Partitioned Active Learning for Heterogeneous Systems

Active learning is a subfield of machine that focuses on improving the data collection efficiency expensive-to-evaluate systems. Especially, active integrated surrogate modeling has shown remarkable performance in computationally demanding engineering However, existence heterogeneity underlying systems may adversely affect learning. In order to improve under this regime, we propose partitioned seeks most informative design points for Gaussian process heterogeneous The proposed consists two systematic subsequent steps: global searching scheme accelerates exploration by investigating uncertain space, and local exploits circumscribed information induced GP. We also Cholesky update driven numerical remedies our address computational complexity challenge. method applied simulations real-world case studies about (i) cost-efficient automatic fuselage shape control aerospace manufacturing; (ii) optimal tribocorrosion-resistant alloys materials science. results show approach outperforms benchmark methods with respect prediction accuracy efficiency.