FSEO: A Few-Shot Evolutionary Optimization Framework for Expensive Multi-Objective Optimization and Constrained Optimization
Keywords: Expensive optimization, few-shot optimization, multi-objective optimization, constrained optimization, surrogate-assisted evolutionary optimization
Abstract: Meta-learning has been demonstrated to be useful to improve the sampling efficiency of Bayesian optimization (BO) and surrogate-assisted evolutionary algorithms (SAEAs) when solving expensive optimization problems (EOPs). However, existing studies focuses on only single-objective optimization, leaving other expensive optimization scenarios unconsidered. We propose a generalized few-shot evolutionary optimization (FSEO) framework and focus on its performance on two common expensive optimization scenarios: multi-objective EOPs (EMOPs) and constrained EOPs (ECOPs). We develop a novel meta-learning modeling approach to train surrogates for our FSEO framework, an accuracy-based update strategy is designed to adapt surrogates during the optimization process. The surrogates in FSEO framework combines neural network with Gaussian Processes (GPs), their network parameters and some parameters of GPs represent useful experience and are meta-learned across related optimization tasks, the remaining GPs parameters are task-specific parameters that represent unique features of the target task. We demonstrate that our FSEO framework is able to improve sampling efficiency on both EMOP and ECOP. Empirical conclusions are made to guide the application of our FSEO framework.
Primary Area: Optimization (convex and non-convex, discrete, stochastic, robust)
Submission Number: 12733
Loading