It is currently Tue Mar 19, 2024 4:06 am

All times are UTC - 8 hours [ DST ]




Post new topic Reply to topic  [ 9 posts ] 
Author Message
Offline
 Post subject: KataGo 1.9.0 released!
Post #1 Posted: Sun Jun 27, 2021 9:26 pm 
Lives in sente

Posts: 757
Liked others: 114
Was liked: 916
Rank: maybe 2d
This is a pretty big release, including:

* Something like 75ish Elo of MCTS search improvements (better search with the same neural net)
* Major performance improvements for top-end GPUs and for larger multi-GPU machines
* New features like symmetric board handling and avoiding repetition of the same joseki in more than one corner for more variety
* A handy python script for computing Elos from SGFs for general use including beyond just KataGo (hopefully solving the issue of https://lifein19x19.com/viewtopic.php?p=237920#p237920), and more.

https://github.com/lightvector/KataGo/releases

Have fun!


This post by lightvector was liked by 9 people: Bill Spight, dfan, ez4u, Farodin, Friday9i, gennan, Kirby, le_4TC, Waylon
Top
 Profile  
 
Offline
 Post subject: Re: KataGo 1.9.0 released!
Post #2 Posted: Tue Jun 29, 2021 8:40 pm 
Lives in sente

Posts: 757
Liked others: 114
Was liked: 916
Rank: maybe 2d
In a quick followup, 1.9.1 is released, fixing a bug in KataGo for usage with KaTrain and a few other clients. Thanks to everyone who helped test and troubleshoot things. :)

https://github.com/lightvector/KataGo/releases

Other than that, see still the 1.9.0 page still for notes on all the big new recent changes.

https://github.com/lightvector/KataGo/r ... tag/v1.9.0


This post by lightvector was liked by: Friday9i
Top
 Profile  
 
Offline
 Post subject: Re: KataGo 1.9.0 released!
Post #3 Posted: Wed Jun 30, 2021 3:04 am 
Judan

Posts: 6082
Liked others: 0
Was liked: 786
lightvector wrote:
* Something like 75ish Elo of MCTS search improvements (better search with the same neural net)
* Major performance improvements for top-end GPUs and for larger multi-GPU machines


Although I might have to wait another year for my first GPU, I very much appreciate your great work!

Top
 Profile  
 
Offline
 Post subject: Re: KataGo 1.9.0 released!
Post #4 Posted: Wed Jun 30, 2021 5:22 am 
Lives in gote

Posts: 577
Liked others: 22
Was liked: 36
Rank: Fox Tygem 6d
KGS: emerus
Tygem: emerus
OGS: emerus
Do the avoidsgfs pattern parameters work for analysis or just gameplay? My first thought was to make a file with some of Katas joseki blindspots and set it to a moderately high negative so that it analyzes them. However, it doesn't seem to have done anything - good chance that I didn't configure properly though.

Top
 Profile  
 
Offline
 Post subject: Re: KataGo 1.9.0 released!
Post #5 Posted: Wed Jun 30, 2021 5:45 am 
Lives in sente

Posts: 757
Liked others: 114
Was liked: 916
Rank: maybe 2d
It should work with analysis. The avoid sgfs thing being negative is a bit tricky - it will work if and only if the move gets searched on its own, so that the pattern is able to be seen to get a bonus.

Effectively, you are creating a blind spot of your own, because the bonus you add is one that the net has never seen in training. E.g. you could add an absurd large bonus to play the 1-1 point in an empty corner (would need to be *very* large to cancel out the normal highly negative judgment of the 1-1 point), but unless the 1-1 point also gets at least one playout, the net will still have no idea that there is a bonus there. Once it does get at least one playout, it would be found though.

The other thing is that it assumes the players are adversarial, so if one player prefers a move a lot, the *other* will try to avoid the situation that lets them make it. But of course, the player whose move that the pattern is about may have a bit more control over whether the pattern happens, if it's a one-way-street, or if it's a move that doesn't require cooperation from the other player. For example, a bonus for the 1-1 point, presuming it gets searched, the other player really doesn't have a way to stop from happening, since there are 4 corners.

You can also check the log file to see if you typoed any parameters, the log file always reports in some of the lines shortly after startup if it found any "unused" parameters in the config.

As a sanity check, instead of trying to use it to add preferences, you could also try to see if you can first use it as intended, and make KataGo dislike something it likes to do. For example, nowadays, it likes to open on 3-4 as black at low playouts, but with a single-move SGF showing a 3-4 point and a high penalty, you should be able to get KataGo to prefer 4-4. Or the other way around - 4-4 is preferred at higher playouts, but a penalty on 4-4 should cause black to stick with 3-4 the entire way.

The above details generally do mean that this feature will work a bit better preventing Kata from mildly dispreferring lines it likes, so long as there are similarly-valued alternatives, rather than adding bonuses to lines, since then you're not fighting against the training of the net, and you are only adding patterns that are guaranteed to be seen in the search and therefore take effect.

Top
 Profile  
 
Offline
 Post subject: Re: KataGo 1.9.0 released!
Post #6 Posted: Wed Jun 30, 2021 6:05 am 
Lives in sente

Posts: 757
Liked others: 114
Was liked: 916
Rank: maybe 2d
By the way - if you have a bunch of known joseki blind spots, please send them to me or post them or whatever. Any reasonable SGFs with variations can be added to the training so that the variations get studied during training.

As I've mentioned before elsewhere, essentially through a series of steps that I run every month or so, you can feed the training process with sgf files and ask KataGo to start games from various positions in them and play them out, to form its own judgment of those positions and moves. If those positions include moves that KataGo doesn't currently see but agrees are good once played out, then this causes KataGo to learn them, exactly the same way that it learns what moves are good in the game as a whole in all of its normal training by playing them out via self-play.

So basically, you can fix blind spots this way.


This post by lightvector was liked by: Bill Spight
Top
 Profile  
 
Offline
 Post subject: Re: KataGo 1.9.0 released!
Post #7 Posted: Wed Jun 30, 2021 6:58 am 
Lives in gote

Posts: 577
Liked others: 22
Was liked: 36
Rank: Fox Tygem 6d
KGS: emerus
Tygem: emerus
OGS: emerus
lightvector wrote:
By the way - if you have a bunch of known joseki blind spots, please send them to me or post them or whatever. Any reasonable SGFs with variations can be added to the training so that the variations get studied during training.

As I've mentioned before elsewhere, essentially through a series of steps that I run every month or so, you can feed the training process with sgf files and ask KataGo to start games from various positions in them and play them out, to form its own judgment of those positions and moves. If those positions include moves that KataGo doesn't currently see but agrees are good once played out, then this causes KataGo to learn them, exactly the same way that it learns what moves are good in the game as a whole in all of its normal training by playing them out via self-play.

So basically, you can fix blind spots this way.


Many of the ones I know have probably been fixed with bigger networks. I might check the current strongest 40b for them, about how many playouts would you say to let it search until it is a blind spot?

Anyway, the adversarial issue messes up any thoughts that I had of analyzing ways to play that avoid the same patterns I want to avoid or the other way around. I guess there would be workarounds, like making the bot play from each position, but they would be a tedious.

Edit: Nevermind, adversary isn't a problem and is good. I believe I have a good idea to work with that. But, I would like to analyze as if I'm playing Black and would never 3-3 invade but it also assumes White wouldn't either. It probably isn't easy to set it up so the parameter only applies to one color though. btw thanks for the help already. I think it can be really cool to play with in analysis.

Top
 Profile  
 
Offline
 Post subject: Re: KataGo 1.9.0 released!
Post #8 Posted: Wed Jun 30, 2021 5:15 pm 
Lives in sente

Posts: 757
Liked others: 114
Was liked: 916
Rank: maybe 2d
emerus wrote:
Many of the ones I know have probably been fixed with bigger networks. I might check the current strongest 40b for them, about how many playouts would you say to let it search until it is a blind spot?

Whatever you feel is reasonable given your hardware. :salute: You're looking for basically moves that either aren't discovered, or that ARE discovered by KataGo, but take an unusual amount of time to be discovered. Especially where the move is an appreciable amount better (e.g. upon the move finally being discovered, the evaluation basically jumps abruptly). In Lizzie or many other tools, there should also be a way to view KataGo's raw policy probability for the move. If the raw policy is below 5%, and especially if it's below 2%, in conjunction with taking time to find and an abrupt value jump, that's a good sign also that the net finds that move to be very surprising.

In 1.9.1, actually due to it using analysis wide root noise 0.04 by default, if you weren't using any analysis wide root noise before, you will find it actually is better at finding some blind-spot moves on its own, but that still doesn't stop them from being good training material that signals that KataGo isn't as experienced with a joseki as it ideally could be, so feel free to share whatever you think reasonable.

emerus wrote:
Edit: Nevermind, adversary isn't a problem and is good. I believe I have a good idea to work with that. But, I would like to analyze as if I'm playing Black and would never 3-3 invade but it also assumes White wouldn't either. It probably isn't easy to set it up so the parameter only applies to one color though. btw thanks for the help already. I think it can be really cool to play with in analysis.


:)


Last edited by lightvector on Wed Jun 30, 2021 5:18 pm, edited 1 time in total.

This post by lightvector was liked by: emerus
Top
 Profile  
 
Offline
 Post subject: Re: KataGo 1.9.0 released!
Post #9 Posted: Wed Jun 30, 2021 5:17 pm 
Lives in sente

Posts: 757
Liked others: 114
Was liked: 916
Rank: maybe 2d
Also, for anyone interested, here's a little image-guided tour about how sometimes the move that KataGo is failing to find is not actually the blind spot, or at least not the only blind spot, rather it could be a deeper move... and showing the interactive process by which one might find that deeper move. https://imgur.com/a/H4D6t7k


This post by lightvector was liked by 2 people: dfan, emerus
Top
 Profile  
 
Display posts from previous:  Sort by  
Post new topic Reply to topic  [ 9 posts ] 

All times are UTC - 8 hours [ DST ]


Who is online

Users browsing this forum: Google [Bot] and 1 guest


You cannot post new topics in this forum
You cannot reply to topics in this forum
You cannot edit your posts in this forum
You cannot delete your posts in this forum
You cannot post attachments in this forum

Search for:
Jump to:  
Powered by phpBB © 2000, 2002, 2005, 2007 phpBB Group