| This dissertation focuses on the development of adaptive feature selection algorithms and sequential decision procedures for pattern recognition systems. A parametric feature selection algorithm based on a pairwise maximum likelihood criterion is developed and evaluated. Characteristics of training-sample estimates of Bhattacharyya coefficients are studied. Asymptotic distributions of high order Prony's estimate for pole locations of exponentially modelled processes are derived. An {dollar}M{dollar}-ary sequential decision algorithm is introduced and an adaptive feature selection decision strategy for sequential classification algorithms is examined. These algorithms are applied to radar target identification systems.; The performance of feature sets selected by the pairwise maximum likelihood criterion, for both coherent and noncoherent radar backscatter data, are evaluated through simulation studies of a radar target identification system. The results demonstrated a significant performance improvement as compared to feature sets selected by other criteria.; Closed-form expressions for the mean and variance of training-sample estimates of the Bhattacharyya coefficients are derived in terms of system dependent parameters. It is shown that the bias of the training-sample estimate may be positive, negative, or, under certain circumstances, begin negative and become positive as the number of training samples increases. The estimates are shown to be asymptotically efficient for a one-dimensional observation space.; The asymptotic probability density function for estimates of pole locations are derived. Closed form results show the density to be multivariate Gaussian. Using the derived density in a maximum likelihood classification scheme, it is shown that performance degrades when the SNR decreases, when energy in the unmodelled dynamics is significant and when only a small number of measurement sets are available.; A study of adaptive feature selection decision strategies for sequential algorithms is carried out. The results reveal that class rejection is not desirable and that adaptive weighted feature selection realizes a significant improvement in the average sample number-probability of error characteristics. |
