Life In 19x19

I don't know whether T Rex played go or not. This title is just to make this thread stand out. There are so many AI-related threads now that I didn't have the mental energy to think of a distinctive yet accurate title.

However, it has a smidgeon of accuracy in that it relates to a T Rex level player of the recent pre-AI past pitted against lesser species armed with AI knowledge.

The fossil record is as follows (Black to play):



This is from a Go Weekly series I have mentioned before here, in which Hirata Tomoya takes a move from a past game, asks the player how he would play now in 2020, and also asks a clutch of players, ranging from 9-dan to 1-dan, how they would play in the AI age.

It's a small sample, so a purely statistical view is not appropriate, but a very significant aspect is that the players explain their choices, and in some cases offer a second or third choice.

T Rex was Cho U here, playing Yuki Satoshi in a Gosei title match in 2009. The reason this episode stood out for me is that all the other players got the right area of the board (as per AI) but all had quite different views about the right pace to play there. It was also interesting that Cho, who won the game as Black, essentially stuck to his guns.

And the reason I want to highlight this reason is that I think it illustrates an approach by Japanese pros to investigating AI that is radically different from the one typical on L19.

I may be wrong in my impression about the tone of L19, but it comes over to me as primarily a concern with finding the best move and determining how much better that is then other moves. It is mostly move-specific. In the case of the pro writings (of which I have seen quite a lot now), I have an impression that the investigation is chiefly area-specific. It's a sort of pruning process before they get down to choosing a move. An alternative interpretation may, of course, be that they do start simply looking for best moves, and they all just happen to end up in the same area. But, either way, it seems clear to me (from this series) that they are having more success in finding the right area than the right move.

And even if my impressions are right, I'm not saying that one approach is automatically better than the other. Pros are starting from a quite different place from amateurs. They can shop on the posh first floor. We lesser mortals have to shop on the ground floor. Still, I do wonder whether a rather more area-specific approach might benefit amateurs. This example may offer some food for thought.

I suggest, as part of the experiment, that you think about the above position first in terms of which area to play in.

I picked the move the bots liked (C) but I did it for reasons totally in the vein of pre-bot era thinking. I rejected the lower side because the low strong white stone in the lower left makes the lower side small for both players. As for the upper left, a play by white would be big territorially and a black play at C erases white's potential while helping to develop the upper side for black. I think I considered the upper left area before choosing the specific move.

Thanks, again, John, for an interesting post.

Since the problem position has at least two stones in each corner, my guess is that there will not be much difference between the AI choices and that of human pros. OTOH, since Hirata chose this position for his survey, my guess is that there is some difference that he noticed that is of some significance to him. Still, it may be that the difference falls within the margin of error for the bots' evaluations, but AFAICT, pros are not concerned with the margin of error.

I am going to check and see how much the human play differs from par, according to the Elf commentary. I am going to guess that it lies between 4% and 10%, i.e., from a bit within what I think Elf's margin of error to be to in the range of a minor error. We'll see if I'm right.

Oh, dear, wrong again.

Elf reckons that Cho U's play lost 11½% to par. Elf's choice was the same as that of the other bots, with 32.3k rollouts. The human play got 0 rollouts, but its evaluation is based upon Elf's reply to it, which got 27.6k rollouts. The difference is greater than 10%, so I owe Hirata an apology.

John Fairbairn wrote:This is from a Go Weekly series I have mentioned before here, in which Hirata Tomoya takes a move from a past game, asks the player how he would play now in 2020, and also asks a clutch of players, ranging from 9-dan to 1-dan, how they would play in the AI age.

It's a small sample, so a purely statistical view is not appropriate, but a very significant aspect is that the players explain their choices, and in some cases offer a second or third choice.

T Rex was Cho U here, playing Yuki Satoshi in a Gosei title match in 2009. The reason this episode stood out for me is that all the other players got the right area of the board (as per AI) but all had quite different views about the right pace to play there. It was also interesting that Cho, who won the game as Black, essentially stuck to his guns.

Another example, perhaps, of starting from where you are.

John Fairbairn wrote:And the reason I want to highlight this reason is that I think it illustrates an approach by Japanese pros to investigating AI that is radically different from the one typical on L19.

I may be wrong in my impression about the tone of L19, but it comes over to me as primarily a concern with finding the best move and determining how much better that is then other moves. It is mostly move-specific. In the case of the pro writings (of which I have seen quite a lot now), I have an impression that the investigation is chiefly area-specific. It's a sort of pruning process before they get down to choosing a move.

As a budding bridge player, I encountered lists of opening leads telling which card to lead from which suit holding, with some holdings being good to lead from, and other holdings being bad. I later learned from Alfred Sheinwold that that approach was wrong. The right idea is to decide first which suit to lead from, and then to decide which card to lead. The main criterion for choosing a suit was the contract and the bidding, with the holding in the suit being a lesser criterion.

Also, as I recall, Lasker, in his Chess Primer, uses the metaphor of attacking a castle to indicate that strategy is not specific, but general, until you get close to the castle (checkmate).

And in golf, if you have a long putt, the advice I was given was not to try to sink it, but to get within 2 feet of the hole (margin of error).

All of which is to say that going from the general to the specific is a successful human heuristic.

As for today's bots, apparently their search is move specific over the whole board, using a form of best first search. This does not mean that there is nothing like first forming a broad, fuzzy focus on areas of the board by the deep neural networks. None of those inner workings are reported to us, so we don't know.

John Fairbairn wrote:An alternative interpretation may, of course, be that they {the humans} do start simply looking for best moves, and they all just happen to end up in the same area. But, either way, it seems clear to me (from this series) that they are having more success in finding the right area than the right move.

I also decided on the top left corner. I was not sure whether it was better to approach the corner or invade it, so I split the difference with the attachment at D-17.

John Fairbairn wrote:And even if my impressions are right, I'm not saying that one approach is automatically better than the other.

Oh, I think we humans are using an approach that is good for us. Searching the whole board is beyond us, except in the sense of performing a number of local searches, and doing a kind of fuzzy search of areas, not specific points. Sometimes a particular move will stand out, but if it doesn't, starting with areas instead of points is a good idea.

John Fairbairn wrote:I was struck by how weird the variation after G looked. Was this weirdness what put the pros off this line? Obviously I can't evaluate this weirdness myself, but pros write game commentaries and insert variation diagrams, and I can tell that this AI diagram looked like nothing I see there.

Weirdness is in the eye of the beholder. FWIW, Elf's mainline variation for G did not seem weird to me. But I think it would have before AlphaGo. Here it is.

I mentioned above a case where a group of pros got the right area as recommended by AI but got flummoxed by the precise spot to play on. This week's follow-up was an amusing contrast.



The pros this week had obviously been reading L19 and were almost unanimous in realising that the press at A would be the move the bots like.

Well, no... Some pros equivocated between A and B, one even going so far as to say it was a toss-up.

That's probably true, but Hirata's bot (unnamed) chose B. My version of Leela originally chose B but after more thought plumped for C. Katago also gave me B. But A, which Cho U chose in real life - the game is GoGoD 2010-02-03b), is only marginally inferior in bottish terms. B is, however, another good example of my new proverb: "a 4-4 needs a reinforcement".

Apart from the amusing contrast, other reasons for mentioning this episode are:

1. The game is a famous example of winning a 4-liberty approach ko as the weaker side and still winning the game. It will probably be new to newer players.

2. Hirata answers a question often posed by amateurs: What is the best book to read about ko? He says it is Cho U's "How tp thnk about k" which he has read "thoroughly". It is unusual for pros tp own up to reading books intended for amateurs!

John Fairbairn wrote:I mentioned above a case where a group of pros got the right area as recommended by AI but got flummoxed by the precise spot to play on. This week's follow-up was an amusing contrast.



The pros this week had obviously been reading L19 and were almost unanimous in realising that the press at A would be the move the bots like.

Well, no... Some pros equivocated between A and B, one even going so far as to say it was a toss-up.

Bully for him.

That's probably true, but Hirata's bot (unnamed) chose B. My version of Leela originally chose B but after more thought plumped for C. Katago also gave me B.

Elf in the GoGoD commentaries also picks B.

But A, which Cho U chose in real life - the game is GoGoD 2010-02-03b), is only marginally inferior in bottish terms.

Indeed. Elf reckons its loss to B as only 2½%, well within Elf's margin of error.

B is, however, another good example of my new proverb: "a 4-4 needs a reinforcement".

Apart from the amusing contrast, other reasons for mentioning this episode are:

1. The game is a famous example of winning a 4-liberty approach ko as the weaker side and still winning the game. It will probably be new to newer players.

I took a quick look. Interesting that Yamashita took the ko at move 163 instead of filling a liberty.

2. Hirata answers a question often posed by amateurs: What is the best book to read about ko? He says it is Cho U's "How tp thnk about k" which he has read "thoroughly". It is unusual for pros tp own up to reading books intended for amateurs!

Cho U's books seem to be quite good. Cho U no Tsumego even got rave reviews from DDKs, who, one would have thought, would have been in over their heads.