主题：【转载】讨论分析博弈论经典游戏--数字游戏 -- alphaboy

My prediction is that this trend will be reversed in favor of an approach in

which the degree of rationality bestowed to the agents depends on the context

being studied. To illustrate how this can work in practice, consider the “guess the

number” game first studied by Rosemarie Nagel (1995). In this game, contestants

are told to guess a number from 0 to 100, with the goal of making their guess as

close as possible to two-thirds of the average guess. In a world where all the players

are known to be fully rational, in the sense that they will form expectations about

the guesses of others can carry out as many levels of deduction as necessary, the

equilibrium in this game is zero.

In any other setting, however, guessing zero is not a good strategy. Recently, I

had the opportunity to play this game for quite large stakes (Thaler, 1997). At my

request, the Financial Times ran a “guess the number” game contest using the rules

described above and offered two business class tickets from London to the United

States as a prize (worth over $10,000). Only integer guesses were permitted.

Although many contestants did guess zero or one, the most popular guesses were

33 (the right guess if everyone else chooses a number at random) and 22 (the right

guess if everyone else picks 33). The average guess was 18.91 and thus the winning

guess was 13. Although modeling how this game is actually played is not easy, some

lessons are clear enough. An appropriate model would have to allow for two kinds

of heterogeneity in sophistication. First, agents differ in how many levels of pro-

cessing they engage in (33 is one level, 22 is two levels, and so on). Second, there

is heterogeneity in how much agents think about the behavior of other agents.

Agents who guess zero are sophisticated on the first dimension and na ve on the

second. Many economists fall into this category (due in part to the False Consensus

Effect and the Curse of Knowledge!) Sophisticated economic models will have

agents that are both more and less sophisticated than the agents we are used to

modeling. I predict this sort of modeling will be the norm in the future.