## Definition **Knowledge Representation (KR)** is the subfield of AI concerned with how to encode information about the world so that a computer system can use it to solve complex tasks. The central concern: a representation must be *expressive* enough to capture the relevant knowledge, *efficient* enough to reason with, and *clear* enough to maintain. ## Representational Forms ### Logic-Based - **[[Propositional Logic]]** — Boolean atoms and connectives. - **[[First-Order Logic]]** — objects, predicates, quantifiers. Most expressive of the classical KR languages. - **Description Logics** — decidable fragments of FOL; underlie OWL. - **Default and non-monotonic logics** — handle revision when new facts arrive (Reiter's default logic, circumscription). ### Structured - **Semantic Networks** — graphs of nodes (concepts) and edges (relations). - **Frames** — structured records with slots, defaults, and inheritance (Minsky 1974). - **Ontologies** — formal specifications of concepts and relations in a domain. See [[Ontologies and Semantic Networks]]. ### Probabilistic - **[[Bayesian Network]]** — represent uncertainty over propositions. - **Markov Logic Networks** — combine FOL with weights on formulas. ### Procedural - **Production rules** — *if-then* rules that fire on a working memory. - **Scripts** — stereotypical sequences of events (Schank & Abelson). ## The Knowledge-Engineering Process 1. **Identify the task.** What questions must the system answer? 2. **Assemble relevant knowledge.** Interview experts; read documents. 3. **Choose a vocabulary** of predicates, constants, functions. 4. **Encode general knowledge** about the domain (axioms). 5. **Encode specific instances** of the problem. 6. **Pose queries.** Run inference. 7. **Debug** — most knowledge bases fail to capture intended meaning on the first pass. ## Trade-offs - **Expressivity vs Tractability.** FOL is more expressive than description logics but reasoning is semi-decidable; DLs are decidable but less expressive. - **Crisp vs Probabilistic.** Real-world knowledge is often uncertain; pure logic fights this. ## Modern Relevance LLMs have absorbed an enormous amount of world knowledge in their weights. But: - They cannot easily be *audited* or *updated* in the way an explicit KB can. - Hybrid systems — LLMs reading from structured KBs via tool use — are an active area in 2026. - See [[Retrieval-Augmented Generation]] for one bridge between LLMs and explicit knowledge. ## Related - [[First-Order Logic]] - [[Ontologies and Semantic Networks]] - [[Bayesian Network]] - [[Inference Rules]] - [[Retrieval-Augmented Generation]]