Machine learning is to study how computers simulate or realize human learning behavior, so as to acquire new knowledge or skills, reorganize the existing knowledge structure and continuously improve their own performance. It is the core of artificial intelligence and the fundamental way to make computers intelligent. Its application covers all fields of artificial intelligence, and it mainly uses induction, synthesis rather than explanation.

Learning ability is a very important feature of intelligent behavior, but the learning mechanism is still unclear. People have given various definitions of machine learning. H.A.Simon believes that learning is an adaptive change made by the system, which makes the system more effective when completing the same or similar tasks next time. According to R.s.Michalski, learning is to construct or modify the representation you have experienced. People engaged in expert system development believe that learning is the acquisition of knowledge. These views have their own emphasis. The first emphasizes the external behavior effect of learning, the second emphasizes the internal process of learning, and the third is mainly from the practical point of view of knowledge engineering.

Machine learning plays a very important role in the research of artificial intelligence. An intelligent system without learning ability can hardly be called a real intelligent system, but previous intelligent systems generally lacked learning ability. For example, they can't correct themselves when they encounter mistakes; Will not improve their performance through experience; Will not automatically acquire and discover the required knowledge. Their reasoning is limited to deduction and lacks induction, so they can only prove existing facts and theorems at most, but can't find new theorems, laws and rules. With the in-depth development of artificial intelligence, these limitations have become more and more prominent. It is in this situation that machine learning has gradually become one of the cores of artificial intelligence research. Its application has been extended to various branches of artificial intelligence, such as expert system, automatic reasoning, natural language understanding, pattern recognition, computer vision, intelligent robots and other fields. Among them, the bottleneck of knowledge acquisition in expert system is particularly typical, and people have been trying to overcome it through machine learning.

The research of machine learning is to establish a computational model or cognitive model of human learning process, develop various learning theories and methods, study general learning algorithms and make theoretical analysis, and establish a task-oriented learning system with specific applications. These research goals influence and promote each other.

Since 1980, the first machine academic seminar was held in Carnegie Mellon University, the research of machine learning has developed rapidly and become one of the central topics.

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

(1) Task-oriented research and analysis A learning system that improves the performance of a set of scheduled tasks.

(2) Cognitive model studies human learning process and carries out computer simulation.

(3) Theoretical analysis explores various possible learning methods and algorithms independent of application fields.

Machine learning is another important research field of artificial intelligence application after expert system, and it is also one of the core research topics of artificial intelligence and neural computing. The existing computer systems and artificial intelligence systems have almost no learning ability, at most, the learning ability is very limited, which can not meet the new requirements of science and technology and production. This chapter will first introduce the definition, significance and brief history of machine learning, then discuss the main strategies and basic structure of machine learning, and finally study various methods and technologies of machine learning one by one, including machine learning, explanation-based 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 even the whole science and technology.

First, the definition and research significance of machine learning

Learning is an important intelligent behavior of human beings, but what is learning has been debated endlessly. Sociologists, logicians and psychologists all have different views. Simon, a master of artificial intelligence, believes that learning is the enhancement or improvement of the system's own ability in repeated work, so that the system can do better or more efficiently when performing the same task or similar task next time. Simon's definition of learning itself illustrates the important role of learning.

Can machines learn like humans? 1959, Samuel of the United States designed a chess program with learning ability, which can improve his chess skills in constant games. Four years later, the program beat the designer himself. Three years later, the program defeated an unbeaten champion in the United States who had won for eight years. This program shows people the ability of machine learning and raises many thought-provoking social and philosophical questions.

One of the main arguments of many people who have negative opinions on whether the machine can surpass human beings is that the machine is man-made, and its performance and action are completely stipulated by the designer, so its ability will not surpass the designer himself in any case. This opinion is true for machines without learning ability, but it is worth considering for machines with learning ability, because the ability of such machines is constantly improving in application. After a period of time, the designer himself doesn't know what level his ability has reached.

What is machine learning? There is no unified definition of "machine learning" so far, and it is difficult to give a recognized and accurate definition. In order to discuss and estimate the progress of this subject, it is necessary to give a definition of machine learning, even if this definition is incomplete and insufficient. As the name implies, machine learning is a subject that studies how to use machines to simulate human learning activities. Strictly speaking, machine learning is the study that machines acquire new knowledge and skills and identify existing knowledge. The "machine" mentioned here refers to the computer; Now it is an electronic computer, and in the future it may be a neutron computer, a photon computer and a neural computer.

Second, the development history of machine learning

Machine learning is a relatively young branch of artificial intelligence research, and its development process can be roughly divided into four periods.

The first stage is from the mid-1950s to the mid-1960s, which is a warm period. …& gt；

The second stage is from the mid-1960s to the mid-1970s, which is called the cooling-off period of machine learning.

The third stage is from the mid-1970s to the mid-1980s, which is called the renaissance period.

The latest stage of machine learning begins at 1986.

The important performance of machine learning entering a new stage is as follows:

(1) Machine learning has become a new frontier discipline and a course in colleges and universities. It integrates applied psychology, biology and neurophysiology as well as mathematics, automation and computer science, and forms the theoretical basis of machine learning.

(2) Various forms of comprehensive learning systems are emerging, which combine various learning methods and learn from each other. In particular, the coupling of connected learning symbol learning can better solve the problem of acquiring and refining knowledge and skills in continuous signal processing, which has been paid attention to.

(3) A unified view of various basic problems of machine learning and artificial intelligence is taking shape. For example, the idea that learning is combined with problem solving, and knowledge expression is simple and easy to learn, has produced chunk learning of SOAR, a general intelligent system. Case teaching method, which combines analogy learning with problem solving, has become an important direction of experiential learning.

(4) The application scope of various learning methods is expanding, and some of them have become commodities. The knowledge acquisition tool of inductive learning has been widely used in diagnostic classification expert system. Connecting learning plays a dominant role in audio-visual recognition. Analytical learning has been used to design a comprehensive expert system. Genetic algorithm and reinforcement learning have a good application prospect in engineering control. Neural network connection learning coupled with symbolic system will play a role in intelligent management of enterprises and motion planning of intelligent robots.

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

Third, the main strategies of machine learning

Learning is a complex intelligent activity, and the learning process is closely related to the reasoning process. According to the number of times reasoning is used in learning, the strategies adopted by machine learning can be roughly divided into four types: mechanical learning, teaching-based learning, analogy learning and example learning. The more reasoning used in learning, the stronger the system's ability.

Fourth, the basic structure of machine learning system

The figure above shows the basic structure of the learning system. The environment provides some information to the learning part of the system, and the learning part uses this information to modify the knowledge base to improve the efficiency of the system execution part to complete the task. The execution part completes the task according to the knowledge base and feeds back the obtained information to the learning part. In the specific application, the environment, knowledge base and execution part determine the specific work content, and the problems to be solved in the learning part are completely determined by the above three parts. Below we describe the influence of these three parts on the design learning system.

The most important factor affecting the design of learning system is the information provided by the environment to the system. Or more specifically, the quality of information. Knowledge base contains general principles to guide the execution of some actions, but the information provided by the environment to the learning system is varied. If the information quality is relatively high and the difference with the general principle is relatively small, then the learning part is easier to handle. If chaotic concrete information is provided to the learning system to guide the implementation of specific actions, the learning system needs to delete unnecessary details after obtaining enough data, summarize and popularize them, form general principles to guide actions, and put them into the knowledge base, so the task of the learning part is heavy and the design is difficult.

Because the information obtained by the learning system is often incomplete, the reasoning carried out by the learning system is not completely reliable, and the rules it summarizes are not necessarily correct. This should be tested by the implementation effect. Correct rules can improve the efficiency of the system 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 learning system. There are many forms of knowledge representation, such as feature vectors, first-order logical statements, production rules, semantic networks and frameworks. These representations have their own characteristics, and the following four aspects should be considered when choosing representations:

(1) has strong expression ability. (2) Easy to reason. (3) It is easy to modify the knowledge base. (4) Knowledge representation is easy to expand.

The last question about the knowledge base is that the learning system cannot acquire knowledge out of thin air without any knowledge at all. Every learning system needs some knowledge to understand the information provided by the environment, make analysis and comparison, make assumptions, and test and modify these assumptions. Therefore, to be more accurate, the learning system is an extension and perfection of the existing knowledge.

The executive part is the core of the whole learning system, because the action of the executive part is the action that the learning part tries to improve. There are three issues related to the operative part: complexity, feedback and transparency.

Five, machine learning classification

1, classification based on learning strategies

Learning strategy refers to the reasoning strategy adopted by the system in the learning process. Learning system always consists of learning and environment. The environment (such as books or teachers) provides information, and the learning part realizes information conversion, remembers it in an understandable form, and obtains useful information from it. In the process of learning, the less reasoning the student (learning part) uses, the greater his dependence on the teacher (environment) and the heavier the burden on the teacher. The classification standard of learning strategies is based on how much reasoning and difficulty students need to realize information conversion. Compliance from simple to complex, from less to more, can be divided into the following five basic types:

1) machine learning (rote learning)

Learners directly absorb the information provided by the environment without any reasoning or other knowledge transformation. Such as Samuel's checkers program, Newell and Simon's LT system. This learning system mainly considers how to index and use the stored knowledge. The learning mode of the system is to learn directly through pre-programmed and constructed programs. Learners do not do any work, or learn directly by receiving established facts and data, and do not make any reasoning about the input information.

2) Learn from guidance or from being told.

Students get information from the environment (teachers or other sources of information such as textbooks, etc. ), transform knowledge into an internally available expression, and organically combine new knowledge with original knowledge. So students are required to have certain reasoning ability, but the environment still needs to do a lot of work. Teachers put forward and organize knowledge in some form, which makes students' knowledge increase continuously. 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 what it has learned. At present, many expert systems use this method to acquire knowledge when building knowledge base. A typical application example of teaching and learning is FOO program.

3) Through deductive learning.

The form of reasoning used by students is deductive reasoning. Reasoning starts from axioms and draws conclusions through logical transformation. This kind of reasoning is a process of "fidelity" transformation and specialization, which enables students to acquire useful knowledge in the process of reasoning. This learning method includes macro operation learning, knowledge editing and chunk technology. The inverse process of deductive reasoning is inductive reasoning.

4) Draw inferences from one example to another.

Based on the similarity of knowledge in two different fields (source field and target field), the corresponding knowledge in the target field is deduced from the knowledge in the source field (including similar features and other properties) through analogy, so as to realize learning. Analogical learning system can transform an existing computer application system into a new field and complete similar functions that were not designed before. Analogy learning needs more reasoning than the above three learning methods. It generally requires that available knowledge be retrieved from the knowledge source (source domain) first, and then transformed into a new form and used in a new situation (target domain). Analogy learning plays an important role in the history of human science and technology development, and many scientific discoveries are obtained through analogy. For example, the famous Rutherford analogy reveals the mystery of atomic structure by comparing the atomic structure (target domain) with the solar system (source domain).

5) Interpretation-based learning (EBL).

According to the target concept provided by the teacher, an example of the concept, domain theory and operational criteria, students first construct an explanation to explain why this example conforms to the target concept, and then summarize this explanation as a sufficient condition for the target concept to conform to the operational criteria. EBL has been widely used to improve knowledge base and system performance. Famous EBL systems include GENESIS by G.DeJong, LEXII and LEAP by T.Mitchell and PRODIGY by S.Minton.

6) Learn from induction.

Inductive learning is to provide some examples or counterexamples of a concept by teachers or the environment, so that students can get a general description of the concept through inductive reasoning. The reasoning workload of this kind of learning is far greater than that of teaching learning and deductive learning, because the environment does not provide general concept description (such as axioms). To some extent, inductive learning is more inferential than analogical learning, because no similar concept can be used as a "source concept". Inductive learning is the most basic and mature learning method, which has been widely studied and applied in the field of artificial intelligence.

2. Classification based on existing knowledge representation

The knowledge acquired by the learning system may include: behavior rules, descriptions of physical objects, problem-solving strategies, various classifications and other types of knowledge used for task realization.

For the knowledge gained in learning, there are mainly the following expressions:

1) Algebraic parameters: The goal of learning is to adjust the algebraic parameters or coefficients of a fixed function to achieve an ideal performance.

2) Decision tree: use decision tree to classify objects. Each internal node in the tree corresponds to an object attribute, each edge corresponds to the optional values of these attributes, and the leaf nodes of the tree correspond to each basic classification of objects.

3) Formal grammar: In the process of learning to recognize a specific language, a series of expressions of the language are summarized to form the formal grammar of the language.

4) Production rules: Production rules are expressed as condition-action pairs and have been widely used. The learning behaviors in the learning system mainly include the generation, generalization, specialization or synthesis of production rules.

5) Formal logical expressions: The basic components of formal logical expressions are propositions, predicates, variables, statements that constrain the range of variables and embedded logical expressions.

6) Graph and network: Some systems use graph matching and graph transformation schemes to effectively compare and index knowledge.

7) Frames and Schemas: Each frame contains a set of slots to describe all aspects of things (concepts and individuals).

8) Computer programs and other process codes: The purpose of acquiring this form of knowledge is to acquire an ability to realize a specific process, not to infer the internal structure of the process.

9) Neural network: This is mainly used for connection learning. The acquired knowledge is finally summed up as neural network.

10) Combination of multiple representations: Sometimes the knowledge acquired in a learning system needs to be integrated with the above knowledge representations.

According to the fineness of representation, knowledge representation can be divided into two categories: coarse-grained symbolic representation with high generalization ability and sub-symbolic representation with low generalization ability. Decision trees, formal grammars, production rules, formal logical expressions, frameworks and patterns belong to symbolic representation classes; Algebraic expression parameters, graphs and networks, neural networks, etc. Belongs to the sub-symbol representation class.

3. Classification by application field

At present, the main application fields are: expert system, cognitive simulation, planning and problem solving, data mining, network information service, image recognition, fault diagnosis, natural language understanding, robots and games.

Judging from the task types embodied in the execution part of machine learning, most of the application research fields are basically concentrated in the following two categories: classification and problem solving.

(1) The classification task requires the system to analyze the unknown input patterns (descriptions of patterns) according to the known classification knowledge, so as to determine the categories of input patterns. The corresponding learning goal is to learn the standards of classification (such as classification rules).

(2) The task of solving problems requires finding an action sequence that transforms the current state into the target state for a given target state; The research work of machine learning in this field mostly focuses on acquiring knowledge (such as search control knowledge, heuristic knowledge and so on). The kind that can improve the efficiency of solving problems through learning.

4. Comprehensive classification

The historical origin, knowledge representation, reasoning strategy, similarity of result evaluation, relative concentration of communication between researchers and application fields of various learning methods are comprehensively considered. Machine learning methods are divided into the following six categories:

1) experiential inductive learning.

Empirical inductive learning uses some data-intensive empirical methods (such as version space method, ID3 method and rule discovery method) to learn examples. Its examples and learning results are generally represented by symbols such as attributes, predicates and relationships. In the classification based on learning strategy, it is equivalent to inductive learning, but the parts of link learning, genetic algorithm and reinforcement learning are deducted.

2) Analytical learning.

Analytical learning method is based on one or several examples, using domain knowledge for analysis. Its main features are:

Reasoning strategy is mainly deduction, not induction;

Use past experience (examples) to guide new problem solving, or generate search control rules that can make more effective use of domain knowledge.

The goal of analytical learning is to improve the performance of the system, not to describe new concepts. Analytical learning includes applied explanatory learning, deductive learning, multi-level structural chunks and macro-operational learning.

3) Analogical learning.

In the classification based on learning strategies, it is equivalent to analogical learning. At present, the striking research in this type of learning is to learn by analogy with specific cases of past experience, which is called case-based learning, or case-based learning for short.

4) Genetic algorithm.

Genetic algorithm simulates the variation and exchange of biological reproduction and Darwin's natural selection (every ecological environment is the survival of the fittest). It encodes the possible solution of the problem into a vector called an individual, and each element of the vector is called a gene. It uses an objective function (corresponding to natural selection criteria) to evaluate each individual in a population (a collection of individuals), and carries out genetic operations such as selection, exchange and mutation on individuals according to the evaluation value (fitness), thus obtaining a new population. Genetic algorithm is suitable for very complex and difficult environment, such as a lot of noise and irrelevant data, things are constantly updated, the problem goal can not be clearly and accurately defined, and the value of current behavior can only be determined through a long implementation process. The research of genetic algorithm, like neural network, has developed into an independent branch of artificial intelligence, and its representative is J.H.Holland.

5) Link learning.

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

6) Strengthen learning.

The characteristic of reinforcement learning is to determine and optimize the choice of actions through trial and error interaction with the environment, so as to realize the so-called sequential decision-making task. In this task, the learning mechanism will change the state of the system by selecting and executing actions, and may get some kind of reinforcement signal (return immediately), thus realizing the interaction with the environment. Strengthening signal is a standardized reward and punishment for system behavior. The goal of system learning is to find a suitable action selection strategy, that is, the method of choosing which action in any given state, so that the generated action sequence can obtain some optimal results (such as the maximum cumulative immediate return).

In the comprehensive classification, empirical inductive learning, genetic algorithm, link learning and reinforcement learning all belong to inductive learning, in which empirical inductive learning is represented by symbols, while genetic algorithm, link learning and reinforcement learning are represented by sub-symbols; Analytical learning belongs to deductive learning.

In fact, analogy strategy can be regarded as the synthesis of inductive and deductive strategies. So the most basic learning strategy is induction and deduction.

From the learning content, the learning with inductive strategy is obviously beyond the scope of the original system knowledge base, and the learning result changes the closed knowledge deduction of the system, so this type of learning can also be called knowledge-level learning; Learning with deductive strategy can improve the efficiency of the system, but it can still be included in the knowledge base of the original system, that is, the learned knowledge can not change the deductive closure of the system, so this type of learning is also called symbol learning.