Self-play series are a standard technique in the testing of game programs. X plays against itself with different resources (search depth d vs. d+1, or thinking time t vs. 2t, or n vs. 2n processors). Often, testers have observed “rather constant” winning quota for large parameter ranges. Also often, diminishing returns are found at the upper end: The side with larger resources is no longer able to keep the winning quota on the previous high level.
I found something new: At (very) small resources the winning quota may also be closer to the 50% level. So in total, for the side with fewer resources the performance curve looks like a basin. For pure Monte Carlo game search (resource parameter = number of random games per move) I found this basin structure in games like “Double Step Races”, Clobber, ConHex, “Fox versus Hounds”, “EinStein wurfelt nicht”. My very long test series give statistical significance. I have made a report with my findings public. It can be downloaded from http://www.althofer.de/monte-carlo-basins-althoefer.pdf Feedback is welcome. Ingo. -- GRATIS: Spider-Man 1-3 sowie 300 weitere Videos! Jetzt freischalten! http://portal.gmx.net/de/go/maxdome _______________________________________________ Computer-go mailing list [email protected] http://dvandva.org/cgi-bin/mailman/listinfo/computer-go
