Machine Learning is the study of how computers can simulate or mimic human learning behavior to acquire new knowledge or skills, reorganizing existing knowledge structures to continuously improve their performance. It is the core of artificial intelligence and the fundamental way to make computers intelligent, with applications across various fields of artificial intelligence, primarily using induction and synthesis rather than deduction.
Machine learning is a discipline focused on understanding and studying the intrinsic mechanisms of learning, establishing theoretical methods for creating computer programs that can automatically improve their performance through learning. In recent years, machine learning theories have been successfully applied and developed in many application fields, becoming one of the foundations and hotspots of computer science. Programs using machine learning methods have been successfully applied in fields such as robotic chess programs, speech recognition, credit card fraud detection, autonomous vehicle driving, intelligent robots, and more. Additionally, the theoretical methods of machine learning have also been used in the data mining field of large datasets. In fact, in any area where experience can be accumulated, machine learning methods can play a role.
Learning ability is a very important characteristic of intelligent behavior, yet the mechanisms of learning remain unclear to this day. Various definitions of machine learning have been proposed. H.A. Simon believes that learning is the adaptive change a system undergoes, making it more effective in completing the same or similar tasks next time. R.s. Michalski considers learning as the construction or modification of representations for experienced events. Those engaged in the development of expert systems view learning as knowledge acquisition. Each perspective emphasizes different aspects: the first focuses on the external behavioral effects of learning, the second emphasizes the internal processes of learning, and the third mainly approaches it from the practical perspective of knowledge engineering.
Machine learning holds a very important position in the study of artificial intelligence. An intelligent system lacking learning capabilities cannot truly be considered intelligent. However, past intelligent systems generally lacked the ability to learn. For instance, they could not self-correct when encountering errors; they did not improve their performance through experience; and they did not automatically acquire and discover the knowledge needed. Their reasoning was limited to deduction and lacked induction, thus they could at most prove existing facts and theorems but could not discover new theorems, laws, and rules. As artificial intelligence has developed further, these limitations have become increasingly prominent. In this context, machine learning has gradually become one of the cores of artificial intelligence research. Its applications span across various branches of artificial intelligence, such as expert systems, automated reasoning, natural language understanding, pattern recognition, computer vision, intelligent robots, and so on. A particularly typical issue is the knowledge acquisition bottleneck problem in expert systems, where efforts have been made to overcome it using machine learning methods.
The research on machine learning is based on understanding the mechanisms of human learning processes through physiology, cognitive science, and other fields, creating computational or cognitive models of human learning processes, developing various learning theories and methods, researching general learning algorithms, and conducting theoretical analyses to establish task-oriented learning systems with specific applications. These research goals influence and promote each other.
Machine learning has already found a wide range of applications, such as in search engines, medical diagnostics, credit card fraud detection, securities market analysis, DNA sequencing, speech and handwriting recognition, strategic games, and robotics.
Since the first academic conference on machine learning was held at Carnegie Mellon University in 1980, research in machine learning has rapidly developed and become a central topic.
Currently, research in the field of machine learning mainly revolves around three aspects:
(1) Task-oriented research: studying and analyzing learning systems that improve the performance of a predetermined set of tasks.
(2) Cognitive models: studying human learning processes and simulating them on computers.
(3) Theoretical analysis: exploring various possible learning methods and algorithms independent of application domains from a theoretical standpoint.
Machine learning is another important research area in artificial intelligence applications following expert systems, and it is also one of the core research topics in artificial intelligence and neural computation. Existing computer systems and artificial intelligence systems have very limited learning capabilities, which cannot meet the new demands posed by technology and production. This chapter will first introduce the definition, significance, and brief history of machine learning, then discuss the main strategies and basic structures of machine learning, and finally study various methods and techniques of machine learning, including mechanical learning, explanation-based learning, case-based learning, concept-based learning, analogy learning, and training neural networks. Discussions on machine learning and advancements in machine learning research will undoubtedly promote the further development of artificial intelligence and the entire field of science and technology.
1. Definition and Significance of Machine Learning
Learning is an important intelligent behavior possessed by humans, but what exactly is learning has been a topic of debate for a long time. Sociologists, logicians, and psychologists all have different views. According to artificial intelligence master Simon, learning is the enhancement or improvement of a system’s abilities through repeated work, making the system perform better or more efficiently in executing the same or similar tasks next time. Simon’s definition of learning itself illustrates the importance of learning.
Can machines possess learning abilities like humans? In 1959, Samuel in the United States designed a chess program that had learning capabilities, allowing it to improve its chess skills through continuous play. Four years later, this program defeated its creator. Three years after that, it defeated a champion who had remained undefeated for eight years. This program demonstrated the capabilities of machine learning and raised many thought-provoking social and philosophical questions.
Can machines surpass human capabilities? A main argument from many skeptics is that machines are artificial, and their performance and actions are entirely dictated by their designers, thus their capabilities cannot exceed those of their creators. This opinion holds true for machines lacking learning abilities, but it is worth considering for machines with learning capabilities, as their abilities improve over time in application, and after a while, even the designers may not know the extent of their capabilities.
What is machine learning? To this day, there is no unified definition of “machine learning,” and it is also quite difficult to provide an accepted and accurate definition. For the sake of discussion and estimating the progress of the discipline, it is necessary to define machine learning, even if this definition is incomplete and insufficient. As the name implies, machine learning is a discipline that studies how to use machines to simulate human learning activities. A more rigorous statement is that machine learning is a field that studies how machines acquire new knowledge and skills and recognize existing knowledge. Here, “machine” refers to computers; currently, they are electronic computers, but in the future, they may also be neutron computers, photon computers, or neural computers, etc.
2. History of Machine Learning Development
Machine learning is a relatively young branch of artificial intelligence research, and its development can be roughly divided into four periods.
The first stage was from the mid-1950s to the mid-1960s, which was a period of enthusiasm.
The second stage was from the mid-1960s to the mid-1970s, known as the calm period of machine learning.
The third stage was from the mid-1970s to the mid-1980s, referred to as the revival period.
The latest stage of machine learning began in 1986.
The important developments marking the new stage of machine learning include the following aspects:
(1) Machine learning has become a new marginal discipline and has formed a course in universities. It integrates psychology, biology, neurophysiology, mathematics, automation, and computer science to form the theoretical foundation of machine learning.
(2) Research on integrated learning systems that combine various learning methods and leverage their strengths is on the rise, especially the coupling of symbolic learning and connectionist learning, which has received attention for better addressing the acquisition and refinement of knowledge and skills in continuous signal processing.
(3) A unified viewpoint on various foundational issues of machine learning and artificial intelligence is forming. For example, the combination of learning and problem-solving, and the viewpoint that knowledge representation facilitates learning has led to the development of the general intelligent system SOAR’s chunk learning. The combination of analogy learning and problem-solving based on case methods has become an important direction in experiential learning.
(4) The application scope of various learning methods continues to expand, with some becoming commercial products. Tools for knowledge acquisition through inductive learning have been widely used in diagnostic classification expert systems. Connectionist learning has an advantage in speech and image recognition. Analytical learning has been applied in the design of integrated expert systems. Genetic algorithms and reinforcement learning show promising applications in engineering control. Connectionist learning coupled with symbolic systems will play a role in intelligent management and intelligent robot motion planning in enterprises.
(5) Academic activities related to machine learning are unprecedentedly active. Internationally, in addition to the annual machine learning symposium, there are also conferences on computer learning theory and genetic algorithms.
3. Main Strategies of Machine Learning
Learning is a complex intelligent activity, and the learning process is closely linked to the reasoning process. Based on the amount of reasoning used in learning, the strategies employed in machine learning can be roughly divided into four types: mechanical learning, learning by instruction, analogy learning, and learning by examples. The more reasoning used in learning, the stronger the system’s capabilities.
4. Basic Structure of Machine Learning Systems
The above diagram represents the basic structure of a learning system. The environment provides certain information to the learning part of the system, which uses this information to modify the knowledge base to enhance the effectiveness of the task completion by the execution part. The execution part completes tasks based on the knowledge base while feeding back the acquired information to the learning part. In specific applications, the environment, knowledge base, and execution part determine the specific work content, and the problems the learning part needs to solve are entirely determined by these three parts. Below we will describe how these three parts influence the design of learning systems.
The most important factor affecting the design of learning systems is the quality of the information provided by the environment to the system. The knowledge base stores general principles guiding the actions of the execution part, but the information provided to the learning system by the environment can be quite diverse. If the quality of the information is high and closely aligns with general principles, then the learning part can handle it more easily. However, if the information provided to the learning system is chaotic and consists of specific instructions for executing actions, then the learning system will need to gather sufficient data, remove unnecessary details, summarize, and generalize to form general principles for guiding actions to be stored in the knowledge base, making the task of the learning part more burdensome and complex to design.
Since the information obtained by learning systems is often incomplete, the reasoning performed by learning systems is not completely reliable, and the rules they summarize may be correct or incorrect. This must be verified through the execution results. Correct rules can enhance the system’s effectiveness and should be retained; incorrect rules should be modified or deleted from the database.
The knowledge base is the second factor affecting the design of learning systems. Knowledge can be represented in various forms, such as feature vectors, first-order logic statements, production rules, semantic networks, and frames. Each representation method has its characteristics, and when selecting a representation method, the following four aspects must be considered:
(1) Strong expressiveness. (2) Ease of reasoning. (3) Ease of modifying the knowledge base. (4) Knowledge representation should be easy to expand.
One final point regarding the knowledge base is that learning systems cannot acquire knowledge out of thin air without any prior knowledge; every learning system requires some existing knowledge to understand the information provided by the environment, analyze and compare it, make hypotheses, and test and modify these hypotheses. Therefore, more accurately, learning systems are extensions and improvements of existing knowledge.
The execution part is the core of the entire learning system, as the actions of the execution part are what the learning part strives to improve. There are three issues related to the execution part: complexity, feedback, and transparency.
5. Classification of Machine Learning
1. Classification Based on Learning Strategies
Learning strategies refer to the reasoning strategies used by the system during the learning process. A learning system always consists of two parts: learning and environment. The environment (such as books or teachers) provides information, and the learning part realizes information transformation, memorizing it in a form that can be understood and extracting useful information from it. In the learning process, the more reasoning the learner (the learning part) uses, the less dependent they are on the teacher (the environment), and the heavier the burden on the teacher becomes. The classification of learning strategies is based on the amount and difficulty of reasoning required for the learner to achieve information transformation, ordered from simple to complex, and from few to many, into the following five basic types:
1) Mechanical Learning (Rote learning)
The learner requires no reasoning or other knowledge transformations, directly absorbing the information provided by the environment. For example, Samuel’s checkers program and Newell and Simon’s LT system. This type of learning system mainly considers how to index and utilize stored knowledge. The system’s learning method is direct, learning through pre-programmed or constructed programs, with the learner doing no work, or learning by directly receiving established facts and data without reasoning about the input information.
2) Learning from Instruction (Learning from instruction or Learning by being told).
The learner acquires information from the environment (teachers or other information sources like textbooks), converting knowledge into an internal representation that can be used, and organically integrating new knowledge with existing knowledge. This requires the learner to have a certain degree of reasoning ability, but the environment still has to do a lot of work. The teacher presents and organizes knowledge in a way that allows the learner to continuously increase their knowledge. This method of learning is similar to the school teaching methods in human society, where the task of learning is to establish a system that can be taught and advised effectively store and apply the knowledge learned. Currently, many expert systems use this method to realize knowledge acquisition when establishing knowledge bases. A typical application of instructive learning is the FOO program.
3) Deductive Learning (Learning by deduction).
The reasoning form used by the learner is deductive reasoning. Reasoning starts from axioms and derives conclusions through logical transformations. This reasoning is a process of “faithful” transformation and specialization, allowing the learner to acquire useful knowledge during 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) Analogy Learning (Learning by analogy).
Using the similarity of knowledge between two different domains (source domain and target domain), it is possible to derive corresponding knowledge in the target domain from knowledge in the source domain (including similar features and other properties) through analogy, thus achieving learning. Analogy learning systems can transform an existing computer application system to adapt to a new domain and accomplish similar functions that were not originally designed. Analogy learning requires more reasoning than the previous three types of learning. It generally requires first retrieving usable knowledge from the knowledge source (source domain), then converting it into a new form to be applied to a new situation (target domain). Analogy learning plays an important role in the history of human scientific and technological development, with many scientific discoveries being made through analogy. For example, the famous Rutherford analogy revealed the mysteries of atomic structure by comparing it to the solar system.
5) Explanation-based Learning (EBL).
The learner constructs an explanation based on the target concept provided by the teacher, an example of that concept, domain theory, and operational criteria, to explain why the example satisfies the target concept, and then generalizes the explanation into a sufficient condition that satisfies the operational criteria for the target concept. EBL has been widely applied in knowledge refinement and improving system performance. Notable EBL systems include G. DeJong’s GENESIS, T. Mitchell’s LEXII and LEAP, as well as S. Minton’s PRODIGY.
6) Inductive Learning (Learning from induction).
Inductive learning is where the teacher or environment provides examples or counterexamples of a certain concept, allowing the learner to derive a general description of that concept through inductive reasoning. The reasoning workload of this type of learning is much greater than that of instructive learning and deductive learning because the environment does not provide general concept descriptions (such as axioms). To some extent, the reasoning required for inductive learning is also greater than that of analogy learning, as there is no similar concept that can be used as a “source concept”. Inductive learning is the most basic and well-developed learning method, and it has been widely researched and applied in the field of artificial intelligence.
2. Classification Based on the Representation of Acquired Knowledge
The knowledge acquired by learning systems may include: behavioral rules, descriptions of physical objects, problem-solving strategies, various classifications, and other types of knowledge used for task implementation.
The knowledge acquired during learning mainly has the following representation forms:
1) Algebraic expression parameters: The goal of learning is to adjust the parameters or coefficients of a fixed functional form of an algebraic expression to achieve ideal performance.
2) Decision trees: Decision trees are used to classify objects, where each internal node corresponds to an object attribute, and each edge corresponds to possible values of these attributes, with the leaf nodes corresponding to each basic classification of the object.
3) Formal grammar: In the study of recognizing a specific language, a formal grammar of that language is formed by inducing a series of expressions of that language.
4) Production rules: Production rules are represented as condition-action pairs and have been widely used. The learning behavior in learning systems mainly involves generating, generalizing, specializing, or synthesizing production rules.
5) Formal logic expressions: The basic components of formal logic expressions are propositions, predicates, variables, statements constraining the range of variables, and embedded logical expressions.
6) Graphs and networks: 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 used to describe various aspects of things (concepts and individuals).
8) Computer programs and other process encodings: The purpose of acquiring knowledge in this form is to obtain the ability to implement specific processes rather than to infer the internal structure of that process.
9) Neural networks: This is mainly used in connectionist learning. The knowledge acquired through learning is ultimately summarized into a neural network.
10) Combination of multiple representation forms: Sometimes, the knowledge acquired in a learning system needs to comprehensively apply several of the above knowledge representation forms.
Based on the fineness of representation, knowledge representation forms can be divided into two major categories: high-level symbolic representation with high generalization and low-level sub-symbolic representation with low generalization. Representation forms such as decision trees, formal grammars, production rules, formal logic expressions, frames, and schemas belong to symbolic representation; while algebraic expression parameters, graphs and networks, neural networks, etc., belong to sub-symbolic representation.
3. Classification by Application Field
The most important application fields currently include: expert systems, cognitive simulation, planning and problem-solving, data mining, web information services, image recognition, fault diagnosis, natural language understanding, robotics, and gaming.
From the perspective of the types of tasks reflected by the execution part of machine learning, most of the current application research fields are mainly concentrated in the following two categories: classification and problem-solving.
(1) Classification tasks require the system to analyze unknown patterns (descriptions of those patterns) based on known classification knowledge to determine the class of the input pattern. The corresponding learning objective is to learn the criteria for classification (such as classification rules).
(2) Problem-solving tasks require finding a sequence of actions that transforms the current state into a target state for a given target state; research in this field of machine learning is mostly focused on acquiring knowledge that can improve the efficiency of problem-solving (such as search control knowledge, heuristic knowledge, etc.).
4. Comprehensive Classification
Considering the historical origins of various learning methods, knowledge representation, reasoning strategies, similarities in result evaluation, the relative concentration of researchers’ exchanges, and application fields, machine learning methods can be distinguished into the following six categories:
1) Empirical Inductive Learning.
Empirical inductive learning uses data-intensive empirical methods (such as version space methods, ID3 methods, law discovery methods) to perform inductive learning on examples. The examples and learning results are generally represented using attributes, predicates, relations, etc. It corresponds to inductive learning in the classification based on learning strategies but excludes parts of connectionist learning, genetic algorithms, and reinforcement learning.
2) Analytical Learning.
Analytical learning methods start from one or a few instances and use domain knowledge for analysis. Its main features are:
·Reasoning strategies are primarily deductive rather than inductive;
·Using past problem-solving experiences (instances) to guide new problem solving or generate search control rules that can utilize domain knowledge more effectively.
The goal of analytical learning is to improve system performance rather than to describe new concepts. Analytical learning includes techniques such as explanation-based learning, deductive learning, multi-level structure chunking, and macro-operation learning.
3) Analogy Learning.
This corresponds to analogy learning in the classification based on learning strategies. Currently, a notable research focus in this type of learning is learning through analogy with specific past experiences, known as case-based learning or simply example learning.
4) Genetic Algorithms.
Genetic algorithms simulate biological reproduction, mutation, exchange, and Darwinian natural selection (survival of the fittest in each ecological environment). They encode possible solutions to problems as vectors, referred to as individuals, with each element of the vector called a gene, and evaluate each individual in the population (set of individuals) according to a target function (corresponding to natural selection criteria), performing selection, exchange, mutation, and other genetic operations based on evaluation values (fitness) to obtain a new population. Genetic algorithms are suitable for very complex and difficult environments, such as those with a lot of noise and irrelevant data, constantly changing things, problems where goals cannot be clearly and precisely defined, and situations where the value of current actions can only be determined after a long execution process. Like neural networks, research on genetic algorithms has developed into an independent branch of artificial intelligence, with J.H. Holland being a representative figure.
5) Connectionist Learning.
A typical connectionist model is realized as an artificial neural network, composed of simple computational units called neurons and weighted connections between the units.
6) Reinforcement Learning.
Reinforcement learning is characterized by determining and optimizing action choices through exploratory interactions with the environment (trial and error) to achieve what is known as sequential decision tasks. In such tasks, the learning mechanism interacts with the environment by selecting and executing actions, leading to changes in the system’s state, and potentially receiving some reinforcement signal (immediate reward), thereby achieving interaction with the environment. The reinforcement signal is a scalarized reward or punishment for the system’s behavior. The goal of system learning is to find a suitable action selection strategy, which means choosing which action to take in any given state, so that the resulting sequence of actions can achieve some optimal result (such as maximizing cumulative immediate rewards).
In comprehensive classification, empirical inductive learning, genetic algorithms, connectionist learning, and reinforcement learning all fall under the category of inductive learning, with empirical inductive learning using symbolic representation, while genetic algorithms, connectionist learning, and reinforcement learning use sub-symbolic representation; analytical learning falls under deductive learning.
In fact, the analogy strategy can be seen as a combination of inductive and deductive strategies. Therefore, the most fundamental learning strategies are only inductive and deductive.
From the perspective of learning content, learning using inductive strategies involves generalizing input, and the knowledge learned obviously exceeds the range that the original system knowledge base can contain, thus changing the system’s deductive closure; this type of learning can be referred to as knowledge-level learning; while learning using deductive strategies, although the learned knowledge can improve the system’s efficiency, it can still be contained within the original system’s knowledge base, meaning the learned knowledge does not change the system’s deductive closure; this type of learning is referred to as symbolic-level learning.
