Skip to content
Tutorials Center
My learning

Topics

Lesson 4 of 37 ยท Getting Started

A First Game: The Beauty Contest

Play the classic guess two thirds of the average game, learn level-k thinking and iterated reasoning, and simulate how guesses shrink toward the equilibrium.

Let us finish the opening chapter with a game you can play with friends, in a classroom or in a simulation. It needs no payoff table. It is called the beauty contest (the name comes from a famous remark about newspaper contests) or the guess two thirds of the average game, and it shows in a few lines the main habit of mind in game theory: think about what the others are thinking.

The rules are simple. Every player secretly picks a whole number from 0 to 100. Once everyone has chosen, compute the average of all guesses and multiply it by 2/3. The player whose guess is closest to that target wins. Think for a moment about what you would pick before you read on.

Reasoning step by step

If you believe everyone else picks at random, the average is around 50, the target around 33, and so you should guess about 33. But if everybody reasons like you, the average will be about 33, so the target is about 22. And if everybody reasons that far, the target is about 15. Each round of reasoning shrinks the answer by a factor of 2/3.

Game theorists describe this as level-k thinking:

  • A level-0 player does not think. Often modeled as picking uniformly at random, with an average of 50.
  • A level-1 player assumes everyone else is level-0 and guesses 2/3 of 50, about 33.
  • A level-2 player assumes everyone else is level-1 and guesses 2/3 of 33, about 22.
  • A level-k player guesses 50 * (2/3)^k.
flowchart LR
    L0["Level 0<br/>random<br/>avg 50"] --> L1["Level 1<br/>guess 33"]
    L1 --> L2["Level 2<br/>guess 22"]
    L2 --> L3["Level 3<br/>guess 15"]
    L3 --> D["..."]
    D --> E["Limit<br/>0"]

Each extra step of reasoning multiplies the guess by 2/3.

Python Level-k guesses

Press Run (or Ctrl + Enter) to see the result here.

Program input ยท what you'd type while the program runs

Type the answers before you press Run, one per line. Each time the program reads a line (with input()), it gets the next line from this box.

The sequence marches toward zero. If every player is perfectly rational and knows that everyone is, then the only stable answer is 0: if everyone picks 0, the average is 0, two thirds of it is 0, and nobody can do better by changing their guess. That is the game's Nash equilibrium, an idea we define properly in Nash Equilibrium. The reasoning that produced it is iterated elimination of dominated guesses: no sensible player picks above 67 because the target can never exceed 67 (two thirds of 100), so guesses above 67 are out; then the target is at most 44, and so on. You will see this exact argument in Iterated Elimination of Dominated Strategies.

What real people do

If you run this game with real people, the winning guess is almost never 0. In many experiments the average guess is around 35 and the winning number is around 20 to 25. A few people choose 0, many choose 33 or 22, and some choose 50 or 67. Real groups contain a mixture of reasoning levels, with most people stopping at level 1 or 2. So the best guess is not the equilibrium. The best guess is a response to who you think is playing.

This is a lasting lesson: equilibrium predicts what happens when everyone is rational and expects the others to be, and it is a great benchmark. When the others are not like that, the best action is a best response to their actual behavior.

Python Simulating a mixed crowd

Press Run (or Ctrl + Enter) to see the result here.

Program input ยท what you'd type while the program runs

Type the answers before you press Run, one per line. Each time the program reads a line (with input()), it gets the next line from this box.

In this crowd, which is mostly level-1 and level-2 thinkers, level-2 players win most often and level-1 players never do. Level-3 players, who think the furthest, win only rarely. The winner is the one who is one step ahead of the crowd, not ten: thinking too far ahead can be as costly as not thinking.

Computing a best response to a belief

Suppose you hold a belief about the other players' guesses, say they come from a given list. Your best response is the guess nearest to two thirds of the average including your own guess. When there are many players, your own influence on the average is small; with few players it matters. Let us compute it exactly.

Python Best response in the beauty contest

Press Run (or Ctrl + Enter) to see the result here.

Program input ยท what you'd type while the program runs

Type the answers before you press Run, one per line. Each time the program reads a line (with input()), it gets the next line from this box.

When the group is large, the best reply to a crowd averaging 30 is about 20, which is 2/3 of 30. With only two players your guess pulls the average noticeably, so the best reply is smaller, because the target is partly made of your own guess. The unique stable state, where guesses equal two thirds of themselves, is again 0.

Equilibrium as a fixed point

There is a compact way to say all this. If every player picks the same number g, the average is g, so the target is (2/3)g. For g to be a winning guess we need g = (2/3)g, and the only solution is g = 0. In other words, equilibrium is a fixed point of the reasoning process: a state where nobody wants to change what they do given what the others do.

Python Iterating toward the fixed point

Press Run (or Ctrl + Enter) to see the result here.

Program input ยท what you'd type while the program runs

Type the answers before you press Run, one per line. Each time the program reads a line (with input()), it gets the next line from this box.

Starting from 100 it takes 12 steps of reasoning to land below 1 and about 23 steps to reach 0.01. This matches how humans behave: stopping after two or three steps is typical.

Variations to try

  • Change the factor from 2/3 to 1/2 or to 3/2. With 3/2 the fixed point is not unique and 100 also becomes an equilibrium; why?
  • Change the upper limit (0 to 1000). The equilibrium stays at 0.
  • Add a few "stubborn" players who always guess 50. How does the winning guess change?

Check your understanding

Quick check ยท 4 questions

of 4 answered You scored / 4 ยท perfect!

  1. 1. A level-1 player in the beauty contest assumes the others are level-0 with an average of 50. What do they guess?

    Level 1 best-responds to level 0: two thirds of 50 is about 33.
  2. 2. Why is 0 the Nash equilibrium of the game?

    At the all-zero profile each player's guess equals two thirds of the average, so no one can do better by deviating.
  3. 3. In a real classroom the winning guess is usually around 20. What does that tell us?

    Real players stop after a few levels of reasoning. Equilibrium is a benchmark; the best action against real behavior is a best response to the real crowd.
  4. 4. If the factor changes from 2/3 to 1/2, what is the level-k sequence starting at 50?

    Each step multiplies by the factor, here 1/2, so the guesses halve every time and still approach 0.

Summary

  • In the beauty contest each player guesses a number, and the winner is closest to two thirds of the average.
  • Level-k thinking models people as taking a limited number of reasoning steps, each multiplying the guess by 2/3.
  • The only guess that survives unlimited iterated reasoning, and the Nash equilibrium, is 0.
  • Against a real crowd, the best choice is a best response to their actual beliefs, which is usually well above 0.
  • Equilibrium is a fixed point: a state nobody wants to leave given what the others do.

Next: Normal-Form Games and Strategies