well done Don and Larry!

[lkaufman]

Wow, +38 elo on IPON. Well done Don and Larry!

Thanks! Now on the list we are much closer to first place Houdini 3 than to third place Stockfish 3 and Critter. I think the reason we came out ahead of expectations is that we made a last-minute change to the parameter "Time usage Aggressiveness", based partly on statements that we move too quickly in the early moves in the recent tournaments. In effect, we moved the zero point default to be three points more aggressive. This seems to have helped.

[IWB]

... Now on the list we are much closer to first place Houdini 3 than to third place Stockfish 3 and Critter....

It is just 150 games but Komodo CCT is the only engine to score more than 40% (45!) vs. H3. That is quite an achievement! On equal hardware and longer time controls I would not bet on the individual outcome of a game!

Congratulations to this engine!

Ingo

[lkaufman]

... Now on the list we are much closer to first place Houdini 3 than to third place Stockfish 3 and Critter....

It is just 150 games but Komodo CCT is the only engine to score more than 40% (45!) vs. H3. That is quite an achievement! On equal hardware and longer time controls I would not bet on the individual outcome of a game!

Congratulations to this engine!

Ingo

I hope that someone will run a long match at a reasonably long time control between the two (on one core). In view of our results in the CCT and TCEC tournaments giving one core vs. multi-core handicap, there is at least some reason for optimism.
I also noticed near the end of your test that the performance rating for Komodo against every single opponent was over 3010, well above the rating of Komodo 5. That degree of consistency surprised me. We don't seem to have any particularly difficult opponents.

[IWB]

... I also noticed near the end of your test that the performance rating for Komodo against every single opponent was over 3010, well above the rating of Komodo 5. That degree of consistency surprised me. We don't seem to have any particularly difficult opponents.

Yes, it seems you made a very nice and even engine:

Komodo CCT - Houdini 3 STD (3079) 68.5 - 81.5 45.67% Perf=3049
Komodo CCT - Stockfish 3 (2978) 88.0 - 62.0 58.67% Perf=3038
Komodo CCT - Critter 1.4a (2977) 83.0 - 67.0 55.33% Perf=3014
Komodo CCT - Deep Rybka 4.1 (2955) 91.0 - 59.0 60.67% Perf=3030
Komodo CCT - Gull II (2928) 96.0 - 54.0 64.00% Perf=3027
Komodo CCT - Chiron 1.5 (2843) 109.0 - 41.0 72.67% Perf=3012
Komodo CCT - Protector 1.5.0 (2839) 114.0 - 36.0 76.00% Perf=3039
Komodo CCT - Naum 4.2 (2832) 118.5 - 31.5 79.00% Perf=3062
Komodo CCT - Hannibal 1.3 (2827) 119.0 - 31.0 79.33% Perf=3060
Komodo CCT - HIARCS 14 WCSC 32b (2819) 118.0 - 32.0 78.67% Perf=3045
Komodo CCT - Deep Shredder 12 (2800) 120.5 - 29.5 80.33% Perf=3044
Komodo CCT - Deep Sjeng c't 2010 32b (2783) 126.0 - 24.0 84.00% Perf=3071
Komodo CCT - Spike 1.4 32b (2769) 123.5 - 26.5 82.33% Perf=3036
Komodo CCT - spark-1.0 (2759) 129.0 - 21.0 86.00% Perf=3074
Komodo CCT - Deep Junior 13.3 (2746) 129.5 - 20.5 86.33% Perf=3066
Komodo CCT - Quazar 0.4 (2729) 130.0 - 20.0 86.67% Perf=3054
Komodo CCT - Toga II 3.0 32b (2713) 135.0 - 15.0 90.00% Perf=3094
Komodo CCT - Zappa Mexico II (2705) 133.5 - 16.5 89.00% Perf=3068


Chiron 1.5 is the worst enemy and Toga II 3.0 the easiest - looking at the performance ... but again, I think 150 games are not enough for this approach.

Bye
Ingo

[Don]

Wow, +38 elo on IPON. Well done Don and Larry!

Thanks! Now on the list we are much closer to first place Houdini 3 than to third place Stockfish 3 and Critter. I think the reason we came out ahead of expectations is that we made a last-minute change to the parameter "Time usage Aggressiveness", based partly on statements that we move too quickly in the early moves in the recent tournaments. In effect, we moved the zero point default to be three points more aggressive. This seems to have helped.

Actually, we did this in addition to reworking the time control algorithm to be a lot smarter about when to take less time and when to take more time. The time control work we did probably added 10 ELO to our results.

[Uri Blass]

Wow, +38 elo on IPON. Well done Don and Larry!

Thanks! Now on the list we are much closer to first place Houdini 3 than to third place Stockfish 3 and Critter. I think the reason we came out ahead of expectations is that we made a last-minute change to the parameter "Time usage Aggressiveness", based partly on statements that we move too quickly in the early moves in the recent tournaments. In effect, we moved the zero point default to be three points more aggressive. This seems to have helped.

Actually, we did this in addition to reworking the time control algorithm to be a lot smarter about when to take less time and when to take more time. The time control work we did probably added 10 ELO to our results.

Note that the rating improvement of komodo has statistical error and 38 elo in the ipon list can be also 28 elo(and it can be also 48 elo).

It is too early to know if you really earned additional rating relative to approximately 30 elo that you expected and we can wait to see if komodo also get more than 30 elo improvement in other rating lists.

[Don]

Wow, +38 elo on IPON. Well done Don and Larry!

Thanks! Now on the list we are much closer to first place Houdini 3 than to third place Stockfish 3 and Critter. I think the reason we came out ahead of expectations is that we made a last-minute change to the parameter "Time usage Aggressiveness", based partly on statements that we move too quickly in the early moves in the recent tournaments. In effect, we moved the zero point default to be three points more aggressive. This seems to have helped.

Actually, we did this in addition to reworking the time control algorithm to be a lot smarter about when to take less time and when to take more time. The time control work we did probably added 10 ELO to our results.

Note that the rating improvement of komodo has statistical error and 38 elo in the ipon list can be also 28 elo(and it can be also 48 elo).

It is too early to know if you really earned additional rating relative to approximately 30 elo that you expected and we can wait to see if komodo also get more than 30 elo improvement in other rating lists.

Our own very thorough private testing indicates over 30 ELO and then another 10 ELO for the time control changes. So at least this is consistent with what we have seen and others have reported to us. For example I can tell you that the gain will not be the 48 ELO gain you suggest is possible. I think it's almost certainly within 5 ELO of what Ingo measured - but of course different tests, different time controls and different conditions add a lot of uncertainty to tests like this.

Don

[Modern Times]

I hope that someone will run a long match at a reasonably long time control between the two (on one core). In view of our results in the CCT and TCEC tournaments giving one core vs. multi-core handicap, there is at least some reason for optimism.

It is doing quite well here so far at 40/40:
16.5 - 17.5 against Houdini 3

It is comfortably beating Stockfish 3, but having a harder time against Rybka 4.1, just 58% against that so far.

That probably doesn't meet your definition of "reasonably long" though.