Trust Region Approaches for Glass Box/Black Box Optimization
Abstract: In this talk, we discuss our work on hybrid glass box/black box optimization, also known as gray box optimization. This problem structure, which bridges classical nonlinear optimization with derivative-free optimization, occurs when a subset of model equations is available only via black box evaluations. κ-fully linear approximation models are used to replace the black box function, and a trust region filter algorithm drives convergence to an optimal solution of the original glass box/black box model. The algorithm is demonstrated on modified optimization benchmarks as well as illustrative examples for chemical process optimization. Then, we discuss our work on extending the use of κ -fully linear models to "short-cut" approximation models. With phase equilibrium as an example, we show how the underlying assumptions of the short-cut model can be "tuned" to build κ -fully linear models that incorporate knowledge about the underlying physics.
Meeting ID: 708 725 647
Participant Passcode: 6397