Introduction to Knowledge Based AI

Topics Conundrums · KBAI · Cognitive Systems · Semantic Networks · Raven's Matrices · Generate and Test
01 //

Conundrums in AI

AI faces fundamental tensions—central challenges not all solvable, but essential when designing intelligent agents.

1 Limited Resources vs. Intractable

Finite compute, memory, speed—yet many AI problems are computationally intractable. How to get near real-time performance?

Example: combinatorial search spaces grow exponentially—we use heuristics, approximations, and anytime algorithms.

Learn more
2 Local vs. Global

Computation is local; most AI problems have global constraints. How do agents address global problems with only local computation?

Agents use constraints, coordination, and incremental refinement to approximate global solutions from local steps.

Learn more
3 Deductive vs. Abductive/Inductive

Logic is deductive; many problems are abductive (best explanation) or inductive (learning). How do we bridge the gap?

We combine probabilistic reasoning, learning from data, and heuristic search with logical frameworks.

Learn more
4 Stale Knowledge vs. New Problems

The world is dynamic; knowledge is limited. Agents start with what they know. How can they address genuinely new problems?

Transfer learning, analogy, and composition of known concepts help agents generalize to novel situations.

Learn more
5 Explanation & Justification

Reasoning and learning are hard; explanation adds complexity. How can an agent explain or justify its decisions?

Trace-based explanations, attention mechanisms, and interpretable representations support explainability.

Learn more
02 //

Characteristics of AI Agents

AI agents have bounded capabilities. Understanding these limits helps design systems that work within them.

Limited computing power

Processing speed, memory, resources are finite.

Limited sensors

Cannot perceive everything in the world.

Limited attention

Cannot focus on everything at once.

Deductive logic

Computational logic is fundamentally deductive.

Incomplete knowledge

Agent's knowledge of the world is incomplete.

Bounded Rationality

How can AI agents with such bounded rationality address open-ended problems in the world?

03 //

What is Knowledge-Based AI

Systems like Watson (Jeopardy) illustrate three fundamental processes that work together.

Reasoning · Learning · Memory — The Deliberation Cycle
Reason Learn Memory

We learn → reason; reasoning yields learning; both read from and write to memory.

Reasoning

Understanding/generating language, decisions, inferences.

Learning

Acquiring and storing knowledge from answers.

Memory

Storing learned knowledge; providing access for reasoning.

Deliberation

Together, reasoning + learning + memory = deliberation. One part of the agent architecture—alongside reaction and metacognition.

04 //

Four Schools of AI

Two spectrums: thinking vs. acting, and humanly vs. rationally. Hover each quadrant to explore.

2×2 Quadrant — Four Schools
Thinking Humanly Cognitive science Models of human cognition Acting Humanly Turing test Mimics human performance Thinking Rationally Formal logic Correct inference Acting Rationally Rational agents Optimal behavior Humanly Rationally

KBAI & Cognitive Systems → human side (thinking & acting humanly)

05 //

Cognitive Systems

The course subtitle breaks down as:

Cognitive

Human-like intelligence. Goal: human-level, human-like intelligence.

Systems

Multiple interacting components: learning, reasoning, memory.

Cognitive Systems

Human-level intelligence through component interaction.

06 //

Three-Layer Architecture

A cognitive system is situated in the world—receives percepts and produces actions. How do we map percepts → actions?

Percepts → Three Layers → Actions
PERCEPTS Layer 1: Reaction Direct mapping · brake lights → brakes Layer 2: Deliberation Reason about world · plan · learn, reason, memory Layer 3: Metacognition Reason about reasoning · repair & reflect ACTIONS
Reaction

Direct mapping. No planning; purely reactive.

Deliberation

Reason about world; form plans. Learning, reasoning, memory.

Metacognition

Reason about the internal mental world; repair & reflect.

07 //

Raven's Progressive Matrices

Written in the 1930s—the most widespread, reliable test of general human intelligence.