An optimizer ensemble algorithm and its application to image registration

The design of effective optimization algorithms is always a hot research topic. An optimizer ensemble where any population-based optimization algorithm can be integrated is proposed in this study. First, the optimizer ensemble framework based on ensemble learning is presented. The learning table consisting of the population members of all optimizers is constructed to share information. The maximum number of iterations is divided into several exchange iterations. Each optimizer exchanges individuals with the learning table in exchange iterations and runs independently in the other iterations. Exchange individuals are generated by a bootstrap sample from the learning table. To maintain a balance between exchange individuals and preserved individuals, the exchange number of each optimizer is adaptively assigned according to its fitness. The output is obtained by the voting approach that selects the highest ranked solution. Second, an optimizer ensemble algorithm (OEA) which combines multiple population-based optimization algorithms is proposed. The computational complexity, convergence, and diversity of OEA are analyzed. Finally, extensive experiments on benchmark functions demonstrate that OEA outperforms several state-ofthe-art algorithms. OEA is used to search the maximum mutual information in image registration. The high performance of OEA is further verified by a large number of registration results on real remote sensing images.