1. What is machine learning
Machine Learning (ML) is a multi-field interdisciplinary subject involving probability theory, statistics, approximation theory, convex analysis, algorithm complexity theory and other disciplines. It specializes in how computers simulate or realize human learning behaviors to acquire new knowledge or skills, and to reorganize existing knowledge structures to continuously improve their performance. It is the core of artificial intelligence and the fundamental way to make computers intelligent. It is applied in all fields of artificial intelligence. It mainly uses induction and synthesis instead of deduction.
The definition of machine learning
Arthur Samuel (1959)：
Areas of study that give computers the ability to learn beyond identifying programming.
Tom Mitchell(1998)：
A program is considered to be able to learn from experience E, solve task T, and achieve a performance measure of P, if and only if, after having experience E, the program's performance in processing T improves as judged by P.
Third, the classification of machine learning
1. Classification based on learning strategies
Learning strategy refers to the reasoning strategy adopted by the system during the learning process. A learning system always consists of two parts, learning and environment. Information is provided by the environment (such as books or teachers), and the learning part realizes the transformation of information, memorizes it in an understandable form, and obtains useful information from it. In the learning process, the less reasoning the student (learning part) uses, the more dependent he is on the teacher (the environment) and the heavier the teacher's burden. The classification standards of learning strategies are classified according to the amount and difficulty of reasoning required by students to realize information conversion. They are divided into the following six basic types in order from simple to complex and from less to more:
1) Rote learning
The learner directly absorbs 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 type of learning system mainly considers how to index the stored knowledge and make use of it. The systematic learning method is to learn directly through pre-programmed and constructed programs. The learner does not do any work, or learns by directly receiving the established facts and data, and does not make any reasoning about the input information.
2）示教学习 (Learning from instruction或Learning by being told)
Students obtain information from the environment (teachers or other sources of information such as textbooks, etc.), convert knowledge into an internally usable representation, and organically combine new knowledge with existing knowledge. So students are required to have a certain level of reasoning ability, but the environment still has a lot of work to do. The teacher proposes and organizes knowledge in some form so that the knowledge possessed by the students 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 that enables it to receive instruction and advice, and to effectively store and apply the knowledge learned. Many expert systems use this method to achieve knowledge acquisition when building a knowledge base. A typical application example of teach-by-learn is the FOO program.
3) Learning by deduction
The form of reasoning used by students is deductive reasoning. Reasoning starts from axioms and deduces conclusions through logical transformations. This kind of reasoning is a process of "fidelity" transformation and specialization, so that students can acquire useful knowledge in the reasoning process. This learning method includes macro-operation learning, knowledge editing and chunking techniques. The inverse process of deductive reasoning is inductive reasoning.
4) Learning by analogy
Using the similarity of knowledge in two different domains (source domain and target domain), the corresponding knowledge of the target domain can be deduced from the knowledge of the source domain (including similar features and other properties) by analogy, so as to realize learning. The analog learning system can transform an existing computer application system into a new field to accomplish similar functions that were not originally designed. Learning by analogy requires more reasoning than the three learning styles above. It generally requires retrieving available knowledge from a knowledge source (source domain) and transforming it into a new form for use in a new situation (target domain). Learning by analogy 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 revealed the mystery of atomic structure by comparing the atomic structure (target domain) with the solar system (source domain).
5) Explanation-based learning (EBL)
Based on 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 the example satisfies the target concept, and then generalize the explanation to a sufficient level of the target concept that satisfies the operational criteria. condition. EBL has been widely used for knowledge base refinement and improving system performance. Famous EBL systems include G.DeJong's GENESIS, T.Mitchell's LEXII and LEAP, and S.Minton's PRODIGY.
6) Learning from induction
Inductive learning is provided by teachers or the environment with some examples or counter-examples of a concept, allowing students to derive a general description of the concept through inductive reasoning. This kind of learning involves far more reasoning than taught and deductive learning because the environment does not provide general conceptual descriptions (such as axioms). To some extent, inductive learning also has a larger amount of reasoning than analogical learning, because there is no similar concept that can be accessed as a "source concept". Inductive learning is the most basic and relatively mature learning method, which has been widely studied and applied in the field of artificial intelligence.
2. Representation classification based on acquired knowledge
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 acquired in learning, there are mainly the following representations:
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 a desired performance.
2) Decision tree
A decision tree is used to divide the categories of objects. Each internal node in the tree corresponds to an object attribute, and each edge corresponds to the optional values of these attributes. The leaf nodes of the tree correspond to each basic classification of the object.
3) Formal grammar
In the study of identifying a particular language, the formal grammar of the language is formed by induction of a series of expressions in the language.
4) Production rules
Production rules, expressed as condition-action pairs, have been widely used. The learning behaviors in the learning system are mainly: generating, generalizing, specializing or synthesizing production rules.
5) Formal logic expressions
The basic components of formal logical expressions are propositions, predicates, variables, statements that constrain the scope of variables, and embedded logical expressions.
6) Graphs and Networks
Some systems employ graph matching and graph transformation schemes to efficiently compare and index knowledge.
7) Framework and schema (schema)
Each frame contains a set of slots that describe various aspects of things (concepts and individuals).
8) Computer programs and other procedural codes
The purpose of acquiring this form of knowledge is to acquire an ability to implement a particular process, not to infer the internal structure of that process.
9) Neural Networks
This is mainly used in join learning. The knowledge acquired by learning is finally summarized into a neural network.
10) Combination of multiple representations
Sometimes the knowledge acquired in a learning system needs to comprehensively apply the above-mentioned knowledge representations.
3. According to the level of detail expressed
Knowledge representations can be divided into two categories: coarse-grained symbolic representations with high generalization and fine-grained sub-symbolic representations with low generalizations. Decision trees, formal grammars, production rules, formal logic expressions, frames and patterns belong to the class of symbolic representation; while parameters of algebraic expressions, graphs and networks, and neural networks belong to the class of subsymbolic representation.
4. Classification by application field
The main application areas 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.
5. From the perspective of the types of tasks reflected in the execution part of machine learning, most of the applied research fields are basically concentrated in the following two categories: classification and problem solving.
(1) The classification task requires the system to analyze the input unknown pattern (the description of the pattern) according to the known classification knowledge to determine the category of the input pattern. The corresponding learning goal is to learn the criteria for classification (such as classification rules).
(2) The problem solving task requires that for a given target state, find an action sequence that converts the current state into the target state; most of the research work in machine learning in this field focuses on learning to obtain information that can improve problem solving. Knowledge of efficiency (eg, search control knowledge, heuristic knowledge, etc.).
6. Comprehensive classification
The historical origin of the emergence of various learning methods, knowledge representation, reasoning strategies, similarity in outcome assessment, relative concentration of researcher exchanges, and application areas are considered comprehensively. Machine learning methods [1] are divided into the following six categories:
1) Empirical inductive learning
Experiential inductive learning adopts some data-intensive empirical methods (such as version space method, ID3 method, law discovery method) to perform inductive learning on examples. The examples and learning results are generally represented by symbols such as attributes, predicates, and relationships. It is equivalent to inductive learning in learning policy-based classification, but deducts the parts of join learning, genetic algorithm, and reinforcement learning.
2) Analytical learning
Analytical learning methods start from one or a few instances and apply domain knowledge for analysis. Its main features are:
1. The reasoning strategy is mainly deduction, not induction;
2. Use past problem solving experience (examples) to guide new problem solving, or generate search control rules that can more effectively use domain knowledge.
The goal of analytical learning is to improve the performance of the system, not to describe new concepts. Analytical learning includes applied interpretive learning, deductive learning, multi-level structural chunking, and macro-operation learning techniques.
3) Learning by analogy
It is equivalent to learning by analogy in learning policy-based classification. The more compelling research in this type of learning is learning by analogy with specific examples of past experiences, known as case-based learning, or simply example learning.
4) Genetic algorithm
Genetic algorithms simulate mutation, exchange and Darwinian natural selection (survival of the fittest in every ecological environment) in biological reproduction. It encodes the possible solutions of the problem as a vector, called an individual, and each element of the vector is called a gene, and uses the objective function (corresponding to the natural selection criterion) to evaluate each individual in the population (collection of individuals), According to the evaluation value (fitness), genetic operations such as selection, exchange, and mutation are performed on individuals to obtain a new population. Genetic algorithms are suitable for very complex and difficult environments, for example, with a lot of noise and irrelevant data, things are constantly updated, problem objectives cannot be clearly and precisely defined, and the value of current behavior can only be determined through a long execution process. Like neural networks, genetic algorithm research has developed into an independent branch of artificial intelligence, and its representative is JH Holland.
5) Connected Learning
A typical connection model is implemented as an artificial neural network, which consists of some simple computational units called neurons and weighted connections between the units.
6) Reinforcement learning
Reinforcement learning is characterized by the determination and optimization of action choices through trial and error interactions with the environment to achieve so-called sequential decision-making tasks. In such tasks, the learning mechanism interacts with the environment by selecting and performing actions that lead to changes in the state of the system, possibly with some kind of reinforcement signal (immediate reward). Reinforcing signals are scalar rewards and punishments for system behavior. The goal of system learning is to find an appropriate action selection strategy, that is, a method to choose which action to choose in any given state, so that the generated action sequence can obtain some optimal result (such as the maximum cumulative immediate reward).
In the comprehensive classification, empirical inductive learning, genetic algorithm, join learning and reinforcement learning all belong to inductive learning, in which empirical inductive learning adopts symbolic representation, while genetic algorithm, join learning and reinforcement learning use subsymbolic representation; analytical learning belongs to Deductive learning.
In fact, analogical strategies can be seen as a synthesis of inductive and deductive strategies. Therefore, the most basic learning strategies are induction and deduction.
From the perspective of learning content, the learning using the induction strategy is to induct the input, the knowledge learned obviously exceeds the scope of the original system knowledge base, and the learned results change the knowledge deduction closure of the system, so this This type of learning can also be called knowledge-level learning; and the learning using deductive strategies can improve the efficiency of the system, but it can still be contained in the knowledge base of the original system, that is, the knowledge learned cannot be changed. The deductive closure of the system, so this type of learning is also called symbolic level learning.
7. Classification of learning forms
1) Supervised learning (supervised learning)
Supervised learning, i.e. providing right and wrong indications during machine learning. Usually the data set contains the final result (0, 1). Algorithms allow the machine to reduce errors by itself. This type of learning is mainly used for classification and prediction (regression & classify). Supervised learning learns a function from a given training data set, and when new data arrives, the result can be predicted based on this function. The training set requirement for supervised learning is to include input and output, which can also be said to be features and targets. Objects in the training set are annotated by humans. Common supervised learning algorithms include regression analysis and statistical classification.
2) Unsupervised learning (unsupervised learning)
Unsupervised learning, also known as inductive learning (clustering), uses the K method (Kmeans), establishes a center (centriole), and reduces the error through looping and decreasing operations (iteration&descent) to achieve the purpose of classification.