12:00, Wed 5 Nov 1997, WeH 7220
AutoRSM: Memory Based Active Learning
for Noisy Optimization and Control
Andrew Moore
Joint work with Jeff Schneider, Justin Boyan, Scott Davies, Mary Soon Lee
The Auton research project is an attempt to build an off-the-shelf
learning software system for highly autonomous use in industrial
control. A key component of Auton is AutoRSM (Autonomous Response
Surface Methods) which is used for actively seeking data to build
models and identify optimal operating regions.
In this talk I will begin by frothing at the mouth about how important
this problem is (and my obsession with it). I will discuss some of the
classes of noisy optimization problems we need to attack. I will
overview an array of possible approaches drawn from numerical
analysis, statistics, controls, reinforcement learning, evolutionary
computation and machine learning.
I'll then discuss memory-based approaches to the problem and the
potential merits. Our previous memory-based approaches have some
drawbacks in terms of the difficulty (computational and statistical)
of choosing the distance metric and size of the local region based on
very limited data. I will introduce our current draft of a new
algorithm, Q2, for robustly identifying regions of interest from
limited experiments and I'll provide some highly preliminary initial
results.