On Self–Enforcement in Extensive–Form Games

A new criterion for self-enforcement is developed which requires that every player's strategy should be at least locally ptimal at each of his information sets under some probabilistic hypothesis about all player's behavior which at least approximately agrees with the tested profile concerning behavior at all points in the game tree where his information set provides no evidence of a deviation. Like sequential equilibrium, this criterion is more stringent than subgame perfection but less stringent than extensive-form perfect equilibrium. The approach builds on an individualistic and set-theoretic consistency condition which is an application of Occam's razor.