Comparison of techniques for modeling uncertainty in a signal detection task

A simulation study was performed to compare the performances of signal detection models based on Bayesian, Dempster-Shafer, and fuzzy set theories. The models combined two different parameters calculated from each waveform. The models were applied to a set of brainstem auditory evoked potential waveforms, and the effects of errors in estimates of waveform parameters were evaluated. With correct parameter estimates, all the models gave similar classification results, and the nonBayesian detectors were able to provide good estimates of the Bayesian log likelihood ratio. The nonBayesian detectors were more robust than the Bayesian detector when parameter statistics were in error. While the generality of these results is limited because of the structured nature of the simulation, they give empirical support to the validity of nonBayesian methods to combine evidence in situations in which Bayesian methods are not appropriate