Knowledge Representation in AI

Knowledge Representation

• Representing knowledge effectively is crucial in Artificial Intelligence (AI).
• Knowledge representation focuses on both storing knowledge efficiently and implementing reasoning processes.
• Knowledge acts as a primary factor in guiding intelligent system decisions.
• To make decisions, AI agents require a knowledge base (KB) which stores facts about the world.

What is Knowledge?

• Data represents primitive verifiable facts, often voluminous and frequently changing.
• Information is interpreted data.
• Knowledge encompasses relationships between data sets (information) which can be used for deducing new information; it is typically general.
• Knowledge represents the behavior of abstract models of the world.

General Knowledge Representations

• Five common schemas for knowledge representation:
  • Logic-based representation (First Order Predicate Logic, FOL)
  • Procedural representation (Production Rules)
  • Network representation (Semantic Networks, Conceptual Graphs)
  • Structural representation (Scripts, Frames, Objects)

Logic-Based Representation

• Propositional Logic:
  • Represents knowledge as propositions (statements that are either true or false).
  • Uses logical operators like AND (), OR (), NOT (), implication (), and equivalence () to combine propositions.
  • Example:
    • "If it is raining (p), then the ground is wet (q)."
    • "If the ground is wet (q), then the ground is slippery (r)."
    • "It is raining (p)."
• First Order Logic (FOL):
  • Extends propositional logic by incorporating relations, objects, and functions applied to objects.
  • Allows reasoning about truths, falsehoods, and unknowns about these relations.
  • Constructs:
    • Constants: Represent objects (e.g., John, apples)
    • Predicates: Represent properties and relations (e.g., likes(john, apples))
    • Functions: Transform objects (e.g., likes(john, fruit_of(apple_tree)))
    • Variables: Represent any object (e.g., likes(X, apples))
    • Quantifiers: Qualify variable values:
      • Universal: ∀X (True for all objects)
      • Existential: ∃X (True for at least one object)

Procedural Representation: Production Rules

• Production rules involve sets of "if condition then action" pairs.
• This results in a match-resolve-act cycle.
  • Match: The agent checks if each rule's condition holds.
  • Resolve: The agent selects a rule to execute from a set of rules whose conditions match (conflict set).
  • Act: If a rule's condition holds, it "fires" and its action is carried out.
• Working memory holds knowledge that can be accessed and written to by production rules.
• This mechanism allows new knowledge generated by a rule to potentially trigger other rules.

Network Representations: Semantic Networks

• Semantic networks represent objects as nodes.
• Relationships between these objects are depicted in a hierarchical network.
• This allows for inheritance of properties, which is analogous to OOP (Object-Oriented Programming).
• Example:
  • Fact-1: Canary is a bird.
  • Fact-2: Birds can fly.
  • Fact-3: Canary is yellow in color.
  • Fact-4: Birds have wings.

Structural Representation: Frame Representations

• A frame is a data structure representing knowledge about a specific object or concept.
• These frames provide a more natural way to organize values than semantic networks, allowing constraints to be imposed on potential values for each slot within a frame.
• Frame representations can be implemented using OOP techniques.
• Inheritance of properties is easily managed in frame-based systems.