Statistical analysis of the effect of input errors on expert system advice

A randomized, statistical test method is presented in which system responses to specific input sequences are compared to those obtained when some system queries are answered incorrectly. Over a representative sample of input sequences this method leads to an overall error sensitivity rate as well as query-specific error rates that reflect key questions in the input sequence. The test method developed here can be used to generate overall and specific analyses of the consequences of user input errors in the application of a medical expert system. A simple example is provided by a nonbranching, three-rule system for which all inputs are yes/no, and three advice-oriented outputs are provided