B1493
Title: Visual diagnostics for constrained optimization with application to guided tours
Authors: Huize Zhang - University of Texas at Austin (United States) [presenting]
Di Cook - Monash University (Australia)
Ursula Laa - BOKU University (Austria)
Nicolas Langrene - BNU - HKBU United International College (China)
Patricia Menendez - Monash University (Australia)
Abstract: Projection pursuit is a technique to find interesting low-dimensional projections of high-dimension data. This is achieved through the optimization of an index function, which assigns an interestingness score to each linear projection. In practice, however, the optimizer does not always work as desired: it may fail unexpectedly, get stuck at a local maximum, or approach the maximum without reaching it. Four diagnostic plots are introduced, designed to track the progress of the optimization and the coverage of the parameter space. When combined with a visualization technique known as the guided tour, different optimization routines are visualized in the high-dimensional space. This allows for the comparison of the search strategies employed by different optimizers.