Life In 19x19http://lifein19x19.com/ KataGo (40x384) 23689108 playoutshttp://lifein19x19.com/viewtopic.php?f=18&t=17715 Page 1 of 1

Author:  goame [ Mon Aug 10, 2020 5:02 am ]
Post subject:  KataGo (40x384) 23689108 playouts

2x RTX 2080 Ti
~60 out of 64 GB RAM
~4 hours and 30 minutes
Japanese rules
Komi 6.5

https://d3dndmfyhecmj0.cloudfront.net/g ... index.html
Large Net Size (never used for self-play)
[740M] g170e-b40c384x2-s2348692992-d1229892979.zip
If you want to see how a super ultra large/slow network performs that nobody has tested until now, try the fat 40-block 384 channel network mentioned a little up above.

Take also a look here: https://lifein19x19.com/viewtopic.php?f ... 51#p254851

 Attachments: 23689108 playouts.jpg [ 1.69 MiB | Viewed 4294 times ]

 Author: goame [ Mon Aug 10, 2020 5:57 am ] Post subject: Re: KataGo (40x384) 23689108 playouts 40x384 vs 30x32041,5% -1,1 58,5% vs 43,3% -1,0 56,7%Q16 = 41,6 20.3m -1,1 vs Q16 = 43,3 6.2m -1,0Circles are not on every cutting point. (A1, A18, A19, B19, S1, T1, T2, T18, T19)vsCircles are on every cutting point.We have another position on the analysis board.Some questions:1. What do you think about the 40% for black vs 60% for white winning chance / fair play at the beginning of the game?2. How can I change the komi? I would like to try with 5.5 komi to see if we can get sometihng like 50% -0 50% or maybe -0,1 instead of -1,1.3. To safe a lot of analysis time and to have better calculations and results, can we get something like Q16 = D16 = D4 = Q4? And also A1, A2, B1 = A19, A18, B19 = T1, T2, S1 = T19, T18, S19. And also A3 = A17 = C19 = R19 = T17 = T3 = R1 = C1.Not for the center point but for every other cutting point we will get the result for at least another 3 cutting points = That means we will have a 1 + (3 for free) analysis = 25% instead of 100% from the analysis time is needed. In other positions we will get 1 + (7 for free) = 12,5% instead of 100% from the analysis time is needed for the same result.This is a great speedup when time doesn't matter and better than to buy 14 other rtx 2080 ti gpus.You can see above how much time this analysis has taken.4. Why we have 41,5% on the left side but on 41,6 on Q16?5. What do you think should be improved?6. What do you thing about the new board position?

 Author: RobertJasiek [ Mon Aug 10, 2020 7:08 am ] Post subject: Re: KataGo (40x384) 23689108 playouts "2x RTX 2080 Ti ~60 out of 64 GB RAM"How much better in practice are 2x RTX 2080 Ti compared to 1x RTX 2080 Ti?How much better in practice are 64 GB RAM compared to 16 or 32 GB, respectively?Is it different for creating a neural net versus applying one?

 Author: lightvector [ Mon Aug 10, 2020 8:53 am ] Post subject: Re: KataGo (40x384) 23689108 playouts Quick answer/observation: I strongly suspect that you have set the rules and/or komi incorrectly. You claim Japanese rules 6.5, but actually your winrates look consistent with Chinese rules 7.5, and even a million playouts probably shouldn't change the winrate that much.I'm not going to implement opening symmetry reduction in KataGo - it causes a lot of additional code complexity and even the potential for artificial training bias for almost no practical gain for real games. Especially when you can capture a lot of this gain via a simple workaround:If you want to save a factor of 4 or 8, then just play the first move manually to break the symmetry. Play the 4-4 point manually, and do the search on the position after that, and then play the 3-4 point manually, and do the search on the position after that, and record those results.

 Author: Bill Spight [ Mon Aug 10, 2020 9:11 am ] Post subject: Re: KataGo (40x384) 23689108 playouts lightvector wrote:I'm not going to implement opening symmetry reduction in KataGo - it causes a lot of additional code complexity and even the potential for artificial training bias for almost no practical gain for real games. Especially when you can capture a lot of this gain via a simple workaround:If you want to save a factor of 4 or 8, then just play the first move manually to break the symmetry. Play the 4-4 point manually, and do the search on the position after that, and then play the 3-4 point manually, and do the search on the position after that, and record those results.Just curious. I have wondered about the utility of dictating a few opening plays for training purposes. For instance, even nirensei vs. nirensei will break rotational symmetry. And doing so might be a way of increasing exploration without jiggering the code. Does that make sense?

 Author: ez4u [ Tue Aug 11, 2020 4:24 am ] Post subject: Re: KataGo (40x384) 23689108 playouts In Lizzie you change the komi by hitting "i" (game info) once your engine loads. Just input the revised komi and save. It is best to change the komi immediately. Otherwise you will get a mixture of results if you move back and forth in a game. If you want to analyze a particular position under different komi, it is better to restart the game with the new komi. With Japanese rules, komi of 6.5 still favors White, but 6.0 favors Black. Try that.

 Author: Uberdude [ Tue Aug 11, 2020 4:34 pm ] Post subject: Re: KataGo (40x384) 23689108 playouts ez4u wrote:With Japanese rules, komi of 6.5 still favors White, but 6.0 favors Black. Try that.How about komi of 2π?

 Author: ez4u [ Tue Aug 11, 2020 6:07 pm ] Post subject: Re: KataGo (40x384) 23689108 playouts Uberdude wrote:ez4u wrote:With Japanese rules, komi of 6.5 still favors White, but 6.0 favors Black. Try that.How about komi of 2π? Good question. However, it will not work at present. I think lightvector wrote somewhere that, in the code, katago checks whether the komi x 2 = an integer (IIRC) and rounds it off if it is not. Hence I believe that it effectively only accepts X.0 or X.5 komi.

 Author: goame [ Wed Aug 12, 2020 12:25 am ] Post subject: Re: KataGo (40x384) 23689108 playouts RobertJasiek wrote:"2x RTX 2080 Ti ~60 out of 64 GB RAM"How much better in practice are 2x RTX 2080 Ti compared to 1x RTX 2080 Ti?How much better in practice are 64 GB RAM compared to 16 or 32 GB, respectively?Is it different for creating a neural net versus applying one?2x RTX 2080 Ti are much better in practice compared tp 1x RTX 2080 Ti.2x RTX 2080 Ti are winning the games in the opening against the strongest human pros on different go servers. 1x RTX 2080 Ti can't do this, it's winning most of the time in the middlegame.It is amazing to see how KataGo on 2x RTX 2080 Ti is destroying KataGo or Leela or ... on only 1x RTX 2080 Ti.I had first only 1x RTX 2080 Ti and it was a big fun. But I decided to buy exactly the same a second RTX 2080 Ti.The playing style is now completely on another level.KataGo is also running much faster and I get now same moves much faster = saving a lot of time.For example the analysis above would take with only 1 RTX 2080 Ti ~9 hours instead of ~4 hours and 30 minutes.Think about a game what you want to be analysed with 1 and with 2 mintues x 300 moves. How long would this take with 1 and with 2x RTX 2080 Ti? And thing about the GB RAM.Someone mentioned that most of the time after 2 or 3 million playouts KataGo and Leela doesn't change their best moves.I know now that this is completely wrong because after 22xxxxxx I have an other variation than after 23xxxxxx playouts.The reason is simple: KataGo is improving every time to think bigger and better and this game is to difficult.In some positions like in the endgame they are forced (only) moves but in the middlegame or opening we can easily take between 25 and 100 million playouts.64 GB RAM are much better than 16 GB RAM and that's the reason why I changed from 16 to 64 GB RAM.For example the analysis has taken ~4 hours and 30 minutes for ~24 million playouts.With only 16 GB RAM you would reach only ~1 hour and 7 minutes and ~6 million playouts.If you are creating a neural net with 2x RTX 2080 Ti, than it should run better on 2x RTX 2080 Ti or higher than on 1x RTX 2080 Ti, because it is designed for 2x RTX 2080 Ti.You will save a lot of time when creating a neural net or get a much stronger net in the same time. It should be clear why many strong GPUs are important, if you have read about all the GPUs KataGo is using to improve.

 Author: goame [ Wed Aug 12, 2020 1:09 am ] Post subject: Re: KataGo (40x384) 23689108 playouts lightvector wrote:Quick answer/observation: I strongly suspect that you have set the rules and/or komi incorrectly. You claim Japanese rules 6.5, but actually your winrates look consistent with Chinese rules 7.5, and even a million playouts probably shouldn't change the winrate that much.I'm not going to implement opening symmetry reduction in KataGo - it causes a lot of additional code complexity and even the potential for artificial training bias for almost no practical gain for real games. Especially when you can capture a lot of this gain via a simple workaround:If you want to save a factor of 4 or 8, then just play the first move manually to break the symmetry. Play the 4-4 point manually, and do the search on the position after that, and then play the 3-4 point manually, and do the search on the position after that, and record those results.I see now when I click i in Lizzie, then there is komi 7.5, so the result above is done with 7.5 komi.When I change the komi to 6.5 or another one and click ok, then it doesn't work. I see it is clicked correctly a little blue button for a moment but this little window won't close...???I can't do other thinks and need to click the red button with x but then the komi is not saved.How to change the rules from Chinese to Japanese or other in Lizzie?What is the difference to the KataGo tuning Japanese rules I have done?Would the code complexity from the opening symmetry reduction slow down KataGo? Let's say 5% slower to the end of the game?2x RTX 2080 Ti are ~2x faster than 1x RTX 2080 Ti.I have seen that the best move have even changed after 23 compared to 22 million playouts.Let's say a better move from KataGo is worth 25% of a stone.With opening symmetry reduction KataGo will be ~4x faster than some slower KataGos. It could be that the faster KataGo with opening symmetry reduction wins every time when this improvement works in the opening between 25% and 100% or a stone.After the first 4 symmetry moves KataGo would be 1 stone or even up to 4 stones better.It's like changing the komi from 7.5 to 6.5 or from 7.5 to 3.5.And after the symmetry KataGo will have much more playouts and play with his move for example with 24 million playouts compared to the KataGo opponent with 6 million playouts.I have loaded an sgf file from my games against CrazyStone DeepLearning and it was played with Japanese rules and 6.5 komi.After running the analysis in the start position I got after 9 million playouts:48,6% -0,2 51,4% and that is much more a fair play now.= 6.5 komi should be used.

 Author: goame [ Wed Aug 12, 2020 1:20 am ] Post subject: Re: KataGo (40x384) 23689108 playouts Bill Spight wrote:lightvector wrote:I'm not going to implement opening symmetry reduction in KataGo - it causes a lot of additional code complexity and even the potential for artificial training bias for almost no practical gain for real games. Especially when you can capture a lot of this gain via a simple workaround:If you want to save a factor of 4 or 8, then just play the first move manually to break the symmetry. Play the 4-4 point manually, and do the search on the position after that, and then play the 3-4 point manually, and do the search on the position after that, and record those results.Just curious. I have wondered about the utility of dictating a few opening plays for training purposes. For instance, even nirensei vs. nirensei will break rotational symmetry. And doing so might be a way of increasing exploration without jiggering the code. Does that make sense?It sounds interesting.I mean the code can work for example with Sanrensei.Special symmetry reducitonFrom left to right D16 vs Q16Up und down D16 vs D4Diagonal D16 vs Q4But also many other things like C16 vs R4 or Q3 or D17

 Author: goame [ Wed Aug 12, 2020 1:22 am ] Post subject: Re: KataGo (40x384) 23689108 playouts ez4u wrote:In Lizzie you change the komi by hitting "i" (game info) once your engine loads. Just input the revised komi and save. It is best to change the komi immediately. Otherwise you will get a mixture of results if you move back and forth in a game. If you want to analyze a particular position under different komi, it is better to restart the game with the new komi. With Japanese rules, komi of 6.5 still favors White, but 6.0 favors Black. Try that.Somehow I can't change the komi after hitting i.

 Author: RobertJasiek [ Wed Aug 12, 2020 1:41 am ] Post subject: Re: KataGo (40x384) 23689108 playouts Many thanks for your hardware experience! It will help me when buying mine but I await the RTX 3xxx series. A computer science professor said that 2x RTX 2080 TI roughly equals the 4 TPUs for AlphaGo Zero.EDIT:Do you use a SLI bridge or would that be counter-productive?

 Author: goame [ Wed Aug 12, 2020 2:35 am ] Post subject: Re: KataGo (40x384) 23689108 playouts RobertJasiek wrote:Many thanks for your hardware experience! It will help me when buying mine but I await the RTX 3xxx series. A computer science professor said that 2x RTX 2080 TI roughly equals the 4 TPUs for AlphaGo Zero.EDIT:Do you use a SLI bridge or would that be counter-productive?I'm using an SLI bridge. I think it's better to have and use it when having 2x RTX 2080 Ti or more GPUs.I have actually seen no information about how much 1 tensor core has improved from RTX 2080 Ti to RTX 3080 Ti. Maybe only 10% faster?Also no information how many tensor cores RTX 3080 Ti would have. If they doubled the cores, then these new cards are interesting for me.

 Author: RobertJasiek [ Wed Aug 12, 2020 3:44 am ] Post subject: Re: KataGo (40x384) 23689108 playouts Currently, there are only rumours about coming Nvidias. One of them claims that 3080 would be similar to 2080 TI but at the price of 2080, which would make them very attractive for buying two cards. Other rumours can't distinguish between ordinary cards and the too expensive Titan series. I am afraid we need to wait. Of course, with your prior investment, you should want at least doubled features before upgrading.

 Author: lightvector [ Wed Aug 12, 2020 6:26 am ] Post subject: Re: KataGo (40x384) 23689108 playouts goame wrote:I see now when I click i in Lizzie, then there is komi 7.5, so the result above is done with 7.5 komi.When I change the komi to 6.5 or another one and click ok, then it doesn't work. I see it is clicked correctly a little blue button for a moment but this little window won't close...???I can't do other thinks and need to click the red button with x but then the komi is not saved.I have no idea, that sounds really weird. Some sort of GUI bug in Lizzie?goame wrote:How to change the rules from Chinese to Japanese or other in Lizzie?Lizzie has no support for changing rules on its own, therefore you need to apply a workaround:https://github.com/lightvector/KataGo#other-questionsgoame wrote:With opening symmetry reduction KataGo will be ~4x faster than some slower KataGos. It could be that the faster KataGo with opening symmetry reduction wins every time when this improvement works in the opening between 25% and 100% or a stone.After the first 4 symmetry moves KataGo would be 1 stone or even up to 4 stones better.It's like changing the komi from 7.5 to 6.5 or from 7.5 to 3.5.And after the symmetry KataGo will have much more playouts and play with his move for example with 24 million playouts compared to the KataGo opponent with 6 million playouts.I think you are off here. As far as I can tell, opening symmetry reduction should be almost worthless for practical playing strength. If the game opens on a 3-4 point, then symmetry is broken on move 1 and stays broken. If it opens on all 4-4s, then symmetry is broken on move 5 when the first 3-3 invasion happens and usually stays broken. After that, there is no difference in the efficiency of search. If a typical game lasts 250 moves, this means we get a speedup on 1-2% of the moves of the game, so that's roughly similar to speeding up the overall bot by 1%-2%... which is negligible. There are waaaay better things to work on than a whole bunch of code complexity for such a tiny gain.Relatively speaking, the game is usually won or lost in the range from around move 8 to the macroendgame. The very early opening and the microendgame probably don't benefit as much from deeper search, the former because you don't need deeper search to play a 4-4 point or a 3-4 point or to start a 3-3 invasion that you would have played anyways, and the microendgame because often the game is already decided and if not, bots are very good at the microendgame already and with enough playouts will often be searching deep enough to solve it, past which you get quickly diminishing returns for more search.So if one were to implement more fancy time management (which *would* be worth doing and would be worth a lot more than 1-2%), in addition to a lot of other important heuristics about choosing how long to search, one would probably greatly downweight the time spent on the first several moves of the game. Which means the gain from symmetry would be even less than 1% - you're getting a speedup on the moves that you would spend less than 1% of your time on anyways.The idea that a bot is going to find a variation that is up to 4 points better by move 5 is silly. (By current top bot standards. obviously in theory we can't prove that something crazy doesn't happen with optimal play, but even if so, as far as practical statistics by real bots at their current level of play, there is no sign that these opening moves differ by more than tiny fractions of a point on average).You might ask - how about 9x9, where the game is shorter and the first few moves actually matter a huge amount? Actually there I think symmetries are still worthless too. Because if you really care about optimizing 9x9, forget about searching the first few moves in real time, you should just build an opening book... and by the time you get out of book almost always you'll have broken symmetry.

 Author: lightvector [ Wed Aug 12, 2020 6:44 am ] Post subject: Re: KataGo (40x384) 23689108 playouts Bill Spight wrote:lightvector wrote:I'm not going to implement opening symmetry reduction in KataGo - it causes a lot of additional code complexity and even the potential for artificial training bias for almost no practical gain for real games. Especially when you can capture a lot of this gain via a simple workaround:If you want to save a factor of 4 or 8, then just play the first move manually to break the symmetry. Play the 4-4 point manually, and do the search on the position after that, and then play the 3-4 point manually, and do the search on the position after that, and record those results.Just curious. I have wondered about the utility of dictating a few opening plays for training purposes. For instance, even nirensei vs. nirensei will break rotational symmetry. And doing so might be a way of increasing exploration without jiggering the code. Does that make sense?As I mentioned above, one should almost completely neglect breaking symmetry as any reason for choosing anything. But as far as forcing openings to increase exploration - yes, that's a good idea. Except for the fact that it's made redundant by even better ideas. Namely: if you're willing to go as far as manually dictating openings, you might as well go further - just take a lot of historical pro and amateur games and play starting from randomly sampled positions from those games. If you wish, you can weight the sampling more heavily towards the opening instead of later in the game. You'll see a huge variety of openings, more than anything you could reasonably input manually. KataGo did this for the last phase of its training, when it departed from its "semi-zero" status. It didn't seem to help with raw playing strength, but I like to hope that it improved KataGo's evaluation quality slightly on the kinds of positions that occur in human games that might not always follow purely bot-favored lines, positions which humans still care about good evaluation quality for teaching and analysis.Also even prior to departing "semi-zero", KataGo used high entropy in the first few moves of the game in self-play (often playing them via raw policy with no search), and occasionally used a technique where the game was initialized by picking a random early move (like move 1, or move 3, or move 10) and forcing the next stone to be played on the best of a random N possible places, for small N. So in some tiny percentage of games you get an opening move on the 5-5 point, or the 3-7 point, or some crazy center move, or whatever, because the N random spots didn't include any of the 4-4, 4-3, 3-3 points. This was also intended to ensure there was a little bit of training data on handling very unusual opening configurations.

 Author: RobertJasiek [ Wed Aug 12, 2020 7:35 am ] Post subject: Re: KataGo (40x384) 23689108 playouts "bots are very good at the microendgame" How do you know? Do they apply endgame theory correctly? Have you verified this?

 Author: Bill Spight [ Wed Aug 12, 2020 7:43 am ] Post subject: Re: KataGo (40x384) 23689108 playouts lightvector wrote:Namely: if you're willing to go as far as manually dictating openings, you might as well go further - just take a lot of historical pro and amateur games and play starting from randomly sampled positions from those games. If you wish, you can weight the sampling more heavily towards the opening instead of later in the game. You'll see a huge variety of openings, more than anything you could reasonably input manually. KataGo did this for the last phase of its training, when it departed from its "semi-zero" status. It didn't seem to help with raw playing strength, but I like to hope that it improved KataGo's evaluation quality slightly on the kinds of positions that occur in human games that might not always follow purely bot-favored lines, positions which humans still care about good evaluation quality for teaching and analysis. (Emphasis mine.)Thanks. My main emphasis is on the use of bots for tools to assist human understanding. OC, right now top humans have their hands full learning to imitate bots, but in 10 or 15 years we will have pros who not only play more like bots but who will be able to explain their own thinking.

 Author: Bill Spight [ Wed Aug 12, 2020 8:43 am ] Post subject: Re: KataGo (40x384) 23689108 playouts RobertJasiek wrote:"bots are very good at the microendgame" How do you know? Do they apply endgame theory correctly? Have you verified this?Depends on what you think is very good. Here is a game position from over 1,000 years ago, GoGoD 1150C&IP01, edited by me. I switched colors and added a comment at move 228. It was played under ancient Chinese rules with territory scoring and a group tax. White played first. Black won by 1 pt., but I discovered, in a few minutes, how White could get jigo in the microendgame. AFAICT, I am the first to have discovered that in all the time the game record has been available. However, I have the advantage of having studied mathematical go. With the switched colors and the fact that each side has three groups, Black to play can win after W228 under modern area scoring with zero komi. I expect that KataGo can find a correct line of play in a few seconds, but I don't know about other bots.

 Page 1 of 1 All times are UTC - 8 hours [ DST ] Powered by phpBB © 2000, 2002, 2005, 2007 phpBB Grouphttp://www.phpbb.com/