Simulation optimization problems arise in many different application areas. For example, in groundwater remediation we want to minimize the cleanup costs subject to a contamination constraint; in combustion, climate, or cosmology applications, we generally want to minimize the error between our observations and our simulations. In these applications, we have to run a computationally expensive simulation model (several minutes to hours per run) in order to obtain a single objective or constraint function value. Analytical descriptions of the objective and its derivatives are not available (black-box). The goal is to find the optimal solution within a very low number of expensive function evaluations. To this end, we use computationally cheap surrogate models to approximate the expensive simulation objective and constraint functions. Throughout the optimization, we use the surrogate models to guide the search for improved sample points at which we then query the expensive simulation. The surrogate models are updated every time a new point has been evaluated with the expensive simulation. In this talk, I will give an overview of surrogate model optimization algorithms and I will showcase application problems that we have solved successfully with this method.

Dr. Mueller is a research scientist in the Computational Research Division of the Computing Sciences Directorate at the Lawrence Berkeley National Laboratory, and affiliated with the Center for Computational Sciences and Engineering (CCSE).