If we are only looking for the shape of the curve, then maybe each datapoint does not need to be as precise as +/- 10 ELO.
On Fri, Jun 17, 2011 at 9:31 AM, "Ingo Althöfer" <[email protected]>wrote: > > -------- Original-Nachricht -------- > > Datum: Fri, 17 Jun 2011 11:58:45 +0100 > > > Von: "Jacques Basaldúa" <[email protected]> > > ... > > When I was experimenting with learning playout weights > > using GAs (something abandoned I have something much > > better in progress) I found easy to make a chain where: > > > > B is 50 Elo point stronger than A > > C is 50 Elo point stronger than B, D than C, E than D > > > > And when you confront E vs A and expect the difference > > to to be 200-ish it is only 30 points. > > I am not a go programmer, but in experiments with very > simple model games I found the following phenomenon, > when looking at the "basic" Monte Carlo algorithm in > 3 levels: MC(1), MC(2), MC(4). > [Here MC(k) is the version which makes k random games for each candidate > move.] > > I found several cases (more than 30 % of the games analysed this way) > with the following phenomenon: > Elo difference between MC(1) and MC(2) = k(1,2) > Elo difference between MC(2) and MC(4) = k(2,4). > Elo difference between MC(1) and MC(4) > k(1,2) + k(2,4) > > The games under investigation were so simple that I was able to run > 100,000 matches in selfplay (or even more), and the inequality above > ( k(1,4) > k(1,2) + k(2,4) ) was statistically significant. > > Ingo > > PS. In this context I would also like to mention (once more) my > technical report on the basin structures in self-play. See at > http://www.althofer.de/monte-carlo-basins-althoefer.pdf > So, the winning quota at double resources does not behave monotonically > (for very quick searches). > -- > NEU: FreePhone - kostenlos mobil telefonieren! > Jetzt informieren: http://www.gmx.net/de/go/freephone > _______________________________________________ > Computer-go mailing list > [email protected] > http://dvandva.org/cgi-bin/mailman/listinfo/computer-go >
_______________________________________________ Computer-go mailing list [email protected] http://dvandva.org/cgi-bin/mailman/listinfo/computer-go
