Self-learning forward neural network to realize adaptive FIR filter

An adaptive filter based on self-learning forward neural network (FNN) is proposed, which is capable of tracking both stationary and nonstationary signals. The neural network and FIR filters are combined with the self-learning mechanism to make the adaptive system operate distributively and parallelly. The important feature of the adaptive system is that the several parallel filters form a supersphere of N-dimensional space in which there is a centre and around it several W try to cover the overall supersphere. The suboptimal W can be quickly selected and then self-learning performance uses LMS to find the W and a new supersphere is built by the knowledge of topological space. Adaptive time is as fast as possible and can be modified as demanded. A new concept in which a supersphere moves according to nonstationary signals in N-dimensional space is presented