Evidence generation and representation for model uncertainty management in nonlinear state estimation

The authors address the problem of managing modeling uncertainty in nonlinear state estimation. A method for assessing model accuracy and providing alternate model selections is developed that involves the application of evidence generation and evidential reasoning. The evidence is the existence of nonzero coefficients of a regression fit to the predicted residuals. A belief function, developed from a classical multiple hypothesis likelihood ratio test, is used to map the evidence into the basic probability assignments for a feature frame of discernment. Dempster-Shafer (D-S) theory of evidential reasoning is used to map the belief in the existence of a feature into belief in model hypotheses and the subsequent measures of support and plausibility. Tracking problem results are presented that compare the performance of the D-S method to that of a traditional model assessment approach