面向数值优化的迁移演化算法及其应用

In the field of evolutionary computation, numerical optimization by using evolutionary algorithms (EAs) obtains great attention in the last decades. However, in evolutionary numerical optimization, almost all approaches start their search from scratch or the ground-zero state of knowledge. This may lead to EAs converge slowly when solving real-world problems. A major impediment of existing evolutionary algorithms for numerical optimization problems (NOPs) in the literature is that the apparent lack of automated useful knowledge transfers and reuses from past problem solving experiences. Therefore, to fill this gap, in this proposal we will investigate the intelligent evolutionary algorithms, which combines transfer learning with EAs for numerical optimization problems, especially for the parameter extraction of different fuel cell models. The intelligent EAs model after how human solves problems through the transfers of building blocks learned from pervious problem-solving experiences in NOPs.. More specifically, we mainly focus on the following key issues: (1) Knowledge learning: How to automatically learn the knowledge from the successes and mistakes of the past problem solving experiences; (2) Similarity metric: How to define the similarity between different NOPs so as to transfer similar knowledge to newly encountered NOPs; (3) Knowledge transferring: How to combine similar knowledge and adaptively transfer to new NOPs; and (4) Algorithm framework and applications: How to design the framework of transfer learning based evolutionary algorithms for NOPs. Moreover, based on the framework, how to design enhanced evolutionary algorithms and use them for the parameter extraction of different fuel cell models. Therefore, we can believe that our research in this proposal is very important and meaningful, and it will promote the cross-research in machine learning, evolutionary computation, and fuel cells.

演化算法求解数值优化问题是当前的研究热点和研究前沿，但其主要不足之一是算法收敛慢，其中一个原因是算法每次从“零知识”开始优化新问题。相似问题间有用信息（知识）的有效重用将是解决此不足的一种有效途径。为此，本项目拟结合迁移学习思想，以燃料电池模型参数提取问题为主要应用背景，系统研究求解数值优化问题的迁移演化算法及其应用。包括：（1）知识自动学习技术，从已求解问题中自动学习有用信息，并建立知识库；（2）数值优化问题的相似度评价准则，有效评价不同问题或不同解之间的相似度，作为有用信息迁移依据；（3）知识自适应迁移技术，把相似问题的有用信息迁移到新问题中，加快新问题的求解；（4）提出求解数值优化问题的迁移演化算法，并应用于不同类型燃料电池模型参数提取问题中。本项目的研究有利于进一步推动机器学习、演化计算、燃料电池等领域的交叉研究，有望提出一系列新方法和新技术，具有重要的理论意义和应用价值。

演化算法求解数值优化问题是当前的研究热点和研究前沿，但其主要不足之一是算法收敛慢，其中一个原因是算法每次从“零知识”开始优化新问题。相似问题间有用信息（知识）的有效重用将是解决此不足的一种有效途径。为此，本项目拟结合迁移学习思想，以燃料电池模型参数提取问题为主要应用背景，系统研究求解数值优化问题的迁移演化算法及其应用。包括：（1）知识自动学习技术，从已求解问题中自动学习有用信息，并建立知识库；（2）数值优化问题的相似度评价准则，有效评价不同问题或不同解之间的相似度，作为有用信息迁移依据；（3）知识自适应迁移技术，把相似问题的有用信息迁移到新问题中，加快新问题的求解；（4）提出求解数值优化问题的迁移演化算法，并应用于不同类型燃料电池模型参数提取问题中。本项目的研究有利于进一步推动机器学习、演化计算、燃料电池等领域的交叉研究，有望提出一系列新方法和新技术，具有重要的理论意义和应用价值。通过本项目的研究，主要取得理论成果如下：（1）设计了基于分类器的多算子集成差分演化算法；（2）提出了带约束的多算子差分演化算法并应用于求解WSN能量分配问题；（3）提出了一般化加权双目标转换技术用于求解NES多根；（4）设计了知识迁移的改进智能优化技术求解燃料电池模型和太阳能电池模型参数提取问题。发表SCI期刊论文8篇，录用SCI期刊论文5篇，录用EI期刊论文2篇，发表国际会议论文3篇。