Title : Visualization for non-convex multi-objective optimization



Project Lead : Grazina Gimbutiene From : Vilnius University (None)

Dates : from -- to 2014-11-28 14:53:18

Description :

Motivation and objectives :
While convex optimization is well explored and standard software packages exist to facilitate it, many emerging real world problems, e. g. in financial decision making, are non-convex. It is broadly admitted that they require further theoretical exploration, as well as efficient algorithms to solve them. The goal of the project is to test the methods and algorithms for multi-objective optimization when criterion functions are non-convex, by visualizing information produced in the process of optimization.

Teams :
The group consists of the researchers of the sector of Optimization at the Institute of Mathematics and Informatics. Main research subjects: global optimization, multi-objective optimization, visualization of multidimensional data.

Dates :
starting date : 15 November, 2014
ending date : 27 November, 2014

Facilities descriptions :
http://visionair-browser.g-scop.grenoble-inp.fr/visionair/Browser/Catalogs/MEXICO.FR.html

Recordings & Results :
The research of this project concentrates on visual support for creating methods for global optimization of non-convex objective functions. Two types of tasks in this reasearch were supported by the VisionAir visualization infrastructure. First, an implementation of an evolutionary multi-objective (EMO) algorithm could be tested visually to see if it produces an expected approximation of a Pareto set and front. Second, the strategy of a single objective optimization algorithm, optimizing a function defined in 3D, could be visualized to test if it produces an expected density of trial points and the relative level of the function values. A conclusion can be made that the visualization using high definition and stereoscopic screens improves the perception of the visualized 3D point arrays in space, as compared to the standard displays.

Conclusions :
The optimization datasets produced by the algorithms being tested have been visualized. It could be seen that one implementation of an EMO works as expected, while the other needs to be corrected. The single objective optimization algorithm parameters could be adjusted in order to increase/decrease the globality of search. The perception of all points in 3D space was admitted to be better than that using commodity monitors.




Project Images :

20141125_115929.jpg
20141125_114655.jpg 20141125_114810.jpg 20141125_115427.jpg
20141125_115056.jpg 20141125_114638.jpg



.



Visionair logo

VISIONAIR / Grenoble INP / 46 avenue Felix Viallet / F-38 031 Grenoble cedex 1 / FRANCE
Project funded by the European Commission under grant agreement 262044