ML Wiki
Machine Learning Wiki - A collection of ML concepts, algorithms, and resources.

Prisoner's Dilemma

Prisoner’s Dilemma

This is a game of Game Theory

The set up:

  • There are two players $p_1$ and $p_2$
  • both have committed a crime
  • they are caught by the police and kept in separate rooms - so they cannot communicate with each other
  • they have two options: cooperate with the police and confess ($C$), or don’t cooperate ($N$)
  • for not cooperating: 1 year in prison
  • for cooperating: the other goes to jail on 10 years and you are set free
  • but if both cooperate, each gets 5 years in prison

This can be depicted by a matrix:

| | $N$ | $C$ | $N$ | (-1, -1) | (-10, 0) || $C$ | (0, -10) | (-5, -5) | The best strategy - the Nash Equilibria is:

  • for $p_2$: $C$ is always better: (0 > -5)
  • for $p_1$: the same
  • so both choose to play $(C, C)$ - the strictly dominating strategy
  • but this strategy clearly is not better than $(N, N)$ - but the Dominance principle misses it
  • which is why it’s called a Game Theory paradox

Variations

TCP blackoff

Set up

  • when TCP correctly implemented, it has “backoff mechanism”
  • if data flow causes congestion, sender reduces speed, until a jam subsides
  • defective implementation doesn’t backoff

There are two strategies

  • $C$ to use correct implementation
  • $D$ to use defective one
  • if both players use $C$, delay is 1
  • if $p_1$ uses $D$, and $p_2$ uses $C$, then $p_1$ has no delays, but $p_2$ has 4ms of delays
  • both players want to minimize the delay time

| | $C$ | $D$ | $C$ | (-1, -1) | (0, -4) || $D$ | (-4, 0) | (-3, -3) | For both players the dominating strategy is $D$

Tickets Price

  • Suppose we have two airline companies $P_1$ and $P_2$
  • They are both thinking about about opening a new destination
  • Both consider two options: either make tickets cheap or make them expensive
  • Clearly if $p_1$ decides to sell cheap tickets while $p_2$ - to sell expensive tickets, everybody will buy from $p_1$

So we can depict it with the following pay-off matrix

  • a cell represents consequences of the decision that both players take

| $p_2 \to$
$p_1 \downarrow$ | 200 | 500 | 200 | (50, 100) | (-100, 200) || 500 | (150, -200) | (-10, -10) | We see that:

  • if both agree on cheap tickets - both will have profits
  • if $p_2$ sells expensive tickets and $p_1$ cheap ones, all go to $p_1$ and $p_2$ will have losses
  • the same with $p_1$ and $p_1$
  • if both decide on expensive tickets - nobody will buy them and they both will experience losses

The profile (500, 500) is the Nash Equilibria

Sources