Direction of arrival estimation using artificial neural networks

The maximum likelihood estimator is the optimal estimator of the direction of sources, but it requires the minimization of a complex, multimodal, multidimensional cost function. A neural optimization procedure is presented that does not require an initial estimate of the direction of the sources and offers the potential of real-time solutions to the direction of arrival problem by utilizing the fast relaxation properties of the Hopfield network. A modification based on an iterated descent procedure is introduced into the Hopfield model dynamic equation to increase the probability of convergence to the global optimum. The algorithms are implemented on an array of closely coupled transputers that perform the random asynchronous neural updates in parallel. The mapping is achieved using a technique called chaotic relaxation. Simulation results are presented to characterize the performance of the neural approach in terms of the variance of the estimates of source directions and the time required for the computation of the estimates