Convergence of perturbation analysis based optimization algorithm with fixed number of customers period (Q1338771)

To find the minimum of \(J(\Theta)\) with respect to \(\Theta\) perturbation analysis proposes a simulation method to estimate \({dJ(\Theta)\over d\Theta}\). It is natural to treat this estimation as a noise-corrupted observation of \({dJ(\Theta)\over d\Theta}\) and put it in the stochastic approximation algorithm. This algorithm is called perturbation analysis Robbins-Monro single run. The simulation results show that the estimate given by that algorithm converges to the zero of \({dJ(\Theta)\over d\Theta}\). In the paper, that fact under the condition that the algorithm is updated every fixed number of customers period is proved. The crucial point of the proof is the verification that the observation noises satisfy the strong law of large numbers.