A new approach to the solution of expert classification problems

Abstract

Above we stated a set of concepts conducive to:

• formation of the set of diagnostic attributes;

• correct elicitation of information from the expert;

• valid dissipation of this information;

• search for the most informative states to be presented to the expert;

• detection and elimination of errors in the expert's response;

• due regard to capacities of human information processing system.

The above set of concepts provides a real opportunity for designing a manmachine system of KB construction meeting a set of requirements (see details in Larichev et al., 1991).

First, it is necessary to structure the problem, identify a set of attributes and scales, determine decision classes. The attributes and scales of their estimates determine a complete list of potential states of the object under study. In line with the available algorithm, computer assesses the potential informativeness of all feasible states and selects the most informative one. This state is presented to the expert. The latter classifies the presented state. Then the expert's answer is verified for consistency (note that verification may be carried out following a series of answers of the expert). Once the inconsistencies arc eliminated, the following informative point is determined, etc. until all states are classified.