Mittagsseminar (in cooperation with A. Steger, D. Steurer and B. Sudakov)

Date and Time: Thursday, October 17, 2013, 12:15 pm

Duration: 30 minutes

Location: OAT S15/S16/S17

Speaker: Sebastian Stich

Natural Gradient in Evolution Strategies

In derivative-free optimization one aims at minimizing an unknown objective function. The only information accessible are algorithm-selected function measurements. Evolution Strategies (ES) are among the state of the art heuristics for this optimization problem. ES typically use parametrized probability distributions to generate correlated samples in promising regions.

Recently, it was shown that applying gradient descent in the parameter space of the search distribution leads to algorithms that are very similar to the most successful ES. The development of those so-called Natural Evolution Strategies (NES) provided new insights into the classical ES and started promising new theoretical investigations.

This still on-going research lead to the Information-Geometric-Optimization (IGO) framework, which tries to capture all ES in an unifying picture.

We present NES and give an introduction to the IGO framework.

Upcoming talks     |     All previous talks     |     Talks by speaker     |     Upcoming talks in iCal format (beta version!)

Previous talks by year:   2023  2022  2021  2020  2019  2018  2017  2016  2015  2014  2013  2012  2011  2010  2009  2008  2007  2006  2005  2004  2003  2002  2001  2000  1999  1998  1997  1996  

Information for students and suggested topics for student talks

Automatic MiSe System Software Version 1.4803M   |   admin login