Stochastic Localization Methods for Discrete Convex Simulation Optimization

We propose a set of new algorithms based on stochastic localization methods for large-scale discrete simulation optimization problems with convexity structure. All proposed algorithms, with the general idea of "localizing" potential good solutions to an adaptively shrinking subset, are guaranteed with high probability to identify a solution that is close enough to the optimal given any precision level. Specifically, for one-dimensional large-scale problems, we propose an enhanced adaptive algorithm with an expected simulation cost asymptotically independent of the problem scale, which is proved to attain the best achievable performance. For multi-dimensional large-scale problems, we propose statistically guaranteed stochastic cutting-plane algorithms, the simulation costs of which have no dependence on model parameters such as the Lipschitz parameter, as well as low polynomial order of dependence on the problem scale and dimension. Numerical experiments are implemented to support our theoretical findings. The theory results, joint the numerical experiments, provide insights and recommendations on which algorithm to use in different real application settings.