Visual analysis of high-dimensional parameter spaces

We present a system called Tuner to systematically analyze the parameter space of com- plex computer simulations, which are time consuming to run and consequently cannot be exhaustively sampled. We begin with a sparse initial sampling of the parameter space, then use these samples to create a fast emulator of the simulation. Analyzing this emulator gives the user insight on further sampling the simulation. Tuner guides the user through sampling and provides tools to find optimal parameter settings of up to two objective functions and perform sensitivity analysis. We present use-cases from the domain of image segmentation algorithms. Since our method must utilize samples of the simulation and relies on an inherently interactive visualization method, we perform a complexity analysis to see how many sam- ples can be rendered while staying interactive. We examined how rendering performance changes with the dimensionality, reconstruction kernel size, and number of sample points. To study this, we decomposed the rendering complexity into a predictive cost function that combines the cost of filtering each data point and then the cost to draw each pixel on screen. This cost function is calibrated to the time to filter and draw for two different hardware configurations. The cost formulation is used to examine the effects on rendering time from using box filtering versus a radial distance measure in high-dimensional data spaces as used for the filtered scatterplot and HyperSlice visualization methods, respectively. We find that for a constant kernel volume, rendering performance increases with dimensionality in the HyperSlice technique while it decreases with the filtered scatterplot technique. We also find that the total number of sample points and not the size of the reconstruction kernel is a much stronger determinant of the rendering time