Throughout human history, people have used technology to model themselves. There is evidence of this from ancient China, Egypt, and Greece, bearing witness to the universality of this activity. Each new technology has, in its turn, been exploited to build intelligent agents or models of mind. Clockwork, hydraulics, telephone switching systems, holograms, analog computers, and digital computers have all been proposed both as technological metaphors for intelligence and as mechanisms for modeling mind.

Hobbes (1588–1679), who has been described by Haugeland [1985, p. 85] as the “Grandfather of AI,” espoused the position that thinking was symbolic reasoning, like talking out loud or working out an answer with pen and paper. The idea of symbolic reasoning was further developed by Descartes (1596–1650), Pascal (1623–1662), Spinoza (1632–1677), Leibniz (1646–1716), and others who were pioneers in the European philosophy of mind.

The idea of symbolic operations became more concrete with the development of computers. Babbage (1792–1871) designed the first general-purpose computer, the Analytical Engine. Leonardo Torres y Quevedo build a chess playing machine based on similar ideas in 1911 [Randell, 1982]. In the early part of the twentieth century, there was much work done on understanding computation. Several models of computation were proposed, including the Turing machine by Alan Turing (1912–1954), a theoretical machine that writes symbols on an infinitely long tape, and the lambda calculus of Church (1903–1995), which is a mathematical formalism for rewriting formulas. It can be shown that these very different formalisms are equivalent in that any function computable by one is computable by the others. This leads to the Church–Turing thesis:

Any effectively computable function can be carried out on a Turing machine (and so also in the lambda calculus or any of the other equivalent formalisms).

Effectively computable means following well-defined operations. In Turing’s day, “computers” were people who followed well-defined steps; computers as known today did not exist. This thesis says that all computation can be carried out on a Turing machine or one of the other equivalent computational machines. The Church–Turing thesis cannot be proved but it is a hypothesis that has stood the test of time. No one has built a machine that has carried out computation that cannot be computed by a Turing machine. There is no evidence that people can compute functions that are not Turing computable. This provides an argument that computation is more than just a metaphor for intelligence; reasoning is computation and computation can be carried out by a computer.

Some of the first applications of computers were AI programs. Samuel [1959] built a checkers program in 1952 and implemented a program that learns to play checkers in the late 1950s. His program beat the Connecticut state checkers champion in 1961. Wang [1960] implemented a program that proved every logic theorem (nearly 400) in Principia Mathematica [Whitehead and Russell, 1925, 1927]. Newell and Simon [1956] built a program, Logic Theorist, that discovers proofs in propositional logic.

In parallel, there was also much work on neural networks learning inspired by how neurons work. McCulloch and Pitts [1943] showed how a simple thresholding “formal neuron” could be the basis for a Turing-complete machine. Learning for artificial neural networks was first described by Minsky [1952]. One of the early significant works was the perceptron of Rosenblatt [1958]. The work on neural networks became less prominent for a number of years after the 1968 book by Minsky and Papert [1988], which argued that the representations learned were inadequate for intelligent action. Many technical foundations for neural networks were laid in the 1980s and 1990s [Rumelhart et al., 1986; Hochreiter and Schmidhuber, 1997; LeCun et al., 1998a]. Widespread adoption followed the success by Krizhevsky et al. [2012] for ImageNet [Deng et al., 2009], a dataset of over 3 million images labelled with over 5000 categories. Subsequent major advances include the introduction of generative adversarial networks (GANs) [Goodfellow et al., 2014] and transformers [Vaswani et al., 2017]. Neural networks in various forms are now the state of the art for predictive models for large perceptual datasets, including images, video, and speech, as well as some tasks for text. They are also used for generative AI, to generate images, text, code, molecules, and other structured output. See Chapter 8.

Neural networks are one of many machine learning tools used for making predictions from data in modern applications. Other methods have been developed though the years, including decision trees [Breiman et al., 1984; Quinlan, 1993] and logistic regression, introduced by Verhulst in 1832 [Cramer, 2002]. These have diverse applications in many areas of science. Combining these algorithms leads to the state-of-the-art gradient-boosted trees [Friedman, 2001; Chen and Guestrin, 2016], which demonstrates the close interconnections between statistics and machine learning.

While useful, making predictions is not sufficient to determine what an agent should do; an agent also needs to plan. Planning in AI was initially based on deterministic actions. Fikes and Nilsson [1971] used deterministic actions to control a mobile robot. Planning under uncertainty has a long history. Markov decision processes (MDPs), the foundation for much of planning under uncertainty, and dynamic programming, a general way to solve them, were invented by Bellman [1957]. These were extended into decision-theoretic planning in the 1990’s [Boutilier et al., 1999]. Decision-theoretic planning with learning is called reinforcement learning. The first reinforcement learning programs were due to Andreae [1963] and Michie [1963]. Major advances came with the inventions of temporal-difference learning [Sutton, 1988] and Q-learning [Watkins and Dayan, 1992]. Work in reinforcement learning has exploded, including superhuman performance in chess, Go and other games [Silver et al., 2017].

Planning requires representations. The need for representations was recognized early.

A computer program capable of acting intelligently in the world must have a general representation of the world in terms of which its inputs are interpreted. Designing such a program requires commitments about what knowledge is and how it is obtained. …More specifically, we want a computer program that decides what to do by inferring in a formal language that a certain strategy will achieve its assigned goal. This requires formalizing concepts of causality, ability, and knowledge.

– McCarthy and Hayes [1969]

Many of the early representations were ad hoc, such as frames [Minsky, 1975], like the schemas of Kant [1787], Bartlett [1932], and Piaget [1953]. Later representations were based on logic [Kowalski, 1979], with knowledge being defined in logic and efficient inference. This resulted in languages such as Prolog [Kowalski, 1988; Colmerauer and Roussel, 1996].

Probabilities were eschewed in AI, because of the number of parameters required, until the breakthrough of Bayesian networks (belief networks) and graphical models [Pearl, 1988], which exploit conditional independence, and form a basis for modeling causality. Combining first-order logic and probability is the topic of statistical relational AI [De Raedt et al., 2016].

There has been a continual tension between how much knowledge is learned and how much is provided by human experts or is innate to an agent. It has long been recognized that learning is needed, and it is known that learning cannot be achieved with data alone. During the 1970s and 1980s, expert systems came to prominence, where the aim was to capture the knowledge of an expert in some domain so that a computer could carry out expert tasks. DENDRAL [Buchanan and Feigenbaum, 1978], developed from 1965 to 1983 in the field of organic chemistry, proposed plausible structures for new organic compounds. MYCIN [Buchanan and Shortliffe, 1984], developed from 1972 to 1980, diagnosed infectious diseases of the blood, prescribed antimicrobial therapy, and explained its reasoning.

An alternative approach, de-emphasizing explicit knowledge representations, emphasized situated embodied agents [Brooks, 1990; Mackworth, 2009]. The hypothesis is that intelligence emerges, in evolution and individual development, through ongoing interaction and coupling with a real environment.

During the 1960s and 1970s, natural language understanding systems were developed for limited domains. For example, the STUDENT program of Bobrow [1967] could solve high-school algebra tasks expressed in natural language. Winograd’s [1972] SHRDLU system could, using restricted natural language, discuss and carry out tasks in a simulated blocks world. CHAT-80 [Warren and Pereira, 1982] could answer geographical questions placed to it in natural language. Figure 1.3 shows some questions that CHAT-80 answered based on a database of facts about countries, rivers, and so on. These systems could only reason in very limited domains using restricted vocabulary and sentence structure. Interestingly, IBM’s Watson, which beat the world champion in the TV game show Jeopardy! in 2011, used a technique similar to CHAT-80 [Lally et al., 2012] for understanding questions; see Section 15.7.

In applications using language in the wild, such as speech recognition and translation in phones, many technologies are combined, including neural networks; see Chapter 8. Large language models, trained on huge datasets, can be used to predict the next word in a text, enabling predictive spelling and the creation of new text.