A stochastic variable metric algorithm for system modelling and identification
H. Anthony Barker, R. K. APPIAH · International Journal of Systems Science · 1971
A variable gain matrix is proposed for the multi-dimensional Robbins-Monro stochastic approximation process for gradient search optimization over an unknown regression surface. The algorithm uses only the gradient or matrix gradient of the objective function and hence is not restricted to the minimum-mean-square-error criterion. It is compared with the recursive least squaros method in identifying a class of discrete non-linear dynamic systems with constant and time-varying parameters.