An evaluation of statistical neural network training algorithms with respect to VLSI implementation for fast adaptive control

This paper evaluates two existing statistical neural network training algorithms developed to overcome the problems associated with VLSI implementation of exact gradient descent algorithms such as backpropagation: the algorithm for pattern extraction (ALOPEX), and the random weight change (RWC) algorithm. The advantages of RWC over ALOPEX for fast VLSI implementation, and for continual online training (COT) applications, such as adaptive control, are explained. Simulation results demonstrate these advantages, and form the basis of a more detailed statistical evaluation of the COT performance of RWC