Interesting Analysis of Ancient Top Professional Players

[illluck]

As a side note, my personal take on the results is that while the methodology appear to make sense and source sgf/analysis are published, it is quite surprising. I'm of the camp that the modern pros are likely stronger due to a much more structured (and likely earlier and more intense) learning process in addition to having multiple very strong opponents to motivate each other. The two pieces that I am wondering about relate to whether "accuracy" is a good enough indicator of strength (given the fairly limited number of the ancient games) and whether "accuracy" between komi and no komi games can be compared directly.

Nonetheless, I have to say that my opinion of the historical players has definitely been changed by the paper (at the very least, the results indicate somewhat comparable strength).

[John Fairbairn]

Nonetheless, I have to say that my opinion of the historical players has definitely been changed by the paper (at the very least, the results indicate somewhat comparable strength).

Over the years I have seen many flattering comments about the skill of old players. I tended to brush them off as either polite remarks or as nationalistic jingoism, despite the fact that I think it's a no-brainer that Leonardo da Vinci was a better painter than, say, David Hockney (whom I admire BTW).

But when I was starting my Museum of Go Theory project, I took a closer look at these comments, and invariably found that they had real substance to them. At around the same time, various people had started running AI tests on old players, and although it would be rash to call them scientific, they too have tended to corroborate the early opinions (of people, like Go Seigen, I might add!).

There are special conditions to do with conditions for go in China in the early Qing (i.e. the Huang/Fan/Shi era), such as the wealth of Yangzhou and its go sponsors), and the peripatetic nature of go masters, spreading the game nationwide in a way that did not happen to the same extent in Japan. There was also the sheer number of players, who did not rig the results of games, as Shusaku did with his Castle Games - most Castle Games were controlled in some way (e.g. "too ill to play, take my pupil instead"). In China the competition was much more honest and, I suggest, forced up the skill level. No-one has yet come up with a convincing explanation of why the level of go in China dipped after Fan and Shi, and it is a shame that too many people (even in China) judge the overall level of old Chinese go by its late Qing standards. This paper seems to confirm that we should be basing our judgements on early Qing.

[Kirby]

Go is a zero sum game. All that matters is the win or loss. Playing an opening that an AI likes may have some correlation to winning the game, but it's by no means a pure indicator.

So an analysis like this is cute, and shows how well some pro openings compare to AIs, but it doesn't do much to indicate "who was stronger". The only sure way to do that is to have direct competition between the pros in question.

You can have a 10k that memorizes a bunch of pro openings, and that 10k can play like a pro for awhile. But that 10k isn't a pro. Have him play an even game against a pro, and that'll become clear very quickly.

[Mike Novack]

Something else to consider, best practical play vs best theoretical play. By which I mean taking into account who is ahead, who is behind, how much time does each player have to work with, etc.

If you are behind, contemplating two alternative plays, A and B, where A is slightly better than B but is simple, few ways for the opponent to go wrong, B may be the better practical play. On the other hand, if ahead, a slightly inferior but simple A might be the better practical choice than a slightly better but very complex line. How much time left relates to this. If one player has lots of time but the other is in time trouble, the better practical choice for the first player might be a very complicated line even if slightly inferior to a theoretically better simple line.

I do not think even our best AIs are yet able to take these factors into account but are returning best theoretical play assuming everything else is equal. But typically NOT equal. Usually one player or the other is ahead and one has more time than the other. Since the ancient pros would be taking these things into account and the AI not, comparing might be misleading.

Or am I wrong? Do we have an AI returning the best PRACTICAL move?

[emerus]

Something else to consider, best practical play vs best theoretical play. By which I mean taking into account who is ahead, who is behind, how much time does each player have to work with, etc.

If you are behind, contemplating two alternative plays, A and B, where A is slightly better than B but is simple, few ways for the opponent to go wrong, B may be the better practical play. On the other hand, if ahead, a slightly inferior but simple A might be the better practical choice than a slightly better but very complex line. How much time left relates to this. If one player has lots of time but the other is in time trouble, the better practical choice for the first player might be a very complicated line even if slightly inferior to a theoretically better simple line.

I do not think even our best AIs are yet able to take these factors into account but are returning best theoretical play assuming everything else is equal. But typically NOT equal. Usually one player or the other is ahead and one has more time than the other. Since the ancient pros would be taking these things into account and the AI not, comparing might be misleading.

I think this is generally where the very top players excelled. Though I think that meaningful positional judgements tend to occur later in the game (maybe median around move 120). As mentioned Lee Changho and Shusaku might be counter examples who played a bit closer to underplay even early in the game. Many of the other tops pros (think Lee Sedol, Cho Hunhyun) were more known for finding more practical ways to shake a losing game up.

Or am I wrong? Do we have an AI returning the best PRACTICAL move?

Kata does a mix. It can avoid capturing races in won games and I believe that Kata's more risk taking play in reverse komi and handicap games is evidence of a practical approach as well. Time control is probably irrelevant to classical players. Definitely a factor in modern Go though.

[go999999]

I found another article, ranking ancient China and ancient Japan. http://home.yikeweiqi.com/#/gonews/detail/49602?type=1