Distribution-Free Learning

Abstract

We select among rules for learning which of two actions in a stationary decision problem achieves a higher expected payoffs when payoffs realized by both actions are known in previous instances. Only a bounded set containing all possible payoffs is known. Rules are evaluated using maximum risk with maximin utility, minimax regret, com- petitive ratio and selection procedures being special cases. A randomized variant of fictitious play attains minimax risk for all risk functions with ex-ante expected payoffs increasing in the number of observations. Fictitious play itself has neither of these two properties. Tight bounds on maximal regret and probability of selecting the best action are included