Takemiya's experiment

[John Fairbairn]

There's been a further instalment in the series and this interested me particularly for a special reason.



It was White to play, and in 1985 Takemiya chose A. Nowadays he would prefer B. LZ (which puts Black a smidgeon ahead) says that loses about three percentage points. All the young players also chose A, but Fukuoka Kotaro actually vacillated between that and B, and he did mention come concern about the left side.

LZ, however, had a full percentage point preference for C, which was the move chosen by Hirata Tomoya 7-dan.

It seems that most of the human pros either had a fixation with the visually dominating wall, or underestimated the aji of the cut at D, or just felt too confused by the left side to be able to pick a move. LZ of course has no psychological hang-ups.

The reason this interested me specially is an emerging finding from my work on the Go Wisdom appendixes I am now adding to books of pro commentaries I do. I index every mention of a term that appears in a commentary and discuss the term, usually in great detail, to help readers of the book think about the games with all that extra information at their disposal. But there is fascinating information simply in the index alone. Sheer numbers of occurrences tell us something, and insights can also be gleaned from how early or how late terms occur (e.g. probes on move 4, boundary plays on move 14, fuseki play on move 83, etc).

Although I haven't yet done a proper analysis, one thing that that keeps striking me over and over again is that the three commonest types of mistakes made by pros are as follows:

Three types seem to stand out.

1. Probes. Strong pros commenting on weak pros often chastise them for not making probes.

2. Order of moves. This is usually of the type that can be analysed by tewari. Wrong timing of forcing moves or other intransitive changes is much less common, and less criticised.

3. Direction of play. And this is how I view the example above.

Analysing why pros so often get the DOP wrong (as adjudged by fellow humans) might be a monumental task, but I'm willing to hazard a guess that psychological factors of the kind mentioned above come into it. I think it is also fair to say that their mistakes are normally of a very small order of magnitude, and certainly way below the level in amateur play. But mistakes are mistakes and I'm sure the pros would also like to know why they make them. My early impression is that they haven't found AI much help yet in that regard.

It's going off at a bit of tangent, but here's an "easy" (10-kyu to dan level) direction of play problem by Rin Kanketsu.


)
This is interesting for various reasons. One is that the White invasion on the lower side is scorned by LZ which prefers either of the 3-3 invasions on the right side - less "direction of play" than "different planet!". But maybe Rin chose this position as something weak amateurs would encounter.

LZ initially agrees with Rin that A is the right. B is plain wrong and doesn't even appear on LZ's radar (although E does in that area).

C is rejected as inferior by Rin (and LZ seems to agree) because forcing contact plays are better to hinder sabaki by White.

But the biggest surprise may be that, after a deeper search, LZ's preference (by a full percentage point) switches to C. I find that hard to talk about in DOP terms. But maybe it's a probe, or even an order of moves issue! In the variation LZ shows, Black still gets a wall across the 4th/5th lines, but gets the left corner as well, whereas in the human commentary Black merely encroaches on the corner.

Another reason I gave this extra problem is that it features sanrensei by Black. I have seen comments here that seem to suggest it is now regarded as bad. I obviously can't say whether it is or it isn't, but FWIW pro comments I have read seem to judge as too difficult rather than wrong (like tengen), though LZ seems to think it loses a fraction over one percentage point (yet is that not within the margin of error Bill has talked about, especially so early in the game; and it may be stylistic?) But I am more intrigued by the fact that pros are still playing it.

[Tryss]

In this position, with 90k playouts, Katago prefer D, with 66k visits, 43% winrate and -2 pts. Followed by E (19k visits, 42.5% wr, -2.2pts).

Katago don't like C, it's barely explored, and when played, the evaluation don't change much (even with 20k playouts). Same with A, even if it get 0.5% better than the initial evaluation with 40k playouts.

[Knotwilg]

A thought.

A major difference between humans and bots may be that we may assign a higher value to something that preserves the past as we are loss averse, while the bots only care about the future prospects and have no notion of a particular loss.

In this case, we see White's "invasion", which makes us aware of the potential loss of something that was temporarily Black's. Now I know the Japanese "invasion" is more about dislodging stones than stealing territory, so the Japanese/pro treatment may be closer to the bots' than "our" Western narrow view on invasions. Be what may, for low dans like me, moves A, C and even B, carry the notion of "preserving what's ours" in their evaluation.

The bots are known to think "corners, sides, centre" so they will naturally evaluate D and E first or higher at first, and evaluate A or C for their merits of strengthening the black stones and weakening White's. From this perspective a pure side territorial move like B is understandably omitted from the analysis.

This may explain why K4 is not highly rated to start with: when White invades, K4 becomes a stone to strengthen. Going for D directly instead will strengthen the approach already, establishing a position in the corner.

[Bill Spight]

The reason this interested me specially is an emerging finding from my work on the Go Wisdom appendixes I am now adding to books of pro commentaries I do. I index every mention of a term that appears in a commentary and discuss the term, usually in great detail, to help readers of the book think about the games with all that extra information at their disposal. But there is fascinating information simply in the index alone. Sheer numbers of occurrences tell us something, and insights can also be gleaned from how early or how late terms occur (e.g. probes on move 4, boundary plays on move 14, fuseki play on move 83, etc).

Interesting and valuable work, John.

Although I haven't yet done a proper analysis, one thing that that keeps striking me over and over again is that the three commonest types of mistakes made by pros are as follows:

Three types seem to stand out.

1. Probes. Strong pros commenting on weak pros often chastise them for not making probes.

2. Order of moves. This is usually of the type that can be analysed by tewari. Wrong timing of forcing moves or other intransitive changes is much less common, and less criticised.

3. Direction of play. And this is how I view the example above.

Analysing why pros so often get the DOP wrong (as adjudged by fellow humans) might be a monumental task, but I'm willing to hazard a guess that psychological factors of the kind mentioned above come into it. I think it is also fair to say that their mistakes are normally of a very small order of magnitude, and certainly way below the level in amateur play.

From the Elf commentaries and the AlphaGo games I think that direction of play is one of those concepts that will undergo quite a revision in coming years. By and large, bots don't seem to care about it, or care very much. And sometimes they violate our understanding of it. Often if I have to guess which side a bot will block a 3-3 invasion on, I will now guess that it's the one that does not conform to direction of play (as I understand it). But the preference is usually quite small, so in those cases they don't really care much at all. I think that this is related to the bots' devaluation of the sides. If the sides don't matter much to the choice of play in one corner, how much do the adjacent corners, twice as far away?

Also, in line with Knotwilg's comment, I suspect that a number of pro mistakes stem from the human psychological factor of loss aversion. Bots are much more willing to throw matters up in the air, confident that they will come down better than they are now. Humans like to hang on to what we've got. Amateurs are obviously more prone to this failing than pros, but I suspect, from what obsevation I have done, that pros are more prone to it than bots.

But mistakes are mistakes and I'm sure the pros would also like to know why they make them. My early impression is that they haven't found AI much help yet in that regard.

Taking a look at pro games from last year, I have been impressed at how few fuseki mistakes they now make, according to Elf. That's really quite rapid progress.

One major problem, OC, is that AI choices do not come with language, i.e., with well thought out concepts. Our first reaction is to try to fit them within our current concepts and to modify our current concepts to fit the bots' plays. Children, OTOH, can just pick things up by imitation. A decade from now it will be interesting to hear what pros who grew up with AI have to say. Exciting times!