计算机科学与探索 ›› 2023, Vol. 17 ›› Issue (6): 1301-1320.DOI: 10.3778/j.issn.1673-9418.2203078

• 理论·算法 • 上一篇    下一篇

自适应约束评估的代理模型辅助演化算法

魏凤凤,陈伟能   

  1. 1. 华南理工大学 计算机科学与工程学院,广州 510006
    2. 华南理工大学 大数据与智能机器人教育部重点实验室,广州 510006
  • 出版日期:2023-06-01 发布日期:2023-06-01

Surrogate-Assisted Evolutionary Algorithms with Adaptive Constraint Evaluation

WEI Fengfeng, CHEN Weineng   

  1. 1. School of Computer Science and Engineering, South China University of Technology, Guangzhou 510006, China
    2. Ministry of Education Key Laboratory for Big Data and Intelligent Robot, South China University of Technology, Guangzhou 510006, China
  • Online:2023-06-01 Published:2023-06-01

摘要: 很多现实优化问题不仅有昂贵目标也有昂贵约束,而现有求解昂贵优化问题的代理模型辅助演化算法(SAEAs)通常对候选解的所有约束进行评估,在评估次数有限的情况下,频繁评估可行域较大的约束不利于种群演化。针对这一问题,研究了求解昂贵约束优化问题的代理模型辅助算法,提出了一种自适应约束评估策略,根据种群演化情况评估可行域信息较少的约束,以节省在可行域较大的约束上的评估次数,在少量昂贵评估次数下自适应进行约束的选择及评估,更好地演化种群;为验证该策略的有效性和通用性,从两个思路设计了两种自适应约束评估的高斯过程回归模型辅助差分进化算法。这两种方法在15个约束优化测试函数中的11个取得显著优异效果;在利用时间延迟模拟昂贵评估次数的情况下,效率提升均在94%以上,其中91.67%的测试例子效率提升在98%以上。另外,这两种方法在4个工业应用问题中均取得优胜效果,表明其在昂贵工业约束优化问题中良好的应用前景。

关键词: 代理模型, 差分进化算法, 昂贵约束优化, 自适应约束评估策略

Abstract: Many real-world optimization problems have not only expensive objectives but also expensive constraints. However, most existing surrogate-assisted evolutionary algorithms (SAEAs) evaluate all constraints of the candidates. With limited number of evaluations, it is wasteful to constantly evaluate constraints whose feasible area is large. To solve this problem, this paper studies SAEAs for expensive constrained optimization problems, and proposes an adaptive constraint evaluation strategy. It can adaptively evaluate constraints with less feasible information according to the population evolution, saving expensive evaluations wasted on constraints with more feasible information. The algorithm can adaptively select and evaluate constraints within limited budget of expensive evaluation to better evolve the population. To verify the effectiveness and scalability of this strategy, this paper designs two Gaussian regression model-assisted differential evolution algorithms cooperated with the adaptive constraint evaluation strategy. Experiments demonstrate the proposed SAEAs perform better in 11 out of 15 problems. Besides, they can achieve more than 94% efficiency improvement with the time delay on evaluations. Specifically, the efficiency improvement is larger than 98% in 91.67% test cases. Experiments in 4 engineering optimization problems demonstrate that SAEAs with the adaptive constraint evaluation strategy have promising applications in real-world optimization problems.

Key words: surrogate model, differential evolution algorithm, expensive constrained optimization, adaptive constraint evaluation strategy