CAM colloquium - Friday, December 1
3:30 p.m.
655 Rhodes Hall

Speaker: Katya Scheinberg, IBM

 

Title: Developing Provably Converging and Practical Methods in Derivative Free Optimization

Abstract: Derivative free optimization (DFO) is the field of nonlinear optimization which targets functions whose derivatives exists but are not available and cannot be approximated efficiently. It is often also the case that such functions are expensive to evaluate and/or are noisy. In the past decade there has been a significant increase in research in the area of DFO. There are several practical algorithms that were proposed most of them lacking global convergence theory. Those methods that do have convergence theory have to resort to impractical extra steps and conditions. But most of the proposed methods have the "right ingredients" for both convergence theory and practical performance. We are attempting to develop a reasonably general framework which can include many practical algorithms and yet is shown to be globally convergent. We will discuss the main ingredients of this framework, such as the requirement on models of the objective function that are used by the algorithms. We will then discuss such models. In particular we will apply simplified results from approximation theory to show that polynomial interpolation and regression models can be made to satisfy the requirements of our framework.

 

Refreshments at 4:30 in 657 Rhodes Hall.