Power Point Slides of the following notes
I'd like to open with a story:
I propose to consider the question, "Can machines think?" This should begin with definitions of the meaning of the terms "machine" and "think." The definitions might be framed so as to reflect so far as possible the normal use of the words, but this attitude is dangerous. If the meaning of the words "machine" and "think" are to be found by examining how they are commonly used it is difficult to escape the conclusion that the meaning and the answer to the question, "Can machines think?" is to be sought in a statistical survey such as a Gallup poll. But this is absurd. Instead of attempting such a definition I shall replace the question by another, which is closely related to it and is expressed in relatively unambiguous words.
The new form of the problem can be described in terms of a game which we call the 'imitation game." It is played with three people, a man (A), a woman (B), and an interrogator (C) who may be of either sex. The interrogator stays in a room apart from the other two. The object of the game for the interrogator is to determine which of the other two is the man and which is the woman. He knows them by labels X and Y, and at the end of the game he says either "X is A and Y is B" or "X is B and Y is A." The interrogator is allowed to put questions to A and B thus:
C: Will X please tell me the length of his or her hair?
Now suppose X is actually A, then A must answer. It is A's object in the game to try and cause C to make the wrong identification. His answer might therefore be:
A: "My hair is shingled, and the longest strands are about nine inches long."
In order that tones of voice may not help the interrogator the answers should be written, or better still, typewritten. The ideal arrangement is to have a teleprinter communicating between the two rooms. Alternatively the question and answers can be repeated by an intermediary. The object of the game for the third player (B) is to help the interrogator. The best strategy for her is probably to give truthful answers. She can add such things as "I am the woman, don't listen to him!" to her answers, but it will avail nothing as the man can make similar remarks.
We now ask the question, "What will happen when a machine takes the part of A in this game?" Will the interrogator decide wrongly as often when the game is played like this as he does when the game is played between a man and a woman? These questions replace our original, "Can machines think?"
The selection that I read to you just now comes from Alan Turing's seminal paper, "Computing Machinery and Intelligence", which appeared in the journal Mind, Volume LIX, Number 236, 1950. Because the paper is now in the public domain, I can make it available to you on-line.
Many people credit this paper, along with Claude Shannon's 1950 paper on the possibility of a chess-playing program, with launching AI as an idea worthy of scientific pursuit. Credit for launching AI as a discipline usually goes to the now-titled Dartmouth Conference of 1956, which gathered most of the great pioneers of AI who created the discipline over the next forty years.
Please read Turing's paper for class Friday. We will use it to launch our investigation into artificial intelligence.
I am sure that you all have some idea of what you think AI is. You better, I asked you to write on this very issue for HW0. What did you come up with? Can you state it in words?
[We took some class time to talk about this].
It's not easy is it? In fact, pick up a half dozen different AI books and you will likely find a half dozen similar yet very different definitions. This is what happened when I tried this:
Artificial intelligence is...
- "a collection of algorithms that are computationally tractable, adequate approximations of intractably specified problems" (Partridge, 1991)
- "the enterprise of constructing a physical symbol system that can reliably pass the Turing Test" (Ginsberg, 1993)
- "the field of computer science that studies how machines can be made to act intelligently" (Jackson, 1986)
- "a field of study that encompasses computational techniques for performing tasks that apparently require intelligence when performed by humans" (Tanimoto, 1990)
- "a very general investigation of the nature of intelligence and the principles and mechanisms required for understanding or repicating it" (Sharples et al., 1989)
- "the getting of computers to do things that seem to be intelligent" (Rowe, 1988)
<sarcasm> That helped didn't it? You have a much more clear picture now about how to define AI, don't you </sarcasm>.
For most people, a definition of AI centers around the original concept of "intelligence" and some form of the question:
How can minds work?
There are, of course, related questions:
These are significant questions, ones that AI didn't invent. (Except, perhaps, when we ask if general-purpose digital computers can have minds.) Philosophers, linguists, neuroscientists, and cognitive psychologists all work on questions that are quite similar to the questions that AI scientists address. Each of these disciplines has its own focus, tools, and methodology.
Computer scientists model the world using computation, the processing of information. It turns out that computation, and its embodiment in the digital computer, give us a tool of unparalleled flexibility. AI scientists use this flexibility as a portal into the age-old problem of minds, in what they consist, and how they work.
AI attempts to go beyond the ordinary limitations we place on the kinds of the problems we try to solve, the kinds of models we try to build. It draws inspiration and motivation from the fact that humans routinely solve problems that computational theory tells us are intractable, computationally infeasible under resource limitations. Consider something as simple as playing a game of chess. In the early 1970s someone wrote the following bit of trivia:
If every man, woman, and child on earth were to spend every waking moment playing chess (16 hours per day) at the rate of one game per minute, it would take 146 billion years to use every variation of the first 10 moves.
Now, the population of our planet has grown by 50% or so since that time, but even still it would take 97 billion years. (Ah, the power of combinatorial explosion!)
Yet humans play chess, sometimes quite well. The thing is, we play chess the way we do most things: We don't often arrive at the best solutions, but we usually do arrive at solutions that are good enough. This idea of satisficing rather than optimizing when confronted with an intractable problem is central to what AI is about, to what separates it from other algorithmics in computer science.
Most people would characterize AI as a subset of CS, saying that it focuses on a specific set of problems and techniques. But given my view of CS as modeling the world, one might turn this relation on its head: CS as we usually practice it is in fact a subset of AI, since it focuses primarily on problems that are tractable, ones that we already know how to solve. An intelligent agent must be able to model the world, to think about whatever is in its environment, and to handle both tractable and intractable problems well enough to achieve its goals.
AI seeks to loosen the constraints currently placed on our programs:
Now I will give you a definition of AI that I have blended from the writing of several different people. It is by no means perfect, but it captures the flavor of what I think AI is about.
Artificial Intelligence is a science that has defined its goal as giving machines the ability to perform tasks that, when performed by humans, require intelligence. These include the ability to solve problems in complex environments, make decisions, to learn and to understand.
In demonstrating the wide range of approaches towards defining AI, the authors of your book, Stuart Russell and Peter Norvig, have devised a heuristic which considers how a definition treats AI on two dimensions -- thinking vs. acting, and humanly vs. rationally. These two dimensions organize definitions into one of four approaches towards AI
| Systems that think like humans | Systems that think rationally |
| Systems that act like humans | Systems that action rationally |
It is worth considering these four approaches
Acting Humanly (The Turing Test)
We started today's class by introducing the concept of the Turing Test. Russell and Norvig (R&N) consider this approach the cornerstone of the "Acts Humanly" category.
The Turing Test evokes the dominant naive view of intelligence in the world, and thus of AI: the ability to act like a human. (Here, "naive" means uninformed or uninitiated in the ways of cognitive science.) It sidesteps the attempt to define intelligence and instead attempts to operationalize it. The Turing Test seems to identify human-like behavior and human-like experience. To play the game well is to use language as we do, to exhibit strengths and weaknesses similar to humans, and to reproduce the peculiarities of human hardware.
Turing's genius in this paper is that he gets to the heart of the naive view in such a way that a most fascinating academic and practical discipline emerges. For example, this paper was so well thought out, that he anticipated all of the major arguments against AI in the 50 years since his paper was first published. He also very aptly identified the major components of AI
There are are limitations to Turing's Test and his overall approach. Despite these, I consider it so fundamental to the start of an AI course, that we will devote an entire period looking at the strengths and weaknesses to this paper (and hence this approach) during our next session (remember, you should read Turing's paper for class Friday).
Thinking Humanly (Cognitive Science)
It is one thing to solve a problem correctly ("act" like a human). It is completely another to consider what reasoning went into the solution for a problem. Approaches towards AI that fit into the "Thinks Humanly" category are more interested in this later more introspective approach to intelligence. These approaches became popular in the 1960s during the "Cognitive Science revolution." This grew out of psychology where the study of information-processing was replacing the prevailing orthodoxy of behaviorism. As an interdisciplinary study it considers the construction of specific and testable theories concerning the internal activities of the brain.
Critics of these approaches point out that we still know too little about the workings of the brain, and that too often researchers make too large a leap in claiming that because a particular algorithm solves a problem will it must be a fair representation of the human brain, or vice versa. Because the connection from one to the other is so difficult to make, Cognitive Science (and a related field, Cognitive Neuroscience) and AI are now distinctly separate fields of study again. Stuart Russell (outside of your textbook) points out however that "both share .. the following characteristic: the available theories do not explain (or engender) anything resembling human-level general intelligence"
Thinking Rationally (Laws of Thought)
When you took 810:080 (Discrete Structures) one of the things you spent a fair amount of time discussing was the field of logic. This field traces its roots back to Aristotle, who tried to approach things from a normative, or prescriptive, way of thinking rather than a descriptive one. The idea is that with proper logic (syllogisms) that a person (or computer) could always yield correct solutions when provided with correct information (of course, all of this depends on the existence of a "correct" solution). That is, if we can codify things precisely (recall (∃ x) (D(x) ^ R(x)) → (∀ x) (D(x) → S(x)) ?), then we can always arrive at "truthful" solutions.
There are several problems with this approach:
Acting Rationally (Rational Agents)
I mostly agree with your author's decision to focus their work and attention on the concept of "Rational Agents." That is, things which act rationally.
Rational behavior is the process of "doing the right thing." Of course, what does that mean? Well, the right thing might be defined as "that which is expected to maximize goal achievement, given the available information." Unlike the previous approach (thinking rationally) the process of "acting" rationally doesn't necessarily require "thinking." Thus blinking would be included in the idea of "acting rationally."
The AI community typically refers to things which act rationally as "agents." Strictly speaking, an agent is simply something that simply "acts." However, we will consider true intelligent, rational, agents as entities which not only act on our behalf, but do so by understanding their own environment (perception), perform over a prolonged period of time (persistence), adjust to changes in their environment and/or new goals (adaptability), and do so without a need for direct intervention (autonomous control).
Thus, for a given class of environments and tasks, we seek to find the agent (abstractly speaking, a function) which performs best when asked to convert a percept or percept history into actions.
Finally, notice that I am not asking for perfection here. Rather we are willing to accept that there are physical and computation limits that will prevent us from ever achieving perfect rationality. Thus, we are willing to settle for the best design given machine resources. We will talk much more about this concept of limited rationality as the course goes on.