Adaptive algorithms and networks for optimal feature extraction from Gaussian data
Pattern Recognition Letters
Strong points of weak convergence: a study using RPA gradient estimation for automatic learning
Automatica (Journal of IFAC)
Hi-index | 754.84 |
Two general classes of stochastic algorithms are considered, including algorithms considered by Ljung as well as algorithms of the formtheta_{n+1} = theta_{n} - gamma_{n+1} V_{n+1}(theta_{n}, Z), whereZis a stationary ergodic process. It is shown how one can apply a lemma of Kushner and Clark to obtain properties of these algorithms. This is done by using in particular Martingale arguments in the generalized Ljung case. In these various situations the convergence is obtained by the method of the associated ordinary differential equation, under the classical boundedness assumptions. In the case of linear algorithms, the boundedness assumptions are dropped.