POLL: Whether Rybka Is An Original Work?

[kingliveson]

do not download this pdf-file.
my avast claims a virus.

Avast is a Czech based company, and guess what citizenship holds Rybka's developer?! Would it surprise me if the conspiracy deepens...

On a more serious note, uploading a virus would be an offense resorting in permanent ban if not temporary. So, no, it contains no virus -- it is just a simple pdf file.

[kingliveson]

As appears in the TalkChess thread, the lousiness of the a8 bishop plays a notable rôle in the above game. I might point out (again) that bad bishops are one of the major differences in method (beyond just parameter tweaking) of evaluation between Rybka 3 and the IPPOLIT series. [I am hesitant to publish the R3 details (and it would take a bit of work to do so completely), but perhaps I will do so if this continues to be an issue].

However I do not think IPPOLIT/IvanHoe has any particular decrement for an a8 bishop blocked by its own pawns [there is the "bishop on back rank penalty" but nothing else], and so Houdini seems to have added an additional evaluation element here [I'm not quite sure it is as important as 40cp -- usually a bishop on a8 can move to b7 where presumably it is would "just" be a bad bishop, but here that is not possible at various junctures due to the White knights].

PS. I didn't Letouzey know was involved in Othello, and was also surprised the 8x8 game is "drawish" (though independent confirmation agrees --- 32-32 is the expected result, though my sources indicate the first player has the chances). And the PDF download was fine for me (not that it did anything besides collate ZW's web pages).

If time permits, please consider it.

[Razor]

Absolutely right Dr Ivannik - as my gran would say, make the most of the best and the least of the worse!

[BB+]

If time permits, please consider it.

Well, given that members of the Rybka team continue to make (admittedly private) claims like IPPOLIT is 90% closer to Rybka (3) than Rybka (1) is to Fruit, I suspect it is worthwhile to post something about the bad/good bishop evaluation.

I still don't think such posts are a great solution (as it only shows one part of the picture), but the alternatives seem little better. For instance, even if (say) I were to reconstruct "source" code that could replicate the Rybka output, there would still be the question of low-level differences (I can't imagine I would have sufficient patience to replicate every nuance in R3 -- indeed, at some point, I'm not sure how it wouldn't just be easier to look at a disassembly rather than such "source" code). The appendices in the original R3/IPPOLIT report should already suffice for most purposes (that is, there are many overall similarities, though at the same time a plethora of differences), but I guess you have to repeat things often enough for them to become accepted.

On the Fruit vs Rybka front, I reiterate my acceptance of the bulk of ZW's analysis, though I think he jumps to conclusions at some points. Though "code" copying [and perhaps GPL violations] have occluded much of the discussion, I still find the nearly identical framework of almost the entire evaluation function [not to mention the exact PST scaling] to be of much more interest when considering whether it is an "independent" engine. List the "positional elements" in evaluation of both Fruit and R1, and you will see what I mean. [Contrariwise, the choice of the 5 components (doubled, isolated, backward, passed, open/closed) in pawn evaluation cannot be considered so "unique" as it were].

Code: Select all

"kingsafety" is always whether something attacks a square attacked by the opposing king
Knights: mobility, kingsafety
Bishops: mobility, kingsafety, and whether trapped
Rooks: mobility, kingsafety, semi-open/open files (with opp king on it another bonus), rook on 7th (under certain conditions), and whether blocked
Queens: mobility, kingsafety, queen on 7th

Find another engine that uses exactly these features, and perhaps I would find it more likely that Fruit and R1 are truly "independent". [There are further similarities (see ZW's analysis) in pawn shelter and passed pawns, though in those cases it is not quite so definitive]. For instance, Crafty has outposts for knights/bishops, "tropisms" for kingsafety, opposing wing bonuses for bishops in endgames, has different criteria for rook on 7th and no bonus for opp king with open files, and doesn't count queen mobility. A list of differences with Glaurung would be similar.

In short, I might agree that R3/IPPOLIT is closer than R1/Fruit [though it depends on how much weight one gives to the various similarities/differences], but not by anywhere near 90%.

[BB+]

For another perspective on the R1/Fruit issue (I always try to refer to Rybka 1 rather than the generic "Rybka", as else the discussion can be changed too easily by those who prefer to stress later developments), one can first note that Fruit 2.1 is 2796 on CCRL 40/40, while Fruit 1.0 seems to pre-date that list (FL estimates Fruit 2.1 is about 150-200 Elo better in evaluation alone).

When you read the technical10.txt (which applies mostly to Fruit 1.0), FL notes the evaluation function is lousy, saying:

There are two (bad) reasons why the evaluation is so "simple":
1) Fruit was designed to experiment with search algorithms (not just for chess)
2) I just can't be bothered with trying to design a "good" evaluation function, as this would be an extremely boring occupation for me.

By version 2.1, he had rewritten the eval routine, and says:

- added (small) king-attack bonus, the last *huge* hole in the eval;
now only large holes remain, "be prepared" says he (to himself)!
[...]
For the first time of its life, after the recent addition of king attacks, Fruit has all major (but admittedly few others) evaluation components. Don't get me wrong: they all need a lot of refinement, but the code layout is there.
[...]
Although I believe I could keep on increasing strength by adding more and more eval terms, I have little interest in doing so. I would not learn anything in the process, unless I develop new tuning/testing techniques. Ideally I would like to spend more time in alternative software, like my own GUI perhaps (specific to engine testing/matches).

Nonetheless, a lot can be done like tuning existing code or building an adapted opening book. Therefore, don't hesitate to contact me if you are interested in giving a hand. Computer testing time is especially welcome, but be warned that I am quite demanding. [...]

[emphasis added -- as noted above, he thought that already 150-200 Elo had been added, but here indicates that more should be readily possible].

Over the next two years, Thomas Gaksch improved Fruit 2.1 in the Toga II series (largely as a hobby, I take it), eventually reaching 2931 with Toga II 1.4.1SE, a gain of 135 rating points over Fruit 2.1. Perhaps half(?) of this 135 could be said to be "easy" (though "boring") to gain via tuning and other tweaks [though you still have to do it, which is nontrivial!].

So once Fruit 2.1 was around for awhile (it was released 2005/06/17), I don't think a ~2860 engine was too extraordinary for an assiduous aspirant. The initial Rybka 1.0 32-bit weighs in at 2887, and it was only in later versions that Rybka really became so far ahead of the field. The main Fruit/R1 difference seem to me to be: bitboards (particularly in 64-bit mode, but already in 32-bit they gain some), tuning, the material evaluation table, and some extra pruning. My speculations on the later history of Rybka appear in this thread. In some sense, beating R3 by 50 rating points (by now I think IvanHoe exceeds R3 by this much, even if IPPOLIT did not already) seems more impressive to me than beating Fruit 2.1 by 100 or more, as my impression (from the above quotations) is that FL didn't always seem overly motivated to max-out the strength.

As in my previous post, the fact that R1 and Fruit both have just the same "major evaluation components" and little else is one of my major causes for suspicion that they are not truly "independent", given the traditional construal of this word in computer chess.

[kingliveson]

But anyway that was Rybka 1.0 beta ... and here we still are talking as if Rybka 3 was a clone of Fruit. I have yet to see one single person say that Rybka 3 has anything to do with Fruit ... not one ... yet we have the continuous babble about ALL Rybka's being Fruit clones and thus Vas must be punished and thus cloning or stealing his work is fine and dandy ... viva la communist revolution!

Fruit is 2626 ELO, and 32 bit Rybka on this same list is 2849 which is a difference of 223 ELO points which is massive.

The above is from talkchess, and I laughed out loud after reading it.

Here are the values from CCRL:

Code: Select all

Rybka 1.0 32-bit 2888
Fruit 2.1 32-bit 2796	

As for the first quote, So it is ok to take an open-source program, improve it, close the source, sell the program, and after 2 or 3 releases, everything should be all dandy?! hahaha.... Anyways, by the same token, ippolit took ideas (not direct code copy/paste) as there are no publicly available Rybka source code, improved on those ideas, realeased the source to the public, and have now released multiple versions that have changed considerably from its initial release. And here we are still talking about IvanHoe as if it was a Rybka 3 clone. Dandy?

These are the same folks who claim this new engine was not stronger than Rybka 3 and yet testings show Rybka 4 and IvanHoe to be about 5 ELO apart within the margin of error.

Code: Select all

Rybka 4 64-bit 4CPU	3292
Rybka 3 64-bit 4CPU	3259

[BB+]

By now I think IvanHoe exceeds R3 by this much [50 Elo], even if IPPOLIT did not already

I should have been more clear that I was talking about on 1 cpu at "blitz" (say 40/4).

My recollection is that Jury Osipov found the material imbalance tables alone in R1 to be enough to gain 68 Elo over Fruit 2.1, but maybe I am conflating some of this (it could be imbalance tables from a later Rybka version, for instance), and I can't find the source of that statement.

[BB+]

My recollection is that Jury Osipov found the material imbalance tables alone in R1 to be enough to gain 68 Elo over Fruit 2.1, but maybe I am conflating some of this (it could be imbalance tables from a later Rybka version, for instance), and I can't find the source of that statement.

Had the right number, but the wrong version. "From my test: Rybka 3 without material imbalance values - minus 68 points Elo." http://talkchess.com/forum/viewtopic.ph ... 02&t=27616

[BB+]

Admittedly the error bars are bigger than those of CCRL, but this (which I found when searching for "Glaurung 0.1.5") has

Code: Select all

   0    21 Rybka 1.0 Beta UCI             : 2757   20  20   743    54.2 %   2728   35.4 %
  -1    33 Fruit 2.1 UCI                  : 2689   22  22   593    49.1 %   2696   40.8 %

for a difference of only 68 rating points. Via going back through some of the history of this [the above link is the 43rd and last(?) update], it seems that most of the Rybka 1.0 Beta UCI testing was done before Strelka/Belka became prevalent (see e.g. this from NIL-23), and so there should be little "enemy effect" from similar engines.

The listed time control at this juncture is: 56'+3" (Celeron 1.7GHz), 52'+3" (Celeron 2GHz), 50'+3" (Celeron 2.6GHz) 48'+3" (Celeron 3GHz), though it seems a 1GHz Celeron was used at other times, and it seems to depend on whether you are a "star" or not as to which computer plays your games (see the Rules). The whole project is Infinite Loop Series by Igor Gorelikov: http://iggor.110mb.com/home.php