“You look for symmetries. You learn to think analytically about the game, and you try to narrow it down to logical rules,” says Jason Laster ’10.

The student is working on an assignment for “Artificial Intelligence,” a computer science (CS) course that introduces principles of machine thinking. Laster has to write an algorithm that can play the strategy game Connect Four. When he succeeds in producing a computer program he cannot beat, he says it feels as if he has created “a sapient being” that is, in one narrow way, smarter than he is.

When the term artificial intelligence (AI) was first introduced to the world at a Dartmouth conference in the summer of 1956, the conference organizers stated the goal of AI research as “making a machine behave in ways that would be called intelligent if a human were so behaving.”

But how does one recognize intelligence? Six years earlier the British mathematician and computer scientist Alan Turing had proposed a test to determine machine intelligence. He suggested that a machine can be considered intelligent if, in a five-minute conversation, it can fool people into believing they are interacting with a fellow human being. He expected by around the year 2000 scientists would have developed a machine that could pass his intelligence test.

The idea of machines with human-like intelligence has taken hold of popular imagination and inspired science fiction literature and movies that depict a world in which computers think and communicate as human beings do, ultimately surpassing human intelligence and rendering humanity obsolete.

Devin Balkcom, who rotates with other CS faculty in teaching “Artificial Intelligence,” laughs away such popular ideas about AI as “100 percent misconception.” He says he doesn’t foresee any machines with human-like intelligence becoming a reality within the next few thousand years. In fact, says Balkcom, we don’t even know what intelligence is.

Since ancient times humans have tried to understand their own minds. Aristotle articulated the first laws of logic 2,500 years ago, and we now have powerful logical systems that allow for the representation of complex problems we can process in superfast computers. But we still don’t understand how we think or how to build machines that approximate human thought, he says.

“One of the difficulties with AI,” says Balkcom, “is that we get very good at writing the algorithms, but we then find they can’t be solved within a reasonable time.” For example, it isn’t very hard to write an algorithm that plays unbeatable chess, but sufficient computer power may not exist to execute it in a timely fashion. AI researchers in the 1950s believed that the chess problem could easily be solved in less than a decade. Programmers still haven’t been able to create a chess computer that never loses a game to a human chess champion.

According to Balkcom, most computer scientists are now skeptical about the predictions made 50 years ago. He says that what is commonly considered AI research is essentially the ambition to make computers do things—such as reasoning, natural language processing and vision recognition—human beings are good at. But once a computer can do something, he says, we tend to no longer consider it intelligence. A regular calculator would certainly have been considered AI 100 years ago, says Balkcom, but now that calculators are part of our everyday life, they no longer seem very smart.

Balkcom doesn’t try to define intelligence. He just has his students focus on what computer scientists do best: creating algorithms to solve logical problems. Although most of his assignments are toy problems—writing programs to play games such as tic-tac-toe, Connect Four, chess and Sudoku—students can apply the skills they learn to real-world logical problems.

Laster, who majors in mathematics modified with CS, thinks everyone should have some background in computer logic. “Whether you end up working for a Fortune 500 company or for the government, there’s no way you should enter the 21st century without knowing how to use computers and think computationally,” he says. “AI helps you think logically and ask the right questions.”

Since “Artificial Intelligence” is an advanced CS course, all students enrolled must have completed a basic algorithms courses and know how to program. The course is offered at an undergraduate as well as a graduate level, and about half of the 18 students are in the master’s or Ph.D. program. Balkcom says the undergraduate students are so motivated they have no trouble keeping up with the graduate students, who tend not to specialize in AI.

Aarathi Prasad, for example, a graduate student from India, hopes that a better understanding of AI will advance her research in wireless sensor networks designed to interpret medical data and monitor hospital patients.

Balkcom’s background is in robotics. In middle and high school he wrote software for fun. When he was in college at Johns Hopkins he figured he might as well make his hobby his profession. He eventually completed a Ph.D. in robotics at Carnegie Mellon University.

“I love machines and what I can make them do,” says Balkcom. He sees computer programming as an art that everyone can learn. “There’s an infinite number of problems, and they become infinitely hard. But at a basic level anyone can program something. It’s this basic set of clean tools with which you can build really complicated, beautiful things,” he says.

“We humans have a very expressive way of representing knowledge,” says Balkcom. “We can phrase almost any question and anything we know about the world in a natural language such as English. It’s much harder to write an algorithm to represent it.”

The problem is that to write something as an algorithm, you first need to understand it perfectly. And in some ways it is also a matter of knowing what to include. One research consortium in Texas, for example, has been trying to encode all human knowledge in first-order logic. “But nobody knows what to do with it,” says Balkcom. “We can encode a lot of things and solve a lot of individual problems, but we don’t know which ones matter and how to hook them together.”

In a way, almost all current CS research relates to problems that can be considered AI. Balkcom goes so far as to assert that AI is actually a marketing term to make computer science seem more appealing to the popular imagination. He is not even committed to the title of the course that he teaches.

“ ‘Artificial Intelligence’ is just a fancy name given to an advanced algorithm class,” says Balkcom. “It could have been called ‘Algorithms-3.’ ”

Even the stated goal of AI research—trying to emulate human intelligence—may actually be misguided, because the human way is not necessarily the best or most efficient way to approach problems, he says. As part of his Ph.D. research Balkcom analyzed how to design robots that can tie knots. The first robots were built to replicate the human way of tying knots, which is, of course, determined by human physiology.

“Well,” says Balkcom, “it turns out that tying a knot isn’t so hard and that humans don’t do it the most efficient way, just the way that’s most convenient for them.”

SUGGESTED READING

Professor Balkcom recommends the following sources for more information on A.I.:

Artificial Intelligence, A Modern Approach by Stuart Russell and Peter Norvig (Prentice Hall, third edition, 2009) is the course textbook

“Computing Machinery and Intelligence,” a 1950 article by Alan Turing in which he proposes a test to determine whether a machine is able to think

“Programming a Computer for Playing Chess” by Claude E. Shannon is a groundbreaking paper that appeared in the March 1950 issue of Philosophical Magazine

Behind Deep Blue: Building the Computer that Defeated the World Chess Champion by Feng-Hsiung Hsu (Princeton University Press, 2004)

“Programs with Common Sense,” a 1959 article by John McCarthy

AI Magazine is a good source for those grounded in CS

Judith Hertog lives in Norwich, Vermont. She is a frequent contributor to DAM.