Re: Go program rating

[Jean-loup Gailly <jloup@xxxxxxxxxxxxxx>]

Martin Mueller writes:

> It would be interesting to put all tournament games played between Go
> programs into one of the ratings programs, as used by AGA or EGF, and see
> what comes out. Does anyone have such a ratings program, and wants to input
> the data for the last couple years? It might make sense to preassign some
> start rating for the long-time competitors.

This preassignment is the problem. Since programs play tournament games
mostly between themselves (I'm not aware of programs entered in human
tournaments), it is difficult to assign them an absolute level meaningful
on a human scale. You are around 5d I think, and you can beat Many Faces at
29 stones(*). Then should Many Faces, one of the long-time competitors, be
pre-assigned a rating of 25k?  That wouldn't be fair because Many Faces
would beat most 25k humans in normal tournament conditions. Should Handtalk
be pre-assigned a rating of 3k? That wouldn't be fair either because 3k
(EGF scale) humans would easily beat Handtalk in normal tournament
conditions. Ideally, programs should be entered in human tournaments to get
a realistic estimate. Games played by Many Faces on IGS or NNGS are not
made in tournament conditions and should not count.

Jean-loup

(*) Game record posted in this list and available in
    http://www.teaser.fr/~jlgailly/go.html