|
|
Foundations of Robotics Seminar, May 4, 2010
Time
and Place | Seminar Abstract
Using Expected Improvement to Effectively Optimize Snake Robot Gaits
Matt Tesch
Robotics Institute
CMU
NSH 1507
Talk 1:30 pm
Several categories of optimization problems suffer from the problem of
expensive cost function evaluation, driving the need for techniques
which choose subsequent experiments in a smart way. One such category
is that involving actual robotic systems, which often require
significant time, effort, and monetary expenditure in order to run
tests. To assist in the selection of the next experiment, there has
been a focus on the idea of \emph{response surfaces} in recent years.
These surfaces interpolate the existing data as well as provide a
measure of confidence in their error, aiming to serve as a
low-fidelity surrogate function that can be used to intelligently
choose the next experiment. In this talk, I discuss using the
response surface methodology with the expected improvement criteria to
optimize open-loop gait parameters for snake robots. First, I give a
brief overview of gaussian processes used to fit the response
surfaces. I then discuss the comparative merits of several methods
for using the response surface to select experimental parameters, and
describe the method I used, along with some subtle but important
modifications. Finally, I will present experimental results showing
this method to effectively optimize snake robot locomotion over
various terrain.
The Robotics Institute is part of the School of Computer Science, Carnegie Mellon University.