Torture Tests: A Quantitative Analysis for the Robustness of Knowledge-Based Systems

The overall aim of this paper is to provide a general setting for quantitative quality measures of Knowledge-Based System behavior which is widely applicable to many Knowledge-Based Systems. We propose a general approach that we call “degradation studies”: an analysis of how system output degrades as a function of degrading system input, such as incomplete or incorrect inputs. Such degradation studies avoid a number of problems that have plagued earlier attempts at defining such quality measures because they do not require a comparison between different (and often incomparable) systems, and they are entirely independent of the internal workings of the particular Knowledge-Based System at hand. To show the feasibility of our approach, we have applied it in a specific casestudy. We have taken a large and realistic vegetation-classification system, and have analyzed its behavior under various varieties of missing input. This casestudy shows that degradation studies can reveal interesting and surprising properties of the system under study.