_________________
That which can be destroyed by the truth should be.
--
My (sadly neglected, but not forgotten) project: http://dailyjoseki.com

I can see nothing in WHR that makes any attempt to handle it actually... I'd be interested to see if I'm mistaken, because it is one of the big flaws I see in any "standard" algorithm.

I believe that allowing the players to have "time-varying strength" will naturally go a long ways towards fixing drift. I haven't tested it though. I could be wrong.

_________________
That which can be destroyed by the truth should be.
--
My (sadly neglected, but not forgotten) project: http://dailyjoseki.com

My understanding was that wouldn't easily "inject" points into the system though. It's just a clever way of making more adaptive rating volatility isn't it?

All rating systems have time varying strengths...that's what makes them rating systems. I assume that's not quite what you meant.

More seriously, all rating systems in general use are susceptible to drift. The probability of player A beating player B is expressed as a function of their rating difference, and that difference doesn't change if both players float up or down by a constant amount.

We understand this intuitively when we realize that different countries have different rating systems, and they all function just fine. There isn't a way to fix the strength of a shodan, so drift is always a problem.

I think there's some term confusion going on here-- by drift I'm specifically referring to the KGS phenomenon where you log in 5 months later to find yourself ranked 4d because everyone you played against earlier got stronger. I'm not referring to the fact that the rating system as a whole can get weaker or stronger. Maybe I should call it "KGS drift", hehe.

pwaldron, as far as I know, the math in most rating systems actually assumes that a player's strength does not change. This is why, for example, KGS weights old games less and less as they age. You get the illusion of change if you ask the algorithm "what am I rated now?" periodically and chart the results, because the algorithm's estimation of your true strength will vary with the data points it has seen. WHR is unique in that it assumes (and tries to calculate) everyone's rating for their whole history.

(An easy way to understand this: KGS adds one pixel to your graph each day with what it thinks your true rating is; it never changes yesterday's pixel. WHR would generate an entirely new graph for you every day.)

_________________
That which can be destroyed by the truth should be.
--
My (sadly neglected, but not forgotten) project: http://dailyjoseki.com

This is probably not actually the case. Even if it were, you need more than one anchor point to stabilize a ranking system.

_________________
That which can be destroyed by the truth should be.
--
My (sadly neglected, but not forgotten) project: http://dailyjoseki.com

The problem is that small errors accumulate and multiply the further you get from the anchor. I don't have a link off hand, sorry

_________________
That which can be destroyed by the truth should be.
--
My (sadly neglected, but not forgotten) project: http://dailyjoseki.com

As far as I understand, the current KGS algorithm determines a single current rank (plus uncertainty) from the whole history, weighing the games by currentness, for each player. A rank development occurs because old data are no longer being considered (or rather, considered less and less). In contrast, the WHR algorithm determines a complete time-dependent rank graph (plus uncertainty) for each player. A rank
development is thus completely modeled at each time.

It seems to me that this means that there is no longer a large drift, because the system understands that a player's rank is not constant.

Many thanks for pointing this out, daniel_the_smith. This is really a great improvement over all ranking and rating algorithms I have known so far. Is there already some work with respect to how handicap should be incorporated?

_________________
A good system naturally covers all corner cases without further effort.

It might work, for a few levels anyway. I'm not sure. Assuming you could get pros to play that many high-handicap games seriously.

@Harleqin, yeah, I think everyone should be using WHR.

_________________
That which can be destroyed by the truth should be.
--
My (sadly neglected, but not forgotten) project: http://dailyjoseki.com

WHR is pretty interesting, and I enjoyed skimming the paper at lunch. I haven't gone through enough detail to form a solid opinion on it, but so far it appears promising as something I'd like.

To date, I have preferred what the paper calls incremental systems, but perhaps WHR provides some good advantages.

This is a very fascinating paper. Thanks for it!

_________________
be immersed

I'm pretty sure I'm not the first person to link that paper from here (or maybe godiscussions), but I'm glad you're all enjoying it

_________________
That which can be destroyed by the truth should be.
--
My (sadly neglected, but not forgotten) project: http://dailyjoseki.com

Alternative:

Instead of issuing traditional amateur dan-certificates we issue pro-style achievement-based ranks for the game unique to this forum: Malkovich games. The rank would be in terms of "m" (1k, 1d, 1p, now... 1m!)

1m = Completed 1 Malkovich game (MG).
2m = Completed 10 MGs as a 1m OR won 5 MGs as a 1m
3m = Completed 20 MGs a 2m OR won 10 MGs as a 2m
4m = Completed 30 MGs as a 3m OR won 15 MGs as a 3m
5m = Completed 40 MGs as a 4m OR won 20 MGs as a 4m OR won a Malkovich tournament (MT) including at least 4 players
6m = Won two MTs including at least 4 players OR won a MT including at least 8 players OR 2nd place in a MT including at least 16 players
7m = Won two MTs including at least 8 players OR won a MT including at least 16 players OR 2nd place in a MT including at least 32 players
8m = Won two MTs including at least 16 players OR won a MT including at least 32 players OR 2nd place in a MT including at least 64 players
9m = Won two MTs including at least 32 players OR won a MT including at least 64 players

Requirements for promotion:
1. At least half the players in the tournament must be KGS SDK or stronger (this is a spectator sport!)
2. Minimum 25 comments per game or 0.25 commments/move for early resignations. (This is not DGS)
3. Tournaments can be of any format (tournament or league), but must be announced and must be open to all interested players. Round-robin leagues can have entry tournaments to limit the size. If the entry tournaments are rather competitive, perhaps the rating won by the league winner could be increased.
4. People who claim promotion post to a thread, and an admin then updates a current official rating list

Other:

• The ranks become hard to reach at the higher levels, but I think that makes it more fun.
• Great contributions to the forum or Malkovich games could be rewarded with a single promotion (up to 5d)
• Certificates should include Magicwand's avatar
• The cost of promotion is one go/bridge book or bottle of Hendrick's Gin sent to me in Minneapolis

_________________
I am Tartuffe from GoDiscussions.com

Magicwand is honorary 10m (ju-malkovich)

OK, lets do it. From now on, every Malkovich game finished will be added to the list and people will be getting promotions. We just have to find a person to do that

_________________