MAPO: Using a Committee of Algorithm-Experts for Parallel Optimization of Costly Functions

摘要

This paper describes a new parallel algorithm for optimizing costly nonconvex functions with box constraints when derivatives are unavailable. MAPO (Multi-Algorithm Parallel Optimization) iteratively uses a committee of optimization algorithms based on response surfaces to generate candidate points for function evaluation. In each iteration, the evaluation points are selected from multiple rankings of all candidate points. Good numerical results for MAPO with 4 radial basis function methods are reported for 8 processors.