Optimal allocation of simulation experiments in discrete stochastic optimization and approximative algorithms

Approximate solutions for discrete stochastic optimization problems are often obtained via simulation. It is reasonable to complement these solutions by confidence regions for the argmin-set. We address the question how a certain total number of random draws should be distributed among the set of alternatives. Two goals are considered: the minimization of the costs caused by using a statistical estimate of the true argmin, and the minimization of the expected size of the confidence sets. We show that an asymptotically optimal sampling strategy in the case of normal errors can be obtained by solving a convex optimization problem. To reduce the computational effort we propose a regularization that leads to a simple one-step allocation rule.