Glossary
What is symbolic AI
In short
Symbolic AI is an approach in which knowledge is written as explicit symbols, rules and logical formulas, and the machine reasons by combining them under strict laws.
AI basics · symbolic AI
MMOOZZGG.com editorsChecked: 1 min read
For a child
When you solve a problem about three apples and two that got eaten, you use the counting rules you learned at school. You can write such rules for a computer too: what an apple is, how to add, how to subtract. Then it can also solve the problem step by step and show how it thought. This is exactly how the first scientists taught machines to reason.
In plain words
Before neural networks learned from examples, AI was built another way. People described the world for the machine themselves: “all birds have wings”, “a penguin is a bird”, “if it rains, the road is wet”. The program stored these statements and drew new ones from them, like a student in a logic class. This approach is called symbolic. Its strengths: you can check and explain every conclusion, and it needs no mountains of data. Its weakness: the real world is too complex to write all the rules by hand, and the system gets lost in a new situation.
For business
In a company, symbolic AI shows up as a rule engine, a product ontology, a planning system or a check of documents against internal policy. It is chosen where a mistake is expensive and a decision must be explained: a tariff calculation, approval of an operation, limit control. Upkeep is expensive, because experts write and update the rules. The practical option today is hybrid: a language model reads free text, and rules and reference lists check that its answer breaks no conditions. This lowers the number of wrong answers that a company later has to answer for to the client and the regulator.
For a specialist
Symbolic AI rests on the physical symbol system hypothesis, which A. Newell and H. Simon set out in the article “Computer Science as Empirical Inquiry: Symbols and Search” (Communications of the ACM, 1976). It says such a system has the necessary and sufficient means for general intelligent action. The basis goes back to the Dartmouth workshop proposal of 31 August 1955: every aspect of learning or intelligence can in principle be described so precisely that a machine can simulate it. The tools of the approach are predicate logic, production rules, semantic networks and ontologies, state-space search, planners and automated theorem proving. The opposite tradition is called connectionism. Neuro-symbolic AI joins the two. AlphaGeometry (Google DeepMind, 2024) pairs a language model that proposes auxiliary constructions with a symbolic deduction engine. It solved 25 of 30 olympiad geometry problems, while the average gold medalist solved 25.9.
Example
A logistics company gives its planner rules: a driver does not work longer than the set time, a refrigerated truck does not carry chemicals, and urgent orders go first. The system builds the routes for the day and shows, for each one, which rules it took into account.
Related terms
Questions and answers
Sources
- A. Newell, H. A. Simon, Computer Science as Empirical Inquiry: Symbols and Search, Communications of the ACM 19(3), 1976computerhistory.org
- J. McCarthy, M. L. Minsky, N. Rochester, C. E. Shannon, A Proposal for the Dartmouth Summer Research Project on Artificial Intelligence, 31.08.1955www-formal.stanford.edu
- Google DeepMind, AlphaGeometry: An Olympiad-level AI system for geometry, 17.01.2024deepmind.google
Not sure whether your company needs this?
AI solutions map