The Prisoner’s Dilemma
Wednesday, February 28th, 2007Via Prospero’s Books, I found this article about robots being used to simulate evolution. I’ve read about similar projects simulating evolution through competing artificial intelligence programs, using the “Prisoner’s Dilemma” scenario as the competitive task. The Prisoner’s Dilemma, for those who are unfamiliar, breaks down as some variation of this:
You and a partner are both correctly arrested for two crimes, one major and one minor, and are put in separate rooms. Executive Assistant District Attorney Jack McCoy comes to visit you and offers you a deal: testify against your partner for the major crime, your partner will get twenty years, and you’ll walk for both crimes. However, his lovely assistant is right now offering the same deal to your partner. If you both confess, you’ll both get five years. If your partner confesses and you don’t, you’ll get the twenty, and he’ll walk. If neither of you confess, McCoy can’t make his case for the major crime, but he’ll make sure you both do two years for the minor one. What’s the right play?
Well, logically speaking, regardless of what your partner ends up doing, you’re better off confessing. But if you both confess, you both end up worse off than if you had both kept your mouths shut. If you had had the chance to communicate with each other, you might have chosen differently. The fact that you don’t know what your idiot partner is going to do while gazing into the eyes of the lovely ADA means that you can’t afford to take any chances, and neither can he. You both end up doing the nickel, even though neither of you had to.
In this example, you only get to play the game once. If you play some version of the Prisoner’s Dilemma with the same person repeatedly, your choices can affect future outcomes. In a sense, the choices you make are a form of communication. Only the very last time you play do you revert back to the original cutthroat scenario. (And since everybody knows this will be the case, the next-to-last iteration can also be cutthroat. How far back does this reasoning work?) There is actually a twenty-year-old Iterated Prisoner’s Dilemma competition for artificial intellegence programs and the winning strategy has long been the simple Tit-for-Tat. But it seems there’s now a new champion, though it seems to me to be a bit of a cheat. Read the article and let me know what you think.
The Prisoner’s Dilemma is an illustration of one of the central concepts of a branch of mathematics called “game theory.” Game theory allows us to make mathematical computations in decision making, even when all of the factors are not known. Think of two generals, one trying to choose a target to attack, the other deciding how to deploy defensive forces. Each knows the other is intelligent and out there making his decision. That’s game theory. If you were to meet someone anywhere in the world outside of the United States, but you couldn’t plan with that person ahead of time, where would you go? Would it surprise you to learn that almost everyone makes the same choice? (Post your answer in the comments section, if you like.) That’s game theory too.
With a branch of mathematics that can take unknown variables into account, a computer’s functionality can be increased significantly. Obviously computers that are powerful enough can play chess, but game theory allows them to play poker as well. There’s already a Texas Hold ‘Em Tournament for Artificial Intelligence programs. Imagine putting all of these programs into a giant simulated Texas Hold ‘Em Tournament where the losing programs died out and the winning programs created offspring with the possibility of mutation. We might evolve the ultimate strategy. And when we do, the first round of drinks are on me!
But as computers get more powerful, imagine other simulations we may be able to run, and what understandings we might be able to gain from these experiments. Evolution has proved itself to be a mighty force in the past. Once all of the data from Web 2.0 is compiled, maybe it will be allowed to evolve into Web 3.0. It’s not about computers becoming super-sentient and ruling over humans. It’s about humans developing and using new tools that can increase our capacity for growth. And if evolution has taught us nothing else, it has taught us that.