Logo image
Analyses and robust algorithms in the membership-set parameter estimation
Dissertation

Analyses and robust algorithms in the membership-set parameter estimation

Hyonyong Cho
University of Iowa
Doctor of Philosophy (PhD), University of Iowa
Autumn 1998
pdf
t1998_c542-o12.27 MB
Delayed release, Embargo ends: 01/01/2099

Abstract

Membership-set parameter estimation is based on a simple assumption that the uncertainty or noise involved in the modeling and measurements is bounded and the bound is known a priori. As long as this assumption is valid, membership set estimation provides a set of estimates which are consistent with input-output measurements. Naturally, an important performance criterion is the size of the membership set. We establish some results on the size of the membership set under both deterministic and stochastic assumptions. Usually, the noise is assumed to be bounded pointwise. However, there are other possibilities in characterizing the noise. We study three different noise models and present the following results: (1) For each noise model, as the amount of data increases, the membership set converges to a singleton, which is the true parameter, under certain conditions. (2) Depending on the characteristic of the noise, some models can be better than the others in terms of the membership-set size or the sensitivity to an erroneous measurement i.e., an outlier. (3) For each model, noise bound and parameter estimates can be obtained by some minimax estimation. We show that the estimates converge to the true values as the amount of data increases. Membership-set estimation is not robust to outliers. With outliers, the re sulting membership set is either a subset of the nominal membership set or, very likely as the amount of data increases, an empty set. In the context of the above- mentioned minimax estimation, we propose a computationally efficient algorithm where the optimization has freedom to exclude a certain number of data points. We prove that applying this algorithm under certain conditions, the estimates converge to the true values even for data containing outliers. Simulations confirm that the proposed algorithm provides accurate estimates in the presence of outliers and that the algorithm is also effective for identifying the outliers.

Details

Metrics

1 File views/ downloads
2 Record Views
Logo image