Comparing Expected Improvement and Kriging Believer for Expensive Bilevel Optimization

摘要

Bilevel optimization refers to a specialized class of problems where one optimization task is nested as a constraint within another. Such problems emerge in a range of real-world scenarios involving hierarchical decision-making, and significant literature exists on classical and evolutionary approaches to solve them. However, computationally expensive bilevel optimization problems remain relatively less explored. Since each evaluation incurs a significant computational cost, one can only perform a limited number of function evaluations during the course of search. Surrogate-assisted strategies provide a promising way forward to deal with such classes of problems. Of particular interest to this study are the steady-state strategies which carefully pre-select a promising solution for true evaluation based on a surrogate model. The main aim of this paper is to compare two widely adopted steady-state infill strategies -Kriging believer (KB) and expected improvement (EI) - through systematic experiments within a nested optimization framework. Our experiments on a set of benchmark problems reveal some interesting and counter-intuitive observations. We discuss some of the underlying reasons and believe that the findings will inform further research on understanding and improving search strategies for expensive bilevel optimization.