hgm wrote:Hmm, I was using this table:
http://gobase.org/studying/articles/elo/
Should I multiply by a factor sqrt(2), because 200 is the standard deviation per player, and what you measure is the difference in strength of two players?
I do not know why their table differs from what I know. They may be using different scaling factors.
Anyway for chess there exists a very simple estimate for an otherwise complicated function which I do not even remember (full of logs and stuff you do no want to meet in person).
If the winning percentage for a program A against program B is between 20% and 80%, take this winning percentage expressed in percents, remove 50 from it, and mutiply the result by seven. You get a fairly good approximate of the elo difference elo(A)-elo(B).
The formula get too inaccurate outside the 20%-80% winning percentage range, so in this case do not use it (or if you do, please don't mention my name).
Examples:
1- program A wins by 60% against program B. Program A is ~70 elo points above program B.
2- program A does only 30% against program B. Program A is ~140 elo points below program B.
3- program A wins by 90% against program B. Play it on the safe side and just say that program A is more than 210 elo points better then program B (who cares anyway
).Please note that this formula doesn't say a word about error margins, which depend mainly on the number of games played, not on the winning percentage. For error margins I use precomputed tables.
// Christophe
