Symbolic Artificial Intelligence

In expert system, symbolic artificial intelligence (also called classical expert system or logic-based expert system) [1] [2] is the term for the collection of all approaches in artificial intelligence research study that are based on high-level symbolic (human-readable) representations of issues, reasoning and search. [3] Symbolic AI used tools such as logic shows, production rules, semantic nets and frames, and it developed applications such as knowledge-based systems (in specific, expert systems), symbolic mathematics, automated theorem provers, ontologies, the semantic web, and automated preparation and scheduling systems. The Symbolic AI paradigm resulted in seminal ideas in search, symbolic programs languages, representatives, multi-agent systems, the semantic web, and the strengths and restrictions of formal understanding and thinking systems.

Symbolic AI was the dominant paradigm of AI research study from the mid-1950s until the mid-1990s. [4] Researchers in the 1960s and the 1970s were encouraged that symbolic approaches would ultimately prosper in producing a maker with artificial general intelligence and considered this the ultimate goal of their field. [citation required] An early boom, with early successes such as the Logic Theorist and Samuel’s Checkers Playing Program, caused unrealistic expectations and guarantees and was followed by the first AI Winter as moneying dried up. [5] [6] A 2nd boom (1969-1986) accompanied the increase of specialist systems, their guarantee of capturing business expertise, and an enthusiastic business accept. [7] [8] That boom, and some early successes, e.g., with XCON at DEC, was followed again by later disappointment. [8] Problems with troubles in knowledge acquisition, preserving big understanding bases, and brittleness in dealing with out-of-domain problems developed. Another, second, AI Winter (1988-2011) followed. [9] Subsequently, AI researchers concentrated on attending to underlying issues in dealing with uncertainty and in understanding acquisition. [10] Uncertainty was resolved with official methods such as concealed Markov designs, Bayesian thinking, and analytical relational learning. [11] [12] Symbolic device discovering dealt with the understanding acquisition problem with contributions consisting of Version Space, Valiant’s PAC learning, Quinlan’s ID3 decision-tree knowing, case-based learning, and programming to learn relations. [13]

Neural networks, a subsymbolic technique, had actually been pursued from early days and reemerged highly in 2012. Early examples are Rosenblatt’s perceptron knowing work, the backpropagation work of Rumelhart, Hinton and Williams, [14] and work in convolutional neural networks by LeCun et al. in 1989. [15] However, neural networks were not considered as successful until about 2012: “Until Big Data became prevalent, the basic agreement in the Al community was that the so-called neural-network technique was helpless. Systems simply didn’t work that well, compared to other methods. … A transformation came in 2012, when a number of people, consisting of a team of scientists dealing with Hinton, worked out a way to use the power of GPUs to tremendously increase the power of neural networks.” [16] Over the next numerous years, deep knowing had amazing success in managing vision, speech recognition, speech synthesis, image generation, and machine translation. However, since 2020, as inherent difficulties with predisposition, description, coherence, and effectiveness ended up being more evident with deep knowing approaches; an increasing variety of AI scientists have required combining the very best of both the symbolic and neural network methods [17] [18] and addressing locations that both techniques have difficulty with, such as common-sense thinking. [16]

A short history of symbolic AI to today day follows below. Time periods and titles are drawn from Henry Kautz’s 2020 AAAI Robert S. Engelmore Memorial Lecture [19] and the longer Wikipedia short article on the History of AI, with dates and titles differing a little for increased clearness.

The first AI summer season: irrational spirit, 1948-1966

Success at early efforts in AI took place in 3 primary locations: synthetic neural networks, knowledge representation, and heuristic search, contributing to high expectations. This area sums up Kautz’s reprise of early AI history.

Approaches inspired by human or animal cognition or habits

Cybernetic methods tried to reproduce the feedback loops in between animals and their environments. A robotic turtle, with sensing units, motors for driving and steering, and seven vacuum tubes for control, based on a preprogrammed neural web, was built as early as 1948. This work can be viewed as an early precursor to later work in neural networks, support knowing, and situated robotics. [20]

An essential early symbolic AI program was the Logic theorist, written by Allen Newell, Herbert Simon and Cliff Shaw in 1955-56, as it had the ability to prove 38 primary theorems from Whitehead and Russell’s Principia Mathematica. Newell, Simon, and Shaw later on generalized this work to produce a domain-independent issue solver, GPS (General Problem Solver). GPS solved issues represented with formal operators by means of state-space search utilizing means-ends analysis. [21]

During the 1960s, symbolic methods accomplished excellent success at simulating smart behavior in structured environments such as game-playing, symbolic mathematics, and theorem-proving. AI research study was concentrated in four organizations in the 1960s: Carnegie Mellon University, Stanford, MIT and (later on) University of Edinburgh. Every one developed its own style of research. Earlier approaches based on cybernetics or synthetic neural networks were abandoned or pressed into the background.

Herbert Simon and Allen Newell studied human analytical abilities and attempted to formalize them, and their work laid the structures of the field of expert system, along with cognitive science, operations research and management science. Their research group utilized the outcomes of mental experiments to establish programs that simulated the techniques that people utilized to fix problems. [22] [23] This tradition, focused at Carnegie Mellon University would ultimately culminate in the advancement of the Soar architecture in the center 1980s. [24] [25]

Heuristic search

In addition to the extremely specialized domain-specific type of knowledge that we will see later on utilized in professional systems, early symbolic AI researchers discovered another more general application of understanding. These were called heuristics, guidelines that guide a search in promising directions: “How can non-enumerative search be useful when the underlying problem is tremendously hard? The approach promoted by Simon and Newell is to employ heuristics: fast algorithms that might stop working on some inputs or output suboptimal options.” [26] Another important advance was to find a way to apply these heuristics that ensures a solution will be found, if there is one, not enduring the periodic fallibility of heuristics: “The A * algorithm supplied a general frame for total and optimum heuristically directed search. A * is used as a subroutine within virtually every AI algorithm today but is still no magic bullet; its guarantee of completeness is purchased the cost of worst-case exponential time. [26]

Early work on knowledge representation and thinking

Early work covered both applications of formal thinking stressing first-order logic, together with attempts to deal with sensible reasoning in a less official manner.

Modeling formal reasoning with reasoning: the “neats”

Unlike Simon and Newell, John McCarthy felt that makers did not need to imitate the exact systems of human idea, however might rather look for the essence of abstract thinking and problem-solving with logic, [27] despite whether individuals utilized the exact same algorithms. [a] His lab at Stanford (SAIL) focused on using formal logic to fix a wide array of issues, consisting of knowledge representation, preparation and knowing. [31] Logic was also the focus of the work at the University of Edinburgh and somewhere else in Europe which led to the advancement of the programs language Prolog and the science of reasoning programs. [32] [33]

Modeling implicit common-sense understanding with frames and scripts: the “scruffies”

Researchers at MIT (such as Marvin Minsky and Seymour Papert) [34] [35] [6] found that solving challenging issues in vision and natural language processing needed advertisement hoc solutions-they argued that no basic and basic principle (like reasoning) would capture all the elements of smart habits. Roger Schank explained their “anti-logic” techniques as “shabby” (instead of the “cool” paradigms at CMU and Stanford). [36] [37] Commonsense understanding bases (such as Doug Lenat’s Cyc) are an example of “scruffy” AI, given that they need to be developed by hand, one complicated concept at a time. [38] [39] [40]

The first AI winter season: crushed dreams, 1967-1977

The first AI winter season was a shock:

During the first AI summer, lots of people thought that maker intelligence could be attained in just a few years. The Defense Advance Research Projects Agency (DARPA) released programs to support AI research study to utilize AI to fix problems of nationwide security; in specific, to automate the translation of Russian to English for intelligence operations and to develop self-governing tanks for the battlefield. Researchers had actually started to realize that achieving AI was going to be much harder than was supposed a years previously, but a combination of hubris and disingenuousness led many university and think-tank scientists to accept financing with promises of deliverables that they ought to have known they could not fulfill. By the mid-1960s neither beneficial natural language translation systems nor self-governing tanks had actually been produced, and a dramatic backlash set in. New DARPA management canceled existing AI funding programs.

Outside of the United States, the most fertile ground for AI research was the UK. The AI winter in the United Kingdom was stimulated on not so much by disappointed military leaders as by rival academics who viewed AI scientists as charlatans and a drain on research study financing. A teacher of applied mathematics, Sir James Lighthill, was commissioned by Parliament to assess the state of AI research in the country. The report stated that all of the problems being dealt with in AI would be better dealt with by researchers from other disciplines-such as applied mathematics. The report likewise declared that AI successes on toy problems might never ever scale to real-world applications due to combinatorial explosion. [41]

The second AI summer season: knowledge is power, 1978-1987

Knowledge-based systems

As limitations with weak, domain-independent approaches ended up being increasingly more obvious, [42] scientists from all 3 customs began to build understanding into AI applications. [43] [7] The understanding transformation was driven by the awareness that understanding underlies high-performance, domain-specific AI applications.

Edward Feigenbaum said:

– “In the understanding lies the power.” [44]
to explain that high efficiency in a specific domain needs both general and extremely domain-specific knowledge. Ed Feigenbaum and Doug Lenat called this The Knowledge Principle:

( 1) The Knowledge Principle: if a program is to carry out an intricate task well, it must understand a good deal about the world in which it operates.
( 2) A possible extension of that concept, called the Breadth Hypothesis: there are 2 additional abilities needed for intelligent habits in unforeseen situations: falling back on progressively general understanding, and analogizing to particular however remote knowledge. [45]

Success with expert systems

This “understanding transformation” caused the development and implementation of expert systems (introduced by Edward Feigenbaum), the very first commercially effective type of AI software. [46] [47] [48]

Key professional systems were:

DENDRAL, which found the structure of organic particles from their chemical formula and mass spectrometer readings.
MYCIN, which detected bacteremia – and recommended additional laboratory tests, when essential – by interpreting lab results, patient history, and doctor observations. “With about 450 rules, MYCIN was able to carry out as well as some specialists, and considerably better than junior doctors.” [49] INTERNIST and CADUCEUS which dealt with internal medicine medical diagnosis. Internist attempted to capture the competence of the chairman of internal medication at the University of Pittsburgh School of Medicine while CADUCEUS could ultimately identify up to 1000 different diseases.
– GUIDON, which revealed how a knowledge base developed for expert problem fixing might be repurposed for mentor. [50] XCON, to set up VAX computer systems, a then laborious process that might take up to 90 days. XCON reduced the time to about 90 minutes. [9]
DENDRAL is considered the first expert system that relied on knowledge-intensive analytical. It is explained below, by Ed Feigenbaum, from a Communications of the ACM interview, Interview with Ed Feigenbaum:

Among the people at Stanford interested in computer-based designs of mind was Joshua Lederberg, the 1958 Nobel Prize winner in genetics. When I told him I desired an induction “sandbox”, he said, “I have just the one for you.” His laboratory was doing mass spectrometry of amino acids. The question was: how do you go from taking a look at the spectrum of an amino acid to the chemical structure of the amino acid? That’s how we started the DENDRAL Project: I was proficient at heuristic search approaches, and he had an algorithm that was good at generating the chemical issue area.

We did not have a grandiose vision. We worked bottom up. Our chemist was Carl Djerassi, inventor of the chemical behind the contraceptive pill, and also among the world’s most appreciated mass spectrometrists. Carl and his postdocs were first-rate specialists in mass spectrometry. We began to add to their knowledge, developing knowledge of engineering as we went along. These experiments totaled up to titrating DENDRAL a growing number of understanding. The more you did that, the smarter the program ended up being. We had extremely great outcomes.

The generalization was: in the knowledge lies the power. That was the huge idea. In my career that is the huge, “Ah ha!,” and it wasn’t the method AI was being done previously. Sounds easy, but it’s probably AI’s most powerful generalization. [51]

The other specialist systems discussed above came after DENDRAL. MYCIN exhibits the classic expert system architecture of a knowledge-base of rules combined to a symbolic reasoning mechanism, consisting of making use of certainty elements to deal with uncertainty. GUIDON shows how an explicit understanding base can be repurposed for a 2nd application, tutoring, and is an example of an intelligent tutoring system, a specific type of knowledge-based application. Clancey showed that it was not sufficient just to use MYCIN’s guidelines for instruction, but that he likewise required to add rules for dialogue management and trainee modeling. [50] XCON is significant due to the fact that of the countless dollars it conserved DEC, which triggered the expert system boom where most all major corporations in the US had professional systems groups, to record business proficiency, maintain it, and automate it:

By 1988, DEC’s AI group had 40 specialist systems released, with more on the method. DuPont had 100 in usage and 500 in advancement. Nearly every significant U.S. corporation had its own Al group and was either using or examining expert systems. [49]

Chess expert knowledge was encoded in Deep Blue. In 1996, this permitted IBM’s Deep Blue, with the aid of symbolic AI, to win in a game of chess against the world champion at that time, Garry Kasparov. [52]

Architecture of knowledge-based and professional systems

A key element of the system architecture for all specialist systems is the understanding base, which shops truths and rules for analytical. [53] The most basic technique for an expert system understanding base is merely a collection or network of production guidelines. Production guidelines connect signs in a relationship similar to an If-Then declaration. The professional system processes the rules to make reductions and to identify what extra information it needs, i.e. what concerns to ask, utilizing human-readable symbols. For instance, OPS5, CLIPS and their followers Jess and Drools run in this style.

Expert systems can operate in either a forward chaining – from proof to conclusions – or backward chaining – from objectives to required information and prerequisites – way. Advanced knowledge-based systems, such as Soar can also perform meta-level reasoning, that is reasoning about their own thinking in regards to deciding how to resolve problems and keeping an eye on the success of problem-solving strategies.

Blackboard systems are a second sort of knowledge-based or skilled system architecture. They design a neighborhood of specialists incrementally contributing, where they can, to solve a problem. The problem is represented in several levels of abstraction or alternate views. The specialists (understanding sources) volunteer their services whenever they recognize they can contribute. Potential analytical actions are represented on an agenda that is upgraded as the problem situation changes. A controller decides how helpful each contribution is, and who should make the next problem-solving action. One example, the BB1 blackboard architecture [54] was originally influenced by studies of how humans prepare to perform multiple jobs in a trip. [55] An innovation of BB1 was to use the same blackboard design to resolving its control problem, i.e., its controller performed meta-level reasoning with understanding sources that kept track of how well a strategy or the analytical was continuing and might change from one technique to another as conditions – such as objectives or times – altered. BB1 has actually been used in multiple domains: construction site preparation, intelligent tutoring systems, and real-time client monitoring.

The second AI winter, 1988-1993

At the height of the AI boom, companies such as Symbolics, LMI, and Texas Instruments were offering LISP makers specifically targeted to accelerate the development of AI applications and research study. In addition, numerous artificial intelligence business, such as Teknowledge and Inference Corporation, were offering professional system shells, training, and speaking with to corporations.

Unfortunately, the AI boom did not last and Kautz best explains the second AI winter season that followed:

Many reasons can be offered for the arrival of the 2nd AI winter season. The hardware companies stopped working when much more cost-effective general Unix workstations from Sun together with good compilers for LISP and Prolog came onto the marketplace. Many industrial implementations of specialist systems were discontinued when they proved too pricey to maintain. Medical specialist systems never ever captured on for several factors: the difficulty in keeping them up to date; the difficulty for doctor to find out how to utilize an overwelming variety of different specialist systems for various medical conditions; and maybe most crucially, the unwillingness of doctors to trust a computer-made medical diagnosis over their gut instinct, even for specific domains where the expert systems might exceed an average medical professional. Equity capital money deserted AI almost overnight. The world AI conference IJCAI hosted a huge and luxurious trade program and thousands of nonacademic guests in 1987 in Vancouver; the main AI conference the list below year, AAAI 1988 in St. Paul, was a little and strictly scholastic affair. [9]

Adding in more extensive foundations, 1993-2011

Uncertain reasoning

Both analytical methods and extensions to reasoning were attempted.

One statistical technique, concealed Markov models, had already been popularized in the 1980s for speech recognition work. [11] Subsequently, in 1988, Judea Pearl promoted the usage of Bayesian Networks as a sound however efficient method of dealing with unsure reasoning with his publication of the book Probabilistic Reasoning in Intelligent Systems: Networks of Plausible Inference. [56] and Bayesian techniques were used successfully in professional systems. [57] Even later, in the 1990s, analytical relational learning, a technique that combines likelihood with sensible formulas, allowed probability to be integrated with first-order logic, e.g., with either Markov Logic Networks or Probabilistic Soft Logic.

Other, non-probabilistic extensions to first-order logic to assistance were also attempted. For instance, non-monotonic thinking could be used with reality maintenance systems. A fact upkeep system tracked presumptions and justifications for all reasonings. It allowed inferences to be withdrawn when presumptions were discovered to be incorrect or a contradiction was obtained. Explanations could be offered a reasoning by describing which rules were applied to create it and then continuing through underlying inferences and rules all the method back to root assumptions. [58] Lofti Zadeh had introduced a different type of extension to handle the representation of vagueness. For instance, in deciding how “heavy” or “high” a male is, there is frequently no clear “yes” or “no” response, and a predicate for heavy or tall would rather return values in between 0 and 1. Those worths represented to what degree the predicates were true. His fuzzy logic even more offered a means for propagating combinations of these values through rational solutions. [59]

Machine learning

Symbolic maker discovering approaches were examined to attend to the knowledge acquisition traffic jam. One of the earliest is Meta-DENDRAL. Meta-DENDRAL used a generate-and-test strategy to produce plausible rule hypotheses to check against spectra. Domain and job knowledge minimized the variety of candidates tested to a workable size. Feigenbaum explained Meta-DENDRAL as

… the conclusion of my dream of the early to mid-1960s having to do with theory development. The conception was that you had a problem solver like DENDRAL that took some inputs and produced an output. In doing so, it used layers of understanding to guide and prune the search. That understanding acted since we interviewed individuals. But how did individuals get the knowledge? By looking at thousands of spectra. So we desired a program that would take a look at thousands of spectra and presume the knowledge of mass spectrometry that DENDRAL could use to fix individual hypothesis formation issues. We did it. We were even able to publish new knowledge of mass spectrometry in the Journal of the American Chemical Society, providing credit only in a footnote that a program, Meta-DENDRAL, really did it. We had the ability to do something that had actually been a dream: to have a computer system program created a new and publishable piece of science. [51]

In contrast to the knowledge-intensive technique of Meta-DENDRAL, Ross Quinlan developed a domain-independent technique to statistical classification, decision tree learning, starting initially with ID3 [60] and after that later on extending its capabilities to C4.5. [61] The decision trees developed are glass box, interpretable classifiers, with human-interpretable classification guidelines.

Advances were made in understanding artificial intelligence theory, too. Tom Mitchell presented variation area learning which describes learning as an explore a space of hypotheses, with upper, more general, and lower, more particular, limits incorporating all practical hypotheses consistent with the examples seen so far. [62] More officially, Valiant introduced Probably Approximately Correct Learning (PAC Learning), a framework for the mathematical analysis of maker learning. [63]

Symbolic device discovering encompassed more than learning by example. E.g., John Anderson provided a cognitive design of human learning where ability practice leads to a collection of rules from a declarative format to a procedural format with his ACT-R cognitive architecture. For example, a student may discover to use “Supplementary angles are 2 angles whose procedures sum 180 degrees” as several different procedural rules. E.g., one guideline might say that if X and Y are extra and you know X, then Y will be 180 – X. He called his approach “understanding collection”. ACT-R has been used effectively to design elements of human cognition, such as discovering and retention. ACT-R is also utilized in intelligent tutoring systems, called cognitive tutors, to effectively teach geometry, computer system programming, and algebra to school children. [64]

Inductive reasoning shows was another method to finding out that permitted reasoning programs to be synthesized from input-output examples. E.g., Ehud Shapiro’s MIS (Model Inference System) might manufacture Prolog programs from examples. [65] John R. Koza used genetic algorithms to program synthesis to create hereditary programming, which he utilized to manufacture LISP programs. Finally, Zohar Manna and Richard Waldinger supplied a more basic method to program synthesis that synthesizes a functional program in the course of proving its requirements to be proper. [66]

As an option to reasoning, Roger Schank introduced case-based reasoning (CBR). The CBR technique laid out in his book, Dynamic Memory, [67] focuses first on keeping in mind essential problem-solving cases for future use and generalizing them where suitable. When confronted with a brand-new problem, CBR retrieves the most similar previous case and adjusts it to the specifics of the current issue. [68] Another option to logic, genetic algorithms and genetic programs are based upon an evolutionary model of learning, where sets of guidelines are encoded into populations, the guidelines govern the behavior of individuals, and choice of the fittest prunes out sets of inappropriate guidelines over many generations. [69]

Symbolic artificial intelligence was applied to discovering concepts, guidelines, heuristics, and problem-solving. Approaches, besides those above, include:

1. Learning from direction or advice-i.e., taking human instruction, posed as advice, and figuring out how to operationalize it in specific situations. For example, in a video game of Hearts, finding out exactly how to play a hand to “avoid taking points.” [70] 2. Learning from exemplars-improving performance by accepting subject-matter professional (SME) feedback throughout training. When problem-solving stops working, querying the professional to either learn a new prototype for analytical or to find out a brand-new description regarding precisely why one prototype is more relevant than another. For instance, the program Protos found out to identify ringing in the ears cases by connecting with an audiologist. [71] 3. Learning by analogy-constructing issue solutions based on comparable problems seen in the past, and then modifying their solutions to fit a new scenario or domain. [72] [73] 4. Apprentice learning systems-learning novel solutions to issues by observing human problem-solving. Domain knowledge describes why novel solutions are correct and how the option can be generalized. LEAP found out how to develop VLSI circuits by observing human designers. [74] 5. Learning by discovery-i.e., developing jobs to perform experiments and then finding out from the outcomes. Doug Lenat’s Eurisko, for example, found out heuristics to beat human players at the Traveller role-playing video game for 2 years in a row. [75] 6. Learning macro-operators-i.e., looking for helpful macro-operators to be gained from series of fundamental analytical actions. Good macro-operators simplify analytical by enabling issues to be fixed at a more abstract level. [76]
Deep learning and neuro-symbolic AI 2011-now

With the increase of deep learning, the symbolic AI method has been compared to deep knowing as complementary “… with parallels having been drawn lot of times by AI scientists in between Kahneman’s research on human reasoning and decision making – shown in his book Thinking, Fast and Slow – and the so-called “AI systems 1 and 2″, which would in principle be designed by deep learning and symbolic reasoning, respectively.” In this view, symbolic reasoning is more apt for deliberative thinking, planning, and description while deep learning is more apt for fast pattern recognition in perceptual applications with loud information. [17] [18]

Neuro-symbolic AI: integrating neural and symbolic techniques

Neuro-symbolic AI efforts to integrate neural and symbolic architectures in a manner that addresses strengths and weaknesses of each, in a complementary style, in order to support robust AI efficient in reasoning, finding out, and cognitive modeling. As argued by Valiant [77] and lots of others, [78] the efficient construction of rich computational cognitive models demands the combination of sound symbolic reasoning and efficient (machine) learning models. Gary Marcus, likewise, argues that: “We can not build abundant cognitive designs in a sufficient, automated method without the triumvirate of hybrid architecture, rich prior knowledge, and sophisticated methods for thinking.”, [79] and in particular: “To construct a robust, knowledge-driven technique to AI we must have the machinery of symbol-manipulation in our toolkit. Excessive of helpful understanding is abstract to make do without tools that represent and control abstraction, and to date, the only machinery that we know of that can manipulate such abstract knowledge reliably is the apparatus of sign control. ” [80]

Henry Kautz, [19] Francesca Rossi, [81] and Bart Selman [82] have also argued for a synthesis. Their arguments are based upon a need to deal with the two type of thinking talked about in Daniel Kahneman’s book, Thinking, Fast and Slow. Kahneman explains human thinking as having 2 components, System 1 and System 2. System 1 is fast, automated, intuitive and unconscious. System 2 is slower, step-by-step, and specific. System 1 is the kind utilized for pattern acknowledgment while System 2 is far much better fit for planning, deduction, and deliberative thinking. In this view, deep knowing best models the very first type of thinking while symbolic thinking finest models the second kind and both are required.

Garcez and Lamb describe research study in this location as being ongoing for at least the previous twenty years, [83] dating from their 2002 book on neurosymbolic learning systems. [84] A series of workshops on neuro-symbolic thinking has actually been held every year because 2005, see http://www.neural-symbolic.org/ for details.

In their 2015 paper, Neural-Symbolic Learning and Reasoning: Contributions and Challenges, Garcez et al. argue that:

The integration of the symbolic and connectionist paradigms of AI has actually been pursued by a fairly small research study neighborhood over the last 20 years and has yielded a number of substantial results. Over the last years, neural symbolic systems have actually been revealed efficient in getting rid of the so-called propositional fixation of neural networks, as McCarthy (1988) put it in response to Smolensky (1988 ); see likewise (Hinton, 1990). Neural networks were shown capable of representing modal and temporal logics (d’Avila Garcez and Lamb, 2006) and fragments of first-order logic (Bader, Hitzler, Hölldobler, 2008; d’Avila Garcez, Lamb, Gabbay, 2009). Further, neural-symbolic systems have actually been used to a number of issues in the locations of bioinformatics, control engineering, software verification and adjustment, visual intelligence, ontology learning, and computer games. [78]

Approaches for combination are differed. Henry Kautz’s taxonomy of neuro-symbolic architectures, along with some examples, follows:

– Symbolic Neural symbolic-is the current technique of many neural designs in natural language processing, where words or subword tokens are both the ultimate input and output of large language models. Examples consist of BERT, RoBERTa, and GPT-3.
– Symbolic [Neural] -is exemplified by AlphaGo, where symbolic strategies are utilized to call neural techniques. In this case the symbolic approach is Monte Carlo tree search and the neural methods find out how to assess video game positions.
– Neural|Symbolic-uses a neural architecture to translate affective information as symbols and relationships that are then reasoned about symbolically.
– Neural: Symbolic → Neural-relies on symbolic thinking to create or label training information that is subsequently found out by a deep knowing model, e.g., to train a neural design for symbolic calculation by using a Macsyma-like symbolic mathematics system to produce or label examples.
– Neural _ Symbolic -uses a neural web that is produced from symbolic guidelines. An example is the Neural Theorem Prover, [85] which constructs a neural network from an AND-OR proof tree produced from understanding base guidelines and terms. Logic Tensor Networks [86] also fall into this category.
– Neural [Symbolic] -enables a neural model to straight call a symbolic thinking engine, e.g., to carry out an action or examine a state.

Many essential research questions remain, such as:

– What is the finest way to integrate neural and symbolic architectures? [87]- How should symbolic structures be represented within neural networks and drawn out from them?
– How should sensible knowledge be learned and reasoned about?
– How can abstract knowledge that is difficult to encode realistically be managed?

Techniques and contributions

This area offers an introduction of techniques and contributions in a total context leading to lots of other, more in-depth articles in Wikipedia. Sections on Artificial Intelligence and Uncertain Reasoning are covered previously in the history area.

AI shows languages

The key AI programming language in the US throughout the last symbolic AI boom period was LISP. LISP is the 2nd earliest programming language after FORTRAN and was developed in 1958 by John McCarthy. LISP offered the first read-eval-print loop to support fast program advancement. Compiled functions could be easily mixed with interpreted functions. Program tracing, stepping, and breakpoints were also supplied, in addition to the capability to alter values or functions and continue from breakpoints or mistakes. It had the very first self-hosting compiler, suggesting that the compiler itself was initially composed in LISP and after that ran interpretively to compile the compiler code.

Other key innovations pioneered by LISP that have spread out to other programs languages include:

Garbage collection
Dynamic typing
Higher-order functions
Recursion
Conditionals

Programs were themselves data structures that other programs might run on, allowing the easy definition of higher-level languages.

In contrast to the US, in Europe the key AI programming language throughout that exact same duration was Prolog. Prolog supplied a built-in shop of facts and clauses that might be queried by a read-eval-print loop. The shop could function as a knowledge base and the clauses could function as guidelines or a limited type of logic. As a subset of first-order logic Prolog was based upon Horn stipulations with a closed-world assumption-any realities not understood were thought about false-and a distinct name presumption for primitive terms-e.g., the identifier barack_obama was thought about to refer to exactly one things. Backtracking and marriage are built-in to Prolog.

Alain Colmerauer and Philippe Roussel are credited as the developers of Prolog. Prolog is a form of logic programs, which was created by Robert Kowalski. Its history was likewise influenced by Carl Hewitt’s PLANNER, an assertional database with pattern-directed invocation of approaches. For more detail see the area on the origins of Prolog in the PLANNER post.

Prolog is also a kind of declarative shows. The logic clauses that describe programs are straight translated to run the programs specified. No specific series of actions is required, as is the case with essential programs languages.

Japan championed Prolog for its Fifth Generation Project, planning to build unique hardware for high efficiency. Similarly, LISP machines were built to run LISP, however as the 2nd AI boom turned to bust these business could not compete with brand-new workstations that might now run LISP or Prolog natively at equivalent speeds. See the history area for more detail.

Smalltalk was another influential AI programs language. For example, it presented metaclasses and, along with Flavors and CommonLoops, affected the Common Lisp Object System, or (CLOS), that is now part of Common Lisp, the existing basic Lisp dialect. CLOS is a Lisp-based object-oriented system that allows several inheritance, in addition to incremental extensions to both classes and metaclasses, therefore supplying a run-time meta-object procedure. [88]

For other AI programs languages see this list of programs languages for synthetic intelligence. Currently, Python, a multi-paradigm programming language, is the most popular shows language, partially due to its comprehensive bundle library that supports data science, natural language processing, and deep learning. Python includes a read-eval-print loop, practical elements such as higher-order functions, and object-oriented programming that includes metaclasses.

Search

Search develops in lots of sort of problem fixing, consisting of preparation, restraint satisfaction, and playing video games such as checkers, chess, and go. The best understood AI-search tree search algorithms are breadth-first search, depth-first search, A *, and Monte Carlo Search. Key search algorithms for Boolean satisfiability are WalkSAT, conflict-driven clause learning, and the DPLL algorithm. For adversarial search when playing video games, alpha-beta pruning, branch and bound, and minimax were early contributions.

Knowledge representation and reasoning

Multiple different techniques to represent understanding and then reason with those representations have actually been investigated. Below is a fast summary of approaches to understanding representation and automated reasoning.

Knowledge representation

Semantic networks, conceptual charts, frames, and reasoning are all methods to modeling knowledge such as domain understanding, analytical understanding, and the semantic meaning of language. Ontologies design crucial concepts and their relationships in a domain. Example ontologies are YAGO, WordNet, and DOLCE. DOLCE is an example of an upper ontology that can be used for any domain while WordNet is a lexical resource that can also be deemed an ontology. YAGO integrates WordNet as part of its ontology, to line up truths extracted from Wikipedia with WordNet synsets. The Disease Ontology is an example of a medical ontology presently being utilized.

Description logic is a logic for automated classification of ontologies and for finding inconsistent category data. OWL is a language used to represent ontologies with description reasoning. Protégé is an ontology editor that can check out in OWL ontologies and after that examine consistency with deductive classifiers such as such as HermiT. [89]

First-order logic is more basic than description logic. The automated theorem provers discussed listed below can prove theorems in first-order reasoning. Horn provision reasoning is more restricted than first-order logic and is utilized in reasoning programming languages such as Prolog. Extensions to first-order reasoning consist of temporal reasoning, to handle time; epistemic logic, to reason about representative understanding; modal logic, to deal with possibility and requirement; and probabilistic reasonings to manage reasoning and possibility together.

Automatic theorem showing

Examples of automated theorem provers for first-order reasoning are:

Prover9.
ACL2.
Vampire.

Prover9 can be utilized in conjunction with the Mace4 model checker. ACL2 is a theorem prover that can handle evidence by induction and is a descendant of the Boyer-Moore Theorem Prover, likewise referred to as Nqthm.

Reasoning in knowledge-based systems

Knowledge-based systems have a specific knowledge base, generally of rules, to boost reusability across domains by separating procedural code and domain understanding. A different reasoning engine processes rules and includes, deletes, or modifies a knowledge shop.

Forward chaining inference engines are the most common, and are seen in CLIPS and OPS5. Backward chaining occurs in Prolog, where a more restricted rational representation is utilized, Horn Clauses. Pattern-matching, specifically unification, is utilized in Prolog.

A more flexible type of analytical occurs when thinking about what to do next occurs, rather than simply selecting one of the available actions. This sort of meta-level reasoning is utilized in Soar and in the BB1 blackboard architecture.

Cognitive architectures such as ACT-R might have extra capabilities, such as the capability to compile often used knowledge into higher-level pieces.

Commonsense thinking

Marvin Minsky initially proposed frames as a method of interpreting common visual situations, such as a workplace, and Roger Schank extended this concept to scripts for common routines, such as dining out. Cyc has tried to record useful sensible knowledge and has “micro-theories” to handle particular kinds of domain-specific reasoning.

Qualitative simulation, such as Benjamin Kuipers’s QSIM, [90] approximates human thinking about naive physics, such as what takes place when we heat a liquid in a pot on the stove. We anticipate it to heat and perhaps boil over, although we may not understand its temperature, its boiling point, or other details, such as climatic pressure.

Similarly, Allen’s temporal period algebra is a simplification of reasoning about time and Region Connection Calculus is a simplification of thinking about spatial relationships. Both can be solved with restraint solvers.

Constraints and constraint-based thinking

Constraint solvers perform a more minimal type of inference than first-order logic. They can simplify sets of spatiotemporal constraints, such as those for RCC or Temporal Algebra, in addition to solving other type of puzzle issues, such as Wordle, Sudoku, cryptarithmetic problems, and so on. Constraint logic programs can be utilized to resolve scheduling issues, for instance with constraint dealing with rules (CHR).

Automated planning

The General Problem Solver (GPS) cast planning as analytical utilized means-ends analysis to develop strategies. STRIPS took a various technique, viewing planning as theorem proving. Graphplan takes a least-commitment method to preparation, rather than sequentially selecting actions from a preliminary state, working forwards, or a goal state if working in reverse. Satplan is a technique to preparing where a preparation problem is decreased to a Boolean satisfiability problem.

Natural language processing

Natural language processing focuses on treating language as data to perform tasks such as determining topics without always understanding the intended significance. Natural language understanding, in contrast, constructs a significance representation and uses that for further processing, such as answering questions.

Parsing, tokenizing, spelling correction, part-of-speech tagging, noun and verb expression chunking are all elements of natural language processing long managed by symbolic AI, however because enhanced by deep learning approaches. In symbolic AI, discourse representation theory and first-order logic have been utilized to represent sentence significances. Latent semantic analysis (LSA) and specific semantic analysis also supplied vector representations of documents. In the latter case, vector elements are interpretable as principles named by Wikipedia short articles.

New deep knowing methods based on Transformer designs have actually now eclipsed these earlier symbolic AI techniques and achieved state-of-the-art performance in natural language processing. However, Transformer models are opaque and do not yet produce human-interpretable semantic representations for sentences and files. Instead, they produce task-specific vectors where the meaning of the vector elements is opaque.

Agents and multi-agent systems

Agents are self-governing systems embedded in an environment they perceive and act on in some sense. Russell and Norvig’s standard textbook on expert system is arranged to reflect agent architectures of increasing elegance. [91] The sophistication of agents varies from easy reactive agents, to those with a model of the world and automated preparation abilities, perhaps a BDI agent, i.e., one with beliefs, desires, and intents – or alternatively a support finding out model learned gradually to pick actions – approximately a combination of alternative architectures, such as a neuro-symbolic architecture [87] that includes deep learning for understanding. [92]

On the other hand, a multi-agent system consists of several agents that interact amongst themselves with some inter-agent communication language such as Knowledge Query and Manipulation Language (KQML). The agents need not all have the very same internal architecture. Advantages of multi-agent systems include the ability to divide work among the representatives and to increase fault tolerance when agents are lost. Research problems consist of how representatives reach agreement, distributed issue solving, multi-agent learning, multi-agent planning, and distributed restraint optimization.

Controversies arose from early in symbolic AI, both within the field-e.g., in between logicists (the pro-logic “neats”) and non-logicists (the anti-logic “scruffies”)- and in between those who welcomed AI however declined symbolic approaches-primarily connectionists-and those outside the field. Critiques from outside of the field were mostly from philosophers, on intellectual premises, however also from funding agencies, particularly throughout the two AI winter seasons.

The Frame Problem: understanding representation obstacles for first-order reasoning

Limitations were discovered in utilizing basic first-order logic to factor about dynamic domains. Problems were found both with regards to identifying the preconditions for an action to prosper and in providing axioms for what did not change after an action was performed.

McCarthy and Hayes introduced the Frame Problem in 1969 in the paper, “Some Philosophical Problems from the Standpoint of Artificial Intelligence.” [93] A basic example takes place in “proving that one person might enter conversation with another”, as an axiom asserting “if a person has a telephone he still has it after searching for a number in the telephone directory” would be required for the deduction to succeed. Similar axioms would be required for other domain actions to specify what did not change.

A comparable issue, called the Qualification Problem, takes place in attempting to specify the preconditions for an action to prosper. A boundless number of pathological conditions can be thought of, e.g., a banana in a tailpipe might avoid a cars and truck from running properly.

McCarthy’s method to repair the frame problem was circumscription, a sort of non-monotonic reasoning where reductions could be made from actions that need just define what would alter while not having to clearly specify whatever that would not change. Other non-monotonic logics provided truth maintenance systems that modified beliefs causing contradictions.

Other ways of dealing with more open-ended domains included probabilistic thinking systems and artificial intelligence to discover new ideas and guidelines. McCarthy’s Advice Taker can be considered as a motivation here, as it could include brand-new knowledge supplied by a human in the kind of assertions or guidelines. For instance, experimental symbolic machine learning systems checked out the ability to take top-level natural language recommendations and to interpret it into domain-specific actionable guidelines.

Similar to the problems in managing dynamic domains, sensible thinking is also tough to record in official reasoning. Examples of sensible thinking consist of implicit thinking about how individuals think or general understanding of daily occasions, items, and living animals. This kind of knowledge is considered approved and not seen as noteworthy. Common-sense thinking is an open location of research and challenging both for symbolic systems (e.g., Cyc has actually attempted to record essential parts of this knowledge over more than a years) and neural systems (e.g., self-driving vehicles that do not understand not to drive into cones or not to strike pedestrians walking a bicycle).

McCarthy viewed his Advice Taker as having sensible, but his definition of common-sense was various than the one above. [94] He specified a program as having good sense “if it automatically deduces for itself an adequately large class of immediate repercussions of anything it is told and what it already understands. “

Connectionist AI: philosophical difficulties and sociological conflicts

Connectionist approaches consist of earlier work on neural networks, [95] such as perceptrons; work in the mid to late 80s, such as Danny Hillis’s Connection Machine and Yann LeCun’s advances in convolutional neural networks; to today’s more advanced methods, such as Transformers, GANs, and other work in deep learning.

Three philosophical positions [96] have been detailed among connectionists:

1. Implementationism-where connectionist architectures implement the abilities for symbolic processing,
2. Radical connectionism-where symbolic processing is declined absolutely, and connectionist architectures underlie intelligence and are totally adequate to discuss it,
3. Moderate connectionism-where symbolic processing and connectionist architectures are viewed as complementary and both are required for intelligence

Olazaran, in his sociological history of the controversies within the neural network community, described the moderate connectionism view as basically compatible with existing research in neuro-symbolic hybrids:

The third and last position I wish to examine here is what I call the moderate connectionist view, a more eclectic view of the current argument in between connectionism and symbolic AI. One of the researchers who has actually elaborated this position most clearly is Andy Clark, a thinker from the School of Cognitive and Computing Sciences of the University of Sussex (Brighton, England). Clark defended hybrid (partly symbolic, partly connectionist) systems. He claimed that (a minimum of) 2 kinds of theories are needed in order to study and design cognition. On the one hand, for some information-processing jobs (such as pattern recognition) connectionism has benefits over symbolic designs. But on the other hand, for other cognitive processes (such as serial, deductive reasoning, and generative sign manipulation procedures) the symbolic paradigm provides appropriate designs, and not just “approximations” (contrary to what extreme connectionists would claim). [97]

Gary Marcus has actually claimed that the animus in the deep learning community against symbolic methods now might be more sociological than philosophical:

To believe that we can just desert symbol-manipulation is to suspend disbelief.

And yet, for the a lot of part, that’s how most current AI proceeds. Hinton and lots of others have actually attempted difficult to banish signs completely. The deep knowing hope-seemingly grounded not so much in science, however in a sort of historical grudge-is that intelligent behavior will emerge simply from the confluence of massive data and deep knowing. Where classical computer systems and software application fix jobs by specifying sets of symbol-manipulating rules devoted to particular jobs, such as editing a line in a word processor or performing a calculation in a spreadsheet, neural networks usually attempt to fix jobs by statistical approximation and discovering from examples.

According to Marcus, Geoffrey Hinton and his colleagues have been emphatically “anti-symbolic”:

When deep learning reemerged in 2012, it was with a sort of take-no-prisoners mindset that has characterized the majority of the last years. By 2015, his hostility toward all things symbols had actually totally crystallized. He provided a talk at an AI workshop at Stanford comparing symbols to aether, one of science’s greatest errors.

…

Since then, his anti-symbolic campaign has actually only increased in strength. In 2016, Yann LeCun, Bengio, and Hinton composed a manifesto for deep knowing in among science’s crucial journals, Nature. It closed with a direct attack on sign manipulation, calling not for reconciliation however for straight-out replacement. Later, Hinton informed an event of European Union leaders that investing any additional money in symbol-manipulating approaches was “a substantial error,” comparing it to buying internal combustion engines in the period of electric cars. [98]

Part of these conflicts might be because of unclear terminology:

Turing award winner Judea Pearl offers a review of maker knowing which, regrettably, conflates the terms artificial intelligence and deep learning. Similarly, when Geoffrey Hinton refers to symbolic AI, the undertone of the term tends to be that of professional systems dispossessed of any capability to discover. Making use of the terminology needs information. Artificial intelligence is not confined to association guideline mining, c.f. the body of work on symbolic ML and relational knowing (the differences to deep learning being the option of representation, localist logical instead of distributed, and the non-use of gradient-based knowing algorithms). Equally, symbolic AI is not practically production rules written by hand. An appropriate meaning of AI concerns understanding representation and reasoning, self-governing multi-agent systems, preparation and argumentation, along with learning. [99]

Situated robotics: the world as a design

Another review of symbolic AI is the embodied cognition method:

The embodied cognition approach declares that it makes no sense to consider the brain separately: cognition takes location within a body, which is embedded in an environment. We require to study the system as a whole; the brain’s operating exploits regularities in its environment, consisting of the rest of its body. Under the embodied cognition technique, robotics, vision, and other sensors become central, not peripheral. [100]

Rodney Brooks created behavior-based robotics, one method to embodied cognition. Nouvelle AI, another name for this approach, is deemed an alternative to both symbolic AI and connectionist AI. His technique rejected representations, either symbolic or distributed, as not only unnecessary, but as detrimental. Instead, he produced the subsumption architecture, a layered architecture for embodied agents. Each layer attains a different function and should work in the real life. For example, the very first robot he describes in Intelligence Without Representation, has 3 layers. The bottom layer translates sonar sensors to prevent things. The middle layer triggers the robot to roam around when there are no obstacles. The top layer causes the robot to go to more far-off locations for further exploration. Each layer can momentarily inhibit or reduce a lower-level layer. He criticized AI scientists for specifying AI issues for their systems, when: “There is no tidy department between perception (abstraction) and reasoning in the real world.” [101] He called his robots “Creatures” and each layer was “made up of a fixed-topology network of simple limited state devices.” [102] In the Nouvelle AI approach, “First, it is critically important to check the Creatures we integrate in the real world; i.e., in the exact same world that we humans live in. It is devastating to fall under the temptation of evaluating them in a simplified world initially, even with the best objectives of later moving activity to an unsimplified world.” [103] His emphasis on real-world testing was in contrast to “Early operate in AI focused on video games, geometrical issues, symbolic algebra, theorem proving, and other formal systems” [104] and the usage of the blocks world in symbolic AI systems such as SHRDLU.

Current views

Each approach-symbolic, connectionist, and behavior-based-has advantages, but has been slammed by the other techniques. Symbolic AI has been slammed as disembodied, responsible to the credentials issue, and poor in handling the affective issues where deep discovering excels. In turn, connectionist AI has been criticized as improperly matched for deliberative step-by-step issue solving, incorporating knowledge, and managing planning. Finally, Nouvelle AI masters reactive and real-world robotics domains however has actually been slammed for problems in incorporating knowing and understanding.

Hybrid AIs including one or more of these techniques are currently seen as the path forward. [19] [81] [82] Russell and Norvig conclude that:

Overall, Dreyfus saw locations where AI did not have total responses and stated that Al is therefore impossible; we now see many of these exact same areas undergoing ongoing research study and advancement leading to increased ability, not impossibility. [100]

Artificial intelligence.
Automated planning and scheduling
Automated theorem proving
Belief revision
Case-based thinking
Cognitive architecture
Cognitive science
Connectionism
Constraint programming
Deep knowing
First-order reasoning
GOFAI
History of synthetic intelligence
Inductive reasoning programming
Knowledge-based systems
Knowledge representation and thinking
Logic programming
Artificial intelligence
Model checking
Model-based reasoning
Multi-agent system
Natural language processing
Neuro-symbolic AI
Ontology
Philosophy of expert system
Physical symbol systems hypothesis
Semantic Web
Sequential pattern mining
Statistical relational learning
Symbolic mathematics
YAGO ontology
WordNet

Notes

^ McCarthy as soon as stated: “This is AI, so we don’t care if it’s mentally real”. [4] McCarthy reiterated his position in 2006 at the AI@50 conference where he said “Expert system is not, by meaning, simulation of human intelligence”. [28] Pamela McCorduck composes that there are “2 significant branches of artificial intelligence: one aimed at producing smart habits despite how it was accomplished, and the other targeted at modeling intelligent procedures discovered in nature, especially human ones.”, [29] Stuart Russell and Peter Norvig composed “Aeronautical engineering texts do not specify the objective of their field as making ‘makers that fly so precisely like pigeons that they can trick even other pigeons.'” [30] Citations

^ Garnelo, Marta; Shanahan, Murray (October 2019). “Reconciling deep knowing with symbolic expert system: representing things and relations”. Current Opinion in Behavioral Sciences. 29: 17-23. doi:10.1016/ j.cobeha.2018.12.010. hdl:10044/ 1/67796.
^ Thomason, Richmond (February 27, 2024). “Logic-Based Expert System”. In Zalta, Edward N. (ed.). Stanford Encyclopedia of Philosophy.
^ Garnelo, Marta; Shanahan, Murray (2019-10-01). “Reconciling deep learning with symbolic artificial intelligence: representing things and relations”. Current Opinion in Behavioral Sciences. 29: 17-23. doi:10.1016/ j.cobeha.2018.12.010. hdl:10044/ 1/67796. S2CID 72336067.
^ a b Kolata 1982.
^ Kautz 2022, pp. 107-109.
^ a b Russell & Norvig 2021, p. 19.
^ a b Russell & Norvig 2021, pp. 22-23.
^ a b Kautz 2022, pp. 109-110.
^ a b c Kautz 2022, p. 110.
^ Kautz 2022, pp. 110-111.
^ a b Russell & Norvig 2021, p. 25.
^ Kautz 2022, p. 111.
^ Kautz 2020, pp. 110-111.
^ Rumelhart, David E.; Hinton, Geoffrey E.; Williams, Ronald J. (1986 ). “Learning representations by back-propagating mistakes”. Nature. 323 (6088 ): 533-536. Bibcode:1986 Natur.323..533 R. doi:10.1038/ 323533a0. ISSN 1476-4687. S2CID 205001834.
^ LeCun, Y.; Boser, B.; Denker, I.; Henderson, D.; Howard, R.; Hubbard, W.; Tackel, L. (1989 ). “Backpropagation Applied to Handwritten Postal Code Recognition”. Neural Computation. 1 (4 ): 541-551. doi:10.1162/ neco.1989.1.4.541. S2CID 41312633.
^ a b Marcus & Davis 2019.
^ a b Rossi, Francesca. “Thinking Fast and Slow in AI”. AAAI. Retrieved 5 July 2022.
^ a b Selman, Bart. “AAAI Presidential Address: The State of AI”. AAAI. Retrieved 5 July 2022.
^ a b c Kautz 2020.
^ Kautz 2022, p. 106.
^ Newell & Simon 1972.
^ & McCorduck 2004, pp. 139-179, 245-250, 322-323 (EPAM).
^ Crevier 1993, pp. 145-149.
^ McCorduck 2004, pp. 450-451.
^ Crevier 1993, pp. 258-263.
^ a b Kautz 2022, p. 108.
^ Russell & Norvig 2021, p. 9 (logicist AI), p. 19 (McCarthy’s work).
^ Maker 2006.
^ McCorduck 2004, pp. 100-101.
^ Russell & Norvig 2021, p. 2.
^ McCorduck 2004, pp. 251-259.
^ Crevier 1993, pp. 193-196.
^ Howe 1994.
^ McCorduck 2004, pp. 259-305.
^ Crevier 1993, pp. 83-102, 163-176.
^ McCorduck 2004, pp. 421-424, 486-489.
^ Crevier 1993, p. 168.
^ McCorduck 2004, p. 489.
^ Crevier 1993, pp. 239-243.
^ Russell & Norvig 2021, p. 316, 340.
^ Kautz 2022, p. 109.
^ Russell & Norvig 2021, p. 22.
^ McCorduck 2004, pp. 266-276, 298-300, 314, 421.
^ Shustek, Len (June 2010). “An interview with Ed Feigenbaum”. Communications of the ACM. 53 (6 ): 41-45. doi:10.1145/ 1743546.1743564. ISSN 0001-0782. S2CID 10239007. Retrieved 2022-07-14.
^ Lenat, Douglas B; Feigenbaum, Edward A (1988 ). “On the thresholds of knowledge”. Proceedings of the International Workshop on Artificial Intelligence for Industrial Applications: 291-300. doi:10.1109/ AIIA.1988.13308. S2CID 11778085.
^ Russell & Norvig 2021, pp. 22-24.
^ McCorduck 2004, pp. 327-335, 434-435.
^ Crevier 1993, pp. 145-62, 197-203.
^ a b Russell & Norvig 2021, p. 23.
^ a b Clancey 1987.
^ a b Shustek, Len (2010 ). “An interview with Ed Feigenbaum”. Communications of the ACM. 53 (6 ): 41-45. doi:10.1145/ 1743546.1743564. ISSN 0001-0782. S2CID 10239007. Retrieved 2022-08-05.
^ “The fascination with AI: what is expert system?”. IONOS Digitalguide. Retrieved 2021-12-02.
^ Hayes-Roth, Murray & Adelman 2015.
^ Hayes-Roth, Barbara (1985 ). “A blackboard architecture for control”. Expert system. 26 (3 ): 251-321. doi:10.1016/ 0004-3702( 85 )90063-3.
^ Hayes-Roth, Barbara (1980 ). Human Planning Processes. RAND.
^ Pearl 1988.
^ Spiegelhalter et al. 1993.
^ Russell & Norvig 2021, pp. 335-337.
^ Russell & Norvig 2021, p. 459.
^ Quinlan, J. Ross. “Chapter 15: Learning Efficient Classification Procedures and their Application to Chess End Games”. In Michalski, Carbonell & Mitchell (1983 ).
^ Quinlan, J. Ross (1992-10-15). C4.5: Programs for Machine Learning (1st ed.). San Mateo, Calif: Morgan Kaufmann. ISBN 978-1-55860-238-0.
^ Mitchell, Tom M.; Utgoff, Paul E.; Banerji, Ranan. “Chapter 6: Learning by Experimentation: Acquiring and Refining Problem-Solving Heuristics”. In Michalski, Carbonell & Mitchell (1983 ).
^ Valiant, L. G. (1984-11-05). “A theory of the learnable”. Communications of the ACM. 27 (11 ): 1134-1142. doi:10.1145/ 1968.1972. ISSN 0001-0782. S2CID 12837541.
^ Koedinger, K. R.; Anderson, J. R.; Hadley, W. H.; Mark, M. A.; others (1997 ). “Intelligent tutoring goes to school in the big city”. International Journal of Artificial Intelligence in Education (IJAIED). 8: 30-43. Retrieved 2012-08-18.
^ Shapiro, Ehud Y (1981 ). “The Model Inference System”. Proceedings of the 7th international joint conference on Artificial intelligence. IJCAI. Vol. 2. p. 1064.
^ Manna, Zohar; Waldinger, Richard (1980-01-01). “A Deductive Approach to Program Synthesis”. ACM Trans. Program. Lang. Syst. 2 (1 ): 90-121. doi:10.1145/ 357084.357090. S2CID 14770735.
^ Schank, Roger C. (1983-01-28). Dynamic Memory: A Theory of Reminding and Learning in Computers and People. Cambridge Cambridgeshire: New York City: Cambridge University Press. ISBN 978-0-521-27029-8.
^ Hammond, Kristian J. (1989-04-11). Case-Based Planning: Viewing Planning as a Memory Task. Boston: Academic Press. ISBN 978-0-12-322060-8.
^ Koza, John R. (1992-12-11). Genetic Programming: On the Programming of Computers by Means of Natural Selection (1st ed.). Cambridge, Mass: A Bradford Book. ISBN 978-0-262-11170-6.
^ Mostow, David Jack. “Chapter 12: Machine Transformation of Advice into a Heuristic Search Procedure”. In Michalski, Carbonell & Mitchell (1983 ).
^ Bareiss, Ray; Porter, Bruce; Wier, Craig. “Chapter 4: Protos: An Exemplar-Based Learning Apprentice”. In Michalski, Carbonell & Mitchell (1986 ), pp. 112-139.
^ Carbonell, Jaime. “Chapter 5: Learning by Analogy: Formulating and Generalizing Plans from Past Experience”. In Michalski, Carbonell & Mitchell (1983 ), pp. 137-162.
^ Carbonell, Jaime. “Chapter 14: Derivational Analogy: A Theory of Reconstructive Problem Solving and Expertise Acquisition”. In Michalski, Carbonell & Mitchell (1986 ), pp. 371-392.
^ Mitchell, Tom; Mabadevan, Sridbar; Steinberg, Louis. “Chapter 10: LEAP: A Learning Apprentice for VLSI Design”. In Kodratoff & Michalski (1990 ), pp. 271-289.
^ Lenat, Douglas. “Chapter 9: The Role of Heuristics in Learning by Discovery: Three Case Studies”. In Michalski, Carbonell & Mitchell (1983 ), pp. 243-306.
^ Korf, Richard E. (1985 ). Learning to Solve Problems by Searching for Macro-Operators. Research Notes in Expert System. Pitman Publishing. ISBN 0-273-08690-1.
^ Valiant 2008.
^ a b Garcez et al. 2015.
^ Marcus 2020, p. 44.
^ Marcus 2020, p. 17.
^ a b Rossi 2022.
^ a b Selman 2022.
^ Garcez & Lamb 2020, p. 2.
^ Garcez et al. 2002.
^ Rocktäschel, Tim; Riedel, Sebastian (2016 ). “Learning Knowledge Base Inference with Neural Theorem Provers”. Proceedings of the 5th Workshop on Automated Knowledge Base Construction. San Diego, CA: Association for Computational Linguistics. pp. 45-50. doi:10.18653/ v1/W16 -1309. Retrieved 2022-08-06.
^ Serafini, Luciano; Garcez, Artur d’Avila (2016 ), Logic Tensor Networks: Deep Learning and Logical Reasoning from Data and Knowledge, arXiv:1606.04422.
^ a b Garcez, Artur d’Avila; Lamb, Luis C.; Gabbay, Dov M. (2009 ). Neural-Symbolic Cognitive Reasoning (1st ed.). Berlin-Heidelberg: Springer. Bibcode:2009 nscr.book … D. doi:10.1007/ 978-3-540-73246-4. ISBN 978-3-540-73245-7. S2CID 14002173.
^ Kiczales, Gregor; Rivieres, Jim des; Bobrow, Daniel G. (1991-07-30). The Art of the Metaobject Protocol (1st ed.). Cambridge, Mass: The MIT Press. ISBN 978-0-262-61074-2.
^ Motik, Boris; Shearer, Rob; Horrocks, Ian (2009-10-28). “Hypertableau Reasoning for Description Logics”. Journal of Artificial Intelligence Research. 36: 165-228. arXiv:1401.3485. doi:10.1613/ jair.2811. ISSN 1076-9757. S2CID 190609.
^ Kuipers, Benjamin (1994 ). Qualitative Reasoning: Modeling and Simulation with Incomplete Knowledge. MIT Press. ISBN 978-0-262-51540-5.
^ Russell & Norvig 2021.
^ Leo de Penning, Artur S. d’Avila Garcez, Luís C. Lamb, John-Jules Ch. Meyer: “A Neural-Symbolic Cognitive Agent for Online Learning and Reasoning.” IJCAI 2011: 1653-1658.
^ McCarthy & Hayes 1969.
^ McCarthy 1959.
^ Nilsson 1998, p. 7.
^ Olazaran 1993, pp. 411-416.
^ Olazaran 1993, pp. 415-416.
^ Marcus 2020, p. 20.
^ Garcez & Lamb 2020, p. 8.
^ a b Russell & Norvig 2021, p. 982.
^ Brooks 1991, p. 143.
^ Brooks 1991, p. 151.
^ Brooks 1991, p. 150.
^ Brooks 1991, p. 142.
References

Brooks, Rodney A. (1991 ). “Intelligence without representation”. Artificial Intelligence. 47 (1 ): 139-159. doi:10.1016/ 0004-3702( 91 )90053-M. ISSN 0004-3702. S2CID 207507849. Retrieved 2022-09-13.
Clancey, William (1987 ). Knowledge-Based Tutoring: The GUIDON Program (MIT Press Series in Artificial Intelligence) (Hardcover ed.).
Crevier, Daniel (1993 ). AI: The Tumultuous Search for Expert System. New York City, NY: BasicBooks. ISBN 0-465-02997-3.
Dreyfus, Hubert L (1981 ). “From micro-worlds to understanding representation: AI at an impasse” (PDF). Mind Design. MIT Press, Cambridge, MA: 161-204.
Garcez, Artur S. d’Avila; Broda, Krysia; Gabbay, Dov M.; Gabbay, Augustus de Morgan Professor of Logic Dov M. (2002 ). Neural-Symbolic Learning Systems: Foundations and Applications. Springer Science & Business Media. ISBN 978-1-85233-512-0.
Garcez, Artur; Besold, Tarek; De Raedt, Luc; Földiák, Peter; Hitzler, Pascal; Icard, Thomas; Kühnberger, Kai-Uwe; Lamb, Luís; Miikkulainen, Risto; Silver, Daniel (2015 ). Neural-Symbolic Learning and Reasoning: Contributions and Challenges. AAI Spring Symposium – Knowledge Representation and Reasoning: Integrating Symbolic and Neural Approaches. Stanford, CA: AAAI Press. doi:10.13140/ 2.1.1779.4243.
Garcez, Artur d’Avila; Gori, Marco; Lamb, Luis C.; Serafini, Luciano; Spranger, Michael; Tran, Son N. (2019 ), Neural-Symbolic Computing: An Effective Methodology for Principled Integration of Artificial Intelligence and Reasoning, arXiv:1905.06088.
Garcez, Artur d’Avila; Lamb, Luis C. (2020 ), Neurosymbolic AI: The 3rd Wave, arXiv:2012.05876.
Haugeland, John (1985 ), Artificial Intelligence: The Very Idea, Cambridge, Mass: MIT Press, ISBN 0-262-08153-9.
Hayes-Roth, Frederick; Murray, William; Adelman, Leonard (2015 ). “Expert systems”. AccessScience. doi:10.1036/ 1097-8542.248550.
Honavar, Vasant; Uhr, Leonard (1994 ). Symbolic Artificial Intelligence, Connectionist Networks & Beyond (Technical report). Iowa State University Digital Repository, Computer Science Technical Reports. 76. p. 6.
Honavar, Vasant (1995 ). Symbolic Artificial Intelligence and Numeric Artificial Neural Networks: Towards a Resolution of the Dichotomy. The Springer International Series In Engineering and Computer Science. Springer US. pp. 351-388. doi:10.1007/ 978-0-585-29599-2_11.
Howe, J. (November 1994). “Artificial Intelligence at Edinburgh University: a Point of view”. Archived from the original on 15 May 2007. Retrieved 30 August 2007.
Kautz, Henry (2020-02-11). The Third AI Summer, Henry Kautz, AAAI 2020 Robert S. Engelmore Memorial Award Lecture. Retrieved 2022-07-06.
Kautz, Henry (2022 ). “The Third AI Summer: AAAI Robert S. Engelmore Memorial Lecture”. AI Magazine. 43 (1 ): 93-104. doi:10.1609/ aimag.v43i1.19122. ISSN 2371-9621. S2CID 248213051. Retrieved 2022-07-12.
Kodratoff, Yves; Michalski, Ryszard, eds. (1990 ). Machine Learning: an Artificial Intelligence Approach. Vol. III. San Mateo, Calif.: Morgan Kaufman. ISBN 0-934613-09-5. OCLC 893488404.
Kolata, G. (1982 ). “How can computers get good sense?”. Science. 217 (4566 ): 1237-1238. Bibcode:1982 Sci … 217.1237 K. doi:10.1126/ science.217.4566.1237. PMID 17837639.
Maker, Meg Houston (2006 ). “AI@50: AI Past, Present, Future”. Dartmouth College. Archived from the original on 3 January 2007. Retrieved 16 October 2008.
Marcus, Gary; Davis, Ernest (2019 ). Rebooting AI: Building Artificial Intelligence We Can Trust. New York: Pantheon Books. ISBN 9781524748258. OCLC 1083223029.
Marcus, Gary (2020 ), The Next Decade in AI: Four Steps Towards Robust Expert system, arXiv:2002.06177.
McCarthy, John (1959 ). PROGRAMS WITH SOUND JUDGMENT. Symposium on Mechanization of Thought Processes. NATIONAL PHYSICAL LABORATORY, TEDDINGTON, UK. p. 8.
McCarthy, John; Hayes, Patrick (1969 ). “Some Philosophical Problems From the Standpoint of Artificial Intelligence”. Machine Intelligence 4. B. Meltzer, Donald Michie (eds.): 463-502.
McCorduck, Pamela (2004 ), Machines Who Think (2nd ed.), Natick, Massachusetts: A. K. Peters, ISBN 1-5688-1205-1.
Michalski, Ryszard; Carbonell, Jaime; Mitchell, Tom, eds. (1983 ). Machine Learning: an Expert System Approach. Vol. I. Palo Alto, Calif.: Tioga Publishing Company. ISBN 0-935382-05-4. OCLC 9262069.
Michalski, Ryszard; Carbonell, Jaime; Mitchell, Tom, eds. (1986 ). Artificial intelligence: an Artificial Intelligence Approach. Vol. II. Los Altos, Calif.: Morgan Kaufman. ISBN 0-934613-00-1.
Newell, Allen; Simon, Herbert A. (1972 ). Human Problem Solving (1st ed.). Englewood Cliffs, New Jersey: Prentice Hall. ISBN 0-13-445403-0.
Newell, Allen; Simon, H. A. (1976 ). “Computer Science as Empirical Inquiry: Symbols and Search”. Communications of the ACM. 19 (3 ): 113-126. doi:10.1145/ 360018.360022.
Nilsson, Nils (1998 ). Artificial Intelligence: A New Synthesis. Morgan Kaufmann. ISBN 978-1-55860-467-4. Archived from the initial on 26 July 2020. Retrieved 18 November 2019.
Olazaran, Mikel (1993-01-01), “A Sociological History of the Neural Network Controversy”, in Yovits, Marshall C. (ed.), Advances in Computers Volume 37, vol. 37, Elsevier, pp. 335-425, doi:10.1016/ S0065-2458( 08 )60408-8, ISBN 9780120121373, recovered 2023-10-31.
Pearl, J. (1988 ). Probabilistic Reasoning in Intelligent Systems: Networks of Plausible Inference. San Mateo, California: Morgan Kaufmann. ISBN 978-1-55860-479-7. OCLC 249625842.
Russell, Stuart J.; Norvig, Peter (2021 ). Artificial Intelligence: A Modern Approach (4th ed.). Hoboken: Pearson. ISBN 978-0-13-461099-3. LCCN 20190474.
Rossi, Francesca (2022-07-06). “AAAI2022: Thinking Fast and Slow in AI (AAAI 2022 Invited Talk)”. Retrieved 2022-07-06.
Selman, Bart (2022-07-06). “AAAI2022: Presidential Address: The State of AI”. Retrieved 2022-07-06.
Serafini, Luciano; Garcez, Artur d’Avila (2016-07-07), Logic Tensor Networks: Deep Learning and Logical Reasoning from Data and Knowledge, arXiv:1606.04422.
Spiegelhalter, David J.; Dawid, A. Philip; Lauritzen, Steffen; Cowell, Robert G. (1993 ). “Bayesian analysis in expert systems”. Statistical Science. 8 (3 ).
Turing, A. M. (1950 ). “I.-Computing Machinery and Intelligence”. Mind. LIX (236 ): 433-460. doi:10.1093/ mind/LIX.236.433. ISSN 0026-4423. Retrieved 2022-09-14.
Valiant, Leslie G (2008 ). “Knowledge Infusion: In Pursuit of Robustness in Artificial Intelligence”. In Hariharan, R.; Mukund, M.; Vinay, V. (eds.). Foundations of Software Technology and Theoretical Computer Technology (Bangalore). pp. 415-422.
Xifan Yao; Jiajun Zhou; Jiangming Zhang; Claudio R. Boer (2017 ). From Intelligent Manufacturing to Smart Manufacturing for Industry 4.0 Driven by Next Generation Artificial Intelligence and Further On.