[computer-go] A new pairing system idea for CGOS

[Don Dailey]

On Thu, 2006-10-05 at 23:29 -0400, Chris Fant wrote:
> Seems like for scheduling purposes you would take the ranking of each
> player and add some amount of noise to it.  That way, you are still
> using the ranking info in a meaningful way but also making it
> non-deterministic.

Yes, this is a good idea.  If I do the pairings as if they were swiss
tournaments I will do it this way.

A similar idea is to walk an initially  sorted list, randomly swapping
pairs, doing a few passes of this.    Each player does something like a
1 dimensional random walk.

- Don





> On 10/5/06, Don Dailey <drd at mit.edu> wrote:
> > I'm not sure all the people here appreciate the beauty of the swiss
> > system or how it actually works.  It has properties that really lend
> > itself to being a nice scheduling algorithm.
> >
> > One property that is very good is that you always play someone with
> > the same score as yourself.  If you've won 4 games in the tournament,
> > you will play someone else who has won 4 games.  If you have lost all
> > your games so far, you will play someone else who has lost all his
> > games.   That means you will gravitate towards players near your own
> > strength, a feature that CGOS aspires to.
> >
> > Normally, you always play the top half against the bottom half within
> > point groups.  The top player in the tournament would play the top
> > player in the bottom half in the first round.  After that, you will
> > essentially have 2 sub-tournaments if you think of this recursively.
> > A tournament of first round winners, and a tournament of first round
> > losers.  But in each sub-tournament you again pit the top half against
> > the bottom half.   The infinite series you mentioned.
> >
> > As someone pointed out, this is a deterministic algorithm.  I do have
> > some concerns about this when CGOS stays static for a long period
> > of time.  A simple solution is to start each tournament with random
> > rankings.
> >
> > This of course creates the possibility that you will get situations
> > like the top 2 players meeting in the first round.  We could take this a
> > step farther by randomizing the top and bottom half separately.  This
> > will effectively kill the determinism while preserving most of the
> > beneficial properties of correct rankings, but not completely.
> >
> > For instance you might still get a first round match between opponents
> > who are very close in strength, which is not what you want in the first
> > round.  A swiss is designed so that the distance in strength between
> > opponents should converge in the later rounds assuming no upsets.  Or
> > looking at if from the other side, it has the characteristic that your
> > greatest mis-matches occur in the first round, assuming no upsets.  With
> > each round, the amount of mismatch is cut in half.
> >
> > With correct pairings you would almost never get a situation where Mogo
> > plays random.   Even with randomized top and bottom halves this would
> > happen only occasionally,  almost soley based on Mogo losing it's first
> > round game to a player in the bottom half or random beating a player in
> > the top half, both events are extremely rare.
> >
> > - Don
> >
> >
> >
> >
> > On Thu, 2006-10-05 at 16:01 -0700, steve uurtamo wrote:
> > > > The Swiss mechanism won't be visible, it will happen
> > > > behind the scenes
> > > > and would be used as a pairing device only.    All
> > > > you would see is
> > > > opponent getting paired from round to round in a
> > > > fair way.
> > >
> > > honestly, i think that this is a great idea.
> > >
> > > it approximates the "infinite sequence of swiss
> > > tournaments", which means that it becomes a metric
> > > that is as good as the swiss tournament system.
> > >
> > > s.
> > >
> > > __________________________________________________
> > > Do You Yahoo!?
> > > Tired of spam?  Yahoo! Mail has the best spam protection around
> > > http://mail.yahoo.com
> >
> > _______________________________________________
> > computer-go mailing list
> > computer-go at computer-go.org
> > http://www.computer-go.org/mailman/listinfo/computer-go/
> >
> _______________________________________________
> computer-go mailing list
> computer-go at computer-go.org
> http://www.computer-go.org/mailman/listinfo/computer-go/