Wildclaw wrote:uPWarrior wrote:2 stone difference* means 50% win probability when the opponent has 2 stones on the board, 3 stones mean 50% when they have 3 stones on the board and so on.
Except that this is actually incorrect. A 2 stone handicap board is actually just 1.5 stones of advantage due to the lack of komi.
As you get closer to ranking up, you should be expected to perform 0.5 ranks better against an average even player and 1.0 ranks better in handicap games. Assuming the numbers on the
http://senseis.xmp.net/?KGSRatingMath are accurate, that means an expected 60-66% (5k-2d) win rate in even games or a 70-79% win rate in handicap games against weaker players.
What would probably make the KGS system feel a lot better is if it did handicap based on exact rating differences instead of rankings. If a strong 2 kyu player plays an average 2 kyu player, it should be a no komi game where the weaker player starts. If a strong 2 kyu plays a weak 3 kyu, it should be a 2 handicap stone game. And so on.
You could also make it even more detailed by mixing handicap stones and komi, but that is probably a bit to unorthodox for a lot of people.
That "*" after "2 stone difference" was there for a reason, I did mention the lack of symmetry by simply removing komi.
Anyway, you missed my point. The 60-66% or 70-79% win rates that go servers are enforcing on players 1 stone apart is pulled out of thin air. There is no mathematical or logical reason for that number to be 65% instead of 55% or 95%. As I said, ranks should be based on stone differences and not on win probabilities in even games. Go servers are optimizing for a metric that should be "learnt", not a constant of the algorithm.
Maybe this idea deserved a bigger and more detailed post as I have not seen this discussed anywhere else, but the way we are applying handicaps right now is totally disconnected to the way we are moving through ranks in go servers with a very large population. Most players are playing even games with each other (look at tygem, the percentage of even games is probably above 99%) and they are moving up and down on ranks based on win percentages in even games but there is no guarantee that players 4 stones apart win 50%* of their games with 4 stones handicap. THAT is what a 4 ranks difference is supposed to mean.
Just think about this hypothetical situation in your go club: a new player comes in and loses 4 games against a 5k and wins 1. Can you predict its rank in any way? Not until they adjust their handicap and get closer to 50%-50%. However, that is what go servers right now are doing, with the added constraint that even the 5k rank was calculated this way.
This is made worse by the fact that rank differences being disconnected from handicaps is self-reinforcing. Players stop playing handicap games because they feel their chances are not even and the risk on their ranking is not well adjusted.
