It is currently Thu Mar 28, 2024 11:47 am

All times are UTC - 8 hours [ DST ]




Post new topic Reply to topic  [ 9 posts ] 
Author Message
Offline
 Post subject: 柯洁能下得过古代的围棋高手吗?
Post #1 Posted: Mon Jan 04, 2021 12:59 am 
Dies with sente

Posts: 117
Liked others: 21
Was liked: 22
(转发)
现在人类的顶尖围棋棋手例如柯洁能下得过古代的围棋高手吗?
作者:BionicMed

长技术贴预警!

先说结论:在棋力上近乎严格相等。范、施在古规则下即有远超普通九段的棋力。

当我们在说“下的过”的时候,我们在说什么?大气磅礴的布局招法?缜密的中盘构思?锱铢必较勾心斗角的官子定型?亦或是场外的故事性,传奇性戏剧性造就的棋圣传说?
很显然都不是。棋就是棋。唯一能评价棋力的,只有更强的棋力。

那么请问什么样的棋力能评价如范施二圣级别的高手呢?业3业4的“贴吧大神”?冲段少年、地方豪强?亦或是职业?如果一位初段的职业推崇古法,一位高段职业对其不以为然,请问后者就一定对吗?

显然,只有棋力最强的棋手,才能得到最精确的答案。我有幸参与而今市面上最强的开源围棋AI——Katago官方权重的续练,并耗时两周,通过计算机每一步十万的模拟量逐个分析了柯、范施、大小李与古力的40张棋谱,力求得到偏差最小的结论。
KatagoS289B40F384与S571B40F256的纯棋感已有职业水平。在十万计算量下,超越了人类的四十段棋力。使用LizzieYzY分析软件,我们可以得出以下分析结论:
Attachment:
1.png
1.png [ 99.6 KiB | Viewed 6612 times ]

(图一)

说明1:当湖棋谱2、柯洁棋谱1 、5,软件出现问题,故不采用数据;
说明2:当湖10局中不确定度最高为45.6,平均最高为26.9,柯洁九段对局中最高为33.6,平均为21.4;
说明3:当湖分析为古棋规则+贴7.5目,以避免虚高;当代为自适应规则;分析计算量为10万/步。

分析结论:
总体来说,柯洁九段在当代规则下全盘评分为66,比范圣的总评62.4略强一些,比施圣的59.7分强出一个档次。
抛开布局,柯洁的中盘评分在58左右,略弱于范圣,略强于施圣。

从高阶数据分析,有开局60手的85%一选的加成,柯洁的正确棋感(一选)比范圣高百分之1.5,施圣高百分之2.8,差距较小,抛除开局因素则略弱于二圣。
而相比之下,7%的三选差距则说明如柯洁九段等当代顶级棋手算路精准,由于长期的训练与竞技犯错相对较少。
然而事实真的是如此吗?从不确定度我们可以看出,古棋的战斗更为激烈,而在这种背景下范、施的恶手所损失平均目数都比柯洁略少。因此究竟谁犯错更少,依旧有待商榷。

在没有AI的年代,二圣能在十番快棋中下出总评不显著弱于柯洁九段的数据,甚至中盘与算路准确程度的数据更是并驾齐驱,这无疑是令人震惊的。在局面较当代显著更复杂的前提下达到这一数据,无疑有力地驳斥了“范施业5论”。
刨除关公战秦琼的“范施学习AI”假设,仅从棋的内容看,两圣已较后AI时代世冠平均水平,有过之而无不及。而面对前AI时代的诸如聂马,六超,大小李等一众高手,则数据多有超越。

得到此结论后,有许多棋友反驳我的结论,我将反驳归纳为以下三点。也欢迎读者提出其他建议,共同探讨。
1、古棋理论多有“玄学”内容,棋理不确。
2、古人虽力大无穷,但发力点不对。
3、战斗较多的棋评分虚高很多。

为了驳斥这三条论点,我额外找来前AI时代人类棋理巅峰李昌镐九段的十局,人类发力精华、鬼手集大成者李世石九段的十局,人类力量巅峰古力九段的十局与棋友一同探讨。
Attachment:
2.png
2.png [ 98.77 KiB | Viewed 6612 times ]

(图二)


分析样例:
Attachment:
3and4.png
3and4.png [ 546.46 KiB | Viewed 6612 times ]

(图三)(图四)

常常听老一辈讲棋说,李昌镐九段的官子是天下一绝,无中生有的本事无出其右。然而,根据AI数据显示,在与下一代的“绝代双骄”——古力九段、李世石九段对比时,李昌镐九段不但官子评分最低,而且平均每步损失的目数也并不怎么少。数据显示李昌镐九段领先天下的也许不是官子,而是中盘战斗。

值得一提的是,在前AI时代总评超过60分的,仅有古力九段一人。无论是从重合度、开局、中盘来看,古力九段都遥遥领先。也许从AI角度来看,在古李十番棋及其他比赛中,古力的棋的内容比李世石九段稍优,最小的目差损失与超高的一选重合无不反映出其胆大技细、直觉敏锐的技术特点。

另一个反直觉的数据显示:李世石九段的官子在三者中较优,开局则与李昌镐九段不相伯仲。从短期趋势来看,李昌镐九段为“上上一代”的人类巅峰,而小李、古力则多有超越,为“前AI时代”人类的最顶峰,因此说“围棋是发展的”并无谬误。

然而,从长期趋势来看,在AI出现后,人类的巅峰水平提升也许并没有想象中那么多。面对范西屏的数据与古力九段并驾齐驱,施襄夏的数据与二李相似这一事实,也许我们应该重新思考围棋发展的结论,或许围棋在数百年间的提升并不显著,仅有少部分天才在缓慢地将人类的极限进行些微的延申,而又被更多有谬误的“进步思想”开历史倒车。

回想起围棋论坛、贴吧,乃至知乎对于古棋二圣(范西屏、施襄夏)多有贬低,依旧觉得有些许可笑。不知以战斗闻名的二圣在数百年前就已经有最顶尖人类(无论AI出现前后)的水平,而将其打成“顶多业5”,实际却可让地方豪强三四子否?抨击古棋开局“不知所云”的业3、业4选手又是否想到二圣开局在AI来看超越大小李呢?虽说眼高手低是手眼协调的必要步骤,“看棋高三段”也是进步的常备动力,但面对未知,也许保持一个开放的观点方能更接近真相。

不知如棋界许多对古棋不了解的人,言之凿凿“古人不可能超过业6,且与其天赋无关”、“古圣必须被让子,业7称霸天下”的“棋史学家”看到在没有经过AI熏陶,没有公开讨论竞争氛围的前提下,二圣仅通过棋才与摸索便实现了对前AI时代最巅峰人类数据的准全包围,并与学习了AI招法后的人类最巅峰柯洁九段在纸面上完全有一战之力时,又该作何表示?数据固然不是棋力,完全达到最强九段数据水平,且与AI素未谋面的古圣又怎么可能仅有业5、业6水平?

注:所有当代棋谱皆选择棋手巅峰时期对局,且大都为大赛决赛有质量番棋。私信作者可获得分析原文件。

_________________
Zhang-hu 章浒
Committed to the restoration Chinese traditional Weiqi
Research on ancient Weiqi rules & Classic (Dunhuang Classic and the Thirteen Chapters Classic)
From Suzhou, Jiangsu Province, China


Last edited by pgwq on Mon Jan 04, 2021 5:11 am, edited 2 times in total.
Top
 Profile  
 
Offline
 Post subject: Re: 柯洁能下得过古代的围棋高手吗?
Post #2 Posted: Mon Jan 04, 2021 3:58 am 
Oza

Posts: 3647
Liked others: 20
Was liked: 4626
Very well done! Although I'm sure we will eventually need a rigorous analysis, by people more used to dealing with statistics than I am, of your method and your interpretation of the results, I already feel confident in saying that your claim that the ancient players are worthier of much more respect than they usually get is one that several of us on this forum will find easy to accept. The Japanese have done a similar exercise for players going back to Honinbo Dosaku and have reached a similar conclusion, and some of us here have done similar things on a small number of games for individual players. The results seem similar to yours, but your exercise goes into impressive detail.

Indeed, I think that in some ways it is easier to make the claim for the old Chinese players, for two reasons. One is that you appear not to have made an adjustment yet for group tax. If that's correct, I think you will find that they score much better (especially, but not just, in the opening) than your figures show. The other reason is that in the case of old Japanese players we have no commentaries or theory books written by the players themselves, and so it is difficult to know how the players thought. In the case of China we have both hundreds of commentaries and theory books. We not only can see how they thought but we can also see that they thought in ways which seem to find a remarkable parallel in AI play. To give one example that is fascinating me at the moment and which I have studied in some detail, there is the term 照应 which is especially common in 不古编 (published in 1682!!). For the benefit of readers not familiar with these commentaries, this refers to a concept of proffering a helping hand to friendly groups, but since the helping hand extends over the whole board and in multiple directions, it works in a very AI-like way. Perhaps a better way of looking at it is like setting up a lighthouse (照 = to shine) offering help to ships in distress, or just plain reassurance, on the vast ocean of the board.

I hope it also needless to add that the enormous technical skill of the old Chinese players has been recognised even outside China. Apart from the tsumego problems of Xuanxuan Qijing, for example, their skill in actual games has been recognised. One good example is 中国古棋譜散歩 (A promenade through old Chinese go manuals) by Watanabe Hideo 7-dan.

What many people here would like to see now is your analysis of Go Seigen (Wu Qingyuan). He is not so ancient, but he did study old Chinese manuals and many people, both professional and amateur, have noticed that he too played in an AI way.

Top
 Profile  
 
Offline
 Post subject: Re: 柯洁能下得过古代的围棋高手吗?
Post #3 Posted: Mon Jan 04, 2021 5:54 pm 
Oza
User avatar

Posts: 2401
Location: Tokyo, Japan
Liked others: 2338
Was liked: 1332
Rank: Jp 6 dan
KGS: ez4u
I guess that more readers here would be interested in an English version of the analysis than any expansion of the OP.

_________________
Dave Sigaty
"Short-lived are both the praiser and the praised, and rememberer and the remembered..."
- Marcus Aurelius; Meditations, VIII 21

Top
 Profile  
 
Offline
 Post subject: Re: 柯洁能下得过古代的围棋高手吗?
Post #4 Posted: Tue Jan 05, 2021 4:47 am 
Gosei
User avatar

Posts: 1810
Liked others: 490
Was liked: 365
Rank: KGS 1-dan
Google Translate wrote:
(Forward)
Can today's top human Go players, such as Ke Jie, have played ancient Go players?
Author: BionicMed

Long technical post warning!

Let me talk about the conclusion first: almost strictly equal in chess power. Fan and Shi are far superior to ordinary nine-duan chess under the ancient rules.

What are we talking about when we are talking about "the next passing"? The majestic layout strategy? A meticulous mid-market idea? Qunzhu must be more stereotyped than intrigue officials? Or is it a story off the court, a legendary chess legend created by legendary drama?
Obviously neither. Chess is chess. The only thing that can evaluate chess power is stronger chess power.

So what kind of chess ability can be evaluated as a master of Fan Shi Ersheng level? Industry 3 Industry 4 "Post Bar God"? Chong Duan boy, local hero? Or is it a profession? If a junior profession respects ancient law and a senior profession disagrees with it, is the latter necessarily correct?

Obviously, only the strongest chess player can get the most accurate answer. I was fortunate enough to participate in the continuation of the official weight of Katago, the strongest open source Go AI on the market. It took two weeks to analyze the simulations of Ke, Fan Shi, Xiaoli and Coulee one by one through computer simulations of 100,000 each step. 40 chess records, and strive to get the least deviation conclusion.
KatagoS289B40F384 and S571B40F256 have a professional sense of pure chess. With one hundred thousand calculations, it surpassed the forty stages of human chess. Using LizzieYzY analysis software, we can draw the following analysis conclusions:
Attachment:
1.png
1.png [99.6 KiB | Viewed 269 times]

(Picture 1)

Note 1: When the Hu chess record 2, Ke Jie chess record 1, 5, there is a problem with the software, so the data is not used;
Note 2: In the 10 rounds of Danghu Lake, the highest uncertainty was 45.6 and the average highest was 26.9. Ke Jie's 9th-dan rounds had the highest uncertainty of 33.6 and the average was 21.4;
Note 3: When the lake analysis is ancient chess rules + paste 7.5 meshes to avoid false height; contemporary is adaptive rules; analysis calculation is 100,000/step.

Analysis conclusion:
In general, Ke Jie's 9th Duan score is 66 under the contemporary rules, which is slightly better than Fan Sheng's overall rating of 62.4, and one grade better than Shi Sheng's 59.7 points.
Regardless of the layout, Ke Jie's mid-game score is around 58, which is slightly weaker than Fan Sheng and slightly stronger than Shi Sheng.

From the analysis of high-level data, there is a bonus of 85% of the first choice for the opening 60 hands. Ke Jie's sense of correctness (first choice) is 1.5% higher than that of Fan Sheng, and Shi Sheng is 2.8% higher. The gap is small. The opening factor is slightly weaker than the second holy.
In contrast, the 7% difference between the three choices indicates that contemporary top players such as Ke Jie 9duan have accurate calculations, due to relatively few mistakes due to long-term training and competition.
But is this really the case? From the uncertainty, we can see that the battle of ancient chess is more intense, and in this context, Fan and Shi's evil hands lose slightly less average points than Ke Jie. So who makes fewer mistakes is still open for discussion.

In the era without AI, it is undoubtedly shocking that Er Sheng was able to play in Jufan Kuaiqi with data that was not significantly weaker than Ke Jie's 9-dan, and even the accuracy of the mid-game and the calculation of the road were even more closely matched. Reaching this data under the premise that the situation is significantly more complicated than contemporary times, undoubtedly strongly refutes the "Fan Shiye 5 Theory".
Excluding the hypothesis of "Fan Shi learns AI" by Guan Gong and Qin Qiong, only from the content of chess, the two saints have exceeded the average level of the world championship in the post-AI era. In the face of the pre-AI era, such as Nie Ma, Liu Chao, Xiao Li and other masters, the data is beyond.

After reaching this conclusion, many chess players refuted my conclusion. I will summarize the refutation into the following three points. Readers are also welcome to make other suggestions and discuss together.
1. Most ancient chess theories have "metaphysics" content, and the theory of chess is not correct.
2. Although the ancients were infinitely powerful, they were wrong.
3. The scores of chess with more battles are much higher.

In order to refute these three arguments, I additionally found the ten rounds of the 9th dan, the peak of human chess in the pre-AI era, the 10 rounds of the 9th dan, the essence of human power and the master of ghost hands, and the ten rounds of the 9th dan of the peak of human power, Gu Li, together with chess friends. Explore.
Attachment:
2.png
2.png [98.77 KiB | Viewed 269 times]

(Figure II)


Analysis sample:
Attachment:
3and4.png
3and4.png [546.46 KiB | Viewed 269 times]

(Picture 3) (Picture 4)

I often hear the older generation talk about chess and say that Li Changhao's 9th dan official is the best in the world, and he has nothing to do with his skills. However, according to AI data, when compared with the next generation of "peerless double arrogance"-Gu Li 9th Duan and Li Shishi 9th Duan, Li Changhao 9th Duan not only has the lowest official score, but also the average number of meshes lost per step is not that small. The data shows that Lee Chang-ho's 9th-dan lead may not be the official, but the mid-game.

It is worth mentioning that in the pre-AI era, only Gu Li Jiudan scored more than 60 points. From the point of view of coincidence, opening and mid-game, Coulee 9th dan is far ahead. Perhaps from the AI ​​point of view, in ancient Li Shifan chess and other games, the content of Gu Li's chess is slightly better than that of Li Shishi's 9th dan. The smallest loss of head and the super-high one-choice overlap all reflect his bold skills. , Intuitive and keen technical characteristics.

Another counter-intuitive statistic shows that Lee Sedol's 9th Dan is better among the three, and the start is not as good as Lee Chang-ho's 9th Dan. From the perspective of short-term trends, the 9th dan of Lee Chang-ho is the pinnacle of the "previous generation" of humans, while Xiao Li and Gu Li are the pinnacles of mankind in the "pre-AI era". Therefore, there is nothing to say that "Go is developed". error.

However, from a long-term trend, after the emergence of AI, the peak level of humanity may not be as much as imagined. Facing the fact that Fan Xiping’s data goes hand-in-hand with Gu Li’s 9th dan and Shi Xiangxia’s data is similar to Er Li, perhaps we should rethink the conclusion of the development of Go. Perhaps the improvement of Go in hundreds of years is not significant, only a few. Some geniuses are slowly extending the limits of mankind slightly, and they are being pushed back by more erroneous "progressive thinking".

I think back to the Go forums, post it, and even know how demeaning the two sages of ancient chess (Fan Xiping, Shi Xiangxia), I still feel a little bit laughable. I wonder if the two saints, known for their battles, already had the top level of human beings (before and after the emergence of AI) hundreds of years ago, and labeling it as "At most industry 5" can actually make the locals powerful? Did the players of Kie 3 and Kie 4 who criticized the “unintelligible” opening of Guqi also think that the opening of Ersheng will surpass the big and small Lee in the AI? Although high-eyes and low-hands are necessary steps for hand-eye coordination, "seeing the three moves of high chess" is also a standing motivation for progress, but in the face of the unknown, perhaps keeping an open perspective can get closer to the truth.

I don’t know that many people in the chess world who don’t understand ancient chess say that "the ancients could not surpass karma 6, and it has nothing to do with their talents." Without the influence of AI and no public discussion of the competitive atmosphere, Ersheng achieved a quasi-full encirclement of the most peak human data in the pre-AI era only through chess and exploration, and it is the best of human beings after learning AI tricks. When the peak of Ke Jie's 9th Duan is completely powerful on paper, what should he say? Although the data is not a chess skill, it has completely reached the level of the strongest nine-segment data, and how can the ancient sage, who has never met with AI, only have the level of karma 5 and karma 6?

Note: All contemporary chess records are played at the peak of the players, and most of them are quality games in the finals of the competition. The author of the private message can obtain the original file for analysis.

_________________
My "guide" to become stronger in Go

Top
 Profile  
 
Offline
 Post subject: Re: 柯洁能下得过古代的围棋高手吗?
Post #5 Posted: Tue Jan 05, 2021 6:00 am 
Dies in gote
User avatar

Posts: 33
Location: Leuven, Belgium
Liked others: 22
Was liked: 10
Rank: EGF 1kyu
Universal go server handle: Farodin
SoDesuNe wrote:
Google Translate wrote:
The data shows that Lee Chang-ho's 9th-dan lead may not be the official, but the mid-game.


Unfortunate that this sentence got mistranslated, as it's a very interesting one. The correct (rough) translation would be:

The data shows that Lee Chang-ho's 9th-dan biggest strength may not be the end-game, but the mid-game.

Other translation notes (after a brief skim over the text):
"Industry X" -> Professional X Dan
"Li shishi" -> Lee Sedol


This post by Farodin was liked by: SoDesuNe
Top
 Profile  
 
Offline
 Post subject: Re: 柯洁能下得过古代的围棋高手吗?
Post #6 Posted: Tue Jan 05, 2021 6:49 am 
Dies in gote

Posts: 54
Liked others: 60
Was liked: 10
pgwq wrote:
...KatagoS289B40F384与S571B40F256...

我在哪里可以下载这些网络?
where can i download these networks?

Top
 Profile  
 
Offline
 Post subject: Re: 柯洁能下得过古代的围棋高手吗?
Post #7 Posted: Tue Jan 05, 2021 9:22 am 
Lives with ko
User avatar

Posts: 251
Location: Waterloo, Ontario (Canada)
Liked others: 110
Was liked: 155
Rank: AGA 1k
GD Posts: 1190
KGS: apetresc
IGS: apetresc
OGS: apetresc
Universal go server handle: apetresc
Amigo wrote:
我在哪里可以下载这些网络?
where can i download these networks?


All published Katago networks are here, with an explanation of the best ones to use for practical purposes here.

_________________
The road to wisdom? Well, it's plain, and simple to express: Err, and err, and err again; but less, and less, and less!
Image Image Image Image

Top
 Profile  
 
Offline
 Post subject: Re: 柯洁能下得过古代的围棋高手吗?
Post #8 Posted: Tue Jan 05, 2021 10:36 am 
Dies in gote

Posts: 54
Liked others: 60
Was liked: 10
apetresc wrote:
Amigo wrote:
我在哪里可以下载这些网络?
where can i download these networks?


All published Katago networks are here, with an explanation of the best ones to use for practical purposes here.

here https://d3dndmfyhecmj0.cloudfront.net/g ... index.html
only one network b40c384: g170e-b40c384x2-s2348692992-d1229892979
and last b40c256: g170-b40c256x2-s5095420928-d1229425124
here https://katagotraining.org/networks/
last b40c256: kata1-b40c256-s5539504384-d1333756242
so if networks S289B40F384 and S571B40F256 exist, I wonder where to download them

Top
 Profile  
 
Offline
 Post subject: Re: 柯洁能下得过古代的围棋高手吗?
Post #9 Posted: Sun Jan 17, 2021 5:23 pm 
Dies with sente

Posts: 117
Liked others: 21
Was liked: 22
John Fairbairn wrote:
Very well done! Although I'm sure we will eventually need a rigorous analysis, by people more used to dealing with statistics than I am, of your method and your interpretation of the results, I already feel confident in saying that your claim that the ancient players are worthier of much more respect than they usually get is one that several of us on this forum will find easy to accept. The Japanese have done a similar exercise for players going back to Honinbo Dosaku and have reached a similar conclusion, and some of us here have done similar things on a small number of games for individual players. The results seem similar to yours, but your exercise goes into impressive detail.

Indeed, I think that in some ways it is easier to make the claim for the old Chinese players, for two reasons. One is that you appear not to have made an adjustment yet for group tax. If that's correct, I think you will find that they score much better (especially, but not just, in the opening) than your figures show. The other reason is that in the case of old Japanese players we have no commentaries or theory books written by the players themselves, and so it is difficult to know how the players thought. In the case of China we have both hundreds of commentaries and theory books. We not only can see how they thought but we can also see that they thought in ways which seem to find a remarkable parallel in AI play. To give one example that is fascinating me at the moment and which I have studied in some detail, there is the term 照应 which is especially common in 不古编 (published in 1682!!). For the benefit of readers not familiar with these commentaries, this refers to a concept of proffering a helping hand to friendly groups, but since the helping hand extends over the whole board and in multiple directions, it works in a very AI-like way. Perhaps a better way of looking at it is like setting up a lighthouse (照 = to shine) offering help to ships in distress, or just plain reassurance, on the vast ocean of the board.

I hope it also needless to add that the enormous technical skill of the old Chinese players has been recognised even outside China. Apart from the tsumego problems of Xuanxuan Qijing, for example, their skill in actual games has been recognised. One good example is 中国古棋譜散歩 (A promenade through old Chinese go manuals) by Watanabe Hideo 7-dan.

What many people here would like to see now is your analysis of Go Seigen (Wu Qingyuan). He is not so ancient, but he did study old Chinese manuals and many people, both professional and amateur, have noticed that he too played in an AI way.


Thank you for your support. We are very impressed by your research on Chinese ancient books.
I'll share your views in the Chinese seated stones WeiQi QQ group.

_________________
Zhang-hu 章浒
Committed to the restoration Chinese traditional Weiqi
Research on ancient Weiqi rules & Classic (Dunhuang Classic and the Thirteen Chapters Classic)
From Suzhou, Jiangsu Province, China

Top
 Profile  
 
Display posts from previous:  Sort by  
Post new topic Reply to topic  [ 9 posts ] 

All times are UTC - 8 hours [ DST ]


Who is online

Users browsing this forum: No registered users and 1 guest


You cannot post new topics in this forum
You cannot reply to topics in this forum
You cannot edit your posts in this forum
You cannot delete your posts in this forum
You cannot post attachments in this forum

Search for:
Jump to:  
Powered by phpBB © 2000, 2002, 2005, 2007 phpBB Group