Improved Topological Niching for Real-Valued Global Optimization

We show how nearest-better clustering, the core component of the NBC-CMA niching evolutionary algorithm, is improved by appyling a second heuristic rule. This leads to enhanced basin identification for higher dimensional (5D to 20D) optimization problems, where the NBC-CMA has previously shown only mediocre performance compared to other niching and global optimization algorithms. The new method is integrated into a niching algorithm (NEA2) and compared to NBC-CMA and BIPOP-CMA-ES via the BBOB benchmarking suite. It performs very well on problems that enable recognizing basins at all with reasonable effort (number of basins not too high, e.g. the Gallagher problems), as expected. Beyond that point, niching is obviously not applicable any more and random restarts as done by the CMA-ES are the method of choice.