Acceleration of Experiment-Based Evolutionary Multi-objective Optimization Using Fitness Estimation

Abstract

Evolutionary Multi-objective Optimization (EMO) is ex-pected to be a powerful optimization framework for real world problems such as engineering design. Recent progress in automatic control and instrumentation provides a smart environment called Hardware In the Loop Simulation (HILS). It is available for our target application, that is, the experiment-based optimization. However, since Multi-objective Evolutionary Algorithms (MOEAs) require a large number of evaluations, it is difficult to apply it to real world problems of costly evaluation. To make experiment-based EMO using the HILS environment feasible, the most important pre-requisite is to reduce the number of necessary fitness evaluations. In the experiment-based EMO, the performance analysis of the evaluation reduction under the uncertainty such as observation noise is highly important, although the previous works assume noise-free environments. In this paper, we propose an evaluation reduction to overcome the above-mentioned problem by selecting the solution candidates by means of the estimated fitness before applying them to the real experiment in MOEAs. We call this technique Pre-selection. For the estimation of fitness, we adopt locally weighted regression. The effectiveness of the proposed method is examined by numerical experiments.