Density-Based Population Initialization Strategy for Continuous Optimization

The population initialization is the first and crucial step in many swarm intelligence and evolutionary algorithms. In this paper, we propose a new density-based population initialization strategy, which is concerned about both uniformity and randomness of the initial population. In experiments, first, the empty space statistic is adopted to indicate its favorable uniformity. Then, the proposed strategy is used to generate an initial population for CMA-ES, and compared with typical initialization strategies over the CEC-2013 multimodal optimization benchmark. The experimental results demonstrate that the density-based initialization strategy could generate more uniform distribution than the random strategy, and such a strategy is beneficial to evolutionary multimodal optimization.