SUPREMAL INFERENCE MACHINE AND MEASURES OF CERTAINTY

Let us assume there exist several expert systems using different infer ence machines, working in parallel on the decision making problem from the same area and supplying for a given object (patient) probabilitie s of different diagnoses. The global results (i.e. number of errors) o n some sample data set of patients determine which of the machines is the best one. Is it possible, using the results (i.e. probabilities of diagnoses for the given patient) of the other machines, to improve de cision power of the best machine? A method, the supremal inference mac hine or algorithm, is introduced attempting to combine different infer ence machines with the help of the random variable ''error content in decision'' whose density is constructed for different measures of cert ainty. Experimental results on a case study from the area of rheumatol ogy are given.