Parallel Cooperative Co-evolution Based Particle Swarm Optimization Algorithm for Solving Conditional Nonlinear Optimal Perturbation

Abstract

Conditional nonlinear optimal perturbation (CNOP) is proposed to study the predictability of numerical weather and climate prediction. Recent researches show that evolutionary algorithms (EAs) could solve CNOP efficiently, such as SAEP and PCAGA. Both of them use dimension reduction methods with EAs to solve CNOP. But these methods always need large scale data samples and their data information are usually incomplete, which sometimes may cause the result unsatisfactory. Another way is to use cooperative co-evolution (CC) method, it adopts multi populations to change the mode of traditional searching optimum solutions. The CC method is applied in the original solution space which could avoid the defects that dimension reduction method has. In this paper, we propose cooperative co-evolution based particle swarm optimization algorithm (CCPSO) for solving CNOP. In our method, we make improvements on PSO with tabu search algorithm. Then we parallelize our method with MPI (PCCPSO). To demonstrate the validity, we compare our method with adjoint-based method, SAEP and PCAGA in ZC model. Experimental results of CNOP magnitudes and patterns show PCCPSO has the satisfactory results that are approximate to the adjoint-based method and better than SAEP and PCAGA. The time consumption of PCCPSO is about 5 min. It is approximate to the adjoint-based method with 15 initial guess fields and faster than SAEP and PCAGA. Our method can reach the speedup of 7.6 times with 12 CPU cores.