A neural network for model order selection in signal processing

The purpose of this paper is to propose the design and use of a neural network for model order selection in a class of parametric estimation method. This neural network uses an original initialization and second order backpropagation. Classical detection tests need an eigen-decomposition of the correlation matrix, which is computationally hard and inefficient in a non-asymptotic case. This paper includes simulations which show the superiority of the neural network approach in comparison to classical tests in term of computational cost and performances.