Title: Adaptive Sampling with Trust-region Optimization for Nonconvex Stochastic Functions.

Abstract: Simulation Optimization problems are often non-convex, derivative-free, and with stochastic noise that may be notably heteroskedastic. There is emerging evidence that stochastic trust-region methods are very effective at solving such problems, especially due to their natural ability to self-tune step sizes. The ASTRO algorithm is one that uses trust-region methods with adaptive sampling for efficiency. The paradigm of adaptive sampling is one where the simulation effort for estimating the function value at each solution is chosen based on its distance to optimality. In this talk, I explain this class of algorithms and its recent trend in improving sample complexity using variance reduction techniques. Lastly, I describe a future direction towards robustness of solving Simulation Optimization problems that can extend to machine learning applications by incorporating estimation bias stemming from input uncertainty.