Tid: 5 december 2016 kl 15.15Seminarierummet 3721, Institutionen för Matematik, KTH, Lindstedtsvägen 25, plan 7. Karta!
Föredragshållare: Alexandre Chotard
Titel: An Overview of Evolution Strategies
Abstract In many numerical optimization problems the objective function is a black box, meaning that information about the structure of the function or derivatives are unavailable, and the function can only be queried for its value point per point. Evolution strategies are comparison-based optimization algorithms for continuous spaces well suited to black-box problems. An Evolution Strategy generates a population of points from a sampling distribution and uses how the objective function ranks the generated points to adapt the sampling distribution and subsequently generate more promising points. These methods offer efficient solutions to difficult problems where the typical difficulties are related to non-convexity, noise, multi-modality, non-separability, ill-conditionning. In this talk I will give an overview of Evolution Strategies from a theoretical as well as practical standpoint. I will discuss how they can be related to stochastic natural gradient descent, invariances they can display to transformations of the search space, their limits, as well as methods to deal with noisy objective functions.
|Sidansvarig: Filip Lindskog