Kai F. Gong

On Training Artificial Neural Networks for Tracking in the Ocean Environment

This paper examines the issues associated with training and applying artificial neural networks (As) for tracking in the ocean environment. To this end, bhe tracking problem and traditional approaches are briefly described with factors affecting performance delineated, The ability to integrate As with other tracking methods to provide the robustness and flexibility desired in undersea tracking systems is expr"ored. In particular, methods for selecting training cases and candidate system architectureS are discumed. As an illustrative example, an ANN trained to map angle-of-arrival measurements into contact state Parameters is presented. Performance is revealed through comparing simulation results relative to the Cramer-Kao lower bound.

Performance characterization of artificial neural networks for contact tracking

Artificial neural networks (ANNs) can be exploited in a variety of information processing applications because they offer simplicity of implementation, possess inherent parallel processing characteristics and are nonlinear and less reliant on modeling of the real process. The paper is concerned with the problem of determining the performance of ANNs trained to provide estimates of contact state variables given a time series of measurements. A method is presented for determining ANN performance. Specifically, performance is shown to be intrinsically related to system observability. A performance analysis of ANNs under various observability conditions is presented along with a methodology for selecting the appropriate ANN-generated solution with a system architecture comprised of multiple clusters of ANNs.<<ETX>>

Adaptive filtering in underwater tracking with correlated measurement noise

The authors describe a multistage, hierarchical adaptive filtering technique derived for a complex, real-time nonlinear state estimation problem where the measurement noise is markedly non-Gaussian. The relative efficiency of the derived mathematical model was demonstrated through a Monte Carlo simulation experiment employing noisy angle-of-arrival measurements for undersea tracking.<<ETX>>

Nonlinear Parameter Estimation With Segmented Data: Trajectory Estimation with Biased Measurements

The problem of extracting parameters from a sequence of data generated by a nonlinear process is examined. The asymptotic normality of the posterior distribution provides a simplified and unified Bayesian and maximum likelihood approach to nonlinear estimation with segmented data. proposed method consists of locally processing data segments to produce a reduced data set. Nonlinear estimation techniques are then applied to combine these local estimates to obtain the desired overall or global estimate. reduced data set results in both computational efficiency and flexible algorithm design. illustrative example, an estimator is developed for a trajectory estimation problem where the data consists of noisy and biased angle-ofarrival measurements. The

Lower Bound Analysis For Large Errors In Nonlinear State Estimation

This paper addresses the problem of estimating a state vector from a sequence of data generated by a nonlinear process. In particular, the effects of nonlinearity on performance bounds for large estimation errors are analyzed via the marginal density function. The use of local processing to produce a reduced data set, under conditions of asymptotic normality, provides a simple mechanism for formulating a likelihood function. Nonlinear estimation techniques are then applied to a reduced order system (reduced dimensionality), and an approximation of the marginal density is obtained with the corresponding error bounds. This method of hierarchical processing results in both computational efficiency and algorithm flexibility. As an illustrative example, simulated results are presented involving noisy and biased data. Deviations from the linear assumptions inherent in the use of the traditional Fisher information matrix (FIM) are explained.

Further Results On Fundamental Limitations Of Conventional Target Motion Analysis

This paper examines the interaction of velocity information with the bearings-only target motion analysis problem. Theoretical results, obtained by a Cramer-Rao lower bound analysis are presented. Experimental results, obtained using a maximum likelihood estimation algorithm and simulated data agree with theory.