Documents of the final AI examination of South China University of Technology

AI review reference

Part 1 thread Theory

1-1. What is AI? This article describes the subject and ability.

Artificial intelligence uses computers to represent and execute human intelligent activities,

Artificial Intelligence (subject): A branch of computer science that involves research, design, and application of intelligent machines. His recent major goal is to study the use of machines to imitate and execute certain intellectual functions of the human brain, and to develop relevant theories and technologies.

Artificial Intelligence (capability): the intelligent behaviors that intelligent machines generally perform related to human intelligence, such as judgment, reasoning, proof, recognition, perception, understanding, communication, design, thinking, planning, learning and problem solving.

 

1-2. What thoughts and trends play an important role in the development of AI?

Mathematical Logic, new thoughts on computing

Control Theory

Expert Systems, machine learning, Computational Intelligence, artificial neural networks, and Behavioral Research

 

1-3. Why can I use machines (computers) to imitate human intelligence?

Human intelligence is a very complex behavior that has not yet been fully explained. However, some of the preliminary cognitive processes of humans currently work on computers in a similar way. Furthermore, it can be viewed as an intelligent information processing system. It is also called a symbolic operating system or a physical symbol system as an information processing system. The so-called symbol is the mode. Any mode is a symbol as long as it can be different from other modes. A complete symbolic system should have the following six basic functions:

(1) Input symbols;

(2) output symbols;

(3) storing symbols;

(4) copy the symbol;

(5) Build a symbolic structure: Identify the relationships between symbols and form a symbolic structure in the symbolic system;

(6) conditional migration: continue the activity process based on the existing symbols.

Assume that any system, if it can show intelligence, must be able to execute the above six functions; otherwise, if any system has these six functions, then it can show intelligence. This intelligence refers to the intelligence that humans possess. This assumption is called the assumption of the physical symbol system. The assumption of the physical symbol system is accompanied by three inferences,

Inference 1: since a person is intelligent, he or she must be a physical symbol system.

Inference 2: since a computer is a physical symbol system, it must be smart.

Conclusion 3: Since humans are physical symbol systems and computers are physical symbol systems, they can use computers to simulate human activities.

 

1-4. What are the main research and application fields of AI? Which of the following are new research hotspots?

Research fields: problem solving, logical reasoning and theorem proof, natural language understanding, automatic programming, expert systems, machine learning, neural networks, robotics, pattern recognition, machine vision, intelligent Control, intelligent retrieval, intelligent scheduling and command, Distributed Artificial Intelligence and agent, Computing Intelligence and evolutionary computing, data mining and Knowledge Discovery, artificial life, System and language tools.

Research hotspots: Expert Systems, machine learning, neural networks, robotics, pattern recognition, Distributed Artificial Intelligence and agent, data mining and knowledge discovery.

 

 

Part 1 Knowledge Representation

2-1. What is knowledge? What are the elements of knowledge? What are the methods for expressing knowledge?

Feigenbaum: knowledge is information that has been cut, shaped, interpreted, and transformed. Simply put, knowledge is processed information.

Bernstein: knowledge is composed of descriptions, relationships, and processes in specific fields.

Hayes-Roth: knowledge is a fact, belief, and heuristic rule. From the knowledge base point of view, knowledge is a symbolic representation of all relevant aspects involved in a certain field.

Knowledge elements: Facts, rules, control, meta-Knowledge

Knowledge Representation: Level 1 logical notation, generative knowledge notation, framework notation, Semantic Network notation, and object-oriented notation

 

2-2. What are the key points of the state space method, question reduction method, predicate logic method, and semantic network method? What are their essential links and similarities and differences?

(PPT none) state space method: This method is used to represent and solve problems based on the state and operator. State Space is generally used to represent the following methods: from an initial state, an operator is added each time, and the test sequence of the operator is established progressively until the target State is reached.

(No ppt) problem Protocol: a description of a known problem. This problem is eventually transformed into a set of subproblems through a series of transformations. The solutions to these subproblems can be obtained directly, this solves the initial problem. Essence of the problem Statute: starts reverse reasoning from the goal (the problem to be solved) and establishes subproblems and subproblems of subproblems, at last, the initial problem becomes a collection of ordinary primitive problems.

Predicate Logic Method: Using predicate formulas and first-order predicate calculus, the problem to be solved becomes a problem to be proved, then, the Solution Theorem and solution inversion are used to prove that a new statement is exported from a known correct statement, which proves that the new statement is also correct.

Semantic Network Method: a structured representation that consists of nodes and arcs or links. Nodes are used to represent objects, concepts, and States, and arcs are used to represent the relationship between nodes. The semantic network solution is a new semantic network with clear results after reasoning and matching. Semantic networks can be used to represent multivariate relationships. Extended semantic networks can represent more complex problems.

 

2-3. How do I use the predicates to express knowledge?

The predicate formula can be used to indicate the fact knowledge of a thing, such as its state, attribute, and concept, or the rule knowledge with a causal relationship between things.

General steps for expressing knowledge using the predicate Formula

1. Define the predicates and individuals to determine the exact meaning of each predicate and individual.

2. Assign specific values to the variables in each predicate based on the objects or concepts to be expressed.

3. According to the meaning of the knowledge to be expressed, connect each predicate with an appropriate connector to form a predicate formula.

 

2-4. What is a generative rule? What are the components of a production system? Describes the functions of each part.

A production system consists of three parts: a total database (or a global database), a production rule, and a control policy.

The general database, also known as the integrated database, context, and blackboard, is used to store the data structure of various current information during the solution process. When the premise of a rule in a production rule matches certain facts in the general database, the rule is activated and its conclusion is stored as a new fact in the general database.

A generative rule is a rule repository used to store a set of rules and exchange rules for a domain knowledge related to problem solving. The integrity, consistency, accuracy, flexibility, and rationality of Knowledge Organization of a rule repository will have an important impact on the operating efficiency and performance of the production system. (Courseware: a set of generative rules used to describe knowledge in the corresponding field .)

The control policy is a reasoning mechanism consisting of a group of programs used to control the operation of the generative system, determine the reasoning line in the Problem Solving Process, and implement the problem solving.

 

2-5. Describe the inference method and process of the inference engine of the generative system.

Generative system: Sometimes, the then part is used to specify actions. In this case, the rule-based system is called a reactive system or a generative system. Generative systems: Positive reasoning, reverse reasoning, and bidirectional reasoning.

Generative rules:

Forward Reasoning: starting from a group of predicates or propositions that represent facts, a group of generative rules are used to prove whether the predicates formula or proposition is true.

Reverse reasoning: Starting from the predicates or propositions that represent the target, a set of generative rules are used to prove that fact predicates or propositions are true. That is, a group of hypothetical targets are first proposed, and then these assumptions are verified one by one.

Two-way reasoning: the two-way reasoning strategy refers to the process from objective to fact reasoning and from fact to target reasoning at the same time, and a certain step in the reasoning process to achieve the matching of facts and goals.

 

2-6. How to express knowledge in framework notation? How can we use Semantic Network to express knowledge?

Framework notation is a structured knowledge representation developed on the basis of framework theory. It is suitable for expressing multiple types of knowledge. The basic idea of the Framework theory is that the human brain has stored a large number of typical scenarios. When new situations arise, the brain selects a basic knowledge structure called the framework from its memory, the specific content changes according to the new situation, forming an understanding of the new situation and remembering it in the human brain.

Semantic Network indicates knowledge: It generally consists of Some Basic Semantic Units. These Basic Semantic Units are called semantic elements and can be expressed as (node 1, arc, node 2) using the following three tuples)

Steps for expressing knowledge with semantic Networks

1. Determine the attributes of all objects and objects in the problem.

2. determine the relationship between the discussed objects.

3. nodes and arcs in the semantic network are organized based on the relationships involved in the semantic network, including adding nodes, arcs, and merging nodes.

4. Each object is used as a node of the semantic network, and the relationship between objects is used as the arc of each node in the network to form a semantic network.

 

Part 1 Knowledge Reasoning

3-1. What is reasoning? Task and classification of reasoning.

Reasoning refers to the thinking process of introducing another judgment based on a certain strategy from known judgment.

The basic task of reasoning is to launch another judgment from one judgment.

Category:

Deductive reasoning: The process of judging and exporting a special name or a single name from the full name

Inductive Reasoning: the reasoning process that summarizes general conclusions from a sufficient number of cases is a kind of reasoning from individual to general.

Default reasoning: it is also called default reasoning. It assumes that certain conditions already have the reasoning when the knowledge is incomplete.

 

3-2. What is replacement? What is unity? What is summary?

Replacement: In Predicate Logic, some inference rules can be applied to certain formulas and combination formula sets to generate new formulas. An important reasoning rule is pseudo-element reasoning, which is the calculation of the formula and the generated formula. Another reasoning rule is full-name reasoning. It is a combination formula generated by a combination formula, where it is any constant symbol. Pseudo-element reasoning and full-name reasoning can be applied at the same time. For example, a combination formula can be generated by a combination formula. This is the replacement of the pair to be searched.

Integration: Find the replacement of the variable to make the two expressions consistent, called Unity. If a replacement acts on each element of the expression set, it is used to represent the set of the replacement instance, indicating that the expression set is integrated. If there is a replacement mechanism: This is called the combination, because the function is to make the set a single form.

Conclusion: Based on the predicate formula, some reasoning rules, and the concept of integration of replacement, we can further study the principle of elimination. Some experts call it the principle of generalization.

 

3-3. What are the steps to form a clause? Use examples to describe.

 

3-4. Convert the following sentences into clauses:

~ ("X) {p (x) → {(" Y) [P (y) → P (f (x, y)] equals ("Y) [Q (x, y) → P (y)]}

~ ("X) {p (x) → {(" Y) [P (y) → P (f (x, y)] equals ("Y) [Q (x, y) → P (y)]}

(1) Eliminate the implication symbols (only apply the escape and ~ Symbol ~ A then B replaces a → B)

~ ("X ){~ P (x) Sums {("Y )[~ P (y) lead P (f (x, y)] lead ("Y )[~ Q (x, y) ∨ P (y)]}

(2) reduce the domain of negative symbols (each negative symbol ~ Only one predicate symbol is used at most, and di Morgan's law is repeatedly applied)

($ X ){~ {~ P (x) Sums {("Y )[~ P (y) lead P (f (x, y)] lead ("Y )[~ Q (x, y) ∨ P (y)]}

($ X) {p (x) bytes {~ {("Y )[~ P (y) lead P (f (x, y)] lead ("Y )[~ Q (x, y) ∨ P (y)]}

($ X) {p (x) rows {{~ ("Y )[~ P (y) ∨ P (f (x, y)] Then {~ ("Y )[~ Q (x, y) ∨ P (y)]}

($ X) {p (x) returns {($ Y) [P (y) returns ~ P (f (x, y)] Then ($ Y) [Q (x, y) Then ~ P (y)]}

(3) Normalize variables (rename dummy elements (fictitious variables) to ensure that each quantizer has its own unique dummy element)

($ X) {p (x) returns {($ Y) [P (y) returns ~ P (f (x, y)] Then ($ ω) [Q (x, ω) Then ~ P (ω)]}

(4) Eliminate the existence quantifiers (replace the constraint variables in the existence quantifiers with the skolem function, and then remove the existing quantifiers)

P (a) returns {[P (B) returns ~ P (f (a, B)] Then [q (a, c) Then ~ P (c)]}

(5) convert it into the first bundle: (move all full quantifiers to the left of the formula, and make the domain of each quantifiers include the entire part of the formula after the quantifiers)

 

(6) Turn the parent formula into a union Paradigm (any parent formula can be written as a union composed of a finite set of some predicates formulas and (or) negative analyses of the predicates formula)

P (a) Lead {[P (B) Lead Q (a, c)] lead [P (B) Lead ~ P (c)] bytes [~ P (f (a, B) ∨ Q (a, c)] Then [~ P (f (a, B) returns ~ P (c)]}

P (a) Lead [P (B) Lead Q (a, c)] lead [P (B) Lead ~ P (c)] bytes [~ P (f (a, B) ∨ Q (a, c)] Then [~ P (f (a, B) returns ~ P (c)]

(7) Remove the full-name quantifiers (all remaining quantifiers are quantified by the Full-name quantifiers. Delete the prefix, that is, remove the full-name quantifiers that obviously appear)

 

(8) Remove the hyphen (replacement with {a, B}) and the hyphen. Finally, a finite set is obtained, where each formula is the extraction of text)

P ()

P (B) ∨ Q (a, c)

P (B) bytes ~ P (c)

~ P (f (a, B) ∨ Q (a, c)

~ P (f (a, B) returns ~ P (c)

(9) change the variable name (you can replace the variable symbol name so that a variable symbol cannot appear in more than one clause)

P (X1)

P (Y1) Qaq (X2, ω 1)

P (Y2) Quit ~ P (ω 2)

~ P (f (X3, Y3) Qaq (X3, ω 3)

~ P (f (X4, Y4) bytes ~ P (ω 4)

 

3-5. Briefly describe the process of proving the theorem using the method of generalization (solving the Inverse Solution Process ).

A formula set S and the target formula L are given, and the objective formula L is verified by reverse evidence or inversion. The proof steps are as follows:

(1) Negative L, get ~ L;

(2) Change ~ L add to S;

(3) Add the new collection {~ L, s} is converted into clause set F;

(4) (previously) Applying the principle of elimination, trying to export an empty sub-sentence that represents a conflict

(Now ppt) the clause in clause set F is repeated. If an empty clause appears, the clause is stopped. At this time, it proves that l is always true.

 

 

3-6. How can I solve the problem and obtain the answer?

The process for obtaining answers to a question from the inversion tree is as follows:

(1) Add each clause generated by the negation of the target formula to the clause of the negation of the target formula;

(2) execute the same resolution as before Based on the inversion tree until a clause is obtained at the root of the tree;

(3) Use the root clause as an answer statement.

Essence: transform an inversion tree with nil at its root into a proof tree with a reply statement at its root.

 

3-7. What types of Reasoning Methods does it have with/or form deduction? Briefly describe the reasoning process

And/or form-based deduction: Positive deduction, Reverse Deduction, and two-way deduction;

Positive deduction:

Starting from known facts, the implication (F rule) is positively used for deductive reasoning until a termination condition of a target formula is obtained.

The sum and/or transformation of fact expressions: 1. Remove "à" in the formula; 2. Move "" to the position close to the predicate; 3. Rename the variable name; 4. The skolem function is introduced to remove the quantifiers. 5. The full quantifiers are eliminated, and the variable names in the major operators are different.

Steps to change the expression of domain knowledge into a prescribed form: 1. Remove "à" in the formula; 2. Move "" to the position close to the predicate; 3. Introduce the skolem function to remove the quantifiers; 4. Remove full-name quantifiers 5. Restore Implication

Reasoning Process:

1. Express known facts with a sum of/or trees

2. Match the left part of the F rule with/or the leaf node of the tree, and add the matched F rule to the/or tree.

3. Repeat Step (2) until a graph containing the target node as the end node is generated.

 

Reverse Deduction reasoning:

Starting from the problem to be proved (GOAL), the implication (rule B) is used in reverse reasoning until the termination conditions containing known facts are obtained.

The transformation process is similar to the transformation of known facts in forward deductive reasoning. First, remove the full-name quantifiers. The method is to use the skolem function of the variable with the specified quantifiers to remove the corresponding variable of the Full-name quantifiers. Then, remove the existing quantifiers.

Reasoning Process:

1. Represent the target formula with a sum of/or a tree.

2. Match the right part of the B rule with the/or leaf node of the tree, and add the B rule that matches successfully to the/or tree.

3. Repeat Step (2) until a consistent solution is generated for the termination on the fact node.

 

And/or two-way deductive reasoning:

Composed of two and/or tree structures that indicate the target and known facts. These and/or trees are operated by the forward deduction of the F rule and the Reverse Deduction of the B rule, respectively, in addition, the F rule is still set to the left of a single text, and the B rule is set to the right of a single text.

 

3-8. What is the significance of uncertainty reasoning? What are the uncertainties?

Problems encountered in the real world and relationships between things are often complicated. The randomness, ambiguity, incompleteness, and inaccuracy of objective things often lead to certain uncertainty in people's understanding. This means that if you still use the classic exact reasoning method for processing, the authenticity of things will inevitably not be reflected. Predicate, which must use Uncertainty Knowledge for reasoning in the case of incomplete and uncertain conditions, that is, uncertainty reasoning.

(Courseware) significance: Make Computer Simulation of human thinking closer to human real thinking process.

Uncertainty Reasoning is a kind of reasoning based on Uncertainty Knowledge Based on non-classical logic. It starts from the initial evidence of uncertainty and uses Uncertainty Knowledge, launch a certain degree of uncertainty and reasonable or almost reasonable conclusion.

There are two types of uncertainty: uncertainty about evidence and uncertainty about conclusion.

 

3-9. Under what circumstances do we need to use uncertain reasoning?

Factors should be taken into account when selecting the uncertainty Representation Method: fully consider the characteristics of domain problems; properly describe the uncertainty of specific problems; meet the actual needs of problem solving; facilitate the estimation of uncertainty in the reasoning process.

 

3-10. Briefly describe the main inference methods of probability method, subjective Bayes method, and credibility Method

Deterministic reasoning: conventional reasoning, elimination of deductive reasoning, and rule-based deductive reasoning.

Subjective Bayes method:

Credibility method: (inference algorithm) Uncertainty algorithm of combined evidence, transfer algorithm of uncertainty, and synthesis algorithm of the same hypothesis for multiple independent evidence.

Evidence Theory: (Inference Model) probability assignment function and probability-like function, representation of Knowledge uncertainty, representation of evidence uncertainty, representation of combined evidence uncertainty, and transfer of uncertainty algorithm.

 

 

Part 1 Machine Learning

4-1 What is learning and machine learning? Why study machine learning?

SIMON: learning is the enhancement or improvement of the system's ability in repetitive work, so that the system can perform the same or similar tasks in the next time, it is better or more efficient than the current one.

(Courseware) learning is an adaptive change made by the system, making the system more effective for the next time to complete the same or similar tasks.

Machine Learning: machine learning is a discipline that studies how to use machines to simulate human learning activities. A more rigorous approach is: machine learning is a learning that studies, acquires new knowledge and new skills, and recognizes existing knowledge.

The importance of machine learning in the new stage is manifested in the following aspects:

(1) machine learning has become a new edge discipline and has formed a course in colleges and universities.

(2) combined with various learning methods, the study of inheritance learning systems in various forms from each other is emerging.

(3) the unity of machine learning and artificial intelligence on various basic issues is being formed.

(4) the application scope of various learning methods is constantly expanding, and some products have been formed.

(5) The Research on Data Mining and knowledge discovery has formed a boom, and has been successfully applied in the fields of biomedicine, financial management, and commercial sales, injecting new vigor into machine learning.

(6) Academic Activities Related to machine learning are unprecedentedly active.

(Courseware) The importance of machine learning: machine learning is one of the main core fields of artificial intelligence, and also a key link and bottleneck of modern intelligent systems. It is hard to imagine: A system without learning functions can be called an intelligent system. Data from biology, finance, and network fields urgently needs to be analyzed or modeled.

 

4-2 explain the basic structure of the Machine Learning System and explain the functions of each part.

The environment provides certain information to the learning part of the system. The learning part uses this information to modify the knowledge base, so as to improve the efficiency of the tasks executed by the system. The execution part completes the tasks based on the knowledge base, at the same time, the obtained information is fed back to the learning part. In specific applications, the environment, knowledge base, and execution determine the specific work content. The problems to be solved by the Learning Department are completely determined by the above three parts.

 

4-3 explain the mode and method of inductive learning.

The general mode of inductive learning is:

Given: ① observation Statement (fact) F, used to indicate specific knowledge about certain objects, States, processes, etc.; ② initial inductive assertions assumed (may be blank); ③ background knowledge, defines knowledge, assumptions, and constraints related to observation statements, candidate assertions, and any relevant problem fields, including priority criteria that can characterize the nature of the inductive assertions.

Sum: assertion (hypothesis) h can repeat the implication or weak implication observation statement and satisfy the background knowledge.

Learning Method:

(1) Example learning: it is also known as instance learning. It is a method of summarizing general concepts through several examples related to a certain concept in the environment. In this learning method, the external environment (teacher) provides a group of examples (positive examples and inverse examples), which are a special set of knowledge, each example expresses knowledge that applies only to this example. The example study aims to summarize general knowledge applicable to a wider range from these special knowledge to cover all positive examples and exclude all counterexamples.

(2) observation Discovery Learning: Observation discovery learning is also known as descriptive generalization. Its goal is to determine a general description of a law or theory, portray the observation set, and define the nature of its class objects. Observation and discovery learning can be divided into observation and machine discovery. The former is used to cluster cases to form a conceptual description. The latter is used to discover rules and generate laws or rules.

 

4-4 briefly describe the basic process of Concept Learning. Use examples to describe it.

Can be considered as an object or event set. It is a subset selected from a larger set or a Boolean function defined in this larger set.

Definition of concept learning problems:

Given a sample set and whether each sample belongs to a certain concept annotation, how can we infer the general definition of this concept. It is also called approximation of Boolean functions from the sample.

Concept Learning refers to the extraction of a Boolean function from the input/output training sample of a Boolean function.

 

Known:

Instance set X: each instance X is described by six attributes. The value range of each attribute is determined.

Hypothesis set h: the combination of the value constraints of each hypothesis H described as six attributes

Objective concept C: A Boolean function with the variable being an instance

Training sample set D: positive and inverse examples of the target function (or target concept)

Solution:

Assume h in H, so that any X, h (x) = C (x) in X)

 

4-5 briefly describe the decision tree method and its application scenarios. In the process of constructing a decision tree, what principles are used for selecting test attributes? How to implement it?

Decision tree learning is a method to approach objective functions with discrete values. The functions learned in this method are represented as a decision tree. The learned decision tree can also be expressed as multiple if-then rules to improve readability.

A decision tree sorts instances from the following nodes to a leaf node. A leaf node is the classification of an instance. Each node in the tree specifies a test of an attribute of the instance, and each subsequent branch of the node corresponds to a possible value of this attribute. The method for classifying instances is to test the attributes specified by the node starting from the root node of the tree, and then move down according to the value corresponding to the attribute value of the given instance. Then, the process repeats in the Child tree with the new node as the root.

Decision Tree Learning is most suitable for issues with the following features:

(1) An instance is represented by an Attribute-value pair;

(2) the target function has discrete output values;

(3) description that may need to be analyzed;

(4) The training data may contain errors;

(5) The training data can contain instances with missing attribute values.

Optimal Classification attributes:

Information Gain:

Used to measure the ability of a given attribute to differentiate training samples.

The ID3 algorithm uses information gain to select attributes from candidate attributes in each step of the growth tree.

Use entropy to measure the uniformity of the sample:

Entropy depicts the purity of any sample set

Select measurement standard for properties-branch indicators:

Information Gain: Use information gain to measure the expected entropy reduction

Gain ratio:

Gini Index:

......

 

4-6 What are the features of Bayesian learning? What are the assumptions of Naive Bayes classifier based on? Use examples to describe.

Bayesian reasoning provides a probability method based on the following assumptions: The amount to be investigated follows a probability distribution and can be inferred based on these probabilities and observed data, bayesian reasoning provides a quantitative method to measure the confidence level of multiple assumptions. Bayesian reasoning provides the foundation for learning algorithms with direct operation probability, it also provides a theoretical framework for analysis of other algorithms.

Features of Bayesian Learning:

(1) Each observed training sample can incrementally reduce or increase the estimated probability of a hypothesis. Other algorithms will completely remove this assumption when a hypothesis is inconsistent with any of the same examples;

(2) The prior knowledge can determine the final probability of the hypothesis together with the observed data. The form of the prior knowledge is: 1) the prior probability of each candidate hypothesis; 2) probability Distribution of each possible hypothesis on the observed data;

(3) Bayesian methods allow assumptions to make predictions of uncertainty;

(4) new instance categories can be predicted by multiple assumptions, weighted by their probabilities;

(5) even when the computational complexity of Bayesian methods is high, they can still be used as an optimal decision-making standard to measure other methods.

 

Naive Bayes classifier introduces a simple assumption to avoid data sparsity: Given a target value, attribute values are mutually independent, that is, the condition.

 

4-7 What are the two elements of Bayes network? What are the conditional independence assumptions contained in Bayes networks? Briefly describe the Bayes network inference mode. Use examples to describe.

Two elements of Bayesian Networks: a set of conditional independence assumptions (which can be expressed as a directed acyclic graph) and a set of Partial conditional probabilities;

Precisely define conditional independence:

For random variables with three discrete values, the probability distribution that the given value follows is independent of the value.

The preceding formula is usually abbreviated

Extend to variable set:

When the following equation is created, it is called a variable set. When a variable set is given, the condition is independent of the variable set.

Inference mode: Bayesian networks can be used to deduce the values of certain target variables when the observed values of other variables are given.

Because the processing is a random variable, it generally does not assign an exact value to the target variable.

What really needs to be inferred is the probability distribution of the target variable, which specifies the probability that the target variable obtains each possible value under the condition that other variables are observed.

This reasoning step is simple when all other variables in the network are exactly known.

In general, Bayesian networks can be used to calculate the probability distribution of another variable in the network when you know the value or distribution of some variables.

 

4-8 What is integrated learning? What is the core idea of the boosting algorithm?

Integration learning is to select a set of assumptions as a whole from the hypothetical space called integration. It combines the classification predictions of new instances and then outputs the results.

Motivation-the error probability of multiple classifiers is always lower than that of a single classifier.

Boosting is the most widely used integrated learning method. Its core idea is weighted training.

The results obtained by a general learning algorithm usually contain many errors (or even a higher accuracy than a random conjecture, that is, as long as the accuracy rate is greater than 1/2) -This algorithm can be called "weak learning" or "basic learning ".

The boosting algorithm calls this weak learning algorithm multiple times and uses different training subsets (different weights are assigned to each sample) each time a weak hypothesis/algorithm is obtained from the called algorithm, the above weak assumptions are combined to generate a more accurate single hypothesis for better prediction results.

 

 

 

 

4-9 What is the learner l can be learned by PAC?

Definition: consider defining a conceptual category on an instance set with a length. The learner uses the hypothetical space. When all and upper distributions are satisfied and satisfied, the learner outputs a hypothesis with a minimum probability. In this case, the learner uses the PAC to learn. The time used is, And the polynomial function.

The definition requirements here meet two conditions. First, a hypothesis with an arbitrary low error rate () must be output with a high probability. Secondly, the learning process must be efficient, with a maximum increase in time in polynomial mode. The polynomial neutralization defines the intensity required for the output hypothesis, and defines the inherent complexity of the Instance space and concept category. Here, it is the length of the medium instance.

 

(Previous exercises)

4-13 when will the EM algorithm be used? Briefly describe the EM algorithm process.

The EM algorithm can be used in situations where the values of variables have never been directly observed, as long as the general form of probability distribution observed by these variables is known. The EM algorithm has been used to train Bayesian Networks and radial basis function networks. The EM algorithm is also the basis of many unsupervised clustering algorithms, and is used to learn the extensive use of Baum-Welch backward algorithms that can observe Markov models.

In the general form of the EM algorithm, it repeats the following two steps until convergence

Step 1: Estimation (e) Step: use the current hypothesis H and observed data X to estimate the probability distribution on y to calculate Q (H '| H)

Step 2: Maximize (m) Step: replace hypothesis h with the hypothesis H that maximizes the Q Function'

When the function q is continuous, the EM algorithm converges to a fixed point of the likelihood function p (Y | H. If the likelihood function has a single maximum value, the EM algorithm can converge to the global maximum likelihood estimation of H. Otherwise, it only guarantees convergence to a local maximum value.

 

Part 1 Outlook

5-1 How do you evaluate the development and debate of artificial intelligence? What is the relationship between debate and development?

The growth of any new things is not smooth sailing. Artificial intelligence is no exception. Since the birth of artificial intelligence in human society, it has aroused controversy. Since its launch in 1956, artificial intelligence has struggled and grown hard in a difficult environment. On the one hand, society has doubts about the scientific nature of artificial intelligence, or fear the development of artificial intelligence. On the other hand, the scientific community also expressed doubts about artificial intelligence.

True science, like any other truth, can never be suppressed. Artificial Intelligence Research is bound to eliminate the danger of rolling over the Yangtze River. In China, artificial intelligence has begun to usher in its spring.

With the development of artificial intelligence, it has evolved from a standalone to a brilliant competition.

As mentioned above, the computer scientific community has different opinions on artificial intelligence. Artificial Intelligence researchers also have different schools of thought. In recent years, they have been debating the basic theories and methods of artificial intelligence. This debate will certainly promote the further development of artificial intelligence.

 

5-2 What are the arguments of different schools of AI in terms of Theories, Methods, and technical routes?

1. Debate on the theory of Artificial Intelligence:

(1) symbolic: the cognitive primitive of a person is a symbol, and the cognitive process is the symbolic operation process. It considers people as a physical symbol system and computers as a physical symbol system. Therefore, we can use computers to simulate human intelligent behaviors, that is, using computer symbols to simulate human cognitive processes. That is to say, human thinking is operable. It also believes that knowledge is a form of information that forms the foundation of intelligence. The core issues of AI are knowledge representation, knowledge reasoning, and knowledge application. Knowledge can be expressed by symbols or used for reasoning. Therefore, it is possible to establish a unified theoretical system of knowledge-based human intelligence and machine intelligence.

(2) connection: the thought primitive of a person is a neuron rather than a symbolic processing process. It opposed the assumption of the physical symbol system and held that the human brain is not used for computers. It also proposed a connectionist brain working mode to replace the computer working mode of symbolic operations.

(3) Racism: it is considered that intelligence depends on perception and action (so it is referred to as racism), and the "Perception-action" model of intelligent behavior is proposed. Activists believe that Intelligence does not require knowledge, representation, or reasoning. Artificial Intelligence can evolve like human intelligence (so it is called evolutionary ); intelligent behavior can only interact with the surrounding environment in the real world. It is also believed that the description of true and objective things and the smart behavior work model of the world are too simplified and abstract, and therefore cannot truly reflect the existence of objective things.

2. Debate Over AI methods:

(1) symbolic: the research method of artificial intelligence should be a functional simulation method. By analyzing the functions and functions of human cognitive systems, we can use computers to simulate these functions to achieve artificial intelligence. The symbolic theory tries to use the mathematical logic method to establish a unified theoretical system of Artificial Intelligence. However, it has encountered many difficulties that cannot be solved temporarily and is rejected by other schools.

(2) connectionist: It is suggested that artificial intelligence should focus on structural simulation, that is, simulating the physiological and neural network structure of human beings, and that functions, structures, and intelligent behaviors are closely related. Different Structures show different functions and behaviors. A variety of Artificial Neural Networks and many learning algorithms have been proposed.

(3) racism: the research method of artificial intelligence should adopt behavior simulation methods, and the functional, structural, and intelligent behaviors should not be divided. Different behaviors show different functions and different control structures. Behavioral research methods are also under suspicion and criticism from other schools. They hold that they can only create intelligent insect behavior, rather than human intelligent behavior.

3. debate over the AI technology line:

(1) Special routes: Special intelligent computers, artificial intelligence software, special development tools, artificial intelligence languages, and other specialized devices should be developed and developed.

(2) general route: it is believed that general computer hardware and software can provide effective support for AI development and can solve a wide range of and general AI problems. The general route emphasizes the development of artificial intelligence application systems and artificial intelligence products. It should be combined with computer technology and mainstream technology, and knowledge engineering should be considered as a branch of software engineering.

(3) Hardware route: it is believed that the development of AI mainly relies on hardware technology. This route also believes that the development of smart machines mainly relies on a variety of smart hardware, smart tools and their solidification technologies.

(4) software route: emphasize that the development of artificial intelligence mainly relies on software technology. The software route holds that the development of intelligent machines mainly lies in the development of various Intelligent Software, tools and their application systems.

 

5-3 What are the impacts of the development of AI on humans? How can we explain the economic, social, and cultural aspects based on my own understanding?

1. Impact of AI on the economy:

(1) benefits of the expert system

(2) AI promotes the development of computer technology

2. Impact of AI on society:

(1) labor employment issues

(2) Social Structure Changes

(3) changes in the way of thinking and concepts

(4) psychological threats

(5) Risks of out-of-control technology

3. Impact of AI on culture:

(1) Improving human knowledge

(2) improving human language

(3) improving cultural life

 

5-4 Review the future development of AI.

1. Updated theoretical framework

2. Better Technology Integration

3. More mature application methods

 

5-5 What are your suggestions for the AI course and its teaching?