The Rybka-Fruit Case
In cases such as the antics of Langer and Espin very little proof is needed to establish the cloning. But in some cases there is a more sophisticated cloning effort, when the clone programmer(s) attempt to hide their actions by making changes to the code of “their” program, presumably hoping to obscure the original source of the algorithms, ideas and the original code itself. The most serious allegations we have come across of this type relate to Rybka, currently the world’s top rated chess program and the winner of the World Computer Chess Championship in 2007, 2008, 2009, 2010. Rybka’s programmer is Vasik Rajlich, an International Master. For more than three years we have been hearing rumours in the computer chess world that Rybka’s engine was derived from the program Fruit, programmed by Fabien Letouzey, which placed second in the 13th World Computer Chess Championship in Reykjavik in 2005. Soon after his success in Reykjavik Fabien Letouzey made his program open source, under a Gnu Public License (GPL), so its copyright is now controlled by the Free Software Foundation.
In order to consider how the published Fruit source code might have influenced the development of Rybka, it is perhaps useful to examine some of the history of both programs. First let us go back a few years, to a time before the Fruit source code was made public. The Hiarcs forum contains the results of the CCCT6 tournament, played on January 31st and February 1st 2004, in which Rybka finished in 53rd place out of 54 contestants. On the Fruit Web site we find the following details of the open source versions of Fruit.
“It made its first appearance to the public in March 2004. Fruit was then just a basic program with a very simple evaluation and basic search. However since then it made skirmish progress adding about 100 Elo to each new release (1.5, 2.0, 2.1 and Fruit 2.2). The latest version from Fabien is “Fruit Beta 05/11/07″ compiled on November, the 3rd 2005. Since then no new versions where released.
Until Version 2.1, Fruit was open source. But with Fruit becoming the strongest engine, the author decided to close the source code to avoid clones which might participate in official tournaments.”
And furthermore, Fruit 2.1 was released with source code on June 17th 2005 under the GNU GPL license.
Let us now consider the point in time when it became clear that Rybka had become enormously strong. From Wikipedia we learn that:
“Vasik Rajlich started working on his chess program at the beginning of 2003. The first Rybka beta was released on December 2, 2005 . . . In December 2005, Rybka participated in the 15th International Paderborn Computer Chess Championship. Rybka won the tournament with a score of 5½ points out of 7, ahead of other engines such as Gandalf, Zappa, Spike, Shredder and Fruit.”
So Rybka’s first outstanding tournament success would seem to have been in December 2005, six months after the date of the release of the open source version of Fruit 2.1. One can understand from this coincidence of timing how many computer chess experts might have been led to think that Rybka’s development owed a considerable dept to the Fruit source code.
But as I have mentioned, at first the Rybka-Fruit case was mere rumour. More recently, however, these rumours have become firm allegations, made by expert chess programmers and supported by evidence which appears on the surface to be rather compelling, both in its nature and in its volume. At this point in time I do not intend to make any definitive statement of my own on these allegations, but will allow the reader to form their own opinion after reading the following.
First, here is a posting by Zach Wegner, who currently develops (with the full permission of Anthony Cozzie, the original Zappa programmer) an upgraded version of Zappa, the World Computer Chess Champion in 2005. Wegner participated in the 2010 World Computer Chess Championship with their program which is called Rondo.
Evaluation
Rybka’s evaluation has been the subject of much speculation ever since its appearance. Various theories have been put forth about the inner workings of the evaluation, but with the publication of Strelka, it was shown just how wrong everyone was. It is perhaps ironic that Rybka’s evaluation is its most similar part to Fruit; it contains, in my opinion, the most damning evidence of all.
General Differences
Simply put, Rybka’s evaluation is virtually identical to Fruit’s. There are a few important changes though, that should be kept in mind when viewing this analysis.
■Most obviously, the translation to Rybka’s bitboard data structures. In some instances, such as in the pawn evaluation, the bitboard version will behave slightly differently than the original. But the high-level functionality is always equivalent in these cases; the changes are brought about because of a more natural representation in bitboards, or for a slight speed gain. In other cases the code has been reorganized a bit; this should be seen more as an optimization than as a real change, since the end result is the same.
■All of the endgame and draw recognition logic in Fruit has been replaced by a large material table in Rybka. This serves mostly the same purpose as the material hash table in Fruit, since it has an evaluation and a flags field.
■All of the weights have been tuned. Due to the unnatural values of Rybka’s evaluation parameters, they were mostly likely tuned in some automated fashion. However, there are a few places where the origin of the values in Fruit is still apparent: piece square tables, passed pawn scores, and the flags in the material table.
Evaluation Detail
In this section, which we skip here for being slightly too technical, the author goes into more depth about the details of each aspect of the evaluations and their similarities and differences. You can read it in the PDF version of this article.
Responses from Vasik Rajlich
When it was suggested in 2007 in an Internet posting that Rybka was a clone of Fruit, Vasik Rajlich strongly denied it.
“Osipov’s speculation is not correct. Rybka is and always was completely original code, with the exception of various low-level snippets which are in the public domain.
Rybka’s scores are minimax score – they are propagated up the search tree. In principle, they should be from the tip of the PV, but because Rybka takes the PV from the hash table, this may not always be the case.
Re. depth, this is simply a tool to drive the iterative search. By conventional I mean ‘in the normal range’.
Vas”
Additionally, when the origins of Strelka became the subject of debate heated in the computer chess forums, Vasik pitched in with his own comments, claiming that Strelka was a clone of Rybka. Vasik posted the following on the Rybka forum.
By Vasik Rajlich Date 2008-01-11 12:26
I’ve taken a look this morning at the Strelka 2.0 sources. The picture is quite clear.
Vast sections of these sources started their life as a decompiled Rybka 1.0. The traces of this are everywhere. The board representation is identical, and all sorts of absolutely unique Rybka code methods, bitboard tricks and even exact data tables are used throughout. Significant portions of the search and evaluation logic are not fully disassembled – the author has left in hardcoded constants and used generic names (such as “PawnStruScore0″ & “PawnStruScore1″, “PassedPawnValue0″ through “PassedPawnValue7″, etc) which show that he hasn’t yet fully understood what is happening.
In some cases, these traces do also extend beyond the inner search and evaluation kernel. For instance, Rybka and Strelka are the only engines which I know about which don’t report “seldepth” and “hashfull”. Rybka’s UCI strings are used throughout.
The author did at first make attempts to hide the Rybka origins, for example by masking the table values in earlier Strelka versions. He also made significant attempts to improve the program. The attempts at improvement are not very original, but they are everywhere. They include PV collection, null verification (and in fact changes to the null implementation itself), some endgame drawishness heuristics, a handful of new evaluation term, a new approach to blending between opening and endgame eval terms, and so on. They also do include various structural changes, such as knight underpromotions, on-the-fly calculations of many tables, the setting of piece-square table values, etc. These changes are extensive and no doubt lead to differences in playing style and perhaps a useful engine for users to have, but they do not change the illegality of the code base.
In light of the above, I am claiming Strelka 2.0 as my own and will release it in the next few days under my own name. The name of the author with the pen name “Osipov” will be included if he comes forward with hiw own real name, otherwise an anonymous contribution will be noted. The contributions of Igor Korshunov will also be confirmed and noted if appropriate. All usage permissions will be granted with this release.
I do not see obvious signs of other code usage, but perhaps this deserves a closer look. Some of the transplanted ideas, such as the null verification search, are rather naive implementations of the approach in Fruit/Toga, although my first impression is that that code itself is original. The Winboard parser from Beowolf which was added to Strelka 1.0 seems to have been completely removed. If someone else does find other signs of code theft, please get in touch with me and I will give proper credit in the upcoming release.
If someone has suggestions about an appropriate license, and in particular the pros and cons of the GPL for a chess engine and for this unusual scenario, or if someone would be willing to help in preparing this code and license for release, please also get in touch with me.
As this code is two years and several hundred Elo old, I am not going to launch any major action. However, ‘Osipov’ has already threatened to repeat the procedure with Rybka 2.3.2a. (He did this after I declined to grant him rights to commercialize Strelka.) If this situation does repeat with a newer Rybka version, I will not just stand and watch any more. In the meantime, if someone has information about ‘Osipov’, please get in touch with me.
Furthermore, when I contacted Vasik a few days before writing this article, inviting him to comment on Zach Wegner’s analysis, he responded as follows:
“Hi David,
I’m not really sure what to say. The Rybka source code is original. I used lots of ideas from Fruit, as I have mentioned many times. Both Fruit and Rybka also use all sorts of common computer chess ideas.
Aside from that, this document is horribly bogus. All that “Rybka code” isn’t Rybka code, it’s just someone’s imagination.
Best regards,
Vas”
And when I asked for clarification as to whether this response meant that the Rybka 1 source code was original, Vasik replied:
“all of the Rybka versions are original, in the sense that I always wrote the source code myself (with the standard exceptions like various low-level snippets, magic numbers, etc).”
Fringe Problems
There is one other type of offence that I would like to mention here in connection with cloning, namely entering a cloned program created by someone other than the entrant, in a tournament, with the entrant knowing it be a clone. One might draw an analogy between the criminal law offence of theft and the crime of handling goods knowing them to be stolen. This offence in the computer chess world is similar to one that recently caused something of a scandal in the Netherlands, when a board member of the Dutch Computer Chess Association (CSVN), the body that organises the prestigious Leiden tournaments entered a pirated copy of Junior in one of the major online annual tournaments. (See here for more details.) Put simply, if someone knows that a program has been ripped off, either by cloning or through piracy, they will not be permitted to use a ripped off copy to compete in any ICGA event.