Thoughts on Honinbo Shusaku

[Knotwilg]

Commentator Sanno describes the game as over or hopeless precisely when KataGo thinks it is very close.

Here again I think we need to defend the human commentators. A very common type of comment is along the lines of "The game is very close but now Black cannot lose." This is because, in their judgement, there are no further areas where mistakes or upsets are likely to occur, and they feel safe in assuming that normal, good play by fellow professionals will guarantee the result, even if it's a win by a mere 1 or 2 points.

If you look at the one but last diagram in the link, KataGo suggests attaching to the lower left in order to attack the top. I can't judge professional capability but that looks complex enough for me for Black to still make a mistake, if the "numbers" say he has a one point lead.

In contrast, in the last diagram there are no more such opportunities and I can agree any pro carries victory home with that position for Black.

So, if I can rely on KataGo, I would say Sanno was a little early announcing victory. Of course KG can be wrong and the humans correctly evaluated the position by 94-96 as a comfortable lead for Black.

[Knotwilg]

Commentator Sanno describes the game as over or hopeless precisely when KataGo thinks it is very close.

As to the repeated assessments that Shusaku gradually loses his initial lead by the middle game, without denying that thats' what the numbers may show, I'd be more inclined to trust the common view that Shusaku opted to invest in thick moves (thick in the endgame sense) with no apparent return in the middle of the game, because he was confident in his ability to turn that kind of investment into cash against the players he usually encountered.

I would be convinced of that explanation if next we saw a gradual point increase with a steady flow of the game. However, in many cases the opponent makes some kind of mistake and it goes downhill from there. I'm more inclined to make psychological inference. Shusaku played thickly and didn't allow for any major upset, however allowed for the game to become potentially close, exhausting his opponents and leading them into risk taking by frustration.

The sample has grown to 6 games. I'll continue to at least 10 and then make a first stab at a bigger scale interpretation.

[Knotwilg]

https://senseis.xmp.net/?InvincibleTheG ... WithKataGo

In this 6th castle game, Shusaku played White against Inoue Matsumoto Inseki and grinded his way to a 2 point victory. Miyamoto Sensei concludes that Matsumoto Inseki has made no mistakes but Shusaku played brilliantly, moving into the lead imperceptibly, the key moment being the skillful settling of his group in the top right.

KataGo indeed doesn't reveal any big point losing mistakes by Black but there are some points in the game where he could have counterattacked to maintain or regain the lead, most notably the cut at B53, which the book dismissed, and the clever play at B115, sacrificing the cutting stones in a flexible manner. Both these moves were rather inventive and required a superior positional judgment. One can't blame Matsumoto for having missed those few occasions of superhuman strength.

The steady graph, with oscillations of at most 2-3 points, confirms Miyamoto's assessment of a "masterpiece".

[John Fairbairn]

I would be convinced of that explanation if next we saw a gradual point increase with a steady flow of the game. However, in many cases the opponent makes some kind of mistake and it goes downhill from there.

Of course, but I think the mistakes are often induced by the style of play, especially a thick style. I say this with the benefit of having just read a commentary on a Go Seigen game where he made a thick shape à la Shusaku (who he studied as a child). His opponent knew he had to choose between a move that counteracted the thickness and a simple tsume (checking move) on the side. But his problem was that he couldn't find a way to evaluate the thickness related move (even though he had lots of time - a two-day game). Basically it was a case of trying to count thickness-associated territory and the collateral effect of the various effect on diverse groups. The sort of thing that is bread and butter for AI.

In the end he chose the thickness-related move but realised within a few more moves that he had made the wrong choice and he resigned after not much more than 100 moves. It seems to me that this is the same trajectory you are describing, with opponents becoming confused or frustrated when facing massive thickness (again, it's the boundary-play type I and the original commentary I'm reading that is being talked about).

I think this trajectory may be the key point that emerges - it's valid strategy if you can exploit just how difficult it is for even the very best pros to evaluate thickness late in the game (i.e. where it is about making territory and eliminating aji rather than attacking on a large scale). Or at least if, like Shusaku, you are better at it than they are.

Another point that I think is worth emphasising is that mistakes are inevitable in ALL human play. It therefore seems legitimate and sensible to factor that into how you play, even if it means a lack of "purity". In his letters and also anecdotes about him, Shusaku appears to be the kind of person who would think deeply about psychological effects. One example is when he was accosted by some brigands but ended up talking to them in such an affable way that they were almost then asking him to rob them!

[Knotwilg]

Castle game 8 is a nightmare. The AI review was very inconsistent and very hard to turn into some kind of reason. Inconsistency is when a sequence of "blue moves" changes its evaluation with several points, hence requires major playouts per move to come to an overall equilibrium. Also, the suggested moves go from one area to another, indicating that some kind of global thickness is key in the game.

Again, the evaluation drops from 6.5 to 2.5 then goes up again towards 6. Commentator Shuko calls the game a success for Shusaku's thickness strategy. At least that is consistent!

So I'm going to change gears here and study the commentary first.

[xela]

Sorry to come late to the party! It's many years since I first read through Invincible. I knew that I'd barely scratched the surface and would need to work through it again one day. Looks like it's time :-)

Back when computers first infiltrated chess analysis, someone coined the term "annotation by result". Pre-computer commentaries (at least for chess) would tend to praise all of the winner's moves, and criticise many of the loser's moves, creating a narrative where the eventual winner of the game was leading all the way. The computers laid bare many reversals of fortunes, and showed that a lot of "easy wins" were not as one-sided as they'd been painted.

Of course temperament is one of the things separating great players from the merely good. The players don't know the computer evaluations during the game. The greats have a deeper well of self-belief which helps them keep looking for resources in a difficult position, complicate things at the right time, snatch victory from the jaws of defeat (as opposed to my own talent for the opposite), ... The computer evaluations don't tell the full story, but I still think it's more interesting if we glance at the decimals enough to notice some of the twists and turns.

[xela]

Castle game 8 is a nightmare...

"Nightmare" is a strong statement! I just spent a couple of hours going through it. Not simple, but surely not that frightening?

For me on a slightly old setup (KataGo v1.12.4 with network KataGo1.12.4-b18-uec.gz), there's quite a lot of moves where Lizzie "changes its mind" in terms of which move is "blue" -- but the overall evaluation (e.g. black 6.2 points ahead or 89% winrate or whatever) is still pretty stable. This suggests to me that there are many opportunities for trades: "if I play on the right side now, then you'll take profit at the top later; if I play at the top, you'll get the right side"...both options are so finely balanced, as KataGo explores the search tree it will prefer now one, now the other, but the margin will never be large either way.

A good example is after move 13, KataGo doesn't like white's cap in response, but can suggest quite a few viable alternatives. I've previously noted the Lizzie comes out in a rash phenomenon: where the algorithm can't find a clear-cut best move, it will widen the search tree a lot.

Similarly at move 30, I was surprised when white tenukied here. KataGo isn't so sure: it doesn't like the game move, but thinks a local response to B29 and a move at the top are about equally good, and keeps flipping between four alternatives.

There's a long stretch from move 33 to 102 where the top centre is unresolved and both players are ignoring it, so you see the score graph bouncing up and down. This sort of thing makes it a harder to interpret the evaluations. (It happens a lot in my own games, both of us overlooking a vital point, and the computer won't give much feedback on the moves actually played, it just keeps reminding me: hey, you really need to sort out that other situation!) Mostly Shusaku had the initiative and Showa never really had the chance to play at the top, but if Shusaku had played K16 earlier, he could have wrapped up the game with less risk (and the computer evaluations would have looked tidier!)

KataGo found what looks like a clear mistake with the lower left reduction: on move 84, it looks like Showa missed an opportunity to make the game close (and the commentary in Invincible doesn't mention this). Also, black 65 seems to be a missed opportunity.

White's move 110 is fascinating if you look carefully at both the score and the winrate, and compare the alternatives. KataGo seems to be making some interesting risk assessments here. It's late at night in Australia right now, so I'll just drop that hint and come back to it another day if anyone is interested...

[Knotwilg]

"Nightmare" is a strong statement! I just spent a couple of hours going through it. Not simple, but surely not that frightening?

I just don't like the game, with all those tenukis and diagonals. Hard to analyze, hard to enjoy.

There's a long stretch from move 33 to 102 where the top centre is unresolved and both players are ignoring it, so you see the score graph bouncing up and down. This sort of thing makes it a harder to interpret the evaluations.

I got another experience: KG kept suggesting Black drawing out his top stone, out of the entanglement. Not only does that look ugly, KG didn't suggest White to forestall it. When the opponent's good point is not your good point, or even the area, I tend to get nervous.

Mostly Shusaku had the initiative and Showa never really had the chance to play at the top ...

Yeah, that's about the only logical explanation for the abovementioned asymmetry.

KataGo found what looks like a clear mistake with the lower left reduction: on move 84, it looks like Showa missed an opportunity to make the game close (and the commentary in Invincible doesn't mention this).

Indeed, interestingly White's improvement chooses thickness above territory. That was a rare enjoyable moment in the analysis

White's move 110 is fascinating if you look carefully at both the score and the winrate, and compare the alternatives. KataGo seems to be making some interesting risk assessments here. It's late at night in Australia right now, so I'll just drop that hint and come back to it another day if anyone is interested...

I don't recall anything that late into the game, or is it about that unanswered small knight's jump in the endgame? Will check later. Please do come back!

[xela]

$$Wc19 Castle game 8, position after black 109, white to play
$$ +---------------------------------------+
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . b . X O . . . . . . . |
$$ | . . . X c . . . . O X O . . O O X . . |
$$ | . . X d . X X X . X O . . O X . X . . |
$$ | . . O O . . O X . X O . . . X . X X . |
$$ | . . . . O . O . O . O . X . . O X O . |
$$ | . . X . . X O . O . . . . . O . O . . |
$$ | . . . O O . . X . . . . . X . . . . . |
$$ | . . X . . . X . X . . . . . . . O . . |
$$ | . . X O . O . X . a X . . . X . . . . |
$$ | . . . X . X X . . O . . O . . X . O . |
$$ | . . X . . . . . . . . O X X . . O . . |
$$ | . . . . . . . . . O . O X . . . . . . |
$$ | . O O . O . X . X O O X . . O . . . . |
$$ | O . O X X X . . . O X X . . . O . . . |
$$ | . O X O O . X . O . O X . X . . O . . |
$$ | . X X . . O X . X . O . . . X . . . . |
$$ | . . O . O . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ +---------------------------------------+

Click Here To Show Diagram Code

[go]$$Wc19 Castle game 8, position after black 109, white to play
$$ +---------------------------------------+
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . b . X O . . . . . . . |
$$ | . . . X c . . . . O X O . . O O X . . |
$$ | . . X d . X X X . X O . . O X . X . . |
$$ | . . O O . . O X . X O . . . X . X X . |
$$ | . . . . O . O . O . O . X . . O X O . |
$$ | . . X . . X O . O . . . . . O . O . . |
$$ | . . . O O . . X . . . . . X . . . . . |
$$ | . . X . . . X . X . . . . . . . O . . |
$$ | . . X O . O . X . a X . . . X . . . . |
$$ | . . . X . X X . . O . . O . . X . O . |
$$ | . . X . . . . . . . . O X X . . O . . |
$$ | . . . . . . . . . O . O X . . . . . . |
$$ | . O O . O . X . X O O X . . O . . . . |
$$ | O . O X X X . . . O X X . . . O . . . |
$$ | . O X O O . X . O . O X . X . . O . . |
$$ | . X X . . O X . X . O . . . X . . . . |
$$ | . . O . O . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ +---------------------------------------+[/go]

No komi; black is a fair bit ahead. Which move would you choose for white?

Here, the game move was a, but KataGo is more interested in b through d: analysing this position, the blue move jumps around. That is to say, a gets a few hundred playouts early on, but then KataGo pretty much gives up on that move and spends most of its later time exploring the other options.

Putting each move on the board in turn and letting it run for 20,000 playouts, I get evaluations of:
a: score B+9.3, winrate 95.4% for black
b: score B+13.6, winrate 90.3% -- this was KataGo's preference after just 2,000 playouts from the previous position
c: score B+11.1, winrate 89.8% -- this one seems to be KataGo's preference after a few more playouts
d: score B+11.6, winrate 90.5%

The details might vary when you re-run, for reasons we've discussed before. (What's after the decimal point is not all that meaningful!) But the overall story is the same. For all of b through d, KataGo is telling us that, compared to the game move, the expected score is worse for white, but the estimated winrate is better. In other words, it can "see" the difference between maximising your score assuming best play versus maximising the chance of the opponent making a mistake and letting you get back into the game. (Am I over-interpreting this?) And in this position, those are two different things, and it's not recommending the move with the best score.

[Knotwilg]

I completed my own review of the 8th castle game and with some extra effort it was already more enjoyable too

https://senseis.xmp.net/?InvincibleTheG ... WithKataGo

Also, Fujisawa Shuko's analysis is more cautious when he seems to be uncertain and spot on on other occasions.

I'll pause these efforts now.

The theme in this series is thickness, thickness, thickness. On the other hand, there's clearly also psychological warfare going on. Shusaku by no means keeps things simple. In this game he seems to invite Showa into a big brawl. Showa resists and takes territory when he feels the pressure mounting. That allows Shusaku to regain his infamous thickness and use it to take victory home in the endgame.