Michael Williams wrote: > It is a very nice graph. I wish we could see the next 11 doublings.
You and me both! Just a couple of other comments: The graph was smoothed with gnuplot's smooth bezier function - but the raw graph looks very similar - just a little more jagged. If you straighten out the line - you get about 160 ELO per doubling, just looking at the graph. This is being a bit conservative and rounding down. I am pretty confident that you would continue to get well over 100 per doubling for many more doublings and that this curve would gradually taper off. I am also confident that if we could run this at 5 or 6 more doublings and play 9x9 matches and this could be done at a reasonable time control, the program would give high dan players a difficult time. This is one of those claims that sounds ludicrous to most players probably. But when chess programs were only 2000 ELO strength, projections were made about what it would take to play grandmaster strength. Those projections were laughed at because nobody believed such a silly thing could happen, but if anything the projections were conservative and by no means exaggerated. It actually happened very quickly due to Moores law. The programs responded dutifully to each new generation of computer with about 80 ELO per doubling or so. Computers are now our masters in chess - matches are only given now with handicaps so that the humans will have a chance. The big surprise is that a doubling is STILL worth about 60 ELO points, the curve seems to be tapering off but it's very gradual. I expect exactly the same in computer go. This assumes the laws of physics and our ingenuity can keep Moores law working for a few more doubling's! I also did enough of a study on 19x19 UCT GO programs to see that the improvement is substantial. It seems to be at least as much as in 9x9. I don't expect the 19x19 curve to taper off for a very long time and I am confident that if Moores law can hang on for just a few more years, we will also be seeing at least mid dan go programs playing 19x19 Go in a few years - assuming they are playing about 3 kyu now and don't improve. Of course a little ingenuity on our part could speed this up! - Don > > > Don Dailey wrote: >> I found the graph, but I can't find the data and the details, although >> it will be on one of the postings. I think this was at least a year >> ago, perhaps 2. >> Here is what I remember: >> >> I played 11 different levels, each a doubling of the previous. The >> weakest level I think was just 1024 play-outs. I ran the study for >> weeks in order to get substantial data points even from the highest >> levels. The highest level, took a significant time to play a single >> game, several times longer than the CGOS time control which was 10 >> minutes at the time. >> The conditions were CGOS 9x9 conditions - komi 7.5, and so on, just >> like CGOS 9x9. >> >> I actually tested 2 basic versions, one with heavy play-outs and one >> with light play-outs. The light play-out version basically plays >> random games. >> >> Both programs were reasonably strong UCT programs - versions of Lazarus >> which probably would play at least 2100 strength on my current computer >> on the current 5 minute server. >> See if this link works to see the graph: >> >> http://greencheeks.homelinux.org:8015/~drd/study.jpg >> The X axis represents the number of doublings and ELO ratings are on the >> Y axis. >> >> - Don >> >> >> >> >> >> >> Michael Williams wrote: >>> Don Dailey wrote: >>>> Mark, >>>> >>>> I wasn't stating a precise value for a doubling when I said 100 >>>> ELO. But it appears that it is actually worth a bit more than 100 >>>> ELO for a >>>> doubling. I did a massive study of this at one point a year or >>>> more ago with thousands of games with UCT based Lazarus program and >>>> the >>>> strength improvement per doubling was very clear and impressive. >>> Don, what komi did you use when you did that study? Looking in the >>> archives, all I can find is you saying that komi=9 is correct. So >>> does that mean 8.5 or 9.5? Or did you allow draws? >>> >>> _______________________________________________ >>> computer-go mailing list >>> [email protected] >>> http://www.computer-go.org/mailman/listinfo/computer-go/ >>> >> _______________________________________________ >> computer-go mailing list >> [email protected] >> http://www.computer-go.org/mailman/listinfo/computer-go/ >> > > _______________________________________________ > computer-go mailing list > [email protected] > http://www.computer-go.org/mailman/listinfo/computer-go/ > _______________________________________________ computer-go mailing list [email protected] http://www.computer-go.org/mailman/listinfo/computer-go/
