On Tue, 2006-12-05 at 20:40 +0100, [EMAIL PROTECTED] wrote: > > On Tue, 2006-12-05 at 09:51 -0800, steve uurtamo wrote: > > > > I'll bet Mogo would give a dan level player fits at > > > > 9x9 if 1 week of > > > > thinking time per move could be compressed enough to > > > > play a 30 minute > > > > game. > > > > > > you could always get a dan player to volunteer for > > > such a game. he would promise not to spend more > > > than 1/2 hour on the game, and mogo would play > > > postally. > > > > > > i'd be very impressed if mogo could "give him fits". > > > > > What do you mean by dan level? I don't mean "high dan", I mean 1 dan. > > So that I can follow this discussion, how would be the "kgs level" of this > player (it is the only level I have access to when looking at the results of > game)?
Wouldn't it be 1 dan on KGS? Maybe the group can help me extrapolate. What estimated (or actual) rating would gnugo 3.7.4 have on KGS? Whatever rating that is, would it play better or worse at 9x9? In other words, if it were EQUAL to a 15 kyu player at 19x19 for example, would it be better or worse at 9x9? On CGOS, gnugu_3.7.4 is rated 1720. MoGo is is in the 2100-2200, presumably it's save to assume it is significantly stronger. But if we can assign an ELO rating to Gnugo 3.7.4, then we can add 300 or 400 to get Mogo's current ELO rating. We will then know approximately how to estimate what a 1 Dan players ELO would be (once we know about what kyu level gnugo 3.7.4 is) by extrapolation. (About 100 ELO per kyu as I understand it.) Then if we estimate the ELO of a 1 Dan player, we can estimate what percentage of games Mogo would win now, and how much it needs to improve to be equal. This would all be somewhat speculative of course but it would give us a rough idea of where we are. I would be willing to be conservative on all the calculations - by assuming Mogo isn't as strong as we think, a doubling in time does not add as much ELO strength as we know it does, etc. We would have to estimate the improvement expected from extra thinking time. The Mogo team probably already has estimates, but I have my own too based on experiments with Lazarus. My studies show that it's about the same as chess used to be, about 100 ELO points for a speed doubling. This may not be accurate for higher levels - I don't really know for sure. I would be willing to assume only a modest 50 points per doubling, to get what I would consider a very safe lower bound. If Mogo is averaging approximately 10 seconds per game on CGOS to achieve 2100, then I've estimated about 16 doublings of speed if it's been given 1 week per move. That is 800 ELO points even using very conservative calculations. Yes, I know a lot of this is supposition but I'm confident that the program given a hypothetical 1 week per move would be pretty impressive. I don't feel these calculations are unreasonable or way out of the question - we have seen that UCT is very nicely scalable. I have not forgotten the computer chess lessons of the past either, where 1 ply gave 250 ELO points but everyone said that formula would not apply to the next "ply". People were silly enough back then to believe that after 6 ply searches going beyond wouldn't help any (or very little.) - Don > > Mogo would also have a memory problem. The UCT programs build trees in > > memory. My own program cannot think more than a few minutes without > > running out of memory - so even the experiment you propose cannot be > > done. > Yes you are right. For MoGo it is even worst. As all the games we have to > play > in are short games, we did not at all "optimize" the memory use, and in fact > MoGo a lot more memory than necessary. I am not totally sure, but I think > even 1 minute/move is already too much :). Of course, we can be more carefull > with memory usage, but for the moment it is not the higher priority. > > > In fact, in honor of Chrilly's laws I will call this "Don's law". > > "What really counts is how bad your bad moves are." > > I think this is a interesting point. I think I agree with that law :). > > Sylvain > > _______________________________________________ > computer-go mailing list > computer-go@computer-go.org > http://www.computer-go.org/mailman/listinfo/computer-go/ _______________________________________________ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/