Applied Math Seminar

---

State-dependent diffusion for global optimization

Speaker: Kui Ren, Columbia University

Location: Warren Weaver Hall 1302

Date: Friday, April 28, 2023, 2:30 p.m.

Synopsis:

We propose and analyze a gradient descent algorithm with added stochastic terms for finding the global optimizers of nonconvex
optimization problems, The main feature of the algorithm is that the size of the randomness is tuned adaptively based on the value of the
objective function. We show, in the setup of the discrete algorithm not its continuous limit, that it is possible for the algorithm to achieve
global convergence with an algebraic rate. We will discuss possible generalizations of the algorithm for applications. This talk is based on joint works with Bjorn Engquist and Yunan Yang.