Not logged inRybka Chess Community Forum
Up Topic Rybka Support & Discussion / Rybka Discussion / The Best Bullet Playing Engine.
- - By Master Om (Bronze) Date 2010-08-13 05:24
The pathetic thing is Rybka 4 sucks in bullet. I think Rybka Dynamic is still the best 1 m player. I today played in PCC with Deep Rybka 4 SSE4.2  6 cores 3.2 Ghz against Rybka 3 Dynamic  ,4 cores 3.0 GHz (both AMD Phenoms II ) banging my head to win with decent book and lost 1-4 draws not counting. Whats happening yaar ?
Is their any settings in bullet.?
Parent - - By Uly (Gold) Date 2010-08-13 08:06
Try playing bullet like "40 moves in 40 seconds" repeating, or similar, Rybka should be much better then.
Parent - - By Master Om (Bronze) Date 2010-08-13 08:30
Actually few guys play 1 m +0 s. You know what settings i should go for Rybka 4? In hash ? In TB cache ?. I use 256 mb hash with 8 mb gui and 8 mb engine tb cache.
Parent - - By Uly (Gold) Date 2010-08-13 09:00
That won't have much of an impact, changing Normal Move Time or Max Move Time will.
Parent - By mocha1961 (***) Date 2010-08-13 11:14
i've been experimenting also on what NM and MM is the best to put in 1/0 games. with his pc spec, he should be doing well with 1min games. any suggestions vytron as to what to put? any input is much appreciated.
Parent - - By Master Om (Bronze) Date 2010-08-13 17:36
For Example ?
Parent - By Uly (Gold) Date 2010-08-13 23:23
I honestly have no idea, I only know that it will very much depend on the hardware and opponent, you can't just use some setting and hope it'll be best for your conditions.

Albert Silver was using Normal Move Time = 72, Max Move Time = 120, TC Buffer = 1, might be worth a try.
Parent - - By mindbreaker (****) Date 2010-08-14 05:26 Edited 2010-08-14 08:01
You do know that Rybka 4 is basically Dynamic Rybka 3 on steroids, well maybe on a protein drink;).  I suspect you need to run more games.  You need several hundred games before you can say anything.  I believe the default material values are identical to the R3 Dynamic rather than R3 Default.  And the material values are what separated Default, Human and Dynamic.

There is still a lot of room for optimization.  I have been working on it.  It really depends which engine the opponent is to make an optimal configuration. I have not been running any variants directly against R4 Default.  I am running with an average of a second a move rather than bullet.

I suspect a lot of people are running [Deleted].  The best I have been able to generate against that engine is my Experiment 28.  It appears to be just about identical in strength head to head (dead even in my tests). I have not been running large pages or SEE4.2 so the configurations should do better with those hardware helps.

Exp 28 (also best against Critter .80)
---------
3
100
150
4
3
10
8
10
8
10
8
17
13
34
26
100

There is still room to optimize the timings which I have not gotten to yet.

Against Naum 4, I recommend Exp 26.  It only lost one game in a hundred in my tests.

Exp 26
--------
3
100
150
1
0
10
8
10
8
10
8
17
13
32
28
100

Experiment 25 is the best against [Deleted]. It does very respectable. Somewhere around 55%.

Exp 25
--------
3
100
150
2
1
10
8
10
8
10
8
17
13
23
28
100

Sorry no anti-StockFish yet.  And of course all this is preliminary and I will also likely come up with improved optimizations.  I think I am over 25,000 games so far.  By about 75,000 games I should be able to give close to optimum general and opponent specific settings I hope (at least for the >3000 Elo engines I have).
And before any administrator kills this because I mentioned clones, consider that I am trying to make people more satisfied with R4 by allowing them to loose less to clones while using R4.
Parent - - By Uly (Gold) Date 2010-08-14 07:04
Oh god, now this thread will be deleted...
Parent - By Master Om (Bronze) Date 2010-08-14 15:39
:grin::grin::grin:
Parent - By Master Om (Bronze) Date 2010-08-14 15:40
WOW
Parent - - By Nelson Hernandez (Gold) Date 2010-08-15 12:22
Admirable.  Systematic experimentation.  Wish everyone had your initiative.
Parent - By Geomusic (*****) Date 2010-08-23 17:10
what happens when you avg  them all out into 1 personality? Should be best overall right? Very nice Work!
Parent - - By MadMax777 Date 2010-08-26 20:09
Terrific job !!! I was wondering if these settings works better playing with W's or B's. Eg: adding more points to W queen than B one...

MadMax
Parent - By mindbreaker (****) Date 2010-08-27 01:41
Thanks. From what I understand W is just self regardless of color and B is opponent.  It would be interesting if we could very by color and opponent.  I am not sure what happens in analysis mode.  Someone said that W does become White and B does become Black in analysis mode.  I would rather things stay the same.  I would have to do some testing to verify that.  Or maybe Vas can clarify it.
Parent - - By mindbreaker (****) Date 2010-08-27 02:43
Here is an update:

Rank    Engine    Rating    # of Games
1    can't say on this site    3356    900
2    Rybka 4 x64 Exp 24 v2    3330    1309
3    Rybka 4 x64 Exp 26 v2    3329    1307
4    Rybka 4 x64 Exp 28 v1    3327    1304
5    Rybka 4 x64 Exp 25 v2    3324    1308
6    Rybka 4 x64 Exp 18 v1    3320    3017
7    Rybka 4 x64 Exp 27 v1    3320    1306
8    Rybka 4 x64 Exp 22 v1    3317    2209
9    Rybka 4 x64 Exp 29 v1    3312    1304
10    Rybka 4 x64 Exp 23 v1    3311    2156
11    Rybka 4 x64 Exp 11 v1    3310    454
12    can't say on this site    3308    2141
13    can't say on this site    3307    716
14    Rybka 4 x64 Exp 8 v1    3304    1315
15    can't say on this site    3301    1876
16    Rybka 4 x64 Exp 14 Human    3302    900
17    can't say on this site    3301    2406
18    Rybka 4 x64 Exp 7 v1    3300    815
19    Rybka 4 x64 Exp 4 v1    3296    900
20    Rybka 4 x64 Exp 20 v1    3295    590
21    Deep Rybka 4 x64 v1    3296    161
22    Rybka 4 x64 Exp 15 v1    3294    1800
23    Rybka 4 x64 Exp 1 v1    3295    916
24    Rybka 4 x64 Exp 9 v1    3295    434
25    Rybka 4 x64 Exp 3 v2    3294    423
26    Rybka 4 x64 Forum v1    3291    1160
27    Rybka 4 x64 Exp 21 v1    3291    900
28    Rybka 4 x64 Exp 16 v1    3291    557
29    Deep Rybka 4 x64    3291    1986
30    Rybka 4 x64 Exp 19 v1    3287    189
31    Deep Rybka 4 x64 Lasker    3283    479
32    can't say on this site    3280    2342
33    Rybka 4 x64 Exp 13 vC13510    3281    750
34    Rybka 4 x64 Exp 2 v1    3280    472
35    Rybka 4 x64 Beta 15 v1    3281    426
36    Rybka 4 x64 Exp 12 v1    3279    900
37    Rybka 4 x64 Exp 10 v1    3278    462
38    Rybka 3    3278    1897
39    Rybka 3 Dynamic    3278    892
40    Rybka 3 Human    3275    1461
41    Rybka 4 x64 Exp 17 v1    3267    355
42    can't say on this site    3262    163
43    can't say on this site    3259    407
44    Rybka 4 x64 Exp 6 v1    3256    121
45    Deep Rybka 4 x64 Human    3254    174
46    can't say on this site    3251    409
47    Stockfish 1.8 JA 64bit    3235    2231
48    Stockfish 1.7.1 JA 64bit 4t    3231    4235
49    Stockfish 1.7 JA 64bit 4t    3221    560
50    can't say on this site    3208    24
51    Stockfish 1.6.3 JA 64bit    3204    384
52    Critter 0.80 64-bit    3168    900
53    Deep Fritz 11    3123    3874
54    Naum 4 4t    3111    3411
55    HIARCS 12 MP    3103    3283
56    Deep Shredder 12    3101    918
57    spark-0.4    3098    902
58    Zappa Mexico II    3087    1619
59    Komodo64 1.1 JA    3073    2338
60    Critter 0.70 64-bit    3073    2326
61    Protector 1.3.3 x64 4t    3067    386
62    Deep Shredder 11 UCI    3058    2410
63    Komodo64 1.0 JA    3051    407
64    bright-0.5c    3029    2350
65    bright-0.4a    3001    21
66    Protector 1.3.6 x64 4t    2962    700
67    Thinker54AInert-MP64-UCI    2958    900
68    spark-0.3a    2933    23
69    Deep Fritz 12    2650    385

Even if "deep" is not in the name it is 4-threads.  Only Komodo is a 1-thread.  Ratings aligned to Rybka 4 at 3291 (What it was at CCRL 40/4 when I checked).

This is not a normal ratings list.  Everything that is not a Rybka 4 variant only played Rybka 4 variants this gives them an advantage in the ratings system because the Rybka 4 variants played considerably weaker opponents on average and the ratings formulas are biased in that situation.  Also engines other than Rybka 4 variants may be a little high or a little low because they did not play a pool of players representative of the engines available.  They may have a clash of personality that may give them better or worse results than they would have generally.

I highly doubt the leader on this chart would have as strong a rating if it had played the same opponents...it is almost certainly inflated by the difference in the pool of opponents.

I test 9 variants at a time.  The batch with Experiments 22 - 29 and 18, is now complete and the next batch has started testing.  The next one should take 3 weeks or so.

For those who are curious here are the parameters for Exp 24:
-------------
3
100
150
3
2
10
8
10
8
10
8
17
13
32
28
100

Sorry if the chart does not line up well...don't know what is up with it...and I can't fix it.
Parent - By MadMax777 Date 2010-08-28 19:02
Thanks Mindbreaker for sharing these deep study with us. Please let us know your conclusions about this new phase, ok? :grin::grin::grin:

Best regards

MadMax777
Parent - - By Geomusic (*****) Date 2010-08-30 09:19
someone has  too much free time my gosh.
Parent - - By mindbreaker (****) Date 2010-08-30 10:01
The machine may be working away but I have other things to do.  Let's hope my cruelty to my machine doesn't precipitate a silicon revolt ;)
Parent - By Geomusic (*****) Date 2010-09-03 06:22
hehe
Parent - - By NATIONAL12 (Gold) Date 2010-09-03 23:03
you actually believe engine x is stronger than R4.of course Nelson loves it,his books are probably made up of micky mouse games played at blitz(sorry Nelson but you know my opinions) if you wish to test your results play against me on corr thread.:smile:
Parent - By mindbreaker (****) Date 2010-09-04 01:00
Evidently you did not read my post: "I highly doubt the leader on this chart would have as strong a rating if it had played the same opponents...it is almost certainly inflated by the difference in the pool of opponents."
Parent - - By NATIONAL12 (Gold) Date 2010-09-03 23:05
you actually believe engine x is stronger than R4.of course Nelson loves it,his books are probably made up of micky mouse games played at blitz(sorry Nelson but you know my opinions) if you wish to test your results play against me on corr thread.:smile:

sorry i did not read title of thread.
Parent - - By Nelson Hernandez (Gold) Date 2010-09-03 23:59
You are comparing two completely different things.  The fact is that opening books (of any size) don't play like correspondence players and vice-versa.  And just as obviously correspondence players, with time and toil, can develop winning combinations more often than a book that merely follows recorded games at a variety of time controls.  That has never been in doubt by anyone who knows computer chess.  But that does not invalidate opening books, because engines playing unassisted still need opening books, and as we all know not all books are created equal in size, design, intent.  Nor are properly done opening books a tool that correspondence players can dismiss.  I have seen many, many games between highly-rated correspondence players who make what appear to be (and are later proven to be) inane moves in the opening.  How do I know?  Because they are making moves that have been demonstrated to be vastly inferior to others through hundreds or thousands of trials from a wide variety of time controls and engines.  More power to them if they can still prevail against those odds, but more often than not, they don't.
Parent - - By Uly (Gold) Date 2010-09-04 18:00

> How do I know?  Because they are making moves that have been demonstrated to be vastly inferior to others through hundreds or thousands of trials from a wide variety of time controls and engines.


Yes, but, assume that with "magic", the correspondence player would be allowed to travel back in time to play from the relevant position against all those engines at their time control WHILE he plays at his correspondence time control. The draws would be turned to wins for the corr player, the games that were lost would be saved, or the tables turned, so if the statistics would be looked at again qith the new games' results, the move he played would be clearly the best.

It makes me wonder how much of the games played in the past are useful at this level, for instance, all the openings that were the best at Rybka 3 times just because they took advantage of a Rybka 3 weakness, but now Rybka 4 plays very differently and such openings are rubbish now, the statistics created are meaningless, and so, I extrapolate that the statistics created by entities weaker than Rybka 3 at faster time controls/slower hardware lose a lot of meaning. It's like chess Monte-Carlo simulations at a bigger scale, those kind of games can be played now at a much faster rate, so I'd think that the usefulness of a given game is short-lived, like, eventually a 40/40 game from the past becomes a 1 min bullet game from the present, and it's pointless to keep it around in the database, as it would be the same as playing lots of such games and adding them.
Parent - By Nelson Hernandez (Gold) Date 2010-09-04 18:54
The same could be said about the usefulness of grandmaster-level games.  We're talking different things again. 

Are the lastest Deep Rybka 4 games the best thing to have in a book?  If you are playing against Deep Rybka 4, yes.  Are Rybka 3 games reduced in value?  Yes.  Worthless?  No.

As for antique games and human games, the overwhelming majority of them follow lines that are subsequently refuted or shown to be ineffective.  If you remove those games from book then you miss the whole cycle of trial-and-error.  You remove potentially useful move diversity.  As I see it probably 99%+ of games are imperfect anyway.  Blunders are made.  Absolute truth is not available.  However a book can give you tendencies, just like a Monte Carlo, as you said.  And those tendencies can be statistically correlated to success.  All we're talking about here, after all, is squeezing out a median +0.05 book-exit premium, after all. 

The real question is whether your book is constrained by size.  If you only want a book to contain 100,000 or 1,000,000 games max, then you must be selective.  But if you have no constraint there is little harm done from keeping the whole collection in book.
Up Topic Rybka Support & Discussion / Rybka Discussion / The Best Bullet Playing Engine.

Powered by mwForum 2.27.4 © 1999-2012 Markus Wichitill