Artificial Intelligence
The Beginnings
In the early Fifties, computer scientists debated the philosophical question—can we say that computers think, are they intelligent? What is an objective scientific test for artificial intelligence (AI) ?
Alan Turing, an English computer scientist, gave us the test for AI that we call the Turing Test for AI, which we still use today. We have not achieved it yet. The Turing Test says:
Put a person in a room with a computer terminal. Let the person type in any question about anything, except mathematical calculation questions that clearly a computer can easily do very much faster than a human. Wait a few minutes and then let the answer appear on the screen. The person then has to guess whether the answer came from a computer or a human being. If the guess is right half the time and wrong half the time, then we cannot tell the difference between a computer’s thinking versus a human’s thinking. We can then say that we have achieved AI.
Alan Turing (BA,
After the war Turing was active in the new field of Computer
Science and Engineering, especially in
Work has been continuous to achieve AI, with major work being done at Harvard, MIT, Carnegie, Berkeley, and Stanford.
Some of the major
problems to overcome
For any computer to exhibit AI, it will have to understand a natural language, like English, just as we do. The problem is that language is learned by watching what people do when people speak, just like you did when you were an infant. For example, when you left the dinner table with your glass of milk and your mother said, “Johnnie, put down your milk, I don’t want you to spill it on my rug.” When your infant brother/sister saw you put your glass of milk down on the table, they got the connection between the words—put down, glass, milk, and Johnnie. Later the concept of rug was learned and the child put it all together. We learn by watching and hearing, with much repetition over long periods of time.
What you know about English took you 18 years to learn. How can a computer do this? Presumably it will have to have a TV camera for an eye, and a microphone for ears. But how do you store the information in the computer’s memory and make any sense of it? These are not simple problems to solve.
Computers can do deductive reasoning, like arithmetic, very quickly, because they “know the rules” of mathematics and are told what steps to take to solve a problem. Solving problems using rules is easy, if you know the rules. Unfortunately we don’t know the rules for most problems in life, so we have to guess what they are. If we guess badly the results can be deadly, like, is that snake in my path venomous, or harmless? People like myself follow the rule “all snakes are dangerous and need to be avoided”. This rule works perfectly, all the time. But we can’t do that with all of life’s problems.
For example, how do you teach a child to safely cross the street? “Look both ways” is a start. But what are we looking for? How do you judge the speed of a vehicle—by repeated observations! After enough experience you learn how long it will take a car to get from the corner to your house and whether you have enough time to get across the street without colliding with the car.
How do you learn to catch a ball thrown to you? You fail many times until you finally learn to estimate where to put you hand to catch the ball and curl your fingers around it. How does a baseball outfielder know where to run to and how fast in order to be there to catch a fly ball hit from the batter at home plate? The answer—practice, practice, practice. Your brain does not have a rule for this, but with sufficient practice it just learns how to do it.
If we knew how the brain does these things we could save a lot of time in programming the solution, but we don’t. The problem of guessing accurately at the rules is called heuristics. The more accurate you guess the rules the more intelligent you will be, and so will a computer exhibit AI.
Another problem with acquiring a natural language is that you can say things that use simple words but they don’t make sense as used. We often call this figurative language, like “I am half dead”, meaning I am very very tired. How will a computer learn the many nuances of our language? How about the movie title “Dead Man Walking”? How will that make sense to a computer without knowing the story?
One advantage a computer will have once it does understand language is that it would be able to read books, magazines, newspapers, scientific articles, etc. at vastly faster speeds than we can. Therefore knowledge will be acquired much faster than we can. Computers would also be able to “watch” movies or TV shows much faster than we. This will speed up the learning of language, because so much of our language is nuanced by body language and inflection which you do not get by reading alone. Therefore the speed of learning, once the trick is figured out, will go much faster.
Another aspect that would be considered good news is that once one computer has achieved AI and a great level of knowledge about our world, it can be reproduced many many times just like we make computers or programs today. Think about it, when we humans reproduce, how much of your parents’ knowledge do you have at birth? None, absolutely none. You have the potential to learn, but you have to go to school for 18 years just to get through High School. But computers and memory can be reproduced easily! This is good news if we ever achieve true AI. My guess is that it is a long way off unless someone makes a significant breakthrough on how to do these things.
Another Direction in AI
In the 1980s,
computer scientists began thinking that while we are waiting for major
breakthroughs in AI to occur, we can do some
interesting things with computers that mimic experts in various fields. The idea, known as Expert Systems, means
creating a program, known as a shell, that is capable of holding a database of
rules, rules that pertain to a particular area of human endeavor, like
medicine. “Knowledge Engineers” then
take the rules of the expert and put them into the shell. When the shell is complete, the program can
then act as if it were an expert in that field, telling you what to do under
circumstances that you feed into the system.
One of the pioneers
in AI and especially expert systems is Edward Feigenbaum (PhD Carnegie Mellon). He has taught at Carnegie, Berkeley, and
Stanford. Feigenbaum won the Turing
Award in 1994 for his pioneering work in AI.
Another name for expert
systems is Knowledge Based Systems=Entering the rules known by experts
related to a useful knowledge area into a database so that the computer can
answer questions like an expert in that area of knowledge.
Besides medicine,
other fields where expert systems have been employed are oil exploration and
examining parts lists for newly ordered computer systems (by DEC).