Conditional Universal Consistency

Abstract

Players choose an action before learning an outcome chosen according to an unknown and history-dependent stochastic rule. Procedures that categorize outcomes, and use a randomized variation on fictitious play within each category are studied. These procedures are “conditionally consistent:” they yield almost as high a time-average payoff as if the player knew the conditional distributions of actions given categories. Moreover, given any alternative procedure, there is a conditionally consistent procedure whose performance is no more than epsilon worse regardless of the discount factor. We also discuss cycles, and argue that the time-average of play should resemble a correlated equilibrium.