Efficient network training for DOA estimation

We treat the estimation of direction of arrival (DOA) in mobile communications as a mapping problem. A multilayer feedforward neural network (FNN) is proposed to establish such a map. The main advantage for a trained FNN is that DOA estimation becomes simple and cost-effective for real-time applications. Since training of FNN by the popular backpropagation algorithm usually requires a large number of training iterations to attain a certain accuracy in terms of network approximation, we propose an efficient network training algorithm based on nonlinear optimization. The FNN is first analyzed to obtain those convex regions containing all local solutions. Then, a search is performed constraining to these convex regions for local minima. Since the search is performed over these convex regions, the proposed algorithm can reduce chances of premature algorithm termination due to low gradient values. Preliminary numerical results are provided to illustrate the potential applications