The mainstream development of artificial intelligence is from “Heuristic Programs” to “Expert Systems”.
1. Heuristic Programs
Heuristic programs are computer programs that simulate human thinking methods and rules, used to simulate and explore human thinking methods and intelligent activities in various problem-solving processes, in order to improve the level of artificial intelligence in computer applications and assist, extend, or expand people’s problem-solving capabilities.
The first famous heuristic program is the “Logic Theory Machine,” abbreviated as “LT”. It was successfully developed in 1956 by Newell (A. Newell), Simon, and Shaw (J. C. Shaw). This was a product of the collaboration between psychologists and computer experts.
The heuristic program simulates certain thinking methods and rules used by mathematicians in the process of proving mathematical theorems, applying psychological methods to design a psychological experiment called “thinking aloud” to record and analyze the thinking process and intelligent activities, studying and exploring methods and rules such as: problem decomposition, variable substitution, symbol replacement, etc. These rules and relevant mathematical axioms were written into a computer program, thus allowing the computer to prove mathematical theorems. Using heuristic programs, it proved 38 mathematical theorems from Chapter 2 of Whitehead and Russell’s famous work “Principia Mathematica” (1910-1913), pioneering the use of computers to simulate advanced human intelligent activities and achieving automation of complex mental labor, which is considered the true beginning of artificial intelligence.
In 1956, another important achievement of heuristic programs was the development of a self-learning “checkers program” by Samuel, which simulated the playing strategies and methods of Samuel himself and other skilled players, capable of estimating several moves in advance, accumulating experience and lessons from playing, learning from skilled opponents or game records, and continuously improving the level of artificial intelligence in checkers. Initially, this checkers program could only play against Samuel himself. Later, after learning from other masters, it defeated its creator in 1959. In 1962, it also defeated the checkers champion of a U.S. state. This research opened up and promoted the study of “machine gaming” and “machine learning” in the field of artificial intelligence.
The representative work of the further development of heuristic programs is the “General Problem Solver” (GPS), developed by Newell, Simon, and Shaw in 1960. They believed that there are certain common methods and rules in people’s thinking processes when solving various problems, which were discovered and verified through psychological experiments. Among these, the most active intelligent activity is the analysis of “goals and methods,” that is, what methods need to be used to achieve the goal of solving a specific problem. Typically, the common thinking methods are as follows:
1. Transformation and Decomposition
Transformation is converting one problem into another isomorphic or homomorphic problem, and transforming a certain goal into another equivalent or equivalent goal; decomposition is breaking down complex large problems into simple small problems, and decomposing the overall goal into local sub-goals.
2. Elimination of Differences
This means using appropriate methods to eliminate the differences between the current state and the goal state, as well as the differences between interim goals and final goals, such as eliminating the differences on both sides of an equation.
3. Use of Operators
Methods are a collection of various operators, which are various operations, calculations, reasoning, actions, tools, means, etc. used to eliminate goal differences.
4. Selection and Matching
In the process of using operators and solving problems by eliminating differences, for various goals, sub-goals, and related methods and operators, selection and matching must be carried out to effectively solve the problem using appropriate methods.
Based on the above analysis, the heuristic program GPS was compiled, capable of solving 11 different types of problems, enhancing the generality of heuristic programs and expanding the application scope of automating mental labor using computers.
In 1960, Chinese-American mathematician Wang Hao proposed a new algorithm for machine theorem proving of propositional logic, using computers to prove over 300 theorems in set theory.
In 1965, Robinson (J. A. Robinson) proposed the “Resolution Principle” of first-order predicate logic, simplifying the decision-making process and advancing machine theorem proving based on predicate logic.
In 1977, Chinese mathematician Wu Wenjun proposed a machine theorem proving method for elementary geometry, further extending to elementary differential geometry and non-Euclidean geometry.
2. Expert Systems
In the field of artificial intelligence, the significant shift in scientific research methods and system development strategies regarding how to use machines to simulate human intelligence, especially how to use computers to simulate human thinking, is marked by the emergence of “Expert Systems”.
Expert systems are computer systems that solve specialized problems based on the professional knowledge and work experience of experts, such as medical expert systems, which are based on doctors’ knowledge and experience for diagnosing and treating diseases. Expert systems simulate the thinking processes of experts themselves, possessing a level of professional knowledge similar to that of experts and the ability to solve specialized problems. Artificial intelligence expert systems can simulate, extend, and enhance the intelligence of experts, and can be used to preserve, disseminate, and gather the professional knowledge and valuable experiences of experts from various fields, facilitating mutual communication and widespread application.
The first expert system, DENDRAL, is a chemical analysis expert system proposed by American scientist Feigenbaum (E. A. Feigenbaum) in 1965 and successfully developed in 1968. It can analyze experimental data from mass spectrometers to infer the molecular structure of unknown compounds, possessing knowledge and analytical capabilities similar to those of chemical experts. The emergence of the chemical analysis expert system DENDRAL marked the birth of a new branch of artificial intelligence—”Expert Systems”.
The medical expert system MYCIN was developed by E. H. Shortliffe and others at Stanford University, starting in 1971, basically completed in 1974, and published in 1976. It possesses the knowledge and experience of internists and can be used for the diagnosis, treatment, and consultation of blood infections. The MYCIN system adopts a system structure of “Knowledge Base” and “Inference Engine,” introducing the concept of “confidence” to perform non-deterministic knowledge reasoning, capable of answering user inquiries and providing estimates of the confidence of answers. MYCIN is a relatively comprehensive and well-structured expert system. Its successful development has provided a model and experience for the research and development of many other expert systems.
Significant achievements in application were made by the geological exploration expert system (PROSPECTOR). It was developed by R. O. Duda and others at SRI, and can be used for analyzing geological survey data, exploring types, locations, and distributions of mineral deposits. Development began in 1976, and it was basically completed in 1981, characterized by multi-expert and multi-professional knowledge and experience. Due to PROSPECTOR’s successful application in aluminum ore exploration, it achieved significant economic and social benefits, attracting widespread attention from the business community and igniting a “gold rush” in artificial intelligence, leading to investments in smart industry companies and research on the application of expert system technologies, promoting the transition of artificial intelligence from academic research to technology market development.
In 1977, Feigenbaum proposed the concept of “Knowledge Engineering,” further promoting the development of knowledge-based expert systems and other knowledge engineering systems. For example, in 1979, at the World Conference of the International Joint Conference on Artificial Intelligence, 80% of the papers were related to expert systems.
To improve the design and development efficiency of expert systems and shorten the development cycle, various types of expert system development tools were developed. For example, skeleton-type expert system development tools like EMYCIN and KAS; knowledge acquisition assistance tools like TEIRESIES and SEEK; general knowledge representation languages like LISP, PROLOG, KRL, OPS5; and combinatorial development tools like AGE.
The first expert system in China was developed by the Control Theory Group of the Institute of Automation, Chinese Academy of Sciences, led by Tu Xuyan and Guo Rongjiang, starting in 1977 and successfully developed in 1979, known as the “Traditional Chinese Medicine Gan Youbo Hepatitis Diagnosis and Treatment Program.” This was also the first traditional Chinese medicine expert system internationally.
Since the 1980s, research and development of expert systems have evolved towards large-scale, multidisciplinary, comprehensive knowledge engineering systems. For example, the HPP’—80 system was jointly developed by Stanford University and SRI as a large-scale knowledge engineering system. It includes application expert systems in various disciplines such as chemical analysis, medical diagnosis, molecular biology, structural mechanics, fault diagnosis, and circuit design, as well as knowledge engineering tools like EMYCIN and AGE.
The extensive research, development, and many successful applications of expert systems have, on one hand, promoted the development of knowledge representation, knowledge reasoning, knowledge acquisition, and knowledge utilization methods and technologies. For example, in the book “Principles of Artificial Intelligence” published by N. J. Nilsson in 1980, the design principles of artificial intelligence systems based on knowledge representation and reasoning were discussed using production systems as the basic structure and predicate calculus methods. As a typical knowledge engineering system, expert systems have become one of the most active and productive branches in the field of artificial intelligence. On the other hand, it has also facilitated the popularization of artificial intelligence, shifting from the exploration of general thinking laws to the utilization of specialized knowledge, from academic theoretical research to technological market application development, allowing the public to gain a better understanding of the significance and value of artificial intelligence. However, expert systems and knowledge engineering still face challenges such as the “bottleneck” of knowledge acquisition and uncertain reasoning, which need further development.