Talks
Fall 2017

Zero-order and Dynamic Sampling Methods for Nonlinear Optimization

Tuesday, October 3rd, 2017, 3:15 pm4:00 pm

Add to Calendar

In the first part of the talk we study the problem of minimizing a noisy function when derivatives are not available. In order to obtain scalability, the algorithm updates a quadratic model of the objective in order O(n) work using noise estimation techniques. Next we discuss a technique for dynamically increasing the accuracy in gradient approximations to achieve optimal complexity as well as efficiency in practice.