U.S. Department of Energy

Pacific Northwest National Laboratory

Discovering Signatures When the Underlying Phenomenon is Poorly Understood

Employing genetic algorithms and a variety of regression models, our methodology globally searches a vast problem space to find relevant biometric features for a re-identification task, tackling the central problem for one of the SDI seed projects. For this seed project, our approach separated important features from unimportant ones, helping to narrow the search and producing robust models that convey reliable information.

For the development of spectroscopic and chromatographic signatures of bio-fuels and unconventional fuels, we have implemented a greedy search algorithm to discover features that may be potentially influential on properties of interest, such as viscosity and cetane number. Our methodology is not deterred by the hundreds of spectroscopic features we have obtained, and the objective function we have developed aims to detect which of the features are important and how do they affect the properties of interest. Next, our work will focus on obtaining many more potentially relevant features and on iteratively quantifying the performance of our methodology, testing its applicability and establishing strengths and weaknesses. A systematic experimental exploration of the problem space for some of these fuels is also being planned.

Impacts

Our approach has the ability to globally explore a large problem space that cannot be exhaustively searched and can detect relationships among features that may not be readily apparent to direct search methods. Our methodology has the ability to evaluate many different combinations of features, and explores the effect of simultaneous changes to multiple features on the phenomenon of interest. The algorithm is not easily deterred by noise, and it offers great flexibility because the objective function can be modified to accommodate changes in user priorities.

Project Staff: 
Tim Bays, John Cort, Alejandro Heredia-Langner, Kris Jarman
| Pacific Northwest National Laboratory