Building of decision function in the structured space of features

Abstract

The task of decision functions construction is considered for a case, when the features space is not represented simply by Cartesian product of sets of values, and has more complex structure; when the sets of allowed values of some variables depend on values taken by another ones. Such features space we call structured. The decision function is interpreted as an estimation of conditional distribution in the space of forecasted variables, under the condition of known values of measured variables. The decision function is constructed on the basis of the empirical data, which can simultaneously contain both sample and probabilistic statements of experts. Thus the expert statements are assumed to be unmatched, and in particular can be inconsistent.