GMM-Guided gradient descent learning of RBF Neural Network with its application on robust GPS satellites selection
Morteza Saraf, Karim Mohammadi, Mohammad Reza Mosavi · 2011
In this paper, a novel gradient descent learning algorithm based on Gaussian Mixture Model (GMM) applied on Radial Basis Function Neural Network (RBFNN) is proposed in order to approximate the functions which have high non-linear order. How we can choose the strategy of gradient descent including learning coefficients selecting and really is it optimized to learn the same for all feature vectors?, are the challenges made us to think precisely on those. In this study, GMM estimates the probability density of the feature space and then the optimal learning rates can be evaluated proportional to these probabilities. These cause the neurons to learn correspondence with the feature distribution likelihoods. Considering robust satellite subset selection, Geometric Dilution of Precision (GDOP) factor is calculated for all subset of satellites and then the subset with lowest value is selected for positioning, but it is so non-linear and has computational burden to navigation systems. We use the proposed method to approximate it. The results on real GPS measurements demonstrate that it significantly track the non-linearity of GPS GDOP.