It is not uncommon, even with top level bots with thousands of visits per candidate move, for a human to come up with a move that is completely, or almost completely off the bot's radar, but which, when played, has a winrate estimate within 1-2% of that of the bot's top choice, and is certainly quite playable. Sometimes the bot will give it an even higher rating than the bot's top choice.

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

Visualize whirled peas.

Everything with love. Stay safe.

I want to comment on the five scenarios:
A, B, D are about big mistakes, actually there can be as many as 20 mistakes (both sides) for -5% mistakes, and several -20% ones for kyu players per game. This is where a kyu players can and need to learn and improve. As for whether AI and human teacher agree on the best move, I am not sure this is too much an issue, at least we need to know (1) this is bad move (2) for what reason, even a human teach can suggest more than one best moves, as well as the AI teacher. The thing a human teacher can do better is to suggest a good enough but not too complex move, which an AI teacher cannot yet.
As for C, the small mistakes (-1% to -3%), don't you think this is mostly for higher level players? Isn't this the difference between kyu and dan/pro players?
Therefore, I don't see why AI is not helpful for kyu players.
I do have a human teacher who helps me review games right now, for big mistakes the human teacher and AI teacher tend to agree that it is a big mistake, as for best moves, as discussed above, it's ok to have more than one, but understanding the concept and develop the ability to tell is important. The limitation of a human teacher is that you have to record the session or you might forget what the teacher has said, while the AI teacher can tell you the story over and over again. On the other hand, if the human teacher can "generalize the way of playing under specific situations" it's a plus, but from time to time, I cannot see the generalization and next time I still might not able to tell the right direction to play, or cannot judge the situation correctly. Another thing a human teacher can do but an AI teacher cannot is to recommend actions of the student to take, such as doing some kind of go problems, studying some books or videos.
Another related observation, it seems I have to study about 100 relevant go books before can become a dan player, but I guess many people can improve much faster than this.

About "C", I don't have a concrete example at hand but let me make an artificial one.

Click Here To Show Diagram Code

[go]$$ DDK mistake
$$ ---------------------------------------
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . O X . . . . . . . . . . . . . |
$$ | . . . O . X . . B , . . . . . X . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . O . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . , . . . . . , . . . . . , . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . O . . . . . , . . . . . X . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ ---------------------------------------[/go]

According to 15-block LeelaZero, the last move is a very small mistake, -1.5%. Yet, if a player makes this mistake, I think it is important to correct it because it means he doesn't know what is the ideal extension from a wall, isn't aware of overconcentration problems, etc.

You might think that 1.5% is small, but in a 250-move game, small mistakes add up. Go game would be simple if just correcting big mistakes was enough to improve. Sometimes it's true, especially at DDK level, but my belief is that we need to pay attention to every detail, be patient and accept that a new piece of knowledge will result in an improvement of a tiny fraction of a stone.

My well informed opinion, having studied winrate differences for a few years, is that, even with 100k rollouts per move by a top bot (which is spread over a number of different candidate moves, so that the top choice usually has several thousand visits), we can't say that a winrate difference of 3% or less between the winrate of a play and the winrate of the bot's top choice is a mistake. If the top choice usually has only a few hundred visits, then I think that we require an even greater winrate difference before we think that a move is a mistake.

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

Visualize whirled peas.

Everything with love. Stay safe.

One thing that might help reduce noise would be to analyze the same position using different bots.

_________________
be immersed

IMHO, even better than the 100 books to shodan heuristic is the 1,000 games to shodan heuristic. But there is a caveat. 1,000 quick games against weak players is unlikely to get you very far. It is important to review your own games, something you are obviously doing. It is also important to play against stronger players, so that you do not pick up bad habits from your opponents, and so that your opponents can punish many of your errors. The usual advice is to play against players three stones stronger than yourself.

Good luck!

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

Visualize whirled peas.

Everything with love. Stay safe.

As you probably know, I agree in general. I also sort of agree about this example. I would agree more if the extension were on the 3d line. But the reason I agree in this case is that humans figured out long ago that this extension is too close to the two stones. We know why it is a mistake.

But if all we know is that a top bot thinks that a move loses 1.5% to par, we don't even know that it is a mistake. We can guess why it might be a mistake, bet we don't know. In this particular case we have corroborating evidence from human thought and experience.

Now, my Last Open Corner heuristic, which strongly favors occupying the last open corner (there being exceptions, OC), is largely based upon 1-2% differences. Sometimes you get differences of 10% or more, more often you get differences between 5 and 10%, and even more often you get small differences. Without the corroboration of the large differences, I would not advance that heuristic.

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

Visualize whirled peas.

Everything with love. Stay safe.

What do we teach a DDK about this move?

Click Here To Show Diagram Code

[go]$$Bc DDK move
$$ ---------------------------------------
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . O X . . . . . . . . . . . . . |
$$ | . . . O . X . . 9 a . . . . . X . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . O . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . , . . . . . , . . . . . , . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . O . . . . . , . . . . . X . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ ---------------------------------------[/go]

15 block LZ with some number of rollouts says that this loses 1.5% to par. As it turns out, there is a recent pro game with the plays up to this point, Tuo Jiaxi, 9 dan (W) vs. Lian Xiao, 9 dan, 2018-08-06, which is in the Elf commentaries. This extension did not show up in the commentaries, getting fewer than 1500 rollouts, so I don't know what winrate estimate it got. The extension to a got 5.2k rollouts, for a loss of 1.3%. (Elf gives larger differences than LZ, so maybe LZ gives it something like a loss of only 1%.) Should we tell the DDK why the extension to a is better? Of course. It is established go knowledge and generalizes to other situations.

But Lian Xiao played in the next diagram, and Elf agrees, with 26.7k rollouts. I also agree.

Click Here To Show Diagram Code

[go]$$Bc Elf's top choice
$$ ---------------------------------------
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . O X . . . . . . . . . . . . . |
$$ | . . . O . X . . . a . . . . . X . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . O . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . 9 . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . , . . . . . , . . . . . , . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . b . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . O . . . . . , . . . . . X . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ ---------------------------------------[/go]

Elf also likes b, with a "loss" of 0.1% and 11.9k rollouts. Should we teach the DDK those moves, as well? I say yes. Go is not for the faint of heart.

But is a playable? Well, Elf prefers the other plays, but not by enough to say that a is a mistake. And in another game vs. Tuo, 2018-05-14, Zhao Chenyu, 6 dan, played that very extension.

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

Visualize whirled peas.

Everything with love. Stay safe.

Ke Jie:

Click Here To Show Diagram Code

[go]$$B
$$ ---------------------------------------
$$ | . . . . . . . . . . . . . . . . . O . |
$$ | . . . . . . . . . . . . O . . . . O X |
$$ | . . . . O X . . B . . . . O . O O X . |
$$ | . . . O . X . . . , . . . . X X X . . |
$$ | . . . . . . . . . . . . O . . . . X . |
$$ | . . . O . . . . . . . . . . . . X . . |
$$ | . . . . . . . . . . . . . O . . . X . |
$$ | . . . . . . . . . . . . . . . O O O . |
$$ | . . . . . . . . . . . . . . . . X . . |
$$ | . . . , . . . . . , . . . . O O X . . |
$$ | . . . . . . . . . . . . . . . X . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . X . . . . . . . . . . . . . . . . . |
$$ | . . . O . . . . . . . . . . . X . . . |
$$ | . . X O . . . . . , . . . . . , . . . |
$$ | . . X O . . . X . . . . . . . X . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ ---------------------------------------[/go]

Cho Chikun:

Click Here To Show Diagram Code

[go]$$B
$$ ---------------------------------------
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . X . . . . . . . . . . . . . |
$$ | . . X X O X . . . . W . . O X . . . . |
$$ | . . O O . . . . . , . . . O . X . . . |
$$ | . . . . . X . . . . . . . . . . . . . |
$$ | . . . O . . . . . . . . . . . X . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . O , . . . . . , . . . . . , . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . O . . . . . . . . . . . . X . . |
$$ | . . . O X . O . . . . . . . . . . . . |
$$ | . O . O . . . . . , . . . . . X . . . |
$$ | . X X . . X . O . . . . . O . . . . . |
$$ | . . . X . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ ---------------------------------------[/go]

I think it is playable

also Ke Jie:

Click Here To Show Diagram Code

[go]$$B
$$ ---------------------------------------
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . O X . . . . . . . . . . . . . . . . |
$$ | . O X O . . . . . . . . . O X . . . . |
$$ | . O X O . . . . . , W . . O . X . . . |
$$ | . X O O . . . . . . . . . . . . . . . |
$$ | X . X . . . . . . . . . . . . X . . . |
$$ | . X . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . , . . . . . , . . . . . , . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . O . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . O . . . . . , . . . . . X . . . |
$$ | . . . . . X . . . . . . . X . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ ---------------------------------------[/go]

Dwyrin was afraid to play a three space extension because of his 9D Pro opponent. Instead of tenuki he could have played like Ke Jie or Cho Chikun. But he could not manage to do so, because of the Know-it-all's advice to never play a two space extension.

Ke(1) and Cho are short being defensive. Ke(2) is short creating thickness reducing Black's sphere of influence. They are different from the early opening position, in which overconcentration is considered bad.

Before I buy a new PC with GPU, I'd like to know if I use Leela Zero correctly.

Is my notebook computer (Intel i5 2.2G, Windows 10 64 bit, 4G RAM) really too slow in terms of visits per seconds?

I have tested it with Lizzie, the following results in 3-5 vps and 12 vps only:

C:/leela-zero-0.17-cpuonly-win64/leelaz.exe -g -b 0 -w C:/leela-zero-0.17-cpuonly-win64/266_40b_15926.gz

C:/leela-zero-0.17-cpuonly-win64/leelaz.exe -g -b 0 -w C:/leela-zero-0.17-cpuonly-win64/117_15b_10214.gz

and Some people told me should be 30 vps or higher, at least for 15 blocks.

Is there any parameter I can try? Thanks again.

If you have an i5 5200u 3.5 playouts a second seems about right. Even 3.5 visits seems ok. if you can use opencl, try using the gpu version https://software.intel.com/en-us/node/540387

_________________
set katago to play at your level https://docdro.id/sHZU1ti or experiment with gtp4zen ( https://rb.gy/kx2ilb )

What do LZ and KataGo say about the close extensions in these cases? It matters what is in the adjacent corner.

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

Visualize whirled peas.

Everything with love. Stay safe.

For the last position, 15-block Katago likes Ke Jie's move:



but other moves are playable as well, everything is within the margin of error (<1%).

15-block LeelaZero prefers the normal low extension, but thinks that Ke Jie's move is playable as well (-0.5%).

P.S. My notebook computer is CPU-only: Intel Core i7-8650u CPU @1.90GHz 2.11 GHz, 32G RAM, 64 bits. On this particular position I get about 80 playouts/second.

Careful with this thinking. In Katago review I noticed it often shows my game move as an acceptable alternative with only slightly lower winrate (like 49% instead of 51%). However, going forward in the game, both me and opponent plays the bots top choice for 1-2 more moves it suddenly drops by 10% or so. So its evaluations doesn't seem reliable for non-top moves (less so than LZ 15b at least).

I know this can happen in complex fights, because some branches are not explored enough at the first move, but in the particular case I explored above, there is no such complex fight, and alternative moves are explored with several thousand visits, so there is no risk of a sudden 10% drop.