b***y
发帖数: 2804 | | i****e
发帖数: 642 | 2 I would say never, maybe.
Chess and Go have certainty: although it is hard to get the best play, but
there is one.
Bridge is probability game. The best bid/play is just to gave the best
chance in long run, but a match has only certain number of hands or
situations.
Also, there are partnership issue, psych-bid and false play in bridge, and
there is no such thing in chess and Go.
【在 b***y 的大作中提到】
: 电脑围棋太强大了。现在就剩桥牌还未攻占~~
: http://tech.sina.com.cn/it/2016-03-09/doc-ifxqafha0527545.shtml
: http://sports.sina.com.cn/go/2016-03-09/doc-ifxqafha0534995.sht
| b***y
发帖数: 2804 | 3 But if computer makes the best probabilistic bid/play, doesn't it become a
winner in the long run? Plus, in a long match, people get tired, but
computers don't.
I think the central issue is understanding of the game, including
psychological aspects. Current game model for bridge needs improvement. Once
it is modeled properly, coupled with learning capability (through something
like neural networks), computers can make the optimal bid/play/defense in a
long run. That doesn't mean they will win every time, but if they win 7
times out of 10, then it is fair to say they are at least as good as (
probably better than) the opponents.
【在 i****e 的大作中提到】
: I would say never, maybe.
: Chess and Go have certainty: although it is hard to get the best play, but
: there is one.
: Bridge is probability game. The best bid/play is just to gave the best
: chance in long run, but a match has only certain number of hands or
: situations.
: Also, there are partnership issue, psych-bid and false play in bridge, and
: there is no such thing in chess and Go.
| b***y
发帖数: 2804 | 4 2:0了。
最可怕的是，AlphaGo的有些职业牌手眼中的“疑问手”或“臭棋”，可能其实是大量
运算后的正招，显示机器对围棋的理解可能已经超过人类。
也许类似的事情会发生在桥牌？一些我们教科书中的“本手”，也许并非最强手。
下面就看柯洁的了。 | i****e
发帖数: 642 | 5 That is true. Besides computer won't get tired, it can calculate the
probability with almost no time, while we can just estimate most of time.
It is not easy to judge whether computer is better than human in bridge, due
to many factors. The major one is the limited number of hands/situations.
Once
something
a
【在 b***y 的大作中提到】
: But if computer makes the best probabilistic bid/play, doesn't it become a
: winner in the long run? Plus, in a long match, people get tired, but
: computers don't.
: I think the central issue is understanding of the game, including
: psychological aspects. Current game model for bridge needs improvement. Once
: it is modeled properly, coupled with learning capability (through something
: like neural networks), computers can make the optimal bid/play/defense in a
: long run. That doesn't mean they will win every time, but if they win 7
: times out of 10, then it is fair to say they are at least as good as (
: probably better than) the opponents.
| b***y
发帖数: 2804 | 6 Computer does have one disadvantage: it doesn't check the manner of play,
hesitation, etc. :-))
【在 i****e 的大作中提到】
: That is true. Besides computer won't get tired, it can calculate the
: probability with almost no time, while we can just estimate most of time.
: It is not easy to judge whether computer is better than human in bridge, due
: to many factors. The major one is the limited number of hands/situations.
:
: Once
: something
: a
| i****e
发帖数: 642 | 7 That is true, but if we use network based system for human to play, we can
also reduce the chance of those unauthorized information.
The other part which could be hard for computer is to justify a play by
human: is a play by a human player a good play from the player's point of
view, or is he making a mistake by choosing an inferior play?
For example, if the third hand defender plays queen, you can assume he has
no jack if he is a reasonable player (unless he plays false card). But if he
is a novice, you will have no idea at all. This will be even harder for a
computer to figure it out.
【在 b***y 的大作中提到】
: Computer does have one disadvantage: it doesn't check the manner of play,
: hesitation, etc. :-))
| o*********1
发帖数: 2608 | 8 桥牌这种游戏 不会是 AI的菜
和棋完全不是一种游戏 | p***r
发帖数: 20570 | 9 AI没戏。规则对人类本来就是不公平的，这种准快棋，对计算机来说是天大的便宜。而
且计算机对于人类棋谱有充分学习和了解，但是人类棋手对计算机的基本算法的了解是
极端有限的，这种程序内部的不自恰和漏洞多了去了。另外，韩国棋手能崛起本质上也
是利用了现代这种准快棋规则的漏洞，而非对围棋本质更加深刻的理解。围棋商业化也
是对围棋的巨大伤害。
桥牌和围棋是完全不同的游戏，因为前者是incomplete information. 桥牌这种游戏至
少在叫牌阶段是没有必胜策略的，任何的体系都存在相对难于处理的大量牌例。这是叫
牌数量的有限性决定的。这就意味着桥牌里面存在着巨大的空间来调整策略，还存在巨
大的可能性利用对手对你体系的不熟悉以至于设立错误的条件而获利。
另外，这里面涉及到迅速寻找对手叫牌和打牌体系中的漏洞，利用这些漏洞进行调整策
略，甚至进行故意的欺骗以获利。人类提供给计算机对手的信息可能根本就不可靠或者
只是某种程度上的可靠信息。这种不可靠的信息从计算机角度来说是极端难以判断的。
这会极大影响设定sampling 的条件。只有计算机能够学会迅速的调整，甚至欺骗并且
能够很大概率的识别这些调整和欺骗之后，才能称得上接近具备某种真正意义上的智能
。利用importance sampling解双明手问题再快也没用。这跟真正意义上的智能无关。
这也是计算机在deep stack no limit Texas holdem上赢不了人类高手的本质原因。这
种战略欺骗能力甚至在machine learning方面都会有问题，因为极端困难的去理解为什
么人类高手会在某种特定形势下做出某种决定。高手可能一直在输小钱造成故意误导而
赢少数的大钱。而且holdem的牌多数情况是不公开的，除了电视转播以外。
【在 b***y 的大作中提到】
: 电脑围棋太强大了。现在就剩桥牌还未攻占~~
: http://tech.sina.com.cn/it/2016-03-09/doc-ifxqafha0527545.shtml
: http://sports.sina.com.cn/go/2016-03-09/doc-ifxqafha0534995.sht
| b***y
发帖数: 2804 | | | | i****e
发帖数: 642 | 11 I would say never, maybe.
Chess and Go have certainty: although it is hard to get the best play, but
there is one.
Bridge is probability game. The best bid/play is just to gave the best
chance in long run, but a match has only certain number of hands or
situations.
Also, there are partnership issue, psych-bid and false play in bridge, and
there is no such thing in chess and Go.
【在 b***y 的大作中提到】
: 电脑围棋太强大了。现在就剩桥牌还未攻占~~
: http://tech.sina.com.cn/it/2016-03-09/doc-ifxqafha0527545.shtml
: http://sports.sina.com.cn/go/2016-03-09/doc-ifxqafha0534995.sht
| b***y
发帖数: 2804 | 12 But if computer makes the best probabilistic bid/play, doesn't it become a
winner in the long run? Plus, in a long match, people get tired, but
computers don't.
I think the central issue is understanding of the game, including
psychological aspects. Current game model for bridge needs improvement. Once
it is modeled properly, coupled with learning capability (through something
like neural networks), computers can make the optimal bid/play/defense in a
long run. That doesn't mean they will win every time, but if they win 7
times out of 10, then it is fair to say they are at least as good as (
probably better than) the opponents.
【在 i****e 的大作中提到】
: I would say never, maybe.
: Chess and Go have certainty: although it is hard to get the best play, but
: there is one.
: Bridge is probability game. The best bid/play is just to gave the best
: chance in long run, but a match has only certain number of hands or
: situations.
: Also, there are partnership issue, psych-bid and false play in bridge, and
: there is no such thing in chess and Go.
| b***y
发帖数: 2804 | 13 2:0了。
最可怕的是，AlphaGo的有些职业牌手眼中的“疑问手”或“臭棋”，可能其实是大量
运算后的正招，显示机器对围棋的理解可能已经超过人类。
也许类似的事情会发生在桥牌？一些我们教科书中的“本手”，也许并非最强手。
下面就看柯洁的了。 | i****e
发帖数: 642 | 14 That is true. Besides computer won't get tired, it can calculate the
probability with almost no time, while we can just estimate most of time.
It is not easy to judge whether computer is better than human in bridge, due
to many factors. The major one is the limited number of hands/situations.
Once
something
a
【在 b***y 的大作中提到】
: But if computer makes the best probabilistic bid/play, doesn't it become a
: winner in the long run? Plus, in a long match, people get tired, but
: computers don't.
: I think the central issue is understanding of the game, including
: psychological aspects. Current game model for bridge needs improvement. Once
: it is modeled properly, coupled with learning capability (through something
: like neural networks), computers can make the optimal bid/play/defense in a
: long run. That doesn't mean they will win every time, but if they win 7
: times out of 10, then it is fair to say they are at least as good as (
: probably better than) the opponents.
| b***y
发帖数: 2804 | 15 Computer does have one disadvantage: it doesn't check the manner of play,
hesitation, etc. :-))
【在 i****e 的大作中提到】
: That is true. Besides computer won't get tired, it can calculate the
: probability with almost no time, while we can just estimate most of time.
: It is not easy to judge whether computer is better than human in bridge, due
: to many factors. The major one is the limited number of hands/situations.
:
: Once
: something
: a
| i****e
发帖数: 642 | 16 That is true, but if we use network based system for human to play, we can
also reduce the chance of those unauthorized information.
The other part which could be hard for computer is to justify a play by
human: is a play by a human player a good play from the player's point of
view, or is he making a mistake by choosing an inferior play?
For example, if the third hand defender plays queen, you can assume he has
no jack if he is a reasonable player (unless he plays false card). But if he
is a novice, you will have no idea at all. This will be even harder for a
computer to figure it out.
【在 b***y 的大作中提到】
: Computer does have one disadvantage: it doesn't check the manner of play,
: hesitation, etc. :-))
| o*********1
发帖数: 2608 | 17 桥牌这种游戏 不会是 AI的菜
和棋完全不是一种游戏 | p***r
发帖数: 20570 | 18 AI没戏。规则对人类本来就是不公平的，这种准快棋，对计算机来说是天大的便宜。而
且计算机对于人类棋谱有充分学习和了解，但是人类棋手对计算机的基本算法的了解是
极端有限的，这种程序内部的不自恰和漏洞多了去了。另外，韩国棋手能崛起本质上也
是利用了现代这种准快棋规则的漏洞，而非对围棋本质更加深刻的理解。围棋商业化也
是对围棋的巨大伤害。
桥牌和围棋是完全不同的游戏，因为前者是incomplete information. 桥牌这种游戏至
少在叫牌阶段是没有必胜策略的，任何的体系都存在相对难于处理的大量牌例。这是叫
牌数量的有限性决定的。这就意味着桥牌里面存在着巨大的空间来调整策略，还存在巨
大的可能性利用对手对你体系的不熟悉以至于设立错误的条件而获利。
另外，这里面涉及到迅速寻找对手叫牌和打牌体系中的漏洞，利用这些漏洞进行调整策
略，甚至进行故意的欺骗以获利。人类提供给计算机对手的信息可能根本就不可靠或者
只是某种程度上的可靠信息。这种不可靠的信息从计算机角度来说是极端难以判断的。
这会极大影响设定sampling 的条件。只有计算机能够学会迅速的调整，甚至欺骗并且
能够很大概率的识别这些调整和欺骗之后，才能称得上接近具备某种真正意义上的智能
。利用importance sampling解双明手问题再快也没用。这跟真正意义上的智能无关。
这也是计算机在deep stack no limit Texas holdem上赢不了人类高手的本质原因。这
种战略欺骗能力甚至在machine learning方面都会有问题，因为极端困难的去理解为什
么人类高手会在某种特定形势下做出某种决定。高手可能一直在输小钱造成故意误导而
赢少数的大钱。而且holdem的牌多数情况是不公开的，除了电视转播以外。
【在 b***y 的大作中提到】
: 电脑围棋太强大了。现在就剩桥牌还未攻占~~
: http://tech.sina.com.cn/it/2016-03-09/doc-ifxqafha0527545.shtml
: http://sports.sina.com.cn/go/2016-03-09/doc-ifxqafha0534995.sht
| o*********1
发帖数: 2608 | 19 AI 首先要解决的就是 perfect information
国际象棋，象棋算是解决了： 不过里面几乎 没有intelligence的含义， 因为原理就
是裸算。 只要资源和复杂度相当， 电脑必胜。
围棋因为复杂度高， 裸算搞不定。 如果有一个AI能够必胜， 这里面一定会包含和核
心intelligence类似的东西
所以AI能在围棋成功还是很了不起的
incomplete information 现在基本上就不到谈AI的时候， 像桥牌， poker这种
【在 p***r 的大作中提到】
: AI没戏。规则对人类本来就是不公平的，这种准快棋，对计算机来说是天大的便宜。而
: 且计算机对于人类棋谱有充分学习和了解，但是人类棋手对计算机的基本算法的了解是
: 极端有限的，这种程序内部的不自恰和漏洞多了去了。另外，韩国棋手能崛起本质上也
: 是利用了现代这种准快棋规则的漏洞，而非对围棋本质更加深刻的理解。围棋商业化也
: 是对围棋的巨大伤害。
: 桥牌和围棋是完全不同的游戏，因为前者是incomplete information. 桥牌这种游戏至
: 少在叫牌阶段是没有必胜策略的，任何的体系都存在相对难于处理的大量牌例。这是叫
: 牌数量的有限性决定的。这就意味着桥牌里面存在着巨大的空间来调整策略，还存在巨
: 大的可能性利用对手对你体系的不熟悉以至于设立错误的条件而获利。
: 另外，这里面涉及到迅速寻找对手叫牌和打牌体系中的漏洞，利用这些漏洞进行调整策
| a****s
发帖数: 524 | 20 如果要问机器能否在桥牌上战胜人类，我觉得答案很简单，一定会。现在缺得仅仅是钱
。如果有人愿意出钱，很快。
桥牌是比围棋简单得多的游戏，起码对人类来说是如此。
对机器而言，桥牌的计算深度只有区区十几步，并且很多分支都是不合法的。机器可以
在几分钟内遍历上百万种变化。
对于围棋，暴力计算是不行的。AlphaGo的办法是学习专家的棋谱，从而在每个分支，
只计算一些推荐的选点和一些半随机的选点。这样大大减少了分支的数目。比如说，在
某个局面下，张九段说走A，李九段说走B, 王九段说走C，AlphaGo说走D，然后分别对A
，B, C, D, 进行深度计算后作出一个价值判断 (围棋的价值判断很难，因为永远没法
推算到终局， 而桥牌的价值判断，简直不值一提）。
和围棋不同，桥牌专家对于相同局面下的攻防，是很少有分歧的。大多数情况下，技术
上正确的打法只有一种。需要分析的情况对机器来说，实在是少得可怜，并且机器永远
不会因为疲劳或情绪波动出错。
剩下的无非就是不完全信息的问题。如果纯算概率，不用说，人肯定不行。你说人会“
搞”（bluffing）， 难道机器不会？ 不懂bluffing，机器能玩儿职业poker？ 实际上
，顶尖的poker player，在网上赌博时，绝对不是仅仅靠直觉的。他们都会用软件来分
析局面，并且对对手的风格和倾向做详细的记录。也就是说，在做机器做的事情。
机器可以纯随机地bluffing，人的行为却很难是纯随机的。机器能对对手的行动做详尽
的记录和实时的分析， Bluffing？ 玩儿死你，你都不知道是怎么死的。
至于这样的机器，是不是真正实现了intelligence， 不知道。
人类自己目前连什么是consciousness（就是动物的脑袋里是怎么回事）都不知道，何
况intelligence？
【在 o*********1 的大作中提到】
: AI 首先要解决的就是 perfect information
: 国际象棋，象棋算是解决了： 不过里面几乎 没有intelligence的含义， 因为原理就
: 是裸算。 只要资源和复杂度相当， 电脑必胜。
: 围棋因为复杂度高， 裸算搞不定。 如果有一个AI能够必胜， 这里面一定会包含和核
: 心intelligence类似的东西
: 所以AI能在围棋成功还是很了不起的
: incomplete information 现在基本上就不到谈AI的时候， 像桥牌， poker这种
| l*****8
发帖数: 16949 | 21 完全同意。我觉得桥牌对机器是很简单的游戏。暴力破解就足够了。
对A
【在 a****s 的大作中提到】
: 如果要问机器能否在桥牌上战胜人类，我觉得答案很简单，一定会。现在缺得仅仅是钱
: 。如果有人愿意出钱，很快。
: 桥牌是比围棋简单得多的游戏，起码对人类来说是如此。
: 对机器而言，桥牌的计算深度只有区区十几步，并且很多分支都是不合法的。机器可以
: 在几分钟内遍历上百万种变化。
: 对于围棋，暴力计算是不行的。AlphaGo的办法是学习专家的棋谱，从而在每个分支，
: 只计算一些推荐的选点和一些半随机的选点。这样大大减少了分支的数目。比如说，在
: 某个局面下，张九段说走A，李九段说走B, 王九段说走C，AlphaGo说走D，然后分别对A
: ，B, C, D, 进行深度计算后作出一个价值判断 (围棋的价值判断很难，因为永远没法
: 推算到终局， 而桥牌的价值判断，简直不值一提）。
| b***y
发帖数: 2804 | 22 暴力破解肯定不行，因为信息不完整，甚至不可靠。举个例子，我方的叫牌是基于对方
叫牌的基础上，你既不能完全相信对方的叫牌，又不能完全不信。做庄也是一样，你不
能全信对方的防守信号，也不能全不信。机器防守也有类似的考虑，机器之间也要打信
号，但也可以故意打假张。这些都不是暴力破解可以提供的，需要一个自我学习的反馈
架构，非常不简单。
【在 l*****8 的大作中提到】
: 完全同意。我觉得桥牌对机器是很简单的游戏。暴力破解就足够了。
:
: 对A
|
|
