Can evolution strategies be enhanced without randomness? This paper introduces two powerful methods for mutation distribution adaptation: derandomization and cumulation. It reviews the shortcomings of mutative strategy parameter control and explores two derandomization levels, establishing fundamental demands for self-adaptation in normal mutation distributions. The paper demonstrates that using arbitrary normal mutation distributions is equivalent to implementing a general linear problem encoding. This rigorously pursued approach results in a completely derandomized self-adaptation scheme, adapting arbitrary normal mutation distributions and meeting previously stated demands. Further improved by cumulation, utilizing an evolution path over single search steps, the developed Covariance Matrix Adaptation (CMA) evolution strategy exhibits local and global search properties, with performance comparable to existing methods on perfectly scaled functions. Simulations reveal significant speed improvements on badly scaled, non-separable functions, indicating that CMA dramatically enhances evolution strategies for complex optimization problems.
Published in Evolutionary Computation, this paper fits squarely within the journal's focus on computational methods inspired by natural evolution. The research on derandomized self-adaptation and covariance matrix adaptation directly addresses the journal's key themes of algorithm optimization and adaptive search strategies. The number of recent citations points to a substantial impact on the field.