To Jeroen and interested minds, re. Tiger node count

[tiger]

Jeroen Noomen in another thread asked me about the node count displayed by Chess Tiger versions 12 to 15.

The thread in question has been partially deleted by the moderators before I had enough time to answer Jeroen's question, so I just start over here.

Historically, Jeroen has been Chess Tiger's book author for the aforementionned version of Chess Tiger, which is the chess engine I have written.

Jeroen is now working for the Rybka team. I have asked him if he could write another opening book for Chess Tiger some time ago, but as I understand he cannot do so anymore because of his exclusive links with the Rybka team.

Jeroen is a fine person and unlike (unfortunately) most people with whom I have been involved in computer chess, including many people posting here, I have meet him in person several times.

OK, I hope you don't mind all these boring historical details, but in view of the recent events here at CCC I thought it would be good to remind that we are real humans and that there are various kind of links between us. Also it is possible that some people participating in the recent discussions did not know much about Jeroen or about me.

So, I was about to write a very precise explanation about the node count in these versions of Chess Tiger, which include some revelations about an algorithm used in Chess Tiger that I have never seen described anywhere else. Describing this algorithm is absolutely necessary in order to answer Jeroen's question.

However now Jeroen is working for the Rybka team. So I was going to answer a question from the Rybka team, but on the other hand this team has never answered this very same question, which has been repeatedly asked since several months, even years.

I do not want to turn this into a childish game of "I asked you first so you answer first and then I answer".

So here is my proposition: I immediately answer Jeroen's question as precisely as possible (which involves revealing a proprietary algorithm used in Chess Tiger) if Vas answers as honestly and as precisely as possible the very same question, which is about Rybka 1.0 (no need to reveal anything about the more recent versions). As we have most of the source code of Rybka 1.0 already, his explanations can refer to this code, so they can be undiscutable.

I hope this can be considered as acceptable.



// Christophe

[chrisw]

Jeroen Noomen in another thread asked me about the node count displayed by Chess Tiger versions 12 to 15.

The thread in question has been partially deleted by the moderators before I had enough time to answer Jeroen's question, so I just start over here.

Historically, Jeroen has been Chess Tiger's book author for the aforementionned version of Chess Tiger, which is the chess engine I have written.

Jeroen is now working for the Rybka team. I have asked him if he could write another opening book for Chess Tiger some time ago, but as I understand he cannot do so anymore because of his exclusive links with the Rybka team.

Jeroen is a fine person and unlike (unfortunately) most people with whom I have been involved in computer chess, including many people posting here, I have meet him in person several times.

OK, I hope you don't mind all these boring historical details, but in view of the recent events here at CCC I thought it would be good to remind that we are real humans and that there are various kind of links between us. Also it is possible that some people participating in the recent discussions did not know much about Jeroen or about me.

So, I was about to write a very precise explanation about the node count in these versions of Chess Tiger, which include some revelations about an algorithm used in Chess Tiger that I have never seen described anywhere else. Describing this algorithm is absolutely necessary in order to answer Jeroen's question.

However now Jeroen is working for the Rybka team. So I was going to answer a question from the Rybka team, but on the other hand this team has never answered this very same question, which has been repeatedly asked since several months, even years.

I do not want to turn this into a childish game of "I asked you first so you answer first and then I answer".

So here is my proposition: I immediately answer Jeroen's question as precisely as possible (which involves revealing a proprietary algorithm used in Chess Tiger) if Vas answers as honestly and as precisely as possible the very same question, which is about Rybka 1.0 (no need to reveal anything about the more recent versions). As we have most of the source code of Rybka 1.0 already, his explanations can refer to this code, so they can be undiscutable.

I hope this can be considered as acceptable.



// Christophe

Actually, I don't feel you have to answer this question at all. I find it a little offensive that what must be private information from a prior collaboration is being threatened to be used against you by the same collaborator, especially now that he works with the 'other side' of the current argument.

Node counter, chess tiger, umpteen years ago? So what?

[kranium]

Chris W. is absolutely right here...and has expressed it perfectly.
I know Rolf is pressuring you, but you owe no explanantion.

[Rolf]

Chris W. is absolutely right here...and has expressed it perfectly.
I know Rolf is pressuring you, but you owe no explanantion.

I was convinced Jeroen had a good question. What I expected that either CT admitted or rejected that he also gave wrong numbers in display. That was only interesting in a relational sense. IMO all programmers can do what they want with their output. Only Bob, CT and Norm saw it different. What is correct asks the lay...

[tiger]

Chris W. is absolutely right here...and has expressed it perfectly.
I know Rolf is pressuring you, but you owe no explanantion.

I think it's not right for me to ask Vas any explanation if I cannot be held by the same standards and be asked the same thing by his team.

Not that I am overly concerned by this node count stuff. When I tried Rybka 1.0 it took me 5 minutes to see that something was obviously obfuscated, so I did not worry much about it after.

I think it has more to do with honesty. Someone in the Tiger team asks the Rybka team about the honesty of the node count. Then someone in the Rybka team asks the Tiger team for a similar explanation.

So OK. Let's go. I'm just waiting for a confirmation that we are going to play with the cards on the table.



// Christophe

[Rolf]

Chris W. is absolutely right here...and has expressed it perfectly.
I know Rolf is pressuring you, but you owe no explanantion.

I think it's not right for me to ask Vas any explanation if I cannot be held by the same standards and be asked the same thing by his team.

Not that I am overly concerned by this node count stuff. When I tried Rybka 1.0 it took me 5 minutes to see that something was obviously obfuscated, so I did not worry much about it after.

I think it has more to do with honesty. Someone in the Tiger team asks the Rybka team about the honesty of the node count. Then someone in the Rybka team asks the Tiger team for a similar explanation.

So OK. Let's go. I'm just waiting for a confirmation that we are going to play with the cards on the table.



// Christophe

Now you exaggerate again. Therefore I cant understand ChrisW.

Fact is Vas has never denied that he had different KN. So, it's not true what Christophe proposes. He isnt known for ALSO giving "false" KN. Vas already admitted itg. But CT in silence again. But he attacks Vas with Rybka. Questions over questions.

[bob]

Chris W. is absolutely right here...and has expressed it perfectly.
I know Rolf is pressuring you, but you owe no explanantion.

I was convinced Jeroen had a good question. What I expected that either CT admitted or rejected that he also gave wrong numbers in display. That was only interesting in a relational sense. IMO all programmers can do what they want with their output. Only Bob, CT and Norm saw it different. What is correct asks the lay...

Let's set the record straight. I never gave diddly-squat about Rybka's node count, obfuscated or not. I didn't discover the obfuscation. I didn't measure the obfuscation. I didn't verify the obfuscation. None of it mattered to me. So let's get that crap out of the sandbox before we play any more.

Someone asked a simple question "why would one do this?" and I gave a very specific and precise answer I didn't make any accusations. I didn't ask any questions. I didn't run any tests. I just explained _why_ one would want to obfuscate specific output of their engine, nothing more, nothing less. I do have the opinion that if one claims that they searched XXX nodes, then they ought to use the _standard_ definition of "node". Even if there is room to count things more than one way. Because this "room" is still a very small percentage of total nodes.

My only interest is in vocabulary and definitions. Everyone counts a ply the same way, with one known exception. Depth is a bit more variable since you could report it as minum full-width depth, average depth over all branches, or even as a multiple-parameter value like 10/15/30 for min, average and max. Nodes are precisely defined, and not just for computer chess, but for _all_ tree-searching applications including checkers, go, and other problems like the travelling salesman example. If we don't use a common vocabulary, with standard and accepted definitions, then we have problems communicating. Much like the spacecraft launched to either mars or some moon out that way, which landed 10 meters below the surface of the planet because one group was measuring in meters, the other was using feet. The way to avoid that is to use absolutely standard definitions. In math, everyone knows what "floor" means, and "ceil", and "1st derivative" We need the same preciseness here.

hope that clears up where _I_ stand on the issue.

[Rolf]

Chris W. is absolutely right here...and has expressed it perfectly.
I know Rolf is pressuring you, but you owe no explanantion.

I was convinced Jeroen had a good question. What I expected that either CT admitted or rejected that he also gave wrong numbers in display. That was only interesting in a relational sense. IMO all programmers can do what they want with their output. Only Bob, CT and Norm saw it different. What is correct asks the lay...

Let's set the record straight. I never gave diddly-squat about Rybka's node count, obfuscated or not. I didn't discover the obfuscation. I didn't measure the obfuscation. I didn't verify the obfuscation. None of it mattered to me. So let's get that crap out of the sandbox before we play any more.

Someone asked a simple question "why would one do this?" and I gave a very specific and precise answer I didn't make any accusations. I didn't ask any questions. I didn't run any tests. I just explained _why_ one would want to obfuscate specific output of their engine, nothing more, nothing less. I do have the opinion that if one claims that they searched XXX nodes, then they ought to use the _standard_ definition of "node". Even if there is room to count things more than one way. Because this "room" is still a very small percentage of total nodes.

My only interest is in vocabulary and definitions. Everyone counts a ply the same way, with one known exception. Depth is a bit more variable since you could report it as minum full-width depth, average depth over all branches, or even as a multiple-parameter value like 10/15/30 for min, average and max. Nodes are precisely defined, and not just for computer chess, but for _all_ tree-searching applications including checkers, go, and other problems like the travelling salesman example. If we don't use a common vocabulary, with standard and accepted definitions, then we have problems communicating. Much like the spacecraft launched to either mars or some moon out that way, which landed 10 meters below the surface of the planet because one group was measuring in meters, the other was using feet. The way to avoid that is to use absolutely standard definitions. In math, everyone knows what "floor" means, and "ceil", and "1st derivative" We need the same preciseness here.

hope that clears up where _I_ stand on the issue.

Some days ago I wrote something about military and secret services and you took that as a an absolute nonsense but I had something concrete in mind. And here I can work it in again.

I still think that in military research, academic research with military impact, research in highest interest of State raison, and for example computerchess you cant expect to reveil secrets of actual standings of the research. Just as examples. The military will never exactly tell you the absolute distance of weapons, of photographic resolution from the Space, of the existence of lethal weapons as such. In computerchess competition someone would be a fool if he gave away exact data.

Of course the tournament organisers could order a sort of doping test for all winners, but this would probably mean few participants because the revelation of exact secrets dont make sense for commercial entities.

In times of the leadership of SMK it was similar. In average tournaments Shredder could lose but in Championships with Stefan as operator Shredder won again. Reason: for such tournaments special - non-public, non-tested versions were in use.

Vas has some other tricks and all others have still others. No chance for reveiling standards for the eval output. And it also wouldnt make sense IMO because different versions of the same engine show a completely different eval score. Just look at Rybka 3 and then its sisters R3 Human and R3 Dynamic. MAnsari reported first results that Dynamic has better results in long time controld. For the same positions the different variants show very different evals! So I would expect that the user should never take numbers as sort of exact copies of human chess game commentary in Informator style and classical numbers for different standings.

[Rolf]

I want to give an interesting observation I have made on Playchess.com

In one of the transmitted GM games, I hope I dont confuse it with a position I saw here in CCC, it was chatted that the free pawn on b5 was evaluated with +14. by Rybka, I forgot what version and variant it was. Of course the comment came in that this is ridiculous and must be a bug.

For me it wasnt a bug, if it was displayed by the Human version what should mean that for a human player it's extremely difficult to defend against the threats with such a free Pawn.

I'm not a computerchess expert, but I for one can imagine that such evals help the engine to plan further play better as if it just had given a +.1 advantage compared with the former position when the free pawn wasnt yet on b5. Could that be commented by real experts like you, Bob?

The topic could say: it makes no sense to behave always in the objectively optimal way if the opponent isnt a Kramnik anyway. So that "speculative" play could be tried, like CSTal for instance. For a short period to find attacks against the King.

[Mike S.]

I'm just waiting for a confirmation that we are going to play with the cards on the table.

I am almost sure you are a better chess player than a poker player. Because with your first reply above, you have actually given the main answer already. Because if Tiger would "like anybody else" simply and normally count nodes which have an (almost) clear definition, as Bob Hyatt has explained, than there would be no need for talk like that...

But I think neither that this is a very interesting question. Also, I do not assume that Vas Rajlich is interested how Tiger counts nodes. But who knows. Anyway, I would not bet on it.

From the user viewpoint, it is of course desireable that any engine (but especially commercial engines) count and report nodes and kN/s in a standard way, if such a quasi or de facto standard exists. We don't want phantasy numbers. But I am afraid the opinions about if such a standard exists and/or if a programmer has to follow it, will be different among different persons.