(145a) Data-Assisted Optimization

In many engineering fields, there is a continuously increasing interest in coupling equation-based first-principle modeling with information that comes in the form of input-output data for optimization. This approach enables the optimization of systems incorporating very detailed information regarding the material, flow, geometry, physical properties and chemistry of the systems. This need has given rise to development of methods that can optimize systems without equations or derivatives, but simply through the exchange of input-output data streams. The typical applications of data-driven optimization consider the system under study entirely as a “black-box”, however, many applications exist (i.e., process synthesis and design of modular manufacturing systems) which can be formulated as hybrid problems. Hybrid systems are comprised of both explicitly known equations (first principles) and data-dependent equations. This talk will present techniques for efficient optimization of such hybrid problems. Specifically, we propose and compare several methods for developing adaptive, flexible and tractable surrogate parametric functions to represent the information that comes in the form of input-output data. Finally, we present a data-assisted algorithm for optimization of these hybrid formulations, which integrates concepts from mathematical programming with machine learning techniques for regression and feature selection.