Introduction to Artificial Intelligence — Lecture 1 Study Materials Summary & Study Notes

🧭 Overview

Artificial Intelligence (AI) is the study and design of systems that exhibit intelligent behavior. AI aims to create programs that can perceive, reason, learn, and act to achieve goals in diverse environments. This course introduction frames AI via definitions, historical milestones, agent models, environment types, and application examples.

🧩 Definitions of AI

AI has many complementary definitions: treating intelligence as thinking like humans, thinking rationally, acting like humans, or acting rationally. Authors differ in emphasis — some focus on computational models of mental faculties, others on the automation of intelligent behavior, and others on designing systems that maximize goal achievement.

🤖 Four High-Level Approaches

• Acting humanly: Evaluated by behavioral indistinguishability, epitomized by the Turing Test (a judge cannot tell machine from human).
• Thinking humanly: Model human cognition and brain processes; intersection with cognitive science.
• Thinking rationally: Emphasizes correct inference and logical reasoning (Aristotelian roots).
• Acting rationally: Agents do the "right thing" to maximize expected goal achievement; this is the rational agent view used by many modern texts.

🗣️ Turing Test and Critiques

The Turing Test is an operational test of human-like behavior. Practical competitions like the Loebner Prize adopt variants. Philosophical critiques include Searle's Chinese Room, which argues that symbol manipulation alone does not imply understanding. Replies include the Systems Reply (the whole system understands) and the Robot Reply (embodiment and perception contribute to understanding).

📚 Foundations and History

AI draws on philosophy, mathematics (Boole), economics (agent payoff views), neuroscience, psychology, and linguistics. The field formally began at the Dartmouth Conference (1956), with early contributions from McCarthy (LISP), Minsky, Shannon, Newell, and Simon.

🧠 Agents and Architectures

An agent perceives via sensors and acts via actuators. An agent's behavior is defined by a function mapping percept sequences to actions. Key agent types (increasing generality): simple reflex agents, reflex agents with state, goal-based agents, utility-based agents, and learning agents.

⚙️ Simple Reflex and Stateful Agents

• Simple reflex agents use condition-action rules ("if-then"). They can be efficient but often short-sighted because they react only to current percepts.
• Reflex agents with state maintain internal state to summarize past percepts, enabling handling of partially observable environments.

🎯 Goal-Based vs Utility-Based Agents

• Goal-based agents plan actions to achieve specified goals; planning may be computationally costly and sensitive to world changes during reasoning.
• Utility-based agents use an evaluation (utility) function to rank states, allowing tradeoffs among competing goals and preferences.

📘 Learning Agents

Learning agents improve performance over time using feedback. Learning components adapt models, rules, or utilities to handle previously unseen situations.

🧾 PEAS Framework

PEAS stands for Performance measure, Environment, Actuators, Sensors. Use PEAS to specify task settings (e.g., taxi driver: performance = safe/fast/comfortable, sensors = cameras/GPS, actuators = steering/brake).

🌍 Environment Properties

Key environment dimensions:

• Fully observable vs partially observable
• Deterministic vs stochastic / strategic
• Episodic vs sequential
• Static vs dynamic
• Discrete vs continuous
• Single-agent vs multiagent Examples map tasks (chess, poker, taxi driving, medical diagnosis) onto these properties to guide agent design.

🧭 Example Environment Mappings

• Chess with a clock: fully observable, strategic, sequential, semi-discrete, multiagent.
• Poker: partially observable, strategic, sequential, discrete, multiagent.
• Taxi driving: partially observable, stochastic, sequential, dynamic, continuous, multiagent.
• Medical diagnosis: partially observable, stochastic, episodic, static, continuous, typically single-agent.

🧰 AI Techniques and Tools

AI systems employ a broad toolkit: search, planning, constraint satisfaction, probabilistic reasoning (Bayesian networks, Monte Carlo), machine learning (reinforcement learning, neural networks), natural language processing, and computer vision. Choice depends on problem structure and environment properties.

🧪 Representative Systems

• TD-Gammon: reinforcement learning + neural networks for backgammon.
• ALVINN: neural-network-driven autonomous driving.
• Pathfinder: medical diagnosis using Bayesian models and simulations.
• Talespin: story-generation using planning and knowledge bases.
• Xavier: mail-delivery robot combining perception and planning.

🧩 AI Questions and Applications

AI research asks: Can we build systems as intelligent as humans or other animals? Can systems be autonomous, evolutionary, or economically valuable (e.g., fraud detection, handwriting recognition)? Real-world AI spans entertainment, robotics, healthcare, manufacturing, space exploration, and finance.

✅ Practical Design Considerations

When building an AI agent, explicitly specify the PEAS description, analyze environment properties, choose appropriate agent architecture, and select algorithms (search, probabilistic models, learning) that suit observability, determinism, and performance criteria.

🔭 Summary

AI blends theoretical foundations and engineering practice. Successful AI requires precise problem framing, selecting the right abstraction (agents, environments), and matching algorithms to the task's uncertainties and constraints. Understanding the tradeoffs among different approaches (human-like vs rational, reactive vs deliberative, learned vs programmed) is central to effective system design.