Komodo - Rybka in Danger?

[Graham Banks]

Again a cheap attempt for an advertisement of a commercial chess project which is a clear violation of the charter.....your post has been reported....
Dr.D

Flagrant commercial exhortation is openly encouraging people to purchase a particular product. I can't see this happening here Wael.

Cheers,
Graham.

[Dr.Wael Deeb]

Again a cheap attempt for an advertisement of a commercial chess project which is a clear violation of the charter.....your post has been reported....
Dr.D

Flagrant commercial exhortation is openly encouraging people to purchase a particular product. I can't see this happening here Wael.

Cheers,
Graham.

Aha,I see....so Vasik is giving it for free I assume
It's an online product for purchasing Graham....
Cheers,
Dr.D

[Graham Banks]

Again a cheap attempt for an advertisement of a commercial chess project which is a clear violation of the charter.....your post has been reported....
Dr.D

Flagrant commercial exhortation is openly encouraging people to purchase a particular product. I can't see this happening here Wael.

Cheers,
Graham.

Aha,I see....so Vasik is giving it for free I assume
It's an online product for purchasing Graham....
Cheers,
Dr.D

So are Shredder, Naum, Fritz and Hiarcs amongst others. We allow them to be discussed here also, but not openly promoted.
Please check your pms.

Cheers,
Graham.

[Dr.Wael Deeb]

Again a cheap attempt for an advertisement of a commercial chess project which is a clear violation of the charter.....your post has been reported....
Dr.D

Flagrant commercial exhortation is openly encouraging people to purchase a particular product. I can't see this happening here Wael.

Cheers,
Graham.

Aha,I see....so Vasik is giving it for free I assume
It's an online product for purchasing Graham....
Cheers,
Dr.D

So are Shredder, Naum, Fritz and Hiarcs amongst others. We allow them to be discussed here also, but not openly promoted.
Please check your pms.

Cheers,
Graham.

I checked my PMs,all clear now
Dr.D

[Graham Banks]

Please check your pms.

Cheers,
Graham.

I checked my PMs,all clear now
Dr.D

Okay. It's not a case of taking sides and I can assure you it's nothing personal. You should know that by now.

Cheers,
Graham.

[mariaclara]

Dr Deeb,

I am shocked you do not understand the situation.

Please slowly read and learn.

our Charter states:

4. Are not flagrant commercial exhortations

Now, Dr. Deeb , what does "flagrant exhortation " mean?

A poster can not post open and clear statements to buy or support "commercial chess products/engines/etc."

However,

Undercover, covert/subtle posts such as discussions/ debates about the legality/validity or superiority of commercial chess engines/products over free engines are allowed.

Do you get it Dr?

Commercial chess products agents can go around our Charter promoting their products by not posting openly "flagrantly exhorting" people to buy their chess engines.
( btw, I hope their agents do not get pro bono chess products like engines, etc.)

All they have to do is start a discussion accusing any engine for that matter, it is a clone of their commercial engine.

(This was one of the sales ads techniques they taught us in our advertising class.)

make a product/movie and or movie star controversial to attract attention.
ex: "Is it true George Clooney dated Angelina Jolie behind Brad's back?"
but behind it all is they are promoting the next movie of Jolie.
Remember all the controversy about Paris? Her publicist said,
"of course, those stories about Paris aren't true. but they make
for good publicity."

In this case, there is a subtle/subconscious message to the buyers, "ah so this commercial engine is good/brilliant/etc." go buy it.

Look at some reactions.

" Bumbletoe free? hmmmmmm............it's strong blabity-bla,
but my version 4444 is much stronger hehehehehe"

They are itching/drooling to buy version 4444.

[slobo]

Dr Deeb,

I am shocked you do not understand the situation.

Please slowly read and learn.

our Charter states:

4. Are not flagrant commercial exhortations

Now, Dr. Deeb , what does "flagrant exhortation " mean?

A poster can not post open and clear statements to buy or support "commercial chess products/engines/etc."

However,

Undercover, covert/subtle posts such as discussions/ debates about the legality/validity or superiority of commercial chess engines/products over free engines are allowed.

Do you get it Dr?

Commercial chess products agents can go around our Charter promoting their products by not posting openly "flagrantly exhorting" people to buy their chess engines.
( btw, I hope their agents do not get pro bono chess products like engines, etc.)

All they have to do is start a discussion accusing any engine for that matter, it is a clone of their commercial engine.

(This was one of the sales ads techniques they taught us in our advertising class.)

make a product/movie and or movie star controversial to attract attention.
ex: "Is it true George Clooney dated Angelina Jolie behind Brad's back?"
but behind it all is they are promoting the next movie of Jolie.
Remember all the controversy about Paris? Her publicist said,
"of course, those stories about Paris aren't true. but they make
for good publicity."

In this case, there is a subtle/subconscious message to the buyers, "ah so this commercial engine is good/brilliant/etc." go buy it.

Look at some reactions.

" Bumbletoe free? hmmmmmm............it's strong blabity-bla,
but my version 4444 is much stronger hehehehehe"

They are itching/drooling to buy version 4444.

Hi Maria Clara,

If the recent posts about the new and free Komodo engine are part of a commercial strategy related to how to sell better a future commercial version of the same engine, then you are certainly right.

It would be a direct link to breaking the CCC Charter.

[Dann Corbit]

I can't even read RobboLito code myself, I'm not a programmer. I based my conclusion on an email from a reliable independent chess programmer (no connection to Don, me, or Rybka) who sent this king table he either found in or derived from an early Ippo version. He had no access to Rybka code himself, he just sent this info along. I compared it with my files of the values sent to Vas for R3, and the resemblance was startling. I didn't ask for other tables to compare, one was enough for me. It was also obvious to me that using Ippo (or later Robbo) for analysis generally produced evals extremely close to R3 in most situations, with rare exceptions. This should be equally obvious to anyone who compares them in analyzing many positions.
I could post the king table I submitted, but as I said I have no proof that it is actually in R3, I only know that Vas told me any time he made even the slightest changes to my terms or values, so I have no personal doubt on this point. So I'll leave it to others to post the relevant section from the R3 code if they wish to do so.
I don't claim that Robbo has taken the entire Rybka eval. Rather it seems that some parts may have been too difficult to decipher and were left out. So of course I won't post the whole eval even if I could prove it to be in R3.

The danger of "one was enough for me" is to compare rybka 1 beta table values to fruit's. More than one is enough there. BTW that has been done with direct disassembly of the piece/square table values. I don't remember the specifics now, but there is some material value difference between the two that requires some sort of multiplier. For example, stockfish uses P=256, so to convert to my pc/sq table values, one would need to multiply mine by 2.56 or divide the stockfish values by the same constant. But that doesn't make the values different, obviously, just duplicated and then scaled properly. I agree that duplicate tables would be a problem. But in both cases here (fruit/rybka 1 and robo/rybka3).

Larry also said the following:

"It was also obvious to me that using Ippo (or later Robbo) for analysis generally produced evals extremely close to R3 in most situations"

I think that it is possible to prove similiarity in this case

take(rybka,robbo,stockfish,Naum,Shredder,toga) and give them to analyze many random positions at small depth(rybka can get depth 1 when other programs get depth that they have similiar strength to rybka depth 1)

If the difference between Rybka and Robbo is significantly smaller than the difference between other pairs then it is going to be an evidence against robbo.

Uri

And since robo* seems to be testing at rybka + 70 elo or so, it is quite likely that in many positions they will be almost identical since they are so strong, while in a few others, they will be quite different due to that 70 elo difference. It is not so easy to compare that way. Pick the right positions and Crafty will appear to be a perfect clone of Rybka.

CCRL has a ponderhit statistic. However, when we consider the addage "Great minds think alike..." it brings to mind that the very top level engines will have very similar analysis (assuming that they can get the right answer and that a position has a right answer). So I would expect the correlation for "cream of the crop" engines to be higher than random engine correlation so I am not sure what a high correlation will tell you.

[bob]

I can't even read RobboLito code myself, I'm not a programmer. I based my conclusion on an email from a reliable independent chess programmer (no connection to Don, me, or Rybka) who sent this king table he either found in or derived from an early Ippo version. He had no access to Rybka code himself, he just sent this info along. I compared it with my files of the values sent to Vas for R3, and the resemblance was startling. I didn't ask for other tables to compare, one was enough for me. It was also obvious to me that using Ippo (or later Robbo) for analysis generally produced evals extremely close to R3 in most situations, with rare exceptions. This should be equally obvious to anyone who compares them in analyzing many positions.
I could post the king table I submitted, but as I said I have no proof that it is actually in R3, I only know that Vas told me any time he made even the slightest changes to my terms or values, so I have no personal doubt on this point. So I'll leave it to others to post the relevant section from the R3 code if they wish to do so.
I don't claim that Robbo has taken the entire Rybka eval. Rather it seems that some parts may have been too difficult to decipher and were left out. So of course I won't post the whole eval even if I could prove it to be in R3.

The danger of "one was enough for me" is to compare rybka 1 beta table values to fruit's. More than one is enough there. BTW that has been done with direct disassembly of the piece/square table values. I don't remember the specifics now, but there is some material value difference between the two that requires some sort of multiplier. For example, stockfish uses P=256, so to convert to my pc/sq table values, one would need to multiply mine by 2.56 or divide the stockfish values by the same constant. But that doesn't make the values different, obviously, just duplicated and then scaled properly. I agree that duplicate tables would be a problem. But in both cases here (fruit/rybka 1 and robo/rybka3).

Larry also said the following:

"It was also obvious to me that using Ippo (or later Robbo) for analysis generally produced evals extremely close to R3 in most situations"

I think that it is possible to prove similiarity in this case

take(rybka,robbo,stockfish,Naum,Shredder,toga) and give them to analyze many random positions at small depth(rybka can get depth 1 when other programs get depth that they have similiar strength to rybka depth 1)

If the difference between Rybka and Robbo is significantly smaller than the difference between other pairs then it is going to be an evidence against robbo.

Uri

And since robo* seems to be testing at rybka + 70 elo or so, it is quite likely that in many positions they will be almost identical since they are so strong, while in a few others, they will be quite different due to that 70 elo difference. It is not so easy to compare that way. Pick the right positions and Crafty will appear to be a perfect clone of Rybka.

CCRL has a ponderhit statistic. However, when we consider the addage "Great minds think alike..." it brings to mind that the very top level engines will have very similar analysis (assuming that they can get the right answer and that a position has a right answer). So I would expect the correlation for "cream of the crop" engines to be higher than random engine correlation so I am not sure what a high correlation will tell you.

I think you could conclude that for decent engines, the higher the correlation, the stronger the group is overall since they are agreeing. If this is a tactical test, this might not be so convincing. But if there are lots of positional test cases to wade through, the more different programs agree on a single move, the more likely that move is to be the best objectively. There will always be some exceptions, since these are computer chess programs with known weaknesses. But overall...

[Rolf]

I can't even read RobboLito code myself, I'm not a programmer. I based my conclusion on an email from a reliable independent chess programmer (no connection to Don, me, or Rybka) who sent this king table he either found in or derived from an early Ippo version. He had no access to Rybka code himself, he just sent this info along. I compared it with my files of the values sent to Vas for R3, and the resemblance was startling. I didn't ask for other tables to compare, one was enough for me. It was also obvious to me that using Ippo (or later Robbo) for analysis generally produced evals extremely close to R3 in most situations, with rare exceptions. This should be equally obvious to anyone who compares them in analyzing many positions.
I could post the king table I submitted, but as I said I have no proof that it is actually in R3, I only know that Vas told me any time he made even the slightest changes to my terms or values, so I have no personal doubt on this point. So I'll leave it to others to post the relevant section from the R3 code if they wish to do so.
I don't claim that Robbo has taken the entire Rybka eval. Rather it seems that some parts may have been too difficult to decipher and were left out. So of course I won't post the whole eval even if I could prove it to be in R3.

The danger of "one was enough for me" is to compare rybka 1 beta table values to fruit's. More than one is enough there. BTW that has been done with direct disassembly of the piece/square table values. I don't remember the specifics now, but there is some material value difference between the two that requires some sort of multiplier. For example, stockfish uses P=256, so to convert to my pc/sq table values, one would need to multiply mine by 2.56 or divide the stockfish values by the same constant. But that doesn't make the values different, obviously, just duplicated and then scaled properly. I agree that duplicate tables would be a problem. But in both cases here (fruit/rybka 1 and robo/rybka3).

Larry also said the following:

"It was also obvious to me that using Ippo (or later Robbo) for analysis generally produced evals extremely close to R3 in most situations"

I think that it is possible to prove similiarity in this case

take(rybka,robbo,stockfish,Naum,Shredder,toga) and give them to analyze many random positions at small depth(rybka can get depth 1 when other programs get depth that they have similiar strength to rybka depth 1)

If the difference between Rybka and Robbo is significantly smaller than the difference between other pairs then it is going to be an evidence against robbo.

Uri

And since robo* seems to be testing at rybka + 70 elo or so, it is quite likely that in many positions they will be almost identical since they are so strong, while in a few others, they will be quite different due to that 70 elo difference. It is not so easy to compare that way. Pick the right positions and Crafty will appear to be a perfect clone of Rybka.

CCRL has a ponderhit statistic. However, when we consider the addage "Great minds think alike..." it brings to mind that the very top level engines will have very similar analysis (assuming that they can get the right answer and that a position has a right answer). So I would expect the correlation for "cream of the crop" engines to be higher than random engine correlation so I am not sure what a high correlation will tell you.

I wished to see the topic discussed, how strength as such is understood in the actual situation of computerchess. How is it possible to substantiate differences in strength (what does it mean exactly?) if a simple correlation analysis shows no differences for clone similarities and strength similarities? A topic that interests me for a decade now, do we have a definition of strength for computerchess or is the assumption that we prolonge the strength of human GM and then assume that a machine program is 300 points ahead of the best humans which is something I would strongly refuse to accept.