Contour location via entropy reduction leveraging multiple - PowerPoint PPT Presentation

Contour location via entropy reduction leveraging multiple information sources (Poster AB#99)  Dr. Alexandre Marques (MIT)  Dr. Remi Lam* (MIT)  Prof. Karen Willcox (ICES, UT Austin) Supported by Air Force CoE on Multi-Fidelity Modeling of Rocket Combustor Dynamics, Award FA9550-17-1-0195, and AFOSR MURI on Managing Multiple Information Sources of Multi-Physics Systems, Awards FA9550-15-1-0038 and FA9550-18-1-0023 * Now at DeepMind 1

Single information source Classification requires many information In many cases IS 0 is expensive, but relatively • • source (IS) evaluations inexpensive (biased) IS are available Active learning based on GP surrogate How to leverage all IS to produce accurate • • produces better classifiers at lower cost classifier at lower cost? 2

Single information source Multi-information source Classification requires many information In many cases IS 0 is expensive, but relatively • • source (IS) evaluations inexpensive (biased) IS are available Active learning based on GP surrogate How to leverage all IS to produce accurate • • produces better classifiers at lower cost classifier at lower cost? 3

Single information source Multi-information source Classification requires many information In many cases IS 0 is expensive, but relatively • • source (IS) evaluations inexpensive (biased) IS are available Active learning based on GP surrogate How to leverage all IS to produce accurate • • produces better classifiers at lower cost classifier at lower cost? 4

Single information source Multi-information source Classification requires many information In many cases IS 0 is expensive, but relatively • • source (IS) evaluations inexpensive (biased) IS are available Active learning based on GP surrogate How to leverage all IS to produce accurate • • produces better classifiers at lower cost classifier at lower cost? 5

Single information source Multi-information source Classification requires many information In many cases IS 0 is expensive, but relatively • • source (IS) evaluations inexpensive (biased) IS are available Active learning based on GP surrogate How to leverage all IS to produce accurate • • produces better classifiers at lower cost classifier at lower cost? 6

Single information source Multi-information source Classification requires many information In many cases IS 0 is expensive, but relatively • • source (IS) evaluations inexpensive (biased) IS are available Active learning based on GP surrogate How to leverage all IS to produce accurate • • produces better classifiers at lower cost classifier at lower cost? 7

Single information source Multi-information source Classification requires many information In many cases IS 0 is expensive, but relatively • • source (IS) evaluations inexpensive (biased) IS are available Active learning based on GP surrogate How to leverage all IS to produce accurate • • produces better classifiers at lower cost classifier at lower cost? 8

Single information source Multi-information source Classification requires many information In many cases IS 0 is expensive, but relatively • • source (IS) evaluations inexpensive (biased) IS are available Active learning based on GP surrogate How to leverage all IS to produce accurate • • produces better classifiers at lower cost classifier at lower cost? 9

Single information source Multi-information source Classification requires many information In many cases IS 0 is expensive, but relatively • • source (IS) evaluations inexpensive (biased) IS are available Active learning based on GP surrogate How to leverage all IS to produce accurate • • produces better classifiers at lower cost classifier at lower cost? 10

Single information source Multi-information source Classification requires many information In many cases IS 0 is expensive, but relatively • • source (IS) evaluations inexpensive (biased) IS are available Active learning based on GP surrogate How to leverage all IS to produce accurate • • produces better classifiers at lower cost classifier at lower cost? 11

Single information source Multi-information source Classification requires many information In many cases IS 0 is expensive, but relatively • • source (IS) evaluations inexpensive (biased) IS are available Active learning based on GP surrogate How to leverage all IS to produce accurate • • produces better classifiers at lower cost classifier at lower cost? 12

Single information source Multi-information source Classification requires many information In many cases IS 0 is expensive, but relatively • • source (IS) evaluations inexpensive (biased) IS are available Active learning based on GP surrogate How to leverage all IS to produce accurate • • produces better classifiers at lower cost classifier at lower cost? 13

Single information source Multi-information source Classification requires many information In many cases IS 0 is expensive, but relatively • • source (IS) evaluations inexpensive (biased) IS are available Active learning based on GP surrogate How to leverage all IS to produce accurate • • produces better classifiers at lower cost classifier at lower cost? 14

Contributions  Contour entropy: measure of uncertainty about the location of the zero contour of function approximated by statistical surrogate model  Decision mechanism: maximizes average reduction of contour entropy via one-step lookahead approach  CLoVER ( C ontour Lo cation V ia E ntropy R eduction): algorithm that combines data from multiple information sources to locate contours of expensive functions at low cost 15

Poster AB #99  CLoVER: Contour location via entropy reduction leveraging multiple information sources  Thursday Dec 6th, Poster Session B, 5-7pm  Room 210 & 230 16