Worst Case Prediction-Based Differential Evolution for Multi-Noisy-Hard-Objective Optimization Problems

摘要

A new multiobjective optimization problem in presence of noise is formulated and called multi-noisy-hard-objective optimization problem (MNHOP). Since considering the worst case performance is important in many real-world optimization problems, each solution of MNHOP is evaluated based on the upper bounds of noisy objective functions' values predicted statistically from multiple samples. Then an Evolutionary Multiobjective Optimization Algorithm (EMOA) based on Differential Evolution is applied to MNHOP. Three sample saving techniques, namely U-cut, C-cut, and resampling, are proposed and introduced into the EMOA for allocating its computing budget only to promising solutions. Finally, the effects of those techniques are examined through numerical experiments.