Font Size: 

We develop a new axiomatic framework to characterize the classical Hurwicz criterion. Our framework is simpler than other characterizations in the literature. We also study the learnability and falsifiability of the Hurwicz axioms. In particular, we compute the Vapnik-Chervonenkis dimension of the class of Hurwicz preferences, show that the Hurwicz class is PAC (probably approximately correct) learnable, provide a lower bound on the sample size required to learn a concept in this class, and provide an efficient polynomial-time algorithm to either learn or falsify a Hurwicz concept based on data.

Hurwicz criterion, machine learning, Vapnik-Chervonenkis dimension, learnability of decision theories