文献类型：期刊文献

英文题名：Contrastive variational auto-encoder driven convergence guidance in evolutionary multitasking

作者：Wang, Ruilin[1,2]; Feng, Xiang[1,2]; Yu, Huiqun[1,2]

机构：[1]Shanghai Engn Res Ctr Smart Energy, Shanghai 200237, Peoples R China

年份：2024

卷号：163

外文期刊名：APPLIED SOFT COMPUTING

收录：;EI(收录号：20242616300303);WOS:【SCI-EXPANDED(收录号:WOS:001259436300001)】；

基金：This work is supported by the National Natural Science Foundation of China (No. 62276097) , Key Program of National Natural Science Foundation of China (No. 62136003) , National Key Research and Development Program of China (No. 2020YFB1711700) , Special Fund for Information Development of Shanghai Economic and Information Commission (No. XX-XXFZ-02-20-2463) and Scientific Research Program of Shanghai Science and Technology Commission (No. 21002411000) .r Information Development of Shanghai Economic and Information Com-mission (No. XX-XXFZ-02-20-2463) and Scientific Research Program of Shanghai Science and Technology Commission (No. 21002411000) .

语种：英文

外文关键词：Multifactorial evolutionary algorithm; Contrastive learning; Variational auto-encoder; Knowledge transfer

摘要：Knowledge transfer is at the core of the Evolutionary Multitasking (EMT) problem, as it exploits the interaction of inter-task common knowledge to accelerate task convergence. However, existing research on EMT is generally based solely on the framework of evolutionary computation, with limited integration of deep learning models. Additionally, few studies have found that utilizing deep learning models to generate individuals for transfer and guide the evolutionary trajectory may promote better convergence of algorithms. To address this research gap, we introduce the MFEA-VC (Multifactorial Evolutionary-Variational Auto-Encoder and Contrastive Learning) algorithm. Individuals are categorized based on task-label and inputted into a VAE, with sampling along feature dimensions. The VAE effectively guides the population towards better search areas by learning the latent trends of the distribution and generating transferred individuals. Simultaneously, a new training objective based on contrastive learning is proposed. This objective regulates the similarity between individuals from the same and different task-label, finely controlling individual features in the latent space. This approach makes the generated superior individuals more interpretable. To verify the superiority of MFEA-VC, we conduct comprehensive empirical studies on multi-task single-objective scenarios. We also validate the effectiveness of our improved loss function through theoretical analysis. The results demonstrate that, compared to state-of-the-art multifactorial algorithms, our method significantly enhances the global search capability during the early evolution stages, achieves excellent convergence results, and exhibits strong adaptability to heterogeneous tasks.