ABSTRACT
In this chapter, the authors provide a discrete probability model. They discuss estimation methods geared to maximizing robust performance. The authors show how the data consistency constraints can be expressed in purely economic terms. They introduce the notion of data constraint relaxation as a way to balance consistency with the data and “prior beliefs,” a dual problem, and discuss the importance of the logarithmic family. In the absence of high-quality data, one might consider modeling the odds ratios, but that introduces additional complexity; moreover, the resulting model, under a general utility function, will be sensitive to the odds ratio model. Power utility functions are used widely in industry. Moreover, power utility functions have constant relative risk aversion and important optimality properties. The authors describe primal problems that were expressed in terms of feature expectation constraints.
