July 25, 2000
It's Only Checkers, but the Computer Taught Itself (right..)
By JAMES GLANZ
|
|
Robert Burroughs for The New York Times
|
|
| Peter Angeline, left, played checkers against a
self-educated computer program last week at the 2000 Congress on Evolutionary Computation
in San Diego. Scientists are debating just how smart the program created by Dr. David
Fogel, right, can teach itself to be. |
Forum
• Science/Health
Forum
• Join a Discussion on
Artifical Life
|
|
ver since an I.B.M. computer defeated
Garry Kasparov, the world chess champion, in 1997, much of the suspense has gone out of
the old battle of man versus machine: the machine won. But the lessons of a much humbler
game have again made that battle fascinating. The game is, of all things, checkers.
Two computer scientists have leveled the playing field by asking a computer program
called a neural network to do something much more difficult than beat a defenseless human
at checkers. Knowing only the rules of checkers and a few basics, and otherwise starting
from scratch, the program must teach itself how to play a good game without help from the
outside world -- including from the programmers.
The program did just that, using the electronic equivalent of natural selection. A sort
of colony of the programs, each slightly different from all the others, played checkers
against one another -- quite ineptly at first -- and chronic losers were killed off.
Slightly mutated versions of the winners were allowed to reproduce. After hundreds of
generations, blind evolution produced an expert checkers player, but not an infallible
one.
The work was described last week in San Diego at the 2000 Congress on Evolutionary
Computation by Dr. David B. Fogel, a computer scientist at Natural Selection Inc., a San
Diego company.
The importance of the work, which Dr. Fogel did with Kumar Chellapilla, a doctoral
student in electrical engineering at the University of California at San Diego, goes well
beyond checkers to a question that has dogged computer science for decades: Can computers
be designed to solve problems under circumstances that their human programmers never
envisioned?
The question is relevant for circumstances as varied as an interstellar probe that
loses contact with Earth and must adapt to unforeseen problems, or an automated factory
scheduler attempting to deal with unexpected shortages or breakdowns.
All this from a good, but not perfect, game of checkers.
"The point here is that it learns to play by itself," said Dr. Christopher
Welty, a professor of computer science at Vassar College.
Because any particular version of Deep Blue, the I.B.M. program that defeated Mr.
Kasparov, was based on strict rules and specific strategies given to it by humans, Dr.
Welty said, "the way it played chess was very predictable -- given the exact same
game, it would play the same way twice." But Dr. Fogel's program can adapt without
intervention by humans, he said.
The program made short work of two dozen casual players at last week's conference,
winning 25 timed "challenge matches," losing none, and battling one player to a
draw.
Whether Dr. Fogel's best checkers program can improve beyond its current skill level is
uncertain; it can hold its own with expert-rated players in official rankings but is
generally blown off the board by players in the higher echelons of master and grand master
players. The question of just how far such an adaptive algorithm can progress is a matter
of intense debate among computer scientists, Dr. Welty said.
But Dr. Ali Zalzala, a computer scientist at Heriot-Watt University in Edinburgh and
the principal organizer of the conference, said that just by taking on such questions, Dr.
Fogel's work may be shedding light on much deeper issues. Dr. Zalzala called the research
"a very serious attempt to understand the operation of biological evolution."
The neural networks that Dr. Fogel has bred into checkers players exist as software
programs on his personal computer. To understand them, however, it helps to visualize the
physical structures that the software is modeling. Those structures consist of a rather
crude representation of the interconnected networks of neurons in the brain.
Living neurons are sometimes called "integrate-and-fire" structures. A neuron
somehow integrates, or adds up, all the electrical stimuli it receives from other neurons
that are connected to its receptors. If the sum rises above a certain threshold, it fires
off its own electrical pulse, stimulating yet another set of neurons that behave in the
same way.
The threshold for firing can be different for each neuron; so can its own firing
intensity compared to the amount of input stimuli it receives, an amplification factor
sometimes called the connection strength. At one end of the network might be sensory input
from a pair of ears, for example, and at the other a conclusion like, "That is Bob's
voice."
Any small part of the network might have connection strengths and thresholds to
recognize one particular feature, like average pitch or the way a single vowel is
pronounced. Because all parts of the network are interconnected, it can put all of these
clues together almost simultaneously to draw a conclusion about whose voice is being
heard.
Humans learn; no one is born able to recognize Bob's voice. Likewise, the thresholds
and connection strengths in Dr. Fogel's network, which are geared toward recognizing
sequences of moves in checkers rather than voices, can change as the program learns.
"As a human being, you are a pattern-recognizing device," said Dr. Fogel, who
added that his simple network could do the same thing.
But he did not give away the game by telling the network how. The network begins with
32 "sensory" inputs, one for each possible position on the board. Those inputs
take on different values depending on whether a square is occupied by nothing, by an
ordinary red or black piece, or by a king.
"Neurons" carrying those signals cross in a complex pattern of nested
connections, reminiscent of a switchboard. Each neuron has a variable connection strength
and each connection has a threshold. For a given configuration of checkers, the output of
the network is a single number.
That number is in some way the network's evaluation of the pattern on the board, a sort
of output of pleasure or pain. At each point of the game, the computer evaluates it for
the existing configuration and for possible ones it could create by making certain moves.
Based on that information, it maximizes the number and makes moves, plays games.
Dr. Fogel began by creating several variants of his program, each with random
connection strengths and thresholds. They played each other, and the regular losers --
those for whom "pleasure" was not a reliable guide to playing a good game --
were eliminated. The others, with random "mutations" that could allow better
players to evolve, were given offspring. Eventually, he arrived at an expert player, a
fact he verified by matching the program against ranked human players.
"The performance is remarkable," said Dr. Michael Conrad, a professor of
computer science at Wayne State University. "I'm still sort of surprised how good it
is."
Aside from getting beat by some humans, however, the program is not in the same league
as Chinook, the computer checkers player devised in the early 1990's by Dr. Jonathan
Schaeffer of the University of Alberta in Canada. Chinook does not have a human equal at
any level, Dr. Schaeffer said. But he said his program, like Big Blue, was infused with
human expertise from the beginning rather than learning on its own.
How good Dr. Fogel's program might someday become, and whether it could ever approach
something like Chinook, is unknown. But Dr. Silvano Colombano, a computer scientist at the
NASA Ames Research Center in California, said the same sort of evolutionary strategy could
be built into computers aboard spacecraft that range too far from Earth to stay in
contact.
"By analogy with the checkers game," Dr. Colombano said, "the computer
can play games with itself and become smarter."
One thing such a spacecraft would have plenty of is time. In the long night of space,
it might even decide to hone its checkers game.
