Stochastic quasigradient methods for optimization of discrete event systems (Q1207835): Difference between revisions

Stochastic quasigradient methods are adapted and further developed for applications to discrete event systems. The construction of the quasigradients is examined for the cases when differentiation schemes encounter difficulties. In particular, the authors deal with discontinuities by developing methods which need only observations of the objective function instead of observations of its gradients: An enhanced finite difference scheme with reduced variance is presented by using random smoothing and common random numbers; furthermore, an algorithm is given which performs on-line approximation of the objective function, on the basis of the current and a number of previous observations, and then differentiation or minimization of the approximating function. The almost sure convergence of the algorithms is shown and a numerical example is presented.