On Sun, Oct 31, 2010 at 3:58 PM, steve uurtamo <[email protected]> wrote:
> thanks for pointing this out again. > > if you think about it, it simply *cannot* be accurate. > Is that logic or is credulity speaking? > > imagine 12 guys in a room who are all better than everyone else at a > boardgame. so they get more ELO than everyone else. imagine also that > they are strictly ordered so that #1 loses to #2 very, very, very > infrequently, and so on. then #1 will have a crazy high ELO compared > with #12, but it says nothing about the percentage of the time that he > would beat #13, since #13 (presumably) almost never plays against #1 > in tournament settings. (since #1 is actually a computer and because > humans wouldn't enjoy being in money tournaments where the best that > they can hope for is 13th place). > > if computers got their ELO points only from ranking organizations by > playing in tournaments against people as well as each other, it might > be a different story. but (again, presumably) that will never happen. > > don, please correct me if i've misstated this. > I'm not sure this explains it fully, but there is definitely something going on. I have been told that some people on the chess servers are masters at manipulating their ratings by picking and choosing who they will play. I think the computer vs computer ratings are inflated somewhat and as I've estimated the true ELO is probably closer to 300, perhaps as high as 400 ELO. The truth of the matter is that humans rarely get a win against good programs except when heavy odds are given. This is purely anecdotal, so take it with a grain of salt, but I have been watching some of the top players at the senior chess championships in Italy, and I've been following along with Komodo (my own chess program which is around 5th or 6th in the rankings) and I am amazed at how many inaccurate moves the humans make and what they miss. We are talking about Grandmasters, although none of them are top world players it's still amazing. I would not stand a chance against these players but with the help of Komodo (and the fact that I am around 1900 ELO myself) I can be a critic and laugh at their moves. I'll see that Komodo likes a certain move and the score will jump to half a pawn. The human will play some other move and the score drops to below a pawn! In each game I have watched I have seen several missed opportunities and outright blunders. In general Komodo will have selected the right move when they differ, or else it doesn't matter. Usually when Komodo is up half a pawn, it is probably going to win the game. It was clear that Komodo (which is over 100 ELO weaker than Rybka) would have taken all of those players apart without much sweat. In fact I think it would win that tournament, if it were a human, given 1 or 2 seconds per move. Most of the weak moves it found almost instantly. Computers are not infallible however. There was one position where the player made an interesting move that Komodo did not like and only after 30 seconds or more did Komodo start to notice that it was actually ok. The human knew it was ok (but he thought a long time before making that move.) It was a move that gives up a pawn but for good compensation and only after a deep search did Komodo decide it was enough compensation. However, there was nothing wrong with holding the pawn either - so even at a fast time control Komodo would not have gone wrong there. Don > > s. > > On Sun, Oct 31, 2010 at 2:42 PM, Don Dailey <[email protected]> wrote: > > > > On Sun, Oct 31, 2010 at 2:30 PM, Jacques Basaldúa <[email protected]> > wrote: > >> > >> Don Dailey wrote: > >> > >>> To the best I can estimate it is something like 300 ELO > >>> now which means in a short match there is almost no chance > >> > >> > for the human. > >> > >> It is more than that, around 500. Only a match with a handheld > >> device could be seen as similar strength: > > > > I'm being ultra conservative when I say 300 and I consider it an absolute > > lower bound. The reason one must be careful is that there is some slop > in > > equating computers to humans. Computers have their own ratings lists > > which attempt to be in line with human ratings but it is widely believed > > that the top programs have some distortion - in other words they are a > bit > > over rated compared to humans. Programs that are a few hundred ELO > weaker > > are not over-rated but the top ones are. Computer vs Computer can > > produce these distortions. > > So 300 is the lowest figure I would be willing to stake my life on - but > I > > believe it's probably more like 400 or maybe even your 500 value. > Larry > > Kaufman has produced a formula that he thinks will give more accurate ELO > > ratings to the top machines, it basically compresses the ratings of > program > > with more than a certain ELO and I forget how it works. I don't think > it > > removes more than 100 ELO so maybe 400 is pretty close to the right value > - > > but in either case I think it's absolutely clear than computers are > > superior and it's not even close. > > Don > > > > > > > > > >> > >> Top humans are near 2800 (ratings.fide.com) > >> > >> Rank Name Title Country Rating Games B-Year > >> 1 Carlsen, Magnus g NOR 2826 0 1990 > >> 2 Topalov, Veselin g BUL 2803 0 1975 > >> 3 Anand, Viswanathan g IND 2800 0 > >> > >> Google for Rybka elo: > >> > >> On a phone: > >> > >> Pocket Rybka 3: 2869 ELO on PocketPC? 23 Aug 2008 > >> > >> 17.01.2009 Deep Rybka 3 x64 2GB Q6600 2.4 3227 > >> > >> I read somewhere (ICGA Journal I think) that it is over > >> 3300 now. > >> > >> So about 500 points over the top human. > >> > >> > >> Jacques. > >> > >> > >> > >> _______________________________________________ > >> Computer-go mailing list > >> [email protected] > >> http://dvandva.org/cgi-bin/mailman/listinfo/computer-go > > > > > > _______________________________________________ > > Computer-go mailing list > > [email protected] > > http://dvandva.org/cgi-bin/mailman/listinfo/computer-go > > > _______________________________________________ > Computer-go mailing list > [email protected] > http://dvandva.org/cgi-bin/mailman/listinfo/computer-go >
_______________________________________________ Computer-go mailing list [email protected] http://dvandva.org/cgi-bin/mailman/listinfo/computer-go
