A framework for managing models in nonlinear optimization of computationally expensive functions

摘要: One of the most significant problems in application standard optimization methods to real-world engineering design is that computation objective function often takes so much computer time (sometimes hours) traditional techniques are not practical. A solution has long been used this situation approximate with something cheaper compute, called a “model” (or surrogate), and optimize model instead actual function. This simple approach succeeds some time, but sometimes it fails because there sufficient priori knowledge build an adequate model. way address problem whatever available, during process sample true at selected points use results monitor progress adapt region interest. We call “model management”. thesis will on fundamental ideas theory pattern search develop rigorous methodology for management. general framework management algorithms be presented along convergence analysis. software implementation framework, which allows reuse existing modeling software, developed several test presented. The potential applications aerospace subject ongoing collaboration between researchers Boeing, IBM, Rice College William & Mary.