Neural Network Based Approaches for Detecting Signals With Unknown Parameters

The detection of gaussian signals with unknown correlation coefficient, rhos, is considered. A strategy for designing mixture of experts in composite hypothesis test is proposed. It is based on designing a single multi-layer perceptron (MLP) trained with rhos varying uniformly in [0,1] to approximate the average likelihood ratio (ALR), and evaluate it for fixed values of rhos, so as to identify different variation subintervals of rhos, attending to the single MLP performance. Taking into consideration the relation that exists between MLP structure and the boundaries it is capable to built, we propose to train different MLPs with different sizes for each subinterval (MLP1 and MLP2, for the lower and higher half, respectively) for improving detection capabilities controlling computational cost. To improve the approximation implemented by MLP1, a radial basis function neural network (RBFNN) trained for the lower subinterval of ps has been combined with MLP2. As the functions approximated by the RBFNN and the MLP are equivalent but different, a combination strategy has been proposed based on thresholding the networks outputs and applying them to an OR logic function. Although this scheme does not outperform the 2MLPs, the reduction in computation cost is very important.