H.G wrote:
So how can that be described as 'handicapping' or 'penalizing' Rybka? That seems silly in the extreme...
Not at all. It penalised the forward thinking and progressive engine authors by denying them the opportunity to showcase their efforts and skills, in this case cluster programming. OK for WCCC there was an open tournament as well, so they *did* get the chance, just not for the official WCCC title, but for Leiden for example, there is no such open tournament. So by setting a hardware limit you are denying them the opportunity to display their skills. It is like saying, "hey, you've spent a lot of time proramming a cluster implementation, but no-one else has been as done it, so we won't let you use it". An important principle, in this case it is cluster technology, but what if it was something else ? A very slippery slope, trying to bring everyone down to the lowest common denominator. Heck, maybe next year they will allow one core only, because Joker is not SMP, Ktulu if it competed is not, neither is say Onno. I don't see how the argument differs. Most people would throw their arms up in the air about this - a 1 core limit - but compare to an 8 core limit, they are both limits on cores, how is it different ? In principle it isn't. And don't tell me the 8-core limit was imposed as a hardware affordability thing with £5,000 W5580 machines being used.
Hehe, but now I beg to re-read what hgm wrote. Is there no sense for irony in computerchess? One of the reasons why I am so often misunderstood. BTW that is the reason why so many good computerchess authors abstain. Because it affords still another talent to communicate on the net. Perhaps this is why some need the eye tp eye contact in operator tournaments whiles others understand the always at least two meanings of a word...
I already have written it somewhere else. What if Vas created a code that tells the others a certain story and behind/besides it does something else? Would that be fraud or cheating? I dont think so in this world of plagiarism.
-Popper and Lakatos are good but I'm stuck on Leibowitz
H.G wrote:
So how can that be described as 'handicapping' or 'penalizing' Rybka? That seems silly in the extreme...
Not at all. It penalised the forward thinking and progressive engine authors by denying them the opportunity to showcase their efforts and skills, in this case cluster programming. OK for WCCC there was an open tournament as well, so they *did* get the chance, just not for the official WCCC title, but for Leiden for example, there is no such open tournament. So by setting a hardware limit you are denying them the opportunity to display their skills. It is like saying, "hey, you've spent a lot of time proramming a cluster implementation, but no-one else has been as done it, so we won't let you use it". An important principle, in this case it is cluster technology, but what if it was something else ? A very slippery slope, trying to bring everyone down to the lowest common denominator. Heck, maybe next year they will allow one core only, because Joker is not SMP, Ktulu if it competed is not, neither is say Onno. I don't see how the argument differs. Most people would throw their arms up in the air about this - a 1 core limit - but compare to an 8 core limit, they are both limits on cores, how is it different ? In principle it isn't. And don't tell me the 8-core limit was imposed as a hardware affordability thing with £5,000 W5580 machines being used.
Rybka is not the first program to use a cluster.
remember cluster toga.
I think that 1 core limit is going to increase Rybka's chances to win tournaments.
The point is that better hardware that both rybka and the opponent use increase the chances of the weaker program to draw the game.
Ryan Benitez wrote:Strelka is not a perfect clone so it is hard to know what was taken from Fruit that was not needed to replicate Rybka.
It's really very, very little. Multi-PV and Verified Null Move were added, and then some really low level stuff was implemented in a slightly different way. For Strelka 2 the algorithms from Fruit to create PSTs were added, whereas in Rybka the results were simply hard-coded (the PSTs are the same anyways). So in a case like that the same code is probably used, but the actual assembly from it isn't in the Rybka 1 binary. There are also a few things where Rybka 1 is closer to Fruit than Strelka.
Uri Blass wrote:If jealousy was the reason for the rule then the people who suggested it did a bad jon because I believe that they only increased rybka's chances to win.
This is the key point. The rules where changed in a way that greatly increased the probability that Rybka would win.
So how can that be described as 'handicapping' or 'penalizing' Rybka? That seems silly in the extreme...
I don't follow this logic. _everyone_ used 8 cores. Without the limit they could have used more. But using more than 8 cores is not an automatic thing, it takes some cleverness to get it to work. Probably some commercial engines do not do very well at > 8 nodes based on occasional test position timings that get reported here. And then there was Rybka with a cluster approach. Not very good, last time it was discussed. But certainly better than nothing. And that advantage was simply taken away.
So those that have lousy parallel searches benefit from restricted hardware.
Limiting the competition to 8 cores meant that Rybka didn't gain a huge advantage over the outdated professional programs, so in this sense, it meant a handicap for Rybka, that made it less likely to win.
Conversely, the 8 core limit also prevented any "surprise" entrants like Cluster Toga, Cluster Sjeng, Hydra, some program running on a supercomputer... Those might have been much more likely to challenge Rybka's supremacy than the outdated professionals. So in this case, limiting the tournament to 8 cores just meant Rybka became even more the favorite to win.
It's a question of handicapping compared to whom...
The point is that wthout the limit, one other program could also have used more cores. But not the others. So it would in effct have become a single-game tournament And the chances for Rybka to win that single game would be lower than the hances it would end ahead of the others in a gauntlet of many games. That is just a matter of reduced standard deviation in the larger umer of games.
I can't see any merit in Ray's argument. How would it help to "show your skills" if you ended only second? Whatever targets the Rybka team could have had in mind for this tournament, they were all subsidiary to winning it. o I don't think that any rue increasing that probability can be called a 'handicap'.
Interesting indeed, I hadn't quite thought of it like that.
Somebody ran a simulation of a gajillion american football seasons and found that the best team won the superbowl infrequently, around 24% of the time. Part of that is due to things that wouldn't affect a chess tourney, like the strengths of the other teams in the division. But it's obvious that with few games and some amount of parity among the upper echelons, the "strongest" will not win a computer chess tourney quite frequently as well.
So if we realign our perception so that the WCCC isn't about crowning the best, it's about crowning the champion, and stop thinking of the champion as "the best", then everything works. In football, this would be problematic because nobody can *really* run the same season a gajillion times, just a simulation of it. But in computer chess, the situation is different... which might explain the relative lack of interest in the results of the WCCC.
hgm wrote:The point is that wthout the limit, one other program could also have used more cores. But not the others. So it would in effct have become a single-game tournament And the chances for Rybka to win that single game would be lower than the hances it would end ahead of the others in a gauntlet of many games. That is just a matter of reduced standard deviation in the larger umer of games.
I can't see any merit in Ray's argument. How would it help to "show your skills" if you ended only second? Whatever targets the Rybka team could have had in mind for this tournament, they were all subsidiary to winning it. o I don't think that any rue increasing that probability can be called a 'handicap'.
I am not sure that only deep sjeng could use the cluster.
It is possible that other participants could also use a cluster but after learning that the important tournament is limited to 8 cores they decided not to work on it and work on improving the smp-code to have better performance with 8 cores.
The main reason that I think that limiting the tournament to 8 cores increase the chances of rybka is because it reduce the level of the tournament and it means smaller probability for draws and bigger probability that the better program wins.
MattieShoes wrote:Interesting indeed, I hadn't quite thought of it like that.
Somebody ran a simulation of a gajillion american football seasons and found that the best team won the superbowl infrequently, around 24% of the time. Part of that is due to things that wouldn't affect a chess tourney, like the strengths of the other teams in the division. But it's obvious that with few games and some amount of parity among the upper echelons, the "strongest" will not win a computer chess tourney quite frequently as well.
So if we realign our perception so that the WCCC isn't about crowning the best, it's about crowning the champion, and stop thinking of the champion as "the best", then everything works. In football, this would be problematic because nobody can *really* run the same season a gajillion times, just a simulation of it. But in computer chess, the situation is different... which might explain the relative lack of interest in the results of the WCCC.
I can only speak for myself, but I have _never_ considered the WCCC to crown "the best". Only "the best on that given day, for that small number of games..."
