An Interactive Approach for Hierarchical Parameter Optimization

Abstract

We propose a novel interactive parameter optimization technique which asks users to determine whether further exploration is required by using a hierarchical visualization technique. Our technique generates a k-d tree and uses the coefficient of multiple determinations (R²) for tree-branching; however, this requires a huge amount of computation since it processes the response surface at every leaf node of the k-d tree, including those that may not have minimum parameters. To solve the problem, we introduce human evaluation of the display on the computer screen, in which every tree-node is evenly displayed regardless of its hierarchical level. We confirmed the effectiveness of the proposed technique by searching for pre-determined parameters in some test functions and biological cell simulation results.