Extracting Specific Signal from Post-nonlinear Mixture Based on Maximum Negentropy

Abstract

To address the problem of extracting specific signal from the post-nonlinear (PNL) mixture, we propose a novel algorithm based on maximum negentropy. Assume that the prior knowledge of the desired source, such as its rough template referred as the references signal, is available. The closeness measurement between the corresponding estimated output and given reference signal is treated as a constraint and incorporated into the negentropy objective function. Therefore, a constrained optimization problem is formed, which is solved by the augmented Lagrange function method with standard gradient descent learning. The inverse of the unknown nonlinear function in the post-nonlinear (PNL) mixture model is approximated by the multilayer perceptions (MLP) network. Experiments on the synthesis dataset demonstrate the validity of our proposed algorithm.