Symbolic artificial intelligence is an approach to artificial intelligence that represents objects, relationships, and knowledge through explicit symbols and manipulates those representations using rules. Its systems include logical reasoners, rule-based programs, and planners that search among possible actions. Unlike approaches centered on learning numerical representations, symbolic methods make important aspects of a problem’s structure explicit. Formal logic is a major foundation, but symbolic AI also encompasses heuristic search and other procedures not organized exclusively around logical deduction. (www-formal.stanford.edu)
Historical development
Symbolic AI emerged prominently during the 1950s. Allen Newell, Herbert A. Simon, and J. Clifford Shaw developed the Logic Theorist, an early program for discovering proofs in symbolic logic. At the 1956 Dartmouth workshop, they described the list-processing language used to implement it. Such projects treated intelligent problem-solving as the manipulation of structured expressions rather than merely numerical calculation. (www-formal.stanford.edu)
John McCarthy developed Lisp to support computations on symbolic expressions; implementation began in autumn 1958. His proposed Advice Taker connected this programming work with the idea of representing information as sentences and using inference to determine actions. Lisp’s list structures supported the representation and transformation of expressions, making it an important language for early AI research. (www-formal.stanford.edu)
In 1976, Newell and Simon articulated the physical symbol system hypothesis, proposing that physical symbol systems provide the necessary and sufficient means for general intelligent action. They presented it as an empirical hypothesis, not a proven mathematical theorem. Their account also emphasized heuristic search: selecting promising transformations instead of exhaustively examining every possibility. (web.cs.wpi.edu)
Knowledge representation
A central concern is knowledge representation and reasoning: choosing structures that express relevant information and support useful computation. A knowledge base may contain facts, general rules, and descriptions of a current situation. For example, LocatedIn(package, warehouse) identifies a relationship between two represented objects. The symbols acquire their intended meaning through the representation’s interpretation; their spelling alone does not establish that meaning. (www-formal.stanford.edu)
Propositional logic represents statements and their combinations, whereas first-order logic adds variables, quantifiers, and predicates describing objects and relationships. Other symbolic structures include semantic networks and frames, which organize information through relationships or collections of attributes. These choices differ in expressive power and in the reasoning procedures they support. (cl.cam.ac.uk)
A computational ontology specifies a domain’s classes, properties, individuals, and relationships. This technical use differs from ontology as a branch of philosophy. The Web Ontology Language, or OWL, provides formally defined semantics that allow reasoners to check consistency and derive information implicit in explicitly stated descriptions. (w3.org)
Inference and problem-solving
An inference engine applies rules to represented information. In forward chaining, computation starts with available facts and repeatedly derives consequences. In backward chaining, it starts with a query or goal and seeks rules and supporting facts that could establish it. The appropriate strategy depends partly on whether the task concerns many consequences or one particular question. (cl.cam.ac.uk)
A simple example of deductive reasoning combines “all inspected packages are cleared” with “package A is inspected” to infer “package A is cleared.” This demonstrates the distinction between stored facts and derived conclusions. A logically sound inference procedure preserves truth relative to its premises, but it does not establish that the premises accurately describe the real world. (cl.cam.ac.uk)
Symbolic problem-solving also searches through states, expressions, or candidate solutions. Heuristics guide which alternatives to examine first. In automated planning, actions are described through their preconditions and effects, and the system seeks a sequence connecting an initial state to a goal. STRIPS, introduced by Richard Fikes and Nils Nilsson in 1971, formalized an influential approach used in the Shakey robot project. (web.cs.wpi.edu)
Expert systems apply encoded domain knowledge to specialized tasks. Their architecture commonly distinguishes the knowledge base from the mechanism interpreting it. This separation permits domain rules to be changed without completely rewriting the reasoning machinery, although obtaining usable knowledge from experts can itself require substantial effort. (cl.cam.ac.uk)
Strengths and limitations
Explicit representations allow assumptions and inference steps to be inspected. They can support explanations showing which facts and rules produced a conclusion. However, inspectability does not automatically make a large system easy to understand: lengthy derivations and complex interactions among rules may still be difficult to interpret. Symbolic structure therefore supports, rather than guarantees, explainable AI. (academic.oup.com)
Knowledge acquisition is a persistent limitation. Experts may struggle to articulate informal judgments, and manually encoded rules can omit circumstances encountered outside their intended domain. Commonsense reasoning presents further difficulties, including exceptions and incomplete information. Nonmonotonic reasoning addresses cases in which additional information can invalidate an earlier conclusion, unlike classical deduction, where adding premises does not remove existing consequences. (aaai-24.aaai.org)
The frame problem concerns representing what remains unchanged when an action occurs without requiring an impractically extensive collection of persistence statements. More generally, symbolic systems must balance expressive representations against the computational cost of reasoning. Restricting a representation language can improve tractability while limiting what it directly expresses. (www-formal.stanford.edu)
Relationship to learning-based AI
Symbolic AI is often contrasted with connectionism, particularly artificial neural networks, which learn distributed numerical representations. The distinction is not equivalent to “reasoning versus learning”: symbolic systems can incorporate learned knowledge, while neural systems can perform tasks involving reasoning. (academic.oup.com)
Neuro-symbolic AI combines neural learning with explicit symbolic representations or reasoning procedures. Designs include neural components that identify entities for a reasoner and symbolic constraints incorporated into learning objectives. These combinations seek complementary capabilities, but integration introduces questions about translating between representations and preserving the meaning of symbolic constraints. (arxiv.org)