Fine-Grained Differential Evolution Algorithm for Multi-Dimensional Function Optimization

摘要

Differential Evolution (DE) algorithm executes its mutation and crossover operator over all dimensions for global optimization. For multi-dimensional function, especially high-dimensional function, the selected candidate solution may differ from its counterpart in multiple dimensions. This can deteriorate DE's intensification ability because the improvement presented in sub-dimensional space may be offset by the deterioration in the rest dimensions. To overcome this intrinsic limitation, this paper presents a fine-grained DE algorithm for multidimensional function optimization with a dimension-by-dimension progressive update and evaluation strategy. Within each dimension, metropolis rule of simulated annealing algorithm is used to decide whether to accept an updated value. This strategy can improve the intensification ability of DE algorithm and also keep it from premature convergence. Simulation experiments were performed on four typical benchmark functions, and the results show that our strategy can significantly improve the performance of DE algorithm.