It is currently Fri Mar 29, 2024 3:56 am

All times are UTC - 8 hours [ DST ]




Post new topic Reply to topic  [ 8 posts ] 
Author Message
Offline
 Post subject: go problems from ai
Post #1 Posted: Mon Mar 09, 2020 11:11 am 
Beginner

Posts: 11
Liked others: 1
Was liked: 0
Rank: KGS 4 kyu
KGS: AaronB
Has anyone ever considered using ai to create tsumego or tesuji problems. It seems like that could be a great thing to do. You might be able to get them from the games they play, or specifically program the ai to create the problems.

What do you all think?

Top
 Profile  
 
Offline
 Post subject: Re: go problems from ai
Post #2 Posted: Mon Mar 09, 2020 11:29 am 
Dies in gote

Posts: 61
Liked others: 12
Was liked: 18
Universal go server handle: Jæja
Nice idea!

- For this you would need to define what tsumego and tesuji problems are. Perhaps both share the property that they involve a local move with no alternatives?
- Note that it might be hard to define what 'local' means. You could put a bounding box around each group and check if there's a single best move in it.
- One approach that will definitely come up with some examples is take it one step further and check if a board position has only one move for the current player. You could then construct a problem set from those moves by checking them manually.

Top
 Profile  
 
Offline
 Post subject: Re: go problems from ai
Post #3 Posted: Mon Mar 09, 2020 3:24 pm 
Judan

Posts: 6725
Location: Cambridge, UK
Liked others: 436
Was liked: 3719
Rank: UK 4 dan
KGS: Uberdude 4d
OGS: Uberdude 7d
Using AI to extract problems from human games (by this forum's lightvector):

https://neuralnetgoproblems.com


This post by Uberdude was liked by: Harleqin
Top
 Profile  
 
Offline
 Post subject: Re: go problems from ai
Post #4 Posted: Mon Mar 09, 2020 6:02 pm 
Lives in sente

Posts: 757
Liked others: 114
Was liked: 916
Rank: maybe 2d
Building that site was a fun experiment.

I hope eventually someone has the time and inclination to do it far better than I did. That site was built back when all I had was pure policy nets, with the one key detail being to give the player rank as an input to the net to try to teach the net about how players of different ranks play and how their moves might differ on average, to try to determine some idea of the "rank" of a problem. I haven't revisited this idea since.

In addition to improving the "rank" heuristics, maybe bigger nets and more data could be much better. The nets weren't that large by today's standards. And it was before I started seriously experimenting with how to integrate a net with search to actually produce high-quality evaluations. So the site is relying heavily on just "this was the move the pro or high-dan amateur played" without the ability to back it up with independent AI evaluation - whether to verify the goodness of the move, or even just to show more context to the user and/or give feedback about answers besides the pro move.

I do think it still came out well given the amount of effort it involved and the limitations I had at the time. :D

And I hope at least a few people have found it useful. A lot of the problems turned out to be good "follow basic instinct" practice (https://senseis.xmp.net/?BasicInstinct) which hopefully can be educational if you're a beginner unfamiliar some of the basic instinct moves in practical open-space situations.


This post by lightvector was liked by 2 people: Bill Spight, Jæja
Top
 Profile  
 
Offline
 Post subject: Re: go problems from ai
Post #5 Posted: Mon Mar 09, 2020 7:00 pm 
Honinbo

Posts: 10905
Liked others: 3651
Was liked: 3374
AFAIK, human pros are at least as good as current top bots in tsumego. However, the bots are not trained to solve or create tsumego. If AI programs were trained to create tsumego, they might create some fantastic ones. :)

Tesuji are another matter. In reviews of pro games, top bots often find (probably) better local plays than the pros did. The positions where that happens could be good for tesuji problems. :)

_________________
The Adkins Principle:
At some point, doesn't thinking have to go on?
— Winona Adkins

Visualize whirled peas.

Everything with love. Stay safe.

Top
 Profile  
 
Offline
 Post subject: Re: go problems from ai
Post #6 Posted: Mon Mar 09, 2020 9:38 pm 
Gosei
User avatar

Posts: 1581
Location: Hong Kong
Liked others: 54
Was liked: 544
GD Posts: 1292
http://lie.math.brocku.ca/GoTools/index ... t=features

Quote:
There is an automatic problem generation feature that can generate about 100 problems on a PC 486 overnight.

_________________
http://tchan001.wordpress.com
A blog on Asian go books, go sightings, and interesting tidbits
Go is such a beautiful game.

Top
 Profile  
 
Offline
 Post subject: Re: go problems from ai
Post #7 Posted: Mon Mar 09, 2020 11:43 pm 
Judan

Posts: 6087
Liked others: 0
Was liked: 786
Every position is a problem but we often want something more specific. The more specific the more difficult constructing problems becomes. It does not require non-AI or AI programs to create problems. If one uses programs for creation, we must evaluate their problem positions whether they meet our specifications.

Hashimoto Utaro was said to have invented a problem every day. That impressed me before I started to create problems.

Nowadays, I create more than one problem per day on average so that I also solve them and describe the solution. Creating a problem takes between seconds and 4 days. The more specific the more time I need. 4 days rarely occur when I need a very particular (counter-)example for given presuppositions, e.g. when I have proved a theorem of existence or need to disprove a conjecture of non-existence.

Creating problems is the easy part. The difficult part is to solve them and verify the solutions by tactical reading (incl. verification of its decision-making) or equivalent means (such as assessing numbers of liberties in a capturing race). Computer-sampling of variations is NOT a verification of a solution! E.g., if a computer tests 99 of 99 failure variations for one particular next move, the program might be wrong if the 100th variation would be successful.

Top
 Profile  
 
Offline
 Post subject: Re: go problems from ai
Post #8 Posted: Tue Mar 10, 2020 2:10 am 
Dies in gote

Posts: 61
Liked others: 12
Was liked: 18
Universal go server handle: Jæja
Uberdude wrote:
Using AI to extract problems from human games (by this forum's lightvector):

https://neuralnetgoproblems.com
Amaaaaazing :bow: I really like the problems. Thanks, lightvector!


This post by Jæja was liked by: lightvector
Top
 Profile  
 
Display posts from previous:  Sort by  
Post new topic Reply to topic  [ 8 posts ] 

All times are UTC - 8 hours [ DST ]


Who is online

Users browsing this forum: No registered users 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