An adaptive dimension level adjustment framework for differential evolution

Differential evolution (DE) has been recognized as one of the most popular evolutionary algorithms. There are numerous DE variants adopting multi-operators based cooperation strategy to improve their performance, but almost all of the adopted cooperation strategies are essentially implemented at the individual level or population level, and the implementation at the dimension level are scarce. In this paper, an adaptive dimension level adjustment (ADLA) framework is designed to relieve the premature convergence or stagnation problem faced by DE algorithm, which can be easily combined with diverse DE variants. When the current optimal individual cannot get improved for a given uninterrupted iterations, ADLA framework will be triggered to select some individuals at random according to specific rule and reinitialize portion of their dimensions from a dynamic search space that adjusted by a population level macroparameter and one individual level microparameter. Moreover, ADLA framework contains two reinitialization operators with different search characteristics, and the coordination between them is executed at the dimension level, which has potential advantages in balancing the global exploration ability and local exploitation ability. Extensive comparison experiments are carried out based on IEEE CEC 2014 test platform, two basic DE algorithms and six outstanding DE variants. The experimental results demonstrate that ADLA framework can memorably enhance the performance of every DE algorithm used for comparison.