Machine Learning Overview

Machine Learning is to study how computers simulate or implement human learning behaviors to acquire new knowledge or skills and reorganize existing knowledge structures to continuously improve their own performance. It is the core of artificial intelligence and the fundamental way to make computers intelligent. It is applied in various fields of artificial intelligence. It mainly uses induction, synthesis, rather than translation.

Machine Learning is a discipline that describes the internal mechanism of understanding and studying learning and establishes theoretical methods for computer programs that can automatically improve their own level through learning. In recent years, machine learning theory has been successfully applied and developed in many application fields and has become one of the foundations and hotspots of computer science. Computer programs using machine learning methods have been successfully used in robot chess programs, speech recognition, credit card fraud monitoring, autonomous vehicle driving, Intelligent Robotics, and other application fields, in addition, the theoretical methods of machine learning are also used in the field of data mining for big datasets. In fact, machine learning methods can play a role in any experience that can be accumulated.

Learning ability is a very important feature of intelligent behavior, but it is still unclear about the learning mechanism. People have made various definitions of machine learning. H.a. Simon believes that 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. R. S. Michalski believes that learning is to construct or modify the representation of the things experienced. People engaged in Expert System Development think that learning is the acquisition of knowledge. These ideas have their own focus. The first focuses on the external behavior of learning, the second focuses on the internal process of learning, and the third focuses on the practicality of Knowledge Engineering.

Machine learning plays an important role in AI research. An intelligent system with no learning ability can hardly be called a real intelligent system. However, in the past, intelligent systems generally lacked the learning ability. For example, they cannot correct themselves when encountering errors; they do not improve their performance through experience; they do not automatically acquire and discover the required knowledge. Their reasoning is limited to deduction and lacks induction. Therefore, they can only prove existing facts and theorems, rather than discovering new theorems, laws, and rules. With the development of artificial intelligence, these limitations have become increasingly prominent. Under such circumstances, machine learning has gradually become one of the core of AI research. It has been applied to various branches of artificial intelligence, such as expert systems, automatic reasoning, natural language understanding, pattern recognition, computer vision, intelligent robots, and other fields. In particular, the bottleneck of knowledge acquisition in the expert system is typical. People have been trying to use machine learning methods to overcome it.

Machine Learning is based on the understanding of human learning mechanisms, such as physiology and cognitive science. It establishes computational models or cognitive models for human learning, and develops various learning theories and methods, study general learning algorithms and perform theoretical analysis to establish a learning system with specific applications for tasks. These research objectives interact with each other.

Machine Learning has been widely used, such as search engines, medical diagnostics, credit card fraud detection, securities market analysis, DNA sequencing, speech and handwriting recognition, strategic games, and robotics applications.

Since the first machine Academic Seminar held at Carnegie Mellon University in 1980, machine learning has developed rapidly and has become one of the central topics.

At present, the research work in the machine learning field focuses on the following three aspects:

(1) task-oriented research and analysis to improve the performance of a group of learning systems for scheduled tasks.
(2) Cognitive Models Study human learning processes and perform computer simulation.
(3) Theoretical Analysis explores various possible learning methods and algorithms independent of the application field theoretically.

Machine Learning is another important research area for AI applications following expert systems and one of the core research topics of artificial intelligence and neural computing. The existing computer systems and artificial intelligence systems have no learning ability, but at most they have very limited learning ability. Therefore, they cannot meet the new requirements of science and technology and production. This chapter first introduces the definition, meaning, and brief history of machine learning, then discusses the main strategies and basic structure of machine learning, and finally studies the methods and technologies of various machine learning methods one by one, including mechanical learning, interpreted learning, case-based learning, concept-based learning, analogy learning, and training neural network learning. The discussion of machine learning and the progress of machine learning research will certainly promote the further development of artificial intelligence and the entire science and technology.
I. Definition and research significance of machine learning

Learning is an important kind of intelligent behavior that humans possess. However, for a long time, there have been many different opinions. The following are some of the opinions of a sociologist, logologist, and psychologist. According to Simon, the AI master, learning is the enhancement or improvement of the system's own capabilities in repetitive work, so that the system will execute the same or similar tasks in the next time, it will be better or more efficient than it is now. Simon's definition of learning itself illustrates the important role of learning.

Can machines be capable of learning like humans? In 1959, Samuel (USA) designed a chess program, which has the ability to learn and can improve his chess skills in constant confrontation. Four years later, the program beat the designer himself. After another three years, this program defeated the United States in an undefeated championship that has remained competitive for eight years. This program demonstrates the capabilities of machine learning and raises many profound social and philosophical questions.

Whether the capabilities of a machine are superior. One of the main arguments of many people who hold negative opinions is that a machine is artificial and its performance and actions are completely defined by the designer, therefore, no matter how powerful it is, it will not surpass the designer himself. This kind of opinion is true for machines that do not have the learning ability, but it is worth considering for machines that have the learning ability, because the ability of such machines is continuously improved in the application, after a while, the designer does not know the level of its ability.

What is machine learning )? So far, there is no uniform definition of "Machine Learning", and it is difficult to give a universally accepted and accurate definition. To facilitate discussion and estimation of the progress of a discipline, it is necessary to give a definition of machine learning, even if the definition is incomplete and inadequate. As its name suggests, machine learning is a discipline that studies how to use machines to simulate human learning activities. Machine Learning is a learning that studies machines to acquire new knowledge and new skills and to recognize existing knowledge. The "machine" here refers to computers. It is now an electronic computer, and it may be a neutron computer, a photon computer, or a neural computer.

Ii. Development History of machine learning

Machine Learning is a young branch of artificial intelligence research. Its development process can be divided into four stages.
The first stage is from the middle of 1950s to the middle of 1960s, which is a warm period.
The second stage is the cool-down period of machine learning from the middle of 1960s to the middle of 1970s.
The third stage is from the middle of 1970s to the middle of 1980s, known as the Renaissance.

The latest stage of machine learning started in 1986.

Machine Learning enters a new stage in the following aspects:

(1) machine learning has become a new edge discipline and has formed a course in colleges and universities. It comprehensively applies psychology, biology and neurophysiology as well as mathematics, automation and computer science to form the theoretical basis of machine learning.

(2) combined with various learning methods, the study of various forms of integrated learning systems from each other is emerging. In particular, the coupling of the link learning symbol learning can better solve the problem of acquiring and improving knowledge and skills in the continuous signal processing.

(3) the unity of machine learning and artificial intelligence on various basic issues is being formed. For example, the combination of learning and problem solving, and the idea of Knowledge Expression for learning lead to the collective learning of the general intelligent system soar. The case-based method combined with analogy learning and problem solving has become an important direction of Experience Learning.

(4) the application scope of various learning methods is constantly expanding, and some products have been formed. The Knowledge Acquisition Tool for inductive learning has been widely used in diagnostic expert systems. Connection learning is dominant in voice and image recognition. Analysis learning has been used to design a comprehensive expert system. Genetic Algorithms and reinforcement learning have good application prospects in engineering control. Neural Networks coupled with the symbolic system will play a role in Intelligent Management and Intelligent Robot Motion Planning of enterprises.

(5) Academic Activities Related to machine learning are unprecedentedly active. In addition to the annual machine learning seminar, there are also computer learning theory conferences and genetic algorithm conferences.

Iii. Main Machine Learning Strategies

Learning is a complex intelligent activity. The learning process is closely related to the reasoning process, according to the amount of Reasoning Used in learning, generally, machine learning strategies can be divided into four types: machine learning, teaching, analogy learning, and case learning. The more Reasoning Used in learning, the stronger the system capability.

Iv. Basic Structure of the Machine Learning System

Indicates the basic structure of the learning system. 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. The following describes the impact of these three parts on the design learning system.

The most important factor affecting the design of the Learning System is the information that the environment provides to the system. Or, more specifically, the quality of information. The knowledge base stores general principles that guide the execution of some actions, but the information provided by the environment to the learning system is various. If the information quality is relatively high and the difference from general principles is small, the learning part is easier to process. If the learning system provides detailed information about the execution of specific actions in a disorganized manner, the learning system needs to delete unnecessary details and make a summary and promotion after obtaining sufficient data, the general principles of guiding actions are formed and put into the knowledge base. In this way, the task of learning is heavy and difficult to design.

Because the information obtained by the learning system is often incomplete, the reasoning performed by the Learning System is not completely reliable. The rules summarized by the learning system may be correct or incorrect. This should be verified through execution results. Correct rules can improve the system performance and should be retained; incorrect rules should be modified or deleted from the database.

Knowledge Base is the second factor that affects the design of the learning system. There are multiple forms of knowledge representation, such as feature vectors, first-order logic statements, generative rules, semantic networks, and frameworks. These representation methods have their own characteristics. You must consider the following four aspects when selecting the representation method:

(1) Strong expressiveness. (2) Easy reasoning. (3) It is easy to modify the knowledge base. (4) Knowledge Representation is easy to expand.

One problem that needs to be explained at the end of the knowledge base is that the learning system cannot obtain knowledge out of thin air without any knowledge. Every learning system requires the information provided by certain knowledge understanding environments, analyze and compare, make assumptions, test and modify these assumptions. Therefore, more specifically, the learning system expands and improves existing knowledge.

The execution part is the core of the entire learning system, because the action of the execution part is the action that the learning part strives to improve. There are three problems related to execution: complexity, feedback, and transparency.

V. Machine Learning Classification

1. Learning Policy-based classification

Learning Strategy refers to the reasoning strategy adopted by the system in the learning process. A learning system is always composed of two parts: Learning and environment. The information provided by the environment (such as books or teachers) is converted by the Learning part, stored in a understandable form, and obtained useful information. In the course of learning, the less reasoning a student (the learning part) uses, the greater his dependence on the teacher (Environment), and the heavier the burden on the teacher. The classification criteria of learning strategies are classified based on the number of reasoning required for information conversion and the difficulty level. There are five basic types of learning strategies:

1) mechanical learning)
Learners do not need any reasoning or other knowledge transformations to directly learn the information provided by the environment. For example, Samuel's checkers, Neel and Simon's lt systems. This type of learning system mainly considers how to index and utilize the stored knowledge. The systematic learning method is to directly learn through prepared and constructed programs in advance. learners do not do any work, or learn by directly receiving established facts and data, no reasoning is made for the input information.

2) learning from instruction or learning by being told ).
Students obtain information from the Environment (teachers or other information sources, such as textbooks), convert knowledge into an internal representation that can be used, and combine new knowledge with original knowledge organically. Therefore, students are required to have a certain degree of reasoning ability, but the environment still needs to do a lot of work. Teachers propose and organize knowledge in some form, so that students' knowledge can be continuously increased. This learning method is similar to the School Teaching Method in human society. The task of learning is to establish a system so that it can accept teaching and suggestions and effectively store and apply the learned knowledge. Currently, many expert systems use this method to acquire knowledge when creating a knowledge base. A typical application of teaching is the foo program.

3) learning by deduction ).
The reasoning used by students is in the form of translation reasoning. Reasoning is based on the principle and the conclusions are derived through logical transformation. This reasoning is a process of "Fidelity" transformation and specialization, so that students can obtain useful knowledge in the reasoning process. This learning method includes macro-operation learning, knowledge editing, and chunking technology. The inverse process of deductive reasoning is inductive reasoning.

4) learning by analogy ).
Knowledge similarity in two different fields (source domain and target domain) can be used to compare the knowledge in the source domain (including similar features and other properties) export the corresponding knowledge of the target domain to achieve learning. An analogy learning system can transform an existing computer application system into a new field to accomplish similar functions that were not designed previously. Analogy learning requires more reasoning than the above three learning methods. It generally requires that you first retrieve available knowledge from the knowledge source (source domain), and then convert it into a new form to use the new situation (target domain. Analogy learning plays an important role in the history of human science and technology development. Many scientific discoveries are achieved through analogy. For example, the famous lutherford analogy reveals the mysteries of the atomic structure by comparing the atomic structure (target domain) with the solar system (source domain.

5) explain-based learning (EBL ).
Based on the target concept provided by instructors, an example of this concept, a domain theory, and operational principles, students first construct an explanation to illustrate why this example satisfies the target concept, it will then be interpreted as a sufficient condition to meet the operational principles of the target concept. EBL has been widely used in improving the knowledge base and improving the system performance. The famous EBL systems include Genesis of Dijon (G. DeJong), lexii and leap of Mitchell (T. Mitchell), and prodigy Of mington (S. Minton.

6) learning from induction ).
Inductive learning is an example or inverse example of a concept provided by a teacher or environment. It allows students to come up with a general description of this concept through inductive reasoning. The reasoning workload of such learning is much higher than that of teaching and deduction, because the environment does not provide general concepts (such as justice ). To some extent, the reasoning of inductive learning is larger than that of analogy learning, because there is no similar concept that can be used as a "Source Concept. Inductive learning is the most basic and mature learning method. It has been widely studied and applied in the AI field.

2. classification based on the representation of acquired knowledge

The knowledge obtained by the learning system may include behavior rules, descriptions of physical objects, problem solving strategies, various categories, and other knowledge types used for task implementation.

The knowledge obtained in learning mainly has the following forms:
1) algebraic expression parameters: the goal of learning is to adjust the parameters or coefficients of an algebraic expression in the form of a fixed function to achieve an ideal performance.
2) Decision Tree: a decision tree is used to divide the class of an object. each node in the tree corresponds to an object attribute, and each side corresponds to the optional values of these attributes, the leaf node of the tree corresponds to each basic classification of objects.
3) Formal Grammar: In the process of identifying a specific language, a series of expressions of the language are summarized to form the formal grammar of the language.
4) generative rules: generative rules are represented as condition-Action pairs and are widely used. The main learning behaviors in the Learning System are generation, generalization, specialization, or synthesis of production rules.
5) Formal logical expressions: the basic components of a formal logical expression are statements that constrain the range of variables, and embedded logical expressions.
6) Graphs and networks: Some systems use Graph Matching and graph conversion schemes to effectively compare and index knowledge.
7) framework and Schema: Each framework contains a set of slots used to describe various aspects of things (Concepts and individuals.
8) computer programs and other process code: This form of knowledge is obtained to gain the ability to implement a specific process, rather than to deduce the internal structure of the process.
9) Neural Networks: This is mainly used in Join learning. The learned knowledge is finally summarized into a neural network.
10) combination of multiple representations: Sometimes the knowledge obtained in a learning system needs to be comprehensively applied to the above knowledge representations.

Based on the precision of representation, knowledge representation can be divided into two categories: coarse-grained Symbol Representation with a high degree of generalization ,?? Sub-symbolic. Such as decision trees, formal grammar, generative rules, formal logical expressions, frameworks, and patterns are symbolic representations; the Algebraic Expression parameters, graphs, networks, and neural networks are sub-symbol representations.

3. categories by application fields

Currently, the main application fields are: expert Systems, cognitive simulation, planning and problem solving, data mining, Network Information Services, image recognition, fault diagnosis, natural language understanding, robotics and games, and other fields.

According to the task types reflected in the execution of machine learning, most of the current application research fields are basically concentrated in the following two areas: Classification and problem solving.
(1) classification tasks require the system to analyze the input unknown mode (description of this mode) based on known classification knowledge to determine the category of the input mode. The learning goal is to learn the criteria for classification (such as classification rules ).
(2) The problem solving task requires the status of the given target ,?? Find an action sequence that converts the current status to the target status; most of the work of machine learning in this field is focused on learning to obtain knowledge that can improve the efficiency of problem solving (such as search control knowledge and heuristic knowledge ).

4. Comprehensive Classification

The historical origins, knowledge representation, reasoning strategies, similarity of results evaluation, the relative concentration of researchers, and application fields of various learning methods are comprehensively considered. Machine learning methods are divided into the following six categories:

1) Empirical inductive learning ).
Empirical inductive learning uses some data-intensive empirical methods (such as version Space Law, ID3 method, and Law Discovery Method) To summarize examples. The examples and learning results are generally represented by attributes, predicates, relationships, and other symbols. It is equivalent to inductive learning based on learning policy classification, but deducts the link learning, genetic algorithm, and reinforcement learning.

2) analytic Learning ?? Learning ).
The analytical learning method is to use domain knowledge for analysis from one or a few examples. Its main features are:
· Reasoning strategies are mainly deductive rather than inductive;
· Use past problem solving experience (examples) to guide new problem solving, or generate search control rules that can use domain knowledge more effectively.
The goal of analysis learning is to improve the system performance, rather than the new concept description. Analytic learning includes application interpretation learning, deductive learning, multi-level structure blocks, and macro operation learning.

3) analogy learning.
It is equivalent to analogy learning based on learning policy classification. Currently, this type of learning is more striking. The research is based on an analogy with the specific examples of past experiences. It is called case_based learning ), or referred to as "sample learning.

4) Genetic Algorithm (genetic ?? Algorithm ).
Genetic Algorithms simulate mutation, exchange, and Darwin's natural selection of biological reproduction (Survival of the fittest in each ecological environment ). It encodes the possible solution of the problem into a vector called an individual. Each element of the vector is called a gene, and the target function (corresponding to the natural selection criteria) is used to group (the set of individual) each individual in the evaluation can be selected, exchanged, mutations, and other genetic operations based on the evaluation value (fitness) to obtain new groups. Genetic algorithms are suitable for complex and difficult environments. For example, they carry a large volume of noise and irrelevant data, constantly update things, and cannot clearly and accurately define the problem objectives, and the value of the current behavior can be determined through a long execution process. Like neural networks, the study of genetic algorithms has evolved into an independent branch of artificial intelligence, represented by J. H. Holland ).

5) Join learning.
A typical connection model is an artificial neural network, which consists of some simple computing units called neurons and weighted connections between units.

6) Reinforcement Learning ).
The feature of enhanced learning is to determine and optimize the action selection through the trial and error interaction with the environment, so as to realize the so-called sequential decision-making task. In such a task, the Learning Mechanism selects and executes actions, which leads to changes in the system status and may receive a reinforcement signal (immediate return) to achieve interaction with the environment. Reinforcement signals are a type of rewards and punishments for system behavior. The objective of system learning is to find an appropriate action selection policy, that is, to select the Action Method in any given State, so that the generated action sequence can obtain an optimal result (for example, the cumulative immediate return is the largest ).

In comprehensive classification, empirical inductive learning, genetic algorithms, Join learning, and enhanced learning are all inductive learning. Empirical inductive learning uses symbolic representation, genetic algorithms, Join learning, and enhanced learning are represented by sub-symbols. Analytical learning is deductive learning.

In fact, analogy strategies can be seen as a combination of induction and deduction strategies. Therefore, the most basic learning strategy is induction and deduction.

From the perspective of the learning content, the induction strategy is used for learning because the input is summarized, and the learned knowledge is obviously beyond the scope of the original system knowledge base, the results have changed the knowledge deduction closure of the system. Therefore, this type of learning can be referred to as knowledge-level learning. However, learning using deduction strategies can improve the efficiency of the system, however, it can still be contained in the knowledge base of the original system, that is, the knowledge learned cannot change the deduction closure of the system. Therefore, this type of learning is also called symbolic learning.