Friday, February 20, 2015

Prisoner's Dilemma

Two prisons accused on a crime are kept in separate confinement.  If A testifies against B and B keep silent, B would get a life sentence and vice versa.  If both testify against each other, both would get 10 years of imprisonment.  If both keep quiet silent, both would be charged with a lesser crime and get 5 years of imprisonment.  What would you do?

The prisoner's dilemma was postulated in 1950s during cold war to study how to achieve international cooperation.  Traditionally, most people believe one can only achieve cooperation unless there is a central effective governing bodies.  That is the backdrop of UN.  The prisoner's dilemma is to study how to cultivate cooperation without that.

Most people tend to choose defection (testify against).  However, if the game is play multiple rounds. the behavior will be affected by past result.  Different algorithms are used to simulate the game play for the best result.  And the best algorithm is called TIT for TAT.  In other words, the player will go for cooperation but if the other party defects, the player will defect to punish the other parties until the other parties cooperate again.  A good algorithm generally is nice and forgiving but also retaliatory - punish an opponent immediately.

One extension to the game is to add social norm.  Each player has a specific value of defect (boldness) and probability to punish a defective (vengefulness).  When a player defect and is witnessed by some other players, he may be punished.  If the vengefulness value is low (no social norm), defection will dominate tje community.  However, vengefulness is not sufficient to maintain cooperation reliably.  Adding punisher to punish player that does not punish defector will on the other hand promote cooperation.  The punisher is playing the role of metanorm.  For example, people giving leering look to parents that does not control their children.

Another extension is to add spatial consideration in the game.  Each player (in a grid) only interact with players around it.  Simulation found that cooperation and defect exist indefinitely without the need for norm or metanorm.

No comments: