2009/6/24 Christian Nentwich <[email protected]> > Don, > > you might have your work cut out if you try to control inflation directly, > that can turn into a black art very quickly. Multiple anchors would be > preferable. An offline, X * 1000 game playoff between gnugo and another > candidate anchor would be enough to fix their rating difference. If their > bilateral winnings drift away during continuous play, the anchor rating > could be tweaked. >
It's all a black art anyway. The anchor itself absorbs (or gives away) rating points into the pool. There is not much difference if I just use it to monitor the inflation/deflation and control it directly - except that I have the ability to control the magnitude of this adjustment. And the advantage is that the anchor player becomes a monitor of the inflation level. Don't worry, I don't plan to change it from what I'm doing. The anchor can still monitor inflation if I track what adjustment I would normally make to it if it were not an anchor. For instance if the opponent adjustments tended to be more negative than positive it would indicate that the entire pool was overrated. A way to help compensate is to adjust the initial rating of new players. However, the first game against a brand new player is not rated for the established player and the K constant is so low (for the new players opponents) that it hardly matters. Each player starts with a high K and it gradually drops to 3. But this K is modified from 0% to 100% depending on the opponents K - so the first game against a player a new player is effectively not rated for his opponent. But I think the initial value does have an impact on deflation/inflation of the entire pool. > > > I'm not sure if the worries voiced on this list about anchors are not > somewhat overdone. > I'm pretty sure it's overdone, but I reserve judgment. I know the phenomenon of self-play intransitivity exists, but it's minor. This is something that can easily be tested privately with a 100,000 games or so to get the amount nailed down - at least for specific trio's of players. I think I may try gnugo vs fuego at 2 different levels. I think that MCTS are all similar and that this is the bigger issue. And as you say, gnugo introduces bias too, it's unavoidable. > Other bots, with improvements, may do just as well against an old version > of Fuego as the full Fuego does, we don't know. Maybe they would do better > than new versions of Fuego. All this reliance on gnugo introduces bias, too, > and after all the anchor player is not a single control variable that > determines the destiny of the server. Players will still play many different > opponents. If Fuego keeps beating the anchor player but losing to everybody > else, it still won't get a higher rank. > > For me, gnugo as an anchor is fine, as I am still experimenting around a > low ELO level. For authors of strong programs: I am quite surprised that you > are not insisting on a much more highly rated anchor. I remember when KGS > was anchored in the kyu ranks, many years ago. I found myself 7 dan one day, > until somebody intervened and reanchored the server. The territory far above > a single anchor player is unsafe. > The thought has occured to me that I should not worry about low resource anchors and that I should simply bite the bullet and insist, as you say, on much stronger anchor players. But the tone of these discussions indicate that few consider that very important. I'm glad to hear that I am not the only one. If I did do this it would not need to disrupt the pool - I would still run the standard gnugo player that I currently use as an anchor and use it as a way to monitor the "new" anchor - at least for the first 100,000 games of the new anchor. I have no problem using programs under heavy development either. What people are missing is that I don't use the latest version, I simply pick a good version and stick with that. For instance I do not upgrade gnugo - I continue to use the same version I started with. So the anchor is not continuously improving - it is a constant. - Don > > > Christian > > > > > On 24/06/2009 05:28, Don Dailey wrote: > > >From what I have discovered so far, there is no compelling reason to > change anchors. What I really was hoping we could do is UPGRADE the > anchor, since many programs are now far stronger than 1800. > > Fuego is pretty strong, but not when it plays at the same CPU intensity as > gnugo. I went up to 5000 simulations and the match is fairly close and the > time is about the same. Going from 3000 to 5000 was quite a remarkable > jump in strength and no doubt we could run at 10,000 and have substantial > superiority - but that's not really what I had in mind. > > So I think I agree with all the comments I have received so far - and my > own observations and testing, there is no compelling reasons to change. > > Now if fuego was substantially stronger using less resources, I would be > more eager to change after carefully calibrating the difference, but that > is not the case, at least not at 19x19. > > There is another way to keep ratings stable and that is to monitor key > players over time and build a deflation/inflation mechanism into the server > to keep it in tune. For instance if there were no anchors, the server > could monitor gnugo and if it were to gradually drop in rating, I could make > minor adjustments to the ratings of winners and losers to compensate > gradually over time. For example the winner could get 1% more ELO and the > loser could lose 1% less ELO when in inflation mode and just the opposite > when in deflation mode. In this way I could feed points into the rating > pool, or gradually extract them as needed. I don't plan to do this, but > there is more than one way to skin a cat. > > If we use more than one player as anchors, I would still pick one player > as the standard, and periodically adjust the "other" anchors based on their > global perormance rating - since they will all tend to drift around relative > to each other and I would not want to make any assumptions about what the > other anchors should be. We cannot just say gnugo is 1800, fuego is > 2000, etc because we don't really know the exact difference between the 2. > But we could refine this over time. > > - Don > > > > > > On Tue, Jun 23, 2009 at 11:34 PM, David Fotland > <[email protected]>wrote: > >> I'd also prefer to use gnugo as an anchor. Since fuego is under >> development, new versions will be playing with an odler version of itself. >> Fuego will win more often against its old version. I don't care about it >> distorting Fuego's rating, but it will distort the rating system. If new >> fuego is on with few other programs it will gain rating points, then when >> other programs come new fuego will give them the other program as its >> rating >> drops. The effect will be to make the rating system less stable, so it's >> hard to use cgos to evaluate new versions of programs to see if they are >> stronger. >> >> I think it's best to use an anchor that's not under active development. I >> like gnugo since there is lots of published results against it, and it is >> not changing rapidly. Also it has a different style than the monte carlo >> programs, so it's more likely to expose bugs in the monte carlo programs. >> >> David >> >> > -----Original Message----- >> > From: [email protected] [mailto:computer-go- >> > [email protected]] On Behalf Of Hideki Kato >> > Sent: Tuesday, June 23, 2009 5:15 PM >> > To: computer-go >> > Subject: [computer-go] Re: fuego strength >> > >> > I'm running Fatman1, GNU Go and GNU Go MC version for 9x9 and two >> > instances of GNU Go for 13x13, five programs in total, on a dual-core >> > Athlon at home. >> > >> > I strongly believe current anchors are resource friendly enough for >> > older pentium 3, 4 or even Celeron processors and not necessary being >> > changed. >> > >> > Changing anchors is a big problem, similar to changing the >> > International prototypes. Also, GNU Go is used as a reference in >> > almost every computer-go research these days. >> > >> > I'm against that idea, especially for 19x19. >> > >> > Hideki >> > >> > Don Dailey: < >> [email protected]>: >> > >I'm trying now to get a rough idea about the strength of fuego and it's >> > >suitablity as the anchor player. >> > > >> > >Right now the numbers are very rough as I need more samples. I'm >> > currently >> > >looking at: >> > > >> > > 1. 9x9 fuego at 1000 simulations >> > > >> > > 2. 19x19 fuego at 3000 simulations. >> > > >> > > >> > >I'm testing against the current CGOS anchors, so FatMan vs fuego at >> 9x9 >> > and >> > >gnugo-3.7.10 at 19x19. >> > > >> > > >> > >At 9x9 fuego appears to be substantially stronger than FatMan, perhaps >> > >100-200 ELO. It also is far faster at 1000 simulation than fatman >> which >> > >requires many more simulations to reach anchor strength. So there is >> no >> > >questions about fuego being a capable anchor for small boards. At this >> > >level on 9x9 FatMan is also stronger than gnugo, so fuego is far >> stronger >> > >than gnugo on 9x9 and is very resource friendly too. >> > > >> > >At 19x19 the story is a bit different. gnugo appears to be >> significantly >> > >stronger, but about twice as slow. There is not enough data to narrow >> > this >> > >down much, but it appears to be over 200 ELO weaker at this level. >> > > >> > >Since fuego is using only about half the CPU resources of gnugo, I can >> > >increase the level. I've only played 30 games at 19x19, so this >> > >conclusion is subject to signficant error, but it's enough to conclude >> > that >> > >it's almost certainly weaker at this level but perhaps not when run at >> the >> > >same CPU intensity as gnugo. >> > > >> > >Of course at higher levels yet, fuego would be far stronger than >> > >gnugo-3.7.10 as seen in the 19x19 cgos tables. But I'm hoping not to >> > push >> > >the anchors too hard - hopefully they can be run on someones older >> spare >> > >computer or set unobtrusively in the background on someones desktop >> > >machine. >> > > >> > > >> > >- Don >> > >---- inline file >> > >_______________________________________________ >> > >computer-go mailing list >> > >[email protected] >> > >http://www.computer-go.org/mailman/listinfo/computer-go/ >> > -- >> > [email protected] (Kato) >> > _______________________________________________ >> > 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 > [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/
