Can someone please post instructions for the following?
Create an empty board and set up an initial pattern of stones in a subrectangle of the board.
Play against a strong AI within the subrectangle, so I can learn the best moves to make in a given local situation.
I often wish I could make my own little puzzles like this, so I can learn the best moves for patterns that actually have come up many times in my real games. I think this procedure would make me a much stronger player, as compared with reading go books (I own 25 go books, but I’m still 10k) or solving tsumego.
More seriously, I’ve heard Katrain and AI Sensei are pretty good for AI-assisted learning. Not sure if either is exactly like you were describing though.
I read them all, years ago. I’m 77 years old and have always learned go slowly. But I love to play go and want to set up situations and learn from them.
Make moves and/or edit the position using the edit tools.
Ensure “AI Review” is enabled in the game dock. (If you have a link that says “Disable AI Review”, then it’s currently enabled. If your link says “Enable AI Review”, click it.)
Wait for the AI to make suggestions.
Step 4 is what you need to be a site supporter for.
(EDIT: In case you didn’t know, site supporters also get AI suggestions for every move and variation when looking at finished games, instead of just the “6 red dots”. The demo board is just an extension of the same feature. Personally, I find looking at my own (and my opponents’) mistakes in game records more straightforward than creating situations in the demo board.)
Katrain allows searching for moves in a restricted area. That is fairly useful for finding the best local move, because in artificially created positions with a largely empty board the AI often wants to tenuki - too many points elsewhere.
AI was trained to find best full board move, not best local move. You need A LOT of playouts to get result that possible to trust when limiting to part of the board. Understanding local pattern may be useful, but don’t do it in actual game. It may be much more efficient to tenuki than to continue local fight.
I think creating artificial positions or limiting AI choices may lead to wrong results or to results that would be very rarely helpful in actual games.
Its better to full board AI review games that you actually played and then make screenshots of full board position each time when AI recommends move in corner in which you are interested in. Make collection of these screenshots.
Then, corner positions that happen most often in actual games would have bigger number of AI recommendations of how play here next. Then you would start to play such corners better.
At first it would look like you get too few AI recommendations from each game. But eventually you would be overwhelmed in the size of collection. So I recommend to review lost games only. Collect only AI moves that you didn’t predict. Don’t collect next AI moves in corners in which you already did mistake. Then your collection would have useful recommendations only. With spaced repetition you would be able to remember it.
Just enter Katrain into google and use the top result. It’s a github page, but don’t be intimidated by that, there is a windows installer for example and instructions for other OSs.
Thank you, but I do not understand how this could be the case. In actual games, there are positions that come up often that I don’t handle well. Some are life and death, but others are not. What they all have in common is being local patterns, with tenuki not involved. I want to isolate such cases as I encounter them, and focus on understanding them. Already I found one such pattern in ensuring that a 3,3 stone could live, but I found it by using Analyze and trying out many possibilities, so I still don’t know the best move in the various variations, which I assume I can only learn reliably from a high dan AI. Trying to learn via actual games is very slow and disorganized. It may be relevant that I play only 9x9 games, so local move patterns are more important here than in full games.
I think its reverse. 9x9 is so small that every move affects everything. While on 19x19 there is such thing as joseki https://online-go.com/joseki/
It’s possible to see same joseki again and again on 19x19 much more often than same opening on 9x9.
But specifying a board of arbitrary length and width seems the wrong way to go about it then.
I would think better to do whole board 9x9 positions including the relevant pattern. Demo board with a subscriber account would work as proposed earlier in this thread.
It’s not exactly the same, but kind of a subset: OJE has “Play mode”.
It’s intended to help learn joseki sequences.
It’s really a “first draft” of that feature, and has had no love or attention since OJE was introduced, but if folk started using it and suggesting improvements it might improve
Rather than this being a problem, I think it is a strength of AI. It teaches us that isolated positions do not exist, period.
Thus, if the AI does not respond to the local position you would like it to respond to, that in itself is also a useful lesson–albeit not the lesson you were looking for.
That being said, when I find myself wanting the AI’s advice on a local position, I simply settle the non-local positions as much as possible, until the local position I’m interested in is the largest or most urgent.
It’s a github page, but don’t be intimidated by that, there is a windows installer for example and instructions for other OSs.