ROBUST SIMULATION BASED OPTIMIZATION WITH INPUT UNCERTAINTY

Simulation-based Optimization (SbO) assumes that the simulation model is valid, and that the probability distributions used therein are accurate. However, in practice, the input probability distributions (input models) are estimated by sampling data from the real system. The errors in such estimates can have a profound impact on the optimal solution obtained by SbO. The existing two-stage framework for SbO under computational budget constraint considers only the stochastic uncertainty. In our variant, we consider the input model parameter uncertainty as well. Our algorithmic procedure is based on the stochastic kriging metamodel-assisted bootstrapping with an efficient global optimization technique which sequentially searches the optimum and incorporates Optimal Computational Budget Allocation (OCBA). This framework is also used for determining tighter worst-case bounds of the SbO with input uncertainty. The proposed framework is illustrated with the M/M/1 queuing model.