Juli Mueller; Artificial Intelligence, Learning, and Intelligent Systems (ALIS) group; National Renewable Energy Laboratory (NREL)
Derivative-free Optimization Algorithms using Surrogate Models and Active Learning
Many scientific domains rely on computational simulation models to approximate physical phenomena under study. In some cases these simulations have parameters that must be tuned such that simulations and experimental observations agree as closely as possible. In other cases, simulations are used for inexpensively examining possible designs and thus enable performance optimization tasks without the need to first build the design and then evaluate its performance. One difficulty that arises with using these simulations in optimization is their computational expense and their blackbox nature, i.e., we do not have access to an analytical description of the objective function or derivatives. In these situations, finding an optimal solution demands that we query the expensive simulation as few times as possible without using derivatives. To this end, we employ computationally cheap surrogate models to approximate the expensive function and we use them to guide the iterative optimization, querying the expensive simulations only for select parameter values. We will discuss the ideas behind these surrogate model algorithms and their extensions to different optimization problem characteristics, as well as discuss their potential for enabling autonomous experimentation in laboratory settings.
More information about this speaker may be found atÌý
Dial-In Information
Ìý
Ìý