Don Dailey: <[email protected]>: >On Wed, Jun 15, 2011 at 9:54 PM, Hideki Kato <[email protected]> wrote: > >> Jacques Basaldúa: <[email protected]>: >> >> >Because people here are used to see programs playing >> >4k with just 2500 sims (like Aya) you may not realize >> >how hard it is until you write your own. With long >> >times you start seeing single digit kyu programs play >> >really good moves. Magnus also commented in the first >> >slow bot tournament that he was very happy how well >> >Valkyria was playing. I don't seen why Zen would be an >> >exception. >> >> That's simple (at least I think). The scaling law converges at some >> point where the speed (or more thinking time) benefits little. > > >It converges at perfect play. Thinking that it just happens to converge at >the exact limits of current hardware is sort of like thinking the earth is >the center of the universe just because this is where we happen to be right >now. In 20 years with hardware 100x faster or more (assuming that >happens) I'm sure people will also be saying that we have reached >convergence. I saw this happen over a period of about 20 years in >computer chess, it was always the same, computers have pretty much >reached their limit and we need to try something else. Eventually people >wised up, but it took a while. Now it starts all over again with Go. >I'm not trying to put you on the spot, but I'm trying to save you from >looking stupid in 10 years should people dig up these old posts.
Why do you think (or believe) so? AFAIK, there is no evidence. I don't say the scaling completely stops but the number, about _100_ Elo for doubling the thinking time, gets much smaller (or, more likely, no more linearly scales to the logarithm of the time, suggesting other mechanisms behind). The global convergence theorem of UCT algorithm is independent here because the theorem tells us nothing about the number of simulations (equivalent to the thinking time) to converge to an optimal move. We now know almost nothing why the Elo rating of a MCTS bot linearly scales with the logarithm of the playouts for a move. Just the variance of the stocastic variable reduces by square root of the number. We'd better to investigate the reason first, I think. The scaling may converge by several reasons. There are so many complex positions that Zen's simulation cannot manage correctly. If the average of the results of the simulations is wrong, the scaling law cannot be kept. With my early experiments, it's clear that the convergence point heavily depends on (the skill of?) the opponents. The nakade problem is a good example I've explained in other posts. >> Zen19D >> already reached that point and Zen19S's rank is lower than Zen19D's. >> > > >> >> Facts: Zen19D and Zen19S share the same binary (version 7.7d9) and >> hardware (6-pc cluster; 26 cores total). Only the time setting is >> different. Zen19D and Zen19S play every move in 13 and 28 second, >> respectively (2 second is a margin for the network delay). Human >> players have 9 x 15 second with Zen19D and 20 minutes + 5 x 30 second >> with Zen19S. Recent ranks on KGS are 5.4 dan and 4.5 dan for Zen19D and >> Zen19S, respectively. >> >> Additionally, since Zen19 (version 7.7) is ranked 4.4 dan with the same >> time setting as Zen19D, 8-core Xeon seems not fast enough to reach that >> point. >> > >So basically what you are showing here is that when both computer and human >play faster, the computer benefits more. > >What is 9 x 15? And what is 20 minutes + 5 x 30, I don't understand the >time controls. How much time is the human getting and how much time is >the computer getting? I assume the computer is getting 13 and 28 right? Those are the style used in KGS. "9 x 15 second" gives human players 15 second for every move and extra 15 second 9 times. "20 minutes + 5 x 30 second" gives players 20 minutes as the main time and after the main time has expired players have to play a move in 30 second but are allowed extra thinking time, 30 second x 5 times. > So the extra time for the computer is just over 2X which isn't that much >perhaps 1/2 dan. But can you estimate how much time on average the human >is getting against each opponent? Because I don't understand the notation >9x15 etc. Hm, that's possible if human players' performance gets 1.5 dan (in average) by 15 second to 20 min and 30 second byoyomi. I have no idea if that's possible because I'm far weak from 4 or 5 dan on KGS. #On my rank (2-3 kyu), I guess up to 1 kyu boost might be possible. >And also what are the error margins? How many relevant games to achieve >these ratings? The margin is just the explanation of the difference between 15 second (allowed) and 13 second (actually used). 2 seconds are left to avoid time-out caused by an communication delay over Internet. Hideki >Don > > > >> >> Hideki >> -- >> Hideki Kato <mailto:[email protected]> >> _______________________________________________ >> Computer-go mailing list >> [email protected] >> http://dvandva.org/cgi-bin/mailman/listinfo/computer-go >> >---- inline file >_______________________________________________ >Computer-go mailing list >[email protected] >http://dvandva.org/cgi-bin/mailman/listinfo/computer-go -- Hideki Kato <mailto:[email protected]> _______________________________________________ Computer-go mailing list [email protected] http://dvandva.org/cgi-bin/mailman/listinfo/computer-go
