It is currently Tue Jan 19, 2021 4:58 pm

All times are UTC - 8 hours [ DST ]




Post new topic Reply to topic  [ 2 posts ] 
Author Message
Online
 Post subject: Joseki choice and frequency according to Elfv2
Post #1 Posted: Tue Jul 07, 2020 9:22 am 
Lives with ko
User avatar

Posts: 195
Liked others: 41
Was liked: 30
Rank: EGF 2k
KGS: MKyle
I've been taking a look at bots and joseki again.

The below sgf is kind of an intermediate result from some of these investigations. While there are issues and limitations to it's usefulness I wonder if some would find it interesting?

The sgf is a tree of joseki sequences with each move annotated with the frequency with which the AI (Elfv2) chooses each of the moves as it's next local move over a wide range of pro board positions.
more detail on the method:
for each joseki position I did a search for the most recent games in my pro database with that position (and with the correct player playing next.)
I used the board position at which the humans played next in that area ie. when the humans judged that this was the most interesting part of the board.
I used 1000 whole board positions for the first few moves and slowly moved down to 50 whole board positions for some of the more obscure sequences. There were also a couple of patterns that humans just don't play and so there were not enough matches to use.
For each whole board position I asked the bot to play a move and counted the frequency with which each move came up. I excluded tenukis when calculating percentages.

I searched for the patterns as a whole quarter board (10x10) and I also categorised moves as tenuki on this basis.


One reason for doing this kind of thing is that although bots are focused on local efficiency, they also play well for the whole board. You can set up a simple board position and show that a bot consideres a specific move x bad as it looses 4% but the bot might consider the move x optimal in a different board position. This aims loosely to capture most of the sequences that the bot thinks are ever top choice.



I will probably post a couple of specific things I found interesting and would encourage anyone interested to share their thoughts


Attachments:
File comment: Elfv2 Joseki frequency tree
summarytree.sgf [9.98 KiB]
Downloaded 152 times

This post by MikeKyle was liked by 2 people: Bill Spight, tchan001
Top
 Profile  
 
Online
 Post subject: Re: Joseki choice and frequency according to Elfv2
Post #2 Posted: Tue Jul 21, 2020 4:29 pm 
Lives with ko
User avatar

Posts: 195
Liked others: 41
Was liked: 30
Rank: EGF 2k
KGS: MKyle
Here's one that I feel like I learned something from:

Click Here To Show Diagram Code
[go]$$c A 35% of the time \nB 17%\nC 13%\nD 13%..
$$ ------------------
$$ | . . . . . . . . .
$$ | . . . . . . . . .
$$ | . . . . . X . . g
$$ | . e d X b . . . .
$$ | . . . . . . . . .
$$ | . . O . c . . . .
$$ | . . . . . . . . .
$$ | . . X . . . . . .
$$ | . . . . . . . . .
$$ | . . a . . . . . .[/go]


I was looking at this tenuki position.
It looks a little bit like move A is a semi-tenuki (although looking at the follow-ups this is definitely part of the local sequence) and so the most frequent move directly relating to the corner would be attaching in the knight's move at move B. I remember this move from handicap games where white is hoping that black will passively allow white to make some nice outside shape:

Click Here To Show Diagram Code
[go]$$Wc White is too happy for the good shape.
$$ ------------------
$$ | . . . . . . . . .
$$ | . . . . . . . . .
$$ | . . 4 6 2 X . . .
$$ | . . 3 X 1 . . . .
$$ | . . . 5 . . . . .
$$ | . . O . . . . . .
$$ | . . . . . . . . .
$$ | . . X . . . . . .
$$ | . . . . . . . . .
$$ | . . . . . . . . .[/go]


I'm not sure I remember the response i was told to try when I was playing handicap games (2 below 1 maybe?)

I tried adding a larger number of board positions and it turns out that move B is a little less popular with the bot as this suggests. it looks like b is a very popular choice when there is a two space extension for black on the left hand side (black has C9)
Here are the board positions sorted by the bot choice:



The bot follow up always starts:
Click Here To Show Diagram Code
[go]$$Wc Bot starts..
$$ ------------------
$$ | . . . . . . . . .
$$ | . . . . . . . . .
$$ | . . . 2 . X . . .
$$ | . . . X 1 . . . .
$$ | . . . 3 . . . . .
$$ | . . O . . . . . .
$$ | . . . . . . . . .
$$ | . . X . . . . . .
$$ | . . . . . . . . .
$$ | . . . . . . . . .[/go]


If black is very strong then he may now choose to play on the outside and try to attack the whole thing, but black most often cuts:

Click Here To Show Diagram Code
[go]$$Bc Usually black cuts
$$ ------------------
$$ | . . . . . . . . .
$$ | . . . . . . . . .
$$ | . . . X . X . . .
$$ | . . . X O 2 . . .
$$ | . . . O 1 . . . .
$$ | . . O . . . . . .
$$ | . . . . . . . . .
$$ | . . X . . . . . .
$$ | . . . . . . . . .
$$ | . . . . . . . . .[/go]


The most common follow up seems to be:

Click Here To Show Diagram Code
[go]$$Bc Most common sequence
$$ ------------------
$$ | . . . . . . . . .
$$ | . . . 5 3 4 . . .
$$ | . . . X 2 X 6 . .
$$ | . c d X O O . . .
$$ | . . . O X . b . .
$$ | . . O . 1 . e . .
$$ | . . . a . . . . .
$$ | . . X . . . . . .
$$ | . . . . . . . . .
$$ | . . . . . . . . .[/go]


Black then usually needs to fix at a. If black chooses instead to take b then white can either play a herself or play c or d for some kind of life. Even if black fixes there seems to be a lot of aji for white to get a ko or seki in the corner as well as that nice ponnuki towards the top side.

Black may instead choose to fix up the corner stones for one of those pushing from behind chicken scenarios:
Click Here To Show Diagram Code
[go]$$Bc A mess
$$ ------------------
$$ | . . . . . . . . .
$$ | . . . . . . . . .
$$ | . . . X 1 X . . .
$$ | . . . X O O 4 . .
$$ | . . . O X 3 . . .
$$ | . . O . 2 5 . . .
$$ | . . . . 6 7 . . .
$$ | . . X . 8 9 . . .
$$ | . . . . 0 . . . .
$$ | . . . . . . . . .[/go]


I may end up using it inappropriately but I have added a tool to my toolkit!


Attachments:
groupedboardpositions.sgf [35.22 KiB]
Downloaded 110 times

This post by MikeKyle was liked by: Bill Spight
Top
 Profile  
 
Display posts from previous:  Sort by  
Post new topic Reply to topic  [ 2 posts ] 

All times are UTC - 8 hours [ DST ]


Who is online

Users browsing this forum: salerno 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