Martin Bryant opinion about Strelka issue

[fern]

Just in case some of you did not see it, Martin Bryant, legendary programmer of Colossus, recently upgrade in strength and to UCI, has seen the proves given by Uri and think strelka is, after all, a clone.
I had rejected the previous -weeks ago- bath of posts about that, but finally I must recognize that the saying of an expert as Martin has much weight. If it is so, we must also accept Uri Blas was right all the time.
I still hope there is only a high degree of similarity, a case not of cloning but of making use of similar or even identical ideas, as any programmer can do, But the doubt has been shed again.

What a pity regards
Fernando

[Steve Maughan]

I've followed this story from a distance and IMHO I don't think it's certain that Strelka is a clone of Rybka Beta 1.0. I acknowledge that there are some strange anomalies but:

• 1. The nps are different
  2. The PVs are different for most positions - normally longer for Strelka
  3. The exe is smaller
  4. Tactics are often solved at different plys - some faster some slower
  5. The original version was not UCI compatible whereas Rybka is UCI compatible

This means that if Strelka is indeed a clone it is not a normal clone. I'd define a normal clone, in the case of an engine that does not come with source code, as one that has simply been renamed using a hex editor. This means that if Strelka is derived from Rybka the author must have somehow obtained the source code, adjusted it and re-compiled it. Since Rybka's source is not public knowledge this is a serious criminal accusation i.e. theft of source code. I'd be extremely cautious of making such an accusation. The differences that I see are sufficient enough for me to give Strelka the benefit of the doubt.

Cheers,

Steve

[Michael Sherwin]

Clone? Well something is fishy about Strelka as Strelka does not even appear to be Strelka anymore! One third the exe size and double the node rate? I can not imagine how I could do that to my program and still have the same program.

Rybka has that huge ~1MB internal data pool that nobody knows what it is for, however, it must affect the eval or why have it. Now if Strelka does not have that ~1MB data pool then how is it evaluating the same as Rybka. There is a program that I will not mention the name of, that I believe is probably based on another engine, that artificially inflates the exe size with usless data to hide its true size. And also modifies its output to hide the fact that it is based on another program. Don't ask, I wont say!

There is a clone of my program RomiChess out there. It did not start out as a RomiChess clone and it is no longer a RomiChess clone, However, for a few versions it was a RomiChess clone and the reason that I discovered it was because I was testing RomiChess against it and kept getting confused by watching the games as to which side RomiChess was playing. It was like watching RomiChess play RomiChess. Then I noticed that the node rate was double what it was in earlier versions and later versions. It also had the exact same analysis bug and ponder bug that RomiChess had.

All this just makes me suspicious about Strelka, just suspicious, nothing more!

[Gandalf]

The author explains the decrease in size as:

transfer of the tables of estimations from the constants into hash, that will lead to the decrease of size of EXE- file

Though I am not quite sure what this is referring to...

[Michael Sherwin]

The author explains the decrease in size as:

transfer of the tables of estimations from the constants into hash, that will lead to the decrease of size of EXE- file

Though I am not quite sure what this is referring to...

The values must be stored somewhere. Is there now an external data file that comes with Strelka?

This gives us a clue as to how the ~1MB table was used. Positions that Rybka's evaluator does not evaluate correctly that are collected through time are hardwired into the exe and then moved to the hash at run time. Or, common pawn formations have a hash key that when combined with material (other than pawn) hashkeys gives a small value that is used in the evaluator. This last one, would greatly help Rybka play better when it first exits its book. Interesting speculation!

[Dr.Wael Deeb]

No Michael,it's only the exe file,nothing else....

[Michael Sherwin]

No Michael,it's only the exe file,nothing else....

Thanks Wael,

Then something is not right! According to the previous poster that I responded to the data was moved into the hash to reduce the exe size. The data still has to be somewhere if it is to be brought into the hash. Maybe the author of Strelka just decided that that data did not really do anything and decided to drop it. Something is not right.

[Uri Blass]

Clone? Well something is fishy about Strelka as Strelka does not even appear to be Strelka anymore! One third the exe size and double the node rate? I can not imagine how I could do that to my program and still have the same program.

Rybka has that huge ~1MB internal data pool that nobody knows what it is for, however, it must affect the eval or why have it. Now if Strelka does not have that ~1MB data pool then how is it evaluating the same as Rybka. There is a program that I will not mention the name of, that I believe is probably based on another engine, that artificially inflates the exe size with usless data to hide its true size. And also modifies its output to hide the fact that it is based on another program. Don't ask, I wont say!

There is a clone of my program RomiChess out there. It did not start out as a RomiChess clone and it is no longer a RomiChess clone, However, for a few versions it was a RomiChess clone and the reason that I discovered it was because I was testing RomiChess against it and kept getting confused by watching the games as to which side RomiChess was playing. It was like watching RomiChess play RomiChess. Then I noticed that the node rate was double what it was in earlier versions and later versions. It also had the exact same analysis bug and ponder bug that RomiChess had.

All this just makes me suspicious about Strelka, just suspicious, nothing more!

Vas explained in the past that the size of the exe file is because of some optimization.

It is possible that rybka has a big table of constants when strelka simply has a function to calculate the big table.

It is possible that the function to calculate the table caused strelka to be slightly slower and significantly smaller.

Uri

[MartinBryant]

Goodness me Fernando, you make me sound like judge, jury and executioner!
Let me re-iterate that I merely expressed my own personal opinion based on the arguments for and against and I cannot and do not claim that it holds more weight than anybody elses.
Also I believe that every person here has a right to express their opinion whichever way it is.

[Dann Corbit]

Martin is wrong.