By Tim Oriani, Regis '12
Imagine you and an accomplice, while attempting a bank
robbery, have been apprehended by the authorities. However, due to legal technicalities,
the courts can only bring petty firearms possession charges, unless one of you confesses.
.JPG)
So each of you is offered two choices: Cooperate or Defect,
i.e. confess or remain silent. Should one of you confess and the other remain
silent, the Cooperator would be set free, and the confession would be used to
sentence the Defector for a long period of time. In the case that both you and
your accomplice choose to confess, both will be sentenced, but also released
early for cooperating. If both remain silent, however, the firearms charge will
put both of you in jail for a very short period of time.
The two of you are kept in separate rooms, unable to
communicate. You must make your decision alone before finding out the results.
Thus was the so called “Prisoner’s Dilemma” presented to
the Senior Seminar by Mark Joinnides, Regis ’03 and a bioengineering
consultant.
Prior to his visit, he had directed us to an online program
in which we competed against an AI for several rounds, choosing in each round to
Cooperate or Defect. We gained points
according to the decision of both parties. Each party attempted to learn the
strategy and techniques of the opponent and adapt his own to win.
For our own minds, this does not seem so daunting a task.
But what about a computer? How does a man-made program, set in its conditions
by the unbreakable rules and restrictions of a computer code, keep up with the
flexibility of an organic mind such as our own? These questions led the Seminar
and Mr. Joinnides to explore the nature of intelligence, trying to discern the
line between fluid thinking and just following rules.
Mr. Joinnides explained the situation from the computer’s
perspective as a two-by-two matrix, because the Prisoners’ Dilemma has two players each capable of two different
choices, and therefore four possible outcomes. Seems simple enough for a
computer.
But now imagine a computer playing chess against a human.
There are twenty possible opening moves in chess. The computer must then be
programmed to incorporate a twenty-by-twenty matrix.
Moreover, unlike the Prisoner’s Dilemma, a game of chess constantly changes move by move.
Every one-by-one unit of the opening matrix leads to another matrix, possibly
much larger than twenty-by-twenty. And every unit of each one of those matrices
leads to another one. And another and another, until a game reaches its
conclusion. But that’s not
intelligence is it? I imagine it as a building, albeit an unimaginably enormous
one, that a computer program travels through, floor by floor, turn by turn, to
reach one of many destinations, hopefully a state of victory.
But what if this building could change and evolve? What
if its floors and pathways changed every time you played against it? Certain
pathways grow stronger, levels rearrange themselves, and the ultimate outcome
becomes more assured as the computer grows more and more accustomed to the
human player’s game strategy.
A program instructed to improve and grow better and
better at its specific task, is that intelligence? As our Seminar discussion
progressed, we searched for an answer, but couldn’t discover one that was clear-cut. As it turns out, it’s not just a question of when
computers gain intelligence, that quality that so defines our species as human
and superior to all others, but of when the human race will be ready to accept
the existence of an intelligent computer.
As society integrates Artificial Intelligence more and
more closely with everyday life, from the genetic algorithm that runs Pandora
to AI robots performing surgery or carrying out military orders, issue of accountability
grow more urgent. Who will be responsible when an Artificially Intelligent
computer makes a mistake, possibly in a life or death situation?
No comments:
Post a Comment