文章目录
摘要
本文旨在呈现《人工智能：一种现代的方法》中的一些经典、有趣的习题。
《人工智能：一种现代的方法》的习题答案可以在我的CSDN资源页下载
更新：课本及习题答案下载
第一次写作时间：2018年07月08日 21:57:42
更新时间：2018年12月3日21:33:05
看到很多网友评论，发现大家对课本及习题答案的需求不小，然而CSDN无法上传这样的资料，所以这里建立一个百度云链接，大家可以自取，祝大家学习进步，考研的同学考研顺利：
链接：https://pan.baidu.com/s/19KGXMLv-WVHHfFfpVJ6VnA
提取码：eku2
第一章——绪论
1.1 用自己的语言定义（1）智能，（2）人工智能，（3）Agent，（4）理性，（5）逻辑推理。
答：1）Intelligence：the ability to apply knowledge in order to perfrom better in an environment.
2)Artificial intelligence: the study and construction of agent programs that perform well in a given environment, for a given agent architecture.
3)Agent: an entity that takes action in response to percepts from an environment.
4）理性：the property of a system which does the “right thing” given what it knows.
5）逻辑推理：the a process of deriving new sentences from old, such that the new sentences are necessarily true if the old ones are true.
1.2 反射行动（如从热炉子上缩手）是理性的嘛？是智能的嘛？
答：是理性的。because slower, deliberative actions would tend to result in more damage to the hand.
不是智能的。If “intelligent” means “applying knowledge” or “using thought and reasoning” then it does not require intelligence to make a reflex action.
第二章——智能Agent
###2.1
 **图2.1 习题2.1**
答：1）错。Perfect rationality refers to the ability to make good decisions given the sensor information received.
2）对。纯反射agent忽略之前的感知信息，所以在部分可感知的环境中是不能达到一个最优的状态估计的。For example, correspondence chess is played by sending moves; if the other player’s move is the current percept, a reflex agent could not keep track of the board state and would have to respond to, say, “a4” in the
same way regardless of the position in which it was played.
3）对。For example, in an environment with a single state, such that all actions have the same reward, it doesn’t matter which action is taken. More generally, any environment that is reward-invariant under permutation of the actions will satisfy this property.
4）错。The agent function, notionally speaking, takes as input the entire percept sequence up to that point, whereas the agent program takes the current percept only.
5）错。For example, the environment may contain Turing machines and input tapes and the agent’s job is to solve the halting problem; there is an agent function that specifies the right answers, but no agent program can implement it. Another example would be an agent function that requires solving intractable problem instances of arbitrary size in constant time.
图灵停机问题参考：[1]
6）对。This is a special case of (3); if it doesn’t matter which action you take, selecting randomly is rational
7）对。For example, we can arbitrarily modify the parts of the environment that are unreachable by any optimal policy as long as they stay unreachable
8）错。Some actions are stupid—and the agent may know this if it has a model of the environment—even if one cannot perceive the environment state.
9）错。Unless it draws the perfect hand, the agent can always lose if an opponent has better cards. This can happen for game after game. The correct statement is that the agent’s expected winnings are nonnegative.
2.2
 **图2.2.1 习题2.2**  **图2.2.2 答案2.2**
2.3
**图2.3.1 问题2.3**
答：
Agent：an entity that perceives and acts; Essentially any object qualifies; the key point is the way the object implements an agent function.
Agent function: a function that specifies the agent’s action in response to every possible percept sequence.
Agent program: that program which, combined with a machine architecture, implements an agent function. In our simple designs, the program takes a new percept on each invocation and returns an action.
Rationality: a property of agents that choose actions that maximize their expected utility, given the percepts to date.
Autonomy: a property of agents whose behavior is determined by their own experience rather than solely by their initial programming.
Reflex agent: an agent whose action depends only on the current percept.
Model-based agent: an agent whose action is derived directly from an internal model of the current world state that is updated over time.
Model-based agent: an agent whose action is derived directly from an internal model of the current world state that is updated over time.
Utility-based agent: an agent that selects actions that it believes will maximize the expected utility of the outcome state.
Learning agent: an agent whose behavior improves over time based on its experience.
3.第三章
###3.1 为什么问题的形式化要在目标的形式化之后？
答：目标形式化中，我们决定对世界的哪些方面感兴趣
这样的话，在问题形式化里面，我们就能直接决定留哪些东西，操作哪些。
###3.2
 **图3.2.1 问题3.2**  **图3.2.2 答案3.2**
之所以记录这个问题，是因为我对这个（b）问题比较感兴趣：有关可采纳函数的判断（不能高估实际花销）。
3.3
 **图3.3.1 问题3.3**  **图3.3.2 答案3.3**
记录这个问题，因为它的推导很巧妙，很有意思，而且非常之详细，让人不禁生出记录称赞的想法。
###3.4 野人和传教士问题
和3.2的形式化问题差不多。多了一个：要检查重复状态的操作。
具体答案+代码在 [2]
3.4
 **图3.4.1 问题3.4**
答：
state 状态：A state is a situation that an agent can find itself in. We distinguish two types of states: world states (the actual concrete situations in the real world) and representational states (the abstract descriptions of the real world that are used by the agent in deliberating about what to
do)
state space 状态空间：A state space is a graph whose nodes are the set of all states, and whose links are actions that transform one state into another
search tree 搜索树： is a tree (a graph with no undirected loops) in which the root node is the start state and the set of children for each node consists of the states reachable by taking any action.
search node 搜索节点：is a node in the search tree.
goal 目标：a state that the agent is trying to reach
action 行动：something that the agent can choose to do.
successor function 后继函数：a successor function described the agent’s options: given a state, it returns a set of (action, state) pairs, where each state is the state reachable by taking the action.
** branching factor 分支因子**：The branching factor in a search tree is the number of actions available to the agent
3.5
 **图3.5.1 问题3.5**  **图3.5.2 答案3.5**
这道题有一个很有意思的想法：抛弃序列结构（压扁状态空间），即把一些动作都组合起来，变成一个超级行动。这样的话整个搜索树深度为1.
坏处，如果发现Go中的第一个行动（unplug your battery）不是解，很难把后面也含这个行动的整合行动给排除掉。
3.6
 **图3.6.1 问题3.6**  **图3.6.2 答案3.6**
这些判断题也很有意思，但是答案的（c）应该是false。作者写错了。
3.7
 **图3.7.1 问题3.7**   **图3.7.2 答案3.7**
3.8
 **图3.8.1 问题3.8**  **图3.8.2 答案3.8**
为什么要放这道题？因为答案有错误的地方，状态空间应该是 n 2 ∗ 2 n 2 n^2*2^{n^2} n2∗2n2
3.9 找出一个状态空间，使用迭代加深搜索比深度优先搜索的性能要差很多（如，一个是 O ( n 2 ) O(n^2) O(n2),另一个是O(n)）。
答：Consider a domain in which every state has a single successor, and there is a single goal
at depth n. Then depth-first search will find the goal in n steps, whereas iterative deepening
search will take 1 + 2 + 3 + · · · + n = O ( n 2 ) O(n^2) O(n2) steps
3.10
 **图3.10.1 问题3.10**  **图3.10.2 答案3.10**
第四章 超越经典搜索
###4.1 在不确定、部分可观察的环境中，Agent怎么达成目标？
答：在完全可观察的、确定的环境下，Agent可以准确地计算出经过任何行动序列之后能达到什么状态，Agent总是知道自己处于什么状态，其传感器一开始告知agent初始状态，而在行动之后无需提供新的信息。
如果部分可观察、不确定，感知信息就变得非常有用了。
部分可观察：每个感知信息都可能缩小Agent可能的状态范围，这样也就使得Agent更容易达到目标。
环境不确定：感知信息告诉Agent某一行动的结果到底是什么。
在这两种情况中。问题的解不是一个序列（因为没法预知未来信息，Agent的未来行动依赖于未来感知信息），而是一个应急规划（也叫：策略）
应急规划描述了根据接收到的感知信息来决定行动。
 **图4.1 与或图搜索**
在图中，在或节点给行动，与节点考虑所有后继。
4.2
 **图4.2.1 问题4.2**
答案：
a)爬山法
b)宽度优先。Local beam search with one initial state and no limit on the number of states retained, resembles breadth-first search in that it adds one complete layer of nodes before adding the next layer. Starting from one state, the algorithm would be essentially identical to breadth-first search except that each layer is generated all at once.
c)Simulated annealing with T = 0 at all times: ignoring the fact that the termination step would be triggered immediately, the search would be identical to first-choice hill climbing because every downward successor would be rejected with probability 1. (Exercise may be modified in future printings.)
d)Simulated annealing with T = ∞ at all times is a random-walk search: it always accepts a new state.
e)Genetic algorithm with population size N = 1: if the population size is 1, then the two selected parents will be the same individual; crossover yields an exact copy of the individual; then there is a small chance of mutation. Thus, the algorithm executes a random walk in the space of individuals.
4.3
 **图4.3.1 问题4.3**  **图4.3.2 答案4.3**
###4.4 用局部搜索中的爬山法求解TSP问题
答：
· Connect all the cities into an arbitrary path.
· Pick two points along the path at random.
· Split the path at those points, producing three pieces.
· Try all six possible ways to connect the three pieces.
· Keep the best one, and reconnect the path accordingly.
· Iterate the steps above until no improvement is observed for a while.
第五章——对抗搜索
5.1 对alpha-beta剪枝的理解
参考[3]（提出了一个下界，钳制值的观点，我觉得比较有意思）
但是主要还是要参考书本上的例子和算法（伪代码），照着算法和例子走一遍程序就比较好理解了。
5.2
 **图1 问题5.2**  **图2 答案5.2-部分**
两个八数码问题。
形式化该问题：1）初始状态：两个随机地8数码 2）后继函数：在一个未解决的puzzle上做移动 3）目标测试：两个问题都达到目标状态 4）cost 移动步数。
这里有必要介绍一下八数码的状态空间 [4]
对于八数码问题的解决，首先要考虑是否有答案。每一个状态可认为是一个1×9的矩阵，问题即通过矩阵的变换，是否可以变换为目标状态对应的矩阵？由数学知识可知，可计算这两个有序数列的逆序值，如果两者都是偶数或奇数，则可通过变换到达，否则，这两个状态不可达。这样，就可以在具体解决问题之前判断出问题是否可解，从而可以避免不必要的搜索。
如果初始状态可以到达目标状态，那么采取什么样的方法呢？
常用的状态空间搜索有深度优先和广度优先。广度优先是从初始状态一层一层向下找，直到找到目标为止。深度优先是按照一定的顺序前查找完一个分支，再查找另一个分支，以至找到目标为止。广度和深度优先搜索有一个很大的缺陷就是他们都是在一个给定的状态空间中穷举。这在状态空间不大的情况下是很合适的算法，可是当状态空间十分大，且不预测的情况下就不可取了。他的效率实在太低，甚至不可完成。由于八数码问题状态空间共有9!个状态，对于八数码问题如果选定了初始状态和目标状态，有9!/2个状态要搜索，考虑到时间和空间的限制，在这里采用A*算法作为搜索策略。在这里就要用到启发式搜索
5.3
 **图1 问题5.3-部分**  **图2 答案5.3-部分**
这个问题也比较经典，所以在这里记录了。
第六章 约束满足问题
6.1 地图着色问题有多少个解？
 **图1 问题6.1** 这个用度启发式，先找到SA，后面的就好取了。 3*2*3=18个解。
6.2
 **图1 问题6.2**  **图2 答案6.2**
这道题有点像八皇后的变形，答案给了A，B两种解法，都很有意思的。
6.3
 **图1 问题6.3**  **图2 答案6.3**
有时候感觉类比过来很麻烦啊，根本想不到。想不通。不是那么简单明了
这道题的（c）我没看懂。记录之。
6.4
 **图1 问题6.4**
答：这道题的答案给出了两种CSP的形式化定义。
1）25个变量，5个颜色，5个宠物，5个国家，5种饮料，5种香烟。 值就是对应的房间号。
2）5个变量，5个房间。值域就是每个房间对应的5个属性。
第七章 逻辑Agent
7.1
 **图1 问题7.1**
答案：只要KB为真的时候，两个语句为真，那么就是真。
7.2
 **图1 问题7.2**
之所以列出这个问题，是希望举一反三，用7.1 的方法罗列出所有世界模型即可。
7.3
 **图1 问题7.3**  **图2 答案7.3**
基本功题目。转成CNF。
7.4
 **图1 问题7.4**  **图2 答案7.4**
这道题也比较基础吧，记录一下。
第八章 一阶逻辑
8.1
 **图1 问题8.1**  **图2 答案8.1**
8.2
 **图1 问题8.2**  **图2 答案8.2**
一阶逻辑推理。
8.3
 **图1 问题8.3**  **图2 答案8.3**
用一阶逻辑表示。
8.4
 **图1 问题8.4**  **图2 答案8.4**
比较有趣的一个逻辑推理吧。
8.5
 **图1 问题8.5**  **图2 答案8.5**
最后一问有些难。
第九章——一阶逻辑的推理
###9.1 什么是前向链接？
从知识库KB中的原子语句出发，在前向推理中应用假言推理规则，增加新的原子语句，直到不能进行任何推理。
9.2
 **图1 问题9.2**  **图2 答案9.2** 一个存在量词的代换。
第十章——经典规划
10.1 问题求解和规划之间的不同和相似之处
相同：Getting from a start state to a goal using a set of defined operations or actions, typically in a deterministic, discrete, observable environmnet.
不同： In planning, however, we open up the representation of states, goals, and plans, which allows for a wider variety of algorithms that decompose the search space, search forwards or backwards, and use automated generation of heuristic functions.
10.2
 **图1 问题10.2**  **图2 答案10.2**
10.3 证明PDDL问题的后向搜索是完备的
答：Briefly, the reason is the same as for forward search: in the absence of function symbols, a PDDL state space is finite. Hence any complete search algorithm will be complete for PDDL planning, whether forward or backward
10.4 为什么在规划问题中去掉每个动作模式的负效果会得到一个松弛问题
答：Goals and preconditions can only be positive literals. So a negative effect can only make it harder to achieve a goal (or a precondition to an action that achieves the goal). Therefore, eliminating all negative effects only makes a problem easier. This would not be true if negative preconditions and goals were allowed.
第十一章——现实世界的规划与行动
11.1 暂无，这一章有点难，我有空再加。
第十三章——不确定性的量化
13.1 贝叶斯规则
**图1 答案13.1**
13.2 使用概率公理证明：一个离散随机变量的任何概率分布，总和等于1
 **图1 答案13.2**
我觉得这道题还有有意思的，这里记录一下。
13.3
 **图1 问题13.3**  **图2 答案13.3**
13.4
 **图1 问题13.4**  **图2 答案13.4**
一个文本分类的简单变形。
第十四章——概率推理
###14.1
 **图1 问题14.1**   **图2 答案14.1**
资源下载
如果想浏览全部答案，可以到我的CSDN资源页下载.
文末诗词
子规夜半犹啼血，不信东风唤不回。 ——王令《送春》
参考文献
[1] 图灵停机问题(halting problem). https://blog.csdn.net/MyLinChi/article/details/79044156
[2] 传教士野人过河问题. https://blog.csdn.net/aidayei/article/details/6768696
[3] α-β剪枝算法. https://blog.csdn.net/luningcsdn/article/details/50930276
[4] 八数码问题. https://www.cnblogs.com/guanghe/p/5485816.html
