CSE 151 Lecture Notes
INTRODUCTION TO ARTIFICIAL INTELLIGENCE

AI: A Modern Approach, Chapter 1: Introduction

If this is not the course you expected to find in this room, stick around anyhow: this will probably be more interesting than the course you were looking for.

There is a web homepage for this course. There's a syllabus that describes the organization of the course. This is a standard upper-division CSE course:

We will take a modern, computer science-oriented view of AI. This is NOT

WHAT THIS COURSE IS ABOUT

What you will learn here are principles and algorithms for representing, reasoning with, and learning knowledge. Knowledge is best defined by its place on a spectrum from low-level to high-level:

data immediately available inputs from the world,
e.g. telephone registration keystrokes
information organized and generalized data,
e.g. "Martha Lee has not taken CSE150"
knowledge organized and generalized information,
e.g. "CSE150 is a prerequisite for CSE151"
wisdom organized and generalized knowledge,
e.g. "Martha is the sort of person who will do well even without satisfying formal prerequisite requirements."

Note the word "formal" above. For knowledge to be processed by computer, it must be represented in a formal way.

DEFINITIONS OF AI

This week we will look at definitions of AI and the history of the field. The purpose is to give you a framework which lets you see WHY we study the algorithms that we do study in the rest of the quarter.

There are four alternative ways of saying what AI is about:

Let's start by noting that the only candidate machines are electronic computers, or electronic computer-controlled mechanical devices, i.e. robots. People have speculated about biological robots etc., but nothing is close to being built.

This slide shows several different definitions of AI grouped along the "thinking" and "behaving" dimensions.

The technical word for thinking or behaving correctly is "rationally". Humans are often irrational. In practice, AI researchers interested in simulating humans only choose one aspect of human thinking or behaviour, and focus on the cases where humans are rational.

"Thinking" includes reasoning and learning. "Behaving" includes perception, language, and physical action as well.

In this course we concentrate on reasoning, learning and perception.

ALAN TURING AND THE "TURING TEST"

Alan Turing invented theoretical computer science, designed some of the first computers, and was also the first scientist to think seriously about computer intelligence, in 1950.

The Turing test is an OPERATIONAL definition of intelligence. An operational definition is one that gives an algorithm for testing objectively whether the definition is satisfied.

Example: "If it looks, walks, and quacks like a duck, then it is a duck."

The Turing test scenario is that a human communicates by typing at a terminal with TWO other agents. The human can say and ask whatever s/he likes, in natural English.

If the human cannot decide which of the two agents is a human and which is a computer, then the computer has achieved AI.

Question: which definition of AI does the Turing test use?

Answer: "behaving humanly".

SCIENCE VERSUS ENGINEERING

A scientist wants to understand the external world. An engineer wants to build useful devices. The scientific aspect of AI is often called "cognitive modeling." It aims at human-like thinking and/or behaviour. The engineering aspect usually aims for rational thinking and/or behaviour. Both aspects usually just examine one part of thinking and/or behaviour. For very narrow domains, e.g. flying a flight simulator or playing chess, it is possible to achieve super-human performance.

KNOWLEDGE-BASED SYSTEMS

Almost all practical engineering successes of AI have come from "knowledge-based systems". A knowledge-based system is a software system that contains a great quantity of explicit knowledge, e.g. facts like "CSE150 is a prerequisite for CSE151". The system uses reasoning algorithms to make deductions from this knowledge to solve particular problems like "Which courses should Martha take this quarter?" The leading advocate of knowledge-based systems is Edward A. Feigenbaum, a professor at Stanford and now Chief Scientist of the U.S. Air Force. Feigenbaum is the most recent winner of the Turing award.

MISINFORMED HOSTILITY TO AI

AI has also always attracted hostility from people who considered it too speculative. Mostly this hostility has been due to the mistaken opinion that there have been no practical successes on the way to the goal of full intelligence. In fact there are thousands of AI systems in daily use in hundreds of companies and hundreds of products. Just one example: automatic transmissions in new cars are usually controlled by small KBSs now (using a particular technology called fuzzy control). In the first part of this class, we will study search algorithms, and then how to represent knowledge formally using logic, and how to reason with this formal knowledge.