A Nonlinear Bilevel Programming Approach for Multifocal Region Creation in Metric Indexing

Abstract

This thesis presents a bilevel programming approach for optimizing region creation in metric indexing techniques. By creating regions from Power transformed distances, region quality is increased compared to the linear alternative. The bilevel algorithm shows a substantially higher performance than nonlinear direct search techniques in benchmarks. Heuristics for the outer level of the bilevel approach are the main focus and contribution of this work.