Author: blass uri
Date: 09:30:28 01/26/00
Go up one level in this thread
On January 25, 2000 at 16:27:11, Thorsten Czub wrote: >On January 25, 2000 at 01:06:05, Harald Faber wrote: >>Because you don't understand. Even Christophe explained that to you. > >even christophe is mistaken too. > >ok - i do explain it to you. > >you can stretch a feather, and it will react dynamically. >but than, suddenly, you come to a point of no return. >and than the feather breakes and the reaction is massive. > >or: you can drive on a motorway to the left, to the right, to the left, >from one side to another. now speed up. it will still work, >but be more difficult. suddenly, you come to the point of no return, >and the car drives out of the path, the dynamically movement >stops and develops into a chaotic movement. >you get a car crash. > >now take 2 chessplayers. they are arround 50 elo different. >the machines are similar. you get reasonable results. like the >car driving left and right. or the feather. > >now take 2 chessprograms. they are arround 100 elo different. >lets call them junior-shredder2. >they are running both on 450 mhz. > >junior kills shredder. but the car is still on track. it will not >crash. its within the range that produce sensible data. > >now take shredder2 from the 450 mhz and put it on >the 200 mhz. >now junior6 450Mhz fights vs. shredder2 200 Mhz. > >and suddenly. you come to a point where your results do not anymore >only show effects of the programs, as the car crash is not only >a reason of the car, or the feather, but of the overstressed >material breaking. > >suddenly the results become worthless. >like letting kasparov playing a 1500 ELO guy. >of course the 20 games match between kasparov produces measurable >data, as much as the car-crash produces measurable data. >but what both things connects is: > >2800 ELO vs. 1500 ELO produces not very sensible data. kasparov against 1500 elo is not the same because it is clear that kasparov is going to get 100% when Junior6(450 Mhz) does not get 100% against shredder2(200Mhz) I think that the problem is that the formulas that the elo is based on are wrong. shredder2 has no learning function and Junior6 has a learning function so the result is that playing more games between Junior6 and shredder2 increase Junior6's elo. The rating is dependent on the number of games that Junior6 and shredder2 played and if the ssdf decide to play 1000 games between them the result is probably going to be something like 970:30 and Junior is going to be number 1 not because of chess strength. The problem is that there are not clear rules(for example counting only the first 20 games for rating against programs with no learning function for rating). Another problem is that all the elo formula is wrong. It is based on some assumptions that are wrong. For example it does not consider the fact that there are players that are more stable Suppose player A gets 45% against 2600 players and get 55% and against 2400 players. players B get 30% against 2600 players and get 70% against 2400 players. both players deserve 2500 but A will get better rating if A plays against more 2600 players when B gets better rating when B plays against more 2400 players. If A plays against 1000 2600 players and against 500 2400 players A's rating is going to be overestimated by the rating system. It is practically impossible to find the best formula and nobody tried to find a better formula(I am sure that it is possible) because people want a simple formula and not a formula when you need a very complicated program to calculate the rating. Uri
This page took 0 seconds to execute
Last modified: Thu, 15 Apr 21 08:11:13 -0700
Current Computer Chess Club Forums at Talkchess. This site by Sean Mintz.