Diffusion adagrad minimum kernel risk sensitive mean p-power loss algorithm

Abstract

The most diffusion algorithms based on the mean square error (MSE) criterion generally have good performance in the presence of Gaussian noise, however suffer from performance deterioration under non-Gaussian noises. To combat non-Gaussian noises, a diffusion minimum kernel risk sensitive mean p-power loss (DMKRSP) algorithm is first designed using a generalized robust kernel risk sensitive mean p-power loss (KRSP) criterion combined with stochastic gradient descent (SGD). Then, due to more error information than SGD, the adaptive gradient (Adagrad) is used in DMKRSP to generate a diffusion Adagrad minimum kernel risk sensitive mean p-power loss (DAMKRSP) algorithm. Finally, the theoretical analysis of DMKRSP and DAMKRSP is presented for steady-state performance analysis. Simulations on system identification show that both DMKRSP and DAMKRSP are superior to other classical algorithms in term of robustness and filtering accuracy.