Well, my idea is not actually that these positions will be problems for bots. They are actually problems for humans. But the human answers are not the answers bots would give. The human answers may not even be on the bots' radars. My thought is that these problems could be resources for people studying the opening until new, AI tested opening books become available.
I invite anyone to post answers for any top AI, and to post new problems.
Pre-AI opening theory makes this almost an only move problem. But I doubt if it is on the bots' radars. Where would a bot play? I don't think that there is one answer.
is the human textbook answer. The reason for not playing at a, is that White has a follow-up at which makes a base, and which also has a follow-up at . The reasoning is sound about choosing plays that have good follow-ups that also have good follow-ups. However, I doubt if any of today's top bots give more than a few rollouts.
Not that is an error. It probably loses a few percentage points to any bot's choice of play, but not enough to make it a clear mistake. It's just that the bots do not value the sides as highly as they do the corners.
Here are my guesses for the bots' answers. They will make a keima enclosure of one of the 4-4 corners. I doubt if they will agree which enclosure to make, but I have covered the bases.
Re: Opening problems for AI: Problem 1
Posted: Sat Jun 13, 2020 2:06 pm
by sorin
I gave it a try with KataGo, turns out that your statement is a human misconception, Bill
Namely: bots like the area around the middle of the right hand side, but just think it is not most efficient to play there directly: rather, KataGo wants to make a contact play against black's shimari first, then extend on the right side (exact location depending on how black answers).
It is true though, that the direct corner enclosures from the hoshis are the top choices for next move.
Re: Opening problems for AI: Problem 1
Posted: Sat Jun 13, 2020 3:47 pm
by Bill Spight
sorin wrote:I gave it a try with KataGo, turns out that your statement is a human misconception, Bill
Well, . . .
Namely: bots like the area around the middle of the right hand side, but just think it is not most efficient to play there directly: rather, KataGo wants to make a contact play against black's shimari first, then extend on the right side (exact location depending on how black answers).
It is true though, that the direct corner enclosures from the hoshis are the top choices for next move.
That jives with my understanding, so I guess I did not express myself very well.
Maybe in looking for another problem I should pick one where I'm pretty sure that the bots will agree that either the consensus choice or the choice of someone like Takagawa or Fujisawa Hideyuki was a definite mistake.
Re: Opening problems for AI: Problem 1
Posted: Sun Jun 14, 2020 4:33 am
by Uberdude
If I wanted to have fun pretending to be Go Seigen...
Uberdude wrote:If I wanted to have fun pretending to be Go Seigen...
Mumonkan wrote:Place yourself in the same freedom as the sky.
You name it neither good nor not good.
Re: Opening problems for AI: Problem 1
Posted: Sun Jun 14, 2020 11:13 am
by sorin
I was thinking yesterday, while reading Bill's new series of threads, that the message I think AI is telling us is: "you can play anything you want, LITERALLY, in the beginning of the game, as long as you follow it up reasonably afterwards, and you don't make any bad mistakes". I know it is not super helpful, but that's what I think I hear it say
Bill Spight wrote:
Uberdude wrote:If I wanted to have fun pretending to be Go Seigen...
Mumonkan wrote:Place yourself in the same freedom as the sky.
You name it neither good nor not good.
Re: Opening problems for AI: Problem 1
Posted: Sun Jun 14, 2020 3:14 pm
by lightvector
KataGo recent stronger unreleased 40 blocks, lots of playouts, analysisWideRootNoise = 0.1 to force evaluation of lots of options. Winrates from black's perspective.
Traditional move is the shoulder hit but we know bots are not too fond of it, definitely not in the beginning. If bots go there, they rather approach A, B, C, D. If elsewhere, they'll enclose one of the UL/LR corner.
I'll try to think about the position without using AI, though I'll inevitably read some of the comments above.
I may be a little out of practice, but I did train quite a bit with Katago in 2020.
Speculatively:
Current position:
1. All the 4-4 6-4 shimaris come to mind first. Last big move in the corners.
2. In the upper left, the attachment is standard human play, but the shoulder hit stands out for me, but mostly to probe whether or not to attach next.
3. "Avoid playing near 3-3 as it is alive" means the lower and left sides are that much less attractive.
4. With direction of upper right shimari, it is traditional human theory to say the right is black's right to expand on whereas the rest of the board is no man's land. Hence a play on the right seems ok.
Past:
Nothing seems strange. The 3-3 is a bit unusual. The small knight's enclosure has come back into fashion with AI. Cross fuseki is entirely normal with AI.
Re: Opening problems for AI: Problem 1
Posted: Fri Jun 04, 2021 11:24 am
by Bill Spight
Wow, Daniel!
Your comments on these problems are quite valuable at this point in time.
Re: Opening problems for AI: Problem 1
Posted: Fri Jun 04, 2021 12:23 pm
by kwhyte
I was surprised to see that the bots are ok with the 4-4 shoulder hit in the lower left. That move is one that I would have played pre-AI but thought I had learned was generally poor.
Re: Opening problems for AI: Problem 1
Posted: Fri Jun 04, 2021 1:48 pm
by Bill Spight
kwhyte wrote:I was surprised to see that the bots are ok with the 4-4 shoulder hit in the lower left. That move is one that I would have played pre-AI but thought I had learned was generally poor.
Well, Honinbo Shusai played it, too, so you're in good company.
is the human textbook answer. The reason for not playing at a, is that White has a follow-up at 
which makes a base, and which also has a follow-up at 
. The reasoning is sound about choosing plays that have good follow-ups that also have good follow-ups. However, I doubt if any of today's top bots give 
