game theroy 博弈论 02 Extensive form games形式广泛的博弈

02 Extensive form games形式广泛的博弈

1. Summary of the chapter 章节摘要

1. Extensive form games  广义形式博弈
2. Perfect information 完美信息
3. The tree of the game 对策树
4. Backward induction 反向归纳
5. Von Neumann theorem 冯·诺依曼定理
6. Different (types of) solutions 不同（类型）的解
7. Combinatorial games  组合对策
8. The Nim game and Bouton theorem 尼姆对策和布顿定理
9. Strategies 策略

2. Extensive form扩展形式

Three politicians are supposed to decide whether to raise their salaries or not. The vote is public and in sequence. They would prefer to receive a salary increase, yet they would also like to vote against it so as not to lose public support

三位政治家应该决定是否加薪。投票是公开的，按顺序进行。他们希望加薪，但也希望投票反对，以免失去公众的支持

Optimal result for each player: having salary increase while voting against it!

每个玩家的最佳结果：在投票反对的同时加薪！

Main features of the game博弈的主要特征：

1. The moves take place in sequence: the politicians vote one after the other行动按顺序进行：政客们一个接一个地投票
2. Every possible situation is known to the players: at any time they know the whole past history, as well as the possible developments每一种可能的情况对玩家来说都是已知的：在任何时候，他们都知道整个过去的历史，以及可能的发展
3. The final outcome is determined by the majority of votes最终结果由多数票决定

This is an example of what is called a game with perfect information这是一个所谓的完美信息游戏的例子：

each player has knowledge about all the events that have previously occurred.

每个玩家都知道之前发生的所有事件。

How can we represent such a game? And how can we solve it?

我们怎样才能代表这样一场比赛？我们该如何解决？

3. The tree 树

3.1 example

Each player’s vote is represented at a branch: YES on the left and NO on the right

每个玩家的投票都在一个分支上代表：左边是YES，右边是NO

Utilities are based on individual votes and possible final outcomes:

1. 1= YES and no raise; 1=是，没有加薪；
2. 2= NO and no raise; 2=NO且无加薪；
3. 3= YES and raise; 3=是并加薪；
4. 4= N0 and raise.4=N0并加薪。
5. 效用基于个人投票和可能的最终结果：

3.2 A game with chance 一场有机会的博弈

Two players 1 and 2 must decide in sequence whether to play or not.

两名玩家1和2必须按顺序决定是否上场。

If both of them decide to play, then a coin is tossed (random component R): the first player wins with heads, whereas the second one with tails.

如果他们两个都决定玩，那么就会扔一枚硬币（随机分量R）：第一个玩家用头获胜，而第二个玩家用尾获胜。

3.3.Directed graphs有向图

a) Definition:

A finite directed graph is a pair (V, E) where 有限有向图是一对（V，E），其中

1. V is a finite set, called the set of vertices ; V是一个有限集，称为顶点集;
2. E ⊂ V × V is a set of ordered pairs of vertices called the set of the (directed) edges. E⊂V×V是一组有序的顶点对，称为（有向）边集

b) Definition:

A path from a vertex v1 to a vertex vk+1 is a finite sequence of vertices-edges v1, e1, v2, …, vk, ek , vk+1 such that ei ≠  ej if i ≠ j and ej= (vj,vj+1).

定义从顶点v1到顶点vk+1的路径是顶点边v1, e1, v2, …, vk, ek , vk+1的有限序列使得ei ≠  ej，如果i ≠ j并且ej= (vj,vj+1)。

k is called the length of the path.

k称为路径的长度。

3.4 tree

a) Definition

An oriented graph is finite directed graph having no bidirected edges, that is for all j, k at most one between (vj,vk) and (vk,vj) may be arrows of the graph.

定义有向图是没有双向边的有限有向图，也就是说，对于所有j，k，(vj,vk)和(vk,vj)之间最多一个可以是图的箭头

b) Definition

A tree is a triple (V, E, x0 )where (V, E) is an oriented graph and x0 is a vertex in V such that there is a unique path from x0to x, where x is any vertex in V

定义树是一个三元组（V，E，x0），其中（V，E）是一个有向图，x0是V中的一个顶点，因此从x0到x有一条唯一的路径，其中x是V中任何顶点

c) Definition

A child of a vertex v is any vertex x such that (v, x) ∈ E.

A vertex is called a leaf if it has no children.

We say that the vertex x follows the vertex v if there is a path from v to x

定义顶点v的子顶点是任何顶点x，使得（v，x）∈E。

如果顶点没有子顶点，则称为叶。

如果存在从v到x的路径，我们说顶点x跟随顶点v

3.5 The tree of a game博弈树

(A complete representation of a game with a tree requires additional conditions)

具有树的博弈的完整表示需要附加条件

Definition:

An Extensive form game with perfect information consists of

一个具有完美信息的广泛形式的博弈包括

1. A finite set N = {1, …, n} of players; 游戏者的有限集N={1，…，N}
2. A game tree (V, E, x0); 游戏树（V，E，x0）
3. A partition of the vertices that are not leaves into sets {P1, P2, . . . , Pn+1}; 非叶顶点划分为集{P1，P2，…，Pn+1}
4. A probability distribution for each vertex in Pn+1, defined on the edges from the vertex to its children; Pn+1中每个顶点的概率分布，定义在从顶点到其子顶点的边上
5. A n-dimensional vector attached to each leaf (i.e. list of possible outcomes). 附加到每个叶的N维向量（即可能结果的列表）

Notes:

The number n denotes the cardinality of the set N.数字n表示集合n的基数。

For a partition into sets Pi one has , and .对于划分为集合Pi的分区，其中，对于

3.6 comments

• The set Pi, for i ≤ n, is the set of the nodes v where Player i must choose a child of v, representing a possible move from him at v；对于i≤n，集合Pi是节点v的集合，其中玩家i必须选择v的子节点，表示在v处从他可能的移动
• Pn+1 is the set of the nodes where a chance move is present: that is n + 1 is the number of players plus the random component. Pn+1 can be empty, meaning that the game does not admit any chance. Pn+1是存在机会移动的节点集：也就是说，n+1是玩家数量加上随机分量。Pn+1可以是空的，这意味着游戏不允许任何机会。
• When Pn+1 is empty, the n players have only preferences on the leaves: a utility function is not required. 当Pn+1为空时，n个玩家在树叶上只有偏好：不需要效用函数

4. Solving the game解决游戏

In order to find the optimal outcome, we employ the rationality axioms

为了找到最优结果，我们使用了合理性公理

1. What one player does when being in positions leading to leaves can be determined by decision theory (rationality assumption 5); 当一名玩家处于导致离开的位置时，他会做什么可以由决策理论决定（合理性假设5）
2. This information is known to all other players (r.a. 4) that can use it; 所有其他可以使用该信息的玩家（r.a.4）都知道该信息
3. Thus players moving to vertices going to leaves use decision theory (r.a. 5); 因此，玩家移动到要离开的顶点使用决策理论（r.a.5）
4. This information is known to all other players (r.a. 4) that can be use it; 所有其他可以使用该信息的玩家（r.a.4）都知道该信息
5. …
6. The player starting the game uses decision theory to make the first move. 开始游戏的玩家使用决策理论做出第一步

Example:

Moving bottom-up, at each vertex a player chooses the edge with highest utility:

自下而上移动，玩家在每个顶点选择具有最高效用的边：

1. Player3: h = (3, 3, 4) over g = (3, 3, 3) ; j = (3, 4, 3) over k = (1, 2, 2); l = (4, 3, 3) over m= (2, 1, 2) ; o = (2, 2, 2) over n= (2, 2, 1)
2. Player2: d = (3, 4, 3) over c = (3, 3, 4) ; e = (4, 3, 3) over f = (2, 2, 2)
3. Player1: b = (4, 3, 3) over a = (3, 4, 3), thereby finding the optimal outcome

5. Backward induction逆向归纳法

5.1 Definition:

Define Length of the game as the length of the longest path in the game

定义将游戏的长度定义为游戏中最长路径的长度

5.2 Algorithm 算法

1. Decision theory, i.e. rationality assumption 5, enables us to solve games of length 1 ; 决策理论，即合理性假设5，使我们能够求解长度为1的博弈。
2. Rationality assumption 4 allows us to solve a game of length i + 1 if the games of length at most i are solved. 合理性假设4使我们能够在长度至多为i的博弈被求解的情况下求解长度为i+1的博弈

Thus, by repeated applications, we can solve games of any finite length

因此，通过反复应用，我们可以求解任何有限长度的对策

This method takes the name of backward induction: it is the process of reasoning backwards in time (that is from the leaves of the tree up to the root), so as to determine a sequence of actions leading one to the optimal outcome.

这种方法被称为反向归纳法：它是一个在时间上向后推理的过程（即从树叶到树根），以确定一系列行动，从而得出最佳结果。

6. The first rationality theorem第一合理性定理

Theorem:

The rational outcomes of a finite, perfect information game are those given by the procedure of backward induction

有限完全信息博弈的合理结果是通过反向归纳过程给出的结果

The method of backwards induction can be applied since every vertex v of the game is the root of a new game, made by all followers of v in the initial game. 可以应用反向归纳的方法，因为博弈的每个顶点v都是新博弈的根，由初始博弈中v的所有追随者创建。

Such a game is called a subgame of the original one.

这样的游戏被称为原始博弈的子博弈。

7. Multiple solutions多解

The outcomes obtained by backward induction are: (4, 3) and (3, 4). In fact, Pl2 does not have any preference between (4, 3) and (0, 3)!

通过反向归纳得到的结果是：（4，3）和（3，4）。事实上，Pl2在（4，3）和（0，3）之间没有任何偏好！

Therefore, in general, uniqueness of solutions is not guaranteed!

因此，一般情况下，无法保证解决方案的唯一性！

8. The chess theorem (von Neumann)国际象棋定理（冯·诺依曼）

Theorem

In the game of chess one and only one of the following alternatives holds在国际象棋比赛中，只有以下一种选择成立：

1. The white has a way to win, no matter what the black does; 无论黑人做什么，白人都有办法获胜
2. The black has a way to win, no matter what the white does; 无论白人做什么，黑人都有办法胜利
3. The white has a way to force at least a draw, no matter what the black does, and the same holds for the black; 无论黑人做了什么，白人至少有办法打平，黑人也一样

Main question: Is the above a true theorem?

主要问题：以上是一个真定理吗？

Why is it impossible to say more than that? Why can’t we determine exactly which one is the correct alternative?

为什么不可能说更多？为什么我们不能确切地确定哪一个是正确的替代方案？

8.1 An interesting proof of the chess theorem 国际象棋定理的有趣证明

Here is a proof of the theorem这是定理的一个证明：

Suppose the length of the game is 2K so each player has K choices to make.

假设游戏的长度是2K，那么每个玩家都有K个选择。

Call ai the move of the White at her i-th stage and bi the one of the Black.

在她的第i个阶段，称ai为白人的动作，称bi为黑人的动作

The first alternative in the chess theorem can be expressed as

国际象棋定理中的第一个备选方案可以表示为

Now suppose this is not true. Then现在假设这不是真的。然后:

But this means exactly that Black has the possibility to get at least a draw

但这恰恰意味着黑人有可能至少获得一场平局

8.2 Summarizing

If White does not have a strategy to win no matter what Black does, then Black has the possibility to get at least the draw.

如果无论布莱克做什么，怀特都没有获胜的策略，那么布莱克至少有可能获得平局。

Symmetrically, if Black does not have a strategy to win no matter what White does, then White has the possibility to get at least the draw

对称地说，如果无论怀特做什么，布莱克都没有获胜的策略，那么怀特至少有可能获得平局

Thus if the first and the second alternatives in the chess theorem are not true, necessarily the third one is true!

因此，如果国际象棋定理中的第一个和第二个备选方案不成立，那么第三个必然成立！

8.3 Extending von Neumann theorem 冯-诺依曼定理的推广

The von Neumann theorem applies to every finite game of perfect information where the possible result is either the victory of one player or a tie. Thus the following corollary holds:

冯·诺依曼定理适用于每一个完美信息的有限博弈，其中可能的结果要么是一方获胜，要么是平局。因此，以下推论成立：

Corollary:

Consider a finite perfect information game with two players, where the only possible outcomes are the victory of one or the other player. Then one and only one of the following alternative holds:

考虑一个有两个玩家的有限完美信息游戏，其中唯一可能的结果是一个或另一个玩家的胜利。则以下备选方案中的一个（也是唯一一个）成立：

1. The first player can win, no matter what the second one does; 无论第二个选手做什么，第一个选手都能赢
2. The second player can win, no matter what the first one does. 无论第一个选手做什么，第二个选手都可以获胜

9. Different types of solutions不同类型的解决方案

1. Very weak solution: The game has a rational outcome, but it is inaccessible, like in chess.非常弱的解决方案：游戏有一个合理的结果，但它是不可访问的，就像国际象棋一样。
2. Weak solution: The outcome of the game is known, but how to get to it is not (in general).弱解决方案：游戏的结果是已知的，但如何实现它（一般来说）还不清楚。
3. Solution: It is possible to provide an algorithm to find a solution. 解决方案：可以提供一种算法来找到解决方案

10. Chomp（游戏）

The player remaining with the last tile at the bottom-left loses the game:

左下角最后一个牌剩余的玩家输掉游戏：

1. when the table is a square with n × n tiles there is a solution当表是一个有n×n个瓦片的正方形时，有一个解
2. when the table is a rectangle with n × m tiles there is a weak solution.当表是具有n×m个瓦片的矩形时，存在弱解。

10.1 Chomp: rectangle矩形

Let us assume, for the sake of reductio ad absurdum, that Player II wins. So, after Player I chooses, say, the orange edge, Player II makes a winning action. Suppose this is the choice of the green edge: then any player starting at the node following the green edge will lose. But the tree starting at that node is exactly the same tree starting at the green edge relative to Player I.

为了简化和荒谬起见，让我们假设玩家II获胜。所以，在玩家I选择了橙色边缘后，玩家II做出了获胜的动作。假设这是对绿色边缘的选择：那么任何从绿色边缘后面的节点开始的玩家都会输。但从该节点开始的树与从玩家I的绿色边缘开始的树完全相同。

Thus Player I has a move available that guarantees victory, whereby Player II would have to lose.

因此，球员I有一个可以保证胜利的移动，而球员II将不得不输

Since we derived a contradiction, the original assumption must be false: it is Player I that wins

既然我们得出了一个矛盾，那么最初的假设肯定是错误的：是玩家I赢了

10.2 Chomp: square方形

11. Impartial combinatorial games公平组合对策

Definition:

An impartial combinatorial game is a game such that公正的组合游戏是这样的游戏：

1. There are two players moving in alternate order有两个玩家按交替的顺序移动
2. There is a finite number of positions in the game游戏中的位置数量有限
3. The players follow the same rules球员们遵循同样的规则
4. The game ends when no further moves are possible当无法进一步移动时，游戏结束
5. The game does not involve chance这场比赛不涉及机会
6. In the classical version, the winner is the player leaving the other player with no available moves, in the mis`ere version the opposite在经典版本中，获胜者是没有可用动作的玩家，而在错误版本中则相反

11.1 Examples of combinatorial games组合对策示例

1. k piles of cards. At her turn each player takes as many cards as she wants (at least one!) from one and only one pile; k堆卡片。轮到她时，每个玩家从一堆牌中拿走她想要的牌（至少一张！）
2. k piles of cards. At her turn each player takes as many cards as she wants (at least one!) from no more than j < k piles; k堆卡片。轮到她时，每个玩家从不超过j<k堆的牌中拿走她想要的牌（至少一张！）
3. k cards in a row. At her turn each player takes either j1 or … or jl cards. 一排k张牌。轮到她时，每个选手要么拿j1，要么…或jl卡

In all these variants of the game, the player remaining without cards loses

在所有这些游戏变体中，没有牌的玩家会输

In the first two cases the positions are (n1, . . . , nk ) where ni is a non negative integer for all i. In the last example positions can be seen as all non negative integers smaller or equal to k

在前两种情况下，位置是（n1，…，nk），其中ni是所有i的非负整数。在最后一个例子中，位置可以看作是小于或等于k的所有非负整数

11.2 How to solve these games如何解决这些游戏

Partition the set of all possible (finitely many) positions into two sets:

将所有可能（有限多个）位置的集合划分为两个集合：

1. P-positions (i.e. ’previous’ player has a winning strategy): loosing; P位置（即“先前”玩家有获胜策略）：失去
2. N-positions (i.e. ’next’ player has a winning strategy): winning. N位置（即“下一个”玩家有获胜策略）：获胜

Note: it is the state of the game that matters, and not who is called to move.

注意：重要的是比赛的状态，而不是谁被要求移动。

Rules for the partition (for the classical version):分区规则（对于经典版本）：

1. Terminal position (0, 0, ..., 0) is a P-position (it is a losing position, as the player does not have any card left); 终端位置（0，0，…，0）是一个P位置（这是一个失败的位置，因为玩家没有任何牌了）
2. From a P-position only N-positions are available; 从P位置只有N位置可用
3. From a N-position it is possible, yet not necessary, to go to a P-position. 从N位置到P位置是可能的，但不是必须的

Therefore, the player starting from a N-position wins!

因此，从N位置开始的玩家获胜！

12. The Nim game (first variant) 尼姆游戏（第一个变体）

The Nim game is defined as (n1，…，nk) where n1 is a positive integer for all i.

尼姆对策定义为（n1，…，nk），其中n1是所有i的正整数。

At her turn any player is supposed to take one (and only one) n1 and substitute it with. The winner is the player who arrives at the position (0, …, 0)

在轮到她的时候，任何球员都应该拿一个（而且只有一个）n1，并用代替。获胜者是到达位置（0，…，0）的玩家

Actions: taking away cards from one pile. 动作：从一堆牌中拿走牌

Goal: to clear the whole table. 目标：清理整张桌子。

12.1 A new operation on the non-negative integers 对非负整数的一个新运算

Define an operation ⊕ on N = {0, 1, 2, …, n, …} in the following way:

用以下方式定义N={0，1，2，…，N，…}上的运算⊕：

For n1, n2 ∈ N 对于n1, n2∈N

1. Write n1, n2 in binary form, denoted by [n1]2, [n2]2; 以二进制形式写入n1, n2，表示为[n1]2, [n2]2
2. Write the sum [n1]2⊕ [n2]2 in binary form where ⊕ is the (usual) sum, but without carry ; 以二进制形式写出和[n1]2⊕ [n2]2，其中⊕是（通常的）和，但不带进位
3. What you get is the result in binary form. 你得到的是二进制形式的结果

12.2 An example一个例子

The ⊕ operation applied to 1,2,4, and 1.

对1,2,4和1进行⊕的运算。

12.3 The group小组

a) Definition

A nonempty set A with a binary operation · defined on it is called a group provided that:定义了二进制运算的非空集A称为群，条件是：

1. for a, b ∈ A the element a · b ∈ A; 对于a, b∈A，元素a·b∈A
2. · is associative: (a · b) · c = a · (b · c); ·是关联的：（a·b）·c=a·（b·c）
3. there is a (unique) element e, called identity, such that a · e = e · a = a for all a ∈ A; 存在一个（唯一的）元素e，称为恒等式，使得对于所有a∈a，a·e=e·a=a
4. for every a ∈ A there is b ∈ A such that a · b = b · a = e: such an element is unique and called inverse of a

对于每个a∈A，都有b∈A使得a·b=b·a=e：这样的元素是唯一的，称为a的逆

If a · b = b · a for all a, b ∈ A the group is called abelian

如果a·b=b·a对于所有a, b∈a，则群称为阿贝尔群

b) Examples and properties示例和特性

 - Examples of abelian groups: 阿贝尔群的例子：

1. The integers Z, equipped with the usual sum; 整数Z，带有通常的和
2. The real numbers excluded 0, equipped with the usual product. 实数不包括0，配备通常的乘积

 -Examples of non-abelian groups: 非阿贝尔群的例子：

1. The n × n matrices with non-zero determinant, equipped with the usual product. 具有非零行列式的n×n矩阵，配备有常乘积

c) Proposition:

Let (A, ·) be a group. Then the cancellation law holds:设（A，·）为一个群。那么，撤销法适用：

Proof : By multiplying by  both sides of the equation a·b = a · c, one obtains  ·a·b =·a·c. Then, insofar  is the inverse of a, this expression reduces to e·b = e·c, which by the property of the identity e is exactly equal to b = c.

证明: 通过将方程a·b=a·c的两边乘以，可以得到·a·b =·a·c。然后，只要是a的倒数，这个表达式就简化为e·b=e·c，根据恒等式e的性质恰好等于b=c

12.4 The Nim group尼姆群

 -Proposition：

  The set of the natural numbers with operation ⊕ is an abelian group

⊕运算的自然数的集合是一个阿贝尔群

 -Proof:

  The identity element is of course 0. The inverse of n is n itself: e.g.

identity元素当然是0。n的逆是n本身：例如

Associativity and commutativity of ⊕ are easy to show;结合性和交换性

Therefore, the cancellation law holds: n1 ⊕ n2 = n1 ⊕ n3 implies n2 = n3

因此，相消定律成立：n1⊕n2=n1⊕n3意味着n2=n3。

12.5 The Bouton theorem定理

-Theorem (Bouton) :

In the Nim game the position (n1, n2, . . . , nk ) is a P-position if and only if

在尼姆对策中，位置（n1，n2，…，nk）是P-位置当且仅当

-Proof

1. Terminal position (0, 0, . . . , 0) is a P-position, with zero Nim-sum.终端位置（0，0，…，0）是一个P位置，Nim和为零。
2. Positions with  go only to positions with non-zero Nim-sum. For, suppose that the next position  is such that: then, by the cancelation law one has , which is impossible since the game requires  具有…的位置仅去往具有非零Nim和的位置。对于，假设下一个位置（…）是这样的：…：那么，根据取消定律，一个有着…，这是不可能的，因为游戏需要…。
3. Positions with  can go to positions with zero Nim-sum Let . Take a pile having 1 in the first column on the left of the expansion of z and put 0 there; then go right, leaving unchanged digits corresponding to 0 and changing them otherwise. Provably, the result is smaller than the original number (see next example). 具有…位置可以到达零Nim和的位置设z:=…。取一个在z展开左侧第一列中有1的桩，在那里放0；然后向右移动，保留与0相对应的不变数字，否则将进行更改。可以证明，结果小于原始数字（见下一个示例）。

Example:

From the following arrangement yielding non-zero Nim-sum

根据以下排列产生非零Nim和

by taking a card out of the first row we go to the next one with zero Nim-sum

从第一排取出一张牌，我们转到下一张零尼姆和的牌

NOTE: there are three initial good moves, one for each row.

注意：最初有三个好动作，每行一个。

13. Conclusions 结论

1. Games with perfect information can be ”solved” by using backward induction具有完美信息的游戏可以通过使用反向归纳来“解决”
2. However backward induction is a concrete solution method only for very simple games, because of limited rationality. 然而，由于有限的合理性，反向归纳法是一种仅适用于非常简单的游戏的具体解决方法。
3. Depending on the game, we can reach different levels of solutions:根据游戏的不同，我们可以达到不同级别的解决方案：

-Very weak solutions: not even the outcome is predictable (chess. . . )

非常弱的解决方案：甚至结果都不可预测（国际象棋…）

-Weak solutions: a logical argument provides the outcome, but how to reach it is not known (chomp, in general)

弱解决方案：逻辑论证提供了结果，但如何达到它还不清楚（chomp，一般来说）

-Solutions: categories of games where it is possible to produce the way to get to the rational outcome

解决方案：可以产生合理结果的游戏类别

14. Strategies

In Backward induction a move must be specified at any node.在反向诱导中，必须在任何节点指定移动。Let Pi be the set of all the nodes where player i is called upon to make a move

让Pi是所有节点的集合，玩家i被要求进行移动.

-Definition:

1. A pure strategy for player i is a function defined on the set Pi , associating to each node v in Pi a child x, or equivalently an edge (v, x); 玩家i的纯策略是在集合Pi上定义的函数，将Pi中的每个节点v与子节点x或等价的边（v，x）相关联
2. A mixed strategy is a probability distribution on the set of the pure strategies. 混合策略是纯策略集合上的概率分布

-When a player has n pure strategies, the set of her mixed strategies is

当一个玩家有n个纯策略时，她的混合策略集为

Σn is the fundamental simplex in n-dimensional space.∑n是n维空间中的基本单纯形

14.1 Strategies in a tree树中的策略

   

All combinations are listed even if equivalent (e.g. strategies b.. for Player II)

列出了所有组合，即使是等效的（例如玩家II的策略b..）

The table has repeated pairs: different strategies can lead to the same outcomes

该表有重复的配对：不同的策略可以导致相同的结果

15. Revisiting von Neumann重温冯·诺依曼

One can reformulate von Neumann’s theorem in terms of strategies as follows:

我们可以用策略来重新表述冯·诺依曼定理，如下所示：

Theorem

In the chess game one of the following alternatives holds: 在国际象棋游戏中，下列备选方案之一成立

1. the white has a winning strategy 白人有获胜策略
2. the black has a winning strategy 黑人有获胜策略
3. both players have a strategy leading them at least to a tie两名棋手都有一个策略，使他们至少打平

15.1 Outcomes chess in strategic form 1战略形式的结果国际象棋1

The White has a winning strategy, i.e. a raw with all W’s. 白人有一个获胜策略，即所有W的原始

15.2 Outcomes chess in strategic form 2战略形式的结果国际象棋2

The Black has a winning strategy, i.e. a column with all B’s. 黑人有一个获胜策略，即一个包含所有B的专栏

15.3 Outcomes chess in strategic form 3战略形态下的结果国际象棋3

The outcome of the game is (at least) a Tie for both players. 这场比赛的结果（至少）是双方打成平局。

It is T for row c and column 3. 第c行第3列为T。

15.4 This is excluded这是排除的

This is the case excluded by von Neumann

冯·诺依曼排除了这种情况

Remark: If Pi = {v1, . . . , vk } and vj has nj children, then the number of strategies of Player i is n1 · n2 · · · · · nk

备注：如果Pi={v1，…，vk}并且vj有nj个孩子，则玩家i的策略数为n1·n2····nk

This shows that the number of strategies even in short games is usually very high.

这表明，即使在短游戏中，策略的数量通常也很高

For instance, if Tic-Tac-Toe is stopped after three moves, the first player has (without exploiting symmetries) 9 · 7 ^(8×9) strategies

例如，如果Tic-Tac-Toe在三次移动后停止，第一个玩家有（不利用对称性）9*7^（8×9）策略

16. Games with imperfect information 信息不完善的游戏

Sometimes players must make moves at the same time, and so they cannot have full knowledge of each other’s moves. This can be still represented with a tree.

有时玩家必须同时做出动作，因此他们无法完全了解对方的动作。这仍然可以用树来表示。

Dashed line: the player does not know exactly which vertex she finds herself in.

虚线：玩家不知道自己到底在哪个顶点。

Figure: The prisoner dilemma repeated twice, after one shot the players see the result

图：囚犯困境重复了两次，一次射击后，玩家看到了结果

16.1 Information set信息集

Definition:

An information set for a player i is a pair (Ui , A(Ui)) with the following properties:

定义玩家i的信息集是一对（Ui，A（Ui）），具有以下属性：

1. Ui ⊂ Pi is a nonempty set of vertices v1, . . . , vk; Ui⊂Pi是顶点v1、…，vk的非空集合
2. each vj ∈ Ui has the same number of children; 每个vj∈Ui都有相同数量的子项
3. Ai(Ui) is a partition of the children of v1 ∪ · · · ∪ vk with the property that each element of the partition contains exactly one child of each vertex vj. Ai（Ui）是v1∪ · · · ∪ vk的子项的一个划分，具有该划分的每个元素恰好包含每个顶点vj的一个子项的性质

Accordingly, player i knows to be in Ui , but not in which vertex she is.

因此，玩家i知道在Ui中，但不知道她在哪个顶点。

The partition yields the choice function, meaning that each set in Ai(Ui) represents an available move for the player (graphically, it is the same choice, i.e. 分区产生了选择函数，这意味着Ai（Ui）中的每个集合代表玩家的可用移动（从图形上讲，这是相同的选择，即:

an edge, coming out of the different vertices: e.g. ’confess’ - ’confess’).

边缘，从不同的顶点出来：例如“坦白”-“坦白”）。

16.2 Formal definition形式定义

Definition:

An Extensive form Game with imperfect information is constituted by

具有不完全信息的广义博弈由

1. A finite set N = {1, . . . , n} of players ; 玩家的有限集合N={1，…，N}
2. A game tree (V, E, x0) ; 游戏树（V，E，x0）
3. A partition made by sets P1, P2, … , Pn+1 of the vertices which are not leaves 由集合P1，P2，…, Pn+1组成的分区，不是叶子的顶点
4. A partition  of the set Pi , for all i, with (U j i , A j i ) information set for all players i for all vertices j (with the same number of children) ;一个分区（Uj i），集合Pi的j＝1，…,  ki，对于所有i，具有（U j i，A j i）对于所有顶点j（具有相同数量的子）的所有玩家i的信息集
5. A probability distribution, for each vertex in Pn+1, defined on the edges going from the vertex to its children ; 对于Pn+1中的每个顶点，在从顶点到其子的边上定义的概率分布
6. An n-dimensional vector attached to each leaf. 附加到每个叶的n维向量

NOTE: If the partition comprises just a single vertex, then a game with imperfect information becomes the same as a game with perfect information.

注：如果分区只包括一个顶点，那么具有不完全信息的博弈与具有完全信息的游戏相同。

16.3 Strategies for imperfect information games不完全信息游戏的策略

Definition:

A pure strategy for player i in an imperfect information game is a function defined on the collection U of his information sets and assigning to each Ui in U an element of the partition A(Ui). A mixed strategy is a probability distribution over the pure strategies.

在不完全信息游戏中，玩家i的纯策略是在他的信息集的集合U上定义的函数，并为U中的每个Ui分配分区A（Ui）的元素。混合策略是在纯策略上的概率分布。

Remark:

A game of perfect information is a particular game of imperfect information where all information sets of all players are singletons (i.e. consist of only one vertex)

注：完全信息博弈是一种特殊的不完全信息博弈，其中所有参与者的所有信息集都是单体（即仅由一个顶点组