In-Depth Notes on Rational Agents, Symbolic vs Non-symbolic AI, and Logic in AI

Readings

• Chapter 7, pages 226-264, Russell and Norvig

Symbolic AI vs Non-Symbolic AI

• Non-symbolic AI
  • Performs calculations based on principles established to solve problems.
  • Examples: Genetic algorithms, neural networks, deep learning.
  • Also called "Connectionist AI"; inspired by human brain neuron interactions.
• Symbolic AI
  • Manipulates symbolic representations to find problem solutions.
  • Develops intelligent systems based on rules and knowledge; actions are interpretable.
  • Processes strings of characters representing real-world entities or concepts in structured formats (lists, hierarchies, networks).

Advantages and Disadvantages of AI Types

• Symbolic AI
  • Advantage: Easy to understand reasoning process; clear explanations of conclusions.
  • Disadvantage: Learning requires hand-coded rules, challenging to implement.
• Non-symbolic AI
  • Disadvantage: Opaque decision-making; hard to trace the reasoning behind conclusions (critical for self-driving cars, medical diagnostics).
  • Symbolic AI is mostly used in academia, limited impact in industry.

Rational Agents: Definition

• A rational agent:
  • Has clear preferences and models uncertainty using expected values.
  • Selects actions based on the optimal expected outcomes from feasible options.
  • Examples of settings: artificial intelligence, decision theory, game theory, cognitive science, philosophy.

Decision Theory in Rational Agents

• Studies decision-making under uncertainty, including the mathematical properties of logic.
• A rational agent aims for advantageous outcomes based on past and present inputs.
• Theoretical models help understand technology impacts on human decisions.

Characteristics of Rational Agents

• Profile and properties of rational behavior; reasonable, sensible judgment.
• Goals drive action selection from distinct options.
• Aim to maximize performance based on assessments of observable reality.

Examples of Rational Agents

• Autonomous Vacuum Cleaners: Cleans efficiently without obstacles.
• Self-Driving Cars: Navigates safely, adheres to road rules, avoids hazards.
• Energy-efficient devices: Select actions minimizing energy usage.

Rational Decisions

• Definition of rationality: achieving predefined goals maximally.
• Decisions depend on outcome utility; rationality maximizes expected utility.
• Computational rationality: balancing decision effectiveness and computation costs.

Rationality Criteria

• Preferences: complete relationships among choices (superiority, inferiority, indifference).
• Logically ordered choices without cyclic inconsistencies.
• Incorporates risk attitudes and emotions into preference rankings.

Designing Rational Agents

• Agents must perceive and act based on maximizing expected utility.
• Selection techniques informed by the characteristics of the environment and percepts.

Pac-Man Example of a Rational Agent

• Actions based on environmental assessment and potential outcomes.
• Goal-driven actions with the consideration of utility at each step.

Performance Measurement of Rational Agents

• Gauged by:
  • Achievement of goals.
  • Quality of environmental assessments.
  • Range of performable actions.

Logic in AI

• Predominant AI paradigm before the 1990s, faced challenges:
  • Deterministic nature limited handling of uncertainty.
  • Rule-based systems lacked adaptability from data.
  • Strengths: Compact expressiveness.

Elements of Logic

• Syntax: Defines sentence structures in a specified language.
• Semantics: Assigns meanings to sentences (truth in models).
• Inference Rules: Mechanisms for deriving new truths from established sentences.

Wumpus World Concept

• PEAS: Performance measure, Environment, Actuators, Sensors.
  • Performance measure includes:
  • +1000 for gold, -1000 for death, -1 per step, -10 for using the arrow.
  • Environment comprises squares with specific properties (e.g., smell near danger).
  • Actuators and sensors help navigate and interact with the environment.

Characteristics of Wumpus World

• Not fully observable; local perception is limited.
• Deterministic outcomes; static elements do not change.
• Single-agent interaction; no environmental mobility.

Knowledge Base and Inference in Propositional Logic

• Knowledge Base (KB): Repository for information where TELL() updates and ASK() queries the agent.
• Propositional Logic:
  • Constructs complex sentences via logical connectives (¬, ∧, ∨, →).
  • Models determine truth assignments for symbols.
  • A model consists of assignments yielding true or false outcomes.

Entailment and Models

• Entailment occurs when one statement must follow from another.
• Example: Knowledge base (KB) entails a statement if true in all models where KB is true.

Inference Methods

• Model Checking: Exhaustive enumeration of possible models.
• Truth Tables: Always exponential in complexity.
• DPLL Algorithm: Enhances efficiency through backtracking and pruning methods.

Modus Ponens Inference Rule

• Example:
  • If it rains (Rain), then it's wet (Rain → Wet).
  • Concludes: It is wet (Wet).

Forward and Backward Chaining

• Strategies for entailment using Horn clauses, determining if one sentence is entailed by the knowledge base.

Horn Clauses

• Limitation: Proves completeness in simpler constructions.
• A definite clause is a conditional statement (p1 ∧ · · · ∧ pk → q).

Resolution in Inference

• Applies the full expressiveness of propositional logic.
• Converts all sentences to Conjunctive Normal Form (CNF).
• Derives conclusions unless derivatives yield false assertions.

Summary of Propositional Logic

• Problems:
  • Poor scalability and lack of expressive power.
• Upcoming focus: First-Order Logic.