Chair: Jose Principe, University of Florida (USA)
Bob X. Li, McMaster University (CANADA)
Simon Haykin, McMaster University (CANADA)
A new method to construct a chaotic generator is presented in this paper. It is based on the fact that sea clutter is chaotic, and it can be reconstructed by the use of a supervised neural network. The constructed chaotic sequence generator is then applied to the direct-sequence coherent BPSK communication system as a pseudo-noise generator. Experimental results are presented. The advantages and disadvantages of the new pseudo- noise generator are discussed.
Shun-Hsyung Chang, National Taiwan Ocean University
Tong-Yao Lee, National Taiwan Ocean University
Wen-Hsien Fang, National Taiwan Institute of Technology (TAIWAN - REPUBLIC OF CHINA)
In this paper, a novel artificial neural network (ANN) called the UNItary Decomposition ANN (UNIDANN), which can perform the unitary (Schur) decomposition of the synaptic weight matrix, is presented. It is shown both analytically and quantitively that if the synaptic weight matrix is positive definite and normal, the dynamic equation involved will converge to a unitary matrix which can transform the weight matrix into an upper triangular one via the Schur decomposition. In particular, if the synaptic weight matrix is also Hermitian (symmetric for real case), the UNIDANN will perform the eigen decomposition. Compared with other existing ANN's, the proposed one possesses several attractive features such as more versatile in the sense that it is capable of performing the Schur decomposition, low computation time and no synchronization problem due to the application of the structure of analog circuit, and faster convergent speed. Some simulations with particular emphasis at the MUSIC bearing estimation algorithm have been provided to justify the validity of the proposed ANN.
Tser-Ya Dai, National Chiao Tung University (TAIWAN - REPUBLIC OF CHINA)
Ta-Sung Lee, National Chiao Tung University (TAIWAN - REPUBLIC OF CHINA)
An application of the radial-basis function neural network (RBF NN) on the angle-of-arrival (AOA) estimation of a desired source in multipath environments is investigated. In conjunction with a set of judiciously constructed beamformers, the RBF NN are used to estimate the desired AOA within an angular sector of interest (ASOI). With a pilot signal emitted from each of the training AOA's within the ASOI, the RBF NN is trained with the higher-order statistics (HOS) estimated from the received array data. In principle, the RBF NN AOA estimator maps the complex HOS into the desired angle response as an function approximator. By matching the HOS to the center vectors associated with the hidden nodes and linearly combining the node values, an AOA estimate results. The efficacy of the proposed AOA estimator is confirmed by computer simulations.
Gerardo J. Melendez, U.S. Army Communications & Electrical Command
Stanislav B. Kesler, Drexel University (USA)
Radar targets with moving components (engines, propellers, etc.) are well characterized in the frequency domain so that the estimation of spectral parameters can be used in the process of extracting features for their classification. With non-imaging scanning radar the application of spectral parameter estimation for target classification is limited by a short time on target and other radar parameters. To alleviate these limitations, articial neural networks have been used bringing with them the long training time issue. This paper suggests a neural network approach that capitalizes on the presence of spectral components and reduces the training time. This approach deivides the training of the multilayer perception (MLO|P) in two steps: pretaining and final training. Pretrainingi guies the MLP to recognize the Fourier basis set of signals. The final training for target classification is then faciliitated.
Jeffrey L. Vaughn, Optical Sciences Company
Neil J. Bershad, University of California at Irvine (USA)
This paper summarizes the results of the simulations and analysis of the learning behavior of a simple two-layer perceptron for a nonlinear system identification problem. Although it is difficult to generalize results for nonlinear systems, the analysis may improve our understanding of neural network training. Numerous sub-optimum stationary points occur for this problem and cause difficulties in the correct identification of the unknown system. The sub-optimum convergence points occur in the saturation regions of the various nonlinearities or for pathological cases. The size of the region of suboptimal convergence points may be reduced by increasing the dimensionality of the input data vector. Also, the range for the rate parameter is computed and an improvement to backpropagation is suggested.