Dimension shifting based intelligent algorithm framework to solve conditional nonlinear optimal perturbation

Abstract

Conditional Nonlinear Optimal Perturbation (CNOP) method is an effective way to study the predictability of oceanic and climatic events. A framework combining the Feature Extraction method and Intelligent Algorithm (FEIA) is frequently used to solve CNOP because it is gradient-free and scalable at high-dimensional scales. However, the fixed latent subspace of FEIA framework makes it challenging to achieve both the quality of solutions and the solving efficiency. To overcome this bottleneck, this paper proposes Dimension Shifting based Intelligent Algorithm (DSIA) framework to solve CNOP in the large-scale model. DSIA framework adopts the dimension shifting strategy, which dynamically shifts search particles in different low-dimensional spaces. To verify the feasibility of DISA framework, we take a Regional Ocean Modeling System (ROMS) model of double-gyre variation as an experimental case. In experiments, we figure out that the selection of feature extraction method and the shifting dimension set are two influential factors for the performance of DSIA framework. Besides, in comparative experiments, DSIA framework yields better objective function values and more valid CNOP than FEIA framework. Moreover, convergence experiments demonstrate DSIA framework can solve CNOP with an appropriate number of function evaluations and has better convergence performance. In conclusion, experimental results prove that DSIA improves both quality of CNOP and solving efficiency.