Good morning. It’s really a great honor to be invited to this exciting forum, and let me get up to thank the organizers for including me in this event.
I’d like to talk about the connection between economics and computer science, between mechanism design and artificial intelligence this morning. Normally, economics and computer science are taught separately, in different departments. It’s possible that a student in one major can learn a little bit at the other. But I think the two subjects share so much in common that they really ought to be talked together and they should be talked together a lot more often.
And to take it a little strict, I want to give an example from my own area of economics, which is called Mechanism Design. Mechanism Design is the engineering part of economic theory. If you know a little bit about economics, you know that most of it is devoted to trying to understand existing economic institutions and trying to predict the outcomes that the institutions generate. But in mechanism design, we do just the opposite. We reverse the direction. We start with the outcomes. We say these are the outcomes one achieve, and then we work backwards to figure out what institutions and mechanism that we can design, which will deliver those happenings.
Let me give you an example of the sort of the mechanism I’m talking about. This is an important example from the last 20 years or so.
Imagine there is a government who wants to transfer the right of broadcasts on particular band of radio frequencies to one of several telecom companies who are interested in this trying. We’ll call the right to use this particular band radio frequencies a license. So you want to transfer the license to one of these telecom companies that are interested in it. And you want to give it to the company that has the biggest value for this license. The problem is, you do not know, as a government, which telecom company actually does value. And so the question is how you make sure that the right company gets the license.
The simplest thing to do is just asking every company how much you value the license. But this is not likely to work because each company has the storming center to exaggerate. Because it understands that by exaggerating, it’s going to increase the likelihood to get the license. So simply asking how much they value the license is not going to work. You can try something a little bit more sophisticated like how each company makes a bid for the license. If it is a statement that how much you’re willing to pay. You can award this license to the company that makes highest bid. And have this high bidder actually pay its bid. This is better than the first mechanism, because now companies want to exaggerate no more. But this is not going to work either.
The reason that is likely to work is that now companies have the incentive to understate their values. Imagine that the license is worth a million to you, if you bid a million and you win, you get something worth a million. But you pay a million, so your net benefit is zero. That means you’re not going to bid a million, you are going to bid something less than a million. And you are going to understate. The problem now is that all the companies are understating, there is no guarantee that the winner will be the company that really does value the license the most. So what is the solution?
The solution is actually very ingenious. It was discovered in 1960s by an American economist. It is very simple but very clever. The idea is that every company is to bid licenses. The winner is the high bidder. So this is the same as before. But now instead of paying its own bid, the winner pays the second placed bid. So this is sometimes called the second-step mechanism.
So for example of the 3 bidders. One bids 10 million, one bids 8 million, and one bids 5 million. The winner is the company that bids 10 million. Because it is the highest. But it pays only 8 million because that’s the second highest. If you think about it, you’ll see that now all companies have no incentive to understate because you don’t pay your bid anyway. If you understate, you’re not going to reduce the amount you pay because that’s determined by another company. You don’t want to understate. If you do understate, you run the risk to lose the license all together. You also don’t want to overstate because if you overstate it, if it is worth 10 million to you, you bid 12 million. If some other company bid 11 million, it’s truly you win, but you have to pay 11 million which is too much. So by overstating you run the risk of paying too much. So you’ll bid exactly what the license is worth for you. It means that the company that really does value the license most will win. You can use the example of elegant solution which is used a lot in practice. Now I describe how this works. While there is only a single license.
It turns out that the same mechanism works with multiple licenses. Now if there are many companies and many licenses, you can have a company make a bid for every license. You can award the licenses so it can maximize the value of the state it is. That could be the efficient thing to do. And it turns out that you can get companies to bid their true values in the way which is with complete elegance to the one license case. What you do is you get up the fact how the company pay its marginal effect on other companies you can vote and how much other companies will be getting if the company questions were excluded, you’ll get what happens to the company with questions that were not excluded, and the differences between the amounts of the two company pays. And it turns out that if it pays this amount, it does have incentive to bid its true value.
Just as in the one license case. The problem with this mechanism and the name of the mechanism is called “broke mechanism”. The problem with this mechanism is that first it’s very compensatory for the bidder itself, because a company has to make a bit on every license and there are lots of licenses. That is going to be a very important pass for them, perhaps even more important. It's a computationally complicated problem for the government itself. Because now it has to figure out which assignment of licenses maximize themselves. And if there are a lot of licenses and a lot of companies. That’s a computationally complex thing to do and it is the end of the hard problem. It turns out though that we can take an idea from artificial intelligence and computer science how is the process solving and drastically simplify the problems.
And let me explain how we can do that. Now every company makes a deal on every license. We run a process in continuing this commerce. We set the price of each license equal to zero. Rather than they can use these things, each company just chooses the license at any given points that it prefer to get its own prices. If the license is just demanded by one company. Each license is just demanded by one company, then supplying equal demand we’re done. But if there is excess demands for some licenses, then we simply raise the price for that license, we are demanding this way. We keep irritating this process. So every license is demanded by just one company. So it is very simple for the companies. They just have to choose which license they like to have. And it’s very simple for the governments, because it doesn’t have to do any complication at all. All it has to do is conduct the prices to the responses they get. Yet it works, it gets to the same answer to a more complicated question, and more complicated the auction does. So it’s brilliant with simple solution, and it is now used by many governments in the world to auction licenses.
I name this point because it shows that economics and computer science are really very close. Here we use the idea in parallel thoughts, and it’s an idea that is developed in computer science to greet the facts in economics.
Unless I come to know about the processes, I would use the thought that is used in the first place. So the two subjects really deserves to be talked together and let me urge that the universities around the world perhaps can consider that combining economics and computer science curriculum in this light. Thank you very much.